1 //===-- PPCISelLowering.cpp - PPC DAG Lowering Implementation -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the PPCISelLowering class.
12 //===----------------------------------------------------------------------===//
14 #include "PPCISelLowering.h"
15 #include "MCTargetDesc/PPCPredicates.h"
16 #include "PPCCallingConv.h"
17 #include "PPCCCState.h"
18 #include "PPCMachineFunctionInfo.h"
19 #include "PPCPerfectShuffle.h"
20 #include "PPCTargetMachine.h"
21 #include "PPCTargetObjectFile.h"
22 #include "llvm/ADT/STLExtras.h"
23 #include "llvm/ADT/Statistic.h"
24 #include "llvm/ADT/StringSwitch.h"
25 #include "llvm/ADT/Triple.h"
26 #include "llvm/CodeGen/CallingConvLower.h"
27 #include "llvm/CodeGen/MachineFrameInfo.h"
28 #include "llvm/CodeGen/MachineFunction.h"
29 #include "llvm/CodeGen/MachineInstrBuilder.h"
30 #include "llvm/CodeGen/MachineLoopInfo.h"
31 #include "llvm/CodeGen/MachineRegisterInfo.h"
32 #include "llvm/CodeGen/SelectionDAG.h"
33 #include "llvm/CodeGen/TargetLoweringObjectFileImpl.h"
34 #include "llvm/IR/CallingConv.h"
35 #include "llvm/IR/Constants.h"
36 #include "llvm/IR/DerivedTypes.h"
37 #include "llvm/IR/Function.h"
38 #include "llvm/IR/Intrinsics.h"
39 #include "llvm/Support/CommandLine.h"
40 #include "llvm/Support/ErrorHandling.h"
41 #include "llvm/Support/Format.h"
42 #include "llvm/Support/MathExtras.h"
43 #include "llvm/Support/raw_ostream.h"
44 #include "llvm/Target/TargetOptions.h"
49 #define DEBUG_TYPE "ppc-lowering"
51 static cl::opt<bool> DisablePPCPreinc("disable-ppc-preinc",
52 cl::desc("disable preincrement load/store generation on PPC"), cl::Hidden);
54 static cl::opt<bool> DisableILPPref("disable-ppc-ilp-pref",
55 cl::desc("disable setting the node scheduling preference to ILP on PPC"), cl::Hidden);
57 static cl::opt<bool> DisablePPCUnaligned("disable-ppc-unaligned",
58 cl::desc("disable unaligned load/store generation on PPC"), cl::Hidden);
60 static cl::opt<bool> DisableSCO("disable-ppc-sco",
61 cl::desc("disable sibling call optimization on ppc"), cl::Hidden);
63 STATISTIC(NumTailCalls, "Number of tail calls");
64 STATISTIC(NumSiblingCalls, "Number of sibling calls");
66 // FIXME: Remove this once the bug has been fixed!
67 extern cl::opt<bool> ANDIGlueBug;
69 PPCTargetLowering::PPCTargetLowering(const PPCTargetMachine &TM,
70 const PPCSubtarget &STI)
71 : TargetLowering(TM), Subtarget(STI) {
72 // Use _setjmp/_longjmp instead of setjmp/longjmp.
73 setUseUnderscoreSetJmp(true);
74 setUseUnderscoreLongJmp(true);
76 // On PPC32/64, arguments smaller than 4/8 bytes are extended, so all
77 // arguments are at least 4/8 bytes aligned.
78 bool isPPC64 = Subtarget.isPPC64();
79 setMinStackArgumentAlignment(isPPC64 ? 8:4);
81 // Set up the register classes.
82 addRegisterClass(MVT::i32, &PPC::GPRCRegClass);
83 if (!useSoftFloat()) {
84 addRegisterClass(MVT::f32, &PPC::F4RCRegClass);
85 addRegisterClass(MVT::f64, &PPC::F8RCRegClass);
88 // PowerPC has an i16 but no i8 (or i1) SEXTLOAD
89 for (MVT VT : MVT::integer_valuetypes()) {
90 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Promote);
91 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i8, Expand);
94 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
96 // PowerPC has pre-inc load and store's.
97 setIndexedLoadAction(ISD::PRE_INC, MVT::i1, Legal);
98 setIndexedLoadAction(ISD::PRE_INC, MVT::i8, Legal);
99 setIndexedLoadAction(ISD::PRE_INC, MVT::i16, Legal);
100 setIndexedLoadAction(ISD::PRE_INC, MVT::i32, Legal);
101 setIndexedLoadAction(ISD::PRE_INC, MVT::i64, Legal);
102 setIndexedLoadAction(ISD::PRE_INC, MVT::f32, Legal);
103 setIndexedLoadAction(ISD::PRE_INC, MVT::f64, Legal);
104 setIndexedStoreAction(ISD::PRE_INC, MVT::i1, Legal);
105 setIndexedStoreAction(ISD::PRE_INC, MVT::i8, Legal);
106 setIndexedStoreAction(ISD::PRE_INC, MVT::i16, Legal);
107 setIndexedStoreAction(ISD::PRE_INC, MVT::i32, Legal);
108 setIndexedStoreAction(ISD::PRE_INC, MVT::i64, Legal);
109 setIndexedStoreAction(ISD::PRE_INC, MVT::f32, Legal);
110 setIndexedStoreAction(ISD::PRE_INC, MVT::f64, Legal);
112 if (Subtarget.useCRBits()) {
113 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
115 if (isPPC64 || Subtarget.hasFPCVT()) {
116 setOperationAction(ISD::SINT_TO_FP, MVT::i1, Promote);
117 AddPromotedToType (ISD::SINT_TO_FP, MVT::i1,
118 isPPC64 ? MVT::i64 : MVT::i32);
119 setOperationAction(ISD::UINT_TO_FP, MVT::i1, Promote);
120 AddPromotedToType(ISD::UINT_TO_FP, MVT::i1,
121 isPPC64 ? MVT::i64 : MVT::i32);
123 setOperationAction(ISD::SINT_TO_FP, MVT::i1, Custom);
124 setOperationAction(ISD::UINT_TO_FP, MVT::i1, Custom);
127 // PowerPC does not support direct load / store of condition registers
128 setOperationAction(ISD::LOAD, MVT::i1, Custom);
129 setOperationAction(ISD::STORE, MVT::i1, Custom);
131 // FIXME: Remove this once the ANDI glue bug is fixed:
133 setOperationAction(ISD::TRUNCATE, MVT::i1, Custom);
135 for (MVT VT : MVT::integer_valuetypes()) {
136 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Promote);
137 setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::i1, Promote);
138 setTruncStoreAction(VT, MVT::i1, Expand);
141 addRegisterClass(MVT::i1, &PPC::CRBITRCRegClass);
144 // This is used in the ppcf128->int sequence. Note it has different semantics
145 // from FP_ROUND: that rounds to nearest, this rounds to zero.
146 setOperationAction(ISD::FP_ROUND_INREG, MVT::ppcf128, Custom);
148 // We do not currently implement these libm ops for PowerPC.
149 setOperationAction(ISD::FFLOOR, MVT::ppcf128, Expand);
150 setOperationAction(ISD::FCEIL, MVT::ppcf128, Expand);
151 setOperationAction(ISD::FTRUNC, MVT::ppcf128, Expand);
152 setOperationAction(ISD::FRINT, MVT::ppcf128, Expand);
153 setOperationAction(ISD::FNEARBYINT, MVT::ppcf128, Expand);
154 setOperationAction(ISD::FREM, MVT::ppcf128, Expand);
156 // PowerPC has no SREM/UREM instructions
157 setOperationAction(ISD::SREM, MVT::i32, Expand);
158 setOperationAction(ISD::UREM, MVT::i32, Expand);
159 setOperationAction(ISD::SREM, MVT::i64, Expand);
160 setOperationAction(ISD::UREM, MVT::i64, Expand);
162 // Don't use SMUL_LOHI/UMUL_LOHI or SDIVREM/UDIVREM to lower SREM/UREM.
163 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
164 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
165 setOperationAction(ISD::UMUL_LOHI, MVT::i64, Expand);
166 setOperationAction(ISD::SMUL_LOHI, MVT::i64, Expand);
167 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
168 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
169 setOperationAction(ISD::UDIVREM, MVT::i64, Expand);
170 setOperationAction(ISD::SDIVREM, MVT::i64, Expand);
172 // We don't support sin/cos/sqrt/fmod/pow
173 setOperationAction(ISD::FSIN , MVT::f64, Expand);
174 setOperationAction(ISD::FCOS , MVT::f64, Expand);
175 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
176 setOperationAction(ISD::FREM , MVT::f64, Expand);
177 setOperationAction(ISD::FPOW , MVT::f64, Expand);
178 setOperationAction(ISD::FMA , MVT::f64, Legal);
179 setOperationAction(ISD::FSIN , MVT::f32, Expand);
180 setOperationAction(ISD::FCOS , MVT::f32, Expand);
181 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
182 setOperationAction(ISD::FREM , MVT::f32, Expand);
183 setOperationAction(ISD::FPOW , MVT::f32, Expand);
184 setOperationAction(ISD::FMA , MVT::f32, Legal);
186 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
188 // If we're enabling GP optimizations, use hardware square root
189 if (!Subtarget.hasFSQRT() &&
190 !(TM.Options.UnsafeFPMath && Subtarget.hasFRSQRTE() &&
192 setOperationAction(ISD::FSQRT, MVT::f64, Expand);
194 if (!Subtarget.hasFSQRT() &&
195 !(TM.Options.UnsafeFPMath && Subtarget.hasFRSQRTES() &&
196 Subtarget.hasFRES()))
197 setOperationAction(ISD::FSQRT, MVT::f32, Expand);
199 if (Subtarget.hasFCPSGN()) {
200 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Legal);
201 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Legal);
203 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
204 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
207 if (Subtarget.hasFPRND()) {
208 setOperationAction(ISD::FFLOOR, MVT::f64, Legal);
209 setOperationAction(ISD::FCEIL, MVT::f64, Legal);
210 setOperationAction(ISD::FTRUNC, MVT::f64, Legal);
211 setOperationAction(ISD::FROUND, MVT::f64, Legal);
213 setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
214 setOperationAction(ISD::FCEIL, MVT::f32, Legal);
215 setOperationAction(ISD::FTRUNC, MVT::f32, Legal);
216 setOperationAction(ISD::FROUND, MVT::f32, Legal);
219 // PowerPC does not have BSWAP, CTPOP or CTTZ
220 setOperationAction(ISD::BSWAP, MVT::i32 , Expand);
221 setOperationAction(ISD::CTTZ , MVT::i32 , Expand);
222 setOperationAction(ISD::BSWAP, MVT::i64 , Expand);
223 setOperationAction(ISD::CTTZ , MVT::i64 , Expand);
225 if (Subtarget.hasPOPCNTD() == PPCSubtarget::POPCNTD_Fast) {
226 setOperationAction(ISD::CTPOP, MVT::i32 , Legal);
227 setOperationAction(ISD::CTPOP, MVT::i64 , Legal);
229 setOperationAction(ISD::CTPOP, MVT::i32 , Expand);
230 setOperationAction(ISD::CTPOP, MVT::i64 , Expand);
233 // PowerPC does not have ROTR
234 setOperationAction(ISD::ROTR, MVT::i32 , Expand);
235 setOperationAction(ISD::ROTR, MVT::i64 , Expand);
237 if (!Subtarget.useCRBits()) {
238 // PowerPC does not have Select
239 setOperationAction(ISD::SELECT, MVT::i32, Expand);
240 setOperationAction(ISD::SELECT, MVT::i64, Expand);
241 setOperationAction(ISD::SELECT, MVT::f32, Expand);
242 setOperationAction(ISD::SELECT, MVT::f64, Expand);
245 // PowerPC wants to turn select_cc of FP into fsel when possible.
246 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
247 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
249 // PowerPC wants to optimize integer setcc a bit
250 if (!Subtarget.useCRBits())
251 setOperationAction(ISD::SETCC, MVT::i32, Custom);
253 // PowerPC does not have BRCOND which requires SetCC
254 if (!Subtarget.useCRBits())
255 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
257 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
259 // PowerPC turns FP_TO_SINT into FCTIWZ and some load/stores.
260 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
262 // PowerPC does not have [U|S]INT_TO_FP
263 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Expand);
264 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Expand);
266 if (Subtarget.hasDirectMove() && isPPC64) {
267 setOperationAction(ISD::BITCAST, MVT::f32, Legal);
268 setOperationAction(ISD::BITCAST, MVT::i32, Legal);
269 setOperationAction(ISD::BITCAST, MVT::i64, Legal);
270 setOperationAction(ISD::BITCAST, MVT::f64, Legal);
272 setOperationAction(ISD::BITCAST, MVT::f32, Expand);
273 setOperationAction(ISD::BITCAST, MVT::i32, Expand);
274 setOperationAction(ISD::BITCAST, MVT::i64, Expand);
275 setOperationAction(ISD::BITCAST, MVT::f64, Expand);
278 // We cannot sextinreg(i1). Expand to shifts.
279 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
281 // NOTE: EH_SJLJ_SETJMP/_LONGJMP supported here is NOT intended to support
282 // SjLj exception handling but a light-weight setjmp/longjmp replacement to
283 // support continuation, user-level threading, and etc.. As a result, no
284 // other SjLj exception interfaces are implemented and please don't build
285 // your own exception handling based on them.
286 // LLVM/Clang supports zero-cost DWARF exception handling.
287 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
288 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
290 // We want to legalize GlobalAddress and ConstantPool nodes into the
291 // appropriate instructions to materialize the address.
292 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
293 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
294 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
295 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
296 setOperationAction(ISD::JumpTable, MVT::i32, Custom);
297 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
298 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
299 setOperationAction(ISD::BlockAddress, MVT::i64, Custom);
300 setOperationAction(ISD::ConstantPool, MVT::i64, Custom);
301 setOperationAction(ISD::JumpTable, MVT::i64, Custom);
304 setOperationAction(ISD::TRAP, MVT::Other, Legal);
306 // TRAMPOLINE is custom lowered.
307 setOperationAction(ISD::INIT_TRAMPOLINE, MVT::Other, Custom);
308 setOperationAction(ISD::ADJUST_TRAMPOLINE, MVT::Other, Custom);
310 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
311 setOperationAction(ISD::VASTART , MVT::Other, Custom);
313 if (Subtarget.isSVR4ABI()) {
315 // VAARG always uses double-word chunks, so promote anything smaller.
316 setOperationAction(ISD::VAARG, MVT::i1, Promote);
317 AddPromotedToType (ISD::VAARG, MVT::i1, MVT::i64);
318 setOperationAction(ISD::VAARG, MVT::i8, Promote);
319 AddPromotedToType (ISD::VAARG, MVT::i8, MVT::i64);
320 setOperationAction(ISD::VAARG, MVT::i16, Promote);
321 AddPromotedToType (ISD::VAARG, MVT::i16, MVT::i64);
322 setOperationAction(ISD::VAARG, MVT::i32, Promote);
323 AddPromotedToType (ISD::VAARG, MVT::i32, MVT::i64);
324 setOperationAction(ISD::VAARG, MVT::Other, Expand);
326 // VAARG is custom lowered with the 32-bit SVR4 ABI.
327 setOperationAction(ISD::VAARG, MVT::Other, Custom);
328 setOperationAction(ISD::VAARG, MVT::i64, Custom);
331 setOperationAction(ISD::VAARG, MVT::Other, Expand);
333 if (Subtarget.isSVR4ABI() && !isPPC64)
334 // VACOPY is custom lowered with the 32-bit SVR4 ABI.
335 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
337 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
339 // Use the default implementation.
340 setOperationAction(ISD::VAEND , MVT::Other, Expand);
341 setOperationAction(ISD::STACKSAVE , MVT::Other, Expand);
342 setOperationAction(ISD::STACKRESTORE , MVT::Other, Custom);
343 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32 , Custom);
344 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64 , Custom);
345 setOperationAction(ISD::GET_DYNAMIC_AREA_OFFSET, MVT::i32, Custom);
346 setOperationAction(ISD::GET_DYNAMIC_AREA_OFFSET, MVT::i64, Custom);
347 setOperationAction(ISD::EH_DWARF_CFA, MVT::i32, Custom);
348 setOperationAction(ISD::EH_DWARF_CFA, MVT::i64, Custom);
350 // We want to custom lower some of our intrinsics.
351 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
353 // To handle counter-based loop conditions.
354 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::i1, Custom);
356 // Comparisons that require checking two conditions.
357 setCondCodeAction(ISD::SETULT, MVT::f32, Expand);
358 setCondCodeAction(ISD::SETULT, MVT::f64, Expand);
359 setCondCodeAction(ISD::SETUGT, MVT::f32, Expand);
360 setCondCodeAction(ISD::SETUGT, MVT::f64, Expand);
361 setCondCodeAction(ISD::SETUEQ, MVT::f32, Expand);
362 setCondCodeAction(ISD::SETUEQ, MVT::f64, Expand);
363 setCondCodeAction(ISD::SETOGE, MVT::f32, Expand);
364 setCondCodeAction(ISD::SETOGE, MVT::f64, Expand);
365 setCondCodeAction(ISD::SETOLE, MVT::f32, Expand);
366 setCondCodeAction(ISD::SETOLE, MVT::f64, Expand);
367 setCondCodeAction(ISD::SETONE, MVT::f32, Expand);
368 setCondCodeAction(ISD::SETONE, MVT::f64, Expand);
370 if (Subtarget.has64BitSupport()) {
371 // They also have instructions for converting between i64 and fp.
372 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
373 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Expand);
374 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
375 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Expand);
376 // This is just the low 32 bits of a (signed) fp->i64 conversion.
377 // We cannot do this with Promote because i64 is not a legal type.
378 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
380 if (Subtarget.hasLFIWAX() || Subtarget.isPPC64())
381 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
383 // PowerPC does not have FP_TO_UINT on 32-bit implementations.
384 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Expand);
387 // With the instructions enabled under FPCVT, we can do everything.
388 if (Subtarget.hasFPCVT()) {
389 if (Subtarget.has64BitSupport()) {
390 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
391 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Custom);
392 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
393 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Custom);
396 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
397 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
398 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
399 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
402 if (Subtarget.use64BitRegs()) {
403 // 64-bit PowerPC implementations can support i64 types directly
404 addRegisterClass(MVT::i64, &PPC::G8RCRegClass);
405 // BUILD_PAIR can't be handled natively, and should be expanded to shl/or
406 setOperationAction(ISD::BUILD_PAIR, MVT::i64, Expand);
407 // 64-bit PowerPC wants to expand i128 shifts itself.
408 setOperationAction(ISD::SHL_PARTS, MVT::i64, Custom);
409 setOperationAction(ISD::SRA_PARTS, MVT::i64, Custom);
410 setOperationAction(ISD::SRL_PARTS, MVT::i64, Custom);
412 // 32-bit PowerPC wants to expand i64 shifts itself.
413 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
414 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
415 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
418 if (Subtarget.hasAltivec()) {
419 // First set operation action for all vector types to expand. Then we
420 // will selectively turn on ones that can be effectively codegen'd.
421 for (MVT VT : MVT::vector_valuetypes()) {
422 // add/sub are legal for all supported vector VT's.
423 setOperationAction(ISD::ADD, VT, Legal);
424 setOperationAction(ISD::SUB, VT, Legal);
426 // Vector instructions introduced in P8
427 if (Subtarget.hasP8Altivec() && (VT.SimpleTy != MVT::v1i128)) {
428 setOperationAction(ISD::CTPOP, VT, Legal);
429 setOperationAction(ISD::CTLZ, VT, Legal);
432 setOperationAction(ISD::CTPOP, VT, Expand);
433 setOperationAction(ISD::CTLZ, VT, Expand);
436 // We promote all shuffles to v16i8.
437 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Promote);
438 AddPromotedToType (ISD::VECTOR_SHUFFLE, VT, MVT::v16i8);
440 // We promote all non-typed operations to v4i32.
441 setOperationAction(ISD::AND , VT, Promote);
442 AddPromotedToType (ISD::AND , VT, MVT::v4i32);
443 setOperationAction(ISD::OR , VT, Promote);
444 AddPromotedToType (ISD::OR , VT, MVT::v4i32);
445 setOperationAction(ISD::XOR , VT, Promote);
446 AddPromotedToType (ISD::XOR , VT, MVT::v4i32);
447 setOperationAction(ISD::LOAD , VT, Promote);
448 AddPromotedToType (ISD::LOAD , VT, MVT::v4i32);
449 setOperationAction(ISD::SELECT, VT, Promote);
450 AddPromotedToType (ISD::SELECT, VT, MVT::v4i32);
451 setOperationAction(ISD::SELECT_CC, VT, Promote);
452 AddPromotedToType (ISD::SELECT_CC, VT, MVT::v4i32);
453 setOperationAction(ISD::STORE, VT, Promote);
454 AddPromotedToType (ISD::STORE, VT, MVT::v4i32);
456 // No other operations are legal.
457 setOperationAction(ISD::MUL , VT, Expand);
458 setOperationAction(ISD::SDIV, VT, Expand);
459 setOperationAction(ISD::SREM, VT, Expand);
460 setOperationAction(ISD::UDIV, VT, Expand);
461 setOperationAction(ISD::UREM, VT, Expand);
462 setOperationAction(ISD::FDIV, VT, Expand);
463 setOperationAction(ISD::FREM, VT, Expand);
464 setOperationAction(ISD::FNEG, VT, Expand);
465 setOperationAction(ISD::FSQRT, VT, Expand);
466 setOperationAction(ISD::FLOG, VT, Expand);
467 setOperationAction(ISD::FLOG10, VT, Expand);
468 setOperationAction(ISD::FLOG2, VT, Expand);
469 setOperationAction(ISD::FEXP, VT, Expand);
470 setOperationAction(ISD::FEXP2, VT, Expand);
471 setOperationAction(ISD::FSIN, VT, Expand);
472 setOperationAction(ISD::FCOS, VT, Expand);
473 setOperationAction(ISD::FABS, VT, Expand);
474 setOperationAction(ISD::FPOWI, VT, Expand);
475 setOperationAction(ISD::FFLOOR, VT, Expand);
476 setOperationAction(ISD::FCEIL, VT, Expand);
477 setOperationAction(ISD::FTRUNC, VT, Expand);
478 setOperationAction(ISD::FRINT, VT, Expand);
479 setOperationAction(ISD::FNEARBYINT, VT, Expand);
480 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Expand);
481 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Expand);
482 setOperationAction(ISD::BUILD_VECTOR, VT, Expand);
483 setOperationAction(ISD::MULHU, VT, Expand);
484 setOperationAction(ISD::MULHS, VT, Expand);
485 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
486 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
487 setOperationAction(ISD::UDIVREM, VT, Expand);
488 setOperationAction(ISD::SDIVREM, VT, Expand);
489 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Expand);
490 setOperationAction(ISD::FPOW, VT, Expand);
491 setOperationAction(ISD::BSWAP, VT, Expand);
492 setOperationAction(ISD::CTTZ, VT, Expand);
493 setOperationAction(ISD::VSELECT, VT, Expand);
494 setOperationAction(ISD::SIGN_EXTEND_INREG, VT, Expand);
495 setOperationAction(ISD::ROTL, VT, Expand);
496 setOperationAction(ISD::ROTR, VT, Expand);
498 for (MVT InnerVT : MVT::vector_valuetypes()) {
499 setTruncStoreAction(VT, InnerVT, Expand);
500 setLoadExtAction(ISD::SEXTLOAD, VT, InnerVT, Expand);
501 setLoadExtAction(ISD::ZEXTLOAD, VT, InnerVT, Expand);
502 setLoadExtAction(ISD::EXTLOAD, VT, InnerVT, Expand);
506 // We can custom expand all VECTOR_SHUFFLEs to VPERM, others we can handle
507 // with merges, splats, etc.
508 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16i8, Custom);
510 setOperationAction(ISD::AND , MVT::v4i32, Legal);
511 setOperationAction(ISD::OR , MVT::v4i32, Legal);
512 setOperationAction(ISD::XOR , MVT::v4i32, Legal);
513 setOperationAction(ISD::LOAD , MVT::v4i32, Legal);
514 setOperationAction(ISD::SELECT, MVT::v4i32,
515 Subtarget.useCRBits() ? Legal : Expand);
516 setOperationAction(ISD::STORE , MVT::v4i32, Legal);
517 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
518 setOperationAction(ISD::FP_TO_UINT, MVT::v4i32, Legal);
519 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
520 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Legal);
521 setOperationAction(ISD::FFLOOR, MVT::v4f32, Legal);
522 setOperationAction(ISD::FCEIL, MVT::v4f32, Legal);
523 setOperationAction(ISD::FTRUNC, MVT::v4f32, Legal);
524 setOperationAction(ISD::FNEARBYINT, MVT::v4f32, Legal);
526 addRegisterClass(MVT::v4f32, &PPC::VRRCRegClass);
527 addRegisterClass(MVT::v4i32, &PPC::VRRCRegClass);
528 addRegisterClass(MVT::v8i16, &PPC::VRRCRegClass);
529 addRegisterClass(MVT::v16i8, &PPC::VRRCRegClass);
531 setOperationAction(ISD::MUL, MVT::v4f32, Legal);
532 setOperationAction(ISD::FMA, MVT::v4f32, Legal);
534 if (TM.Options.UnsafeFPMath || Subtarget.hasVSX()) {
535 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
536 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
539 if (Subtarget.hasP8Altivec())
540 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
542 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
544 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
545 setOperationAction(ISD::MUL, MVT::v16i8, Custom);
547 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4f32, Custom);
548 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i32, Custom);
550 setOperationAction(ISD::BUILD_VECTOR, MVT::v16i8, Custom);
551 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i16, Custom);
552 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i32, Custom);
553 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
555 // Altivec does not contain unordered floating-point compare instructions
556 setCondCodeAction(ISD::SETUO, MVT::v4f32, Expand);
557 setCondCodeAction(ISD::SETUEQ, MVT::v4f32, Expand);
558 setCondCodeAction(ISD::SETO, MVT::v4f32, Expand);
559 setCondCodeAction(ISD::SETONE, MVT::v4f32, Expand);
561 if (Subtarget.hasVSX()) {
562 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2f64, Legal);
563 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Legal);
564 if (Subtarget.hasP8Vector()) {
565 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4f32, Legal);
566 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Legal);
568 if (Subtarget.hasDirectMove() && isPPC64) {
569 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Legal);
570 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Legal);
571 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i32, Legal);
572 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2i64, Legal);
573 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Legal);
574 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Legal);
575 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Legal);
576 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Legal);
578 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Legal);
580 setOperationAction(ISD::FFLOOR, MVT::v2f64, Legal);
581 setOperationAction(ISD::FCEIL, MVT::v2f64, Legal);
582 setOperationAction(ISD::FTRUNC, MVT::v2f64, Legal);
583 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Legal);
584 setOperationAction(ISD::FROUND, MVT::v2f64, Legal);
586 setOperationAction(ISD::FROUND, MVT::v4f32, Legal);
588 setOperationAction(ISD::MUL, MVT::v2f64, Legal);
589 setOperationAction(ISD::FMA, MVT::v2f64, Legal);
591 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
592 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
594 setOperationAction(ISD::VSELECT, MVT::v16i8, Legal);
595 setOperationAction(ISD::VSELECT, MVT::v8i16, Legal);
596 setOperationAction(ISD::VSELECT, MVT::v4i32, Legal);
597 setOperationAction(ISD::VSELECT, MVT::v4f32, Legal);
598 setOperationAction(ISD::VSELECT, MVT::v2f64, Legal);
600 // Share the Altivec comparison restrictions.
601 setCondCodeAction(ISD::SETUO, MVT::v2f64, Expand);
602 setCondCodeAction(ISD::SETUEQ, MVT::v2f64, Expand);
603 setCondCodeAction(ISD::SETO, MVT::v2f64, Expand);
604 setCondCodeAction(ISD::SETONE, MVT::v2f64, Expand);
606 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
607 setOperationAction(ISD::STORE, MVT::v2f64, Legal);
609 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Legal);
611 if (Subtarget.hasP8Vector())
612 addRegisterClass(MVT::f32, &PPC::VSSRCRegClass);
614 addRegisterClass(MVT::f64, &PPC::VSFRCRegClass);
616 addRegisterClass(MVT::v4i32, &PPC::VSRCRegClass);
617 addRegisterClass(MVT::v4f32, &PPC::VSRCRegClass);
618 addRegisterClass(MVT::v2f64, &PPC::VSRCRegClass);
620 if (Subtarget.hasP8Altivec()) {
621 setOperationAction(ISD::SHL, MVT::v2i64, Legal);
622 setOperationAction(ISD::SRA, MVT::v2i64, Legal);
623 setOperationAction(ISD::SRL, MVT::v2i64, Legal);
625 setOperationAction(ISD::SETCC, MVT::v2i64, Legal);
628 setOperationAction(ISD::SHL, MVT::v2i64, Expand);
629 setOperationAction(ISD::SRA, MVT::v2i64, Expand);
630 setOperationAction(ISD::SRL, MVT::v2i64, Expand);
632 setOperationAction(ISD::SETCC, MVT::v2i64, Custom);
634 // VSX v2i64 only supports non-arithmetic operations.
635 setOperationAction(ISD::ADD, MVT::v2i64, Expand);
636 setOperationAction(ISD::SUB, MVT::v2i64, Expand);
639 setOperationAction(ISD::LOAD, MVT::v2i64, Promote);
640 AddPromotedToType (ISD::LOAD, MVT::v2i64, MVT::v2f64);
641 setOperationAction(ISD::STORE, MVT::v2i64, Promote);
642 AddPromotedToType (ISD::STORE, MVT::v2i64, MVT::v2f64);
644 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Legal);
646 setOperationAction(ISD::SINT_TO_FP, MVT::v2i64, Legal);
647 setOperationAction(ISD::UINT_TO_FP, MVT::v2i64, Legal);
648 setOperationAction(ISD::FP_TO_SINT, MVT::v2i64, Legal);
649 setOperationAction(ISD::FP_TO_UINT, MVT::v2i64, Legal);
651 // Vector operation legalization checks the result type of
652 // SIGN_EXTEND_INREG, overall legalization checks the inner type.
653 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i64, Legal);
654 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i32, Legal);
655 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i16, Custom);
656 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i8, Custom);
658 setOperationAction(ISD::FNEG, MVT::v4f32, Legal);
659 setOperationAction(ISD::FNEG, MVT::v2f64, Legal);
660 setOperationAction(ISD::FABS, MVT::v4f32, Legal);
661 setOperationAction(ISD::FABS, MVT::v2f64, Legal);
663 addRegisterClass(MVT::v2i64, &PPC::VSRCRegClass);
666 if (Subtarget.hasP8Altivec()) {
667 addRegisterClass(MVT::v2i64, &PPC::VRRCRegClass);
668 addRegisterClass(MVT::v1i128, &PPC::VRRCRegClass);
670 if (Subtarget.hasP9Vector()) {
671 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Legal);
672 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Legal);
676 if (Subtarget.hasQPX()) {
677 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
678 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
679 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
680 setOperationAction(ISD::FREM, MVT::v4f64, Expand);
682 setOperationAction(ISD::FCOPYSIGN, MVT::v4f64, Legal);
683 setOperationAction(ISD::FGETSIGN, MVT::v4f64, Expand);
685 setOperationAction(ISD::LOAD , MVT::v4f64, Custom);
686 setOperationAction(ISD::STORE , MVT::v4f64, Custom);
688 setTruncStoreAction(MVT::v4f64, MVT::v4f32, Custom);
689 setLoadExtAction(ISD::EXTLOAD, MVT::v4f64, MVT::v4f32, Custom);
691 if (!Subtarget.useCRBits())
692 setOperationAction(ISD::SELECT, MVT::v4f64, Expand);
693 setOperationAction(ISD::VSELECT, MVT::v4f64, Legal);
695 setOperationAction(ISD::EXTRACT_VECTOR_ELT , MVT::v4f64, Legal);
696 setOperationAction(ISD::INSERT_VECTOR_ELT , MVT::v4f64, Expand);
697 setOperationAction(ISD::CONCAT_VECTORS , MVT::v4f64, Expand);
698 setOperationAction(ISD::EXTRACT_SUBVECTOR , MVT::v4f64, Expand);
699 setOperationAction(ISD::VECTOR_SHUFFLE , MVT::v4f64, Custom);
700 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4f64, Legal);
701 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f64, Custom);
703 setOperationAction(ISD::FP_TO_SINT , MVT::v4f64, Legal);
704 setOperationAction(ISD::FP_TO_UINT , MVT::v4f64, Expand);
706 setOperationAction(ISD::FP_ROUND , MVT::v4f32, Legal);
707 setOperationAction(ISD::FP_ROUND_INREG , MVT::v4f32, Expand);
708 setOperationAction(ISD::FP_EXTEND, MVT::v4f64, Legal);
710 setOperationAction(ISD::FNEG , MVT::v4f64, Legal);
711 setOperationAction(ISD::FABS , MVT::v4f64, Legal);
712 setOperationAction(ISD::FSIN , MVT::v4f64, Expand);
713 setOperationAction(ISD::FCOS , MVT::v4f64, Expand);
714 setOperationAction(ISD::FPOWI , MVT::v4f64, Expand);
715 setOperationAction(ISD::FPOW , MVT::v4f64, Expand);
716 setOperationAction(ISD::FLOG , MVT::v4f64, Expand);
717 setOperationAction(ISD::FLOG2 , MVT::v4f64, Expand);
718 setOperationAction(ISD::FLOG10 , MVT::v4f64, Expand);
719 setOperationAction(ISD::FEXP , MVT::v4f64, Expand);
720 setOperationAction(ISD::FEXP2 , MVT::v4f64, Expand);
722 setOperationAction(ISD::FMINNUM, MVT::v4f64, Legal);
723 setOperationAction(ISD::FMAXNUM, MVT::v4f64, Legal);
725 setIndexedLoadAction(ISD::PRE_INC, MVT::v4f64, Legal);
726 setIndexedStoreAction(ISD::PRE_INC, MVT::v4f64, Legal);
728 addRegisterClass(MVT::v4f64, &PPC::QFRCRegClass);
730 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
731 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
732 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
733 setOperationAction(ISD::FREM, MVT::v4f32, Expand);
735 setOperationAction(ISD::FCOPYSIGN, MVT::v4f32, Legal);
736 setOperationAction(ISD::FGETSIGN, MVT::v4f32, Expand);
738 setOperationAction(ISD::LOAD , MVT::v4f32, Custom);
739 setOperationAction(ISD::STORE , MVT::v4f32, Custom);
741 if (!Subtarget.useCRBits())
742 setOperationAction(ISD::SELECT, MVT::v4f32, Expand);
743 setOperationAction(ISD::VSELECT, MVT::v4f32, Legal);
745 setOperationAction(ISD::EXTRACT_VECTOR_ELT , MVT::v4f32, Legal);
746 setOperationAction(ISD::INSERT_VECTOR_ELT , MVT::v4f32, Expand);
747 setOperationAction(ISD::CONCAT_VECTORS , MVT::v4f32, Expand);
748 setOperationAction(ISD::EXTRACT_SUBVECTOR , MVT::v4f32, Expand);
749 setOperationAction(ISD::VECTOR_SHUFFLE , MVT::v4f32, Custom);
750 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4f32, Legal);
751 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
753 setOperationAction(ISD::FP_TO_SINT , MVT::v4f32, Legal);
754 setOperationAction(ISD::FP_TO_UINT , MVT::v4f32, Expand);
756 setOperationAction(ISD::FNEG , MVT::v4f32, Legal);
757 setOperationAction(ISD::FABS , MVT::v4f32, Legal);
758 setOperationAction(ISD::FSIN , MVT::v4f32, Expand);
759 setOperationAction(ISD::FCOS , MVT::v4f32, Expand);
760 setOperationAction(ISD::FPOWI , MVT::v4f32, Expand);
761 setOperationAction(ISD::FPOW , MVT::v4f32, Expand);
762 setOperationAction(ISD::FLOG , MVT::v4f32, Expand);
763 setOperationAction(ISD::FLOG2 , MVT::v4f32, Expand);
764 setOperationAction(ISD::FLOG10 , MVT::v4f32, Expand);
765 setOperationAction(ISD::FEXP , MVT::v4f32, Expand);
766 setOperationAction(ISD::FEXP2 , MVT::v4f32, Expand);
768 setOperationAction(ISD::FMINNUM, MVT::v4f32, Legal);
769 setOperationAction(ISD::FMAXNUM, MVT::v4f32, Legal);
771 setIndexedLoadAction(ISD::PRE_INC, MVT::v4f32, Legal);
772 setIndexedStoreAction(ISD::PRE_INC, MVT::v4f32, Legal);
774 addRegisterClass(MVT::v4f32, &PPC::QSRCRegClass);
776 setOperationAction(ISD::AND , MVT::v4i1, Legal);
777 setOperationAction(ISD::OR , MVT::v4i1, Legal);
778 setOperationAction(ISD::XOR , MVT::v4i1, Legal);
780 if (!Subtarget.useCRBits())
781 setOperationAction(ISD::SELECT, MVT::v4i1, Expand);
782 setOperationAction(ISD::VSELECT, MVT::v4i1, Legal);
784 setOperationAction(ISD::LOAD , MVT::v4i1, Custom);
785 setOperationAction(ISD::STORE , MVT::v4i1, Custom);
787 setOperationAction(ISD::EXTRACT_VECTOR_ELT , MVT::v4i1, Custom);
788 setOperationAction(ISD::INSERT_VECTOR_ELT , MVT::v4i1, Expand);
789 setOperationAction(ISD::CONCAT_VECTORS , MVT::v4i1, Expand);
790 setOperationAction(ISD::EXTRACT_SUBVECTOR , MVT::v4i1, Expand);
791 setOperationAction(ISD::VECTOR_SHUFFLE , MVT::v4i1, Custom);
792 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i1, Expand);
793 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i1, Custom);
795 setOperationAction(ISD::SINT_TO_FP, MVT::v4i1, Custom);
796 setOperationAction(ISD::UINT_TO_FP, MVT::v4i1, Custom);
798 addRegisterClass(MVT::v4i1, &PPC::QBRCRegClass);
800 setOperationAction(ISD::FFLOOR, MVT::v4f64, Legal);
801 setOperationAction(ISD::FCEIL, MVT::v4f64, Legal);
802 setOperationAction(ISD::FTRUNC, MVT::v4f64, Legal);
803 setOperationAction(ISD::FROUND, MVT::v4f64, Legal);
805 setOperationAction(ISD::FFLOOR, MVT::v4f32, Legal);
806 setOperationAction(ISD::FCEIL, MVT::v4f32, Legal);
807 setOperationAction(ISD::FTRUNC, MVT::v4f32, Legal);
808 setOperationAction(ISD::FROUND, MVT::v4f32, Legal);
810 setOperationAction(ISD::FNEARBYINT, MVT::v4f64, Expand);
811 setOperationAction(ISD::FNEARBYINT, MVT::v4f32, Expand);
813 // These need to set FE_INEXACT, and so cannot be vectorized here.
814 setOperationAction(ISD::FRINT, MVT::v4f64, Expand);
815 setOperationAction(ISD::FRINT, MVT::v4f32, Expand);
817 if (TM.Options.UnsafeFPMath) {
818 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
819 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
821 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
822 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
824 setOperationAction(ISD::FDIV, MVT::v4f64, Expand);
825 setOperationAction(ISD::FSQRT, MVT::v4f64, Expand);
827 setOperationAction(ISD::FDIV, MVT::v4f32, Expand);
828 setOperationAction(ISD::FSQRT, MVT::v4f32, Expand);
832 if (Subtarget.has64BitSupport())
833 setOperationAction(ISD::PREFETCH, MVT::Other, Legal);
835 setOperationAction(ISD::READCYCLECOUNTER, MVT::i64, isPPC64 ? Legal : Custom);
838 setOperationAction(ISD::ATOMIC_LOAD, MVT::i64, Expand);
839 setOperationAction(ISD::ATOMIC_STORE, MVT::i64, Expand);
842 setBooleanContents(ZeroOrOneBooleanContent);
844 if (Subtarget.hasAltivec()) {
845 // Altivec instructions set fields to all zeros or all ones.
846 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
850 // These libcalls are not available in 32-bit.
851 setLibcallName(RTLIB::SHL_I128, nullptr);
852 setLibcallName(RTLIB::SRL_I128, nullptr);
853 setLibcallName(RTLIB::SRA_I128, nullptr);
856 setStackPointerRegisterToSaveRestore(isPPC64 ? PPC::X1 : PPC::R1);
858 // We have target-specific dag combine patterns for the following nodes:
859 setTargetDAGCombine(ISD::SINT_TO_FP);
860 setTargetDAGCombine(ISD::BUILD_VECTOR);
861 if (Subtarget.hasFPCVT())
862 setTargetDAGCombine(ISD::UINT_TO_FP);
863 setTargetDAGCombine(ISD::LOAD);
864 setTargetDAGCombine(ISD::STORE);
865 setTargetDAGCombine(ISD::BR_CC);
866 if (Subtarget.useCRBits())
867 setTargetDAGCombine(ISD::BRCOND);
868 setTargetDAGCombine(ISD::BSWAP);
869 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
870 setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN);
871 setTargetDAGCombine(ISD::INTRINSIC_VOID);
873 setTargetDAGCombine(ISD::SIGN_EXTEND);
874 setTargetDAGCombine(ISD::ZERO_EXTEND);
875 setTargetDAGCombine(ISD::ANY_EXTEND);
877 if (Subtarget.useCRBits()) {
878 setTargetDAGCombine(ISD::TRUNCATE);
879 setTargetDAGCombine(ISD::SETCC);
880 setTargetDAGCombine(ISD::SELECT_CC);
883 // Use reciprocal estimates.
884 if (TM.Options.UnsafeFPMath) {
885 setTargetDAGCombine(ISD::FDIV);
886 setTargetDAGCombine(ISD::FSQRT);
889 // Darwin long double math library functions have $LDBL128 appended.
890 if (Subtarget.isDarwin()) {
891 setLibcallName(RTLIB::COS_PPCF128, "cosl$LDBL128");
892 setLibcallName(RTLIB::POW_PPCF128, "powl$LDBL128");
893 setLibcallName(RTLIB::REM_PPCF128, "fmodl$LDBL128");
894 setLibcallName(RTLIB::SIN_PPCF128, "sinl$LDBL128");
895 setLibcallName(RTLIB::SQRT_PPCF128, "sqrtl$LDBL128");
896 setLibcallName(RTLIB::LOG_PPCF128, "logl$LDBL128");
897 setLibcallName(RTLIB::LOG2_PPCF128, "log2l$LDBL128");
898 setLibcallName(RTLIB::LOG10_PPCF128, "log10l$LDBL128");
899 setLibcallName(RTLIB::EXP_PPCF128, "expl$LDBL128");
900 setLibcallName(RTLIB::EXP2_PPCF128, "exp2l$LDBL128");
903 // With 32 condition bits, we don't need to sink (and duplicate) compares
904 // aggressively in CodeGenPrep.
905 if (Subtarget.useCRBits()) {
906 setHasMultipleConditionRegisters();
907 setJumpIsExpensive();
910 setMinFunctionAlignment(2);
911 if (Subtarget.isDarwin())
912 setPrefFunctionAlignment(4);
914 switch (Subtarget.getDarwinDirective()) {
918 case PPC::DIR_E500mc:
928 setPrefFunctionAlignment(4);
929 setPrefLoopAlignment(4);
933 if (Subtarget.enableMachineScheduler())
934 setSchedulingPreference(Sched::Source);
936 setSchedulingPreference(Sched::Hybrid);
938 computeRegisterProperties(STI.getRegisterInfo());
940 // The Freescale cores do better with aggressive inlining of memcpy and
941 // friends. GCC uses same threshold of 128 bytes (= 32 word stores).
942 if (Subtarget.getDarwinDirective() == PPC::DIR_E500mc ||
943 Subtarget.getDarwinDirective() == PPC::DIR_E5500) {
944 MaxStoresPerMemset = 32;
945 MaxStoresPerMemsetOptSize = 16;
946 MaxStoresPerMemcpy = 32;
947 MaxStoresPerMemcpyOptSize = 8;
948 MaxStoresPerMemmove = 32;
949 MaxStoresPerMemmoveOptSize = 8;
950 } else if (Subtarget.getDarwinDirective() == PPC::DIR_A2) {
951 // The A2 also benefits from (very) aggressive inlining of memcpy and
952 // friends. The overhead of a the function call, even when warm, can be
953 // over one hundred cycles.
954 MaxStoresPerMemset = 128;
955 MaxStoresPerMemcpy = 128;
956 MaxStoresPerMemmove = 128;
960 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
961 /// the desired ByVal argument alignment.
962 static void getMaxByValAlign(Type *Ty, unsigned &MaxAlign,
963 unsigned MaxMaxAlign) {
964 if (MaxAlign == MaxMaxAlign)
966 if (VectorType *VTy = dyn_cast<VectorType>(Ty)) {
967 if (MaxMaxAlign >= 32 && VTy->getBitWidth() >= 256)
969 else if (VTy->getBitWidth() >= 128 && MaxAlign < 16)
971 } else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
972 unsigned EltAlign = 0;
973 getMaxByValAlign(ATy->getElementType(), EltAlign, MaxMaxAlign);
974 if (EltAlign > MaxAlign)
976 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
977 for (auto *EltTy : STy->elements()) {
978 unsigned EltAlign = 0;
979 getMaxByValAlign(EltTy, EltAlign, MaxMaxAlign);
980 if (EltAlign > MaxAlign)
982 if (MaxAlign == MaxMaxAlign)
988 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
989 /// function arguments in the caller parameter area.
990 unsigned PPCTargetLowering::getByValTypeAlignment(Type *Ty,
991 const DataLayout &DL) const {
992 // Darwin passes everything on 4 byte boundary.
993 if (Subtarget.isDarwin())
996 // 16byte and wider vectors are passed on 16byte boundary.
997 // The rest is 8 on PPC64 and 4 on PPC32 boundary.
998 unsigned Align = Subtarget.isPPC64() ? 8 : 4;
999 if (Subtarget.hasAltivec() || Subtarget.hasQPX())
1000 getMaxByValAlign(Ty, Align, Subtarget.hasQPX() ? 32 : 16);
1004 bool PPCTargetLowering::useSoftFloat() const {
1005 return Subtarget.useSoftFloat();
1008 const char *PPCTargetLowering::getTargetNodeName(unsigned Opcode) const {
1009 switch ((PPCISD::NodeType)Opcode) {
1010 case PPCISD::FIRST_NUMBER: break;
1011 case PPCISD::FSEL: return "PPCISD::FSEL";
1012 case PPCISD::FCFID: return "PPCISD::FCFID";
1013 case PPCISD::FCFIDU: return "PPCISD::FCFIDU";
1014 case PPCISD::FCFIDS: return "PPCISD::FCFIDS";
1015 case PPCISD::FCFIDUS: return "PPCISD::FCFIDUS";
1016 case PPCISD::FCTIDZ: return "PPCISD::FCTIDZ";
1017 case PPCISD::FCTIWZ: return "PPCISD::FCTIWZ";
1018 case PPCISD::FCTIDUZ: return "PPCISD::FCTIDUZ";
1019 case PPCISD::FCTIWUZ: return "PPCISD::FCTIWUZ";
1020 case PPCISD::FRE: return "PPCISD::FRE";
1021 case PPCISD::FRSQRTE: return "PPCISD::FRSQRTE";
1022 case PPCISD::STFIWX: return "PPCISD::STFIWX";
1023 case PPCISD::VMADDFP: return "PPCISD::VMADDFP";
1024 case PPCISD::VNMSUBFP: return "PPCISD::VNMSUBFP";
1025 case PPCISD::VPERM: return "PPCISD::VPERM";
1026 case PPCISD::XXSPLT: return "PPCISD::XXSPLT";
1027 case PPCISD::XXINSERT: return "PPCISD::XXINSERT";
1028 case PPCISD::VECSHL: return "PPCISD::VECSHL";
1029 case PPCISD::CMPB: return "PPCISD::CMPB";
1030 case PPCISD::Hi: return "PPCISD::Hi";
1031 case PPCISD::Lo: return "PPCISD::Lo";
1032 case PPCISD::TOC_ENTRY: return "PPCISD::TOC_ENTRY";
1033 case PPCISD::DYNALLOC: return "PPCISD::DYNALLOC";
1034 case PPCISD::DYNAREAOFFSET: return "PPCISD::DYNAREAOFFSET";
1035 case PPCISD::GlobalBaseReg: return "PPCISD::GlobalBaseReg";
1036 case PPCISD::SRL: return "PPCISD::SRL";
1037 case PPCISD::SRA: return "PPCISD::SRA";
1038 case PPCISD::SHL: return "PPCISD::SHL";
1039 case PPCISD::SRA_ADDZE: return "PPCISD::SRA_ADDZE";
1040 case PPCISD::CALL: return "PPCISD::CALL";
1041 case PPCISD::CALL_NOP: return "PPCISD::CALL_NOP";
1042 case PPCISD::MTCTR: return "PPCISD::MTCTR";
1043 case PPCISD::BCTRL: return "PPCISD::BCTRL";
1044 case PPCISD::BCTRL_LOAD_TOC: return "PPCISD::BCTRL_LOAD_TOC";
1045 case PPCISD::RET_FLAG: return "PPCISD::RET_FLAG";
1046 case PPCISD::READ_TIME_BASE: return "PPCISD::READ_TIME_BASE";
1047 case PPCISD::EH_SJLJ_SETJMP: return "PPCISD::EH_SJLJ_SETJMP";
1048 case PPCISD::EH_SJLJ_LONGJMP: return "PPCISD::EH_SJLJ_LONGJMP";
1049 case PPCISD::MFOCRF: return "PPCISD::MFOCRF";
1050 case PPCISD::MFVSR: return "PPCISD::MFVSR";
1051 case PPCISD::MTVSRA: return "PPCISD::MTVSRA";
1052 case PPCISD::MTVSRZ: return "PPCISD::MTVSRZ";
1053 case PPCISD::SINT_VEC_TO_FP: return "PPCISD::SINT_VEC_TO_FP";
1054 case PPCISD::UINT_VEC_TO_FP: return "PPCISD::UINT_VEC_TO_FP";
1055 case PPCISD::ANDIo_1_EQ_BIT: return "PPCISD::ANDIo_1_EQ_BIT";
1056 case PPCISD::ANDIo_1_GT_BIT: return "PPCISD::ANDIo_1_GT_BIT";
1057 case PPCISD::VCMP: return "PPCISD::VCMP";
1058 case PPCISD::VCMPo: return "PPCISD::VCMPo";
1059 case PPCISD::LBRX: return "PPCISD::LBRX";
1060 case PPCISD::STBRX: return "PPCISD::STBRX";
1061 case PPCISD::LFIWAX: return "PPCISD::LFIWAX";
1062 case PPCISD::LFIWZX: return "PPCISD::LFIWZX";
1063 case PPCISD::LXVD2X: return "PPCISD::LXVD2X";
1064 case PPCISD::STXVD2X: return "PPCISD::STXVD2X";
1065 case PPCISD::COND_BRANCH: return "PPCISD::COND_BRANCH";
1066 case PPCISD::BDNZ: return "PPCISD::BDNZ";
1067 case PPCISD::BDZ: return "PPCISD::BDZ";
1068 case PPCISD::MFFS: return "PPCISD::MFFS";
1069 case PPCISD::FADDRTZ: return "PPCISD::FADDRTZ";
1070 case PPCISD::TC_RETURN: return "PPCISD::TC_RETURN";
1071 case PPCISD::CR6SET: return "PPCISD::CR6SET";
1072 case PPCISD::CR6UNSET: return "PPCISD::CR6UNSET";
1073 case PPCISD::PPC32_GOT: return "PPCISD::PPC32_GOT";
1074 case PPCISD::PPC32_PICGOT: return "PPCISD::PPC32_PICGOT";
1075 case PPCISD::ADDIS_GOT_TPREL_HA: return "PPCISD::ADDIS_GOT_TPREL_HA";
1076 case PPCISD::LD_GOT_TPREL_L: return "PPCISD::LD_GOT_TPREL_L";
1077 case PPCISD::ADD_TLS: return "PPCISD::ADD_TLS";
1078 case PPCISD::ADDIS_TLSGD_HA: return "PPCISD::ADDIS_TLSGD_HA";
1079 case PPCISD::ADDI_TLSGD_L: return "PPCISD::ADDI_TLSGD_L";
1080 case PPCISD::GET_TLS_ADDR: return "PPCISD::GET_TLS_ADDR";
1081 case PPCISD::ADDI_TLSGD_L_ADDR: return "PPCISD::ADDI_TLSGD_L_ADDR";
1082 case PPCISD::ADDIS_TLSLD_HA: return "PPCISD::ADDIS_TLSLD_HA";
1083 case PPCISD::ADDI_TLSLD_L: return "PPCISD::ADDI_TLSLD_L";
1084 case PPCISD::GET_TLSLD_ADDR: return "PPCISD::GET_TLSLD_ADDR";
1085 case PPCISD::ADDI_TLSLD_L_ADDR: return "PPCISD::ADDI_TLSLD_L_ADDR";
1086 case PPCISD::ADDIS_DTPREL_HA: return "PPCISD::ADDIS_DTPREL_HA";
1087 case PPCISD::ADDI_DTPREL_L: return "PPCISD::ADDI_DTPREL_L";
1088 case PPCISD::VADD_SPLAT: return "PPCISD::VADD_SPLAT";
1089 case PPCISD::SC: return "PPCISD::SC";
1090 case PPCISD::CLRBHRB: return "PPCISD::CLRBHRB";
1091 case PPCISD::MFBHRBE: return "PPCISD::MFBHRBE";
1092 case PPCISD::RFEBB: return "PPCISD::RFEBB";
1093 case PPCISD::XXSWAPD: return "PPCISD::XXSWAPD";
1094 case PPCISD::SWAP_NO_CHAIN: return "PPCISD::SWAP_NO_CHAIN";
1095 case PPCISD::QVFPERM: return "PPCISD::QVFPERM";
1096 case PPCISD::QVGPCI: return "PPCISD::QVGPCI";
1097 case PPCISD::QVALIGNI: return "PPCISD::QVALIGNI";
1098 case PPCISD::QVESPLATI: return "PPCISD::QVESPLATI";
1099 case PPCISD::QBFLT: return "PPCISD::QBFLT";
1100 case PPCISD::QVLFSb: return "PPCISD::QVLFSb";
1105 EVT PPCTargetLowering::getSetCCResultType(const DataLayout &DL, LLVMContext &C,
1108 return Subtarget.useCRBits() ? MVT::i1 : MVT::i32;
1110 if (Subtarget.hasQPX())
1111 return EVT::getVectorVT(C, MVT::i1, VT.getVectorNumElements());
1113 return VT.changeVectorElementTypeToInteger();
1116 bool PPCTargetLowering::enableAggressiveFMAFusion(EVT VT) const {
1117 assert(VT.isFloatingPoint() && "Non-floating-point FMA?");
1121 //===----------------------------------------------------------------------===//
1122 // Node matching predicates, for use by the tblgen matching code.
1123 //===----------------------------------------------------------------------===//
1125 /// isFloatingPointZero - Return true if this is 0.0 or -0.0.
1126 static bool isFloatingPointZero(SDValue Op) {
1127 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
1128 return CFP->getValueAPF().isZero();
1129 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
1130 // Maybe this has already been legalized into the constant pool?
1131 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(Op.getOperand(1)))
1132 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
1133 return CFP->getValueAPF().isZero();
1138 /// isConstantOrUndef - Op is either an undef node or a ConstantSDNode. Return
1139 /// true if Op is undef or if it matches the specified value.
1140 static bool isConstantOrUndef(int Op, int Val) {
1141 return Op < 0 || Op == Val;
1144 /// isVPKUHUMShuffleMask - Return true if this is the shuffle mask for a
1145 /// VPKUHUM instruction.
1146 /// The ShuffleKind distinguishes between big-endian operations with
1147 /// two different inputs (0), either-endian operations with two identical
1148 /// inputs (1), and little-endian operations with two different inputs (2).
1149 /// For the latter, the input operands are swapped (see PPCInstrAltivec.td).
1150 bool PPC::isVPKUHUMShuffleMask(ShuffleVectorSDNode *N, unsigned ShuffleKind,
1151 SelectionDAG &DAG) {
1152 bool IsLE = DAG.getDataLayout().isLittleEndian();
1153 if (ShuffleKind == 0) {
1156 for (unsigned i = 0; i != 16; ++i)
1157 if (!isConstantOrUndef(N->getMaskElt(i), i*2+1))
1159 } else if (ShuffleKind == 2) {
1162 for (unsigned i = 0; i != 16; ++i)
1163 if (!isConstantOrUndef(N->getMaskElt(i), i*2))
1165 } else if (ShuffleKind == 1) {
1166 unsigned j = IsLE ? 0 : 1;
1167 for (unsigned i = 0; i != 8; ++i)
1168 if (!isConstantOrUndef(N->getMaskElt(i), i*2+j) ||
1169 !isConstantOrUndef(N->getMaskElt(i+8), i*2+j))
1175 /// isVPKUWUMShuffleMask - Return true if this is the shuffle mask for a
1176 /// VPKUWUM instruction.
1177 /// The ShuffleKind distinguishes between big-endian operations with
1178 /// two different inputs (0), either-endian operations with two identical
1179 /// inputs (1), and little-endian operations with two different inputs (2).
1180 /// For the latter, the input operands are swapped (see PPCInstrAltivec.td).
1181 bool PPC::isVPKUWUMShuffleMask(ShuffleVectorSDNode *N, unsigned ShuffleKind,
1182 SelectionDAG &DAG) {
1183 bool IsLE = DAG.getDataLayout().isLittleEndian();
1184 if (ShuffleKind == 0) {
1187 for (unsigned i = 0; i != 16; i += 2)
1188 if (!isConstantOrUndef(N->getMaskElt(i ), i*2+2) ||
1189 !isConstantOrUndef(N->getMaskElt(i+1), i*2+3))
1191 } else if (ShuffleKind == 2) {
1194 for (unsigned i = 0; i != 16; i += 2)
1195 if (!isConstantOrUndef(N->getMaskElt(i ), i*2) ||
1196 !isConstantOrUndef(N->getMaskElt(i+1), i*2+1))
1198 } else if (ShuffleKind == 1) {
1199 unsigned j = IsLE ? 0 : 2;
1200 for (unsigned i = 0; i != 8; i += 2)
1201 if (!isConstantOrUndef(N->getMaskElt(i ), i*2+j) ||
1202 !isConstantOrUndef(N->getMaskElt(i+1), i*2+j+1) ||
1203 !isConstantOrUndef(N->getMaskElt(i+8), i*2+j) ||
1204 !isConstantOrUndef(N->getMaskElt(i+9), i*2+j+1))
1210 /// isVPKUDUMShuffleMask - Return true if this is the shuffle mask for a
1211 /// VPKUDUM instruction, AND the VPKUDUM instruction exists for the
1212 /// current subtarget.
1214 /// The ShuffleKind distinguishes between big-endian operations with
1215 /// two different inputs (0), either-endian operations with two identical
1216 /// inputs (1), and little-endian operations with two different inputs (2).
1217 /// For the latter, the input operands are swapped (see PPCInstrAltivec.td).
1218 bool PPC::isVPKUDUMShuffleMask(ShuffleVectorSDNode *N, unsigned ShuffleKind,
1219 SelectionDAG &DAG) {
1220 const PPCSubtarget& Subtarget =
1221 static_cast<const PPCSubtarget&>(DAG.getSubtarget());
1222 if (!Subtarget.hasP8Vector())
1225 bool IsLE = DAG.getDataLayout().isLittleEndian();
1226 if (ShuffleKind == 0) {
1229 for (unsigned i = 0; i != 16; i += 4)
1230 if (!isConstantOrUndef(N->getMaskElt(i ), i*2+4) ||
1231 !isConstantOrUndef(N->getMaskElt(i+1), i*2+5) ||
1232 !isConstantOrUndef(N->getMaskElt(i+2), i*2+6) ||
1233 !isConstantOrUndef(N->getMaskElt(i+3), i*2+7))
1235 } else if (ShuffleKind == 2) {
1238 for (unsigned i = 0; i != 16; i += 4)
1239 if (!isConstantOrUndef(N->getMaskElt(i ), i*2) ||
1240 !isConstantOrUndef(N->getMaskElt(i+1), i*2+1) ||
1241 !isConstantOrUndef(N->getMaskElt(i+2), i*2+2) ||
1242 !isConstantOrUndef(N->getMaskElt(i+3), i*2+3))
1244 } else if (ShuffleKind == 1) {
1245 unsigned j = IsLE ? 0 : 4;
1246 for (unsigned i = 0; i != 8; i += 4)
1247 if (!isConstantOrUndef(N->getMaskElt(i ), i*2+j) ||
1248 !isConstantOrUndef(N->getMaskElt(i+1), i*2+j+1) ||
1249 !isConstantOrUndef(N->getMaskElt(i+2), i*2+j+2) ||
1250 !isConstantOrUndef(N->getMaskElt(i+3), i*2+j+3) ||
1251 !isConstantOrUndef(N->getMaskElt(i+8), i*2+j) ||
1252 !isConstantOrUndef(N->getMaskElt(i+9), i*2+j+1) ||
1253 !isConstantOrUndef(N->getMaskElt(i+10), i*2+j+2) ||
1254 !isConstantOrUndef(N->getMaskElt(i+11), i*2+j+3))
1260 /// isVMerge - Common function, used to match vmrg* shuffles.
1262 static bool isVMerge(ShuffleVectorSDNode *N, unsigned UnitSize,
1263 unsigned LHSStart, unsigned RHSStart) {
1264 if (N->getValueType(0) != MVT::v16i8)
1266 assert((UnitSize == 1 || UnitSize == 2 || UnitSize == 4) &&
1267 "Unsupported merge size!");
1269 for (unsigned i = 0; i != 8/UnitSize; ++i) // Step over units
1270 for (unsigned j = 0; j != UnitSize; ++j) { // Step over bytes within unit
1271 if (!isConstantOrUndef(N->getMaskElt(i*UnitSize*2+j),
1272 LHSStart+j+i*UnitSize) ||
1273 !isConstantOrUndef(N->getMaskElt(i*UnitSize*2+UnitSize+j),
1274 RHSStart+j+i*UnitSize))
1280 /// isVMRGLShuffleMask - Return true if this is a shuffle mask suitable for
1281 /// a VMRGL* instruction with the specified unit size (1,2 or 4 bytes).
1282 /// The ShuffleKind distinguishes between big-endian merges with two
1283 /// different inputs (0), either-endian merges with two identical inputs (1),
1284 /// and little-endian merges with two different inputs (2). For the latter,
1285 /// the input operands are swapped (see PPCInstrAltivec.td).
1286 bool PPC::isVMRGLShuffleMask(ShuffleVectorSDNode *N, unsigned UnitSize,
1287 unsigned ShuffleKind, SelectionDAG &DAG) {
1288 if (DAG.getDataLayout().isLittleEndian()) {
1289 if (ShuffleKind == 1) // unary
1290 return isVMerge(N, UnitSize, 0, 0);
1291 else if (ShuffleKind == 2) // swapped
1292 return isVMerge(N, UnitSize, 0, 16);
1296 if (ShuffleKind == 1) // unary
1297 return isVMerge(N, UnitSize, 8, 8);
1298 else if (ShuffleKind == 0) // normal
1299 return isVMerge(N, UnitSize, 8, 24);
1305 /// isVMRGHShuffleMask - Return true if this is a shuffle mask suitable for
1306 /// a VMRGH* instruction with the specified unit size (1,2 or 4 bytes).
1307 /// The ShuffleKind distinguishes between big-endian merges with two
1308 /// different inputs (0), either-endian merges with two identical inputs (1),
1309 /// and little-endian merges with two different inputs (2). For the latter,
1310 /// the input operands are swapped (see PPCInstrAltivec.td).
