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"
17 #include "ARMCallingConv.h"
18 #include "ARMConstantPoolValue.h"
19 #include "ARMISelLowering.h"
20 #include "ARMMachineFunctionInfo.h"
21 #include "ARMPerfectShuffle.h"
22 #include "ARMRegisterInfo.h"
23 #include "ARMSubtarget.h"
24 #include "ARMTargetMachine.h"
25 #include "ARMTargetObjectFile.h"
26 #include "MCTargetDesc/ARMAddressingModes.h"
27 #include "llvm/CallingConv.h"
28 #include "llvm/Constants.h"
29 #include "llvm/Function.h"
30 #include "llvm/GlobalValue.h"
31 #include "llvm/Instruction.h"
32 #include "llvm/Instructions.h"
33 #include "llvm/Intrinsics.h"
34 #include "llvm/Type.h"
35 #include "llvm/CodeGen/CallingConvLower.h"
36 #include "llvm/CodeGen/IntrinsicLowering.h"
37 #include "llvm/CodeGen/MachineBasicBlock.h"
38 #include "llvm/CodeGen/MachineFrameInfo.h"
39 #include "llvm/CodeGen/MachineFunction.h"
40 #include "llvm/CodeGen/MachineInstrBuilder.h"
41 #include "llvm/CodeGen/MachineModuleInfo.h"
42 #include "llvm/CodeGen/MachineRegisterInfo.h"
43 #include "llvm/CodeGen/PseudoSourceValue.h"
44 #include "llvm/CodeGen/SelectionDAG.h"
45 #include "llvm/MC/MCSectionMachO.h"
46 #include "llvm/Target/TargetOptions.h"
47 #include "llvm/ADT/VectorExtras.h"
48 #include "llvm/ADT/StringExtras.h"
49 #include "llvm/ADT/Statistic.h"
50 #include "llvm/Support/CommandLine.h"
51 #include "llvm/Support/ErrorHandling.h"
52 #include "llvm/Support/MathExtras.h"
53 #include "llvm/Support/raw_ostream.h"
57 STATISTIC(NumTailCalls, "Number of tail calls");
58 STATISTIC(NumMovwMovt, "Number of GAs materialized with movw + movt");
60 // This option should go away when tail calls fully work.
62 EnableARMTailCalls("arm-tail-calls", cl::Hidden,
63 cl::desc("Generate tail calls (TEMPORARY OPTION)."),
67 EnableARMLongCalls("arm-long-calls", cl::Hidden,
68 cl::desc("Generate calls via indirect call instructions"),
72 ARMInterworking("arm-interworking", cl::Hidden,
73 cl::desc("Enable / disable ARM interworking (for debugging only)"),
77 class ARMCCState : public CCState {
79 ARMCCState(CallingConv::ID CC, bool isVarArg, MachineFunction &MF,
80 const TargetMachine &TM, SmallVector<CCValAssign, 16> &locs,
81 LLVMContext &C, ParmContext PC)
82 : CCState(CC, isVarArg, MF, TM, locs, C) {
83 assert(((PC == Call) || (PC == Prologue)) &&
84 "ARMCCState users must specify whether their context is call"
85 "or prologue generation.");
91 // The APCS parameter registers.
92 static const unsigned GPRArgRegs[] = {
93 ARM::R0, ARM::R1, ARM::R2, ARM::R3
96 void ARMTargetLowering::addTypeForNEON(EVT VT, EVT PromotedLdStVT,
97 EVT PromotedBitwiseVT) {
98 if (VT != PromotedLdStVT) {
99 setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
100 AddPromotedToType (ISD::LOAD, VT.getSimpleVT(),
101 PromotedLdStVT.getSimpleVT());
103 setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
104 AddPromotedToType (ISD::STORE, VT.getSimpleVT(),
105 PromotedLdStVT.getSimpleVT());
108 EVT ElemTy = VT.getVectorElementType();
109 if (ElemTy != MVT::i64 && ElemTy != MVT::f64)
110 setOperationAction(ISD::SETCC, VT.getSimpleVT(), Custom);
111 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT.getSimpleVT(), Custom);
112 if (ElemTy != MVT::i32) {
113 setOperationAction(ISD::SINT_TO_FP, VT.getSimpleVT(), Expand);
114 setOperationAction(ISD::UINT_TO_FP, VT.getSimpleVT(), Expand);
115 setOperationAction(ISD::FP_TO_SINT, VT.getSimpleVT(), Expand);
116 setOperationAction(ISD::FP_TO_UINT, VT.getSimpleVT(), Expand);
118 setOperationAction(ISD::BUILD_VECTOR, VT.getSimpleVT(), Custom);
119 setOperationAction(ISD::VECTOR_SHUFFLE, VT.getSimpleVT(), Custom);
120 setOperationAction(ISD::CONCAT_VECTORS, VT.getSimpleVT(), Legal);
121 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT.getSimpleVT(), Legal);
122 setOperationAction(ISD::SELECT, VT.getSimpleVT(), Expand);
123 setOperationAction(ISD::SELECT_CC, VT.getSimpleVT(), Expand);
124 if (VT.isInteger()) {
125 setOperationAction(ISD::SHL, VT.getSimpleVT(), Custom);
126 setOperationAction(ISD::SRA, VT.getSimpleVT(), Custom);
127 setOperationAction(ISD::SRL, VT.getSimpleVT(), Custom);
128 setLoadExtAction(ISD::SEXTLOAD, VT.getSimpleVT(), Expand);
129 setLoadExtAction(ISD::ZEXTLOAD, VT.getSimpleVT(), Expand);
130 for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
131 InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
132 setTruncStoreAction(VT.getSimpleVT(),
133 (MVT::SimpleValueType)InnerVT, Expand);
135 setLoadExtAction(ISD::EXTLOAD, VT.getSimpleVT(), Expand);
137 // Promote all bit-wise operations.
138 if (VT.isInteger() && VT != PromotedBitwiseVT) {
139 setOperationAction(ISD::AND, VT.getSimpleVT(), Promote);
140 AddPromotedToType (ISD::AND, VT.getSimpleVT(),
141 PromotedBitwiseVT.getSimpleVT());
142 setOperationAction(ISD::OR, VT.getSimpleVT(), Promote);
143 AddPromotedToType (ISD::OR, VT.getSimpleVT(),
144 PromotedBitwiseVT.getSimpleVT());
145 setOperationAction(ISD::XOR, VT.getSimpleVT(), Promote);
146 AddPromotedToType (ISD::XOR, VT.getSimpleVT(),
147 PromotedBitwiseVT.getSimpleVT());
150 // Neon does not support vector divide/remainder operations.
151 setOperationAction(ISD::SDIV, VT.getSimpleVT(), Expand);
152 setOperationAction(ISD::UDIV, VT.getSimpleVT(), Expand);
153 setOperationAction(ISD::FDIV, VT.getSimpleVT(), Expand);
154 setOperationAction(ISD::SREM, VT.getSimpleVT(), Expand);
155 setOperationAction(ISD::UREM, VT.getSimpleVT(), Expand);
156 setOperationAction(ISD::FREM, VT.getSimpleVT(), Expand);
159 void ARMTargetLowering::addDRTypeForNEON(EVT VT) {
160 addRegisterClass(VT, ARM::DPRRegisterClass);
161 addTypeForNEON(VT, MVT::f64, MVT::v2i32);
164 void ARMTargetLowering::addQRTypeForNEON(EVT VT) {
165 addRegisterClass(VT, ARM::QPRRegisterClass);
166 addTypeForNEON(VT, MVT::v2f64, MVT::v4i32);
169 static TargetLoweringObjectFile *createTLOF(TargetMachine &TM) {
170 if (TM.getSubtarget<ARMSubtarget>().isTargetDarwin())
171 return new TargetLoweringObjectFileMachO();
173 return new ARMElfTargetObjectFile();
176 ARMTargetLowering::ARMTargetLowering(TargetMachine &TM)
177 : TargetLowering(TM, createTLOF(TM)) {
178 Subtarget = &TM.getSubtarget<ARMSubtarget>();
179 RegInfo = TM.getRegisterInfo();
180 Itins = TM.getInstrItineraryData();
182 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
184 if (Subtarget->isTargetDarwin()) {
185 // Uses VFP for Thumb libfuncs if available.
186 if (Subtarget->isThumb() && Subtarget->hasVFP2()) {
187 // Single-precision floating-point arithmetic.
188 setLibcallName(RTLIB::ADD_F32, "__addsf3vfp");
189 setLibcallName(RTLIB::SUB_F32, "__subsf3vfp");
190 setLibcallName(RTLIB::MUL_F32, "__mulsf3vfp");
191 setLibcallName(RTLIB::DIV_F32, "__divsf3vfp");
193 // Double-precision floating-point arithmetic.
194 setLibcallName(RTLIB::ADD_F64, "__adddf3vfp");
195 setLibcallName(RTLIB::SUB_F64, "__subdf3vfp");
196 setLibcallName(RTLIB::MUL_F64, "__muldf3vfp");
197 setLibcallName(RTLIB::DIV_F64, "__divdf3vfp");
199 // Single-precision comparisons.
200 setLibcallName(RTLIB::OEQ_F32, "__eqsf2vfp");
201 setLibcallName(RTLIB::UNE_F32, "__nesf2vfp");
202 setLibcallName(RTLIB::OLT_F32, "__ltsf2vfp");
203 setLibcallName(RTLIB::OLE_F32, "__lesf2vfp");
204 setLibcallName(RTLIB::OGE_F32, "__gesf2vfp");
205 setLibcallName(RTLIB::OGT_F32, "__gtsf2vfp");
206 setLibcallName(RTLIB::UO_F32, "__unordsf2vfp");
207 setLibcallName(RTLIB::O_F32, "__unordsf2vfp");
209 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
210 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETNE);
211 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
212 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
213 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
214 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
215 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
216 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
218 // Double-precision comparisons.
219 setLibcallName(RTLIB::OEQ_F64, "__eqdf2vfp");
220 setLibcallName(RTLIB::UNE_F64, "__nedf2vfp");
221 setLibcallName(RTLIB::OLT_F64, "__ltdf2vfp");
222 setLibcallName(RTLIB::OLE_F64, "__ledf2vfp");
223 setLibcallName(RTLIB::OGE_F64, "__gedf2vfp");
224 setLibcallName(RTLIB::OGT_F64, "__gtdf2vfp");
225 setLibcallName(RTLIB::UO_F64, "__unorddf2vfp");
226 setLibcallName(RTLIB::O_F64, "__unorddf2vfp");
228 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
229 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETNE);
230 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
231 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
232 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
233 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
234 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
235 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
237 // Floating-point to integer conversions.
238 // i64 conversions are done via library routines even when generating VFP
239 // instructions, so use the same ones.
240 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__fixdfsivfp");
241 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__fixunsdfsivfp");
242 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__fixsfsivfp");
243 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__fixunssfsivfp");
245 // Conversions between floating types.
246 setLibcallName(RTLIB::FPROUND_F64_F32, "__truncdfsf2vfp");
247 setLibcallName(RTLIB::FPEXT_F32_F64, "__extendsfdf2vfp");
249 // Integer to floating-point conversions.
250 // i64 conversions are done via library routines even when generating VFP
251 // instructions, so use the same ones.
252 // FIXME: There appears to be some naming inconsistency in ARM libgcc:
253 // e.g., __floatunsidf vs. __floatunssidfvfp.
254 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__floatsidfvfp");
255 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__floatunssidfvfp");
256 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__floatsisfvfp");
257 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__floatunssisfvfp");
261 // These libcalls are not available in 32-bit.
262 setLibcallName(RTLIB::SHL_I128, 0);
263 setLibcallName(RTLIB::SRL_I128, 0);
264 setLibcallName(RTLIB::SRA_I128, 0);
266 if (Subtarget->isAAPCS_ABI()) {
267 // Double-precision floating-point arithmetic helper functions
268 // RTABI chapter 4.1.2, Table 2
269 setLibcallName(RTLIB::ADD_F64, "__aeabi_dadd");
270 setLibcallName(RTLIB::DIV_F64, "__aeabi_ddiv");
271 setLibcallName(RTLIB::MUL_F64, "__aeabi_dmul");
272 setLibcallName(RTLIB::SUB_F64, "__aeabi_dsub");
273 setLibcallCallingConv(RTLIB::ADD_F64, CallingConv::ARM_AAPCS);
274 setLibcallCallingConv(RTLIB::DIV_F64, CallingConv::ARM_AAPCS);
275 setLibcallCallingConv(RTLIB::MUL_F64, CallingConv::ARM_AAPCS);
276 setLibcallCallingConv(RTLIB::SUB_F64, CallingConv::ARM_AAPCS);
278 // Double-precision floating-point comparison helper functions
279 // RTABI chapter 4.1.2, Table 3
280 setLibcallName(RTLIB::OEQ_F64, "__aeabi_dcmpeq");
281 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
282 setLibcallName(RTLIB::UNE_F64, "__aeabi_dcmpeq");
283 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETEQ);
284 setLibcallName(RTLIB::OLT_F64, "__aeabi_dcmplt");
285 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
286 setLibcallName(RTLIB::OLE_F64, "__aeabi_dcmple");
287 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
288 setLibcallName(RTLIB::OGE_F64, "__aeabi_dcmpge");
289 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
290 setLibcallName(RTLIB::OGT_F64, "__aeabi_dcmpgt");
291 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
292 setLibcallName(RTLIB::UO_F64, "__aeabi_dcmpun");
293 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
294 setLibcallName(RTLIB::O_F64, "__aeabi_dcmpun");
295 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
296 setLibcallCallingConv(RTLIB::OEQ_F64, CallingConv::ARM_AAPCS);
297 setLibcallCallingConv(RTLIB::UNE_F64, CallingConv::ARM_AAPCS);
298 setLibcallCallingConv(RTLIB::OLT_F64, CallingConv::ARM_AAPCS);
299 setLibcallCallingConv(RTLIB::OLE_F64, CallingConv::ARM_AAPCS);
300 setLibcallCallingConv(RTLIB::OGE_F64, CallingConv::ARM_AAPCS);
301 setLibcallCallingConv(RTLIB::OGT_F64, CallingConv::ARM_AAPCS);
302 setLibcallCallingConv(RTLIB::UO_F64, CallingConv::ARM_AAPCS);
303 setLibcallCallingConv(RTLIB::O_F64, CallingConv::ARM_AAPCS);
305 // Single-precision floating-point arithmetic helper functions
306 // RTABI chapter 4.1.2, Table 4
307 setLibcallName(RTLIB::ADD_F32, "__aeabi_fadd");
308 setLibcallName(RTLIB::DIV_F32, "__aeabi_fdiv");
309 setLibcallName(RTLIB::MUL_F32, "__aeabi_fmul");
310 setLibcallName(RTLIB::SUB_F32, "__aeabi_fsub");
311 setLibcallCallingConv(RTLIB::ADD_F32, CallingConv::ARM_AAPCS);
312 setLibcallCallingConv(RTLIB::DIV_F32, CallingConv::ARM_AAPCS);
313 setLibcallCallingConv(RTLIB::MUL_F32, CallingConv::ARM_AAPCS);
314 setLibcallCallingConv(RTLIB::SUB_F32, CallingConv::ARM_AAPCS);
316 // Single-precision floating-point comparison helper functions
317 // RTABI chapter 4.1.2, Table 5
318 setLibcallName(RTLIB::OEQ_F32, "__aeabi_fcmpeq");
319 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
320 setLibcallName(RTLIB::UNE_F32, "__aeabi_fcmpeq");
321 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETEQ);
322 setLibcallName(RTLIB::OLT_F32, "__aeabi_fcmplt");
323 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
324 setLibcallName(RTLIB::OLE_F32, "__aeabi_fcmple");
325 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
326 setLibcallName(RTLIB::OGE_F32, "__aeabi_fcmpge");
327 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
328 setLibcallName(RTLIB::OGT_F32, "__aeabi_fcmpgt");
329 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
330 setLibcallName(RTLIB::UO_F32, "__aeabi_fcmpun");
331 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
332 setLibcallName(RTLIB::O_F32, "__aeabi_fcmpun");
333 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
334 setLibcallCallingConv(RTLIB::OEQ_F32, CallingConv::ARM_AAPCS);
335 setLibcallCallingConv(RTLIB::UNE_F32, CallingConv::ARM_AAPCS);
336 setLibcallCallingConv(RTLIB::OLT_F32, CallingConv::ARM_AAPCS);
337 setLibcallCallingConv(RTLIB::OLE_F32, CallingConv::ARM_AAPCS);
338 setLibcallCallingConv(RTLIB::OGE_F32, CallingConv::ARM_AAPCS);
339 setLibcallCallingConv(RTLIB::OGT_F32, CallingConv::ARM_AAPCS);
340 setLibcallCallingConv(RTLIB::UO_F32, CallingConv::ARM_AAPCS);
341 setLibcallCallingConv(RTLIB::O_F32, CallingConv::ARM_AAPCS);
343 // Floating-point to integer conversions.
344 // RTABI chapter 4.1.2, Table 6
345 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__aeabi_d2iz");
346 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__aeabi_d2uiz");
347 setLibcallName(RTLIB::FPTOSINT_F64_I64, "__aeabi_d2lz");
348 setLibcallName(RTLIB::FPTOUINT_F64_I64, "__aeabi_d2ulz");
349 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__aeabi_f2iz");
350 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__aeabi_f2uiz");
351 setLibcallName(RTLIB::FPTOSINT_F32_I64, "__aeabi_f2lz");
352 setLibcallName(RTLIB::FPTOUINT_F32_I64, "__aeabi_f2ulz");
353 setLibcallCallingConv(RTLIB::FPTOSINT_F64_I32, CallingConv::ARM_AAPCS);
354 setLibcallCallingConv(RTLIB::FPTOUINT_F64_I32, CallingConv::ARM_AAPCS);
355 setLibcallCallingConv(RTLIB::FPTOSINT_F64_I64, CallingConv::ARM_AAPCS);
356 setLibcallCallingConv(RTLIB::FPTOUINT_F64_I64, CallingConv::ARM_AAPCS);
357 setLibcallCallingConv(RTLIB::FPTOSINT_F32_I32, CallingConv::ARM_AAPCS);
358 setLibcallCallingConv(RTLIB::FPTOUINT_F32_I32, CallingConv::ARM_AAPCS);
359 setLibcallCallingConv(RTLIB::FPTOSINT_F32_I64, CallingConv::ARM_AAPCS);
360 setLibcallCallingConv(RTLIB::FPTOUINT_F32_I64, CallingConv::ARM_AAPCS);
362 // Conversions between floating types.
363 // RTABI chapter 4.1.2, Table 7
364 setLibcallName(RTLIB::FPROUND_F64_F32, "__aeabi_d2f");
365 setLibcallName(RTLIB::FPEXT_F32_F64, "__aeabi_f2d");
366 setLibcallCallingConv(RTLIB::FPROUND_F64_F32, CallingConv::ARM_AAPCS);
367 setLibcallCallingConv(RTLIB::FPEXT_F32_F64, CallingConv::ARM_AAPCS);
369 // Integer to floating-point conversions.
370 // RTABI chapter 4.1.2, Table 8
371 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__aeabi_i2d");
372 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__aeabi_ui2d");
373 setLibcallName(RTLIB::SINTTOFP_I64_F64, "__aeabi_l2d");
374 setLibcallName(RTLIB::UINTTOFP_I64_F64, "__aeabi_ul2d");
375 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__aeabi_i2f");
376 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__aeabi_ui2f");
377 setLibcallName(RTLIB::SINTTOFP_I64_F32, "__aeabi_l2f");
378 setLibcallName(RTLIB::UINTTOFP_I64_F32, "__aeabi_ul2f");
379 setLibcallCallingConv(RTLIB::SINTTOFP_I32_F64, CallingConv::ARM_AAPCS);
380 setLibcallCallingConv(RTLIB::UINTTOFP_I32_F64, CallingConv::ARM_AAPCS);
381 setLibcallCallingConv(RTLIB::SINTTOFP_I64_F64, CallingConv::ARM_AAPCS);
382 setLibcallCallingConv(RTLIB::UINTTOFP_I64_F64, CallingConv::ARM_AAPCS);
383 setLibcallCallingConv(RTLIB::SINTTOFP_I32_F32, CallingConv::ARM_AAPCS);
384 setLibcallCallingConv(RTLIB::UINTTOFP_I32_F32, CallingConv::ARM_AAPCS);
385 setLibcallCallingConv(RTLIB::SINTTOFP_I64_F32, CallingConv::ARM_AAPCS);
386 setLibcallCallingConv(RTLIB::UINTTOFP_I64_F32, CallingConv::ARM_AAPCS);
388 // Long long helper functions
389 // RTABI chapter 4.2, Table 9
390 setLibcallName(RTLIB::MUL_I64, "__aeabi_lmul");
391 setLibcallName(RTLIB::SDIV_I64, "__aeabi_ldivmod");
392 setLibcallName(RTLIB::UDIV_I64, "__aeabi_uldivmod");
393 setLibcallName(RTLIB::SHL_I64, "__aeabi_llsl");
394 setLibcallName(RTLIB::SRL_I64, "__aeabi_llsr");
395 setLibcallName(RTLIB::SRA_I64, "__aeabi_lasr");
396 setLibcallCallingConv(RTLIB::MUL_I64, CallingConv::ARM_AAPCS);
397 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::ARM_AAPCS);
398 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::ARM_AAPCS);
399 setLibcallCallingConv(RTLIB::SHL_I64, CallingConv::ARM_AAPCS);
400 setLibcallCallingConv(RTLIB::SRL_I64, CallingConv::ARM_AAPCS);
401 setLibcallCallingConv(RTLIB::SRA_I64, CallingConv::ARM_AAPCS);
403 // Integer division functions
404 // RTABI chapter 4.3.1
405 setLibcallName(RTLIB::SDIV_I8, "__aeabi_idiv");
406 setLibcallName(RTLIB::SDIV_I16, "__aeabi_idiv");
407 setLibcallName(RTLIB::SDIV_I32, "__aeabi_idiv");
408 setLibcallName(RTLIB::UDIV_I8, "__aeabi_uidiv");
409 setLibcallName(RTLIB::UDIV_I16, "__aeabi_uidiv");
410 setLibcallName(RTLIB::UDIV_I32, "__aeabi_uidiv");
411 setLibcallCallingConv(RTLIB::SDIV_I8, CallingConv::ARM_AAPCS);
412 setLibcallCallingConv(RTLIB::SDIV_I16, CallingConv::ARM_AAPCS);
413 setLibcallCallingConv(RTLIB::SDIV_I32, CallingConv::ARM_AAPCS);
414 setLibcallCallingConv(RTLIB::UDIV_I8, CallingConv::ARM_AAPCS);
415 setLibcallCallingConv(RTLIB::UDIV_I16, CallingConv::ARM_AAPCS);
416 setLibcallCallingConv(RTLIB::UDIV_I32, CallingConv::ARM_AAPCS);
419 // RTABI chapter 4.3.4
420 setLibcallName(RTLIB::MEMCPY, "__aeabi_memcpy");
421 setLibcallName(RTLIB::MEMMOVE, "__aeabi_memmove");
422 setLibcallName(RTLIB::MEMSET, "__aeabi_memset");
425 // Use divmod compiler-rt calls for iOS 5.0 and later.
426 if (Subtarget->getTargetTriple().getOS() == Triple::IOS &&
427 !Subtarget->getTargetTriple().isOSVersionLT(5, 0)) {
428 setLibcallName(RTLIB::SDIVREM_I32, "__divmodsi4");
429 setLibcallName(RTLIB::UDIVREM_I32, "__udivmodsi4");
432 if (Subtarget->isThumb1Only())
433 addRegisterClass(MVT::i32, ARM::tGPRRegisterClass);
435 addRegisterClass(MVT::i32, ARM::GPRRegisterClass);
436 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
437 addRegisterClass(MVT::f32, ARM::SPRRegisterClass);
438 if (!Subtarget->isFPOnlySP())
439 addRegisterClass(MVT::f64, ARM::DPRRegisterClass);
441 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
444 if (Subtarget->hasNEON()) {
445 addDRTypeForNEON(MVT::v2f32);
446 addDRTypeForNEON(MVT::v8i8);
447 addDRTypeForNEON(MVT::v4i16);
448 addDRTypeForNEON(MVT::v2i32);
449 addDRTypeForNEON(MVT::v1i64);
451 addQRTypeForNEON(MVT::v4f32);
452 addQRTypeForNEON(MVT::v2f64);
453 addQRTypeForNEON(MVT::v16i8);
454 addQRTypeForNEON(MVT::v8i16);
455 addQRTypeForNEON(MVT::v4i32);
456 addQRTypeForNEON(MVT::v2i64);
458 // v2f64 is legal so that QR subregs can be extracted as f64 elements, but
459 // neither Neon nor VFP support any arithmetic operations on it.
460 setOperationAction(ISD::FADD, MVT::v2f64, Expand);
461 setOperationAction(ISD::FSUB, MVT::v2f64, Expand);
462 setOperationAction(ISD::FMUL, MVT::v2f64, Expand);
463 setOperationAction(ISD::FDIV, MVT::v2f64, Expand);
464 setOperationAction(ISD::FREM, MVT::v2f64, Expand);
465 setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand);
466 setOperationAction(ISD::SETCC, MVT::v2f64, Expand);
467 setOperationAction(ISD::FNEG, MVT::v2f64, Expand);
468 setOperationAction(ISD::FABS, MVT::v2f64, Expand);
469 setOperationAction(ISD::FSQRT, MVT::v2f64, Expand);
470 setOperationAction(ISD::FSIN, MVT::v2f64, Expand);
471 setOperationAction(ISD::FCOS, MVT::v2f64, Expand);
472 setOperationAction(ISD::FPOWI, MVT::v2f64, Expand);
473 setOperationAction(ISD::FPOW, MVT::v2f64, Expand);
474 setOperationAction(ISD::FLOG, MVT::v2f64, Expand);
475 setOperationAction(ISD::FLOG2, MVT::v2f64, Expand);
476 setOperationAction(ISD::FLOG10, MVT::v2f64, Expand);
477 setOperationAction(ISD::FEXP, MVT::v2f64, Expand);
478 setOperationAction(ISD::FEXP2, MVT::v2f64, Expand);
479 setOperationAction(ISD::FCEIL, MVT::v2f64, Expand);
480 setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand);
481 setOperationAction(ISD::FRINT, MVT::v2f64, Expand);
482 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand);
483 setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand);
485 setTruncStoreAction(MVT::v2f64, MVT::v2f32, Expand);
487 // Neon does not support some operations on v1i64 and v2i64 types.
488 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
489 // Custom handling for some quad-vector types to detect VMULL.
