1 //===-- ARMISelLowering.cpp - ARM 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 defines the interfaces that ARM uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
15 #define DEBUG_TYPE "arm-isel"
16 #include "ARMISelLowering.h"
18 #include "ARMCallingConv.h"
19 #include "ARMConstantPoolValue.h"
20 #include "ARMMachineFunctionInfo.h"
21 #include "ARMPerfectShuffle.h"
22 #include "ARMSubtarget.h"
23 #include "ARMTargetMachine.h"
24 #include "ARMTargetObjectFile.h"
25 #include "MCTargetDesc/ARMAddressingModes.h"
26 #include "llvm/CallingConv.h"
27 #include "llvm/Constants.h"
28 #include "llvm/Function.h"
29 #include "llvm/GlobalValue.h"
30 #include "llvm/Instruction.h"
31 #include "llvm/Instructions.h"
32 #include "llvm/Intrinsics.h"
33 #include "llvm/Type.h"
34 #include "llvm/CodeGen/CallingConvLower.h"
35 #include "llvm/CodeGen/IntrinsicLowering.h"
36 #include "llvm/CodeGen/MachineBasicBlock.h"
37 #include "llvm/CodeGen/MachineFrameInfo.h"
38 #include "llvm/CodeGen/MachineFunction.h"
39 #include "llvm/CodeGen/MachineInstrBuilder.h"
40 #include "llvm/CodeGen/MachineModuleInfo.h"
41 #include "llvm/CodeGen/MachineRegisterInfo.h"
42 #include "llvm/CodeGen/SelectionDAG.h"
43 #include "llvm/MC/MCSectionMachO.h"
44 #include "llvm/Target/TargetOptions.h"
45 #include "llvm/ADT/StringExtras.h"
46 #include "llvm/ADT/Statistic.h"
47 #include "llvm/Support/CommandLine.h"
48 #include "llvm/Support/ErrorHandling.h"
49 #include "llvm/Support/MathExtras.h"
50 #include "llvm/Support/raw_ostream.h"
53 STATISTIC(NumTailCalls, "Number of tail calls");
54 STATISTIC(NumMovwMovt, "Number of GAs materialized with movw + movt");
55 STATISTIC(NumLoopByVals, "Number of loops generated for byval arguments");
57 // This option should go away when tail calls fully work.
59 EnableARMTailCalls("arm-tail-calls", cl::Hidden,
60 cl::desc("Generate tail calls (TEMPORARY OPTION)."),
64 EnableARMLongCalls("arm-long-calls", cl::Hidden,
65 cl::desc("Generate calls via indirect call instructions"),
69 ARMInterworking("arm-interworking", cl::Hidden,
70 cl::desc("Enable / disable ARM interworking (for debugging only)"),
74 class ARMCCState : public CCState {
76 ARMCCState(CallingConv::ID CC, bool isVarArg, MachineFunction &MF,
77 const TargetMachine &TM, SmallVector<CCValAssign, 16> &locs,
78 LLVMContext &C, ParmContext PC)
79 : CCState(CC, isVarArg, MF, TM, locs, C) {
80 assert(((PC == Call) || (PC == Prologue)) &&
81 "ARMCCState users must specify whether their context is call"
82 "or prologue generation.");
88 // The APCS parameter registers.
89 static const uint16_t GPRArgRegs[] = {
90 ARM::R0, ARM::R1, ARM::R2, ARM::R3
93 void ARMTargetLowering::addTypeForNEON(MVT VT, MVT PromotedLdStVT,
94 MVT PromotedBitwiseVT) {
95 if (VT != PromotedLdStVT) {
96 setOperationAction(ISD::LOAD, VT, Promote);
97 AddPromotedToType (ISD::LOAD, VT, PromotedLdStVT);
99 setOperationAction(ISD::STORE, VT, Promote);
100 AddPromotedToType (ISD::STORE, VT, PromotedLdStVT);
103 MVT ElemTy = VT.getVectorElementType();
104 if (ElemTy != MVT::i64 && ElemTy != MVT::f64)
105 setOperationAction(ISD::SETCC, VT, Custom);
106 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
107 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
108 if (ElemTy == MVT::i32) {
109 setOperationAction(ISD::SINT_TO_FP, VT, Custom);
110 setOperationAction(ISD::UINT_TO_FP, VT, Custom);
111 setOperationAction(ISD::FP_TO_SINT, VT, Custom);
112 setOperationAction(ISD::FP_TO_UINT, VT, Custom);
114 setOperationAction(ISD::SINT_TO_FP, VT, Expand);
115 setOperationAction(ISD::UINT_TO_FP, VT, Expand);
116 setOperationAction(ISD::FP_TO_SINT, VT, Expand);
117 setOperationAction(ISD::FP_TO_UINT, VT, Expand);
119 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
120 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
121 setOperationAction(ISD::CONCAT_VECTORS, VT, Legal);
122 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Legal);
123 setOperationAction(ISD::SELECT, VT, Expand);
124 setOperationAction(ISD::SELECT_CC, VT, Expand);
125 setOperationAction(ISD::VSELECT, VT, Expand);
126 setOperationAction(ISD::SIGN_EXTEND_INREG, VT, Expand);
127 if (VT.isInteger()) {
128 setOperationAction(ISD::SHL, VT, Custom);
129 setOperationAction(ISD::SRA, VT, Custom);
130 setOperationAction(ISD::SRL, VT, Custom);
133 // Promote all bit-wise operations.
134 if (VT.isInteger() && VT != PromotedBitwiseVT) {
135 setOperationAction(ISD::AND, VT, Promote);
136 AddPromotedToType (ISD::AND, VT, PromotedBitwiseVT);
137 setOperationAction(ISD::OR, VT, Promote);
138 AddPromotedToType (ISD::OR, VT, PromotedBitwiseVT);
139 setOperationAction(ISD::XOR, VT, Promote);
140 AddPromotedToType (ISD::XOR, VT, PromotedBitwiseVT);
143 // Neon does not support vector divide/remainder operations.
144 setOperationAction(ISD::SDIV, VT, Expand);
145 setOperationAction(ISD::UDIV, VT, Expand);
146 setOperationAction(ISD::FDIV, VT, Expand);
147 setOperationAction(ISD::SREM, VT, Expand);
148 setOperationAction(ISD::UREM, VT, Expand);
149 setOperationAction(ISD::FREM, VT, Expand);
152 void ARMTargetLowering::addDRTypeForNEON(MVT VT) {
153 addRegisterClass(VT, &ARM::DPRRegClass);
154 addTypeForNEON(VT, MVT::f64, MVT::v2i32);
157 void ARMTargetLowering::addQRTypeForNEON(MVT VT) {
158 addRegisterClass(VT, &ARM::QPRRegClass);
159 addTypeForNEON(VT, MVT::v2f64, MVT::v4i32);
162 static TargetLoweringObjectFile *createTLOF(TargetMachine &TM) {
163 if (TM.getSubtarget<ARMSubtarget>().isTargetDarwin())
164 return new TargetLoweringObjectFileMachO();
166 return new ARMElfTargetObjectFile();
169 ARMTargetLowering::ARMTargetLowering(TargetMachine &TM)
170 : TargetLowering(TM, createTLOF(TM)) {
171 Subtarget = &TM.getSubtarget<ARMSubtarget>();
172 RegInfo = TM.getRegisterInfo();
173 Itins = TM.getInstrItineraryData();
175 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
177 if (Subtarget->isTargetDarwin()) {
178 // Uses VFP for Thumb libfuncs if available.
179 if (Subtarget->isThumb() && Subtarget->hasVFP2()) {
180 // Single-precision floating-point arithmetic.
181 setLibcallName(RTLIB::ADD_F32, "__addsf3vfp");
182 setLibcallName(RTLIB::SUB_F32, "__subsf3vfp");
183 setLibcallName(RTLIB::MUL_F32, "__mulsf3vfp");
184 setLibcallName(RTLIB::DIV_F32, "__divsf3vfp");
186 // Double-precision floating-point arithmetic.
187 setLibcallName(RTLIB::ADD_F64, "__adddf3vfp");
188 setLibcallName(RTLIB::SUB_F64, "__subdf3vfp");
189 setLibcallName(RTLIB::MUL_F64, "__muldf3vfp");
190 setLibcallName(RTLIB::DIV_F64, "__divdf3vfp");
192 // Single-precision comparisons.
193 setLibcallName(RTLIB::OEQ_F32, "__eqsf2vfp");
194 setLibcallName(RTLIB::UNE_F32, "__nesf2vfp");
195 setLibcallName(RTLIB::OLT_F32, "__ltsf2vfp");
196 setLibcallName(RTLIB::OLE_F32, "__lesf2vfp");
197 setLibcallName(RTLIB::OGE_F32, "__gesf2vfp");
198 setLibcallName(RTLIB::OGT_F32, "__gtsf2vfp");
199 setLibcallName(RTLIB::UO_F32, "__unordsf2vfp");
200 setLibcallName(RTLIB::O_F32, "__unordsf2vfp");
202 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
203 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETNE);
204 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
205 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
206 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
207 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
208 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
209 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
211 // Double-precision comparisons.
212 setLibcallName(RTLIB::OEQ_F64, "__eqdf2vfp");
213 setLibcallName(RTLIB::UNE_F64, "__nedf2vfp");
214 setLibcallName(RTLIB::OLT_F64, "__ltdf2vfp");
215 setLibcallName(RTLIB::OLE_F64, "__ledf2vfp");
216 setLibcallName(RTLIB::OGE_F64, "__gedf2vfp");
217 setLibcallName(RTLIB::OGT_F64, "__gtdf2vfp");
218 setLibcallName(RTLIB::UO_F64, "__unorddf2vfp");
219 setLibcallName(RTLIB::O_F64, "__unorddf2vfp");
221 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
222 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETNE);
223 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
224 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
225 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
226 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
227 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
228 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
230 // Floating-point to integer conversions.
231 // i64 conversions are done via library routines even when generating VFP
232 // instructions, so use the same ones.
233 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__fixdfsivfp");
234 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__fixunsdfsivfp");
235 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__fixsfsivfp");
236 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__fixunssfsivfp");
238 // Conversions between floating types.
239 setLibcallName(RTLIB::FPROUND_F64_F32, "__truncdfsf2vfp");
240 setLibcallName(RTLIB::FPEXT_F32_F64, "__extendsfdf2vfp");
242 // Integer to floating-point conversions.
243 // i64 conversions are done via library routines even when generating VFP
244 // instructions, so use the same ones.
245 // FIXME: There appears to be some naming inconsistency in ARM libgcc:
246 // e.g., __floatunsidf vs. __floatunssidfvfp.
247 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__floatsidfvfp");
248 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__floatunssidfvfp");
249 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__floatsisfvfp");
250 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__floatunssisfvfp");
254 // These libcalls are not available in 32-bit.
255 setLibcallName(RTLIB::SHL_I128, 0);
256 setLibcallName(RTLIB::SRL_I128, 0);
257 setLibcallName(RTLIB::SRA_I128, 0);
259 if (Subtarget->isAAPCS_ABI() && !Subtarget->isTargetDarwin()) {
260 // Double-precision floating-point arithmetic helper functions
261 // RTABI chapter 4.1.2, Table 2
262 setLibcallName(RTLIB::ADD_F64, "__aeabi_dadd");
263 setLibcallName(RTLIB::DIV_F64, "__aeabi_ddiv");
264 setLibcallName(RTLIB::MUL_F64, "__aeabi_dmul");
265 setLibcallName(RTLIB::SUB_F64, "__aeabi_dsub");
266 setLibcallCallingConv(RTLIB::ADD_F64, CallingConv::ARM_AAPCS);
267 setLibcallCallingConv(RTLIB::DIV_F64, CallingConv::ARM_AAPCS);
268 setLibcallCallingConv(RTLIB::MUL_F64, CallingConv::ARM_AAPCS);
269 setLibcallCallingConv(RTLIB::SUB_F64, CallingConv::ARM_AAPCS);
271 // Double-precision floating-point comparison helper functions
272 // RTABI chapter 4.1.2, Table 3
273 setLibcallName(RTLIB::OEQ_F64, "__aeabi_dcmpeq");
274 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
275 setLibcallName(RTLIB::UNE_F64, "__aeabi_dcmpeq");
276 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETEQ);
277 setLibcallName(RTLIB::OLT_F64, "__aeabi_dcmplt");
278 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
279 setLibcallName(RTLIB::OLE_F64, "__aeabi_dcmple");
280 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
281 setLibcallName(RTLIB::OGE_F64, "__aeabi_dcmpge");
282 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
283 setLibcallName(RTLIB::OGT_F64, "__aeabi_dcmpgt");
284 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
285 setLibcallName(RTLIB::UO_F64, "__aeabi_dcmpun");
286 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
287 setLibcallName(RTLIB::O_F64, "__aeabi_dcmpun");
288 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
289 setLibcallCallingConv(RTLIB::OEQ_F64, CallingConv::ARM_AAPCS);
290 setLibcallCallingConv(RTLIB::UNE_F64, CallingConv::ARM_AAPCS);
291 setLibcallCallingConv(RTLIB::OLT_F64, CallingConv::ARM_AAPCS);
292 setLibcallCallingConv(RTLIB::OLE_F64, CallingConv::ARM_AAPCS);
293 setLibcallCallingConv(RTLIB::OGE_F64, CallingConv::ARM_AAPCS);
294 setLibcallCallingConv(RTLIB::OGT_F64, CallingConv::ARM_AAPCS);
295 setLibcallCallingConv(RTLIB::UO_F64, CallingConv::ARM_AAPCS);
296 setLibcallCallingConv(RTLIB::O_F64, CallingConv::ARM_AAPCS);
298 // Single-precision floating-point arithmetic helper functions
299 // RTABI chapter 4.1.2, Table 4
300 setLibcallName(RTLIB::ADD_F32, "__aeabi_fadd");
301 setLibcallName(RTLIB::DIV_F32, "__aeabi_fdiv");
302 setLibcallName(RTLIB::MUL_F32, "__aeabi_fmul");
303 setLibcallName(RTLIB::SUB_F32, "__aeabi_fsub");
304 setLibcallCallingConv(RTLIB::ADD_F32, CallingConv::ARM_AAPCS);
305 setLibcallCallingConv(RTLIB::DIV_F32, CallingConv::ARM_AAPCS);
306 setLibcallCallingConv(RTLIB::MUL_F32, CallingConv::ARM_AAPCS);
307 setLibcallCallingConv(RTLIB::SUB_F32, CallingConv::ARM_AAPCS);
309 // Single-precision floating-point comparison helper functions
310 // RTABI chapter 4.1.2, Table 5
311 setLibcallName(RTLIB::OEQ_F32, "__aeabi_fcmpeq");
312 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
313 setLibcallName(RTLIB::UNE_F32, "__aeabi_fcmpeq");
314 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETEQ);
315 setLibcallName(RTLIB::OLT_F32, "__aeabi_fcmplt");
316 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
317 setLibcallName(RTLIB::OLE_F32, "__aeabi_fcmple");
318 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
319 setLibcallName(RTLIB::OGE_F32, "__aeabi_fcmpge");
320 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
321 setLibcallName(RTLIB::OGT_F32, "__aeabi_fcmpgt");
322 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
323 setLibcallName(RTLIB::UO_F32, "__aeabi_fcmpun");
324 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
325 setLibcallName(RTLIB::O_F32, "__aeabi_fcmpun");
326 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
327 setLibcallCallingConv(RTLIB::OEQ_F32, CallingConv::ARM_AAPCS);
328 setLibcallCallingConv(RTLIB::UNE_F32, CallingConv::ARM_AAPCS);
329 setLibcallCallingConv(RTLIB::OLT_F32, CallingConv::ARM_AAPCS);
330 setLibcallCallingConv(RTLIB::OLE_F32, CallingConv::ARM_AAPCS);
331 setLibcallCallingConv(RTLIB::OGE_F32, CallingConv::ARM_AAPCS);
332 setLibcallCallingConv(RTLIB::OGT_F32, CallingConv::ARM_AAPCS);
333 setLibcallCallingConv(RTLIB::UO_F32, CallingConv::ARM_AAPCS);
334 setLibcallCallingConv(RTLIB::O_F32, CallingConv::ARM_AAPCS);
336 // Floating-point to integer conversions.
337 // RTABI chapter 4.1.2, Table 6
338 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__aeabi_d2iz");
339 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__aeabi_d2uiz");
340 setLibcallName(RTLIB::FPTOSINT_F64_I64, "__aeabi_d2lz");
341 setLibcallName(RTLIB::FPTOUINT_F64_I64, "__aeabi_d2ulz");
342 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__aeabi_f2iz");
343 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__aeabi_f2uiz");
344 setLibcallName(RTLIB::FPTOSINT_F32_I64, "__aeabi_f2lz");
345 setLibcallName(RTLIB::FPTOUINT_F32_I64, "__aeabi_f2ulz");
346 setLibcallCallingConv(RTLIB::FPTOSINT_F64_I32, CallingConv::ARM_AAPCS);
347 setLibcallCallingConv(RTLIB::FPTOUINT_F64_I32, CallingConv::ARM_AAPCS);
348 setLibcallCallingConv(RTLIB::FPTOSINT_F64_I64, CallingConv::ARM_AAPCS);
349 setLibcallCallingConv(RTLIB::FPTOUINT_F64_I64, CallingConv::ARM_AAPCS);
350 setLibcallCallingConv(RTLIB::FPTOSINT_F32_I32, CallingConv::ARM_AAPCS);
351 setLibcallCallingConv(RTLIB::FPTOUINT_F32_I32, CallingConv::ARM_AAPCS);
352 setLibcallCallingConv(RTLIB::FPTOSINT_F32_I64, CallingConv::ARM_AAPCS);
353 setLibcallCallingConv(RTLIB::FPTOUINT_F32_I64, CallingConv::ARM_AAPCS);
355 // Conversions between floating types.
356 // RTABI chapter 4.1.2, Table 7
357 setLibcallName(RTLIB::FPROUND_F64_F32, "__aeabi_d2f");
358 setLibcallName(RTLIB::FPEXT_F32_F64, "__aeabi_f2d");
359 setLibcallCallingConv(RTLIB::FPROUND_F64_F32, CallingConv::ARM_AAPCS);
360 setLibcallCallingConv(RTLIB::FPEXT_F32_F64, CallingConv::ARM_AAPCS);
362 // Integer to floating-point conversions.
363 // RTABI chapter 4.1.2, Table 8
364 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__aeabi_i2d");
365 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__aeabi_ui2d");
366 setLibcallName(RTLIB::SINTTOFP_I64_F64, "__aeabi_l2d");
367 setLibcallName(RTLIB::UINTTOFP_I64_F64, "__aeabi_ul2d");
368 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__aeabi_i2f");
369 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__aeabi_ui2f");
370 setLibcallName(RTLIB::SINTTOFP_I64_F32, "__aeabi_l2f");
371 setLibcallName(RTLIB::UINTTOFP_I64_F32, "__aeabi_ul2f");
372 setLibcallCallingConv(RTLIB::SINTTOFP_I32_F64, CallingConv::ARM_AAPCS);
373 setLibcallCallingConv(RTLIB::UINTTOFP_I32_F64, CallingConv::ARM_AAPCS);
374 setLibcallCallingConv(RTLIB::SINTTOFP_I64_F64, CallingConv::ARM_AAPCS);
375 setLibcallCallingConv(RTLIB::UINTTOFP_I64_F64, CallingConv::ARM_AAPCS);
376 setLibcallCallingConv(RTLIB::SINTTOFP_I32_F32, CallingConv::ARM_AAPCS);
377 setLibcallCallingConv(RTLIB::UINTTOFP_I32_F32, CallingConv::ARM_AAPCS);
378 setLibcallCallingConv(RTLIB::SINTTOFP_I64_F32, CallingConv::ARM_AAPCS);
379 setLibcallCallingConv(RTLIB::UINTTOFP_I64_F32, CallingConv::ARM_AAPCS);
381 // Long long helper functions
382 // RTABI chapter 4.2, Table 9
383 setLibcallName(RTLIB::MUL_I64, "__aeabi_lmul");
384 setLibcallName(RTLIB::SHL_I64, "__aeabi_llsl");
385 setLibcallName(RTLIB::SRL_I64, "__aeabi_llsr");
386 setLibcallName(RTLIB::SRA_I64, "__aeabi_lasr");
387 setLibcallCallingConv(RTLIB::MUL_I64, CallingConv::ARM_AAPCS);
388 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::ARM_AAPCS);
389 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::ARM_AAPCS);
390 setLibcallCallingConv(RTLIB::SHL_I64, CallingConv::ARM_AAPCS);
391 setLibcallCallingConv(RTLIB::SRL_I64, CallingConv::ARM_AAPCS);
392 setLibcallCallingConv(RTLIB::SRA_I64, CallingConv::ARM_AAPCS);
394 // Integer division functions
395 // RTABI chapter 4.3.1
396 setLibcallName(RTLIB::SDIV_I8, "__aeabi_idiv");
397 setLibcallName(RTLIB::SDIV_I16, "__aeabi_idiv");
398 setLibcallName(RTLIB::SDIV_I32, "__aeabi_idiv");
399 setLibcallName(RTLIB::SDIV_I64, "__aeabi_ldivmod");
400 setLibcallName(RTLIB::UDIV_I8, "__aeabi_uidiv");
401 setLibcallName(RTLIB::UDIV_I16, "__aeabi_uidiv");
402 setLibcallName(RTLIB::UDIV_I32, "__aeabi_uidiv");
403 setLibcallName(RTLIB::UDIV_I64, "__aeabi_uldivmod");
404 setLibcallCallingConv(RTLIB::SDIV_I8, CallingConv::ARM_AAPCS);
405 setLibcallCallingConv(RTLIB::SDIV_I16, CallingConv::ARM_AAPCS);
406 setLibcallCallingConv(RTLIB::SDIV_I32, CallingConv::ARM_AAPCS);
407 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::ARM_AAPCS);
408 setLibcallCallingConv(RTLIB::UDIV_I8, CallingConv::ARM_AAPCS);
409 setLibcallCallingConv(RTLIB::UDIV_I16, CallingConv::ARM_AAPCS);
410 setLibcallCallingConv(RTLIB::UDIV_I32, CallingConv::ARM_AAPCS);
411 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::ARM_AAPCS);
414 // RTABI chapter 4.3.4
415 setLibcallName(RTLIB::MEMCPY, "__aeabi_memcpy");
416 setLibcallName(RTLIB::MEMMOVE, "__aeabi_memmove");
417 setLibcallName(RTLIB::MEMSET, "__aeabi_memset");
418 setLibcallCallingConv(RTLIB::MEMCPY, CallingConv::ARM_AAPCS);
419 setLibcallCallingConv(RTLIB::MEMMOVE, CallingConv::ARM_AAPCS);
420 setLibcallCallingConv(RTLIB::MEMSET, CallingConv::ARM_AAPCS);
423 // Use divmod compiler-rt calls for iOS 5.0 and later.
424 if (Subtarget->getTargetTriple().getOS() == Triple::IOS &&
425 !Subtarget->getTargetTriple().isOSVersionLT(5, 0)) {
426 setLibcallName(RTLIB::SDIVREM_I32, "__divmodsi4");
427 setLibcallName(RTLIB::UDIVREM_I32, "__udivmodsi4");
430 if (Subtarget->isThumb1Only())
431 addRegisterClass(MVT::i32, &ARM::tGPRRegClass);
433 addRegisterClass(MVT::i32, &ARM::GPRRegClass);
434 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
435 !Subtarget->isThumb1Only()) {
436 addRegisterClass(MVT::f32, &ARM::SPRRegClass);
437 if (!Subtarget->isFPOnlySP())
438 addRegisterClass(MVT::f64, &ARM::DPRRegClass);
440 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
443 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
444 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
445 for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
446 InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
447 setTruncStoreAction((MVT::SimpleValueType)VT,
448 (MVT::SimpleValueType)InnerVT, Expand);
449 setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT, Expand);
450 setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT, Expand);
451 setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT, Expand);
454 setOperationAction(ISD::ConstantFP, MVT::f32, Custom);
456 if (Subtarget->hasNEON()) {
457 addDRTypeForNEON(MVT::v2f32);
458 addDRTypeForNEON(MVT::v8i8);
459 addDRTypeForNEON(MVT::v4i16);
460 addDRTypeForNEON(MVT::v2i32);
461 addDRTypeForNEON(MVT::v1i64);
463 addQRTypeForNEON(MVT::v4f32);
464 addQRTypeForNEON(MVT::v2f64);
465 addQRTypeForNEON(MVT::v16i8);
466 addQRTypeForNEON(MVT::v8i16);
467 addQRTypeForNEON(MVT::v4i32);
468 addQRTypeForNEON(MVT::v2i64);
470 // v2f64 is legal so that QR subregs can be extracted as f64 elements, but
471 // neither Neon nor VFP support any arithmetic operations on it.
472 // The same with v4f32. But keep in mind that vadd, vsub, vmul are natively
473 // supported for v4f32.
474 setOperationAction(ISD::FADD, MVT::v2f64, Expand);
475 setOperationAction(ISD::FSUB, MVT::v2f64, Expand);
476 setOperationAction(ISD::FMUL, MVT::v2f64, Expand);
477 // FIXME: Code duplication: FDIV and FREM are expanded always, see
478 // ARMTargetLowering::addTypeForNEON method for details.
479 setOperationAction(ISD::FDIV, MVT::v2f64, Expand);
480 setOperationAction(ISD::FREM, MVT::v2f64, Expand);
481 // FIXME: Create unittest.
482 // In another words, find a way when "copysign" appears in DAG with vector
484 setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand);
485 // FIXME: Code duplication: SETCC has custom operation action, see
486 // ARMTargetLowering::addTypeForNEON method for details.
487 setOperationAction(ISD::SETCC, MVT::v2f64, Expand);
488 // FIXME: Create unittest for FNEG and for FABS.
489 setOperationAction(ISD::FNEG, MVT::v2f64, Expand);
490 setOperationAction(ISD::FABS, MVT::v2f64, Expand);
491 setOperationAction(ISD::FSQRT, MVT::v2f64, Expand);
492 setOperationAction(ISD::FSIN, MVT::v2f64, Expand);
493 setOperationAction(ISD::FCOS, MVT::v2f64, Expand);
494 setOperationAction(ISD::FPOWI, MVT::v2f64, Expand);
495 setOperationAction(ISD::FPOW, MVT::v2f64, Expand);
496 setOperationAction(ISD::FLOG, MVT::v2f64, Expand);
497 setOperationAction(ISD::FLOG2, MVT::v2f64, Expand);
498 setOperationAction(ISD::FLOG10, MVT::v2f64, Expand);
499 setOperationAction(ISD::FEXP, MVT::v2f64, Expand);
500 setOperationAction(ISD::FEXP2, MVT::v2f64, Expand);
501 // FIXME: Create unittest for FCEIL, FTRUNC, FRINT, FNEARBYINT, FFLOOR.
502 setOperationAction(ISD::FCEIL, MVT::v2f64, Expand);
503 setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand);
504 setOperationAction(ISD::FRINT, MVT::v2f64, Expand);
505 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand);
506 setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand);
508 setOperationAction(ISD::FSQRT, MVT::v4f32, Expand);
509 setOperationAction(ISD::FSIN, MVT::v4f32, Expand);
510 setOperationAction(ISD::FCOS, MVT::v4f32, Expand);
511 setOperationAction(ISD::FPOWI, MVT::v4f32, Expand);
512 setOperationAction(ISD::FPOW, MVT::v4f32, Expand);
513 setOperationAction(ISD::FLOG, MVT::v4f32, Expand);
514 setOperationAction(ISD::FLOG2, MVT::v4f32, Expand);
515 setOperationAction(ISD::FLOG10, MVT::v4f32, Expand);
516 setOperationAction(ISD::FEXP, MVT::v4f32, Expand);
517 setOperationAction(ISD::FEXP2, MVT::v4f32, Expand);
518 setOperationAction(ISD::FFLOOR, MVT::v4f32, Expand);
520 // Neon does not support some operations on v1i64 and v2i64 types.
521 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
522 // Custom handling for some quad-vector types to detect VMULL.
523 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
524 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
525 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
526 // Custom handling for some vector types to avoid expensive expansions
527 setOperationAction(ISD::SDIV, MVT::v4i16, Custom);
528 setOperationAction(ISD::SDIV, MVT::v8i8, Custom);
529 setOperationAction(ISD::UDIV, MVT::v4i16, Custom);
530 setOperationAction(ISD::UDIV, MVT::v8i8, Custom);
531 setOperationAction(ISD::SETCC, MVT::v1i64, Expand);
532 setOperationAction(ISD::SETCC, MVT::v2i64, Expand);
533 // Neon does not have single instruction SINT_TO_FP and UINT_TO_FP with
534 // a destination type that is wider than the source, and nor does
535 // it have a FP_TO_[SU]INT instruction with a narrower destination than
537 setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Custom);
538 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
539 setOperationAction(ISD::FP_TO_UINT, MVT::v4i16, Custom);
540 setOperationAction(ISD::FP_TO_SINT, MVT::v4i16, Custom);
542 setTargetDAGCombine(ISD::INTRINSIC_VOID);
543 setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN);
544 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
545 setTargetDAGCombine(ISD::SHL);
546 setTargetDAGCombine(ISD::SRL);
547 setTargetDAGCombine(ISD::SRA);
548 setTargetDAGCombine(ISD::SIGN_EXTEND);
549 setTargetDAGCombine(ISD::ZERO_EXTEND);
550 setTargetDAGCombine(ISD::ANY_EXTEND);
551 setTargetDAGCombine(ISD::SELECT_CC);
552 setTargetDAGCombine(ISD::BUILD_VECTOR);
553 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
554 setTargetDAGCombine(ISD::INSERT_VECTOR_ELT);
555 setTargetDAGCombine(ISD::STORE);
556 setTargetDAGCombine(ISD::FP_TO_SINT);
557 setTargetDAGCombine(ISD::FP_TO_UINT);
558 setTargetDAGCombine(ISD::FDIV);
560 // It is legal to extload from v4i8 to v4i16 or v4i32.
561 MVT Tys[6] = {MVT::v8i8, MVT::v4i8, MVT::v2i8,
562 MVT::v4i16, MVT::v2i16,
564 for (unsigned i = 0; i < 6; ++i) {
565 setLoadExtAction(ISD::EXTLOAD, Tys[i], Legal);
566 setLoadExtAction(ISD::ZEXTLOAD, Tys[i], Legal);
567 setLoadExtAction(ISD::SEXTLOAD, Tys[i], Legal);
571 // ARM and Thumb2 support UMLAL/SMLAL.
572 if (!Subtarget->isThumb1Only())
573 setTargetDAGCombine(ISD::ADDC);
576 computeRegisterProperties();
578 // ARM does not have f32 extending load.
579 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
581 // ARM does not have i1 sign extending load.
582 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
584 // ARM supports all 4 flavors of integer indexed load / store.
585 if (!Subtarget->isThumb1Only()) {
586 for (unsigned im = (unsigned)ISD::PRE_INC;
587 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
588 setIndexedLoadAction(im, MVT::i1, Legal);
589 setIndexedLoadAction(im, MVT::i8, Legal);
590 setIndexedLoadAction(im, MVT::i16, Legal);
591 setIndexedLoadAction(im, MVT::i32, Legal);
592 setIndexedStoreAction(im, MVT::i1, Legal);
593 setIndexedStoreAction(im, MVT::i8, Legal);
594 setIndexedStoreAction(im, MVT::i16, Legal);
595 setIndexedStoreAction(im, MVT::i32, Legal);
599 // i64 operation support.
600 setOperationAction(ISD::MUL, MVT::i64, Expand);
601 setOperationAction(ISD::MULHU, MVT::i32, Expand);
602 if (Subtarget->isThumb1Only()) {
603 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
604 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
606 if (Subtarget->isThumb1Only() || !Subtarget->hasV6Ops()
607 || (Subtarget->isThumb2() && !Subtarget->hasThumb2DSP()))
608 setOperationAction(ISD::MULHS, MVT::i32, Expand);
610 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
611 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
612 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
613 setOperationAction(ISD::SRL, MVT::i64, Custom);
614 setOperationAction(ISD::SRA, MVT::i64, Custom);
616 if (!Subtarget->isThumb1Only()) {
617 // FIXME: We should do this for Thumb1 as well.
618 setOperationAction(ISD::ADDC, MVT::i32, Custom);
619 setOperationAction(ISD::ADDE, MVT::i32, Custom);
620 setOperationAction(ISD::SUBC, MVT::i32, Custom);
621 setOperationAction(ISD::SUBE, MVT::i32, Custom);
624 // ARM does not have ROTL.
625 setOperationAction(ISD::ROTL, MVT::i32, Expand);
626 setOperationAction(ISD::CTTZ, MVT::i32, Custom);
627 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
628 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
629 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
631 // These just redirect to CTTZ and CTLZ on ARM.
632 setOperationAction(ISD::CTTZ_ZERO_UNDEF , MVT::i32 , Expand);
633 setOperationAction(ISD::CTLZ_ZERO_UNDEF , MVT::i32 , Expand);
635 // Only ARMv6 has BSWAP.
636 if (!Subtarget->hasV6Ops())
637 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
639 if (!(Subtarget->hasDivide() && Subtarget->isThumb2()) &&
640 !(Subtarget->hasDivideInARMMode() && !Subtarget->isThumb())) {
641 // These are expanded into libcalls if the cpu doesn't have HW divider.
642 setOperationAction(ISD::SDIV, MVT::i32, Expand);
643 setOperationAction(ISD::UDIV, MVT::i32, Expand);
645 setOperationAction(ISD::SREM, MVT::i32, Expand);
646 setOperationAction(ISD::UREM, MVT::i32, Expand);
647 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
648 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
650 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
651 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
652 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
653 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
654 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
656 setOperationAction(ISD::TRAP, MVT::Other, Legal);
658 // Use the default implementation.
659 setOperationAction(ISD::VASTART, MVT::Other, Custom);
660 setOperationAction(ISD::VAARG, MVT::Other, Expand);
661 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
662 setOperationAction(ISD::VAEND, MVT::Other, Expand);
663 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
664 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
666 if (!Subtarget->isTargetDarwin()) {
667 // Non-Darwin platforms may return values in these registers via the
668 // personality function.
669 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
670 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
671 setExceptionPointerRegister(ARM::R0);
672 setExceptionSelectorRegister(ARM::R1);
675 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
676 // ARMv6 Thumb1 (except for CPUs that support dmb / dsb) and earlier use
677 // the default expansion.
678 // FIXME: This should be checking for v6k, not just v6.
679 if (Subtarget->hasDataBarrier() ||
680 (Subtarget->hasV6Ops() && !Subtarget->isThumb())) {
681 // membarrier needs custom lowering; the rest are legal and handled
683 setOperationAction(ISD::MEMBARRIER, MVT::Other, Custom);
684 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Custom);
685 // Custom lowering for 64-bit ops
686 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
687 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
688 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);
689 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom);
690 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
691 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
692 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Custom);
693 // Automatically insert fences (dmb ist) around ATOMIC_SWAP etc.
694 setInsertFencesForAtomic(true);
696 // Set them all for expansion, which will force libcalls.
697 setOperationAction(ISD::MEMBARRIER, MVT::Other, Expand);
698 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Expand);
699 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Expand);
700 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Expand);
701 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Expand);
702 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Expand);
703 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Expand);
704 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Expand);
705 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Expand);
706 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Expand);
707 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i32, Expand);
708 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i32, Expand);
709 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i32, Expand);
710 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i32, Expand);
711 // Mark ATOMIC_LOAD and ATOMIC_STORE custom so we can handle the
712 // Unordered/Monotonic case.
713 setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Custom);
714 setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Custom);
715 // Since the libcalls include locking, fold in the fences
716 setShouldFoldAtomicFences(true);
719 setOperationAction(ISD::PREFETCH, MVT::Other, Custom);
721 // Requires SXTB/SXTH, available on v6 and up in both ARM and Thumb modes.
722 if (!Subtarget->hasV6Ops()) {
723 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
724 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
726 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
728 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
729 !Subtarget->isThumb1Only()) {
730 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
731 // iff target supports vfp2.
732 setOperationAction(ISD::BITCAST, MVT::i64, Custom);
733 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
736 // We want to custom lower some of our intrinsics.
737 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
738 if (Subtarget->isTargetDarwin()) {
739 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
740 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
741 setLibcallName(RTLIB::UNWIND_RESUME, "_Unwind_SjLj_Resume");
744 setOperationAction(ISD::SETCC, MVT::i32, Expand);
745 setOperationAction(ISD::SETCC, MVT::f32, Expand);
746 setOperationAction(ISD::SETCC, MVT::f64, Expand);
747 setOperationAction(ISD::SELECT, MVT::i32, Custom);
748 setOperationAction(ISD::SELECT, MVT::f32, Custom);
749 setOperationAction(ISD::SELECT, MVT::f64, Custom);
750 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
751 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
752 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
754 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
755 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
756 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
757 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
758 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
760 // We don't support sin/cos/fmod/copysign/pow
761 setOperationAction(ISD::FSIN, MVT::f64, Expand);
762 setOperationAction(ISD::FSIN, MVT::f32, Expand);
763 setOperationAction(ISD::FCOS, MVT::f32, Expand);
764 setOperationAction(ISD::FCOS, MVT::f64, Expand);
765 setOperationAction(ISD::FREM, MVT::f64, Expand);
766 setOperationAction(ISD::FREM, MVT::f32, Expand);
767 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
768 !Subtarget->isThumb1Only()) {
769 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
770 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
772 setOperationAction(ISD::FPOW, MVT::f64, Expand);
773 setOperationAction(ISD::FPOW, MVT::f32, Expand);
775 if (!Subtarget->hasVFP4()) {
776 setOperationAction(ISD::FMA, MVT::f64, Expand);
777 setOperationAction(ISD::FMA, MVT::f32, Expand);
780 // Various VFP goodness
781 if (!TM.Options.UseSoftFloat && !Subtarget->isThumb1Only()) {
782 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
783 if (Subtarget->hasVFP2()) {
784 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
785 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
786 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
787 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
789 // Special handling for half-precision FP.
790 if (!Subtarget->hasFP16()) {
791 setOperationAction(ISD::FP16_TO_FP32, MVT::f32, Expand);
792 setOperationAction(ISD::FP32_TO_FP16, MVT::i32, Expand);
796 // We have target-specific dag combine patterns for the following nodes:
797 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
798 setTargetDAGCombine(ISD::ADD);
799 setTargetDAGCombine(ISD::SUB);
800 setTargetDAGCombine(ISD::MUL);
801 setTargetDAGCombine(ISD::AND);
802 setTargetDAGCombine(ISD::OR);
803 setTargetDAGCombine(ISD::XOR);
805 if (Subtarget->hasV6Ops())
806 setTargetDAGCombine(ISD::SRL);
808 setStackPointerRegisterToSaveRestore(ARM::SP);
810 if (TM.Options.UseSoftFloat || Subtarget->isThumb1Only() ||
811 !Subtarget->hasVFP2())
812 setSchedulingPreference(Sched::RegPressure);
814 setSchedulingPreference(Sched::Hybrid);
816 //// temporary - rewrite interface to use type
817 maxStoresPerMemcpy = maxStoresPerMemcpyOptSize = 1;
818 maxStoresPerMemset = 16;
819 maxStoresPerMemsetOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
821 // On ARM arguments smaller than 4 bytes are extended, so all arguments
822 // are at least 4 bytes aligned.
823 setMinStackArgumentAlignment(4);
825 benefitFromCodePlacementOpt = true;
827 // Prefer likely predicted branches to selects on out-of-order cores.
828 predictableSelectIsExpensive = Subtarget->isLikeA9();
830 setMinFunctionAlignment(Subtarget->isThumb() ? 1 : 2);
833 // FIXME: It might make sense to define the representative register class as the
834 // nearest super-register that has a non-null superset. For example, DPR_VFP2 is
835 // a super-register of SPR, and DPR is a superset if DPR_VFP2. Consequently,
836 // SPR's representative would be DPR_VFP2. This should work well if register
837 // pressure tracking were modified such that a register use would increment the
838 // pressure of the register class's representative and all of it's super
839 // classes' representatives transitively. We have not implemented this because
840 // of the difficulty prior to coalescing of modeling operand register classes
841 // due to the common occurrence of cross class copies and subregister insertions
843 std::pair<const TargetRegisterClass*, uint8_t>
844 ARMTargetLowering::findRepresentativeClass(EVT VT) const{
845 const TargetRegisterClass *RRC = 0;
847 switch (VT.getSimpleVT().SimpleTy) {
849 return TargetLowering::findRepresentativeClass(VT);
850 // Use DPR as representative register class for all floating point
851 // and vector types. Since there are 32 SPR registers and 32 DPR registers so
852 // the cost is 1 for both f32 and f64.
853 case MVT::f32: case MVT::f64: case MVT::v8i8: case MVT::v4i16:
854 case MVT::v2i32: case MVT::v1i64: case MVT::v2f32:
855 RRC = &ARM::DPRRegClass;
856 // When NEON is used for SP, only half of the register file is available
857 // because operations that define both SP and DP results will be constrained
858 // to the VFP2 class (D0-D15). We currently model this constraint prior to
859 // coalescing by double-counting the SP regs. See the FIXME above.