1311 bool PPC::isVMRGHShuffleMask(ShuffleVectorSDNode *N, unsigned UnitSize,
1312 unsigned ShuffleKind, SelectionDAG &DAG) {
1313 if (DAG.getDataLayout().isLittleEndian()) {
1314 if (ShuffleKind == 1) // unary
1315 return isVMerge(N, UnitSize, 8, 8);
1316 else if (ShuffleKind == 2) // swapped
1317 return isVMerge(N, UnitSize, 8, 24);
1321 if (ShuffleKind == 1) // unary
1322 return isVMerge(N, UnitSize, 0, 0);
1323 else if (ShuffleKind == 0) // normal
1324 return isVMerge(N, UnitSize, 0, 16);
1331 * \brief Common function used to match vmrgew and vmrgow shuffles
1333 * The indexOffset determines whether to look for even or odd words in
1334 * the shuffle mask. This is based on the of the endianness of the target
1337 * - Use offset of 0 to check for odd elements
1338 * - Use offset of 4 to check for even elements
1340 * - Use offset of 0 to check for even elements
1341 * - Use offset of 4 to check for odd elements
1342 * A detailed description of the vector element ordering for little endian and
1343 * big endian can be found at
1344 * http://www.ibm.com/developerworks/library/l-ibm-xl-c-cpp-compiler/index.html
1345 * Targeting your applications - what little endian and big endian IBM XL C/C++
1346 * compiler differences mean to you
1348 * The mask to the shuffle vector instruction specifies the indices of the
1349 * elements from the two input vectors to place in the result. The elements are
1350 * numbered in array-access order, starting with the first vector. These vectors
1351 * are always of type v16i8, thus each vector will contain 16 elements of size
1352 * 8. More info on the shuffle vector can be found in the
1353 * http://llvm.org/docs/LangRef.html#shufflevector-instruction
1354 * Language Reference.
1356 * The RHSStartValue indicates whether the same input vectors are used (unary)
1357 * or two different input vectors are used, based on the following:
1358 * - If the instruction uses the same vector for both inputs, the range of the
1359 * indices will be 0 to 15. In this case, the RHSStart value passed should
1361 * - If the instruction has two different vectors then the range of the
1362 * indices will be 0 to 31. In this case, the RHSStart value passed should
1363 * be 16 (indices 0-15 specify elements in the first vector while indices 16
1364 * to 31 specify elements in the second vector).
1366 * \param[in] N The shuffle vector SD Node to analyze
1367 * \param[in] IndexOffset Specifies whether to look for even or odd elements
1368 * \param[in] RHSStartValue Specifies the starting index for the righthand input
1369 * vector to the shuffle_vector instruction
1370 * \return true iff this shuffle vector represents an even or odd word merge
1372 static bool isVMerge(ShuffleVectorSDNode *N, unsigned IndexOffset,
1373 unsigned RHSStartValue) {
1374 if (N->getValueType(0) != MVT::v16i8)
1377 for (unsigned i = 0; i < 2; ++i)
1378 for (unsigned j = 0; j < 4; ++j)
1379 if (!isConstantOrUndef(N->getMaskElt(i*4+j),
1380 i*RHSStartValue+j+IndexOffset) ||
1381 !isConstantOrUndef(N->getMaskElt(i*4+j+8),
1382 i*RHSStartValue+j+IndexOffset+8))
1388 * \brief Determine if the specified shuffle mask is suitable for the vmrgew or
1389 * vmrgow instructions.
1391 * \param[in] N The shuffle vector SD Node to analyze
1392 * \param[in] CheckEven Check for an even merge (true) or an odd merge (false)
1393 * \param[in] ShuffleKind Identify the type of merge:
1394 * - 0 = big-endian merge with two different inputs;
1395 * - 1 = either-endian merge with two identical inputs;
1396 * - 2 = little-endian merge with two different inputs (inputs are swapped for
1397 * little-endian merges).
1398 * \param[in] DAG The current SelectionDAG
1399 * \return true iff this shuffle mask
1401 bool PPC::isVMRGEOShuffleMask(ShuffleVectorSDNode *N, bool CheckEven,
1402 unsigned ShuffleKind, SelectionDAG &DAG) {
1403 if (DAG.getDataLayout().isLittleEndian()) {
1404 unsigned indexOffset = CheckEven ? 4 : 0;
1405 if (ShuffleKind == 1) // Unary
1406 return isVMerge(N, indexOffset, 0);
1407 else if (ShuffleKind == 2) // swapped
1408 return isVMerge(N, indexOffset, 16);
1413 unsigned indexOffset = CheckEven ? 0 : 4;
1414 if (ShuffleKind == 1) // Unary
1415 return isVMerge(N, indexOffset, 0);
1416 else if (ShuffleKind == 0) // Normal
1417 return isVMerge(N, indexOffset, 16);
1424 /// isVSLDOIShuffleMask - If this is a vsldoi shuffle mask, return the shift
1425 /// amount, otherwise return -1.
1426 /// The ShuffleKind distinguishes between big-endian operations with two
1427 /// different inputs (0), either-endian operations with two identical inputs
1428 /// (1), and little-endian operations with two different inputs (2). For the
1429 /// latter, the input operands are swapped (see PPCInstrAltivec.td).
1430 int PPC::isVSLDOIShuffleMask(SDNode *N, unsigned ShuffleKind,
1431 SelectionDAG &DAG) {
1432 if (N->getValueType(0) != MVT::v16i8)
1435 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
1437 // Find the first non-undef value in the shuffle mask.
1439 for (i = 0; i != 16 && SVOp->getMaskElt(i) < 0; ++i)
1442 if (i == 16) return -1; // all undef.
1444 // Otherwise, check to see if the rest of the elements are consecutively
1445 // numbered from this value.
1446 unsigned ShiftAmt = SVOp->getMaskElt(i);
1447 if (ShiftAmt < i) return -1;
1450 bool isLE = DAG.getDataLayout().isLittleEndian();
1452 if ((ShuffleKind == 0 && !isLE) || (ShuffleKind == 2 && isLE)) {
1453 // Check the rest of the elements to see if they are consecutive.
1454 for (++i; i != 16; ++i)
1455 if (!isConstantOrUndef(SVOp->getMaskElt(i), ShiftAmt+i))
1457 } else if (ShuffleKind == 1) {
1458 // Check the rest of the elements to see if they are consecutive.
1459 for (++i; i != 16; ++i)
1460 if (!isConstantOrUndef(SVOp->getMaskElt(i), (ShiftAmt+i) & 15))
1466 ShiftAmt = 16 - ShiftAmt;
1471 /// isSplatShuffleMask - Return true if the specified VECTOR_SHUFFLE operand
1472 /// specifies a splat of a single element that is suitable for input to
1473 /// VSPLTB/VSPLTH/VSPLTW.
1474 bool PPC::isSplatShuffleMask(ShuffleVectorSDNode *N, unsigned EltSize) {
1475 assert(N->getValueType(0) == MVT::v16i8 &&
1476 (EltSize == 1 || EltSize == 2 || EltSize == 4));
1478 // The consecutive indices need to specify an element, not part of two
1479 // different elements. So abandon ship early if this isn't the case.
1480 if (N->getMaskElt(0) % EltSize != 0)
1483 // This is a splat operation if each element of the permute is the same, and
1484 // if the value doesn't reference the second vector.
1485 unsigned ElementBase = N->getMaskElt(0);
1487 // FIXME: Handle UNDEF elements too!
1488 if (ElementBase >= 16)
1491 // Check that the indices are consecutive, in the case of a multi-byte element
1492 // splatted with a v16i8 mask.
1493 for (unsigned i = 1; i != EltSize; ++i)
1494 if (N->getMaskElt(i) < 0 || N->getMaskElt(i) != (int)(i+ElementBase))
1497 for (unsigned i = EltSize, e = 16; i != e; i += EltSize) {
1498 if (N->getMaskElt(i) < 0) continue;
1499 for (unsigned j = 0; j != EltSize; ++j)
1500 if (N->getMaskElt(i+j) != N->getMaskElt(j))
1506 bool PPC::isXXINSERTWMask(ShuffleVectorSDNode *N, unsigned &ShiftElts,
1507 unsigned &InsertAtByte, bool &Swap, bool IsLE) {
1509 // Check that the mask is shuffling words
1510 for (unsigned i = 0; i < 4; ++i) {
1511 unsigned B0 = N->getMaskElt(i*4);
1512 unsigned B1 = N->getMaskElt(i*4+1);
1513 unsigned B2 = N->getMaskElt(i*4+2);
1514 unsigned B3 = N->getMaskElt(i*4+3);
1517 if (B1 != B0+1 || B2 != B1+1 || B3 != B2+1)
1521 // Now we look at mask elements 0,4,8,12
1522 unsigned M0 = N->getMaskElt(0) / 4;
1523 unsigned M1 = N->getMaskElt(4) / 4;
1524 unsigned M2 = N->getMaskElt(8) / 4;
1525 unsigned M3 = N->getMaskElt(12) / 4;
1526 unsigned LittleEndianShifts[] = { 2, 1, 0, 3 };
1527 unsigned BigEndianShifts[] = { 3, 0, 1, 2 };
1529 // Below, let H and L be arbitrary elements of the shuffle mask
1530 // where H is in the range [4,7] and L is in the range [0,3].
1531 // H, 1, 2, 3 or L, 5, 6, 7
1532 if ((M0 > 3 && M1 == 1 && M2 == 2 && M3 == 3) ||
1533 (M0 < 4 && M1 == 5 && M2 == 6 && M3 == 7)) {
1534 ShiftElts = IsLE ? LittleEndianShifts[M0 & 0x3] : BigEndianShifts[M0 & 0x3];
1535 InsertAtByte = IsLE ? 12 : 0;
1539 // 0, H, 2, 3 or 4, L, 6, 7
1540 if ((M1 > 3 && M0 == 0 && M2 == 2 && M3 == 3) ||
1541 (M1 < 4 && M0 == 4 && M2 == 6 && M3 == 7)) {
1542 ShiftElts = IsLE ? LittleEndianShifts[M1 & 0x3] : BigEndianShifts[M1 & 0x3];
1543 InsertAtByte = IsLE ? 8 : 4;
1547 // 0, 1, H, 3 or 4, 5, L, 7
1548 if ((M2 > 3 && M0 == 0 && M1 == 1 && M3 == 3) ||
1549 (M2 < 4 && M0 == 4 && M1 == 5 && M3 == 7)) {
1550 ShiftElts = IsLE ? LittleEndianShifts[M2 & 0x3] : BigEndianShifts[M2 & 0x3];
1551 InsertAtByte = IsLE ? 4 : 8;
1555 // 0, 1, 2, H or 4, 5, 6, L
1556 if ((M3 > 3 && M0 == 0 && M1 == 1 && M2 == 2) ||
1557 (M3 < 4 && M0 == 4 && M1 == 5 && M2 == 6)) {
1558 ShiftElts = IsLE ? LittleEndianShifts[M3 & 0x3] : BigEndianShifts[M3 & 0x3];
1559 InsertAtByte = IsLE ? 0 : 12;
1564 // If both vector operands for the shuffle are the same vector, the mask will
1565 // contain only elements from the first one and the second one will be undef.
1566 if (N->getOperand(1).isUndef()) {
1569 unsigned XXINSERTWSrcElem = IsLE ? 2 : 1;
1570 if (M0 == XXINSERTWSrcElem && M1 == 1 && M2 == 2 && M3 == 3) {
1571 InsertAtByte = IsLE ? 12 : 0;
1574 if (M0 == 0 && M1 == XXINSERTWSrcElem && M2 == 2 && M3 == 3) {
1575 InsertAtByte = IsLE ? 8 : 4;
1578 if (M0 == 0 && M1 == 1 && M2 == XXINSERTWSrcElem && M3 == 3) {
1579 InsertAtByte = IsLE ? 4 : 8;
1582 if (M0 == 0 && M1 == 1 && M2 == 2 && M3 == XXINSERTWSrcElem) {
1583 InsertAtByte = IsLE ? 0 : 12;
1591 /// getVSPLTImmediate - Return the appropriate VSPLT* immediate to splat the
1592 /// specified isSplatShuffleMask VECTOR_SHUFFLE mask.
1593 unsigned PPC::getVSPLTImmediate(SDNode *N, unsigned EltSize,
1594 SelectionDAG &DAG) {
1595 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
1596 assert(isSplatShuffleMask(SVOp, EltSize));
1597 if (DAG.getDataLayout().isLittleEndian())
1598 return (16 / EltSize) - 1 - (SVOp->getMaskElt(0) / EltSize);
1600 return SVOp->getMaskElt(0) / EltSize;
1603 /// get_VSPLTI_elt - If this is a build_vector of constants which can be formed
1604 /// by using a vspltis[bhw] instruction of the specified element size, return
1605 /// the constant being splatted. The ByteSize field indicates the number of
1606 /// bytes of each element [124] -> [bhw].
1607 SDValue PPC::get_VSPLTI_elt(SDNode *N, unsigned ByteSize, SelectionDAG &DAG) {
1608 SDValue OpVal(nullptr, 0);
1610 // If ByteSize of the splat is bigger than the element size of the
1611 // build_vector, then we have a case where we are checking for a splat where
1612 // multiple elements of the buildvector are folded together into a single
1613 // logical element of the splat (e.g. "vsplish 1" to splat {0,1}*8).
1614 unsigned EltSize = 16/N->getNumOperands();
1615 if (EltSize < ByteSize) {
1616 unsigned Multiple = ByteSize/EltSize; // Number of BV entries per spltval.
1617 SDValue UniquedVals[4];
1618 assert(Multiple > 1 && Multiple <= 4 && "How can this happen?");
1620 // See if all of the elements in the buildvector agree across.
1621 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
1622 if (N->getOperand(i).isUndef()) continue;
1623 // If the element isn't a constant, bail fully out.
1624 if (!isa<ConstantSDNode>(N->getOperand(i))) return SDValue();
1627 if (!UniquedVals[i&(Multiple-1)].getNode())
1628 UniquedVals[i&(Multiple-1)] = N->getOperand(i);
1629 else if (UniquedVals[i&(Multiple-1)] != N->getOperand(i))
1630 return SDValue(); // no match.
1633 // Okay, if we reached this point, UniquedVals[0..Multiple-1] contains
1634 // either constant or undef values that are identical for each chunk. See
1635 // if these chunks can form into a larger vspltis*.
1637 // Check to see if all of the leading entries are either 0 or -1. If
1638 // neither, then this won't fit into the immediate field.
1639 bool LeadingZero = true;
1640 bool LeadingOnes = true;
1641 for (unsigned i = 0; i != Multiple-1; ++i) {
1642 if (!UniquedVals[i].getNode()) continue; // Must have been undefs.
1644 LeadingZero &= isNullConstant(UniquedVals[i]);
1645 LeadingOnes &= isAllOnesConstant(UniquedVals[i]);
1647 // Finally, check the least significant entry.
1649 if (!UniquedVals[Multiple-1].getNode())
1650 return DAG.getTargetConstant(0, SDLoc(N), MVT::i32); // 0,0,0,undef
1651 int Val = cast<ConstantSDNode>(UniquedVals[Multiple-1])->getZExtValue();
1652 if (Val < 16) // 0,0,0,4 -> vspltisw(4)
1653 return DAG.getTargetConstant(Val, SDLoc(N), MVT::i32);
1656 if (!UniquedVals[Multiple-1].getNode())
1657 return DAG.getTargetConstant(~0U, SDLoc(N), MVT::i32); // -1,-1,-1,undef
1658 int Val =cast<ConstantSDNode>(UniquedVals[Multiple-1])->getSExtValue();
1659 if (Val >= -16) // -1,-1,-1,-2 -> vspltisw(-2)
1660 return DAG.getTargetConstant(Val, SDLoc(N), MVT::i32);
1666 // Check to see if this buildvec has a single non-undef value in its elements.
1667 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
1668 if (N->getOperand(i).isUndef()) continue;
1669 if (!OpVal.getNode())
1670 OpVal = N->getOperand(i);
1671 else if (OpVal != N->getOperand(i))
1675 if (!OpVal.getNode()) return SDValue(); // All UNDEF: use implicit def.
1677 unsigned ValSizeInBytes = EltSize;
1679 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(OpVal)) {
1680 Value = CN->getZExtValue();
1681 } else if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(OpVal)) {
1682 assert(CN->getValueType(0) == MVT::f32 && "Only one legal FP vector type!");
1683 Value = FloatToBits(CN->getValueAPF().convertToFloat());
1686 // If the splat value is larger than the element value, then we can never do
1687 // this splat. The only case that we could fit the replicated bits into our
1688 // immediate field for would be zero, and we prefer to use vxor for it.
1689 if (ValSizeInBytes < ByteSize) return SDValue();
1691 // If the element value is larger than the splat value, check if it consists
1692 // of a repeated bit pattern of size ByteSize.
1693 if (!APInt(ValSizeInBytes * 8, Value).isSplat(ByteSize * 8))
1696 // Properly sign extend the value.
1697 int MaskVal = SignExtend32(Value, ByteSize * 8);
1699 // If this is zero, don't match, zero matches ISD::isBuildVectorAllZeros.
1700 if (MaskVal == 0) return SDValue();
1702 // Finally, if this value fits in a 5 bit sext field, return it
1703 if (SignExtend32<5>(MaskVal) == MaskVal)
1704 return DAG.getTargetConstant(MaskVal, SDLoc(N), MVT::i32);
1708 /// isQVALIGNIShuffleMask - If this is a qvaligni shuffle mask, return the shift
1709 /// amount, otherwise return -1.
1710 int PPC::isQVALIGNIShuffleMask(SDNode *N) {
1711 EVT VT = N->getValueType(0);
1712 if (VT != MVT::v4f64 && VT != MVT::v4f32 && VT != MVT::v4i1)
1715 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
1717 // Find the first non-undef value in the shuffle mask.
1719 for (i = 0; i != 4 && SVOp->getMaskElt(i) < 0; ++i)
1722 if (i == 4) return -1; // all undef.
1724 // Otherwise, check to see if the rest of the elements are consecutively
1725 // numbered from this value.
1726 unsigned ShiftAmt = SVOp->getMaskElt(i);
1727 if (ShiftAmt < i) return -1;
1730 // Check the rest of the elements to see if they are consecutive.
1731 for (++i; i != 4; ++i)
1732 if (!isConstantOrUndef(SVOp->getMaskElt(i), ShiftAmt+i))
1738 //===----------------------------------------------------------------------===//
1739 // Addressing Mode Selection
1740 //===----------------------------------------------------------------------===//
1742 /// isIntS16Immediate - This method tests to see if the node is either a 32-bit
1743 /// or 64-bit immediate, and if the value can be accurately represented as a
1744 /// sign extension from a 16-bit value. If so, this returns true and the
1746 static bool isIntS16Immediate(SDNode *N, short &Imm) {
1747 if (!isa<ConstantSDNode>(N))
1750 Imm = (short)cast<ConstantSDNode>(N)->getZExtValue();
1751 if (N->getValueType(0) == MVT::i32)
1752 return Imm == (int32_t)cast<ConstantSDNode>(N)->getZExtValue();
1754 return Imm == (int64_t)cast<ConstantSDNode>(N)->getZExtValue();
1756 static bool isIntS16Immediate(SDValue Op, short &Imm) {
1757 return isIntS16Immediate(Op.getNode(), Imm);
1760 /// SelectAddressRegReg - Given the specified addressed, check to see if it
1761 /// can be represented as an indexed [r+r] operation. Returns false if it
1762 /// can be more efficiently represented with [r+imm].
1763 bool PPCTargetLowering::SelectAddressRegReg(SDValue N, SDValue &Base,
1765 SelectionDAG &DAG) const {
1767 if (N.getOpcode() == ISD::ADD) {
1768 if (isIntS16Immediate(N.getOperand(1), imm))
1769 return false; // r+i
1770 if (N.getOperand(1).getOpcode() == PPCISD::Lo)
1771 return false; // r+i
1773 Base = N.getOperand(0);
1774 Index = N.getOperand(1);
1776 } else if (N.getOpcode() == ISD::OR) {
1777 if (isIntS16Immediate(N.getOperand(1), imm))
1778 return false; // r+i can fold it if we can.
1780 // If this is an or of disjoint bitfields, we can codegen this as an add
1781 // (for better address arithmetic) if the LHS and RHS of the OR are provably
1783 APInt LHSKnownZero, LHSKnownOne;
1784 APInt RHSKnownZero, RHSKnownOne;
1785 DAG.computeKnownBits(N.getOperand(0),
1786 LHSKnownZero, LHSKnownOne);
1788 if (LHSKnownZero.getBoolValue()) {
1789 DAG.computeKnownBits(N.getOperand(1),
1790 RHSKnownZero, RHSKnownOne);
1791 // If all of the bits are known zero on the LHS or RHS, the add won't
1793 if (~(LHSKnownZero | RHSKnownZero) == 0) {
1794 Base = N.getOperand(0);
1795 Index = N.getOperand(1);
1804 // If we happen to be doing an i64 load or store into a stack slot that has
1805 // less than a 4-byte alignment, then the frame-index elimination may need to
1806 // use an indexed load or store instruction (because the offset may not be a
1807 // multiple of 4). The extra register needed to hold the offset comes from the
1808 // register scavenger, and it is possible that the scavenger will need to use
1809 // an emergency spill slot. As a result, we need to make sure that a spill slot
1810 // is allocated when doing an i64 load/store into a less-than-4-byte-aligned
1812 static void fixupFuncForFI(SelectionDAG &DAG, int FrameIdx, EVT VT) {
1813 // FIXME: This does not handle the LWA case.
1817 // NOTE: We'll exclude negative FIs here, which come from argument
1818 // lowering, because there are no known test cases triggering this problem
1819 // using packed structures (or similar). We can remove this exclusion if
1820 // we find such a test case. The reason why this is so test-case driven is
1821 // because this entire 'fixup' is only to prevent crashes (from the
1822 // register scavenger) on not-really-valid inputs. For example, if we have:
1824 // %b = bitcast i1* %a to i64*
1825 // store i64* a, i64 b
1826 // then the store should really be marked as 'align 1', but is not. If it
1827 // were marked as 'align 1' then the indexed form would have been
1828 // instruction-selected initially, and the problem this 'fixup' is preventing
1829 // won't happen regardless.
1833 MachineFunction &MF = DAG.getMachineFunction();
1834 MachineFrameInfo &MFI = MF.getFrameInfo();
1836 unsigned Align = MFI.getObjectAlignment(FrameIdx);
1840 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
1841 FuncInfo->setHasNonRISpills();
1844 /// Returns true if the address N can be represented by a base register plus
1845 /// a signed 16-bit displacement [r+imm], and if it is not better
1846 /// represented as reg+reg. If Aligned is true, only accept displacements
1847 /// suitable for STD and friends, i.e. multiples of 4.
1848 bool PPCTargetLowering::SelectAddressRegImm(SDValue N, SDValue &Disp,
1851 bool Aligned) const {
1852 // FIXME dl should come from parent load or store, not from address
1854 // If this can be more profitably realized as r+r, fail.
1855 if (SelectAddressRegReg(N, Disp, Base, DAG))
1858 if (N.getOpcode() == ISD::ADD) {
1860 if (isIntS16Immediate(N.getOperand(1), imm) &&
1861 (!Aligned || (imm & 3) == 0)) {
1862 Disp = DAG.getTargetConstant(imm, dl, N.getValueType());
1863 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
1864 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
1865 fixupFuncForFI(DAG, FI->getIndex(), N.getValueType());
1867 Base = N.getOperand(0);
1869 return true; // [r+i]
1870 } else if (N.getOperand(1).getOpcode() == PPCISD::Lo) {
1871 // Match LOAD (ADD (X, Lo(G))).
1872 assert(!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getZExtValue()
1873 && "Cannot handle constant offsets yet!");
1874 Disp = N.getOperand(1).getOperand(0); // The global address.
1875 assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||
1876 Disp.getOpcode() == ISD::TargetGlobalTLSAddress ||
1877 Disp.getOpcode() == ISD::TargetConstantPool ||
1878 Disp.getOpcode() == ISD::TargetJumpTable);
1879 Base = N.getOperand(0);
1880 return true; // [&g+r]
1882 } else if (N.getOpcode() == ISD::OR) {
1884 if (isIntS16Immediate(N.getOperand(1), imm) &&
1885 (!Aligned || (imm & 3) == 0)) {
1886 // If this is an or of disjoint bitfields, we can codegen this as an add
1887 // (for better address arithmetic) if the LHS and RHS of the OR are
1888 // provably disjoint.
1889 APInt LHSKnownZero, LHSKnownOne;
1890 DAG.computeKnownBits(N.getOperand(0), LHSKnownZero, LHSKnownOne);
1892 if ((LHSKnownZero.getZExtValue()|~(uint64_t)imm) == ~0ULL) {
1893 // If all of the bits are known zero on the LHS or RHS, the add won't
1895 if (FrameIndexSDNode *FI =
1896 dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
1897 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
1898 fixupFuncForFI(DAG, FI->getIndex(), N.getValueType());
1900 Base = N.getOperand(0);
1902 Disp = DAG.getTargetConstant(imm, dl, N.getValueType());
1906 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
1907 // Loading from a constant address.
1909 // If this address fits entirely in a 16-bit sext immediate field, codegen
1912 if (isIntS16Immediate(CN, Imm) && (!Aligned || (Imm & 3) == 0)) {
1913 Disp = DAG.getTargetConstant(Imm, dl, CN->getValueType(0));
1914 Base = DAG.getRegister(Subtarget.isPPC64() ? PPC::ZERO8 : PPC::ZERO,
1915 CN->getValueType(0));
1919 // Handle 32-bit sext immediates with LIS + addr mode.
1920 if ((CN->getValueType(0) == MVT::i32 ||
1921 (int64_t)CN->getZExtValue() == (int)CN->getZExtValue()) &&
1922 (!Aligned || (CN->getZExtValue() & 3) == 0)) {
1923 int Addr = (int)CN->getZExtValue();
1925 // Otherwise, break this down into an LIS + disp.
1926 Disp = DAG.getTargetConstant((short)Addr, dl, MVT::i32);
1928 Base = DAG.getTargetConstant((Addr - (signed short)Addr) >> 16, dl,
1930 unsigned Opc = CN->getValueType(0) == MVT::i32 ? PPC::LIS : PPC::LIS8;
1931 Base = SDValue(DAG.getMachineNode(Opc, dl, CN->getValueType(0), Base), 0);
1936 Disp = DAG.getTargetConstant(0, dl, getPointerTy(DAG.getDataLayout()));
1937 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N)) {
1938 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
1939 fixupFuncForFI(DAG, FI->getIndex(), N.getValueType());
1942 return true; // [r+0]
1945 /// SelectAddressRegRegOnly - Given the specified addressed, force it to be
1946 /// represented as an indexed [r+r] operation.
1947 bool PPCTargetLowering::SelectAddressRegRegOnly(SDValue N, SDValue &Base,
1949 SelectionDAG &DAG) const {
1950 // Check to see if we can easily represent this as an [r+r] address. This
1951 // will fail if it thinks that the address is more profitably represented as
1952 // reg+imm, e.g. where imm = 0.
1953 if (SelectAddressRegReg(N, Base, Index, DAG))
1956 // If the operand is an addition, always emit this as [r+r], since this is
1957 // better (for code size, and execution, as the memop does the add for free)
1958 // than emitting an explicit add.
1959 if (N.getOpcode() == ISD::ADD) {
1960 Base = N.getOperand(0);
1961 Index = N.getOperand(1);
1965 // Otherwise, do it the hard way, using R0 as the base register.
1966 Base = DAG.getRegister(Subtarget.isPPC64() ? PPC::ZERO8 : PPC::ZERO,
1972 /// getPreIndexedAddressParts - returns true by value, base pointer and
1973 /// offset pointer and addressing mode by reference if the node's address
1974 /// can be legally represented as pre-indexed load / store address.
1975 bool PPCTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
1977 ISD::MemIndexedMode &AM,
1978 SelectionDAG &DAG) const {
1979 if (DisablePPCPreinc) return false;
1985 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
1986 Ptr = LD->getBasePtr();
1987 VT = LD->getMemoryVT();
1988 Alignment = LD->getAlignment();
1989 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
1990 Ptr = ST->getBasePtr();
1991 VT = ST->getMemoryVT();
1992 Alignment = ST->getAlignment();
1997 // PowerPC doesn't have preinc load/store instructions for vectors (except
1998 // for QPX, which does have preinc r+r forms).
1999 if (VT.isVector()) {
2000 if (!Subtarget.hasQPX() || (VT != MVT::v4f64 && VT != MVT::v4f32)) {
2002 } else if (SelectAddressRegRegOnly(Ptr, Offset, Base, DAG)) {
2008 if (SelectAddressRegReg(Ptr, Base, Offset, DAG)) {
2010 // Common code will reject creating a pre-inc form if the base pointer
2011 // is a frame index, or if N is a store and the base pointer is either
2012 // the same as or a predecessor of the value being stored. Check for
2013 // those situations here, and try with swapped Base/Offset instead.
2016 if (isa<FrameIndexSDNode>(Base) || isa<RegisterSDNode>(Base))
2019 SDValue Val = cast<StoreSDNode>(N)->getValue();
2020 if (Val == Base || Base.getNode()->isPredecessorOf(Val.getNode()))
2025 std::swap(Base, Offset);
2031 // LDU/STU can only handle immediates that are a multiple of 4.
2032 if (VT != MVT::i64) {
2033 if (!SelectAddressRegImm(Ptr, Offset, Base, DAG, false))
2036 // LDU/STU need an address with at least 4-byte alignment.
2040 if (!SelectAddressRegImm(Ptr, Offset, Base, DAG, true))
2044 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
2045 // PPC64 doesn't have lwau, but it does have lwaux. Reject preinc load of
2046 // sext i32 to i64 when addr mode is r+i.
2047 if (LD->getValueType(0) == MVT::i64 && LD->getMemoryVT() == MVT::i32 &&
2048 LD->getExtensionType() == ISD::SEXTLOAD &&
2049 isa<ConstantSDNode>(Offset))
2057 //===----------------------------------------------------------------------===//
2058 // LowerOperation implementation
2059 //===----------------------------------------------------------------------===//
2061 /// Return true if we should reference labels using a PICBase, set the HiOpFlags
2062 /// and LoOpFlags to the target MO flags.
2063 static void getLabelAccessInfo(bool IsPIC, const PPCSubtarget &Subtarget,
2064 unsigned &HiOpFlags, unsigned &LoOpFlags,
2065 const GlobalValue *GV = nullptr) {
2066 HiOpFlags = PPCII::MO_HA;
2067 LoOpFlags = PPCII::MO_LO;
2069 // Don't use the pic base if not in PIC relocation model.
2071 HiOpFlags |= PPCII::MO_PIC_FLAG;
2072 LoOpFlags |= PPCII::MO_PIC_FLAG;
2075 // If this is a reference to a global value that requires a non-lazy-ptr, make
2076 // sure that instruction lowering adds it.
2077 if (GV && Subtarget.hasLazyResolverStub(GV)) {
2078 HiOpFlags |= PPCII::MO_NLP_FLAG;
2079 LoOpFlags |= PPCII::MO_NLP_FLAG;
2081 if (GV->hasHiddenVisibility()) {
2082 HiOpFlags |= PPCII::MO_NLP_HIDDEN_FLAG;
2083 LoOpFlags |= PPCII::MO_NLP_HIDDEN_FLAG;
2088 static SDValue LowerLabelRef(SDValue HiPart, SDValue LoPart, bool isPIC,
2089 SelectionDAG &DAG) {
2091 EVT PtrVT = HiPart.getValueType();
2092 SDValue Zero = DAG.getConstant(0, DL, PtrVT);
2094 SDValue Hi = DAG.getNode(PPCISD::Hi, DL, PtrVT, HiPart, Zero);
2095 SDValue Lo = DAG.getNode(PPCISD::Lo, DL, PtrVT, LoPart, Zero);
2097 // With PIC, the first instruction is actually "GR+hi(&G)".
2099 Hi = DAG.getNode(ISD::ADD, DL, PtrVT,
2100 DAG.getNode(PPCISD::GlobalBaseReg, DL, PtrVT), Hi);
2102 // Generate non-pic code that has direct accesses to the constant pool.
2103 // The address of the global is just (hi(&g)+lo(&g)).
2104 return DAG.getNode(ISD::ADD, DL, PtrVT, Hi, Lo);
2107 static void setUsesTOCBasePtr(MachineFunction &MF) {
2108 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
2109 FuncInfo->setUsesTOCBasePtr();
2112 static void setUsesTOCBasePtr(SelectionDAG &DAG) {
2113 setUsesTOCBasePtr(DAG.getMachineFunction());
2116 static SDValue getTOCEntry(SelectionDAG &DAG, const SDLoc &dl, bool Is64Bit,
2118 EVT VT = Is64Bit ? MVT::i64 : MVT::i32;
2119 SDValue Reg = Is64Bit ? DAG.getRegister(PPC::X2, VT) :
2120 DAG.getNode(PPCISD::GlobalBaseReg, dl, VT);
2122 SDValue Ops[] = { GA, Reg };
2123 return DAG.getMemIntrinsicNode(
2124 PPCISD::TOC_ENTRY, dl, DAG.getVTList(VT, MVT::Other), Ops, VT,
2125 MachinePointerInfo::getGOT(DAG.getMachineFunction()), 0, false, true,
2129 SDValue PPCTargetLowering::LowerConstantPool(SDValue Op,
2130 SelectionDAG &DAG) const {
2131 EVT PtrVT = Op.getValueType();
2132 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
2133 const Constant *C = CP->getConstVal();
2135 // 64-bit SVR4 ABI code is always position-independent.
2136 // The actual address of the GlobalValue is stored in the TOC.
2137 if (Subtarget.isSVR4ABI() && Subtarget.isPPC64()) {
2138 setUsesTOCBasePtr(DAG);
2139 SDValue GA = DAG.getTargetConstantPool(C, PtrVT, CP->getAlignment(), 0);
2140 return getTOCEntry(DAG, SDLoc(CP), true, GA);
2143 unsigned MOHiFlag, MOLoFlag;
2144 bool IsPIC = isPositionIndependent();
2145 getLabelAccessInfo(IsPIC, Subtarget, MOHiFlag, MOLoFlag);
2147 if (IsPIC && Subtarget.isSVR4ABI()) {
2148 SDValue GA = DAG.getTargetConstantPool(C, PtrVT, CP->getAlignment(),
2149 PPCII::MO_PIC_FLAG);
2150 return getTOCEntry(DAG, SDLoc(CP), false, GA);
2154 DAG.getTargetConstantPool(C, PtrVT, CP->getAlignment(), 0, MOHiFlag);
2156 DAG.getTargetConstantPool(C, PtrVT, CP->getAlignment(), 0, MOLoFlag);
2157 return LowerLabelRef(CPIHi, CPILo, IsPIC, DAG);
2160 SDValue PPCTargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
2161 EVT PtrVT = Op.getValueType();
2162 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
2164 // 64-bit SVR4 ABI code is always position-independent.
2165 // The actual address of the GlobalValue is stored in the TOC.
2166 if (Subtarget.isSVR4ABI() && Subtarget.isPPC64()) {
2167 setUsesTOCBasePtr(DAG);
2168 SDValue GA = DAG.getTargetJumpTable(JT->getIndex(), PtrVT);
2169 return getTOCEntry(DAG, SDLoc(JT), true, GA);
2172 unsigned MOHiFlag, MOLoFlag;
2173 bool IsPIC = isPositionIndependent();
2174 getLabelAccessInfo(IsPIC, Subtarget, MOHiFlag, MOLoFlag);
2176 if (IsPIC && Subtarget.isSVR4ABI()) {
2177 SDValue GA = DAG.getTargetJumpTable(JT->getIndex(), PtrVT,
2178 PPCII::MO_PIC_FLAG);
2179 return getTOCEntry(DAG, SDLoc(GA), false, GA);
2182 SDValue JTIHi = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, MOHiFlag);
2183 SDValue JTILo = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, MOLoFlag);
2184 return LowerLabelRef(JTIHi, JTILo, IsPIC, DAG);
2187 SDValue PPCTargetLowering::LowerBlockAddress(SDValue Op,
2188 SelectionDAG &DAG) const {
2189 EVT PtrVT = Op.getValueType();
2190 BlockAddressSDNode *BASDN = cast<BlockAddressSDNode>(Op);
2191 const BlockAddress *BA = BASDN->getBlockAddress();
2193 // 64-bit SVR4 ABI code is always position-independent.
2194 // The actual BlockAddress is stored in the TOC.
2195 if (Subtarget.isSVR4ABI() && Subtarget.isPPC64()) {
2196 setUsesTOCBasePtr(DAG);
2197 SDValue GA = DAG.getTargetBlockAddress(BA, PtrVT, BASDN->getOffset());
2198 return getTOCEntry(DAG, SDLoc(BASDN), true, GA);
2201 unsigned MOHiFlag, MOLoFlag;
2202 bool IsPIC = isPositionIndependent();
2203 getLabelAccessInfo(IsPIC, Subtarget, MOHiFlag, MOLoFlag);
2204 SDValue TgtBAHi = DAG.getTargetBlockAddress(BA, PtrVT, 0, MOHiFlag);
2205 SDValue TgtBALo = DAG.getTargetBlockAddress(BA, PtrVT, 0, MOLoFlag);
2206 return LowerLabelRef(TgtBAHi, TgtBALo, IsPIC, DAG);
2209 SDValue PPCTargetLowering::LowerGlobalTLSAddress(SDValue Op,
2210 SelectionDAG &DAG) const {
2212 // FIXME: TLS addresses currently use medium model code sequences,
2213 // which is the most useful form. Eventually support for small and
2214 // large models could be added if users need it, at the cost of
2215 // additional complexity.
2216 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
2217 if (DAG.getTarget().Options.EmulatedTLS)
2218 return LowerToTLSEmulatedModel(GA, DAG);
2221 const GlobalValue *GV = GA->getGlobal();
2222 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2223 bool is64bit = Subtarget.isPPC64();
2224 const Module *M = DAG.getMachineFunction().getFunction()->getParent();
2225 PICLevel::Level picLevel = M->getPICLevel();
2227 TLSModel::Model Model = getTargetMachine().getTLSModel(GV);
2229 if (Model == TLSModel::LocalExec) {
2230 SDValue TGAHi = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
2231 PPCII::MO_TPREL_HA);
2232 SDValue TGALo = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
2233 PPCII::MO_TPREL_LO);
2234 SDValue TLSReg = DAG.getRegister(is64bit ? PPC::X13 : PPC::R2,
2235 is64bit ? MVT::i64 : MVT::i32);
2236 SDValue Hi = DAG.getNode(PPCISD::Hi, dl, PtrVT, TGAHi, TLSReg);
2237 return DAG.getNode(PPCISD::Lo, dl, PtrVT, TGALo, Hi);
2240 if (Model == TLSModel::InitialExec) {
2241 SDValue TGA = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, 0);
2242 SDValue TGATLS = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
2246 setUsesTOCBasePtr(DAG);
2247 SDValue GOTReg = DAG.getRegister(PPC::X2, MVT::i64);
2248 GOTPtr = DAG.getNode(PPCISD::ADDIS_GOT_TPREL_HA, dl,
2249 PtrVT, GOTReg, TGA);
2251 GOTPtr = DAG.getNode(PPCISD::PPC32_GOT, dl, PtrVT);
2252 SDValue TPOffset = DAG.getNode(PPCISD::LD_GOT_TPREL_L, dl,
2253 PtrVT, TGA, GOTPtr);
2254 return DAG.getNode(PPCISD::ADD_TLS, dl, PtrVT, TPOffset, TGATLS);
2257 if (Model == TLSModel::GeneralDynamic) {
2258 SDValue TGA = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, 0);
2261 setUsesTOCBasePtr(DAG);
2262 SDValue GOTReg = DAG.getRegister(PPC::X2, MVT::i64);
2263 GOTPtr = DAG.getNode(PPCISD::ADDIS_TLSGD_HA, dl, PtrVT,
2266 if (picLevel == PICLevel::SmallPIC)
2267 GOTPtr = DAG.getNode(PPCISD::GlobalBaseReg, dl, PtrVT);
2269 GOTPtr = DAG.getNode(PPCISD::PPC32_PICGOT, dl, PtrVT);
2271 return DAG.getNode(PPCISD::ADDI_TLSGD_L_ADDR, dl, PtrVT,
2275 if (Model == TLSModel::LocalDynamic) {
2276 SDValue TGA = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, 0);
2279 setUsesTOCBasePtr(DAG);
2280 SDValue GOTReg = DAG.getRegister(PPC::X2, MVT::i64);
2281 GOTPtr = DAG.getNode(PPCISD::ADDIS_TLSLD_HA, dl, PtrVT,
2284 if (picLevel == PICLevel::SmallPIC)
2285 GOTPtr = DAG.getNode(PPCISD::GlobalBaseReg, dl, PtrVT);
2287 GOTPtr = DAG.getNode(PPCISD::PPC32_PICGOT, dl, PtrVT);
2289 SDValue TLSAddr = DAG.getNode(PPCISD::ADDI_TLSLD_L_ADDR, dl,
2290 PtrVT, GOTPtr, TGA, TGA);
2291 SDValue DtvOffsetHi = DAG.getNode(PPCISD::ADDIS_DTPREL_HA, dl,
2292 PtrVT, TLSAddr, TGA);
2293 return DAG.getNode(PPCISD::ADDI_DTPREL_L, dl, PtrVT, DtvOffsetHi, TGA);
2296 llvm_unreachable("Unknown TLS model!");
2299 SDValue PPCTargetLowering::LowerGlobalAddress(SDValue Op,
2300 SelectionDAG &DAG) const {
2301 EVT PtrVT = Op.getValueType();
2302 GlobalAddressSDNode *GSDN = cast<GlobalAddressSDNode>(Op);
2304 const GlobalValue *GV = GSDN->getGlobal();
2306 // 64-bit SVR4 ABI code is always position-independent.
2307 // The actual address of the GlobalValue is stored in the TOC.
2308 if (Subtarget.isSVR4ABI() && Subtarget.isPPC64()) {
2309 setUsesTOCBasePtr(DAG);
2310 SDValue GA = DAG.getTargetGlobalAddress(GV, DL, PtrVT, GSDN->getOffset());
2311 return getTOCEntry(DAG, DL, true, GA);
2314 unsigned MOHiFlag, MOLoFlag;
2315 bool IsPIC = isPositionIndependent();
2316 getLabelAccessInfo(IsPIC, Subtarget, MOHiFlag, MOLoFlag, GV);
2318 if (IsPIC && Subtarget.isSVR4ABI()) {
2319 SDValue GA = DAG.getTargetGlobalAddress(GV, DL, PtrVT,
2321 PPCII::MO_PIC_FLAG);
2322 return getTOCEntry(DAG, DL, false, GA);
2326 DAG.getTargetGlobalAddress(GV, DL, PtrVT, GSDN->getOffset(), MOHiFlag);
2328 DAG.getTargetGlobalAddress(GV, DL, PtrVT, GSDN->getOffset(), MOLoFlag);
2330 SDValue Ptr = LowerLabelRef(GAHi, GALo, IsPIC, DAG);
2332 // If the global reference is actually to a non-lazy-pointer, we have to do an
2333 // extra load to get the address of the global.
2334 if (MOHiFlag & PPCII::MO_NLP_FLAG)
2335 Ptr = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), Ptr, MachinePointerInfo());
2339 SDValue PPCTargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
2340 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
2343 if (Op.getValueType() == MVT::v2i64) {
2344 // When the operands themselves are v2i64 values, we need to do something
2345 // special because VSX has no underlying comparison operations for these.
2346 if (Op.getOperand(0).getValueType() == MVT::v2i64) {
2347 // Equality can be handled by casting to the legal type for Altivec
2348 // comparisons, everything else needs to be expanded.
2349 if (CC == ISD::SETEQ || CC == ISD::SETNE) {
2350 return DAG.getNode(ISD::BITCAST, dl, MVT::v2i64,
2351 DAG.getSetCC(dl, MVT::v4i32,
2352 DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Op.getOperand(0)),
2353 DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Op.getOperand(1)),
2360 // We handle most of these in the usual way.
2364 // If we're comparing for equality to zero, expose the fact that this is
2365 // implemented as a ctlz/srl pair on ppc, so that the dag combiner can
2366 // fold the new nodes.
2367 if (SDValue V = lowerCmpEqZeroToCtlzSrl(Op, DAG))
2370 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
2371 // Leave comparisons against 0 and -1 alone for now, since they're usually
2372 // optimized. FIXME: revisit this when we can custom lower all setcc
2374 if (C->isAllOnesValue() || C->isNullValue())
2378 // If we have an integer seteq/setne, turn it into a compare against zero
2379 // by xor'ing the rhs with the lhs, which is faster than setting a
2380 // condition register, reading it back out, and masking the correct bit. The
2381 // normal approach here uses sub to do this instead of xor. Using xor exposes
2382 // the result to other bit-twiddling opportunities.
2383 EVT LHSVT = Op.getOperand(0).getValueType();
2384 if (LHSVT.isInteger() && (CC == ISD::SETEQ || CC == ISD::SETNE)) {
2385 EVT VT = Op.getValueType();
2386 SDValue Sub = DAG.getNode(ISD::XOR, dl, LHSVT, Op.getOperand(0),
2388 return DAG.getSetCC(dl, VT, Sub, DAG.getConstant(0, dl, LHSVT), CC);
2393 SDValue PPCTargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
2394 SDNode *Node = Op.getNode();
2395 EVT VT = Node->getValueType(0);
2396 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2397 SDValue InChain = Node->getOperand(0);
2398 SDValue VAListPtr = Node->getOperand(1);
2399 const Value *SV = cast<SrcValueSDNode>(Node->getOperand(2))->getValue();
2402 assert(!Subtarget.isPPC64() && "LowerVAARG is PPC32 only");
2405 SDValue GprIndex = DAG.getExtLoad(ISD::ZEXTLOAD, dl, MVT::i32, InChain,
2406 VAListPtr, MachinePointerInfo(SV), MVT::i8);
2407 InChain = GprIndex.getValue(1);
2409 if (VT == MVT::i64) {
2410 // Check if GprIndex is even
2411 SDValue GprAnd = DAG.getNode(ISD::AND, dl, MVT::i32, GprIndex,
2412 DAG.getConstant(1, dl, MVT::i32));
2413 SDValue CC64 = DAG.getSetCC(dl, MVT::i32, GprAnd,
2414 DAG.getConstant(0, dl, MVT::i32), ISD::SETNE);
2415 SDValue GprIndexPlusOne = DAG.getNode(ISD::ADD, dl, MVT::i32, GprIndex,
2416 DAG.getConstant(1, dl, MVT::i32));
2417 // Align GprIndex to be even if it isn't
2418 GprIndex = DAG.getNode(ISD::SELECT, dl, MVT::i32, CC64, GprIndexPlusOne,
2422 // fpr index is 1 byte after gpr
2423 SDValue FprPtr = DAG.getNode(ISD::ADD, dl, PtrVT, VAListPtr,
2424 DAG.getConstant(1, dl, MVT::i32));
2427 SDValue FprIndex = DAG.getExtLoad(ISD::ZEXTLOAD, dl, MVT::i32, InChain,
2428 FprPtr, MachinePointerInfo(SV), MVT::i8);
2429 InChain = FprIndex.getValue(1);
2431 SDValue RegSaveAreaPtr = DAG.getNode(ISD::ADD, dl, PtrVT, VAListPtr,
2432 DAG.getConstant(8, dl, MVT::i32));
2434 SDValue OverflowAreaPtr = DAG.getNode(ISD::ADD, dl, PtrVT, VAListPtr,
2435 DAG.getConstant(4, dl, MVT::i32));
2438 SDValue OverflowArea =
2439 DAG.getLoad(MVT::i32, dl, InChain, OverflowAreaPtr, MachinePointerInfo());
2440 InChain = OverflowArea.getValue(1);
2442 SDValue RegSaveArea =
2443 DAG.getLoad(MVT::i32, dl, InChain, RegSaveAreaPtr, MachinePointerInfo());
2444 InChain = RegSaveArea.getValue(1);
2446 // select overflow_area if index > 8
2447 SDValue CC = DAG.getSetCC(dl, MVT::i32, VT.isInteger() ? GprIndex : FprIndex,
2448 DAG.getConstant(8, dl, MVT::i32), ISD::SETLT);
2450 // adjustment constant gpr_index * 4/8
2451 SDValue RegConstant = DAG.getNode(ISD::MUL, dl, MVT::i32,
2452 VT.isInteger() ? GprIndex : FprIndex,
2453 DAG.getConstant(VT.isInteger() ? 4 : 8, dl,
2456 // OurReg = RegSaveArea + RegConstant
2457 SDValue OurReg = DAG.getNode(ISD::ADD, dl, PtrVT, RegSaveArea,
2460 // Floating types are 32 bytes into RegSaveArea
2461 if (VT.isFloatingPoint())
2462 OurReg = DAG.getNode(ISD::ADD, dl, PtrVT, OurReg,
2463 DAG.getConstant(32, dl, MVT::i32));
2465 // increase {f,g}pr_index by 1 (or 2 if VT is i64)
2466 SDValue IndexPlus1 = DAG.getNode(ISD::ADD, dl, MVT::i32,
2467 VT.isInteger() ? GprIndex : FprIndex,
2468 DAG.getConstant(VT == MVT::i64 ? 2 : 1, dl,
2471 InChain = DAG.getTruncStore(InChain, dl, IndexPlus1,
2472 VT.isInteger() ? VAListPtr : FprPtr,
2473 MachinePointerInfo(SV), MVT::i8);
2475 // determine if we should load from reg_save_area or overflow_area
2476 SDValue Result = DAG.getNode(ISD::SELECT, dl, PtrVT, CC, OurReg, OverflowArea);
2478 // increase overflow_area by 4/8 if gpr/fpr > 8
2479 SDValue OverflowAreaPlusN = DAG.getNode(ISD::ADD, dl, PtrVT, OverflowArea,
2480 DAG.getConstant(VT.isInteger() ? 4 : 8,
2483 OverflowArea = DAG.getNode(ISD::SELECT, dl, MVT::i32, CC, OverflowArea,
2486 InChain = DAG.getTruncStore(InChain, dl, OverflowArea, OverflowAreaPtr,
2487 MachinePointerInfo(), MVT::i32);
2489 return DAG.getLoad(VT, dl, InChain, Result, MachinePointerInfo());
2492 SDValue PPCTargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) const {
2493 assert(!Subtarget.isPPC64() && "LowerVACOPY is PPC32 only");
2495 // We have to copy the entire va_list struct:
2496 // 2*sizeof(char) + 2 Byte alignment + 2*sizeof(char*) = 12 Byte
2497 return DAG.getMemcpy(Op.getOperand(0), Op,
2498 Op.getOperand(1), Op.getOperand(2),
2499 DAG.getConstant(12, SDLoc(Op), MVT::i32), 8, false, true,
2500 false, MachinePointerInfo(), MachinePointerInfo());
2503 SDValue PPCTargetLowering::LowerADJUST_TRAMPOLINE(SDValue Op,
2504 SelectionDAG &DAG) const {
2505 return Op.getOperand(0);
2508 SDValue PPCTargetLowering::LowerINIT_TRAMPOLINE(SDValue Op,
2509 SelectionDAG &DAG) const {
2510 SDValue Chain = Op.getOperand(0);
2511 SDValue Trmp = Op.getOperand(1); // trampoline
2512 SDValue FPtr = Op.getOperand(2); // nested function
2513 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
2516 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2517 bool isPPC64 = (PtrVT == MVT::i64);
2518 Type *IntPtrTy = DAG.getDataLayout().getIntPtrType(*DAG.getContext());
2520 TargetLowering::ArgListTy Args;
2521 TargetLowering::ArgListEntry Entry;
2523 Entry.Ty = IntPtrTy;
2524 Entry.Node = Trmp; Args.push_back(Entry);
2526 // TrampSize == (isPPC64 ? 48 : 40);
2527 Entry.Node = DAG.getConstant(isPPC64 ? 48 : 40, dl,
2528 isPPC64 ? MVT::i64 : MVT::i32);
2529 Args.push_back(Entry);
2531 Entry.Node = FPtr; Args.push_back(Entry);
2532 Entry.Node = Nest; Args.push_back(Entry);
2534 // Lower to a call to __trampoline_setup(Trmp, TrampSize, FPtr, ctx_reg)
2535 TargetLowering::CallLoweringInfo CLI(DAG);
2536 CLI.setDebugLoc(dl).setChain(Chain)
2537 .setCallee(CallingConv::C, Type::getVoidTy(*DAG.getContext()),
2538 DAG.getExternalSymbol("__trampoline_setup", PtrVT),
2541 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
2542 return CallResult.second;
2545 SDValue PPCTargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
2546 MachineFunction &MF = DAG.getMachineFunction();
2547 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
2548 EVT PtrVT = getPointerTy(MF.getDataLayout());
2552 if (Subtarget.isDarwinABI() || Subtarget.isPPC64()) {
2553 // vastart just stores the address of the VarArgsFrameIndex slot into the
2554 // memory location argument.
2555 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2556 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
2557 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
2558 MachinePointerInfo(SV));
2561 // For the 32-bit SVR4 ABI we follow the layout of the va_list struct.
2562 // We suppose the given va_list is already allocated.
2565 // char gpr; /* index into the array of 8 GPRs
2566 // * stored in the register save area
2567 // * gpr=0 corresponds to r3,
2568 // * gpr=1 to r4, etc.
2570 // char fpr; /* index into the array of 8 FPRs
2571 // * stored in the register save area
2572 // * fpr=0 corresponds to f1,
2573 // * fpr=1 to f2, etc.
2575 // char *overflow_arg_area;
2576 // /* location on stack that holds
2577 // * the next overflow argument
2579 // char *reg_save_area;
2580 // /* where r3:r10 and f1:f8 (if saved)
2585 SDValue ArgGPR = DAG.getConstant(FuncInfo->getVarArgsNumGPR(), dl, MVT::i32);
2586 SDValue ArgFPR = DAG.getConstant(FuncInfo->getVarArgsNumFPR(), dl, MVT::i32);
2587 SDValue StackOffsetFI = DAG.getFrameIndex(FuncInfo->getVarArgsStackOffset(),
2589 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
2592 uint64_t FrameOffset = PtrVT.getSizeInBits()/8;
2593 SDValue ConstFrameOffset = DAG.getConstant(FrameOffset, dl, PtrVT);
2595 uint64_t StackOffset = PtrVT.getSizeInBits()/8 - 1;
2596 SDValue ConstStackOffset = DAG.getConstant(StackOffset, dl, PtrVT);
2598 uint64_t FPROffset = 1;
2599 SDValue ConstFPROffset = DAG.getConstant(FPROffset, dl, PtrVT);
2601 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
2603 // Store first byte : number of int regs
2604 SDValue firstStore =
2605 DAG.getTruncStore(Op.getOperand(0), dl, ArgGPR, Op.getOperand(1),
2606 MachinePointerInfo(SV), MVT::i8);
2607 uint64_t nextOffset = FPROffset;
2608 SDValue nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, Op.getOperand(1),
2611 // Store second byte : number of float regs
2612 SDValue secondStore =
2613 DAG.getTruncStore(firstStore, dl, ArgFPR, nextPtr,
2614 MachinePointerInfo(SV, nextOffset), MVT::i8);
2615 nextOffset += StackOffset;
2616 nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, nextPtr, ConstStackOffset);
2618 // Store second word : arguments given on stack
2619 SDValue thirdStore = DAG.getStore(secondStore, dl, StackOffsetFI, nextPtr,
2620 MachinePointerInfo(SV, nextOffset));
2621 nextOffset += FrameOffset;
2622 nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, nextPtr, ConstFrameOffset);
2624 // Store third word : arguments given in registers
2625 return DAG.getStore(thirdStore, dl, FR, nextPtr,
2626 MachinePointerInfo(SV, nextOffset));
2629 #include "PPCGenCallingConv.inc"
2631 // Function whose sole purpose is to kill compiler warnings
2632 // stemming from unused functions included from PPCGenCallingConv.inc.
2633 CCAssignFn *PPCTargetLowering::useFastISelCCs(unsigned Flag) const {
2634 return Flag ? CC_PPC64_ELF_FIS : RetCC_PPC64_ELF_FIS;
2637 bool llvm::CC_PPC32_SVR4_Custom_Dummy(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
2638 CCValAssign::LocInfo &LocInfo,
2639 ISD::ArgFlagsTy &ArgFlags,
2644 bool llvm::CC_PPC32_SVR4_Custom_AlignArgRegs(unsigned &ValNo, MVT &ValVT,
2646 CCValAssign::LocInfo &LocInfo,
2647 ISD::ArgFlagsTy &ArgFlags,
2649 static const MCPhysReg ArgRegs[] = {
2650 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
2651 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
2653 const unsigned NumArgRegs = array_lengthof(ArgRegs);
2655 unsigned RegNum = State.getFirstUnallocated(ArgRegs);
2657 // Skip one register if the first unallocated register has an even register
2658 // number and there are still argument registers available which have not been
2659 // allocated yet. RegNum is actually an index into ArgRegs, which means we
2660 // need to skip a register if RegNum is odd.
2661 if (RegNum != NumArgRegs && RegNum % 2 == 1) {
2662 State.AllocateReg(ArgRegs[RegNum]);
2665 // Always return false here, as this function only makes sure that the first
2666 // unallocated register has an odd register number and does not actually
2667 // allocate a register for the current argument.
2672 llvm::CC_PPC32_SVR4_Custom_SkipLastArgRegsPPCF128(unsigned &ValNo, MVT &ValVT,
2674 CCValAssign::LocInfo &LocInfo,
2675 ISD::ArgFlagsTy &ArgFlags,
2677 static const MCPhysReg ArgRegs[] = {
2678 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
2679 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
2681 const unsigned NumArgRegs = array_lengthof(ArgRegs);
2683 unsigned RegNum = State.getFirstUnallocated(ArgRegs);
2684 int RegsLeft = NumArgRegs - RegNum;
2686 // Skip if there is not enough registers left for long double type (4 gpr regs
2687 // in soft float mode) and put long double argument on the stack.
2688 if (RegNum != NumArgRegs && RegsLeft < 4) {
2689 for (int i = 0; i < RegsLeft; i++) {
2690 State.AllocateReg(ArgRegs[RegNum + i]);
2697 bool llvm::CC_PPC32_SVR4_Custom_AlignFPArgRegs(unsigned &ValNo, MVT &ValVT,
2699 CCValAssign::LocInfo &LocInfo,
2700 ISD::ArgFlagsTy &ArgFlags,
2702 static const MCPhysReg ArgRegs[] = {
2703 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
2707 const unsigned NumArgRegs = array_lengthof(ArgRegs);
2709 unsigned RegNum = State.getFirstUnallocated(ArgRegs);
2711 // If there is only one Floating-point register left we need to put both f64
2712 // values of a split ppc_fp128 value on the stack.
2713 if (RegNum != NumArgRegs && ArgRegs[RegNum] == PPC::F8) {
2714 State.AllocateReg(ArgRegs[RegNum]);
2717 // Always return false here, as this function only makes sure that the two f64
2718 // values a ppc_fp128 value is split into are both passed in registers or both
2719 // passed on the stack and does not actually allocate a register for the
2720 // current argument.
2724 /// FPR - The set of FP registers that should be allocated for arguments,
2726 static const MCPhysReg FPR[] = {PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5,
2727 PPC::F6, PPC::F7, PPC::F8, PPC::F9, PPC::F10,
2728 PPC::F11, PPC::F12, PPC::F13};
2730 /// QFPR - The set of QPX registers that should be allocated for arguments.
2731 static const MCPhysReg QFPR[] = {
2732 PPC::QF1, PPC::QF2, PPC::QF3, PPC::QF4, PPC::QF5, PPC::QF6, PPC::QF7,
2733 PPC::QF8, PPC::QF9, PPC::QF10, PPC::QF11, PPC::QF12, PPC::QF13};
2735 /// CalculateStackSlotSize - Calculates the size reserved for this argument on
2737 static unsigned CalculateStackSlotSize(EVT ArgVT, ISD::ArgFlagsTy Flags,
2738 unsigned PtrByteSize) {
2739 unsigned ArgSize = ArgVT.getStoreSize();
2740 if (Flags.isByVal())
2741 ArgSize = Flags.getByValSize();
2743 // Round up to multiples of the pointer size, except for array members,
2744 // which are always packed.
2745 if (!Flags.isInConsecutiveRegs())
2746 ArgSize = ((ArgSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
2751 /// CalculateStackSlotAlignment - Calculates the alignment of this argument
2753 static unsigned CalculateStackSlotAlignment(EVT ArgVT, EVT OrigVT,
2754 ISD::ArgFlagsTy Flags,
2755 unsigned PtrByteSize) {
2756 unsigned Align = PtrByteSize;
2758 // Altivec parameters are padded to a 16 byte boundary.
2759 if (ArgVT == MVT::v4f32 || ArgVT == MVT::v4i32 ||
2760 ArgVT == MVT::v8i16 || ArgVT == MVT::v16i8 ||
2761 ArgVT == MVT::v2f64 || ArgVT == MVT::v2i64 ||
2762 ArgVT == MVT::v1i128)
2764 // QPX vector types stored in double-precision are padded to a 32 byte
2766 else if (ArgVT == MVT::v4f64 || ArgVT == MVT::v4i1)
2769 // ByVal parameters are aligned as requested.
2770 if (Flags.isByVal()) {
2771 unsigned BVAlign = Flags.getByValAlign();
2772 if (BVAlign > PtrByteSize) {
2773 if (BVAlign % PtrByteSize != 0)
2775 "ByVal alignment is not a multiple of the pointer size");
2781 // Array members are always packed to their original alignment.
2782 if (Flags.isInConsecutiveRegs()) {
2783 // If the array member was split into multiple registers, the first
2784 // needs to be aligned to the size of the full type. (Except for
2785 // ppcf128, which is only aligned as its f64 components.)
2786 if (Flags.isSplit() && OrigVT != MVT::ppcf128)
2787 Align = OrigVT.getStoreSize();
2789 Align = ArgVT.getStoreSize();
2795 /// CalculateStackSlotUsed - Return whether this argument will use its
2796 /// stack slot (instead of being passed in registers). ArgOffset,
2797 /// AvailableFPRs, and AvailableVRs must hold the current argument
2798 /// position, and will be updated to account for this argument.
2799 static bool CalculateStackSlotUsed(EVT ArgVT, EVT OrigVT,
2800 ISD::ArgFlagsTy Flags,
2801 unsigned PtrByteSize,
2802 unsigned LinkageSize,
2803 unsigned ParamAreaSize,
2804 unsigned &ArgOffset,
2805 unsigned &AvailableFPRs,
2806 unsigned &AvailableVRs, bool HasQPX) {
2807 bool UseMemory = false;
2809 // Respect alignment of argument on the stack.
2811 CalculateStackSlotAlignment(ArgVT, OrigVT, Flags, PtrByteSize);
2812 ArgOffset = ((ArgOffset + Align - 1) / Align) * Align;
2813 // If there's no space left in the argument save area, we must
2814 // use memory (this check also catches zero-sized arguments).
2815 if (ArgOffset >= LinkageSize + ParamAreaSize)
2818 // Allocate argument on the stack.
2819 ArgOffset += CalculateStackSlotSize(ArgVT, Flags, PtrByteSize);
2820 if (Flags.isInConsecutiveRegsLast())
2821 ArgOffset = ((ArgOffset + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
2822 // If we overran the argument save area, we must use memory
2823 // (this check catches arguments passed partially in memory)
2824 if (ArgOffset > LinkageSize + ParamAreaSize)
2827 // However, if the argument is actually passed in an FPR or a VR,
2828 // we don't use memory after all.
2829 if (!Flags.isByVal()) {
2830 if (ArgVT == MVT::f32 || ArgVT == MVT::f64 ||
2831 // QPX registers overlap with the scalar FP registers.
2832 (HasQPX && (ArgVT == MVT::v4f32 ||
2833 ArgVT == MVT::v4f64 ||
2834 ArgVT == MVT::v4i1)))
2835 if (AvailableFPRs > 0) {
2839 if (ArgVT == MVT::v4f32 || ArgVT == MVT::v4i32 ||
2840 ArgVT == MVT::v8i16 || ArgVT == MVT::v16i8 ||
2841 ArgVT == MVT::v2f64 || ArgVT == MVT::v2i64 ||
2842 ArgVT == MVT::v1i128)
2843 if (AvailableVRs > 0) {
2852 /// EnsureStackAlignment - Round stack frame size up from NumBytes to
2853 /// ensure minimum alignment required for target.