490 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
491 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
492 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
493 // Custom handling for some vector types to avoid expensive expansions
494 setOperationAction(ISD::SDIV, MVT::v4i16, Custom);
495 setOperationAction(ISD::SDIV, MVT::v8i8, Custom);
496 setOperationAction(ISD::UDIV, MVT::v4i16, Custom);
497 setOperationAction(ISD::UDIV, MVT::v8i8, Custom);
498 setOperationAction(ISD::SETCC, MVT::v1i64, Expand);
499 setOperationAction(ISD::SETCC, MVT::v2i64, Expand);
500 // Neon does not have single instruction SINT_TO_FP and UINT_TO_FP with
501 // a destination type that is wider than the source.
502 setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Custom);
503 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
505 setTargetDAGCombine(ISD::INTRINSIC_VOID);
506 setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN);
507 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
508 setTargetDAGCombine(ISD::SHL);
509 setTargetDAGCombine(ISD::SRL);
510 setTargetDAGCombine(ISD::SRA);
511 setTargetDAGCombine(ISD::SIGN_EXTEND);
512 setTargetDAGCombine(ISD::ZERO_EXTEND);
513 setTargetDAGCombine(ISD::ANY_EXTEND);
514 setTargetDAGCombine(ISD::SELECT_CC);
515 setTargetDAGCombine(ISD::BUILD_VECTOR);
516 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
517 setTargetDAGCombine(ISD::INSERT_VECTOR_ELT);
518 setTargetDAGCombine(ISD::STORE);
519 setTargetDAGCombine(ISD::FP_TO_SINT);
520 setTargetDAGCombine(ISD::FP_TO_UINT);
521 setTargetDAGCombine(ISD::FDIV);
523 setLoadExtAction(ISD::EXTLOAD, MVT::v4i8, Expand);
526 computeRegisterProperties();
528 // ARM does not have f32 extending load.
529 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
531 // ARM does not have i1 sign extending load.
532 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
534 // ARM supports all 4 flavors of integer indexed load / store.
535 if (!Subtarget->isThumb1Only()) {
536 for (unsigned im = (unsigned)ISD::PRE_INC;
537 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
538 setIndexedLoadAction(im, MVT::i1, Legal);
539 setIndexedLoadAction(im, MVT::i8, Legal);
540 setIndexedLoadAction(im, MVT::i16, Legal);
541 setIndexedLoadAction(im, MVT::i32, Legal);
542 setIndexedStoreAction(im, MVT::i1, Legal);
543 setIndexedStoreAction(im, MVT::i8, Legal);
544 setIndexedStoreAction(im, MVT::i16, Legal);
545 setIndexedStoreAction(im, MVT::i32, Legal);
549 // i64 operation support.
550 setOperationAction(ISD::MUL, MVT::i64, Expand);
551 setOperationAction(ISD::MULHU, MVT::i32, Expand);
552 if (Subtarget->isThumb1Only()) {
553 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
554 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
556 if (Subtarget->isThumb1Only() || !Subtarget->hasV6Ops()
557 || (Subtarget->isThumb2() && !Subtarget->hasThumb2DSP()))
558 setOperationAction(ISD::MULHS, MVT::i32, Expand);
560 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
561 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
562 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
563 setOperationAction(ISD::SRL, MVT::i64, Custom);
564 setOperationAction(ISD::SRA, MVT::i64, Custom);
566 if (!Subtarget->isThumb1Only()) {
567 // FIXME: We should do this for Thumb1 as well.
568 setOperationAction(ISD::ADDC, MVT::i32, Custom);
569 setOperationAction(ISD::ADDE, MVT::i32, Custom);
570 setOperationAction(ISD::SUBC, MVT::i32, Custom);
571 setOperationAction(ISD::SUBE, MVT::i32, Custom);
574 // ARM does not have ROTL.
575 setOperationAction(ISD::ROTL, MVT::i32, Expand);
576 setOperationAction(ISD::CTTZ, MVT::i32, Custom);
577 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
578 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
579 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
581 // Only ARMv6 has BSWAP.
582 if (!Subtarget->hasV6Ops())
583 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
585 // These are expanded into libcalls.
586 if (!Subtarget->hasDivide() || !Subtarget->isThumb2()) {
587 // v7M has a hardware divider
588 setOperationAction(ISD::SDIV, MVT::i32, Expand);
589 setOperationAction(ISD::UDIV, MVT::i32, Expand);
591 setOperationAction(ISD::SREM, MVT::i32, Expand);
592 setOperationAction(ISD::UREM, MVT::i32, Expand);
593 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
594 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
596 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
597 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
598 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
599 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
600 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
602 setOperationAction(ISD::TRAP, MVT::Other, Legal);
604 // Use the default implementation.
605 setOperationAction(ISD::VASTART, MVT::Other, Custom);
606 setOperationAction(ISD::VAARG, MVT::Other, Expand);
607 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
608 setOperationAction(ISD::VAEND, MVT::Other, Expand);
609 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
610 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
611 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
612 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
613 setExceptionPointerRegister(ARM::R0);
614 setExceptionSelectorRegister(ARM::R1);
616 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
617 // ARMv6 Thumb1 (except for CPUs that support dmb / dsb) and earlier use
618 // the default expansion.
619 // FIXME: This should be checking for v6k, not just v6.
620 if (Subtarget->hasDataBarrier() ||
621 (Subtarget->hasV6Ops() && !Subtarget->isThumb())) {
622 // membarrier needs custom lowering; the rest are legal and handled
624 setOperationAction(ISD::MEMBARRIER, MVT::Other, Custom);
625 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Custom);
626 // Custom lowering for 64-bit ops
627 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
628 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
629 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);
630 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom);
631 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
632 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
633 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Custom);
634 // Automatically insert fences (dmb ist) around ATOMIC_SWAP etc.
635 setInsertFencesForAtomic(true);
637 // Set them all for expansion, which will force libcalls.
638 setOperationAction(ISD::MEMBARRIER, MVT::Other, Expand);
639 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Expand);
640 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Expand);
641 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Expand);
642 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Expand);
643 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Expand);
644 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Expand);
645 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Expand);
646 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Expand);
647 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Expand);
648 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i32, Expand);
649 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i32, Expand);
650 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i32, Expand);
651 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i32, Expand);
652 // Mark ATOMIC_LOAD and ATOMIC_STORE custom so we can handle the
653 // Unordered/Monotonic case.
654 setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Custom);
655 setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Custom);
656 // Since the libcalls include locking, fold in the fences
657 setShouldFoldAtomicFences(true);
660 setOperationAction(ISD::PREFETCH, MVT::Other, Custom);
662 // Requires SXTB/SXTH, available on v6 and up in both ARM and Thumb modes.
663 if (!Subtarget->hasV6Ops()) {
664 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
665 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
667 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
669 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
670 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
671 // iff target supports vfp2.
672 setOperationAction(ISD::BITCAST, MVT::i64, Custom);
673 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
676 // We want to custom lower some of our intrinsics.
677 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
678 if (Subtarget->isTargetDarwin()) {
679 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
680 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
681 setOperationAction(ISD::EH_SJLJ_DISPATCHSETUP, MVT::Other, Custom);
682 setLibcallName(RTLIB::UNWIND_RESUME, "_Unwind_SjLj_Resume");
685 setOperationAction(ISD::SETCC, MVT::i32, Expand);
686 setOperationAction(ISD::SETCC, MVT::f32, Expand);
687 setOperationAction(ISD::SETCC, MVT::f64, Expand);
688 setOperationAction(ISD::SELECT, MVT::i32, Custom);
689 setOperationAction(ISD::SELECT, MVT::f32, Custom);
690 setOperationAction(ISD::SELECT, MVT::f64, Custom);
691 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
692 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
693 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
695 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
696 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
697 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
698 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
699 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
701 // We don't support sin/cos/fmod/copysign/pow
702 setOperationAction(ISD::FSIN, MVT::f64, Expand);
703 setOperationAction(ISD::FSIN, MVT::f32, Expand);
704 setOperationAction(ISD::FCOS, MVT::f32, Expand);
705 setOperationAction(ISD::FCOS, MVT::f64, Expand);
706 setOperationAction(ISD::FREM, MVT::f64, Expand);
707 setOperationAction(ISD::FREM, MVT::f32, Expand);
708 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
709 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
710 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
712 setOperationAction(ISD::FPOW, MVT::f64, Expand);
713 setOperationAction(ISD::FPOW, MVT::f32, Expand);
715 setOperationAction(ISD::FMA, MVT::f64, Expand);
716 setOperationAction(ISD::FMA, MVT::f32, Expand);
718 // Various VFP goodness
719 if (!UseSoftFloat && !Subtarget->isThumb1Only()) {
720 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
721 if (Subtarget->hasVFP2()) {
722 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
723 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
724 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
725 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
727 // Special handling for half-precision FP.
728 if (!Subtarget->hasFP16()) {
729 setOperationAction(ISD::FP16_TO_FP32, MVT::f32, Expand);
730 setOperationAction(ISD::FP32_TO_FP16, MVT::i32, Expand);
734 // We have target-specific dag combine patterns for the following nodes:
735 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
736 setTargetDAGCombine(ISD::ADD);
737 setTargetDAGCombine(ISD::SUB);
738 setTargetDAGCombine(ISD::MUL);
740 if (Subtarget->hasV6T2Ops() || Subtarget->hasNEON())
741 setTargetDAGCombine(ISD::OR);
742 if (Subtarget->hasNEON())
743 setTargetDAGCombine(ISD::AND);
745 setStackPointerRegisterToSaveRestore(ARM::SP);
747 if (UseSoftFloat || Subtarget->isThumb1Only() || !Subtarget->hasVFP2())
748 setSchedulingPreference(Sched::RegPressure);
750 setSchedulingPreference(Sched::Hybrid);
752 //// temporary - rewrite interface to use type
753 maxStoresPerMemcpy = maxStoresPerMemcpyOptSize = 1;
755 // On ARM arguments smaller than 4 bytes are extended, so all arguments
756 // are at least 4 bytes aligned.
757 setMinStackArgumentAlignment(4);
759 benefitFromCodePlacementOpt = true;
761 setMinFunctionAlignment(Subtarget->isThumb() ? 1 : 2);
764 // FIXME: It might make sense to define the representative register class as the
765 // nearest super-register that has a non-null superset. For example, DPR_VFP2 is
766 // a super-register of SPR, and DPR is a superset if DPR_VFP2. Consequently,
767 // SPR's representative would be DPR_VFP2. This should work well if register
768 // pressure tracking were modified such that a register use would increment the
769 // pressure of the register class's representative and all of it's super
770 // classes' representatives transitively. We have not implemented this because
771 // of the difficulty prior to coalescing of modeling operand register classes
772 // due to the common occurrence of cross class copies and subregister insertions
774 std::pair<const TargetRegisterClass*, uint8_t>
775 ARMTargetLowering::findRepresentativeClass(EVT VT) const{
776 const TargetRegisterClass *RRC = 0;
778 switch (VT.getSimpleVT().SimpleTy) {
780 return TargetLowering::findRepresentativeClass(VT);
781 // Use DPR as representative register class for all floating point
782 // and vector types. Since there are 32 SPR registers and 32 DPR registers so
783 // the cost is 1 for both f32 and f64.
784 case MVT::f32: case MVT::f64: case MVT::v8i8: case MVT::v4i16:
785 case MVT::v2i32: case MVT::v1i64: case MVT::v2f32:
786 RRC = ARM::DPRRegisterClass;
787 // When NEON is used for SP, only half of the register file is available
788 // because operations that define both SP and DP results will be constrained
789 // to the VFP2 class (D0-D15). We currently model this constraint prior to
790 // coalescing by double-counting the SP regs. See the FIXME above.
791 if (Subtarget->useNEONForSinglePrecisionFP())
794 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
795 case MVT::v4f32: case MVT::v2f64:
796 RRC = ARM::DPRRegisterClass;
800 RRC = ARM::DPRRegisterClass;
804 RRC = ARM::DPRRegisterClass;
808 return std::make_pair(RRC, Cost);
811 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
814 case ARMISD::Wrapper: return "ARMISD::Wrapper";
815 case ARMISD::WrapperDYN: return "ARMISD::WrapperDYN";
816 case ARMISD::WrapperPIC: return "ARMISD::WrapperPIC";
817 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
818 case ARMISD::CALL: return "ARMISD::CALL";
819 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
820 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
821 case ARMISD::tCALL: return "ARMISD::tCALL";
822 case ARMISD::BRCOND: return "ARMISD::BRCOND";
823 case ARMISD::BR_JT: return "ARMISD::BR_JT";
824 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
825 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
826 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
827 case ARMISD::CMP: return "ARMISD::CMP";
828 case ARMISD::CMPZ: return "ARMISD::CMPZ";
829 case ARMISD::CMPFP: return "ARMISD::CMPFP";
830 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
831 case ARMISD::BCC_i64: return "ARMISD::BCC_i64";
832 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
833 case ARMISD::CMOV: return "ARMISD::CMOV";
835 case ARMISD::RBIT: return "ARMISD::RBIT";
837 case ARMISD::FTOSI: return "ARMISD::FTOSI";
838 case ARMISD::FTOUI: return "ARMISD::FTOUI";
839 case ARMISD::SITOF: return "ARMISD::SITOF";
840 case ARMISD::UITOF: return "ARMISD::UITOF";
842 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
843 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
844 case ARMISD::RRX: return "ARMISD::RRX";
846 case ARMISD::ADDC: return "ARMISD::ADDC";
847 case ARMISD::ADDE: return "ARMISD::ADDE";
848 case ARMISD::SUBC: return "ARMISD::SUBC";
849 case ARMISD::SUBE: return "ARMISD::SUBE";
851 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
852 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
854 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
855 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
856 case ARMISD::EH_SJLJ_DISPATCHSETUP:return "ARMISD::EH_SJLJ_DISPATCHSETUP";
858 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN";
860 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
862 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
864 case ARMISD::MEMBARRIER: return "ARMISD::MEMBARRIER";
865 case ARMISD::MEMBARRIER_MCR: return "ARMISD::MEMBARRIER_MCR";
867 case ARMISD::PRELOAD: return "ARMISD::PRELOAD";
869 case ARMISD::VCEQ: return "ARMISD::VCEQ";
870 case ARMISD::VCEQZ: return "ARMISD::VCEQZ";
871 case ARMISD::VCGE: return "ARMISD::VCGE";
872 case ARMISD::VCGEZ: return "ARMISD::VCGEZ";
873 case ARMISD::VCLEZ: return "ARMISD::VCLEZ";
874 case ARMISD::VCGEU: return "ARMISD::VCGEU";
875 case ARMISD::VCGT: return "ARMISD::VCGT";
876 case ARMISD::VCGTZ: return "ARMISD::VCGTZ";
877 case ARMISD::VCLTZ: return "ARMISD::VCLTZ";
878 case ARMISD::VCGTU: return "ARMISD::VCGTU";
879 case ARMISD::VTST: return "ARMISD::VTST";
881 case ARMISD::VSHL: return "ARMISD::VSHL";
882 case ARMISD::VSHRs: return "ARMISD::VSHRs";
883 case ARMISD::VSHRu: return "ARMISD::VSHRu";
884 case ARMISD::VSHLLs: return "ARMISD::VSHLLs";
885 case ARMISD::VSHLLu: return "ARMISD::VSHLLu";
886 case ARMISD::VSHLLi: return "ARMISD::VSHLLi";
887 case ARMISD::VSHRN: return "ARMISD::VSHRN";
888 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
889 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
890 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
891 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
892 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
893 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
894 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
895 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
896 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
897 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
898 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
899 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
900 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
901 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
902 case ARMISD::VMOVIMM: return "ARMISD::VMOVIMM";
903 case ARMISD::VMVNIMM: return "ARMISD::VMVNIMM";
904 case ARMISD::VDUP: return "ARMISD::VDUP";
905 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
906 case ARMISD::VEXT: return "ARMISD::VEXT";
907 case ARMISD::VREV64: return "ARMISD::VREV64";
908 case ARMISD::VREV32: return "ARMISD::VREV32";
909 case ARMISD::VREV16: return "ARMISD::VREV16";
910 case ARMISD::VZIP: return "ARMISD::VZIP";
911 case ARMISD::VUZP: return "ARMISD::VUZP";
912 case ARMISD::VTRN: return "ARMISD::VTRN";
913 case ARMISD::VTBL1: return "ARMISD::VTBL1";
914 case ARMISD::VTBL2: return "ARMISD::VTBL2";
915 case ARMISD::VMULLs: return "ARMISD::VMULLs";
916 case ARMISD::VMULLu: return "ARMISD::VMULLu";
917 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
918 case ARMISD::FMAX: return "ARMISD::FMAX";
919 case ARMISD::FMIN: return "ARMISD::FMIN";
920 case ARMISD::BFI: return "ARMISD::BFI";
921 case ARMISD::VORRIMM: return "ARMISD::VORRIMM";
922 case ARMISD::VBICIMM: return "ARMISD::VBICIMM";
923 case ARMISD::VBSL: return "ARMISD::VBSL";
924 case ARMISD::VLD2DUP: return "ARMISD::VLD2DUP";
925 case ARMISD::VLD3DUP: return "ARMISD::VLD3DUP";
926 case ARMISD::VLD4DUP: return "ARMISD::VLD4DUP";
927 case ARMISD::VLD1_UPD: return "ARMISD::VLD1_UPD";
928 case ARMISD::VLD2_UPD: return "ARMISD::VLD2_UPD";
929 case ARMISD::VLD3_UPD: return "ARMISD::VLD3_UPD";
930 case ARMISD::VLD4_UPD: return "ARMISD::VLD4_UPD";
931 case ARMISD::VLD2LN_UPD: return "ARMISD::VLD2LN_UPD";
932 case ARMISD::VLD3LN_UPD: return "ARMISD::VLD3LN_UPD";
933 case ARMISD::VLD4LN_UPD: return "ARMISD::VLD4LN_UPD";
934 case ARMISD::VLD2DUP_UPD: return "ARMISD::VLD2DUP_UPD";
935 case ARMISD::VLD3DUP_UPD: return "ARMISD::VLD3DUP_UPD";
936 case ARMISD::VLD4DUP_UPD: return "ARMISD::VLD4DUP_UPD";
937 case ARMISD::VST1_UPD: return "ARMISD::VST1_UPD";
938 case ARMISD::VST2_UPD: return "ARMISD::VST2_UPD";
939 case ARMISD::VST3_UPD: return "ARMISD::VST3_UPD";
940 case ARMISD::VST4_UPD: return "ARMISD::VST4_UPD";
941 case ARMISD::VST2LN_UPD: return "ARMISD::VST2LN_UPD";
942 case ARMISD::VST3LN_UPD: return "ARMISD::VST3LN_UPD";
943 case ARMISD::VST4LN_UPD: return "ARMISD::VST4LN_UPD";
947 EVT ARMTargetLowering::getSetCCResultType(EVT VT) const {
948 if (!VT.isVector()) return getPointerTy();
949 return VT.changeVectorElementTypeToInteger();
952 /// getRegClassFor - Return the register class that should be used for the
953 /// specified value type.
954 TargetRegisterClass *ARMTargetLowering::getRegClassFor(EVT VT) const {
955 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
956 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
957 // load / store 4 to 8 consecutive D registers.
958 if (Subtarget->hasNEON()) {
959 if (VT == MVT::v4i64)
960 return ARM::QQPRRegisterClass;
961 else if (VT == MVT::v8i64)
962 return ARM::QQQQPRRegisterClass;
964 return TargetLowering::getRegClassFor(VT);
967 // Create a fast isel object.
969 ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo) const {
970 return ARM::createFastISel(funcInfo);
973 /// getMaximalGlobalOffset - Returns the maximal possible offset which can
974 /// be used for loads / stores from the global.
975 unsigned ARMTargetLowering::getMaximalGlobalOffset() const {
976 return (Subtarget->isThumb1Only() ? 127 : 4095);
979 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
980 unsigned NumVals = N->getNumValues();
982 return Sched::RegPressure;
984 for (unsigned i = 0; i != NumVals; ++i) {
985 EVT VT = N->getValueType(i);
986 if (VT == MVT::Glue || VT == MVT::Other)
988 if (VT.isFloatingPoint() || VT.isVector())
989 return Sched::Latency;
992 if (!N->isMachineOpcode())
993 return Sched::RegPressure;
995 // Load are scheduled for latency even if there instruction itinerary
997 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
998 const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
1000 if (MCID.getNumDefs() == 0)
1001 return Sched::RegPressure;
1002 if (!Itins->isEmpty() &&
1003 Itins->getOperandCycle(MCID.getSchedClass(), 0) > 2)
1004 return Sched::Latency;
1006 return Sched::RegPressure;
1009 //===----------------------------------------------------------------------===//
1011 //===----------------------------------------------------------------------===//
1013 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
1014 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
1016 default: llvm_unreachable("Unknown condition code!");
1017 case ISD::SETNE: return ARMCC::NE;
1018 case ISD::SETEQ: return ARMCC::EQ;
1019 case ISD::SETGT: return ARMCC::GT;
1020 case ISD::SETGE: return ARMCC::GE;
1021 case ISD::SETLT: return ARMCC::LT;
1022 case ISD::SETLE: return ARMCC::LE;
1023 case ISD::SETUGT: return ARMCC::HI;
1024 case ISD::SETUGE: return ARMCC::HS;
1025 case ISD::SETULT: return ARMCC::LO;
1026 case ISD::SETULE: return ARMCC::LS;
1030 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
1031 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
1032 ARMCC::CondCodes &CondCode2) {
1033 CondCode2 = ARMCC::AL;
1035 default: llvm_unreachable("Unknown FP condition!");
1037 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
1039 case ISD::SETOGT: CondCode = ARMCC::GT; break;
1041 case ISD::SETOGE: CondCode = ARMCC::GE; break;
1042 case ISD::SETOLT: CondCode = ARMCC::MI; break;
1043 case ISD::SETOLE: CondCode = ARMCC::LS; break;
1044 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
1045 case ISD::SETO: CondCode = ARMCC::VC; break;
1046 case ISD::SETUO: CondCode = ARMCC::VS; break;
1047 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
1048 case ISD::SETUGT: CondCode = ARMCC::HI; break;
1049 case ISD::SETUGE: CondCode = ARMCC::PL; break;
1051 case ISD::SETULT: CondCode = ARMCC::LT; break;
1053 case ISD::SETULE: CondCode = ARMCC::LE; break;
1055 case ISD::SETUNE: CondCode = ARMCC::NE; break;
1059 //===----------------------------------------------------------------------===//
1060 // Calling Convention Implementation
1061 //===----------------------------------------------------------------------===//
1063 #include "ARMGenCallingConv.inc"
1065 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1066 /// given CallingConvention value.
1067 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
1069 bool isVarArg) const {
1072 llvm_unreachable("Unsupported calling convention");
1073 case CallingConv::Fast:
1074 if (Subtarget->hasVFP2() && !isVarArg) {
1075 if (!Subtarget->isAAPCS_ABI())
1076 return (Return ? RetFastCC_ARM_APCS : FastCC_ARM_APCS);
1077 // For AAPCS ABI targets, just use VFP variant of the calling convention.
1078 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1081 case CallingConv::C: {
1082 // Use target triple & subtarget features to do actual dispatch.
1083 if (!Subtarget->isAAPCS_ABI())
1084 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1085 else if (Subtarget->hasVFP2() &&
1086 FloatABIType == FloatABI::Hard && !isVarArg)
1087 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1088 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1090 case CallingConv::ARM_AAPCS_VFP:
1091 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1092 case CallingConv::ARM_AAPCS:
1093 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1094 case CallingConv::ARM_APCS:
1095 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1099 /// LowerCallResult - Lower the result values of a call into the
1100 /// appropriate copies out of appropriate physical registers.
1102 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1103 CallingConv::ID CallConv, bool isVarArg,
1104 const SmallVectorImpl<ISD::InputArg> &Ins,
1105 DebugLoc dl, SelectionDAG &DAG,
1106 SmallVectorImpl<SDValue> &InVals) const {
1108 // Assign locations to each value returned by this call.
1109 SmallVector<CCValAssign, 16> RVLocs;
1110 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1111 getTargetMachine(), RVLocs, *DAG.getContext(), Call);
1112 CCInfo.AnalyzeCallResult(Ins,
1113 CCAssignFnForNode(CallConv, /* Return*/ true,
1116 // Copy all of the result registers out of their specified physreg.
1117 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1118 CCValAssign VA = RVLocs[i];
1121 if (VA.needsCustom()) {
1122 // Handle f64 or half of a v2f64.
1123 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1125 Chain = Lo.getValue(1);
1126 InFlag = Lo.getValue(2);
1127 VA = RVLocs[++i]; // skip ahead to next loc
1128 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1130 Chain = Hi.getValue(1);
1131 InFlag = Hi.getValue(2);
1132 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1134 if (VA.getLocVT() == MVT::v2f64) {
1135 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1136 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1137 DAG.getConstant(0, MVT::i32));
1139 VA = RVLocs[++i]; // skip ahead to next loc
1140 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1141 Chain = Lo.getValue(1);
1142 InFlag = Lo.getValue(2);
1143 VA = RVLocs[++i]; // skip ahead to next loc
1144 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1145 Chain = Hi.getValue(1);
1146 InFlag = Hi.getValue(2);
1147 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1148 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1149 DAG.getConstant(1, MVT::i32));
1152 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
1154 Chain = Val.getValue(1);
1155 InFlag = Val.getValue(2);
1158 switch (VA.getLocInfo()) {
1159 default: llvm_unreachable("Unknown loc info!");
1160 case CCValAssign::Full: break;
1161 case CCValAssign::BCvt:
1162 Val = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), Val);
1166 InVals.push_back(Val);
1172 /// LowerMemOpCallTo - Store the argument to the stack.
1174 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
1175 SDValue StackPtr, SDValue Arg,
1176 DebugLoc dl, SelectionDAG &DAG,
1177 const CCValAssign &VA,
1178 ISD::ArgFlagsTy Flags) const {
1179 unsigned LocMemOffset = VA.getLocMemOffset();
1180 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1181 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1182 return DAG.getStore(Chain, dl, Arg, PtrOff,
1183 MachinePointerInfo::getStack(LocMemOffset),
1187 void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
1188 SDValue Chain, SDValue &Arg,
1189 RegsToPassVector &RegsToPass,
1190 CCValAssign &VA, CCValAssign &NextVA,
1192 SmallVector<SDValue, 8> &MemOpChains,
1193 ISD::ArgFlagsTy Flags) const {
1195 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1196 DAG.getVTList(MVT::i32, MVT::i32), Arg);
1197 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
1199 if (NextVA.isRegLoc())
1200 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
1202 assert(NextVA.isMemLoc());
1203 if (StackPtr.getNode() == 0)
1204 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1206 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
1212 /// LowerCall - Lowering a call into a callseq_start <-
1213 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
1216 ARMTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1217 CallingConv::ID CallConv, bool isVarArg,
1219 const SmallVectorImpl<ISD::OutputArg> &Outs,
1220 const SmallVectorImpl<SDValue> &OutVals,
1221 const SmallVectorImpl<ISD::InputArg> &Ins,
1222 DebugLoc dl, SelectionDAG &DAG,
1223 SmallVectorImpl<SDValue> &InVals) const {
1224 MachineFunction &MF = DAG.getMachineFunction();
1225 bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1226 bool IsSibCall = false;
1227 // Disable tail calls if they're not supported.