860 if (Subtarget->useNEONForSinglePrecisionFP())
863 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
864 case MVT::v4f32: case MVT::v2f64:
865 RRC = &ARM::DPRRegClass;
869 RRC = &ARM::DPRRegClass;
873 RRC = &ARM::DPRRegClass;
877 return std::make_pair(RRC, Cost);
880 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
883 case ARMISD::Wrapper: return "ARMISD::Wrapper";
884 case ARMISD::WrapperDYN: return "ARMISD::WrapperDYN";
885 case ARMISD::WrapperPIC: return "ARMISD::WrapperPIC";
886 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
887 case ARMISD::CALL: return "ARMISD::CALL";
888 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
889 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
890 case ARMISD::tCALL: return "ARMISD::tCALL";
891 case ARMISD::BRCOND: return "ARMISD::BRCOND";
892 case ARMISD::BR_JT: return "ARMISD::BR_JT";
893 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
894 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
895 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
896 case ARMISD::CMP: return "ARMISD::CMP";
897 case ARMISD::CMN: return "ARMISD::CMN";
898 case ARMISD::CMPZ: return "ARMISD::CMPZ";
899 case ARMISD::CMPFP: return "ARMISD::CMPFP";
900 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
901 case ARMISD::BCC_i64: return "ARMISD::BCC_i64";
902 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
904 case ARMISD::CMOV: return "ARMISD::CMOV";
906 case ARMISD::RBIT: return "ARMISD::RBIT";
908 case ARMISD::FTOSI: return "ARMISD::FTOSI";
909 case ARMISD::FTOUI: return "ARMISD::FTOUI";
910 case ARMISD::SITOF: return "ARMISD::SITOF";
911 case ARMISD::UITOF: return "ARMISD::UITOF";
913 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
914 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
915 case ARMISD::RRX: return "ARMISD::RRX";
917 case ARMISD::ADDC: return "ARMISD::ADDC";
918 case ARMISD::ADDE: return "ARMISD::ADDE";
919 case ARMISD::SUBC: return "ARMISD::SUBC";
920 case ARMISD::SUBE: return "ARMISD::SUBE";
922 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
923 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
925 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
926 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
928 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN";
930 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
932 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
934 case ARMISD::MEMBARRIER: return "ARMISD::MEMBARRIER";
935 case ARMISD::MEMBARRIER_MCR: return "ARMISD::MEMBARRIER_MCR";
937 case ARMISD::PRELOAD: return "ARMISD::PRELOAD";
939 case ARMISD::VCEQ: return "ARMISD::VCEQ";
940 case ARMISD::VCEQZ: return "ARMISD::VCEQZ";
941 case ARMISD::VCGE: return "ARMISD::VCGE";
942 case ARMISD::VCGEZ: return "ARMISD::VCGEZ";
943 case ARMISD::VCLEZ: return "ARMISD::VCLEZ";
944 case ARMISD::VCGEU: return "ARMISD::VCGEU";
945 case ARMISD::VCGT: return "ARMISD::VCGT";
946 case ARMISD::VCGTZ: return "ARMISD::VCGTZ";
947 case ARMISD::VCLTZ: return "ARMISD::VCLTZ";
948 case ARMISD::VCGTU: return "ARMISD::VCGTU";
949 case ARMISD::VTST: return "ARMISD::VTST";
951 case ARMISD::VSHL: return "ARMISD::VSHL";
952 case ARMISD::VSHRs: return "ARMISD::VSHRs";
953 case ARMISD::VSHRu: return "ARMISD::VSHRu";
954 case ARMISD::VSHLLs: return "ARMISD::VSHLLs";
955 case ARMISD::VSHLLu: return "ARMISD::VSHLLu";
956 case ARMISD::VSHLLi: return "ARMISD::VSHLLi";
957 case ARMISD::VSHRN: return "ARMISD::VSHRN";
958 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
959 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
960 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
961 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
962 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
963 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
964 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
965 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
966 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
967 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
968 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
969 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
970 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
971 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
972 case ARMISD::VMOVIMM: return "ARMISD::VMOVIMM";
973 case ARMISD::VMVNIMM: return "ARMISD::VMVNIMM";
974 case ARMISD::VMOVFPIMM: return "ARMISD::VMOVFPIMM";
975 case ARMISD::VDUP: return "ARMISD::VDUP";
976 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
977 case ARMISD::VEXT: return "ARMISD::VEXT";
978 case ARMISD::VREV64: return "ARMISD::VREV64";
979 case ARMISD::VREV32: return "ARMISD::VREV32";
980 case ARMISD::VREV16: return "ARMISD::VREV16";
981 case ARMISD::VZIP: return "ARMISD::VZIP";
982 case ARMISD::VUZP: return "ARMISD::VUZP";
983 case ARMISD::VTRN: return "ARMISD::VTRN";
984 case ARMISD::VTBL1: return "ARMISD::VTBL1";
985 case ARMISD::VTBL2: return "ARMISD::VTBL2";
986 case ARMISD::VMULLs: return "ARMISD::VMULLs";
987 case ARMISD::VMULLu: return "ARMISD::VMULLu";
988 case ARMISD::UMLAL: return "ARMISD::UMLAL";
989 case ARMISD::SMLAL: return "ARMISD::SMLAL";
990 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
991 case ARMISD::FMAX: return "ARMISD::FMAX";
992 case ARMISD::FMIN: return "ARMISD::FMIN";
993 case ARMISD::BFI: return "ARMISD::BFI";
994 case ARMISD::VORRIMM: return "ARMISD::VORRIMM";
995 case ARMISD::VBICIMM: return "ARMISD::VBICIMM";
996 case ARMISD::VBSL: return "ARMISD::VBSL";
997 case ARMISD::VLD2DUP: return "ARMISD::VLD2DUP";
998 case ARMISD::VLD3DUP: return "ARMISD::VLD3DUP";
999 case ARMISD::VLD4DUP: return "ARMISD::VLD4DUP";
1000 case ARMISD::VLD1_UPD: return "ARMISD::VLD1_UPD";
1001 case ARMISD::VLD2_UPD: return "ARMISD::VLD2_UPD";
1002 case ARMISD::VLD3_UPD: return "ARMISD::VLD3_UPD";
1003 case ARMISD::VLD4_UPD: return "ARMISD::VLD4_UPD";
1004 case ARMISD::VLD2LN_UPD: return "ARMISD::VLD2LN_UPD";
1005 case ARMISD::VLD3LN_UPD: return "ARMISD::VLD3LN_UPD";
1006 case ARMISD::VLD4LN_UPD: return "ARMISD::VLD4LN_UPD";
1007 case ARMISD::VLD2DUP_UPD: return "ARMISD::VLD2DUP_UPD";
1008 case ARMISD::VLD3DUP_UPD: return "ARMISD::VLD3DUP_UPD";
1009 case ARMISD::VLD4DUP_UPD: return "ARMISD::VLD4DUP_UPD";
1010 case ARMISD::VST1_UPD: return "ARMISD::VST1_UPD";
1011 case ARMISD::VST2_UPD: return "ARMISD::VST2_UPD";
1012 case ARMISD::VST3_UPD: return "ARMISD::VST3_UPD";
1013 case ARMISD::VST4_UPD: return "ARMISD::VST4_UPD";
1014 case ARMISD::VST2LN_UPD: return "ARMISD::VST2LN_UPD";
1015 case ARMISD::VST3LN_UPD: return "ARMISD::VST3LN_UPD";
1016 case ARMISD::VST4LN_UPD: return "ARMISD::VST4LN_UPD";
1020 EVT ARMTargetLowering::getSetCCResultType(EVT VT) const {
1021 if (!VT.isVector()) return getPointerTy();
1022 return VT.changeVectorElementTypeToInteger();
1025 /// getRegClassFor - Return the register class that should be used for the
1026 /// specified value type.
1027 const TargetRegisterClass *ARMTargetLowering::getRegClassFor(EVT VT) const {
1028 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
1029 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
1030 // load / store 4 to 8 consecutive D registers.
1031 if (Subtarget->hasNEON()) {
1032 if (VT == MVT::v4i64)
1033 return &ARM::QQPRRegClass;
1034 if (VT == MVT::v8i64)
1035 return &ARM::QQQQPRRegClass;
1037 return TargetLowering::getRegClassFor(VT);
1040 // Create a fast isel object.
1042 ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
1043 const TargetLibraryInfo *libInfo) const {
1044 return ARM::createFastISel(funcInfo, libInfo);
1047 /// getMaximalGlobalOffset - Returns the maximal possible offset which can
1048 /// be used for loads / stores from the global.
1049 unsigned ARMTargetLowering::getMaximalGlobalOffset() const {
1050 return (Subtarget->isThumb1Only() ? 127 : 4095);
1053 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
1054 unsigned NumVals = N->getNumValues();
1056 return Sched::RegPressure;
1058 for (unsigned i = 0; i != NumVals; ++i) {
1059 EVT VT = N->getValueType(i);
1060 if (VT == MVT::Glue || VT == MVT::Other)
1062 if (VT.isFloatingPoint() || VT.isVector())
1066 if (!N->isMachineOpcode())
1067 return Sched::RegPressure;
1069 // Load are scheduled for latency even if there instruction itinerary
1070 // is not available.
1071 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1072 const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
1074 if (MCID.getNumDefs() == 0)
1075 return Sched::RegPressure;
1076 if (!Itins->isEmpty() &&
1077 Itins->getOperandCycle(MCID.getSchedClass(), 0) > 2)
1080 return Sched::RegPressure;
1083 //===----------------------------------------------------------------------===//
1085 //===----------------------------------------------------------------------===//
1087 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
1088 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
1090 default: llvm_unreachable("Unknown condition code!");
1091 case ISD::SETNE: return ARMCC::NE;
1092 case ISD::SETEQ: return ARMCC::EQ;
1093 case ISD::SETGT: return ARMCC::GT;
1094 case ISD::SETGE: return ARMCC::GE;
1095 case ISD::SETLT: return ARMCC::LT;
1096 case ISD::SETLE: return ARMCC::LE;
1097 case ISD::SETUGT: return ARMCC::HI;
1098 case ISD::SETUGE: return ARMCC::HS;
1099 case ISD::SETULT: return ARMCC::LO;
1100 case ISD::SETULE: return ARMCC::LS;
1104 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
1105 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
1106 ARMCC::CondCodes &CondCode2) {
1107 CondCode2 = ARMCC::AL;
1109 default: llvm_unreachable("Unknown FP condition!");
1111 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
1113 case ISD::SETOGT: CondCode = ARMCC::GT; break;
1115 case ISD::SETOGE: CondCode = ARMCC::GE; break;
1116 case ISD::SETOLT: CondCode = ARMCC::MI; break;
1117 case ISD::SETOLE: CondCode = ARMCC::LS; break;
1118 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
1119 case ISD::SETO: CondCode = ARMCC::VC; break;
1120 case ISD::SETUO: CondCode = ARMCC::VS; break;
1121 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
1122 case ISD::SETUGT: CondCode = ARMCC::HI; break;
1123 case ISD::SETUGE: CondCode = ARMCC::PL; break;
1125 case ISD::SETULT: CondCode = ARMCC::LT; break;
1127 case ISD::SETULE: CondCode = ARMCC::LE; break;
1129 case ISD::SETUNE: CondCode = ARMCC::NE; break;
1133 //===----------------------------------------------------------------------===//
1134 // Calling Convention Implementation
1135 //===----------------------------------------------------------------------===//
1137 #include "ARMGenCallingConv.inc"
1139 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1140 /// given CallingConvention value.
1141 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
1143 bool isVarArg) const {
1146 llvm_unreachable("Unsupported calling convention");
1147 case CallingConv::Fast:
1148 if (Subtarget->hasVFP2() && !isVarArg) {
1149 if (!Subtarget->isAAPCS_ABI())
1150 return (Return ? RetFastCC_ARM_APCS : FastCC_ARM_APCS);
1151 // For AAPCS ABI targets, just use VFP variant of the calling convention.
1152 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1155 case CallingConv::C: {
1156 // Use target triple & subtarget features to do actual dispatch.
1157 if (!Subtarget->isAAPCS_ABI())
1158 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1159 else if (Subtarget->hasVFP2() &&
1160 getTargetMachine().Options.FloatABIType == FloatABI::Hard &&
1162 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1163 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1165 case CallingConv::ARM_AAPCS_VFP:
1167 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1169 case CallingConv::ARM_AAPCS:
1170 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1171 case CallingConv::ARM_APCS:
1172 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1173 case CallingConv::GHC:
1174 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS_GHC);
1178 /// LowerCallResult - Lower the result values of a call into the
1179 /// appropriate copies out of appropriate physical registers.
1181 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1182 CallingConv::ID CallConv, bool isVarArg,
1183 const SmallVectorImpl<ISD::InputArg> &Ins,
1184 DebugLoc dl, SelectionDAG &DAG,
1185 SmallVectorImpl<SDValue> &InVals) const {
1187 // Assign locations to each value returned by this call.
1188 SmallVector<CCValAssign, 16> RVLocs;
1189 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1190 getTargetMachine(), RVLocs, *DAG.getContext(), Call);
1191 CCInfo.AnalyzeCallResult(Ins,
1192 CCAssignFnForNode(CallConv, /* Return*/ true,
1195 // Copy all of the result registers out of their specified physreg.
1196 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1197 CCValAssign VA = RVLocs[i];
1200 if (VA.needsCustom()) {
1201 // Handle f64 or half of a v2f64.
1202 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1204 Chain = Lo.getValue(1);
1205 InFlag = Lo.getValue(2);
1206 VA = RVLocs[++i]; // skip ahead to next loc
1207 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1209 Chain = Hi.getValue(1);
1210 InFlag = Hi.getValue(2);
1211 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1213 if (VA.getLocVT() == MVT::v2f64) {
1214 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1215 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1216 DAG.getConstant(0, MVT::i32));
1218 VA = RVLocs[++i]; // skip ahead to next loc
1219 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1220 Chain = Lo.getValue(1);
1221 InFlag = Lo.getValue(2);
1222 VA = RVLocs[++i]; // skip ahead to next loc
1223 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1224 Chain = Hi.getValue(1);
1225 InFlag = Hi.getValue(2);
1226 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1227 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1228 DAG.getConstant(1, MVT::i32));
1231 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
1233 Chain = Val.getValue(1);
1234 InFlag = Val.getValue(2);
1237 switch (VA.getLocInfo()) {
1238 default: llvm_unreachable("Unknown loc info!");
1239 case CCValAssign::Full: break;
1240 case CCValAssign::BCvt:
1241 Val = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), Val);
1245 InVals.push_back(Val);
1251 /// LowerMemOpCallTo - Store the argument to the stack.
1253 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
1254 SDValue StackPtr, SDValue Arg,
1255 DebugLoc dl, SelectionDAG &DAG,
1256 const CCValAssign &VA,
1257 ISD::ArgFlagsTy Flags) const {
1258 unsigned LocMemOffset = VA.getLocMemOffset();
1259 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1260 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1261 return DAG.getStore(Chain, dl, Arg, PtrOff,
1262 MachinePointerInfo::getStack(LocMemOffset),
1266 void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
1267 SDValue Chain, SDValue &Arg,
1268 RegsToPassVector &RegsToPass,
1269 CCValAssign &VA, CCValAssign &NextVA,
1271 SmallVector<SDValue, 8> &MemOpChains,
1272 ISD::ArgFlagsTy Flags) const {
1274 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1275 DAG.getVTList(MVT::i32, MVT::i32), Arg);
1276 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
1278 if (NextVA.isRegLoc())
1279 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
1281 assert(NextVA.isMemLoc());
1282 if (StackPtr.getNode() == 0)
1283 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1285 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
1291 /// LowerCall - Lowering a call into a callseq_start <-
1292 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
1295 ARMTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
1296 SmallVectorImpl<SDValue> &InVals) const {
1297 SelectionDAG &DAG = CLI.DAG;
1298 DebugLoc &dl = CLI.DL;
1299 SmallVector<ISD::OutputArg, 32> &Outs = CLI.Outs;
1300 SmallVector<SDValue, 32> &OutVals = CLI.OutVals;
1301 SmallVector<ISD::InputArg, 32> &Ins = CLI.Ins;
1302 SDValue Chain = CLI.Chain;
1303 SDValue Callee = CLI.Callee;
1304 bool &isTailCall = CLI.IsTailCall;
1305 CallingConv::ID CallConv = CLI.CallConv;
1306 bool doesNotRet = CLI.DoesNotReturn;
1307 bool isVarArg = CLI.IsVarArg;
1309 MachineFunction &MF = DAG.getMachineFunction();
1310 bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1311 bool IsSibCall = false;
1312 // Disable tail calls if they're not supported.
1313 if (!EnableARMTailCalls && !Subtarget->supportsTailCall())
1316 // Check if it's really possible to do a tail call.
1317 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1318 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1319 Outs, OutVals, Ins, DAG);
1320 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1321 // detected sibcalls.
1328 // Analyze operands of the call, assigning locations to each operand.
1329 SmallVector<CCValAssign, 16> ArgLocs;
1330 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1331 getTargetMachine(), ArgLocs, *DAG.getContext(), Call);
1332 CCInfo.AnalyzeCallOperands(Outs,
1333 CCAssignFnForNode(CallConv, /* Return*/ false,
1336 // Get a count of how many bytes are to be pushed on the stack.
1337 unsigned NumBytes = CCInfo.getNextStackOffset();
1339 // For tail calls, memory operands are available in our caller's stack.
1343 // Adjust the stack pointer for the new arguments...
1344 // These operations are automatically eliminated by the prolog/epilog pass
1346 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1348 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1350 RegsToPassVector RegsToPass;
1351 SmallVector<SDValue, 8> MemOpChains;
1353 // Walk the register/memloc assignments, inserting copies/loads. In the case
1354 // of tail call optimization, arguments are handled later.
1355 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1357 ++i, ++realArgIdx) {
1358 CCValAssign &VA = ArgLocs[i];
1359 SDValue Arg = OutVals[realArgIdx];
1360 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1361 bool isByVal = Flags.isByVal();
1363 // Promote the value if needed.
1364 switch (VA.getLocInfo()) {
1365 default: llvm_unreachable("Unknown loc info!");
1366 case CCValAssign::Full: break;
1367 case CCValAssign::SExt:
1368 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1370 case CCValAssign::ZExt:
1371 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1373 case CCValAssign::AExt:
1374 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1376 case CCValAssign::BCvt:
1377 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1381 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1382 if (VA.needsCustom()) {
1383 if (VA.getLocVT() == MVT::v2f64) {
1384 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1385 DAG.getConstant(0, MVT::i32));
1386 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1387 DAG.getConstant(1, MVT::i32));
1389 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1390 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1392 VA = ArgLocs[++i]; // skip ahead to next loc
1393 if (VA.isRegLoc()) {
1394 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1395 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1397 assert(VA.isMemLoc());
1399 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1400 dl, DAG, VA, Flags));
1403 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1404 StackPtr, MemOpChains, Flags);
1406 } else if (VA.isRegLoc()) {
1407 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1408 } else if (isByVal) {
1409 assert(VA.isMemLoc());
1410 unsigned offset = 0;
1412 // True if this byval aggregate will be split between registers
1414 if (CCInfo.isFirstByValRegValid()) {
1415 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1417 for (i = 0, j = CCInfo.getFirstByValReg(); j < ARM::R4; i++, j++) {
1418 SDValue Const = DAG.getConstant(4*i, MVT::i32);
1419 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
1420 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
1421 MachinePointerInfo(),
1422 false, false, false, 0);
1423 MemOpChains.push_back(Load.getValue(1));
1424 RegsToPass.push_back(std::make_pair(j, Load));
1426 offset = ARM::R4 - CCInfo.getFirstByValReg();
1427 CCInfo.clearFirstByValReg();
1430 if (Flags.getByValSize() - 4*offset > 0) {
1431 unsigned LocMemOffset = VA.getLocMemOffset();
1432 SDValue StkPtrOff = DAG.getIntPtrConstant(LocMemOffset);
1433 SDValue Dst = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr,
1435 SDValue SrcOffset = DAG.getIntPtrConstant(4*offset);
1436 SDValue Src = DAG.getNode(ISD::ADD, dl, getPointerTy(), Arg, SrcOffset);
1437 SDValue SizeNode = DAG.getConstant(Flags.getByValSize() - 4*offset,
1439 SDValue AlignNode = DAG.getConstant(Flags.getByValAlign(), MVT::i32);
1441 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
1442 SDValue Ops[] = { Chain, Dst, Src, SizeNode, AlignNode};
1443 MemOpChains.push_back(DAG.getNode(ARMISD::COPY_STRUCT_BYVAL, dl, VTs,
1444 Ops, array_lengthof(Ops)));
1446 } else if (!IsSibCall) {
1447 assert(VA.isMemLoc());
1449 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1450 dl, DAG, VA, Flags));
1454 if (!MemOpChains.empty())
1455 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1456 &MemOpChains[0], MemOpChains.size());
1458 // Build a sequence of copy-to-reg nodes chained together with token chain
1459 // and flag operands which copy the outgoing args into the appropriate regs.
1461 // Tail call byval lowering might overwrite argument registers so in case of
1462 // tail call optimization the copies to registers are lowered later.
1464 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1465 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1466 RegsToPass[i].second, InFlag);
1467 InFlag = Chain.getValue(1);
1470 // For tail calls lower the arguments to the 'real' stack slot.
1472 // Force all the incoming stack arguments to be loaded from the stack
1473 // before any new outgoing arguments are stored to the stack, because the
1474 // outgoing stack slots may alias the incoming argument stack slots, and
1475 // the alias isn't otherwise explicit. This is slightly more conservative
1476 // than necessary, because it means that each store effectively depends
1477 // on every argument instead of just those arguments it would clobber.
1479 // Do not flag preceding copytoreg stuff together with the following stuff.
1481 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1482 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1483 RegsToPass[i].second, InFlag);
1484 InFlag = Chain.getValue(1);
1489 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1490 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1491 // node so that legalize doesn't hack it.
1492 bool isDirect = false;
1493 bool isARMFunc = false;
1494 bool isLocalARMFunc = false;
1495 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1497 if (EnableARMLongCalls) {
1498 assert (getTargetMachine().getRelocationModel() == Reloc::Static
1499 && "long-calls with non-static relocation model!");
1500 // Handle a global address or an external symbol. If it's not one of
1501 // those, the target's already in a register, so we don't need to do
1503 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1504 const GlobalValue *GV = G->getGlobal();
1505 // Create a constant pool entry for the callee address
1506 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1507 ARMConstantPoolValue *CPV =
1508 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 0);
1510 // Get the address of the callee into a register
1511 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1512 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1513 Callee = DAG.getLoad(getPointerTy(), dl,
1514 DAG.getEntryNode(), CPAddr,
1515 MachinePointerInfo::getConstantPool(),
1516 false, false, false, 0);
1517 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1518 const char *Sym = S->getSymbol();
1520 // Create a constant pool entry for the callee address
1521 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1522 ARMConstantPoolValue *CPV =
1523 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1524 ARMPCLabelIndex, 0);
1525 // Get the address of the callee into a register
1526 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1527 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1528 Callee = DAG.getLoad(getPointerTy(), dl,
1529 DAG.getEntryNode(), CPAddr,
1530 MachinePointerInfo::getConstantPool(),
1531 false, false, false, 0);
1533 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1534 const GlobalValue *GV = G->getGlobal();
1536 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1537 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
1538 getTargetMachine().getRelocationModel() != Reloc::Static;
1539 isARMFunc = !Subtarget->isThumb() || isStub;
1540 // ARM call to a local ARM function is predicable.
1541 isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking);
1542 // tBX takes a register source operand.
1543 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1544 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1545 ARMConstantPoolValue *CPV =
1546 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 4);
1547 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1548 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1549 Callee = DAG.getLoad(getPointerTy(), dl,
1550 DAG.getEntryNode(), CPAddr,
1551 MachinePointerInfo::getConstantPool(),
1552 false, false, false, 0);
1553 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1554 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1555 getPointerTy(), Callee, PICLabel);
1557 // On ELF targets for PIC code, direct calls should go through the PLT
1558 unsigned OpFlags = 0;
1559 if (Subtarget->isTargetELF() &&
1560 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1561 OpFlags = ARMII::MO_PLT;
1562 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
1564 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1566 bool isStub = Subtarget->isTargetDarwin() &&
1567 getTargetMachine().getRelocationModel() != Reloc::Static;
1568 isARMFunc = !Subtarget->isThumb() || isStub;
1569 // tBX takes a register source operand.
1570 const char *Sym = S->getSymbol();
1571 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1572 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1573 ARMConstantPoolValue *CPV =
1574 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1575 ARMPCLabelIndex, 4);
1576 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1577 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1578 Callee = DAG.getLoad(getPointerTy(), dl,
1579 DAG.getEntryNode(), CPAddr,
1580 MachinePointerInfo::getConstantPool(),
1581 false, false, false, 0);
1582 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1583 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1584 getPointerTy(), Callee, PICLabel);
1586 unsigned OpFlags = 0;
1587 // On ELF targets for PIC code, direct calls should go through the PLT
1588 if (Subtarget->isTargetELF() &&
1589 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1590 OpFlags = ARMII::MO_PLT;
1591 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlags);
1595 // FIXME: handle tail calls differently.
1597 if (Subtarget->isThumb()) {
1598 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1599 CallOpc = ARMISD::CALL_NOLINK;
1600 else if (doesNotRet && isDirect && !isARMFunc &&
1601 Subtarget->hasRAS() && !Subtarget->isThumb1Only())
1602 // "mov lr, pc; b _foo" to avoid confusing the RSP
1603 CallOpc = ARMISD::CALL_NOLINK;
1605 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1607 if (!isDirect && !Subtarget->hasV5TOps()) {
1608 CallOpc = ARMISD::CALL_NOLINK;
1609 } else if (doesNotRet && isDirect && Subtarget->hasRAS())
1610 // "mov lr, pc; b _foo" to avoid confusing the RSP
1611 CallOpc = ARMISD::CALL_NOLINK;
1613 CallOpc = isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL;
1616 std::vector<SDValue> Ops;
1617 Ops.push_back(Chain);
1618 Ops.push_back(Callee);
1620 // Add argument registers to the end of the list so that they are known live
1622 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1623 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1624 RegsToPass[i].second.getValueType()));
1626 // Add a register mask operand representing the call-preserved registers.
1627 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
1628 const uint32_t *Mask = TRI->getCallPreservedMask(CallConv);
1629 assert(Mask && "Missing call preserved mask for calling convention");
1630 Ops.push_back(DAG.getRegisterMask(Mask));
1632 if (InFlag.getNode())
1633 Ops.push_back(InFlag);
1635 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
1637 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, &Ops[0], Ops.size());
1639 // Returns a chain and a flag for retval copy to use.
1640 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
1641 InFlag = Chain.getValue(1);
1643 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1644 DAG.getIntPtrConstant(0, true), InFlag);
1646 InFlag = Chain.getValue(1);
1648 // Handle result values, copying them out of physregs into vregs that we
1650 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins,
1654 /// HandleByVal - Every parameter *after* a byval parameter is passed
1655 /// on the stack. Remember the next parameter register to allocate,
1656 /// and then confiscate the rest of the parameter registers to insure
1659 ARMTargetLowering::HandleByVal(CCState *State, unsigned &size) const {
1660 unsigned reg = State->AllocateReg(GPRArgRegs, 4);
1661 assert((State->getCallOrPrologue() == Prologue ||
1662 State->getCallOrPrologue() == Call) &&
1663 "unhandled ParmContext");
1664 if ((!State->isFirstByValRegValid()) &&
1665 (ARM::R0 <= reg) && (reg <= ARM::R3)) {
1666 State->setFirstByValReg(reg);
1667 // At a call site, a byval parameter that is split between
1668 // registers and memory needs its size truncated here. In a
1669 // function prologue, such byval parameters are reassembled in
1670 // memory, and are not truncated.
1671 if (State->getCallOrPrologue() == Call) {
1672 unsigned excess = 4 * (ARM::R4 - reg);
1673 assert(size >= excess && "expected larger existing stack allocation");
1677 // Confiscate any remaining parameter registers to preclude their
1678 // assignment to subsequent parameters.
1679 while (State->AllocateReg(GPRArgRegs, 4))
1683 /// MatchingStackOffset - Return true if the given stack call argument is
1684 /// already available in the same position (relatively) of the caller's
1685 /// incoming argument stack.
1687 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1688 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1689 const TargetInstrInfo *TII) {
1690 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1692 if (Arg.getOpcode() == ISD::CopyFromReg) {
1693 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1694 if (!TargetRegisterInfo::isVirtualRegister(VR))
1696 MachineInstr *Def = MRI->getVRegDef(VR);
1699 if (!Flags.isByVal()) {
1700 if (!TII->isLoadFromStackSlot(Def, FI))
1705 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
1706 if (Flags.isByVal())
1707 // ByVal argument is passed in as a pointer but it's now being
1708 // dereferenced. e.g.
1709 // define @foo(%struct.X* %A) {
1710 // tail call @bar(%struct.X* byval %A)
1713 SDValue Ptr = Ld->getBasePtr();
1714 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
1717 FI = FINode->getIndex();
1721 assert(FI != INT_MAX);
1722 if (!MFI->isFixedObjectIndex(FI))
1724 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
1727 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
1728 /// for tail call optimization. Targets which want to do tail call
1729 /// optimization should implement this function.
1731 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1732 CallingConv::ID CalleeCC,
1734 bool isCalleeStructRet,
1735 bool isCallerStructRet,
1736 const SmallVectorImpl<ISD::OutputArg> &Outs,
1737 const SmallVectorImpl<SDValue> &OutVals,
1738 const SmallVectorImpl<ISD::InputArg> &Ins,
1739 SelectionDAG& DAG) const {
1740 const Function *CallerF = DAG.getMachineFunction().getFunction();
1741 CallingConv::ID CallerCC = CallerF->getCallingConv();
1742 bool CCMatch = CallerCC == CalleeCC;
1744 // Look for obvious safe cases to perform tail call optimization that do not
1745 // require ABI changes. This is what gcc calls sibcall.
1747 // Do not sibcall optimize vararg calls unless the call site is not passing
1749 if (isVarArg && !Outs.empty())
1752 // Also avoid sibcall optimization if either caller or callee uses struct
1753 // return semantics.
1754 if (isCalleeStructRet || isCallerStructRet)
1757 // FIXME: Completely disable sibcall for Thumb1 since Thumb1RegisterInfo::
1758 // emitEpilogue is not ready for them. Thumb tail calls also use t2B, as
1759 // the Thumb1 16-bit unconditional branch doesn't have sufficient relocation
1760 // support in the assembler and linker to be used. This would need to be
1761 // fixed to fully support tail calls in Thumb1.
1763 // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take
1764 // LR. This means if we need to reload LR, it takes an extra instructions,
1765 // which outweighs the value of the tail call; but here we don't know yet
1766 // whether LR is going to be used. Probably the right approach is to
1767 // generate the tail call here and turn it back into CALL/RET in
1768 // emitEpilogue if LR is used.
1770 // Thumb1 PIC calls to external symbols use BX, so they can be tail calls,
1771 // but we need to make sure there are enough registers; the only valid
1772 // registers are the 4 used for parameters. We don't currently do this
1774 if (Subtarget->isThumb1Only())
1777 // If the calling conventions do not match, then we'd better make sure the
1778 // results are returned in the same way as what the caller expects.
1780 SmallVector<CCValAssign, 16> RVLocs1;
1781 ARMCCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(),
1782 getTargetMachine(), RVLocs1, *DAG.getContext(), Call);
1783 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
1785 SmallVector<CCValAssign, 16> RVLocs2;
1786 ARMCCState CCInfo2(CallerCC, false, DAG.getMachineFunction(),
1787 getTargetMachine(), RVLocs2, *DAG.getContext(), Call);
1788 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
1790 if (RVLocs1.size() != RVLocs2.size())
1792 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
1793 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
1795 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
1797 if (RVLocs1[i].isRegLoc()) {
1798 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
1801 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
1807 // If Caller's vararg or byval argument has been split between registers and
1808 // stack, do not perform tail call, since part of the argument is in caller's
1810 const ARMFunctionInfo *AFI_Caller = DAG.getMachineFunction().
1811 getInfo<ARMFunctionInfo>();
1812 if (AFI_Caller->getVarArgsRegSaveSize())
1815 // If the callee takes no arguments then go on to check the results of the
1817 if (!Outs.empty()) {
1818 // Check if stack adjustment is needed. For now, do not do this if any
1819 // argument is passed on the stack.
1820 SmallVector<CCValAssign, 16> ArgLocs;
1821 ARMCCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(),
1822 getTargetMachine(), ArgLocs, *DAG.getContext(), Call);
1823 CCInfo.AnalyzeCallOperands(Outs,
1824 CCAssignFnForNode(CalleeCC, false, isVarArg));
1825 if (CCInfo.getNextStackOffset()) {
1826 MachineFunction &MF = DAG.getMachineFunction();
1828 // Check if the arguments are already laid out in the right way as
1829 // the caller's fixed stack objects.
1830 MachineFrameInfo *MFI = MF.getFrameInfo();
1831 const MachineRegisterInfo *MRI = &MF.getRegInfo();
1832 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1833 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1835 ++i, ++realArgIdx) {
1836 CCValAssign &VA = ArgLocs[i];
1837 EVT RegVT = VA.getLocVT();
1838 SDValue Arg = OutVals[realArgIdx];
1839 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1840 if (VA.getLocInfo() == CCValAssign::Indirect)
1842 if (VA.needsCustom()) {
1843 // f64 and vector types are split into multiple registers or
1844 // register/stack-slot combinations. The types will not match
1845 // the registers; give up on memory f64 refs until we figure
1846 // out what to do about this.
1849 if (!ArgLocs[++i].isRegLoc())
1851 if (RegVT == MVT::v2f64) {
1852 if (!ArgLocs[++i].isRegLoc())
1854 if (!ArgLocs[++i].isRegLoc())
1857 } else if (!VA.isRegLoc()) {
1858 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
1870 ARMTargetLowering::LowerReturn(SDValue Chain,
1871 CallingConv::ID CallConv, bool isVarArg,
1872 const SmallVectorImpl<ISD::OutputArg> &Outs,
1873 const SmallVectorImpl<SDValue> &OutVals,
1874 DebugLoc dl, SelectionDAG &DAG) const {
1876 // CCValAssign - represent the assignment of the return value to a location.
1877 SmallVector<CCValAssign, 16> RVLocs;
1879 // CCState - Info about the registers and stack slots.
1880 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1881 getTargetMachine(), RVLocs, *DAG.getContext(), Call);
1883 // Analyze outgoing return values.
1884 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1887 // If this is the first return lowered for this function, add
1888 // the regs to the liveout set for the function.
1889 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1890 for (unsigned i = 0; i != RVLocs.size(); ++i)
1891 if (RVLocs[i].isRegLoc())
1892 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1897 // Copy the result values into the output registers.
1898 for (unsigned i = 0, realRVLocIdx = 0;
1900 ++i, ++realRVLocIdx) {
1901 CCValAssign &VA = RVLocs[i];
1902 assert(VA.isRegLoc() && "Can only return in registers!");
1904 SDValue Arg = OutVals[realRVLocIdx];
1906 switch (VA.getLocInfo()) {
1907 default: llvm_unreachable("Unknown loc info!");
1908 case CCValAssign::Full: break;
1909 case CCValAssign::BCvt:
1910 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1914 if (VA.needsCustom()) {
1915 if (VA.getLocVT() == MVT::v2f64) {
1916 // Extract the first half and return it in two registers.
1917 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1918 DAG.getConstant(0, MVT::i32));
1919 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
1920 DAG.getVTList(MVT::i32, MVT::i32), Half);
1922 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1923 Flag = Chain.getValue(1);
1924 VA = RVLocs[++i]; // skip ahead to next loc
1925 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1926 HalfGPRs.getValue(1), Flag);
1927 Flag = Chain.getValue(1);
1928 VA = RVLocs[++i]; // skip ahead to next loc
1930 // Extract the 2nd half and fall through to handle it as an f64 value.
1931 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1932 DAG.getConstant(1, MVT::i32));
1934 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
1936 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1937 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
1938 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
1939 Flag = Chain.getValue(1);
1940 VA = RVLocs[++i]; // skip ahead to next loc
1941 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
1944 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1946 // Guarantee that all emitted copies are
1947 // stuck together, avoiding something bad.
1948 Flag = Chain.getValue(1);
1953 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1955 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
1960 bool ARMTargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
1961 if (N->getNumValues() != 1)
1963 if (!N->hasNUsesOfValue(1, 0))
1966 SDValue TCChain = Chain;
1967 SDNode *Copy = *N->use_begin();
1968 if (Copy->getOpcode() == ISD::CopyToReg) {
1969 // If the copy has a glue operand, we conservatively assume it isn't safe to
1970 // perform a tail call.
1971 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
1973 TCChain = Copy->getOperand(0);
1974 } else if (Copy->getOpcode() == ARMISD::VMOVRRD) {
1975 SDNode *VMov = Copy;
1976 // f64 returned in a pair of GPRs.
1977 SmallPtrSet<SDNode*, 2> Copies;
1978 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
1980 if (UI->getOpcode() != ISD::CopyToReg)
1984 if (Copies.size() > 2)
1987 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
1989 SDValue UseChain = UI->getOperand(0);
1990 if (Copies.count(UseChain.getNode()))
1997 } else if (Copy->getOpcode() == ISD::BITCAST) {
1998 // f32 returned in a single GPR.
1999 if (!Copy->hasOneUse())
2001 Copy = *Copy->use_begin();
2002 if (Copy->getOpcode() != ISD::CopyToReg || !Copy->hasNUsesOfValue(1, 0))
2004 Chain = Copy->getOperand(0);
2009 bool HasRet = false;
2010 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
2012 if (UI->getOpcode() != ARMISD::RET_FLAG)
2024 bool ARMTargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
2025 if (!EnableARMTailCalls && !Subtarget->supportsTailCall())
2028 if (!CI->isTailCall())
2031 return !Subtarget->isThumb1Only();
2034 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
2035 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
2036 // one of the above mentioned nodes. It has to be wrapped because otherwise
2037 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
2038 // be used to form addressing mode. These wrapped nodes will be selected
2040 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
2041 EVT PtrVT = Op.getValueType();
2042 // FIXME there is no actual debug info here
2043 DebugLoc dl = Op.getDebugLoc();
2044 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
2046 if (CP->isMachineConstantPoolEntry())
2047 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
2048 CP->getAlignment());
2050 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
2051 CP->getAlignment());
2052 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
2055 unsigned ARMTargetLowering::getJumpTableEncoding() const {
2056 return MachineJumpTableInfo::EK_Inline;
2059 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
2060 SelectionDAG &DAG) const {
2061 MachineFunction &MF = DAG.getMachineFunction();
2062 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2063 unsigned ARMPCLabelIndex = 0;
2064 DebugLoc DL = Op.getDebugLoc();
2065 EVT PtrVT = getPointerTy();
2066 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
2067 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2069 if (RelocM == Reloc::Static) {
2070 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
2072 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2073 ARMPCLabelIndex = AFI->createPICLabelUId();
2074 ARMConstantPoolValue *CPV =
2075 ARMConstantPoolConstant::Create(BA, ARMPCLabelIndex,
2076 ARMCP::CPBlockAddress, PCAdj);
2077 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2079 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
2080 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
2081 MachinePointerInfo::getConstantPool(),
2082 false, false, false, 0);
2083 if (RelocM == Reloc::Static)
2085 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2086 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
2089 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
2091 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
2092 SelectionDAG &DAG) const {
2093 DebugLoc dl = GA->getDebugLoc();
2094 EVT PtrVT = getPointerTy();
2095 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2096 MachineFunction &MF = DAG.getMachineFunction();
2097 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2098 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2099 ARMConstantPoolValue *CPV =
2100 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2101 ARMCP::CPValue, PCAdj, ARMCP::TLSGD, true);
2102 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2103 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
2104 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
2105 MachinePointerInfo::getConstantPool(),
2106 false, false, false, 0);
2107 SDValue Chain = Argument.getValue(1);
2109 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2110 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
2112 // call __tls_get_addr.
2115 Entry.Node = Argument;
2116 Entry.Ty = (Type *) Type::getInt32Ty(*DAG.getContext());
2117 Args.push_back(Entry);
2118 // FIXME: is there useful debug info available here?
2119 TargetLowering::CallLoweringInfo CLI(Chain,
2120 (Type *) Type::getInt32Ty(*DAG.getContext()),
2121 false, false, false, false,
2122 0, CallingConv::C, /*isTailCall=*/false,
2123 /*doesNotRet=*/false, /*isReturnValueUsed=*/true,
2124 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
2125 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
2126 return CallResult.first;
2129 // Lower ISD::GlobalTLSAddress using the "initial exec" or
2130 // "local exec" model.
2132 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
2134 TLSModel::Model model) const {
2135 const GlobalValue *GV = GA->getGlobal();
2136 DebugLoc dl = GA->getDebugLoc();
2138 SDValue Chain = DAG.getEntryNode();
2139 EVT PtrVT = getPointerTy();
2140 // Get the Thread Pointer
2141 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2143 if (model == TLSModel::InitialExec) {
2144 MachineFunction &MF = DAG.getMachineFunction();
2145 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2146 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2147 // Initial exec model.
2148 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2149 ARMConstantPoolValue *CPV =
2150 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2151 ARMCP::CPValue, PCAdj, ARMCP::GOTTPOFF,
2153 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2154 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2155 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2156 MachinePointerInfo::getConstantPool(),
2157 false, false, false, 0);
2158 Chain = Offset.getValue(1);
2160 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2161 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
2163 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2164 MachinePointerInfo::getConstantPool(),
2165 false, false, false, 0);
2168 assert(model == TLSModel::LocalExec);
2169 ARMConstantPoolValue *CPV =
2170 ARMConstantPoolConstant::Create(GV, ARMCP::TPOFF);
2171 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2172 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2173 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2174 MachinePointerInfo::getConstantPool(),
2175 false, false, false, 0);
2178 // The address of the thread local variable is the add of the thread
2179 // pointer with the offset of the variable.
2180 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
2184 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
2185 // TODO: implement the "local dynamic" model
2186 assert(Subtarget->isTargetELF() &&
2187 "TLS not implemented for non-ELF targets");
2188 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
2190 TLSModel::Model model = getTargetMachine().getTLSModel(GA->getGlobal());
2193 case TLSModel::GeneralDynamic:
2194 case TLSModel::LocalDynamic:
2195 return LowerToTLSGeneralDynamicModel(GA, DAG);
2196 case TLSModel::InitialExec:
2197 case TLSModel::LocalExec:
2198 return LowerToTLSExecModels(GA, DAG, model);
2200 llvm_unreachable("bogus TLS model");
2203 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
2204 SelectionDAG &DAG) const {
2205 EVT PtrVT = getPointerTy();
2206 DebugLoc dl = Op.getDebugLoc();
2207 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2208 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2209 if (RelocM == Reloc::PIC_) {
2210 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
2211 ARMConstantPoolValue *CPV =
2212 ARMConstantPoolConstant::Create(GV,
2213 UseGOTOFF ? ARMCP::GOTOFF : ARMCP::GOT);
2214 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2215 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2216 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
2218 MachinePointerInfo::getConstantPool(),
2219 false, false, false, 0);
2220 SDValue Chain = Result.getValue(1);
2221 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
2222 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
2224 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
2225 MachinePointerInfo::getGOT(),
2226 false, false, false, 0);
2230 // If we have T2 ops, we can materialize the address directly via movt/movw
2231 // pair. This is always cheaper.