2854 static unsigned EnsureStackAlignment(const PPCFrameLowering *Lowering,
2855 unsigned NumBytes) {
2856 unsigned TargetAlign = Lowering->getStackAlignment();
2857 unsigned AlignMask = TargetAlign - 1;
2858 NumBytes = (NumBytes + AlignMask) & ~AlignMask;
2862 SDValue PPCTargetLowering::LowerFormalArguments(
2863 SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
2864 const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl,
2865 SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const {
2866 if (Subtarget.isSVR4ABI()) {
2867 if (Subtarget.isPPC64())
2868 return LowerFormalArguments_64SVR4(Chain, CallConv, isVarArg, Ins,
2871 return LowerFormalArguments_32SVR4(Chain, CallConv, isVarArg, Ins,
2874 return LowerFormalArguments_Darwin(Chain, CallConv, isVarArg, Ins,
2879 SDValue PPCTargetLowering::LowerFormalArguments_32SVR4(
2880 SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
2881 const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl,
2882 SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const {
2884 // 32-bit SVR4 ABI Stack Frame Layout:
2885 // +-----------------------------------+
2886 // +--> | Back chain |
2887 // | +-----------------------------------+
2888 // | | Floating-point register save area |
2889 // | +-----------------------------------+
2890 // | | General register save area |
2891 // | +-----------------------------------+
2892 // | | CR save word |
2893 // | +-----------------------------------+
2894 // | | VRSAVE save word |
2895 // | +-----------------------------------+
2896 // | | Alignment padding |
2897 // | +-----------------------------------+
2898 // | | Vector register save area |
2899 // | +-----------------------------------+
2900 // | | Local variable space |
2901 // | +-----------------------------------+
2902 // | | Parameter list area |
2903 // | +-----------------------------------+
2904 // | | LR save word |
2905 // | +-----------------------------------+
2906 // SP--> +--- | Back chain |
2907 // +-----------------------------------+
2910 // System V Application Binary Interface PowerPC Processor Supplement
2911 // AltiVec Technology Programming Interface Manual
2913 MachineFunction &MF = DAG.getMachineFunction();
2914 MachineFrameInfo &MFI = MF.getFrameInfo();
2915 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
2917 EVT PtrVT = getPointerTy(MF.getDataLayout());
2918 // Potential tail calls could cause overwriting of argument stack slots.
2919 bool isImmutable = !(getTargetMachine().Options.GuaranteedTailCallOpt &&
2920 (CallConv == CallingConv::Fast));
2921 unsigned PtrByteSize = 4;
2923 // Assign locations to all of the incoming arguments.
2924 SmallVector<CCValAssign, 16> ArgLocs;
2925 PPCCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs,
2928 // Reserve space for the linkage area on the stack.
2929 unsigned LinkageSize = Subtarget.getFrameLowering()->getLinkageSize();
2930 CCInfo.AllocateStack(LinkageSize, PtrByteSize);
2932 CCInfo.PreAnalyzeFormalArguments(Ins);
2934 CCInfo.AnalyzeFormalArguments(Ins, CC_PPC32_SVR4);
2935 CCInfo.clearWasPPCF128();
2937 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2938 CCValAssign &VA = ArgLocs[i];
2940 // Arguments stored in registers.
2941 if (VA.isRegLoc()) {
2942 const TargetRegisterClass *RC;
2943 EVT ValVT = VA.getValVT();
2945 switch (ValVT.getSimpleVT().SimpleTy) {
2947 llvm_unreachable("ValVT not supported by formal arguments Lowering");
2950 RC = &PPC::GPRCRegClass;
2953 if (Subtarget.hasP8Vector())
2954 RC = &PPC::VSSRCRegClass;
2956 RC = &PPC::F4RCRegClass;
2959 if (Subtarget.hasVSX())
2960 RC = &PPC::VSFRCRegClass;
2962 RC = &PPC::F8RCRegClass;
2967 RC = &PPC::VRRCRegClass;
2970 RC = Subtarget.hasQPX() ? &PPC::QSRCRegClass : &PPC::VRRCRegClass;
2974 RC = &PPC::VSHRCRegClass;
2977 RC = &PPC::QFRCRegClass;
2980 RC = &PPC::QBRCRegClass;
2984 // Transform the arguments stored in physical registers into virtual ones.
2985 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2986 SDValue ArgValue = DAG.getCopyFromReg(Chain, dl, Reg,
2987 ValVT == MVT::i1 ? MVT::i32 : ValVT);
2989 if (ValVT == MVT::i1)
2990 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, ArgValue);
2992 InVals.push_back(ArgValue);
2994 // Argument stored in memory.
2995 assert(VA.isMemLoc());
2997 unsigned ArgSize = VA.getLocVT().getStoreSize();
2998 int FI = MFI.CreateFixedObject(ArgSize, VA.getLocMemOffset(),
3001 // Create load nodes to retrieve arguments from the stack.
3002 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
3004 DAG.getLoad(VA.getValVT(), dl, Chain, FIN, MachinePointerInfo()));
3008 // Assign locations to all of the incoming aggregate by value arguments.
3009 // Aggregates passed by value are stored in the local variable space of the
3010 // caller's stack frame, right above the parameter list area.
3011 SmallVector<CCValAssign, 16> ByValArgLocs;
3012 CCState CCByValInfo(CallConv, isVarArg, DAG.getMachineFunction(),
3013 ByValArgLocs, *DAG.getContext());
3015 // Reserve stack space for the allocations in CCInfo.
3016 CCByValInfo.AllocateStack(CCInfo.getNextStackOffset(), PtrByteSize);
3018 CCByValInfo.AnalyzeFormalArguments(Ins, CC_PPC32_SVR4_ByVal);
3020 // Area that is at least reserved in the caller of this function.
3021 unsigned MinReservedArea = CCByValInfo.getNextStackOffset();
3022 MinReservedArea = std::max(MinReservedArea, LinkageSize);
3024 // Set the size that is at least reserved in caller of this function. Tail
3025 // call optimized function's reserved stack space needs to be aligned so that
3026 // taking the difference between two stack areas will result in an aligned
3029 EnsureStackAlignment(Subtarget.getFrameLowering(), MinReservedArea);
3030 FuncInfo->setMinReservedArea(MinReservedArea);
3032 SmallVector<SDValue, 8> MemOps;
3034 // If the function takes variable number of arguments, make a frame index for
3035 // the start of the first vararg value... for expansion of llvm.va_start.
3037 static const MCPhysReg GPArgRegs[] = {
3038 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
3039 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
3041 const unsigned NumGPArgRegs = array_lengthof(GPArgRegs);
3043 static const MCPhysReg FPArgRegs[] = {
3044 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
3047 unsigned NumFPArgRegs = array_lengthof(FPArgRegs);
3052 FuncInfo->setVarArgsNumGPR(CCInfo.getFirstUnallocated(GPArgRegs));
3053 FuncInfo->setVarArgsNumFPR(CCInfo.getFirstUnallocated(FPArgRegs));
3055 // Make room for NumGPArgRegs and NumFPArgRegs.
3056 int Depth = NumGPArgRegs * PtrVT.getSizeInBits()/8 +
3057 NumFPArgRegs * MVT(MVT::f64).getSizeInBits()/8;
3059 FuncInfo->setVarArgsStackOffset(
3060 MFI.CreateFixedObject(PtrVT.getSizeInBits()/8,
3061 CCInfo.getNextStackOffset(), true));
3063 FuncInfo->setVarArgsFrameIndex(MFI.CreateStackObject(Depth, 8, false));
3064 SDValue FIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
3066 // The fixed integer arguments of a variadic function are stored to the
3067 // VarArgsFrameIndex on the stack so that they may be loaded by
3068 // dereferencing the result of va_next.
3069 for (unsigned GPRIndex = 0; GPRIndex != NumGPArgRegs; ++GPRIndex) {
3070 // Get an existing live-in vreg, or add a new one.
3071 unsigned VReg = MF.getRegInfo().getLiveInVirtReg(GPArgRegs[GPRIndex]);
3073 VReg = MF.addLiveIn(GPArgRegs[GPRIndex], &PPC::GPRCRegClass);
3075 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
3077 DAG.getStore(Val.getValue(1), dl, Val, FIN, MachinePointerInfo());
3078 MemOps.push_back(Store);
3079 // Increment the address by four for the next argument to store
3080 SDValue PtrOff = DAG.getConstant(PtrVT.getSizeInBits()/8, dl, PtrVT);
3081 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
3084 // FIXME 32-bit SVR4: We only need to save FP argument registers if CR bit 6
3086 // The double arguments are stored to the VarArgsFrameIndex
3088 for (unsigned FPRIndex = 0; FPRIndex != NumFPArgRegs; ++FPRIndex) {
3089 // Get an existing live-in vreg, or add a new one.
3090 unsigned VReg = MF.getRegInfo().getLiveInVirtReg(FPArgRegs[FPRIndex]);
3092 VReg = MF.addLiveIn(FPArgRegs[FPRIndex], &PPC::F8RCRegClass);
3094 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::f64);
3096 DAG.getStore(Val.getValue(1), dl, Val, FIN, MachinePointerInfo());
3097 MemOps.push_back(Store);
3098 // Increment the address by eight for the next argument to store
3099 SDValue PtrOff = DAG.getConstant(MVT(MVT::f64).getSizeInBits()/8, dl,
3101 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
3105 if (!MemOps.empty())
3106 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOps);
3111 // PPC64 passes i8, i16, and i32 values in i64 registers. Promote
3112 // value to MVT::i64 and then truncate to the correct register size.
3113 SDValue PPCTargetLowering::extendArgForPPC64(ISD::ArgFlagsTy Flags,
3114 EVT ObjectVT, SelectionDAG &DAG,
3116 const SDLoc &dl) const {
3118 ArgVal = DAG.getNode(ISD::AssertSext, dl, MVT::i64, ArgVal,
3119 DAG.getValueType(ObjectVT));
3120 else if (Flags.isZExt())
3121 ArgVal = DAG.getNode(ISD::AssertZext, dl, MVT::i64, ArgVal,
3122 DAG.getValueType(ObjectVT));
3124 return DAG.getNode(ISD::TRUNCATE, dl, ObjectVT, ArgVal);
3127 SDValue PPCTargetLowering::LowerFormalArguments_64SVR4(
3128 SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
3129 const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl,
3130 SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const {
3131 // TODO: add description of PPC stack frame format, or at least some docs.
3133 bool isELFv2ABI = Subtarget.isELFv2ABI();
3134 bool isLittleEndian = Subtarget.isLittleEndian();
3135 MachineFunction &MF = DAG.getMachineFunction();
3136 MachineFrameInfo &MFI = MF.getFrameInfo();
3137 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
3139 assert(!(CallConv == CallingConv::Fast && isVarArg) &&
3140 "fastcc not supported on varargs functions");
3142 EVT PtrVT = getPointerTy(MF.getDataLayout());
3143 // Potential tail calls could cause overwriting of argument stack slots.
3144 bool isImmutable = !(getTargetMachine().Options.GuaranteedTailCallOpt &&
3145 (CallConv == CallingConv::Fast));
3146 unsigned PtrByteSize = 8;
3147 unsigned LinkageSize = Subtarget.getFrameLowering()->getLinkageSize();
3149 static const MCPhysReg GPR[] = {
3150 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
3151 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
3153 static const MCPhysReg VR[] = {
3154 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
3155 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
3157 static const MCPhysReg VSRH[] = {
3158 PPC::VSH2, PPC::VSH3, PPC::VSH4, PPC::VSH5, PPC::VSH6, PPC::VSH7, PPC::VSH8,
3159 PPC::VSH9, PPC::VSH10, PPC::VSH11, PPC::VSH12, PPC::VSH13
3162 const unsigned Num_GPR_Regs = array_lengthof(GPR);
3163 const unsigned Num_FPR_Regs = 13;
3164 const unsigned Num_VR_Regs = array_lengthof(VR);
3165 const unsigned Num_QFPR_Regs = Num_FPR_Regs;
3167 // Do a first pass over the arguments to determine whether the ABI
3168 // guarantees that our caller has allocated the parameter save area
3169 // on its stack frame. In the ELFv1 ABI, this is always the case;
3170 // in the ELFv2 ABI, it is true if this is a vararg function or if
3171 // any parameter is located in a stack slot.
3173 bool HasParameterArea = !isELFv2ABI || isVarArg;
3174 unsigned ParamAreaSize = Num_GPR_Regs * PtrByteSize;
3175 unsigned NumBytes = LinkageSize;
3176 unsigned AvailableFPRs = Num_FPR_Regs;
3177 unsigned AvailableVRs = Num_VR_Regs;
3178 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
3179 if (Ins[i].Flags.isNest())
3182 if (CalculateStackSlotUsed(Ins[i].VT, Ins[i].ArgVT, Ins[i].Flags,
3183 PtrByteSize, LinkageSize, ParamAreaSize,
3184 NumBytes, AvailableFPRs, AvailableVRs,
3185 Subtarget.hasQPX()))
3186 HasParameterArea = true;
3189 // Add DAG nodes to load the arguments or copy them out of registers. On
3190 // entry to a function on PPC, the arguments start after the linkage area,
3191 // although the first ones are often in registers.
3193 unsigned ArgOffset = LinkageSize;
3194 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
3195 unsigned &QFPR_idx = FPR_idx;
3196 SmallVector<SDValue, 8> MemOps;
3197 Function::const_arg_iterator FuncArg = MF.getFunction()->arg_begin();
3198 unsigned CurArgIdx = 0;
3199 for (unsigned ArgNo = 0, e = Ins.size(); ArgNo != e; ++ArgNo) {
3201 bool needsLoad = false;
3202 EVT ObjectVT = Ins[ArgNo].VT;
3203 EVT OrigVT = Ins[ArgNo].ArgVT;
3204 unsigned ObjSize = ObjectVT.getStoreSize();
3205 unsigned ArgSize = ObjSize;
3206 ISD::ArgFlagsTy Flags = Ins[ArgNo].Flags;
3207 if (Ins[ArgNo].isOrigArg()) {
3208 std::advance(FuncArg, Ins[ArgNo].getOrigArgIndex() - CurArgIdx);
3209 CurArgIdx = Ins[ArgNo].getOrigArgIndex();
3211 // We re-align the argument offset for each argument, except when using the
3212 // fast calling convention, when we need to make sure we do that only when
3213 // we'll actually use a stack slot.
3214 unsigned CurArgOffset, Align;
3215 auto ComputeArgOffset = [&]() {
3216 /* Respect alignment of argument on the stack. */
3217 Align = CalculateStackSlotAlignment(ObjectVT, OrigVT, Flags, PtrByteSize);
3218 ArgOffset = ((ArgOffset + Align - 1) / Align) * Align;
3219 CurArgOffset = ArgOffset;
3222 if (CallConv != CallingConv::Fast) {
3225 /* Compute GPR index associated with argument offset. */
3226 GPR_idx = (ArgOffset - LinkageSize) / PtrByteSize;
3227 GPR_idx = std::min(GPR_idx, Num_GPR_Regs);
3230 // FIXME the codegen can be much improved in some cases.
3231 // We do not have to keep everything in memory.
3232 if (Flags.isByVal()) {
3233 assert(Ins[ArgNo].isOrigArg() && "Byval arguments cannot be implicit");
3235 if (CallConv == CallingConv::Fast)
3238 // ObjSize is the true size, ArgSize rounded up to multiple of registers.
3239 ObjSize = Flags.getByValSize();
3240 ArgSize = ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
3241 // Empty aggregate parameters do not take up registers. Examples:
3245 // etc. However, we have to provide a place-holder in InVals, so
3246 // pretend we have an 8-byte item at the current address for that
3249 int FI = MFI.CreateFixedObject(PtrByteSize, ArgOffset, true);
3250 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
3251 InVals.push_back(FIN);
3255 // Create a stack object covering all stack doublewords occupied
3256 // by the argument. If the argument is (fully or partially) on
3257 // the stack, or if the argument is fully in registers but the
3258 // caller has allocated the parameter save anyway, we can refer
3259 // directly to the caller's stack frame. Otherwise, create a
3260 // local copy in our own frame.
3262 if (HasParameterArea ||
3263 ArgSize + ArgOffset > LinkageSize + Num_GPR_Regs * PtrByteSize)
3264 FI = MFI.CreateFixedObject(ArgSize, ArgOffset, false, true);
3266 FI = MFI.CreateStackObject(ArgSize, Align, false);
3267 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
3269 // Handle aggregates smaller than 8 bytes.
3270 if (ObjSize < PtrByteSize) {
3271 // The value of the object is its address, which differs from the
3272 // address of the enclosing doubleword on big-endian systems.
3274 if (!isLittleEndian) {
3275 SDValue ArgOff = DAG.getConstant(PtrByteSize - ObjSize, dl, PtrVT);
3276 Arg = DAG.getNode(ISD::ADD, dl, ArgOff.getValueType(), Arg, ArgOff);
3278 InVals.push_back(Arg);
3280 if (GPR_idx != Num_GPR_Regs) {
3281 unsigned VReg = MF.addLiveIn(GPR[GPR_idx++], &PPC::G8RCRegClass);
3282 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
3285 if (ObjSize==1 || ObjSize==2 || ObjSize==4) {
3286 EVT ObjType = (ObjSize == 1 ? MVT::i8 :
3287 (ObjSize == 2 ? MVT::i16 : MVT::i32));
3288 Store = DAG.getTruncStore(Val.getValue(1), dl, Val, Arg,
3289 MachinePointerInfo(&*FuncArg), ObjType);
3291 // For sizes that don't fit a truncating store (3, 5, 6, 7),
3292 // store the whole register as-is to the parameter save area
3294 Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
3295 MachinePointerInfo(&*FuncArg));
3298 MemOps.push_back(Store);
3300 // Whether we copied from a register or not, advance the offset
3301 // into the parameter save area by a full doubleword.
3302 ArgOffset += PtrByteSize;
3306 // The value of the object is its address, which is the address of
3307 // its first stack doubleword.
3308 InVals.push_back(FIN);
3310 // Store whatever pieces of the object are in registers to memory.
3311 for (unsigned j = 0; j < ArgSize; j += PtrByteSize) {
3312 if (GPR_idx == Num_GPR_Regs)
3315 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
3316 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
3319 SDValue Off = DAG.getConstant(j, dl, PtrVT);
3320 Addr = DAG.getNode(ISD::ADD, dl, Off.getValueType(), Addr, Off);
3322 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, Addr,
3323 MachinePointerInfo(&*FuncArg, j));
3324 MemOps.push_back(Store);
3327 ArgOffset += ArgSize;
3331 switch (ObjectVT.getSimpleVT().SimpleTy) {
3332 default: llvm_unreachable("Unhandled argument type!");
3336 if (Flags.isNest()) {
3337 // The 'nest' parameter, if any, is passed in R11.
3338 unsigned VReg = MF.addLiveIn(PPC::X11, &PPC::G8RCRegClass);
3339 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
3341 if (ObjectVT == MVT::i32 || ObjectVT == MVT::i1)
3342 ArgVal = extendArgForPPC64(Flags, ObjectVT, DAG, ArgVal, dl);
3347 // These can be scalar arguments or elements of an integer array type
3348 // passed directly. Clang may use those instead of "byval" aggregate
3349 // types to avoid forcing arguments to memory unnecessarily.
3350 if (GPR_idx != Num_GPR_Regs) {
3351 unsigned VReg = MF.addLiveIn(GPR[GPR_idx++], &PPC::G8RCRegClass);
3352 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
3354 if (ObjectVT == MVT::i32 || ObjectVT == MVT::i1)
3355 // PPC64 passes i8, i16, and i32 values in i64 registers. Promote
3356 // value to MVT::i64 and then truncate to the correct register size.
3357 ArgVal = extendArgForPPC64(Flags, ObjectVT, DAG, ArgVal, dl);
3359 if (CallConv == CallingConv::Fast)
3363 ArgSize = PtrByteSize;
3365 if (CallConv != CallingConv::Fast || needsLoad)
3371 // These can be scalar arguments or elements of a float array type
3372 // passed directly. The latter are used to implement ELFv2 homogenous
3373 // float aggregates.
3374 if (FPR_idx != Num_FPR_Regs) {
3377 if (ObjectVT == MVT::f32)
3378 VReg = MF.addLiveIn(FPR[FPR_idx],
3379 Subtarget.hasP8Vector()
3380 ? &PPC::VSSRCRegClass
3381 : &PPC::F4RCRegClass);
3383 VReg = MF.addLiveIn(FPR[FPR_idx], Subtarget.hasVSX()
3384 ? &PPC::VSFRCRegClass
3385 : &PPC::F8RCRegClass);
3387 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT);
3389 } else if (GPR_idx != Num_GPR_Regs && CallConv != CallingConv::Fast) {
3390 // FIXME: We may want to re-enable this for CallingConv::Fast on the P8
3391 // once we support fp <-> gpr moves.
3393 // This can only ever happen in the presence of f32 array types,
3394 // since otherwise we never run out of FPRs before running out
3396 unsigned VReg = MF.addLiveIn(GPR[GPR_idx++], &PPC::G8RCRegClass);
3397 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
3399 if (ObjectVT == MVT::f32) {
3400 if ((ArgOffset % PtrByteSize) == (isLittleEndian ? 4 : 0))
3401 ArgVal = DAG.getNode(ISD::SRL, dl, MVT::i64, ArgVal,
3402 DAG.getConstant(32, dl, MVT::i32));
3403 ArgVal = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, ArgVal);
3406 ArgVal = DAG.getNode(ISD::BITCAST, dl, ObjectVT, ArgVal);
3408 if (CallConv == CallingConv::Fast)
3414 // When passing an array of floats, the array occupies consecutive
3415 // space in the argument area; only round up to the next doubleword
3416 // at the end of the array. Otherwise, each float takes 8 bytes.
3417 if (CallConv != CallingConv::Fast || needsLoad) {
3418 ArgSize = Flags.isInConsecutiveRegs() ? ObjSize : PtrByteSize;
3419 ArgOffset += ArgSize;
3420 if (Flags.isInConsecutiveRegsLast())
3421 ArgOffset = ((ArgOffset + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
3431 if (!Subtarget.hasQPX()) {
3432 // These can be scalar arguments or elements of a vector array type
3433 // passed directly. The latter are used to implement ELFv2 homogenous
3434 // vector aggregates.
3435 if (VR_idx != Num_VR_Regs) {
3436 unsigned VReg = (ObjectVT == MVT::v2f64 || ObjectVT == MVT::v2i64) ?
3437 MF.addLiveIn(VSRH[VR_idx], &PPC::VSHRCRegClass) :
3438 MF.addLiveIn(VR[VR_idx], &PPC::VRRCRegClass);
3439 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT);
3442 if (CallConv == CallingConv::Fast)
3447 if (CallConv != CallingConv::Fast || needsLoad)
3452 assert(ObjectVT.getSimpleVT().SimpleTy == MVT::v4f32 &&
3453 "Invalid QPX parameter type");
3458 // QPX vectors are treated like their scalar floating-point subregisters
3459 // (except that they're larger).
3460 unsigned Sz = ObjectVT.getSimpleVT().SimpleTy == MVT::v4f32 ? 16 : 32;
3461 if (QFPR_idx != Num_QFPR_Regs) {
3462 const TargetRegisterClass *RC;
3463 switch (ObjectVT.getSimpleVT().SimpleTy) {
3464 case MVT::v4f64: RC = &PPC::QFRCRegClass; break;
3465 case MVT::v4f32: RC = &PPC::QSRCRegClass; break;
3466 default: RC = &PPC::QBRCRegClass; break;
3469 unsigned VReg = MF.addLiveIn(QFPR[QFPR_idx], RC);
3470 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT);
3473 if (CallConv == CallingConv::Fast)
3477 if (CallConv != CallingConv::Fast || needsLoad)
3482 // We need to load the argument to a virtual register if we determined
3483 // above that we ran out of physical registers of the appropriate type.
3485 if (ObjSize < ArgSize && !isLittleEndian)
3486 CurArgOffset += ArgSize - ObjSize;
3487 int FI = MFI.CreateFixedObject(ObjSize, CurArgOffset, isImmutable);
3488 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
3489 ArgVal = DAG.getLoad(ObjectVT, dl, Chain, FIN, MachinePointerInfo());
3492 InVals.push_back(ArgVal);
3495 // Area that is at least reserved in the caller of this function.
3496 unsigned MinReservedArea;
3497 if (HasParameterArea)
3498 MinReservedArea = std::max(ArgOffset, LinkageSize + 8 * PtrByteSize);
3500 MinReservedArea = LinkageSize;
3502 // Set the size that is at least reserved in caller of this function. Tail
3503 // call optimized functions' reserved stack space needs to be aligned so that
3504 // taking the difference between two stack areas will result in an aligned
3507 EnsureStackAlignment(Subtarget.getFrameLowering(), MinReservedArea);
3508 FuncInfo->setMinReservedArea(MinReservedArea);
3510 // If the function takes variable number of arguments, make a frame index for
3511 // the start of the first vararg value... for expansion of llvm.va_start.
3513 int Depth = ArgOffset;
3515 FuncInfo->setVarArgsFrameIndex(
3516 MFI.CreateFixedObject(PtrByteSize, Depth, true));
3517 SDValue FIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
3519 // If this function is vararg, store any remaining integer argument regs
3520 // to their spots on the stack so that they may be loaded by dereferencing
3521 // the result of va_next.
3522 for (GPR_idx = (ArgOffset - LinkageSize) / PtrByteSize;
3523 GPR_idx < Num_GPR_Regs; ++GPR_idx) {
3524 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
3525 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
3527 DAG.getStore(Val.getValue(1), dl, Val, FIN, MachinePointerInfo());
3528 MemOps.push_back(Store);
3529 // Increment the address by four for the next argument to store
3530 SDValue PtrOff = DAG.getConstant(PtrByteSize, dl, PtrVT);
3531 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
3535 if (!MemOps.empty())
3536 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOps);
3541 SDValue PPCTargetLowering::LowerFormalArguments_Darwin(
3542 SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
3543 const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl,
3544 SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const {
3545 // TODO: add description of PPC stack frame format, or at least some docs.
3547 MachineFunction &MF = DAG.getMachineFunction();
3548 MachineFrameInfo &MFI = MF.getFrameInfo();
3549 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
3551 EVT PtrVT = getPointerTy(MF.getDataLayout());
3552 bool isPPC64 = PtrVT == MVT::i64;
3553 // Potential tail calls could cause overwriting of argument stack slots.
3554 bool isImmutable = !(getTargetMachine().Options.GuaranteedTailCallOpt &&
3555 (CallConv == CallingConv::Fast));
3556 unsigned PtrByteSize = isPPC64 ? 8 : 4;
3557 unsigned LinkageSize = Subtarget.getFrameLowering()->getLinkageSize();
3558 unsigned ArgOffset = LinkageSize;
3559 // Area that is at least reserved in caller of this function.
3560 unsigned MinReservedArea = ArgOffset;
3562 static const MCPhysReg GPR_32[] = { // 32-bit registers.
3563 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
3564 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
3566 static const MCPhysReg GPR_64[] = { // 64-bit registers.
3567 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
3568 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
3570 static const MCPhysReg VR[] = {
3571 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
3572 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
3575 const unsigned Num_GPR_Regs = array_lengthof(GPR_32);
3576 const unsigned Num_FPR_Regs = 13;
3577 const unsigned Num_VR_Regs = array_lengthof( VR);
3579 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
3581 const MCPhysReg *GPR = isPPC64 ? GPR_64 : GPR_32;
3583 // In 32-bit non-varargs functions, the stack space for vectors is after the
3584 // stack space for non-vectors. We do not use this space unless we have
3585 // too many vectors to fit in registers, something that only occurs in
3586 // constructed examples:), but we have to walk the arglist to figure
3587 // that out...for the pathological case, compute VecArgOffset as the
3588 // start of the vector parameter area. Computing VecArgOffset is the
3589 // entire point of the following loop.
3590 unsigned VecArgOffset = ArgOffset;
3591 if (!isVarArg && !isPPC64) {
3592 for (unsigned ArgNo = 0, e = Ins.size(); ArgNo != e;
3594 EVT ObjectVT = Ins[ArgNo].VT;
3595 ISD::ArgFlagsTy Flags = Ins[ArgNo].Flags;
3597 if (Flags.isByVal()) {
3598 // ObjSize is the true size, ArgSize rounded up to multiple of regs.
3599 unsigned ObjSize = Flags.getByValSize();
3601 ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
3602 VecArgOffset += ArgSize;
3606 switch(ObjectVT.getSimpleVT().SimpleTy) {
3607 default: llvm_unreachable("Unhandled argument type!");
3613 case MVT::i64: // PPC64
3615 // FIXME: We are guaranteed to be !isPPC64 at this point.
3616 // Does MVT::i64 apply?
3623 // Nothing to do, we're only looking at Nonvector args here.
3628 // We've found where the vector parameter area in memory is. Skip the
3629 // first 12 parameters; these don't use that memory.
3630 VecArgOffset = ((VecArgOffset+15)/16)*16;
3631 VecArgOffset += 12*16;
3633 // Add DAG nodes to load the arguments or copy them out of registers. On
3634 // entry to a function on PPC, the arguments start after the linkage area,
3635 // although the first ones are often in registers.
3637 SmallVector<SDValue, 8> MemOps;
3638 unsigned nAltivecParamsAtEnd = 0;
3639 Function::const_arg_iterator FuncArg = MF.getFunction()->arg_begin();
3640 unsigned CurArgIdx = 0;
3641 for (unsigned ArgNo = 0, e = Ins.size(); ArgNo != e; ++ArgNo) {
3643 bool needsLoad = false;
3644 EVT ObjectVT = Ins[ArgNo].VT;
3645 unsigned ObjSize = ObjectVT.getSizeInBits()/8;
3646 unsigned ArgSize = ObjSize;
3647 ISD::ArgFlagsTy Flags = Ins[ArgNo].Flags;
3648 if (Ins[ArgNo].isOrigArg()) {
3649 std::advance(FuncArg, Ins[ArgNo].getOrigArgIndex() - CurArgIdx);
3650 CurArgIdx = Ins[ArgNo].getOrigArgIndex();
3652 unsigned CurArgOffset = ArgOffset;
3654 // Varargs or 64 bit Altivec parameters are padded to a 16 byte boundary.
3655 if (ObjectVT==MVT::v4f32 || ObjectVT==MVT::v4i32 ||
3656 ObjectVT==MVT::v8i16 || ObjectVT==MVT::v16i8) {
3657 if (isVarArg || isPPC64) {
3658 MinReservedArea = ((MinReservedArea+15)/16)*16;
3659 MinReservedArea += CalculateStackSlotSize(ObjectVT,
3662 } else nAltivecParamsAtEnd++;
3664 // Calculate min reserved area.
3665 MinReservedArea += CalculateStackSlotSize(Ins[ArgNo].VT,
3669 // FIXME the codegen can be much improved in some cases.
3670 // We do not have to keep everything in memory.
3671 if (Flags.isByVal()) {
3672 assert(Ins[ArgNo].isOrigArg() && "Byval arguments cannot be implicit");
3674 // ObjSize is the true size, ArgSize rounded up to multiple of registers.
3675 ObjSize = Flags.getByValSize();
3676 ArgSize = ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
3677 // Objects of size 1 and 2 are right justified, everything else is
3678 // left justified. This means the memory address is adjusted forwards.
3679 if (ObjSize==1 || ObjSize==2) {
3680 CurArgOffset = CurArgOffset + (4 - ObjSize);
3682 // The value of the object is its address.
3683 int FI = MFI.CreateFixedObject(ObjSize, CurArgOffset, false, true);
3684 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
3685 InVals.push_back(FIN);
3686 if (ObjSize==1 || ObjSize==2) {
3687 if (GPR_idx != Num_GPR_Regs) {
3690 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
3692 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
3693 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
3694 EVT ObjType = ObjSize == 1 ? MVT::i8 : MVT::i16;
3696 DAG.getTruncStore(Val.getValue(1), dl, Val, FIN,
3697 MachinePointerInfo(&*FuncArg), ObjType);
3698 MemOps.push_back(Store);
3702 ArgOffset += PtrByteSize;
3706 for (unsigned j = 0; j < ArgSize; j += PtrByteSize) {
3707 // Store whatever pieces of the object are in registers
3708 // to memory. ArgOffset will be the address of the beginning
3710 if (GPR_idx != Num_GPR_Regs) {
3713 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
3715 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
3716 int FI = MFI.CreateFixedObject(PtrByteSize, ArgOffset, true);
3717 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
3718 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
3719 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
3720 MachinePointerInfo(&*FuncArg, j));
3721 MemOps.push_back(Store);
3723 ArgOffset += PtrByteSize;
3725 ArgOffset += ArgSize - (ArgOffset-CurArgOffset);
3732 switch (ObjectVT.getSimpleVT().SimpleTy) {
3733 default: llvm_unreachable("Unhandled argument type!");
3737 if (GPR_idx != Num_GPR_Regs) {
3738 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
3739 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
3741 if (ObjectVT == MVT::i1)
3742 ArgVal = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, ArgVal);
3747 ArgSize = PtrByteSize;
3749 // All int arguments reserve stack space in the Darwin ABI.
3750 ArgOffset += PtrByteSize;
3754 case MVT::i64: // PPC64
3755 if (GPR_idx != Num_GPR_Regs) {
3756 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
3757 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
3759 if (ObjectVT == MVT::i32 || ObjectVT == MVT::i1)
3760 // PPC64 passes i8, i16, and i32 values in i64 registers. Promote
3761 // value to MVT::i64 and then truncate to the correct register size.
3762 ArgVal = extendArgForPPC64(Flags, ObjectVT, DAG, ArgVal, dl);
3767 ArgSize = PtrByteSize;
3769 // All int arguments reserve stack space in the Darwin ABI.
3775 // Every 4 bytes of argument space consumes one of the GPRs available for
3776 // argument passing.
3777 if (GPR_idx != Num_GPR_Regs) {
3779 if (ObjSize == 8 && GPR_idx != Num_GPR_Regs && !isPPC64)
3782 if (FPR_idx != Num_FPR_Regs) {
3785 if (ObjectVT == MVT::f32)
3786 VReg = MF.addLiveIn(FPR[FPR_idx], &PPC::F4RCRegClass);
3788 VReg = MF.addLiveIn(FPR[FPR_idx], &PPC::F8RCRegClass);
3790 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT);
3796 // All FP arguments reserve stack space in the Darwin ABI.
3797 ArgOffset += isPPC64 ? 8 : ObjSize;
3803 // Note that vector arguments in registers don't reserve stack space,
3804 // except in varargs functions.
3805 if (VR_idx != Num_VR_Regs) {
3806 unsigned VReg = MF.addLiveIn(VR[VR_idx], &PPC::VRRCRegClass);
3807 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT);
3809 while ((ArgOffset % 16) != 0) {
3810 ArgOffset += PtrByteSize;
3811 if (GPR_idx != Num_GPR_Regs)
3815 GPR_idx = std::min(GPR_idx+4, Num_GPR_Regs); // FIXME correct for ppc64?
3819 if (!isVarArg && !isPPC64) {
3820 // Vectors go after all the nonvectors.
3821 CurArgOffset = VecArgOffset;
3824 // Vectors are aligned.
3825 ArgOffset = ((ArgOffset+15)/16)*16;
3826 CurArgOffset = ArgOffset;
3834 // We need to load the argument to a virtual register if we determined above
3835 // that we ran out of physical registers of the appropriate type.
3837 int FI = MFI.CreateFixedObject(ObjSize,
3838 CurArgOffset + (ArgSize - ObjSize),
3840 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
3841 ArgVal = DAG.getLoad(ObjectVT, dl, Chain, FIN, MachinePointerInfo());
3844 InVals.push_back(ArgVal);
3847 // Allow for Altivec parameters at the end, if needed.
3848 if (nAltivecParamsAtEnd) {
3849 MinReservedArea = ((MinReservedArea+15)/16)*16;
3850 MinReservedArea += 16*nAltivecParamsAtEnd;
3853 // Area that is at least reserved in the caller of this function.
3854 MinReservedArea = std::max(MinReservedArea, LinkageSize + 8 * PtrByteSize);
3856 // Set the size that is at least reserved in caller of this function. Tail
3857 // call optimized functions' reserved stack space needs to be aligned so that
3858 // taking the difference between two stack areas will result in an aligned
3861 EnsureStackAlignment(Subtarget.getFrameLowering(), MinReservedArea);
3862 FuncInfo->setMinReservedArea(MinReservedArea);
3864 // If the function takes variable number of arguments, make a frame index for
3865 // the start of the first vararg value... for expansion of llvm.va_start.
3867 int Depth = ArgOffset;
3869 FuncInfo->setVarArgsFrameIndex(
3870 MFI.CreateFixedObject(PtrVT.getSizeInBits()/8,
3872 SDValue FIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
3874 // If this function is vararg, store any remaining integer argument regs
3875 // to their spots on the stack so that they may be loaded by dereferencing
3876 // the result of va_next.
3877 for (; GPR_idx != Num_GPR_Regs; ++GPR_idx) {
3881 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
3883 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
3885 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
3887 DAG.getStore(Val.getValue(1), dl, Val, FIN, MachinePointerInfo());
3888 MemOps.push_back(Store);
3889 // Increment the address by four for the next argument to store
3890 SDValue PtrOff = DAG.getConstant(PtrVT.getSizeInBits()/8, dl, PtrVT);
3891 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
3895 if (!MemOps.empty())
3896 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOps);
3901 /// CalculateTailCallSPDiff - Get the amount the stack pointer has to be
3902 /// adjusted to accommodate the arguments for the tailcall.
3903 static int CalculateTailCallSPDiff(SelectionDAG& DAG, bool isTailCall,
3904 unsigned ParamSize) {
3906 if (!isTailCall) return 0;
3908 PPCFunctionInfo *FI = DAG.getMachineFunction().getInfo<PPCFunctionInfo>();
3909 unsigned CallerMinReservedArea = FI->getMinReservedArea();
3910 int SPDiff = (int)CallerMinReservedArea - (int)ParamSize;
3911 // Remember only if the new adjustement is bigger.
3912 if (SPDiff < FI->getTailCallSPDelta())
3913 FI->setTailCallSPDelta(SPDiff);
3918 static bool isFunctionGlobalAddress(SDValue Callee);
3921 resideInSameModule(SDValue Callee, Reloc::Model RelMod) {
3922 // If !G, Callee can be an external symbol.
3923 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
3924 if (!G) return false;
3926 const GlobalValue *GV = G->getGlobal();
3928 if (GV->isDeclaration()) return false;
3930 switch(GV->getLinkage()) {
3931 default: llvm_unreachable("unknow linkage type");
3932 case GlobalValue::AvailableExternallyLinkage:
3933 case GlobalValue::ExternalWeakLinkage:
3936 // Callee with weak linkage is allowed if it has hidden or protected
3938 case GlobalValue::LinkOnceAnyLinkage:
3939 case GlobalValue::LinkOnceODRLinkage: // e.g. c++ inline functions
3940 case GlobalValue::WeakAnyLinkage:
3941 case GlobalValue::WeakODRLinkage: // e.g. c++ template instantiation
3942 if (GV->hasDefaultVisibility())
3945 case GlobalValue::ExternalLinkage:
3946 case GlobalValue::InternalLinkage:
3947 case GlobalValue::PrivateLinkage:
3951 // With '-fPIC', calling default visiblity function need insert 'nop' after
3952 // function call, no matter that function resides in same module or not, so
3953 // we treat it as in different module.
3954 if (RelMod == Reloc::PIC_ && GV->hasDefaultVisibility())
3961 needStackSlotPassParameters(const PPCSubtarget &Subtarget,
3962 const SmallVectorImpl<ISD::OutputArg> &Outs) {
3963 assert(Subtarget.isSVR4ABI() && Subtarget.isPPC64());
3965 const unsigned PtrByteSize = 8;
3966 const unsigned LinkageSize = Subtarget.getFrameLowering()->getLinkageSize();
3968 static const MCPhysReg GPR[] = {
3969 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
3970 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
3972 static const MCPhysReg VR[] = {
3973 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
3974 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
3977 const unsigned NumGPRs = array_lengthof(GPR);
3978 const unsigned NumFPRs = 13;
3979 const unsigned NumVRs = array_lengthof(VR);
3980 const unsigned ParamAreaSize = NumGPRs * PtrByteSize;
3982 unsigned NumBytes = LinkageSize;
3983 unsigned AvailableFPRs = NumFPRs;
3984 unsigned AvailableVRs = NumVRs;
3986 for (const ISD::OutputArg& Param : Outs) {
3987 if (Param.Flags.isNest()) continue;
3989 if (CalculateStackSlotUsed(Param.VT, Param.ArgVT, Param.Flags,
3990 PtrByteSize, LinkageSize, ParamAreaSize,
3991 NumBytes, AvailableFPRs, AvailableVRs,
3992 Subtarget.hasQPX()))
3999 hasSameArgumentList(const Function *CallerFn, ImmutableCallSite *CS) {
4000 if (CS->arg_size() != CallerFn->getArgumentList().size())
4003 ImmutableCallSite::arg_iterator CalleeArgIter = CS->arg_begin();
4004 ImmutableCallSite::arg_iterator CalleeArgEnd = CS->arg_end();
4005 Function::const_arg_iterator CallerArgIter = CallerFn->arg_begin();
4007 for (; CalleeArgIter != CalleeArgEnd; ++CalleeArgIter, ++CallerArgIter) {
4008 const Value* CalleeArg = *CalleeArgIter;
4009 const Value* CallerArg = &(*CallerArgIter);
4010 if (CalleeArg == CallerArg)
4013 // e.g. @caller([4 x i64] %a, [4 x i64] %b) {
4014 // tail call @callee([4 x i64] undef, [4 x i64] %b)
4016 // 1st argument of callee is undef and has the same type as caller.
4017 if (CalleeArg->getType() == CallerArg->getType() &&
4018 isa<UndefValue>(CalleeArg))
4028 PPCTargetLowering::IsEligibleForTailCallOptimization_64SVR4(
4030 CallingConv::ID CalleeCC,
4031 ImmutableCallSite *CS,
4033 const SmallVectorImpl<ISD::OutputArg> &Outs,
4034 const SmallVectorImpl<ISD::InputArg> &Ins,
4035 SelectionDAG& DAG) const {
4036 bool TailCallOpt = getTargetMachine().Options.GuaranteedTailCallOpt;
4038 if (DisableSCO && !TailCallOpt) return false;
4040 // Variadic argument functions are not supported.
4041 if (isVarArg) return false;
4043 MachineFunction &MF = DAG.getMachineFunction();
4044 CallingConv::ID CallerCC = MF.getFunction()->getCallingConv();
4046 // Tail or Sibling call optimization (TCO/SCO) needs callee and caller has
4047 // the same calling convention
4048 if (CallerCC != CalleeCC) return false;
4050 // SCO support C calling convention
4051 if (CalleeCC != CallingConv::Fast && CalleeCC != CallingConv::C)
4054 // Caller contains any byval parameter is not supported.
4055 if (any_of(Ins, [](const ISD::InputArg &IA) { return IA.Flags.isByVal(); }))
4058 // Callee contains any byval parameter is not supported, too.
4059 // Note: This is a quick work around, because in some cases, e.g.
4060 // caller's stack size > callee's stack size, we are still able to apply
4061 // sibling call optimization. See: https://reviews.llvm.org/D23441#513574
4062 if (any_of(Outs, [](const ISD::OutputArg& OA) { return OA.Flags.isByVal(); }))
4065 // No TCO/SCO on indirect call because Caller have to restore its TOC
4066 if (!isFunctionGlobalAddress(Callee) &&
4067 !isa<ExternalSymbolSDNode>(Callee))
4070 // Check if Callee resides in the same module, because for now, PPC64 SVR4 ABI
4071 // (ELFv1/ELFv2) doesn't allow tail calls to a symbol resides in another
4073 // ref: https://bugzilla.mozilla.org/show_bug.cgi?id=973977
4074 if (!resideInSameModule(Callee, getTargetMachine().getRelocationModel()))
4077 // TCO allows altering callee ABI, so we don't have to check further.
4078 if (CalleeCC == CallingConv::Fast && TailCallOpt)
4081 if (DisableSCO) return false;
4083 // If callee use the same argument list that caller is using, then we can
4084 // apply SCO on this case. If it is not, then we need to check if callee needs
4085 // stack for passing arguments.
4086 if (!hasSameArgumentList(MF.getFunction(), CS) &&
4087 needStackSlotPassParameters(Subtarget, Outs)) {
4094 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
4095 /// for tail call optimization. Targets which want to do tail call
4096 /// optimization should implement this function.
4098 PPCTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
4099 CallingConv::ID CalleeCC,
4101 const SmallVectorImpl<ISD::InputArg> &Ins,
4102 SelectionDAG& DAG) const {
4103 if (!getTargetMachine().Options.GuaranteedTailCallOpt)
4106 // Variable argument functions are not supported.
4110 MachineFunction &MF = DAG.getMachineFunction();
4111 CallingConv::ID CallerCC = MF.getFunction()->getCallingConv();
4112 if (CalleeCC == CallingConv::Fast && CallerCC == CalleeCC) {
4113 // Functions containing by val parameters are not supported.
4114 for (unsigned i = 0; i != Ins.size(); i++) {
4115 ISD::ArgFlagsTy Flags = Ins[i].Flags;
4116 if (Flags.isByVal()) return false;
4119 // Non-PIC/GOT tail calls are supported.
4120 if (getTargetMachine().getRelocationModel() != Reloc::PIC_)
4123 // At the moment we can only do local tail calls (in same module, hidden
4124 // or protected) if we are generating PIC.
4125 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
4126 return G->getGlobal()->hasHiddenVisibility()
4127 || G->getGlobal()->hasProtectedVisibility();
4133 /// isCallCompatibleAddress - Return the immediate to use if the specified
4134 /// 32-bit value is representable in the immediate field of a BxA instruction.
4135 static SDNode *isBLACompatibleAddress(SDValue Op, SelectionDAG &DAG) {
4136 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
4137 if (!C) return nullptr;
4139 int Addr = C->getZExtValue();
4140 if ((Addr & 3) != 0 || // Low 2 bits are implicitly zero.
4141 SignExtend32<26>(Addr) != Addr)
4142 return nullptr; // Top 6 bits have to be sext of immediate.
4146 (int)C->getZExtValue() >> 2, SDLoc(Op),
4147 DAG.getTargetLoweringInfo().getPointerTy(DAG.getDataLayout()))
4153 struct TailCallArgumentInfo {
4158 TailCallArgumentInfo() : FrameIdx(0) {}
4162 /// StoreTailCallArgumentsToStackSlot - Stores arguments to their stack slot.
4163 static void StoreTailCallArgumentsToStackSlot(
4164 SelectionDAG &DAG, SDValue Chain,
4165 const SmallVectorImpl<TailCallArgumentInfo> &TailCallArgs,
4166 SmallVectorImpl<SDValue> &MemOpChains, const SDLoc &dl) {
4167 for (unsigned i = 0, e = TailCallArgs.size(); i != e; ++i) {
4168 SDValue Arg = TailCallArgs[i].Arg;
4169 SDValue FIN = TailCallArgs[i].FrameIdxOp;
4170 int FI = TailCallArgs[i].FrameIdx;
4171 // Store relative to framepointer.
4172 MemOpChains.push_back(DAG.getStore(
4173 Chain, dl, Arg, FIN,
4174 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI)));
4178 /// EmitTailCallStoreFPAndRetAddr - Move the frame pointer and return address to
4179 /// the appropriate stack slot for the tail call optimized function call.
4180 static SDValue EmitTailCallStoreFPAndRetAddr(SelectionDAG &DAG, SDValue Chain,
4181 SDValue OldRetAddr, SDValue OldFP,
4182 int SPDiff, const SDLoc &dl) {
4184 // Calculate the new stack slot for the return address.
4185 MachineFunction &MF = DAG.getMachineFunction();
4186 const PPCSubtarget &Subtarget = MF.getSubtarget<PPCSubtarget>();
4187 const PPCFrameLowering *FL = Subtarget.getFrameLowering();
4188 bool isPPC64 = Subtarget.isPPC64();
4189 int SlotSize = isPPC64 ? 8 : 4;
4190 int NewRetAddrLoc = SPDiff + FL->getReturnSaveOffset();
4191 int NewRetAddr = MF.getFrameInfo().CreateFixedObject(SlotSize,
4192 NewRetAddrLoc, true);
4193 EVT VT = isPPC64 ? MVT::i64 : MVT::i32;
4194 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewRetAddr, VT);
4195 Chain = DAG.getStore(Chain, dl, OldRetAddr, NewRetAddrFrIdx,
4196 MachinePointerInfo::getFixedStack(MF, NewRetAddr));
4198 // When using the 32/64-bit SVR4 ABI there is no need to move the FP stack
4199 // slot as the FP is never overwritten.
4200 if (Subtarget.isDarwinABI()) {
4201 int NewFPLoc = SPDiff + FL->getFramePointerSaveOffset();
4202 int NewFPIdx = MF.getFrameInfo().CreateFixedObject(SlotSize, NewFPLoc,
4204 SDValue NewFramePtrIdx = DAG.getFrameIndex(NewFPIdx, VT);
4205 Chain = DAG.getStore(Chain, dl, OldFP, NewFramePtrIdx,
4206 MachinePointerInfo::getFixedStack(
4207 DAG.getMachineFunction(), NewFPIdx));
4213 /// CalculateTailCallArgDest - Remember Argument for later processing. Calculate
4214 /// the position of the argument.
4216 CalculateTailCallArgDest(SelectionDAG &DAG, MachineFunction &MF, bool isPPC64,
4217 SDValue Arg, int SPDiff, unsigned ArgOffset,
4218 SmallVectorImpl<TailCallArgumentInfo>& TailCallArguments) {
4219 int Offset = ArgOffset + SPDiff;
4220 uint32_t OpSize = (Arg.getValueType().getSizeInBits()+7)/8;
4221 int FI = MF.getFrameInfo().CreateFixedObject(OpSize, Offset, true);
4222 EVT VT = isPPC64 ? MVT::i64 : MVT::i32;
4223 SDValue FIN = DAG.getFrameIndex(FI, VT);
4224 TailCallArgumentInfo Info;
4226 Info.FrameIdxOp = FIN;
4228 TailCallArguments.push_back(Info);
4231 /// EmitTCFPAndRetAddrLoad - Emit load from frame pointer and return address
4232 /// stack slot. Returns the chain as result and the loaded frame pointers in
4233 /// LROpOut/FPOpout. Used when tail calling.
4234 SDValue PPCTargetLowering::EmitTailCallLoadFPAndRetAddr(
4235 SelectionDAG &DAG, int SPDiff, SDValue Chain, SDValue &LROpOut,
4236 SDValue &FPOpOut, const SDLoc &dl) const {
4238 // Load the LR and FP stack slot for later adjusting.
4239 EVT VT = Subtarget.isPPC64() ? MVT::i64 : MVT::i32;
4240 LROpOut = getReturnAddrFrameIndex(DAG);
4241 LROpOut = DAG.getLoad(VT, dl, Chain, LROpOut, MachinePointerInfo());
4242 Chain = SDValue(LROpOut.getNode(), 1);
4244 // When using the 32/64-bit SVR4 ABI there is no need to load the FP stack
4245 // slot as the FP is never overwritten.
4246 if (Subtarget.isDarwinABI()) {
4247 FPOpOut = getFramePointerFrameIndex(DAG);
4248 FPOpOut = DAG.getLoad(VT, dl, Chain, FPOpOut, MachinePointerInfo());
4249 Chain = SDValue(FPOpOut.getNode(), 1);
4255 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
4256 /// by "Src" to address "Dst" of size "Size". Alignment information is
4257 /// specified by the specific parameter attribute. The copy will be passed as
4258 /// a byval function parameter.
4259 /// Sometimes what we are copying is the end of a larger object, the part that
4260 /// does not fit in registers.
4261 static SDValue CreateCopyOfByValArgument(SDValue Src, SDValue Dst,
4262 SDValue Chain, ISD::ArgFlagsTy Flags,
4263 SelectionDAG &DAG, const SDLoc &dl) {
4264 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), dl, MVT::i32);
4265 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
4266 false, false, false, MachinePointerInfo(),
4267 MachinePointerInfo());
4270 /// LowerMemOpCallTo - Store the argument to the stack or remember it in case of
4272 static void LowerMemOpCallTo(
4273 SelectionDAG &DAG, MachineFunction &MF, SDValue Chain, SDValue Arg,
4274 SDValue PtrOff, int SPDiff, unsigned ArgOffset, bool isPPC64,
4275 bool isTailCall, bool isVector, SmallVectorImpl<SDValue> &MemOpChains,
4276 SmallVectorImpl<TailCallArgumentInfo> &TailCallArguments, const SDLoc &dl) {
4277 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy(DAG.getDataLayout());
4282 StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
4284 StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
4285 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr,
4286 DAG.getConstant(ArgOffset, dl, PtrVT));
4288 MemOpChains.push_back(
4289 DAG.getStore(Chain, dl, Arg, PtrOff, MachinePointerInfo()));
4290 // Calculate and remember argument location.
4291 } else CalculateTailCallArgDest(DAG, MF, isPPC64, Arg, SPDiff, ArgOffset,
4296 PrepareTailCall(SelectionDAG &DAG, SDValue &InFlag, SDValue &Chain,
4297 const SDLoc &dl, int SPDiff, unsigned NumBytes, SDValue LROp,
4299 SmallVectorImpl<TailCallArgumentInfo> &TailCallArguments) {
4300 // Emit a sequence of copyto/copyfrom virtual registers for arguments that
4301 // might overwrite each other in case of tail call optimization.
4302 SmallVector<SDValue, 8> MemOpChains2;
4303 // Do not flag preceding copytoreg stuff together with the following stuff.
4305 StoreTailCallArgumentsToStackSlot(DAG, Chain, TailCallArguments,
4307 if (!MemOpChains2.empty())
4308 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains2);
4310 // Store the return address to the appropriate stack slot.
4311 Chain = EmitTailCallStoreFPAndRetAddr(DAG, Chain, LROp, FPOp, SPDiff, dl);
4313 // Emit callseq_end just before tailcall node.
4314 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, dl, true),
4315 DAG.getIntPtrConstant(0, dl, true), InFlag, dl);
4316 InFlag = Chain.getValue(1);
4319 // Is this global address that of a function that can be called by name? (as
4320 // opposed to something that must hold a descriptor for an indirect call).
4321 static bool isFunctionGlobalAddress(SDValue Callee) {
4322 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
4323 if (Callee.getOpcode() == ISD::GlobalTLSAddress ||
4324 Callee.getOpcode() == ISD::TargetGlobalTLSAddress)
4327 return G->getGlobal()->getValueType()->isFunctionTy();
4334 PrepareCall(SelectionDAG &DAG, SDValue &Callee, SDValue &InFlag, SDValue &Chain,
4335 SDValue CallSeqStart, const SDLoc &dl, int SPDiff, bool isTailCall,
4336 bool isPatchPoint, bool hasNest,
4337 SmallVectorImpl<std::pair<unsigned, SDValue>> &RegsToPass,
4338 SmallVectorImpl<SDValue> &Ops, std::vector<EVT> &NodeTys,
4339 ImmutableCallSite *CS, const PPCSubtarget &Subtarget) {
4341 bool isPPC64 = Subtarget.isPPC64();
4342 bool isSVR4ABI = Subtarget.isSVR4ABI();
4343 bool isELFv2ABI = Subtarget.isELFv2ABI();
4345 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy(DAG.getDataLayout());
4346 NodeTys.push_back(MVT::Other); // Returns a chain
4347 NodeTys.push_back(MVT::Glue); // Returns a flag for retval copy to use.
4349 unsigned CallOpc = PPCISD::CALL;
4351 bool needIndirectCall = true;
4352 if (!isSVR4ABI || !isPPC64)
4353 if (SDNode *Dest = isBLACompatibleAddress(Callee, DAG)) {
4354 // If this is an absolute destination address, use the munged value.
4355 Callee = SDValue(Dest, 0);
4356 needIndirectCall = false;
4359 // PC-relative references to external symbols should go through $stub, unless
4360 // we're building with the leopard linker or later, which automatically
4361 // synthesizes these stubs.
4362 const TargetMachine &TM = DAG.getTarget();
4363 const Module *Mod = DAG.getMachineFunction().getFunction()->getParent();
4364 const GlobalValue *GV = nullptr;
4365 if (auto *G = dyn_cast<GlobalAddressSDNode>(Callee))
4366 GV = G->getGlobal();
4367 bool Local = TM.shouldAssumeDSOLocal(*Mod, GV);
4368 bool UsePlt = !Local && Subtarget.isTargetELF() && !isPPC64;
4370 if (isFunctionGlobalAddress(Callee)) {
4371 GlobalAddressSDNode *G = cast<GlobalAddressSDNode>(Callee);
4372 // A call to a TLS address is actually an indirect call to a
4373 // thread-specific pointer.
4374 unsigned OpFlags = 0;
4376 OpFlags = PPCII::MO_PLT;
4378 // If the callee is a GlobalAddress/ExternalSymbol node (quite common,
4379 // every direct call is) turn it into a TargetGlobalAddress /
4380 // TargetExternalSymbol node so that legalize doesn't hack it.
4381 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), dl,
4382 Callee.getValueType(), 0, OpFlags);
4383 needIndirectCall = false;
4386 if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
4387 unsigned char OpFlags = 0;
4390 OpFlags = PPCII::MO_PLT;
4392 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), Callee.getValueType(),
4394 needIndirectCall = false;
4398 // We'll form an invalid direct call when lowering a patchpoint; the full
4399 // sequence for an indirect call is complicated, and many of the
4400 // instructions introduced might have side effects (and, thus, can't be
4401 // removed later). The call itself will be removed as soon as the
4402 // argument/return lowering is complete, so the fact that it has the wrong
4403 // kind of operands should not really matter.
4404 needIndirectCall = false;
4407 if (needIndirectCall) {
4408 // Otherwise, this is an indirect call. We have to use a MTCTR/BCTRL pair
4409 // to do the call, we can't use PPCISD::CALL.
4410 SDValue MTCTROps[] = {Chain, Callee, InFlag};
4412 if (isSVR4ABI && isPPC64 && !isELFv2ABI) {
4413 // Function pointers in the 64-bit SVR4 ABI do not point to the function
4414 // entry point, but to the function descriptor (the function entry point
4415 // address is part of the function descriptor though).
4416 // The function descriptor is a three doubleword structure with the
4417 // following fields: function entry point, TOC base address and
4418 // environment pointer.
4419 // Thus for a call through a function pointer, the following actions need
4421 // 1. Save the TOC of the caller in the TOC save area of its stack
4422 // frame (this is done in LowerCall_Darwin() or LowerCall_64SVR4()).
4423 // 2. Load the address of the function entry point from the function
4425 // 3. Load the TOC of the callee from the function descriptor into r2.
4426 // 4. Load the environment pointer from the function descriptor into
4428 // 5. Branch to the function entry point address.
4429 // 6. On return of the callee, the TOC of the caller needs to be
4430 // restored (this is done in FinishCall()).
4432 // The loads are scheduled at the beginning of the call sequence, and the
4433 // register copies are flagged together to ensure that no other
4434 // operations can be scheduled in between. E.g. without flagging the
4435 // copies together, a TOC access in the caller could be scheduled between
4436 // the assignment of the callee TOC and the branch to the callee, which
4437 // results in the TOC access going through the TOC of the callee instead
4438 // of going through the TOC of the caller, which leads to incorrect code.
4440 // Load the address of the function entry point from the function
4442 SDValue LDChain = CallSeqStart.getValue(CallSeqStart->getNumValues()-1);
4443 if (LDChain.getValueType() == MVT::Glue)
4444 LDChain = CallSeqStart.getValue(CallSeqStart->getNumValues()-2);
4446 auto MMOFlags = Subtarget.hasInvariantFunctionDescriptors()
4447 ? (MachineMemOperand::MODereferenceable |
4448 MachineMemOperand::MOInvariant)
4449 : MachineMemOperand::MONone;
4451 MachinePointerInfo MPI(CS ? CS->getCalledValue() : nullptr);
4452 SDValue LoadFuncPtr = DAG.getLoad(MVT::i64, dl, LDChain, Callee, MPI,
4453 /* Alignment = */ 8, MMOFlags);
4455 // Load environment pointer into r11.
4456 SDValue PtrOff = DAG.getIntPtrConstant(16, dl);
4457 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, MVT::i64, Callee, PtrOff);
4458 SDValue LoadEnvPtr =
4459 DAG.getLoad(MVT::i64, dl, LDChain, AddPtr, MPI.getWithOffset(16),
4460 /* Alignment = */ 8, MMOFlags);
4462 SDValue TOCOff = DAG.getIntPtrConstant(8, dl);
4463 SDValue AddTOC = DAG.getNode(ISD::ADD, dl, MVT::i64, Callee, TOCOff);
4465 DAG.getLoad(MVT::i64, dl, LDChain, AddTOC, MPI.getWithOffset(8),
4466 /* Alignment = */ 8, MMOFlags);
4468 setUsesTOCBasePtr(DAG);
4469 SDValue TOCVal = DAG.getCopyToReg(Chain, dl, PPC::X2, TOCPtr,
4471 Chain = TOCVal.getValue(0);
4472 InFlag = TOCVal.getValue(1);
4474 // If the function call has an explicit 'nest' parameter, it takes the
4475 // place of the environment pointer.
4477 SDValue EnvVal = DAG.getCopyToReg(Chain, dl, PPC::X11, LoadEnvPtr,
4480 Chain = EnvVal.getValue(0);
4481 InFlag = EnvVal.getValue(1);
4484 MTCTROps[0] = Chain;
4485 MTCTROps[1] = LoadFuncPtr;
4486 MTCTROps[2] = InFlag;
4489 Chain = DAG.getNode(PPCISD::MTCTR, dl, NodeTys,
4490 makeArrayRef(MTCTROps, InFlag.getNode() ? 3 : 2));
4491 InFlag = Chain.getValue(1);
4494 NodeTys.push_back(MVT::Other);
4495 NodeTys.push_back(MVT::Glue);
4496 Ops.push_back(Chain);
4497 CallOpc = PPCISD::BCTRL;
4498 Callee.setNode(nullptr);
4499 // Add use of X11 (holding environment pointer)
4500 if (isSVR4ABI && isPPC64 && !isELFv2ABI && !hasNest)
4501 Ops.push_back(DAG.getRegister(PPC::X11, PtrVT));
4502 // Add CTR register as callee so a bctr can be emitted later.
4504 Ops.push_back(DAG.getRegister(isPPC64 ? PPC::CTR8 : PPC::CTR, PtrVT));
4507 // If this is a direct call, pass the chain and the callee.
4508 if (Callee.getNode()) {
4509 Ops.push_back(Chain);
4510 Ops.push_back(Callee);
4512 // If this is a tail call add stack pointer delta.
4514 Ops.push_back(DAG.getConstant(SPDiff, dl, MVT::i32));
4516 // Add argument registers to the end of the list so that they are known live
4518 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
4519 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
4520 RegsToPass[i].second.getValueType()));
4522 // All calls, in both the ELF V1 and V2 ABIs, need the TOC register live
4524 if (isSVR4ABI && isPPC64 && !isPatchPoint) {
4525 setUsesTOCBasePtr(DAG);
4526 Ops.push_back(DAG.getRegister(PPC::X2, PtrVT));
4533 bool isLocalCall(const SDValue &Callee)
4535 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
4536 return G->getGlobal()->isStrongDefinitionForLinker();
4540 SDValue PPCTargetLowering::LowerCallResult(
4541 SDValue Chain, SDValue InFlag, CallingConv::ID CallConv, bool isVarArg,
4542 const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl,
4543 SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const {
4545 SmallVector<CCValAssign, 16> RVLocs;
4546 CCState CCRetInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
4548 CCRetInfo.AnalyzeCallResult(Ins, RetCC_PPC);
4550 // Copy all of the result registers out of their specified physreg.