1228 if (!EnableARMTailCalls && !Subtarget->supportsTailCall())
1231 // Check if it's really possible to do a tail call.
1232 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1233 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1234 Outs, OutVals, Ins, DAG);
1235 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1236 // detected sibcalls.
1243 // Analyze operands of the call, assigning locations to each operand.
1244 SmallVector<CCValAssign, 16> ArgLocs;
1245 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1246 getTargetMachine(), ArgLocs, *DAG.getContext(), Call);
1247 CCInfo.AnalyzeCallOperands(Outs,
1248 CCAssignFnForNode(CallConv, /* Return*/ false,
1251 // Get a count of how many bytes are to be pushed on the stack.
1252 unsigned NumBytes = CCInfo.getNextStackOffset();
1254 // For tail calls, memory operands are available in our caller's stack.
1258 // Adjust the stack pointer for the new arguments...
1259 // These operations are automatically eliminated by the prolog/epilog pass
1261 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1263 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1265 RegsToPassVector RegsToPass;
1266 SmallVector<SDValue, 8> MemOpChains;
1268 // Walk the register/memloc assignments, inserting copies/loads. In the case
1269 // of tail call optimization, arguments are handled later.
1270 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1272 ++i, ++realArgIdx) {
1273 CCValAssign &VA = ArgLocs[i];
1274 SDValue Arg = OutVals[realArgIdx];
1275 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1276 bool isByVal = Flags.isByVal();
1278 // Promote the value if needed.
1279 switch (VA.getLocInfo()) {
1280 default: llvm_unreachable("Unknown loc info!");
1281 case CCValAssign::Full: break;
1282 case CCValAssign::SExt:
1283 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1285 case CCValAssign::ZExt:
1286 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1288 case CCValAssign::AExt:
1289 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1291 case CCValAssign::BCvt:
1292 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1296 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1297 if (VA.needsCustom()) {
1298 if (VA.getLocVT() == MVT::v2f64) {
1299 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1300 DAG.getConstant(0, MVT::i32));
1301 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1302 DAG.getConstant(1, MVT::i32));
1304 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1305 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1307 VA = ArgLocs[++i]; // skip ahead to next loc
1308 if (VA.isRegLoc()) {
1309 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1310 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1312 assert(VA.isMemLoc());
1314 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1315 dl, DAG, VA, Flags));
1318 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1319 StackPtr, MemOpChains, Flags);
1321 } else if (VA.isRegLoc()) {
1322 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1323 } else if (isByVal) {
1324 assert(VA.isMemLoc());
1325 unsigned offset = 0;
1327 // True if this byval aggregate will be split between registers
1329 if (CCInfo.isFirstByValRegValid()) {
1330 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1332 for (i = 0, j = CCInfo.getFirstByValReg(); j < ARM::R4; i++, j++) {
1333 SDValue Const = DAG.getConstant(4*i, MVT::i32);
1334 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
1335 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
1336 MachinePointerInfo(),
1338 MemOpChains.push_back(Load.getValue(1));
1339 RegsToPass.push_back(std::make_pair(j, Load));
1341 offset = ARM::R4 - CCInfo.getFirstByValReg();
1342 CCInfo.clearFirstByValReg();
1345 unsigned LocMemOffset = VA.getLocMemOffset();
1346 SDValue StkPtrOff = DAG.getIntPtrConstant(LocMemOffset);
1347 SDValue Dst = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr,
1349 SDValue SrcOffset = DAG.getIntPtrConstant(4*offset);
1350 SDValue Src = DAG.getNode(ISD::ADD, dl, getPointerTy(), Arg, SrcOffset);
1351 SDValue SizeNode = DAG.getConstant(Flags.getByValSize() - 4*offset,
1353 // TODO: Disable AlwaysInline when it becomes possible
1354 // to emit a nested call sequence.
1355 MemOpChains.push_back(DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode,
1356 Flags.getByValAlign(),
1357 /*isVolatile=*/false,
1358 /*AlwaysInline=*/true,
1359 MachinePointerInfo(0),
1360 MachinePointerInfo(0)));
1362 } else if (!IsSibCall) {
1363 assert(VA.isMemLoc());
1365 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1366 dl, DAG, VA, Flags));
1370 if (!MemOpChains.empty())
1371 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1372 &MemOpChains[0], MemOpChains.size());
1374 // Build a sequence of copy-to-reg nodes chained together with token chain
1375 // and flag operands which copy the outgoing args into the appropriate regs.
1377 // Tail call byval lowering might overwrite argument registers so in case of
1378 // tail call optimization the copies to registers are lowered later.
1380 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1381 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1382 RegsToPass[i].second, InFlag);
1383 InFlag = Chain.getValue(1);
1386 // For tail calls lower the arguments to the 'real' stack slot.
1388 // Force all the incoming stack arguments to be loaded from the stack
1389 // before any new outgoing arguments are stored to the stack, because the
1390 // outgoing stack slots may alias the incoming argument stack slots, and
1391 // the alias isn't otherwise explicit. This is slightly more conservative
1392 // than necessary, because it means that each store effectively depends
1393 // on every argument instead of just those arguments it would clobber.
1395 // Do not flag preceding copytoreg stuff together with the following stuff.
1397 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1398 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1399 RegsToPass[i].second, InFlag);
1400 InFlag = Chain.getValue(1);
1405 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1406 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1407 // node so that legalize doesn't hack it.
1408 bool isDirect = false;
1409 bool isARMFunc = false;
1410 bool isLocalARMFunc = false;
1411 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1413 if (EnableARMLongCalls) {
1414 assert (getTargetMachine().getRelocationModel() == Reloc::Static
1415 && "long-calls with non-static relocation model!");
1416 // Handle a global address or an external symbol. If it's not one of
1417 // those, the target's already in a register, so we don't need to do
1419 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1420 const GlobalValue *GV = G->getGlobal();
1421 // Create a constant pool entry for the callee address
1422 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1423 ARMConstantPoolValue *CPV =
1424 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 0);
1426 // Get the address of the callee into a register
1427 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1428 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1429 Callee = DAG.getLoad(getPointerTy(), dl,
1430 DAG.getEntryNode(), CPAddr,
1431 MachinePointerInfo::getConstantPool(),
1433 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1434 const char *Sym = S->getSymbol();
1436 // Create a constant pool entry for the callee address
1437 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1438 ARMConstantPoolValue *CPV =
1439 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1440 ARMPCLabelIndex, 0);
1441 // Get the address of the callee into a register
1442 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1443 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1444 Callee = DAG.getLoad(getPointerTy(), dl,
1445 DAG.getEntryNode(), CPAddr,
1446 MachinePointerInfo::getConstantPool(),
1449 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1450 const GlobalValue *GV = G->getGlobal();
1452 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1453 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
1454 getTargetMachine().getRelocationModel() != Reloc::Static;
1455 isARMFunc = !Subtarget->isThumb() || isStub;
1456 // ARM call to a local ARM function is predicable.
1457 isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking);
1458 // tBX takes a register source operand.
1459 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1460 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1461 ARMConstantPoolValue *CPV =
1462 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 4);
1463 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1464 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1465 Callee = DAG.getLoad(getPointerTy(), dl,
1466 DAG.getEntryNode(), CPAddr,
1467 MachinePointerInfo::getConstantPool(),
1469 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1470 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1471 getPointerTy(), Callee, PICLabel);
1473 // On ELF targets for PIC code, direct calls should go through the PLT
1474 unsigned OpFlags = 0;
1475 if (Subtarget->isTargetELF() &&
1476 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1477 OpFlags = ARMII::MO_PLT;
1478 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
1480 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1482 bool isStub = Subtarget->isTargetDarwin() &&
1483 getTargetMachine().getRelocationModel() != Reloc::Static;
1484 isARMFunc = !Subtarget->isThumb() || isStub;
1485 // tBX takes a register source operand.
1486 const char *Sym = S->getSymbol();
1487 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1488 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1489 ARMConstantPoolValue *CPV =
1490 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1491 ARMPCLabelIndex, 4);
1492 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1493 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1494 Callee = DAG.getLoad(getPointerTy(), dl,
1495 DAG.getEntryNode(), CPAddr,
1496 MachinePointerInfo::getConstantPool(),
1498 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1499 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1500 getPointerTy(), Callee, PICLabel);
1502 unsigned OpFlags = 0;
1503 // On ELF targets for PIC code, direct calls should go through the PLT
1504 if (Subtarget->isTargetELF() &&
1505 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1506 OpFlags = ARMII::MO_PLT;
1507 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlags);
1511 // FIXME: handle tail calls differently.
1513 if (Subtarget->isThumb()) {
1514 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1515 CallOpc = ARMISD::CALL_NOLINK;
1517 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1519 CallOpc = (isDirect || Subtarget->hasV5TOps())
1520 ? (isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL)
1521 : ARMISD::CALL_NOLINK;
1524 std::vector<SDValue> Ops;
1525 Ops.push_back(Chain);
1526 Ops.push_back(Callee);
1528 // Add argument registers to the end of the list so that they are known live
1530 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1531 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1532 RegsToPass[i].second.getValueType()));
1534 if (InFlag.getNode())
1535 Ops.push_back(InFlag);
1537 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
1539 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, &Ops[0], Ops.size());
1541 // Returns a chain and a flag for retval copy to use.
1542 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
1543 InFlag = Chain.getValue(1);
1545 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1546 DAG.getIntPtrConstant(0, true), InFlag);
1548 InFlag = Chain.getValue(1);
1550 // Handle result values, copying them out of physregs into vregs that we
1552 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins,
1556 /// HandleByVal - Every parameter *after* a byval parameter is passed
1557 /// on the stack. Remember the next parameter register to allocate,
1558 /// and then confiscate the rest of the parameter registers to insure
1561 llvm::ARMTargetLowering::HandleByVal(CCState *State, unsigned &size) const {
1562 unsigned reg = State->AllocateReg(GPRArgRegs, 4);
1563 assert((State->getCallOrPrologue() == Prologue ||
1564 State->getCallOrPrologue() == Call) &&
1565 "unhandled ParmContext");
1566 if ((!State->isFirstByValRegValid()) &&
1567 (ARM::R0 <= reg) && (reg <= ARM::R3)) {
1568 State->setFirstByValReg(reg);
1569 // At a call site, a byval parameter that is split between
1570 // registers and memory needs its size truncated here. In a
1571 // function prologue, such byval parameters are reassembled in
1572 // memory, and are not truncated.
1573 if (State->getCallOrPrologue() == Call) {
1574 unsigned excess = 4 * (ARM::R4 - reg);
1575 assert(size >= excess && "expected larger existing stack allocation");
1579 // Confiscate any remaining parameter registers to preclude their
1580 // assignment to subsequent parameters.
1581 while (State->AllocateReg(GPRArgRegs, 4))
1585 /// MatchingStackOffset - Return true if the given stack call argument is
1586 /// already available in the same position (relatively) of the caller's
1587 /// incoming argument stack.
1589 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1590 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1591 const ARMInstrInfo *TII) {
1592 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1594 if (Arg.getOpcode() == ISD::CopyFromReg) {
1595 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1596 if (!TargetRegisterInfo::isVirtualRegister(VR))
1598 MachineInstr *Def = MRI->getVRegDef(VR);
1601 if (!Flags.isByVal()) {
1602 if (!TII->isLoadFromStackSlot(Def, FI))
1607 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
1608 if (Flags.isByVal())
1609 // ByVal argument is passed in as a pointer but it's now being
1610 // dereferenced. e.g.
1611 // define @foo(%struct.X* %A) {
1612 // tail call @bar(%struct.X* byval %A)
1615 SDValue Ptr = Ld->getBasePtr();
1616 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
1619 FI = FINode->getIndex();
1623 assert(FI != INT_MAX);
1624 if (!MFI->isFixedObjectIndex(FI))
1626 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
1629 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
1630 /// for tail call optimization. Targets which want to do tail call
1631 /// optimization should implement this function.
1633 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1634 CallingConv::ID CalleeCC,
1636 bool isCalleeStructRet,
1637 bool isCallerStructRet,
1638 const SmallVectorImpl<ISD::OutputArg> &Outs,
1639 const SmallVectorImpl<SDValue> &OutVals,
1640 const SmallVectorImpl<ISD::InputArg> &Ins,
1641 SelectionDAG& DAG) const {
1642 const Function *CallerF = DAG.getMachineFunction().getFunction();
1643 CallingConv::ID CallerCC = CallerF->getCallingConv();
1644 bool CCMatch = CallerCC == CalleeCC;
1646 // Look for obvious safe cases to perform tail call optimization that do not
1647 // require ABI changes. This is what gcc calls sibcall.
1649 // Do not sibcall optimize vararg calls unless the call site is not passing
1651 if (isVarArg && !Outs.empty())
1654 // Also avoid sibcall optimization if either caller or callee uses struct
1655 // return semantics.
1656 if (isCalleeStructRet || isCallerStructRet)
1659 // FIXME: Completely disable sibcall for Thumb1 since Thumb1RegisterInfo::
1660 // emitEpilogue is not ready for them. Thumb tail calls also use t2B, as
1661 // the Thumb1 16-bit unconditional branch doesn't have sufficient relocation
1662 // support in the assembler and linker to be used. This would need to be
1663 // fixed to fully support tail calls in Thumb1.
1665 // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take
1666 // LR. This means if we need to reload LR, it takes an extra instructions,
1667 // which outweighs the value of the tail call; but here we don't know yet
1668 // whether LR is going to be used. Probably the right approach is to
1669 // generate the tail call here and turn it back into CALL/RET in
1670 // emitEpilogue if LR is used.
1672 // Thumb1 PIC calls to external symbols use BX, so they can be tail calls,
1673 // but we need to make sure there are enough registers; the only valid
1674 // registers are the 4 used for parameters. We don't currently do this
1676 if (Subtarget->isThumb1Only())
1679 // If the calling conventions do not match, then we'd better make sure the
1680 // results are returned in the same way as what the caller expects.
1682 SmallVector<CCValAssign, 16> RVLocs1;
1683 ARMCCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(),
1684 getTargetMachine(), RVLocs1, *DAG.getContext(), Call);
1685 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
1687 SmallVector<CCValAssign, 16> RVLocs2;
1688 ARMCCState CCInfo2(CallerCC, false, DAG.getMachineFunction(),
1689 getTargetMachine(), RVLocs2, *DAG.getContext(), Call);
1690 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
1692 if (RVLocs1.size() != RVLocs2.size())
1694 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
1695 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
1697 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
1699 if (RVLocs1[i].isRegLoc()) {
1700 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
1703 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
1709 // If the callee takes no arguments then go on to check the results of the
1711 if (!Outs.empty()) {
1712 // Check if stack adjustment is needed. For now, do not do this if any
1713 // argument is passed on the stack.
1714 SmallVector<CCValAssign, 16> ArgLocs;
1715 ARMCCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(),
1716 getTargetMachine(), ArgLocs, *DAG.getContext(), Call);
1717 CCInfo.AnalyzeCallOperands(Outs,
1718 CCAssignFnForNode(CalleeCC, false, isVarArg));
1719 if (CCInfo.getNextStackOffset()) {
1720 MachineFunction &MF = DAG.getMachineFunction();
1722 // Check if the arguments are already laid out in the right way as
1723 // the caller's fixed stack objects.
1724 MachineFrameInfo *MFI = MF.getFrameInfo();
1725 const MachineRegisterInfo *MRI = &MF.getRegInfo();
1726 const ARMInstrInfo *TII =
1727 ((ARMTargetMachine&)getTargetMachine()).getInstrInfo();
1728 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1730 ++i, ++realArgIdx) {
1731 CCValAssign &VA = ArgLocs[i];
1732 EVT RegVT = VA.getLocVT();
1733 SDValue Arg = OutVals[realArgIdx];
1734 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1735 if (VA.getLocInfo() == CCValAssign::Indirect)
1737 if (VA.needsCustom()) {
1738 // f64 and vector types are split into multiple registers or
1739 // register/stack-slot combinations. The types will not match
1740 // the registers; give up on memory f64 refs until we figure
1741 // out what to do about this.
1744 if (!ArgLocs[++i].isRegLoc())
1746 if (RegVT == MVT::v2f64) {
1747 if (!ArgLocs[++i].isRegLoc())
1749 if (!ArgLocs[++i].isRegLoc())
1752 } else if (!VA.isRegLoc()) {
1753 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
1765 ARMTargetLowering::LowerReturn(SDValue Chain,
1766 CallingConv::ID CallConv, bool isVarArg,
1767 const SmallVectorImpl<ISD::OutputArg> &Outs,
1768 const SmallVectorImpl<SDValue> &OutVals,
1769 DebugLoc dl, SelectionDAG &DAG) const {
1771 // CCValAssign - represent the assignment of the return value to a location.
1772 SmallVector<CCValAssign, 16> RVLocs;
1774 // CCState - Info about the registers and stack slots.
1775 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1776 getTargetMachine(), RVLocs, *DAG.getContext(), Call);
1778 // Analyze outgoing return values.
1779 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1782 // If this is the first return lowered for this function, add
1783 // the regs to the liveout set for the function.
1784 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1785 for (unsigned i = 0; i != RVLocs.size(); ++i)
1786 if (RVLocs[i].isRegLoc())
1787 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1792 // Copy the result values into the output registers.
1793 for (unsigned i = 0, realRVLocIdx = 0;
1795 ++i, ++realRVLocIdx) {
1796 CCValAssign &VA = RVLocs[i];
1797 assert(VA.isRegLoc() && "Can only return in registers!");
1799 SDValue Arg = OutVals[realRVLocIdx];
1801 switch (VA.getLocInfo()) {
1802 default: llvm_unreachable("Unknown loc info!");
1803 case CCValAssign::Full: break;
1804 case CCValAssign::BCvt:
1805 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1809 if (VA.needsCustom()) {
1810 if (VA.getLocVT() == MVT::v2f64) {
1811 // Extract the first half and return it in two registers.
1812 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1813 DAG.getConstant(0, MVT::i32));
1814 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
1815 DAG.getVTList(MVT::i32, MVT::i32), Half);
1817 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1818 Flag = Chain.getValue(1);
1819 VA = RVLocs[++i]; // skip ahead to next loc
1820 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1821 HalfGPRs.getValue(1), Flag);
1822 Flag = Chain.getValue(1);
1823 VA = RVLocs[++i]; // skip ahead to next loc
1825 // Extract the 2nd half and fall through to handle it as an f64 value.
1826 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1827 DAG.getConstant(1, MVT::i32));
1829 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
1831 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1832 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
1833 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
1834 Flag = Chain.getValue(1);
1835 VA = RVLocs[++i]; // skip ahead to next loc
1836 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
1839 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1841 // Guarantee that all emitted copies are
1842 // stuck together, avoiding something bad.
1843 Flag = Chain.getValue(1);
1848 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1850 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
1855 bool ARMTargetLowering::isUsedByReturnOnly(SDNode *N) const {
1856 if (N->getNumValues() != 1)
1858 if (!N->hasNUsesOfValue(1, 0))
1861 unsigned NumCopies = 0;
1863 SDNode *Use = *N->use_begin();
1864 if (Use->getOpcode() == ISD::CopyToReg) {
1865 Copies[NumCopies++] = Use;
1866 } else if (Use->getOpcode() == ARMISD::VMOVRRD) {
1867 // f64 returned in a pair of GPRs.
1868 for (SDNode::use_iterator UI = Use->use_begin(), UE = Use->use_end();
1870 if (UI->getOpcode() != ISD::CopyToReg)
1872 Copies[UI.getUse().getResNo()] = *UI;
1875 } else if (Use->getOpcode() == ISD::BITCAST) {
1876 // f32 returned in a single GPR.
1877 if (!Use->hasNUsesOfValue(1, 0))
1879 Use = *Use->use_begin();
1880 if (Use->getOpcode() != ISD::CopyToReg || !Use->hasNUsesOfValue(1, 0))
1882 Copies[NumCopies++] = Use;
1887 if (NumCopies != 1 && NumCopies != 2)
1890 bool HasRet = false;
1891 for (unsigned i = 0; i < NumCopies; ++i) {
1892 SDNode *Copy = Copies[i];
1893 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
1895 if (UI->getOpcode() == ISD::CopyToReg) {
1897 if (Use == Copies[0] || Use == Copies[1])
1901 if (UI->getOpcode() != ARMISD::RET_FLAG)
1910 bool ARMTargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
1911 if (!EnableARMTailCalls)
1914 if (!CI->isTailCall())
1917 return !Subtarget->isThumb1Only();
1920 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
1921 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
1922 // one of the above mentioned nodes. It has to be wrapped because otherwise
1923 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
1924 // be used to form addressing mode. These wrapped nodes will be selected
1926 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
1927 EVT PtrVT = Op.getValueType();
1928 // FIXME there is no actual debug info here
1929 DebugLoc dl = Op.getDebugLoc();
1930 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1932 if (CP->isMachineConstantPoolEntry())
1933 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
1934 CP->getAlignment());
1936 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
1937 CP->getAlignment());
1938 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
1941 unsigned ARMTargetLowering::getJumpTableEncoding() const {
1942 return MachineJumpTableInfo::EK_Inline;
1945 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
1946 SelectionDAG &DAG) const {
1947 MachineFunction &MF = DAG.getMachineFunction();
1948 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1949 unsigned ARMPCLabelIndex = 0;
1950 DebugLoc DL = Op.getDebugLoc();
1951 EVT PtrVT = getPointerTy();
1952 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1953 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1955 if (RelocM == Reloc::Static) {
1956 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
1958 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1959 ARMPCLabelIndex = AFI->createPICLabelUId();
1960 ARMConstantPoolValue *CPV =
1961 ARMConstantPoolConstant::Create(BA, ARMPCLabelIndex,
1962 ARMCP::CPBlockAddress, PCAdj);
1963 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1965 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
1966 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
1967 MachinePointerInfo::getConstantPool(),
1969 if (RelocM == Reloc::Static)
1971 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1972 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
1975 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
1977 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
1978 SelectionDAG &DAG) const {
1979 DebugLoc dl = GA->getDebugLoc();
1980 EVT PtrVT = getPointerTy();
1981 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1982 MachineFunction &MF = DAG.getMachineFunction();
1983 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1984 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1985 ARMConstantPoolValue *CPV =
1986 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
1987 ARMCP::CPValue, PCAdj, ARMCP::TLSGD, true);
1988 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1989 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
1990 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
1991 MachinePointerInfo::getConstantPool(),
1993 SDValue Chain = Argument.getValue(1);
1995 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1996 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
1998 // call __tls_get_addr.
2001 Entry.Node = Argument;
2002 Entry.Ty = (Type *) Type::getInt32Ty(*DAG.getContext());
2003 Args.push_back(Entry);
2004 // FIXME: is there useful debug info available here?
2005 std::pair<SDValue, SDValue> CallResult =
2006 LowerCallTo(Chain, (Type *) Type::getInt32Ty(*DAG.getContext()),
2007 false, false, false, false,
2008 0, CallingConv::C, false, /*isReturnValueUsed=*/true,
2009 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
2010 return CallResult.first;
2013 // Lower ISD::GlobalTLSAddress using the "initial exec" or
2014 // "local exec" model.
2016 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
2017 SelectionDAG &DAG) const {
2018 const GlobalValue *GV = GA->getGlobal();
2019 DebugLoc dl = GA->getDebugLoc();
2021 SDValue Chain = DAG.getEntryNode();
2022 EVT PtrVT = getPointerTy();
2023 // Get the Thread Pointer
2024 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2026 if (GV->isDeclaration()) {
2027 MachineFunction &MF = DAG.getMachineFunction();
2028 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2029 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2030 // Initial exec model.
2031 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2032 ARMConstantPoolValue *CPV =
2033 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2034 ARMCP::CPValue, PCAdj, ARMCP::GOTTPOFF,
2036 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2037 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2038 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2039 MachinePointerInfo::getConstantPool(),
2041 Chain = Offset.getValue(1);
2043 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2044 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
2046 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2047 MachinePointerInfo::getConstantPool(),
2051 ARMConstantPoolValue *CPV =
2052 ARMConstantPoolConstant::Create(GV, ARMCP::TPOFF);
2053 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2054 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2055 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2056 MachinePointerInfo::getConstantPool(),
2060 // The address of the thread local variable is the add of the thread
2061 // pointer with the offset of the variable.
2062 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
2066 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
2067 // TODO: implement the "local dynamic" model
2068 assert(Subtarget->isTargetELF() &&
2069 "TLS not implemented for non-ELF targets");
2070 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
2071 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
2072 // otherwise use the "Local Exec" TLS Model
2073 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
2074 return LowerToTLSGeneralDynamicModel(GA, DAG);
2076 return LowerToTLSExecModels(GA, DAG);
2079 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
2080 SelectionDAG &DAG) const {
2081 EVT PtrVT = getPointerTy();
2082 DebugLoc dl = Op.getDebugLoc();
2083 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2084 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2085 if (RelocM == Reloc::PIC_) {
2086 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
2087 ARMConstantPoolValue *CPV =
2088 ARMConstantPoolConstant::Create(GV,
2089 UseGOTOFF ? ARMCP::GOTOFF : ARMCP::GOT);
2090 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2091 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2092 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
2094 MachinePointerInfo::getConstantPool(),
2096 SDValue Chain = Result.getValue(1);
2097 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
2098 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
2100 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
2101 MachinePointerInfo::getGOT(), false, false, 0);
2105 // If we have T2 ops, we can materialize the address directly via movt/movw
2106 // pair. This is always cheaper in terms of performance, but uses at least 2
2108 if (Subtarget->useMovt() &&
2109 !DAG.getMachineFunction().getFunction()->hasFnAttr(Attribute::OptimizeForSize)) {
2111 // FIXME: Once remat is capable of dealing with instructions with register
2112 // operands, expand this into two nodes.
2113 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2114 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2116 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2117 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2118 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2119 MachinePointerInfo::getConstantPool(),
2124 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
2125 SelectionDAG &DAG) const {
2126 EVT PtrVT = getPointerTy();
2127 DebugLoc dl = Op.getDebugLoc();
2128 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2129 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2130 MachineFunction &MF = DAG.getMachineFunction();
2131 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2133 // FIXME: Enable this for static codegen when tool issues are fixed.
2134 if (Subtarget->useMovt() && RelocM != Reloc::Static &&
2135 !DAG.getMachineFunction().getFunction()->hasFnAttr(Attribute::OptimizeForSize)) {
2137 // FIXME: Once remat is capable of dealing with instructions with register
2138 // operands, expand this into two nodes.