2232 if (Subtarget->useMovt()) {
2234 // FIXME: Once remat is capable of dealing with instructions with register
2235 // operands, expand this into two nodes.
2236 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2237 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2239 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2240 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2241 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2242 MachinePointerInfo::getConstantPool(),
2243 false, false, false, 0);
2247 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
2248 SelectionDAG &DAG) const {
2249 EVT PtrVT = getPointerTy();
2250 DebugLoc dl = Op.getDebugLoc();
2251 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2252 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2253 MachineFunction &MF = DAG.getMachineFunction();
2254 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2256 // FIXME: Enable this for static codegen when tool issues are fixed. Also
2257 // update ARMFastISel::ARMMaterializeGV.
2258 if (Subtarget->useMovt() && RelocM != Reloc::Static) {
2260 // FIXME: Once remat is capable of dealing with instructions with register
2261 // operands, expand this into two nodes.
2262 if (RelocM == Reloc::Static)
2263 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2264 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2266 unsigned Wrapper = (RelocM == Reloc::PIC_)
2267 ? ARMISD::WrapperPIC : ARMISD::WrapperDYN;
2268 SDValue Result = DAG.getNode(Wrapper, dl, PtrVT,
2269 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2270 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2271 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
2272 MachinePointerInfo::getGOT(),
2273 false, false, false, 0);
2277 unsigned ARMPCLabelIndex = 0;
2279 if (RelocM == Reloc::Static) {
2280 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2282 ARMPCLabelIndex = AFI->createPICLabelUId();
2283 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
2284 ARMConstantPoolValue *CPV =
2285 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue,
2287 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2289 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2291 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2292 MachinePointerInfo::getConstantPool(),
2293 false, false, false, 0);
2294 SDValue Chain = Result.getValue(1);
2296 if (RelocM == Reloc::PIC_) {
2297 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2298 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2301 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2302 Result = DAG.getLoad(PtrVT, dl, Chain, Result, MachinePointerInfo::getGOT(),
2303 false, false, false, 0);
2308 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
2309 SelectionDAG &DAG) const {
2310 assert(Subtarget->isTargetELF() &&
2311 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
2312 MachineFunction &MF = DAG.getMachineFunction();
2313 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2314 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2315 EVT PtrVT = getPointerTy();
2316 DebugLoc dl = Op.getDebugLoc();
2317 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2318 ARMConstantPoolValue *CPV =
2319 ARMConstantPoolSymbol::Create(*DAG.getContext(), "_GLOBAL_OFFSET_TABLE_",
2320 ARMPCLabelIndex, PCAdj);
2321 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2322 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2323 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2324 MachinePointerInfo::getConstantPool(),
2325 false, false, false, 0);
2326 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2327 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2331 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
2332 DebugLoc dl = Op.getDebugLoc();
2333 SDValue Val = DAG.getConstant(0, MVT::i32);
2334 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl,
2335 DAG.getVTList(MVT::i32, MVT::Other), Op.getOperand(0),
2336 Op.getOperand(1), Val);
2340 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
2341 DebugLoc dl = Op.getDebugLoc();
2342 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
2343 Op.getOperand(1), DAG.getConstant(0, MVT::i32));
2347 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
2348 const ARMSubtarget *Subtarget) const {
2349 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2350 DebugLoc dl = Op.getDebugLoc();
2352 default: return SDValue(); // Don't custom lower most intrinsics.
2353 case Intrinsic::arm_thread_pointer: {
2354 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2355 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2357 case Intrinsic::eh_sjlj_lsda: {
2358 MachineFunction &MF = DAG.getMachineFunction();
2359 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2360 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2361 EVT PtrVT = getPointerTy();
2362 DebugLoc dl = Op.getDebugLoc();
2363 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2365 unsigned PCAdj = (RelocM != Reloc::PIC_)
2366 ? 0 : (Subtarget->isThumb() ? 4 : 8);
2367 ARMConstantPoolValue *CPV =
2368 ARMConstantPoolConstant::Create(MF.getFunction(), ARMPCLabelIndex,
2369 ARMCP::CPLSDA, PCAdj);
2370 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2371 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2373 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2374 MachinePointerInfo::getConstantPool(),
2375 false, false, false, 0);
2377 if (RelocM == Reloc::PIC_) {
2378 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2379 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2383 case Intrinsic::arm_neon_vmulls:
2384 case Intrinsic::arm_neon_vmullu: {
2385 unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vmulls)
2386 ? ARMISD::VMULLs : ARMISD::VMULLu;
2387 return DAG.getNode(NewOpc, Op.getDebugLoc(), Op.getValueType(),
2388 Op.getOperand(1), Op.getOperand(2));
2393 static SDValue LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG,
2394 const ARMSubtarget *Subtarget) {
2395 DebugLoc dl = Op.getDebugLoc();
2396 if (!Subtarget->hasDataBarrier()) {
2397 // Some ARMv6 cpus can support data barriers with an mcr instruction.
2398 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2400 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2401 "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
2402 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2403 DAG.getConstant(0, MVT::i32));
2406 SDValue Op5 = Op.getOperand(5);
2407 bool isDeviceBarrier = cast<ConstantSDNode>(Op5)->getZExtValue() != 0;
2408 unsigned isLL = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
2409 unsigned isLS = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
2410 bool isOnlyStoreBarrier = (isLL == 0 && isLS == 0);
2412 ARM_MB::MemBOpt DMBOpt;
2413 if (isDeviceBarrier)
2414 DMBOpt = isOnlyStoreBarrier ? ARM_MB::ST : ARM_MB::SY;
2416 DMBOpt = isOnlyStoreBarrier ? ARM_MB::ISHST : ARM_MB::ISH;
2417 return DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0),
2418 DAG.getConstant(DMBOpt, MVT::i32));
2422 static SDValue LowerATOMIC_FENCE(SDValue Op, SelectionDAG &DAG,
2423 const ARMSubtarget *Subtarget) {
2424 // FIXME: handle "fence singlethread" more efficiently.
2425 DebugLoc dl = Op.getDebugLoc();
2426 if (!Subtarget->hasDataBarrier()) {
2427 // Some ARMv6 cpus can support data barriers with an mcr instruction.
2428 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2430 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2431 "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
2432 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2433 DAG.getConstant(0, MVT::i32));
2436 return DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0),
2437 DAG.getConstant(ARM_MB::ISH, MVT::i32));
2440 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG,
2441 const ARMSubtarget *Subtarget) {
2442 // ARM pre v5TE and Thumb1 does not have preload instructions.
2443 if (!(Subtarget->isThumb2() ||
2444 (!Subtarget->isThumb1Only() && Subtarget->hasV5TEOps())))
2445 // Just preserve the chain.
2446 return Op.getOperand(0);
2448 DebugLoc dl = Op.getDebugLoc();
2449 unsigned isRead = ~cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() & 1;
2451 (!Subtarget->hasV7Ops() || !Subtarget->hasMPExtension()))
2452 // ARMv7 with MP extension has PLDW.
2453 return Op.getOperand(0);
2455 unsigned isData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
2456 if (Subtarget->isThumb()) {
2458 isRead = ~isRead & 1;
2459 isData = ~isData & 1;
2462 return DAG.getNode(ARMISD::PRELOAD, dl, MVT::Other, Op.getOperand(0),
2463 Op.getOperand(1), DAG.getConstant(isRead, MVT::i32),
2464 DAG.getConstant(isData, MVT::i32));
2467 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
2468 MachineFunction &MF = DAG.getMachineFunction();
2469 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
2471 // vastart just stores the address of the VarArgsFrameIndex slot into the
2472 // memory location argument.
2473 DebugLoc dl = Op.getDebugLoc();
2474 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2475 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2476 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
2477 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
2478 MachinePointerInfo(SV), false, false, 0);
2482 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
2483 SDValue &Root, SelectionDAG &DAG,
2484 DebugLoc dl) const {
2485 MachineFunction &MF = DAG.getMachineFunction();
2486 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2488 const TargetRegisterClass *RC;
2489 if (AFI->isThumb1OnlyFunction())
2490 RC = &ARM::tGPRRegClass;
2492 RC = &ARM::GPRRegClass;
2494 // Transform the arguments stored in physical registers into virtual ones.
2495 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2496 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2499 if (NextVA.isMemLoc()) {
2500 MachineFrameInfo *MFI = MF.getFrameInfo();
2501 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true);
2503 // Create load node to retrieve arguments from the stack.
2504 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2505 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
2506 MachinePointerInfo::getFixedStack(FI),
2507 false, false, false, 0);
2509 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
2510 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2513 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
2517 ARMTargetLowering::computeRegArea(CCState &CCInfo, MachineFunction &MF,
2518 unsigned &VARegSize, unsigned &VARegSaveSize)
2521 if (CCInfo.isFirstByValRegValid())
2522 NumGPRs = ARM::R4 - CCInfo.getFirstByValReg();
2524 unsigned int firstUnalloced;
2525 firstUnalloced = CCInfo.getFirstUnallocated(GPRArgRegs,
2526 sizeof(GPRArgRegs) /
2527 sizeof(GPRArgRegs[0]));
2528 NumGPRs = (firstUnalloced <= 3) ? (4 - firstUnalloced) : 0;
2531 unsigned Align = MF.getTarget().getFrameLowering()->getStackAlignment();
2532 VARegSize = NumGPRs * 4;
2533 VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
2536 // The remaining GPRs hold either the beginning of variable-argument
2537 // data, or the beginning of an aggregate passed by value (usuall
2538 // byval). Either way, we allocate stack slots adjacent to the data
2539 // provided by our caller, and store the unallocated registers there.
2540 // If this is a variadic function, the va_list pointer will begin with
2541 // these values; otherwise, this reassembles a (byval) structure that
2542 // was split between registers and memory.
2544 ARMTargetLowering::VarArgStyleRegisters(CCState &CCInfo, SelectionDAG &DAG,
2545 DebugLoc dl, SDValue &Chain,
2546 const Value *OrigArg,
2547 unsigned OffsetFromOrigArg,
2548 unsigned ArgOffset) const {
2549 MachineFunction &MF = DAG.getMachineFunction();
2550 MachineFrameInfo *MFI = MF.getFrameInfo();
2551 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2552 unsigned firstRegToSaveIndex;
2553 if (CCInfo.isFirstByValRegValid())
2554 firstRegToSaveIndex = CCInfo.getFirstByValReg() - ARM::R0;
2556 firstRegToSaveIndex = CCInfo.getFirstUnallocated
2557 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
2560 unsigned VARegSize, VARegSaveSize;
2561 computeRegArea(CCInfo, MF, VARegSize, VARegSaveSize);
2562 if (VARegSaveSize) {
2563 // If this function is vararg, store any remaining integer argument regs
2564 // to their spots on the stack so that they may be loaded by deferencing
2565 // the result of va_next.
2566 AFI->setVarArgsRegSaveSize(VARegSaveSize);
2567 AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(VARegSaveSize,
2568 ArgOffset + VARegSaveSize
2571 SDValue FIN = DAG.getFrameIndex(AFI->getVarArgsFrameIndex(),
2574 SmallVector<SDValue, 4> MemOps;
2575 for (unsigned i = 0; firstRegToSaveIndex < 4; ++firstRegToSaveIndex, ++i) {
2576 const TargetRegisterClass *RC;
2577 if (AFI->isThumb1OnlyFunction())
2578 RC = &ARM::tGPRRegClass;
2580 RC = &ARM::GPRRegClass;
2582 unsigned VReg = MF.addLiveIn(GPRArgRegs[firstRegToSaveIndex], RC);
2583 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2585 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2586 MachinePointerInfo(OrigArg, OffsetFromOrigArg + 4*i),
2588 MemOps.push_back(Store);
2589 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
2590 DAG.getConstant(4, getPointerTy()));
2592 if (!MemOps.empty())
2593 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2594 &MemOps[0], MemOps.size());
2596 // This will point to the next argument passed via stack.
2597 AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(4, ArgOffset, true));
2601 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
2602 CallingConv::ID CallConv, bool isVarArg,
2603 const SmallVectorImpl<ISD::InputArg>
2605 DebugLoc dl, SelectionDAG &DAG,
2606 SmallVectorImpl<SDValue> &InVals)
2608 MachineFunction &MF = DAG.getMachineFunction();
2609 MachineFrameInfo *MFI = MF.getFrameInfo();
2611 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2613 // Assign locations to all of the incoming arguments.
2614 SmallVector<CCValAssign, 16> ArgLocs;
2615 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2616 getTargetMachine(), ArgLocs, *DAG.getContext(), Prologue);
2617 CCInfo.AnalyzeFormalArguments(Ins,
2618 CCAssignFnForNode(CallConv, /* Return*/ false,
2621 SmallVector<SDValue, 16> ArgValues;
2622 int lastInsIndex = -1;
2624 Function::const_arg_iterator CurOrigArg = MF.getFunction()->arg_begin();
2625 unsigned CurArgIdx = 0;
2626 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2627 CCValAssign &VA = ArgLocs[i];
2628 std::advance(CurOrigArg, Ins[VA.getValNo()].OrigArgIndex - CurArgIdx);
2629 CurArgIdx = Ins[VA.getValNo()].OrigArgIndex;
2630 // Arguments stored in registers.
2631 if (VA.isRegLoc()) {
2632 EVT RegVT = VA.getLocVT();
2634 if (VA.needsCustom()) {
2635 // f64 and vector types are split up into multiple registers or
2636 // combinations of registers and stack slots.
2637 if (VA.getLocVT() == MVT::v2f64) {
2638 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
2640 VA = ArgLocs[++i]; // skip ahead to next loc
2642 if (VA.isMemLoc()) {
2643 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true);
2644 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2645 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
2646 MachinePointerInfo::getFixedStack(FI),
2647 false, false, false, 0);
2649 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
2652 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
2653 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2654 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
2655 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2656 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
2658 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
2661 const TargetRegisterClass *RC;
2663 if (RegVT == MVT::f32)
2664 RC = &ARM::SPRRegClass;
2665 else if (RegVT == MVT::f64)
2666 RC = &ARM::DPRRegClass;
2667 else if (RegVT == MVT::v2f64)
2668 RC = &ARM::QPRRegClass;
2669 else if (RegVT == MVT::i32)
2670 RC = AFI->isThumb1OnlyFunction() ?
2671 (const TargetRegisterClass*)&ARM::tGPRRegClass :
2672 (const TargetRegisterClass*)&ARM::GPRRegClass;
2674 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
2676 // Transform the arguments in physical registers into virtual ones.
2677 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2678 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2681 // If this is an 8 or 16-bit value, it is really passed promoted
2682 // to 32 bits. Insert an assert[sz]ext to capture this, then
2683 // truncate to the right size.
2684 switch (VA.getLocInfo()) {
2685 default: llvm_unreachable("Unknown loc info!");
2686 case CCValAssign::Full: break;
2687 case CCValAssign::BCvt:
2688 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
2690 case CCValAssign::SExt:
2691 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2692 DAG.getValueType(VA.getValVT()));
2693 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2695 case CCValAssign::ZExt:
2696 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2697 DAG.getValueType(VA.getValVT()));
2698 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2702 InVals.push_back(ArgValue);
2704 } else { // VA.isRegLoc()
2707 assert(VA.isMemLoc());
2708 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
2710 int index = ArgLocs[i].getValNo();
2712 // Some Ins[] entries become multiple ArgLoc[] entries.
2713 // Process them only once.
2714 if (index != lastInsIndex)
2716 ISD::ArgFlagsTy Flags = Ins[index].Flags;
2717 // FIXME: For now, all byval parameter objects are marked mutable.
2718 // This can be changed with more analysis.
2719 // In case of tail call optimization mark all arguments mutable.
2720 // Since they could be overwritten by lowering of arguments in case of
2722 if (Flags.isByVal()) {
2723 unsigned VARegSize, VARegSaveSize;
2724 computeRegArea(CCInfo, MF, VARegSize, VARegSaveSize);
2725 VarArgStyleRegisters(CCInfo, DAG,
2726 dl, Chain, CurOrigArg, Ins[VA.getValNo()].PartOffset, 0);
2727 unsigned Bytes = Flags.getByValSize() - VARegSize;
2728 if (Bytes == 0) Bytes = 1; // Don't create zero-sized stack objects.
2729 int FI = MFI->CreateFixedObject(Bytes,
2730 VA.getLocMemOffset(), false);
2731 InVals.push_back(DAG.getFrameIndex(FI, getPointerTy()));
2733 int FI = MFI->CreateFixedObject(VA.getLocVT().getSizeInBits()/8,
2734 VA.getLocMemOffset(), true);
2736 // Create load nodes to retrieve arguments from the stack.
2737 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2738 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
2739 MachinePointerInfo::getFixedStack(FI),
2740 false, false, false, 0));
2742 lastInsIndex = index;
2749 VarArgStyleRegisters(CCInfo, DAG, dl, Chain, 0, 0,
2750 CCInfo.getNextStackOffset());
2755 /// isFloatingPointZero - Return true if this is +0.0.
2756 static bool isFloatingPointZero(SDValue Op) {
2757 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
2758 return CFP->getValueAPF().isPosZero();
2759 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
2760 // Maybe this has already been legalized into the constant pool?
2761 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
2762 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
2763 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
2764 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
2765 return CFP->getValueAPF().isPosZero();
2771 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
2772 /// the given operands.
2774 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
2775 SDValue &ARMcc, SelectionDAG &DAG,
2776 DebugLoc dl) const {
2777 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
2778 unsigned C = RHSC->getZExtValue();
2779 if (!isLegalICmpImmediate(C)) {
2780 // Constant does not fit, try adjusting it by one?
2785 if (C != 0x80000000 && isLegalICmpImmediate(C-1)) {
2786 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
2787 RHS = DAG.getConstant(C-1, MVT::i32);
2792 if (C != 0 && isLegalICmpImmediate(C-1)) {
2793 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
2794 RHS = DAG.getConstant(C-1, MVT::i32);
2799 if (C != 0x7fffffff && isLegalICmpImmediate(C+1)) {
2800 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
2801 RHS = DAG.getConstant(C+1, MVT::i32);
2806 if (C != 0xffffffff && isLegalICmpImmediate(C+1)) {
2807 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
2808 RHS = DAG.getConstant(C+1, MVT::i32);
2815 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2816 ARMISD::NodeType CompareType;
2819 CompareType = ARMISD::CMP;
2824 CompareType = ARMISD::CMPZ;
2827 ARMcc = DAG.getConstant(CondCode, MVT::i32);
2828 return DAG.getNode(CompareType, dl, MVT::Glue, LHS, RHS);
2831 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
2833 ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
2834 DebugLoc dl) const {
2836 if (!isFloatingPointZero(RHS))
2837 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Glue, LHS, RHS);
2839 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Glue, LHS);
2840 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Glue, Cmp);
2843 /// duplicateCmp - Glue values can have only one use, so this function
2844 /// duplicates a comparison node.
2846 ARMTargetLowering::duplicateCmp(SDValue Cmp, SelectionDAG &DAG) const {
2847 unsigned Opc = Cmp.getOpcode();
2848 DebugLoc DL = Cmp.getDebugLoc();
2849 if (Opc == ARMISD::CMP || Opc == ARMISD::CMPZ)
2850 return DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
2852 assert(Opc == ARMISD::FMSTAT && "unexpected comparison operation");
2853 Cmp = Cmp.getOperand(0);
2854 Opc = Cmp.getOpcode();
2855 if (Opc == ARMISD::CMPFP)
2856 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
2858 assert(Opc == ARMISD::CMPFPw0 && "unexpected operand of FMSTAT");
2859 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0));
2861 return DAG.getNode(ARMISD::FMSTAT, DL, MVT::Glue, Cmp);
2864 SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
2865 SDValue Cond = Op.getOperand(0);
2866 SDValue SelectTrue = Op.getOperand(1);
2867 SDValue SelectFalse = Op.getOperand(2);
2868 DebugLoc dl = Op.getDebugLoc();
2872 // (select (cmov 1, 0, cond), t, f) -> (cmov t, f, cond)
2873 // (select (cmov 0, 1, cond), t, f) -> (cmov f, t, cond)
2875 if (Cond.getOpcode() == ARMISD::CMOV && Cond.hasOneUse()) {
2876 const ConstantSDNode *CMOVTrue =
2877 dyn_cast<ConstantSDNode>(Cond.getOperand(0));
2878 const ConstantSDNode *CMOVFalse =
2879 dyn_cast<ConstantSDNode>(Cond.getOperand(1));
2881 if (CMOVTrue && CMOVFalse) {
2882 unsigned CMOVTrueVal = CMOVTrue->getZExtValue();
2883 unsigned CMOVFalseVal = CMOVFalse->getZExtValue();
2887 if (CMOVTrueVal == 1 && CMOVFalseVal == 0) {
2889 False = SelectFalse;
2890 } else if (CMOVTrueVal == 0 && CMOVFalseVal == 1) {
2895 if (True.getNode() && False.getNode()) {
2896 EVT VT = Op.getValueType();
2897 SDValue ARMcc = Cond.getOperand(2);
2898 SDValue CCR = Cond.getOperand(3);
2899 SDValue Cmp = duplicateCmp(Cond.getOperand(4), DAG);
2900 assert(True.getValueType() == VT);
2901 return DAG.getNode(ARMISD::CMOV, dl, VT, True, False, ARMcc, CCR, Cmp);
2906 // ARM's BooleanContents value is UndefinedBooleanContent. Mask out the
2907 // undefined bits before doing a full-word comparison with zero.
2908 Cond = DAG.getNode(ISD::AND, dl, Cond.getValueType(), Cond,
2909 DAG.getConstant(1, Cond.getValueType()));
2911 return DAG.getSelectCC(dl, Cond,
2912 DAG.getConstant(0, Cond.getValueType()),
2913 SelectTrue, SelectFalse, ISD::SETNE);
2916 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
2917 EVT VT = Op.getValueType();
2918 SDValue LHS = Op.getOperand(0);
2919 SDValue RHS = Op.getOperand(1);
2920 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
2921 SDValue TrueVal = Op.getOperand(2);
2922 SDValue FalseVal = Op.getOperand(3);
2923 DebugLoc dl = Op.getDebugLoc();
2925 if (LHS.getValueType() == MVT::i32) {
2927 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2928 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2929 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,Cmp);
2932 ARMCC::CondCodes CondCode, CondCode2;
2933 FPCCToARMCC(CC, CondCode, CondCode2);
2935 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
2936 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2937 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2938 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
2940 if (CondCode2 != ARMCC::AL) {
2941 SDValue ARMcc2 = DAG.getConstant(CondCode2, MVT::i32);
2942 // FIXME: Needs another CMP because flag can have but one use.
2943 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
2944 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
2945 Result, TrueVal, ARMcc2, CCR, Cmp2);
2950 /// canChangeToInt - Given the fp compare operand, return true if it is suitable
2951 /// to morph to an integer compare sequence.
2952 static bool canChangeToInt(SDValue Op, bool &SeenZero,
2953 const ARMSubtarget *Subtarget) {
2954 SDNode *N = Op.getNode();
2955 if (!N->hasOneUse())
2956 // Otherwise it requires moving the value from fp to integer registers.
2958 if (!N->getNumValues())
2960 EVT VT = Op.getValueType();
2961 if (VT != MVT::f32 && !Subtarget->isFPBrccSlow())
2962 // f32 case is generally profitable. f64 case only makes sense when vcmpe +
2963 // vmrs are very slow, e.g. cortex-a8.
2966 if (isFloatingPointZero(Op)) {
2970 return ISD::isNormalLoad(N);
2973 static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) {
2974 if (isFloatingPointZero(Op))
2975 return DAG.getConstant(0, MVT::i32);
2977 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op))
2978 return DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2979 Ld->getChain(), Ld->getBasePtr(), Ld->getPointerInfo(),
2980 Ld->isVolatile(), Ld->isNonTemporal(),
2981 Ld->isInvariant(), Ld->getAlignment());
2983 llvm_unreachable("Unknown VFP cmp argument!");
2986 static void expandf64Toi32(SDValue Op, SelectionDAG &DAG,
2987 SDValue &RetVal1, SDValue &RetVal2) {
2988 if (isFloatingPointZero(Op)) {
2989 RetVal1 = DAG.getConstant(0, MVT::i32);
2990 RetVal2 = DAG.getConstant(0, MVT::i32);
2994 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) {
2995 SDValue Ptr = Ld->getBasePtr();
2996 RetVal1 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2997 Ld->getChain(), Ptr,
2998 Ld->getPointerInfo(),
2999 Ld->isVolatile(), Ld->isNonTemporal(),
3000 Ld->isInvariant(), Ld->getAlignment());
3002 EVT PtrType = Ptr.getValueType();
3003 unsigned NewAlign = MinAlign(Ld->getAlignment(), 4);
3004 SDValue NewPtr = DAG.getNode(ISD::ADD, Op.getDebugLoc(),
3005 PtrType, Ptr, DAG.getConstant(4, PtrType));
3006 RetVal2 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
3007 Ld->getChain(), NewPtr,
3008 Ld->getPointerInfo().getWithOffset(4),
3009 Ld->isVolatile(), Ld->isNonTemporal(),
3010 Ld->isInvariant(), NewAlign);
3014 llvm_unreachable("Unknown VFP cmp argument!");
3017 /// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some
3018 /// f32 and even f64 comparisons to integer ones.
3020 ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const {
3021 SDValue Chain = Op.getOperand(0);
3022 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3023 SDValue LHS = Op.getOperand(2);
3024 SDValue RHS = Op.getOperand(3);
3025 SDValue Dest = Op.getOperand(4);
3026 DebugLoc dl = Op.getDebugLoc();
3028 bool LHSSeenZero = false;
3029 bool LHSOk = canChangeToInt(LHS, LHSSeenZero, Subtarget);
3030 bool RHSSeenZero = false;
3031 bool RHSOk = canChangeToInt(RHS, RHSSeenZero, Subtarget);
3032 if (LHSOk && RHSOk && (LHSSeenZero || RHSSeenZero)) {
3033 // If unsafe fp math optimization is enabled and there are no other uses of
3034 // the CMP operands, and the condition code is EQ or NE, we can optimize it
3035 // to an integer comparison.
3036 if (CC == ISD::SETOEQ)
3038 else if (CC == ISD::SETUNE)
3041 SDValue Mask = DAG.getConstant(0x7fffffff, MVT::i32);
3043 if (LHS.getValueType() == MVT::f32) {
3044 LHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3045 bitcastf32Toi32(LHS, DAG), Mask);
3046 RHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3047 bitcastf32Toi32(RHS, DAG), Mask);
3048 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3049 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3050 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3051 Chain, Dest, ARMcc, CCR, Cmp);
3056 expandf64Toi32(LHS, DAG, LHS1, LHS2);
3057 expandf64Toi32(RHS, DAG, RHS1, RHS2);
3058 LHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, LHS2, Mask);
3059 RHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, RHS2, Mask);
3060 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3061 ARMcc = DAG.getConstant(CondCode, MVT::i32);
3062 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3063 SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest };
3064 return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops, 7);
3070 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
3071 SDValue Chain = Op.getOperand(0);
3072 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3073 SDValue LHS = Op.getOperand(2);
3074 SDValue RHS = Op.getOperand(3);
3075 SDValue Dest = Op.getOperand(4);
3076 DebugLoc dl = Op.getDebugLoc();
3078 if (LHS.getValueType() == MVT::i32) {
3080 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3081 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3082 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3083 Chain, Dest, ARMcc, CCR, Cmp);
3086 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3088 if (getTargetMachine().Options.UnsafeFPMath &&
3089 (CC == ISD::SETEQ || CC == ISD::SETOEQ ||
3090 CC == ISD::SETNE || CC == ISD::SETUNE)) {
3091 SDValue Result = OptimizeVFPBrcond(Op, DAG);
3092 if (Result.getNode())
3096 ARMCC::CondCodes CondCode, CondCode2;
3097 FPCCToARMCC(CC, CondCode, CondCode2);
3099 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
3100 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
3101 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3102 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3103 SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp };
3104 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
3105 if (CondCode2 != ARMCC::AL) {
3106 ARMcc = DAG.getConstant(CondCode2, MVT::i32);
3107 SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) };
3108 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
3113 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
3114 SDValue Chain = Op.getOperand(0);
3115 SDValue Table = Op.getOperand(1);
3116 SDValue Index = Op.getOperand(2);
3117 DebugLoc dl = Op.getDebugLoc();
3119 EVT PTy = getPointerTy();
3120 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
3121 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
3122 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
3123 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
3124 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
3125 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
3126 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
3127 if (Subtarget->isThumb2()) {
3128 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
3129 // which does another jump to the destination. This also makes it easier
3130 // to translate it to TBB / TBH later.
3131 // FIXME: This might not work if the function is extremely large.
3132 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
3133 Addr, Op.getOperand(2), JTI, UId);
3135 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
3136 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
3137 MachinePointerInfo::getJumpTable(),
3138 false, false, false, 0);
3139 Chain = Addr.getValue(1);
3140 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
3141 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3143 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
3144 MachinePointerInfo::getJumpTable(),
3145 false, false, false, 0);
3146 Chain = Addr.getValue(1);
3147 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3151 static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3152 EVT VT = Op.getValueType();
3153 DebugLoc dl = Op.getDebugLoc();
3155 if (Op.getValueType().getVectorElementType() == MVT::i32) {
3156 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::f32)
3158 return DAG.UnrollVectorOp(Op.getNode());
3161 assert(Op.getOperand(0).getValueType() == MVT::v4f32 &&
3162 "Invalid type for custom lowering!");
3163 if (VT != MVT::v4i16)
3164 return DAG.UnrollVectorOp(Op.getNode());
3166 Op = DAG.getNode(Op.getOpcode(), dl, MVT::v4i32, Op.getOperand(0));
3167 return DAG.getNode(ISD::TRUNCATE, dl, VT, Op);
3170 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3171 EVT VT = Op.getValueType();
3173 return LowerVectorFP_TO_INT(Op, DAG);
3175 DebugLoc dl = Op.getDebugLoc();
3178 switch (Op.getOpcode()) {
3179 default: llvm_unreachable("Invalid opcode!");
3180 case ISD::FP_TO_SINT:
3181 Opc = ARMISD::FTOSI;
3183 case ISD::FP_TO_UINT:
3184 Opc = ARMISD::FTOUI;
3187 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
3188 return DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
3191 static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3192 EVT VT = Op.getValueType();
3193 DebugLoc dl = Op.getDebugLoc();
3195 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::i32) {
3196 if (VT.getVectorElementType() == MVT::f32)
3198 return DAG.UnrollVectorOp(Op.getNode());
3201 assert(Op.getOperand(0).getValueType() == MVT::v4i16 &&
3202 "Invalid type for custom lowering!");
3203 if (VT != MVT::v4f32)
3204 return DAG.UnrollVectorOp(Op.getNode());
3208 switch (Op.getOpcode()) {
3209 default: llvm_unreachable("Invalid opcode!");
3210 case ISD::SINT_TO_FP:
3211 CastOpc = ISD::SIGN_EXTEND;
3212 Opc = ISD::SINT_TO_FP;
3214 case ISD::UINT_TO_FP:
3215 CastOpc = ISD::ZERO_EXTEND;
3216 Opc = ISD::UINT_TO_FP;
3220 Op = DAG.getNode(CastOpc, dl, MVT::v4i32, Op.getOperand(0));
3221 return DAG.getNode(Opc, dl, VT, Op);
3224 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3225 EVT VT = Op.getValueType();
3227 return LowerVectorINT_TO_FP(Op, DAG);
3229 DebugLoc dl = Op.getDebugLoc();
3232 switch (Op.getOpcode()) {
3233 default: llvm_unreachable("Invalid opcode!");
3234 case ISD::SINT_TO_FP:
3235 Opc = ARMISD::SITOF;
3237 case ISD::UINT_TO_FP:
3238 Opc = ARMISD::UITOF;
3242 Op = DAG.getNode(ISD::BITCAST, dl, MVT::f32, Op.getOperand(0));
3243 return DAG.getNode(Opc, dl, VT, Op);
3246 SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
3247 // Implement fcopysign with a fabs and a conditional fneg.
3248 SDValue Tmp0 = Op.getOperand(0);
3249 SDValue Tmp1 = Op.getOperand(1);
3250 DebugLoc dl = Op.getDebugLoc();
3251 EVT VT = Op.getValueType();
3252 EVT SrcVT = Tmp1.getValueType();
3253 bool InGPR = Tmp0.getOpcode() == ISD::BITCAST ||
3254 Tmp0.getOpcode() == ARMISD::VMOVDRR;
3255 bool UseNEON = !InGPR && Subtarget->hasNEON();
3258 // Use VBSL to copy the sign bit.
3259 unsigned EncodedVal = ARM_AM::createNEONModImm(0x6, 0x80);
3260 SDValue Mask = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v2i32,
3261 DAG.getTargetConstant(EncodedVal, MVT::i32));
3262 EVT OpVT = (VT == MVT::f32) ? MVT::v2i32 : MVT::v1i64;
3264 Mask = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3265 DAG.getNode(ISD::BITCAST, dl, OpVT, Mask),
3266 DAG.getConstant(32, MVT::i32));
3267 else /*if (VT == MVT::f32)*/
3268 Tmp0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp0);
3269 if (SrcVT == MVT::f32) {
3270 Tmp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp1);
3272 Tmp1 = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3273 DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1),
3274 DAG.getConstant(32, MVT::i32));
3275 } else if (VT == MVT::f32)
3276 Tmp1 = DAG.getNode(ARMISD::VSHRu, dl, MVT::v1i64,
3277 DAG.getNode(ISD::BITCAST, dl, MVT::v1i64, Tmp1),
3278 DAG.getConstant(32, MVT::i32));
3279 Tmp0 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp0);
3280 Tmp1 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1);
3282 SDValue AllOnes = DAG.getTargetConstant(ARM_AM::createNEONModImm(0xe, 0xff),
3284 AllOnes = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v8i8, AllOnes);
3285 SDValue MaskNot = DAG.getNode(ISD::XOR, dl, OpVT, Mask,
3286 DAG.getNode(ISD::BITCAST, dl, OpVT, AllOnes));
3288 SDValue Res = DAG.getNode(ISD::OR, dl, OpVT,
3289 DAG.getNode(ISD::AND, dl, OpVT, Tmp1, Mask),
3290 DAG.getNode(ISD::AND, dl, OpVT, Tmp0, MaskNot));
3291 if (VT == MVT::f32) {
3292 Res = DAG.getNode(ISD::BITCAST, dl, MVT::v2f32, Res);
3293 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, Res,
3294 DAG.getConstant(0, MVT::i32));
3296 Res = DAG.getNode(ISD::BITCAST, dl, MVT::f64, Res);
3302 // Bitcast operand 1 to i32.
3303 if (SrcVT == MVT::f64)
3304 Tmp1 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
3305 &Tmp1, 1).getValue(1);
3306 Tmp1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp1);
3308 // Or in the signbit with integer operations.
3309 SDValue Mask1 = DAG.getConstant(0x80000000, MVT::i32);
3310 SDValue Mask2 = DAG.getConstant(0x7fffffff, MVT::i32);
3311 Tmp1 = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp1, Mask1);
3312 if (VT == MVT::f32) {
3313 Tmp0 = DAG.getNode(ISD::AND, dl, MVT::i32,
3314 DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp0), Mask2);
3315 return DAG.getNode(ISD::BITCAST, dl, MVT::f32,
3316 DAG.getNode(ISD::OR, dl, MVT::i32, Tmp0, Tmp1));
3319 // f64: Or the high part with signbit and then combine two parts.
3320 Tmp0 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
3322 SDValue Lo = Tmp0.getValue(0);
3323 SDValue Hi = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp0.getValue(1), Mask2);
3324 Hi = DAG.getNode(ISD::OR, dl, MVT::i32, Hi, Tmp1);
3325 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
3328 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
3329 MachineFunction &MF = DAG.getMachineFunction();
3330 MachineFrameInfo *MFI = MF.getFrameInfo();
3331 MFI->setReturnAddressIsTaken(true);
3333 EVT VT = Op.getValueType();
3334 DebugLoc dl = Op.getDebugLoc();
3335 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3337 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
3338 SDValue Offset = DAG.getConstant(4, MVT::i32);
3339 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
3340 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
3341 MachinePointerInfo(), false, false, false, 0);
3344 // Return LR, which contains the return address. Mark it an implicit live-in.
3345 unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32));
3346 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
3349 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
3350 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3351 MFI->setFrameAddressIsTaken(true);
3353 EVT VT = Op.getValueType();
3354 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
3355 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3356 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
3357 ? ARM::R7 : ARM::R11;
3358 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
3360 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
3361 MachinePointerInfo(),
3362 false, false, false, 0);
3366 /// ExpandBITCAST - If the target supports VFP, this function is called to
3367 /// expand a bit convert where either the source or destination type is i64 to
3368 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
3369 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
3370 /// vectors), since the legalizer won't know what to do with that.
3371 static SDValue ExpandBITCAST(SDNode *N, SelectionDAG &DAG) {
3372 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3373 DebugLoc dl = N->getDebugLoc();
3374 SDValue Op = N->getOperand(0);
3376 // This function is only supposed to be called for i64 types, either as the
3377 // source or destination of the bit convert.
3378 EVT SrcVT = Op.getValueType();
3379 EVT DstVT = N->getValueType(0);
3380 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
3381 "ExpandBITCAST called for non-i64 type");
3383 // Turn i64->f64 into VMOVDRR.
3384 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
3385 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
3386 DAG.getConstant(0, MVT::i32));
3387 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
3388 DAG.getConstant(1, MVT::i32));
3389 return DAG.getNode(ISD::BITCAST, dl, DstVT,
3390 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
3393 // Turn f64->i64 into VMOVRRD.
3394 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
3395 SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
3396 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
3397 // Merge the pieces into a single i64 value.
3398 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
3404 /// getZeroVector - Returns a vector of specified type with all zero elements.
3405 /// Zero vectors are used to represent vector negation and in those cases
3406 /// will be implemented with the NEON VNEG instruction. However, VNEG does
3407 /// not support i64 elements, so sometimes the zero vectors will need to be
3408 /// explicitly constructed. Regardless, use a canonical VMOV to create the
3410 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3411 assert(VT.isVector() && "Expected a vector type");
3412 // The canonical modified immediate encoding of a zero vector is....0!
3413 SDValue EncodedVal = DAG.getTargetConstant(0, MVT::i32);
3414 EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
3415 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal);
3416 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
3419 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
3420 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
3421 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
3422 SelectionDAG &DAG) const {
3423 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3424 EVT VT = Op.getValueType();
3425 unsigned VTBits = VT.getSizeInBits();
3426 DebugLoc dl = Op.getDebugLoc();
3427 SDValue ShOpLo = Op.getOperand(0);
3428 SDValue ShOpHi = Op.getOperand(1);
3429 SDValue ShAmt = Op.getOperand(2);
3431 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
3433 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
3435 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
3436 DAG.getConstant(VTBits, MVT::i32), ShAmt);
3437 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
3438 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
3439 DAG.getConstant(VTBits, MVT::i32));
3440 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
3441 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3442 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
3444 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3445 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
3447 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
3448 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc,
3451 SDValue Ops[2] = { Lo, Hi };
3452 return DAG.getMergeValues(Ops, 2, dl);
3455 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
3456 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
3457 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
3458 SelectionDAG &DAG) const {
3459 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3460 EVT VT = Op.getValueType();
3461 unsigned VTBits = VT.getSizeInBits();
3462 DebugLoc dl = Op.getDebugLoc();
3463 SDValue ShOpLo = Op.getOperand(0);
3464 SDValue ShOpHi = Op.getOperand(1);
3465 SDValue ShAmt = Op.getOperand(2);
3468 assert(Op.getOpcode() == ISD::SHL_PARTS);
3469 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
3470 DAG.getConstant(VTBits, MVT::i32), ShAmt);
3471 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
3472 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
3473 DAG.getConstant(VTBits, MVT::i32));
3474 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
3475 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
3477 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3478 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3479 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
3481 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
3482 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc,
3485 SDValue Ops[2] = { Lo, Hi };
3486 return DAG.getMergeValues(Ops, 2, dl);
3489 SDValue ARMTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
3490 SelectionDAG &DAG) const {
3491 // The rounding mode is in bits 23:22 of the FPSCR.
3492 // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0
3493 // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3)
3494 // so that the shift + and get folded into a bitfield extract.
3495 DebugLoc dl = Op.getDebugLoc();
3496 SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::i32,
3497 DAG.getConstant(Intrinsic::arm_get_fpscr,
3499 SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPSCR,
3500 DAG.getConstant(1U << 22, MVT::i32));
3501 SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds,
3502 DAG.getConstant(22, MVT::i32));
3503 return DAG.getNode(ISD::AND, dl, MVT::i32, RMODE,
3504 DAG.getConstant(3, MVT::i32));
3507 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
3508 const ARMSubtarget *ST) {
3509 EVT VT = N->getValueType(0);
3510 DebugLoc dl = N->getDebugLoc();
3512 if (!ST->hasV6T2Ops())
3515 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
3516 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
3519 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
3520 const ARMSubtarget *ST) {
3521 EVT VT = N->getValueType(0);
3522 DebugLoc dl = N->getDebugLoc();
3527 // Lower vector shifts on NEON to use VSHL.
3528 assert(ST->hasNEON() && "unexpected vector shift");
3530 // Left shifts translate directly to the vshiftu intrinsic.
3531 if (N->getOpcode() == ISD::SHL)
3532 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
3533 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
3534 N->getOperand(0), N->getOperand(1));
3536 assert((N->getOpcode() == ISD::SRA ||
3537 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
3539 // NEON uses the same intrinsics for both left and right shifts. For
3540 // right shifts, the shift amounts are negative, so negate the vector of
3542 EVT ShiftVT = N->getOperand(1).getValueType();
3543 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
3544 getZeroVector(ShiftVT, DAG, dl),
3546 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
3547 Intrinsic::arm_neon_vshifts :
3548 Intrinsic::arm_neon_vshiftu);
3549 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
3550 DAG.getConstant(vshiftInt, MVT::i32),
3551 N->getOperand(0), NegatedCount);
3554 static SDValue Expand64BitShift(SDNode *N, SelectionDAG &DAG,
3555 const ARMSubtarget *ST) {
3556 EVT VT = N->getValueType(0);
3557 DebugLoc dl = N->getDebugLoc();
3559 // We can get here for a node like i32 = ISD::SHL i32, i64
3563 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
3564 "Unknown shift to lower!");
3566 // We only lower SRA, SRL of 1 here, all others use generic lowering.