4551 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
4552 CCValAssign &VA = RVLocs[i];
4553 assert(VA.isRegLoc() && "Can only return in registers!");
4555 SDValue Val = DAG.getCopyFromReg(Chain, dl,
4556 VA.getLocReg(), VA.getLocVT(), InFlag);
4557 Chain = Val.getValue(1);
4558 InFlag = Val.getValue(2);
4560 switch (VA.getLocInfo()) {
4561 default: llvm_unreachable("Unknown loc info!");
4562 case CCValAssign::Full: break;
4563 case CCValAssign::AExt:
4564 Val = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Val);
4566 case CCValAssign::ZExt:
4567 Val = DAG.getNode(ISD::AssertZext, dl, VA.getLocVT(), Val,
4568 DAG.getValueType(VA.getValVT()));
4569 Val = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Val);
4571 case CCValAssign::SExt:
4572 Val = DAG.getNode(ISD::AssertSext, dl, VA.getLocVT(), Val,
4573 DAG.getValueType(VA.getValVT()));
4574 Val = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Val);
4578 InVals.push_back(Val);
4584 SDValue PPCTargetLowering::FinishCall(
4585 CallingConv::ID CallConv, const SDLoc &dl, bool isTailCall, bool isVarArg,
4586 bool isPatchPoint, bool hasNest, SelectionDAG &DAG,
4587 SmallVector<std::pair<unsigned, SDValue>, 8> &RegsToPass, SDValue InFlag,
4588 SDValue Chain, SDValue CallSeqStart, SDValue &Callee, int SPDiff,
4589 unsigned NumBytes, const SmallVectorImpl<ISD::InputArg> &Ins,
4590 SmallVectorImpl<SDValue> &InVals, ImmutableCallSite *CS) const {
4592 std::vector<EVT> NodeTys;
4593 SmallVector<SDValue, 8> Ops;
4594 unsigned CallOpc = PrepareCall(DAG, Callee, InFlag, Chain, CallSeqStart, dl,
4595 SPDiff, isTailCall, isPatchPoint, hasNest,
4596 RegsToPass, Ops, NodeTys, CS, Subtarget);
4598 // Add implicit use of CR bit 6 for 32-bit SVR4 vararg calls
4599 if (isVarArg && Subtarget.isSVR4ABI() && !Subtarget.isPPC64())
4600 Ops.push_back(DAG.getRegister(PPC::CR1EQ, MVT::i32));
4602 // When performing tail call optimization the callee pops its arguments off
4603 // the stack. Account for this here so these bytes can be pushed back on in
4604 // PPCFrameLowering::eliminateCallFramePseudoInstr.
4605 int BytesCalleePops =
4606 (CallConv == CallingConv::Fast &&
4607 getTargetMachine().Options.GuaranteedTailCallOpt) ? NumBytes : 0;
4609 // Add a register mask operand representing the call-preserved registers.
4610 const TargetRegisterInfo *TRI = Subtarget.getRegisterInfo();
4611 const uint32_t *Mask =
4612 TRI->getCallPreservedMask(DAG.getMachineFunction(), CallConv);
4613 assert(Mask && "Missing call preserved mask for calling convention");
4614 Ops.push_back(DAG.getRegisterMask(Mask));
4616 if (InFlag.getNode())
4617 Ops.push_back(InFlag);
4621 assert(((Callee.getOpcode() == ISD::Register &&
4622 cast<RegisterSDNode>(Callee)->getReg() == PPC::CTR) ||
4623 Callee.getOpcode() == ISD::TargetExternalSymbol ||
4624 Callee.getOpcode() == ISD::TargetGlobalAddress ||
4625 isa<ConstantSDNode>(Callee)) &&
4626 "Expecting an global address, external symbol, absolute value or register");
4628 DAG.getMachineFunction().getFrameInfo().setHasTailCall();
4629 return DAG.getNode(PPCISD::TC_RETURN, dl, MVT::Other, Ops);
4632 // Add a NOP immediately after the branch instruction when using the 64-bit
4633 // SVR4 ABI. At link time, if caller and callee are in a different module and
4634 // thus have a different TOC, the call will be replaced with a call to a stub
4635 // function which saves the current TOC, loads the TOC of the callee and
4636 // branches to the callee. The NOP will be replaced with a load instruction
4637 // which restores the TOC of the caller from the TOC save slot of the current
4638 // stack frame. If caller and callee belong to the same module (and have the
4639 // same TOC), the NOP will remain unchanged.
4641 if (!isTailCall && Subtarget.isSVR4ABI()&& Subtarget.isPPC64() &&
4643 if (CallOpc == PPCISD::BCTRL) {
4644 // This is a call through a function pointer.
4645 // Restore the caller TOC from the save area into R2.
4646 // See PrepareCall() for more information about calls through function
4647 // pointers in the 64-bit SVR4 ABI.
4648 // We are using a target-specific load with r2 hard coded, because the
4649 // result of a target-independent load would never go directly into r2,
4650 // since r2 is a reserved register (which prevents the register allocator
4651 // from allocating it), resulting in an additional register being
4652 // allocated and an unnecessary move instruction being generated.
4653 CallOpc = PPCISD::BCTRL_LOAD_TOC;
4655 EVT PtrVT = getPointerTy(DAG.getDataLayout());
4656 SDValue StackPtr = DAG.getRegister(PPC::X1, PtrVT);
4657 unsigned TOCSaveOffset = Subtarget.getFrameLowering()->getTOCSaveOffset();
4658 SDValue TOCOff = DAG.getIntPtrConstant(TOCSaveOffset, dl);
4659 SDValue AddTOC = DAG.getNode(ISD::ADD, dl, MVT::i64, StackPtr, TOCOff);
4661 // The address needs to go after the chain input but before the flag (or
4662 // any other variadic arguments).
4663 Ops.insert(std::next(Ops.begin()), AddTOC);
4664 } else if ((CallOpc == PPCISD::CALL) &&
4665 (!isLocalCall(Callee) ||
4666 DAG.getTarget().getRelocationModel() == Reloc::PIC_))
4667 // Otherwise insert NOP for non-local calls.
4668 CallOpc = PPCISD::CALL_NOP;
4671 Chain = DAG.getNode(CallOpc, dl, NodeTys, Ops);
4672 InFlag = Chain.getValue(1);
4674 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, dl, true),
4675 DAG.getIntPtrConstant(BytesCalleePops, dl, true),
4678 InFlag = Chain.getValue(1);
4680 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
4681 Ins, dl, DAG, InVals);
4685 PPCTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
4686 SmallVectorImpl<SDValue> &InVals) const {
4687 SelectionDAG &DAG = CLI.DAG;
4689 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
4690 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
4691 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
4692 SDValue Chain = CLI.Chain;
4693 SDValue Callee = CLI.Callee;
4694 bool &isTailCall = CLI.IsTailCall;
4695 CallingConv::ID CallConv = CLI.CallConv;
4696 bool isVarArg = CLI.IsVarArg;
4697 bool isPatchPoint = CLI.IsPatchPoint;
4698 ImmutableCallSite *CS = CLI.CS;
4701 if (Subtarget.useLongCalls() && !(CS && CS->isMustTailCall()))
4703 else if (Subtarget.isSVR4ABI() && Subtarget.isPPC64())
4705 IsEligibleForTailCallOptimization_64SVR4(Callee, CallConv, CS,
4706 isVarArg, Outs, Ins, DAG);
4708 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv, isVarArg,
4712 if (!getTargetMachine().Options.GuaranteedTailCallOpt)
4715 assert(isa<GlobalAddressSDNode>(Callee) &&
4716 "Callee should be an llvm::Function object.");
4718 const GlobalValue *GV = cast<GlobalAddressSDNode>(Callee)->getGlobal();
4719 const unsigned Width = 80 - strlen("TCO caller: ")
4720 - strlen(", callee linkage: 0, 0");
4721 dbgs() << "TCO caller: "
4722 << left_justify(DAG.getMachineFunction().getName(), Width)
4723 << ", callee linkage: "
4724 << GV->getVisibility() << ", " << GV->getLinkage() << "\n"
4729 if (!isTailCall && CS && CS->isMustTailCall())
4730 report_fatal_error("failed to perform tail call elimination on a call "
4731 "site marked musttail");
4733 // When long calls (i.e. indirect calls) are always used, calls are always
4734 // made via function pointer. If we have a function name, first translate it
4736 if (Subtarget.useLongCalls() && isa<GlobalAddressSDNode>(Callee) &&
4738 Callee = LowerGlobalAddress(Callee, DAG);
4740 if (Subtarget.isSVR4ABI()) {
4741 if (Subtarget.isPPC64())
4742 return LowerCall_64SVR4(Chain, Callee, CallConv, isVarArg,
4743 isTailCall, isPatchPoint, Outs, OutVals, Ins,
4744 dl, DAG, InVals, CS);
4746 return LowerCall_32SVR4(Chain, Callee, CallConv, isVarArg,
4747 isTailCall, isPatchPoint, Outs, OutVals, Ins,
4748 dl, DAG, InVals, CS);
4751 return LowerCall_Darwin(Chain, Callee, CallConv, isVarArg,
4752 isTailCall, isPatchPoint, Outs, OutVals, Ins,
4753 dl, DAG, InVals, CS);
4756 SDValue PPCTargetLowering::LowerCall_32SVR4(
4757 SDValue Chain, SDValue Callee, CallingConv::ID CallConv, bool isVarArg,
4758 bool isTailCall, bool isPatchPoint,
4759 const SmallVectorImpl<ISD::OutputArg> &Outs,
4760 const SmallVectorImpl<SDValue> &OutVals,
4761 const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl,
4762 SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals,
4763 ImmutableCallSite *CS) const {
4764 // See PPCTargetLowering::LowerFormalArguments_32SVR4() for a description
4765 // of the 32-bit SVR4 ABI stack frame layout.
4767 assert((CallConv == CallingConv::C ||
4768 CallConv == CallingConv::Fast) && "Unknown calling convention!");
4770 unsigned PtrByteSize = 4;
4772 MachineFunction &MF = DAG.getMachineFunction();
4774 // Mark this function as potentially containing a function that contains a
4775 // tail call. As a consequence the frame pointer will be used for dynamicalloc
4776 // and restoring the callers stack pointer in this functions epilog. This is
4777 // done because by tail calling the called function might overwrite the value
4778 // in this function's (MF) stack pointer stack slot 0(SP).
4779 if (getTargetMachine().Options.GuaranteedTailCallOpt &&
4780 CallConv == CallingConv::Fast)
4781 MF.getInfo<PPCFunctionInfo>()->setHasFastCall();
4783 // Count how many bytes are to be pushed on the stack, including the linkage
4784 // area, parameter list area and the part of the local variable space which
4785 // contains copies of aggregates which are passed by value.
4787 // Assign locations to all of the outgoing arguments.
4788 SmallVector<CCValAssign, 16> ArgLocs;
4789 PPCCCState CCInfo(CallConv, isVarArg, MF, ArgLocs, *DAG.getContext());
4791 // Reserve space for the linkage area on the stack.
4792 CCInfo.AllocateStack(Subtarget.getFrameLowering()->getLinkageSize(),
4795 CCInfo.PreAnalyzeCallOperands(Outs);
4798 // Handle fixed and variable vector arguments differently.
4799 // Fixed vector arguments go into registers as long as registers are
4800 // available. Variable vector arguments always go into memory.
4801 unsigned NumArgs = Outs.size();
4803 for (unsigned i = 0; i != NumArgs; ++i) {
4804 MVT ArgVT = Outs[i].VT;
4805 ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;
4808 if (Outs[i].IsFixed) {
4809 Result = CC_PPC32_SVR4(i, ArgVT, ArgVT, CCValAssign::Full, ArgFlags,
4812 Result = CC_PPC32_SVR4_VarArg(i, ArgVT, ArgVT, CCValAssign::Full,
4818 errs() << "Call operand #" << i << " has unhandled type "
4819 << EVT(ArgVT).getEVTString() << "\n";
4821 llvm_unreachable(nullptr);
4825 // All arguments are treated the same.
4826 CCInfo.AnalyzeCallOperands(Outs, CC_PPC32_SVR4);
4828 CCInfo.clearWasPPCF128();
4830 // Assign locations to all of the outgoing aggregate by value arguments.
4831 SmallVector<CCValAssign, 16> ByValArgLocs;
4832 CCState CCByValInfo(CallConv, isVarArg, MF, ByValArgLocs, *DAG.getContext());
4834 // Reserve stack space for the allocations in CCInfo.
4835 CCByValInfo.AllocateStack(CCInfo.getNextStackOffset(), PtrByteSize);
4837 CCByValInfo.AnalyzeCallOperands(Outs, CC_PPC32_SVR4_ByVal);
4839 // Size of the linkage area, parameter list area and the part of the local
4840 // space variable where copies of aggregates which are passed by value are
4842 unsigned NumBytes = CCByValInfo.getNextStackOffset();
4844 // Calculate by how many bytes the stack has to be adjusted in case of tail
4845 // call optimization.
4846 int SPDiff = CalculateTailCallSPDiff(DAG, isTailCall, NumBytes);
4848 // Adjust the stack pointer for the new arguments...
4849 // These operations are automatically eliminated by the prolog/epilog pass
4850 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, dl, true),
4852 SDValue CallSeqStart = Chain;
4854 // Load the return address and frame pointer so it can be moved somewhere else
4857 Chain = EmitTailCallLoadFPAndRetAddr(DAG, SPDiff, Chain, LROp, FPOp, dl);
4859 // Set up a copy of the stack pointer for use loading and storing any
4860 // arguments that may not fit in the registers available for argument
4862 SDValue StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
4864 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
4865 SmallVector<TailCallArgumentInfo, 8> TailCallArguments;
4866 SmallVector<SDValue, 8> MemOpChains;
4868 bool seenFloatArg = false;
4869 // Walk the register/memloc assignments, inserting copies/loads.
4870 for (unsigned i = 0, j = 0, e = ArgLocs.size();
4873 CCValAssign &VA = ArgLocs[i];
4874 SDValue Arg = OutVals[i];
4875 ISD::ArgFlagsTy Flags = Outs[i].Flags;
4877 if (Flags.isByVal()) {
4878 // Argument is an aggregate which is passed by value, thus we need to
4879 // create a copy of it in the local variable space of the current stack
4880 // frame (which is the stack frame of the caller) and pass the address of
4881 // this copy to the callee.
4882 assert((j < ByValArgLocs.size()) && "Index out of bounds!");
4883 CCValAssign &ByValVA = ByValArgLocs[j++];
4884 assert((VA.getValNo() == ByValVA.getValNo()) && "ValNo mismatch!");
4886 // Memory reserved in the local variable space of the callers stack frame.
4887 unsigned LocMemOffset = ByValVA.getLocMemOffset();
4889 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset, dl);
4890 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(MF.getDataLayout()),
4893 // Create a copy of the argument in the local area of the current
4895 SDValue MemcpyCall =
4896 CreateCopyOfByValArgument(Arg, PtrOff,
4897 CallSeqStart.getNode()->getOperand(0),
4900 // This must go outside the CALLSEQ_START..END.
4901 SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
4902 CallSeqStart.getNode()->getOperand(1),
4904 DAG.ReplaceAllUsesWith(CallSeqStart.getNode(),
4905 NewCallSeqStart.getNode());
4906 Chain = CallSeqStart = NewCallSeqStart;
4908 // Pass the address of the aggregate copy on the stack either in a
4909 // physical register or in the parameter list area of the current stack
4910 // frame to the callee.
4914 if (VA.isRegLoc()) {
4915 if (Arg.getValueType() == MVT::i1)
4916 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Arg);
4918 seenFloatArg |= VA.getLocVT().isFloatingPoint();
4919 // Put argument in a physical register.
4920 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
4922 // Put argument in the parameter list area of the current stack frame.
4923 assert(VA.isMemLoc());
4924 unsigned LocMemOffset = VA.getLocMemOffset();
4927 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset, dl);
4928 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(MF.getDataLayout()),
4931 MemOpChains.push_back(
4932 DAG.getStore(Chain, dl, Arg, PtrOff, MachinePointerInfo()));
4934 // Calculate and remember argument location.
4935 CalculateTailCallArgDest(DAG, MF, false, Arg, SPDiff, LocMemOffset,
4941 if (!MemOpChains.empty())
4942 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains);
4944 // Build a sequence of copy-to-reg nodes chained together with token chain
4945 // and flag operands which copy the outgoing args into the appropriate regs.
4947 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
4948 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
4949 RegsToPass[i].second, InFlag);
4950 InFlag = Chain.getValue(1);
4953 // Set CR bit 6 to true if this is a vararg call with floating args passed in
4956 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
4957 SDValue Ops[] = { Chain, InFlag };
4959 Chain = DAG.getNode(seenFloatArg ? PPCISD::CR6SET : PPCISD::CR6UNSET,
4960 dl, VTs, makeArrayRef(Ops, InFlag.getNode() ? 2 : 1));
4962 InFlag = Chain.getValue(1);
4966 PrepareTailCall(DAG, InFlag, Chain, dl, SPDiff, NumBytes, LROp, FPOp,
4969 return FinishCall(CallConv, dl, isTailCall, isVarArg, isPatchPoint,
4970 /* unused except on PPC64 ELFv1 */ false, DAG,
4971 RegsToPass, InFlag, Chain, CallSeqStart, Callee, SPDiff,
4972 NumBytes, Ins, InVals, CS);
4975 // Copy an argument into memory, being careful to do this outside the
4976 // call sequence for the call to which the argument belongs.
4977 SDValue PPCTargetLowering::createMemcpyOutsideCallSeq(
4978 SDValue Arg, SDValue PtrOff, SDValue CallSeqStart, ISD::ArgFlagsTy Flags,
4979 SelectionDAG &DAG, const SDLoc &dl) const {
4980 SDValue MemcpyCall = CreateCopyOfByValArgument(Arg, PtrOff,
4981 CallSeqStart.getNode()->getOperand(0),
4983 // The MEMCPY must go outside the CALLSEQ_START..END.
4984 SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
4985 CallSeqStart.getNode()->getOperand(1),
4987 DAG.ReplaceAllUsesWith(CallSeqStart.getNode(),
4988 NewCallSeqStart.getNode());
4989 return NewCallSeqStart;
4992 SDValue PPCTargetLowering::LowerCall_64SVR4(
4993 SDValue Chain, SDValue Callee, CallingConv::ID CallConv, bool isVarArg,
4994 bool isTailCall, bool isPatchPoint,
4995 const SmallVectorImpl<ISD::OutputArg> &Outs,
4996 const SmallVectorImpl<SDValue> &OutVals,
4997 const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl,
4998 SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals,
4999 ImmutableCallSite *CS) const {
5001 bool isELFv2ABI = Subtarget.isELFv2ABI();
5002 bool isLittleEndian = Subtarget.isLittleEndian();
5003 unsigned NumOps = Outs.size();
5004 bool hasNest = false;
5005 bool IsSibCall = false;
5007 EVT PtrVT = getPointerTy(DAG.getDataLayout());
5008 unsigned PtrByteSize = 8;
5010 MachineFunction &MF = DAG.getMachineFunction();
5012 if (isTailCall && !getTargetMachine().Options.GuaranteedTailCallOpt)
5015 // Mark this function as potentially containing a function that contains a
5016 // tail call. As a consequence the frame pointer will be used for dynamicalloc
5017 // and restoring the callers stack pointer in this functions epilog. This is
5018 // done because by tail calling the called function might overwrite the value
5019 // in this function's (MF) stack pointer stack slot 0(SP).
5020 if (getTargetMachine().Options.GuaranteedTailCallOpt &&
5021 CallConv == CallingConv::Fast)
5022 MF.getInfo<PPCFunctionInfo>()->setHasFastCall();
5024 assert(!(CallConv == CallingConv::Fast && isVarArg) &&
5025 "fastcc not supported on varargs functions");
5027 // Count how many bytes are to be pushed on the stack, including the linkage
5028 // area, and parameter passing area. On ELFv1, the linkage area is 48 bytes
5029 // reserved space for [SP][CR][LR][2 x unused][TOC]; on ELFv2, the linkage
5030 // area is 32 bytes reserved space for [SP][CR][LR][TOC].
5031 unsigned LinkageSize = Subtarget.getFrameLowering()->getLinkageSize();
5032 unsigned NumBytes = LinkageSize;
5033 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
5034 unsigned &QFPR_idx = FPR_idx;
5036 static const MCPhysReg GPR[] = {
5037 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
5038 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
5040 static const MCPhysReg VR[] = {
5041 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
5042 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
5044 static const MCPhysReg VSRH[] = {
5045 PPC::VSH2, PPC::VSH3, PPC::VSH4, PPC::VSH5, PPC::VSH6, PPC::VSH7, PPC::VSH8,
5046 PPC::VSH9, PPC::VSH10, PPC::VSH11, PPC::VSH12, PPC::VSH13
5049 const unsigned NumGPRs = array_lengthof(GPR);
5050 const unsigned NumFPRs = 13;
5051 const unsigned NumVRs = array_lengthof(VR);
5052 const unsigned NumQFPRs = NumFPRs;
5054 // When using the fast calling convention, we don't provide backing for
5055 // arguments that will be in registers.
5056 unsigned NumGPRsUsed = 0, NumFPRsUsed = 0, NumVRsUsed = 0;
5058 // Add up all the space actually used.
5059 for (unsigned i = 0; i != NumOps; ++i) {
5060 ISD::ArgFlagsTy Flags = Outs[i].Flags;
5061 EVT ArgVT = Outs[i].VT;
5062 EVT OrigVT = Outs[i].ArgVT;
5067 if (CallConv == CallingConv::Fast) {
5068 if (Flags.isByVal())
5069 NumGPRsUsed += (Flags.getByValSize()+7)/8;
5071 switch (ArgVT.getSimpleVT().SimpleTy) {
5072 default: llvm_unreachable("Unexpected ValueType for argument!");
5076 if (++NumGPRsUsed <= NumGPRs)
5085 if (++NumVRsUsed <= NumVRs)
5089 // When using QPX, this is handled like a FP register, otherwise, it
5090 // is an Altivec register.
5091 if (Subtarget.hasQPX()) {
5092 if (++NumFPRsUsed <= NumFPRs)
5095 if (++NumVRsUsed <= NumVRs)
5101 case MVT::v4f64: // QPX
5102 case MVT::v4i1: // QPX
5103 if (++NumFPRsUsed <= NumFPRs)
5109 /* Respect alignment of argument on the stack. */
5111 CalculateStackSlotAlignment(ArgVT, OrigVT, Flags, PtrByteSize);
5112 NumBytes = ((NumBytes + Align - 1) / Align) * Align;
5114 NumBytes += CalculateStackSlotSize(ArgVT, Flags, PtrByteSize);
5115 if (Flags.isInConsecutiveRegsLast())
5116 NumBytes = ((NumBytes + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
5119 unsigned NumBytesActuallyUsed = NumBytes;
5121 // The prolog code of the callee may store up to 8 GPR argument registers to
5122 // the stack, allowing va_start to index over them in memory if its varargs.
5123 // Because we cannot tell if this is needed on the caller side, we have to
5124 // conservatively assume that it is needed. As such, make sure we have at
5125 // least enough stack space for the caller to store the 8 GPRs.
5126 // FIXME: On ELFv2, it may be unnecessary to allocate the parameter area.
5127 NumBytes = std::max(NumBytes, LinkageSize + 8 * PtrByteSize);
5129 // Tail call needs the stack to be aligned.
5130 if (getTargetMachine().Options.GuaranteedTailCallOpt &&
5131 CallConv == CallingConv::Fast)
5132 NumBytes = EnsureStackAlignment(Subtarget.getFrameLowering(), NumBytes);
5136 // Calculate by how many bytes the stack has to be adjusted in case of tail
5137 // call optimization.
5139 SPDiff = CalculateTailCallSPDiff(DAG, isTailCall, NumBytes);
5141 // To protect arguments on the stack from being clobbered in a tail call,
5142 // force all the loads to happen before doing any other lowering.
5144 Chain = DAG.getStackArgumentTokenFactor(Chain);
5146 // Adjust the stack pointer for the new arguments...
5147 // These operations are automatically eliminated by the prolog/epilog pass
5149 Chain = DAG.getCALLSEQ_START(Chain,
5150 DAG.getIntPtrConstant(NumBytes, dl, true), dl);
5151 SDValue CallSeqStart = Chain;
5153 // Load the return address and frame pointer so it can be move somewhere else
5156 Chain = EmitTailCallLoadFPAndRetAddr(DAG, SPDiff, Chain, LROp, FPOp, dl);
5158 // Set up a copy of the stack pointer for use loading and storing any
5159 // arguments that may not fit in the registers available for argument
5161 SDValue StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
5163 // Figure out which arguments are going to go in registers, and which in
5164 // memory. Also, if this is a vararg function, floating point operations
5165 // must be stored to our stack, and loaded into integer regs as well, if
5166 // any integer regs are available for argument passing.
5167 unsigned ArgOffset = LinkageSize;
5169 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
5170 SmallVector<TailCallArgumentInfo, 8> TailCallArguments;
5172 SmallVector<SDValue, 8> MemOpChains;
5173 for (unsigned i = 0; i != NumOps; ++i) {
5174 SDValue Arg = OutVals[i];
5175 ISD::ArgFlagsTy Flags = Outs[i].Flags;
5176 EVT ArgVT = Outs[i].VT;
5177 EVT OrigVT = Outs[i].ArgVT;
5179 // PtrOff will be used to store the current argument to the stack if a
5180 // register cannot be found for it.
5183 // We re-align the argument offset for each argument, except when using the
5184 // fast calling convention, when we need to make sure we do that only when
5185 // we'll actually use a stack slot.
5186 auto ComputePtrOff = [&]() {
5187 /* Respect alignment of argument on the stack. */
5189 CalculateStackSlotAlignment(ArgVT, OrigVT, Flags, PtrByteSize);
5190 ArgOffset = ((ArgOffset + Align - 1) / Align) * Align;
5192 PtrOff = DAG.getConstant(ArgOffset, dl, StackPtr.getValueType());
5194 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, PtrOff);
5197 if (CallConv != CallingConv::Fast) {
5200 /* Compute GPR index associated with argument offset. */
5201 GPR_idx = (ArgOffset - LinkageSize) / PtrByteSize;
5202 GPR_idx = std::min(GPR_idx, NumGPRs);
5205 // Promote integers to 64-bit values.
5206 if (Arg.getValueType() == MVT::i32 || Arg.getValueType() == MVT::i1) {
5207 // FIXME: Should this use ANY_EXTEND if neither sext nor zext?
5208 unsigned ExtOp = Flags.isSExt() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
5209 Arg = DAG.getNode(ExtOp, dl, MVT::i64, Arg);
5212 // FIXME memcpy is used way more than necessary. Correctness first.
5213 // Note: "by value" is code for passing a structure by value, not
5215 if (Flags.isByVal()) {
5216 // Note: Size includes alignment padding, so
5217 // struct x { short a; char b; }
5218 // will have Size = 4. With #pragma pack(1), it will have Size = 3.
5219 // These are the proper values we need for right-justifying the
5220 // aggregate in a parameter register.
5221 unsigned Size = Flags.getByValSize();
5223 // An empty aggregate parameter takes up no storage and no
5228 if (CallConv == CallingConv::Fast)
5231 // All aggregates smaller than 8 bytes must be passed right-justified.
5232 if (Size==1 || Size==2 || Size==4) {
5233 EVT VT = (Size==1) ? MVT::i8 : ((Size==2) ? MVT::i16 : MVT::i32);
5234 if (GPR_idx != NumGPRs) {
5235 SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, PtrVT, Chain, Arg,
5236 MachinePointerInfo(), VT);
5237 MemOpChains.push_back(Load.getValue(1));
5238 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
5240 ArgOffset += PtrByteSize;
5245 if (GPR_idx == NumGPRs && Size < 8) {
5246 SDValue AddPtr = PtrOff;
5247 if (!isLittleEndian) {
5248 SDValue Const = DAG.getConstant(PtrByteSize - Size, dl,
5249 PtrOff.getValueType());
5250 AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, Const);
5252 Chain = CallSeqStart = createMemcpyOutsideCallSeq(Arg, AddPtr,
5255 ArgOffset += PtrByteSize;
5258 // Copy entire object into memory. There are cases where gcc-generated
5259 // code assumes it is there, even if it could be put entirely into
5260 // registers. (This is not what the doc says.)
5262 // FIXME: The above statement is likely due to a misunderstanding of the
5263 // documents. All arguments must be copied into the parameter area BY
5264 // THE CALLEE in the event that the callee takes the address of any
5265 // formal argument. That has not yet been implemented. However, it is
5266 // reasonable to use the stack area as a staging area for the register
5269 // Skip this for small aggregates, as we will use the same slot for a
5270 // right-justified copy, below.
5272 Chain = CallSeqStart = createMemcpyOutsideCallSeq(Arg, PtrOff,
5276 // When a register is available, pass a small aggregate right-justified.
5277 if (Size < 8 && GPR_idx != NumGPRs) {
5278 // The easiest way to get this right-justified in a register
5279 // is to copy the structure into the rightmost portion of a
5280 // local variable slot, then load the whole slot into the
5282 // FIXME: The memcpy seems to produce pretty awful code for
5283 // small aggregates, particularly for packed ones.
5284 // FIXME: It would be preferable to use the slot in the
5285 // parameter save area instead of a new local variable.
5286 SDValue AddPtr = PtrOff;
5287 if (!isLittleEndian) {
5288 SDValue Const = DAG.getConstant(8 - Size, dl, PtrOff.getValueType());
5289 AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, Const);
5291 Chain = CallSeqStart = createMemcpyOutsideCallSeq(Arg, AddPtr,
5295 // Load the slot into the register.
5297 DAG.getLoad(PtrVT, dl, Chain, PtrOff, MachinePointerInfo());
5298 MemOpChains.push_back(Load.getValue(1));
5299 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
5301 // Done with this argument.
5302 ArgOffset += PtrByteSize;
5306 // For aggregates larger than PtrByteSize, copy the pieces of the
5307 // object that fit into registers from the parameter save area.
5308 for (unsigned j=0; j<Size; j+=PtrByteSize) {
5309 SDValue Const = DAG.getConstant(j, dl, PtrOff.getValueType());
5310 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
5311 if (GPR_idx != NumGPRs) {
5313 DAG.getLoad(PtrVT, dl, Chain, AddArg, MachinePointerInfo());
5314 MemOpChains.push_back(Load.getValue(1));
5315 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
5316 ArgOffset += PtrByteSize;
5318 ArgOffset += ((Size - j + PtrByteSize-1)/PtrByteSize)*PtrByteSize;
5325 switch (Arg.getSimpleValueType().SimpleTy) {
5326 default: llvm_unreachable("Unexpected ValueType for argument!");
5330 if (Flags.isNest()) {
5331 // The 'nest' parameter, if any, is passed in R11.
5332 RegsToPass.push_back(std::make_pair(PPC::X11, Arg));
5337 // These can be scalar arguments or elements of an integer array type
5338 // passed directly. Clang may use those instead of "byval" aggregate
5339 // types to avoid forcing arguments to memory unnecessarily.
5340 if (GPR_idx != NumGPRs) {
5341 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Arg));
5343 if (CallConv == CallingConv::Fast)
5346 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
5347 true, isTailCall, false, MemOpChains,
5348 TailCallArguments, dl);
5349 if (CallConv == CallingConv::Fast)
5350 ArgOffset += PtrByteSize;
5352 if (CallConv != CallingConv::Fast)
5353 ArgOffset += PtrByteSize;
5357 // These can be scalar arguments or elements of a float array type
5358 // passed directly. The latter are used to implement ELFv2 homogenous
5359 // float aggregates.
5361 // Named arguments go into FPRs first, and once they overflow, the
5362 // remaining arguments go into GPRs and then the parameter save area.
5363 // Unnamed arguments for vararg functions always go to GPRs and
5364 // then the parameter save area. For now, put all arguments to vararg
5365 // routines always in both locations (FPR *and* GPR or stack slot).
5366 bool NeedGPROrStack = isVarArg || FPR_idx == NumFPRs;
5367 bool NeededLoad = false;
5369 // First load the argument into the next available FPR.
5370 if (FPR_idx != NumFPRs)
5371 RegsToPass.push_back(std::make_pair(FPR[FPR_idx++], Arg));
5373 // Next, load the argument into GPR or stack slot if needed.
5374 if (!NeedGPROrStack)
5376 else if (GPR_idx != NumGPRs && CallConv != CallingConv::Fast) {
5377 // FIXME: We may want to re-enable this for CallingConv::Fast on the P8
5378 // once we support fp <-> gpr moves.
5380 // In the non-vararg case, this can only ever happen in the
5381 // presence of f32 array types, since otherwise we never run
5382 // out of FPRs before running out of GPRs.
5385 // Double values are always passed in a single GPR.
5386 if (Arg.getValueType() != MVT::f32) {
5387 ArgVal = DAG.getNode(ISD::BITCAST, dl, MVT::i64, Arg);
5389 // Non-array float values are extended and passed in a GPR.
5390 } else if (!Flags.isInConsecutiveRegs()) {
5391 ArgVal = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Arg);
5392 ArgVal = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i64, ArgVal);
5394 // If we have an array of floats, we collect every odd element
5395 // together with its predecessor into one GPR.
5396 } else if (ArgOffset % PtrByteSize != 0) {
5398 Lo = DAG.getNode(ISD::BITCAST, dl, MVT::i32, OutVals[i - 1]);
5399 Hi = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Arg);
5400 if (!isLittleEndian)
5402 ArgVal = DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
5404 // The final element, if even, goes into the first half of a GPR.
5405 } else if (Flags.isInConsecutiveRegsLast()) {
5406 ArgVal = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Arg);
5407 ArgVal = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i64, ArgVal);
5408 if (!isLittleEndian)
5409 ArgVal = DAG.getNode(ISD::SHL, dl, MVT::i64, ArgVal,
5410 DAG.getConstant(32, dl, MVT::i32));
5412 // Non-final even elements are skipped; they will be handled
5413 // together the with subsequent argument on the next go-around.
5417 if (ArgVal.getNode())
5418 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], ArgVal));
5420 if (CallConv == CallingConv::Fast)
5423 // Single-precision floating-point values are mapped to the
5424 // second (rightmost) word of the stack doubleword.
5425 if (Arg.getValueType() == MVT::f32 &&
5426 !isLittleEndian && !Flags.isInConsecutiveRegs()) {
5427 SDValue ConstFour = DAG.getConstant(4, dl, PtrOff.getValueType());
5428 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, ConstFour);
5431 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
5432 true, isTailCall, false, MemOpChains,
5433 TailCallArguments, dl);
5437 // When passing an array of floats, the array occupies consecutive
5438 // space in the argument area; only round up to the next doubleword
5439 // at the end of the array. Otherwise, each float takes 8 bytes.
5440 if (CallConv != CallingConv::Fast || NeededLoad) {
5441 ArgOffset += (Arg.getValueType() == MVT::f32 &&
5442 Flags.isInConsecutiveRegs()) ? 4 : 8;
5443 if (Flags.isInConsecutiveRegsLast())
5444 ArgOffset = ((ArgOffset + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
5455 if (!Subtarget.hasQPX()) {
5456 // These can be scalar arguments or elements of a vector array type
5457 // passed directly. The latter are used to implement ELFv2 homogenous
5458 // vector aggregates.
5460 // For a varargs call, named arguments go into VRs or on the stack as
5461 // usual; unnamed arguments always go to the stack or the corresponding
5462 // GPRs when within range. For now, we always put the value in both
5463 // locations (or even all three).
5465 // We could elide this store in the case where the object fits
5466 // entirely in R registers. Maybe later.
5468 DAG.getStore(Chain, dl, Arg, PtrOff, MachinePointerInfo());
5469 MemOpChains.push_back(Store);
5470 if (VR_idx != NumVRs) {
5472 DAG.getLoad(MVT::v4f32, dl, Store, PtrOff, MachinePointerInfo());
5473 MemOpChains.push_back(Load.getValue(1));
5475 unsigned VReg = (Arg.getSimpleValueType() == MVT::v2f64 ||
5476 Arg.getSimpleValueType() == MVT::v2i64) ?
5477 VSRH[VR_idx] : VR[VR_idx];
5480 RegsToPass.push_back(std::make_pair(VReg, Load));
5483 for (unsigned i=0; i<16; i+=PtrByteSize) {
5484 if (GPR_idx == NumGPRs)
5486 SDValue Ix = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff,
5487 DAG.getConstant(i, dl, PtrVT));
5489 DAG.getLoad(PtrVT, dl, Store, Ix, MachinePointerInfo());
5490 MemOpChains.push_back(Load.getValue(1));
5491 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
5496 // Non-varargs Altivec params go into VRs or on the stack.
5497 if (VR_idx != NumVRs) {
5498 unsigned VReg = (Arg.getSimpleValueType() == MVT::v2f64 ||
5499 Arg.getSimpleValueType() == MVT::v2i64) ?
5500 VSRH[VR_idx] : VR[VR_idx];
5503 RegsToPass.push_back(std::make_pair(VReg, Arg));
5505 if (CallConv == CallingConv::Fast)
5508 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
5509 true, isTailCall, true, MemOpChains,
5510 TailCallArguments, dl);
5511 if (CallConv == CallingConv::Fast)
5515 if (CallConv != CallingConv::Fast)
5520 assert(Arg.getValueType().getSimpleVT().SimpleTy == MVT::v4f32 &&
5521 "Invalid QPX parameter type");
5526 bool IsF32 = Arg.getValueType().getSimpleVT().SimpleTy == MVT::v4f32;
5528 // We could elide this store in the case where the object fits
5529 // entirely in R registers. Maybe later.
5531 DAG.getStore(Chain, dl, Arg, PtrOff, MachinePointerInfo());
5532 MemOpChains.push_back(Store);
5533 if (QFPR_idx != NumQFPRs) {
5534 SDValue Load = DAG.getLoad(IsF32 ? MVT::v4f32 : MVT::v4f64, dl, Store,
5535 PtrOff, MachinePointerInfo());
5536 MemOpChains.push_back(Load.getValue(1));
5537 RegsToPass.push_back(std::make_pair(QFPR[QFPR_idx++], Load));
5539 ArgOffset += (IsF32 ? 16 : 32);
5540 for (unsigned i = 0; i < (IsF32 ? 16U : 32U); i += PtrByteSize) {
5541 if (GPR_idx == NumGPRs)
5543 SDValue Ix = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff,
5544 DAG.getConstant(i, dl, PtrVT));
5546 DAG.getLoad(PtrVT, dl, Store, Ix, MachinePointerInfo());
5547 MemOpChains.push_back(Load.getValue(1));
5548 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
5553 // Non-varargs QPX params go into registers or on the stack.
5554 if (QFPR_idx != NumQFPRs) {
5555 RegsToPass.push_back(std::make_pair(QFPR[QFPR_idx++], Arg));
5557 if (CallConv == CallingConv::Fast)
5560 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
5561 true, isTailCall, true, MemOpChains,
5562 TailCallArguments, dl);
5563 if (CallConv == CallingConv::Fast)
5564 ArgOffset += (IsF32 ? 16 : 32);
5567 if (CallConv != CallingConv::Fast)
5568 ArgOffset += (IsF32 ? 16 : 32);
5574 assert(NumBytesActuallyUsed == ArgOffset);
5575 (void)NumBytesActuallyUsed;
5577 if (!MemOpChains.empty())
5578 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains);
5580 // Check if this is an indirect call (MTCTR/BCTRL).
5581 // See PrepareCall() for more information about calls through function
5582 // pointers in the 64-bit SVR4 ABI.
5583 if (!isTailCall && !isPatchPoint &&
5584 !isFunctionGlobalAddress(Callee) &&
5585 !isa<ExternalSymbolSDNode>(Callee)) {
5586 // Load r2 into a virtual register and store it to the TOC save area.
5587 setUsesTOCBasePtr(DAG);
5588 SDValue Val = DAG.getCopyFromReg(Chain, dl, PPC::X2, MVT::i64);
5589 // TOC save area offset.
5590 unsigned TOCSaveOffset = Subtarget.getFrameLowering()->getTOCSaveOffset();
5591 SDValue PtrOff = DAG.getIntPtrConstant(TOCSaveOffset, dl);
5592 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, PtrOff);
5593 Chain = DAG.getStore(
5594 Val.getValue(1), dl, Val, AddPtr,
5595 MachinePointerInfo::getStack(DAG.getMachineFunction(), TOCSaveOffset));
5596 // In the ELFv2 ABI, R12 must contain the address of an indirect callee.
5597 // This does not mean the MTCTR instruction must use R12; it's easier
5598 // to model this as an extra parameter, so do that.
5599 if (isELFv2ABI && !isPatchPoint)
5600 RegsToPass.push_back(std::make_pair((unsigned)PPC::X12, Callee));
5603 // Build a sequence of copy-to-reg nodes chained together with token chain
5604 // and flag operands which copy the outgoing args into the appropriate regs.
5606 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
5607 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
5608 RegsToPass[i].second, InFlag);
5609 InFlag = Chain.getValue(1);
5612 if (isTailCall && !IsSibCall)
5613 PrepareTailCall(DAG, InFlag, Chain, dl, SPDiff, NumBytes, LROp, FPOp,
5616 return FinishCall(CallConv, dl, isTailCall, isVarArg, isPatchPoint, hasNest,
5617 DAG, RegsToPass, InFlag, Chain, CallSeqStart, Callee,
5618 SPDiff, NumBytes, Ins, InVals, CS);
5621 SDValue PPCTargetLowering::LowerCall_Darwin(
5622 SDValue Chain, SDValue Callee, CallingConv::ID CallConv, bool isVarArg,
5623 bool isTailCall, bool isPatchPoint,
5624 const SmallVectorImpl<ISD::OutputArg> &Outs,
5625 const SmallVectorImpl<SDValue> &OutVals,
5626 const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl,
5627 SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals,
5628 ImmutableCallSite *CS) const {
5630 unsigned NumOps = Outs.size();
5632 EVT PtrVT = getPointerTy(DAG.getDataLayout());
5633 bool isPPC64 = PtrVT == MVT::i64;
5634 unsigned PtrByteSize = isPPC64 ? 8 : 4;
5636 MachineFunction &MF = DAG.getMachineFunction();
5638 // Mark this function as potentially containing a function that contains a
5639 // tail call. As a consequence the frame pointer will be used for dynamicalloc
5640 // and restoring the callers stack pointer in this functions epilog. This is
5641 // done because by tail calling the called function might overwrite the value
5642 // in this function's (MF) stack pointer stack slot 0(SP).
5643 if (getTargetMachine().Options.GuaranteedTailCallOpt &&
5644 CallConv == CallingConv::Fast)
5645 MF.getInfo<PPCFunctionInfo>()->setHasFastCall();
5647 // Count how many bytes are to be pushed on the stack, including the linkage
5648 // area, and parameter passing area. We start with 24/48 bytes, which is
5649 // prereserved space for [SP][CR][LR][3 x unused].
5650 unsigned LinkageSize = Subtarget.getFrameLowering()->getLinkageSize();
5651 unsigned NumBytes = LinkageSize;
5653 // Add up all the space actually used.
5654 // In 32-bit non-varargs calls, Altivec parameters all go at the end; usually
5655 // they all go in registers, but we must reserve stack space for them for
5656 // possible use by the caller. In varargs or 64-bit calls, parameters are
5657 // assigned stack space in order, with padding so Altivec parameters are
5659 unsigned nAltivecParamsAtEnd = 0;
5660 for (unsigned i = 0; i != NumOps; ++i) {
5661 ISD::ArgFlagsTy Flags = Outs[i].Flags;
5662 EVT ArgVT = Outs[i].VT;
5663 // Varargs Altivec parameters are padded to a 16 byte boundary.
5664 if (ArgVT == MVT::v4f32 || ArgVT == MVT::v4i32 ||
5665 ArgVT == MVT::v8i16 || ArgVT == MVT::v16i8 ||
5666 ArgVT == MVT::v2f64 || ArgVT == MVT::v2i64) {
5667 if (!isVarArg && !isPPC64) {
5668 // Non-varargs Altivec parameters go after all the non-Altivec
5669 // parameters; handle those later so we know how much padding we need.
5670 nAltivecParamsAtEnd++;
5673 // Varargs and 64-bit Altivec parameters are padded to 16 byte boundary.
5674 NumBytes = ((NumBytes+15)/16)*16;
5676 NumBytes += CalculateStackSlotSize(ArgVT, Flags, PtrByteSize);
5679 // Allow for Altivec parameters at the end, if needed.
5680 if (nAltivecParamsAtEnd) {
5681 NumBytes = ((NumBytes+15)/16)*16;
5682 NumBytes += 16*nAltivecParamsAtEnd;
5685 // The prolog code of the callee may store up to 8 GPR argument registers to
5686 // the stack, allowing va_start to index over them in memory if its varargs.
5687 // Because we cannot tell if this is needed on the caller side, we have to
5688 // conservatively assume that it is needed. As such, make sure we have at
5689 // least enough stack space for the caller to store the 8 GPRs.
5690 NumBytes = std::max(NumBytes, LinkageSize + 8 * PtrByteSize);
5692 // Tail call needs the stack to be aligned.
5693 if (getTargetMachine().Options.GuaranteedTailCallOpt &&
5694 CallConv == CallingConv::Fast)
5695 NumBytes = EnsureStackAlignment(Subtarget.getFrameLowering(), NumBytes);
5697 // Calculate by how many bytes the stack has to be adjusted in case of tail
5698 // call optimization.
5699 int SPDiff = CalculateTailCallSPDiff(DAG, isTailCall, NumBytes);
5701 // To protect arguments on the stack from being clobbered in a tail call,
5702 // force all the loads to happen before doing any other lowering.
5704 Chain = DAG.getStackArgumentTokenFactor(Chain);
5706 // Adjust the stack pointer for the new arguments...
5707 // These operations are automatically eliminated by the prolog/epilog pass
5708 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, dl, true),
5710 SDValue CallSeqStart = Chain;
5712 // Load the return address and frame pointer so it can be move somewhere else
5715 Chain = EmitTailCallLoadFPAndRetAddr(DAG, SPDiff, Chain, LROp, FPOp, dl);
5717 // Set up a copy of the stack pointer for use loading and storing any
5718 // arguments that may not fit in the registers available for argument
5722 StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
5724 StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
5726 // Figure out which arguments are going to go in registers, and which in
5727 // memory. Also, if this is a vararg function, floating point operations
5728 // must be stored to our stack, and loaded into integer regs as well, if
5729 // any integer regs are available for argument passing.
5730 unsigned ArgOffset = LinkageSize;
5731 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
5733 static const MCPhysReg GPR_32[] = { // 32-bit registers.
5734 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
5735 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
5737 static const MCPhysReg GPR_64[] = { // 64-bit registers.
5738 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
5739 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
5741 static const MCPhysReg VR[] = {
5742 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
5743 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
5745 const unsigned NumGPRs = array_lengthof(GPR_32);
5746 const unsigned NumFPRs = 13;
5747 const unsigned NumVRs = array_lengthof(VR);
5749 const MCPhysReg *GPR = isPPC64 ? GPR_64 : GPR_32;
5751 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
5752 SmallVector<TailCallArgumentInfo, 8> TailCallArguments;
5754 SmallVector<SDValue, 8> MemOpChains;
5755 for (unsigned i = 0; i != NumOps; ++i) {
5756 SDValue Arg = OutVals[i];
5757 ISD::ArgFlagsTy Flags = Outs[i].Flags;
5759 // PtrOff will be used to store the current argument to the stack if a
5760 // register cannot be found for it.
5763 PtrOff = DAG.getConstant(ArgOffset, dl, StackPtr.getValueType());
5765 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, PtrOff);
5767 // On PPC64, promote integers to 64-bit values.
5768 if (isPPC64 && Arg.getValueType() == MVT::i32) {
5769 // FIXME: Should this use ANY_EXTEND if neither sext nor zext?
5770 unsigned ExtOp = Flags.isSExt() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
5771 Arg = DAG.getNode(ExtOp, dl, MVT::i64, Arg);
5774 // FIXME memcpy is used way more than necessary. Correctness first.
5775 // Note: "by value" is code for passing a structure by value, not
5777 if (Flags.isByVal()) {
5778 unsigned Size = Flags.getByValSize();
5779 // Very small objects are passed right-justified. Everything else is
5780 // passed left-justified.
5781 if (Size==1 || Size==2) {
5782 EVT VT = (Size==1) ? MVT::i8 : MVT::i16;
5783 if (GPR_idx != NumGPRs) {
5784 SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, PtrVT, Chain, Arg,
5785 MachinePointerInfo(), VT);
5786 MemOpChains.push_back(Load.getValue(1));
5787 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
5789 ArgOffset += PtrByteSize;
5791 SDValue Const = DAG.getConstant(PtrByteSize - Size, dl,
5792 PtrOff.getValueType());
5793 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, Const);
5794 Chain = CallSeqStart = createMemcpyOutsideCallSeq(Arg, AddPtr,
5797 ArgOffset += PtrByteSize;
5801 // Copy entire object into memory. There are cases where gcc-generated
5802 // code assumes it is there, even if it could be put entirely into
5803 // registers. (This is not what the doc says.)
5804 Chain = CallSeqStart = createMemcpyOutsideCallSeq(Arg, PtrOff,
5808 // For small aggregates (Darwin only) and aggregates >= PtrByteSize,
5809 // copy the pieces of the object that fit into registers from the
5810 // parameter save area.
5811 for (unsigned j=0; j<Size; j+=PtrByteSize) {
5812 SDValue Const = DAG.getConstant(j, dl, PtrOff.getValueType());
5813 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
5814 if (GPR_idx != NumGPRs) {
5816 DAG.getLoad(PtrVT, dl, Chain, AddArg, MachinePointerInfo());
5817 MemOpChains.push_back(Load.getValue(1));
5818 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
5819 ArgOffset += PtrByteSize;
5821 ArgOffset += ((Size - j + PtrByteSize-1)/PtrByteSize)*PtrByteSize;
5828 switch (Arg.getSimpleValueType().SimpleTy) {
5829 default: llvm_unreachable("Unexpected ValueType for argument!");
5833 if (GPR_idx != NumGPRs) {
5834 if (Arg.getValueType() == MVT::i1)
5835 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, PtrVT, Arg);
5837 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Arg));
5839 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
5840 isPPC64, isTailCall, false, MemOpChains,
5841 TailCallArguments, dl);
5843 ArgOffset += PtrByteSize;
5847 if (FPR_idx != NumFPRs) {
5848 RegsToPass.push_back(std::make_pair(FPR[FPR_idx++], Arg));
5852 DAG.getStore(Chain, dl, Arg, PtrOff, MachinePointerInfo());
5853 MemOpChains.push_back(Store);
5855 // Float varargs are always shadowed in available integer registers
5856 if (GPR_idx != NumGPRs) {
5858 DAG.getLoad(PtrVT, dl, Store, PtrOff, MachinePointerInfo());
5859 MemOpChains.push_back(Load.getValue(1));
5860 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
5862 if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 && !isPPC64){
5863 SDValue ConstFour = DAG.getConstant(4, dl, PtrOff.getValueType());
5864 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, ConstFour);
5866 DAG.getLoad(PtrVT, dl, Store, PtrOff, MachinePointerInfo());
5867 MemOpChains.push_back(Load.getValue(1));
5868 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
5871 // If we have any FPRs remaining, we may also have GPRs remaining.
5872 // Args passed in FPRs consume either 1 (f32) or 2 (f64) available
5874 if (GPR_idx != NumGPRs)
5876 if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 &&
5877 !isPPC64) // PPC64 has 64-bit GPR's obviously :)
5881 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
5882 isPPC64, isTailCall, false, MemOpChains,
5883 TailCallArguments, dl);
5887 ArgOffset += Arg.getValueType() == MVT::f32 ? 4 : 8;
5894 // These go aligned on the stack, or in the corresponding R registers
5895 // when within range. The Darwin PPC ABI doc claims they also go in
5896 // V registers; in fact gcc does this only for arguments that are
5897 // prototyped, not for those that match the ... We do it for all
5898 // arguments, seems to work.
5899 while (ArgOffset % 16 !=0) {
5900 ArgOffset += PtrByteSize;
5901 if (GPR_idx != NumGPRs)
5904 // We could elide this store in the case where the object fits
5905 // entirely in R registers. Maybe later.
5906 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr,
5907 DAG.getConstant(ArgOffset, dl, PtrVT));
5909 DAG.getStore(Chain, dl, Arg, PtrOff, MachinePointerInfo());
5910 MemOpChains.push_back(Store);
5911 if (VR_idx != NumVRs) {
5913 DAG.getLoad(MVT::v4f32, dl, Store, PtrOff, MachinePointerInfo());
5914 MemOpChains.push_back(Load.getValue(1));
5915 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Load));
5918 for (unsigned i=0; i<16; i+=PtrByteSize) {
5919 if (GPR_idx == NumGPRs)
5921 SDValue Ix = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff,
5922 DAG.getConstant(i, dl, PtrVT));
5924 DAG.getLoad(PtrVT, dl, Store, Ix, MachinePointerInfo());
5925 MemOpChains.push_back(Load.getValue(1));
5926 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
5931 // Non-varargs Altivec params generally go in registers, but have
5932 // stack space allocated at the end.
5933 if (VR_idx != NumVRs) {
5934 // Doesn't have GPR space allocated.
5935 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Arg));
5936 } else if (nAltivecParamsAtEnd==0) {
5937 // We are emitting Altivec params in order.
5938 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
5939 isPPC64, isTailCall, true, MemOpChains,
5940 TailCallArguments, dl);
5946 // If all Altivec parameters fit in registers, as they usually do,
5947 // they get stack space following the non-Altivec parameters. We
5948 // don't track this here because nobody below needs it.
5949 // If there are more Altivec parameters than fit in registers emit
5951 if (!isVarArg && nAltivecParamsAtEnd > NumVRs) {
5953 // Offset is aligned; skip 1st 12 params which go in V registers.
5954 ArgOffset = ((ArgOffset+15)/16)*16;
5956 for (unsigned i = 0; i != NumOps; ++i) {
5957 SDValue Arg = OutVals[i];
5958 EVT ArgType = Outs[i].VT;
5959 if (ArgType==MVT::v4f32 || ArgType==MVT::v4i32 ||
5960 ArgType==MVT::v8i16 || ArgType==MVT::v16i8) {
5963 // We are emitting Altivec params in order.
5964 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
5965 isPPC64, isTailCall, true, MemOpChains,
5966 TailCallArguments, dl);
5973 if (!MemOpChains.empty())
5974 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains);
5976 // On Darwin, R12 must contain the address of an indirect callee. This does
5977 // not mean the MTCTR instruction must use R12; it's easier to model this as
5978 // an extra parameter, so do that.
5980 !isFunctionGlobalAddress(Callee) &&
5981 !isa<ExternalSymbolSDNode>(Callee) &&
5982 !isBLACompatibleAddress(Callee, DAG))
5983 RegsToPass.push_back(std::make_pair((unsigned)(isPPC64 ? PPC::X12 :
5984 PPC::R12), Callee));
5986 // Build a sequence of copy-to-reg nodes chained together with token chain
5987 // and flag operands which copy the outgoing args into the appropriate regs.
5989 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
5990 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
5991 RegsToPass[i].second, InFlag);
5992 InFlag = Chain.getValue(1);
5996 PrepareTailCall(DAG, InFlag, Chain, dl, SPDiff, NumBytes, LROp, FPOp,
5999 return FinishCall(CallConv, dl, isTailCall, isVarArg, isPatchPoint,
6000 /* unused except on PPC64 ELFv1 */ false, DAG,
6001 RegsToPass, InFlag, Chain, CallSeqStart, Callee, SPDiff,
6002 NumBytes, Ins, InVals, CS);
6006 PPCTargetLowering::CanLowerReturn(CallingConv::ID CallConv,
6007 MachineFunction &MF, bool isVarArg,
6008 const SmallVectorImpl<ISD::OutputArg> &Outs,
6009 LLVMContext &Context) const {
6010 SmallVector<CCValAssign, 16> RVLocs;
6011 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, Context);
6012 return CCInfo.CheckReturn(Outs, RetCC_PPC);
6016 PPCTargetLowering::LowerReturn(SDValue Chain, CallingConv::ID CallConv,
6018 const SmallVectorImpl<ISD::OutputArg> &Outs,
6019 const SmallVectorImpl<SDValue> &OutVals,
6020 const SDLoc &dl, SelectionDAG &DAG) const {
6022 SmallVector<CCValAssign, 16> RVLocs;
6023 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
6025 CCInfo.AnalyzeReturn(Outs, RetCC_PPC);
6028 SmallVector<SDValue, 4> RetOps(1, Chain);
6030 // Copy the result values into the output registers.
6031 for (unsigned i = 0; i != RVLocs.size(); ++i) {
6032 CCValAssign &VA = RVLocs[i];
6033 assert(VA.isRegLoc() && "Can only return in registers!");
6035 SDValue Arg = OutVals[i];
6037 switch (VA.getLocInfo()) {
6038 default: llvm_unreachable("Unknown loc info!");
6039 case CCValAssign::Full: break;
6040 case CCValAssign::AExt:
6041 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
6043 case CCValAssign::ZExt:
6044 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
6046 case CCValAssign::SExt:
6047 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
6051 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
6052 Flag = Chain.getValue(1);
6053 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
6056 const PPCRegisterInfo *TRI = Subtarget.getRegisterInfo();
6057 const MCPhysReg *I =
6058 TRI->getCalleeSavedRegsViaCopy(&DAG.getMachineFunction());
6062 if (PPC::G8RCRegClass.contains(*I))
6063 RetOps.push_back(DAG.getRegister(*I, MVT::i64));
6064 else if (PPC::F8RCRegClass.contains(*I))
6065 RetOps.push_back(DAG.getRegister(*I, MVT::getFloatingPointVT(64)));
6066 else if (PPC::CRRCRegClass.contains(*I))
6067 RetOps.push_back(DAG.getRegister(*I, MVT::i1));
6068 else if (PPC::VRRCRegClass.contains(*I))
6069 RetOps.push_back(DAG.getRegister(*I, MVT::Other));
6071 llvm_unreachable("Unexpected register class in CSRsViaCopy!");
6075 RetOps[0] = Chain; // Update chain.
6077 // Add the flag if we have it.
6079 RetOps.push_back(Flag);
6081 return DAG.getNode(PPCISD::RET_FLAG, dl, MVT::Other, RetOps);
6085 PPCTargetLowering::LowerGET_DYNAMIC_AREA_OFFSET(SDValue Op,
6086 SelectionDAG &DAG) const {
6089 // Get the corect type for integers.
6090 EVT IntVT = Op.getValueType();
6093 SDValue Chain = Op.getOperand(0);
6094 SDValue FPSIdx = getFramePointerFrameIndex(DAG);
6095 // Build a DYNAREAOFFSET node.
6096 SDValue Ops[2] = {Chain, FPSIdx};
6097 SDVTList VTs = DAG.getVTList(IntVT);
6098 return DAG.getNode(PPCISD::DYNAREAOFFSET, dl, VTs, Ops);
6101 SDValue PPCTargetLowering::LowerSTACKRESTORE(SDValue Op,
6102 SelectionDAG &DAG) const {
6103 // When we pop the dynamic allocation we need to restore the SP link.
6106 // Get the corect type for pointers.
6107 EVT PtrVT = getPointerTy(DAG.getDataLayout());
6109 // Construct the stack pointer operand.
6110 bool isPPC64 = Subtarget.isPPC64();
6111 unsigned SP = isPPC64 ? PPC::X1 : PPC::R1;
6112 SDValue StackPtr = DAG.getRegister(SP, PtrVT);
6114 // Get the operands for the STACKRESTORE.
6115 SDValue Chain = Op.getOperand(0);
6116 SDValue SaveSP = Op.getOperand(1);
6118 // Load the old link SP.
6119 SDValue LoadLinkSP =
6120 DAG.getLoad(PtrVT, dl, Chain, StackPtr, MachinePointerInfo());
6122 // Restore the stack pointer.
6123 Chain = DAG.getCopyToReg(LoadLinkSP.getValue(1), dl, SP, SaveSP);
6125 // Store the old link SP.
6126 return DAG.getStore(Chain, dl, LoadLinkSP, StackPtr, MachinePointerInfo());
6129 SDValue PPCTargetLowering::getReturnAddrFrameIndex(SelectionDAG &DAG) const {
6130 MachineFunction &MF = DAG.getMachineFunction();
6131 bool isPPC64 = Subtarget.isPPC64();
6132 EVT PtrVT = getPointerTy(MF.getDataLayout());
6134 // Get current frame pointer save index. The users of this index will be
6135 // primarily DYNALLOC instructions.
6136 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
6137 int RASI = FI->getReturnAddrSaveIndex();
6139 // If the frame pointer save index hasn't been defined yet.
6141 // Find out what the fix offset of the frame pointer save area.
6142 int LROffset = Subtarget.getFrameLowering()->getReturnSaveOffset();
6143 // Allocate the frame index for frame pointer save area.
6144 RASI = MF.getFrameInfo().CreateFixedObject(isPPC64? 8 : 4, LROffset, false);
6146 FI->setReturnAddrSaveIndex(RASI);
6148 return DAG.getFrameIndex(RASI, PtrVT);
6152 PPCTargetLowering::getFramePointerFrameIndex(SelectionDAG & DAG) const {
6153 MachineFunction &MF = DAG.getMachineFunction();
6154 bool isPPC64 = Subtarget.isPPC64();
6155 EVT PtrVT = getPointerTy(MF.getDataLayout());
6157 // Get current frame pointer save index. The users of this index will be
6158 // primarily DYNALLOC instructions.
6159 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
6160 int FPSI = FI->getFramePointerSaveIndex();
6162 // If the frame pointer save index hasn't been defined yet.
6164 // Find out what the fix offset of the frame pointer save area.
6165 int FPOffset = Subtarget.getFrameLowering()->getFramePointerSaveOffset();
6166 // Allocate the frame index for frame pointer save area.
6167 FPSI = MF.getFrameInfo().CreateFixedObject(isPPC64? 8 : 4, FPOffset, true);
6169 FI->setFramePointerSaveIndex(FPSI);
6171 return DAG.getFrameIndex(FPSI, PtrVT);
6174 SDValue PPCTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
6175 SelectionDAG &DAG) const {
6177 SDValue Chain = Op.getOperand(0);
6178 SDValue Size = Op.getOperand(1);
6181 // Get the corect type for pointers.
6182 EVT PtrVT = getPointerTy(DAG.getDataLayout());
6184 SDValue NegSize = DAG.getNode(ISD::SUB, dl, PtrVT,
6185 DAG.getConstant(0, dl, PtrVT), Size);
6186 // Construct a node for the frame pointer save index.
6187 SDValue FPSIdx = getFramePointerFrameIndex(DAG);
6188 // Build a DYNALLOC node.
6189 SDValue Ops[3] = { Chain, NegSize, FPSIdx };
6190 SDVTList VTs = DAG.getVTList(PtrVT, MVT::Other);
6191 return DAG.getNode(PPCISD::DYNALLOC, dl, VTs, Ops);
6194 SDValue PPCTargetLowering::LowerEH_DWARF_CFA(SDValue Op,
6195 SelectionDAG &DAG) const {
6196 MachineFunction &MF = DAG.getMachineFunction();
6198 bool isPPC64 = Subtarget.isPPC64();
6199 EVT PtrVT = getPointerTy(DAG.getDataLayout());
6201 int FI = MF.getFrameInfo().CreateFixedObject(isPPC64 ? 8 : 4, 0, false);
6202 return DAG.getFrameIndex(FI, PtrVT);
6205 SDValue PPCTargetLowering::lowerEH_SJLJ_SETJMP(SDValue Op,
6206 SelectionDAG &DAG) const {
6208 return DAG.getNode(PPCISD::EH_SJLJ_SETJMP, DL,
6209 DAG.getVTList(MVT::i32, MVT::Other),
6210 Op.getOperand(0), Op.getOperand(1));
6213 SDValue PPCTargetLowering::lowerEH_SJLJ_LONGJMP(SDValue Op,
6214 SelectionDAG &DAG) const {
6216 return DAG.getNode(PPCISD::EH_SJLJ_LONGJMP, DL, MVT::Other,
6217 Op.getOperand(0), Op.getOperand(1));
6220 SDValue PPCTargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const {
6221 if (Op.getValueType().isVector())
6222 return LowerVectorLoad(Op, DAG);
6224 assert(Op.getValueType() == MVT::i1 &&
6225 "Custom lowering only for i1 loads");
6227 // First, load 8 bits into 32 bits, then truncate to 1 bit.