2139 if (RelocM == Reloc::Static)
2140 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2141 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2143 unsigned Wrapper = (RelocM == Reloc::PIC_)
2144 ? ARMISD::WrapperPIC : ARMISD::WrapperDYN;
2145 SDValue Result = DAG.getNode(Wrapper, dl, PtrVT,
2146 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2147 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2148 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
2149 MachinePointerInfo::getGOT(), false, false, 0);
2153 unsigned ARMPCLabelIndex = 0;
2155 if (RelocM == Reloc::Static) {
2156 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2158 ARMPCLabelIndex = AFI->createPICLabelUId();
2159 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
2160 ARMConstantPoolValue *CPV =
2161 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue,
2163 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2165 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2167 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2168 MachinePointerInfo::getConstantPool(),
2170 SDValue Chain = Result.getValue(1);
2172 if (RelocM == Reloc::PIC_) {
2173 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2174 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2177 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2178 Result = DAG.getLoad(PtrVT, dl, Chain, Result, MachinePointerInfo::getGOT(),
2184 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
2185 SelectionDAG &DAG) const {
2186 assert(Subtarget->isTargetELF() &&
2187 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
2188 MachineFunction &MF = DAG.getMachineFunction();
2189 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2190 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2191 EVT PtrVT = getPointerTy();
2192 DebugLoc dl = Op.getDebugLoc();
2193 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2194 ARMConstantPoolValue *CPV =
2195 ARMConstantPoolSymbol::Create(*DAG.getContext(), "_GLOBAL_OFFSET_TABLE_",
2196 ARMPCLabelIndex, PCAdj);
2197 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2198 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2199 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2200 MachinePointerInfo::getConstantPool(),
2202 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2203 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2207 ARMTargetLowering::LowerEH_SJLJ_DISPATCHSETUP(SDValue Op, SelectionDAG &DAG)
2209 DebugLoc dl = Op.getDebugLoc();
2210 return DAG.getNode(ARMISD::EH_SJLJ_DISPATCHSETUP, dl, MVT::Other,
2211 Op.getOperand(0), Op.getOperand(1));
2215 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
2216 DebugLoc dl = Op.getDebugLoc();
2217 SDValue Val = DAG.getConstant(0, MVT::i32);
2218 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl,
2219 DAG.getVTList(MVT::i32, MVT::Other), Op.getOperand(0),
2220 Op.getOperand(1), Val);
2224 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
2225 DebugLoc dl = Op.getDebugLoc();
2226 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
2227 Op.getOperand(1), DAG.getConstant(0, MVT::i32));
2231 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
2232 const ARMSubtarget *Subtarget) const {
2233 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2234 DebugLoc dl = Op.getDebugLoc();
2236 default: return SDValue(); // Don't custom lower most intrinsics.
2237 case Intrinsic::arm_thread_pointer: {
2238 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2239 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2241 case Intrinsic::eh_sjlj_lsda: {
2242 MachineFunction &MF = DAG.getMachineFunction();
2243 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2244 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2245 EVT PtrVT = getPointerTy();
2246 DebugLoc dl = Op.getDebugLoc();
2247 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2249 unsigned PCAdj = (RelocM != Reloc::PIC_)
2250 ? 0 : (Subtarget->isThumb() ? 4 : 8);
2251 ARMConstantPoolValue *CPV =
2252 ARMConstantPoolConstant::Create(MF.getFunction(), ARMPCLabelIndex,
2253 ARMCP::CPLSDA, PCAdj);
2254 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2255 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2257 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2258 MachinePointerInfo::getConstantPool(),
2261 if (RelocM == Reloc::PIC_) {
2262 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2263 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2267 case Intrinsic::arm_neon_vmulls:
2268 case Intrinsic::arm_neon_vmullu: {
2269 unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vmulls)
2270 ? ARMISD::VMULLs : ARMISD::VMULLu;
2271 return DAG.getNode(NewOpc, Op.getDebugLoc(), Op.getValueType(),
2272 Op.getOperand(1), Op.getOperand(2));
2277 static SDValue LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG,
2278 const ARMSubtarget *Subtarget) {
2279 DebugLoc dl = Op.getDebugLoc();
2280 if (!Subtarget->hasDataBarrier()) {
2281 // Some ARMv6 cpus can support data barriers with an mcr instruction.
2282 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2284 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2285 "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
2286 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2287 DAG.getConstant(0, MVT::i32));
2290 SDValue Op5 = Op.getOperand(5);
2291 bool isDeviceBarrier = cast<ConstantSDNode>(Op5)->getZExtValue() != 0;
2292 unsigned isLL = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
2293 unsigned isLS = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
2294 bool isOnlyStoreBarrier = (isLL == 0 && isLS == 0);
2296 ARM_MB::MemBOpt DMBOpt;
2297 if (isDeviceBarrier)
2298 DMBOpt = isOnlyStoreBarrier ? ARM_MB::ST : ARM_MB::SY;
2300 DMBOpt = isOnlyStoreBarrier ? ARM_MB::ISHST : ARM_MB::ISH;
2301 return DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0),
2302 DAG.getConstant(DMBOpt, MVT::i32));
2306 static SDValue LowerATOMIC_FENCE(SDValue Op, SelectionDAG &DAG,
2307 const ARMSubtarget *Subtarget) {
2308 // FIXME: handle "fence singlethread" more efficiently.
2309 DebugLoc dl = Op.getDebugLoc();
2310 if (!Subtarget->hasDataBarrier()) {
2311 // Some ARMv6 cpus can support data barriers with an mcr instruction.
2312 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2314 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2315 "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
2316 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2317 DAG.getConstant(0, MVT::i32));
2320 return DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0),
2321 DAG.getConstant(ARM_MB::ISH, MVT::i32));
2324 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG,
2325 const ARMSubtarget *Subtarget) {
2326 // ARM pre v5TE and Thumb1 does not have preload instructions.
2327 if (!(Subtarget->isThumb2() ||
2328 (!Subtarget->isThumb1Only() && Subtarget->hasV5TEOps())))
2329 // Just preserve the chain.
2330 return Op.getOperand(0);
2332 DebugLoc dl = Op.getDebugLoc();
2333 unsigned isRead = ~cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() & 1;
2335 (!Subtarget->hasV7Ops() || !Subtarget->hasMPExtension()))
2336 // ARMv7 with MP extension has PLDW.
2337 return Op.getOperand(0);
2339 unsigned isData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
2340 if (Subtarget->isThumb()) {
2342 isRead = ~isRead & 1;
2343 isData = ~isData & 1;
2346 return DAG.getNode(ARMISD::PRELOAD, dl, MVT::Other, Op.getOperand(0),
2347 Op.getOperand(1), DAG.getConstant(isRead, MVT::i32),
2348 DAG.getConstant(isData, MVT::i32));
2351 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
2352 MachineFunction &MF = DAG.getMachineFunction();
2353 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
2355 // vastart just stores the address of the VarArgsFrameIndex slot into the
2356 // memory location argument.
2357 DebugLoc dl = Op.getDebugLoc();
2358 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2359 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2360 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
2361 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
2362 MachinePointerInfo(SV), false, false, 0);
2366 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
2367 SDValue &Root, SelectionDAG &DAG,
2368 DebugLoc dl) const {
2369 MachineFunction &MF = DAG.getMachineFunction();
2370 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2372 TargetRegisterClass *RC;
2373 if (AFI->isThumb1OnlyFunction())
2374 RC = ARM::tGPRRegisterClass;
2376 RC = ARM::GPRRegisterClass;
2378 // Transform the arguments stored in physical registers into virtual ones.
2379 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2380 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2383 if (NextVA.isMemLoc()) {
2384 MachineFrameInfo *MFI = MF.getFrameInfo();
2385 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true);
2387 // Create load node to retrieve arguments from the stack.
2388 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2389 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
2390 MachinePointerInfo::getFixedStack(FI),
2393 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
2394 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2397 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
2401 ARMTargetLowering::computeRegArea(CCState &CCInfo, MachineFunction &MF,
2402 unsigned &VARegSize, unsigned &VARegSaveSize)
2405 if (CCInfo.isFirstByValRegValid())
2406 NumGPRs = ARM::R4 - CCInfo.getFirstByValReg();
2408 unsigned int firstUnalloced;
2409 firstUnalloced = CCInfo.getFirstUnallocated(GPRArgRegs,
2410 sizeof(GPRArgRegs) /
2411 sizeof(GPRArgRegs[0]));
2412 NumGPRs = (firstUnalloced <= 3) ? (4 - firstUnalloced) : 0;
2415 unsigned Align = MF.getTarget().getFrameLowering()->getStackAlignment();
2416 VARegSize = NumGPRs * 4;
2417 VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
2420 // The remaining GPRs hold either the beginning of variable-argument
2421 // data, or the beginning of an aggregate passed by value (usuall
2422 // byval). Either way, we allocate stack slots adjacent to the data
2423 // provided by our caller, and store the unallocated registers there.
2424 // If this is a variadic function, the va_list pointer will begin with
2425 // these values; otherwise, this reassembles a (byval) structure that
2426 // was split between registers and memory.
2428 ARMTargetLowering::VarArgStyleRegisters(CCState &CCInfo, SelectionDAG &DAG,
2429 DebugLoc dl, SDValue &Chain,
2430 unsigned ArgOffset) const {
2431 MachineFunction &MF = DAG.getMachineFunction();
2432 MachineFrameInfo *MFI = MF.getFrameInfo();
2433 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2434 unsigned firstRegToSaveIndex;
2435 if (CCInfo.isFirstByValRegValid())
2436 firstRegToSaveIndex = CCInfo.getFirstByValReg() - ARM::R0;
2438 firstRegToSaveIndex = CCInfo.getFirstUnallocated
2439 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
2442 unsigned VARegSize, VARegSaveSize;
2443 computeRegArea(CCInfo, MF, VARegSize, VARegSaveSize);
2444 if (VARegSaveSize) {
2445 // If this function is vararg, store any remaining integer argument regs
2446 // to their spots on the stack so that they may be loaded by deferencing
2447 // the result of va_next.
2448 AFI->setVarArgsRegSaveSize(VARegSaveSize);
2449 AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(VARegSaveSize,
2450 ArgOffset + VARegSaveSize
2453 SDValue FIN = DAG.getFrameIndex(AFI->getVarArgsFrameIndex(),
2456 SmallVector<SDValue, 4> MemOps;
2457 for (; firstRegToSaveIndex < 4; ++firstRegToSaveIndex) {
2458 TargetRegisterClass *RC;
2459 if (AFI->isThumb1OnlyFunction())
2460 RC = ARM::tGPRRegisterClass;
2462 RC = ARM::GPRRegisterClass;
2464 unsigned VReg = MF.addLiveIn(GPRArgRegs[firstRegToSaveIndex], RC);
2465 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2467 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2468 MachinePointerInfo::getFixedStack(AFI->getVarArgsFrameIndex()),
2470 MemOps.push_back(Store);
2471 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
2472 DAG.getConstant(4, getPointerTy()));
2474 if (!MemOps.empty())
2475 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2476 &MemOps[0], MemOps.size());
2478 // This will point to the next argument passed via stack.
2479 AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(4, ArgOffset, true));
2483 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
2484 CallingConv::ID CallConv, bool isVarArg,
2485 const SmallVectorImpl<ISD::InputArg>
2487 DebugLoc dl, SelectionDAG &DAG,
2488 SmallVectorImpl<SDValue> &InVals)
2490 MachineFunction &MF = DAG.getMachineFunction();
2491 MachineFrameInfo *MFI = MF.getFrameInfo();
2493 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2495 // Assign locations to all of the incoming arguments.
2496 SmallVector<CCValAssign, 16> ArgLocs;
2497 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2498 getTargetMachine(), ArgLocs, *DAG.getContext(), Prologue);
2499 CCInfo.AnalyzeFormalArguments(Ins,
2500 CCAssignFnForNode(CallConv, /* Return*/ false,
2503 SmallVector<SDValue, 16> ArgValues;
2504 int lastInsIndex = -1;
2507 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2508 CCValAssign &VA = ArgLocs[i];
2510 // Arguments stored in registers.
2511 if (VA.isRegLoc()) {
2512 EVT RegVT = VA.getLocVT();
2514 if (VA.needsCustom()) {
2515 // f64 and vector types are split up into multiple registers or
2516 // combinations of registers and stack slots.
2517 if (VA.getLocVT() == MVT::v2f64) {
2518 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
2520 VA = ArgLocs[++i]; // skip ahead to next loc
2522 if (VA.isMemLoc()) {
2523 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true);
2524 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2525 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
2526 MachinePointerInfo::getFixedStack(FI),
2529 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
2532 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
2533 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2534 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
2535 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2536 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
2538 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
2541 TargetRegisterClass *RC;
2543 if (RegVT == MVT::f32)
2544 RC = ARM::SPRRegisterClass;
2545 else if (RegVT == MVT::f64)
2546 RC = ARM::DPRRegisterClass;
2547 else if (RegVT == MVT::v2f64)
2548 RC = ARM::QPRRegisterClass;
2549 else if (RegVT == MVT::i32)
2550 RC = (AFI->isThumb1OnlyFunction() ?
2551 ARM::tGPRRegisterClass : ARM::GPRRegisterClass);
2553 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
2555 // Transform the arguments in physical registers into virtual ones.
2556 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2557 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2560 // If this is an 8 or 16-bit value, it is really passed promoted
2561 // to 32 bits. Insert an assert[sz]ext to capture this, then
2562 // truncate to the right size.
2563 switch (VA.getLocInfo()) {
2564 default: llvm_unreachable("Unknown loc info!");
2565 case CCValAssign::Full: break;
2566 case CCValAssign::BCvt:
2567 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
2569 case CCValAssign::SExt:
2570 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2571 DAG.getValueType(VA.getValVT()));
2572 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2574 case CCValAssign::ZExt:
2575 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2576 DAG.getValueType(VA.getValVT()));
2577 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2581 InVals.push_back(ArgValue);
2583 } else { // VA.isRegLoc()
2586 assert(VA.isMemLoc());
2587 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
2589 int index = ArgLocs[i].getValNo();
2591 // Some Ins[] entries become multiple ArgLoc[] entries.
2592 // Process them only once.
2593 if (index != lastInsIndex)
2595 ISD::ArgFlagsTy Flags = Ins[index].Flags;
2596 // FIXME: For now, all byval parameter objects are marked mutable.
2597 // This can be changed with more analysis.
2598 // In case of tail call optimization mark all arguments mutable.
2599 // Since they could be overwritten by lowering of arguments in case of
2601 if (Flags.isByVal()) {
2602 unsigned VARegSize, VARegSaveSize;
2603 computeRegArea(CCInfo, MF, VARegSize, VARegSaveSize);
2604 VarArgStyleRegisters(CCInfo, DAG, dl, Chain, 0);
2605 unsigned Bytes = Flags.getByValSize() - VARegSize;
2606 if (Bytes == 0) Bytes = 1; // Don't create zero-sized stack objects.
2607 int FI = MFI->CreateFixedObject(Bytes,
2608 VA.getLocMemOffset(), false);
2609 InVals.push_back(DAG.getFrameIndex(FI, getPointerTy()));
2611 int FI = MFI->CreateFixedObject(VA.getLocVT().getSizeInBits()/8,
2612 VA.getLocMemOffset(), true);
2614 // Create load nodes to retrieve arguments from the stack.
2615 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2616 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
2617 MachinePointerInfo::getFixedStack(FI),
2620 lastInsIndex = index;
2627 VarArgStyleRegisters(CCInfo, DAG, dl, Chain, CCInfo.getNextStackOffset());
2632 /// isFloatingPointZero - Return true if this is +0.0.
2633 static bool isFloatingPointZero(SDValue Op) {
2634 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
2635 return CFP->getValueAPF().isPosZero();
2636 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
2637 // Maybe this has already been legalized into the constant pool?
2638 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
2639 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
2640 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
2641 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
2642 return CFP->getValueAPF().isPosZero();
2648 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
2649 /// the given operands.
2651 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
2652 SDValue &ARMcc, SelectionDAG &DAG,
2653 DebugLoc dl) const {
2654 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
2655 unsigned C = RHSC->getZExtValue();
2656 if (!isLegalICmpImmediate(C)) {
2657 // Constant does not fit, try adjusting it by one?
2662 if (C != 0x80000000 && isLegalICmpImmediate(C-1)) {
2663 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
2664 RHS = DAG.getConstant(C-1, MVT::i32);
2669 if (C != 0 && isLegalICmpImmediate(C-1)) {
2670 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
2671 RHS = DAG.getConstant(C-1, MVT::i32);
2676 if (C != 0x7fffffff && isLegalICmpImmediate(C+1)) {
2677 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
2678 RHS = DAG.getConstant(C+1, MVT::i32);
2683 if (C != 0xffffffff && isLegalICmpImmediate(C+1)) {
2684 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
2685 RHS = DAG.getConstant(C+1, MVT::i32);
2692 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2693 ARMISD::NodeType CompareType;
2696 CompareType = ARMISD::CMP;
2701 CompareType = ARMISD::CMPZ;
2704 ARMcc = DAG.getConstant(CondCode, MVT::i32);
2705 return DAG.getNode(CompareType, dl, MVT::Glue, LHS, RHS);
2708 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
2710 ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
2711 DebugLoc dl) const {
2713 if (!isFloatingPointZero(RHS))
2714 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Glue, LHS, RHS);
2716 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Glue, LHS);
2717 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Glue, Cmp);
2720 /// duplicateCmp - Glue values can have only one use, so this function
2721 /// duplicates a comparison node.
2723 ARMTargetLowering::duplicateCmp(SDValue Cmp, SelectionDAG &DAG) const {
2724 unsigned Opc = Cmp.getOpcode();
2725 DebugLoc DL = Cmp.getDebugLoc();
2726 if (Opc == ARMISD::CMP || Opc == ARMISD::CMPZ)
2727 return DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
2729 assert(Opc == ARMISD::FMSTAT && "unexpected comparison operation");
2730 Cmp = Cmp.getOperand(0);
2731 Opc = Cmp.getOpcode();
2732 if (Opc == ARMISD::CMPFP)
2733 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
2735 assert(Opc == ARMISD::CMPFPw0 && "unexpected operand of FMSTAT");
2736 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0));
2738 return DAG.getNode(ARMISD::FMSTAT, DL, MVT::Glue, Cmp);
2741 SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
2742 SDValue Cond = Op.getOperand(0);
2743 SDValue SelectTrue = Op.getOperand(1);
2744 SDValue SelectFalse = Op.getOperand(2);
2745 DebugLoc dl = Op.getDebugLoc();
2749 // (select (cmov 1, 0, cond), t, f) -> (cmov t, f, cond)
2750 // (select (cmov 0, 1, cond), t, f) -> (cmov f, t, cond)
2752 if (Cond.getOpcode() == ARMISD::CMOV && Cond.hasOneUse()) {
2753 const ConstantSDNode *CMOVTrue =
2754 dyn_cast<ConstantSDNode>(Cond.getOperand(0));
2755 const ConstantSDNode *CMOVFalse =
2756 dyn_cast<ConstantSDNode>(Cond.getOperand(1));
2758 if (CMOVTrue && CMOVFalse) {
2759 unsigned CMOVTrueVal = CMOVTrue->getZExtValue();
2760 unsigned CMOVFalseVal = CMOVFalse->getZExtValue();
2764 if (CMOVTrueVal == 1 && CMOVFalseVal == 0) {
2766 False = SelectFalse;
2767 } else if (CMOVTrueVal == 0 && CMOVFalseVal == 1) {
2772 if (True.getNode() && False.getNode()) {
2773 EVT VT = Op.getValueType();
2774 SDValue ARMcc = Cond.getOperand(2);
2775 SDValue CCR = Cond.getOperand(3);
2776 SDValue Cmp = duplicateCmp(Cond.getOperand(4), DAG);
2777 assert(True.getValueType() == VT);
2778 return DAG.getNode(ARMISD::CMOV, dl, VT, True, False, ARMcc, CCR, Cmp);
2783 return DAG.getSelectCC(dl, Cond,
2784 DAG.getConstant(0, Cond.getValueType()),
2785 SelectTrue, SelectFalse, ISD::SETNE);
2788 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
2789 EVT VT = Op.getValueType();
2790 SDValue LHS = Op.getOperand(0);
2791 SDValue RHS = Op.getOperand(1);
2792 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
2793 SDValue TrueVal = Op.getOperand(2);
2794 SDValue FalseVal = Op.getOperand(3);
2795 DebugLoc dl = Op.getDebugLoc();
2797 if (LHS.getValueType() == MVT::i32) {
2799 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2800 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2801 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,Cmp);
2804 ARMCC::CondCodes CondCode, CondCode2;
2805 FPCCToARMCC(CC, CondCode, CondCode2);
2807 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
2808 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2809 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2810 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
2812 if (CondCode2 != ARMCC::AL) {
2813 SDValue ARMcc2 = DAG.getConstant(CondCode2, MVT::i32);
2814 // FIXME: Needs another CMP because flag can have but one use.
2815 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
2816 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
2817 Result, TrueVal, ARMcc2, CCR, Cmp2);
2822 /// canChangeToInt - Given the fp compare operand, return true if it is suitable
2823 /// to morph to an integer compare sequence.
2824 static bool canChangeToInt(SDValue Op, bool &SeenZero,
2825 const ARMSubtarget *Subtarget) {
2826 SDNode *N = Op.getNode();
2827 if (!N->hasOneUse())
2828 // Otherwise it requires moving the value from fp to integer registers.
2830 if (!N->getNumValues())
2832 EVT VT = Op.getValueType();
2833 if (VT != MVT::f32 && !Subtarget->isFPBrccSlow())
2834 // f32 case is generally profitable. f64 case only makes sense when vcmpe +
2835 // vmrs are very slow, e.g. cortex-a8.
2838 if (isFloatingPointZero(Op)) {
2842 return ISD::isNormalLoad(N);
2845 static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) {
2846 if (isFloatingPointZero(Op))
2847 return DAG.getConstant(0, MVT::i32);
2849 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op))
2850 return DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2851 Ld->getChain(), Ld->getBasePtr(), Ld->getPointerInfo(),
2852 Ld->isVolatile(), Ld->isNonTemporal(),
2853 Ld->getAlignment());
2855 llvm_unreachable("Unknown VFP cmp argument!");
2858 static void expandf64Toi32(SDValue Op, SelectionDAG &DAG,
2859 SDValue &RetVal1, SDValue &RetVal2) {
2860 if (isFloatingPointZero(Op)) {
2861 RetVal1 = DAG.getConstant(0, MVT::i32);
2862 RetVal2 = DAG.getConstant(0, MVT::i32);
2866 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) {
2867 SDValue Ptr = Ld->getBasePtr();
2868 RetVal1 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2869 Ld->getChain(), Ptr,
2870 Ld->getPointerInfo(),
2871 Ld->isVolatile(), Ld->isNonTemporal(),
2872 Ld->getAlignment());
2874 EVT PtrType = Ptr.getValueType();
2875 unsigned NewAlign = MinAlign(Ld->getAlignment(), 4);
2876 SDValue NewPtr = DAG.getNode(ISD::ADD, Op.getDebugLoc(),
2877 PtrType, Ptr, DAG.getConstant(4, PtrType));
2878 RetVal2 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2879 Ld->getChain(), NewPtr,
2880 Ld->getPointerInfo().getWithOffset(4),
2881 Ld->isVolatile(), Ld->isNonTemporal(),
2886 llvm_unreachable("Unknown VFP cmp argument!");
2889 /// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some
2890 /// f32 and even f64 comparisons to integer ones.
2892 ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const {
2893 SDValue Chain = Op.getOperand(0);
2894 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2895 SDValue LHS = Op.getOperand(2);
2896 SDValue RHS = Op.getOperand(3);
2897 SDValue Dest = Op.getOperand(4);
2898 DebugLoc dl = Op.getDebugLoc();
2900 bool SeenZero = false;
2901 if (canChangeToInt(LHS, SeenZero, Subtarget) &&
2902 canChangeToInt(RHS, SeenZero, Subtarget) &&
2903 // If one of the operand is zero, it's safe to ignore the NaN case since
2904 // we only care about equality comparisons.
2905 (SeenZero || (DAG.isKnownNeverNaN(LHS) && DAG.isKnownNeverNaN(RHS)))) {
2906 // If unsafe fp math optimization is enabled and there are no other uses of
2907 // the CMP operands, and the condition code is EQ or NE, we can optimize it
2908 // to an integer comparison.
2909 if (CC == ISD::SETOEQ)
2911 else if (CC == ISD::SETUNE)
2915 if (LHS.getValueType() == MVT::f32) {
2916 LHS = bitcastf32Toi32(LHS, DAG);
2917 RHS = bitcastf32Toi32(RHS, DAG);
2918 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2919 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2920 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
2921 Chain, Dest, ARMcc, CCR, Cmp);
2926 expandf64Toi32(LHS, DAG, LHS1, LHS2);
2927 expandf64Toi32(RHS, DAG, RHS1, RHS2);
2928 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2929 ARMcc = DAG.getConstant(CondCode, MVT::i32);
2930 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
2931 SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest };
2932 return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops, 7);
2938 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
2939 SDValue Chain = Op.getOperand(0);
2940 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2941 SDValue LHS = Op.getOperand(2);
2942 SDValue RHS = Op.getOperand(3);
2943 SDValue Dest = Op.getOperand(4);
2944 DebugLoc dl = Op.getDebugLoc();
2946 if (LHS.getValueType() == MVT::i32) {
2948 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2949 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2950 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
2951 Chain, Dest, ARMcc, CCR, Cmp);
2954 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
2957 (CC == ISD::SETEQ || CC == ISD::SETOEQ ||
2958 CC == ISD::SETNE || CC == ISD::SETUNE)) {
2959 SDValue Result = OptimizeVFPBrcond(Op, DAG);
2960 if (Result.getNode())
2964 ARMCC::CondCodes CondCode, CondCode2;
2965 FPCCToARMCC(CC, CondCode, CondCode2);
2967 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
2968 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2969 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2970 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
2971 SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp };
2972 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2973 if (CondCode2 != ARMCC::AL) {
2974 ARMcc = DAG.getConstant(CondCode2, MVT::i32);
2975 SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) };
2976 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2981 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
2982 SDValue Chain = Op.getOperand(0);
2983 SDValue Table = Op.getOperand(1);
2984 SDValue Index = Op.getOperand(2);
2985 DebugLoc dl = Op.getDebugLoc();
2987 EVT PTy = getPointerTy();
2988 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
2989 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
2990 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
2991 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
2992 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
2993 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
2994 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
2995 if (Subtarget->isThumb2()) {
2996 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
2997 // which does another jump to the destination. This also makes it easier
2998 // to translate it to TBB / TBH later.
2999 // FIXME: This might not work if the function is extremely large.