3567 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
3568 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
3571 // If we are in thumb mode, we don't have RRX.
3572 if (ST->isThumb1Only()) return SDValue();
3574 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
3575 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
3576 DAG.getConstant(0, MVT::i32));
3577 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
3578 DAG.getConstant(1, MVT::i32));
3580 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
3581 // captures the result into a carry flag.
3582 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
3583 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Glue), &Hi, 1);
3585 // The low part is an ARMISD::RRX operand, which shifts the carry in.
3586 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
3588 // Merge the pieces into a single i64 value.
3589 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
3592 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
3593 SDValue TmpOp0, TmpOp1;
3594 bool Invert = false;
3598 SDValue Op0 = Op.getOperand(0);
3599 SDValue Op1 = Op.getOperand(1);
3600 SDValue CC = Op.getOperand(2);
3601 EVT VT = Op.getValueType();
3602 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
3603 DebugLoc dl = Op.getDebugLoc();
3605 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
3606 switch (SetCCOpcode) {
3607 default: llvm_unreachable("Illegal FP comparison");
3609 case ISD::SETNE: Invert = true; // Fallthrough
3611 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
3613 case ISD::SETLT: Swap = true; // Fallthrough
3615 case ISD::SETGT: Opc = ARMISD::VCGT; break;
3617 case ISD::SETLE: Swap = true; // Fallthrough
3619 case ISD::SETGE: Opc = ARMISD::VCGE; break;
3620 case ISD::SETUGE: Swap = true; // Fallthrough
3621 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
3622 case ISD::SETUGT: Swap = true; // Fallthrough
3623 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
3624 case ISD::SETUEQ: Invert = true; // Fallthrough
3626 // Expand this to (OLT | OGT).
3630 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
3631 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
3633 case ISD::SETUO: Invert = true; // Fallthrough
3635 // Expand this to (OLT | OGE).
3639 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
3640 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
3644 // Integer comparisons.
3645 switch (SetCCOpcode) {
3646 default: llvm_unreachable("Illegal integer comparison");
3647 case ISD::SETNE: Invert = true;
3648 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
3649 case ISD::SETLT: Swap = true;
3650 case ISD::SETGT: Opc = ARMISD::VCGT; break;
3651 case ISD::SETLE: Swap = true;
3652 case ISD::SETGE: Opc = ARMISD::VCGE; break;
3653 case ISD::SETULT: Swap = true;
3654 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
3655 case ISD::SETULE: Swap = true;
3656 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
3659 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
3660 if (Opc == ARMISD::VCEQ) {
3663 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
3665 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
3668 // Ignore bitconvert.
3669 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BITCAST)
3670 AndOp = AndOp.getOperand(0);
3672 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
3674 Op0 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(0));
3675 Op1 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(1));
3682 std::swap(Op0, Op1);
3684 // If one of the operands is a constant vector zero, attempt to fold the
3685 // comparison to a specialized compare-against-zero form.
3687 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
3689 else if (ISD::isBuildVectorAllZeros(Op0.getNode())) {
3690 if (Opc == ARMISD::VCGE)
3691 Opc = ARMISD::VCLEZ;
3692 else if (Opc == ARMISD::VCGT)
3693 Opc = ARMISD::VCLTZ;
3698 if (SingleOp.getNode()) {
3701 Result = DAG.getNode(ARMISD::VCEQZ, dl, VT, SingleOp); break;
3703 Result = DAG.getNode(ARMISD::VCGEZ, dl, VT, SingleOp); break;
3705 Result = DAG.getNode(ARMISD::VCLEZ, dl, VT, SingleOp); break;
3707 Result = DAG.getNode(ARMISD::VCGTZ, dl, VT, SingleOp); break;
3709 Result = DAG.getNode(ARMISD::VCLTZ, dl, VT, SingleOp); break;
3711 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
3714 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
3718 Result = DAG.getNOT(dl, Result, VT);
3723 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
3724 /// valid vector constant for a NEON instruction with a "modified immediate"
3725 /// operand (e.g., VMOV). If so, return the encoded value.
3726 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
3727 unsigned SplatBitSize, SelectionDAG &DAG,
3728 EVT &VT, bool is128Bits, NEONModImmType type) {
3729 unsigned OpCmode, Imm;
3731 // SplatBitSize is set to the smallest size that splats the vector, so a
3732 // zero vector will always have SplatBitSize == 8. However, NEON modified
3733 // immediate instructions others than VMOV do not support the 8-bit encoding
3734 // of a zero vector, and the default encoding of zero is supposed to be the
3739 switch (SplatBitSize) {
3741 if (type != VMOVModImm)
3743 // Any 1-byte value is OK. Op=0, Cmode=1110.
3744 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
3747 VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
3751 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
3752 VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
3753 if ((SplatBits & ~0xff) == 0) {
3754 // Value = 0x00nn: Op=x, Cmode=100x.
3759 if ((SplatBits & ~0xff00) == 0) {
3760 // Value = 0xnn00: Op=x, Cmode=101x.
3762 Imm = SplatBits >> 8;
3768 // NEON's 32-bit VMOV supports splat values where:
3769 // * only one byte is nonzero, or
3770 // * the least significant byte is 0xff and the second byte is nonzero, or
3771 // * the least significant 2 bytes are 0xff and the third is nonzero.
3772 VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
3773 if ((SplatBits & ~0xff) == 0) {
3774 // Value = 0x000000nn: Op=x, Cmode=000x.
3779 if ((SplatBits & ~0xff00) == 0) {
3780 // Value = 0x0000nn00: Op=x, Cmode=001x.
3782 Imm = SplatBits >> 8;
3785 if ((SplatBits & ~0xff0000) == 0) {
3786 // Value = 0x00nn0000: Op=x, Cmode=010x.
3788 Imm = SplatBits >> 16;
3791 if ((SplatBits & ~0xff000000) == 0) {
3792 // Value = 0xnn000000: Op=x, Cmode=011x.
3794 Imm = SplatBits >> 24;
3798 // cmode == 0b1100 and cmode == 0b1101 are not supported for VORR or VBIC
3799 if (type == OtherModImm) return SDValue();
3801 if ((SplatBits & ~0xffff) == 0 &&
3802 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
3803 // Value = 0x0000nnff: Op=x, Cmode=1100.
3805 Imm = SplatBits >> 8;
3810 if ((SplatBits & ~0xffffff) == 0 &&
3811 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
3812 // Value = 0x00nnffff: Op=x, Cmode=1101.
3814 Imm = SplatBits >> 16;
3815 SplatBits |= 0xffff;
3819 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
3820 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
3821 // VMOV.I32. A (very) minor optimization would be to replicate the value
3822 // and fall through here to test for a valid 64-bit splat. But, then the
3823 // caller would also need to check and handle the change in size.
3827 if (type != VMOVModImm)
3829 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
3830 uint64_t BitMask = 0xff;
3832 unsigned ImmMask = 1;
3834 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
3835 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
3838 } else if ((SplatBits & BitMask) != 0) {
3844 // Op=1, Cmode=1110.
3847 VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
3852 llvm_unreachable("unexpected size for isNEONModifiedImm");
3855 unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm);
3856 return DAG.getTargetConstant(EncodedVal, MVT::i32);
3859 SDValue ARMTargetLowering::LowerConstantFP(SDValue Op, SelectionDAG &DAG,
3860 const ARMSubtarget *ST) const {
3861 if (!ST->useNEONForSinglePrecisionFP() || !ST->hasVFP3() || ST->hasD16())
3864 ConstantFPSDNode *CFP = cast<ConstantFPSDNode>(Op);
3865 assert(Op.getValueType() == MVT::f32 &&
3866 "ConstantFP custom lowering should only occur for f32.");
3868 // Try splatting with a VMOV.f32...
3869 APFloat FPVal = CFP->getValueAPF();
3870 int ImmVal = ARM_AM::getFP32Imm(FPVal);
3872 DebugLoc DL = Op.getDebugLoc();
3873 SDValue NewVal = DAG.getTargetConstant(ImmVal, MVT::i32);
3874 SDValue VecConstant = DAG.getNode(ARMISD::VMOVFPIMM, DL, MVT::v2f32,
3876 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecConstant,
3877 DAG.getConstant(0, MVT::i32));
3880 // If that fails, try a VMOV.i32
3882 unsigned iVal = FPVal.bitcastToAPInt().getZExtValue();
3883 SDValue NewVal = isNEONModifiedImm(iVal, 0, 32, DAG, VMovVT, false,
3885 if (NewVal != SDValue()) {
3886 DebugLoc DL = Op.getDebugLoc();
3887 SDValue VecConstant = DAG.getNode(ARMISD::VMOVIMM, DL, VMovVT,
3889 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
3891 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
3892 DAG.getConstant(0, MVT::i32));
3895 // Finally, try a VMVN.i32
3896 NewVal = isNEONModifiedImm(~iVal & 0xffffffff, 0, 32, DAG, VMovVT, false,
3898 if (NewVal != SDValue()) {
3899 DebugLoc DL = Op.getDebugLoc();
3900 SDValue VecConstant = DAG.getNode(ARMISD::VMVNIMM, DL, VMovVT, NewVal);
3901 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
3903 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
3904 DAG.getConstant(0, MVT::i32));
3911 static bool isVEXTMask(ArrayRef<int> M, EVT VT,
3912 bool &ReverseVEXT, unsigned &Imm) {
3913 unsigned NumElts = VT.getVectorNumElements();
3914 ReverseVEXT = false;
3916 // Assume that the first shuffle index is not UNDEF. Fail if it is.
3922 // If this is a VEXT shuffle, the immediate value is the index of the first
3923 // element. The other shuffle indices must be the successive elements after
3925 unsigned ExpectedElt = Imm;
3926 for (unsigned i = 1; i < NumElts; ++i) {
3927 // Increment the expected index. If it wraps around, it may still be
3928 // a VEXT but the source vectors must be swapped.
3930 if (ExpectedElt == NumElts * 2) {
3935 if (M[i] < 0) continue; // ignore UNDEF indices
3936 if (ExpectedElt != static_cast<unsigned>(M[i]))
3940 // Adjust the index value if the source operands will be swapped.
3947 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
3948 /// instruction with the specified blocksize. (The order of the elements
3949 /// within each block of the vector is reversed.)
3950 static bool isVREVMask(ArrayRef<int> M, EVT VT, unsigned BlockSize) {
3951 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
3952 "Only possible block sizes for VREV are: 16, 32, 64");
3954 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3958 unsigned NumElts = VT.getVectorNumElements();
3959 unsigned BlockElts = M[0] + 1;
3960 // If the first shuffle index is UNDEF, be optimistic.
3962 BlockElts = BlockSize / EltSz;
3964 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
3967 for (unsigned i = 0; i < NumElts; ++i) {
3968 if (M[i] < 0) continue; // ignore UNDEF indices
3969 if ((unsigned) M[i] != (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
3976 static bool isVTBLMask(ArrayRef<int> M, EVT VT) {
3977 // We can handle <8 x i8> vector shuffles. If the index in the mask is out of
3978 // range, then 0 is placed into the resulting vector. So pretty much any mask
3979 // of 8 elements can work here.
3980 return VT == MVT::v8i8 && M.size() == 8;
3983 static bool isVTRNMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
3984 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3988 unsigned NumElts = VT.getVectorNumElements();
3989 WhichResult = (M[0] == 0 ? 0 : 1);
3990 for (unsigned i = 0; i < NumElts; i += 2) {
3991 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
3992 (M[i+1] >= 0 && (unsigned) M[i+1] != i + NumElts + WhichResult))
3998 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
3999 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4000 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
4001 static bool isVTRN_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4002 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4006 unsigned NumElts = VT.getVectorNumElements();
4007 WhichResult = (M[0] == 0 ? 0 : 1);
4008 for (unsigned i = 0; i < NumElts; i += 2) {
4009 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
4010 (M[i+1] >= 0 && (unsigned) M[i+1] != i + WhichResult))
4016 static bool isVUZPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4017 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4021 unsigned NumElts = VT.getVectorNumElements();
4022 WhichResult = (M[0] == 0 ? 0 : 1);
4023 for (unsigned i = 0; i != NumElts; ++i) {
4024 if (M[i] < 0) continue; // ignore UNDEF indices
4025 if ((unsigned) M[i] != 2 * i + WhichResult)
4029 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4030 if (VT.is64BitVector() && EltSz == 32)
4036 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
4037 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4038 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
4039 static bool isVUZP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4040 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4044 unsigned Half = VT.getVectorNumElements() / 2;
4045 WhichResult = (M[0] == 0 ? 0 : 1);
4046 for (unsigned j = 0; j != 2; ++j) {
4047 unsigned Idx = WhichResult;
4048 for (unsigned i = 0; i != Half; ++i) {
4049 int MIdx = M[i + j * Half];
4050 if (MIdx >= 0 && (unsigned) MIdx != Idx)
4056 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4057 if (VT.is64BitVector() && EltSz == 32)
4063 static bool isVZIPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4064 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4068 unsigned NumElts = VT.getVectorNumElements();
4069 WhichResult = (M[0] == 0 ? 0 : 1);
4070 unsigned Idx = WhichResult * NumElts / 2;
4071 for (unsigned i = 0; i != NumElts; i += 2) {
4072 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
4073 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx + NumElts))
4078 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4079 if (VT.is64BitVector() && EltSz == 32)
4085 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
4086 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4087 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
4088 static bool isVZIP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4089 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4093 unsigned NumElts = VT.getVectorNumElements();
4094 WhichResult = (M[0] == 0 ? 0 : 1);
4095 unsigned Idx = WhichResult * NumElts / 2;
4096 for (unsigned i = 0; i != NumElts; i += 2) {
4097 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
4098 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx))
4103 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4104 if (VT.is64BitVector() && EltSz == 32)
4110 // If N is an integer constant that can be moved into a register in one
4111 // instruction, return an SDValue of such a constant (will become a MOV
4112 // instruction). Otherwise return null.
4113 static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG,
4114 const ARMSubtarget *ST, DebugLoc dl) {
4116 if (!isa<ConstantSDNode>(N))
4118 Val = cast<ConstantSDNode>(N)->getZExtValue();
4120 if (ST->isThumb1Only()) {
4121 if (Val <= 255 || ~Val <= 255)
4122 return DAG.getConstant(Val, MVT::i32);
4124 if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1)
4125 return DAG.getConstant(Val, MVT::i32);
4130 // If this is a case we can't handle, return null and let the default
4131 // expansion code take care of it.
4132 SDValue ARMTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
4133 const ARMSubtarget *ST) const {
4134 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
4135 DebugLoc dl = Op.getDebugLoc();
4136 EVT VT = Op.getValueType();
4138 APInt SplatBits, SplatUndef;
4139 unsigned SplatBitSize;
4141 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
4142 if (SplatBitSize <= 64) {
4143 // Check if an immediate VMOV works.
4145 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
4146 SplatUndef.getZExtValue(), SplatBitSize,
4147 DAG, VmovVT, VT.is128BitVector(),
4149 if (Val.getNode()) {
4150 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
4151 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4154 // Try an immediate VMVN.
4155 uint64_t NegatedImm = (~SplatBits).getZExtValue();
4156 Val = isNEONModifiedImm(NegatedImm,
4157 SplatUndef.getZExtValue(), SplatBitSize,
4158 DAG, VmovVT, VT.is128BitVector(),
4160 if (Val.getNode()) {
4161 SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
4162 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4165 // Use vmov.f32 to materialize other v2f32 and v4f32 splats.
4166 if ((VT == MVT::v2f32 || VT == MVT::v4f32) && SplatBitSize == 32) {
4167 int ImmVal = ARM_AM::getFP32Imm(SplatBits);
4169 SDValue Val = DAG.getTargetConstant(ImmVal, MVT::i32);
4170 return DAG.getNode(ARMISD::VMOVFPIMM, dl, VT, Val);
4176 // Scan through the operands to see if only one value is used.
4178 // As an optimisation, even if more than one value is used it may be more
4179 // profitable to splat with one value then change some lanes.
4181 // Heuristically we decide to do this if the vector has a "dominant" value,
4182 // defined as splatted to more than half of the lanes.
4183 unsigned NumElts = VT.getVectorNumElements();
4184 bool isOnlyLowElement = true;
4185 bool usesOnlyOneValue = true;
4186 bool hasDominantValue = false;
4187 bool isConstant = true;
4189 // Map of the number of times a particular SDValue appears in the
4191 DenseMap<SDValue, unsigned> ValueCounts;
4193 for (unsigned i = 0; i < NumElts; ++i) {
4194 SDValue V = Op.getOperand(i);
4195 if (V.getOpcode() == ISD::UNDEF)
4198 isOnlyLowElement = false;
4199 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
4202 ValueCounts.insert(std::make_pair(V, 0));
4203 unsigned &Count = ValueCounts[V];
4205 // Is this value dominant? (takes up more than half of the lanes)
4206 if (++Count > (NumElts / 2)) {
4207 hasDominantValue = true;
4211 if (ValueCounts.size() != 1)
4212 usesOnlyOneValue = false;
4213 if (!Value.getNode() && ValueCounts.size() > 0)
4214 Value = ValueCounts.begin()->first;
4216 if (ValueCounts.size() == 0)
4217 return DAG.getUNDEF(VT);
4219 if (isOnlyLowElement)
4220 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
4222 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4224 // Use VDUP for non-constant splats. For f32 constant splats, reduce to
4225 // i32 and try again.
4226 if (hasDominantValue && EltSize <= 32) {
4230 // If we are VDUPing a value that comes directly from a vector, that will
4231 // cause an unnecessary move to and from a GPR, where instead we could
4232 // just use VDUPLANE.
4233 if (Value->getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
4234 // We need to create a new undef vector to use for the VDUPLANE if the
4235 // size of the vector from which we get the value is different than the
4236 // size of the vector that we need to create. We will insert the element
4237 // such that the register coalescer will remove unnecessary copies.
4238 if (VT != Value->getOperand(0).getValueType()) {
4239 ConstantSDNode *constIndex;
4240 constIndex = dyn_cast<ConstantSDNode>(Value->getOperand(1));
4241 assert(constIndex && "The index is not a constant!");
4242 unsigned index = constIndex->getAPIntValue().getLimitedValue() %
4243 VT.getVectorNumElements();
4244 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
4245 DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DAG.getUNDEF(VT),
4246 Value, DAG.getConstant(index, MVT::i32)),
4247 DAG.getConstant(index, MVT::i32));
4249 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
4250 Value->getOperand(0), Value->getOperand(1));
4254 N = DAG.getNode(ARMISD::VDUP, dl, VT, Value);
4256 if (!usesOnlyOneValue) {
4257 // The dominant value was splatted as 'N', but we now have to insert
4258 // all differing elements.
4259 for (unsigned I = 0; I < NumElts; ++I) {
4260 if (Op.getOperand(I) == Value)
4262 SmallVector<SDValue, 3> Ops;
4264 Ops.push_back(Op.getOperand(I));
4265 Ops.push_back(DAG.getConstant(I, MVT::i32));
4266 N = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, &Ops[0], 3);
4271 if (VT.getVectorElementType().isFloatingPoint()) {
4272 SmallVector<SDValue, 8> Ops;
4273 for (unsigned i = 0; i < NumElts; ++i)
4274 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, MVT::i32,
4276 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
4277 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, &Ops[0], NumElts);
4278 Val = LowerBUILD_VECTOR(Val, DAG, ST);
4280 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4282 if (usesOnlyOneValue) {
4283 SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
4284 if (isConstant && Val.getNode())
4285 return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
4289 // If all elements are constants and the case above didn't get hit, fall back
4290 // to the default expansion, which will generate a load from the constant
4295 // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
4297 SDValue shuffle = ReconstructShuffle(Op, DAG);
4298 if (shuffle != SDValue())
4302 // Vectors with 32- or 64-bit elements can be built by directly assigning
4303 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
4304 // will be legalized.
4305 if (EltSize >= 32) {
4306 // Do the expansion with floating-point types, since that is what the VFP
4307 // registers are defined to use, and since i64 is not legal.
4308 EVT EltVT = EVT::getFloatingPointVT(EltSize);
4309 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
4310 SmallVector<SDValue, 8> Ops;
4311 for (unsigned i = 0; i < NumElts; ++i)
4312 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, EltVT, Op.getOperand(i)));
4313 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
4314 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4320 // Gather data to see if the operation can be modelled as a
4321 // shuffle in combination with VEXTs.
4322 SDValue ARMTargetLowering::ReconstructShuffle(SDValue Op,
4323 SelectionDAG &DAG) const {
4324 DebugLoc dl = Op.getDebugLoc();
4325 EVT VT = Op.getValueType();
4326 unsigned NumElts = VT.getVectorNumElements();
4328 SmallVector<SDValue, 2> SourceVecs;
4329 SmallVector<unsigned, 2> MinElts;
4330 SmallVector<unsigned, 2> MaxElts;
4332 for (unsigned i = 0; i < NumElts; ++i) {
4333 SDValue V = Op.getOperand(i);
4334 if (V.getOpcode() == ISD::UNDEF)
4336 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
4337 // A shuffle can only come from building a vector from various
4338 // elements of other vectors.
4340 } else if (V.getOperand(0).getValueType().getVectorElementType() !=
4341 VT.getVectorElementType()) {
4342 // This code doesn't know how to handle shuffles where the vector
4343 // element types do not match (this happens because type legalization
4344 // promotes the return type of EXTRACT_VECTOR_ELT).
4345 // FIXME: It might be appropriate to extend this code to handle
4346 // mismatched types.
4350 // Record this extraction against the appropriate vector if possible...
4351 SDValue SourceVec = V.getOperand(0);
4352 // If the element number isn't a constant, we can't effectively
4353 // analyze what's going on.
4354 if (!isa<ConstantSDNode>(V.getOperand(1)))
4356 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
4357 bool FoundSource = false;
4358 for (unsigned j = 0; j < SourceVecs.size(); ++j) {
4359 if (SourceVecs[j] == SourceVec) {
4360 if (MinElts[j] > EltNo)
4362 if (MaxElts[j] < EltNo)
4369 // Or record a new source if not...
4371 SourceVecs.push_back(SourceVec);
4372 MinElts.push_back(EltNo);
4373 MaxElts.push_back(EltNo);
4377 // Currently only do something sane when at most two source vectors
4379 if (SourceVecs.size() > 2)
4382 SDValue ShuffleSrcs[2] = {DAG.getUNDEF(VT), DAG.getUNDEF(VT) };
4383 int VEXTOffsets[2] = {0, 0};
4385 // This loop extracts the usage patterns of the source vectors
4386 // and prepares appropriate SDValues for a shuffle if possible.
4387 for (unsigned i = 0; i < SourceVecs.size(); ++i) {
4388 if (SourceVecs[i].getValueType() == VT) {
4389 // No VEXT necessary
4390 ShuffleSrcs[i] = SourceVecs[i];
4393 } else if (SourceVecs[i].getValueType().getVectorNumElements() < NumElts) {
4394 // It probably isn't worth padding out a smaller vector just to
4395 // break it down again in a shuffle.
4399 // Since only 64-bit and 128-bit vectors are legal on ARM and
4400 // we've eliminated the other cases...
4401 assert(SourceVecs[i].getValueType().getVectorNumElements() == 2*NumElts &&
4402 "unexpected vector sizes in ReconstructShuffle");
4404 if (MaxElts[i] - MinElts[i] >= NumElts) {
4405 // Span too large for a VEXT to cope
4409 if (MinElts[i] >= NumElts) {
4410 // The extraction can just take the second half
4411 VEXTOffsets[i] = NumElts;
4412 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4414 DAG.getIntPtrConstant(NumElts));
4415 } else if (MaxElts[i] < NumElts) {
4416 // The extraction can just take the first half
4418 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4420 DAG.getIntPtrConstant(0));
4422 // An actual VEXT is needed
4423 VEXTOffsets[i] = MinElts[i];
4424 SDValue VEXTSrc1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4426 DAG.getIntPtrConstant(0));
4427 SDValue VEXTSrc2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4429 DAG.getIntPtrConstant(NumElts));
4430 ShuffleSrcs[i] = DAG.getNode(ARMISD::VEXT, dl, VT, VEXTSrc1, VEXTSrc2,
4431 DAG.getConstant(VEXTOffsets[i], MVT::i32));
4435 SmallVector<int, 8> Mask;
4437 for (unsigned i = 0; i < NumElts; ++i) {
4438 SDValue Entry = Op.getOperand(i);
4439 if (Entry.getOpcode() == ISD::UNDEF) {
4444 SDValue ExtractVec = Entry.getOperand(0);
4445 int ExtractElt = cast<ConstantSDNode>(Op.getOperand(i)
4446 .getOperand(1))->getSExtValue();
4447 if (ExtractVec == SourceVecs[0]) {
4448 Mask.push_back(ExtractElt - VEXTOffsets[0]);
4450 Mask.push_back(ExtractElt + NumElts - VEXTOffsets[1]);
4454 // Final check before we try to produce nonsense...
4455 if (isShuffleMaskLegal(Mask, VT))
4456 return DAG.getVectorShuffle(VT, dl, ShuffleSrcs[0], ShuffleSrcs[1],
4462 /// isShuffleMaskLegal - Targets can use this to indicate that they only
4463 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
4464 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
4465 /// are assumed to be legal.
4467 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
4469 if (VT.getVectorNumElements() == 4 &&
4470 (VT.is128BitVector() || VT.is64BitVector())) {
4471 unsigned PFIndexes[4];
4472 for (unsigned i = 0; i != 4; ++i) {
4476 PFIndexes[i] = M[i];
4479 // Compute the index in the perfect shuffle table.
4480 unsigned PFTableIndex =
4481 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
4482 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4483 unsigned Cost = (PFEntry >> 30);
4490 unsigned Imm, WhichResult;
4492 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4493 return (EltSize >= 32 ||
4494 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
4495 isVREVMask(M, VT, 64) ||
4496 isVREVMask(M, VT, 32) ||
4497 isVREVMask(M, VT, 16) ||
4498 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
4499 isVTBLMask(M, VT) ||
4500 isVTRNMask(M, VT, WhichResult) ||
4501 isVUZPMask(M, VT, WhichResult) ||
4502 isVZIPMask(M, VT, WhichResult) ||
4503 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
4504 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
4505 isVZIP_v_undef_Mask(M, VT, WhichResult));
4508 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
4509 /// the specified operations to build the shuffle.
4510 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
4511 SDValue RHS, SelectionDAG &DAG,
4513 unsigned OpNum = (PFEntry >> 26) & 0x0F;
4514 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
4515 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
4518 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
4527 OP_VUZPL, // VUZP, left result
4528 OP_VUZPR, // VUZP, right result
4529 OP_VZIPL, // VZIP, left result
4530 OP_VZIPR, // VZIP, right result
4531 OP_VTRNL, // VTRN, left result
4532 OP_VTRNR // VTRN, right result
4535 if (OpNum == OP_COPY) {
4536 if (LHSID == (1*9+2)*9+3) return LHS;
4537 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
4541 SDValue OpLHS, OpRHS;
4542 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
4543 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
4544 EVT VT = OpLHS.getValueType();
4547 default: llvm_unreachable("Unknown shuffle opcode!");
4549 // VREV divides the vector in half and swaps within the half.
4550 if (VT.getVectorElementType() == MVT::i32 ||
4551 VT.getVectorElementType() == MVT::f32)
4552 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
4553 // vrev <4 x i16> -> VREV32
4554 if (VT.getVectorElementType() == MVT::i16)
4555 return DAG.getNode(ARMISD::VREV32, dl, VT, OpLHS);
4556 // vrev <4 x i8> -> VREV16
4557 assert(VT.getVectorElementType() == MVT::i8);
4558 return DAG.getNode(ARMISD::VREV16, dl, VT, OpLHS);
4563 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
4564 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
4568 return DAG.getNode(ARMISD::VEXT, dl, VT,
4570 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
4573 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4574 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
4577 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4578 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
4581 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4582 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
4586 static SDValue LowerVECTOR_SHUFFLEv8i8(SDValue Op,
4587 ArrayRef<int> ShuffleMask,
4588 SelectionDAG &DAG) {
4589 // Check to see if we can use the VTBL instruction.
4590 SDValue V1 = Op.getOperand(0);
4591 SDValue V2 = Op.getOperand(1);
4592 DebugLoc DL = Op.getDebugLoc();
4594 SmallVector<SDValue, 8> VTBLMask;
4595 for (ArrayRef<int>::iterator
4596 I = ShuffleMask.begin(), E = ShuffleMask.end(); I != E; ++I)
4597 VTBLMask.push_back(DAG.getConstant(*I, MVT::i32));
4599 if (V2.getNode()->getOpcode() == ISD::UNDEF)
4600 return DAG.getNode(ARMISD::VTBL1, DL, MVT::v8i8, V1,
4601 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8,
4604 return DAG.getNode(ARMISD::VTBL2, DL, MVT::v8i8, V1, V2,
4605 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8,
4609 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
4610 SDValue V1 = Op.getOperand(0);
4611 SDValue V2 = Op.getOperand(1);
4612 DebugLoc dl = Op.getDebugLoc();
4613 EVT VT = Op.getValueType();
4614 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
4616 // Convert shuffles that are directly supported on NEON to target-specific
4617 // DAG nodes, instead of keeping them as shuffles and matching them again
4618 // during code selection. This is more efficient and avoids the possibility
4619 // of inconsistencies between legalization and selection.
4620 // FIXME: floating-point vectors should be canonicalized to integer vectors
4621 // of the same time so that they get CSEd properly.
4622 ArrayRef<int> ShuffleMask = SVN->getMask();
4624 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4625 if (EltSize <= 32) {
4626 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
4627 int Lane = SVN->getSplatIndex();
4628 // If this is undef splat, generate it via "just" vdup, if possible.
4629 if (Lane == -1) Lane = 0;
4631 // Test if V1 is a SCALAR_TO_VECTOR.
4632 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
4633 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
4635 // Test if V1 is a BUILD_VECTOR which is equivalent to a SCALAR_TO_VECTOR
4636 // (and probably will turn into a SCALAR_TO_VECTOR once legalization
4638 if (Lane == 0 && V1.getOpcode() == ISD::BUILD_VECTOR &&
4639 !isa<ConstantSDNode>(V1.getOperand(0))) {
4640 bool IsScalarToVector = true;
4641 for (unsigned i = 1, e = V1.getNumOperands(); i != e; ++i)
4642 if (V1.getOperand(i).getOpcode() != ISD::UNDEF) {
4643 IsScalarToVector = false;
4646 if (IsScalarToVector)
4647 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
4649 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
4650 DAG.getConstant(Lane, MVT::i32));
4655 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
4658 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
4659 DAG.getConstant(Imm, MVT::i32));
4662 if (isVREVMask(ShuffleMask, VT, 64))
4663 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
4664 if (isVREVMask(ShuffleMask, VT, 32))
4665 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
4666 if (isVREVMask(ShuffleMask, VT, 16))
4667 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
4669 // Check for Neon shuffles that modify both input vectors in place.
4670 // If both results are used, i.e., if there are two shuffles with the same
4671 // source operands and with masks corresponding to both results of one of
4672 // these operations, DAG memoization will ensure that a single node is
4673 // used for both shuffles.
4674 unsigned WhichResult;
4675 if (isVTRNMask(ShuffleMask, VT, WhichResult))
4676 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4677 V1, V2).getValue(WhichResult);
4678 if (isVUZPMask(ShuffleMask, VT, WhichResult))
4679 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4680 V1, V2).getValue(WhichResult);
4681 if (isVZIPMask(ShuffleMask, VT, WhichResult))
4682 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4683 V1, V2).getValue(WhichResult);
4685 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
4686 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4687 V1, V1).getValue(WhichResult);
4688 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
4689 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4690 V1, V1).getValue(WhichResult);
4691 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
4692 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4693 V1, V1).getValue(WhichResult);
4696 // If the shuffle is not directly supported and it has 4 elements, use
4697 // the PerfectShuffle-generated table to synthesize it from other shuffles.
4698 unsigned NumElts = VT.getVectorNumElements();
4700 unsigned PFIndexes[4];
4701 for (unsigned i = 0; i != 4; ++i) {
4702 if (ShuffleMask[i] < 0)
4705 PFIndexes[i] = ShuffleMask[i];
4708 // Compute the index in the perfect shuffle table.
4709 unsigned PFTableIndex =
4710 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
4711 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4712 unsigned Cost = (PFEntry >> 30);
4715 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
4718 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
4719 if (EltSize >= 32) {
4720 // Do the expansion with floating-point types, since that is what the VFP
4721 // registers are defined to use, and since i64 is not legal.
4722 EVT EltVT = EVT::getFloatingPointVT(EltSize);
4723 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
4724 V1 = DAG.getNode(ISD::BITCAST, dl, VecVT, V1);
4725 V2 = DAG.getNode(ISD::BITCAST, dl, VecVT, V2);
4726 SmallVector<SDValue, 8> Ops;
4727 for (unsigned i = 0; i < NumElts; ++i) {
4728 if (ShuffleMask[i] < 0)
4729 Ops.push_back(DAG.getUNDEF(EltVT));
4731 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
4732 ShuffleMask[i] < (int)NumElts ? V1 : V2,
4733 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
4736 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
4737 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4740 if (VT == MVT::v8i8) {
4741 SDValue NewOp = LowerVECTOR_SHUFFLEv8i8(Op, ShuffleMask, DAG);
4742 if (NewOp.getNode())
4749 static SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4750 // INSERT_VECTOR_ELT is legal only for immediate indexes.
4751 SDValue Lane = Op.getOperand(2);
4752 if (!isa<ConstantSDNode>(Lane))
4758 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4759 // EXTRACT_VECTOR_ELT is legal only for immediate indexes.
4760 SDValue Lane = Op.getOperand(1);
4761 if (!isa<ConstantSDNode>(Lane))
4764 SDValue Vec = Op.getOperand(0);
4765 if (Op.getValueType() == MVT::i32 &&
4766 Vec.getValueType().getVectorElementType().getSizeInBits() < 32) {
4767 DebugLoc dl = Op.getDebugLoc();
4768 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
4774 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
4775 // The only time a CONCAT_VECTORS operation can have legal types is when
4776 // two 64-bit vectors are concatenated to a 128-bit vector.
4777 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
4778 "unexpected CONCAT_VECTORS");
4779 DebugLoc dl = Op.getDebugLoc();
4780 SDValue Val = DAG.getUNDEF(MVT::v2f64);
4781 SDValue Op0 = Op.getOperand(0);
4782 SDValue Op1 = Op.getOperand(1);
4783 if (Op0.getOpcode() != ISD::UNDEF)
4784 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
4785 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op0),
4786 DAG.getIntPtrConstant(0));
4787 if (Op1.getOpcode() != ISD::UNDEF)
4788 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
4789 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op1),
4790 DAG.getIntPtrConstant(1));
4791 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Val);
4794 /// isExtendedBUILD_VECTOR - Check if N is a constant BUILD_VECTOR where each
4795 /// element has been zero/sign-extended, depending on the isSigned parameter,
4796 /// from an integer type half its size.
4797 static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG,
4799 // A v2i64 BUILD_VECTOR will have been legalized to a BITCAST from v4i32.
4800 EVT VT = N->getValueType(0);
4801 if (VT == MVT::v2i64 && N->getOpcode() == ISD::BITCAST) {
4802 SDNode *BVN = N->getOperand(0).getNode();
4803 if (BVN->getValueType(0) != MVT::v4i32 ||
4804 BVN->getOpcode() != ISD::BUILD_VECTOR)
4806 unsigned LoElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
4807 unsigned HiElt = 1 - LoElt;
4808 ConstantSDNode *Lo0 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt));
4809 ConstantSDNode *Hi0 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt));
4810 ConstantSDNode *Lo1 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt+2));
4811 ConstantSDNode *Hi1 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt+2));
4812 if (!Lo0 || !Hi0 || !Lo1 || !Hi1)
4815 if (Hi0->getSExtValue() == Lo0->getSExtValue() >> 32 &&
4816 Hi1->getSExtValue() == Lo1->getSExtValue() >> 32)
4819 if (Hi0->isNullValue() && Hi1->isNullValue())
4825 if (N->getOpcode() != ISD::BUILD_VECTOR)
4828 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
4829 SDNode *Elt = N->getOperand(i).getNode();
4830 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
4831 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4832 unsigned HalfSize = EltSize / 2;
4834 if (!isIntN(HalfSize, C->getSExtValue()))
4837 if (!isUIntN(HalfSize, C->getZExtValue()))
4848 /// isSignExtended - Check if a node is a vector value that is sign-extended
4849 /// or a constant BUILD_VECTOR with sign-extended elements.
4850 static bool isSignExtended(SDNode *N, SelectionDAG &DAG) {
4851 if (N->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N))
4853 if (isExtendedBUILD_VECTOR(N, DAG, true))
4858 /// isZeroExtended - Check if a node is a vector value that is zero-extended
4859 /// or a constant BUILD_VECTOR with zero-extended elements.
4860 static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) {
4861 if (N->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N))
4863 if (isExtendedBUILD_VECTOR(N, DAG, false))
4868 /// SkipExtension - For a node that is a SIGN_EXTEND, ZERO_EXTEND, extending
4869 /// load, or BUILD_VECTOR with extended elements, return the unextended value.
4870 static SDValue SkipExtension(SDNode *N, SelectionDAG &DAG) {
4871 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
4872 return N->getOperand(0);
4873 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
4874 return DAG.getLoad(LD->getMemoryVT(), N->getDebugLoc(), LD->getChain(),
4875 LD->getBasePtr(), LD->getPointerInfo(), LD->isVolatile(),
4876 LD->isNonTemporal(), LD->isInvariant(),
4877 LD->getAlignment());
4878 // Otherwise, the value must be a BUILD_VECTOR. For v2i64, it will
4879 // have been legalized as a BITCAST from v4i32.
4880 if (N->getOpcode() == ISD::BITCAST) {
4881 SDNode *BVN = N->getOperand(0).getNode();
4882 assert(BVN->getOpcode() == ISD::BUILD_VECTOR &&
4883 BVN->getValueType(0) == MVT::v4i32 && "expected v4i32 BUILD_VECTOR");
4884 unsigned LowElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
4885 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(), MVT::v2i32,
4886 BVN->getOperand(LowElt), BVN->getOperand(LowElt+2));
4888 // Construct a new BUILD_VECTOR with elements truncated to half the size.
4889 assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR");
4890 EVT VT = N->getValueType(0);
4891 unsigned EltSize = VT.getVectorElementType().getSizeInBits() / 2;
4892 unsigned NumElts = VT.getVectorNumElements();
4893 MVT TruncVT = MVT::getIntegerVT(EltSize);
4894 SmallVector<SDValue, 8> Ops;
4895 for (unsigned i = 0; i != NumElts; ++i) {
4896 ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i));
4897 const APInt &CInt = C->getAPIntValue();
4898 // Element types smaller than 32 bits are not legal, so use i32 elements.
4899 // The values are implicitly truncated so sext vs. zext doesn't matter.
4900 Ops.push_back(DAG.getConstant(CInt.zextOrTrunc(32), MVT::i32));
4902 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
4903 MVT::getVectorVT(TruncVT, NumElts), Ops.data(), NumElts);
4906 static bool isAddSubSExt(SDNode *N, SelectionDAG &DAG) {
4907 unsigned Opcode = N->getOpcode();
4908 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
4909 SDNode *N0 = N->getOperand(0).getNode();
4910 SDNode *N1 = N->getOperand(1).getNode();
4911 return N0->hasOneUse() && N1->hasOneUse() &&
4912 isSignExtended(N0, DAG) && isSignExtended(N1, DAG);
4917 static bool isAddSubZExt(SDNode *N, SelectionDAG &DAG) {
4918 unsigned Opcode = N->getOpcode();
4919 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
4920 SDNode *N0 = N->getOperand(0).getNode();
4921 SDNode *N1 = N->getOperand(1).getNode();
4922 return N0->hasOneUse() && N1->hasOneUse() &&
4923 isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG);
4928 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
4929 // Multiplications are only custom-lowered for 128-bit vectors so that
4930 // VMULL can be detected. Otherwise v2i64 multiplications are not legal.
4931 EVT VT = Op.getValueType();
4932 assert(VT.is128BitVector() && "unexpected type for custom-lowering ISD::MUL");
4933 SDNode *N0 = Op.getOperand(0).getNode();
4934 SDNode *N1 = Op.getOperand(1).getNode();
4935 unsigned NewOpc = 0;
4937 bool isN0SExt = isSignExtended(N0, DAG);
4938 bool isN1SExt = isSignExtended(N1, DAG);
4939 if (isN0SExt && isN1SExt)
4940 NewOpc = ARMISD::VMULLs;
4942 bool isN0ZExt = isZeroExtended(N0, DAG);
4943 bool isN1ZExt = isZeroExtended(N1, DAG);
4944 if (isN0ZExt && isN1ZExt)
4945 NewOpc = ARMISD::VMULLu;
4946 else if (isN1SExt || isN1ZExt) {
4947 // Look for (s/zext A + s/zext B) * (s/zext C). We want to turn these
4948 // into (s/zext A * s/zext C) + (s/zext B * s/zext C)
4949 if (isN1SExt && isAddSubSExt(N0, DAG)) {
4950 NewOpc = ARMISD::VMULLs;
4952 } else if (isN1ZExt && isAddSubZExt(N0, DAG)) {
4953 NewOpc = ARMISD::VMULLu;
4955 } else if (isN0ZExt && isAddSubZExt(N1, DAG)) {
4957 NewOpc = ARMISD::VMULLu;
4963 if (VT == MVT::v2i64)
4964 // Fall through to expand this. It is not legal.
4967 // Other vector multiplications are legal.
4972 // Legalize to a VMULL instruction.
4973 DebugLoc DL = Op.getDebugLoc();
4975 SDValue Op1 = SkipExtension(N1, DAG);
4977 Op0 = SkipExtension(N0, DAG);
4978 assert(Op0.getValueType().is64BitVector() &&
4979 Op1.getValueType().is64BitVector() &&
4980 "unexpected types for extended operands to VMULL");
4981 return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
4984 // Optimizing (zext A + zext B) * C, to (VMULL A, C) + (VMULL B, C) during
4985 // isel lowering to take advantage of no-stall back to back vmul + vmla.