6230 LoadSDNode *LD = cast<LoadSDNode>(Op);
6232 SDValue Chain = LD->getChain();
6233 SDValue BasePtr = LD->getBasePtr();
6234 MachineMemOperand *MMO = LD->getMemOperand();
6237 DAG.getExtLoad(ISD::EXTLOAD, dl, getPointerTy(DAG.getDataLayout()), Chain,
6238 BasePtr, MVT::i8, MMO);
6239 SDValue Result = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, NewLD);
6241 SDValue Ops[] = { Result, SDValue(NewLD.getNode(), 1) };
6242 return DAG.getMergeValues(Ops, dl);
6245 SDValue PPCTargetLowering::LowerSTORE(SDValue Op, SelectionDAG &DAG) const {
6246 if (Op.getOperand(1).getValueType().isVector())
6247 return LowerVectorStore(Op, DAG);
6249 assert(Op.getOperand(1).getValueType() == MVT::i1 &&
6250 "Custom lowering only for i1 stores");
6252 // First, zero extend to 32 bits, then use a truncating store to 8 bits.
6255 StoreSDNode *ST = cast<StoreSDNode>(Op);
6257 SDValue Chain = ST->getChain();
6258 SDValue BasePtr = ST->getBasePtr();
6259 SDValue Value = ST->getValue();
6260 MachineMemOperand *MMO = ST->getMemOperand();
6262 Value = DAG.getNode(ISD::ZERO_EXTEND, dl, getPointerTy(DAG.getDataLayout()),
6264 return DAG.getTruncStore(Chain, dl, Value, BasePtr, MVT::i8, MMO);
6267 // FIXME: Remove this once the ANDI glue bug is fixed:
6268 SDValue PPCTargetLowering::LowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const {
6269 assert(Op.getValueType() == MVT::i1 &&
6270 "Custom lowering only for i1 results");
6273 return DAG.getNode(PPCISD::ANDIo_1_GT_BIT, DL, MVT::i1,
6277 /// LowerSELECT_CC - Lower floating point select_cc's into fsel instruction when
6279 SDValue PPCTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
6280 // Not FP? Not a fsel.
6281 if (!Op.getOperand(0).getValueType().isFloatingPoint() ||
6282 !Op.getOperand(2).getValueType().isFloatingPoint())
6285 // We might be able to do better than this under some circumstances, but in
6286 // general, fsel-based lowering of select is a finite-math-only optimization.
6287 // For more information, see section F.3 of the 2.06 ISA specification.
6288 if (!DAG.getTarget().Options.NoInfsFPMath ||
6289 !DAG.getTarget().Options.NoNaNsFPMath)
6291 // TODO: Propagate flags from the select rather than global settings.
6293 Flags.setNoInfs(true);
6294 Flags.setNoNaNs(true);
6296 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
6298 EVT ResVT = Op.getValueType();
6299 EVT CmpVT = Op.getOperand(0).getValueType();
6300 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
6301 SDValue TV = Op.getOperand(2), FV = Op.getOperand(3);
6304 // If the RHS of the comparison is a 0.0, we don't need to do the
6305 // subtraction at all.
6307 if (isFloatingPointZero(RHS))
6309 default: break; // SETUO etc aren't handled by fsel.
6313 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
6314 LHS = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, LHS);
6315 Sel1 = DAG.getNode(PPCISD::FSEL, dl, ResVT, LHS, TV, FV);
6316 if (Sel1.getValueType() == MVT::f32) // Comparison is always 64-bits
6317 Sel1 = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Sel1);
6318 return DAG.getNode(PPCISD::FSEL, dl, ResVT,
6319 DAG.getNode(ISD::FNEG, dl, MVT::f64, LHS), Sel1, FV);
6322 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
6325 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
6326 LHS = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, LHS);
6327 return DAG.getNode(PPCISD::FSEL, dl, ResVT, LHS, TV, FV);
6330 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
6333 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
6334 LHS = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, LHS);
6335 return DAG.getNode(PPCISD::FSEL, dl, ResVT,
6336 DAG.getNode(ISD::FNEG, dl, MVT::f64, LHS), TV, FV);
6341 default: break; // SETUO etc aren't handled by fsel.
6345 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, LHS, RHS, &Flags);
6346 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
6347 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
6348 Sel1 = DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, TV, FV);
6349 if (Sel1.getValueType() == MVT::f32) // Comparison is always 64-bits
6350 Sel1 = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Sel1);
6351 return DAG.getNode(PPCISD::FSEL, dl, ResVT,
6352 DAG.getNode(ISD::FNEG, dl, MVT::f64, Cmp), Sel1, FV);
6355 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, LHS, RHS, &Flags);
6356 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
6357 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
6358 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, FV, TV);
6361 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, LHS, RHS, &Flags);
6362 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
6363 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
6364 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, TV, FV);
6367 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, RHS, LHS, &Flags);
6368 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
6369 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
6370 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, FV, TV);
6373 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, RHS, LHS, &Flags);
6374 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
6375 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
6376 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, TV, FV);
6381 void PPCTargetLowering::LowerFP_TO_INTForReuse(SDValue Op, ReuseLoadInfo &RLI,
6383 const SDLoc &dl) const {
6384 assert(Op.getOperand(0).getValueType().isFloatingPoint());
6385 SDValue Src = Op.getOperand(0);
6386 if (Src.getValueType() == MVT::f32)
6387 Src = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Src);
6390 switch (Op.getSimpleValueType().SimpleTy) {
6391 default: llvm_unreachable("Unhandled FP_TO_INT type in custom expander!");
6394 Op.getOpcode() == ISD::FP_TO_SINT
6396 : (Subtarget.hasFPCVT() ? PPCISD::FCTIWUZ : PPCISD::FCTIDZ),
6400 assert((Op.getOpcode() == ISD::FP_TO_SINT || Subtarget.hasFPCVT()) &&
6401 "i64 FP_TO_UINT is supported only with FPCVT");
6402 Tmp = DAG.getNode(Op.getOpcode()==ISD::FP_TO_SINT ? PPCISD::FCTIDZ :
6408 // Convert the FP value to an int value through memory.
6409 bool i32Stack = Op.getValueType() == MVT::i32 && Subtarget.hasSTFIWX() &&
6410 (Op.getOpcode() == ISD::FP_TO_SINT || Subtarget.hasFPCVT());
6411 SDValue FIPtr = DAG.CreateStackTemporary(i32Stack ? MVT::i32 : MVT::f64);
6412 int FI = cast<FrameIndexSDNode>(FIPtr)->getIndex();
6413 MachinePointerInfo MPI =
6414 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI);
6416 // Emit a store to the stack slot.
6419 MachineFunction &MF = DAG.getMachineFunction();
6420 MachineMemOperand *MMO =
6421 MF.getMachineMemOperand(MPI, MachineMemOperand::MOStore, 4, 4);
6422 SDValue Ops[] = { DAG.getEntryNode(), Tmp, FIPtr };
6423 Chain = DAG.getMemIntrinsicNode(PPCISD::STFIWX, dl,
6424 DAG.getVTList(MVT::Other), Ops, MVT::i32, MMO);
6426 Chain = DAG.getStore(DAG.getEntryNode(), dl, Tmp, FIPtr, MPI);
6428 // Result is a load from the stack slot. If loading 4 bytes, make sure to
6429 // add in a bias on big endian.
6430 if (Op.getValueType() == MVT::i32 && !i32Stack) {
6431 FIPtr = DAG.getNode(ISD::ADD, dl, FIPtr.getValueType(), FIPtr,
6432 DAG.getConstant(4, dl, FIPtr.getValueType()));
6433 MPI = MPI.getWithOffset(Subtarget.isLittleEndian() ? 0 : 4);
6441 /// \brief Custom lowers floating point to integer conversions to use
6442 /// the direct move instructions available in ISA 2.07 to avoid the
6443 /// need for load/store combinations.
6444 SDValue PPCTargetLowering::LowerFP_TO_INTDirectMove(SDValue Op,
6446 const SDLoc &dl) const {
6447 assert(Op.getOperand(0).getValueType().isFloatingPoint());
6448 SDValue Src = Op.getOperand(0);
6450 if (Src.getValueType() == MVT::f32)
6451 Src = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Src);
6454 switch (Op.getSimpleValueType().SimpleTy) {
6455 default: llvm_unreachable("Unhandled FP_TO_INT type in custom expander!");
6458 Op.getOpcode() == ISD::FP_TO_SINT
6460 : (Subtarget.hasFPCVT() ? PPCISD::FCTIWUZ : PPCISD::FCTIDZ),
6462 Tmp = DAG.getNode(PPCISD::MFVSR, dl, MVT::i32, Tmp);
6465 assert((Op.getOpcode() == ISD::FP_TO_SINT || Subtarget.hasFPCVT()) &&
6466 "i64 FP_TO_UINT is supported only with FPCVT");
6467 Tmp = DAG.getNode(Op.getOpcode()==ISD::FP_TO_SINT ? PPCISD::FCTIDZ :
6470 Tmp = DAG.getNode(PPCISD::MFVSR, dl, MVT::i64, Tmp);
6476 SDValue PPCTargetLowering::LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG,
6477 const SDLoc &dl) const {
6478 if (Subtarget.hasDirectMove() && Subtarget.isPPC64())
6479 return LowerFP_TO_INTDirectMove(Op, DAG, dl);
6482 LowerFP_TO_INTForReuse(Op, RLI, DAG, dl);
6484 return DAG.getLoad(Op.getValueType(), dl, RLI.Chain, RLI.Ptr, RLI.MPI,
6485 RLI.Alignment, RLI.MMOFlags(), RLI.AAInfo, RLI.Ranges);
6488 // We're trying to insert a regular store, S, and then a load, L. If the
6489 // incoming value, O, is a load, we might just be able to have our load use the
6490 // address used by O. However, we don't know if anything else will store to
6491 // that address before we can load from it. To prevent this situation, we need
6492 // to insert our load, L, into the chain as a peer of O. To do this, we give L
6493 // the same chain operand as O, we create a token factor from the chain results
6494 // of O and L, and we replace all uses of O's chain result with that token
6495 // factor (see spliceIntoChain below for this last part).
6496 bool PPCTargetLowering::canReuseLoadAddress(SDValue Op, EVT MemVT,
6499 ISD::LoadExtType ET) const {
6501 if (ET == ISD::NON_EXTLOAD &&
6502 (Op.getOpcode() == ISD::FP_TO_UINT ||
6503 Op.getOpcode() == ISD::FP_TO_SINT) &&
6504 isOperationLegalOrCustom(Op.getOpcode(),
6505 Op.getOperand(0).getValueType())) {
6507 LowerFP_TO_INTForReuse(Op, RLI, DAG, dl);
6511 LoadSDNode *LD = dyn_cast<LoadSDNode>(Op);
6512 if (!LD || LD->getExtensionType() != ET || LD->isVolatile() ||
6513 LD->isNonTemporal())
6515 if (LD->getMemoryVT() != MemVT)
6518 RLI.Ptr = LD->getBasePtr();
6519 if (LD->isIndexed() && !LD->getOffset().isUndef()) {
6520 assert(LD->getAddressingMode() == ISD::PRE_INC &&
6521 "Non-pre-inc AM on PPC?");
6522 RLI.Ptr = DAG.getNode(ISD::ADD, dl, RLI.Ptr.getValueType(), RLI.Ptr,
6526 RLI.Chain = LD->getChain();
6527 RLI.MPI = LD->getPointerInfo();
6528 RLI.IsDereferenceable = LD->isDereferenceable();
6529 RLI.IsInvariant = LD->isInvariant();
6530 RLI.Alignment = LD->getAlignment();
6531 RLI.AAInfo = LD->getAAInfo();
6532 RLI.Ranges = LD->getRanges();
6534 RLI.ResChain = SDValue(LD, LD->isIndexed() ? 2 : 1);
6538 // Given the head of the old chain, ResChain, insert a token factor containing
6539 // it and NewResChain, and make users of ResChain now be users of that token
6541 void PPCTargetLowering::spliceIntoChain(SDValue ResChain,
6542 SDValue NewResChain,
6543 SelectionDAG &DAG) const {
6547 SDLoc dl(NewResChain);
6549 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6550 NewResChain, DAG.getUNDEF(MVT::Other));
6551 assert(TF.getNode() != NewResChain.getNode() &&
6552 "A new TF really is required here");
6554 DAG.ReplaceAllUsesOfValueWith(ResChain, TF);
6555 DAG.UpdateNodeOperands(TF.getNode(), ResChain, NewResChain);
6558 /// \brief Analyze profitability of direct move
6559 /// prefer float load to int load plus direct move
6560 /// when there is no integer use of int load
6561 static bool directMoveIsProfitable(const SDValue &Op) {
6562 SDNode *Origin = Op.getOperand(0).getNode();
6563 if (Origin->getOpcode() != ISD::LOAD)
6566 for (SDNode::use_iterator UI = Origin->use_begin(),
6567 UE = Origin->use_end();
6570 // Only look at the users of the loaded value.
6571 if (UI.getUse().get().getResNo() != 0)
6574 if (UI->getOpcode() != ISD::SINT_TO_FP &&
6575 UI->getOpcode() != ISD::UINT_TO_FP)
6582 /// \brief Custom lowers integer to floating point conversions to use
6583 /// the direct move instructions available in ISA 2.07 to avoid the
6584 /// need for load/store combinations.
6585 SDValue PPCTargetLowering::LowerINT_TO_FPDirectMove(SDValue Op,
6587 const SDLoc &dl) const {
6588 assert((Op.getValueType() == MVT::f32 ||
6589 Op.getValueType() == MVT::f64) &&
6590 "Invalid floating point type as target of conversion");
6591 assert(Subtarget.hasFPCVT() &&
6592 "Int to FP conversions with direct moves require FPCVT");
6594 SDValue Src = Op.getOperand(0);
6595 bool SinglePrec = Op.getValueType() == MVT::f32;
6596 bool WordInt = Src.getSimpleValueType().SimpleTy == MVT::i32;
6597 bool Signed = Op.getOpcode() == ISD::SINT_TO_FP;
6598 unsigned ConvOp = Signed ? (SinglePrec ? PPCISD::FCFIDS : PPCISD::FCFID) :
6599 (SinglePrec ? PPCISD::FCFIDUS : PPCISD::FCFIDU);
6602 FP = DAG.getNode(Signed ? PPCISD::MTVSRA : PPCISD::MTVSRZ,
6604 FP = DAG.getNode(ConvOp, dl, SinglePrec ? MVT::f32 : MVT::f64, FP);
6607 FP = DAG.getNode(PPCISD::MTVSRA, dl, MVT::f64, Src);
6608 FP = DAG.getNode(ConvOp, dl, SinglePrec ? MVT::f32 : MVT::f64, FP);
6614 SDValue PPCTargetLowering::LowerINT_TO_FP(SDValue Op,
6615 SelectionDAG &DAG) const {
6618 if (Subtarget.hasQPX() && Op.getOperand(0).getValueType() == MVT::v4i1) {
6619 if (Op.getValueType() != MVT::v4f32 && Op.getValueType() != MVT::v4f64)
6622 SDValue Value = Op.getOperand(0);
6623 // The values are now known to be -1 (false) or 1 (true). To convert this
6624 // into 0 (false) and 1 (true), add 1 and then divide by 2 (multiply by 0.5).
6625 // This can be done with an fma and the 0.5 constant: (V+1.0)*0.5 = 0.5*V+0.5
6626 Value = DAG.getNode(PPCISD::QBFLT, dl, MVT::v4f64, Value);
6628 SDValue FPHalfs = DAG.getConstantFP(0.5, dl, MVT::v4f64);
6630 Value = DAG.getNode(ISD::FMA, dl, MVT::v4f64, Value, FPHalfs, FPHalfs);
6632 if (Op.getValueType() != MVT::v4f64)
6633 Value = DAG.getNode(ISD::FP_ROUND, dl,
6634 Op.getValueType(), Value,
6635 DAG.getIntPtrConstant(1, dl));
6639 // Don't handle ppc_fp128 here; let it be lowered to a libcall.
6640 if (Op.getValueType() != MVT::f32 && Op.getValueType() != MVT::f64)
6643 if (Op.getOperand(0).getValueType() == MVT::i1)
6644 return DAG.getNode(ISD::SELECT, dl, Op.getValueType(), Op.getOperand(0),
6645 DAG.getConstantFP(1.0, dl, Op.getValueType()),
6646 DAG.getConstantFP(0.0, dl, Op.getValueType()));
6648 // If we have direct moves, we can do all the conversion, skip the store/load
6649 // however, without FPCVT we can't do most conversions.
6650 if (Subtarget.hasDirectMove() && directMoveIsProfitable(Op) &&
6651 Subtarget.isPPC64() && Subtarget.hasFPCVT())
6652 return LowerINT_TO_FPDirectMove(Op, DAG, dl);
6654 assert((Op.getOpcode() == ISD::SINT_TO_FP || Subtarget.hasFPCVT()) &&
6655 "UINT_TO_FP is supported only with FPCVT");
6657 // If we have FCFIDS, then use it when converting to single-precision.
6658 // Otherwise, convert to double-precision and then round.
6659 unsigned FCFOp = (Subtarget.hasFPCVT() && Op.getValueType() == MVT::f32)
6660 ? (Op.getOpcode() == ISD::UINT_TO_FP ? PPCISD::FCFIDUS
6662 : (Op.getOpcode() == ISD::UINT_TO_FP ? PPCISD::FCFIDU
6664 MVT FCFTy = (Subtarget.hasFPCVT() && Op.getValueType() == MVT::f32)
6668 if (Op.getOperand(0).getValueType() == MVT::i64) {
6669 SDValue SINT = Op.getOperand(0);
6670 // When converting to single-precision, we actually need to convert
6671 // to double-precision first and then round to single-precision.
6672 // To avoid double-rounding effects during that operation, we have
6673 // to prepare the input operand. Bits that might be truncated when
6674 // converting to double-precision are replaced by a bit that won't
6675 // be lost at this stage, but is below the single-precision rounding
6678 // However, if -enable-unsafe-fp-math is in effect, accept double
6679 // rounding to avoid the extra overhead.
6680 if (Op.getValueType() == MVT::f32 &&
6681 !Subtarget.hasFPCVT() &&
6682 !DAG.getTarget().Options.UnsafeFPMath) {
6684 // Twiddle input to make sure the low 11 bits are zero. (If this
6685 // is the case, we are guaranteed the value will fit into the 53 bit
6686 // mantissa of an IEEE double-precision value without rounding.)
6687 // If any of those low 11 bits were not zero originally, make sure
6688 // bit 12 (value 2048) is set instead, so that the final rounding
6689 // to single-precision gets the correct result.
6690 SDValue Round = DAG.getNode(ISD::AND, dl, MVT::i64,
6691 SINT, DAG.getConstant(2047, dl, MVT::i64));
6692 Round = DAG.getNode(ISD::ADD, dl, MVT::i64,
6693 Round, DAG.getConstant(2047, dl, MVT::i64));
6694 Round = DAG.getNode(ISD::OR, dl, MVT::i64, Round, SINT);
6695 Round = DAG.getNode(ISD::AND, dl, MVT::i64,
6696 Round, DAG.getConstant(-2048, dl, MVT::i64));
6698 // However, we cannot use that value unconditionally: if the magnitude
6699 // of the input value is small, the bit-twiddling we did above might
6700 // end up visibly changing the output. Fortunately, in that case, we
6701 // don't need to twiddle bits since the original input will convert
6702 // exactly to double-precision floating-point already. Therefore,
6703 // construct a conditional to use the original value if the top 11
6704 // bits are all sign-bit copies, and use the rounded value computed
6706 SDValue Cond = DAG.getNode(ISD::SRA, dl, MVT::i64,
6707 SINT, DAG.getConstant(53, dl, MVT::i32));
6708 Cond = DAG.getNode(ISD::ADD, dl, MVT::i64,
6709 Cond, DAG.getConstant(1, dl, MVT::i64));
6710 Cond = DAG.getSetCC(dl, MVT::i32,
6711 Cond, DAG.getConstant(1, dl, MVT::i64), ISD::SETUGT);
6713 SINT = DAG.getNode(ISD::SELECT, dl, MVT::i64, Cond, Round, SINT);
6719 MachineFunction &MF = DAG.getMachineFunction();
6720 if (canReuseLoadAddress(SINT, MVT::i64, RLI, DAG)) {
6721 Bits = DAG.getLoad(MVT::f64, dl, RLI.Chain, RLI.Ptr, RLI.MPI,
6722 RLI.Alignment, RLI.MMOFlags(), RLI.AAInfo, RLI.Ranges);
6723 spliceIntoChain(RLI.ResChain, Bits.getValue(1), DAG);
6724 } else if (Subtarget.hasLFIWAX() &&
6725 canReuseLoadAddress(SINT, MVT::i32, RLI, DAG, ISD::SEXTLOAD)) {
6726 MachineMemOperand *MMO =
6727 MF.getMachineMemOperand(RLI.MPI, MachineMemOperand::MOLoad, 4,
6728 RLI.Alignment, RLI.AAInfo, RLI.Ranges);
6729 SDValue Ops[] = { RLI.Chain, RLI.Ptr };
6730 Bits = DAG.getMemIntrinsicNode(PPCISD::LFIWAX, dl,
6731 DAG.getVTList(MVT::f64, MVT::Other),
6732 Ops, MVT::i32, MMO);
6733 spliceIntoChain(RLI.ResChain, Bits.getValue(1), DAG);
6734 } else if (Subtarget.hasFPCVT() &&
6735 canReuseLoadAddress(SINT, MVT::i32, RLI, DAG, ISD::ZEXTLOAD)) {
6736 MachineMemOperand *MMO =
6737 MF.getMachineMemOperand(RLI.MPI, MachineMemOperand::MOLoad, 4,
6738 RLI.Alignment, RLI.AAInfo, RLI.Ranges);
6739 SDValue Ops[] = { RLI.Chain, RLI.Ptr };
6740 Bits = DAG.getMemIntrinsicNode(PPCISD::LFIWZX, dl,
6741 DAG.getVTList(MVT::f64, MVT::Other),
6742 Ops, MVT::i32, MMO);
6743 spliceIntoChain(RLI.ResChain, Bits.getValue(1), DAG);
6744 } else if (((Subtarget.hasLFIWAX() &&
6745 SINT.getOpcode() == ISD::SIGN_EXTEND) ||
6746 (Subtarget.hasFPCVT() &&
6747 SINT.getOpcode() == ISD::ZERO_EXTEND)) &&
6748 SINT.getOperand(0).getValueType() == MVT::i32) {
6749 MachineFrameInfo &MFI = MF.getFrameInfo();
6750 EVT PtrVT = getPointerTy(DAG.getDataLayout());
6752 int FrameIdx = MFI.CreateStackObject(4, 4, false);
6753 SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
6756 DAG.getStore(DAG.getEntryNode(), dl, SINT.getOperand(0), FIdx,
6757 MachinePointerInfo::getFixedStack(
6758 DAG.getMachineFunction(), FrameIdx));
6760 assert(cast<StoreSDNode>(Store)->getMemoryVT() == MVT::i32 &&
6761 "Expected an i32 store");
6766 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FrameIdx);
6769 MachineMemOperand *MMO =
6770 MF.getMachineMemOperand(RLI.MPI, MachineMemOperand::MOLoad, 4,
6771 RLI.Alignment, RLI.AAInfo, RLI.Ranges);
6772 SDValue Ops[] = { RLI.Chain, RLI.Ptr };
6773 Bits = DAG.getMemIntrinsicNode(SINT.getOpcode() == ISD::ZERO_EXTEND ?
6774 PPCISD::LFIWZX : PPCISD::LFIWAX,
6775 dl, DAG.getVTList(MVT::f64, MVT::Other),
6776 Ops, MVT::i32, MMO);
6778 Bits = DAG.getNode(ISD::BITCAST, dl, MVT::f64, SINT);
6780 SDValue FP = DAG.getNode(FCFOp, dl, FCFTy, Bits);
6782 if (Op.getValueType() == MVT::f32 && !Subtarget.hasFPCVT())
6783 FP = DAG.getNode(ISD::FP_ROUND, dl,
6784 MVT::f32, FP, DAG.getIntPtrConstant(0, dl));
6788 assert(Op.getOperand(0).getValueType() == MVT::i32 &&
6789 "Unhandled INT_TO_FP type in custom expander!");
6790 // Since we only generate this in 64-bit mode, we can take advantage of
6791 // 64-bit registers. In particular, sign extend the input value into the
6792 // 64-bit register with extsw, store the WHOLE 64-bit value into the stack
6793 // then lfd it and fcfid it.
6794 MachineFunction &MF = DAG.getMachineFunction();
6795 MachineFrameInfo &MFI = MF.getFrameInfo();
6796 EVT PtrVT = getPointerTy(MF.getDataLayout());
6799 if (Subtarget.hasLFIWAX() || Subtarget.hasFPCVT()) {
6802 if (!(ReusingLoad = canReuseLoadAddress(Op.getOperand(0), MVT::i32, RLI,
6804 int FrameIdx = MFI.CreateStackObject(4, 4, false);
6805 SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
6808 DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0), FIdx,
6809 MachinePointerInfo::getFixedStack(
6810 DAG.getMachineFunction(), FrameIdx));
6812 assert(cast<StoreSDNode>(Store)->getMemoryVT() == MVT::i32 &&
6813 "Expected an i32 store");
6818 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FrameIdx);
6822 MachineMemOperand *MMO =
6823 MF.getMachineMemOperand(RLI.MPI, MachineMemOperand::MOLoad, 4,
6824 RLI.Alignment, RLI.AAInfo, RLI.Ranges);
6825 SDValue Ops[] = { RLI.Chain, RLI.Ptr };
6826 Ld = DAG.getMemIntrinsicNode(Op.getOpcode() == ISD::UINT_TO_FP ?
6827 PPCISD::LFIWZX : PPCISD::LFIWAX,
6828 dl, DAG.getVTList(MVT::f64, MVT::Other),
6829 Ops, MVT::i32, MMO);
6831 spliceIntoChain(RLI.ResChain, Ld.getValue(1), DAG);
6833 assert(Subtarget.isPPC64() &&
6834 "i32->FP without LFIWAX supported only on PPC64");
6836 int FrameIdx = MFI.CreateStackObject(8, 8, false);
6837 SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
6839 SDValue Ext64 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::i64,
6842 // STD the extended value into the stack slot.
6843 SDValue Store = DAG.getStore(
6844 DAG.getEntryNode(), dl, Ext64, FIdx,
6845 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FrameIdx));
6847 // Load the value as a double.
6849 MVT::f64, dl, Store, FIdx,
6850 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FrameIdx));
6853 // FCFID it and return it.
6854 SDValue FP = DAG.getNode(FCFOp, dl, FCFTy, Ld);
6855 if (Op.getValueType() == MVT::f32 && !Subtarget.hasFPCVT())
6856 FP = DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, FP,
6857 DAG.getIntPtrConstant(0, dl));
6861 SDValue PPCTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
6862 SelectionDAG &DAG) const {
6865 The rounding mode is in bits 30:31 of FPSR, and has the following
6872 FLT_ROUNDS, on the other hand, expects the following:
6879 To perform the conversion, we do:
6880 ((FPSCR & 0x3) ^ ((~FPSCR & 0x3) >> 1))
6883 MachineFunction &MF = DAG.getMachineFunction();
6884 EVT VT = Op.getValueType();
6885 EVT PtrVT = getPointerTy(MF.getDataLayout());
6887 // Save FP Control Word to register
6889 MVT::f64, // return register
6890 MVT::Glue // unused in this context
6892 SDValue Chain = DAG.getNode(PPCISD::MFFS, dl, NodeTys, None);
6894 // Save FP register to stack slot
6895 int SSFI = MF.getFrameInfo().CreateStackObject(8, 8, false);
6896 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
6897 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Chain, StackSlot,
6898 MachinePointerInfo());
6900 // Load FP Control Word from low 32 bits of stack slot.
6901 SDValue Four = DAG.getConstant(4, dl, PtrVT);
6902 SDValue Addr = DAG.getNode(ISD::ADD, dl, PtrVT, StackSlot, Four);
6903 SDValue CWD = DAG.getLoad(MVT::i32, dl, Store, Addr, MachinePointerInfo());
6905 // Transform as necessary
6907 DAG.getNode(ISD::AND, dl, MVT::i32,
6908 CWD, DAG.getConstant(3, dl, MVT::i32));
6910 DAG.getNode(ISD::SRL, dl, MVT::i32,
6911 DAG.getNode(ISD::AND, dl, MVT::i32,
6912 DAG.getNode(ISD::XOR, dl, MVT::i32,
6913 CWD, DAG.getConstant(3, dl, MVT::i32)),
6914 DAG.getConstant(3, dl, MVT::i32)),
6915 DAG.getConstant(1, dl, MVT::i32));
6918 DAG.getNode(ISD::XOR, dl, MVT::i32, CWD1, CWD2);
6920 return DAG.getNode((VT.getSizeInBits() < 16 ?
6921 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
6924 SDValue PPCTargetLowering::LowerSHL_PARTS(SDValue Op, SelectionDAG &DAG) const {
6925 EVT VT = Op.getValueType();
6926 unsigned BitWidth = VT.getSizeInBits();
6928 assert(Op.getNumOperands() == 3 &&
6929 VT == Op.getOperand(1).getValueType() &&
6932 // Expand into a bunch of logical ops. Note that these ops
6933 // depend on the PPC behavior for oversized shift amounts.
6934 SDValue Lo = Op.getOperand(0);
6935 SDValue Hi = Op.getOperand(1);
6936 SDValue Amt = Op.getOperand(2);
6937 EVT AmtVT = Amt.getValueType();
6939 SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
6940 DAG.getConstant(BitWidth, dl, AmtVT), Amt);
6941 SDValue Tmp2 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Amt);
6942 SDValue Tmp3 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Tmp1);
6943 SDValue Tmp4 = DAG.getNode(ISD::OR , dl, VT, Tmp2, Tmp3);
6944 SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
6945 DAG.getConstant(-BitWidth, dl, AmtVT));
6946 SDValue Tmp6 = DAG.getNode(PPCISD::SHL, dl, VT, Lo, Tmp5);
6947 SDValue OutHi = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp6);
6948 SDValue OutLo = DAG.getNode(PPCISD::SHL, dl, VT, Lo, Amt);
6949 SDValue OutOps[] = { OutLo, OutHi };
6950 return DAG.getMergeValues(OutOps, dl);
6953 SDValue PPCTargetLowering::LowerSRL_PARTS(SDValue Op, SelectionDAG &DAG) const {
6954 EVT VT = Op.getValueType();
6956 unsigned BitWidth = VT.getSizeInBits();
6957 assert(Op.getNumOperands() == 3 &&
6958 VT == Op.getOperand(1).getValueType() &&
6961 // Expand into a bunch of logical ops. Note that these ops
6962 // depend on the PPC behavior for oversized shift amounts.
6963 SDValue Lo = Op.getOperand(0);
6964 SDValue Hi = Op.getOperand(1);
6965 SDValue Amt = Op.getOperand(2);
6966 EVT AmtVT = Amt.getValueType();
6968 SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
6969 DAG.getConstant(BitWidth, dl, AmtVT), Amt);
6970 SDValue Tmp2 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Amt);
6971 SDValue Tmp3 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Tmp1);
6972 SDValue Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp3);
6973 SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
6974 DAG.getConstant(-BitWidth, dl, AmtVT));
6975 SDValue Tmp6 = DAG.getNode(PPCISD::SRL, dl, VT, Hi, Tmp5);
6976 SDValue OutLo = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp6);
6977 SDValue OutHi = DAG.getNode(PPCISD::SRL, dl, VT, Hi, Amt);
6978 SDValue OutOps[] = { OutLo, OutHi };
6979 return DAG.getMergeValues(OutOps, dl);
6982 SDValue PPCTargetLowering::LowerSRA_PARTS(SDValue Op, SelectionDAG &DAG) const {
6984 EVT VT = Op.getValueType();
6985 unsigned BitWidth = VT.getSizeInBits();
6986 assert(Op.getNumOperands() == 3 &&
6987 VT == Op.getOperand(1).getValueType() &&
6990 // Expand into a bunch of logical ops, followed by a select_cc.
6991 SDValue Lo = Op.getOperand(0);
6992 SDValue Hi = Op.getOperand(1);
6993 SDValue Amt = Op.getOperand(2);
6994 EVT AmtVT = Amt.getValueType();
6996 SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
6997 DAG.getConstant(BitWidth, dl, AmtVT), Amt);
6998 SDValue Tmp2 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Amt);
6999 SDValue Tmp3 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Tmp1);
7000 SDValue Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp3);
7001 SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
7002 DAG.getConstant(-BitWidth, dl, AmtVT));
7003 SDValue Tmp6 = DAG.getNode(PPCISD::SRA, dl, VT, Hi, Tmp5);
7004 SDValue OutHi = DAG.getNode(PPCISD::SRA, dl, VT, Hi, Amt);
7005 SDValue OutLo = DAG.getSelectCC(dl, Tmp5, DAG.getConstant(0, dl, AmtVT),
7006 Tmp4, Tmp6, ISD::SETLE);
7007 SDValue OutOps[] = { OutLo, OutHi };
7008 return DAG.getMergeValues(OutOps, dl);
7011 //===----------------------------------------------------------------------===//
7012 // Vector related lowering.
7015 /// BuildSplatI - Build a canonical splati of Val with an element size of
7016 /// SplatSize. Cast the result to VT.
7017 static SDValue BuildSplatI(int Val, unsigned SplatSize, EVT VT,
7018 SelectionDAG &DAG, const SDLoc &dl) {
7019 assert(Val >= -16 && Val <= 15 && "vsplti is out of range!");
7021 static const MVT VTys[] = { // canonical VT to use for each size.
7022 MVT::v16i8, MVT::v8i16, MVT::Other, MVT::v4i32
7025 EVT ReqVT = VT != MVT::Other ? VT : VTys[SplatSize-1];
7027 // Force vspltis[hw] -1 to vspltisb -1 to canonicalize.
7031 EVT CanonicalVT = VTys[SplatSize-1];
7033 // Build a canonical splat for this value.
7034 return DAG.getBitcast(ReqVT, DAG.getConstant(Val, dl, CanonicalVT));
7037 /// BuildIntrinsicOp - Return a unary operator intrinsic node with the
7038 /// specified intrinsic ID.
7039 static SDValue BuildIntrinsicOp(unsigned IID, SDValue Op, SelectionDAG &DAG,
7040 const SDLoc &dl, EVT DestVT = MVT::Other) {
7041 if (DestVT == MVT::Other) DestVT = Op.getValueType();
7042 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, DestVT,
7043 DAG.getConstant(IID, dl, MVT::i32), Op);
7046 /// BuildIntrinsicOp - Return a binary operator intrinsic node with the
7047 /// specified intrinsic ID.
7048 static SDValue BuildIntrinsicOp(unsigned IID, SDValue LHS, SDValue RHS,
7049 SelectionDAG &DAG, const SDLoc &dl,
7050 EVT DestVT = MVT::Other) {
7051 if (DestVT == MVT::Other) DestVT = LHS.getValueType();
7052 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, DestVT,
7053 DAG.getConstant(IID, dl, MVT::i32), LHS, RHS);
7056 /// BuildIntrinsicOp - Return a ternary operator intrinsic node with the
7057 /// specified intrinsic ID.
7058 static SDValue BuildIntrinsicOp(unsigned IID, SDValue Op0, SDValue Op1,
7059 SDValue Op2, SelectionDAG &DAG, const SDLoc &dl,
7060 EVT DestVT = MVT::Other) {
7061 if (DestVT == MVT::Other) DestVT = Op0.getValueType();
7062 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, DestVT,
7063 DAG.getConstant(IID, dl, MVT::i32), Op0, Op1, Op2);
7066 /// BuildVSLDOI - Return a VECTOR_SHUFFLE that is a vsldoi of the specified
7067 /// amount. The result has the specified value type.
7068 static SDValue BuildVSLDOI(SDValue LHS, SDValue RHS, unsigned Amt, EVT VT,
7069 SelectionDAG &DAG, const SDLoc &dl) {
7070 // Force LHS/RHS to be the right type.
7071 LHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, LHS);
7072 RHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, RHS);
7075 for (unsigned i = 0; i != 16; ++i)
7077 SDValue T = DAG.getVectorShuffle(MVT::v16i8, dl, LHS, RHS, Ops);
7078 return DAG.getNode(ISD::BITCAST, dl, VT, T);
7081 // If this is a case we can't handle, return null and let the default
7082 // expansion code take care of it. If we CAN select this case, and if it
7083 // selects to a single instruction, return Op. Otherwise, if we can codegen
7084 // this case more efficiently than a constant pool load, lower it to the
7085 // sequence of ops that should be used.
7086 SDValue PPCTargetLowering::LowerBUILD_VECTOR(SDValue Op,
7087 SelectionDAG &DAG) const {
7089 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
7090 assert(BVN && "Expected a BuildVectorSDNode in LowerBUILD_VECTOR");
7092 if (Subtarget.hasQPX() && Op.getValueType() == MVT::v4i1) {
7093 // We first build an i32 vector, load it into a QPX register,
7094 // then convert it to a floating-point vector and compare it
7095 // to a zero vector to get the boolean result.
7096 MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo();
7097 int FrameIdx = MFI.CreateStackObject(16, 16, false);
7098 MachinePointerInfo PtrInfo =
7099 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FrameIdx);
7100 EVT PtrVT = getPointerTy(DAG.getDataLayout());
7101 SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
7103 assert(BVN->getNumOperands() == 4 &&
7104 "BUILD_VECTOR for v4i1 does not have 4 operands");
7106 bool IsConst = true;
7107 for (unsigned i = 0; i < 4; ++i) {
7108 if (BVN->getOperand(i).isUndef()) continue;
7109 if (!isa<ConstantSDNode>(BVN->getOperand(i))) {
7117 ConstantFP::get(Type::getFloatTy(*DAG.getContext()), 1.0);
7119 ConstantFP::get(Type::getFloatTy(*DAG.getContext()), -1.0);
7122 for (unsigned i = 0; i < 4; ++i) {
7123 if (BVN->getOperand(i).isUndef())
7124 CV[i] = UndefValue::get(Type::getFloatTy(*DAG.getContext()));
7125 else if (isNullConstant(BVN->getOperand(i)))
7131 Constant *CP = ConstantVector::get(CV);
7132 SDValue CPIdx = DAG.getConstantPool(CP, getPointerTy(DAG.getDataLayout()),
7133 16 /* alignment */);
7135 SDValue Ops[] = {DAG.getEntryNode(), CPIdx};
7136 SDVTList VTs = DAG.getVTList({MVT::v4i1, /*chain*/ MVT::Other});
7137 return DAG.getMemIntrinsicNode(
7138 PPCISD::QVLFSb, dl, VTs, Ops, MVT::v4f32,
7139 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()));
7142 SmallVector<SDValue, 4> Stores;
7143 for (unsigned i = 0; i < 4; ++i) {
7144 if (BVN->getOperand(i).isUndef()) continue;
7146 unsigned Offset = 4*i;
7147 SDValue Idx = DAG.getConstant(Offset, dl, FIdx.getValueType());
7148 Idx = DAG.getNode(ISD::ADD, dl, FIdx.getValueType(), FIdx, Idx);
7150 unsigned StoreSize = BVN->getOperand(i).getValueType().getStoreSize();
7151 if (StoreSize > 4) {
7153 DAG.getTruncStore(DAG.getEntryNode(), dl, BVN->getOperand(i), Idx,
7154 PtrInfo.getWithOffset(Offset), MVT::i32));
7156 SDValue StoreValue = BVN->getOperand(i);
7158 StoreValue = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, StoreValue);
7160 Stores.push_back(DAG.getStore(DAG.getEntryNode(), dl, StoreValue, Idx,
7161 PtrInfo.getWithOffset(Offset)));
7166 if (!Stores.empty())
7167 StoreChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Stores);
7169 StoreChain = DAG.getEntryNode();
7171 // Now load from v4i32 into the QPX register; this will extend it to
7172 // v4i64 but not yet convert it to a floating point. Nevertheless, this
7173 // is typed as v4f64 because the QPX register integer states are not
7174 // explicitly represented.
7176 SDValue Ops[] = {StoreChain,
7177 DAG.getConstant(Intrinsic::ppc_qpx_qvlfiwz, dl, MVT::i32),
7179 SDVTList VTs = DAG.getVTList({MVT::v4f64, /*chain*/ MVT::Other});
7181 SDValue LoadedVect = DAG.getMemIntrinsicNode(ISD::INTRINSIC_W_CHAIN,
7182 dl, VTs, Ops, MVT::v4i32, PtrInfo);
7183 LoadedVect = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f64,
7184 DAG.getConstant(Intrinsic::ppc_qpx_qvfcfidu, dl, MVT::i32),
7187 SDValue FPZeros = DAG.getConstantFP(0.0, dl, MVT::v4f64);
7189 return DAG.getSetCC(dl, MVT::v4i1, LoadedVect, FPZeros, ISD::SETEQ);
7192 // All other QPX vectors are handled by generic code.
7193 if (Subtarget.hasQPX())
7196 // Check if this is a splat of a constant value.
7197 APInt APSplatBits, APSplatUndef;
7198 unsigned SplatBitSize;
7200 if (! BVN->isConstantSplat(APSplatBits, APSplatUndef, SplatBitSize,
7201 HasAnyUndefs, 0, !Subtarget.isLittleEndian()) ||
7205 unsigned SplatBits = APSplatBits.getZExtValue();
7206 unsigned SplatUndef = APSplatUndef.getZExtValue();
7207 unsigned SplatSize = SplatBitSize / 8;
7209 // First, handle single instruction cases.
7212 if (SplatBits == 0) {
7213 // Canonicalize all zero vectors to be v4i32.
7214 if (Op.getValueType() != MVT::v4i32 || HasAnyUndefs) {
7215 SDValue Z = DAG.getConstant(0, dl, MVT::v4i32);
7216 Op = DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Z);
7221 // If the sign extended value is in the range [-16,15], use VSPLTI[bhw].
7222 int32_t SextVal= (int32_t(SplatBits << (32-SplatBitSize)) >>
7224 if (SextVal >= -16 && SextVal <= 15)
7225 return BuildSplatI(SextVal, SplatSize, Op.getValueType(), DAG, dl);
7227 // Two instruction sequences.
7229 // If this value is in the range [-32,30] and is even, use:
7230 // VSPLTI[bhw](val/2) + VSPLTI[bhw](val/2)
7231 // If this value is in the range [17,31] and is odd, use:
7232 // VSPLTI[bhw](val-16) - VSPLTI[bhw](-16)
7233 // If this value is in the range [-31,-17] and is odd, use:
7234 // VSPLTI[bhw](val+16) + VSPLTI[bhw](-16)
7235 // Note the last two are three-instruction sequences.
7236 if (SextVal >= -32 && SextVal <= 31) {
7237 // To avoid having these optimizations undone by constant folding,
7238 // we convert to a pseudo that will be expanded later into one of
7240 SDValue Elt = DAG.getConstant(SextVal, dl, MVT::i32);
7241 EVT VT = (SplatSize == 1 ? MVT::v16i8 :
7242 (SplatSize == 2 ? MVT::v8i16 : MVT::v4i32));
7243 SDValue EltSize = DAG.getConstant(SplatSize, dl, MVT::i32);
7244 SDValue RetVal = DAG.getNode(PPCISD::VADD_SPLAT, dl, VT, Elt, EltSize);
7245 if (VT == Op.getValueType())
7248 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), RetVal);
7251 // If this is 0x8000_0000 x 4, turn into vspltisw + vslw. If it is
7252 // 0x7FFF_FFFF x 4, turn it into not(0x8000_0000). This is important
7254 if (SplatSize == 4 && SplatBits == (0x7FFFFFFF&~SplatUndef)) {
7255 // Make -1 and vspltisw -1:
7256 SDValue OnesV = BuildSplatI(-1, 4, MVT::v4i32, DAG, dl);
7258 // Make the VSLW intrinsic, computing 0x8000_0000.
7259 SDValue Res = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, OnesV,
7262 // xor by OnesV to invert it.
7263 Res = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Res, OnesV);
7264 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
7267 // Check to see if this is a wide variety of vsplti*, binop self cases.
7268 static const signed char SplatCsts[] = {
7269 -1, 1, -2, 2, -3, 3, -4, 4, -5, 5, -6, 6, -7, 7,
7270 -8, 8, -9, 9, -10, 10, -11, 11, -12, 12, -13, 13, 14, -14, 15, -15, -16
7273 for (unsigned idx = 0; idx < array_lengthof(SplatCsts); ++idx) {
7274 // Indirect through the SplatCsts array so that we favor 'vsplti -1' for
7275 // cases which are ambiguous (e.g. formation of 0x8000_0000). 'vsplti -1'
7276 int i = SplatCsts[idx];
7278 // Figure out what shift amount will be used by altivec if shifted by i in
7280 unsigned TypeShiftAmt = i & (SplatBitSize-1);
7282 // vsplti + shl self.
7283 if (SextVal == (int)((unsigned)i << TypeShiftAmt)) {
7284 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
7285 static const unsigned IIDs[] = { // Intrinsic to use for each size.
7286 Intrinsic::ppc_altivec_vslb, Intrinsic::ppc_altivec_vslh, 0,
7287 Intrinsic::ppc_altivec_vslw
7289 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
7290 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
7293 // vsplti + srl self.
7294 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
7295 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
7296 static const unsigned IIDs[] = { // Intrinsic to use for each size.
7297 Intrinsic::ppc_altivec_vsrb, Intrinsic::ppc_altivec_vsrh, 0,
7298 Intrinsic::ppc_altivec_vsrw
7300 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
7301 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
7304 // vsplti + sra self.
7305 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
7306 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
7307 static const unsigned IIDs[] = { // Intrinsic to use for each size.
7308 Intrinsic::ppc_altivec_vsrab, Intrinsic::ppc_altivec_vsrah, 0,
7309 Intrinsic::ppc_altivec_vsraw
7311 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
7312 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
7315 // vsplti + rol self.
7316 if (SextVal == (int)(((unsigned)i << TypeShiftAmt) |
7317 ((unsigned)i >> (SplatBitSize-TypeShiftAmt)))) {
7318 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
7319 static const unsigned IIDs[] = { // Intrinsic to use for each size.
7320 Intrinsic::ppc_altivec_vrlb, Intrinsic::ppc_altivec_vrlh, 0,
7321 Intrinsic::ppc_altivec_vrlw
7323 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
7324 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
7327 // t = vsplti c, result = vsldoi t, t, 1
7328 if (SextVal == (int)(((unsigned)i << 8) | (i < 0 ? 0xFF : 0))) {
7329 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
7330 unsigned Amt = Subtarget.isLittleEndian() ? 15 : 1;
7331 return BuildVSLDOI(T, T, Amt, Op.getValueType(), DAG, dl);
7333 // t = vsplti c, result = vsldoi t, t, 2
7334 if (SextVal == (int)(((unsigned)i << 16) | (i < 0 ? 0xFFFF : 0))) {
7335 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
7336 unsigned Amt = Subtarget.isLittleEndian() ? 14 : 2;
7337 return BuildVSLDOI(T, T, Amt, Op.getValueType(), DAG, dl);
7339 // t = vsplti c, result = vsldoi t, t, 3
7340 if (SextVal == (int)(((unsigned)i << 24) | (i < 0 ? 0xFFFFFF : 0))) {
7341 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
7342 unsigned Amt = Subtarget.isLittleEndian() ? 13 : 3;
7343 return BuildVSLDOI(T, T, Amt, Op.getValueType(), DAG, dl);
7350 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
7351 /// the specified operations to build the shuffle.
7352 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
7353 SDValue RHS, SelectionDAG &DAG,
7355 unsigned OpNum = (PFEntry >> 26) & 0x0F;
7356 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
7357 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
7360 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
7372 if (OpNum == OP_COPY) {
7373 if (LHSID == (1*9+2)*9+3) return LHS;
7374 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
7378 SDValue OpLHS, OpRHS;
7379 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
7380 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
7384 default: llvm_unreachable("Unknown i32 permute!");
7386 ShufIdxs[ 0] = 0; ShufIdxs[ 1] = 1; ShufIdxs[ 2] = 2; ShufIdxs[ 3] = 3;
7387 ShufIdxs[ 4] = 16; ShufIdxs[ 5] = 17; ShufIdxs[ 6] = 18; ShufIdxs[ 7] = 19;
7388 ShufIdxs[ 8] = 4; ShufIdxs[ 9] = 5; ShufIdxs[10] = 6; ShufIdxs[11] = 7;
7389 ShufIdxs[12] = 20; ShufIdxs[13] = 21; ShufIdxs[14] = 22; ShufIdxs[15] = 23;
7392 ShufIdxs[ 0] = 8; ShufIdxs[ 1] = 9; ShufIdxs[ 2] = 10; ShufIdxs[ 3] = 11;
7393 ShufIdxs[ 4] = 24; ShufIdxs[ 5] = 25; ShufIdxs[ 6] = 26; ShufIdxs[ 7] = 27;
7394 ShufIdxs[ 8] = 12; ShufIdxs[ 9] = 13; ShufIdxs[10] = 14; ShufIdxs[11] = 15;
7395 ShufIdxs[12] = 28; ShufIdxs[13] = 29; ShufIdxs[14] = 30; ShufIdxs[15] = 31;
7398 for (unsigned i = 0; i != 16; ++i)
7399 ShufIdxs[i] = (i&3)+0;
7402 for (unsigned i = 0; i != 16; ++i)
7403 ShufIdxs[i] = (i&3)+4;
7406 for (unsigned i = 0; i != 16; ++i)
7407 ShufIdxs[i] = (i&3)+8;
7410 for (unsigned i = 0; i != 16; ++i)
7411 ShufIdxs[i] = (i&3)+12;
7414 return BuildVSLDOI(OpLHS, OpRHS, 4, OpLHS.getValueType(), DAG, dl);
7416 return BuildVSLDOI(OpLHS, OpRHS, 8, OpLHS.getValueType(), DAG, dl);
7418 return BuildVSLDOI(OpLHS, OpRHS, 12, OpLHS.getValueType(), DAG, dl);
7420 EVT VT = OpLHS.getValueType();
7421 OpLHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OpLHS);
7422 OpRHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OpRHS);
7423 SDValue T = DAG.getVectorShuffle(MVT::v16i8, dl, OpLHS, OpRHS, ShufIdxs);
7424 return DAG.getNode(ISD::BITCAST, dl, VT, T);
7427 /// LowerVECTOR_SHUFFLE - Return the code we lower for VECTOR_SHUFFLE. If this
7428 /// is a shuffle we can handle in a single instruction, return it. Otherwise,
7429 /// return the code it can be lowered into. Worst case, it can always be
7430 /// lowered into a vperm.
7431 SDValue PPCTargetLowering::LowerVECTOR_SHUFFLE(SDValue Op,
7432 SelectionDAG &DAG) const {
7434 SDValue V1 = Op.getOperand(0);
7435 SDValue V2 = Op.getOperand(1);
7436 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
7437 EVT VT = Op.getValueType();
7438 bool isLittleEndian = Subtarget.isLittleEndian();
7440 unsigned ShiftElts, InsertAtByte;
7442 if (Subtarget.hasP9Vector() &&
7443 PPC::isXXINSERTWMask(SVOp, ShiftElts, InsertAtByte, Swap,
7447 SDValue Conv1 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, V1);
7448 SDValue Conv2 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, V2);
7450 SDValue Shl = DAG.getNode(PPCISD::VECSHL, dl, MVT::v4i32, Conv2, Conv2,
7451 DAG.getConstant(ShiftElts, dl, MVT::i32));
7452 SDValue Ins = DAG.getNode(PPCISD::XXINSERT, dl, MVT::v4i32, Conv1, Shl,
7453 DAG.getConstant(InsertAtByte, dl, MVT::i32));
7454 return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, Ins);
7456 SDValue Ins = DAG.getNode(PPCISD::XXINSERT, dl, MVT::v4i32, Conv1, Conv2,
7457 DAG.getConstant(InsertAtByte, dl, MVT::i32));
7458 return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, Ins);
7461 if (Subtarget.hasVSX()) {
7462 if (V2.isUndef() && PPC::isSplatShuffleMask(SVOp, 4)) {
7463 int SplatIdx = PPC::getVSPLTImmediate(SVOp, 4, DAG);
7464 SDValue Conv = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, V1);
7465 SDValue Splat = DAG.getNode(PPCISD::XXSPLT, dl, MVT::v4i32, Conv,
7466 DAG.getConstant(SplatIdx, dl, MVT::i32));
7467 return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, Splat);
7470 // Left shifts of 8 bytes are actually swaps. Convert accordingly.
7471 if (V2.isUndef() && PPC::isVSLDOIShuffleMask(SVOp, 1, DAG) == 8) {
7472 SDValue Conv = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, V1);
7473 SDValue Swap = DAG.getNode(PPCISD::SWAP_NO_CHAIN, dl, MVT::v2f64, Conv);
7474 return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, Swap);
7479 if (Subtarget.hasQPX()) {
7480 if (VT.getVectorNumElements() != 4)
7483 if (V2.isUndef()) V2 = V1;
7485 int AlignIdx = PPC::isQVALIGNIShuffleMask(SVOp);
7486 if (AlignIdx != -1) {
7487 return DAG.getNode(PPCISD::QVALIGNI, dl, VT, V1, V2,
7488 DAG.getConstant(AlignIdx, dl, MVT::i32));
7489 } else if (SVOp->isSplat()) {
7490 int SplatIdx = SVOp->getSplatIndex();
7491 if (SplatIdx >= 4) {
7496 return DAG.getNode(PPCISD::QVESPLATI, dl, VT, V1,
7497 DAG.getConstant(SplatIdx, dl, MVT::i32));
7500 // Lower this into a qvgpci/qvfperm pair.
7502 // Compute the qvgpci literal
7504 for (unsigned i = 0; i < 4; ++i) {
7505 int m = SVOp->getMaskElt(i);
7506 unsigned mm = m >= 0 ? (unsigned) m : i;
7507 idx |= mm << (3-i)*3;
7510 SDValue V3 = DAG.getNode(PPCISD::QVGPCI, dl, MVT::v4f64,
7511 DAG.getConstant(idx, dl, MVT::i32));
7512 return DAG.getNode(PPCISD::QVFPERM, dl, VT, V1, V2, V3);
7515 // Cases that are handled by instructions that take permute immediates
7516 // (such as vsplt*) should be left as VECTOR_SHUFFLE nodes so they can be
7517 // selected by the instruction selector.
7519 if (PPC::isSplatShuffleMask(SVOp, 1) ||
7520 PPC::isSplatShuffleMask(SVOp, 2) ||
7521 PPC::isSplatShuffleMask(SVOp, 4) ||
7522 PPC::isVPKUWUMShuffleMask(SVOp, 1, DAG) ||
7523 PPC::isVPKUHUMShuffleMask(SVOp, 1, DAG) ||
7524 PPC::isVSLDOIShuffleMask(SVOp, 1, DAG) != -1 ||
7525 PPC::isVMRGLShuffleMask(SVOp, 1, 1, DAG) ||
7526 PPC::isVMRGLShuffleMask(SVOp, 2, 1, DAG) ||
7527 PPC::isVMRGLShuffleMask(SVOp, 4, 1, DAG) ||
7528 PPC::isVMRGHShuffleMask(SVOp, 1, 1, DAG) ||
7529 PPC::isVMRGHShuffleMask(SVOp, 2, 1, DAG) ||
7530 PPC::isVMRGHShuffleMask(SVOp, 4, 1, DAG) ||
7531 (Subtarget.hasP8Altivec() && (
7532 PPC::isVPKUDUMShuffleMask(SVOp, 1, DAG) ||
7533 PPC::isVMRGEOShuffleMask(SVOp, true, 1, DAG) ||
7534 PPC::isVMRGEOShuffleMask(SVOp, false, 1, DAG)))) {
7539 // Altivec has a variety of "shuffle immediates" that take two vector inputs
7540 // and produce a fixed permutation. If any of these match, do not lower to
7542 unsigned int ShuffleKind = isLittleEndian ? 2 : 0;
7543 if (PPC::isVPKUWUMShuffleMask(SVOp, ShuffleKind, DAG) ||
7544 PPC::isVPKUHUMShuffleMask(SVOp, ShuffleKind, DAG) ||
7545 PPC::isVSLDOIShuffleMask(SVOp, ShuffleKind, DAG) != -1 ||
7546 PPC::isVMRGLShuffleMask(SVOp, 1, ShuffleKind, DAG) ||
7547 PPC::isVMRGLShuffleMask(SVOp, 2, ShuffleKind, DAG) ||
7548 PPC::isVMRGLShuffleMask(SVOp, 4, ShuffleKind, DAG) ||
7549 PPC::isVMRGHShuffleMask(SVOp, 1, ShuffleKind, DAG) ||
7550 PPC::isVMRGHShuffleMask(SVOp, 2, ShuffleKind, DAG) ||
7551 PPC::isVMRGHShuffleMask(SVOp, 4, ShuffleKind, DAG) ||
7552 (Subtarget.hasP8Altivec() && (
7553 PPC::isVPKUDUMShuffleMask(SVOp, ShuffleKind, DAG) ||
7554 PPC::isVMRGEOShuffleMask(SVOp, true, ShuffleKind, DAG) ||
7555 PPC::isVMRGEOShuffleMask(SVOp, false, ShuffleKind, DAG))))
7558 // Check to see if this is a shuffle of 4-byte values. If so, we can use our
7559 // perfect shuffle table to emit an optimal matching sequence.
7560 ArrayRef<int> PermMask = SVOp->getMask();
7562 unsigned PFIndexes[4];
7563 bool isFourElementShuffle = true;
7564 for (unsigned i = 0; i != 4 && isFourElementShuffle; ++i) { // Element number
7565 unsigned EltNo = 8; // Start out undef.
7566 for (unsigned j = 0; j != 4; ++j) { // Intra-element byte.
7567 if (PermMask[i*4+j] < 0)
7568 continue; // Undef, ignore it.
7570 unsigned ByteSource = PermMask[i*4+j];
7571 if ((ByteSource & 3) != j) {
7572 isFourElementShuffle = false;
7577 EltNo = ByteSource/4;
7578 } else if (EltNo != ByteSource/4) {
7579 isFourElementShuffle = false;
7583 PFIndexes[i] = EltNo;
7586 // If this shuffle can be expressed as a shuffle of 4-byte elements, use the
7587 // perfect shuffle vector to determine if it is cost effective to do this as
7588 // discrete instructions, or whether we should use a vperm.
7589 // For now, we skip this for little endian until such time as we have a
7590 // little-endian perfect shuffle table.
7591 if (isFourElementShuffle && !isLittleEndian) {
7592 // Compute the index in the perfect shuffle table.
7593 unsigned PFTableIndex =
7594 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
7596 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
7597 unsigned Cost = (PFEntry >> 30);
7599 // Determining when to avoid vperm is tricky. Many things affect the cost
7600 // of vperm, particularly how many times the perm mask needs to be computed.
7601 // For example, if the perm mask can be hoisted out of a loop or is already
7602 // used (perhaps because there are multiple permutes with the same shuffle
7603 // mask?) the vperm has a cost of 1. OTOH, hoisting the permute mask out of
7604 // the loop requires an extra register.
7606 // As a compromise, we only emit discrete instructions if the shuffle can be
7607 // generated in 3 or fewer operations. When we have loop information
7608 // available, if this block is within a loop, we should avoid using vperm
7609 // for 3-operation perms and use a constant pool load instead.
7611 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
7614 // Lower this to a VPERM(V1, V2, V3) expression, where V3 is a constant
7615 // vector that will get spilled to the constant pool.
7616 if (V2.isUndef()) V2 = V1;
7618 // The SHUFFLE_VECTOR mask is almost exactly what we want for vperm, except
7619 // that it is in input element units, not in bytes. Convert now.
7621 // For little endian, the order of the input vectors is reversed, and
7622 // the permutation mask is complemented with respect to 31. This is
7623 // necessary to produce proper semantics with the big-endian-biased vperm
7625 EVT EltVT = V1.getValueType().getVectorElementType();
7626 unsigned BytesPerElement = EltVT.getSizeInBits()/8;
7628 SmallVector<SDValue, 16> ResultMask;
7629 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i) {
7630 unsigned SrcElt = PermMask[i] < 0 ? 0 : PermMask[i];
7632 for (unsigned j = 0; j != BytesPerElement; ++j)
7634 ResultMask.push_back(DAG.getConstant(31 - (SrcElt*BytesPerElement + j),
7637 ResultMask.push_back(DAG.getConstant(SrcElt*BytesPerElement + j, dl,
7641 SDValue VPermMask = DAG.getBuildVector(MVT::v16i8, dl, ResultMask);
7643 return DAG.getNode(PPCISD::VPERM, dl, V1.getValueType(),
7646 return DAG.getNode(PPCISD::VPERM, dl, V1.getValueType(),
7650 /// getVectorCompareInfo - Given an intrinsic, return false if it is not a
7651 /// vector comparison. If it is, return true and fill in Opc/isDot with
7652 /// information about the intrinsic.
7653 static bool getVectorCompareInfo(SDValue Intrin, int &CompareOpc,
7654 bool &isDot, const PPCSubtarget &Subtarget) {
7655 unsigned IntrinsicID =
7656 cast<ConstantSDNode>(Intrin.getOperand(0))->getZExtValue();
7659 switch (IntrinsicID) {
7660 default: return false;
7661 // Comparison predicates.
7662 case Intrinsic::ppc_altivec_vcmpbfp_p: CompareOpc = 966; isDot = 1; break;
7663 case Intrinsic::ppc_altivec_vcmpeqfp_p: CompareOpc = 198; isDot = 1; break;
7664 case Intrinsic::ppc_altivec_vcmpequb_p: CompareOpc = 6; isDot = 1; break;
7665 case Intrinsic::ppc_altivec_vcmpequh_p: CompareOpc = 70; isDot = 1; break;
7666 case Intrinsic::ppc_altivec_vcmpequw_p: CompareOpc = 134; isDot = 1; break;
7667 case Intrinsic::ppc_altivec_vcmpequd_p:
7668 if (Subtarget.hasP8Altivec()) {
7675 case Intrinsic::ppc_altivec_vcmpgefp_p: CompareOpc = 454; isDot = 1; break;
7676 case Intrinsic::ppc_altivec_vcmpgtfp_p: CompareOpc = 710; isDot = 1; break;
7677 case Intrinsic::ppc_altivec_vcmpgtsb_p: CompareOpc = 774; isDot = 1; break;
7678 case Intrinsic::ppc_altivec_vcmpgtsh_p: CompareOpc = 838; isDot = 1; break;
7679 case Intrinsic::ppc_altivec_vcmpgtsw_p: CompareOpc = 902; isDot = 1; break;
7680 case Intrinsic::ppc_altivec_vcmpgtsd_p:
7681 if (Subtarget.hasP8Altivec()) {
7688 case Intrinsic::ppc_altivec_vcmpgtub_p: CompareOpc = 518; isDot = 1; break;
7689 case Intrinsic::ppc_altivec_vcmpgtuh_p: CompareOpc = 582; isDot = 1; break;
7690 case Intrinsic::ppc_altivec_vcmpgtuw_p: CompareOpc = 646; isDot = 1; break;
7691 case Intrinsic::ppc_altivec_vcmpgtud_p:
7692 if (Subtarget.hasP8Altivec()) {
7699 // VSX predicate comparisons use the same infrastructure
7700 case Intrinsic::ppc_vsx_xvcmpeqdp_p:
7701 case Intrinsic::ppc_vsx_xvcmpgedp_p:
7702 case Intrinsic::ppc_vsx_xvcmpgtdp_p:
7703 case Intrinsic::ppc_vsx_xvcmpeqsp_p:
7704 case Intrinsic::ppc_vsx_xvcmpgesp_p:
7705 case Intrinsic::ppc_vsx_xvcmpgtsp_p:
7706 if (Subtarget.hasVSX()) {
7707 switch (IntrinsicID) {
7708 case Intrinsic::ppc_vsx_xvcmpeqdp_p: CompareOpc = 99; break;
7709 case Intrinsic::ppc_vsx_xvcmpgedp_p: CompareOpc = 115; break;
7710 case Intrinsic::ppc_vsx_xvcmpgtdp_p: CompareOpc = 107; break;
7711 case Intrinsic::ppc_vsx_xvcmpeqsp_p: CompareOpc = 67; break;
7712 case Intrinsic::ppc_vsx_xvcmpgesp_p: CompareOpc = 83; break;
7713 case Intrinsic::ppc_vsx_xvcmpgtsp_p: CompareOpc = 75; break;
7722 // Normal Comparisons.