3000 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
3001 Addr, Op.getOperand(2), JTI, UId);
3003 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
3004 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
3005 MachinePointerInfo::getJumpTable(),
3007 Chain = Addr.getValue(1);
3008 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
3009 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3011 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
3012 MachinePointerInfo::getJumpTable(), false, false, 0);
3013 Chain = Addr.getValue(1);
3014 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3018 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3019 DebugLoc dl = Op.getDebugLoc();
3022 switch (Op.getOpcode()) {
3024 assert(0 && "Invalid opcode!");
3025 case ISD::FP_TO_SINT:
3026 Opc = ARMISD::FTOSI;
3028 case ISD::FP_TO_UINT:
3029 Opc = ARMISD::FTOUI;
3032 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
3033 return DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
3036 static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3037 EVT VT = Op.getValueType();
3038 DebugLoc dl = Op.getDebugLoc();
3040 assert(Op.getOperand(0).getValueType() == MVT::v4i16 &&
3041 "Invalid type for custom lowering!");
3042 if (VT != MVT::v4f32)
3043 return DAG.UnrollVectorOp(Op.getNode());
3047 switch (Op.getOpcode()) {
3049 assert(0 && "Invalid opcode!");
3050 case ISD::SINT_TO_FP:
3051 CastOpc = ISD::SIGN_EXTEND;
3052 Opc = ISD::SINT_TO_FP;
3054 case ISD::UINT_TO_FP:
3055 CastOpc = ISD::ZERO_EXTEND;
3056 Opc = ISD::UINT_TO_FP;
3060 Op = DAG.getNode(CastOpc, dl, MVT::v4i32, Op.getOperand(0));
3061 return DAG.getNode(Opc, dl, VT, Op);
3064 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3065 EVT VT = Op.getValueType();
3067 return LowerVectorINT_TO_FP(Op, DAG);
3069 DebugLoc dl = Op.getDebugLoc();
3072 switch (Op.getOpcode()) {
3074 assert(0 && "Invalid opcode!");
3075 case ISD::SINT_TO_FP:
3076 Opc = ARMISD::SITOF;
3078 case ISD::UINT_TO_FP:
3079 Opc = ARMISD::UITOF;
3083 Op = DAG.getNode(ISD::BITCAST, dl, MVT::f32, Op.getOperand(0));
3084 return DAG.getNode(Opc, dl, VT, Op);
3087 SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
3088 // Implement fcopysign with a fabs and a conditional fneg.
3089 SDValue Tmp0 = Op.getOperand(0);
3090 SDValue Tmp1 = Op.getOperand(1);
3091 DebugLoc dl = Op.getDebugLoc();
3092 EVT VT = Op.getValueType();
3093 EVT SrcVT = Tmp1.getValueType();
3094 bool InGPR = Tmp0.getOpcode() == ISD::BITCAST ||
3095 Tmp0.getOpcode() == ARMISD::VMOVDRR;
3096 bool UseNEON = !InGPR && Subtarget->hasNEON();
3099 // Use VBSL to copy the sign bit.
3100 unsigned EncodedVal = ARM_AM::createNEONModImm(0x6, 0x80);
3101 SDValue Mask = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v2i32,
3102 DAG.getTargetConstant(EncodedVal, MVT::i32));
3103 EVT OpVT = (VT == MVT::f32) ? MVT::v2i32 : MVT::v1i64;
3105 Mask = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3106 DAG.getNode(ISD::BITCAST, dl, OpVT, Mask),
3107 DAG.getConstant(32, MVT::i32));
3108 else /*if (VT == MVT::f32)*/
3109 Tmp0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp0);
3110 if (SrcVT == MVT::f32) {
3111 Tmp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp1);
3113 Tmp1 = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3114 DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1),
3115 DAG.getConstant(32, MVT::i32));
3116 } else if (VT == MVT::f32)
3117 Tmp1 = DAG.getNode(ARMISD::VSHRu, dl, MVT::v1i64,
3118 DAG.getNode(ISD::BITCAST, dl, MVT::v1i64, Tmp1),
3119 DAG.getConstant(32, MVT::i32));
3120 Tmp0 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp0);
3121 Tmp1 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1);
3123 SDValue AllOnes = DAG.getTargetConstant(ARM_AM::createNEONModImm(0xe, 0xff),
3125 AllOnes = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v8i8, AllOnes);
3126 SDValue MaskNot = DAG.getNode(ISD::XOR, dl, OpVT, Mask,
3127 DAG.getNode(ISD::BITCAST, dl, OpVT, AllOnes));
3129 SDValue Res = DAG.getNode(ISD::OR, dl, OpVT,
3130 DAG.getNode(ISD::AND, dl, OpVT, Tmp1, Mask),
3131 DAG.getNode(ISD::AND, dl, OpVT, Tmp0, MaskNot));
3132 if (VT == MVT::f32) {
3133 Res = DAG.getNode(ISD::BITCAST, dl, MVT::v2f32, Res);
3134 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, Res,
3135 DAG.getConstant(0, MVT::i32));
3137 Res = DAG.getNode(ISD::BITCAST, dl, MVT::f64, Res);
3143 // Bitcast operand 1 to i32.
3144 if (SrcVT == MVT::f64)
3145 Tmp1 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
3146 &Tmp1, 1).getValue(1);
3147 Tmp1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp1);
3149 // Or in the signbit with integer operations.
3150 SDValue Mask1 = DAG.getConstant(0x80000000, MVT::i32);
3151 SDValue Mask2 = DAG.getConstant(0x7fffffff, MVT::i32);
3152 Tmp1 = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp1, Mask1);
3153 if (VT == MVT::f32) {
3154 Tmp0 = DAG.getNode(ISD::AND, dl, MVT::i32,
3155 DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp0), Mask2);
3156 return DAG.getNode(ISD::BITCAST, dl, MVT::f32,
3157 DAG.getNode(ISD::OR, dl, MVT::i32, Tmp0, Tmp1));
3160 // f64: Or the high part with signbit and then combine two parts.
3161 Tmp0 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
3163 SDValue Lo = Tmp0.getValue(0);
3164 SDValue Hi = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp0.getValue(1), Mask2);
3165 Hi = DAG.getNode(ISD::OR, dl, MVT::i32, Hi, Tmp1);
3166 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
3169 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
3170 MachineFunction &MF = DAG.getMachineFunction();
3171 MachineFrameInfo *MFI = MF.getFrameInfo();
3172 MFI->setReturnAddressIsTaken(true);
3174 EVT VT = Op.getValueType();
3175 DebugLoc dl = Op.getDebugLoc();
3176 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3178 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
3179 SDValue Offset = DAG.getConstant(4, MVT::i32);
3180 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
3181 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
3182 MachinePointerInfo(), false, false, 0);
3185 // Return LR, which contains the return address. Mark it an implicit live-in.
3186 unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32));
3187 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
3190 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
3191 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3192 MFI->setFrameAddressIsTaken(true);
3194 EVT VT = Op.getValueType();
3195 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
3196 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3197 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
3198 ? ARM::R7 : ARM::R11;
3199 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
3201 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
3202 MachinePointerInfo(),
3207 /// ExpandBITCAST - If the target supports VFP, this function is called to
3208 /// expand a bit convert where either the source or destination type is i64 to
3209 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
3210 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
3211 /// vectors), since the legalizer won't know what to do with that.
3212 static SDValue ExpandBITCAST(SDNode *N, SelectionDAG &DAG) {
3213 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3214 DebugLoc dl = N->getDebugLoc();
3215 SDValue Op = N->getOperand(0);
3217 // This function is only supposed to be called for i64 types, either as the
3218 // source or destination of the bit convert.
3219 EVT SrcVT = Op.getValueType();
3220 EVT DstVT = N->getValueType(0);
3221 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
3222 "ExpandBITCAST called for non-i64 type");
3224 // Turn i64->f64 into VMOVDRR.
3225 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
3226 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
3227 DAG.getConstant(0, MVT::i32));
3228 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
3229 DAG.getConstant(1, MVT::i32));
3230 return DAG.getNode(ISD::BITCAST, dl, DstVT,
3231 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
3234 // Turn f64->i64 into VMOVRRD.
3235 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
3236 SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
3237 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
3238 // Merge the pieces into a single i64 value.
3239 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
3245 /// getZeroVector - Returns a vector of specified type with all zero elements.
3246 /// Zero vectors are used to represent vector negation and in those cases
3247 /// will be implemented with the NEON VNEG instruction. However, VNEG does
3248 /// not support i64 elements, so sometimes the zero vectors will need to be
3249 /// explicitly constructed. Regardless, use a canonical VMOV to create the
3251 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3252 assert(VT.isVector() && "Expected a vector type");
3253 // The canonical modified immediate encoding of a zero vector is....0!
3254 SDValue EncodedVal = DAG.getTargetConstant(0, MVT::i32);
3255 EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
3256 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal);
3257 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
3260 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
3261 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
3262 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
3263 SelectionDAG &DAG) const {
3264 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3265 EVT VT = Op.getValueType();
3266 unsigned VTBits = VT.getSizeInBits();
3267 DebugLoc dl = Op.getDebugLoc();
3268 SDValue ShOpLo = Op.getOperand(0);
3269 SDValue ShOpHi = Op.getOperand(1);
3270 SDValue ShAmt = Op.getOperand(2);
3272 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
3274 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
3276 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
3277 DAG.getConstant(VTBits, MVT::i32), ShAmt);
3278 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
3279 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
3280 DAG.getConstant(VTBits, MVT::i32));
3281 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
3282 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3283 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
3285 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3286 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
3288 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
3289 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc,
3292 SDValue Ops[2] = { Lo, Hi };
3293 return DAG.getMergeValues(Ops, 2, dl);
3296 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
3297 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
3298 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
3299 SelectionDAG &DAG) const {
3300 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3301 EVT VT = Op.getValueType();
3302 unsigned VTBits = VT.getSizeInBits();
3303 DebugLoc dl = Op.getDebugLoc();
3304 SDValue ShOpLo = Op.getOperand(0);
3305 SDValue ShOpHi = Op.getOperand(1);
3306 SDValue ShAmt = Op.getOperand(2);
3309 assert(Op.getOpcode() == ISD::SHL_PARTS);
3310 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
3311 DAG.getConstant(VTBits, MVT::i32), ShAmt);
3312 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
3313 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
3314 DAG.getConstant(VTBits, MVT::i32));
3315 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
3316 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
3318 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3319 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3320 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
3322 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
3323 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc,
3326 SDValue Ops[2] = { Lo, Hi };
3327 return DAG.getMergeValues(Ops, 2, dl);
3330 SDValue ARMTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
3331 SelectionDAG &DAG) const {
3332 // The rounding mode is in bits 23:22 of the FPSCR.
3333 // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0
3334 // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3)
3335 // so that the shift + and get folded into a bitfield extract.
3336 DebugLoc dl = Op.getDebugLoc();
3337 SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::i32,
3338 DAG.getConstant(Intrinsic::arm_get_fpscr,
3340 SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPSCR,
3341 DAG.getConstant(1U << 22, MVT::i32));
3342 SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds,
3343 DAG.getConstant(22, MVT::i32));
3344 return DAG.getNode(ISD::AND, dl, MVT::i32, RMODE,
3345 DAG.getConstant(3, MVT::i32));
3348 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
3349 const ARMSubtarget *ST) {
3350 EVT VT = N->getValueType(0);
3351 DebugLoc dl = N->getDebugLoc();
3353 if (!ST->hasV6T2Ops())
3356 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
3357 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
3360 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
3361 const ARMSubtarget *ST) {
3362 EVT VT = N->getValueType(0);
3363 DebugLoc dl = N->getDebugLoc();
3368 // Lower vector shifts on NEON to use VSHL.
3369 assert(ST->hasNEON() && "unexpected vector shift");
3371 // Left shifts translate directly to the vshiftu intrinsic.
3372 if (N->getOpcode() == ISD::SHL)
3373 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
3374 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
3375 N->getOperand(0), N->getOperand(1));
3377 assert((N->getOpcode() == ISD::SRA ||
3378 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
3380 // NEON uses the same intrinsics for both left and right shifts. For
3381 // right shifts, the shift amounts are negative, so negate the vector of
3383 EVT ShiftVT = N->getOperand(1).getValueType();
3384 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
3385 getZeroVector(ShiftVT, DAG, dl),
3387 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
3388 Intrinsic::arm_neon_vshifts :
3389 Intrinsic::arm_neon_vshiftu);
3390 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
3391 DAG.getConstant(vshiftInt, MVT::i32),
3392 N->getOperand(0), NegatedCount);
3395 static SDValue Expand64BitShift(SDNode *N, SelectionDAG &DAG,
3396 const ARMSubtarget *ST) {
3397 EVT VT = N->getValueType(0);
3398 DebugLoc dl = N->getDebugLoc();
3400 // We can get here for a node like i32 = ISD::SHL i32, i64
3404 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
3405 "Unknown shift to lower!");
3407 // We only lower SRA, SRL of 1 here, all others use generic lowering.
3408 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
3409 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
3412 // If we are in thumb mode, we don't have RRX.
3413 if (ST->isThumb1Only()) return SDValue();
3415 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
3416 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
3417 DAG.getConstant(0, MVT::i32));
3418 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
3419 DAG.getConstant(1, MVT::i32));
3421 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
3422 // captures the result into a carry flag.
3423 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
3424 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Glue), &Hi, 1);
3426 // The low part is an ARMISD::RRX operand, which shifts the carry in.
3427 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
3429 // Merge the pieces into a single i64 value.
3430 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
3433 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
3434 SDValue TmpOp0, TmpOp1;
3435 bool Invert = false;
3439 SDValue Op0 = Op.getOperand(0);
3440 SDValue Op1 = Op.getOperand(1);
3441 SDValue CC = Op.getOperand(2);
3442 EVT VT = Op.getValueType();
3443 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
3444 DebugLoc dl = Op.getDebugLoc();
3446 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
3447 switch (SetCCOpcode) {
3448 default: llvm_unreachable("Illegal FP comparison"); break;
3450 case ISD::SETNE: Invert = true; // Fallthrough
3452 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
3454 case ISD::SETLT: Swap = true; // Fallthrough
3456 case ISD::SETGT: Opc = ARMISD::VCGT; break;
3458 case ISD::SETLE: Swap = true; // Fallthrough
3460 case ISD::SETGE: Opc = ARMISD::VCGE; break;
3461 case ISD::SETUGE: Swap = true; // Fallthrough
3462 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
3463 case ISD::SETUGT: Swap = true; // Fallthrough
3464 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
3465 case ISD::SETUEQ: Invert = true; // Fallthrough
3467 // Expand this to (OLT | OGT).
3471 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
3472 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
3474 case ISD::SETUO: Invert = true; // Fallthrough
3476 // Expand this to (OLT | OGE).
3480 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
3481 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
3485 // Integer comparisons.
3486 switch (SetCCOpcode) {
3487 default: llvm_unreachable("Illegal integer comparison"); break;
3488 case ISD::SETNE: Invert = true;
3489 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
3490 case ISD::SETLT: Swap = true;
3491 case ISD::SETGT: Opc = ARMISD::VCGT; break;
3492 case ISD::SETLE: Swap = true;
3493 case ISD::SETGE: Opc = ARMISD::VCGE; break;
3494 case ISD::SETULT: Swap = true;
3495 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
3496 case ISD::SETULE: Swap = true;
3497 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
3500 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
3501 if (Opc == ARMISD::VCEQ) {
3504 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
3506 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
3509 // Ignore bitconvert.
3510 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BITCAST)
3511 AndOp = AndOp.getOperand(0);
3513 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
3515 Op0 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(0));
3516 Op1 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(1));
3523 std::swap(Op0, Op1);
3525 // If one of the operands is a constant vector zero, attempt to fold the
3526 // comparison to a specialized compare-against-zero form.
3528 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
3530 else if (ISD::isBuildVectorAllZeros(Op0.getNode())) {
3531 if (Opc == ARMISD::VCGE)
3532 Opc = ARMISD::VCLEZ;
3533 else if (Opc == ARMISD::VCGT)
3534 Opc = ARMISD::VCLTZ;
3539 if (SingleOp.getNode()) {
3542 Result = DAG.getNode(ARMISD::VCEQZ, dl, VT, SingleOp); break;
3544 Result = DAG.getNode(ARMISD::VCGEZ, dl, VT, SingleOp); break;
3546 Result = DAG.getNode(ARMISD::VCLEZ, dl, VT, SingleOp); break;
3548 Result = DAG.getNode(ARMISD::VCGTZ, dl, VT, SingleOp); break;
3550 Result = DAG.getNode(ARMISD::VCLTZ, dl, VT, SingleOp); break;
3552 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
3555 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
3559 Result = DAG.getNOT(dl, Result, VT);
3564 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
3565 /// valid vector constant for a NEON instruction with a "modified immediate"
3566 /// operand (e.g., VMOV). If so, return the encoded value.
3567 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
3568 unsigned SplatBitSize, SelectionDAG &DAG,
3569 EVT &VT, bool is128Bits, NEONModImmType type) {
3570 unsigned OpCmode, Imm;
3572 // SplatBitSize is set to the smallest size that splats the vector, so a
3573 // zero vector will always have SplatBitSize == 8. However, NEON modified
3574 // immediate instructions others than VMOV do not support the 8-bit encoding
3575 // of a zero vector, and the default encoding of zero is supposed to be the
3580 switch (SplatBitSize) {
3582 if (type != VMOVModImm)
3584 // Any 1-byte value is OK. Op=0, Cmode=1110.
3585 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
3588 VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
3592 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
3593 VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
3594 if ((SplatBits & ~0xff) == 0) {
3595 // Value = 0x00nn: Op=x, Cmode=100x.
3600 if ((SplatBits & ~0xff00) == 0) {
3601 // Value = 0xnn00: Op=x, Cmode=101x.
3603 Imm = SplatBits >> 8;
3609 // NEON's 32-bit VMOV supports splat values where:
3610 // * only one byte is nonzero, or
3611 // * the least significant byte is 0xff and the second byte is nonzero, or
3612 // * the least significant 2 bytes are 0xff and the third is nonzero.
3613 VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
3614 if ((SplatBits & ~0xff) == 0) {
3615 // Value = 0x000000nn: Op=x, Cmode=000x.
3620 if ((SplatBits & ~0xff00) == 0) {
3621 // Value = 0x0000nn00: Op=x, Cmode=001x.
3623 Imm = SplatBits >> 8;
3626 if ((SplatBits & ~0xff0000) == 0) {
3627 // Value = 0x00nn0000: Op=x, Cmode=010x.
3629 Imm = SplatBits >> 16;
3632 if ((SplatBits & ~0xff000000) == 0) {
3633 // Value = 0xnn000000: Op=x, Cmode=011x.
3635 Imm = SplatBits >> 24;
3639 // cmode == 0b1100 and cmode == 0b1101 are not supported for VORR or VBIC
3640 if (type == OtherModImm) return SDValue();
3642 if ((SplatBits & ~0xffff) == 0 &&
3643 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
3644 // Value = 0x0000nnff: Op=x, Cmode=1100.
3646 Imm = SplatBits >> 8;
3651 if ((SplatBits & ~0xffffff) == 0 &&
3652 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
3653 // Value = 0x00nnffff: Op=x, Cmode=1101.
3655 Imm = SplatBits >> 16;
3656 SplatBits |= 0xffff;
3660 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
3661 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
3662 // VMOV.I32. A (very) minor optimization would be to replicate the value
3663 // and fall through here to test for a valid 64-bit splat. But, then the
3664 // caller would also need to check and handle the change in size.
3668 if (type != VMOVModImm)
3670 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
3671 uint64_t BitMask = 0xff;
3673 unsigned ImmMask = 1;
3675 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
3676 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
3679 } else if ((SplatBits & BitMask) != 0) {
3685 // Op=1, Cmode=1110.
3688 VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
3693 llvm_unreachable("unexpected size for isNEONModifiedImm");
3697 unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm);
3698 return DAG.getTargetConstant(EncodedVal, MVT::i32);
3701 static bool isVEXTMask(const SmallVectorImpl<int> &M, EVT VT,
3702 bool &ReverseVEXT, unsigned &Imm) {
3703 unsigned NumElts = VT.getVectorNumElements();
3704 ReverseVEXT = false;
3706 // Assume that the first shuffle index is not UNDEF. Fail if it is.
3712 // If this is a VEXT shuffle, the immediate value is the index of the first
3713 // element. The other shuffle indices must be the successive elements after
3715 unsigned ExpectedElt = Imm;
3716 for (unsigned i = 1; i < NumElts; ++i) {
3717 // Increment the expected index. If it wraps around, it may still be
3718 // a VEXT but the source vectors must be swapped.
3720 if (ExpectedElt == NumElts * 2) {
3725 if (M[i] < 0) continue; // ignore UNDEF indices
3726 if (ExpectedElt != static_cast<unsigned>(M[i]))
3730 // Adjust the index value if the source operands will be swapped.
3737 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
3738 /// instruction with the specified blocksize. (The order of the elements
3739 /// within each block of the vector is reversed.)
3740 static bool isVREVMask(const SmallVectorImpl<int> &M, EVT VT,
3741 unsigned BlockSize) {
3742 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
3743 "Only possible block sizes for VREV are: 16, 32, 64");
3745 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3749 unsigned NumElts = VT.getVectorNumElements();
3750 unsigned BlockElts = M[0] + 1;
3751 // If the first shuffle index is UNDEF, be optimistic.
3753 BlockElts = BlockSize / EltSz;
3755 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
3758 for (unsigned i = 0; i < NumElts; ++i) {
3759 if (M[i] < 0) continue; // ignore UNDEF indices
3760 if ((unsigned) M[i] != (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
3767 static bool isVTBLMask(const SmallVectorImpl<int> &M, EVT VT) {
3768 // We can handle <8 x i8> vector shuffles. If the index in the mask is out of
3769 // range, then 0 is placed into the resulting vector. So pretty much any mask
3770 // of 8 elements can work here.
3771 return VT == MVT::v8i8 && M.size() == 8;
3774 static bool isVTRNMask(const SmallVectorImpl<int> &M, EVT VT,
3775 unsigned &WhichResult) {
3776 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3780 unsigned NumElts = VT.getVectorNumElements();
3781 WhichResult = (M[0] == 0 ? 0 : 1);
3782 for (unsigned i = 0; i < NumElts; i += 2) {
3783 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
3784 (M[i+1] >= 0 && (unsigned) M[i+1] != i + NumElts + WhichResult))
3790 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
3791 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3792 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
3793 static bool isVTRN_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3794 unsigned &WhichResult) {
3795 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3799 unsigned NumElts = VT.getVectorNumElements();
3800 WhichResult = (M[0] == 0 ? 0 : 1);
3801 for (unsigned i = 0; i < NumElts; i += 2) {
3802 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
3803 (M[i+1] >= 0 && (unsigned) M[i+1] != i + WhichResult))
3809 static bool isVUZPMask(const SmallVectorImpl<int> &M, EVT VT,
3810 unsigned &WhichResult) {
3811 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3815 unsigned NumElts = VT.getVectorNumElements();
3816 WhichResult = (M[0] == 0 ? 0 : 1);
3817 for (unsigned i = 0; i != NumElts; ++i) {
3818 if (M[i] < 0) continue; // ignore UNDEF indices
3819 if ((unsigned) M[i] != 2 * i + WhichResult)
3823 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3824 if (VT.is64BitVector() && EltSz == 32)
3830 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
3831 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3832 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
3833 static bool isVUZP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3834 unsigned &WhichResult) {
3835 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3839 unsigned Half = VT.getVectorNumElements() / 2;
3840 WhichResult = (M[0] == 0 ? 0 : 1);
3841 for (unsigned j = 0; j != 2; ++j) {
3842 unsigned Idx = WhichResult;
3843 for (unsigned i = 0; i != Half; ++i) {
3844 int MIdx = M[i + j * Half];
3845 if (MIdx >= 0 && (unsigned) MIdx != Idx)
3851 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3852 if (VT.is64BitVector() && EltSz == 32)
3858 static bool isVZIPMask(const SmallVectorImpl<int> &M, EVT VT,
3859 unsigned &WhichResult) {
3860 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3864 unsigned NumElts = VT.getVectorNumElements();
3865 WhichResult = (M[0] == 0 ? 0 : 1);
3866 unsigned Idx = WhichResult * NumElts / 2;
3867 for (unsigned i = 0; i != NumElts; i += 2) {
3868 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
3869 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx + NumElts))
3874 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3875 if (VT.is64BitVector() && EltSz == 32)
3881 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
3882 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3883 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
3884 static bool isVZIP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3885 unsigned &WhichResult) {
3886 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3890 unsigned NumElts = VT.getVectorNumElements();
3891 WhichResult = (M[0] == 0 ? 0 : 1);
3892 unsigned Idx = WhichResult * NumElts / 2;
3893 for (unsigned i = 0; i != NumElts; i += 2) {
3894 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
3895 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx))
3900 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3901 if (VT.is64BitVector() && EltSz == 32)
3907 // If N is an integer constant that can be moved into a register in one
3908 // instruction, return an SDValue of such a constant (will become a MOV
3909 // instruction). Otherwise return null.
3910 static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG,
3911 const ARMSubtarget *ST, DebugLoc dl) {
3913 if (!isa<ConstantSDNode>(N))
3915 Val = cast<ConstantSDNode>(N)->getZExtValue();
3917 if (ST->isThumb1Only()) {
3918 if (Val <= 255 || ~Val <= 255)
3919 return DAG.getConstant(Val, MVT::i32);
3921 if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1)
3922 return DAG.getConstant(Val, MVT::i32);
3927 // If this is a case we can't handle, return null and let the default
3928 // expansion code take care of it.
3929 SDValue ARMTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
3930 const ARMSubtarget *ST) const {
3931 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
3932 DebugLoc dl = Op.getDebugLoc();
3933 EVT VT = Op.getValueType();
3935 APInt SplatBits, SplatUndef;
3936 unsigned SplatBitSize;
3938 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
3939 if (SplatBitSize <= 64) {
3940 // Check if an immediate VMOV works.
3942 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
3943 SplatUndef.getZExtValue(), SplatBitSize,
3944 DAG, VmovVT, VT.is128BitVector(),
3946 if (Val.getNode()) {
3947 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
3948 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
3951 // Try an immediate VMVN.
3952 uint64_t NegatedImm = (~SplatBits).getZExtValue();
3953 Val = isNEONModifiedImm(NegatedImm,
3954 SplatUndef.getZExtValue(), SplatBitSize,
3955 DAG, VmovVT, VT.is128BitVector(),
3957 if (Val.getNode()) {
3958 SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
3959 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
3964 // Scan through the operands to see if only one value is used.