4992 SDValue N00 = SkipExtension(N0->getOperand(0).getNode(), DAG);
4993 SDValue N01 = SkipExtension(N0->getOperand(1).getNode(), DAG);
4994 EVT Op1VT = Op1.getValueType();
4995 return DAG.getNode(N0->getOpcode(), DL, VT,
4996 DAG.getNode(NewOpc, DL, VT,
4997 DAG.getNode(ISD::BITCAST, DL, Op1VT, N00), Op1),
4998 DAG.getNode(NewOpc, DL, VT,
4999 DAG.getNode(ISD::BITCAST, DL, Op1VT, N01), Op1));
5003 LowerSDIV_v4i8(SDValue X, SDValue Y, DebugLoc dl, SelectionDAG &DAG) {
5005 // float4 xf = vcvt_f32_s32(vmovl_s16(a.lo));
5006 // float4 yf = vcvt_f32_s32(vmovl_s16(b.lo));
5007 X = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, X);
5008 Y = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, Y);
5009 X = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, X);
5010 Y = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, Y);
5011 // Get reciprocal estimate.
5012 // float4 recip = vrecpeq_f32(yf);
5013 Y = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5014 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), Y);
5015 // Because char has a smaller range than uchar, we can actually get away
5016 // without any newton steps. This requires that we use a weird bias
5017 // of 0xb000, however (again, this has been exhaustively tested).
5018 // float4 result = as_float4(as_int4(xf*recip) + 0xb000);
5019 X = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, X, Y);
5020 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, X);
5021 Y = DAG.getConstant(0xb000, MVT::i32);
5022 Y = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Y, Y, Y, Y);
5023 X = DAG.getNode(ISD::ADD, dl, MVT::v4i32, X, Y);
5024 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, X);
5025 // Convert back to short.
5026 X = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, X);
5027 X = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, X);
5032 LowerSDIV_v4i16(SDValue N0, SDValue N1, DebugLoc dl, SelectionDAG &DAG) {
5034 // Convert to float.
5035 // float4 yf = vcvt_f32_s32(vmovl_s16(y));
5036 // float4 xf = vcvt_f32_s32(vmovl_s16(x));
5037 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N0);
5038 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N1);
5039 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
5040 N1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
5042 // Use reciprocal estimate and one refinement step.
5043 // float4 recip = vrecpeq_f32(yf);
5044 // recip *= vrecpsq_f32(yf, recip);
5045 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5046 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), N1);
5047 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5048 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
5050 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
5051 // Because short has a smaller range than ushort, we can actually get away
5052 // with only a single newton step. This requires that we use a weird bias
5053 // of 89, however (again, this has been exhaustively tested).
5054 // float4 result = as_float4(as_int4(xf*recip) + 0x89);
5055 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
5056 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
5057 N1 = DAG.getConstant(0x89, MVT::i32);
5058 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
5059 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
5060 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
5061 // Convert back to integer and return.
5062 // return vmovn_s32(vcvt_s32_f32(result));
5063 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
5064 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
5068 static SDValue LowerSDIV(SDValue Op, SelectionDAG &DAG) {
5069 EVT VT = Op.getValueType();
5070 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
5071 "unexpected type for custom-lowering ISD::SDIV");
5073 DebugLoc dl = Op.getDebugLoc();
5074 SDValue N0 = Op.getOperand(0);
5075 SDValue N1 = Op.getOperand(1);
5078 if (VT == MVT::v8i8) {
5079 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N0);
5080 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N1);
5082 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5083 DAG.getIntPtrConstant(4));
5084 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5085 DAG.getIntPtrConstant(4));
5086 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5087 DAG.getIntPtrConstant(0));
5088 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5089 DAG.getIntPtrConstant(0));
5091 N0 = LowerSDIV_v4i8(N0, N1, dl, DAG); // v4i16
5092 N2 = LowerSDIV_v4i8(N2, N3, dl, DAG); // v4i16
5094 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
5095 N0 = LowerCONCAT_VECTORS(N0, DAG);
5097 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v8i8, N0);
5100 return LowerSDIV_v4i16(N0, N1, dl, DAG);
5103 static SDValue LowerUDIV(SDValue Op, SelectionDAG &DAG) {
5104 EVT VT = Op.getValueType();
5105 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
5106 "unexpected type for custom-lowering ISD::UDIV");
5108 DebugLoc dl = Op.getDebugLoc();
5109 SDValue N0 = Op.getOperand(0);
5110 SDValue N1 = Op.getOperand(1);
5113 if (VT == MVT::v8i8) {
5114 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N0);
5115 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N1);
5117 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5118 DAG.getIntPtrConstant(4));
5119 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5120 DAG.getIntPtrConstant(4));
5121 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5122 DAG.getIntPtrConstant(0));
5123 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5124 DAG.getIntPtrConstant(0));
5126 N0 = LowerSDIV_v4i16(N0, N1, dl, DAG); // v4i16
5127 N2 = LowerSDIV_v4i16(N2, N3, dl, DAG); // v4i16
5129 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
5130 N0 = LowerCONCAT_VECTORS(N0, DAG);
5132 N0 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v8i8,
5133 DAG.getConstant(Intrinsic::arm_neon_vqmovnsu, MVT::i32),
5138 // v4i16 sdiv ... Convert to float.
5139 // float4 yf = vcvt_f32_s32(vmovl_u16(y));
5140 // float4 xf = vcvt_f32_s32(vmovl_u16(x));
5141 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N0);
5142 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N1);
5143 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
5144 SDValue BN1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
5146 // Use reciprocal estimate and two refinement steps.
5147 // float4 recip = vrecpeq_f32(yf);
5148 // recip *= vrecpsq_f32(yf, recip);
5149 // recip *= vrecpsq_f32(yf, recip);
5150 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5151 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), BN1);
5152 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5153 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
5155 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
5156 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5157 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
5159 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
5160 // Simply multiplying by the reciprocal estimate can leave us a few ulps
5161 // too low, so we add 2 ulps (exhaustive testing shows that this is enough,
5162 // and that it will never cause us to return an answer too large).
5163 // float4 result = as_float4(as_int4(xf*recip) + 2);
5164 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
5165 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
5166 N1 = DAG.getConstant(2, MVT::i32);
5167 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
5168 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
5169 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
5170 // Convert back to integer and return.
5171 // return vmovn_u32(vcvt_s32_f32(result));
5172 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
5173 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
5177 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
5178 EVT VT = Op.getNode()->getValueType(0);
5179 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
5182 bool ExtraOp = false;
5183 switch (Op.getOpcode()) {
5184 default: llvm_unreachable("Invalid code");
5185 case ISD::ADDC: Opc = ARMISD::ADDC; break;
5186 case ISD::ADDE: Opc = ARMISD::ADDE; ExtraOp = true; break;
5187 case ISD::SUBC: Opc = ARMISD::SUBC; break;
5188 case ISD::SUBE: Opc = ARMISD::SUBE; ExtraOp = true; break;
5192 return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
5194 return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
5195 Op.getOperand(1), Op.getOperand(2));
5198 static SDValue LowerAtomicLoadStore(SDValue Op, SelectionDAG &DAG) {
5199 // Monotonic load/store is legal for all targets
5200 if (cast<AtomicSDNode>(Op)->getOrdering() <= Monotonic)
5203 // Aquire/Release load/store is not legal for targets without a
5204 // dmb or equivalent available.
5210 ReplaceATOMIC_OP_64(SDNode *Node, SmallVectorImpl<SDValue>& Results,
5211 SelectionDAG &DAG, unsigned NewOp) {
5212 DebugLoc dl = Node->getDebugLoc();
5213 assert (Node->getValueType(0) == MVT::i64 &&
5214 "Only know how to expand i64 atomics");
5216 SmallVector<SDValue, 6> Ops;
5217 Ops.push_back(Node->getOperand(0)); // Chain
5218 Ops.push_back(Node->getOperand(1)); // Ptr
5220 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5221 Node->getOperand(2), DAG.getIntPtrConstant(0)));
5222 // High part of Val1
5223 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5224 Node->getOperand(2), DAG.getIntPtrConstant(1)));
5225 if (NewOp == ARMISD::ATOMCMPXCHG64_DAG) {
5226 // High part of Val1
5227 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5228 Node->getOperand(3), DAG.getIntPtrConstant(0)));
5229 // High part of Val2
5230 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5231 Node->getOperand(3), DAG.getIntPtrConstant(1)));
5233 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
5235 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops.data(), Ops.size(), MVT::i64,
5236 cast<MemSDNode>(Node)->getMemOperand());
5237 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1) };
5238 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
5239 Results.push_back(Result.getValue(2));
5242 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
5243 switch (Op.getOpcode()) {
5244 default: llvm_unreachable("Don't know how to custom lower this!");
5245 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
5246 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
5247 case ISD::GlobalAddress:
5248 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
5249 LowerGlobalAddressELF(Op, DAG);
5250 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
5251 case ISD::SELECT: return LowerSELECT(Op, DAG);
5252 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
5253 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
5254 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
5255 case ISD::VASTART: return LowerVASTART(Op, DAG);
5256 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG, Subtarget);
5257 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, DAG, Subtarget);
5258 case ISD::PREFETCH: return LowerPREFETCH(Op, DAG, Subtarget);
5259 case ISD::SINT_TO_FP:
5260 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
5261 case ISD::FP_TO_SINT:
5262 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
5263 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
5264 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
5265 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
5266 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
5267 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
5268 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
5269 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
5271 case ISD::BITCAST: return ExpandBITCAST(Op.getNode(), DAG);
5274 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
5275 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
5276 case ISD::SRL_PARTS:
5277 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
5278 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
5279 case ISD::SETCC: return LowerVSETCC(Op, DAG);
5280 case ISD::ConstantFP: return LowerConstantFP(Op, DAG, Subtarget);
5281 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG, Subtarget);
5282 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
5283 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
5284 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
5285 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
5286 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
5287 case ISD::MUL: return LowerMUL(Op, DAG);
5288 case ISD::SDIV: return LowerSDIV(Op, DAG);
5289 case ISD::UDIV: return LowerUDIV(Op, DAG);
5293 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
5294 case ISD::ATOMIC_LOAD:
5295 case ISD::ATOMIC_STORE: return LowerAtomicLoadStore(Op, DAG);
5299 /// ReplaceNodeResults - Replace the results of node with an illegal result
5300 /// type with new values built out of custom code.
5301 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
5302 SmallVectorImpl<SDValue>&Results,
5303 SelectionDAG &DAG) const {
5305 switch (N->getOpcode()) {
5307 llvm_unreachable("Don't know how to custom expand this!");
5309 Res = ExpandBITCAST(N, DAG);
5313 Res = Expand64BitShift(N, DAG, Subtarget);
5315 case ISD::ATOMIC_LOAD_ADD:
5316 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMADD64_DAG);
5318 case ISD::ATOMIC_LOAD_AND:
5319 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMAND64_DAG);
5321 case ISD::ATOMIC_LOAD_NAND:
5322 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMNAND64_DAG);
5324 case ISD::ATOMIC_LOAD_OR:
5325 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMOR64_DAG);
5327 case ISD::ATOMIC_LOAD_SUB:
5328 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMSUB64_DAG);
5330 case ISD::ATOMIC_LOAD_XOR:
5331 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMXOR64_DAG);
5333 case ISD::ATOMIC_SWAP:
5334 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMSWAP64_DAG);
5336 case ISD::ATOMIC_CMP_SWAP:
5337 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMCMPXCHG64_DAG);
5341 Results.push_back(Res);
5344 //===----------------------------------------------------------------------===//
5345 // ARM Scheduler Hooks
5346 //===----------------------------------------------------------------------===//
5349 ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
5350 MachineBasicBlock *BB,
5351 unsigned Size) const {
5352 unsigned dest = MI->getOperand(0).getReg();
5353 unsigned ptr = MI->getOperand(1).getReg();
5354 unsigned oldval = MI->getOperand(2).getReg();
5355 unsigned newval = MI->getOperand(3).getReg();
5356 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5357 DebugLoc dl = MI->getDebugLoc();
5358 bool isThumb2 = Subtarget->isThumb2();
5360 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5361 unsigned scratch = MRI.createVirtualRegister(isThumb2 ?
5362 (const TargetRegisterClass*)&ARM::rGPRRegClass :
5363 (const TargetRegisterClass*)&ARM::GPRRegClass);
5366 MRI.constrainRegClass(dest, &ARM::rGPRRegClass);
5367 MRI.constrainRegClass(oldval, &ARM::rGPRRegClass);
5368 MRI.constrainRegClass(newval, &ARM::rGPRRegClass);
5371 unsigned ldrOpc, strOpc;
5373 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
5375 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
5376 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
5379 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
5380 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
5383 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
5384 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
5388 MachineFunction *MF = BB->getParent();
5389 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5390 MachineFunction::iterator It = BB;
5391 ++It; // insert the new blocks after the current block
5393 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
5394 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
5395 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5396 MF->insert(It, loop1MBB);
5397 MF->insert(It, loop2MBB);
5398 MF->insert(It, exitMBB);
5400 // Transfer the remainder of BB and its successor edges to exitMBB.
5401 exitMBB->splice(exitMBB->begin(), BB,
5402 llvm::next(MachineBasicBlock::iterator(MI)),
5404 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5408 // fallthrough --> loop1MBB
5409 BB->addSuccessor(loop1MBB);
5412 // ldrex dest, [ptr]
5416 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
5417 if (ldrOpc == ARM::t2LDREX)
5419 AddDefaultPred(MIB);
5420 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
5421 .addReg(dest).addReg(oldval));
5422 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5423 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5424 BB->addSuccessor(loop2MBB);
5425 BB->addSuccessor(exitMBB);
5428 // strex scratch, newval, [ptr]
5432 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval).addReg(ptr);
5433 if (strOpc == ARM::t2STREX)
5435 AddDefaultPred(MIB);
5436 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5437 .addReg(scratch).addImm(0));
5438 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5439 .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5440 BB->addSuccessor(loop1MBB);
5441 BB->addSuccessor(exitMBB);
5447 MI->eraseFromParent(); // The instruction is gone now.
5453 ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
5454 unsigned Size, unsigned BinOpcode) const {
5455 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
5456 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5458 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5459 MachineFunction *MF = BB->getParent();
5460 MachineFunction::iterator It = BB;
5463 unsigned dest = MI->getOperand(0).getReg();
5464 unsigned ptr = MI->getOperand(1).getReg();
5465 unsigned incr = MI->getOperand(2).getReg();
5466 DebugLoc dl = MI->getDebugLoc();
5467 bool isThumb2 = Subtarget->isThumb2();
5469 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5471 MRI.constrainRegClass(dest, &ARM::rGPRRegClass);
5472 MRI.constrainRegClass(ptr, &ARM::rGPRRegClass);
5475 unsigned ldrOpc, strOpc;
5477 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
5479 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
5480 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
5483 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
5484 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
5487 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
5488 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
5492 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5493 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5494 MF->insert(It, loopMBB);
5495 MF->insert(It, exitMBB);
5497 // Transfer the remainder of BB and its successor edges to exitMBB.
5498 exitMBB->splice(exitMBB->begin(), BB,
5499 llvm::next(MachineBasicBlock::iterator(MI)),
5501 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5503 const TargetRegisterClass *TRC = isThumb2 ?
5504 (const TargetRegisterClass*)&ARM::rGPRRegClass :
5505 (const TargetRegisterClass*)&ARM::GPRRegClass;
5506 unsigned scratch = MRI.createVirtualRegister(TRC);
5507 unsigned scratch2 = (!BinOpcode) ? incr : MRI.createVirtualRegister(TRC);
5511 // fallthrough --> loopMBB
5512 BB->addSuccessor(loopMBB);
5516 // <binop> scratch2, dest, incr
5517 // strex scratch, scratch2, ptr
5520 // fallthrough --> exitMBB
5522 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
5523 if (ldrOpc == ARM::t2LDREX)
5525 AddDefaultPred(MIB);
5527 // operand order needs to go the other way for NAND
5528 if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr)
5529 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
5530 addReg(incr).addReg(dest)).addReg(0);
5532 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
5533 addReg(dest).addReg(incr)).addReg(0);
5536 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2).addReg(ptr);
5537 if (strOpc == ARM::t2STREX)
5539 AddDefaultPred(MIB);
5540 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5541 .addReg(scratch).addImm(0));
5542 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5543 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5545 BB->addSuccessor(loopMBB);
5546 BB->addSuccessor(exitMBB);
5552 MI->eraseFromParent(); // The instruction is gone now.
5558 ARMTargetLowering::EmitAtomicBinaryMinMax(MachineInstr *MI,
5559 MachineBasicBlock *BB,
5562 ARMCC::CondCodes Cond) const {
5563 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5565 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5566 MachineFunction *MF = BB->getParent();
5567 MachineFunction::iterator It = BB;
5570 unsigned dest = MI->getOperand(0).getReg();
5571 unsigned ptr = MI->getOperand(1).getReg();
5572 unsigned incr = MI->getOperand(2).getReg();
5573 unsigned oldval = dest;
5574 DebugLoc dl = MI->getDebugLoc();
5575 bool isThumb2 = Subtarget->isThumb2();
5577 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5579 MRI.constrainRegClass(dest, &ARM::rGPRRegClass);
5580 MRI.constrainRegClass(ptr, &ARM::rGPRRegClass);
5583 unsigned ldrOpc, strOpc, extendOpc;
5585 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
5587 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
5588 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
5589 extendOpc = isThumb2 ? ARM::t2SXTB : ARM::SXTB;
5592 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
5593 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
5594 extendOpc = isThumb2 ? ARM::t2SXTH : ARM::SXTH;
5597 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
5598 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
5603 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5604 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5605 MF->insert(It, loopMBB);
5606 MF->insert(It, exitMBB);
5608 // Transfer the remainder of BB and its successor edges to exitMBB.
5609 exitMBB->splice(exitMBB->begin(), BB,
5610 llvm::next(MachineBasicBlock::iterator(MI)),
5612 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5614 const TargetRegisterClass *TRC = isThumb2 ?
5615 (const TargetRegisterClass*)&ARM::rGPRRegClass :
5616 (const TargetRegisterClass*)&ARM::GPRRegClass;
5617 unsigned scratch = MRI.createVirtualRegister(TRC);
5618 unsigned scratch2 = MRI.createVirtualRegister(TRC);
5622 // fallthrough --> loopMBB
5623 BB->addSuccessor(loopMBB);
5627 // (sign extend dest, if required)
5629 // cmov.cond scratch2, incr, dest
5630 // strex scratch, scratch2, ptr
5633 // fallthrough --> exitMBB
5635 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
5636 if (ldrOpc == ARM::t2LDREX)
5638 AddDefaultPred(MIB);
5640 // Sign extend the value, if necessary.
5641 if (signExtend && extendOpc) {
5642 oldval = MRI.createVirtualRegister(&ARM::GPRRegClass);
5643 AddDefaultPred(BuildMI(BB, dl, TII->get(extendOpc), oldval)
5648 // Build compare and cmov instructions.
5649 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
5650 .addReg(oldval).addReg(incr));
5651 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2MOVCCr : ARM::MOVCCr), scratch2)
5652 .addReg(incr).addReg(oldval).addImm(Cond).addReg(ARM::CPSR);
5654 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2).addReg(ptr);
5655 if (strOpc == ARM::t2STREX)
5657 AddDefaultPred(MIB);
5658 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5659 .addReg(scratch).addImm(0));
5660 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5661 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5663 BB->addSuccessor(loopMBB);
5664 BB->addSuccessor(exitMBB);
5670 MI->eraseFromParent(); // The instruction is gone now.
5676 ARMTargetLowering::EmitAtomicBinary64(MachineInstr *MI, MachineBasicBlock *BB,
5677 unsigned Op1, unsigned Op2,
5678 bool NeedsCarry, bool IsCmpxchg) const {
5679 // This also handles ATOMIC_SWAP, indicated by Op1==0.
5680 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5682 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5683 MachineFunction *MF = BB->getParent();
5684 MachineFunction::iterator It = BB;
5687 unsigned destlo = MI->getOperand(0).getReg();
5688 unsigned desthi = MI->getOperand(1).getReg();
5689 unsigned ptr = MI->getOperand(2).getReg();
5690 unsigned vallo = MI->getOperand(3).getReg();
5691 unsigned valhi = MI->getOperand(4).getReg();
5692 DebugLoc dl = MI->getDebugLoc();
5693 bool isThumb2 = Subtarget->isThumb2();
5695 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5697 MRI.constrainRegClass(destlo, &ARM::rGPRRegClass);
5698 MRI.constrainRegClass(desthi, &ARM::rGPRRegClass);
5699 MRI.constrainRegClass(ptr, &ARM::rGPRRegClass);
5702 unsigned ldrOpc = isThumb2 ? ARM::t2LDREXD : ARM::LDREXD;
5703 unsigned strOpc = isThumb2 ? ARM::t2STREXD : ARM::STREXD;
5705 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5706 MachineBasicBlock *contBB = 0, *cont2BB = 0;
5708 contBB = MF->CreateMachineBasicBlock(LLVM_BB);
5709 cont2BB = MF->CreateMachineBasicBlock(LLVM_BB);
5711 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5712 MF->insert(It, loopMBB);
5714 MF->insert(It, contBB);
5715 MF->insert(It, cont2BB);
5717 MF->insert(It, exitMBB);
5719 // Transfer the remainder of BB and its successor edges to exitMBB.
5720 exitMBB->splice(exitMBB->begin(), BB,
5721 llvm::next(MachineBasicBlock::iterator(MI)),
5723 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5725 const TargetRegisterClass *TRC = isThumb2 ?
5726 (const TargetRegisterClass*)&ARM::tGPRRegClass :
5727 (const TargetRegisterClass*)&ARM::GPRRegClass;
5728 unsigned storesuccess = MRI.createVirtualRegister(TRC);
5732 // fallthrough --> loopMBB
5733 BB->addSuccessor(loopMBB);
5736 // ldrexd r2, r3, ptr
5737 // <binopa> r0, r2, incr
5738 // <binopb> r1, r3, incr
5739 // strexd storesuccess, r0, r1, ptr
5740 // cmp storesuccess, #0
5742 // fallthrough --> exitMBB
5744 // Note that the registers are explicitly specified because there is not any
5745 // way to force the register allocator to allocate a register pair.
5747 // FIXME: The hardcoded registers are not necessary for Thumb2, but we
5748 // need to properly enforce the restriction that the two output registers
5749 // for ldrexd must be different.
5752 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc))
5753 .addReg(ARM::R2, RegState::Define)
5754 .addReg(ARM::R3, RegState::Define).addReg(ptr));
5755 // Copy r2/r3 into dest. (This copy will normally be coalesced.)
5756 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), destlo).addReg(ARM::R2);
5757 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), desthi).addReg(ARM::R3);
5761 for (unsigned i = 0; i < 2; i++) {
5762 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr :
5764 .addReg(i == 0 ? destlo : desthi)
5765 .addReg(i == 0 ? vallo : valhi));
5766 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5767 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5768 BB->addSuccessor(exitMBB);
5769 BB->addSuccessor(i == 0 ? contBB : cont2BB);
5770 BB = (i == 0 ? contBB : cont2BB);
5773 // Copy to physregs for strexd
5774 unsigned setlo = MI->getOperand(5).getReg();
5775 unsigned sethi = MI->getOperand(6).getReg();
5776 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R0).addReg(setlo);
5777 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R1).addReg(sethi);
5779 // Perform binary operation
5780 AddDefaultPred(BuildMI(BB, dl, TII->get(Op1), ARM::R0)
5781 .addReg(destlo).addReg(vallo))
5782 .addReg(NeedsCarry ? ARM::CPSR : 0, getDefRegState(NeedsCarry));
5783 AddDefaultPred(BuildMI(BB, dl, TII->get(Op2), ARM::R1)
5784 .addReg(desthi).addReg(valhi)).addReg(0);
5786 // Copy to physregs for strexd
5787 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R0).addReg(vallo);
5788 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R1).addReg(valhi);
5792 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), storesuccess)
5793 .addReg(ARM::R0).addReg(ARM::R1).addReg(ptr));
5795 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5796 .addReg(storesuccess).addImm(0));
5797 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5798 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5800 BB->addSuccessor(loopMBB);
5801 BB->addSuccessor(exitMBB);
5807 MI->eraseFromParent(); // The instruction is gone now.
5812 /// SetupEntryBlockForSjLj - Insert code into the entry block that creates and
5813 /// registers the function context.
5814 void ARMTargetLowering::
5815 SetupEntryBlockForSjLj(MachineInstr *MI, MachineBasicBlock *MBB,
5816 MachineBasicBlock *DispatchBB, int FI) const {
5817 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5818 DebugLoc dl = MI->getDebugLoc();
5819 MachineFunction *MF = MBB->getParent();
5820 MachineRegisterInfo *MRI = &MF->getRegInfo();
5821 MachineConstantPool *MCP = MF->getConstantPool();
5822 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
5823 const Function *F = MF->getFunction();
5825 bool isThumb = Subtarget->isThumb();
5826 bool isThumb2 = Subtarget->isThumb2();
5828 unsigned PCLabelId = AFI->createPICLabelUId();
5829 unsigned PCAdj = (isThumb || isThumb2) ? 4 : 8;
5830 ARMConstantPoolValue *CPV =
5831 ARMConstantPoolMBB::Create(F->getContext(), DispatchBB, PCLabelId, PCAdj);
5832 unsigned CPI = MCP->getConstantPoolIndex(CPV, 4);
5834 const TargetRegisterClass *TRC = isThumb ?
5835 (const TargetRegisterClass*)&ARM::tGPRRegClass :
5836 (const TargetRegisterClass*)&ARM::GPRRegClass;
5838 // Grab constant pool and fixed stack memory operands.
5839 MachineMemOperand *CPMMO =
5840 MF->getMachineMemOperand(MachinePointerInfo::getConstantPool(),
5841 MachineMemOperand::MOLoad, 4, 4);
5843 MachineMemOperand *FIMMOSt =
5844 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
5845 MachineMemOperand::MOStore, 4, 4);
5847 // Load the address of the dispatch MBB into the jump buffer.
5849 // Incoming value: jbuf
5850 // ldr.n r5, LCPI1_1
5853 // str r5, [$jbuf, #+4] ; &jbuf[1]
5854 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5855 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2LDRpci), NewVReg1)
5856 .addConstantPoolIndex(CPI)
5857 .addMemOperand(CPMMO));
5858 // Set the low bit because of thumb mode.
5859 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
5861 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2ORRri), NewVReg2)
5862 .addReg(NewVReg1, RegState::Kill)
5864 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
5865 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg3)
5866 .addReg(NewVReg2, RegState::Kill)
5868 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2STRi12))
5869 .addReg(NewVReg3, RegState::Kill)
5871 .addImm(36) // &jbuf[1] :: pc
5872 .addMemOperand(FIMMOSt));
5873 } else if (isThumb) {
5874 // Incoming value: jbuf
5875 // ldr.n r1, LCPI1_4
5879 // add r2, $jbuf, #+4 ; &jbuf[1]
5881 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5882 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tLDRpci), NewVReg1)
5883 .addConstantPoolIndex(CPI)
5884 .addMemOperand(CPMMO));
5885 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
5886 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg2)
5887 .addReg(NewVReg1, RegState::Kill)
5889 // Set the low bit because of thumb mode.
5890 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
5891 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tMOVi8), NewVReg3)
5892 .addReg(ARM::CPSR, RegState::Define)
5894 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
5895 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tORR), NewVReg4)
5896 .addReg(ARM::CPSR, RegState::Define)
5897 .addReg(NewVReg2, RegState::Kill)
5898 .addReg(NewVReg3, RegState::Kill));
5899 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
5900 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tADDrSPi), NewVReg5)
5902 .addImm(36)); // &jbuf[1] :: pc
5903 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tSTRi))
5904 .addReg(NewVReg4, RegState::Kill)
5905 .addReg(NewVReg5, RegState::Kill)
5907 .addMemOperand(FIMMOSt));
5909 // Incoming value: jbuf
5912 // str r1, [$jbuf, #+4] ; &jbuf[1]
5913 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5914 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::LDRi12), NewVReg1)
5915 .addConstantPoolIndex(CPI)
5917 .addMemOperand(CPMMO));
5918 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
5919 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::PICADD), NewVReg2)
5920 .addReg(NewVReg1, RegState::Kill)
5921 .addImm(PCLabelId));
5922 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::STRi12))
5923 .addReg(NewVReg2, RegState::Kill)
5925 .addImm(36) // &jbuf[1] :: pc
5926 .addMemOperand(FIMMOSt));
5930 MachineBasicBlock *ARMTargetLowering::
5931 EmitSjLjDispatchBlock(MachineInstr *MI, MachineBasicBlock *MBB) const {
5932 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5933 DebugLoc dl = MI->getDebugLoc();
5934 MachineFunction *MF = MBB->getParent();
5935 MachineRegisterInfo *MRI = &MF->getRegInfo();
5936 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
5937 MachineFrameInfo *MFI = MF->getFrameInfo();
5938 int FI = MFI->getFunctionContextIndex();
5940 const TargetRegisterClass *TRC = Subtarget->isThumb() ?
5941 (const TargetRegisterClass*)&ARM::tGPRRegClass :
5942 (const TargetRegisterClass*)&ARM::GPRnopcRegClass;
5944 // Get a mapping of the call site numbers to all of the landing pads they're
5946 DenseMap<unsigned, SmallVector<MachineBasicBlock*, 2> > CallSiteNumToLPad;
5947 unsigned MaxCSNum = 0;
5948 MachineModuleInfo &MMI = MF->getMMI();
5949 for (MachineFunction::iterator BB = MF->begin(), E = MF->end(); BB != E;
5951 if (!BB->isLandingPad()) continue;
5953 // FIXME: We should assert that the EH_LABEL is the first MI in the landing
5955 for (MachineBasicBlock::iterator
5956 II = BB->begin(), IE = BB->end(); II != IE; ++II) {
5957 if (!II->isEHLabel()) continue;
5959 MCSymbol *Sym = II->getOperand(0).getMCSymbol();
5960 if (!MMI.hasCallSiteLandingPad(Sym)) continue;
5962 SmallVectorImpl<unsigned> &CallSiteIdxs = MMI.getCallSiteLandingPad(Sym);
5963 for (SmallVectorImpl<unsigned>::iterator
5964 CSI = CallSiteIdxs.begin(), CSE = CallSiteIdxs.end();
5965 CSI != CSE; ++CSI) {
5966 CallSiteNumToLPad[*CSI].push_back(BB);
5967 MaxCSNum = std::max(MaxCSNum, *CSI);
5973 // Get an ordered list of the machine basic blocks for the jump table.
5974 std::vector<MachineBasicBlock*> LPadList;
5975 SmallPtrSet<MachineBasicBlock*, 64> InvokeBBs;
5976 LPadList.reserve(CallSiteNumToLPad.size());
5977 for (unsigned I = 1; I <= MaxCSNum; ++I) {
5978 SmallVectorImpl<MachineBasicBlock*> &MBBList = CallSiteNumToLPad[I];
5979 for (SmallVectorImpl<MachineBasicBlock*>::iterator
5980 II = MBBList.begin(), IE = MBBList.end(); II != IE; ++II) {
5981 LPadList.push_back(*II);
5982 InvokeBBs.insert((*II)->pred_begin(), (*II)->pred_end());
5986 assert(!LPadList.empty() &&
5987 "No landing pad destinations for the dispatch jump table!");
5989 // Create the jump table and associated information.
5990 MachineJumpTableInfo *JTI =
5991 MF->getOrCreateJumpTableInfo(MachineJumpTableInfo::EK_Inline);
5992 unsigned MJTI = JTI->createJumpTableIndex(LPadList);
5993 unsigned UId = AFI->createJumpTableUId();
5995 // Create the MBBs for the dispatch code.
5997 // Shove the dispatch's address into the return slot in the function context.
5998 MachineBasicBlock *DispatchBB = MF->CreateMachineBasicBlock();
5999 DispatchBB->setIsLandingPad();
6001 MachineBasicBlock *TrapBB = MF->CreateMachineBasicBlock();
6002 BuildMI(TrapBB, dl, TII->get(Subtarget->isThumb() ? ARM::tTRAP : ARM::TRAP));
6003 DispatchBB->addSuccessor(TrapBB);
6005 MachineBasicBlock *DispContBB = MF->CreateMachineBasicBlock();
6006 DispatchBB->addSuccessor(DispContBB);
6009 MF->insert(MF->end(), DispatchBB);
6010 MF->insert(MF->end(), DispContBB);
6011 MF->insert(MF->end(), TrapBB);
6013 // Insert code into the entry block that creates and registers the function
6015 SetupEntryBlockForSjLj(MI, MBB, DispatchBB, FI);
6017 MachineMemOperand *FIMMOLd =
6018 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
6019 MachineMemOperand::MOLoad |
6020 MachineMemOperand::MOVolatile, 4, 4);
6022 if (AFI->isThumb1OnlyFunction())
6023 BuildMI(DispatchBB, dl, TII->get(ARM::tInt_eh_sjlj_dispatchsetup));
6024 else if (!Subtarget->hasVFP2())
6025 BuildMI(DispatchBB, dl, TII->get(ARM::Int_eh_sjlj_dispatchsetup_nofp));
6027 BuildMI(DispatchBB, dl, TII->get(ARM::Int_eh_sjlj_dispatchsetup));
6029 unsigned NumLPads = LPadList.size();
6030 if (Subtarget->isThumb2()) {
6031 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6032 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2LDRi12), NewVReg1)
6035 .addMemOperand(FIMMOLd));
6037 if (NumLPads < 256) {
6038 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPri))
6040 .addImm(LPadList.size()));
6042 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6043 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVi16), VReg1)
6044 .addImm(NumLPads & 0xFFFF));
6046 unsigned VReg2 = VReg1;
6047 if ((NumLPads & 0xFFFF0000) != 0) {
6048 VReg2 = MRI->createVirtualRegister(TRC);
6049 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVTi16), VReg2)
6051 .addImm(NumLPads >> 16));
6054 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPrr))
6059 BuildMI(DispatchBB, dl, TII->get(ARM::t2Bcc))
6064 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6065 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::t2LEApcrelJT),NewVReg3)
6066 .addJumpTableIndex(MJTI)
6069 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6072 BuildMI(DispContBB, dl, TII->get(ARM::t2ADDrs), NewVReg4)
6073 .addReg(NewVReg3, RegState::Kill)
6075 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
6077 BuildMI(DispContBB, dl, TII->get(ARM::t2BR_JT))
6078 .addReg(NewVReg4, RegState::Kill)
6080 .addJumpTableIndex(MJTI)
6082 } else if (Subtarget->isThumb()) {
6083 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6084 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRspi), NewVReg1)
6087 .addMemOperand(FIMMOLd));
6089 if (NumLPads < 256) {
6090 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPi8))
6094 MachineConstantPool *ConstantPool = MF->getConstantPool();
6095 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
6096 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
6098 // MachineConstantPool wants an explicit alignment.
6099 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
6101 Align = getDataLayout()->getTypeAllocSize(C->getType());
6102 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6104 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6105 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRpci))
6106 .addReg(VReg1, RegState::Define)
6107 .addConstantPoolIndex(Idx));
6108 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPr))
6113 BuildMI(DispatchBB, dl, TII->get(ARM::tBcc))
6118 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6119 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLSLri), NewVReg2)
6120 .addReg(ARM::CPSR, RegState::Define)
6124 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6125 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLEApcrelJT), NewVReg3)
6126 .addJumpTableIndex(MJTI)
6129 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6130 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg4)
6131 .addReg(ARM::CPSR, RegState::Define)
6132 .addReg(NewVReg2, RegState::Kill)
6135 MachineMemOperand *JTMMOLd =
6136 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
6137 MachineMemOperand::MOLoad, 4, 4);
6139 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6140 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLDRi), NewVReg5)
6141 .addReg(NewVReg4, RegState::Kill)
6143 .addMemOperand(JTMMOLd));
6145 unsigned NewVReg6 = MRI->createVirtualRegister(TRC);
6146 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg6)
6147 .addReg(ARM::CPSR, RegState::Define)
6148 .addReg(NewVReg5, RegState::Kill)
6151 BuildMI(DispContBB, dl, TII->get(ARM::tBR_JTr))
6152 .addReg(NewVReg6, RegState::Kill)
6153 .addJumpTableIndex(MJTI)
6156 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6157 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRi12), NewVReg1)
6160 .addMemOperand(FIMMOLd));
6162 if (NumLPads < 256) {
6163 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPri))
6166 } else if (Subtarget->hasV6T2Ops() && isUInt<16>(NumLPads)) {
6167 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6168 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVi16), VReg1)
6169 .addImm(NumLPads & 0xFFFF));
6171 unsigned VReg2 = VReg1;
6172 if ((NumLPads & 0xFFFF0000) != 0) {
6173 VReg2 = MRI->createVirtualRegister(TRC);
6174 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVTi16), VReg2)
6176 .addImm(NumLPads >> 16));
6179 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
6183 MachineConstantPool *ConstantPool = MF->getConstantPool();
6184 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
6185 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
6187 // MachineConstantPool wants an explicit alignment.
6188 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
6190 Align = getDataLayout()->getTypeAllocSize(C->getType());
6191 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6193 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6194 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRcp))
6195 .addReg(VReg1, RegState::Define)
6196 .addConstantPoolIndex(Idx)
6198 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
6200 .addReg(VReg1, RegState::Kill));
6203 BuildMI(DispatchBB, dl, TII->get(ARM::Bcc))
6208 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6210 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::MOVsi), NewVReg3)
6212 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
6213 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6214 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::LEApcrelJT), NewVReg4)
6215 .addJumpTableIndex(MJTI)
6218 MachineMemOperand *JTMMOLd =
6219 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
6220 MachineMemOperand::MOLoad, 4, 4);
6221 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6223 BuildMI(DispContBB, dl, TII->get(ARM::LDRrs), NewVReg5)
6224 .addReg(NewVReg3, RegState::Kill)
6227 .addMemOperand(JTMMOLd));
6229 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTadd))
6230 .addReg(NewVReg5, RegState::Kill)
6232 .addJumpTableIndex(MJTI)
6236 // Add the jump table entries as successors to the MBB.
6237 SmallPtrSet<MachineBasicBlock*, 8> SeenMBBs;
6238 for (std::vector<MachineBasicBlock*>::iterator
6239 I = LPadList.begin(), E = LPadList.end(); I != E; ++I) {
6240 MachineBasicBlock *CurMBB = *I;
6241 if (SeenMBBs.insert(CurMBB))
6242 DispContBB->addSuccessor(CurMBB);
6245 // N.B. the order the invoke BBs are processed in doesn't matter here.
6246 const ARMBaseInstrInfo *AII = static_cast<const ARMBaseInstrInfo*>(TII);
6247 const ARMBaseRegisterInfo &RI = AII->getRegisterInfo();
6248 const uint16_t *SavedRegs = RI.getCalleeSavedRegs(MF);
6249 SmallVector<MachineBasicBlock*, 64> MBBLPads;
6250 for (SmallPtrSet<MachineBasicBlock*, 64>::iterator
6251 I = InvokeBBs.begin(), E = InvokeBBs.end(); I != E; ++I) {
6252 MachineBasicBlock *BB = *I;
6254 // Remove the landing pad successor from the invoke block and replace it
6255 // with the new dispatch block.
6256 SmallVector<MachineBasicBlock*, 4> Successors(BB->succ_begin(),
6258 while (!Successors.empty()) {
6259 MachineBasicBlock *SMBB = Successors.pop_back_val();
6260 if (SMBB->isLandingPad()) {
6261 BB->removeSuccessor(SMBB);
6262 MBBLPads.push_back(SMBB);
6266 BB->addSuccessor(DispatchBB);
6268 // Find the invoke call and mark all of the callee-saved registers as
6269 // 'implicit defined' so that they're spilled. This prevents code from
6270 // moving instructions to before the EH block, where they will never be
6272 for (MachineBasicBlock::reverse_iterator
6273 II = BB->rbegin(), IE = BB->rend(); II != IE; ++II) {
6274 if (!II->isCall()) continue;
6276 DenseMap<unsigned, bool> DefRegs;
6277 for (MachineInstr::mop_iterator
6278 OI = II->operands_begin(), OE = II->operands_end();
6280 if (!OI->isReg()) continue;
6281 DefRegs[OI->getReg()] = true;
6284 MachineInstrBuilder MIB(&*II);
6286 for (unsigned i = 0; SavedRegs[i] != 0; ++i) {
6287 unsigned Reg = SavedRegs[i];
6288 if (Subtarget->isThumb2() &&
6289 !ARM::tGPRRegClass.contains(Reg) &&
6290 !ARM::hGPRRegClass.contains(Reg))
6292 if (Subtarget->isThumb1Only() && !ARM::tGPRRegClass.contains(Reg))
6294 if (!Subtarget->isThumb() && !ARM::GPRRegClass.contains(Reg))
6297 MIB.addReg(Reg, RegState::ImplicitDefine | RegState::Dead);
6304 // Mark all former landing pads as non-landing pads. The dispatch is the only
6306 for (SmallVectorImpl<MachineBasicBlock*>::iterator
6307 I = MBBLPads.begin(), E = MBBLPads.end(); I != E; ++I)
6308 (*I)->setIsLandingPad(false);
6310 // The instruction is gone now.
6311 MI->eraseFromParent();
6317 MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) {
6318 for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
6319 E = MBB->succ_end(); I != E; ++I)
6322 llvm_unreachable("Expecting a BB with two successors!");
6325 MachineBasicBlock *ARMTargetLowering::
6326 EmitStructByval(MachineInstr *MI, MachineBasicBlock *BB) const {
6327 // This pseudo instruction has 3 operands: dst, src, size
6328 // We expand it to a loop if size > Subtarget->getMaxInlineSizeThreshold().
6329 // Otherwise, we will generate unrolled scalar copies.
6330 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6331 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6332 MachineFunction::iterator It = BB;
6335 unsigned dest = MI->getOperand(0).getReg();
6336 unsigned src = MI->getOperand(1).getReg();
6337 unsigned SizeVal = MI->getOperand(2).getImm();
6338 unsigned Align = MI->getOperand(3).getImm();
6339 DebugLoc dl = MI->getDebugLoc();
6341 bool isThumb2 = Subtarget->isThumb2();
6342 MachineFunction *MF = BB->getParent();
6343 MachineRegisterInfo &MRI = MF->getRegInfo();
6344 unsigned ldrOpc, strOpc, UnitSize = 0;
6346 const TargetRegisterClass *TRC = isThumb2 ?