7723 case Intrinsic::ppc_altivec_vcmpbfp: CompareOpc = 966; isDot = 0; break;
7724 case Intrinsic::ppc_altivec_vcmpeqfp: CompareOpc = 198; isDot = 0; break;
7725 case Intrinsic::ppc_altivec_vcmpequb: CompareOpc = 6; isDot = 0; break;
7726 case Intrinsic::ppc_altivec_vcmpequh: CompareOpc = 70; isDot = 0; break;
7727 case Intrinsic::ppc_altivec_vcmpequw: CompareOpc = 134; isDot = 0; break;
7728 case Intrinsic::ppc_altivec_vcmpequd:
7729 if (Subtarget.hasP8Altivec()) {
7736 case Intrinsic::ppc_altivec_vcmpgefp: CompareOpc = 454; isDot = 0; break;
7737 case Intrinsic::ppc_altivec_vcmpgtfp: CompareOpc = 710; isDot = 0; break;
7738 case Intrinsic::ppc_altivec_vcmpgtsb: CompareOpc = 774; isDot = 0; break;
7739 case Intrinsic::ppc_altivec_vcmpgtsh: CompareOpc = 838; isDot = 0; break;
7740 case Intrinsic::ppc_altivec_vcmpgtsw: CompareOpc = 902; isDot = 0; break;
7741 case Intrinsic::ppc_altivec_vcmpgtsd:
7742 if (Subtarget.hasP8Altivec()) {
7749 case Intrinsic::ppc_altivec_vcmpgtub: CompareOpc = 518; isDot = 0; break;
7750 case Intrinsic::ppc_altivec_vcmpgtuh: CompareOpc = 582; isDot = 0; break;
7751 case Intrinsic::ppc_altivec_vcmpgtuw: CompareOpc = 646; isDot = 0; break;
7752 case Intrinsic::ppc_altivec_vcmpgtud:
7753 if (Subtarget.hasP8Altivec()) {
7764 /// LowerINTRINSIC_WO_CHAIN - If this is an intrinsic that we want to custom
7765 /// lower, do it, otherwise return null.
7766 SDValue PPCTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
7767 SelectionDAG &DAG) const {
7768 unsigned IntrinsicID =
7769 cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7771 if (IntrinsicID == Intrinsic::thread_pointer) {
7772 // Reads the thread pointer register, used for __builtin_thread_pointer.
7773 bool is64bit = Subtarget.isPPC64();
7774 return DAG.getRegister(is64bit ? PPC::X13 : PPC::R2,
7775 is64bit ? MVT::i64 : MVT::i32);
7778 // If this is a lowered altivec predicate compare, CompareOpc is set to the
7779 // opcode number of the comparison.
7783 if (!getVectorCompareInfo(Op, CompareOpc, isDot, Subtarget))
7784 return SDValue(); // Don't custom lower most intrinsics.
7786 // If this is a non-dot comparison, make the VCMP node and we are done.
7788 SDValue Tmp = DAG.getNode(PPCISD::VCMP, dl, Op.getOperand(2).getValueType(),
7789 Op.getOperand(1), Op.getOperand(2),
7790 DAG.getConstant(CompareOpc, dl, MVT::i32));
7791 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Tmp);
7794 // Create the PPCISD altivec 'dot' comparison node.
7796 Op.getOperand(2), // LHS
7797 Op.getOperand(3), // RHS
7798 DAG.getConstant(CompareOpc, dl, MVT::i32)
7800 EVT VTs[] = { Op.getOperand(2).getValueType(), MVT::Glue };
7801 SDValue CompNode = DAG.getNode(PPCISD::VCMPo, dl, VTs, Ops);
7803 // Now that we have the comparison, emit a copy from the CR to a GPR.
7804 // This is flagged to the above dot comparison.
7805 SDValue Flags = DAG.getNode(PPCISD::MFOCRF, dl, MVT::i32,
7806 DAG.getRegister(PPC::CR6, MVT::i32),
7807 CompNode.getValue(1));
7809 // Unpack the result based on how the target uses it.
7810 unsigned BitNo; // Bit # of CR6.
7811 bool InvertBit; // Invert result?
7812 switch (cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue()) {
7813 default: // Can't happen, don't crash on invalid number though.
7814 case 0: // Return the value of the EQ bit of CR6.
7815 BitNo = 0; InvertBit = false;
7817 case 1: // Return the inverted value of the EQ bit of CR6.
7818 BitNo = 0; InvertBit = true;
7820 case 2: // Return the value of the LT bit of CR6.
7821 BitNo = 2; InvertBit = false;
7823 case 3: // Return the inverted value of the LT bit of CR6.
7824 BitNo = 2; InvertBit = true;
7828 // Shift the bit into the low position.
7829 Flags = DAG.getNode(ISD::SRL, dl, MVT::i32, Flags,
7830 DAG.getConstant(8 - (3 - BitNo), dl, MVT::i32));
7832 Flags = DAG.getNode(ISD::AND, dl, MVT::i32, Flags,
7833 DAG.getConstant(1, dl, MVT::i32));
7835 // If we are supposed to, toggle the bit.
7837 Flags = DAG.getNode(ISD::XOR, dl, MVT::i32, Flags,
7838 DAG.getConstant(1, dl, MVT::i32));
7842 SDValue PPCTargetLowering::LowerSIGN_EXTEND_INREG(SDValue Op,
7843 SelectionDAG &DAG) const {
7845 // For v2i64 (VSX), we can pattern patch the v2i32 case (using fp <-> int
7846 // instructions), but for smaller types, we need to first extend up to v2i32
7847 // before doing going farther.
7848 if (Op.getValueType() == MVT::v2i64) {
7849 EVT ExtVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
7850 if (ExtVT != MVT::v2i32) {
7851 Op = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Op.getOperand(0));
7852 Op = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, MVT::v4i32, Op,
7853 DAG.getValueType(EVT::getVectorVT(*DAG.getContext(),
7854 ExtVT.getVectorElementType(), 4)));
7855 Op = DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, Op);
7856 Op = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, MVT::v2i64, Op,
7857 DAG.getValueType(MVT::v2i32));
7866 SDValue PPCTargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op,
7867 SelectionDAG &DAG) const {
7869 // Create a stack slot that is 16-byte aligned.
7870 MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo();
7871 int FrameIdx = MFI.CreateStackObject(16, 16, false);
7872 EVT PtrVT = getPointerTy(DAG.getDataLayout());
7873 SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
7875 // Store the input value into Value#0 of the stack slot.
7876 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0), FIdx,
7877 MachinePointerInfo());
7879 return DAG.getLoad(Op.getValueType(), dl, Store, FIdx, MachinePointerInfo());
7882 SDValue PPCTargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
7883 SelectionDAG &DAG) const {
7885 SDNode *N = Op.getNode();
7887 assert(N->getOperand(0).getValueType() == MVT::v4i1 &&
7888 "Unknown extract_vector_elt type");
7890 SDValue Value = N->getOperand(0);
7892 // The first part of this is like the store lowering except that we don't
7893 // need to track the chain.
7895 // The values are now known to be -1 (false) or 1 (true). To convert this
7896 // into 0 (false) and 1 (true), add 1 and then divide by 2 (multiply by 0.5).
7897 // This can be done with an fma and the 0.5 constant: (V+1.0)*0.5 = 0.5*V+0.5
7898 Value = DAG.getNode(PPCISD::QBFLT, dl, MVT::v4f64, Value);
7900 // FIXME: We can make this an f32 vector, but the BUILD_VECTOR code needs to
7901 // understand how to form the extending load.
7902 SDValue FPHalfs = DAG.getConstantFP(0.5, dl, MVT::v4f64);
7904 Value = DAG.getNode(ISD::FMA, dl, MVT::v4f64, Value, FPHalfs, FPHalfs);
7906 // Now convert to an integer and store.
7907 Value = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f64,
7908 DAG.getConstant(Intrinsic::ppc_qpx_qvfctiwu, dl, MVT::i32),
7911 MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo();
7912 int FrameIdx = MFI.CreateStackObject(16, 16, false);
7913 MachinePointerInfo PtrInfo =
7914 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FrameIdx);
7915 EVT PtrVT = getPointerTy(DAG.getDataLayout());
7916 SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
7918 SDValue StoreChain = DAG.getEntryNode();
7919 SDValue Ops[] = {StoreChain,
7920 DAG.getConstant(Intrinsic::ppc_qpx_qvstfiw, dl, MVT::i32),
7922 SDVTList VTs = DAG.getVTList(/*chain*/ MVT::Other);
7924 StoreChain = DAG.getMemIntrinsicNode(ISD::INTRINSIC_VOID,
7925 dl, VTs, Ops, MVT::v4i32, PtrInfo);
7927 // Extract the value requested.
7928 unsigned Offset = 4*cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
7929 SDValue Idx = DAG.getConstant(Offset, dl, FIdx.getValueType());
7930 Idx = DAG.getNode(ISD::ADD, dl, FIdx.getValueType(), FIdx, Idx);
7933 DAG.getLoad(MVT::i32, dl, StoreChain, Idx, PtrInfo.getWithOffset(Offset));
7935 if (!Subtarget.useCRBits())
7938 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, IntVal);
7941 /// Lowering for QPX v4i1 loads
7942 SDValue PPCTargetLowering::LowerVectorLoad(SDValue Op,
7943 SelectionDAG &DAG) const {
7945 LoadSDNode *LN = cast<LoadSDNode>(Op.getNode());
7946 SDValue LoadChain = LN->getChain();
7947 SDValue BasePtr = LN->getBasePtr();
7949 if (Op.getValueType() == MVT::v4f64 ||
7950 Op.getValueType() == MVT::v4f32) {
7951 EVT MemVT = LN->getMemoryVT();
7952 unsigned Alignment = LN->getAlignment();
7954 // If this load is properly aligned, then it is legal.
7955 if (Alignment >= MemVT.getStoreSize())
7958 EVT ScalarVT = Op.getValueType().getScalarType(),
7959 ScalarMemVT = MemVT.getScalarType();
7960 unsigned Stride = ScalarMemVT.getStoreSize();
7962 SDValue Vals[4], LoadChains[4];
7963 for (unsigned Idx = 0; Idx < 4; ++Idx) {
7965 if (ScalarVT != ScalarMemVT)
7966 Load = DAG.getExtLoad(LN->getExtensionType(), dl, ScalarVT, LoadChain,
7968 LN->getPointerInfo().getWithOffset(Idx * Stride),
7969 ScalarMemVT, MinAlign(Alignment, Idx * Stride),
7970 LN->getMemOperand()->getFlags(), LN->getAAInfo());
7972 Load = DAG.getLoad(ScalarVT, dl, LoadChain, BasePtr,
7973 LN->getPointerInfo().getWithOffset(Idx * Stride),
7974 MinAlign(Alignment, Idx * Stride),
7975 LN->getMemOperand()->getFlags(), LN->getAAInfo());
7977 if (Idx == 0 && LN->isIndexed()) {
7978 assert(LN->getAddressingMode() == ISD::PRE_INC &&
7979 "Unknown addressing mode on vector load");
7980 Load = DAG.getIndexedLoad(Load, dl, BasePtr, LN->getOffset(),
7981 LN->getAddressingMode());
7985 LoadChains[Idx] = Load.getValue(1);
7987 BasePtr = DAG.getNode(ISD::ADD, dl, BasePtr.getValueType(), BasePtr,
7988 DAG.getConstant(Stride, dl,
7989 BasePtr.getValueType()));
7992 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, LoadChains);
7993 SDValue Value = DAG.getBuildVector(Op.getValueType(), dl, Vals);
7995 if (LN->isIndexed()) {
7996 SDValue RetOps[] = { Value, Vals[0].getValue(1), TF };
7997 return DAG.getMergeValues(RetOps, dl);
8000 SDValue RetOps[] = { Value, TF };
8001 return DAG.getMergeValues(RetOps, dl);
8004 assert(Op.getValueType() == MVT::v4i1 && "Unknown load to lower");
8005 assert(LN->isUnindexed() && "Indexed v4i1 loads are not supported");
8007 // To lower v4i1 from a byte array, we load the byte elements of the
8008 // vector and then reuse the BUILD_VECTOR logic.
8010 SDValue VectElmts[4], VectElmtChains[4];
8011 for (unsigned i = 0; i < 4; ++i) {
8012 SDValue Idx = DAG.getConstant(i, dl, BasePtr.getValueType());
8013 Idx = DAG.getNode(ISD::ADD, dl, BasePtr.getValueType(), BasePtr, Idx);
8015 VectElmts[i] = DAG.getExtLoad(
8016 ISD::EXTLOAD, dl, MVT::i32, LoadChain, Idx,
8017 LN->getPointerInfo().getWithOffset(i), MVT::i8,
8018 /* Alignment = */ 1, LN->getMemOperand()->getFlags(), LN->getAAInfo());
8019 VectElmtChains[i] = VectElmts[i].getValue(1);
8022 LoadChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, VectElmtChains);
8023 SDValue Value = DAG.getBuildVector(MVT::v4i1, dl, VectElmts);
8025 SDValue RVals[] = { Value, LoadChain };
8026 return DAG.getMergeValues(RVals, dl);
8029 /// Lowering for QPX v4i1 stores
8030 SDValue PPCTargetLowering::LowerVectorStore(SDValue Op,
8031 SelectionDAG &DAG) const {
8033 StoreSDNode *SN = cast<StoreSDNode>(Op.getNode());
8034 SDValue StoreChain = SN->getChain();
8035 SDValue BasePtr = SN->getBasePtr();
8036 SDValue Value = SN->getValue();
8038 if (Value.getValueType() == MVT::v4f64 ||
8039 Value.getValueType() == MVT::v4f32) {
8040 EVT MemVT = SN->getMemoryVT();
8041 unsigned Alignment = SN->getAlignment();
8043 // If this store is properly aligned, then it is legal.
8044 if (Alignment >= MemVT.getStoreSize())
8047 EVT ScalarVT = Value.getValueType().getScalarType(),
8048 ScalarMemVT = MemVT.getScalarType();
8049 unsigned Stride = ScalarMemVT.getStoreSize();
8052 for (unsigned Idx = 0; Idx < 4; ++Idx) {
8053 SDValue Ex = DAG.getNode(
8054 ISD::EXTRACT_VECTOR_ELT, dl, ScalarVT, Value,
8055 DAG.getConstant(Idx, dl, getVectorIdxTy(DAG.getDataLayout())));
8057 if (ScalarVT != ScalarMemVT)
8059 DAG.getTruncStore(StoreChain, dl, Ex, BasePtr,
8060 SN->getPointerInfo().getWithOffset(Idx * Stride),
8061 ScalarMemVT, MinAlign(Alignment, Idx * Stride),
8062 SN->getMemOperand()->getFlags(), SN->getAAInfo());
8064 Store = DAG.getStore(StoreChain, dl, Ex, BasePtr,
8065 SN->getPointerInfo().getWithOffset(Idx * Stride),
8066 MinAlign(Alignment, Idx * Stride),
8067 SN->getMemOperand()->getFlags(), SN->getAAInfo());
8069 if (Idx == 0 && SN->isIndexed()) {
8070 assert(SN->getAddressingMode() == ISD::PRE_INC &&
8071 "Unknown addressing mode on vector store");
8072 Store = DAG.getIndexedStore(Store, dl, BasePtr, SN->getOffset(),
8073 SN->getAddressingMode());
8076 BasePtr = DAG.getNode(ISD::ADD, dl, BasePtr.getValueType(), BasePtr,
8077 DAG.getConstant(Stride, dl,
8078 BasePtr.getValueType()));
8079 Stores[Idx] = Store;
8082 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Stores);
8084 if (SN->isIndexed()) {
8085 SDValue RetOps[] = { TF, Stores[0].getValue(1) };
8086 return DAG.getMergeValues(RetOps, dl);
8092 assert(SN->isUnindexed() && "Indexed v4i1 stores are not supported");
8093 assert(Value.getValueType() == MVT::v4i1 && "Unknown store to lower");
8095 // The values are now known to be -1 (false) or 1 (true). To convert this
8096 // into 0 (false) and 1 (true), add 1 and then divide by 2 (multiply by 0.5).
8097 // This can be done with an fma and the 0.5 constant: (V+1.0)*0.5 = 0.5*V+0.5
8098 Value = DAG.getNode(PPCISD::QBFLT, dl, MVT::v4f64, Value);
8100 // FIXME: We can make this an f32 vector, but the BUILD_VECTOR code needs to
8101 // understand how to form the extending load.
8102 SDValue FPHalfs = DAG.getConstantFP(0.5, dl, MVT::v4f64);
8104 Value = DAG.getNode(ISD::FMA, dl, MVT::v4f64, Value, FPHalfs, FPHalfs);
8106 // Now convert to an integer and store.
8107 Value = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f64,
8108 DAG.getConstant(Intrinsic::ppc_qpx_qvfctiwu, dl, MVT::i32),
8111 MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo();
8112 int FrameIdx = MFI.CreateStackObject(16, 16, false);
8113 MachinePointerInfo PtrInfo =
8114 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FrameIdx);
8115 EVT PtrVT = getPointerTy(DAG.getDataLayout());
8116 SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
8118 SDValue Ops[] = {StoreChain,
8119 DAG.getConstant(Intrinsic::ppc_qpx_qvstfiw, dl, MVT::i32),
8121 SDVTList VTs = DAG.getVTList(/*chain*/ MVT::Other);
8123 StoreChain = DAG.getMemIntrinsicNode(ISD::INTRINSIC_VOID,
8124 dl, VTs, Ops, MVT::v4i32, PtrInfo);
8126 // Move data into the byte array.
8127 SDValue Loads[4], LoadChains[4];
8128 for (unsigned i = 0; i < 4; ++i) {
8129 unsigned Offset = 4*i;
8130 SDValue Idx = DAG.getConstant(Offset, dl, FIdx.getValueType());
8131 Idx = DAG.getNode(ISD::ADD, dl, FIdx.getValueType(), FIdx, Idx);
8133 Loads[i] = DAG.getLoad(MVT::i32, dl, StoreChain, Idx,
8134 PtrInfo.getWithOffset(Offset));
8135 LoadChains[i] = Loads[i].getValue(1);
8138 StoreChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, LoadChains);
8141 for (unsigned i = 0; i < 4; ++i) {
8142 SDValue Idx = DAG.getConstant(i, dl, BasePtr.getValueType());
8143 Idx = DAG.getNode(ISD::ADD, dl, BasePtr.getValueType(), BasePtr, Idx);
8145 Stores[i] = DAG.getTruncStore(
8146 StoreChain, dl, Loads[i], Idx, SN->getPointerInfo().getWithOffset(i),
8147 MVT::i8, /* Alignment = */ 1, SN->getMemOperand()->getFlags(),
8151 StoreChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Stores);
8156 SDValue PPCTargetLowering::LowerMUL(SDValue Op, SelectionDAG &DAG) const {
8158 if (Op.getValueType() == MVT::v4i32) {
8159 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
8161 SDValue Zero = BuildSplatI( 0, 1, MVT::v4i32, DAG, dl);
8162 SDValue Neg16 = BuildSplatI(-16, 4, MVT::v4i32, DAG, dl);//+16 as shift amt.
8164 SDValue RHSSwap = // = vrlw RHS, 16
8165 BuildIntrinsicOp(Intrinsic::ppc_altivec_vrlw, RHS, Neg16, DAG, dl);
8167 // Shrinkify inputs to v8i16.
8168 LHS = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, LHS);
8169 RHS = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, RHS);
8170 RHSSwap = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, RHSSwap);
8172 // Low parts multiplied together, generating 32-bit results (we ignore the
8174 SDValue LoProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmulouh,
8175 LHS, RHS, DAG, dl, MVT::v4i32);
8177 SDValue HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmsumuhm,
8178 LHS, RHSSwap, Zero, DAG, dl, MVT::v4i32);
8179 // Shift the high parts up 16 bits.
8180 HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, HiProd,
8182 return DAG.getNode(ISD::ADD, dl, MVT::v4i32, LoProd, HiProd);
8183 } else if (Op.getValueType() == MVT::v8i16) {
8184 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
8186 SDValue Zero = BuildSplatI(0, 1, MVT::v8i16, DAG, dl);
8188 return BuildIntrinsicOp(Intrinsic::ppc_altivec_vmladduhm,
8189 LHS, RHS, Zero, DAG, dl);
8190 } else if (Op.getValueType() == MVT::v16i8) {
8191 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
8192 bool isLittleEndian = Subtarget.isLittleEndian();
8194 // Multiply the even 8-bit parts, producing 16-bit sums.
8195 SDValue EvenParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuleub,
8196 LHS, RHS, DAG, dl, MVT::v8i16);
8197 EvenParts = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, EvenParts);
8199 // Multiply the odd 8-bit parts, producing 16-bit sums.
8200 SDValue OddParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuloub,
8201 LHS, RHS, DAG, dl, MVT::v8i16);
8202 OddParts = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OddParts);
8204 // Merge the results together. Because vmuleub and vmuloub are
8205 // instructions with a big-endian bias, we must reverse the
8206 // element numbering and reverse the meaning of "odd" and "even"
8207 // when generating little endian code.
8209 for (unsigned i = 0; i != 8; ++i) {
8210 if (isLittleEndian) {
8212 Ops[i*2+1] = 2*i+16;
8215 Ops[i*2+1] = 2*i+1+16;
8219 return DAG.getVectorShuffle(MVT::v16i8, dl, OddParts, EvenParts, Ops);
8221 return DAG.getVectorShuffle(MVT::v16i8, dl, EvenParts, OddParts, Ops);
8223 llvm_unreachable("Unknown mul to lower!");
8227 /// LowerOperation - Provide custom lowering hooks for some operations.
8229 SDValue PPCTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
8230 switch (Op.getOpcode()) {
8231 default: llvm_unreachable("Wasn't expecting to be able to lower this!");
8232 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
8233 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
8234 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
8235 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
8236 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
8237 case ISD::SETCC: return LowerSETCC(Op, DAG);
8238 case ISD::INIT_TRAMPOLINE: return LowerINIT_TRAMPOLINE(Op, DAG);
8239 case ISD::ADJUST_TRAMPOLINE: return LowerADJUST_TRAMPOLINE(Op, DAG);
8241 return LowerVASTART(Op, DAG);
8244 return LowerVAARG(Op, DAG);
8247 return LowerVACOPY(Op, DAG);
8249 case ISD::STACKRESTORE:
8250 return LowerSTACKRESTORE(Op, DAG);
8252 case ISD::DYNAMIC_STACKALLOC:
8253 return LowerDYNAMIC_STACKALLOC(Op, DAG);
8255 case ISD::GET_DYNAMIC_AREA_OFFSET:
8256 return LowerGET_DYNAMIC_AREA_OFFSET(Op, DAG);
8258 case ISD::EH_DWARF_CFA:
8259 return LowerEH_DWARF_CFA(Op, DAG);
8261 case ISD::EH_SJLJ_SETJMP: return lowerEH_SJLJ_SETJMP(Op, DAG);
8262 case ISD::EH_SJLJ_LONGJMP: return lowerEH_SJLJ_LONGJMP(Op, DAG);
8264 case ISD::LOAD: return LowerLOAD(Op, DAG);
8265 case ISD::STORE: return LowerSTORE(Op, DAG);
8266 case ISD::TRUNCATE: return LowerTRUNCATE(Op, DAG);
8267 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
8268 case ISD::FP_TO_UINT:
8269 case ISD::FP_TO_SINT: return LowerFP_TO_INT(Op, DAG,
8271 case ISD::UINT_TO_FP:
8272 case ISD::SINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
8273 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
8275 // Lower 64-bit shifts.
8276 case ISD::SHL_PARTS: return LowerSHL_PARTS(Op, DAG);
8277 case ISD::SRL_PARTS: return LowerSRL_PARTS(Op, DAG);
8278 case ISD::SRA_PARTS: return LowerSRA_PARTS(Op, DAG);
8280 // Vector-related lowering.
8281 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
8282 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
8283 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
8284 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
8285 case ISD::SIGN_EXTEND_INREG: return LowerSIGN_EXTEND_INREG(Op, DAG);
8286 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
8287 case ISD::MUL: return LowerMUL(Op, DAG);
8289 // For counter-based loop handling.
8290 case ISD::INTRINSIC_W_CHAIN: return SDValue();
8292 // Frame & Return address.
8293 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
8294 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
8298 void PPCTargetLowering::ReplaceNodeResults(SDNode *N,
8299 SmallVectorImpl<SDValue>&Results,
8300 SelectionDAG &DAG) const {
8302 switch (N->getOpcode()) {
8304 llvm_unreachable("Do not know how to custom type legalize this operation!");
8305 case ISD::READCYCLECOUNTER: {
8306 SDVTList VTs = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
8307 SDValue RTB = DAG.getNode(PPCISD::READ_TIME_BASE, dl, VTs, N->getOperand(0));
8309 Results.push_back(RTB);
8310 Results.push_back(RTB.getValue(1));
8311 Results.push_back(RTB.getValue(2));
8314 case ISD::INTRINSIC_W_CHAIN: {
8315 if (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() !=
8316 Intrinsic::ppc_is_decremented_ctr_nonzero)
8319 assert(N->getValueType(0) == MVT::i1 &&
8320 "Unexpected result type for CTR decrement intrinsic");
8321 EVT SVT = getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(),
8322 N->getValueType(0));
8323 SDVTList VTs = DAG.getVTList(SVT, MVT::Other);
8324 SDValue NewInt = DAG.getNode(N->getOpcode(), dl, VTs, N->getOperand(0),
8327 Results.push_back(NewInt);
8328 Results.push_back(NewInt.getValue(1));
8332 if (!Subtarget.isSVR4ABI() || Subtarget.isPPC64())
8335 EVT VT = N->getValueType(0);
8337 if (VT == MVT::i64) {
8338 SDValue NewNode = LowerVAARG(SDValue(N, 1), DAG);
8340 Results.push_back(NewNode);
8341 Results.push_back(NewNode.getValue(1));
8345 case ISD::FP_ROUND_INREG: {
8346 assert(N->getValueType(0) == MVT::ppcf128);
8347 assert(N->getOperand(0).getValueType() == MVT::ppcf128);
8348 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl,
8349 MVT::f64, N->getOperand(0),
8350 DAG.getIntPtrConstant(0, dl));
8351 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl,
8352 MVT::f64, N->getOperand(0),
8353 DAG.getIntPtrConstant(1, dl));
8355 // Add the two halves of the long double in round-to-zero mode.
8356 SDValue FPreg = DAG.getNode(PPCISD::FADDRTZ, dl, MVT::f64, Lo, Hi);
8358 // We know the low half is about to be thrown away, so just use something
8360 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::ppcf128,
8364 case ISD::FP_TO_SINT:
8365 case ISD::FP_TO_UINT:
8366 // LowerFP_TO_INT() can only handle f32 and f64.
8367 if (N->getOperand(0).getValueType() == MVT::ppcf128)
8369 Results.push_back(LowerFP_TO_INT(SDValue(N, 0), DAG, dl));
8374 //===----------------------------------------------------------------------===//
8375 // Other Lowering Code
8376 //===----------------------------------------------------------------------===//
8378 static Instruction* callIntrinsic(IRBuilder<> &Builder, Intrinsic::ID Id) {
8379 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
8380 Function *Func = Intrinsic::getDeclaration(M, Id);
8381 return Builder.CreateCall(Func, {});
8384 // The mappings for emitLeading/TrailingFence is taken from
8385 // http://www.cl.cam.ac.uk/~pes20/cpp/cpp0xmappings.html
8386 Instruction* PPCTargetLowering::emitLeadingFence(IRBuilder<> &Builder,
8387 AtomicOrdering Ord, bool IsStore,
8388 bool IsLoad) const {
8389 if (Ord == AtomicOrdering::SequentiallyConsistent)
8390 return callIntrinsic(Builder, Intrinsic::ppc_sync);
8391 if (isReleaseOrStronger(Ord))
8392 return callIntrinsic(Builder, Intrinsic::ppc_lwsync);
8396 Instruction* PPCTargetLowering::emitTrailingFence(IRBuilder<> &Builder,
8397 AtomicOrdering Ord, bool IsStore,
8398 bool IsLoad) const {
8399 if (IsLoad && isAcquireOrStronger(Ord))
8400 return callIntrinsic(Builder, Intrinsic::ppc_lwsync);
8401 // FIXME: this is too conservative, a dependent branch + isync is enough.
8402 // See http://www.cl.cam.ac.uk/~pes20/cpp/cpp0xmappings.html and
8403 // http://www.rdrop.com/users/paulmck/scalability/paper/N2745r.2011.03.04a.html
8404 // and http://www.cl.cam.ac.uk/~pes20/cppppc/ for justification.
8409 PPCTargetLowering::EmitAtomicBinary(MachineInstr &MI, MachineBasicBlock *BB,
8410 unsigned AtomicSize,
8413 unsigned CmpPred) const {
8414 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
8415 const TargetInstrInfo *TII = Subtarget.getInstrInfo();
8417 auto LoadMnemonic = PPC::LDARX;
8418 auto StoreMnemonic = PPC::STDCX;
8419 switch (AtomicSize) {
8421 llvm_unreachable("Unexpected size of atomic entity");
8423 LoadMnemonic = PPC::LBARX;
8424 StoreMnemonic = PPC::STBCX;
8425 assert(Subtarget.hasPartwordAtomics() && "Call this only with size >=4");
8428 LoadMnemonic = PPC::LHARX;
8429 StoreMnemonic = PPC::STHCX;
8430 assert(Subtarget.hasPartwordAtomics() && "Call this only with size >=4");
8433 LoadMnemonic = PPC::LWARX;
8434 StoreMnemonic = PPC::STWCX;
8437 LoadMnemonic = PPC::LDARX;
8438 StoreMnemonic = PPC::STDCX;
8442 const BasicBlock *LLVM_BB = BB->getBasicBlock();
8443 MachineFunction *F = BB->getParent();
8444 MachineFunction::iterator It = ++BB->getIterator();
8446 unsigned dest = MI.getOperand(0).getReg();
8447 unsigned ptrA = MI.getOperand(1).getReg();
8448 unsigned ptrB = MI.getOperand(2).getReg();
8449 unsigned incr = MI.getOperand(3).getReg();
8450 DebugLoc dl = MI.getDebugLoc();
8452 MachineBasicBlock *loopMBB = F->CreateMachineBasicBlock(LLVM_BB);
8453 MachineBasicBlock *loop2MBB =
8454 CmpOpcode ? F->CreateMachineBasicBlock(LLVM_BB) : nullptr;
8455 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
8456 F->insert(It, loopMBB);
8458 F->insert(It, loop2MBB);
8459 F->insert(It, exitMBB);
8460 exitMBB->splice(exitMBB->begin(), BB,
8461 std::next(MachineBasicBlock::iterator(MI)), BB->end());
8462 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
8464 MachineRegisterInfo &RegInfo = F->getRegInfo();
8465 unsigned TmpReg = (!BinOpcode) ? incr :
8466 RegInfo.createVirtualRegister( AtomicSize == 8 ? &PPC::G8RCRegClass
8467 : &PPC::GPRCRegClass);
8471 // fallthrough --> loopMBB
8472 BB->addSuccessor(loopMBB);
8475 // l[wd]arx dest, ptr
8476 // add r0, dest, incr
8477 // st[wd]cx. r0, ptr
8479 // fallthrough --> exitMBB
8483 // l[wd]arx dest, ptr
8484 // cmpl?[wd] incr, dest
8487 // st[wd]cx. dest, ptr
8489 // fallthrough --> exitMBB
8492 BuildMI(BB, dl, TII->get(LoadMnemonic), dest)
8493 .addReg(ptrA).addReg(ptrB);
8495 BuildMI(BB, dl, TII->get(BinOpcode), TmpReg).addReg(incr).addReg(dest);
8497 BuildMI(BB, dl, TII->get(CmpOpcode), PPC::CR0)
8498 .addReg(incr).addReg(dest);
8499 BuildMI(BB, dl, TII->get(PPC::BCC))
8500 .addImm(CmpPred).addReg(PPC::CR0).addMBB(exitMBB);
8501 BB->addSuccessor(loop2MBB);
8502 BB->addSuccessor(exitMBB);
8505 BuildMI(BB, dl, TII->get(StoreMnemonic))
8506 .addReg(TmpReg).addReg(ptrA).addReg(ptrB);
8507 BuildMI(BB, dl, TII->get(PPC::BCC))
8508 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loopMBB);
8509 BB->addSuccessor(loopMBB);
8510 BB->addSuccessor(exitMBB);
8519 PPCTargetLowering::EmitPartwordAtomicBinary(MachineInstr &MI,
8520 MachineBasicBlock *BB,
8521 bool is8bit, // operation
8524 unsigned CmpPred) const {
8525 // If we support part-word atomic mnemonics, just use them
8526 if (Subtarget.hasPartwordAtomics())
8527 return EmitAtomicBinary(MI, BB, is8bit ? 1 : 2, BinOpcode,
8528 CmpOpcode, CmpPred);
8530 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
8531 const TargetInstrInfo *TII = Subtarget.getInstrInfo();
8532 // In 64 bit mode we have to use 64 bits for addresses, even though the
8533 // lwarx/stwcx are 32 bits. With the 32-bit atomics we can use address
8534 // registers without caring whether they're 32 or 64, but here we're
8535 // doing actual arithmetic on the addresses.
8536 bool is64bit = Subtarget.isPPC64();
8537 bool isLittleEndian = Subtarget.isLittleEndian();
8538 unsigned ZeroReg = is64bit ? PPC::ZERO8 : PPC::ZERO;
8540 const BasicBlock *LLVM_BB = BB->getBasicBlock();
8541 MachineFunction *F = BB->getParent();
8542 MachineFunction::iterator It = ++BB->getIterator();
8544 unsigned dest = MI.getOperand(0).getReg();
8545 unsigned ptrA = MI.getOperand(1).getReg();
8546 unsigned ptrB = MI.getOperand(2).getReg();
8547 unsigned incr = MI.getOperand(3).getReg();
8548 DebugLoc dl = MI.getDebugLoc();
8550 MachineBasicBlock *loopMBB = F->CreateMachineBasicBlock(LLVM_BB);
8551 MachineBasicBlock *loop2MBB =
8552 CmpOpcode ? F->CreateMachineBasicBlock(LLVM_BB) : nullptr;
8553 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
8554 F->insert(It, loopMBB);
8556 F->insert(It, loop2MBB);
8557 F->insert(It, exitMBB);
8558 exitMBB->splice(exitMBB->begin(), BB,
8559 std::next(MachineBasicBlock::iterator(MI)), BB->end());
8560 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
8562 MachineRegisterInfo &RegInfo = F->getRegInfo();
8563 const TargetRegisterClass *RC = is64bit ? &PPC::G8RCRegClass
8564 : &PPC::GPRCRegClass;
8565 unsigned PtrReg = RegInfo.createVirtualRegister(RC);
8566 unsigned Shift1Reg = RegInfo.createVirtualRegister(RC);
8568 isLittleEndian ? Shift1Reg : RegInfo.createVirtualRegister(RC);
8569 unsigned Incr2Reg = RegInfo.createVirtualRegister(RC);
8570 unsigned MaskReg = RegInfo.createVirtualRegister(RC);
8571 unsigned Mask2Reg = RegInfo.createVirtualRegister(RC);
8572 unsigned Mask3Reg = RegInfo.createVirtualRegister(RC);
8573 unsigned Tmp2Reg = RegInfo.createVirtualRegister(RC);
8574 unsigned Tmp3Reg = RegInfo.createVirtualRegister(RC);
8575 unsigned Tmp4Reg = RegInfo.createVirtualRegister(RC);
8576 unsigned TmpDestReg = RegInfo.createVirtualRegister(RC);
8578 unsigned TmpReg = (!BinOpcode) ? Incr2Reg : RegInfo.createVirtualRegister(RC);
8582 // fallthrough --> loopMBB
8583 BB->addSuccessor(loopMBB);
8585 // The 4-byte load must be aligned, while a char or short may be
8586 // anywhere in the word. Hence all this nasty bookkeeping code.
8587 // add ptr1, ptrA, ptrB [copy if ptrA==0]
8588 // rlwinm shift1, ptr1, 3, 27, 28 [3, 27, 27]
8589 // xori shift, shift1, 24 [16]
8590 // rlwinm ptr, ptr1, 0, 0, 29
8591 // slw incr2, incr, shift
8592 // li mask2, 255 [li mask3, 0; ori mask2, mask3, 65535]
8593 // slw mask, mask2, shift
8595 // lwarx tmpDest, ptr
8596 // add tmp, tmpDest, incr2
8597 // andc tmp2, tmpDest, mask
8598 // and tmp3, tmp, mask
8599 // or tmp4, tmp3, tmp2
8602 // fallthrough --> exitMBB
8603 // srw dest, tmpDest, shift
8604 if (ptrA != ZeroReg) {
8605 Ptr1Reg = RegInfo.createVirtualRegister(RC);
8606 BuildMI(BB, dl, TII->get(is64bit ? PPC::ADD8 : PPC::ADD4), Ptr1Reg)
8607 .addReg(ptrA).addReg(ptrB);
8611 BuildMI(BB, dl, TII->get(PPC::RLWINM), Shift1Reg).addReg(Ptr1Reg)
8612 .addImm(3).addImm(27).addImm(is8bit ? 28 : 27);
8613 if (!isLittleEndian)
8614 BuildMI(BB, dl, TII->get(is64bit ? PPC::XORI8 : PPC::XORI), ShiftReg)
8615 .addReg(Shift1Reg).addImm(is8bit ? 24 : 16);
8617 BuildMI(BB, dl, TII->get(PPC::RLDICR), PtrReg)
8618 .addReg(Ptr1Reg).addImm(0).addImm(61);
8620 BuildMI(BB, dl, TII->get(PPC::RLWINM), PtrReg)
8621 .addReg(Ptr1Reg).addImm(0).addImm(0).addImm(29);
8622 BuildMI(BB, dl, TII->get(PPC::SLW), Incr2Reg)
8623 .addReg(incr).addReg(ShiftReg);
8625 BuildMI(BB, dl, TII->get(PPC::LI), Mask2Reg).addImm(255);
8627 BuildMI(BB, dl, TII->get(PPC::LI), Mask3Reg).addImm(0);
8628 BuildMI(BB, dl, TII->get(PPC::ORI),Mask2Reg).addReg(Mask3Reg).addImm(65535);
8630 BuildMI(BB, dl, TII->get(PPC::SLW), MaskReg)
8631 .addReg(Mask2Reg).addReg(ShiftReg);
8634 BuildMI(BB, dl, TII->get(PPC::LWARX), TmpDestReg)
8635 .addReg(ZeroReg).addReg(PtrReg);
8637 BuildMI(BB, dl, TII->get(BinOpcode), TmpReg)
8638 .addReg(Incr2Reg).addReg(TmpDestReg);
8639 BuildMI(BB, dl, TII->get(is64bit ? PPC::ANDC8 : PPC::ANDC), Tmp2Reg)
8640 .addReg(TmpDestReg).addReg(MaskReg);
8641 BuildMI(BB, dl, TII->get(is64bit ? PPC::AND8 : PPC::AND), Tmp3Reg)
8642 .addReg(TmpReg).addReg(MaskReg);
8644 // For unsigned comparisons, we can directly compare the shifted values.
8645 // For signed comparisons we shift and sign extend.
8646 unsigned SReg = RegInfo.createVirtualRegister(RC);
8647 BuildMI(BB, dl, TII->get(is64bit ? PPC::AND8 : PPC::AND), SReg)
8648 .addReg(TmpDestReg).addReg(MaskReg);
8649 unsigned ValueReg = SReg;
8650 unsigned CmpReg = Incr2Reg;
8651 if (CmpOpcode == PPC::CMPW) {
8652 ValueReg = RegInfo.createVirtualRegister(RC);
8653 BuildMI(BB, dl, TII->get(PPC::SRW), ValueReg)
8654 .addReg(SReg).addReg(ShiftReg);
8655 unsigned ValueSReg = RegInfo.createVirtualRegister(RC);
8656 BuildMI(BB, dl, TII->get(is8bit ? PPC::EXTSB : PPC::EXTSH), ValueSReg)
8658 ValueReg = ValueSReg;
8661 BuildMI(BB, dl, TII->get(CmpOpcode), PPC::CR0)
8662 .addReg(CmpReg).addReg(ValueReg);
8663 BuildMI(BB, dl, TII->get(PPC::BCC))
8664 .addImm(CmpPred).addReg(PPC::CR0).addMBB(exitMBB);
8665 BB->addSuccessor(loop2MBB);
8666 BB->addSuccessor(exitMBB);
8669 BuildMI(BB, dl, TII->get(is64bit ? PPC::OR8 : PPC::OR), Tmp4Reg)
8670 .addReg(Tmp3Reg).addReg(Tmp2Reg);
8671 BuildMI(BB, dl, TII->get(PPC::STWCX))
8672 .addReg(Tmp4Reg).addReg(ZeroReg).addReg(PtrReg);
8673 BuildMI(BB, dl, TII->get(PPC::BCC))
8674 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loopMBB);
8675 BB->addSuccessor(loopMBB);
8676 BB->addSuccessor(exitMBB);
8681 BuildMI(*BB, BB->begin(), dl, TII->get(PPC::SRW), dest).addReg(TmpDestReg)
8686 llvm::MachineBasicBlock *
8687 PPCTargetLowering::emitEHSjLjSetJmp(MachineInstr &MI,
8688 MachineBasicBlock *MBB) const {
8689 DebugLoc DL = MI.getDebugLoc();
8690 const TargetInstrInfo *TII = Subtarget.getInstrInfo();
8692 MachineFunction *MF = MBB->getParent();
8693 MachineRegisterInfo &MRI = MF->getRegInfo();
8695 const BasicBlock *BB = MBB->getBasicBlock();
8696 MachineFunction::iterator I = ++MBB->getIterator();
8699 MachineInstr::mmo_iterator MMOBegin = MI.memoperands_begin();
8700 MachineInstr::mmo_iterator MMOEnd = MI.memoperands_end();
8702 unsigned DstReg = MI.getOperand(0).getReg();
8703 const TargetRegisterClass *RC = MRI.getRegClass(DstReg);
8704 assert(RC->hasType(MVT::i32) && "Invalid destination!");
8705 unsigned mainDstReg = MRI.createVirtualRegister(RC);
8706 unsigned restoreDstReg = MRI.createVirtualRegister(RC);
8708 MVT PVT = getPointerTy(MF->getDataLayout());
8709 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
8710 "Invalid Pointer Size!");
8711 // For v = setjmp(buf), we generate
8714 // SjLjSetup mainMBB
8720 // buf[LabelOffset] = LR
8724 // v = phi(main, restore)
8727 MachineBasicBlock *thisMBB = MBB;
8728 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
8729 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
8730 MF->insert(I, mainMBB);
8731 MF->insert(I, sinkMBB);
8733 MachineInstrBuilder MIB;
8735 // Transfer the remainder of BB and its successor edges to sinkMBB.
8736 sinkMBB->splice(sinkMBB->begin(), MBB,
8737 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
8738 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
8740 // Note that the structure of the jmp_buf used here is not compatible
8741 // with that used by libc, and is not designed to be. Specifically, it
8742 // stores only those 'reserved' registers that LLVM does not otherwise
8743 // understand how to spill. Also, by convention, by the time this
8744 // intrinsic is called, Clang has already stored the frame address in the
8745 // first slot of the buffer and stack address in the third. Following the
8746 // X86 target code, we'll store the jump address in the second slot. We also
8747 // need to save the TOC pointer (R2) to handle jumps between shared
8748 // libraries, and that will be stored in the fourth slot. The thread
8749 // identifier (R13) is not affected.
8752 const int64_t LabelOffset = 1 * PVT.getStoreSize();
8753 const int64_t TOCOffset = 3 * PVT.getStoreSize();
8754 const int64_t BPOffset = 4 * PVT.getStoreSize();
8756 // Prepare IP either in reg.
8757 const TargetRegisterClass *PtrRC = getRegClassFor(PVT);
8758 unsigned LabelReg = MRI.createVirtualRegister(PtrRC);
8759 unsigned BufReg = MI.getOperand(1).getReg();
8761 if (Subtarget.isPPC64() && Subtarget.isSVR4ABI()) {
8762 setUsesTOCBasePtr(*MBB->getParent());
8763 MIB = BuildMI(*thisMBB, MI, DL, TII->get(PPC::STD))
8767 MIB.setMemRefs(MMOBegin, MMOEnd);
8770 // Naked functions never have a base pointer, and so we use r1. For all
8771 // other functions, this decision must be delayed until during PEI.
8773 if (MF->getFunction()->hasFnAttribute(Attribute::Naked))
8774 BaseReg = Subtarget.isPPC64() ? PPC::X1 : PPC::R1;
8776 BaseReg = Subtarget.isPPC64() ? PPC::BP8 : PPC::BP;
8778 MIB = BuildMI(*thisMBB, MI, DL,
8779 TII->get(Subtarget.isPPC64() ? PPC::STD : PPC::STW))
8783 MIB.setMemRefs(MMOBegin, MMOEnd);
8786 MIB = BuildMI(*thisMBB, MI, DL, TII->get(PPC::BCLalways)).addMBB(mainMBB);
8787 const PPCRegisterInfo *TRI = Subtarget.getRegisterInfo();
8788 MIB.addRegMask(TRI->getNoPreservedMask());
8790 BuildMI(*thisMBB, MI, DL, TII->get(PPC::LI), restoreDstReg).addImm(1);
8792 MIB = BuildMI(*thisMBB, MI, DL, TII->get(PPC::EH_SjLj_Setup))
8794 MIB = BuildMI(*thisMBB, MI, DL, TII->get(PPC::B)).addMBB(sinkMBB);
8796 thisMBB->addSuccessor(mainMBB, BranchProbability::getZero());
8797 thisMBB->addSuccessor(sinkMBB, BranchProbability::getOne());
8802 BuildMI(mainMBB, DL,
8803 TII->get(Subtarget.isPPC64() ? PPC::MFLR8 : PPC::MFLR), LabelReg);
8806 if (Subtarget.isPPC64()) {
8807 MIB = BuildMI(mainMBB, DL, TII->get(PPC::STD))
8809 .addImm(LabelOffset)
8812 MIB = BuildMI(mainMBB, DL, TII->get(PPC::STW))
8814 .addImm(LabelOffset)
8818 MIB.setMemRefs(MMOBegin, MMOEnd);
8820 BuildMI(mainMBB, DL, TII->get(PPC::LI), mainDstReg).addImm(0);
8821 mainMBB->addSuccessor(sinkMBB);
8824 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
8825 TII->get(PPC::PHI), DstReg)
8826 .addReg(mainDstReg).addMBB(mainMBB)
8827 .addReg(restoreDstReg).addMBB(thisMBB);
8829 MI.eraseFromParent();
8834 PPCTargetLowering::emitEHSjLjLongJmp(MachineInstr &MI,
8835 MachineBasicBlock *MBB) const {
8836 DebugLoc DL = MI.getDebugLoc();
8837 const TargetInstrInfo *TII = Subtarget.getInstrInfo();
8839 MachineFunction *MF = MBB->getParent();
8840 MachineRegisterInfo &MRI = MF->getRegInfo();
8843 MachineInstr::mmo_iterator MMOBegin = MI.memoperands_begin();
8844 MachineInstr::mmo_iterator MMOEnd = MI.memoperands_end();
8846 MVT PVT = getPointerTy(MF->getDataLayout());
8847 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
8848 "Invalid Pointer Size!");
8850 const TargetRegisterClass *RC =
8851 (PVT == MVT::i64) ? &PPC::G8RCRegClass : &PPC::GPRCRegClass;
8852 unsigned Tmp = MRI.createVirtualRegister(RC);
8853 // Since FP is only updated here but NOT referenced, it's treated as GPR.
8854 unsigned FP = (PVT == MVT::i64) ? PPC::X31 : PPC::R31;
8855 unsigned SP = (PVT == MVT::i64) ? PPC::X1 : PPC::R1;
8859 : (Subtarget.isSVR4ABI() && isPositionIndependent() ? PPC::R29
8862 MachineInstrBuilder MIB;
8864 const int64_t LabelOffset = 1 * PVT.getStoreSize();
8865 const int64_t SPOffset = 2 * PVT.getStoreSize();
8866 const int64_t TOCOffset = 3 * PVT.getStoreSize();
8867 const int64_t BPOffset = 4 * PVT.getStoreSize();
8869 unsigned BufReg = MI.getOperand(0).getReg();
8871 // Reload FP (the jumped-to function may not have had a
8872 // frame pointer, and if so, then its r31 will be restored
8874 if (PVT == MVT::i64) {
8875 MIB = BuildMI(*MBB, MI, DL, TII->get(PPC::LD), FP)
8879 MIB = BuildMI(*MBB, MI, DL, TII->get(PPC::LWZ), FP)
8883 MIB.setMemRefs(MMOBegin, MMOEnd);
8886 if (PVT == MVT::i64) {
8887 MIB = BuildMI(*MBB, MI, DL, TII->get(PPC::LD), Tmp)
8888 .addImm(LabelOffset)
8891 MIB = BuildMI(*MBB, MI, DL, TII->get(PPC::LWZ), Tmp)
8892 .addImm(LabelOffset)
8895 MIB.setMemRefs(MMOBegin, MMOEnd);
8898 if (PVT == MVT::i64) {
8899 MIB = BuildMI(*MBB, MI, DL, TII->get(PPC::LD), SP)
8903 MIB = BuildMI(*MBB, MI, DL, TII->get(PPC::LWZ), SP)
8907 MIB.setMemRefs(MMOBegin, MMOEnd);
8910 if (PVT == MVT::i64) {
8911 MIB = BuildMI(*MBB, MI, DL, TII->get(PPC::LD), BP)
8915 MIB = BuildMI(*MBB, MI, DL, TII->get(PPC::LWZ), BP)
8919 MIB.setMemRefs(MMOBegin, MMOEnd);
8922 if (PVT == MVT::i64 && Subtarget.isSVR4ABI()) {
8923 setUsesTOCBasePtr(*MBB->getParent());
8924 MIB = BuildMI(*MBB, MI, DL, TII->get(PPC::LD), PPC::X2)
8928 MIB.setMemRefs(MMOBegin, MMOEnd);
8932 BuildMI(*MBB, MI, DL,
8933 TII->get(PVT == MVT::i64 ? PPC::MTCTR8 : PPC::MTCTR)).addReg(Tmp);
8934 BuildMI(*MBB, MI, DL, TII->get(PVT == MVT::i64 ? PPC::BCTR8 : PPC::BCTR));
8936 MI.eraseFromParent();
8941 PPCTargetLowering::EmitInstrWithCustomInserter(MachineInstr &MI,
8942 MachineBasicBlock *BB) const {
8943 if (MI.getOpcode() == TargetOpcode::STACKMAP ||
8944 MI.getOpcode() == TargetOpcode::PATCHPOINT) {
8945 if (Subtarget.isPPC64() && Subtarget.isSVR4ABI() &&
8946 MI.getOpcode() == TargetOpcode::PATCHPOINT) {
8947 // Call lowering should have added an r2 operand to indicate a dependence
8948 // on the TOC base pointer value. It can't however, because there is no
8949 // way to mark the dependence as implicit there, and so the stackmap code
8950 // will confuse it with a regular operand. Instead, add the dependence
8952 setUsesTOCBasePtr(*BB->getParent());
8953 MI.addOperand(MachineOperand::CreateReg(PPC::X2, false, true));
8956 return emitPatchPoint(MI, BB);
8959 if (MI.getOpcode() == PPC::EH_SjLj_SetJmp32 ||
8960 MI.getOpcode() == PPC::EH_SjLj_SetJmp64) {
8961 return emitEHSjLjSetJmp(MI, BB);
8962 } else if (MI.getOpcode() == PPC::EH_SjLj_LongJmp32 ||
8963 MI.getOpcode() == PPC::EH_SjLj_LongJmp64) {
8964 return emitEHSjLjLongJmp(MI, BB);
8967 const TargetInstrInfo *TII = Subtarget.getInstrInfo();
8969 // To "insert" these instructions we actually have to insert their
8970 // control-flow patterns.
8971 const BasicBlock *LLVM_BB = BB->getBasicBlock();
8972 MachineFunction::iterator It = ++BB->getIterator();
8974 MachineFunction *F = BB->getParent();
8976 if (Subtarget.hasISEL() &&
8977 (MI.getOpcode() == PPC::SELECT_CC_I4 ||
8978 MI.getOpcode() == PPC::SELECT_CC_I8 ||
8979 MI.getOpcode() == PPC::SELECT_I4 || MI.getOpcode() == PPC::SELECT_I8)) {
8980 SmallVector<MachineOperand, 2> Cond;
8981 if (MI.getOpcode() == PPC::SELECT_CC_I4 ||
8982 MI.getOpcode() == PPC::SELECT_CC_I8)
8983 Cond.push_back(MI.getOperand(4));
8985 Cond.push_back(MachineOperand::CreateImm(PPC::PRED_BIT_SET));
8986 Cond.push_back(MI.getOperand(1));
8988 DebugLoc dl = MI.getDebugLoc();
8989 TII->insertSelect(*BB, MI, dl, MI.getOperand(0).getReg(), Cond,
8990 MI.getOperand(2).getReg(), MI.getOperand(3).getReg());
8991 } else if (MI.getOpcode() == PPC::SELECT_CC_I4 ||
8992 MI.getOpcode() == PPC::SELECT_CC_I8 ||
8993 MI.getOpcode() == PPC::SELECT_CC_F4 ||
8994 MI.getOpcode() == PPC::SELECT_CC_F8 ||
8995 MI.getOpcode() == PPC::SELECT_CC_QFRC ||
8996 MI.getOpcode() == PPC::SELECT_CC_QSRC ||
8997 MI.getOpcode() == PPC::SELECT_CC_QBRC ||
8998 MI.getOpcode() == PPC::SELECT_CC_VRRC ||
8999 MI.getOpcode() == PPC::SELECT_CC_VSFRC ||
9000 MI.getOpcode() == PPC::SELECT_CC_VSSRC ||
9001 MI.getOpcode() == PPC::SELECT_CC_VSRC ||
9002 MI.getOpcode() == PPC::SELECT_I4 ||
9003 MI.getOpcode() == PPC::SELECT_I8 ||
9004 MI.getOpcode() == PPC::SELECT_F4 ||
9005 MI.getOpcode() == PPC::SELECT_F8 ||
9006 MI.getOpcode() == PPC::SELECT_QFRC ||
9007 MI.getOpcode() == PPC::SELECT_QSRC ||
9008 MI.getOpcode() == PPC::SELECT_QBRC ||
9009 MI.getOpcode() == PPC::SELECT_VRRC ||
9010 MI.getOpcode() == PPC::SELECT_VSFRC ||
9011 MI.getOpcode() == PPC::SELECT_VSSRC ||
9012 MI.getOpcode() == PPC::SELECT_VSRC) {
9013 // The incoming instruction knows the destination vreg to set, the
9014 // condition code register to branch on, the true/false values to
9015 // select between, and a branch opcode to use.
9020 // cmpTY ccX, r1, r2
9022 // fallthrough --> copy0MBB
9023 MachineBasicBlock *thisMBB = BB;
9024 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
9025 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
9026 DebugLoc dl = MI.getDebugLoc();
9027 F->insert(It, copy0MBB);
9028 F->insert(It, sinkMBB);
9030 // Transfer the remainder of BB and its successor edges to sinkMBB.
9031 sinkMBB->splice(sinkMBB->begin(), BB,
9032 std::next(MachineBasicBlock::iterator(MI)), BB->end());
9033 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
9035 // Next, add the true and fallthrough blocks as its successors.
9036 BB->addSuccessor(copy0MBB);
9037 BB->addSuccessor(sinkMBB);
9039 if (MI.getOpcode() == PPC::SELECT_I4 || MI.getOpcode() == PPC::SELECT_I8 ||
9040 MI.getOpcode() == PPC::SELECT_F4 || MI.getOpcode() == PPC::SELECT_F8 ||
9041 MI.getOpcode() == PPC::SELECT_QFRC ||
9042 MI.getOpcode() == PPC::SELECT_QSRC ||
9043 MI.getOpcode() == PPC::SELECT_QBRC ||
9044 MI.getOpcode() == PPC::SELECT_VRRC ||
9045 MI.getOpcode() == PPC::SELECT_VSFRC ||
9046 MI.getOpcode() == PPC::SELECT_VSSRC ||
9047 MI.getOpcode() == PPC::SELECT_VSRC) {
9048 BuildMI(BB, dl, TII->get(PPC::BC))
9049 .addReg(MI.getOperand(1).getReg())
9052 unsigned SelectPred = MI.getOperand(4).getImm();
9053 BuildMI(BB, dl, TII->get(PPC::BCC))
9055 .addReg(MI.getOperand(1).getReg())
9060 // %FalseValue = ...
9061 // # fallthrough to sinkMBB
9064 // Update machine-CFG edges
9065 BB->addSuccessor(sinkMBB);
9068 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
9071 BuildMI(*BB, BB->begin(), dl, TII->get(PPC::PHI), MI.getOperand(0).getReg())
9072 .addReg(MI.getOperand(3).getReg())
9074 .addReg(MI.getOperand(2).getReg())
9076 } else if (MI.getOpcode() == PPC::ReadTB) {
9077 // To read the 64-bit time-base register on a 32-bit target, we read the
9078 // two halves. Should the counter have wrapped while it was being read, we
9079 // need to try again.
9082 // mfspr Rx,TBU # load from TBU
9083 // mfspr Ry,TB # load from TB
9084 // mfspr Rz,TBU # load from TBU
9085 // cmpw crX,Rx,Rz # check if 'old'='new'
9086 // bne readLoop # branch if they're not equal
9089 MachineBasicBlock *readMBB = F->CreateMachineBasicBlock(LLVM_BB);
9090 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
9091 DebugLoc dl = MI.getDebugLoc();
9092 F->insert(It, readMBB);
9093 F->insert(It, sinkMBB);
9095 // Transfer the remainder of BB and its successor edges to sinkMBB.