3965 unsigned NumElts = VT.getVectorNumElements();
3966 bool isOnlyLowElement = true;
3967 bool usesOnlyOneValue = true;
3968 bool isConstant = true;
3970 for (unsigned i = 0; i < NumElts; ++i) {
3971 SDValue V = Op.getOperand(i);
3972 if (V.getOpcode() == ISD::UNDEF)
3975 isOnlyLowElement = false;
3976 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
3979 if (!Value.getNode())
3981 else if (V != Value)
3982 usesOnlyOneValue = false;
3985 if (!Value.getNode())
3986 return DAG.getUNDEF(VT);
3988 if (isOnlyLowElement)
3989 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
3991 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3993 // Use VDUP for non-constant splats. For f32 constant splats, reduce to
3994 // i32 and try again.
3995 if (usesOnlyOneValue && EltSize <= 32) {
3997 return DAG.getNode(ARMISD::VDUP, dl, VT, Value);
3998 if (VT.getVectorElementType().isFloatingPoint()) {
3999 SmallVector<SDValue, 8> Ops;
4000 for (unsigned i = 0; i < NumElts; ++i)
4001 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, MVT::i32,
4003 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
4004 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, &Ops[0], NumElts);
4005 Val = LowerBUILD_VECTOR(Val, DAG, ST);
4007 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4009 SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
4011 return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
4014 // If all elements are constants and the case above didn't get hit, fall back
4015 // to the default expansion, which will generate a load from the constant
4020 // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
4022 SDValue shuffle = ReconstructShuffle(Op, DAG);
4023 if (shuffle != SDValue())
4027 // Vectors with 32- or 64-bit elements can be built by directly assigning
4028 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
4029 // will be legalized.
4030 if (EltSize >= 32) {
4031 // Do the expansion with floating-point types, since that is what the VFP
4032 // registers are defined to use, and since i64 is not legal.
4033 EVT EltVT = EVT::getFloatingPointVT(EltSize);
4034 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
4035 SmallVector<SDValue, 8> Ops;
4036 for (unsigned i = 0; i < NumElts; ++i)
4037 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, EltVT, Op.getOperand(i)));
4038 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
4039 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4045 // Gather data to see if the operation can be modelled as a
4046 // shuffle in combination with VEXTs.
4047 SDValue ARMTargetLowering::ReconstructShuffle(SDValue Op,
4048 SelectionDAG &DAG) const {
4049 DebugLoc dl = Op.getDebugLoc();
4050 EVT VT = Op.getValueType();
4051 unsigned NumElts = VT.getVectorNumElements();
4053 SmallVector<SDValue, 2> SourceVecs;
4054 SmallVector<unsigned, 2> MinElts;
4055 SmallVector<unsigned, 2> MaxElts;
4057 for (unsigned i = 0; i < NumElts; ++i) {
4058 SDValue V = Op.getOperand(i);
4059 if (V.getOpcode() == ISD::UNDEF)
4061 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
4062 // A shuffle can only come from building a vector from various
4063 // elements of other vectors.
4065 } else if (V.getOperand(0).getValueType().getVectorElementType() !=
4066 VT.getVectorElementType()) {
4067 // This code doesn't know how to handle shuffles where the vector
4068 // element types do not match (this happens because type legalization
4069 // promotes the return type of EXTRACT_VECTOR_ELT).
4070 // FIXME: It might be appropriate to extend this code to handle
4071 // mismatched types.
4075 // Record this extraction against the appropriate vector if possible...
4076 SDValue SourceVec = V.getOperand(0);
4077 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
4078 bool FoundSource = false;
4079 for (unsigned j = 0; j < SourceVecs.size(); ++j) {
4080 if (SourceVecs[j] == SourceVec) {
4081 if (MinElts[j] > EltNo)
4083 if (MaxElts[j] < EltNo)
4090 // Or record a new source if not...
4092 SourceVecs.push_back(SourceVec);
4093 MinElts.push_back(EltNo);
4094 MaxElts.push_back(EltNo);
4098 // Currently only do something sane when at most two source vectors
4100 if (SourceVecs.size() > 2)
4103 SDValue ShuffleSrcs[2] = {DAG.getUNDEF(VT), DAG.getUNDEF(VT) };
4104 int VEXTOffsets[2] = {0, 0};
4106 // This loop extracts the usage patterns of the source vectors
4107 // and prepares appropriate SDValues for a shuffle if possible.
4108 for (unsigned i = 0; i < SourceVecs.size(); ++i) {
4109 if (SourceVecs[i].getValueType() == VT) {
4110 // No VEXT necessary
4111 ShuffleSrcs[i] = SourceVecs[i];
4114 } else if (SourceVecs[i].getValueType().getVectorNumElements() < NumElts) {
4115 // It probably isn't worth padding out a smaller vector just to
4116 // break it down again in a shuffle.
4120 // Since only 64-bit and 128-bit vectors are legal on ARM and
4121 // we've eliminated the other cases...
4122 assert(SourceVecs[i].getValueType().getVectorNumElements() == 2*NumElts &&
4123 "unexpected vector sizes in ReconstructShuffle");
4125 if (MaxElts[i] - MinElts[i] >= NumElts) {
4126 // Span too large for a VEXT to cope
4130 if (MinElts[i] >= NumElts) {
4131 // The extraction can just take the second half
4132 VEXTOffsets[i] = NumElts;
4133 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4135 DAG.getIntPtrConstant(NumElts));
4136 } else if (MaxElts[i] < NumElts) {
4137 // The extraction can just take the first half
4139 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4141 DAG.getIntPtrConstant(0));
4143 // An actual VEXT is needed
4144 VEXTOffsets[i] = MinElts[i];
4145 SDValue VEXTSrc1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4147 DAG.getIntPtrConstant(0));
4148 SDValue VEXTSrc2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4150 DAG.getIntPtrConstant(NumElts));
4151 ShuffleSrcs[i] = DAG.getNode(ARMISD::VEXT, dl, VT, VEXTSrc1, VEXTSrc2,
4152 DAG.getConstant(VEXTOffsets[i], MVT::i32));
4156 SmallVector<int, 8> Mask;
4158 for (unsigned i = 0; i < NumElts; ++i) {
4159 SDValue Entry = Op.getOperand(i);
4160 if (Entry.getOpcode() == ISD::UNDEF) {
4165 SDValue ExtractVec = Entry.getOperand(0);
4166 int ExtractElt = cast<ConstantSDNode>(Op.getOperand(i)
4167 .getOperand(1))->getSExtValue();
4168 if (ExtractVec == SourceVecs[0]) {
4169 Mask.push_back(ExtractElt - VEXTOffsets[0]);
4171 Mask.push_back(ExtractElt + NumElts - VEXTOffsets[1]);
4175 // Final check before we try to produce nonsense...
4176 if (isShuffleMaskLegal(Mask, VT))
4177 return DAG.getVectorShuffle(VT, dl, ShuffleSrcs[0], ShuffleSrcs[1],
4183 /// isShuffleMaskLegal - Targets can use this to indicate that they only
4184 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
4185 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
4186 /// are assumed to be legal.
4188 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
4190 if (VT.getVectorNumElements() == 4 &&
4191 (VT.is128BitVector() || VT.is64BitVector())) {
4192 unsigned PFIndexes[4];
4193 for (unsigned i = 0; i != 4; ++i) {
4197 PFIndexes[i] = M[i];
4200 // Compute the index in the perfect shuffle table.
4201 unsigned PFTableIndex =
4202 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
4203 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4204 unsigned Cost = (PFEntry >> 30);
4211 unsigned Imm, WhichResult;
4213 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4214 return (EltSize >= 32 ||
4215 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
4216 isVREVMask(M, VT, 64) ||
4217 isVREVMask(M, VT, 32) ||
4218 isVREVMask(M, VT, 16) ||
4219 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
4220 isVTBLMask(M, VT) ||
4221 isVTRNMask(M, VT, WhichResult) ||
4222 isVUZPMask(M, VT, WhichResult) ||
4223 isVZIPMask(M, VT, WhichResult) ||
4224 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
4225 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
4226 isVZIP_v_undef_Mask(M, VT, WhichResult));
4229 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
4230 /// the specified operations to build the shuffle.
4231 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
4232 SDValue RHS, SelectionDAG &DAG,
4234 unsigned OpNum = (PFEntry >> 26) & 0x0F;
4235 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
4236 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
4239 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
4248 OP_VUZPL, // VUZP, left result
4249 OP_VUZPR, // VUZP, right result
4250 OP_VZIPL, // VZIP, left result
4251 OP_VZIPR, // VZIP, right result
4252 OP_VTRNL, // VTRN, left result
4253 OP_VTRNR // VTRN, right result
4256 if (OpNum == OP_COPY) {
4257 if (LHSID == (1*9+2)*9+3) return LHS;
4258 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
4262 SDValue OpLHS, OpRHS;
4263 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
4264 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
4265 EVT VT = OpLHS.getValueType();
4268 default: llvm_unreachable("Unknown shuffle opcode!");
4270 // VREV divides the vector in half and swaps within the half.
4271 if (VT.getVectorElementType() == MVT::i32 ||
4272 VT.getVectorElementType() == MVT::f32)
4273 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
4274 // vrev <4 x i16> -> VREV32
4275 if (VT.getVectorElementType() == MVT::i16)
4276 return DAG.getNode(ARMISD::VREV32, dl, VT, OpLHS);
4277 // vrev <4 x i8> -> VREV16
4278 assert(VT.getVectorElementType() == MVT::i8);
4279 return DAG.getNode(ARMISD::VREV16, dl, VT, OpLHS);
4284 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
4285 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
4289 return DAG.getNode(ARMISD::VEXT, dl, VT,
4291 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
4294 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4295 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
4298 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4299 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
4302 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4303 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
4307 static SDValue LowerVECTOR_SHUFFLEv8i8(SDValue Op,
4308 SmallVectorImpl<int> &ShuffleMask,
4309 SelectionDAG &DAG) {
4310 // Check to see if we can use the VTBL instruction.
4311 SDValue V1 = Op.getOperand(0);
4312 SDValue V2 = Op.getOperand(1);
4313 DebugLoc DL = Op.getDebugLoc();
4315 SmallVector<SDValue, 8> VTBLMask;
4316 for (SmallVectorImpl<int>::iterator
4317 I = ShuffleMask.begin(), E = ShuffleMask.end(); I != E; ++I)
4318 VTBLMask.push_back(DAG.getConstant(*I, MVT::i32));
4320 if (V2.getNode()->getOpcode() == ISD::UNDEF)
4321 return DAG.getNode(ARMISD::VTBL1, DL, MVT::v8i8, V1,
4322 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8,
4325 return DAG.getNode(ARMISD::VTBL2, DL, MVT::v8i8, V1, V2,
4326 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8,
4330 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
4331 SDValue V1 = Op.getOperand(0);
4332 SDValue V2 = Op.getOperand(1);
4333 DebugLoc dl = Op.getDebugLoc();
4334 EVT VT = Op.getValueType();
4335 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
4336 SmallVector<int, 8> ShuffleMask;
4338 // Convert shuffles that are directly supported on NEON to target-specific
4339 // DAG nodes, instead of keeping them as shuffles and matching them again
4340 // during code selection. This is more efficient and avoids the possibility
4341 // of inconsistencies between legalization and selection.
4342 // FIXME: floating-point vectors should be canonicalized to integer vectors
4343 // of the same time so that they get CSEd properly.
4344 SVN->getMask(ShuffleMask);
4346 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4347 if (EltSize <= 32) {
4348 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
4349 int Lane = SVN->getSplatIndex();
4350 // If this is undef splat, generate it via "just" vdup, if possible.
4351 if (Lane == -1) Lane = 0;
4353 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
4354 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
4356 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
4357 DAG.getConstant(Lane, MVT::i32));
4362 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
4365 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
4366 DAG.getConstant(Imm, MVT::i32));
4369 if (isVREVMask(ShuffleMask, VT, 64))
4370 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
4371 if (isVREVMask(ShuffleMask, VT, 32))
4372 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
4373 if (isVREVMask(ShuffleMask, VT, 16))
4374 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
4376 // Check for Neon shuffles that modify both input vectors in place.
4377 // If both results are used, i.e., if there are two shuffles with the same
4378 // source operands and with masks corresponding to both results of one of
4379 // these operations, DAG memoization will ensure that a single node is
4380 // used for both shuffles.
4381 unsigned WhichResult;
4382 if (isVTRNMask(ShuffleMask, VT, WhichResult))
4383 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4384 V1, V2).getValue(WhichResult);
4385 if (isVUZPMask(ShuffleMask, VT, WhichResult))
4386 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4387 V1, V2).getValue(WhichResult);
4388 if (isVZIPMask(ShuffleMask, VT, WhichResult))
4389 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4390 V1, V2).getValue(WhichResult);
4392 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
4393 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4394 V1, V1).getValue(WhichResult);
4395 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
4396 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4397 V1, V1).getValue(WhichResult);
4398 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
4399 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4400 V1, V1).getValue(WhichResult);
4403 // If the shuffle is not directly supported and it has 4 elements, use
4404 // the PerfectShuffle-generated table to synthesize it from other shuffles.
4405 unsigned NumElts = VT.getVectorNumElements();
4407 unsigned PFIndexes[4];
4408 for (unsigned i = 0; i != 4; ++i) {
4409 if (ShuffleMask[i] < 0)
4412 PFIndexes[i] = ShuffleMask[i];
4415 // Compute the index in the perfect shuffle table.
4416 unsigned PFTableIndex =
4417 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
4418 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4419 unsigned Cost = (PFEntry >> 30);
4422 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
4425 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
4426 if (EltSize >= 32) {
4427 // Do the expansion with floating-point types, since that is what the VFP
4428 // registers are defined to use, and since i64 is not legal.
4429 EVT EltVT = EVT::getFloatingPointVT(EltSize);
4430 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
4431 V1 = DAG.getNode(ISD::BITCAST, dl, VecVT, V1);
4432 V2 = DAG.getNode(ISD::BITCAST, dl, VecVT, V2);
4433 SmallVector<SDValue, 8> Ops;
4434 for (unsigned i = 0; i < NumElts; ++i) {
4435 if (ShuffleMask[i] < 0)
4436 Ops.push_back(DAG.getUNDEF(EltVT));
4438 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
4439 ShuffleMask[i] < (int)NumElts ? V1 : V2,
4440 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
4443 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
4444 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4447 if (VT == MVT::v8i8) {
4448 SDValue NewOp = LowerVECTOR_SHUFFLEv8i8(Op, ShuffleMask, DAG);
4449 if (NewOp.getNode())
4456 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4457 // EXTRACT_VECTOR_ELT is legal only for immediate indexes.
4458 SDValue Lane = Op.getOperand(1);
4459 if (!isa<ConstantSDNode>(Lane))
4462 SDValue Vec = Op.getOperand(0);
4463 if (Op.getValueType() == MVT::i32 &&
4464 Vec.getValueType().getVectorElementType().getSizeInBits() < 32) {
4465 DebugLoc dl = Op.getDebugLoc();
4466 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
4472 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
4473 // The only time a CONCAT_VECTORS operation can have legal types is when
4474 // two 64-bit vectors are concatenated to a 128-bit vector.
4475 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
4476 "unexpected CONCAT_VECTORS");
4477 DebugLoc dl = Op.getDebugLoc();
4478 SDValue Val = DAG.getUNDEF(MVT::v2f64);
4479 SDValue Op0 = Op.getOperand(0);
4480 SDValue Op1 = Op.getOperand(1);
4481 if (Op0.getOpcode() != ISD::UNDEF)
4482 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
4483 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op0),
4484 DAG.getIntPtrConstant(0));
4485 if (Op1.getOpcode() != ISD::UNDEF)
4486 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
4487 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op1),
4488 DAG.getIntPtrConstant(1));
4489 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Val);
4492 /// isExtendedBUILD_VECTOR - Check if N is a constant BUILD_VECTOR where each
4493 /// element has been zero/sign-extended, depending on the isSigned parameter,
4494 /// from an integer type half its size.
4495 static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG,
4497 // A v2i64 BUILD_VECTOR will have been legalized to a BITCAST from v4i32.
4498 EVT VT = N->getValueType(0);
4499 if (VT == MVT::v2i64 && N->getOpcode() == ISD::BITCAST) {
4500 SDNode *BVN = N->getOperand(0).getNode();
4501 if (BVN->getValueType(0) != MVT::v4i32 ||
4502 BVN->getOpcode() != ISD::BUILD_VECTOR)
4504 unsigned LoElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
4505 unsigned HiElt = 1 - LoElt;
4506 ConstantSDNode *Lo0 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt));
4507 ConstantSDNode *Hi0 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt));
4508 ConstantSDNode *Lo1 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt+2));
4509 ConstantSDNode *Hi1 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt+2));
4510 if (!Lo0 || !Hi0 || !Lo1 || !Hi1)
4513 if (Hi0->getSExtValue() == Lo0->getSExtValue() >> 32 &&
4514 Hi1->getSExtValue() == Lo1->getSExtValue() >> 32)
4517 if (Hi0->isNullValue() && Hi1->isNullValue())
4523 if (N->getOpcode() != ISD::BUILD_VECTOR)
4526 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
4527 SDNode *Elt = N->getOperand(i).getNode();
4528 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
4529 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4530 unsigned HalfSize = EltSize / 2;
4532 if (!isIntN(HalfSize, C->getSExtValue()))
4535 if (!isUIntN(HalfSize, C->getZExtValue()))
4546 /// isSignExtended - Check if a node is a vector value that is sign-extended
4547 /// or a constant BUILD_VECTOR with sign-extended elements.
4548 static bool isSignExtended(SDNode *N, SelectionDAG &DAG) {
4549 if (N->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N))
4551 if (isExtendedBUILD_VECTOR(N, DAG, true))
4556 /// isZeroExtended - Check if a node is a vector value that is zero-extended
4557 /// or a constant BUILD_VECTOR with zero-extended elements.
4558 static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) {
4559 if (N->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N))
4561 if (isExtendedBUILD_VECTOR(N, DAG, false))
4566 /// SkipExtension - For a node that is a SIGN_EXTEND, ZERO_EXTEND, extending
4567 /// load, or BUILD_VECTOR with extended elements, return the unextended value.
4568 static SDValue SkipExtension(SDNode *N, SelectionDAG &DAG) {
4569 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
4570 return N->getOperand(0);
4571 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
4572 return DAG.getLoad(LD->getMemoryVT(), N->getDebugLoc(), LD->getChain(),
4573 LD->getBasePtr(), LD->getPointerInfo(), LD->isVolatile(),
4574 LD->isNonTemporal(), LD->getAlignment());
4575 // Otherwise, the value must be a BUILD_VECTOR. For v2i64, it will
4576 // have been legalized as a BITCAST from v4i32.
4577 if (N->getOpcode() == ISD::BITCAST) {
4578 SDNode *BVN = N->getOperand(0).getNode();
4579 assert(BVN->getOpcode() == ISD::BUILD_VECTOR &&
4580 BVN->getValueType(0) == MVT::v4i32 && "expected v4i32 BUILD_VECTOR");
4581 unsigned LowElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
4582 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(), MVT::v2i32,
4583 BVN->getOperand(LowElt), BVN->getOperand(LowElt+2));
4585 // Construct a new BUILD_VECTOR with elements truncated to half the size.
4586 assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR");
4587 EVT VT = N->getValueType(0);
4588 unsigned EltSize = VT.getVectorElementType().getSizeInBits() / 2;
4589 unsigned NumElts = VT.getVectorNumElements();
4590 MVT TruncVT = MVT::getIntegerVT(EltSize);
4591 SmallVector<SDValue, 8> Ops;
4592 for (unsigned i = 0; i != NumElts; ++i) {
4593 ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i));
4594 const APInt &CInt = C->getAPIntValue();
4595 Ops.push_back(DAG.getConstant(CInt.trunc(EltSize), TruncVT));
4597 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
4598 MVT::getVectorVT(TruncVT, NumElts), Ops.data(), NumElts);
4601 static bool isAddSubSExt(SDNode *N, SelectionDAG &DAG) {
4602 unsigned Opcode = N->getOpcode();
4603 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
4604 SDNode *N0 = N->getOperand(0).getNode();
4605 SDNode *N1 = N->getOperand(1).getNode();
4606 return N0->hasOneUse() && N1->hasOneUse() &&
4607 isSignExtended(N0, DAG) && isSignExtended(N1, DAG);
4612 static bool isAddSubZExt(SDNode *N, SelectionDAG &DAG) {
4613 unsigned Opcode = N->getOpcode();
4614 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
4615 SDNode *N0 = N->getOperand(0).getNode();
4616 SDNode *N1 = N->getOperand(1).getNode();
4617 return N0->hasOneUse() && N1->hasOneUse() &&
4618 isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG);
4623 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
4624 // Multiplications are only custom-lowered for 128-bit vectors so that
4625 // VMULL can be detected. Otherwise v2i64 multiplications are not legal.
4626 EVT VT = Op.getValueType();
4627 assert(VT.is128BitVector() && "unexpected type for custom-lowering ISD::MUL");
4628 SDNode *N0 = Op.getOperand(0).getNode();
4629 SDNode *N1 = Op.getOperand(1).getNode();
4630 unsigned NewOpc = 0;
4632 bool isN0SExt = isSignExtended(N0, DAG);
4633 bool isN1SExt = isSignExtended(N1, DAG);
4634 if (isN0SExt && isN1SExt)
4635 NewOpc = ARMISD::VMULLs;
4637 bool isN0ZExt = isZeroExtended(N0, DAG);
4638 bool isN1ZExt = isZeroExtended(N1, DAG);
4639 if (isN0ZExt && isN1ZExt)
4640 NewOpc = ARMISD::VMULLu;
4641 else if (isN1SExt || isN1ZExt) {
4642 // Look for (s/zext A + s/zext B) * (s/zext C). We want to turn these
4643 // into (s/zext A * s/zext C) + (s/zext B * s/zext C)
4644 if (isN1SExt && isAddSubSExt(N0, DAG)) {
4645 NewOpc = ARMISD::VMULLs;
4647 } else if (isN1ZExt && isAddSubZExt(N0, DAG)) {
4648 NewOpc = ARMISD::VMULLu;
4650 } else if (isN0ZExt && isAddSubZExt(N1, DAG)) {
4652 NewOpc = ARMISD::VMULLu;
4658 if (VT == MVT::v2i64)
4659 // Fall through to expand this. It is not legal.
4662 // Other vector multiplications are legal.
4667 // Legalize to a VMULL instruction.
4668 DebugLoc DL = Op.getDebugLoc();
4670 SDValue Op1 = SkipExtension(N1, DAG);
4672 Op0 = SkipExtension(N0, DAG);
4673 assert(Op0.getValueType().is64BitVector() &&
4674 Op1.getValueType().is64BitVector() &&
4675 "unexpected types for extended operands to VMULL");
4676 return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
4679 // Optimizing (zext A + zext B) * C, to (VMULL A, C) + (VMULL B, C) during
4680 // isel lowering to take advantage of no-stall back to back vmul + vmla.
4687 SDValue N00 = SkipExtension(N0->getOperand(0).getNode(), DAG);
4688 SDValue N01 = SkipExtension(N0->getOperand(1).getNode(), DAG);
4689 EVT Op1VT = Op1.getValueType();
4690 return DAG.getNode(N0->getOpcode(), DL, VT,
4691 DAG.getNode(NewOpc, DL, VT,
4692 DAG.getNode(ISD::BITCAST, DL, Op1VT, N00), Op1),
4693 DAG.getNode(NewOpc, DL, VT,
4694 DAG.getNode(ISD::BITCAST, DL, Op1VT, N01), Op1));
4698 LowerSDIV_v4i8(SDValue X, SDValue Y, DebugLoc dl, SelectionDAG &DAG) {
4700 // float4 xf = vcvt_f32_s32(vmovl_s16(a.lo));
4701 // float4 yf = vcvt_f32_s32(vmovl_s16(b.lo));
4702 X = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, X);
4703 Y = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, Y);
4704 X = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, X);
4705 Y = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, Y);
4706 // Get reciprocal estimate.
4707 // float4 recip = vrecpeq_f32(yf);
4708 Y = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4709 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), Y);
4710 // Because char has a smaller range than uchar, we can actually get away
4711 // without any newton steps. This requires that we use a weird bias
4712 // of 0xb000, however (again, this has been exhaustively tested).
4713 // float4 result = as_float4(as_int4(xf*recip) + 0xb000);
4714 X = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, X, Y);
4715 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, X);
4716 Y = DAG.getConstant(0xb000, MVT::i32);
4717 Y = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Y, Y, Y, Y);
4718 X = DAG.getNode(ISD::ADD, dl, MVT::v4i32, X, Y);
4719 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, X);
4720 // Convert back to short.
4721 X = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, X);
4722 X = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, X);
4727 LowerSDIV_v4i16(SDValue N0, SDValue N1, DebugLoc dl, SelectionDAG &DAG) {
4729 // Convert to float.
4730 // float4 yf = vcvt_f32_s32(vmovl_s16(y));
4731 // float4 xf = vcvt_f32_s32(vmovl_s16(x));
4732 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N0);
4733 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N1);
4734 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
4735 N1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
4737 // Use reciprocal estimate and one refinement step.
4738 // float4 recip = vrecpeq_f32(yf);
4739 // recip *= vrecpsq_f32(yf, recip);
4740 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4741 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), N1);
4742 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4743 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
4745 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
4746 // Because short has a smaller range than ushort, we can actually get away
4747 // with only a single newton step. This requires that we use a weird bias
4748 // of 89, however (again, this has been exhaustively tested).
4749 // float4 result = as_float4(as_int4(xf*recip) + 0x89);
4750 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
4751 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
4752 N1 = DAG.getConstant(0x89, MVT::i32);
4753 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
4754 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
4755 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
4756 // Convert back to integer and return.
4757 // return vmovn_s32(vcvt_s32_f32(result));
4758 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
4759 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
4763 static SDValue LowerSDIV(SDValue Op, SelectionDAG &DAG) {
4764 EVT VT = Op.getValueType();
4765 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
4766 "unexpected type for custom-lowering ISD::SDIV");
4768 DebugLoc dl = Op.getDebugLoc();
4769 SDValue N0 = Op.getOperand(0);
4770 SDValue N1 = Op.getOperand(1);
4773 if (VT == MVT::v8i8) {
4774 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N0);
4775 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N1);
4777 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
4778 DAG.getIntPtrConstant(4));
4779 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
4780 DAG.getIntPtrConstant(4));
4781 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
4782 DAG.getIntPtrConstant(0));
4783 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
4784 DAG.getIntPtrConstant(0));
4786 N0 = LowerSDIV_v4i8(N0, N1, dl, DAG); // v4i16
4787 N2 = LowerSDIV_v4i8(N2, N3, dl, DAG); // v4i16
4789 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
4790 N0 = LowerCONCAT_VECTORS(N0, DAG);
4792 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v8i8, N0);
4795 return LowerSDIV_v4i16(N0, N1, dl, DAG);
4798 static SDValue LowerUDIV(SDValue Op, SelectionDAG &DAG) {
4799 EVT VT = Op.getValueType();
4800 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
4801 "unexpected type for custom-lowering ISD::UDIV");
4803 DebugLoc dl = Op.getDebugLoc();
4804 SDValue N0 = Op.getOperand(0);
4805 SDValue N1 = Op.getOperand(1);
4808 if (VT == MVT::v8i8) {
4809 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N0);
4810 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N1);
4812 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
4813 DAG.getIntPtrConstant(4));
4814 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
4815 DAG.getIntPtrConstant(4));
4816 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
4817 DAG.getIntPtrConstant(0));
4818 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
4819 DAG.getIntPtrConstant(0));
4821 N0 = LowerSDIV_v4i16(N0, N1, dl, DAG); // v4i16
4822 N2 = LowerSDIV_v4i16(N2, N3, dl, DAG); // v4i16
4824 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
4825 N0 = LowerCONCAT_VECTORS(N0, DAG);
4827 N0 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v8i8,
4828 DAG.getConstant(Intrinsic::arm_neon_vqmovnsu, MVT::i32),
4833 // v4i16 sdiv ... Convert to float.