6347 (const TargetRegisterClass*)&ARM::tGPRRegClass :
6348 (const TargetRegisterClass*)&ARM::GPRRegClass;
6349 const TargetRegisterClass *TRC_Vec = 0;
6352 ldrOpc = isThumb2 ? ARM::t2LDRB_POST : ARM::LDRB_POST_IMM;
6353 strOpc = isThumb2 ? ARM::t2STRB_POST : ARM::STRB_POST_IMM;
6355 } else if (Align & 2) {
6356 ldrOpc = isThumb2 ? ARM::t2LDRH_POST : ARM::LDRH_POST;
6357 strOpc = isThumb2 ? ARM::t2STRH_POST : ARM::STRH_POST;
6360 // Check whether we can use NEON instructions.
6361 if (!MF->getFunction()->getFnAttributes().
6362 hasAttribute(Attributes::NoImplicitFloat) &&
6363 Subtarget->hasNEON()) {
6364 if ((Align % 16 == 0) && SizeVal >= 16) {
6365 ldrOpc = ARM::VLD1q32wb_fixed;
6366 strOpc = ARM::VST1q32wb_fixed;
6368 TRC_Vec = (const TargetRegisterClass*)&ARM::DPairRegClass;
6370 else if ((Align % 8 == 0) && SizeVal >= 8) {
6371 ldrOpc = ARM::VLD1d32wb_fixed;
6372 strOpc = ARM::VST1d32wb_fixed;
6374 TRC_Vec = (const TargetRegisterClass*)&ARM::DPRRegClass;
6377 // Can't use NEON instructions.
6378 if (UnitSize == 0) {
6379 ldrOpc = isThumb2 ? ARM::t2LDR_POST : ARM::LDR_POST_IMM;
6380 strOpc = isThumb2 ? ARM::t2STR_POST : ARM::STR_POST_IMM;
6385 unsigned BytesLeft = SizeVal % UnitSize;
6386 unsigned LoopSize = SizeVal - BytesLeft;
6388 if (SizeVal <= Subtarget->getMaxInlineSizeThreshold()) {
6389 // Use LDR and STR to copy.
6390 // [scratch, srcOut] = LDR_POST(srcIn, UnitSize)
6391 // [destOut] = STR_POST(scratch, destIn, UnitSize)
6392 unsigned srcIn = src;
6393 unsigned destIn = dest;
6394 for (unsigned i = 0; i < LoopSize; i+=UnitSize) {
6395 unsigned scratch = MRI.createVirtualRegister(UnitSize >= 8 ? TRC_Vec:TRC);
6396 unsigned srcOut = MRI.createVirtualRegister(TRC);
6397 unsigned destOut = MRI.createVirtualRegister(TRC);
6398 if (UnitSize >= 8) {
6399 AddDefaultPred(BuildMI(*BB, MI, dl,
6400 TII->get(ldrOpc), scratch)
6401 .addReg(srcOut, RegState::Define).addReg(srcIn).addImm(0));
6403 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut)
6404 .addReg(destIn).addImm(0).addReg(scratch));
6405 } else if (isThumb2) {
6406 AddDefaultPred(BuildMI(*BB, MI, dl,
6407 TII->get(ldrOpc), scratch)
6408 .addReg(srcOut, RegState::Define).addReg(srcIn).addImm(UnitSize));
6410 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut)
6411 .addReg(scratch).addReg(destIn)
6414 AddDefaultPred(BuildMI(*BB, MI, dl,
6415 TII->get(ldrOpc), scratch)
6416 .addReg(srcOut, RegState::Define).addReg(srcIn).addReg(0)
6419 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut)
6420 .addReg(scratch).addReg(destIn)
6421 .addReg(0).addImm(UnitSize));
6427 // Handle the leftover bytes with LDRB and STRB.
6428 // [scratch, srcOut] = LDRB_POST(srcIn, 1)
6429 // [destOut] = STRB_POST(scratch, destIn, 1)
6430 ldrOpc = isThumb2 ? ARM::t2LDRB_POST : ARM::LDRB_POST_IMM;
6431 strOpc = isThumb2 ? ARM::t2STRB_POST : ARM::STRB_POST_IMM;
6432 for (unsigned i = 0; i < BytesLeft; i++) {
6433 unsigned scratch = MRI.createVirtualRegister(TRC);
6434 unsigned srcOut = MRI.createVirtualRegister(TRC);
6435 unsigned destOut = MRI.createVirtualRegister(TRC);
6437 AddDefaultPred(BuildMI(*BB, MI, dl,
6438 TII->get(ldrOpc),scratch)
6439 .addReg(srcOut, RegState::Define).addReg(srcIn).addImm(1));
6441 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut)
6442 .addReg(scratch).addReg(destIn)
6443 .addReg(0).addImm(1));
6445 AddDefaultPred(BuildMI(*BB, MI, dl,
6446 TII->get(ldrOpc),scratch)
6447 .addReg(srcOut, RegState::Define).addReg(srcIn)
6448 .addReg(0).addImm(1));
6450 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(strOpc), destOut)
6451 .addReg(scratch).addReg(destIn)
6452 .addReg(0).addImm(1));
6457 MI->eraseFromParent(); // The instruction is gone now.
6461 // Expand the pseudo op to a loop.
6464 // movw varEnd, # --> with thumb2
6466 // ldrcp varEnd, idx --> without thumb2
6467 // fallthrough --> loopMBB
6469 // PHI varPhi, varEnd, varLoop
6470 // PHI srcPhi, src, srcLoop
6471 // PHI destPhi, dst, destLoop
6472 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
6473 // [destLoop] = STR_POST(scratch, destPhi, UnitSize)
6474 // subs varLoop, varPhi, #UnitSize
6476 // fallthrough --> exitMBB
6478 // epilogue to handle left-over bytes
6479 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
6480 // [destOut] = STRB_POST(scratch, destLoop, 1)
6481 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
6482 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
6483 MF->insert(It, loopMBB);
6484 MF->insert(It, exitMBB);
6486 // Transfer the remainder of BB and its successor edges to exitMBB.
6487 exitMBB->splice(exitMBB->begin(), BB,
6488 llvm::next(MachineBasicBlock::iterator(MI)),
6490 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
6492 // Load an immediate to varEnd.
6493 unsigned varEnd = MRI.createVirtualRegister(TRC);
6495 unsigned VReg1 = varEnd;
6496 if ((LoopSize & 0xFFFF0000) != 0)
6497 VReg1 = MRI.createVirtualRegister(TRC);
6498 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2MOVi16), VReg1)
6499 .addImm(LoopSize & 0xFFFF));
6501 if ((LoopSize & 0xFFFF0000) != 0)
6502 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2MOVTi16), varEnd)
6504 .addImm(LoopSize >> 16));
6506 MachineConstantPool *ConstantPool = MF->getConstantPool();
6507 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
6508 const Constant *C = ConstantInt::get(Int32Ty, LoopSize);
6510 // MachineConstantPool wants an explicit alignment.
6511 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
6513 Align = getDataLayout()->getTypeAllocSize(C->getType());
6514 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6516 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::LDRcp))
6517 .addReg(varEnd, RegState::Define)
6518 .addConstantPoolIndex(Idx)
6521 BB->addSuccessor(loopMBB);
6523 // Generate the loop body:
6524 // varPhi = PHI(varLoop, varEnd)
6525 // srcPhi = PHI(srcLoop, src)
6526 // destPhi = PHI(destLoop, dst)
6527 MachineBasicBlock *entryBB = BB;
6529 unsigned varLoop = MRI.createVirtualRegister(TRC);
6530 unsigned varPhi = MRI.createVirtualRegister(TRC);
6531 unsigned srcLoop = MRI.createVirtualRegister(TRC);
6532 unsigned srcPhi = MRI.createVirtualRegister(TRC);
6533 unsigned destLoop = MRI.createVirtualRegister(TRC);
6534 unsigned destPhi = MRI.createVirtualRegister(TRC);
6536 BuildMI(*BB, BB->begin(), dl, TII->get(ARM::PHI), varPhi)
6537 .addReg(varLoop).addMBB(loopMBB)
6538 .addReg(varEnd).addMBB(entryBB);
6539 BuildMI(BB, dl, TII->get(ARM::PHI), srcPhi)
6540 .addReg(srcLoop).addMBB(loopMBB)
6541 .addReg(src).addMBB(entryBB);
6542 BuildMI(BB, dl, TII->get(ARM::PHI), destPhi)
6543 .addReg(destLoop).addMBB(loopMBB)
6544 .addReg(dest).addMBB(entryBB);
6546 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
6547 // [destLoop] = STR_POST(scratch, destPhi, UnitSiz)
6548 unsigned scratch = MRI.createVirtualRegister(UnitSize >= 8 ? TRC_Vec:TRC);
6549 if (UnitSize >= 8) {
6550 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), scratch)
6551 .addReg(srcLoop, RegState::Define).addReg(srcPhi).addImm(0));
6553 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), destLoop)
6554 .addReg(destPhi).addImm(0).addReg(scratch));
6555 } else if (isThumb2) {
6556 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), scratch)
6557 .addReg(srcLoop, RegState::Define).addReg(srcPhi).addImm(UnitSize));
6559 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), destLoop)
6560 .addReg(scratch).addReg(destPhi)
6563 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), scratch)
6564 .addReg(srcLoop, RegState::Define).addReg(srcPhi).addReg(0)
6567 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), destLoop)
6568 .addReg(scratch).addReg(destPhi)
6569 .addReg(0).addImm(UnitSize));
6572 // Decrement loop variable by UnitSize.
6573 MachineInstrBuilder MIB = BuildMI(BB, dl,
6574 TII->get(isThumb2 ? ARM::t2SUBri : ARM::SUBri), varLoop);
6575 AddDefaultCC(AddDefaultPred(MIB.addReg(varPhi).addImm(UnitSize)));
6576 MIB->getOperand(5).setReg(ARM::CPSR);
6577 MIB->getOperand(5).setIsDef(true);
6579 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6580 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
6582 // loopMBB can loop back to loopMBB or fall through to exitMBB.
6583 BB->addSuccessor(loopMBB);
6584 BB->addSuccessor(exitMBB);
6586 // Add epilogue to handle BytesLeft.
6588 MachineInstr *StartOfExit = exitMBB->begin();
6589 ldrOpc = isThumb2 ? ARM::t2LDRB_POST : ARM::LDRB_POST_IMM;
6590 strOpc = isThumb2 ? ARM::t2STRB_POST : ARM::STRB_POST_IMM;
6592 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
6593 // [destOut] = STRB_POST(scratch, destLoop, 1)
6594 unsigned srcIn = srcLoop;
6595 unsigned destIn = destLoop;
6596 for (unsigned i = 0; i < BytesLeft; i++) {
6597 unsigned scratch = MRI.createVirtualRegister(TRC);
6598 unsigned srcOut = MRI.createVirtualRegister(TRC);
6599 unsigned destOut = MRI.createVirtualRegister(TRC);
6601 AddDefaultPred(BuildMI(*BB, StartOfExit, dl,
6602 TII->get(ldrOpc),scratch)
6603 .addReg(srcOut, RegState::Define).addReg(srcIn).addImm(1));
6605 AddDefaultPred(BuildMI(*BB, StartOfExit, dl, TII->get(strOpc), destOut)
6606 .addReg(scratch).addReg(destIn)
6609 AddDefaultPred(BuildMI(*BB, StartOfExit, dl,
6610 TII->get(ldrOpc),scratch)
6611 .addReg(srcOut, RegState::Define).addReg(srcIn).addReg(0).addImm(1));
6613 AddDefaultPred(BuildMI(*BB, StartOfExit, dl, TII->get(strOpc), destOut)
6614 .addReg(scratch).addReg(destIn)
6615 .addReg(0).addImm(1));
6621 MI->eraseFromParent(); // The instruction is gone now.
6626 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
6627 MachineBasicBlock *BB) const {
6628 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6629 DebugLoc dl = MI->getDebugLoc();
6630 bool isThumb2 = Subtarget->isThumb2();
6631 switch (MI->getOpcode()) {
6634 llvm_unreachable("Unexpected instr type to insert");
6636 // The Thumb2 pre-indexed stores have the same MI operands, they just
6637 // define them differently in the .td files from the isel patterns, so
6638 // they need pseudos.
6639 case ARM::t2STR_preidx:
6640 MI->setDesc(TII->get(ARM::t2STR_PRE));
6642 case ARM::t2STRB_preidx:
6643 MI->setDesc(TII->get(ARM::t2STRB_PRE));
6645 case ARM::t2STRH_preidx:
6646 MI->setDesc(TII->get(ARM::t2STRH_PRE));
6649 case ARM::STRi_preidx:
6650 case ARM::STRBi_preidx: {
6651 unsigned NewOpc = MI->getOpcode() == ARM::STRi_preidx ?
6652 ARM::STR_PRE_IMM : ARM::STRB_PRE_IMM;
6653 // Decode the offset.
6654 unsigned Offset = MI->getOperand(4).getImm();
6655 bool isSub = ARM_AM::getAM2Op(Offset) == ARM_AM::sub;
6656 Offset = ARM_AM::getAM2Offset(Offset);
6660 MachineMemOperand *MMO = *MI->memoperands_begin();
6661 BuildMI(*BB, MI, dl, TII->get(NewOpc))
6662 .addOperand(MI->getOperand(0)) // Rn_wb
6663 .addOperand(MI->getOperand(1)) // Rt
6664 .addOperand(MI->getOperand(2)) // Rn
6665 .addImm(Offset) // offset (skip GPR==zero_reg)
6666 .addOperand(MI->getOperand(5)) // pred
6667 .addOperand(MI->getOperand(6))
6668 .addMemOperand(MMO);
6669 MI->eraseFromParent();
6672 case ARM::STRr_preidx:
6673 case ARM::STRBr_preidx:
6674 case ARM::STRH_preidx: {
6676 switch (MI->getOpcode()) {
6677 default: llvm_unreachable("unexpected opcode!");
6678 case ARM::STRr_preidx: NewOpc = ARM::STR_PRE_REG; break;
6679 case ARM::STRBr_preidx: NewOpc = ARM::STRB_PRE_REG; break;
6680 case ARM::STRH_preidx: NewOpc = ARM::STRH_PRE; break;
6682 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(NewOpc));
6683 for (unsigned i = 0; i < MI->getNumOperands(); ++i)
6684 MIB.addOperand(MI->getOperand(i));
6685 MI->eraseFromParent();
6688 case ARM::ATOMIC_LOAD_ADD_I8:
6689 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
6690 case ARM::ATOMIC_LOAD_ADD_I16:
6691 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
6692 case ARM::ATOMIC_LOAD_ADD_I32:
6693 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
6695 case ARM::ATOMIC_LOAD_AND_I8:
6696 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6697 case ARM::ATOMIC_LOAD_AND_I16:
6698 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6699 case ARM::ATOMIC_LOAD_AND_I32:
6700 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6702 case ARM::ATOMIC_LOAD_OR_I8:
6703 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6704 case ARM::ATOMIC_LOAD_OR_I16:
6705 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6706 case ARM::ATOMIC_LOAD_OR_I32:
6707 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6709 case ARM::ATOMIC_LOAD_XOR_I8:
6710 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6711 case ARM::ATOMIC_LOAD_XOR_I16:
6712 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6713 case ARM::ATOMIC_LOAD_XOR_I32:
6714 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6716 case ARM::ATOMIC_LOAD_NAND_I8:
6717 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
6718 case ARM::ATOMIC_LOAD_NAND_I16:
6719 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
6720 case ARM::ATOMIC_LOAD_NAND_I32:
6721 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
6723 case ARM::ATOMIC_LOAD_SUB_I8:
6724 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
6725 case ARM::ATOMIC_LOAD_SUB_I16:
6726 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
6727 case ARM::ATOMIC_LOAD_SUB_I32:
6728 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
6730 case ARM::ATOMIC_LOAD_MIN_I8:
6731 return EmitAtomicBinaryMinMax(MI, BB, 1, true, ARMCC::LT);
6732 case ARM::ATOMIC_LOAD_MIN_I16:
6733 return EmitAtomicBinaryMinMax(MI, BB, 2, true, ARMCC::LT);
6734 case ARM::ATOMIC_LOAD_MIN_I32:
6735 return EmitAtomicBinaryMinMax(MI, BB, 4, true, ARMCC::LT);
6737 case ARM::ATOMIC_LOAD_MAX_I8:
6738 return EmitAtomicBinaryMinMax(MI, BB, 1, true, ARMCC::GT);
6739 case ARM::ATOMIC_LOAD_MAX_I16:
6740 return EmitAtomicBinaryMinMax(MI, BB, 2, true, ARMCC::GT);
6741 case ARM::ATOMIC_LOAD_MAX_I32:
6742 return EmitAtomicBinaryMinMax(MI, BB, 4, true, ARMCC::GT);
6744 case ARM::ATOMIC_LOAD_UMIN_I8:
6745 return EmitAtomicBinaryMinMax(MI, BB, 1, false, ARMCC::LO);
6746 case ARM::ATOMIC_LOAD_UMIN_I16:
6747 return EmitAtomicBinaryMinMax(MI, BB, 2, false, ARMCC::LO);
6748 case ARM::ATOMIC_LOAD_UMIN_I32:
6749 return EmitAtomicBinaryMinMax(MI, BB, 4, false, ARMCC::LO);
6751 case ARM::ATOMIC_LOAD_UMAX_I8:
6752 return EmitAtomicBinaryMinMax(MI, BB, 1, false, ARMCC::HI);
6753 case ARM::ATOMIC_LOAD_UMAX_I16:
6754 return EmitAtomicBinaryMinMax(MI, BB, 2, false, ARMCC::HI);
6755 case ARM::ATOMIC_LOAD_UMAX_I32:
6756 return EmitAtomicBinaryMinMax(MI, BB, 4, false, ARMCC::HI);
6758 case ARM::ATOMIC_SWAP_I8: return EmitAtomicBinary(MI, BB, 1, 0);
6759 case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0);
6760 case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0);
6762 case ARM::ATOMIC_CMP_SWAP_I8: return EmitAtomicCmpSwap(MI, BB, 1);
6763 case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2);
6764 case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4);
6767 case ARM::ATOMADD6432:
6768 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr,
6769 isThumb2 ? ARM::t2ADCrr : ARM::ADCrr,
6770 /*NeedsCarry*/ true);
6771 case ARM::ATOMSUB6432:
6772 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
6773 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
6774 /*NeedsCarry*/ true);
6775 case ARM::ATOMOR6432:
6776 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr,
6777 isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6778 case ARM::ATOMXOR6432:
6779 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2EORrr : ARM::EORrr,
6780 isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6781 case ARM::ATOMAND6432:
6782 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr,
6783 isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6784 case ARM::ATOMSWAP6432:
6785 return EmitAtomicBinary64(MI, BB, 0, 0, false);
6786 case ARM::ATOMCMPXCHG6432:
6787 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
6788 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
6789 /*NeedsCarry*/ false, /*IsCmpxchg*/true);
6791 case ARM::tMOVCCr_pseudo: {
6792 // To "insert" a SELECT_CC instruction, we actually have to insert the
6793 // diamond control-flow pattern. The incoming instruction knows the
6794 // destination vreg to set, the condition code register to branch on, the
6795 // true/false values to select between, and a branch opcode to use.
6796 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6797 MachineFunction::iterator It = BB;
6803 // cmpTY ccX, r1, r2
6805 // fallthrough --> copy0MBB
6806 MachineBasicBlock *thisMBB = BB;
6807 MachineFunction *F = BB->getParent();
6808 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
6809 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
6810 F->insert(It, copy0MBB);
6811 F->insert(It, sinkMBB);
6813 // Transfer the remainder of BB and its successor edges to sinkMBB.
6814 sinkMBB->splice(sinkMBB->begin(), BB,
6815 llvm::next(MachineBasicBlock::iterator(MI)),
6817 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
6819 BB->addSuccessor(copy0MBB);
6820 BB->addSuccessor(sinkMBB);
6822 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
6823 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
6826 // %FalseValue = ...
6827 // # fallthrough to sinkMBB
6830 // Update machine-CFG edges
6831 BB->addSuccessor(sinkMBB);
6834 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
6837 BuildMI(*BB, BB->begin(), dl,
6838 TII->get(ARM::PHI), MI->getOperand(0).getReg())
6839 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
6840 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
6842 MI->eraseFromParent(); // The pseudo instruction is gone now.
6847 case ARM::BCCZi64: {
6848 // If there is an unconditional branch to the other successor, remove it.
6849 BB->erase(llvm::next(MachineBasicBlock::iterator(MI)), BB->end());
6851 // Compare both parts that make up the double comparison separately for
6853 bool RHSisZero = MI->getOpcode() == ARM::BCCZi64;
6855 unsigned LHS1 = MI->getOperand(1).getReg();
6856 unsigned LHS2 = MI->getOperand(2).getReg();
6858 AddDefaultPred(BuildMI(BB, dl,
6859 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
6860 .addReg(LHS1).addImm(0));
6861 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
6862 .addReg(LHS2).addImm(0)
6863 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
6865 unsigned RHS1 = MI->getOperand(3).getReg();
6866 unsigned RHS2 = MI->getOperand(4).getReg();
6867 AddDefaultPred(BuildMI(BB, dl,
6868 TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
6869 .addReg(LHS1).addReg(RHS1));
6870 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
6871 .addReg(LHS2).addReg(RHS2)
6872 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
6875 MachineBasicBlock *destMBB = MI->getOperand(RHSisZero ? 3 : 5).getMBB();
6876 MachineBasicBlock *exitMBB = OtherSucc(BB, destMBB);
6877 if (MI->getOperand(0).getImm() == ARMCC::NE)
6878 std::swap(destMBB, exitMBB);
6880 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6881 .addMBB(destMBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
6883 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2B)).addMBB(exitMBB));
6885 BuildMI(BB, dl, TII->get(ARM::B)) .addMBB(exitMBB);
6887 MI->eraseFromParent(); // The pseudo instruction is gone now.
6891 case ARM::Int_eh_sjlj_setjmp:
6892 case ARM::Int_eh_sjlj_setjmp_nofp:
6893 case ARM::tInt_eh_sjlj_setjmp:
6894 case ARM::t2Int_eh_sjlj_setjmp:
6895 case ARM::t2Int_eh_sjlj_setjmp_nofp:
6896 EmitSjLjDispatchBlock(MI, BB);
6901 // To insert an ABS instruction, we have to insert the
6902 // diamond control-flow pattern. The incoming instruction knows the
6903 // source vreg to test against 0, the destination vreg to set,
6904 // the condition code register to branch on, the
6905 // true/false values to select between, and a branch opcode to use.
6910 // BCC (branch to SinkBB if V0 >= 0)
6911 // RSBBB: V3 = RSBri V2, 0 (compute ABS if V2 < 0)
6912 // SinkBB: V1 = PHI(V2, V3)
6913 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6914 MachineFunction::iterator BBI = BB;
6916 MachineFunction *Fn = BB->getParent();
6917 MachineBasicBlock *RSBBB = Fn->CreateMachineBasicBlock(LLVM_BB);
6918 MachineBasicBlock *SinkBB = Fn->CreateMachineBasicBlock(LLVM_BB);
6919 Fn->insert(BBI, RSBBB);
6920 Fn->insert(BBI, SinkBB);
6922 unsigned int ABSSrcReg = MI->getOperand(1).getReg();
6923 unsigned int ABSDstReg = MI->getOperand(0).getReg();
6924 bool isThumb2 = Subtarget->isThumb2();
6925 MachineRegisterInfo &MRI = Fn->getRegInfo();
6926 // In Thumb mode S must not be specified if source register is the SP or
6927 // PC and if destination register is the SP, so restrict register class
6928 unsigned NewRsbDstReg = MRI.createVirtualRegister(isThumb2 ?
6929 (const TargetRegisterClass*)&ARM::rGPRRegClass :
6930 (const TargetRegisterClass*)&ARM::GPRRegClass);
6932 // Transfer the remainder of BB and its successor edges to sinkMBB.
6933 SinkBB->splice(SinkBB->begin(), BB,
6934 llvm::next(MachineBasicBlock::iterator(MI)),
6936 SinkBB->transferSuccessorsAndUpdatePHIs(BB);
6938 BB->addSuccessor(RSBBB);
6939 BB->addSuccessor(SinkBB);
6941 // fall through to SinkMBB
6942 RSBBB->addSuccessor(SinkBB);
6944 // insert a cmp at the end of BB
6945 AddDefaultPred(BuildMI(BB, dl,
6946 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
6947 .addReg(ABSSrcReg).addImm(0));
6949 // insert a bcc with opposite CC to ARMCC::MI at the end of BB
6951 TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)).addMBB(SinkBB)
6952 .addImm(ARMCC::getOppositeCondition(ARMCC::MI)).addReg(ARM::CPSR);
6954 // insert rsbri in RSBBB
6955 // Note: BCC and rsbri will be converted into predicated rsbmi
6956 // by if-conversion pass
6957 BuildMI(*RSBBB, RSBBB->begin(), dl,
6958 TII->get(isThumb2 ? ARM::t2RSBri : ARM::RSBri), NewRsbDstReg)
6959 .addReg(ABSSrcReg, RegState::Kill)
6960 .addImm(0).addImm((unsigned)ARMCC::AL).addReg(0).addReg(0);
6962 // insert PHI in SinkBB,
6963 // reuse ABSDstReg to not change uses of ABS instruction
6964 BuildMI(*SinkBB, SinkBB->begin(), dl,
6965 TII->get(ARM::PHI), ABSDstReg)
6966 .addReg(NewRsbDstReg).addMBB(RSBBB)
6967 .addReg(ABSSrcReg).addMBB(BB);
6969 // remove ABS instruction
6970 MI->eraseFromParent();
6972 // return last added BB
6975 case ARM::COPY_STRUCT_BYVAL_I32:
6977 return EmitStructByval(MI, BB);
6981 void ARMTargetLowering::AdjustInstrPostInstrSelection(MachineInstr *MI,
6982 SDNode *Node) const {
6983 if (!MI->hasPostISelHook()) {
6984 assert(!convertAddSubFlagsOpcode(MI->getOpcode()) &&
6985 "Pseudo flag-setting opcodes must be marked with 'hasPostISelHook'");
6989 const MCInstrDesc *MCID = &MI->getDesc();
6990 // Adjust potentially 's' setting instructions after isel, i.e. ADC, SBC, RSB,
6991 // RSC. Coming out of isel, they have an implicit CPSR def, but the optional
6992 // operand is still set to noreg. If needed, set the optional operand's
6993 // register to CPSR, and remove the redundant implicit def.
6995 // e.g. ADCS (..., CPSR<imp-def>) -> ADC (... opt:CPSR<def>).
6997 // Rename pseudo opcodes.
6998 unsigned NewOpc = convertAddSubFlagsOpcode(MI->getOpcode());
7000 const ARMBaseInstrInfo *TII =
7001 static_cast<const ARMBaseInstrInfo*>(getTargetMachine().getInstrInfo());
7002 MCID = &TII->get(NewOpc);
7004 assert(MCID->getNumOperands() == MI->getDesc().getNumOperands() + 1 &&
7005 "converted opcode should be the same except for cc_out");
7009 // Add the optional cc_out operand
7010 MI->addOperand(MachineOperand::CreateReg(0, /*isDef=*/true));
7012 unsigned ccOutIdx = MCID->getNumOperands() - 1;
7014 // Any ARM instruction that sets the 's' bit should specify an optional
7015 // "cc_out" operand in the last operand position.
7016 if (!MI->hasOptionalDef() || !MCID->OpInfo[ccOutIdx].isOptionalDef()) {
7017 assert(!NewOpc && "Optional cc_out operand required");
7020 // Look for an implicit def of CPSR added by MachineInstr ctor. Remove it
7021 // since we already have an optional CPSR def.
7022 bool definesCPSR = false;
7023 bool deadCPSR = false;
7024 for (unsigned i = MCID->getNumOperands(), e = MI->getNumOperands();
7026 const MachineOperand &MO = MI->getOperand(i);
7027 if (MO.isReg() && MO.isDef() && MO.getReg() == ARM::CPSR) {
7031 MI->RemoveOperand(i);
7036 assert(!NewOpc && "Optional cc_out operand required");
7039 assert(deadCPSR == !Node->hasAnyUseOfValue(1) && "inconsistent dead flag");
7041 assert(!MI->getOperand(ccOutIdx).getReg() &&
7042 "expect uninitialized optional cc_out operand");
7046 // If this instruction was defined with an optional CPSR def and its dag node
7047 // had a live implicit CPSR def, then activate the optional CPSR def.
7048 MachineOperand &MO = MI->getOperand(ccOutIdx);
7049 MO.setReg(ARM::CPSR);
7053 //===----------------------------------------------------------------------===//
7054 // ARM Optimization Hooks
7055 //===----------------------------------------------------------------------===//
7057 // Helper function that checks if N is a null or all ones constant.
7058 static inline bool isZeroOrAllOnes(SDValue N, bool AllOnes) {
7059 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N);
7062 return AllOnes ? C->isAllOnesValue() : C->isNullValue();
7065 // Return true if N is conditionally 0 or all ones.
7066 // Detects these expressions where cc is an i1 value:
7068 // (select cc 0, y) [AllOnes=0]
7069 // (select cc y, 0) [AllOnes=0]
7070 // (zext cc) [AllOnes=0]
7071 // (sext cc) [AllOnes=0/1]
7072 // (select cc -1, y) [AllOnes=1]
7073 // (select cc y, -1) [AllOnes=1]
7075 // Invert is set when N is the null/all ones constant when CC is false.
7076 // OtherOp is set to the alternative value of N.
7077 static bool isConditionalZeroOrAllOnes(SDNode *N, bool AllOnes,
7078 SDValue &CC, bool &Invert,
7080 SelectionDAG &DAG) {
7081 switch (N->getOpcode()) {
7082 default: return false;
7084 CC = N->getOperand(0);
7085 SDValue N1 = N->getOperand(1);
7086 SDValue N2 = N->getOperand(2);
7087 if (isZeroOrAllOnes(N1, AllOnes)) {
7092 if (isZeroOrAllOnes(N2, AllOnes)) {
7099 case ISD::ZERO_EXTEND:
7100 // (zext cc) can never be the all ones value.
7104 case ISD::SIGN_EXTEND: {
7105 EVT VT = N->getValueType(0);
7106 CC = N->getOperand(0);
7107 if (CC.getValueType() != MVT::i1)
7111 // When looking for an AllOnes constant, N is an sext, and the 'other'
7113 OtherOp = DAG.getConstant(0, VT);
7114 else if (N->getOpcode() == ISD::ZERO_EXTEND)
7115 // When looking for a 0 constant, N can be zext or sext.
7116 OtherOp = DAG.getConstant(1, VT);
7118 OtherOp = DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), VT);
7124 // Combine a constant select operand into its use:
7126 // (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
7127 // (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
7128 // (and (select cc, -1, c), x) -> (select cc, x, (and, x, c)) [AllOnes=1]
7129 // (or (select cc, 0, c), x) -> (select cc, x, (or, x, c))
7130 // (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c))
7132 // The transform is rejected if the select doesn't have a constant operand that
7133 // is null, or all ones when AllOnes is set.
7135 // Also recognize sext/zext from i1:
7137 // (add (zext cc), x) -> (select cc (add x, 1), x)
7138 // (add (sext cc), x) -> (select cc (add x, -1), x)
7140 // These transformations eventually create predicated instructions.
7142 // @param N The node to transform.
7143 // @param Slct The N operand that is a select.
7144 // @param OtherOp The other N operand (x above).
7145 // @param DCI Context.
7146 // @param AllOnes Require the select constant to be all ones instead of null.
7147 // @returns The new node, or SDValue() on failure.
7149 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
7150 TargetLowering::DAGCombinerInfo &DCI,
7151 bool AllOnes = false) {
7152 SelectionDAG &DAG = DCI.DAG;
7153 EVT VT = N->getValueType(0);
7154 SDValue NonConstantVal;
7157 if (!isConditionalZeroOrAllOnes(Slct.getNode(), AllOnes, CCOp, SwapSelectOps,
7158 NonConstantVal, DAG))
7161 // Slct is now know to be the desired identity constant when CC is true.
7162 SDValue TrueVal = OtherOp;
7163 SDValue FalseVal = DAG.getNode(N->getOpcode(), N->getDebugLoc(), VT,
7164 OtherOp, NonConstantVal);
7165 // Unless SwapSelectOps says CC should be false.
7167 std::swap(TrueVal, FalseVal);
7169 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
7170 CCOp, TrueVal, FalseVal);
7173 // Attempt combineSelectAndUse on each operand of a commutative operator N.
7175 SDValue combineSelectAndUseCommutative(SDNode *N, bool AllOnes,
7176 TargetLowering::DAGCombinerInfo &DCI) {
7177 SDValue N0 = N->getOperand(0);
7178 SDValue N1 = N->getOperand(1);
7179 if (N0.getNode()->hasOneUse()) {
7180 SDValue Result = combineSelectAndUse(N, N0, N1, DCI, AllOnes);
7181 if (Result.getNode())
7184 if (N1.getNode()->hasOneUse()) {
7185 SDValue Result = combineSelectAndUse(N, N1, N0, DCI, AllOnes);
7186 if (Result.getNode())
7192 // AddCombineToVPADDL- For pair-wise add on neon, use the vpaddl instruction
7193 // (only after legalization).
7194 static SDValue AddCombineToVPADDL(SDNode *N, SDValue N0, SDValue N1,
7195 TargetLowering::DAGCombinerInfo &DCI,
7196 const ARMSubtarget *Subtarget) {
7198 // Only perform optimization if after legalize, and if NEON is available. We
7199 // also expected both operands to be BUILD_VECTORs.
7200 if (DCI.isBeforeLegalize() || !Subtarget->hasNEON()
7201 || N0.getOpcode() != ISD::BUILD_VECTOR
7202 || N1.getOpcode() != ISD::BUILD_VECTOR)
7205 // Check output type since VPADDL operand elements can only be 8, 16, or 32.
7206 EVT VT = N->getValueType(0);
7207 if (!VT.isInteger() || VT.getVectorElementType() == MVT::i64)
7210 // Check that the vector operands are of the right form.
7211 // N0 and N1 are BUILD_VECTOR nodes with N number of EXTRACT_VECTOR
7212 // operands, where N is the size of the formed vector.
7213 // Each EXTRACT_VECTOR should have the same input vector and odd or even
7214 // index such that we have a pair wise add pattern.
7216 // Grab the vector that all EXTRACT_VECTOR nodes should be referencing.
7217 if (N0->getOperand(0)->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
7219 SDValue Vec = N0->getOperand(0)->getOperand(0);
7220 SDNode *V = Vec.getNode();
7221 unsigned nextIndex = 0;
7223 // For each operands to the ADD which are BUILD_VECTORs,
7224 // check to see if each of their operands are an EXTRACT_VECTOR with
7225 // the same vector and appropriate index.
7226 for (unsigned i = 0, e = N0->getNumOperands(); i != e; ++i) {
7227 if (N0->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT
7228 && N1->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
7230 SDValue ExtVec0 = N0->getOperand(i);
7231 SDValue ExtVec1 = N1->getOperand(i);
7233 // First operand is the vector, verify its the same.
7234 if (V != ExtVec0->getOperand(0).getNode() ||
7235 V != ExtVec1->getOperand(0).getNode())
7238 // Second is the constant, verify its correct.
7239 ConstantSDNode *C0 = dyn_cast<ConstantSDNode>(ExtVec0->getOperand(1));
7240 ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(ExtVec1->getOperand(1));
7242 // For the constant, we want to see all the even or all the odd.
7243 if (!C0 || !C1 || C0->getZExtValue() != nextIndex
7244 || C1->getZExtValue() != nextIndex+1)
7253 // Create VPADDL node.
7254 SelectionDAG &DAG = DCI.DAG;
7255 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7257 // Build operand list.
7258 SmallVector<SDValue, 8> Ops;
7259 Ops.push_back(DAG.getConstant(Intrinsic::arm_neon_vpaddls,
7260 TLI.getPointerTy()));
7262 // Input is the vector.
7265 // Get widened type and narrowed type.
7267 unsigned numElem = VT.getVectorNumElements();
7268 switch (VT.getVectorElementType().getSimpleVT().SimpleTy) {
7269 case MVT::i8: widenType = MVT::getVectorVT(MVT::i16, numElem); break;
7270 case MVT::i16: widenType = MVT::getVectorVT(MVT::i32, numElem); break;
7271 case MVT::i32: widenType = MVT::getVectorVT(MVT::i64, numElem); break;
7273 llvm_unreachable("Invalid vector element type for padd optimization.");
7276 SDValue tmp = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(),
7277 widenType, &Ops[0], Ops.size());
7278 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, tmp);
7281 static SDValue findMUL_LOHI(SDValue V) {
7282 if (V->getOpcode() == ISD::UMUL_LOHI ||
7283 V->getOpcode() == ISD::SMUL_LOHI)
7288 static SDValue AddCombineTo64bitMLAL(SDNode *AddcNode,
7289 TargetLowering::DAGCombinerInfo &DCI,
7290 const ARMSubtarget *Subtarget) {
7292 if (Subtarget->isThumb1Only()) return SDValue();
7294 // Only perform the checks after legalize when the pattern is available.
7295 if (DCI.isBeforeLegalize()) return SDValue();
7297 // Look for multiply add opportunities.
7298 // The pattern is a ISD::UMUL_LOHI followed by two add nodes, where
7299 // each add nodes consumes a value from ISD::UMUL_LOHI and there is
7300 // a glue link from the first add to the second add.
7301 // If we find this pattern, we can replace the U/SMUL_LOHI, ADDC, and ADDE by
7302 // a S/UMLAL instruction.
7305 // \ / \ [no multiline comment]
7311 assert(AddcNode->getOpcode() == ISD::ADDC && "Expect an ADDC");
7312 SDValue AddcOp0 = AddcNode->getOperand(0);
7313 SDValue AddcOp1 = AddcNode->getOperand(1);
7315 // Check if the two operands are from the same mul_lohi node.
7316 if (AddcOp0.getNode() == AddcOp1.getNode())
7319 assert(AddcNode->getNumValues() == 2 &&
7320 AddcNode->getValueType(0) == MVT::i32 &&
7321 AddcNode->getValueType(1) == MVT::Glue &&
7322 "Expect ADDC with two result values: i32, glue");
7324 // Check that the ADDC adds the low result of the S/UMUL_LOHI.
7325 if (AddcOp0->getOpcode() != ISD::UMUL_LOHI &&
7326 AddcOp0->getOpcode() != ISD::SMUL_LOHI &&
7327 AddcOp1->getOpcode() != ISD::UMUL_LOHI &&
7328 AddcOp1->getOpcode() != ISD::SMUL_LOHI)
7331 // Look for the glued ADDE.
7332 SDNode* AddeNode = AddcNode->getGluedUser();
7333 if (AddeNode == NULL)
7336 // Make sure it is really an ADDE.
7337 if (AddeNode->getOpcode() != ISD::ADDE)
7340 assert(AddeNode->getNumOperands() == 3 &&
7341 AddeNode->getOperand(2).getValueType() == MVT::Glue &&
7342 "ADDE node has the wrong inputs");
7344 // Check for the triangle shape.
7345 SDValue AddeOp0 = AddeNode->getOperand(0);
7346 SDValue AddeOp1 = AddeNode->getOperand(1);
7348 // Make sure that the ADDE operands are not coming from the same node.
7349 if (AddeOp0.getNode() == AddeOp1.getNode())
7352 // Find the MUL_LOHI node walking up ADDE's operands.
7353 bool IsLeftOperandMUL = false;
7354 SDValue MULOp = findMUL_LOHI(AddeOp0);
7355 if (MULOp == SDValue())
7356 MULOp = findMUL_LOHI(AddeOp1);
7358 IsLeftOperandMUL = true;
7359 if (MULOp == SDValue())
7362 // Figure out the right opcode.
7363 unsigned Opc = MULOp->getOpcode();
7364 unsigned FinalOpc = (Opc == ISD::SMUL_LOHI) ? ARMISD::SMLAL : ARMISD::UMLAL;
7366 // Figure out the high and low input values to the MLAL node.
7367 SDValue* HiMul = &MULOp;
7368 SDValue* HiAdd = NULL;
7369 SDValue* LoMul = NULL;
7370 SDValue* LowAdd = NULL;
7372 if (IsLeftOperandMUL)
7378 if (AddcOp0->getOpcode() == Opc) {
7382 if (AddcOp1->getOpcode() == Opc) {
7390 if (LoMul->getNode() != HiMul->getNode())
7393 // Create the merged node.
7394 SelectionDAG &DAG = DCI.DAG;
7396 // Build operand list.
7397 SmallVector<SDValue, 8> Ops;
7398 Ops.push_back(LoMul->getOperand(0));
7399 Ops.push_back(LoMul->getOperand(1));
7400 Ops.push_back(*LowAdd);
7401 Ops.push_back(*HiAdd);
7403 SDValue MLALNode = DAG.getNode(FinalOpc, AddcNode->getDebugLoc(),
7404 DAG.getVTList(MVT::i32, MVT::i32),
7405 &Ops[0], Ops.size());
7407 // Replace the ADDs' nodes uses by the MLA node's values.
7408 SDValue HiMLALResult(MLALNode.getNode(), 1);
7409 DAG.ReplaceAllUsesOfValueWith(SDValue(AddeNode, 0), HiMLALResult);
7411 SDValue LoMLALResult(MLALNode.getNode(), 0);
7412 DAG.ReplaceAllUsesOfValueWith(SDValue(AddcNode, 0), LoMLALResult);
7414 // Return original node to notify the driver to stop replacing.
7415 SDValue resNode(AddcNode, 0);
7419 /// PerformADDCCombine - Target-specific dag combine transform from
7420 /// ISD::ADDC, ISD::ADDE, and ISD::MUL_LOHI to MLAL.
7421 static SDValue PerformADDCCombine(SDNode *N,
7422 TargetLowering::DAGCombinerInfo &DCI,
7423 const ARMSubtarget *Subtarget) {
7425 return AddCombineTo64bitMLAL(N, DCI, Subtarget);
7429 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
7430 /// operands N0 and N1. This is a helper for PerformADDCombine that is
7431 /// called with the default operands, and if that fails, with commuted
7433 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
7434 TargetLowering::DAGCombinerInfo &DCI,
7435 const ARMSubtarget *Subtarget){
7437 // Attempt to create vpaddl for this add.