9096 sinkMBB->splice(sinkMBB->begin(), BB,
9097 std::next(MachineBasicBlock::iterator(MI)), BB->end());
9098 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
9100 BB->addSuccessor(readMBB);
9103 MachineRegisterInfo &RegInfo = F->getRegInfo();
9104 unsigned ReadAgainReg = RegInfo.createVirtualRegister(&PPC::GPRCRegClass);
9105 unsigned LoReg = MI.getOperand(0).getReg();
9106 unsigned HiReg = MI.getOperand(1).getReg();
9108 BuildMI(BB, dl, TII->get(PPC::MFSPR), HiReg).addImm(269);
9109 BuildMI(BB, dl, TII->get(PPC::MFSPR), LoReg).addImm(268);
9110 BuildMI(BB, dl, TII->get(PPC::MFSPR), ReadAgainReg).addImm(269);
9112 unsigned CmpReg = RegInfo.createVirtualRegister(&PPC::CRRCRegClass);
9114 BuildMI(BB, dl, TII->get(PPC::CMPW), CmpReg)
9115 .addReg(HiReg).addReg(ReadAgainReg);
9116 BuildMI(BB, dl, TII->get(PPC::BCC))
9117 .addImm(PPC::PRED_NE).addReg(CmpReg).addMBB(readMBB);
9119 BB->addSuccessor(readMBB);
9120 BB->addSuccessor(sinkMBB);
9121 } else if (MI.getOpcode() == PPC::ATOMIC_LOAD_ADD_I8)
9122 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::ADD4);
9123 else if (MI.getOpcode() == PPC::ATOMIC_LOAD_ADD_I16)
9124 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::ADD4);
9125 else if (MI.getOpcode() == PPC::ATOMIC_LOAD_ADD_I32)
9126 BB = EmitAtomicBinary(MI, BB, 4, PPC::ADD4);
9127 else if (MI.getOpcode() == PPC::ATOMIC_LOAD_ADD_I64)
9128 BB = EmitAtomicBinary(MI, BB, 8, PPC::ADD8);
9130 else if (MI.getOpcode() == PPC::ATOMIC_LOAD_AND_I8)
9131 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::AND);
9132 else if (MI.getOpcode() == PPC::ATOMIC_LOAD_AND_I16)
9133 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::AND);
9134 else if (MI.getOpcode() == PPC::ATOMIC_LOAD_AND_I32)
9135 BB = EmitAtomicBinary(MI, BB, 4, PPC::AND);
9136 else if (MI.getOpcode() == PPC::ATOMIC_LOAD_AND_I64)
9137 BB = EmitAtomicBinary(MI, BB, 8, PPC::AND8);
9139 else if (MI.getOpcode() == PPC::ATOMIC_LOAD_OR_I8)
9140 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::OR);
9141 else if (MI.getOpcode() == PPC::ATOMIC_LOAD_OR_I16)
9142 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::OR);
9143 else if (MI.getOpcode() == PPC::ATOMIC_LOAD_OR_I32)
9144 BB = EmitAtomicBinary(MI, BB, 4, PPC::OR);
9145 else if (MI.getOpcode() == PPC::ATOMIC_LOAD_OR_I64)
9146 BB = EmitAtomicBinary(MI, BB, 8, PPC::OR8);
9148 else if (MI.getOpcode() == PPC::ATOMIC_LOAD_XOR_I8)
9149 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::XOR);
9150 else if (MI.getOpcode() == PPC::ATOMIC_LOAD_XOR_I16)
9151 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::XOR);
9152 else if (MI.getOpcode() == PPC::ATOMIC_LOAD_XOR_I32)
9153 BB = EmitAtomicBinary(MI, BB, 4, PPC::XOR);
9154 else if (MI.getOpcode() == PPC::ATOMIC_LOAD_XOR_I64)
9155 BB = EmitAtomicBinary(MI, BB, 8, PPC::XOR8);
9157 else if (MI.getOpcode() == PPC::ATOMIC_LOAD_NAND_I8)
9158 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::NAND);
9159 else if (MI.getOpcode() == PPC::ATOMIC_LOAD_NAND_I16)
9160 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::NAND);
9161 else if (MI.getOpcode() == PPC::ATOMIC_LOAD_NAND_I32)
9162 BB = EmitAtomicBinary(MI, BB, 4, PPC::NAND);
9163 else if (MI.getOpcode() == PPC::ATOMIC_LOAD_NAND_I64)
9164 BB = EmitAtomicBinary(MI, BB, 8, PPC::NAND8);
9166 else if (MI.getOpcode() == PPC::ATOMIC_LOAD_SUB_I8)
9167 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::SUBF);
9168 else if (MI.getOpcode() == PPC::ATOMIC_LOAD_SUB_I16)
9169 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::SUBF);
9170 else if (MI.getOpcode() == PPC::ATOMIC_LOAD_SUB_I32)
9171 BB = EmitAtomicBinary(MI, BB, 4, PPC::SUBF);
9172 else if (MI.getOpcode() == PPC::ATOMIC_LOAD_SUB_I64)
9173 BB = EmitAtomicBinary(MI, BB, 8, PPC::SUBF8);
9175 else if (MI.getOpcode() == PPC::ATOMIC_LOAD_MIN_I8)
9176 BB = EmitPartwordAtomicBinary(MI, BB, true, 0, PPC::CMPW, PPC::PRED_GE);
9177 else if (MI.getOpcode() == PPC::ATOMIC_LOAD_MIN_I16)
9178 BB = EmitPartwordAtomicBinary(MI, BB, false, 0, PPC::CMPW, PPC::PRED_GE);
9179 else if (MI.getOpcode() == PPC::ATOMIC_LOAD_MIN_I32)
9180 BB = EmitAtomicBinary(MI, BB, 4, 0, PPC::CMPW, PPC::PRED_GE);
9181 else if (MI.getOpcode() == PPC::ATOMIC_LOAD_MIN_I64)
9182 BB = EmitAtomicBinary(MI, BB, 8, 0, PPC::CMPD, PPC::PRED_GE);
9184 else if (MI.getOpcode() == PPC::ATOMIC_LOAD_MAX_I8)
9185 BB = EmitPartwordAtomicBinary(MI, BB, true, 0, PPC::CMPW, PPC::PRED_LE);
9186 else if (MI.getOpcode() == PPC::ATOMIC_LOAD_MAX_I16)
9187 BB = EmitPartwordAtomicBinary(MI, BB, false, 0, PPC::CMPW, PPC::PRED_LE);
9188 else if (MI.getOpcode() == PPC::ATOMIC_LOAD_MAX_I32)
9189 BB = EmitAtomicBinary(MI, BB, 4, 0, PPC::CMPW, PPC::PRED_LE);
9190 else if (MI.getOpcode() == PPC::ATOMIC_LOAD_MAX_I64)
9191 BB = EmitAtomicBinary(MI, BB, 8, 0, PPC::CMPD, PPC::PRED_LE);
9193 else if (MI.getOpcode() == PPC::ATOMIC_LOAD_UMIN_I8)
9194 BB = EmitPartwordAtomicBinary(MI, BB, true, 0, PPC::CMPLW, PPC::PRED_GE);
9195 else if (MI.getOpcode() == PPC::ATOMIC_LOAD_UMIN_I16)
9196 BB = EmitPartwordAtomicBinary(MI, BB, false, 0, PPC::CMPLW, PPC::PRED_GE);
9197 else if (MI.getOpcode() == PPC::ATOMIC_LOAD_UMIN_I32)
9198 BB = EmitAtomicBinary(MI, BB, 4, 0, PPC::CMPLW, PPC::PRED_GE);
9199 else if (MI.getOpcode() == PPC::ATOMIC_LOAD_UMIN_I64)
9200 BB = EmitAtomicBinary(MI, BB, 8, 0, PPC::CMPLD, PPC::PRED_GE);
9202 else if (MI.getOpcode() == PPC::ATOMIC_LOAD_UMAX_I8)
9203 BB = EmitPartwordAtomicBinary(MI, BB, true, 0, PPC::CMPLW, PPC::PRED_LE);
9204 else if (MI.getOpcode() == PPC::ATOMIC_LOAD_UMAX_I16)
9205 BB = EmitPartwordAtomicBinary(MI, BB, false, 0, PPC::CMPLW, PPC::PRED_LE);
9206 else if (MI.getOpcode() == PPC::ATOMIC_LOAD_UMAX_I32)
9207 BB = EmitAtomicBinary(MI, BB, 4, 0, PPC::CMPLW, PPC::PRED_LE);
9208 else if (MI.getOpcode() == PPC::ATOMIC_LOAD_UMAX_I64)
9209 BB = EmitAtomicBinary(MI, BB, 8, 0, PPC::CMPLD, PPC::PRED_LE);
9211 else if (MI.getOpcode() == PPC::ATOMIC_SWAP_I8)
9212 BB = EmitPartwordAtomicBinary(MI, BB, true, 0);
9213 else if (MI.getOpcode() == PPC::ATOMIC_SWAP_I16)
9214 BB = EmitPartwordAtomicBinary(MI, BB, false, 0);
9215 else if (MI.getOpcode() == PPC::ATOMIC_SWAP_I32)
9216 BB = EmitAtomicBinary(MI, BB, 4, 0);
9217 else if (MI.getOpcode() == PPC::ATOMIC_SWAP_I64)
9218 BB = EmitAtomicBinary(MI, BB, 8, 0);
9220 else if (MI.getOpcode() == PPC::ATOMIC_CMP_SWAP_I32 ||
9221 MI.getOpcode() == PPC::ATOMIC_CMP_SWAP_I64 ||
9222 (Subtarget.hasPartwordAtomics() &&
9223 MI.getOpcode() == PPC::ATOMIC_CMP_SWAP_I8) ||
9224 (Subtarget.hasPartwordAtomics() &&
9225 MI.getOpcode() == PPC::ATOMIC_CMP_SWAP_I16)) {
9226 bool is64bit = MI.getOpcode() == PPC::ATOMIC_CMP_SWAP_I64;
9228 auto LoadMnemonic = PPC::LDARX;
9229 auto StoreMnemonic = PPC::STDCX;
9230 switch (MI.getOpcode()) {
9232 llvm_unreachable("Compare and swap of unknown size");
9233 case PPC::ATOMIC_CMP_SWAP_I8:
9234 LoadMnemonic = PPC::LBARX;
9235 StoreMnemonic = PPC::STBCX;
9236 assert(Subtarget.hasPartwordAtomics() && "No support partword atomics.");
9238 case PPC::ATOMIC_CMP_SWAP_I16:
9239 LoadMnemonic = PPC::LHARX;
9240 StoreMnemonic = PPC::STHCX;
9241 assert(Subtarget.hasPartwordAtomics() && "No support partword atomics.");
9243 case PPC::ATOMIC_CMP_SWAP_I32:
9244 LoadMnemonic = PPC::LWARX;
9245 StoreMnemonic = PPC::STWCX;
9247 case PPC::ATOMIC_CMP_SWAP_I64:
9248 LoadMnemonic = PPC::LDARX;
9249 StoreMnemonic = PPC::STDCX;
9252 unsigned dest = MI.getOperand(0).getReg();
9253 unsigned ptrA = MI.getOperand(1).getReg();
9254 unsigned ptrB = MI.getOperand(2).getReg();
9255 unsigned oldval = MI.getOperand(3).getReg();
9256 unsigned newval = MI.getOperand(4).getReg();
9257 DebugLoc dl = MI.getDebugLoc();
9259 MachineBasicBlock *loop1MBB = F->CreateMachineBasicBlock(LLVM_BB);
9260 MachineBasicBlock *loop2MBB = F->CreateMachineBasicBlock(LLVM_BB);
9261 MachineBasicBlock *midMBB = F->CreateMachineBasicBlock(LLVM_BB);
9262 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
9263 F->insert(It, loop1MBB);
9264 F->insert(It, loop2MBB);
9265 F->insert(It, midMBB);
9266 F->insert(It, exitMBB);
9267 exitMBB->splice(exitMBB->begin(), BB,
9268 std::next(MachineBasicBlock::iterator(MI)), BB->end());
9269 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
9273 // fallthrough --> loopMBB
9274 BB->addSuccessor(loop1MBB);
9277 // l[bhwd]arx dest, ptr
9278 // cmp[wd] dest, oldval
9281 // st[bhwd]cx. newval, ptr
9285 // st[bhwd]cx. dest, ptr
9288 BuildMI(BB, dl, TII->get(LoadMnemonic), dest)
9289 .addReg(ptrA).addReg(ptrB);
9290 BuildMI(BB, dl, TII->get(is64bit ? PPC::CMPD : PPC::CMPW), PPC::CR0)
9291 .addReg(oldval).addReg(dest);
9292 BuildMI(BB, dl, TII->get(PPC::BCC))
9293 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(midMBB);
9294 BB->addSuccessor(loop2MBB);
9295 BB->addSuccessor(midMBB);
9298 BuildMI(BB, dl, TII->get(StoreMnemonic))
9299 .addReg(newval).addReg(ptrA).addReg(ptrB);
9300 BuildMI(BB, dl, TII->get(PPC::BCC))
9301 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loop1MBB);
9302 BuildMI(BB, dl, TII->get(PPC::B)).addMBB(exitMBB);
9303 BB->addSuccessor(loop1MBB);
9304 BB->addSuccessor(exitMBB);
9307 BuildMI(BB, dl, TII->get(StoreMnemonic))
9308 .addReg(dest).addReg(ptrA).addReg(ptrB);
9309 BB->addSuccessor(exitMBB);
9314 } else if (MI.getOpcode() == PPC::ATOMIC_CMP_SWAP_I8 ||
9315 MI.getOpcode() == PPC::ATOMIC_CMP_SWAP_I16) {
9316 // We must use 64-bit registers for addresses when targeting 64-bit,
9317 // since we're actually doing arithmetic on them. Other registers
9319 bool is64bit = Subtarget.isPPC64();
9320 bool isLittleEndian = Subtarget.isLittleEndian();
9321 bool is8bit = MI.getOpcode() == PPC::ATOMIC_CMP_SWAP_I8;
9323 unsigned dest = MI.getOperand(0).getReg();
9324 unsigned ptrA = MI.getOperand(1).getReg();
9325 unsigned ptrB = MI.getOperand(2).getReg();
9326 unsigned oldval = MI.getOperand(3).getReg();
9327 unsigned newval = MI.getOperand(4).getReg();
9328 DebugLoc dl = MI.getDebugLoc();
9330 MachineBasicBlock *loop1MBB = F->CreateMachineBasicBlock(LLVM_BB);
9331 MachineBasicBlock *loop2MBB = F->CreateMachineBasicBlock(LLVM_BB);
9332 MachineBasicBlock *midMBB = F->CreateMachineBasicBlock(LLVM_BB);
9333 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
9334 F->insert(It, loop1MBB);
9335 F->insert(It, loop2MBB);
9336 F->insert(It, midMBB);
9337 F->insert(It, exitMBB);
9338 exitMBB->splice(exitMBB->begin(), BB,
9339 std::next(MachineBasicBlock::iterator(MI)), BB->end());
9340 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
9342 MachineRegisterInfo &RegInfo = F->getRegInfo();
9343 const TargetRegisterClass *RC = is64bit ? &PPC::G8RCRegClass
9344 : &PPC::GPRCRegClass;
9345 unsigned PtrReg = RegInfo.createVirtualRegister(RC);
9346 unsigned Shift1Reg = RegInfo.createVirtualRegister(RC);
9348 isLittleEndian ? Shift1Reg : RegInfo.createVirtualRegister(RC);
9349 unsigned NewVal2Reg = RegInfo.createVirtualRegister(RC);
9350 unsigned NewVal3Reg = RegInfo.createVirtualRegister(RC);
9351 unsigned OldVal2Reg = RegInfo.createVirtualRegister(RC);
9352 unsigned OldVal3Reg = RegInfo.createVirtualRegister(RC);
9353 unsigned MaskReg = RegInfo.createVirtualRegister(RC);
9354 unsigned Mask2Reg = RegInfo.createVirtualRegister(RC);
9355 unsigned Mask3Reg = RegInfo.createVirtualRegister(RC);
9356 unsigned Tmp2Reg = RegInfo.createVirtualRegister(RC);
9357 unsigned Tmp4Reg = RegInfo.createVirtualRegister(RC);
9358 unsigned TmpDestReg = RegInfo.createVirtualRegister(RC);
9360 unsigned TmpReg = RegInfo.createVirtualRegister(RC);
9361 unsigned ZeroReg = is64bit ? PPC::ZERO8 : PPC::ZERO;
9364 // fallthrough --> loopMBB
9365 BB->addSuccessor(loop1MBB);
9367 // The 4-byte load must be aligned, while a char or short may be
9368 // anywhere in the word. Hence all this nasty bookkeeping code.
9369 // add ptr1, ptrA, ptrB [copy if ptrA==0]
9370 // rlwinm shift1, ptr1, 3, 27, 28 [3, 27, 27]
9371 // xori shift, shift1, 24 [16]
9372 // rlwinm ptr, ptr1, 0, 0, 29
9373 // slw newval2, newval, shift
9374 // slw oldval2, oldval,shift
9375 // li mask2, 255 [li mask3, 0; ori mask2, mask3, 65535]
9376 // slw mask, mask2, shift
9377 // and newval3, newval2, mask
9378 // and oldval3, oldval2, mask
9380 // lwarx tmpDest, ptr
9381 // and tmp, tmpDest, mask
9382 // cmpw tmp, oldval3
9385 // andc tmp2, tmpDest, mask
9386 // or tmp4, tmp2, newval3
9391 // stwcx. tmpDest, ptr
9393 // srw dest, tmpDest, shift
9394 if (ptrA != ZeroReg) {
9395 Ptr1Reg = RegInfo.createVirtualRegister(RC);
9396 BuildMI(BB, dl, TII->get(is64bit ? PPC::ADD8 : PPC::ADD4), Ptr1Reg)
9397 .addReg(ptrA).addReg(ptrB);
9401 BuildMI(BB, dl, TII->get(PPC::RLWINM), Shift1Reg).addReg(Ptr1Reg)
9402 .addImm(3).addImm(27).addImm(is8bit ? 28 : 27);
9403 if (!isLittleEndian)
9404 BuildMI(BB, dl, TII->get(is64bit ? PPC::XORI8 : PPC::XORI), ShiftReg)
9405 .addReg(Shift1Reg).addImm(is8bit ? 24 : 16);
9407 BuildMI(BB, dl, TII->get(PPC::RLDICR), PtrReg)
9408 .addReg(Ptr1Reg).addImm(0).addImm(61);
9410 BuildMI(BB, dl, TII->get(PPC::RLWINM), PtrReg)
9411 .addReg(Ptr1Reg).addImm(0).addImm(0).addImm(29);
9412 BuildMI(BB, dl, TII->get(PPC::SLW), NewVal2Reg)
9413 .addReg(newval).addReg(ShiftReg);
9414 BuildMI(BB, dl, TII->get(PPC::SLW), OldVal2Reg)
9415 .addReg(oldval).addReg(ShiftReg);
9417 BuildMI(BB, dl, TII->get(PPC::LI), Mask2Reg).addImm(255);
9419 BuildMI(BB, dl, TII->get(PPC::LI), Mask3Reg).addImm(0);
9420 BuildMI(BB, dl, TII->get(PPC::ORI), Mask2Reg)
9421 .addReg(Mask3Reg).addImm(65535);
9423 BuildMI(BB, dl, TII->get(PPC::SLW), MaskReg)
9424 .addReg(Mask2Reg).addReg(ShiftReg);
9425 BuildMI(BB, dl, TII->get(PPC::AND), NewVal3Reg)
9426 .addReg(NewVal2Reg).addReg(MaskReg);
9427 BuildMI(BB, dl, TII->get(PPC::AND), OldVal3Reg)
9428 .addReg(OldVal2Reg).addReg(MaskReg);
9431 BuildMI(BB, dl, TII->get(PPC::LWARX), TmpDestReg)
9432 .addReg(ZeroReg).addReg(PtrReg);
9433 BuildMI(BB, dl, TII->get(PPC::AND),TmpReg)
9434 .addReg(TmpDestReg).addReg(MaskReg);
9435 BuildMI(BB, dl, TII->get(PPC::CMPW), PPC::CR0)
9436 .addReg(TmpReg).addReg(OldVal3Reg);
9437 BuildMI(BB, dl, TII->get(PPC::BCC))
9438 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(midMBB);
9439 BB->addSuccessor(loop2MBB);
9440 BB->addSuccessor(midMBB);
9443 BuildMI(BB, dl, TII->get(PPC::ANDC),Tmp2Reg)
9444 .addReg(TmpDestReg).addReg(MaskReg);
9445 BuildMI(BB, dl, TII->get(PPC::OR),Tmp4Reg)
9446 .addReg(Tmp2Reg).addReg(NewVal3Reg);
9447 BuildMI(BB, dl, TII->get(PPC::STWCX)).addReg(Tmp4Reg)
9448 .addReg(ZeroReg).addReg(PtrReg);
9449 BuildMI(BB, dl, TII->get(PPC::BCC))
9450 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loop1MBB);
9451 BuildMI(BB, dl, TII->get(PPC::B)).addMBB(exitMBB);
9452 BB->addSuccessor(loop1MBB);
9453 BB->addSuccessor(exitMBB);
9456 BuildMI(BB, dl, TII->get(PPC::STWCX)).addReg(TmpDestReg)
9457 .addReg(ZeroReg).addReg(PtrReg);
9458 BB->addSuccessor(exitMBB);
9463 BuildMI(*BB, BB->begin(), dl, TII->get(PPC::SRW),dest).addReg(TmpReg)
9465 } else if (MI.getOpcode() == PPC::FADDrtz) {
9466 // This pseudo performs an FADD with rounding mode temporarily forced
9467 // to round-to-zero. We emit this via custom inserter since the FPSCR
9468 // is not modeled at the SelectionDAG level.
9469 unsigned Dest = MI.getOperand(0).getReg();
9470 unsigned Src1 = MI.getOperand(1).getReg();
9471 unsigned Src2 = MI.getOperand(2).getReg();
9472 DebugLoc dl = MI.getDebugLoc();
9474 MachineRegisterInfo &RegInfo = F->getRegInfo();
9475 unsigned MFFSReg = RegInfo.createVirtualRegister(&PPC::F8RCRegClass);
9477 // Save FPSCR value.
9478 BuildMI(*BB, MI, dl, TII->get(PPC::MFFS), MFFSReg);
9480 // Set rounding mode to round-to-zero.
9481 BuildMI(*BB, MI, dl, TII->get(PPC::MTFSB1)).addImm(31);
9482 BuildMI(*BB, MI, dl, TII->get(PPC::MTFSB0)).addImm(30);
9484 // Perform addition.
9485 BuildMI(*BB, MI, dl, TII->get(PPC::FADD), Dest).addReg(Src1).addReg(Src2);
9487 // Restore FPSCR value.
9488 BuildMI(*BB, MI, dl, TII->get(PPC::MTFSFb)).addImm(1).addReg(MFFSReg);
9489 } else if (MI.getOpcode() == PPC::ANDIo_1_EQ_BIT ||
9490 MI.getOpcode() == PPC::ANDIo_1_GT_BIT ||
9491 MI.getOpcode() == PPC::ANDIo_1_EQ_BIT8 ||
9492 MI.getOpcode() == PPC::ANDIo_1_GT_BIT8) {
9493 unsigned Opcode = (MI.getOpcode() == PPC::ANDIo_1_EQ_BIT8 ||
9494 MI.getOpcode() == PPC::ANDIo_1_GT_BIT8)
9497 bool isEQ = (MI.getOpcode() == PPC::ANDIo_1_EQ_BIT ||
9498 MI.getOpcode() == PPC::ANDIo_1_EQ_BIT8);
9500 MachineRegisterInfo &RegInfo = F->getRegInfo();
9501 unsigned Dest = RegInfo.createVirtualRegister(Opcode == PPC::ANDIo ?
9502 &PPC::GPRCRegClass :
9503 &PPC::G8RCRegClass);
9505 DebugLoc dl = MI.getDebugLoc();
9506 BuildMI(*BB, MI, dl, TII->get(Opcode), Dest)
9507 .addReg(MI.getOperand(1).getReg())
9509 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY),
9510 MI.getOperand(0).getReg())
9511 .addReg(isEQ ? PPC::CR0EQ : PPC::CR0GT);
9512 } else if (MI.getOpcode() == PPC::TCHECK_RET) {
9513 DebugLoc Dl = MI.getDebugLoc();
9514 MachineRegisterInfo &RegInfo = F->getRegInfo();
9515 unsigned CRReg = RegInfo.createVirtualRegister(&PPC::CRRCRegClass);
9516 BuildMI(*BB, MI, Dl, TII->get(PPC::TCHECK), CRReg);
9519 llvm_unreachable("Unexpected instr type to insert");
9522 MI.eraseFromParent(); // The pseudo instruction is gone now.
9526 //===----------------------------------------------------------------------===//
9527 // Target Optimization Hooks
9528 //===----------------------------------------------------------------------===//
9530 static std::string getRecipOp(const char *Base, EVT VT) {
9531 std::string RecipOp(Base);
9532 if (VT.getScalarType() == MVT::f64)
9538 RecipOp = "vec-" + RecipOp;
9543 SDValue PPCTargetLowering::getRsqrtEstimate(SDValue Operand,
9544 DAGCombinerInfo &DCI,
9545 unsigned &RefinementSteps,
9546 bool &UseOneConstNR) const {
9547 EVT VT = Operand.getValueType();
9548 if ((VT == MVT::f32 && Subtarget.hasFRSQRTES()) ||
9549 (VT == MVT::f64 && Subtarget.hasFRSQRTE()) ||
9550 (VT == MVT::v4f32 && Subtarget.hasAltivec()) ||
9551 (VT == MVT::v2f64 && Subtarget.hasVSX()) ||
9552 (VT == MVT::v4f32 && Subtarget.hasQPX()) ||
9553 (VT == MVT::v4f64 && Subtarget.hasQPX())) {
9554 TargetRecip Recips = DCI.DAG.getTarget().Options.Reciprocals;
9555 std::string RecipOp = getRecipOp("sqrt", VT);
9556 if (!Recips.isEnabled(RecipOp))
9559 RefinementSteps = Recips.getRefinementSteps(RecipOp);
9560 UseOneConstNR = true;
9561 return DCI.DAG.getNode(PPCISD::FRSQRTE, SDLoc(Operand), VT, Operand);
9566 SDValue PPCTargetLowering::getRecipEstimate(SDValue Operand,
9567 DAGCombinerInfo &DCI,
9568 unsigned &RefinementSteps) const {
9569 EVT VT = Operand.getValueType();
9570 if ((VT == MVT::f32 && Subtarget.hasFRES()) ||
9571 (VT == MVT::f64 && Subtarget.hasFRE()) ||
9572 (VT == MVT::v4f32 && Subtarget.hasAltivec()) ||
9573 (VT == MVT::v2f64 && Subtarget.hasVSX()) ||
9574 (VT == MVT::v4f32 && Subtarget.hasQPX()) ||
9575 (VT == MVT::v4f64 && Subtarget.hasQPX())) {
9576 TargetRecip Recips = DCI.DAG.getTarget().Options.Reciprocals;
9577 std::string RecipOp = getRecipOp("div", VT);
9578 if (!Recips.isEnabled(RecipOp))
9581 RefinementSteps = Recips.getRefinementSteps(RecipOp);
9582 return DCI.DAG.getNode(PPCISD::FRE, SDLoc(Operand), VT, Operand);
9587 unsigned PPCTargetLowering::combineRepeatedFPDivisors() const {
9588 // Note: This functionality is used only when unsafe-fp-math is enabled, and
9589 // on cores with reciprocal estimates (which are used when unsafe-fp-math is
9590 // enabled for division), this functionality is redundant with the default
9591 // combiner logic (once the division -> reciprocal/multiply transformation
9592 // has taken place). As a result, this matters more for older cores than for
9595 // Combine multiple FDIVs with the same divisor into multiple FMULs by the
9596 // reciprocal if there are two or more FDIVs (for embedded cores with only
9597 // one FP pipeline) for three or more FDIVs (for generic OOO cores).
9598 switch (Subtarget.getDarwinDirective()) {
9603 case PPC::DIR_E500mc:
9604 case PPC::DIR_E5500:
9609 // isConsecutiveLSLoc needs to work even if all adds have not yet been
9610 // collapsed, and so we need to look through chains of them.
9611 static void getBaseWithConstantOffset(SDValue Loc, SDValue &Base,
9612 int64_t& Offset, SelectionDAG &DAG) {
9613 if (DAG.isBaseWithConstantOffset(Loc)) {
9614 Base = Loc.getOperand(0);
9615 Offset += cast<ConstantSDNode>(Loc.getOperand(1))->getSExtValue();
9617 // The base might itself be a base plus an offset, and if so, accumulate
9619 getBaseWithConstantOffset(Loc.getOperand(0), Base, Offset, DAG);
9623 static bool isConsecutiveLSLoc(SDValue Loc, EVT VT, LSBaseSDNode *Base,
9624 unsigned Bytes, int Dist,
9625 SelectionDAG &DAG) {
9626 if (VT.getSizeInBits() / 8 != Bytes)
9629 SDValue BaseLoc = Base->getBasePtr();
9630 if (Loc.getOpcode() == ISD::FrameIndex) {
9631 if (BaseLoc.getOpcode() != ISD::FrameIndex)
9633 const MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo();
9634 int FI = cast<FrameIndexSDNode>(Loc)->getIndex();
9635 int BFI = cast<FrameIndexSDNode>(BaseLoc)->getIndex();
9636 int FS = MFI.getObjectSize(FI);
9637 int BFS = MFI.getObjectSize(BFI);
9638 if (FS != BFS || FS != (int)Bytes) return false;
9639 return MFI.getObjectOffset(FI) == (MFI.getObjectOffset(BFI) + Dist*Bytes);
9642 SDValue Base1 = Loc, Base2 = BaseLoc;
9643 int64_t Offset1 = 0, Offset2 = 0;
9644 getBaseWithConstantOffset(Loc, Base1, Offset1, DAG);
9645 getBaseWithConstantOffset(BaseLoc, Base2, Offset2, DAG);
9646 if (Base1 == Base2 && Offset1 == (Offset2 + Dist * Bytes))
9649 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9650 const GlobalValue *GV1 = nullptr;
9651 const GlobalValue *GV2 = nullptr;
9654 bool isGA1 = TLI.isGAPlusOffset(Loc.getNode(), GV1, Offset1);
9655 bool isGA2 = TLI.isGAPlusOffset(BaseLoc.getNode(), GV2, Offset2);
9656 if (isGA1 && isGA2 && GV1 == GV2)
9657 return Offset1 == (Offset2 + Dist*Bytes);
9661 // Like SelectionDAG::isConsecutiveLoad, but also works for stores, and does
9662 // not enforce equality of the chain operands.
9663 static bool isConsecutiveLS(SDNode *N, LSBaseSDNode *Base,
9664 unsigned Bytes, int Dist,
9665 SelectionDAG &DAG) {
9666 if (LSBaseSDNode *LS = dyn_cast<LSBaseSDNode>(N)) {
9667 EVT VT = LS->getMemoryVT();
9668 SDValue Loc = LS->getBasePtr();
9669 return isConsecutiveLSLoc(Loc, VT, Base, Bytes, Dist, DAG);
9672 if (N->getOpcode() == ISD::INTRINSIC_W_CHAIN) {
9674 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
9675 default: return false;
9676 case Intrinsic::ppc_qpx_qvlfd:
9677 case Intrinsic::ppc_qpx_qvlfda:
9680 case Intrinsic::ppc_qpx_qvlfs:
9681 case Intrinsic::ppc_qpx_qvlfsa:
9684 case Intrinsic::ppc_qpx_qvlfcd:
9685 case Intrinsic::ppc_qpx_qvlfcda:
9688 case Intrinsic::ppc_qpx_qvlfcs:
9689 case Intrinsic::ppc_qpx_qvlfcsa:
9692 case Intrinsic::ppc_qpx_qvlfiwa:
9693 case Intrinsic::ppc_qpx_qvlfiwz:
9694 case Intrinsic::ppc_altivec_lvx:
9695 case Intrinsic::ppc_altivec_lvxl:
9696 case Intrinsic::ppc_vsx_lxvw4x:
9699 case Intrinsic::ppc_vsx_lxvd2x:
9702 case Intrinsic::ppc_altivec_lvebx:
9705 case Intrinsic::ppc_altivec_lvehx:
9708 case Intrinsic::ppc_altivec_lvewx:
9713 return isConsecutiveLSLoc(N->getOperand(2), VT, Base, Bytes, Dist, DAG);
9716 if (N->getOpcode() == ISD::INTRINSIC_VOID) {
9718 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
9719 default: return false;
9720 case Intrinsic::ppc_qpx_qvstfd:
9721 case Intrinsic::ppc_qpx_qvstfda:
9724 case Intrinsic::ppc_qpx_qvstfs:
9725 case Intrinsic::ppc_qpx_qvstfsa:
9728 case Intrinsic::ppc_qpx_qvstfcd:
9729 case Intrinsic::ppc_qpx_qvstfcda:
9732 case Intrinsic::ppc_qpx_qvstfcs:
9733 case Intrinsic::ppc_qpx_qvstfcsa:
9736 case Intrinsic::ppc_qpx_qvstfiw:
9737 case Intrinsic::ppc_qpx_qvstfiwa:
9738 case Intrinsic::ppc_altivec_stvx:
9739 case Intrinsic::ppc_altivec_stvxl:
9740 case Intrinsic::ppc_vsx_stxvw4x:
9743 case Intrinsic::ppc_vsx_stxvd2x:
9746 case Intrinsic::ppc_altivec_stvebx:
9749 case Intrinsic::ppc_altivec_stvehx:
9752 case Intrinsic::ppc_altivec_stvewx:
9757 return isConsecutiveLSLoc(N->getOperand(3), VT, Base, Bytes, Dist, DAG);
9763 // Return true is there is a nearyby consecutive load to the one provided
9764 // (regardless of alignment). We search up and down the chain, looking though
9765 // token factors and other loads (but nothing else). As a result, a true result
9766 // indicates that it is safe to create a new consecutive load adjacent to the
9768 static bool findConsecutiveLoad(LoadSDNode *LD, SelectionDAG &DAG) {
9769 SDValue Chain = LD->getChain();
9770 EVT VT = LD->getMemoryVT();
9772 SmallSet<SDNode *, 16> LoadRoots;
9773 SmallVector<SDNode *, 8> Queue(1, Chain.getNode());
9774 SmallSet<SDNode *, 16> Visited;
9776 // First, search up the chain, branching to follow all token-factor operands.
9777 // If we find a consecutive load, then we're done, otherwise, record all
9778 // nodes just above the top-level loads and token factors.
9779 while (!Queue.empty()) {
9780 SDNode *ChainNext = Queue.pop_back_val();
9781 if (!Visited.insert(ChainNext).second)
9784 if (MemSDNode *ChainLD = dyn_cast<MemSDNode>(ChainNext)) {
9785 if (isConsecutiveLS(ChainLD, LD, VT.getStoreSize(), 1, DAG))
9788 if (!Visited.count(ChainLD->getChain().getNode()))
9789 Queue.push_back(ChainLD->getChain().getNode());
9790 } else if (ChainNext->getOpcode() == ISD::TokenFactor) {
9791 for (const SDUse &O : ChainNext->ops())
9792 if (!Visited.count(O.getNode()))
9793 Queue.push_back(O.getNode());
9795 LoadRoots.insert(ChainNext);
9798 // Second, search down the chain, starting from the top-level nodes recorded
9799 // in the first phase. These top-level nodes are the nodes just above all
9800 // loads and token factors. Starting with their uses, recursively look though
9801 // all loads (just the chain uses) and token factors to find a consecutive
9806 for (SmallSet<SDNode *, 16>::iterator I = LoadRoots.begin(),
9807 IE = LoadRoots.end(); I != IE; ++I) {
9808 Queue.push_back(*I);
9810 while (!Queue.empty()) {
9811 SDNode *LoadRoot = Queue.pop_back_val();
9812 if (!Visited.insert(LoadRoot).second)
9815 if (MemSDNode *ChainLD = dyn_cast<MemSDNode>(LoadRoot))
9816 if (isConsecutiveLS(ChainLD, LD, VT.getStoreSize(), 1, DAG))
9819 for (SDNode::use_iterator UI = LoadRoot->use_begin(),
9820 UE = LoadRoot->use_end(); UI != UE; ++UI)
9821 if (((isa<MemSDNode>(*UI) &&
9822 cast<MemSDNode>(*UI)->getChain().getNode() == LoadRoot) ||
9823 UI->getOpcode() == ISD::TokenFactor) && !Visited.count(*UI))
9824 Queue.push_back(*UI);
9831 SDValue PPCTargetLowering::DAGCombineTruncBoolExt(SDNode *N,
9832 DAGCombinerInfo &DCI) const {
9833 SelectionDAG &DAG = DCI.DAG;
9836 assert(Subtarget.useCRBits() && "Expecting to be tracking CR bits");
9837 // If we're tracking CR bits, we need to be careful that we don't have:
9838 // trunc(binary-ops(zext(x), zext(y)))
9840 // trunc(binary-ops(binary-ops(zext(x), zext(y)), ...)
9841 // such that we're unnecessarily moving things into GPRs when it would be
9842 // better to keep them in CR bits.
9844 // Note that trunc here can be an actual i1 trunc, or can be the effective
9845 // truncation that comes from a setcc or select_cc.
9846 if (N->getOpcode() == ISD::TRUNCATE &&
9847 N->getValueType(0) != MVT::i1)
9850 if (N->getOperand(0).getValueType() != MVT::i32 &&
9851 N->getOperand(0).getValueType() != MVT::i64)
9854 if (N->getOpcode() == ISD::SETCC ||
9855 N->getOpcode() == ISD::SELECT_CC) {
9856 // If we're looking at a comparison, then we need to make sure that the
9857 // high bits (all except for the first) don't matter the result.
9859 cast<CondCodeSDNode>(N->getOperand(
9860 N->getOpcode() == ISD::SETCC ? 2 : 4))->get();
9861 unsigned OpBits = N->getOperand(0).getValueSizeInBits();
9863 if (ISD::isSignedIntSetCC(CC)) {
9864 if (DAG.ComputeNumSignBits(N->getOperand(0)) != OpBits ||
9865 DAG.ComputeNumSignBits(N->getOperand(1)) != OpBits)
9867 } else if (ISD::isUnsignedIntSetCC(CC)) {
9868 if (!DAG.MaskedValueIsZero(N->getOperand(0),
9869 APInt::getHighBitsSet(OpBits, OpBits-1)) ||
9870 !DAG.MaskedValueIsZero(N->getOperand(1),
9871 APInt::getHighBitsSet(OpBits, OpBits-1)))
9874 // This is neither a signed nor an unsigned comparison, just make sure
9875 // that the high bits are equal.
9876 APInt Op1Zero, Op1One;
9877 APInt Op2Zero, Op2One;
9878 DAG.computeKnownBits(N->getOperand(0), Op1Zero, Op1One);
9879 DAG.computeKnownBits(N->getOperand(1), Op2Zero, Op2One);
9881 // We don't really care about what is known about the first bit (if
9882 // anything), so clear it in all masks prior to comparing them.
9883 Op1Zero.clearBit(0); Op1One.clearBit(0);
9884 Op2Zero.clearBit(0); Op2One.clearBit(0);
9886 if (Op1Zero != Op2Zero || Op1One != Op2One)
9891 // We now know that the higher-order bits are irrelevant, we just need to
9892 // make sure that all of the intermediate operations are bit operations, and
9893 // all inputs are extensions.
9894 if (N->getOperand(0).getOpcode() != ISD::AND &&
9895 N->getOperand(0).getOpcode() != ISD::OR &&
9896 N->getOperand(0).getOpcode() != ISD::XOR &&
9897 N->getOperand(0).getOpcode() != ISD::SELECT &&
9898 N->getOperand(0).getOpcode() != ISD::SELECT_CC &&
9899 N->getOperand(0).getOpcode() != ISD::TRUNCATE &&
9900 N->getOperand(0).getOpcode() != ISD::SIGN_EXTEND &&
9901 N->getOperand(0).getOpcode() != ISD::ZERO_EXTEND &&
9902 N->getOperand(0).getOpcode() != ISD::ANY_EXTEND)
9905 if ((N->getOpcode() == ISD::SETCC || N->getOpcode() == ISD::SELECT_CC) &&
9906 N->getOperand(1).getOpcode() != ISD::AND &&
9907 N->getOperand(1).getOpcode() != ISD::OR &&
9908 N->getOperand(1).getOpcode() != ISD::XOR &&
9909 N->getOperand(1).getOpcode() != ISD::SELECT &&
9910 N->getOperand(1).getOpcode() != ISD::SELECT_CC &&
9911 N->getOperand(1).getOpcode() != ISD::TRUNCATE &&
9912 N->getOperand(1).getOpcode() != ISD::SIGN_EXTEND &&
9913 N->getOperand(1).getOpcode() != ISD::ZERO_EXTEND &&
9914 N->getOperand(1).getOpcode() != ISD::ANY_EXTEND)
9917 SmallVector<SDValue, 4> Inputs;
9918 SmallVector<SDValue, 8> BinOps, PromOps;
9919 SmallPtrSet<SDNode *, 16> Visited;
9921 for (unsigned i = 0; i < 2; ++i) {
9922 if (((N->getOperand(i).getOpcode() == ISD::SIGN_EXTEND ||
9923 N->getOperand(i).getOpcode() == ISD::ZERO_EXTEND ||
9924 N->getOperand(i).getOpcode() == ISD::ANY_EXTEND) &&
9925 N->getOperand(i).getOperand(0).getValueType() == MVT::i1) ||
9926 isa<ConstantSDNode>(N->getOperand(i)))
9927 Inputs.push_back(N->getOperand(i));
9929 BinOps.push_back(N->getOperand(i));
9931 if (N->getOpcode() == ISD::TRUNCATE)
9935 // Visit all inputs, collect all binary operations (and, or, xor and
9936 // select) that are all fed by extensions.
9937 while (!BinOps.empty()) {
9938 SDValue BinOp = BinOps.back();
9941 if (!Visited.insert(BinOp.getNode()).second)
9944 PromOps.push_back(BinOp);
9946 for (unsigned i = 0, ie = BinOp.getNumOperands(); i != ie; ++i) {
9947 // The condition of the select is not promoted.
9948 if (BinOp.getOpcode() == ISD::SELECT && i == 0)
9950 if (BinOp.getOpcode() == ISD::SELECT_CC && i != 2 && i != 3)
9953 if (((BinOp.getOperand(i).getOpcode() == ISD::SIGN_EXTEND ||
9954 BinOp.getOperand(i).getOpcode() == ISD::ZERO_EXTEND ||
9955 BinOp.getOperand(i).getOpcode() == ISD::ANY_EXTEND) &&
9956 BinOp.getOperand(i).getOperand(0).getValueType() == MVT::i1) ||
9957 isa<ConstantSDNode>(BinOp.getOperand(i))) {
9958 Inputs.push_back(BinOp.getOperand(i));
9959 } else if (BinOp.getOperand(i).getOpcode() == ISD::AND ||
9960 BinOp.getOperand(i).getOpcode() == ISD::OR ||
9961 BinOp.getOperand(i).getOpcode() == ISD::XOR ||
9962 BinOp.getOperand(i).getOpcode() == ISD::SELECT ||
9963 BinOp.getOperand(i).getOpcode() == ISD::SELECT_CC ||
9964 BinOp.getOperand(i).getOpcode() == ISD::TRUNCATE ||
9965 BinOp.getOperand(i).getOpcode() == ISD::SIGN_EXTEND ||
9966 BinOp.getOperand(i).getOpcode() == ISD::ZERO_EXTEND ||
9967 BinOp.getOperand(i).getOpcode() == ISD::ANY_EXTEND) {
9968 BinOps.push_back(BinOp.getOperand(i));
9970 // We have an input that is not an extension or another binary
9971 // operation; we'll abort this transformation.
9977 // Make sure that this is a self-contained cluster of operations (which
9978 // is not quite the same thing as saying that everything has only one
9980 for (unsigned i = 0, ie = Inputs.size(); i != ie; ++i) {
9981 if (isa<ConstantSDNode>(Inputs[i]))
9984 for (SDNode::use_iterator UI = Inputs[i].getNode()->use_begin(),
9985 UE = Inputs[i].getNode()->use_end();
9988 if (User != N && !Visited.count(User))
9991 // Make sure that we're not going to promote the non-output-value
9992 // operand(s) or SELECT or SELECT_CC.
9993 // FIXME: Although we could sometimes handle this, and it does occur in
9994 // practice that one of the condition inputs to the select is also one of
9995 // the outputs, we currently can't deal with this.
9996 if (User->getOpcode() == ISD::SELECT) {
9997 if (User->getOperand(0) == Inputs[i])
9999 } else if (User->getOpcode() == ISD::SELECT_CC) {
10000 if (User->getOperand(0) == Inputs[i] ||
10001 User->getOperand(1) == Inputs[i])
10007 for (unsigned i = 0, ie = PromOps.size(); i != ie; ++i) {
10008 for (SDNode::use_iterator UI = PromOps[i].getNode()->use_begin(),
10009 UE = PromOps[i].getNode()->use_end();
10011 SDNode *User = *UI;
10012 if (User != N && !Visited.count(User))
10015 // Make sure that we're not going to promote the non-output-value
10016 // operand(s) or SELECT or SELECT_CC.
10017 // FIXME: Although we could sometimes handle this, and it does occur in
10018 // practice that one of the condition inputs to the select is also one of
10019 // the outputs, we currently can't deal with this.
10020 if (User->getOpcode() == ISD::SELECT) {
10021 if (User->getOperand(0) == PromOps[i])
10023 } else if (User->getOpcode() == ISD::SELECT_CC) {
10024 if (User->getOperand(0) == PromOps[i] ||
10025 User->getOperand(1) == PromOps[i])
10031 // Replace all inputs with the extension operand.
10032 for (unsigned i = 0, ie = Inputs.size(); i != ie; ++i) {
10033 // Constants may have users outside the cluster of to-be-promoted nodes,
10034 // and so we need to replace those as we do the promotions.
10035 if (isa<ConstantSDNode>(Inputs[i]))
10038 DAG.ReplaceAllUsesOfValueWith(Inputs[i], Inputs[i].getOperand(0));
10041 std::list<HandleSDNode> PromOpHandles;
10042 for (auto &PromOp : PromOps)
10043 PromOpHandles.emplace_back(PromOp);
10045 // Replace all operations (these are all the same, but have a different
10046 // (i1) return type). DAG.getNode will validate that the types of
10047 // a binary operator match, so go through the list in reverse so that
10048 // we've likely promoted both operands first. Any intermediate truncations or
10049 // extensions disappear.
10050 while (!PromOpHandles.empty()) {
10051 SDValue PromOp = PromOpHandles.back().getValue();
10052 PromOpHandles.pop_back();
10054 if (PromOp.getOpcode() == ISD::TRUNCATE ||
10055 PromOp.getOpcode() == ISD::SIGN_EXTEND ||
10056 PromOp.getOpcode() == ISD::ZERO_EXTEND ||
10057 PromOp.getOpcode() == ISD::ANY_EXTEND) {
10058 if (!isa<ConstantSDNode>(PromOp.getOperand(0)) &&
10059 PromOp.getOperand(0).getValueType() != MVT::i1) {
10060 // The operand is not yet ready (see comment below).
10061 PromOpHandles.emplace_front(PromOp);
10065 SDValue RepValue = PromOp.getOperand(0);
10066 if (isa<ConstantSDNode>(RepValue))
10067 RepValue = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, RepValue);
10069 DAG.ReplaceAllUsesOfValueWith(PromOp, RepValue);
10074 switch (PromOp.getOpcode()) {
10075 default: C = 0; break;
10076 case ISD::SELECT: C = 1; break;
10077 case ISD::SELECT_CC: C = 2; break;
10080 if ((!isa<ConstantSDNode>(PromOp.getOperand(C)) &&
10081 PromOp.getOperand(C).getValueType() != MVT::i1) ||
10082 (!isa<ConstantSDNode>(PromOp.getOperand(C+1)) &&
10083 PromOp.getOperand(C+1).getValueType() != MVT::i1)) {
10084 // The to-be-promoted operands of this node have not yet been
10085 // promoted (this should be rare because we're going through the
10086 // list backward, but if one of the operands has several users in
10087 // this cluster of to-be-promoted nodes, it is possible).
10088 PromOpHandles.emplace_front(PromOp);
10092 SmallVector<SDValue, 3> Ops(PromOp.getNode()->op_begin(),
10093 PromOp.getNode()->op_end());
10095 // If there are any constant inputs, make sure they're replaced now.
10096 for (unsigned i = 0; i < 2; ++i)
10097 if (isa<ConstantSDNode>(Ops[C+i]))
10098 Ops[C+i] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, Ops[C+i]);
10100 DAG.ReplaceAllUsesOfValueWith(PromOp,
10101 DAG.getNode(PromOp.getOpcode(), dl, MVT::i1, Ops));
10104 // Now we're left with the initial truncation itself.
10105 if (N->getOpcode() == ISD::TRUNCATE)
10106 return N->getOperand(0);
10108 // Otherwise, this is a comparison. The operands to be compared have just
10109 // changed type (to i1), but everything else is the same.
10110 return SDValue(N, 0);
10113 SDValue PPCTargetLowering::DAGCombineExtBoolTrunc(SDNode *N,
10114 DAGCombinerInfo &DCI) const {
10115 SelectionDAG &DAG = DCI.DAG;
10118 // If we're tracking CR bits, we need to be careful that we don't have:
10119 // zext(binary-ops(trunc(x), trunc(y)))
10121 // zext(binary-ops(binary-ops(trunc(x), trunc(y)), ...)
10122 // such that we're unnecessarily moving things into CR bits that can more
10123 // efficiently stay in GPRs. Note that if we're not certain that the high
10124 // bits are set as required by the final extension, we still may need to do
10125 // some masking to get the proper behavior.
10127 // This same functionality is important on PPC64 when dealing with
10128 // 32-to-64-bit extensions; these occur often when 32-bit values are used as
10129 // the return values of functions. Because it is so similar, it is handled
10132 if (N->getValueType(0) != MVT::i32 &&
10133 N->getValueType(0) != MVT::i64)
10136 if (!((N->getOperand(0).getValueType() == MVT::i1 && Subtarget.useCRBits()) ||
10137 (N->getOperand(0).getValueType() == MVT::i32 && Subtarget.isPPC64())))
10140 if (N->getOperand(0).getOpcode() != ISD::AND &&
10141 N->getOperand(0).getOpcode() != ISD::OR &&
10142 N->getOperand(0).getOpcode() != ISD::XOR &&
10143 N->getOperand(0).getOpcode() != ISD::SELECT &&
10144 N->getOperand(0).getOpcode() != ISD::SELECT_CC)
10147 SmallVector<SDValue, 4> Inputs;
10148 SmallVector<SDValue, 8> BinOps(1, N->getOperand(0)), PromOps;
10149 SmallPtrSet<SDNode *, 16> Visited;
10151 // Visit all inputs, collect all binary operations (and, or, xor and
10152 // select) that are all fed by truncations.
10153 while (!BinOps.empty()) {
10154 SDValue BinOp = BinOps.back();
10157 if (!Visited.insert(BinOp.getNode()).second)
10160 PromOps.push_back(BinOp);
10162 for (unsigned i = 0, ie = BinOp.getNumOperands(); i != ie; ++i) {
10163 // The condition of the select is not promoted.
10164 if (BinOp.getOpcode() == ISD::SELECT && i == 0)
10166 if (BinOp.getOpcode() == ISD::SELECT_CC && i != 2 && i != 3)
10169 if (BinOp.getOperand(i).getOpcode() == ISD::TRUNCATE ||
10170 isa<ConstantSDNode>(BinOp.getOperand(i))) {
10171 Inputs.push_back(BinOp.getOperand(i));
10172 } else if (BinOp.getOperand(i).getOpcode() == ISD::AND ||
10173 BinOp.getOperand(i).getOpcode() == ISD::OR ||
10174 BinOp.getOperand(i).getOpcode() == ISD::XOR ||
10175 BinOp.getOperand(i).getOpcode() == ISD::SELECT ||
10176 BinOp.getOperand(i).getOpcode() == ISD::SELECT_CC) {
10177 BinOps.push_back(BinOp.getOperand(i));
10179 // We have an input that is not a truncation or another binary
10180 // operation; we'll abort this transformation.
10186 // The operands of a select that must be truncated when the select is
10187 // promoted because the operand is actually part of the to-be-promoted set.
10188 DenseMap<SDNode *, EVT> SelectTruncOp[2];
10190 // Make sure that this is a self-contained cluster of operations (which
10191 // is not quite the same thing as saying that everything has only one
10193 for (unsigned i = 0, ie = Inputs.size(); i != ie; ++i) {
10194 if (isa<ConstantSDNode>(Inputs[i]))
10197 for (SDNode::use_iterator UI = Inputs[i].getNode()->use_begin(),
10198 UE = Inputs[i].getNode()->use_end();
10200 SDNode *User = *UI;
10201 if (User != N && !Visited.count(User))
10204 // If we're going to promote the non-output-value operand(s) or SELECT or
10205 // SELECT_CC, record them for truncation.
10206 if (User->getOpcode() == ISD::SELECT) {
10207 if (User->getOperand(0) == Inputs[i])
10208 SelectTruncOp[0].insert(std::make_pair(User,
10209 User->getOperand(0).getValueType()));
10210 } else if (User->getOpcode() == ISD::SELECT_CC) {
10211 if (User->getOperand(0) == Inputs[i])
10212 SelectTruncOp[0].insert(std::make_pair(User,
10213 User->getOperand(0).getValueType()));
10214 if (User->getOperand(1) == Inputs[i])
10215 SelectTruncOp[1].insert(std::make_pair(User,
10216 User->getOperand(1).getValueType()));
10221 for (unsigned i = 0, ie = PromOps.size(); i != ie; ++i) {
10222 for (SDNode::use_iterator UI = PromOps[i].getNode()->use_begin(),
10223 UE = PromOps[i].getNode()->use_end();
10225 SDNode *User = *UI;
10226 if (User != N && !Visited.count(User))
10229 // If we're going to promote the non-output-value operand(s) or SELECT or
10230 // SELECT_CC, record them for truncation.
10231 if (User->getOpcode() == ISD::SELECT) {
10232 if (User->getOperand(0) == PromOps[i])
10233 SelectTruncOp[0].insert(std::make_pair(User,
10234 User->getOperand(0).getValueType()));
10235 } else if (User->getOpcode() == ISD::SELECT_CC) {
10236 if (User->getOperand(0) == PromOps[i])
10237 SelectTruncOp[0].insert(std::make_pair(User,
10238 User->getOperand(0).getValueType()));
10239 if (User->getOperand(1) == PromOps[i])
10240 SelectTruncOp[1].insert(std::make_pair(User,
10241 User->getOperand(1).getValueType()));
10246 unsigned PromBits = N->getOperand(0).getValueSizeInBits();
10247 bool ReallyNeedsExt = false;
10248 if (N->getOpcode() != ISD::ANY_EXTEND) {
10249 // If all of the inputs are not already sign/zero extended, then
10250 // we'll still need to do that at the end.
10251 for (unsigned i = 0, ie = Inputs.size(); i != ie; ++i) {
10252 if (isa<ConstantSDNode>(Inputs[i]))
10256 Inputs[i].getOperand(0).getValueSizeInBits();
10257 assert(PromBits < OpBits && "Truncation not to a smaller bit count?");
10259 if ((N->getOpcode() == ISD::ZERO_EXTEND &&
10260 !DAG.MaskedValueIsZero(Inputs[i].getOperand(0),
10261 APInt::getHighBitsSet(OpBits,
10262 OpBits-PromBits))) ||
10263 (N->getOpcode() == ISD::SIGN_EXTEND &&
10264 DAG.ComputeNumSignBits(Inputs[i].getOperand(0)) <
10265 (OpBits-(PromBits-1)))) {
10266 ReallyNeedsExt = true;
10272 // Replace all inputs, either with the truncation operand, or a
10273 // truncation or extension to the final output type.
10274 for (unsigned i = 0, ie = Inputs.size(); i != ie; ++i) {
10275 // Constant inputs need to be replaced with the to-be-promoted nodes that
10276 // use them because they might have users outside of the cluster of
10278 if (isa<ConstantSDNode>(Inputs[i]))
10281 SDValue InSrc = Inputs[i].getOperand(0);
10282 if (Inputs[i].getValueType() == N->getValueType(0))
10283 DAG.ReplaceAllUsesOfValueWith(Inputs[i], InSrc);
10284 else if (N->getOpcode() == ISD::SIGN_EXTEND)
10285 DAG.ReplaceAllUsesOfValueWith(Inputs[i],
10286 DAG.getSExtOrTrunc(InSrc, dl, N->getValueType(0)));
10287 else if (N->getOpcode() == ISD::ZERO_EXTEND)
10288 DAG.ReplaceAllUsesOfValueWith(Inputs[i],
10289 DAG.getZExtOrTrunc(InSrc, dl, N->getValueType(0)));
10291 DAG.ReplaceAllUsesOfValueWith(Inputs[i],
10292 DAG.getAnyExtOrTrunc(InSrc, dl, N->getValueType(0)));
10295 std::list<HandleSDNode> PromOpHandles;
10296 for (auto &PromOp : PromOps)
10297 PromOpHandles.emplace_back(PromOp);
10299 // Replace all operations (these are all the same, but have a different
10300 // (promoted) return type). DAG.getNode will validate that the types of
10301 // a binary operator match, so go through the list in reverse so that
10302 // we've likely promoted both operands first.
10303 while (!PromOpHandles.empty()) {
10304 SDValue PromOp = PromOpHandles.back().getValue();
10305 PromOpHandles.pop_back();
10308 switch (PromOp.getOpcode()) {
10309 default: C = 0; break;
10310 case ISD::SELECT: C = 1; break;
10311 case ISD::SELECT_CC: C = 2; break;
10314 if ((!isa<ConstantSDNode>(PromOp.getOperand(C)) &&
10315 PromOp.getOperand(C).getValueType() != N->getValueType(0)) ||
10316 (!isa<ConstantSDNode>(PromOp.getOperand(C+1)) &&
10317 PromOp.getOperand(C+1).getValueType() != N->getValueType(0))) {
10318 // The to-be-promoted operands of this node have not yet been
10319 // promoted (this should be rare because we're going through the
10320 // list backward, but if one of the operands has several users in
10321 // this cluster of to-be-promoted nodes, it is possible).
10322 PromOpHandles.emplace_front(PromOp);
10326 // For SELECT and SELECT_CC nodes, we do a similar check for any
10327 // to-be-promoted comparison inputs.
10328 if (PromOp.getOpcode() == ISD::SELECT ||
10329 PromOp.getOpcode() == ISD::SELECT_CC) {
10330 if ((SelectTruncOp[0].count(PromOp.getNode()) &&
10331 PromOp.getOperand(0).getValueType() != N->getValueType(0)) ||
10332 (SelectTruncOp[1].count(PromOp.getNode()) &&
10333 PromOp.getOperand(1).getValueType() != N->getValueType(0))) {
10334 PromOpHandles.emplace_front(PromOp);
10339 SmallVector<SDValue, 3> Ops(PromOp.getNode()->op_begin(),
10340 PromOp.getNode()->op_end());
10342 // If this node has constant inputs, then they'll need to be promoted here.
10343 for (unsigned i = 0; i < 2; ++i) {
10344 if (!isa<ConstantSDNode>(Ops[C+i]))
10346 if (Ops[C+i].getValueType() == N->getValueType(0))
10349 if (N->getOpcode() == ISD::SIGN_EXTEND)
10350 Ops[C+i] = DAG.getSExtOrTrunc(Ops[C+i], dl, N->getValueType(0));
10351 else if (N->getOpcode() == ISD::ZERO_EXTEND)
10352 Ops[C+i] = DAG.getZExtOrTrunc(Ops[C+i], dl, N->getValueType(0));
10354 Ops[C+i] = DAG.getAnyExtOrTrunc(Ops[C+i], dl, N->getValueType(0));
10357 // If we've promoted the comparison inputs of a SELECT or SELECT_CC,
10358 // truncate them again to the original value type.
10359 if (PromOp.getOpcode() == ISD::SELECT ||
10360 PromOp.getOpcode() == ISD::SELECT_CC) {
10361 auto SI0 = SelectTruncOp[0].find(PromOp.getNode());
10362 if (SI0 != SelectTruncOp[0].end())
10363 Ops[0] = DAG.getNode(ISD::TRUNCATE, dl, SI0->second, Ops[0]);
10364 auto SI1 = SelectTruncOp[1].find(PromOp.getNode());
10365 if (SI1 != SelectTruncOp[1].end())
10366 Ops[1] = DAG.getNode(ISD::TRUNCATE, dl, SI1->second, Ops[1]);
10369 DAG.ReplaceAllUsesOfValueWith(PromOp,
10370 DAG.getNode(PromOp.getOpcode(), dl, N->getValueType(0), Ops));
10373 // Now we're left with the initial extension itself.
10374 if (!ReallyNeedsExt)
10375 return N->getOperand(0);
10377 // To zero extend, just mask off everything except for the first bit (in the
10379 if (N->getOpcode() == ISD::ZERO_EXTEND)
10380 return DAG.getNode(ISD::AND, dl, N->getValueType(0), N->getOperand(0),
10381 DAG.getConstant(APInt::getLowBitsSet(
10382 N->getValueSizeInBits(0), PromBits),
10383 dl, N->getValueType(0)));
10385 assert(N->getOpcode() == ISD::SIGN_EXTEND &&
10386 "Invalid extension type");
10387 EVT ShiftAmountTy = getShiftAmountTy(N->getValueType(0), DAG.getDataLayout());
10389 DAG.getConstant(N->getValueSizeInBits(0) - PromBits, dl, ShiftAmountTy);
10390 return DAG.getNode(
10391 ISD::SRA, dl, N->getValueType(0),
10392 DAG.getNode(ISD::SHL, dl, N->getValueType(0), N->getOperand(0), ShiftCst),
10396 SDValue PPCTargetLowering::DAGCombineBuildVector(SDNode *N,
10397 DAGCombinerInfo &DCI) const {
10398 assert(N->getOpcode() == ISD::BUILD_VECTOR &&
10399 "Should be called with a BUILD_VECTOR node");
10401 SelectionDAG &DAG = DCI.DAG;
10403 if (N->getValueType(0) != MVT::v2f64 || !Subtarget.hasVSX())
10407 // (build_vector ([su]int_to_fp (extractelt 0)), [su]int_to_fp (extractelt 1))
10408 if (N->getOperand(0).getOpcode() != ISD::SINT_TO_FP &&
10409 N->getOperand(0).getOpcode() != ISD::UINT_TO_FP)
10411 if (N->getOperand(1).getOpcode() != ISD::SINT_TO_FP &&
10412 N->getOperand(1).getOpcode() != ISD::UINT_TO_FP)
10414 if (N->getOperand(0).getOpcode() != N->getOperand(1).getOpcode())
10417 SDValue Ext1 = N->getOperand(0).getOperand(0);
10418 SDValue Ext2 = N->getOperand(1).getOperand(0);
10419 if(Ext1.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
10420 Ext2.getOpcode() != ISD::EXTRACT_VECTOR_ELT)
10423 ConstantSDNode *Ext1Op = dyn_cast<ConstantSDNode>(Ext1.getOperand(1));
10424 ConstantSDNode *Ext2Op = dyn_cast<ConstantSDNode>(Ext2.getOperand(1));
10425 if (!Ext1Op || !Ext2Op)
10427 if (Ext1.getValueType() != MVT::i32 ||
10428 Ext2.getValueType() != MVT::i32)
10429 if (Ext1.getOperand(0) != Ext2.getOperand(0))
10432 int FirstElem = Ext1Op->getZExtValue();
10433 int SecondElem = Ext2Op->getZExtValue();
10435 if (FirstElem == 0 && SecondElem == 1)
10436 SubvecIdx = Subtarget.isLittleEndian() ? 1 : 0;
10437 else if (FirstElem == 2 && SecondElem == 3)
10438 SubvecIdx = Subtarget.isLittleEndian() ? 0 : 1;
10442 SDValue SrcVec = Ext1.getOperand(0);
10443 auto NodeType = (N->getOperand(1).getOpcode() == ISD::SINT_TO_FP) ?
10444 PPCISD::SINT_VEC_TO_FP : PPCISD::UINT_VEC_TO_FP;
10445 return DAG.getNode(NodeType, dl, MVT::v2f64,
10446 SrcVec, DAG.getIntPtrConstant(SubvecIdx, dl));
10449 SDValue PPCTargetLowering::combineFPToIntToFP(SDNode *N,
10450 DAGCombinerInfo &DCI) const {
10451 assert((N->getOpcode() == ISD::SINT_TO_FP ||
10452 N->getOpcode() == ISD::UINT_TO_FP) &&
10453 "Need an int -> FP conversion node here");
10455 if (!Subtarget.has64BitSupport())
10458 SelectionDAG &DAG = DCI.DAG;
10462 // Don't handle ppc_fp128 here or i1 conversions.
10463 if (Op.getValueType() != MVT::f32 && Op.getValueType() != MVT::f64)
10465 if (Op.getOperand(0).getValueType() == MVT::i1)
10468 // For i32 intermediate values, unfortunately, the conversion functions
10469 // leave the upper 32 bits of the value are undefined. Within the set of
10470 // scalar instructions, we have no method for zero- or sign-extending the
10471 // value. Thus, we cannot handle i32 intermediate values here.
10472 if (Op.getOperand(0).getValueType() == MVT::i32)
10475 assert((Op.getOpcode() == ISD::SINT_TO_FP || Subtarget.hasFPCVT()) &&
10476 "UINT_TO_FP is supported only with FPCVT");
10478 // If we have FCFIDS, then use it when converting to single-precision.
10479 // Otherwise, convert to double-precision and then round.
10480 unsigned FCFOp = (Subtarget.hasFPCVT() && Op.getValueType() == MVT::f32)
10481 ? (Op.getOpcode() == ISD::UINT_TO_FP ? PPCISD::FCFIDUS
10483 : (Op.getOpcode() == ISD::UINT_TO_FP ? PPCISD::FCFIDU
10485 MVT FCFTy = (Subtarget.hasFPCVT() && Op.getValueType() == MVT::f32)
10489 // If we're converting from a float, to an int, and back to a float again,
10490 // then we don't need the store/load pair at all.
10491 if ((Op.getOperand(0).getOpcode() == ISD::FP_TO_UINT &&
10492 Subtarget.hasFPCVT()) ||
10493 (Op.getOperand(0).getOpcode() == ISD::FP_TO_SINT)) {
10494 SDValue Src = Op.getOperand(0).getOperand(0);
10495 if (Src.getValueType() == MVT::f32) {
10496 Src = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Src);
10497 DCI.AddToWorklist(Src.getNode());
10498 } else if (Src.getValueType() != MVT::f64) {
10499 // Make sure that we don't pick up a ppc_fp128 source value.
10504 Op.getOperand(0).getOpcode() == ISD::FP_TO_SINT ? PPCISD::FCTIDZ :
10507 SDValue Tmp = DAG.getNode(FCTOp, dl, MVT::f64, Src);
10508 SDValue FP = DAG.getNode(FCFOp, dl, FCFTy, Tmp);
10510 if (Op.getValueType() == MVT::f32 && !Subtarget.hasFPCVT()) {
10511 FP = DAG.getNode(ISD::FP_ROUND, dl,
10512 MVT::f32, FP, DAG.getIntPtrConstant(0, dl));
10513 DCI.AddToWorklist(FP.getNode());
10522 // expandVSXLoadForLE - Convert VSX loads (which may be intrinsics for
10523 // builtins) into loads with swaps.
10524 SDValue PPCTargetLowering::expandVSXLoadForLE(SDNode *N,
10525 DAGCombinerInfo &DCI) const {
10526 SelectionDAG &DAG = DCI.DAG;
10530 MachineMemOperand *MMO;
10532 switch (N->getOpcode()) {
10534 llvm_unreachable("Unexpected opcode for little endian VSX load");
10536 LoadSDNode *LD = cast<LoadSDNode>(N);
10537 Chain = LD->getChain();
10538 Base = LD->getBasePtr();
10539 MMO = LD->getMemOperand();
10540 // If the MMO suggests this isn't a load of a full vector, leave
10541 // things alone. For a built-in, we have to make the change for
10542 // correctness, so if there is a size problem that will be a bug.
10543 if (MMO->getSize() < 16)
10547 case ISD::INTRINSIC_W_CHAIN: {
10548 MemIntrinsicSDNode *Intrin = cast<MemIntrinsicSDNode>(N);
10549 Chain = Intrin->getChain();
10550 // Similarly to the store case below, Intrin->getBasePtr() doesn't get
10551 // us what we want. Get operand 2 instead.
10552 Base = Intrin->getOperand(2);
10553 MMO = Intrin->getMemOperand();
10558 MVT VecTy = N->getValueType(0).getSimpleVT();
10559 SDValue LoadOps[] = { Chain, Base };
10560 SDValue Load = DAG.getMemIntrinsicNode(PPCISD::LXVD2X, dl,
10561 DAG.getVTList(MVT::v2f64, MVT::Other),
10562 LoadOps, MVT::v2f64, MMO);
10564 DCI.AddToWorklist(Load.getNode());
10565 Chain = Load.getValue(1);
10566 SDValue Swap = DAG.getNode(
10567 PPCISD::XXSWAPD, dl, DAG.getVTList(MVT::v2f64, MVT::Other), Chain, Load);
10568 DCI.AddToWorklist(Swap.getNode());
10570 // Add a bitcast if the resulting load type doesn't match v2f64.