4834 // float4 yf = vcvt_f32_s32(vmovl_u16(y));
4835 // float4 xf = vcvt_f32_s32(vmovl_u16(x));
4836 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N0);
4837 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N1);
4838 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
4839 SDValue BN1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
4841 // Use reciprocal estimate and two refinement steps.
4842 // float4 recip = vrecpeq_f32(yf);
4843 // recip *= vrecpsq_f32(yf, recip);
4844 // recip *= vrecpsq_f32(yf, recip);
4845 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4846 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), BN1);
4847 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4848 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
4850 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
4851 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4852 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
4854 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
4855 // Simply multiplying by the reciprocal estimate can leave us a few ulps
4856 // too low, so we add 2 ulps (exhaustive testing shows that this is enough,
4857 // and that it will never cause us to return an answer too large).
4858 // float4 result = as_float4(as_int4(xf*recip) + 2);
4859 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
4860 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
4861 N1 = DAG.getConstant(2, MVT::i32);
4862 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
4863 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
4864 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
4865 // Convert back to integer and return.
4866 // return vmovn_u32(vcvt_s32_f32(result));
4867 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
4868 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
4872 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
4873 EVT VT = Op.getNode()->getValueType(0);
4874 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
4877 bool ExtraOp = false;
4878 switch (Op.getOpcode()) {
4879 default: assert(0 && "Invalid code");
4880 case ISD::ADDC: Opc = ARMISD::ADDC; break;
4881 case ISD::ADDE: Opc = ARMISD::ADDE; ExtraOp = true; break;
4882 case ISD::SUBC: Opc = ARMISD::SUBC; break;
4883 case ISD::SUBE: Opc = ARMISD::SUBE; ExtraOp = true; break;
4887 return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
4889 return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
4890 Op.getOperand(1), Op.getOperand(2));
4893 static SDValue LowerAtomicLoadStore(SDValue Op, SelectionDAG &DAG) {
4894 // Monotonic load/store is legal for all targets
4895 if (cast<AtomicSDNode>(Op)->getOrdering() <= Monotonic)
4898 // Aquire/Release load/store is not legal for targets without a
4899 // dmb or equivalent available.
4905 ReplaceATOMIC_OP_64(SDNode *Node, SmallVectorImpl<SDValue>& Results,
4906 SelectionDAG &DAG, unsigned NewOp) {
4907 DebugLoc dl = Node->getDebugLoc();
4908 assert (Node->getValueType(0) == MVT::i64 &&
4909 "Only know how to expand i64 atomics");
4911 SmallVector<SDValue, 6> Ops;
4912 Ops.push_back(Node->getOperand(0)); // Chain
4913 Ops.push_back(Node->getOperand(1)); // Ptr
4915 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
4916 Node->getOperand(2), DAG.getIntPtrConstant(0)));
4917 // High part of Val1
4918 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
4919 Node->getOperand(2), DAG.getIntPtrConstant(1)));
4920 if (NewOp == ARMISD::ATOMCMPXCHG64_DAG) {
4921 // High part of Val1
4922 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
4923 Node->getOperand(3), DAG.getIntPtrConstant(0)));
4924 // High part of Val2
4925 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
4926 Node->getOperand(3), DAG.getIntPtrConstant(1)));
4928 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
4930 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops.data(), Ops.size(), MVT::i64,
4931 cast<MemSDNode>(Node)->getMemOperand());
4932 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1) };
4933 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
4934 Results.push_back(Result.getValue(2));
4937 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
4938 switch (Op.getOpcode()) {
4939 default: llvm_unreachable("Don't know how to custom lower this!");
4940 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
4941 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
4942 case ISD::GlobalAddress:
4943 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
4944 LowerGlobalAddressELF(Op, DAG);
4945 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
4946 case ISD::SELECT: return LowerSELECT(Op, DAG);
4947 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
4948 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
4949 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
4950 case ISD::VASTART: return LowerVASTART(Op, DAG);
4951 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG, Subtarget);
4952 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, DAG, Subtarget);
4953 case ISD::PREFETCH: return LowerPREFETCH(Op, DAG, Subtarget);
4954 case ISD::SINT_TO_FP:
4955 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
4956 case ISD::FP_TO_SINT:
4957 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
4958 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
4959 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
4960 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
4961 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
4962 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
4963 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
4964 case ISD::EH_SJLJ_DISPATCHSETUP: return LowerEH_SJLJ_DISPATCHSETUP(Op, DAG);
4965 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
4967 case ISD::BITCAST: return ExpandBITCAST(Op.getNode(), DAG);
4970 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
4971 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
4972 case ISD::SRL_PARTS:
4973 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
4974 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
4975 case ISD::SETCC: return LowerVSETCC(Op, DAG);
4976 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG, Subtarget);
4977 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
4978 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
4979 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
4980 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
4981 case ISD::MUL: return LowerMUL(Op, DAG);
4982 case ISD::SDIV: return LowerSDIV(Op, DAG);
4983 case ISD::UDIV: return LowerUDIV(Op, DAG);
4987 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
4988 case ISD::ATOMIC_LOAD:
4989 case ISD::ATOMIC_STORE: return LowerAtomicLoadStore(Op, DAG);
4994 /// ReplaceNodeResults - Replace the results of node with an illegal result
4995 /// type with new values built out of custom code.
4996 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
4997 SmallVectorImpl<SDValue>&Results,
4998 SelectionDAG &DAG) const {
5000 switch (N->getOpcode()) {
5002 llvm_unreachable("Don't know how to custom expand this!");
5005 Res = ExpandBITCAST(N, DAG);
5009 Res = Expand64BitShift(N, DAG, Subtarget);
5011 case ISD::ATOMIC_LOAD_ADD:
5012 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMADD64_DAG);
5014 case ISD::ATOMIC_LOAD_AND:
5015 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMAND64_DAG);
5017 case ISD::ATOMIC_LOAD_NAND:
5018 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMNAND64_DAG);
5020 case ISD::ATOMIC_LOAD_OR:
5021 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMOR64_DAG);
5023 case ISD::ATOMIC_LOAD_SUB:
5024 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMSUB64_DAG);
5026 case ISD::ATOMIC_LOAD_XOR:
5027 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMXOR64_DAG);
5029 case ISD::ATOMIC_SWAP:
5030 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMSWAP64_DAG);
5032 case ISD::ATOMIC_CMP_SWAP:
5033 ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMCMPXCHG64_DAG);
5037 Results.push_back(Res);
5040 //===----------------------------------------------------------------------===//
5041 // ARM Scheduler Hooks
5042 //===----------------------------------------------------------------------===//
5045 ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
5046 MachineBasicBlock *BB,
5047 unsigned Size) const {
5048 unsigned dest = MI->getOperand(0).getReg();
5049 unsigned ptr = MI->getOperand(1).getReg();
5050 unsigned oldval = MI->getOperand(2).getReg();
5051 unsigned newval = MI->getOperand(3).getReg();
5052 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5053 DebugLoc dl = MI->getDebugLoc();
5054 bool isThumb2 = Subtarget->isThumb2();
5056 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5058 MRI.createVirtualRegister(isThumb2 ? ARM::rGPRRegisterClass
5059 : ARM::GPRRegisterClass);
5062 MRI.constrainRegClass(dest, ARM::rGPRRegisterClass);
5063 MRI.constrainRegClass(oldval, ARM::rGPRRegisterClass);
5064 MRI.constrainRegClass(newval, ARM::rGPRRegisterClass);
5067 unsigned ldrOpc, strOpc;
5069 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
5071 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
5072 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
5075 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
5076 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
5079 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
5080 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
5084 MachineFunction *MF = BB->getParent();
5085 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5086 MachineFunction::iterator It = BB;
5087 ++It; // insert the new blocks after the current block
5089 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
5090 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
5091 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5092 MF->insert(It, loop1MBB);
5093 MF->insert(It, loop2MBB);
5094 MF->insert(It, exitMBB);
5096 // Transfer the remainder of BB and its successor edges to exitMBB.
5097 exitMBB->splice(exitMBB->begin(), BB,
5098 llvm::next(MachineBasicBlock::iterator(MI)),
5100 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5104 // fallthrough --> loop1MBB
5105 BB->addSuccessor(loop1MBB);
5108 // ldrex dest, [ptr]
5112 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
5113 if (ldrOpc == ARM::t2LDREX)
5115 AddDefaultPred(MIB);
5116 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
5117 .addReg(dest).addReg(oldval));
5118 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5119 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5120 BB->addSuccessor(loop2MBB);
5121 BB->addSuccessor(exitMBB);
5124 // strex scratch, newval, [ptr]
5128 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval).addReg(ptr);
5129 if (strOpc == ARM::t2STREX)
5131 AddDefaultPred(MIB);
5132 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5133 .addReg(scratch).addImm(0));
5134 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5135 .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5136 BB->addSuccessor(loop1MBB);
5137 BB->addSuccessor(exitMBB);
5143 MI->eraseFromParent(); // The instruction is gone now.
5149 ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
5150 unsigned Size, unsigned BinOpcode) const {
5151 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
5152 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5154 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5155 MachineFunction *MF = BB->getParent();
5156 MachineFunction::iterator It = BB;
5159 unsigned dest = MI->getOperand(0).getReg();
5160 unsigned ptr = MI->getOperand(1).getReg();
5161 unsigned incr = MI->getOperand(2).getReg();
5162 DebugLoc dl = MI->getDebugLoc();
5163 bool isThumb2 = Subtarget->isThumb2();
5165 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5167 MRI.constrainRegClass(dest, ARM::rGPRRegisterClass);
5168 MRI.constrainRegClass(ptr, ARM::rGPRRegisterClass);
5171 unsigned ldrOpc, strOpc;
5173 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
5175 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
5176 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
5179 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
5180 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
5183 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
5184 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
5188 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5189 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5190 MF->insert(It, loopMBB);
5191 MF->insert(It, exitMBB);
5193 // Transfer the remainder of BB and its successor edges to exitMBB.
5194 exitMBB->splice(exitMBB->begin(), BB,
5195 llvm::next(MachineBasicBlock::iterator(MI)),
5197 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5199 TargetRegisterClass *TRC =
5200 isThumb2 ? ARM::tGPRRegisterClass : ARM::GPRRegisterClass;
5201 unsigned scratch = MRI.createVirtualRegister(TRC);
5202 unsigned scratch2 = (!BinOpcode) ? incr : MRI.createVirtualRegister(TRC);
5206 // fallthrough --> loopMBB
5207 BB->addSuccessor(loopMBB);
5211 // <binop> scratch2, dest, incr
5212 // strex scratch, scratch2, ptr
5215 // fallthrough --> exitMBB
5217 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
5218 if (ldrOpc == ARM::t2LDREX)
5220 AddDefaultPred(MIB);
5222 // operand order needs to go the other way for NAND
5223 if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr)
5224 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
5225 addReg(incr).addReg(dest)).addReg(0);
5227 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
5228 addReg(dest).addReg(incr)).addReg(0);
5231 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2).addReg(ptr);
5232 if (strOpc == ARM::t2STREX)
5234 AddDefaultPred(MIB);
5235 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5236 .addReg(scratch).addImm(0));
5237 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5238 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5240 BB->addSuccessor(loopMBB);
5241 BB->addSuccessor(exitMBB);
5247 MI->eraseFromParent(); // The instruction is gone now.
5253 ARMTargetLowering::EmitAtomicBinaryMinMax(MachineInstr *MI,
5254 MachineBasicBlock *BB,
5257 ARMCC::CondCodes Cond) const {
5258 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5260 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5261 MachineFunction *MF = BB->getParent();
5262 MachineFunction::iterator It = BB;
5265 unsigned dest = MI->getOperand(0).getReg();
5266 unsigned ptr = MI->getOperand(1).getReg();
5267 unsigned incr = MI->getOperand(2).getReg();
5268 unsigned oldval = dest;
5269 DebugLoc dl = MI->getDebugLoc();
5270 bool isThumb2 = Subtarget->isThumb2();
5272 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5274 MRI.constrainRegClass(dest, ARM::rGPRRegisterClass);
5275 MRI.constrainRegClass(ptr, ARM::rGPRRegisterClass);
5278 unsigned ldrOpc, strOpc, extendOpc;
5280 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
5282 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
5283 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
5284 extendOpc = isThumb2 ? ARM::t2SXTB : ARM::SXTB;
5287 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
5288 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
5289 extendOpc = isThumb2 ? ARM::t2SXTH : ARM::SXTH;
5292 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
5293 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
5298 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5299 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5300 MF->insert(It, loopMBB);
5301 MF->insert(It, exitMBB);
5303 // Transfer the remainder of BB and its successor edges to exitMBB.
5304 exitMBB->splice(exitMBB->begin(), BB,
5305 llvm::next(MachineBasicBlock::iterator(MI)),
5307 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5309 TargetRegisterClass *TRC =
5310 isThumb2 ? ARM::tGPRRegisterClass : ARM::GPRRegisterClass;
5311 unsigned scratch = MRI.createVirtualRegister(TRC);
5312 unsigned scratch2 = MRI.createVirtualRegister(TRC);
5316 // fallthrough --> loopMBB
5317 BB->addSuccessor(loopMBB);
5321 // (sign extend dest, if required)
5323 // cmov.cond scratch2, dest, incr
5324 // strex scratch, scratch2, ptr
5327 // fallthrough --> exitMBB
5329 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
5330 if (ldrOpc == ARM::t2LDREX)
5332 AddDefaultPred(MIB);
5334 // Sign extend the value, if necessary.
5335 if (signExtend && extendOpc) {
5336 oldval = MRI.createVirtualRegister(ARM::GPRRegisterClass);
5337 AddDefaultPred(BuildMI(BB, dl, TII->get(extendOpc), oldval)
5342 // Build compare and cmov instructions.
5343 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
5344 .addReg(oldval).addReg(incr));
5345 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2MOVCCr : ARM::MOVCCr), scratch2)
5346 .addReg(oldval).addReg(incr).addImm(Cond).addReg(ARM::CPSR);
5348 MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2).addReg(ptr);
5349 if (strOpc == ARM::t2STREX)
5351 AddDefaultPred(MIB);
5352 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5353 .addReg(scratch).addImm(0));
5354 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5355 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5357 BB->addSuccessor(loopMBB);
5358 BB->addSuccessor(exitMBB);
5364 MI->eraseFromParent(); // The instruction is gone now.
5370 ARMTargetLowering::EmitAtomicBinary64(MachineInstr *MI, MachineBasicBlock *BB,
5371 unsigned Op1, unsigned Op2,
5372 bool NeedsCarry, bool IsCmpxchg) const {
5373 // This also handles ATOMIC_SWAP, indicated by Op1==0.
5374 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5376 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5377 MachineFunction *MF = BB->getParent();
5378 MachineFunction::iterator It = BB;
5381 unsigned destlo = MI->getOperand(0).getReg();
5382 unsigned desthi = MI->getOperand(1).getReg();
5383 unsigned ptr = MI->getOperand(2).getReg();
5384 unsigned vallo = MI->getOperand(3).getReg();
5385 unsigned valhi = MI->getOperand(4).getReg();
5386 DebugLoc dl = MI->getDebugLoc();
5387 bool isThumb2 = Subtarget->isThumb2();
5389 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5391 MRI.constrainRegClass(destlo, ARM::rGPRRegisterClass);
5392 MRI.constrainRegClass(desthi, ARM::rGPRRegisterClass);
5393 MRI.constrainRegClass(ptr, ARM::rGPRRegisterClass);
5396 unsigned ldrOpc = isThumb2 ? ARM::t2LDREXD : ARM::LDREXD;
5397 unsigned strOpc = isThumb2 ? ARM::t2STREXD : ARM::STREXD;
5399 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5400 MachineBasicBlock *contBB = 0, *cont2BB = 0;
5402 contBB = MF->CreateMachineBasicBlock(LLVM_BB);
5403 cont2BB = MF->CreateMachineBasicBlock(LLVM_BB);
5405 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5406 MF->insert(It, loopMBB);
5408 MF->insert(It, contBB);
5409 MF->insert(It, cont2BB);
5411 MF->insert(It, exitMBB);
5413 // Transfer the remainder of BB and its successor edges to exitMBB.
5414 exitMBB->splice(exitMBB->begin(), BB,
5415 llvm::next(MachineBasicBlock::iterator(MI)),
5417 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5419 TargetRegisterClass *TRC =
5420 isThumb2 ? ARM::tGPRRegisterClass : ARM::GPRRegisterClass;
5421 unsigned storesuccess = MRI.createVirtualRegister(TRC);
5425 // fallthrough --> loopMBB
5426 BB->addSuccessor(loopMBB);
5429 // ldrexd r2, r3, ptr
5430 // <binopa> r0, r2, incr
5431 // <binopb> r1, r3, incr
5432 // strexd storesuccess, r0, r1, ptr
5433 // cmp storesuccess, #0
5435 // fallthrough --> exitMBB
5437 // Note that the registers are explicitly specified because there is not any
5438 // way to force the register allocator to allocate a register pair.
5440 // FIXME: The hardcoded registers are not necessary for Thumb2, but we
5441 // need to properly enforce the restriction that the two output registers
5442 // for ldrexd must be different.
5445 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc))
5446 .addReg(ARM::R2, RegState::Define)
5447 .addReg(ARM::R3, RegState::Define).addReg(ptr));
5448 // Copy r2/r3 into dest. (This copy will normally be coalesced.)
5449 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), destlo).addReg(ARM::R2);
5450 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), desthi).addReg(ARM::R3);
5454 for (unsigned i = 0; i < 2; i++) {
5455 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr :
5457 .addReg(i == 0 ? destlo : desthi)
5458 .addReg(i == 0 ? vallo : valhi));
5459 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5460 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5461 BB->addSuccessor(exitMBB);
5462 BB->addSuccessor(i == 0 ? contBB : cont2BB);
5463 BB = (i == 0 ? contBB : cont2BB);
5466 // Copy to physregs for strexd
5467 unsigned setlo = MI->getOperand(5).getReg();
5468 unsigned sethi = MI->getOperand(6).getReg();
5469 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R0).addReg(setlo);
5470 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R1).addReg(sethi);
5472 // Perform binary operation
5473 AddDefaultPred(BuildMI(BB, dl, TII->get(Op1), ARM::R0)
5474 .addReg(destlo).addReg(vallo))
5475 .addReg(NeedsCarry ? ARM::CPSR : 0, getDefRegState(NeedsCarry));
5476 AddDefaultPred(BuildMI(BB, dl, TII->get(Op2), ARM::R1)
5477 .addReg(desthi).addReg(valhi)).addReg(0);
5479 // Copy to physregs for strexd
5480 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R0).addReg(vallo);
5481 BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R1).addReg(valhi);
5485 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), storesuccess)
5486 .addReg(ARM::R0).addReg(ARM::R1).addReg(ptr));
5488 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5489 .addReg(storesuccess).addImm(0));
5490 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5491 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5493 BB->addSuccessor(loopMBB);
5494 BB->addSuccessor(exitMBB);
5500 MI->eraseFromParent(); // The instruction is gone now.
5505 /// EmitBasePointerRecalculation - For functions using a base pointer, we
5506 /// rematerialize it (via the frame pointer).
5507 void ARMTargetLowering::
5508 EmitBasePointerRecalculation(MachineInstr *MI, MachineBasicBlock *MBB,
5509 MachineBasicBlock *DispatchBB) const {
5510 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5511 const ARMBaseInstrInfo *AII = static_cast<const ARMBaseInstrInfo*>(TII);
5512 MachineFunction &MF = *MI->getParent()->getParent();
5513 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
5514 const ARMBaseRegisterInfo &RI = AII->getRegisterInfo();
5516 if (!RI.hasBasePointer(MF)) return;
5518 MachineBasicBlock::iterator MBBI = MI;
5520 int32_t NumBytes = AFI->getFramePtrSpillOffset();
5521 unsigned FramePtr = RI.getFrameRegister(MF);
5522 assert(MF.getTarget().getFrameLowering()->hasFP(MF) &&
5523 "Base pointer without frame pointer?");
5525 if (AFI->isThumb2Function())
5526 llvm::emitT2RegPlusImmediate(*MBB, MBBI, MI->getDebugLoc(), ARM::R6,
5527 FramePtr, -NumBytes, ARMCC::AL, 0, *AII);
5528 else if (AFI->isThumbFunction())
5529 llvm::emitThumbRegPlusImmediate(*MBB, MBBI, MI->getDebugLoc(), ARM::R6,
5530 FramePtr, -NumBytes, *AII, RI);
5532 llvm::emitARMRegPlusImmediate(*MBB, MBBI, MI->getDebugLoc(), ARM::R6,
5533 FramePtr, -NumBytes, ARMCC::AL, 0, *AII);
5535 if (!RI.needsStackRealignment(MF)) return;
5537 // If there's dynamic realignment, adjust for it.
5538 MachineFrameInfo *MFI = MF.getFrameInfo();
5539 unsigned MaxAlign = MFI->getMaxAlignment();
5540 assert(!AFI->isThumb1OnlyFunction());
5542 // Emit bic r6, r6, MaxAlign
5543 unsigned bicOpc = AFI->isThumbFunction() ? ARM::t2BICri : ARM::BICri;
5546 BuildMI(*MBB, MBBI, MI->getDebugLoc(), TII->get(bicOpc), ARM::R6)
5547 .addReg(ARM::R6, RegState::Kill)
5548 .addImm(MaxAlign - 1)));
5551 /// SetupEntryBlockForSjLj - Insert code into the entry block that creates and
5552 /// registers the function context.
5553 void ARMTargetLowering::
5554 SetupEntryBlockForSjLj(MachineInstr *MI, MachineBasicBlock *MBB,
5555 MachineBasicBlock *DispatchBB, int FI) const {
5556 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5557 DebugLoc dl = MI->getDebugLoc();
5558 MachineFunction *MF = MBB->getParent();
5559 MachineRegisterInfo *MRI = &MF->getRegInfo();
5560 MachineConstantPool *MCP = MF->getConstantPool();
5561 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
5562 const Function *F = MF->getFunction();
5564 bool isThumb = Subtarget->isThumb();
5565 bool isThumb2 = Subtarget->isThumb2();
5567 unsigned PCLabelId = AFI->createPICLabelUId();
5568 unsigned PCAdj = (isThumb || isThumb2) ? 4 : 8;
5569 ARMConstantPoolValue *CPV =
5570 ARMConstantPoolMBB::Create(F->getContext(), DispatchBB, PCLabelId, PCAdj);
5571 unsigned CPI = MCP->getConstantPoolIndex(CPV, 4);
5573 const TargetRegisterClass *TRC =
5574 isThumb ? ARM::tGPRRegisterClass : ARM::GPRRegisterClass;
5576 // Grab constant pool and fixed stack memory operands.
5577 MachineMemOperand *CPMMO =
5578 MF->getMachineMemOperand(MachinePointerInfo::getConstantPool(),
5579 MachineMemOperand::MOLoad, 4, 4);
5581 MachineMemOperand *FIMMOSt =
5582 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
5583 MachineMemOperand::MOStore, 4, 4);
5585 EmitBasePointerRecalculation(MI, MBB, DispatchBB);
5587 // Load the address of the dispatch MBB into the jump buffer.
5589 // Incoming value: jbuf
5590 // ldr.n r5, LCPI1_1
5593 // str r5, [$jbuf, #+4] ; &jbuf[1]
5594 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5595 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2LDRpci), NewVReg1)
5596 .addConstantPoolIndex(CPI)
5597 .addMemOperand(CPMMO));
5598 // Set the low bit because of thumb mode.
5599 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
5601 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2ORRri), NewVReg2)
5602 .addReg(NewVReg1, RegState::Kill)
5604 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
5605 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg3)
5606 .addReg(NewVReg2, RegState::Kill)
5608 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2STRi12))
5609 .addReg(NewVReg3, RegState::Kill)
5611 .addImm(36) // &jbuf[1] :: pc
5612 .addMemOperand(FIMMOSt));
5613 } else if (isThumb) {
5614 // Incoming value: jbuf
5615 // ldr.n r1, LCPI1_4
5619 // add r2, $jbuf, #+4 ; &jbuf[1]
5621 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5622 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tLDRpci), NewVReg1)
5623 .addConstantPoolIndex(CPI)
5624 .addMemOperand(CPMMO));
5625 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
5626 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg2)
5627 .addReg(NewVReg1, RegState::Kill)
5629 // Set the low bit because of thumb mode.