7438 SDValue Result = AddCombineToVPADDL(N, N0, N1, DCI, Subtarget);
7439 if (Result.getNode())
7442 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
7443 if (N0.getNode()->hasOneUse()) {
7444 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
7445 if (Result.getNode()) return Result;
7450 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
7452 static SDValue PerformADDCombine(SDNode *N,
7453 TargetLowering::DAGCombinerInfo &DCI,
7454 const ARMSubtarget *Subtarget) {
7455 SDValue N0 = N->getOperand(0);
7456 SDValue N1 = N->getOperand(1);
7458 // First try with the default operand order.
7459 SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI, Subtarget);
7460 if (Result.getNode())
7463 // If that didn't work, try again with the operands commuted.
7464 return PerformADDCombineWithOperands(N, N1, N0, DCI, Subtarget);
7467 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
7469 static SDValue PerformSUBCombine(SDNode *N,
7470 TargetLowering::DAGCombinerInfo &DCI) {
7471 SDValue N0 = N->getOperand(0);
7472 SDValue N1 = N->getOperand(1);
7474 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
7475 if (N1.getNode()->hasOneUse()) {
7476 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
7477 if (Result.getNode()) return Result;
7483 /// PerformVMULCombine
7484 /// Distribute (A + B) * C to (A * C) + (B * C) to take advantage of the
7485 /// special multiplier accumulator forwarding.
7491 static SDValue PerformVMULCombine(SDNode *N,
7492 TargetLowering::DAGCombinerInfo &DCI,
7493 const ARMSubtarget *Subtarget) {
7494 if (!Subtarget->hasVMLxForwarding())
7497 SelectionDAG &DAG = DCI.DAG;
7498 SDValue N0 = N->getOperand(0);
7499 SDValue N1 = N->getOperand(1);
7500 unsigned Opcode = N0.getOpcode();
7501 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
7502 Opcode != ISD::FADD && Opcode != ISD::FSUB) {
7503 Opcode = N1.getOpcode();
7504 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
7505 Opcode != ISD::FADD && Opcode != ISD::FSUB)
7510 EVT VT = N->getValueType(0);
7511 DebugLoc DL = N->getDebugLoc();
7512 SDValue N00 = N0->getOperand(0);
7513 SDValue N01 = N0->getOperand(1);
7514 return DAG.getNode(Opcode, DL, VT,
7515 DAG.getNode(ISD::MUL, DL, VT, N00, N1),
7516 DAG.getNode(ISD::MUL, DL, VT, N01, N1));
7519 static SDValue PerformMULCombine(SDNode *N,
7520 TargetLowering::DAGCombinerInfo &DCI,
7521 const ARMSubtarget *Subtarget) {
7522 SelectionDAG &DAG = DCI.DAG;
7524 if (Subtarget->isThumb1Only())
7527 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
7530 EVT VT = N->getValueType(0);
7531 if (VT.is64BitVector() || VT.is128BitVector())
7532 return PerformVMULCombine(N, DCI, Subtarget);
7536 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
7540 int64_t MulAmt = C->getSExtValue();
7541 unsigned ShiftAmt = CountTrailingZeros_64(MulAmt);
7543 ShiftAmt = ShiftAmt & (32 - 1);
7544 SDValue V = N->getOperand(0);
7545 DebugLoc DL = N->getDebugLoc();
7548 MulAmt >>= ShiftAmt;
7551 if (isPowerOf2_32(MulAmt - 1)) {
7552 // (mul x, 2^N + 1) => (add (shl x, N), x)
7553 Res = DAG.getNode(ISD::ADD, DL, VT,
7555 DAG.getNode(ISD::SHL, DL, VT,
7557 DAG.getConstant(Log2_32(MulAmt - 1),
7559 } else if (isPowerOf2_32(MulAmt + 1)) {
7560 // (mul x, 2^N - 1) => (sub (shl x, N), x)
7561 Res = DAG.getNode(ISD::SUB, DL, VT,
7562 DAG.getNode(ISD::SHL, DL, VT,
7564 DAG.getConstant(Log2_32(MulAmt + 1),
7570 uint64_t MulAmtAbs = -MulAmt;
7571 if (isPowerOf2_32(MulAmtAbs + 1)) {
7572 // (mul x, -(2^N - 1)) => (sub x, (shl x, N))
7573 Res = DAG.getNode(ISD::SUB, DL, VT,
7575 DAG.getNode(ISD::SHL, DL, VT,
7577 DAG.getConstant(Log2_32(MulAmtAbs + 1),
7579 } else if (isPowerOf2_32(MulAmtAbs - 1)) {
7580 // (mul x, -(2^N + 1)) => - (add (shl x, N), x)
7581 Res = DAG.getNode(ISD::ADD, DL, VT,
7583 DAG.getNode(ISD::SHL, DL, VT,
7585 DAG.getConstant(Log2_32(MulAmtAbs-1),
7587 Res = DAG.getNode(ISD::SUB, DL, VT,
7588 DAG.getConstant(0, MVT::i32),Res);
7595 Res = DAG.getNode(ISD::SHL, DL, VT,
7596 Res, DAG.getConstant(ShiftAmt, MVT::i32));
7598 // Do not add new nodes to DAG combiner worklist.
7599 DCI.CombineTo(N, Res, false);
7603 static SDValue PerformANDCombine(SDNode *N,
7604 TargetLowering::DAGCombinerInfo &DCI,
7605 const ARMSubtarget *Subtarget) {
7607 // Attempt to use immediate-form VBIC
7608 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
7609 DebugLoc dl = N->getDebugLoc();
7610 EVT VT = N->getValueType(0);
7611 SelectionDAG &DAG = DCI.DAG;
7613 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
7616 APInt SplatBits, SplatUndef;
7617 unsigned SplatBitSize;
7620 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
7621 if (SplatBitSize <= 64) {
7623 SDValue Val = isNEONModifiedImm((~SplatBits).getZExtValue(),
7624 SplatUndef.getZExtValue(), SplatBitSize,
7625 DAG, VbicVT, VT.is128BitVector(),
7627 if (Val.getNode()) {
7629 DAG.getNode(ISD::BITCAST, dl, VbicVT, N->getOperand(0));
7630 SDValue Vbic = DAG.getNode(ARMISD::VBICIMM, dl, VbicVT, Input, Val);
7631 return DAG.getNode(ISD::BITCAST, dl, VT, Vbic);
7636 if (!Subtarget->isThumb1Only()) {
7637 // fold (and (select cc, -1, c), x) -> (select cc, x, (and, x, c))
7638 SDValue Result = combineSelectAndUseCommutative(N, true, DCI);
7639 if (Result.getNode())
7646 /// PerformORCombine - Target-specific dag combine xforms for ISD::OR
7647 static SDValue PerformORCombine(SDNode *N,
7648 TargetLowering::DAGCombinerInfo &DCI,
7649 const ARMSubtarget *Subtarget) {
7650 // Attempt to use immediate-form VORR
7651 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
7652 DebugLoc dl = N->getDebugLoc();
7653 EVT VT = N->getValueType(0);
7654 SelectionDAG &DAG = DCI.DAG;
7656 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
7659 APInt SplatBits, SplatUndef;
7660 unsigned SplatBitSize;
7662 if (BVN && Subtarget->hasNEON() &&
7663 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
7664 if (SplatBitSize <= 64) {
7666 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
7667 SplatUndef.getZExtValue(), SplatBitSize,
7668 DAG, VorrVT, VT.is128BitVector(),
7670 if (Val.getNode()) {
7672 DAG.getNode(ISD::BITCAST, dl, VorrVT, N->getOperand(0));
7673 SDValue Vorr = DAG.getNode(ARMISD::VORRIMM, dl, VorrVT, Input, Val);
7674 return DAG.getNode(ISD::BITCAST, dl, VT, Vorr);
7679 if (!Subtarget->isThumb1Only()) {
7680 // fold (or (select cc, 0, c), x) -> (select cc, x, (or, x, c))
7681 SDValue Result = combineSelectAndUseCommutative(N, false, DCI);
7682 if (Result.getNode())
7686 // The code below optimizes (or (and X, Y), Z).
7687 // The AND operand needs to have a single user to make these optimizations
7689 SDValue N0 = N->getOperand(0);
7690 if (N0.getOpcode() != ISD::AND || !N0.hasOneUse())
7692 SDValue N1 = N->getOperand(1);
7694 // (or (and B, A), (and C, ~A)) => (VBSL A, B, C) when A is a constant.
7695 if (Subtarget->hasNEON() && N1.getOpcode() == ISD::AND && VT.isVector() &&
7696 DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
7698 unsigned SplatBitSize;
7701 BuildVectorSDNode *BVN0 = dyn_cast<BuildVectorSDNode>(N0->getOperand(1));
7703 if (BVN0 && BVN0->isConstantSplat(SplatBits0, SplatUndef, SplatBitSize,
7704 HasAnyUndefs) && !HasAnyUndefs) {
7705 BuildVectorSDNode *BVN1 = dyn_cast<BuildVectorSDNode>(N1->getOperand(1));
7707 if (BVN1 && BVN1->isConstantSplat(SplatBits1, SplatUndef, SplatBitSize,
7708 HasAnyUndefs) && !HasAnyUndefs &&
7709 SplatBits0 == ~SplatBits1) {
7710 // Canonicalize the vector type to make instruction selection simpler.
7711 EVT CanonicalVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
7712 SDValue Result = DAG.getNode(ARMISD::VBSL, dl, CanonicalVT,
7713 N0->getOperand(1), N0->getOperand(0),
7715 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
7720 // Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when
7723 // BFI is only available on V6T2+
7724 if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops())
7727 DebugLoc DL = N->getDebugLoc();
7728 // 1) or (and A, mask), val => ARMbfi A, val, mask
7729 // iff (val & mask) == val
7731 // 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
7732 // 2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2)
7733 // && mask == ~mask2
7734 // 2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2)
7735 // && ~mask == mask2
7736 // (i.e., copy a bitfield value into another bitfield of the same width)
7741 SDValue N00 = N0.getOperand(0);
7743 // The value and the mask need to be constants so we can verify this is
7744 // actually a bitfield set. If the mask is 0xffff, we can do better
7745 // via a movt instruction, so don't use BFI in that case.
7746 SDValue MaskOp = N0.getOperand(1);
7747 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(MaskOp);
7750 unsigned Mask = MaskC->getZExtValue();
7754 // Case (1): or (and A, mask), val => ARMbfi A, val, mask
7755 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
7757 unsigned Val = N1C->getZExtValue();
7758 if ((Val & ~Mask) != Val)
7761 if (ARM::isBitFieldInvertedMask(Mask)) {
7762 Val >>= CountTrailingZeros_32(~Mask);
7764 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00,
7765 DAG.getConstant(Val, MVT::i32),
7766 DAG.getConstant(Mask, MVT::i32));
7768 // Do not add new nodes to DAG combiner worklist.
7769 DCI.CombineTo(N, Res, false);
7772 } else if (N1.getOpcode() == ISD::AND) {
7773 // case (2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
7774 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
7777 unsigned Mask2 = N11C->getZExtValue();
7779 // Mask and ~Mask2 (or reverse) must be equivalent for the BFI pattern
7781 if (ARM::isBitFieldInvertedMask(Mask) &&
7783 // The pack halfword instruction works better for masks that fit it,
7784 // so use that when it's available.
7785 if (Subtarget->hasT2ExtractPack() &&
7786 (Mask == 0xffff || Mask == 0xffff0000))
7789 unsigned amt = CountTrailingZeros_32(Mask2);
7790 Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0),
7791 DAG.getConstant(amt, MVT::i32));
7792 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00, Res,
7793 DAG.getConstant(Mask, MVT::i32));
7794 // Do not add new nodes to DAG combiner worklist.
7795 DCI.CombineTo(N, Res, false);
7797 } else if (ARM::isBitFieldInvertedMask(~Mask) &&
7799 // The pack halfword instruction works better for masks that fit it,
7800 // so use that when it's available.
7801 if (Subtarget->hasT2ExtractPack() &&
7802 (Mask2 == 0xffff || Mask2 == 0xffff0000))
7805 unsigned lsb = CountTrailingZeros_32(Mask);
7806 Res = DAG.getNode(ISD::SRL, DL, VT, N00,
7807 DAG.getConstant(lsb, MVT::i32));
7808 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res,
7809 DAG.getConstant(Mask2, MVT::i32));
7810 // Do not add new nodes to DAG combiner worklist.
7811 DCI.CombineTo(N, Res, false);
7816 if (DAG.MaskedValueIsZero(N1, MaskC->getAPIntValue()) &&
7817 N00.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N00.getOperand(1)) &&
7818 ARM::isBitFieldInvertedMask(~Mask)) {
7819 // Case (3): or (and (shl A, #shamt), mask), B => ARMbfi B, A, ~mask
7820 // where lsb(mask) == #shamt and masked bits of B are known zero.
7821 SDValue ShAmt = N00.getOperand(1);
7822 unsigned ShAmtC = cast<ConstantSDNode>(ShAmt)->getZExtValue();
7823 unsigned LSB = CountTrailingZeros_32(Mask);
7827 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1, N00.getOperand(0),
7828 DAG.getConstant(~Mask, MVT::i32));
7830 // Do not add new nodes to DAG combiner worklist.
7831 DCI.CombineTo(N, Res, false);
7837 static SDValue PerformXORCombine(SDNode *N,
7838 TargetLowering::DAGCombinerInfo &DCI,
7839 const ARMSubtarget *Subtarget) {
7840 EVT VT = N->getValueType(0);
7841 SelectionDAG &DAG = DCI.DAG;
7843 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
7846 if (!Subtarget->isThumb1Only()) {
7847 // fold (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c))
7848 SDValue Result = combineSelectAndUseCommutative(N, false, DCI);
7849 if (Result.getNode())
7856 /// PerformBFICombine - (bfi A, (and B, Mask1), Mask2) -> (bfi A, B, Mask2) iff
7857 /// the bits being cleared by the AND are not demanded by the BFI.
7858 static SDValue PerformBFICombine(SDNode *N,
7859 TargetLowering::DAGCombinerInfo &DCI) {
7860 SDValue N1 = N->getOperand(1);
7861 if (N1.getOpcode() == ISD::AND) {
7862 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
7865 unsigned InvMask = cast<ConstantSDNode>(N->getOperand(2))->getZExtValue();
7866 unsigned LSB = CountTrailingZeros_32(~InvMask);
7867 unsigned Width = (32 - CountLeadingZeros_32(~InvMask)) - LSB;
7868 unsigned Mask = (1 << Width)-1;
7869 unsigned Mask2 = N11C->getZExtValue();
7870 if ((Mask & (~Mask2)) == 0)
7871 return DCI.DAG.getNode(ARMISD::BFI, N->getDebugLoc(), N->getValueType(0),
7872 N->getOperand(0), N1.getOperand(0),
7878 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
7879 /// ARMISD::VMOVRRD.
7880 static SDValue PerformVMOVRRDCombine(SDNode *N,
7881 TargetLowering::DAGCombinerInfo &DCI) {
7882 // vmovrrd(vmovdrr x, y) -> x,y
7883 SDValue InDouble = N->getOperand(0);
7884 if (InDouble.getOpcode() == ARMISD::VMOVDRR)
7885 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
7887 // vmovrrd(load f64) -> (load i32), (load i32)
7888 SDNode *InNode = InDouble.getNode();
7889 if (ISD::isNormalLoad(InNode) && InNode->hasOneUse() &&
7890 InNode->getValueType(0) == MVT::f64 &&
7891 InNode->getOperand(1).getOpcode() == ISD::FrameIndex &&
7892 !cast<LoadSDNode>(InNode)->isVolatile()) {
7893 // TODO: Should this be done for non-FrameIndex operands?
7894 LoadSDNode *LD = cast<LoadSDNode>(InNode);
7896 SelectionDAG &DAG = DCI.DAG;
7897 DebugLoc DL = LD->getDebugLoc();
7898 SDValue BasePtr = LD->getBasePtr();
7899 SDValue NewLD1 = DAG.getLoad(MVT::i32, DL, LD->getChain(), BasePtr,
7900 LD->getPointerInfo(), LD->isVolatile(),
7901 LD->isNonTemporal(), LD->isInvariant(),
7902 LD->getAlignment());
7904 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
7905 DAG.getConstant(4, MVT::i32));
7906 SDValue NewLD2 = DAG.getLoad(MVT::i32, DL, NewLD1.getValue(1), OffsetPtr,
7907 LD->getPointerInfo(), LD->isVolatile(),
7908 LD->isNonTemporal(), LD->isInvariant(),
7909 std::min(4U, LD->getAlignment() / 2));
7911 DAG.ReplaceAllUsesOfValueWith(SDValue(LD, 1), NewLD2.getValue(1));
7912 SDValue Result = DCI.CombineTo(N, NewLD1, NewLD2);
7913 DCI.RemoveFromWorklist(LD);
7921 /// PerformVMOVDRRCombine - Target-specific dag combine xforms for
7922 /// ARMISD::VMOVDRR. This is also used for BUILD_VECTORs with 2 operands.
7923 static SDValue PerformVMOVDRRCombine(SDNode *N, SelectionDAG &DAG) {
7924 // N=vmovrrd(X); vmovdrr(N:0, N:1) -> bit_convert(X)
7925 SDValue Op0 = N->getOperand(0);
7926 SDValue Op1 = N->getOperand(1);
7927 if (Op0.getOpcode() == ISD::BITCAST)
7928 Op0 = Op0.getOperand(0);
7929 if (Op1.getOpcode() == ISD::BITCAST)
7930 Op1 = Op1.getOperand(0);
7931 if (Op0.getOpcode() == ARMISD::VMOVRRD &&
7932 Op0.getNode() == Op1.getNode() &&
7933 Op0.getResNo() == 0 && Op1.getResNo() == 1)
7934 return DAG.getNode(ISD::BITCAST, N->getDebugLoc(),
7935 N->getValueType(0), Op0.getOperand(0));
7939 /// PerformSTORECombine - Target-specific dag combine xforms for
7941 static SDValue PerformSTORECombine(SDNode *N,
7942 TargetLowering::DAGCombinerInfo &DCI) {
7943 StoreSDNode *St = cast<StoreSDNode>(N);
7944 if (St->isVolatile())
7947 // Optimize trunc store (of multiple scalars) to shuffle and store. First,
7948 // pack all of the elements in one place. Next, store to memory in fewer
7950 SDValue StVal = St->getValue();
7951 EVT VT = StVal.getValueType();
7952 if (St->isTruncatingStore() && VT.isVector()) {
7953 SelectionDAG &DAG = DCI.DAG;
7954 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7955 EVT StVT = St->getMemoryVT();
7956 unsigned NumElems = VT.getVectorNumElements();
7957 assert(StVT != VT && "Cannot truncate to the same type");
7958 unsigned FromEltSz = VT.getVectorElementType().getSizeInBits();
7959 unsigned ToEltSz = StVT.getVectorElementType().getSizeInBits();
7961 // From, To sizes and ElemCount must be pow of two
7962 if (!isPowerOf2_32(NumElems * FromEltSz * ToEltSz)) return SDValue();
7964 // We are going to use the original vector elt for storing.
7965 // Accumulated smaller vector elements must be a multiple of the store size.
7966 if (0 != (NumElems * FromEltSz) % ToEltSz) return SDValue();
7968 unsigned SizeRatio = FromEltSz / ToEltSz;
7969 assert(SizeRatio * NumElems * ToEltSz == VT.getSizeInBits());
7971 // Create a type on which we perform the shuffle.
7972 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(), StVT.getScalarType(),
7973 NumElems*SizeRatio);
7974 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
7976 DebugLoc DL = St->getDebugLoc();
7977 SDValue WideVec = DAG.getNode(ISD::BITCAST, DL, WideVecVT, StVal);
7978 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
7979 for (unsigned i = 0; i < NumElems; ++i) ShuffleVec[i] = i * SizeRatio;
7981 // Can't shuffle using an illegal type.
7982 if (!TLI.isTypeLegal(WideVecVT)) return SDValue();
7984 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, DL, WideVec,
7985 DAG.getUNDEF(WideVec.getValueType()),
7987 // At this point all of the data is stored at the bottom of the
7988 // register. We now need to save it to mem.
7990 // Find the largest store unit
7991 MVT StoreType = MVT::i8;
7992 for (unsigned tp = MVT::FIRST_INTEGER_VALUETYPE;
7993 tp < MVT::LAST_INTEGER_VALUETYPE; ++tp) {
7994 MVT Tp = (MVT::SimpleValueType)tp;
7995 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToEltSz)
7998 // Didn't find a legal store type.
7999 if (!TLI.isTypeLegal(StoreType))
8002 // Bitcast the original vector into a vector of store-size units
8003 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
8004 StoreType, VT.getSizeInBits()/EVT(StoreType).getSizeInBits());
8005 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
8006 SDValue ShuffWide = DAG.getNode(ISD::BITCAST, DL, StoreVecVT, Shuff);
8007 SmallVector<SDValue, 8> Chains;
8008 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits()/8,
8009 TLI.getPointerTy());
8010 SDValue BasePtr = St->getBasePtr();
8012 // Perform one or more big stores into memory.
8013 unsigned E = (ToEltSz*NumElems)/StoreType.getSizeInBits();
8014 for (unsigned I = 0; I < E; I++) {
8015 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
8016 StoreType, ShuffWide,
8017 DAG.getIntPtrConstant(I));
8018 SDValue Ch = DAG.getStore(St->getChain(), DL, SubVec, BasePtr,
8019 St->getPointerInfo(), St->isVolatile(),
8020 St->isNonTemporal(), St->getAlignment());
8021 BasePtr = DAG.getNode(ISD::ADD, DL, BasePtr.getValueType(), BasePtr,
8023 Chains.push_back(Ch);
8025 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, &Chains[0],
8029 if (!ISD::isNormalStore(St))
8032 // Split a store of a VMOVDRR into two integer stores to avoid mixing NEON and
8033 // ARM stores of arguments in the same cache line.
8034 if (StVal.getNode()->getOpcode() == ARMISD::VMOVDRR &&
8035 StVal.getNode()->hasOneUse()) {
8036 SelectionDAG &DAG = DCI.DAG;
8037 DebugLoc DL = St->getDebugLoc();
8038 SDValue BasePtr = St->getBasePtr();
8039 SDValue NewST1 = DAG.getStore(St->getChain(), DL,
8040 StVal.getNode()->getOperand(0), BasePtr,
8041 St->getPointerInfo(), St->isVolatile(),
8042 St->isNonTemporal(), St->getAlignment());
8044 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
8045 DAG.getConstant(4, MVT::i32));
8046 return DAG.getStore(NewST1.getValue(0), DL, StVal.getNode()->getOperand(1),
8047 OffsetPtr, St->getPointerInfo(), St->isVolatile(),
8048 St->isNonTemporal(),
8049 std::min(4U, St->getAlignment() / 2));
8052 if (StVal.getValueType() != MVT::i64 ||
8053 StVal.getNode()->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
8056 // Bitcast an i64 store extracted from a vector to f64.
8057 // Otherwise, the i64 value will be legalized to a pair of i32 values.
8058 SelectionDAG &DAG = DCI.DAG;
8059 DebugLoc dl = StVal.getDebugLoc();
8060 SDValue IntVec = StVal.getOperand(0);
8061 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
8062 IntVec.getValueType().getVectorNumElements());
8063 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, IntVec);
8064 SDValue ExtElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
8065 Vec, StVal.getOperand(1));
8066 dl = N->getDebugLoc();
8067 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ExtElt);
8068 // Make the DAGCombiner fold the bitcasts.
8069 DCI.AddToWorklist(Vec.getNode());
8070 DCI.AddToWorklist(ExtElt.getNode());
8071 DCI.AddToWorklist(V.getNode());
8072 return DAG.getStore(St->getChain(), dl, V, St->getBasePtr(),
8073 St->getPointerInfo(), St->isVolatile(),
8074 St->isNonTemporal(), St->getAlignment(),
8078 /// hasNormalLoadOperand - Check if any of the operands of a BUILD_VECTOR node
8079 /// are normal, non-volatile loads. If so, it is profitable to bitcast an
8080 /// i64 vector to have f64 elements, since the value can then be loaded
8081 /// directly into a VFP register.
8082 static bool hasNormalLoadOperand(SDNode *N) {
8083 unsigned NumElts = N->getValueType(0).getVectorNumElements();
8084 for (unsigned i = 0; i < NumElts; ++i) {
8085 SDNode *Elt = N->getOperand(i).getNode();
8086 if (ISD::isNormalLoad(Elt) && !cast<LoadSDNode>(Elt)->isVolatile())
8092 /// PerformBUILD_VECTORCombine - Target-specific dag combine xforms for
8093 /// ISD::BUILD_VECTOR.
8094 static SDValue PerformBUILD_VECTORCombine(SDNode *N,
8095 TargetLowering::DAGCombinerInfo &DCI){
8096 // build_vector(N=ARMISD::VMOVRRD(X), N:1) -> bit_convert(X):
8097 // VMOVRRD is introduced when legalizing i64 types. It forces the i64 value
8098 // into a pair of GPRs, which is fine when the value is used as a scalar,
8099 // but if the i64 value is converted to a vector, we need to undo the VMOVRRD.
8100 SelectionDAG &DAG = DCI.DAG;
8101 if (N->getNumOperands() == 2) {
8102 SDValue RV = PerformVMOVDRRCombine(N, DAG);
8107 // Load i64 elements as f64 values so that type legalization does not split
8108 // them up into i32 values.
8109 EVT VT = N->getValueType(0);
8110 if (VT.getVectorElementType() != MVT::i64 || !hasNormalLoadOperand(N))
8112 DebugLoc dl = N->getDebugLoc();
8113 SmallVector<SDValue, 8> Ops;
8114 unsigned NumElts = VT.getVectorNumElements();
8115 for (unsigned i = 0; i < NumElts; ++i) {
8116 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(i));
8118 // Make the DAGCombiner fold the bitcast.
8119 DCI.AddToWorklist(V.getNode());
8121 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, NumElts);
8122 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, FloatVT, Ops.data(), NumElts);
8123 return DAG.getNode(ISD::BITCAST, dl, VT, BV);
8126 /// PerformInsertEltCombine - Target-specific dag combine xforms for
8127 /// ISD::INSERT_VECTOR_ELT.
8128 static SDValue PerformInsertEltCombine(SDNode *N,
8129 TargetLowering::DAGCombinerInfo &DCI) {
8130 // Bitcast an i64 load inserted into a vector to f64.
8131 // Otherwise, the i64 value will be legalized to a pair of i32 values.
8132 EVT VT = N->getValueType(0);
8133 SDNode *Elt = N->getOperand(1).getNode();
8134 if (VT.getVectorElementType() != MVT::i64 ||
8135 !ISD::isNormalLoad(Elt) || cast<LoadSDNode>(Elt)->isVolatile())
8138 SelectionDAG &DAG = DCI.DAG;
8139 DebugLoc dl = N->getDebugLoc();
8140 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
8141 VT.getVectorNumElements());
8142 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, N->getOperand(0));
8143 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(1));
8144 // Make the DAGCombiner fold the bitcasts.
8145 DCI.AddToWorklist(Vec.getNode());
8146 DCI.AddToWorklist(V.getNode());
8147 SDValue InsElt = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, FloatVT,
8148 Vec, V, N->getOperand(2));
8149 return DAG.getNode(ISD::BITCAST, dl, VT, InsElt);
8152 /// PerformVECTOR_SHUFFLECombine - Target-specific dag combine xforms for
8153 /// ISD::VECTOR_SHUFFLE.
8154 static SDValue PerformVECTOR_SHUFFLECombine(SDNode *N, SelectionDAG &DAG) {
8155 // The LLVM shufflevector instruction does not require the shuffle mask
8156 // length to match the operand vector length, but ISD::VECTOR_SHUFFLE does
8157 // have that requirement. When translating to ISD::VECTOR_SHUFFLE, if the
8158 // operands do not match the mask length, they are extended by concatenating
8159 // them with undef vectors. That is probably the right thing for other
8160 // targets, but for NEON it is better to concatenate two double-register
8161 // size vector operands into a single quad-register size vector. Do that
8162 // transformation here:
8163 // shuffle(concat(v1, undef), concat(v2, undef)) ->
8164 // shuffle(concat(v1, v2), undef)
8165 SDValue Op0 = N->getOperand(0);
8166 SDValue Op1 = N->getOperand(1);
8167 if (Op0.getOpcode() != ISD::CONCAT_VECTORS ||
8168 Op1.getOpcode() != ISD::CONCAT_VECTORS ||
8169 Op0.getNumOperands() != 2 ||
8170 Op1.getNumOperands() != 2)
8172 SDValue Concat0Op1 = Op0.getOperand(1);
8173 SDValue Concat1Op1 = Op1.getOperand(1);
8174 if (Concat0Op1.getOpcode() != ISD::UNDEF ||
8175 Concat1Op1.getOpcode() != ISD::UNDEF)
8177 // Skip the transformation if any of the types are illegal.
8178 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8179 EVT VT = N->getValueType(0);
8180 if (!TLI.isTypeLegal(VT) ||
8181 !TLI.isTypeLegal(Concat0Op1.getValueType()) ||
8182 !TLI.isTypeLegal(Concat1Op1.getValueType()))
8185 SDValue NewConcat = DAG.getNode(ISD::CONCAT_VECTORS, N->getDebugLoc(), VT,
8186 Op0.getOperand(0), Op1.getOperand(0));
8187 // Translate the shuffle mask.
8188 SmallVector<int, 16> NewMask;
8189 unsigned NumElts = VT.getVectorNumElements();
8190 unsigned HalfElts = NumElts/2;
8191 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
8192 for (unsigned n = 0; n < NumElts; ++n) {
8193 int MaskElt = SVN->getMaskElt(n);
8195 if (MaskElt < (int)HalfElts)
8197 else if (MaskElt >= (int)NumElts && MaskElt < (int)(NumElts + HalfElts))
8198 NewElt = HalfElts + MaskElt - NumElts;
8199 NewMask.push_back(NewElt);
8201 return DAG.getVectorShuffle(VT, N->getDebugLoc(), NewConcat,
8202 DAG.getUNDEF(VT), NewMask.data());
8205 /// CombineBaseUpdate - Target-specific DAG combine function for VLDDUP and
8206 /// NEON load/store intrinsics to merge base address updates.
8207 static SDValue CombineBaseUpdate(SDNode *N,
8208 TargetLowering::DAGCombinerInfo &DCI) {
8209 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
8212 SelectionDAG &DAG = DCI.DAG;
8213 bool isIntrinsic = (N->getOpcode() == ISD::INTRINSIC_VOID ||
8214 N->getOpcode() == ISD::INTRINSIC_W_CHAIN);
8215 unsigned AddrOpIdx = (isIntrinsic ? 2 : 1);
8216 SDValue Addr = N->getOperand(AddrOpIdx);
8218 // Search for a use of the address operand that is an increment.
8219 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(),
8220 UE = Addr.getNode()->use_end(); UI != UE; ++UI) {
8222 if (User->getOpcode() != ISD::ADD ||
8223 UI.getUse().getResNo() != Addr.getResNo())
8226 // Check that the add is independent of the load/store. Otherwise, folding
8227 // it would create a cycle.
8228 if (User->isPredecessorOf(N) || N->isPredecessorOf(User))
8231 // Find the new opcode for the updating load/store.
8233 bool isLaneOp = false;
8234 unsigned NewOpc = 0;
8235 unsigned NumVecs = 0;
8237 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
8239 default: llvm_unreachable("unexpected intrinsic for Neon base update");
8240 case Intrinsic::arm_neon_vld1: NewOpc = ARMISD::VLD1_UPD;
8242 case Intrinsic::arm_neon_vld2: NewOpc = ARMISD::VLD2_UPD;
8244 case Intrinsic::arm_neon_vld3: NewOpc = ARMISD::VLD3_UPD;
8246 case Intrinsic::arm_neon_vld4: NewOpc = ARMISD::VLD4_UPD;
8248 case Intrinsic::arm_neon_vld2lane: NewOpc = ARMISD::VLD2LN_UPD;
8249 NumVecs = 2; isLaneOp = true; break;
8250 case Intrinsic::arm_neon_vld3lane: NewOpc = ARMISD::VLD3LN_UPD;
8251 NumVecs = 3; isLaneOp = true; break;
8252 case Intrinsic::arm_neon_vld4lane: NewOpc = ARMISD::VLD4LN_UPD;
8253 NumVecs = 4; isLaneOp = true; break;
8254 case Intrinsic::arm_neon_vst1: NewOpc = ARMISD::VST1_UPD;
8255 NumVecs = 1; isLoad = false; break;
8256 case Intrinsic::arm_neon_vst2: NewOpc = ARMISD::VST2_UPD;
8257 NumVecs = 2; isLoad = false; break;
8258 case Intrinsic::arm_neon_vst3: NewOpc = ARMISD::VST3_UPD;
8259 NumVecs = 3; isLoad = false; break;
8260 case Intrinsic::arm_neon_vst4: NewOpc = ARMISD::VST4_UPD;
8261 NumVecs = 4; isLoad = false; break;
8262 case Intrinsic::arm_neon_vst2lane: NewOpc = ARMISD::VST2LN_UPD;
8263 NumVecs = 2; isLoad = false; isLaneOp = true; break;
8264 case Intrinsic::arm_neon_vst3lane: NewOpc = ARMISD::VST3LN_UPD;
8265 NumVecs = 3; isLoad = false; isLaneOp = true; break;
8266 case Intrinsic::arm_neon_vst4lane: NewOpc = ARMISD::VST4LN_UPD;
8267 NumVecs = 4; isLoad = false; isLaneOp = true; break;
8271 switch (N->getOpcode()) {
8272 default: llvm_unreachable("unexpected opcode for Neon base update");
8273 case ARMISD::VLD2DUP: NewOpc = ARMISD::VLD2DUP_UPD; NumVecs = 2; break;
8274 case ARMISD::VLD3DUP: NewOpc = ARMISD::VLD3DUP_UPD; NumVecs = 3; break;
8275 case ARMISD::VLD4DUP: NewOpc = ARMISD::VLD4DUP_UPD; NumVecs = 4; break;
8279 // Find the size of memory referenced by the load/store.
8282 VecTy = N->getValueType(0);
8284 VecTy = N->getOperand(AddrOpIdx+1).getValueType();
8285 unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8;
8287 NumBytes /= VecTy.getVectorNumElements();
8289 // If the increment is a constant, it must match the memory ref size.
8290 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
8291 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
8292 uint64_t IncVal = CInc->getZExtValue();
8293 if (IncVal != NumBytes)
8295 } else if (NumBytes >= 3 * 16) {
8296 // VLD3/4 and VST3/4 for 128-bit vectors are implemented with two
8297 // separate instructions that make it harder to use a non-constant update.
8301 // Create the new updating load/store node.
8303 unsigned NumResultVecs = (isLoad ? NumVecs : 0);
8305 for (n = 0; n < NumResultVecs; ++n)
8307 Tys[n++] = MVT::i32;
8308 Tys[n] = MVT::Other;
8309 SDVTList SDTys = DAG.getVTList(Tys, NumResultVecs+2);
8310 SmallVector<SDValue, 8> Ops;
8311 Ops.push_back(N->getOperand(0)); // incoming chain
8312 Ops.push_back(N->getOperand(AddrOpIdx));
8314 for (unsigned i = AddrOpIdx + 1; i < N->getNumOperands(); ++i) {
8315 Ops.push_back(N->getOperand(i));
8317 MemIntrinsicSDNode *MemInt = cast<MemIntrinsicSDNode>(N);
8318 SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, N->getDebugLoc(), SDTys,
8319 Ops.data(), Ops.size(),
8320 MemInt->getMemoryVT(),
8321 MemInt->getMemOperand());
8324 std::vector<SDValue> NewResults;
8325 for (unsigned i = 0; i < NumResultVecs; ++i) {
8326 NewResults.push_back(SDValue(UpdN.getNode(), i));
8328 NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs+1)); // chain
8329 DCI.CombineTo(N, NewResults);
8330 DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs));
8337 /// CombineVLDDUP - For a VDUPLANE node N, check if its source operand is a
8338 /// vldN-lane (N > 1) intrinsic, and if all the other uses of that intrinsic
8339 /// are also VDUPLANEs. If so, combine them to a vldN-dup operation and
8341 static bool CombineVLDDUP(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
8342 SelectionDAG &DAG = DCI.DAG;
8343 EVT VT = N->getValueType(0);
8344 // vldN-dup instructions only support 64-bit vectors for N > 1.
8345 if (!VT.is64BitVector())
8348 // Check if the VDUPLANE operand is a vldN-dup intrinsic.
8349 SDNode *VLD = N->getOperand(0).getNode();
8350 if (VLD->getOpcode() != ISD::INTRINSIC_W_CHAIN)
8352 unsigned NumVecs = 0;
8353 unsigned NewOpc = 0;
8354 unsigned IntNo = cast<ConstantSDNode>(VLD->getOperand(1))->getZExtValue();
8355 if (IntNo == Intrinsic::arm_neon_vld2lane) {
8357 NewOpc = ARMISD::VLD2DUP;
8358 } else if (IntNo == Intrinsic::arm_neon_vld3lane) {
8360 NewOpc = ARMISD::VLD3DUP;
8361 } else if (IntNo == Intrinsic::arm_neon_vld4lane) {
8363 NewOpc = ARMISD::VLD4DUP;
8368 // First check that all the vldN-lane uses are VDUPLANEs and that the lane
8369 // numbers match the load.
8370 unsigned VLDLaneNo =
8371 cast<ConstantSDNode>(VLD->getOperand(NumVecs+3))->getZExtValue();
8372 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
8374 // Ignore uses of the chain result.
8375 if (UI.getUse().getResNo() == NumVecs)
8378 if (User->getOpcode() != ARMISD::VDUPLANE ||
8379 VLDLaneNo != cast<ConstantSDNode>(User->getOperand(1))->getZExtValue())
8383 // Create the vldN-dup node.
8386 for (n = 0; n < NumVecs; ++n)
8388 Tys[n] = MVT::Other;
8389 SDVTList SDTys = DAG.getVTList(Tys, NumVecs+1);
8390 SDValue Ops[] = { VLD->getOperand(0), VLD->getOperand(2) };
8391 MemIntrinsicSDNode *VLDMemInt = cast<MemIntrinsicSDNode>(VLD);
8392 SDValue VLDDup = DAG.getMemIntrinsicNode(NewOpc, VLD->getDebugLoc(), SDTys,
8393 Ops, 2, VLDMemInt->getMemoryVT(),
8394 VLDMemInt->getMemOperand());
8397 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
8399 unsigned ResNo = UI.getUse().getResNo();
8400 // Ignore uses of the chain result.
8401 if (ResNo == NumVecs)
8404 DCI.CombineTo(User, SDValue(VLDDup.getNode(), ResNo));
8407 // Now the vldN-lane intrinsic is dead except for its chain result.
8408 // Update uses of the chain.
8409 std::vector<SDValue> VLDDupResults;
8410 for (unsigned n = 0; n < NumVecs; ++n)
8411 VLDDupResults.push_back(SDValue(VLDDup.getNode(), n));
8412 VLDDupResults.push_back(SDValue(VLDDup.getNode(), NumVecs));
8413 DCI.CombineTo(VLD, VLDDupResults);
8418 /// PerformVDUPLANECombine - Target-specific dag combine xforms for
8419 /// ARMISD::VDUPLANE.
8420 static SDValue PerformVDUPLANECombine(SDNode *N,
8421 TargetLowering::DAGCombinerInfo &DCI) {
8422 SDValue Op = N->getOperand(0);
8424 // If the source is a vldN-lane (N > 1) intrinsic, and all the other uses
8425 // of that intrinsic are also VDUPLANEs, combine them to a vldN-dup operation.
8426 if (CombineVLDDUP(N, DCI))
8427 return SDValue(N, 0);
8429 // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is
8430 // redundant. Ignore bit_converts for now; element sizes are checked below.
8431 while (Op.getOpcode() == ISD::BITCAST)
8432 Op = Op.getOperand(0);
8433 if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM)
8436 // Make sure the VMOV element size is not bigger than the VDUPLANE elements.
8437 unsigned EltSize = Op.getValueType().getVectorElementType().getSizeInBits();
8438 // The canonical VMOV for a zero vector uses a 32-bit element size.
8439 unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
8441 if (ARM_AM::decodeNEONModImm(Imm, EltBits) == 0)
8443 EVT VT = N->getValueType(0);
8444 if (EltSize > VT.getVectorElementType().getSizeInBits())
8447 return DCI.DAG.getNode(ISD::BITCAST, N->getDebugLoc(), VT, Op);
8450 // isConstVecPow2 - Return true if each vector element is a power of 2, all
8451 // elements are the same constant, C, and Log2(C) ranges from 1 to 32.
8452 static bool isConstVecPow2(SDValue ConstVec, bool isSigned, uint64_t &C)
8456 for (unsigned I = 0, E = ConstVec.getValueType().getVectorNumElements();
8458 ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(ConstVec.getOperand(I));
8463 APFloat APF = C->getValueAPF();
8464 if (APF.convertToInteger(&cN, 64, isSigned, APFloat::rmTowardZero, &isExact)
8465 != APFloat::opOK || !isExact)
8468 c0 = (I == 0) ? cN : c0;
8469 if (!isPowerOf2_64(cN) || c0 != cN || Log2_64(c0) < 1 || Log2_64(c0) > 32)
8476 /// PerformVCVTCombine - VCVT (floating-point to fixed-point, Advanced SIMD)
8477 /// can replace combinations of VMUL and VCVT (floating-point to integer)
8478 /// when the VMUL has a constant operand that is a power of 2.