10571 if (VecTy != MVT::v2f64) {
10572 SDValue N = DAG.getNode(ISD::BITCAST, dl, VecTy, Swap);
10573 DCI.AddToWorklist(N.getNode());
10574 // Package {bitcast value, swap's chain} to match Load's shape.
10575 return DAG.getNode(ISD::MERGE_VALUES, dl, DAG.getVTList(VecTy, MVT::Other),
10576 N, Swap.getValue(1));
10582 // expandVSXStoreForLE - Convert VSX stores (which may be intrinsics for
10583 // builtins) into stores with swaps.
10584 SDValue PPCTargetLowering::expandVSXStoreForLE(SDNode *N,
10585 DAGCombinerInfo &DCI) const {
10586 SelectionDAG &DAG = DCI.DAG;
10591 MachineMemOperand *MMO;
10593 switch (N->getOpcode()) {
10595 llvm_unreachable("Unexpected opcode for little endian VSX store");
10597 StoreSDNode *ST = cast<StoreSDNode>(N);
10598 Chain = ST->getChain();
10599 Base = ST->getBasePtr();
10600 MMO = ST->getMemOperand();
10602 // If the MMO suggests this isn't a store of a full vector, leave
10603 // things alone. For a built-in, we have to make the change for
10604 // correctness, so if there is a size problem that will be a bug.
10605 if (MMO->getSize() < 16)
10609 case ISD::INTRINSIC_VOID: {
10610 MemIntrinsicSDNode *Intrin = cast<MemIntrinsicSDNode>(N);
10611 Chain = Intrin->getChain();
10612 // Intrin->getBasePtr() oddly does not get what we want.
10613 Base = Intrin->getOperand(3);
10614 MMO = Intrin->getMemOperand();
10620 SDValue Src = N->getOperand(SrcOpnd);
10621 MVT VecTy = Src.getValueType().getSimpleVT();
10623 // All stores are done as v2f64 and possible bit cast.
10624 if (VecTy != MVT::v2f64) {
10625 Src = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Src);
10626 DCI.AddToWorklist(Src.getNode());
10629 SDValue Swap = DAG.getNode(PPCISD::XXSWAPD, dl,
10630 DAG.getVTList(MVT::v2f64, MVT::Other), Chain, Src);
10631 DCI.AddToWorklist(Swap.getNode());
10632 Chain = Swap.getValue(1);
10633 SDValue StoreOps[] = { Chain, Swap, Base };
10634 SDValue Store = DAG.getMemIntrinsicNode(PPCISD::STXVD2X, dl,
10635 DAG.getVTList(MVT::Other),
10636 StoreOps, VecTy, MMO);
10637 DCI.AddToWorklist(Store.getNode());
10641 SDValue PPCTargetLowering::PerformDAGCombine(SDNode *N,
10642 DAGCombinerInfo &DCI) const {
10643 SelectionDAG &DAG = DCI.DAG;
10645 switch (N->getOpcode()) {
10648 if (isNullConstant(N->getOperand(0))) // 0 << V -> 0.
10649 return N->getOperand(0);
10652 if (isNullConstant(N->getOperand(0))) // 0 >>u V -> 0.
10653 return N->getOperand(0);
10656 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
10657 if (C->isNullValue() || // 0 >>s V -> 0.
10658 C->isAllOnesValue()) // -1 >>s V -> -1.
10659 return N->getOperand(0);
10662 case ISD::SIGN_EXTEND:
10663 case ISD::ZERO_EXTEND:
10664 case ISD::ANY_EXTEND:
10665 return DAGCombineExtBoolTrunc(N, DCI);
10666 case ISD::TRUNCATE:
10668 case ISD::SELECT_CC:
10669 return DAGCombineTruncBoolExt(N, DCI);
10670 case ISD::SINT_TO_FP:
10671 case ISD::UINT_TO_FP:
10672 return combineFPToIntToFP(N, DCI);
10674 // Turn STORE (FP_TO_SINT F) -> STFIWX(FCTIWZ(F)).
10675 if (Subtarget.hasSTFIWX() && !cast<StoreSDNode>(N)->isTruncatingStore() &&
10676 N->getOperand(1).getOpcode() == ISD::FP_TO_SINT &&
10677 N->getOperand(1).getValueType() == MVT::i32 &&
10678 N->getOperand(1).getOperand(0).getValueType() != MVT::ppcf128) {
10679 SDValue Val = N->getOperand(1).getOperand(0);
10680 if (Val.getValueType() == MVT::f32) {
10681 Val = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Val);
10682 DCI.AddToWorklist(Val.getNode());
10684 Val = DAG.getNode(PPCISD::FCTIWZ, dl, MVT::f64, Val);
10685 DCI.AddToWorklist(Val.getNode());
10688 N->getOperand(0), Val, N->getOperand(2),
10689 DAG.getValueType(N->getOperand(1).getValueType())
10692 Val = DAG.getMemIntrinsicNode(PPCISD::STFIWX, dl,
10693 DAG.getVTList(MVT::Other), Ops,
10694 cast<StoreSDNode>(N)->getMemoryVT(),
10695 cast<StoreSDNode>(N)->getMemOperand());
10696 DCI.AddToWorklist(Val.getNode());
10700 // Turn STORE (BSWAP) -> sthbrx/stwbrx.
10701 if (cast<StoreSDNode>(N)->isUnindexed() &&
10702 N->getOperand(1).getOpcode() == ISD::BSWAP &&
10703 N->getOperand(1).getNode()->hasOneUse() &&
10704 (N->getOperand(1).getValueType() == MVT::i32 ||
10705 N->getOperand(1).getValueType() == MVT::i16 ||
10706 (Subtarget.hasLDBRX() && Subtarget.isPPC64() &&
10707 N->getOperand(1).getValueType() == MVT::i64))) {
10708 SDValue BSwapOp = N->getOperand(1).getOperand(0);
10709 // Do an any-extend to 32-bits if this is a half-word input.
10710 if (BSwapOp.getValueType() == MVT::i16)
10711 BSwapOp = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, BSwapOp);
10714 N->getOperand(0), BSwapOp, N->getOperand(2),
10715 DAG.getValueType(N->getOperand(1).getValueType())
10718 DAG.getMemIntrinsicNode(PPCISD::STBRX, dl, DAG.getVTList(MVT::Other),
10719 Ops, cast<StoreSDNode>(N)->getMemoryVT(),
10720 cast<StoreSDNode>(N)->getMemOperand());
10723 // For little endian, VSX stores require generating xxswapd/lxvd2x.
10724 EVT VT = N->getOperand(1).getValueType();
10725 if (VT.isSimple()) {
10726 MVT StoreVT = VT.getSimpleVT();
10727 if (Subtarget.hasVSX() && Subtarget.isLittleEndian() &&
10728 (StoreVT == MVT::v2f64 || StoreVT == MVT::v2i64 ||
10729 StoreVT == MVT::v4f32 || StoreVT == MVT::v4i32))
10730 return expandVSXStoreForLE(N, DCI);
10735 LoadSDNode *LD = cast<LoadSDNode>(N);
10736 EVT VT = LD->getValueType(0);
10738 // For little endian, VSX loads require generating lxvd2x/xxswapd.
10739 if (VT.isSimple()) {
10740 MVT LoadVT = VT.getSimpleVT();
10741 if (Subtarget.hasVSX() && Subtarget.isLittleEndian() &&
10742 (LoadVT == MVT::v2f64 || LoadVT == MVT::v2i64 ||
10743 LoadVT == MVT::v4f32 || LoadVT == MVT::v4i32))
10744 return expandVSXLoadForLE(N, DCI);
10747 // We sometimes end up with a 64-bit integer load, from which we extract
10748 // two single-precision floating-point numbers. This happens with
10749 // std::complex<float>, and other similar structures, because of the way we
10750 // canonicalize structure copies. However, if we lack direct moves,
10751 // then the final bitcasts from the extracted integer values to the
10752 // floating-point numbers turn into store/load pairs. Even with direct moves,
10753 // just loading the two floating-point numbers is likely better.
10754 auto ReplaceTwoFloatLoad = [&]() {
10755 if (VT != MVT::i64)
10758 if (LD->getExtensionType() != ISD::NON_EXTLOAD ||
10762 // We're looking for a sequence like this:
10763 // t13: i64,ch = load<LD8[%ref.tmp]> t0, t6, undef:i64
10764 // t16: i64 = srl t13, Constant:i32<32>
10765 // t17: i32 = truncate t16
10766 // t18: f32 = bitcast t17
10767 // t19: i32 = truncate t13
10768 // t20: f32 = bitcast t19
10770 if (!LD->hasNUsesOfValue(2, 0))
10773 auto UI = LD->use_begin();
10774 while (UI.getUse().getResNo() != 0) ++UI;
10775 SDNode *Trunc = *UI++;
10776 while (UI.getUse().getResNo() != 0) ++UI;
10777 SDNode *RightShift = *UI;
10778 if (Trunc->getOpcode() != ISD::TRUNCATE)
10779 std::swap(Trunc, RightShift);
10781 if (Trunc->getOpcode() != ISD::TRUNCATE ||
10782 Trunc->getValueType(0) != MVT::i32 ||
10783 !Trunc->hasOneUse())
10785 if (RightShift->getOpcode() != ISD::SRL ||
10786 !isa<ConstantSDNode>(RightShift->getOperand(1)) ||
10787 RightShift->getConstantOperandVal(1) != 32 ||
10788 !RightShift->hasOneUse())
10791 SDNode *Trunc2 = *RightShift->use_begin();
10792 if (Trunc2->getOpcode() != ISD::TRUNCATE ||
10793 Trunc2->getValueType(0) != MVT::i32 ||
10794 !Trunc2->hasOneUse())
10797 SDNode *Bitcast = *Trunc->use_begin();
10798 SDNode *Bitcast2 = *Trunc2->use_begin();
10800 if (Bitcast->getOpcode() != ISD::BITCAST ||
10801 Bitcast->getValueType(0) != MVT::f32)
10803 if (Bitcast2->getOpcode() != ISD::BITCAST ||
10804 Bitcast2->getValueType(0) != MVT::f32)
10807 if (Subtarget.isLittleEndian())
10808 std::swap(Bitcast, Bitcast2);
10810 // Bitcast has the second float (in memory-layout order) and Bitcast2
10811 // has the first one.
10813 SDValue BasePtr = LD->getBasePtr();
10814 if (LD->isIndexed()) {
10815 assert(LD->getAddressingMode() == ISD::PRE_INC &&
10816 "Non-pre-inc AM on PPC?");
10818 DAG.getNode(ISD::ADD, dl, BasePtr.getValueType(), BasePtr,
10823 LD->getMemOperand()->getFlags() & ~MachineMemOperand::MOVolatile;
10824 SDValue FloatLoad = DAG.getLoad(MVT::f32, dl, LD->getChain(), BasePtr,
10825 LD->getPointerInfo(), LD->getAlignment(),
10826 MMOFlags, LD->getAAInfo());
10828 DAG.getNode(ISD::ADD, dl, BasePtr.getValueType(),
10829 BasePtr, DAG.getIntPtrConstant(4, dl));
10830 SDValue FloatLoad2 = DAG.getLoad(
10831 MVT::f32, dl, SDValue(FloatLoad.getNode(), 1), AddPtr,
10832 LD->getPointerInfo().getWithOffset(4),
10833 MinAlign(LD->getAlignment(), 4), MMOFlags, LD->getAAInfo());
10835 if (LD->isIndexed()) {
10836 // Note that DAGCombine should re-form any pre-increment load(s) from
10837 // what is produced here if that makes sense.
10838 DAG.ReplaceAllUsesOfValueWith(SDValue(LD, 1), BasePtr);
10841 DCI.CombineTo(Bitcast2, FloatLoad);
10842 DCI.CombineTo(Bitcast, FloatLoad2);
10844 DAG.ReplaceAllUsesOfValueWith(SDValue(LD, LD->isIndexed() ? 2 : 1),
10845 SDValue(FloatLoad2.getNode(), 1));
10849 if (ReplaceTwoFloatLoad())
10850 return SDValue(N, 0);
10852 EVT MemVT = LD->getMemoryVT();
10853 Type *Ty = MemVT.getTypeForEVT(*DAG.getContext());
10854 unsigned ABIAlignment = DAG.getDataLayout().getABITypeAlignment(Ty);
10855 Type *STy = MemVT.getScalarType().getTypeForEVT(*DAG.getContext());
10856 unsigned ScalarABIAlignment = DAG.getDataLayout().getABITypeAlignment(STy);
10857 if (LD->isUnindexed() && VT.isVector() &&
10858 ((Subtarget.hasAltivec() && ISD::isNON_EXTLoad(N) &&
10859 // P8 and later hardware should just use LOAD.
10860 !Subtarget.hasP8Vector() && (VT == MVT::v16i8 || VT == MVT::v8i16 ||
10861 VT == MVT::v4i32 || VT == MVT::v4f32)) ||
10862 (Subtarget.hasQPX() && (VT == MVT::v4f64 || VT == MVT::v4f32) &&
10863 LD->getAlignment() >= ScalarABIAlignment)) &&
10864 LD->getAlignment() < ABIAlignment) {
10865 // This is a type-legal unaligned Altivec or QPX load.
10866 SDValue Chain = LD->getChain();
10867 SDValue Ptr = LD->getBasePtr();
10868 bool isLittleEndian = Subtarget.isLittleEndian();
10870 // This implements the loading of unaligned vectors as described in
10871 // the venerable Apple Velocity Engine overview. Specifically:
10872 // https://developer.apple.com/hardwaredrivers/ve/alignment.html
10873 // https://developer.apple.com/hardwaredrivers/ve/code_optimization.html
10875 // The general idea is to expand a sequence of one or more unaligned
10876 // loads into an alignment-based permutation-control instruction (lvsl
10877 // or lvsr), a series of regular vector loads (which always truncate
10878 // their input address to an aligned address), and a series of
10879 // permutations. The results of these permutations are the requested
10880 // loaded values. The trick is that the last "extra" load is not taken
10881 // from the address you might suspect (sizeof(vector) bytes after the
10882 // last requested load), but rather sizeof(vector) - 1 bytes after the
10883 // last requested vector. The point of this is to avoid a page fault if
10884 // the base address happened to be aligned. This works because if the
10885 // base address is aligned, then adding less than a full vector length
10886 // will cause the last vector in the sequence to be (re)loaded.
10887 // Otherwise, the next vector will be fetched as you might suspect was
10890 // We might be able to reuse the permutation generation from
10891 // a different base address offset from this one by an aligned amount.
10892 // The INTRINSIC_WO_CHAIN DAG combine will attempt to perform this
10893 // optimization later.
10894 Intrinsic::ID Intr, IntrLD, IntrPerm;
10895 MVT PermCntlTy, PermTy, LDTy;
10896 if (Subtarget.hasAltivec()) {
10897 Intr = isLittleEndian ? Intrinsic::ppc_altivec_lvsr :
10898 Intrinsic::ppc_altivec_lvsl;
10899 IntrLD = Intrinsic::ppc_altivec_lvx;
10900 IntrPerm = Intrinsic::ppc_altivec_vperm;
10901 PermCntlTy = MVT::v16i8;
10902 PermTy = MVT::v4i32;
10905 Intr = MemVT == MVT::v4f64 ? Intrinsic::ppc_qpx_qvlpcld :
10906 Intrinsic::ppc_qpx_qvlpcls;
10907 IntrLD = MemVT == MVT::v4f64 ? Intrinsic::ppc_qpx_qvlfd :
10908 Intrinsic::ppc_qpx_qvlfs;
10909 IntrPerm = Intrinsic::ppc_qpx_qvfperm;
10910 PermCntlTy = MVT::v4f64;
10911 PermTy = MVT::v4f64;
10912 LDTy = MemVT.getSimpleVT();
10915 SDValue PermCntl = BuildIntrinsicOp(Intr, Ptr, DAG, dl, PermCntlTy);
10917 // Create the new MMO for the new base load. It is like the original MMO,
10918 // but represents an area in memory almost twice the vector size centered
10919 // on the original address. If the address is unaligned, we might start
10920 // reading up to (sizeof(vector)-1) bytes below the address of the
10921 // original unaligned load.
10922 MachineFunction &MF = DAG.getMachineFunction();
10923 MachineMemOperand *BaseMMO =
10924 MF.getMachineMemOperand(LD->getMemOperand(),
10925 -(long)MemVT.getStoreSize()+1,
10926 2*MemVT.getStoreSize()-1);
10928 // Create the new base load.
10930 DAG.getTargetConstant(IntrLD, dl, getPointerTy(MF.getDataLayout()));
10931 SDValue BaseLoadOps[] = { Chain, LDXIntID, Ptr };
10933 DAG.getMemIntrinsicNode(ISD::INTRINSIC_W_CHAIN, dl,
10934 DAG.getVTList(PermTy, MVT::Other),
10935 BaseLoadOps, LDTy, BaseMMO);
10937 // Note that the value of IncOffset (which is provided to the next
10938 // load's pointer info offset value, and thus used to calculate the
10939 // alignment), and the value of IncValue (which is actually used to
10940 // increment the pointer value) are different! This is because we
10941 // require the next load to appear to be aligned, even though it
10942 // is actually offset from the base pointer by a lesser amount.
10943 int IncOffset = VT.getSizeInBits() / 8;
10944 int IncValue = IncOffset;
10946 // Walk (both up and down) the chain looking for another load at the real
10947 // (aligned) offset (the alignment of the other load does not matter in
10948 // this case). If found, then do not use the offset reduction trick, as
10949 // that will prevent the loads from being later combined (as they would
10950 // otherwise be duplicates).
10951 if (!findConsecutiveLoad(LD, DAG))
10954 SDValue Increment =
10955 DAG.getConstant(IncValue, dl, getPointerTy(MF.getDataLayout()));
10956 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
10958 MachineMemOperand *ExtraMMO =
10959 MF.getMachineMemOperand(LD->getMemOperand(),
10960 1, 2*MemVT.getStoreSize()-1);
10961 SDValue ExtraLoadOps[] = { Chain, LDXIntID, Ptr };
10962 SDValue ExtraLoad =
10963 DAG.getMemIntrinsicNode(ISD::INTRINSIC_W_CHAIN, dl,
10964 DAG.getVTList(PermTy, MVT::Other),
10965 ExtraLoadOps, LDTy, ExtraMMO);
10967 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
10968 BaseLoad.getValue(1), ExtraLoad.getValue(1));
10970 // Because vperm has a big-endian bias, we must reverse the order
10971 // of the input vectors and complement the permute control vector
10972 // when generating little endian code. We have already handled the
10973 // latter by using lvsr instead of lvsl, so just reverse BaseLoad
10974 // and ExtraLoad here.
10976 if (isLittleEndian)
10977 Perm = BuildIntrinsicOp(IntrPerm,
10978 ExtraLoad, BaseLoad, PermCntl, DAG, dl);
10980 Perm = BuildIntrinsicOp(IntrPerm,
10981 BaseLoad, ExtraLoad, PermCntl, DAG, dl);
10984 Perm = Subtarget.hasAltivec() ?
10985 DAG.getNode(ISD::BITCAST, dl, VT, Perm) :
10986 DAG.getNode(ISD::FP_ROUND, dl, VT, Perm, // QPX
10987 DAG.getTargetConstant(1, dl, MVT::i64));
10988 // second argument is 1 because this rounding
10989 // is always exact.
10991 // The output of the permutation is our loaded result, the TokenFactor is
10993 DCI.CombineTo(N, Perm, TF);
10994 return SDValue(N, 0);
10998 case ISD::INTRINSIC_WO_CHAIN: {
10999 bool isLittleEndian = Subtarget.isLittleEndian();
11000 unsigned IID = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
11001 Intrinsic::ID Intr = (isLittleEndian ? Intrinsic::ppc_altivec_lvsr
11002 : Intrinsic::ppc_altivec_lvsl);
11003 if ((IID == Intr ||
11004 IID == Intrinsic::ppc_qpx_qvlpcld ||
11005 IID == Intrinsic::ppc_qpx_qvlpcls) &&
11006 N->getOperand(1)->getOpcode() == ISD::ADD) {
11007 SDValue Add = N->getOperand(1);
11009 int Bits = IID == Intrinsic::ppc_qpx_qvlpcld ?
11010 5 /* 32 byte alignment */ : 4 /* 16 byte alignment */;
11012 if (DAG.MaskedValueIsZero(
11013 Add->getOperand(1),
11014 APInt::getAllOnesValue(Bits /* alignment */)
11016 Add.getValueType().getScalarType().getSizeInBits()))) {
11017 SDNode *BasePtr = Add->getOperand(0).getNode();
11018 for (SDNode::use_iterator UI = BasePtr->use_begin(),
11019 UE = BasePtr->use_end();
11021 if (UI->getOpcode() == ISD::INTRINSIC_WO_CHAIN &&
11022 cast<ConstantSDNode>(UI->getOperand(0))->getZExtValue() == IID) {
11023 // We've found another LVSL/LVSR, and this address is an aligned
11024 // multiple of that one. The results will be the same, so use the
11025 // one we've just found instead.
11027 return SDValue(*UI, 0);
11032 if (isa<ConstantSDNode>(Add->getOperand(1))) {
11033 SDNode *BasePtr = Add->getOperand(0).getNode();
11034 for (SDNode::use_iterator UI = BasePtr->use_begin(),
11035 UE = BasePtr->use_end(); UI != UE; ++UI) {
11036 if (UI->getOpcode() == ISD::ADD &&
11037 isa<ConstantSDNode>(UI->getOperand(1)) &&
11038 (cast<ConstantSDNode>(Add->getOperand(1))->getZExtValue() -
11039 cast<ConstantSDNode>(UI->getOperand(1))->getZExtValue()) %
11040 (1ULL << Bits) == 0) {
11041 SDNode *OtherAdd = *UI;
11042 for (SDNode::use_iterator VI = OtherAdd->use_begin(),
11043 VE = OtherAdd->use_end(); VI != VE; ++VI) {
11044 if (VI->getOpcode() == ISD::INTRINSIC_WO_CHAIN &&
11045 cast<ConstantSDNode>(VI->getOperand(0))->getZExtValue() == IID) {
11046 return SDValue(*VI, 0);
11056 case ISD::INTRINSIC_W_CHAIN: {
11057 // For little endian, VSX loads require generating lxvd2x/xxswapd.
11058 if (Subtarget.hasVSX() && Subtarget.isLittleEndian()) {
11059 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
11062 case Intrinsic::ppc_vsx_lxvw4x:
11063 case Intrinsic::ppc_vsx_lxvd2x:
11064 return expandVSXLoadForLE(N, DCI);
11069 case ISD::INTRINSIC_VOID: {
11070 // For little endian, VSX stores require generating xxswapd/stxvd2x.
11071 if (Subtarget.hasVSX() && Subtarget.isLittleEndian()) {
11072 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
11075 case Intrinsic::ppc_vsx_stxvw4x:
11076 case Intrinsic::ppc_vsx_stxvd2x:
11077 return expandVSXStoreForLE(N, DCI);
11083 // Turn BSWAP (LOAD) -> lhbrx/lwbrx.
11084 if (ISD::isNON_EXTLoad(N->getOperand(0).getNode()) &&
11085 N->getOperand(0).hasOneUse() &&
11086 (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i16 ||
11087 (Subtarget.hasLDBRX() && Subtarget.isPPC64() &&
11088 N->getValueType(0) == MVT::i64))) {
11089 SDValue Load = N->getOperand(0);
11090 LoadSDNode *LD = cast<LoadSDNode>(Load);
11091 // Create the byte-swapping load.
11093 LD->getChain(), // Chain
11094 LD->getBasePtr(), // Ptr
11095 DAG.getValueType(N->getValueType(0)) // VT
11098 DAG.getMemIntrinsicNode(PPCISD::LBRX, dl,
11099 DAG.getVTList(N->getValueType(0) == MVT::i64 ?
11100 MVT::i64 : MVT::i32, MVT::Other),
11101 Ops, LD->getMemoryVT(), LD->getMemOperand());
11103 // If this is an i16 load, insert the truncate.
11104 SDValue ResVal = BSLoad;
11105 if (N->getValueType(0) == MVT::i16)
11106 ResVal = DAG.getNode(ISD::TRUNCATE, dl, MVT::i16, BSLoad);
11108 // First, combine the bswap away. This makes the value produced by the
11110 DCI.CombineTo(N, ResVal);
11112 // Next, combine the load away, we give it a bogus result value but a real
11113 // chain result. The result value is dead because the bswap is dead.
11114 DCI.CombineTo(Load.getNode(), ResVal, BSLoad.getValue(1));
11116 // Return N so it doesn't get rechecked!
11117 return SDValue(N, 0);
11121 case PPCISD::VCMP: {
11122 // If a VCMPo node already exists with exactly the same operands as this
11123 // node, use its result instead of this node (VCMPo computes both a CR6 and
11124 // a normal output).
11126 if (!N->getOperand(0).hasOneUse() &&
11127 !N->getOperand(1).hasOneUse() &&
11128 !N->getOperand(2).hasOneUse()) {
11130 // Scan all of the users of the LHS, looking for VCMPo's that match.
11131 SDNode *VCMPoNode = nullptr;
11133 SDNode *LHSN = N->getOperand(0).getNode();
11134 for (SDNode::use_iterator UI = LHSN->use_begin(), E = LHSN->use_end();
11136 if (UI->getOpcode() == PPCISD::VCMPo &&
11137 UI->getOperand(1) == N->getOperand(1) &&
11138 UI->getOperand(2) == N->getOperand(2) &&
11139 UI->getOperand(0) == N->getOperand(0)) {
11144 // If there is no VCMPo node, or if the flag value has a single use, don't
11146 if (!VCMPoNode || VCMPoNode->hasNUsesOfValue(0, 1))
11149 // Look at the (necessarily single) use of the flag value. If it has a
11150 // chain, this transformation is more complex. Note that multiple things
11151 // could use the value result, which we should ignore.
11152 SDNode *FlagUser = nullptr;
11153 for (SDNode::use_iterator UI = VCMPoNode->use_begin();
11154 FlagUser == nullptr; ++UI) {
11155 assert(UI != VCMPoNode->use_end() && "Didn't find user!");
11156 SDNode *User = *UI;
11157 for (unsigned i = 0, e = User->getNumOperands(); i != e; ++i) {
11158 if (User->getOperand(i) == SDValue(VCMPoNode, 1)) {
11165 // If the user is a MFOCRF instruction, we know this is safe.
11166 // Otherwise we give up for right now.
11167 if (FlagUser->getOpcode() == PPCISD::MFOCRF)
11168 return SDValue(VCMPoNode, 0);
11172 case ISD::BRCOND: {
11173 SDValue Cond = N->getOperand(1);
11174 SDValue Target = N->getOperand(2);
11176 if (Cond.getOpcode() == ISD::INTRINSIC_W_CHAIN &&
11177 cast<ConstantSDNode>(Cond.getOperand(1))->getZExtValue() ==
11178 Intrinsic::ppc_is_decremented_ctr_nonzero) {
11180 // We now need to make the intrinsic dead (it cannot be instruction
11182 DAG.ReplaceAllUsesOfValueWith(Cond.getValue(1), Cond.getOperand(0));
11183 assert(Cond.getNode()->hasOneUse() &&
11184 "Counter decrement has more than one use");
11186 return DAG.getNode(PPCISD::BDNZ, dl, MVT::Other,
11187 N->getOperand(0), Target);
11192 // If this is a branch on an altivec predicate comparison, lower this so
11193 // that we don't have to do a MFOCRF: instead, branch directly on CR6. This
11194 // lowering is done pre-legalize, because the legalizer lowers the predicate
11195 // compare down to code that is difficult to reassemble.
11196 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(1))->get();
11197 SDValue LHS = N->getOperand(2), RHS = N->getOperand(3);
11199 // Sometimes the promoted value of the intrinsic is ANDed by some non-zero
11200 // value. If so, pass-through the AND to get to the intrinsic.
11201 if (LHS.getOpcode() == ISD::AND &&
11202 LHS.getOperand(0).getOpcode() == ISD::INTRINSIC_W_CHAIN &&
11203 cast<ConstantSDNode>(LHS.getOperand(0).getOperand(1))->getZExtValue() ==
11204 Intrinsic::ppc_is_decremented_ctr_nonzero &&
11205 isa<ConstantSDNode>(LHS.getOperand(1)) &&
11206 !isNullConstant(LHS.getOperand(1)))
11207 LHS = LHS.getOperand(0);
11209 if (LHS.getOpcode() == ISD::INTRINSIC_W_CHAIN &&
11210 cast<ConstantSDNode>(LHS.getOperand(1))->getZExtValue() ==
11211 Intrinsic::ppc_is_decremented_ctr_nonzero &&
11212 isa<ConstantSDNode>(RHS)) {
11213 assert((CC == ISD::SETEQ || CC == ISD::SETNE) &&
11214 "Counter decrement comparison is not EQ or NE");
11216 unsigned Val = cast<ConstantSDNode>(RHS)->getZExtValue();
11217 bool isBDNZ = (CC == ISD::SETEQ && Val) ||
11218 (CC == ISD::SETNE && !Val);
11220 // We now need to make the intrinsic dead (it cannot be instruction
11222 DAG.ReplaceAllUsesOfValueWith(LHS.getValue(1), LHS.getOperand(0));
11223 assert(LHS.getNode()->hasOneUse() &&
11224 "Counter decrement has more than one use");
11226 return DAG.getNode(isBDNZ ? PPCISD::BDNZ : PPCISD::BDZ, dl, MVT::Other,
11227 N->getOperand(0), N->getOperand(4));
11233 if (LHS.getOpcode() == ISD::INTRINSIC_WO_CHAIN &&
11234 isa<ConstantSDNode>(RHS) && (CC == ISD::SETEQ || CC == ISD::SETNE) &&
11235 getVectorCompareInfo(LHS, CompareOpc, isDot, Subtarget)) {
11236 assert(isDot && "Can't compare against a vector result!");
11238 // If this is a comparison against something other than 0/1, then we know
11239 // that the condition is never/always true.
11240 unsigned Val = cast<ConstantSDNode>(RHS)->getZExtValue();
11241 if (Val != 0 && Val != 1) {
11242 if (CC == ISD::SETEQ) // Cond never true, remove branch.
11243 return N->getOperand(0);
11244 // Always !=, turn it into an unconditional branch.
11245 return DAG.getNode(ISD::BR, dl, MVT::Other,
11246 N->getOperand(0), N->getOperand(4));
11249 bool BranchOnWhenPredTrue = (CC == ISD::SETEQ) ^ (Val == 0);
11251 // Create the PPCISD altivec 'dot' comparison node.
11253 LHS.getOperand(2), // LHS of compare
11254 LHS.getOperand(3), // RHS of compare
11255 DAG.getConstant(CompareOpc, dl, MVT::i32)
11257 EVT VTs[] = { LHS.getOperand(2).getValueType(), MVT::Glue };
11258 SDValue CompNode = DAG.getNode(PPCISD::VCMPo, dl, VTs, Ops);
11260 // Unpack the result based on how the target uses it.
11261 PPC::Predicate CompOpc;
11262 switch (cast<ConstantSDNode>(LHS.getOperand(1))->getZExtValue()) {
11263 default: // Can't happen, don't crash on invalid number though.
11264 case 0: // Branch on the value of the EQ bit of CR6.
11265 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_EQ : PPC::PRED_NE;
11267 case 1: // Branch on the inverted value of the EQ bit of CR6.
11268 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_NE : PPC::PRED_EQ;
11270 case 2: // Branch on the value of the LT bit of CR6.
11271 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_LT : PPC::PRED_GE;
11273 case 3: // Branch on the inverted value of the LT bit of CR6.
11274 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_GE : PPC::PRED_LT;
11278 return DAG.getNode(PPCISD::COND_BRANCH, dl, MVT::Other, N->getOperand(0),
11279 DAG.getConstant(CompOpc, dl, MVT::i32),
11280 DAG.getRegister(PPC::CR6, MVT::i32),
11281 N->getOperand(4), CompNode.getValue(1));
11285 case ISD::BUILD_VECTOR:
11286 return DAGCombineBuildVector(N, DCI);
11293 PPCTargetLowering::BuildSDIVPow2(SDNode *N, const APInt &Divisor,
11295 std::vector<SDNode *> *Created) const {
11296 // fold (sdiv X, pow2)
11297 EVT VT = N->getValueType(0);
11298 if (VT == MVT::i64 && !Subtarget.isPPC64())
11300 if ((VT != MVT::i32 && VT != MVT::i64) ||
11301 !(Divisor.isPowerOf2() || (-Divisor).isPowerOf2()))
11305 SDValue N0 = N->getOperand(0);
11307 bool IsNegPow2 = (-Divisor).isPowerOf2();
11308 unsigned Lg2 = (IsNegPow2 ? -Divisor : Divisor).countTrailingZeros();
11309 SDValue ShiftAmt = DAG.getConstant(Lg2, DL, VT);
11311 SDValue Op = DAG.getNode(PPCISD::SRA_ADDZE, DL, VT, N0, ShiftAmt);
11313 Created->push_back(Op.getNode());
11316 Op = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, DL, VT), Op);
11318 Created->push_back(Op.getNode());
11324 //===----------------------------------------------------------------------===//
11325 // Inline Assembly Support
11326 //===----------------------------------------------------------------------===//
11328 void PPCTargetLowering::computeKnownBitsForTargetNode(const SDValue Op,
11331 const SelectionDAG &DAG,
11332 unsigned Depth) const {
11333 KnownZero = KnownOne = APInt(KnownZero.getBitWidth(), 0);
11334 switch (Op.getOpcode()) {
11336 case PPCISD::LBRX: {
11337 // lhbrx is known to have the top bits cleared out.
11338 if (cast<VTSDNode>(Op.getOperand(2))->getVT() == MVT::i16)
11339 KnownZero = 0xFFFF0000;
11342 case ISD::INTRINSIC_WO_CHAIN: {
11343 switch (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue()) {
11345 case Intrinsic::ppc_altivec_vcmpbfp_p:
11346 case Intrinsic::ppc_altivec_vcmpeqfp_p:
11347 case Intrinsic::ppc_altivec_vcmpequb_p:
11348 case Intrinsic::ppc_altivec_vcmpequh_p:
11349 case Intrinsic::ppc_altivec_vcmpequw_p:
11350 case Intrinsic::ppc_altivec_vcmpequd_p:
11351 case Intrinsic::ppc_altivec_vcmpgefp_p:
11352 case Intrinsic::ppc_altivec_vcmpgtfp_p:
11353 case Intrinsic::ppc_altivec_vcmpgtsb_p:
11354 case Intrinsic::ppc_altivec_vcmpgtsh_p:
11355 case Intrinsic::ppc_altivec_vcmpgtsw_p:
11356 case Intrinsic::ppc_altivec_vcmpgtsd_p:
11357 case Intrinsic::ppc_altivec_vcmpgtub_p:
11358 case Intrinsic::ppc_altivec_vcmpgtuh_p:
11359 case Intrinsic::ppc_altivec_vcmpgtuw_p:
11360 case Intrinsic::ppc_altivec_vcmpgtud_p:
11361 KnownZero = ~1U; // All bits but the low one are known to be zero.
11368 unsigned PPCTargetLowering::getPrefLoopAlignment(MachineLoop *ML) const {
11369 switch (Subtarget.getDarwinDirective()) {
11372 case PPC::DIR_PWR4:
11373 case PPC::DIR_PWR5:
11374 case PPC::DIR_PWR5X:
11375 case PPC::DIR_PWR6:
11376 case PPC::DIR_PWR6X:
11377 case PPC::DIR_PWR7:
11378 case PPC::DIR_PWR8:
11379 case PPC::DIR_PWR9: {
11383 const PPCInstrInfo *TII = Subtarget.getInstrInfo();
11385 // For small loops (between 5 and 8 instructions), align to a 32-byte
11386 // boundary so that the entire loop fits in one instruction-cache line.
11387 uint64_t LoopSize = 0;
11388 for (auto I = ML->block_begin(), IE = ML->block_end(); I != IE; ++I)
11389 for (auto J = (*I)->begin(), JE = (*I)->end(); J != JE; ++J) {
11390 LoopSize += TII->getInstSizeInBytes(*J);
11395 if (LoopSize > 16 && LoopSize <= 32)
11402 return TargetLowering::getPrefLoopAlignment(ML);
11405 /// getConstraintType - Given a constraint, return the type of
11406 /// constraint it is for this target.
11407 PPCTargetLowering::ConstraintType
11408 PPCTargetLowering::getConstraintType(StringRef Constraint) const {
11409 if (Constraint.size() == 1) {
11410 switch (Constraint[0]) {
11418 return C_RegisterClass;
11420 // FIXME: While Z does indicate a memory constraint, it specifically
11421 // indicates an r+r address (used in conjunction with the 'y' modifier
11422 // in the replacement string). Currently, we're forcing the base
11423 // register to be r0 in the asm printer (which is interpreted as zero)
11424 // and forming the complete address in the second register. This is
11428 } else if (Constraint == "wc") { // individual CR bits.
11429 return C_RegisterClass;
11430 } else if (Constraint == "wa" || Constraint == "wd" ||
11431 Constraint == "wf" || Constraint == "ws") {
11432 return C_RegisterClass; // VSX registers.
11434 return TargetLowering::getConstraintType(Constraint);
11437 /// Examine constraint type and operand type and determine a weight value.
11438 /// This object must already have been set up with the operand type
11439 /// and the current alternative constraint selected.
11440 TargetLowering::ConstraintWeight
11441 PPCTargetLowering::getSingleConstraintMatchWeight(
11442 AsmOperandInfo &info, const char *constraint) const {
11443 ConstraintWeight weight = CW_Invalid;
11444 Value *CallOperandVal = info.CallOperandVal;
11445 // If we don't have a value, we can't do a match,
11446 // but allow it at the lowest weight.
11447 if (!CallOperandVal)
11449 Type *type = CallOperandVal->getType();
11451 // Look at the constraint type.
11452 if (StringRef(constraint) == "wc" && type->isIntegerTy(1))
11453 return CW_Register; // an individual CR bit.
11454 else if ((StringRef(constraint) == "wa" ||
11455 StringRef(constraint) == "wd" ||
11456 StringRef(constraint) == "wf") &&
11457 type->isVectorTy())
11458 return CW_Register;
11459 else if (StringRef(constraint) == "ws" && type->isDoubleTy())
11460 return CW_Register;
11462 switch (*constraint) {
11464 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
11467 if (type->isIntegerTy())
11468 weight = CW_Register;
11471 if (type->isFloatTy())
11472 weight = CW_Register;
11475 if (type->isDoubleTy())
11476 weight = CW_Register;
11479 if (type->isVectorTy())
11480 weight = CW_Register;
11483 weight = CW_Register;
11486 weight = CW_Memory;
11492 std::pair<unsigned, const TargetRegisterClass *>
11493 PPCTargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
11494 StringRef Constraint,
11496 if (Constraint.size() == 1) {
11497 // GCC RS6000 Constraint Letters
11498 switch (Constraint[0]) {
11499 case 'b': // R1-R31
11500 if (VT == MVT::i64 && Subtarget.isPPC64())
11501 return std::make_pair(0U, &PPC::G8RC_NOX0RegClass);
11502 return std::make_pair(0U, &PPC::GPRC_NOR0RegClass);
11503 case 'r': // R0-R31
11504 if (VT == MVT::i64 && Subtarget.isPPC64())
11505 return std::make_pair(0U, &PPC::G8RCRegClass);
11506 return std::make_pair(0U, &PPC::GPRCRegClass);
11507 // 'd' and 'f' constraints are both defined to be "the floating point
11508 // registers", where one is for 32-bit and the other for 64-bit. We don't
11509 // really care overly much here so just give them all the same reg classes.
11512 if (VT == MVT::f32 || VT == MVT::i32)
11513 return std::make_pair(0U, &PPC::F4RCRegClass);
11514 if (VT == MVT::f64 || VT == MVT::i64)
11515 return std::make_pair(0U, &PPC::F8RCRegClass);
11516 if (VT == MVT::v4f64 && Subtarget.hasQPX())
11517 return std::make_pair(0U, &PPC::QFRCRegClass);
11518 if (VT == MVT::v4f32 && Subtarget.hasQPX())
11519 return std::make_pair(0U, &PPC::QSRCRegClass);
11522 if (VT == MVT::v4f64 && Subtarget.hasQPX())
11523 return std::make_pair(0U, &PPC::QFRCRegClass);
11524 if (VT == MVT::v4f32 && Subtarget.hasQPX())
11525 return std::make_pair(0U, &PPC::QSRCRegClass);
11526 if (Subtarget.hasAltivec())
11527 return std::make_pair(0U, &PPC::VRRCRegClass);
11529 return std::make_pair(0U, &PPC::CRRCRegClass);
11531 } else if (Constraint == "wc" && Subtarget.useCRBits()) {
11532 // An individual CR bit.
11533 return std::make_pair(0U, &PPC::CRBITRCRegClass);
11534 } else if ((Constraint == "wa" || Constraint == "wd" ||
11535 Constraint == "wf") && Subtarget.hasVSX()) {
11536 return std::make_pair(0U, &PPC::VSRCRegClass);
11537 } else if (Constraint == "ws" && Subtarget.hasVSX()) {
11538 if (VT == MVT::f32 && Subtarget.hasP8Vector())
11539 return std::make_pair(0U, &PPC::VSSRCRegClass);
11541 return std::make_pair(0U, &PPC::VSFRCRegClass);
11544 std::pair<unsigned, const TargetRegisterClass *> R =
11545 TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
11547 // r[0-9]+ are used, on PPC64, to refer to the corresponding 64-bit registers
11548 // (which we call X[0-9]+). If a 64-bit value has been requested, and a
11549 // 32-bit GPR has been selected, then 'upgrade' it to the 64-bit parent
11551 // FIXME: If TargetLowering::getRegForInlineAsmConstraint could somehow use
11552 // the AsmName field from *RegisterInfo.td, then this would not be necessary.
11553 if (R.first && VT == MVT::i64 && Subtarget.isPPC64() &&
11554 PPC::GPRCRegClass.contains(R.first))
11555 return std::make_pair(TRI->getMatchingSuperReg(R.first,
11556 PPC::sub_32, &PPC::G8RCRegClass),
11557 &PPC::G8RCRegClass);
11559 // GCC accepts 'cc' as an alias for 'cr0', and we need to do the same.
11560 if (!R.second && StringRef("{cc}").equals_lower(Constraint)) {
11561 R.first = PPC::CR0;
11562 R.second = &PPC::CRRCRegClass;
11568 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
11569 /// vector. If it is invalid, don't add anything to Ops.
11570 void PPCTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
11571 std::string &Constraint,
11572 std::vector<SDValue>&Ops,
11573 SelectionDAG &DAG) const {
11576 // Only support length 1 constraints.
11577 if (Constraint.length() > 1) return;
11579 char Letter = Constraint[0];
11590 ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op);
11591 if (!CST) return; // Must be an immediate to match.
11593 int64_t Value = CST->getSExtValue();
11594 EVT TCVT = MVT::i64; // All constants taken to be 64 bits so that negative
11595 // numbers are printed as such.
11597 default: llvm_unreachable("Unknown constraint letter!");
11598 case 'I': // "I" is a signed 16-bit constant.
11599 if (isInt<16>(Value))
11600 Result = DAG.getTargetConstant(Value, dl, TCVT);
11602 case 'J': // "J" is a constant with only the high-order 16 bits nonzero.
11603 if (isShiftedUInt<16, 16>(Value))
11604 Result = DAG.getTargetConstant(Value, dl, TCVT);
11606 case 'L': // "L" is a signed 16-bit constant shifted left 16 bits.
11607 if (isShiftedInt<16, 16>(Value))
11608 Result = DAG.getTargetConstant(Value, dl, TCVT);
11610 case 'K': // "K" is a constant with only the low-order 16 bits nonzero.
11611 if (isUInt<16>(Value))
11612 Result = DAG.getTargetConstant(Value, dl, TCVT);
11614 case 'M': // "M" is a constant that is greater than 31.
11616 Result = DAG.getTargetConstant(Value, dl, TCVT);
11618 case 'N': // "N" is a positive constant that is an exact power of two.
11619 if (Value > 0 && isPowerOf2_64(Value))
11620 Result = DAG.getTargetConstant(Value, dl, TCVT);
11622 case 'O': // "O" is the constant zero.
11624 Result = DAG.getTargetConstant(Value, dl, TCVT);
11626 case 'P': // "P" is a constant whose negation is a signed 16-bit constant.
11627 if (isInt<16>(-Value))
11628 Result = DAG.getTargetConstant(Value, dl, TCVT);
11635 if (Result.getNode()) {
11636 Ops.push_back(Result);
11640 // Handle standard constraint letters.
11641 TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
11644 // isLegalAddressingMode - Return true if the addressing mode represented
11645 // by AM is legal for this target, for a load/store of the specified type.
11646 bool PPCTargetLowering::isLegalAddressingMode(const DataLayout &DL,
11647 const AddrMode &AM, Type *Ty,
11648 unsigned AS) const {
11649 // PPC does not allow r+i addressing modes for vectors!
11650 if (Ty->isVectorTy() && AM.BaseOffs != 0)
11653 // PPC allows a sign-extended 16-bit immediate field.
11654 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1)
11657 // No global is ever allowed as a base.
11661 // PPC only support r+r,
11662 switch (AM.Scale) {
11663 case 0: // "r+i" or just "i", depending on HasBaseReg.
11666 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed.
11668 // Otherwise we have r+r or r+i.
11671 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed.
11673 // Allow 2*r as r+r.
11676 // No other scales are supported.
11683 SDValue PPCTargetLowering::LowerRETURNADDR(SDValue Op,
11684 SelectionDAG &DAG) const {
11685 MachineFunction &MF = DAG.getMachineFunction();
11686 MachineFrameInfo &MFI = MF.getFrameInfo();
11687 MFI.setReturnAddressIsTaken(true);
11689 if (verifyReturnAddressArgumentIsConstant(Op, DAG))
11693 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
11695 // Make sure the function does not optimize away the store of the RA to
11697 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
11698 FuncInfo->setLRStoreRequired();
11699 bool isPPC64 = Subtarget.isPPC64();
11700 auto PtrVT = getPointerTy(MF.getDataLayout());
11703 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
11705 DAG.getConstant(Subtarget.getFrameLowering()->getReturnSaveOffset(), dl,
11706 isPPC64 ? MVT::i64 : MVT::i32);
11707 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
11708 DAG.getNode(ISD::ADD, dl, PtrVT, FrameAddr, Offset),
11709 MachinePointerInfo());
11712 // Just load the return address off the stack.
11713 SDValue RetAddrFI = getReturnAddrFrameIndex(DAG);
11714 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), RetAddrFI,
11715 MachinePointerInfo());
11718 SDValue PPCTargetLowering::LowerFRAMEADDR(SDValue Op,
11719 SelectionDAG &DAG) const {
11721 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
11723 MachineFunction &MF = DAG.getMachineFunction();
11724 MachineFrameInfo &MFI = MF.getFrameInfo();
11725 MFI.setFrameAddressIsTaken(true);
11727 EVT PtrVT = getPointerTy(MF.getDataLayout());
11728 bool isPPC64 = PtrVT == MVT::i64;
11730 // Naked functions never have a frame pointer, and so we use r1. For all
11731 // other functions, this decision must be delayed until during PEI.
11733 if (MF.getFunction()->hasFnAttribute(Attribute::Naked))
11734 FrameReg = isPPC64 ? PPC::X1 : PPC::R1;
11736 FrameReg = isPPC64 ? PPC::FP8 : PPC::FP;
11738 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg,
11741 FrameAddr = DAG.getLoad(Op.getValueType(), dl, DAG.getEntryNode(),
11742 FrameAddr, MachinePointerInfo());
11746 // FIXME? Maybe this could be a TableGen attribute on some registers and
11747 // this table could be generated automatically from RegInfo.
11748 unsigned PPCTargetLowering::getRegisterByName(const char* RegName, EVT VT,
11749 SelectionDAG &DAG) const {
11750 bool isPPC64 = Subtarget.isPPC64();
11751 bool isDarwinABI = Subtarget.isDarwinABI();
11753 if ((isPPC64 && VT != MVT::i64 && VT != MVT::i32) ||
11754 (!isPPC64 && VT != MVT::i32))
11755 report_fatal_error("Invalid register global variable type");
11757 bool is64Bit = isPPC64 && VT == MVT::i64;
11758 unsigned Reg = StringSwitch<unsigned>(RegName)
11759 .Case("r1", is64Bit ? PPC::X1 : PPC::R1)
11760 .Case("r2", (isDarwinABI || isPPC64) ? 0 : PPC::R2)
11761 .Case("r13", (!isPPC64 && isDarwinABI) ? 0 :
11762 (is64Bit ? PPC::X13 : PPC::R13))
11767 report_fatal_error("Invalid register name global variable");
11771 PPCTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
11772 // The PowerPC target isn't yet aware of offsets.
11776 bool PPCTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
11778 unsigned Intrinsic) const {
11780 switch (Intrinsic) {
11781 case Intrinsic::ppc_qpx_qvlfd:
11782 case Intrinsic::ppc_qpx_qvlfs:
11783 case Intrinsic::ppc_qpx_qvlfcd:
11784 case Intrinsic::ppc_qpx_qvlfcs:
11785 case Intrinsic::ppc_qpx_qvlfiwa:
11786 case Intrinsic::ppc_qpx_qvlfiwz:
11787 case Intrinsic::ppc_altivec_lvx:
11788 case Intrinsic::ppc_altivec_lvxl:
11789 case Intrinsic::ppc_altivec_lvebx:
11790 case Intrinsic::ppc_altivec_lvehx:
11791 case Intrinsic::ppc_altivec_lvewx:
11792 case Intrinsic::ppc_vsx_lxvd2x:
11793 case Intrinsic::ppc_vsx_lxvw4x: {
11795 switch (Intrinsic) {
11796 case Intrinsic::ppc_altivec_lvebx:
11799 case Intrinsic::ppc_altivec_lvehx:
11802 case Intrinsic::ppc_altivec_lvewx:
11805 case Intrinsic::ppc_vsx_lxvd2x:
11808 case Intrinsic::ppc_qpx_qvlfd:
11811 case Intrinsic::ppc_qpx_qvlfs:
11814 case Intrinsic::ppc_qpx_qvlfcd:
11817 case Intrinsic::ppc_qpx_qvlfcs:
11825 Info.opc = ISD::INTRINSIC_W_CHAIN;
11827 Info.ptrVal = I.getArgOperand(0);
11828 Info.offset = -VT.getStoreSize()+1;
11829 Info.size = 2*VT.getStoreSize()-1;
11832 Info.readMem = true;
11833 Info.writeMem = false;
11836 case Intrinsic::ppc_qpx_qvlfda:
11837 case Intrinsic::ppc_qpx_qvlfsa:
11838 case Intrinsic::ppc_qpx_qvlfcda:
11839 case Intrinsic::ppc_qpx_qvlfcsa:
11840 case Intrinsic::ppc_qpx_qvlfiwaa:
11841 case Intrinsic::ppc_qpx_qvlfiwza: {
11843 switch (Intrinsic) {
11844 case Intrinsic::ppc_qpx_qvlfda:
11847 case Intrinsic::ppc_qpx_qvlfsa:
11850 case Intrinsic::ppc_qpx_qvlfcda:
11853 case Intrinsic::ppc_qpx_qvlfcsa:
11861 Info.opc = ISD::INTRINSIC_W_CHAIN;
11863 Info.ptrVal = I.getArgOperand(0);
11865 Info.size = VT.getStoreSize();
11868 Info.readMem = true;
11869 Info.writeMem = false;
11872 case Intrinsic::ppc_qpx_qvstfd:
11873 case Intrinsic::ppc_qpx_qvstfs:
11874 case Intrinsic::ppc_qpx_qvstfcd:
11875 case Intrinsic::ppc_qpx_qvstfcs:
11876 case Intrinsic::ppc_qpx_qvstfiw:
11877 case Intrinsic::ppc_altivec_stvx:
11878 case Intrinsic::ppc_altivec_stvxl:
11879 case Intrinsic::ppc_altivec_stvebx:
11880 case Intrinsic::ppc_altivec_stvehx:
11881 case Intrinsic::ppc_altivec_stvewx:
11882 case Intrinsic::ppc_vsx_stxvd2x:
11883 case Intrinsic::ppc_vsx_stxvw4x: {
11885 switch (Intrinsic) {
11886 case Intrinsic::ppc_altivec_stvebx:
11889 case Intrinsic::ppc_altivec_stvehx:
11892 case Intrinsic::ppc_altivec_stvewx:
11895 case Intrinsic::ppc_vsx_stxvd2x:
11898 case Intrinsic::ppc_qpx_qvstfd:
11901 case Intrinsic::ppc_qpx_qvstfs:
11904 case Intrinsic::ppc_qpx_qvstfcd:
11907 case Intrinsic::ppc_qpx_qvstfcs:
11915 Info.opc = ISD::INTRINSIC_VOID;
11917 Info.ptrVal = I.getArgOperand(1);
11918 Info.offset = -VT.getStoreSize()+1;
11919 Info.size = 2*VT.getStoreSize()-1;
11922 Info.readMem = false;
11923 Info.writeMem = true;
11926 case Intrinsic::ppc_qpx_qvstfda:
11927 case Intrinsic::ppc_qpx_qvstfsa:
11928 case Intrinsic::ppc_qpx_qvstfcda:
11929 case Intrinsic::ppc_qpx_qvstfcsa:
11930 case Intrinsic::ppc_qpx_qvstfiwa: {
11932 switch (Intrinsic) {
11933 case Intrinsic::ppc_qpx_qvstfda:
11936 case Intrinsic::ppc_qpx_qvstfsa:
11939 case Intrinsic::ppc_qpx_qvstfcda:
11942 case Intrinsic::ppc_qpx_qvstfcsa:
11950 Info.opc = ISD::INTRINSIC_VOID;
11952 Info.ptrVal = I.getArgOperand(1);
11954 Info.size = VT.getStoreSize();
11957 Info.readMem = false;
11958 Info.writeMem = true;
11968 /// getOptimalMemOpType - Returns the target specific optimal type for load
11969 /// and store operations as a result of memset, memcpy, and memmove
11970 /// lowering. If DstAlign is zero that means it's safe to destination
11971 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
11972 /// means there isn't a need to check it against alignment requirement,
11973 /// probably because the source does not need to be loaded. If 'IsMemset' is
11974 /// true, that means it's expanding a memset. If 'ZeroMemset' is true, that
11975 /// means it's a memset of zero. 'MemcpyStrSrc' indicates whether the memcpy
11976 /// source is constant so it does not need to be loaded.
11977 /// It returns EVT::Other if the type should be determined using generic
11978 /// target-independent logic.
11979 EVT PPCTargetLowering::getOptimalMemOpType(uint64_t Size,
11980 unsigned DstAlign, unsigned SrcAlign,
11981 bool IsMemset, bool ZeroMemset,
11983 MachineFunction &MF) const {
11984 if (getTargetMachine().getOptLevel() != CodeGenOpt::None) {
11985 const Function *F = MF.getFunction();
11986 // When expanding a memset, require at least two QPX instructions to cover
11987 // the cost of loading the value to be stored from the constant pool.
11988 if (Subtarget.hasQPX() && Size >= 32 && (!IsMemset || Size >= 64) &&
11989 (!SrcAlign || SrcAlign >= 32) && (!DstAlign || DstAlign >= 32) &&
11990 !F->hasFnAttribute(Attribute::NoImplicitFloat)) {
11994 // We should use Altivec/VSX loads and stores when available. For unaligned
11995 // addresses, unaligned VSX loads are only fast starting with the P8.
11996 if (Subtarget.hasAltivec() && Size >= 16 &&
11997 (((!SrcAlign || SrcAlign >= 16) && (!DstAlign || DstAlign >= 16)) ||
11998 ((IsMemset && Subtarget.hasVSX()) || Subtarget.hasP8Vector())))
12002 if (Subtarget.isPPC64()) {
12009 /// \brief Returns true if it is beneficial to convert a load of a constant
12010 /// to just the constant itself.
12011 bool PPCTargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
12013 assert(Ty->isIntegerTy());
12015 unsigned BitSize = Ty->getPrimitiveSizeInBits();
12016 return !(BitSize == 0 || BitSize > 64);
12019 bool PPCTargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const {
12020 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
12022 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
12023 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
12024 return NumBits1 == 64 && NumBits2 == 32;
12027 bool PPCTargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
12028 if (!VT1.isInteger() || !VT2.isInteger())
12030 unsigned NumBits1 = VT1.getSizeInBits();
12031 unsigned NumBits2 = VT2.getSizeInBits();
12032 return NumBits1 == 64 && NumBits2 == 32;
12035 bool PPCTargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
12036 // Generally speaking, zexts are not free, but they are free when they can be
12037 // folded with other operations.
12038 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(Val)) {
12039 EVT MemVT = LD->getMemoryVT();
12040 if ((MemVT == MVT::i1 || MemVT == MVT::i8 || MemVT == MVT::i16 ||
12041 (Subtarget.isPPC64() && MemVT == MVT::i32)) &&
12042 (LD->getExtensionType() == ISD::NON_EXTLOAD ||
12043 LD->getExtensionType() == ISD::ZEXTLOAD))
12047 // FIXME: Add other cases...
12048 // - 32-bit shifts with a zext to i64
12049 // - zext after ctlz, bswap, etc.
12050 // - zext after and by a constant mask
12052 return TargetLowering::isZExtFree(Val, VT2);
12055 bool PPCTargetLowering::isFPExtFree(EVT VT) const {
12056 assert(VT.isFloatingPoint());
12060 bool PPCTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
12061 return isInt<16>(Imm) || isUInt<16>(Imm);
12064 bool PPCTargetLowering::isLegalAddImmediate(int64_t Imm) const {
12065 return isInt<16>(Imm) || isUInt<16>(Imm);
12068 bool PPCTargetLowering::allowsMisalignedMemoryAccesses(EVT VT,
12071 bool *Fast) const {
12072 if (DisablePPCUnaligned)
12075 // PowerPC supports unaligned memory access for simple non-vector types.
12076 // Although accessing unaligned addresses is not as efficient as accessing
12077 // aligned addresses, it is generally more efficient than manual expansion,
12078 // and generally only traps for software emulation when crossing page
12081 if (!VT.isSimple())
12084 if (VT.getSimpleVT().isVector()) {
12085 if (Subtarget.hasVSX()) {
12086 if (VT != MVT::v2f64 && VT != MVT::v2i64 &&
12087 VT != MVT::v4f32 && VT != MVT::v4i32)
12094 if (VT == MVT::ppcf128)
12103 bool PPCTargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
12104 VT = VT.getScalarType();
12106 if (!VT.isSimple())
12109 switch (VT.getSimpleVT().SimpleTy) {
12121 PPCTargetLowering::getScratchRegisters(CallingConv::ID) const {
12122 // LR is a callee-save register, but we must treat it as clobbered by any call
12123 // site. Hence we include LR in the scratch registers, which are in turn added
12124 // as implicit-defs for stackmaps and patchpoints. The same reasoning applies
12125 // to CTR, which is used by any indirect call.
12126 static const MCPhysReg ScratchRegs[] = {
12127 PPC::X12, PPC::LR8, PPC::CTR8, 0
12130 return ScratchRegs;
12133 unsigned PPCTargetLowering::getExceptionPointerRegister(
12134 const Constant *PersonalityFn) const {
12135 return Subtarget.isPPC64() ? PPC::X3 : PPC::R3;
12138 unsigned PPCTargetLowering::getExceptionSelectorRegister(
12139 const Constant *PersonalityFn) const {
12140 return Subtarget.isPPC64() ? PPC::X4 : PPC::R4;
12144 PPCTargetLowering::shouldExpandBuildVectorWithShuffles(
12145 EVT VT , unsigned DefinedValues) const {
12146 if (VT == MVT::v2i64)
12147 return Subtarget.hasDirectMove(); // Don't need stack ops with direct moves
12149 if (Subtarget.hasVSX() || Subtarget.hasQPX())
12152 return TargetLowering::shouldExpandBuildVectorWithShuffles(VT, DefinedValues);
12155 Sched::Preference PPCTargetLowering::getSchedulingPreference(SDNode *N) const {
12156 if (DisableILPPref || Subtarget.enableMachineScheduler())
12157 return TargetLowering::getSchedulingPreference(N);
12162 // Create a fast isel object.
12164 PPCTargetLowering::createFastISel(FunctionLoweringInfo &FuncInfo,
12165 const TargetLibraryInfo *LibInfo) const {
12166 return PPC::createFastISel(FuncInfo, LibInfo);
12169 void PPCTargetLowering::initializeSplitCSR(MachineBasicBlock *Entry) const {
12170 if (Subtarget.isDarwinABI()) return;
12171 if (!Subtarget.isPPC64()) return;
12173 // Update IsSplitCSR in PPCFunctionInfo
12174 PPCFunctionInfo *PFI = Entry->getParent()->getInfo<PPCFunctionInfo>();
12175 PFI->setIsSplitCSR(true);
12178 void PPCTargetLowering::insertCopiesSplitCSR(
12179 MachineBasicBlock *Entry,
12180 const SmallVectorImpl<MachineBasicBlock *> &Exits) const {
12181 const PPCRegisterInfo *TRI = Subtarget.getRegisterInfo();
12182 const MCPhysReg *IStart = TRI->getCalleeSavedRegsViaCopy(Entry->getParent());
12186 const TargetInstrInfo *TII = Subtarget.getInstrInfo();
12187 MachineRegisterInfo *MRI = &Entry->getParent()->getRegInfo();
12188 MachineBasicBlock::iterator MBBI = Entry->begin();
12189 for (const MCPhysReg *I = IStart; *I; ++I) {
12190 const TargetRegisterClass *RC = nullptr;
12191 if (PPC::G8RCRegClass.contains(*I))
12192 RC = &PPC::G8RCRegClass;
12193 else if (PPC::F8RCRegClass.contains(*I))
12194 RC = &PPC::F8RCRegClass;
12195 else if (PPC::CRRCRegClass.contains(*I))
12196 RC = &PPC::CRRCRegClass;
12197 else if (PPC::VRRCRegClass.contains(*I))
12198 RC = &PPC::VRRCRegClass;
12200 llvm_unreachable("Unexpected register class in CSRsViaCopy!");
12202 unsigned NewVR = MRI->createVirtualRegister(RC);
12203 // Create copy from CSR to a virtual register.
12204 // FIXME: this currently does not emit CFI pseudo-instructions, it works
12205 // fine for CXX_FAST_TLS since the C++-style TLS access functions should be
12206 // nounwind. If we want to generalize this later, we may need to emit
12207 // CFI pseudo-instructions.
12208 assert(Entry->getParent()->getFunction()->hasFnAttribute(
12209 Attribute::NoUnwind) &&
12210 "Function should be nounwind in insertCopiesSplitCSR!");
12211 Entry->addLiveIn(*I);
12212 BuildMI(*Entry, MBBI, DebugLoc(), TII->get(TargetOpcode::COPY), NewVR)
12215 // Insert the copy-back instructions right before the terminator
12216 for (auto *Exit : Exits)
12217 BuildMI(*Exit, Exit->getFirstTerminator(), DebugLoc(),
12218 TII->get(TargetOpcode::COPY), *I)
12223 // Override to enable LOAD_STACK_GUARD lowering on Linux.
12224 bool PPCTargetLowering::useLoadStackGuardNode() const {
12225 if (!Subtarget.isTargetLinux())
12226 return TargetLowering::useLoadStackGuardNode();
12230 // Override to disable global variable loading on Linux.
12231 void PPCTargetLowering::insertSSPDeclarations(Module &M) const {
12232 if (!Subtarget.isTargetLinux())
12233 return TargetLowering::insertSSPDeclarations(M);