8480 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
8481 /// vmul.f32 d16, d17, d16
8482 /// vcvt.s32.f32 d16, d16
8484 /// vcvt.s32.f32 d16, d16, #3
8485 static SDValue PerformVCVTCombine(SDNode *N,
8486 TargetLowering::DAGCombinerInfo &DCI,
8487 const ARMSubtarget *Subtarget) {
8488 SelectionDAG &DAG = DCI.DAG;
8489 SDValue Op = N->getOperand(0);
8491 if (!Subtarget->hasNEON() || !Op.getValueType().isVector() ||
8492 Op.getOpcode() != ISD::FMUL)
8496 SDValue N0 = Op->getOperand(0);
8497 SDValue ConstVec = Op->getOperand(1);
8498 bool isSigned = N->getOpcode() == ISD::FP_TO_SINT;
8500 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
8501 !isConstVecPow2(ConstVec, isSigned, C))
8504 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfp2fxs :
8505 Intrinsic::arm_neon_vcvtfp2fxu;
8506 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(),
8508 DAG.getConstant(IntrinsicOpcode, MVT::i32), N0,
8509 DAG.getConstant(Log2_64(C), MVT::i32));
8512 /// PerformVDIVCombine - VCVT (fixed-point to floating-point, Advanced SIMD)
8513 /// can replace combinations of VCVT (integer to floating-point) and VDIV
8514 /// when the VDIV has a constant operand that is a power of 2.
8516 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
8517 /// vcvt.f32.s32 d16, d16
8518 /// vdiv.f32 d16, d17, d16
8520 /// vcvt.f32.s32 d16, d16, #3
8521 static SDValue PerformVDIVCombine(SDNode *N,
8522 TargetLowering::DAGCombinerInfo &DCI,
8523 const ARMSubtarget *Subtarget) {
8524 SelectionDAG &DAG = DCI.DAG;
8525 SDValue Op = N->getOperand(0);
8526 unsigned OpOpcode = Op.getNode()->getOpcode();
8528 if (!Subtarget->hasNEON() || !N->getValueType(0).isVector() ||
8529 (OpOpcode != ISD::SINT_TO_FP && OpOpcode != ISD::UINT_TO_FP))
8533 SDValue ConstVec = N->getOperand(1);
8534 bool isSigned = OpOpcode == ISD::SINT_TO_FP;
8536 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
8537 !isConstVecPow2(ConstVec, isSigned, C))
8540 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfxs2fp :
8541 Intrinsic::arm_neon_vcvtfxu2fp;
8542 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(),
8544 DAG.getConstant(IntrinsicOpcode, MVT::i32),
8545 Op.getOperand(0), DAG.getConstant(Log2_64(C), MVT::i32));
8548 /// Getvshiftimm - Check if this is a valid build_vector for the immediate
8549 /// operand of a vector shift operation, where all the elements of the
8550 /// build_vector must have the same constant integer value.
8551 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
8552 // Ignore bit_converts.
8553 while (Op.getOpcode() == ISD::BITCAST)
8554 Op = Op.getOperand(0);
8555 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
8556 APInt SplatBits, SplatUndef;
8557 unsigned SplatBitSize;
8559 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
8560 HasAnyUndefs, ElementBits) ||
8561 SplatBitSize > ElementBits)
8563 Cnt = SplatBits.getSExtValue();
8567 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
8568 /// operand of a vector shift left operation. That value must be in the range:
8569 /// 0 <= Value < ElementBits for a left shift; or
8570 /// 0 <= Value <= ElementBits for a long left shift.
8571 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
8572 assert(VT.isVector() && "vector shift count is not a vector type");
8573 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
8574 if (! getVShiftImm(Op, ElementBits, Cnt))
8576 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
8579 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
8580 /// operand of a vector shift right operation. For a shift opcode, the value
8581 /// is positive, but for an intrinsic the value count must be negative. The
8582 /// absolute value must be in the range:
8583 /// 1 <= |Value| <= ElementBits for a right shift; or
8584 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
8585 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
8587 assert(VT.isVector() && "vector shift count is not a vector type");
8588 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
8589 if (! getVShiftImm(Op, ElementBits, Cnt))
8593 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
8596 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
8597 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
8598 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
8601 // Don't do anything for most intrinsics.
8604 // Vector shifts: check for immediate versions and lower them.
8605 // Note: This is done during DAG combining instead of DAG legalizing because
8606 // the build_vectors for 64-bit vector element shift counts are generally
8607 // not legal, and it is hard to see their values after they get legalized to
8608 // loads from a constant pool.
8609 case Intrinsic::arm_neon_vshifts:
8610 case Intrinsic::arm_neon_vshiftu:
8611 case Intrinsic::arm_neon_vshiftls:
8612 case Intrinsic::arm_neon_vshiftlu:
8613 case Intrinsic::arm_neon_vshiftn:
8614 case Intrinsic::arm_neon_vrshifts:
8615 case Intrinsic::arm_neon_vrshiftu:
8616 case Intrinsic::arm_neon_vrshiftn:
8617 case Intrinsic::arm_neon_vqshifts:
8618 case Intrinsic::arm_neon_vqshiftu:
8619 case Intrinsic::arm_neon_vqshiftsu:
8620 case Intrinsic::arm_neon_vqshiftns:
8621 case Intrinsic::arm_neon_vqshiftnu:
8622 case Intrinsic::arm_neon_vqshiftnsu:
8623 case Intrinsic::arm_neon_vqrshiftns:
8624 case Intrinsic::arm_neon_vqrshiftnu:
8625 case Intrinsic::arm_neon_vqrshiftnsu: {
8626 EVT VT = N->getOperand(1).getValueType();
8628 unsigned VShiftOpc = 0;
8631 case Intrinsic::arm_neon_vshifts:
8632 case Intrinsic::arm_neon_vshiftu:
8633 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
8634 VShiftOpc = ARMISD::VSHL;
8637 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
8638 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
8639 ARMISD::VSHRs : ARMISD::VSHRu);
8644 case Intrinsic::arm_neon_vshiftls:
8645 case Intrinsic::arm_neon_vshiftlu:
8646 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
8648 llvm_unreachable("invalid shift count for vshll intrinsic");
8650 case Intrinsic::arm_neon_vrshifts:
8651 case Intrinsic::arm_neon_vrshiftu:
8652 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
8656 case Intrinsic::arm_neon_vqshifts:
8657 case Intrinsic::arm_neon_vqshiftu:
8658 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
8662 case Intrinsic::arm_neon_vqshiftsu:
8663 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
8665 llvm_unreachable("invalid shift count for vqshlu intrinsic");
8667 case Intrinsic::arm_neon_vshiftn:
8668 case Intrinsic::arm_neon_vrshiftn:
8669 case Intrinsic::arm_neon_vqshiftns:
8670 case Intrinsic::arm_neon_vqshiftnu:
8671 case Intrinsic::arm_neon_vqshiftnsu:
8672 case Intrinsic::arm_neon_vqrshiftns:
8673 case Intrinsic::arm_neon_vqrshiftnu:
8674 case Intrinsic::arm_neon_vqrshiftnsu:
8675 // Narrowing shifts require an immediate right shift.
8676 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
8678 llvm_unreachable("invalid shift count for narrowing vector shift "
8682 llvm_unreachable("unhandled vector shift");
8686 case Intrinsic::arm_neon_vshifts:
8687 case Intrinsic::arm_neon_vshiftu:
8688 // Opcode already set above.
8690 case Intrinsic::arm_neon_vshiftls:
8691 case Intrinsic::arm_neon_vshiftlu:
8692 if (Cnt == VT.getVectorElementType().getSizeInBits())
8693 VShiftOpc = ARMISD::VSHLLi;
8695 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
8696 ARMISD::VSHLLs : ARMISD::VSHLLu);
8698 case Intrinsic::arm_neon_vshiftn:
8699 VShiftOpc = ARMISD::VSHRN; break;
8700 case Intrinsic::arm_neon_vrshifts:
8701 VShiftOpc = ARMISD::VRSHRs; break;
8702 case Intrinsic::arm_neon_vrshiftu:
8703 VShiftOpc = ARMISD::VRSHRu; break;
8704 case Intrinsic::arm_neon_vrshiftn:
8705 VShiftOpc = ARMISD::VRSHRN; break;
8706 case Intrinsic::arm_neon_vqshifts:
8707 VShiftOpc = ARMISD::VQSHLs; break;
8708 case Intrinsic::arm_neon_vqshiftu:
8709 VShiftOpc = ARMISD::VQSHLu; break;
8710 case Intrinsic::arm_neon_vqshiftsu:
8711 VShiftOpc = ARMISD::VQSHLsu; break;
8712 case Intrinsic::arm_neon_vqshiftns:
8713 VShiftOpc = ARMISD::VQSHRNs; break;
8714 case Intrinsic::arm_neon_vqshiftnu:
8715 VShiftOpc = ARMISD::VQSHRNu; break;
8716 case Intrinsic::arm_neon_vqshiftnsu:
8717 VShiftOpc = ARMISD::VQSHRNsu; break;
8718 case Intrinsic::arm_neon_vqrshiftns:
8719 VShiftOpc = ARMISD::VQRSHRNs; break;
8720 case Intrinsic::arm_neon_vqrshiftnu:
8721 VShiftOpc = ARMISD::VQRSHRNu; break;
8722 case Intrinsic::arm_neon_vqrshiftnsu:
8723 VShiftOpc = ARMISD::VQRSHRNsu; break;
8726 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
8727 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
8730 case Intrinsic::arm_neon_vshiftins: {
8731 EVT VT = N->getOperand(1).getValueType();
8733 unsigned VShiftOpc = 0;
8735 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
8736 VShiftOpc = ARMISD::VSLI;
8737 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
8738 VShiftOpc = ARMISD::VSRI;
8740 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
8743 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
8744 N->getOperand(1), N->getOperand(2),
8745 DAG.getConstant(Cnt, MVT::i32));
8748 case Intrinsic::arm_neon_vqrshifts:
8749 case Intrinsic::arm_neon_vqrshiftu:
8750 // No immediate versions of these to check for.
8757 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
8758 /// lowers them. As with the vector shift intrinsics, this is done during DAG
8759 /// combining instead of DAG legalizing because the build_vectors for 64-bit
8760 /// vector element shift counts are generally not legal, and it is hard to see
8761 /// their values after they get legalized to loads from a constant pool.
8762 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
8763 const ARMSubtarget *ST) {
8764 EVT VT = N->getValueType(0);
8765 if (N->getOpcode() == ISD::SRL && VT == MVT::i32 && ST->hasV6Ops()) {
8766 // Canonicalize (srl (bswap x), 16) to (rotr (bswap x), 16) if the high
8767 // 16-bits of x is zero. This optimizes rev + lsr 16 to rev16.
8768 SDValue N1 = N->getOperand(1);
8769 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) {
8770 SDValue N0 = N->getOperand(0);
8771 if (C->getZExtValue() == 16 && N0.getOpcode() == ISD::BSWAP &&
8772 DAG.MaskedValueIsZero(N0.getOperand(0),
8773 APInt::getHighBitsSet(32, 16)))
8774 return DAG.getNode(ISD::ROTR, N->getDebugLoc(), VT, N0, N1);
8778 // Nothing to be done for scalar shifts.
8779 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8780 if (!VT.isVector() || !TLI.isTypeLegal(VT))
8783 assert(ST->hasNEON() && "unexpected vector shift");
8786 switch (N->getOpcode()) {
8787 default: llvm_unreachable("unexpected shift opcode");
8790 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
8791 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
8792 DAG.getConstant(Cnt, MVT::i32));
8797 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
8798 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
8799 ARMISD::VSHRs : ARMISD::VSHRu);
8800 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
8801 DAG.getConstant(Cnt, MVT::i32));
8807 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
8808 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
8809 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
8810 const ARMSubtarget *ST) {
8811 SDValue N0 = N->getOperand(0);
8813 // Check for sign- and zero-extensions of vector extract operations of 8-
8814 // and 16-bit vector elements. NEON supports these directly. They are
8815 // handled during DAG combining because type legalization will promote them
8816 // to 32-bit types and it is messy to recognize the operations after that.
8817 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
8818 SDValue Vec = N0.getOperand(0);
8819 SDValue Lane = N0.getOperand(1);
8820 EVT VT = N->getValueType(0);
8821 EVT EltVT = N0.getValueType();
8822 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8824 if (VT == MVT::i32 &&
8825 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
8826 TLI.isTypeLegal(Vec.getValueType()) &&
8827 isa<ConstantSDNode>(Lane)) {
8830 switch (N->getOpcode()) {
8831 default: llvm_unreachable("unexpected opcode");
8832 case ISD::SIGN_EXTEND:
8833 Opc = ARMISD::VGETLANEs;
8835 case ISD::ZERO_EXTEND:
8836 case ISD::ANY_EXTEND:
8837 Opc = ARMISD::VGETLANEu;
8840 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
8847 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
8848 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
8849 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
8850 const ARMSubtarget *ST) {
8851 // If the target supports NEON, try to use vmax/vmin instructions for f32
8852 // selects like "x < y ? x : y". Unless the NoNaNsFPMath option is set,
8853 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
8854 // a NaN; only do the transformation when it matches that behavior.
8856 // For now only do this when using NEON for FP operations; if using VFP, it
8857 // is not obvious that the benefit outweighs the cost of switching to the
8859 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
8860 N->getValueType(0) != MVT::f32)
8863 SDValue CondLHS = N->getOperand(0);
8864 SDValue CondRHS = N->getOperand(1);
8865 SDValue LHS = N->getOperand(2);
8866 SDValue RHS = N->getOperand(3);
8867 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
8869 unsigned Opcode = 0;
8871 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
8872 IsReversed = false; // x CC y ? x : y
8873 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
8874 IsReversed = true ; // x CC y ? y : x
8888 // If LHS is NaN, an ordered comparison will be false and the result will
8889 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
8890 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
8891 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
8892 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
8894 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
8895 // will return -0, so vmin can only be used for unsafe math or if one of
8896 // the operands is known to be nonzero.
8897 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
8898 !DAG.getTarget().Options.UnsafeFPMath &&
8899 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
8901 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
8910 // If LHS is NaN, an ordered comparison will be false and the result will
8911 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
8912 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
8913 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
8914 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
8916 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
8917 // will return +0, so vmax can only be used for unsafe math or if one of
8918 // the operands is known to be nonzero.
8919 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
8920 !DAG.getTarget().Options.UnsafeFPMath &&
8921 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
8923 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
8929 return DAG.getNode(Opcode, N->getDebugLoc(), N->getValueType(0), LHS, RHS);
8932 /// PerformCMOVCombine - Target-specific DAG combining for ARMISD::CMOV.
8934 ARMTargetLowering::PerformCMOVCombine(SDNode *N, SelectionDAG &DAG) const {
8935 SDValue Cmp = N->getOperand(4);
8936 if (Cmp.getOpcode() != ARMISD::CMPZ)
8937 // Only looking at EQ and NE cases.
8940 EVT VT = N->getValueType(0);
8941 DebugLoc dl = N->getDebugLoc();
8942 SDValue LHS = Cmp.getOperand(0);
8943 SDValue RHS = Cmp.getOperand(1);
8944 SDValue FalseVal = N->getOperand(0);
8945 SDValue TrueVal = N->getOperand(1);
8946 SDValue ARMcc = N->getOperand(2);
8947 ARMCC::CondCodes CC =
8948 (ARMCC::CondCodes)cast<ConstantSDNode>(ARMcc)->getZExtValue();
8966 /// FIXME: Turn this into a target neutral optimization?
8968 if (CC == ARMCC::NE && FalseVal == RHS && FalseVal != LHS) {
8969 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, TrueVal, ARMcc,
8970 N->getOperand(3), Cmp);
8971 } else if (CC == ARMCC::EQ && TrueVal == RHS) {
8973 SDValue NewCmp = getARMCmp(LHS, RHS, ISD::SETNE, ARMcc, DAG, dl);
8974 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, FalseVal, ARMcc,
8975 N->getOperand(3), NewCmp);
8978 if (Res.getNode()) {
8979 APInt KnownZero, KnownOne;
8980 DAG.ComputeMaskedBits(SDValue(N,0), KnownZero, KnownOne);
8981 // Capture demanded bits information that would be otherwise lost.
8982 if (KnownZero == 0xfffffffe)
8983 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
8984 DAG.getValueType(MVT::i1));
8985 else if (KnownZero == 0xffffff00)
8986 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
8987 DAG.getValueType(MVT::i8));
8988 else if (KnownZero == 0xffff0000)
8989 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
8990 DAG.getValueType(MVT::i16));
8996 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
8997 DAGCombinerInfo &DCI) const {
8998 switch (N->getOpcode()) {
9000 case ISD::ADDC: return PerformADDCCombine(N, DCI, Subtarget);
9001 case ISD::ADD: return PerformADDCombine(N, DCI, Subtarget);
9002 case ISD::SUB: return PerformSUBCombine(N, DCI);
9003 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
9004 case ISD::OR: return PerformORCombine(N, DCI, Subtarget);
9005 case ISD::XOR: return PerformXORCombine(N, DCI, Subtarget);
9006 case ISD::AND: return PerformANDCombine(N, DCI, Subtarget);
9007 case ARMISD::BFI: return PerformBFICombine(N, DCI);
9008 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
9009 case ARMISD::VMOVDRR: return PerformVMOVDRRCombine(N, DCI.DAG);
9010 case ISD::STORE: return PerformSTORECombine(N, DCI);
9011 case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DCI);
9012 case ISD::INSERT_VECTOR_ELT: return PerformInsertEltCombine(N, DCI);
9013 case ISD::VECTOR_SHUFFLE: return PerformVECTOR_SHUFFLECombine(N, DCI.DAG);
9014 case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI);
9015 case ISD::FP_TO_SINT:
9016 case ISD::FP_TO_UINT: return PerformVCVTCombine(N, DCI, Subtarget);
9017 case ISD::FDIV: return PerformVDIVCombine(N, DCI, Subtarget);
9018 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
9021 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
9022 case ISD::SIGN_EXTEND:
9023 case ISD::ZERO_EXTEND:
9024 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
9025 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
9026 case ARMISD::CMOV: return PerformCMOVCombine(N, DCI.DAG);
9027 case ARMISD::VLD2DUP:
9028 case ARMISD::VLD3DUP:
9029 case ARMISD::VLD4DUP:
9030 return CombineBaseUpdate(N, DCI);
9031 case ISD::INTRINSIC_VOID:
9032 case ISD::INTRINSIC_W_CHAIN:
9033 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
9034 case Intrinsic::arm_neon_vld1:
9035 case Intrinsic::arm_neon_vld2:
9036 case Intrinsic::arm_neon_vld3:
9037 case Intrinsic::arm_neon_vld4:
9038 case Intrinsic::arm_neon_vld2lane:
9039 case Intrinsic::arm_neon_vld3lane:
9040 case Intrinsic::arm_neon_vld4lane:
9041 case Intrinsic::arm_neon_vst1:
9042 case Intrinsic::arm_neon_vst2:
9043 case Intrinsic::arm_neon_vst3:
9044 case Intrinsic::arm_neon_vst4:
9045 case Intrinsic::arm_neon_vst2lane:
9046 case Intrinsic::arm_neon_vst3lane:
9047 case Intrinsic::arm_neon_vst4lane:
9048 return CombineBaseUpdate(N, DCI);
9056 bool ARMTargetLowering::isDesirableToTransformToIntegerOp(unsigned Opc,
9058 return (VT == MVT::f32) && (Opc == ISD::LOAD || Opc == ISD::STORE);
9061 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
9062 // The AllowsUnaliged flag models the SCTLR.A setting in ARM cpus
9063 bool AllowsUnaligned = Subtarget->allowsUnalignedMem();
9065 switch (VT.getSimpleVT().SimpleTy) {
9071 // Unaligned access can use (for example) LRDB, LRDH, LDR
9072 return AllowsUnaligned;
9075 // For any little-endian targets with neon, we can support unaligned ld/st
9076 // of D and Q (e.g. {D0,D1}) registers by using vld1.i8/vst1.i8.
9077 // A big-endian target may also explictly support unaligned accesses
9078 return Subtarget->hasNEON() && (AllowsUnaligned || isLittleEndian());
9082 static bool memOpAlign(unsigned DstAlign, unsigned SrcAlign,
9083 unsigned AlignCheck) {
9084 return ((SrcAlign == 0 || SrcAlign % AlignCheck == 0) &&
9085 (DstAlign == 0 || DstAlign % AlignCheck == 0));
9088 EVT ARMTargetLowering::getOptimalMemOpType(uint64_t Size,
9089 unsigned DstAlign, unsigned SrcAlign,
9092 MachineFunction &MF) const {
9093 const Function *F = MF.getFunction();
9095 // See if we can use NEON instructions for this...
9097 !F->getFnAttributes().hasAttribute(Attributes::NoImplicitFloat) &&
9098 Subtarget->hasNEON()) {
9099 if (memOpAlign(SrcAlign, DstAlign, 16) && Size >= 16) {
9101 } else if (memOpAlign(SrcAlign, DstAlign, 8) && Size >= 8) {
9106 // Lowering to i32/i16 if the size permits.
9109 } else if (Size >= 2) {
9113 // Let the target-independent logic figure it out.
9117 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
9122 switch (VT.getSimpleVT().SimpleTy) {
9123 default: return false;
9138 if ((V & (Scale - 1)) != 0)
9141 return V == (V & ((1LL << 5) - 1));
9144 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
9145 const ARMSubtarget *Subtarget) {
9152 switch (VT.getSimpleVT().SimpleTy) {
9153 default: return false;
9158 // + imm12 or - imm8
9160 return V == (V & ((1LL << 8) - 1));
9161 return V == (V & ((1LL << 12) - 1));
9164 // Same as ARM mode. FIXME: NEON?
9165 if (!Subtarget->hasVFP2())
9170 return V == (V & ((1LL << 8) - 1));
9174 /// isLegalAddressImmediate - Return true if the integer value can be used
9175 /// as the offset of the target addressing mode for load / store of the
9177 static bool isLegalAddressImmediate(int64_t V, EVT VT,
9178 const ARMSubtarget *Subtarget) {
9185 if (Subtarget->isThumb1Only())
9186 return isLegalT1AddressImmediate(V, VT);
9187 else if (Subtarget->isThumb2())
9188 return isLegalT2AddressImmediate(V, VT, Subtarget);
9193 switch (VT.getSimpleVT().SimpleTy) {
9194 default: return false;
9199 return V == (V & ((1LL << 12) - 1));
9202 return V == (V & ((1LL << 8) - 1));
9205 if (!Subtarget->hasVFP2()) // FIXME: NEON?
9210 return V == (V & ((1LL << 8) - 1));
9214 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
9216 int Scale = AM.Scale;
9220 switch (VT.getSimpleVT().SimpleTy) {
9221 default: return false;
9230 return Scale == 2 || Scale == 4 || Scale == 8;
9233 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
9237 // Note, we allow "void" uses (basically, uses that aren't loads or
9238 // stores), because arm allows folding a scale into many arithmetic
9239 // operations. This should be made more precise and revisited later.
9241 // Allow r << imm, but the imm has to be a multiple of two.
9242 if (Scale & 1) return false;
9243 return isPowerOf2_32(Scale);
9247 /// isLegalAddressingMode - Return true if the addressing mode represented
9248 /// by AM is legal for this target, for a load/store of the specified type.
9249 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
9251 EVT VT = getValueType(Ty, true);
9252 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
9255 // Can never fold addr of global into load/store.
9260 case 0: // no scale reg, must be "r+i" or "r", or "i".
9263 if (Subtarget->isThumb1Only())
9267 // ARM doesn't support any R+R*scale+imm addr modes.
9274 if (Subtarget->isThumb2())
9275 return isLegalT2ScaledAddressingMode(AM, VT);
9277 int Scale = AM.Scale;
9278 switch (VT.getSimpleVT().SimpleTy) {
9279 default: return false;
9283 if (Scale < 0) Scale = -Scale;
9287 return isPowerOf2_32(Scale & ~1);
9291 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
9296 // Note, we allow "void" uses (basically, uses that aren't loads or
9297 // stores), because arm allows folding a scale into many arithmetic
9298 // operations. This should be made more precise and revisited later.
9300 // Allow r << imm, but the imm has to be a multiple of two.
9301 if (Scale & 1) return false;
9302 return isPowerOf2_32(Scale);
9308 /// isLegalICmpImmediate - Return true if the specified immediate is legal
9309 /// icmp immediate, that is the target has icmp instructions which can compare
9310 /// a register against the immediate without having to materialize the
9311 /// immediate into a register.
9312 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
9313 // Thumb2 and ARM modes can use cmn for negative immediates.
9314 if (!Subtarget->isThumb())
9315 return ARM_AM::getSOImmVal(llvm::abs64(Imm)) != -1;
9316 if (Subtarget->isThumb2())
9317 return ARM_AM::getT2SOImmVal(llvm::abs64(Imm)) != -1;
9318 // Thumb1 doesn't have cmn, and only 8-bit immediates.
9319 return Imm >= 0 && Imm <= 255;
9322 /// isLegalAddImmediate - Return true if the specified immediate is a legal add
9323 /// *or sub* immediate, that is the target has add or sub instructions which can
9324 /// add a register with the immediate without having to materialize the
9325 /// immediate into a register.
9326 bool ARMTargetLowering::isLegalAddImmediate(int64_t Imm) const {
9327 // Same encoding for add/sub, just flip the sign.
9328 int64_t AbsImm = llvm::abs64(Imm);
9329 if (!Subtarget->isThumb())
9330 return ARM_AM::getSOImmVal(AbsImm) != -1;
9331 if (Subtarget->isThumb2())
9332 return ARM_AM::getT2SOImmVal(AbsImm) != -1;
9333 // Thumb1 only has 8-bit unsigned immediate.
9334 return AbsImm >= 0 && AbsImm <= 255;
9337 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
9338 bool isSEXTLoad, SDValue &Base,
9339 SDValue &Offset, bool &isInc,
9340 SelectionDAG &DAG) {
9341 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
9344 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
9346 Base = Ptr->getOperand(0);
9347 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
9348 int RHSC = (int)RHS->getZExtValue();
9349 if (RHSC < 0 && RHSC > -256) {
9350 assert(Ptr->getOpcode() == ISD::ADD);
9352 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
9356 isInc = (Ptr->getOpcode() == ISD::ADD);
9357 Offset = Ptr->getOperand(1);
9359 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
9361 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
9362 int RHSC = (int)RHS->getZExtValue();
9363 if (RHSC < 0 && RHSC > -0x1000) {
9364 assert(Ptr->getOpcode() == ISD::ADD);
9366 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
9367 Base = Ptr->getOperand(0);
9372 if (Ptr->getOpcode() == ISD::ADD) {
9374 ARM_AM::ShiftOpc ShOpcVal=
9375 ARM_AM::getShiftOpcForNode(Ptr->getOperand(0).getOpcode());
9376 if (ShOpcVal != ARM_AM::no_shift) {
9377 Base = Ptr->getOperand(1);
9378 Offset = Ptr->getOperand(0);
9380 Base = Ptr->getOperand(0);
9381 Offset = Ptr->getOperand(1);
9386 isInc = (Ptr->getOpcode() == ISD::ADD);
9387 Base = Ptr->getOperand(0);
9388 Offset = Ptr->getOperand(1);
9392 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
9396 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
9397 bool isSEXTLoad, SDValue &Base,
9398 SDValue &Offset, bool &isInc,
9399 SelectionDAG &DAG) {
9400 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
9403 Base = Ptr->getOperand(0);
9404 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
9405 int RHSC = (int)RHS->getZExtValue();
9406 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
9407 assert(Ptr->getOpcode() == ISD::ADD);
9409 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
9411 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
9412 isInc = Ptr->getOpcode() == ISD::ADD;
9413 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
9421 /// getPreIndexedAddressParts - returns true by value, base pointer and
9422 /// offset pointer and addressing mode by reference if the node's address
9423 /// can be legally represented as pre-indexed load / store address.
9425 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
9427 ISD::MemIndexedMode &AM,
9428 SelectionDAG &DAG) const {
9429 if (Subtarget->isThumb1Only())
9434 bool isSEXTLoad = false;
9435 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
9436 Ptr = LD->getBasePtr();
9437 VT = LD->getMemoryVT();
9438 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
9439 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
9440 Ptr = ST->getBasePtr();
9441 VT = ST->getMemoryVT();
9446 bool isLegal = false;
9447 if (Subtarget->isThumb2())
9448 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
9449 Offset, isInc, DAG);
9451 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
9452 Offset, isInc, DAG);
9456 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
9460 /// getPostIndexedAddressParts - returns true by value, base pointer and
9461 /// offset pointer and addressing mode by reference if this node can be
9462 /// combined with a load / store to form a post-indexed load / store.
9463 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
9466 ISD::MemIndexedMode &AM,
9467 SelectionDAG &DAG) const {
9468 if (Subtarget->isThumb1Only())
9473 bool isSEXTLoad = false;
9474 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
9475 VT = LD->getMemoryVT();
9476 Ptr = LD->getBasePtr();
9477 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
9478 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
9479 VT = ST->getMemoryVT();
9480 Ptr = ST->getBasePtr();
9485 bool isLegal = false;
9486 if (Subtarget->isThumb2())
9487 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
9490 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
9496 // Swap base ptr and offset to catch more post-index load / store when
9497 // it's legal. In Thumb2 mode, offset must be an immediate.
9498 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
9499 !Subtarget->isThumb2())
9500 std::swap(Base, Offset);
9502 // Post-indexed load / store update the base pointer.
9507 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
9511 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
9514 const SelectionDAG &DAG,
9515 unsigned Depth) const {
9516 KnownZero = KnownOne = APInt(KnownOne.getBitWidth(), 0);
9517 switch (Op.getOpcode()) {
9519 case ARMISD::CMOV: {
9520 // Bits are known zero/one if known on the LHS and RHS.
9521 DAG.ComputeMaskedBits(Op.getOperand(0), KnownZero, KnownOne, Depth+1);
9522 if (KnownZero == 0 && KnownOne == 0) return;
9524 APInt KnownZeroRHS, KnownOneRHS;
9525 DAG.ComputeMaskedBits(Op.getOperand(1), KnownZeroRHS, KnownOneRHS, Depth+1);
9526 KnownZero &= KnownZeroRHS;
9527 KnownOne &= KnownOneRHS;
9533 //===----------------------------------------------------------------------===//
9534 // ARM Inline Assembly Support
9535 //===----------------------------------------------------------------------===//
9537 bool ARMTargetLowering::ExpandInlineAsm(CallInst *CI) const {
9538 // Looking for "rev" which is V6+.
9539 if (!Subtarget->hasV6Ops())
9542 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
9543 std::string AsmStr = IA->getAsmString();
9544 SmallVector<StringRef, 4> AsmPieces;
9545 SplitString(AsmStr, AsmPieces, ";\n");
9547 switch (AsmPieces.size()) {
9548 default: return false;
9550 AsmStr = AsmPieces[0];
9552 SplitString(AsmStr, AsmPieces, " \t,");
9555 if (AsmPieces.size() == 3 &&
9556 AsmPieces[0] == "rev" && AsmPieces[1] == "$0" && AsmPieces[2] == "$1" &&
9557 IA->getConstraintString().compare(0, 4, "=l,l") == 0) {
9558 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
9559 if (Ty && Ty->getBitWidth() == 32)
9560 return IntrinsicLowering::LowerToByteSwap(CI);
9568 /// getConstraintType - Given a constraint letter, return the type of
9569 /// constraint it is for this target.
9570 ARMTargetLowering::ConstraintType
9571 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
9572 if (Constraint.size() == 1) {
9573 switch (Constraint[0]) {
9575 case 'l': return C_RegisterClass;
9576 case 'w': return C_RegisterClass;
9577 case 'h': return C_RegisterClass;
9578 case 'x': return C_RegisterClass;
9579 case 't': return C_RegisterClass;
9580 case 'j': return C_Other; // Constant for movw.
9581 // An address with a single base register. Due to the way we
9582 // currently handle addresses it is the same as an 'r' memory constraint.
9583 case 'Q': return C_Memory;
9585 } else if (Constraint.size() == 2) {
9586 switch (Constraint[0]) {
9588 // All 'U+' constraints are addresses.
9589 case 'U': return C_Memory;
9592 return TargetLowering::getConstraintType(Constraint);
9595 /// Examine constraint type and operand type and determine a weight value.
9596 /// This object must already have been set up with the operand type
9597 /// and the current alternative constraint selected.
9598 TargetLowering::ConstraintWeight
9599 ARMTargetLowering::getSingleConstraintMatchWeight(
9600 AsmOperandInfo &info, const char *constraint) const {
9601 ConstraintWeight weight = CW_Invalid;
9602 Value *CallOperandVal = info.CallOperandVal;
9603 // If we don't have a value, we can't do a match,
9604 // but allow it at the lowest weight.
9605 if (CallOperandVal == NULL)
9607 Type *type = CallOperandVal->getType();
9608 // Look at the constraint type.
9609 switch (*constraint) {
9611 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
9614 if (type->isIntegerTy()) {
9615 if (Subtarget->isThumb())
9616 weight = CW_SpecificReg;
9618 weight = CW_Register;
9622 if (type->isFloatingPointTy())
9623 weight = CW_Register;
9629 typedef std::pair<unsigned, const TargetRegisterClass*> RCPair;
9631 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
9633 if (Constraint.size() == 1) {
9634 // GCC ARM Constraint Letters
9635 switch (Constraint[0]) {
9636 case 'l': // Low regs or general regs.
9637 if (Subtarget->isThumb())
9638 return RCPair(0U, &ARM::tGPRRegClass);
9639 return RCPair(0U, &ARM::GPRRegClass);
9640 case 'h': // High regs or no regs.
9641 if (Subtarget->isThumb())
9642 return RCPair(0U, &ARM::hGPRRegClass);
9645 return RCPair(0U, &ARM::GPRRegClass);
9648 return RCPair(0U, &ARM::SPRRegClass);
9649 if (VT.getSizeInBits() == 64)
9650 return RCPair(0U, &ARM::DPRRegClass);
9651 if (VT.getSizeInBits() == 128)
9652 return RCPair(0U, &ARM::QPRRegClass);
9656 return RCPair(0U, &ARM::SPR_8RegClass);
9657 if (VT.getSizeInBits() == 64)
9658 return RCPair(0U, &ARM::DPR_8RegClass);
9659 if (VT.getSizeInBits() == 128)
9660 return RCPair(0U, &ARM::QPR_8RegClass);
9664 return RCPair(0U, &ARM::SPRRegClass);
9668 if (StringRef("{cc}").equals_lower(Constraint))
9669 return std::make_pair(unsigned(ARM::CPSR), &ARM::CCRRegClass);
9671 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
9674 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
9675 /// vector. If it is invalid, don't add anything to Ops.
9676 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
9677 std::string &Constraint,
9678 std::vector<SDValue>&Ops,
9679 SelectionDAG &DAG) const {
9680 SDValue Result(0, 0);
9682 // Currently only support length 1 constraints.
9683 if (Constraint.length() != 1) return;
9685 char ConstraintLetter = Constraint[0];
9686 switch (ConstraintLetter) {
9689 case 'I': case 'J': case 'K': case 'L':
9690 case 'M': case 'N': case 'O':
9691 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
9695 int64_t CVal64 = C->getSExtValue();
9696 int CVal = (int) CVal64;
9697 // None of these constraints allow values larger than 32 bits. Check
9698 // that the value fits in an int.
9702 switch (ConstraintLetter) {
9704 // Constant suitable for movw, must be between 0 and
9706 if (Subtarget->hasV6T2Ops())
9707 if (CVal >= 0 && CVal <= 65535)
9711 if (Subtarget->isThumb1Only()) {
9712 // This must be a constant between 0 and 255, for ADD
9714 if (CVal >= 0 && CVal <= 255)
9716 } else if (Subtarget->isThumb2()) {
9717 // A constant that can be used as an immediate value in a
9718 // data-processing instruction.
9719 if (ARM_AM::getT2SOImmVal(CVal) != -1)
9722 // A constant that can be used as an immediate value in a
9723 // data-processing instruction.
9724 if (ARM_AM::getSOImmVal(CVal) != -1)
9730 if (Subtarget->isThumb()) { // FIXME thumb2
9731 // This must be a constant between -255 and -1, for negated ADD
9732 // immediates. This can be used in GCC with an "n" modifier that
9733 // prints the negated value, for use with SUB instructions. It is
9734 // not useful otherwise but is implemented for compatibility.
9735 if (CVal >= -255 && CVal <= -1)
9738 // This must be a constant between -4095 and 4095. It is not clear
9739 // what this constraint is intended for. Implemented for
9740 // compatibility with GCC.
9741 if (CVal >= -4095 && CVal <= 4095)
9747 if (Subtarget->isThumb1Only()) {
9748 // A 32-bit value where only one byte has a nonzero value. Exclude
9749 // zero to match GCC. This constraint is used by GCC internally for
9750 // constants that can be loaded with a move/shift combination.
9751 // It is not useful otherwise but is implemented for compatibility.
9752 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
9754 } else if (Subtarget->isThumb2()) {
9755 // A constant whose bitwise inverse can be used as an immediate
9756 // value in a data-processing instruction. This can be used in GCC
9757 // with a "B" modifier that prints the inverted value, for use with
9758 // BIC and MVN instructions. It is not useful otherwise but is
9759 // implemented for compatibility.
9760 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
9763 // A constant whose bitwise inverse can be used as an immediate
9764 // value in a data-processing instruction. This can be used in GCC
9765 // with a "B" modifier that prints the inverted value, for use with
9766 // BIC and MVN instructions. It is not useful otherwise but is
9767 // implemented for compatibility.
9768 if (ARM_AM::getSOImmVal(~CVal) != -1)
9774 if (Subtarget->isThumb1Only()) {
9775 // This must be a constant between -7 and 7,
9776 // for 3-operand ADD/SUB immediate instructions.
9777 if (CVal >= -7 && CVal < 7)
9779 } else if (Subtarget->isThumb2()) {
9780 // A constant whose negation can be used as an immediate value in a
9781 // data-processing instruction. This can be used in GCC with an "n"
9782 // modifier that prints the negated value, for use with SUB
9783 // instructions. It is not useful otherwise but is implemented for
9785 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
9788 // A constant whose negation can be used as an immediate value in a
9789 // data-processing instruction. This can be used in GCC with an "n"
9790 // modifier that prints the negated value, for use with SUB
9791 // instructions. It is not useful otherwise but is implemented for
9793 if (ARM_AM::getSOImmVal(-CVal) != -1)
9799 if (Subtarget->isThumb()) { // FIXME thumb2
9800 // This must be a multiple of 4 between 0 and 1020, for
9801 // ADD sp + immediate.
9802 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
9805 // A power of two or a constant between 0 and 32. This is used in
9806 // GCC for the shift amount on shifted register operands, but it is
9807 // useful in general for any shift amounts.
9808 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
9814 if (Subtarget->isThumb()) { // FIXME thumb2
9815 // This must be a constant between 0 and 31, for shift amounts.
9816 if (CVal >= 0 && CVal <= 31)
9822 if (Subtarget->isThumb()) { // FIXME thumb2
9823 // This must be a multiple of 4 between -508 and 508, for
9824 // ADD/SUB sp = sp + immediate.
9825 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
9830 Result = DAG.getTargetConstant(CVal, Op.getValueType());
9834 if (Result.getNode()) {
9835 Ops.push_back(Result);
9838 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
9842 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
9843 // The ARM target isn't yet aware of offsets.
9847 bool ARM::isBitFieldInvertedMask(unsigned v) {
9848 if (v == 0xffffffff)
9850 // there can be 1's on either or both "outsides", all the "inside"
9852 unsigned int lsb = 0, msb = 31;
9853 while (v & (1 << msb)) --msb;
9854 while (v & (1 << lsb)) ++lsb;
9855 for (unsigned int i = lsb; i <= msb; ++i) {
9862 /// isFPImmLegal - Returns true if the target can instruction select the
9863 /// specified FP immediate natively. If false, the legalizer will
9864 /// materialize the FP immediate as a load from a constant pool.
9865 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
9866 if (!Subtarget->hasVFP3())
9869 return ARM_AM::getFP32Imm(Imm) != -1;
9871 return ARM_AM::getFP64Imm(Imm) != -1;
9875 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
9876 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
9877 /// specified in the intrinsic calls.
9878 bool ARMTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
9880 unsigned Intrinsic) const {
9881 switch (Intrinsic) {
9882 case Intrinsic::arm_neon_vld1:
9883 case Intrinsic::arm_neon_vld2:
9884 case Intrinsic::arm_neon_vld3:
9885 case Intrinsic::arm_neon_vld4:
9886 case Intrinsic::arm_neon_vld2lane:
9887 case Intrinsic::arm_neon_vld3lane:
9888 case Intrinsic::arm_neon_vld4lane: {
9889 Info.opc = ISD::INTRINSIC_W_CHAIN;
9890 // Conservatively set memVT to the entire set of vectors loaded.
9891 uint64_t NumElts = getDataLayout()->getTypeAllocSize(I.getType()) / 8;
9892 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
9893 Info.ptrVal = I.getArgOperand(0);
9895 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
9896 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
9897 Info.vol = false; // volatile loads with NEON intrinsics not supported
9898 Info.readMem = true;
9899 Info.writeMem = false;
9902 case Intrinsic::arm_neon_vst1:
9903 case Intrinsic::arm_neon_vst2:
9904 case Intrinsic::arm_neon_vst3:
9905 case Intrinsic::arm_neon_vst4:
9906 case Intrinsic::arm_neon_vst2lane:
9907 case Intrinsic::arm_neon_vst3lane:
9908 case Intrinsic::arm_neon_vst4lane: {
9909 Info.opc = ISD::INTRINSIC_VOID;
9910 // Conservatively set memVT to the entire set of vectors stored.
9911 unsigned NumElts = 0;
9912 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
9913 Type *ArgTy = I.getArgOperand(ArgI)->getType();
9914 if (!ArgTy->isVectorTy())
9916 NumElts += getDataLayout()->getTypeAllocSize(ArgTy) / 8;
9918 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
9919 Info.ptrVal = I.getArgOperand(0);
9921 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
9922 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
9923 Info.vol = false; // volatile stores with NEON intrinsics not supported
9924 Info.readMem = false;
9925 Info.writeMem = true;
9928 case Intrinsic::arm_strexd: {
9929 Info.opc = ISD::INTRINSIC_W_CHAIN;
9930 Info.memVT = MVT::i64;
9931 Info.ptrVal = I.getArgOperand(2);
9935 Info.readMem = false;
9936 Info.writeMem = true;
9939 case Intrinsic::arm_ldrexd: {
9940 Info.opc = ISD::INTRINSIC_W_CHAIN;
9941 Info.memVT = MVT::i64;
9942 Info.ptrVal = I.getArgOperand(0);
9946 Info.readMem = true;
9947 Info.writeMem = false;