5630 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
5631 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tMOVi8), NewVReg3)
5632 .addReg(ARM::CPSR, RegState::Define)
5634 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
5635 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tORR), NewVReg4)
5636 .addReg(ARM::CPSR, RegState::Define)
5637 .addReg(NewVReg2, RegState::Kill)
5638 .addReg(NewVReg3, RegState::Kill));
5639 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
5640 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tADDrSPi), NewVReg5)
5642 .addImm(36)); // &jbuf[1] :: pc
5643 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tSTRi))
5644 .addReg(NewVReg4, RegState::Kill)
5645 .addReg(NewVReg5, RegState::Kill)
5647 .addMemOperand(FIMMOSt));
5649 // Incoming value: jbuf
5652 // str r1, [$jbuf, #+4] ; &jbuf[1]
5653 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5654 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::LDRi12), NewVReg1)
5655 .addConstantPoolIndex(CPI)
5657 .addMemOperand(CPMMO));
5658 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
5659 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::PICADD), NewVReg2)
5660 .addReg(NewVReg1, RegState::Kill)
5661 .addImm(PCLabelId));
5662 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::STRi12))
5663 .addReg(NewVReg2, RegState::Kill)
5665 .addImm(36) // &jbuf[1] :: pc
5666 .addMemOperand(FIMMOSt));
5670 MachineBasicBlock *ARMTargetLowering::
5671 EmitSjLjDispatchBlock(MachineInstr *MI, MachineBasicBlock *MBB) const {
5672 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5673 DebugLoc dl = MI->getDebugLoc();
5674 MachineFunction *MF = MBB->getParent();
5675 MachineRegisterInfo *MRI = &MF->getRegInfo();
5676 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
5677 MachineFrameInfo *MFI = MF->getFrameInfo();
5678 int FI = MFI->getFunctionContextIndex();
5680 const TargetRegisterClass *TRC =
5681 Subtarget->isThumb() ? ARM::tGPRRegisterClass : ARM::GPRRegisterClass;
5683 // Get a mapping of the call site numbers to all of the landing pads they're
5685 DenseMap<unsigned, SmallVector<MachineBasicBlock*, 2> > CallSiteNumToLPad;
5686 unsigned MaxCSNum = 0;
5687 MachineModuleInfo &MMI = MF->getMMI();
5688 for (MachineFunction::iterator BB = MF->begin(), E = MF->end(); BB != E; ++BB) {
5689 if (!BB->isLandingPad()) continue;
5691 // FIXME: We should assert that the EH_LABEL is the first MI in the landing
5693 for (MachineBasicBlock::iterator
5694 II = BB->begin(), IE = BB->end(); II != IE; ++II) {
5695 if (!II->isEHLabel()) continue;
5697 MCSymbol *Sym = II->getOperand(0).getMCSymbol();
5698 if (!MMI.hasCallSiteLandingPad(Sym)) continue;
5700 SmallVectorImpl<unsigned> &CallSiteIdxs = MMI.getCallSiteLandingPad(Sym);
5701 for (SmallVectorImpl<unsigned>::iterator
5702 CSI = CallSiteIdxs.begin(), CSE = CallSiteIdxs.end();
5703 CSI != CSE; ++CSI) {
5704 CallSiteNumToLPad[*CSI].push_back(BB);
5705 MaxCSNum = std::max(MaxCSNum, *CSI);
5711 // Get an ordered list of the machine basic blocks for the jump table.
5712 std::vector<MachineBasicBlock*> LPadList;
5713 SmallPtrSet<MachineBasicBlock*, 64> InvokeBBs;
5714 LPadList.reserve(CallSiteNumToLPad.size());
5715 for (unsigned I = 1; I <= MaxCSNum; ++I) {
5716 SmallVectorImpl<MachineBasicBlock*> &MBBList = CallSiteNumToLPad[I];
5717 for (SmallVectorImpl<MachineBasicBlock*>::iterator
5718 II = MBBList.begin(), IE = MBBList.end(); II != IE; ++II) {
5719 LPadList.push_back(*II);
5720 InvokeBBs.insert((*II)->pred_begin(), (*II)->pred_end());
5724 assert(!LPadList.empty() &&
5725 "No landing pad destinations for the dispatch jump table!");
5727 // Create the jump table and associated information.
5728 MachineJumpTableInfo *JTI =
5729 MF->getOrCreateJumpTableInfo(MachineJumpTableInfo::EK_Inline);
5730 unsigned MJTI = JTI->createJumpTableIndex(LPadList);
5731 unsigned UId = AFI->createJumpTableUId();
5733 // Create the MBBs for the dispatch code.
5735 // Shove the dispatch's address into the return slot in the function context.
5736 MachineBasicBlock *DispatchBB = MF->CreateMachineBasicBlock();
5737 DispatchBB->setIsLandingPad();
5739 MachineBasicBlock *TrapBB = MF->CreateMachineBasicBlock();
5740 BuildMI(TrapBB, dl, TII->get(Subtarget->isThumb() ? ARM::tTRAP : ARM::TRAP));
5741 DispatchBB->addSuccessor(TrapBB);
5743 MachineBasicBlock *DispContBB = MF->CreateMachineBasicBlock();
5744 DispatchBB->addSuccessor(DispContBB);
5747 MF->insert(MF->end(), DispatchBB);
5748 MF->insert(MF->end(), DispContBB);
5749 MF->insert(MF->end(), TrapBB);
5751 // Insert code into the entry block that creates and registers the function
5753 SetupEntryBlockForSjLj(MI, MBB, DispatchBB, FI);
5755 MachineMemOperand *FIMMOLd =
5756 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
5757 MachineMemOperand::MOLoad |
5758 MachineMemOperand::MOVolatile, 4, 4);
5760 unsigned NumLPads = LPadList.size();
5761 if (Subtarget->isThumb2()) {
5762 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5763 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2LDRi12), NewVReg1)
5766 .addMemOperand(FIMMOLd));
5768 if (NumLPads < 256) {
5769 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPri))
5771 .addImm(LPadList.size()));
5773 unsigned VReg1 = MRI->createVirtualRegister(TRC);
5774 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVi16), VReg1)
5775 .addImm(NumLPads & 0xFFFF));
5777 unsigned VReg2 = VReg1;
5778 if ((NumLPads & 0xFFFF0000) != 0) {
5779 VReg2 = MRI->createVirtualRegister(TRC);
5780 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVTi16), VReg2)
5782 .addImm(NumLPads >> 16));
5785 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPrr))
5790 BuildMI(DispatchBB, dl, TII->get(ARM::t2Bcc))
5795 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
5796 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::t2LEApcrelJT),NewVReg3)
5797 .addJumpTableIndex(MJTI)
5800 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
5803 BuildMI(DispContBB, dl, TII->get(ARM::t2ADDrs), NewVReg4)
5804 .addReg(NewVReg3, RegState::Kill)
5806 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
5808 BuildMI(DispContBB, dl, TII->get(ARM::t2BR_JT))
5809 .addReg(NewVReg4, RegState::Kill)
5811 .addJumpTableIndex(MJTI)
5813 } else if (Subtarget->isThumb()) {
5814 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5815 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRspi), NewVReg1)
5818 .addMemOperand(FIMMOLd));
5820 if (NumLPads < 256) {
5821 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPi8))
5825 MachineConstantPool *ConstantPool = MF->getConstantPool();
5826 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
5827 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
5829 // MachineConstantPool wants an explicit alignment.
5830 unsigned Align = getTargetData()->getPrefTypeAlignment(Int32Ty);
5832 Align = getTargetData()->getTypeAllocSize(C->getType());
5833 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
5835 unsigned VReg1 = MRI->createVirtualRegister(TRC);
5836 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRpci))
5837 .addReg(VReg1, RegState::Define)
5838 .addConstantPoolIndex(Idx));
5839 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPr))
5844 BuildMI(DispatchBB, dl, TII->get(ARM::tBcc))
5849 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
5850 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLSLri), NewVReg2)
5851 .addReg(ARM::CPSR, RegState::Define)
5855 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
5856 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLEApcrelJT), NewVReg3)
5857 .addJumpTableIndex(MJTI)
5860 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
5861 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg4)
5862 .addReg(ARM::CPSR, RegState::Define)
5863 .addReg(NewVReg2, RegState::Kill)
5866 MachineMemOperand *JTMMOLd =
5867 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
5868 MachineMemOperand::MOLoad, 4, 4);
5870 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
5871 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLDRi), NewVReg5)
5872 .addReg(NewVReg4, RegState::Kill)
5874 .addMemOperand(JTMMOLd));
5876 unsigned NewVReg6 = MRI->createVirtualRegister(TRC);
5877 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg6)
5878 .addReg(ARM::CPSR, RegState::Define)
5879 .addReg(NewVReg5, RegState::Kill)
5882 BuildMI(DispContBB, dl, TII->get(ARM::tBR_JTr))
5883 .addReg(NewVReg6, RegState::Kill)
5884 .addJumpTableIndex(MJTI)
5887 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5888 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRi12), NewVReg1)
5891 .addMemOperand(FIMMOLd));
5893 if (NumLPads < 256) {
5894 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPri))
5897 } else if (Subtarget->hasV6T2Ops() && isUInt<16>(NumLPads)) {
5898 unsigned VReg1 = MRI->createVirtualRegister(TRC);
5899 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVi16), VReg1)
5900 .addImm(NumLPads & 0xFFFF));
5902 unsigned VReg2 = VReg1;
5903 if ((NumLPads & 0xFFFF0000) != 0) {
5904 VReg2 = MRI->createVirtualRegister(TRC);
5905 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVTi16), VReg2)
5907 .addImm(NumLPads >> 16));
5910 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
5914 MachineConstantPool *ConstantPool = MF->getConstantPool();
5915 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
5916 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
5918 // MachineConstantPool wants an explicit alignment.
5919 unsigned Align = getTargetData()->getPrefTypeAlignment(Int32Ty);
5921 Align = getTargetData()->getTypeAllocSize(C->getType());
5922 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
5924 unsigned VReg1 = MRI->createVirtualRegister(TRC);
5925 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRcp))
5926 .addReg(VReg1, RegState::Define)
5927 .addConstantPoolIndex(Idx));
5928 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
5930 .addReg(VReg1, RegState::Kill));
5933 BuildMI(DispatchBB, dl, TII->get(ARM::Bcc))
5938 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
5940 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::MOVsi), NewVReg3)
5942 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
5943 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
5944 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::LEApcrelJT), NewVReg4)
5945 .addJumpTableIndex(MJTI)
5948 MachineMemOperand *JTMMOLd =
5949 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
5950 MachineMemOperand::MOLoad, 4, 4);
5951 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
5953 BuildMI(DispContBB, dl, TII->get(ARM::LDRrs), NewVReg5)
5954 .addReg(NewVReg3, RegState::Kill)
5957 .addMemOperand(JTMMOLd));
5959 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTadd))
5960 .addReg(NewVReg5, RegState::Kill)
5962 .addJumpTableIndex(MJTI)
5966 // Add the jump table entries as successors to the MBB.
5967 MachineBasicBlock *PrevMBB = 0;
5968 for (std::vector<MachineBasicBlock*>::iterator
5969 I = LPadList.begin(), E = LPadList.end(); I != E; ++I) {
5970 MachineBasicBlock *CurMBB = *I;
5971 if (PrevMBB != CurMBB)
5972 DispContBB->addSuccessor(CurMBB);
5976 // N.B. the order the invoke BBs are processed in doesn't matter here.
5977 const ARMBaseInstrInfo *AII = static_cast<const ARMBaseInstrInfo*>(TII);
5978 const ARMBaseRegisterInfo &RI = AII->getRegisterInfo();
5979 const unsigned *SavedRegs = RI.getCalleeSavedRegs(MF);
5980 SmallVector<MachineBasicBlock*, 64> MBBLPads;
5981 for (SmallPtrSet<MachineBasicBlock*, 64>::iterator
5982 I = InvokeBBs.begin(), E = InvokeBBs.end(); I != E; ++I) {
5983 MachineBasicBlock *BB = *I;
5985 // Remove the landing pad successor from the invoke block and replace it
5986 // with the new dispatch block.
5987 for (MachineBasicBlock::succ_iterator
5988 SI = BB->succ_begin(), SE = BB->succ_end(); SI != SE; ++SI) {
5989 MachineBasicBlock *SMBB = *SI;
5990 if (SMBB->isLandingPad()) {
5991 BB->removeSuccessor(SMBB);
5992 MBBLPads.push_back(SMBB);
5996 BB->addSuccessor(DispatchBB);
5998 // Find the invoke call and mark all of the callee-saved registers as
5999 // 'implicit defined' so that they're spilled. This prevents code from
6000 // moving instructions to before the EH block, where they will never be
6002 for (MachineBasicBlock::reverse_iterator
6003 II = BB->rbegin(), IE = BB->rend(); II != IE; ++II) {
6004 if (!II->getDesc().isCall()) continue;
6006 DenseMap<unsigned, bool> DefRegs;
6007 for (MachineInstr::mop_iterator
6008 OI = II->operands_begin(), OE = II->operands_end();
6010 if (!OI->isReg()) continue;
6011 DefRegs[OI->getReg()] = true;
6014 MachineInstrBuilder MIB(&*II);
6016 for (unsigned i = 0; SavedRegs[i] != 0; ++i) {
6017 if (!TRC->contains(SavedRegs[i])) continue;
6018 if (!DefRegs[SavedRegs[i]])
6019 MIB.addReg(SavedRegs[i], RegState::ImplicitDefine | RegState::Dead);
6026 // Mark all former landing pads as non-landing pads. The dispatch is the only
6028 for (SmallVectorImpl<MachineBasicBlock*>::iterator
6029 I = MBBLPads.begin(), E = MBBLPads.end(); I != E; ++I)
6030 (*I)->setIsLandingPad(false);
6032 // The instruction is gone now.
6033 MI->eraseFromParent();
6039 MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) {
6040 for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
6041 E = MBB->succ_end(); I != E; ++I)
6044 llvm_unreachable("Expecting a BB with two successors!");
6048 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
6049 MachineBasicBlock *BB) const {
6050 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6051 DebugLoc dl = MI->getDebugLoc();
6052 bool isThumb2 = Subtarget->isThumb2();
6053 switch (MI->getOpcode()) {
6056 llvm_unreachable("Unexpected instr type to insert");
6058 // The Thumb2 pre-indexed stores have the same MI operands, they just
6059 // define them differently in the .td files from the isel patterns, so
6060 // they need pseudos.
6061 case ARM::t2STR_preidx:
6062 MI->setDesc(TII->get(ARM::t2STR_PRE));
6064 case ARM::t2STRB_preidx:
6065 MI->setDesc(TII->get(ARM::t2STRB_PRE));
6067 case ARM::t2STRH_preidx:
6068 MI->setDesc(TII->get(ARM::t2STRH_PRE));
6071 case ARM::STRi_preidx:
6072 case ARM::STRBi_preidx: {
6073 unsigned NewOpc = MI->getOpcode() == ARM::STRi_preidx ?
6074 ARM::STR_PRE_IMM : ARM::STRB_PRE_IMM;
6075 // Decode the offset.
6076 unsigned Offset = MI->getOperand(4).getImm();
6077 bool isSub = ARM_AM::getAM2Op(Offset) == ARM_AM::sub;
6078 Offset = ARM_AM::getAM2Offset(Offset);
6082 MachineMemOperand *MMO = *MI->memoperands_begin();
6083 BuildMI(*BB, MI, dl, TII->get(NewOpc))
6084 .addOperand(MI->getOperand(0)) // Rn_wb
6085 .addOperand(MI->getOperand(1)) // Rt
6086 .addOperand(MI->getOperand(2)) // Rn
6087 .addImm(Offset) // offset (skip GPR==zero_reg)
6088 .addOperand(MI->getOperand(5)) // pred
6089 .addOperand(MI->getOperand(6))
6090 .addMemOperand(MMO);
6091 MI->eraseFromParent();
6094 case ARM::STRr_preidx:
6095 case ARM::STRBr_preidx:
6096 case ARM::STRH_preidx: {
6098 switch (MI->getOpcode()) {
6099 default: llvm_unreachable("unexpected opcode!");
6100 case ARM::STRr_preidx: NewOpc = ARM::STR_PRE_REG; break;
6101 case ARM::STRBr_preidx: NewOpc = ARM::STRB_PRE_REG; break;
6102 case ARM::STRH_preidx: NewOpc = ARM::STRH_PRE; break;
6104 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(NewOpc));
6105 for (unsigned i = 0; i < MI->getNumOperands(); ++i)
6106 MIB.addOperand(MI->getOperand(i));
6107 MI->eraseFromParent();
6110 case ARM::ATOMIC_LOAD_ADD_I8:
6111 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
6112 case ARM::ATOMIC_LOAD_ADD_I16:
6113 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
6114 case ARM::ATOMIC_LOAD_ADD_I32:
6115 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
6117 case ARM::ATOMIC_LOAD_AND_I8:
6118 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6119 case ARM::ATOMIC_LOAD_AND_I16:
6120 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6121 case ARM::ATOMIC_LOAD_AND_I32:
6122 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6124 case ARM::ATOMIC_LOAD_OR_I8:
6125 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6126 case ARM::ATOMIC_LOAD_OR_I16:
6127 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6128 case ARM::ATOMIC_LOAD_OR_I32:
6129 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6131 case ARM::ATOMIC_LOAD_XOR_I8:
6132 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6133 case ARM::ATOMIC_LOAD_XOR_I16:
6134 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6135 case ARM::ATOMIC_LOAD_XOR_I32:
6136 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6138 case ARM::ATOMIC_LOAD_NAND_I8:
6139 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
6140 case ARM::ATOMIC_LOAD_NAND_I16:
6141 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
6142 case ARM::ATOMIC_LOAD_NAND_I32:
6143 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
6145 case ARM::ATOMIC_LOAD_SUB_I8:
6146 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
6147 case ARM::ATOMIC_LOAD_SUB_I16:
6148 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
6149 case ARM::ATOMIC_LOAD_SUB_I32:
6150 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
6152 case ARM::ATOMIC_LOAD_MIN_I8:
6153 return EmitAtomicBinaryMinMax(MI, BB, 1, true, ARMCC::LT);
6154 case ARM::ATOMIC_LOAD_MIN_I16:
6155 return EmitAtomicBinaryMinMax(MI, BB, 2, true, ARMCC::LT);
6156 case ARM::ATOMIC_LOAD_MIN_I32:
6157 return EmitAtomicBinaryMinMax(MI, BB, 4, true, ARMCC::LT);
6159 case ARM::ATOMIC_LOAD_MAX_I8:
6160 return EmitAtomicBinaryMinMax(MI, BB, 1, true, ARMCC::GT);
6161 case ARM::ATOMIC_LOAD_MAX_I16:
6162 return EmitAtomicBinaryMinMax(MI, BB, 2, true, ARMCC::GT);
6163 case ARM::ATOMIC_LOAD_MAX_I32:
6164 return EmitAtomicBinaryMinMax(MI, BB, 4, true, ARMCC::GT);
6166 case ARM::ATOMIC_LOAD_UMIN_I8:
6167 return EmitAtomicBinaryMinMax(MI, BB, 1, false, ARMCC::LO);
6168 case ARM::ATOMIC_LOAD_UMIN_I16:
6169 return EmitAtomicBinaryMinMax(MI, BB, 2, false, ARMCC::LO);
6170 case ARM::ATOMIC_LOAD_UMIN_I32:
6171 return EmitAtomicBinaryMinMax(MI, BB, 4, false, ARMCC::LO);
6173 case ARM::ATOMIC_LOAD_UMAX_I8:
6174 return EmitAtomicBinaryMinMax(MI, BB, 1, false, ARMCC::HI);
6175 case ARM::ATOMIC_LOAD_UMAX_I16:
6176 return EmitAtomicBinaryMinMax(MI, BB, 2, false, ARMCC::HI);
6177 case ARM::ATOMIC_LOAD_UMAX_I32:
6178 return EmitAtomicBinaryMinMax(MI, BB, 4, false, ARMCC::HI);
6180 case ARM::ATOMIC_SWAP_I8: return EmitAtomicBinary(MI, BB, 1, 0);
6181 case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0);
6182 case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0);
6184 case ARM::ATOMIC_CMP_SWAP_I8: return EmitAtomicCmpSwap(MI, BB, 1);
6185 case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2);
6186 case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4);
6189 case ARM::ATOMADD6432:
6190 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr,
6191 isThumb2 ? ARM::t2ADCrr : ARM::ADCrr,
6192 /*NeedsCarry*/ true);
6193 case ARM::ATOMSUB6432:
6194 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
6195 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
6196 /*NeedsCarry*/ true);
6197 case ARM::ATOMOR6432:
6198 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr,
6199 isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6200 case ARM::ATOMXOR6432:
6201 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2EORrr : ARM::EORrr,
6202 isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6203 case ARM::ATOMAND6432:
6204 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr,
6205 isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6206 case ARM::ATOMSWAP6432:
6207 return EmitAtomicBinary64(MI, BB, 0, 0, false);
6208 case ARM::ATOMCMPXCHG6432:
6209 return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
6210 isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
6211 /*NeedsCarry*/ false, /*IsCmpxchg*/true);
6213 case ARM::tMOVCCr_pseudo: {
6214 // To "insert" a SELECT_CC instruction, we actually have to insert the
6215 // diamond control-flow pattern. The incoming instruction knows the
6216 // destination vreg to set, the condition code register to branch on, the
6217 // true/false values to select between, and a branch opcode to use.
6218 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6219 MachineFunction::iterator It = BB;
6225 // cmpTY ccX, r1, r2
6227 // fallthrough --> copy0MBB
6228 MachineBasicBlock *thisMBB = BB;
6229 MachineFunction *F = BB->getParent();
6230 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
6231 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
6232 F->insert(It, copy0MBB);
6233 F->insert(It, sinkMBB);
6235 // Transfer the remainder of BB and its successor edges to sinkMBB.
6236 sinkMBB->splice(sinkMBB->begin(), BB,
6237 llvm::next(MachineBasicBlock::iterator(MI)),
6239 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
6241 BB->addSuccessor(copy0MBB);
6242 BB->addSuccessor(sinkMBB);
6244 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
6245 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
6248 // %FalseValue = ...
6249 // # fallthrough to sinkMBB
6252 // Update machine-CFG edges
6253 BB->addSuccessor(sinkMBB);
6256 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
6259 BuildMI(*BB, BB->begin(), dl,
6260 TII->get(ARM::PHI), MI->getOperand(0).getReg())
6261 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
6262 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
6264 MI->eraseFromParent(); // The pseudo instruction is gone now.
6269 case ARM::BCCZi64: {
6270 // If there is an unconditional branch to the other successor, remove it.
6271 BB->erase(llvm::next(MachineBasicBlock::iterator(MI)), BB->end());
6273 // Compare both parts that make up the double comparison separately for
6275 bool RHSisZero = MI->getOpcode() == ARM::BCCZi64;
6277 unsigned LHS1 = MI->getOperand(1).getReg();
6278 unsigned LHS2 = MI->getOperand(2).getReg();
6280 AddDefaultPred(BuildMI(BB, dl,
6281 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
6282 .addReg(LHS1).addImm(0));
6283 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
6284 .addReg(LHS2).addImm(0)
6285 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
6287 unsigned RHS1 = MI->getOperand(3).getReg();
6288 unsigned RHS2 = MI->getOperand(4).getReg();
6289 AddDefaultPred(BuildMI(BB, dl,
6290 TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
6291 .addReg(LHS1).addReg(RHS1));
6292 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
6293 .addReg(LHS2).addReg(RHS2)
6294 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
6297 MachineBasicBlock *destMBB = MI->getOperand(RHSisZero ? 3 : 5).getMBB();
6298 MachineBasicBlock *exitMBB = OtherSucc(BB, destMBB);
6299 if (MI->getOperand(0).getImm() == ARMCC::NE)
6300 std::swap(destMBB, exitMBB);
6302 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6303 .addMBB(destMBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
6305 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2B)).addMBB(exitMBB));
6307 BuildMI(BB, dl, TII->get(ARM::B)) .addMBB(exitMBB);
6309 MI->eraseFromParent(); // The pseudo instruction is gone now.
6313 case ARM::Int_eh_sjlj_setjmp:
6314 case ARM::Int_eh_sjlj_setjmp_nofp:
6315 case ARM::tInt_eh_sjlj_setjmp:
6316 case ARM::t2Int_eh_sjlj_setjmp:
6317 case ARM::t2Int_eh_sjlj_setjmp_nofp:
6318 EmitSjLjDispatchBlock(MI, BB);
6323 // To insert an ABS instruction, we have to insert the
6324 // diamond control-flow pattern. The incoming instruction knows the
6325 // source vreg to test against 0, the destination vreg to set,
6326 // the condition code register to branch on, the
6327 // true/false values to select between, and a branch opcode to use.
6332 // BCC (branch to SinkBB if V0 >= 0)
6333 // RSBBB: V3 = RSBri V2, 0 (compute ABS if V2 < 0)
6334 // SinkBB: V1 = PHI(V2, V3)
6335 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6336 MachineFunction::iterator BBI = BB;
6338 MachineFunction *Fn = BB->getParent();
6339 MachineBasicBlock *RSBBB = Fn->CreateMachineBasicBlock(LLVM_BB);
6340 MachineBasicBlock *SinkBB = Fn->CreateMachineBasicBlock(LLVM_BB);
6341 Fn->insert(BBI, RSBBB);
6342 Fn->insert(BBI, SinkBB);
6344 unsigned int ABSSrcReg = MI->getOperand(1).getReg();
6345 unsigned int ABSDstReg = MI->getOperand(0).getReg();
6346 bool isThumb2 = Subtarget->isThumb2();
6347 MachineRegisterInfo &MRI = Fn->getRegInfo();
6348 // In Thumb mode S must not be specified if source register is the SP or
6349 // PC and if destination register is the SP, so restrict register class
6350 unsigned NewMovDstReg = MRI.createVirtualRegister(
6351 isThumb2 ? ARM::rGPRRegisterClass : ARM::GPRRegisterClass);
6352 unsigned NewRsbDstReg = MRI.createVirtualRegister(
6353 isThumb2 ? ARM::rGPRRegisterClass : ARM::GPRRegisterClass);
6355 // Transfer the remainder of BB and its successor edges to sinkMBB.
6356 SinkBB->splice(SinkBB->begin(), BB,
6357 llvm::next(MachineBasicBlock::iterator(MI)),
6359 SinkBB->transferSuccessorsAndUpdatePHIs(BB);
6361 BB->addSuccessor(RSBBB);
6362 BB->addSuccessor(SinkBB);
6364 // fall through to SinkMBB
6365 RSBBB->addSuccessor(SinkBB);
6367 // insert a movs at the end of BB
6368 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2MOVr : ARM::MOVr),
6370 .addReg(ABSSrcReg, RegState::Kill)
6371 .addImm((unsigned)ARMCC::AL).addReg(0)
6372 .addReg(ARM::CPSR, RegState::Define);
6374 // insert a bcc with opposite CC to ARMCC::MI at the end of BB
6376 TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)).addMBB(SinkBB)
6377 .addImm(ARMCC::getOppositeCondition(ARMCC::MI)).addReg(ARM::CPSR);
6379 // insert rsbri in RSBBB
6380 // Note: BCC and rsbri will be converted into predicated rsbmi
6381 // by if-conversion pass
6382 BuildMI(*RSBBB, RSBBB->begin(), dl,
6383 TII->get(isThumb2 ? ARM::t2RSBri : ARM::RSBri), NewRsbDstReg)
6384 .addReg(NewMovDstReg, RegState::Kill)
6385 .addImm(0).addImm((unsigned)ARMCC::AL).addReg(0).addReg(0);
6387 // insert PHI in SinkBB,
6388 // reuse ABSDstReg to not change uses of ABS instruction
6389 BuildMI(*SinkBB, SinkBB->begin(), dl,
6390 TII->get(ARM::PHI), ABSDstReg)
6391 .addReg(NewRsbDstReg).addMBB(RSBBB)
6392 .addReg(NewMovDstReg).addMBB(BB);
6394 // remove ABS instruction
6395 MI->eraseFromParent();
6397 // return last added BB
6403 void ARMTargetLowering::AdjustInstrPostInstrSelection(MachineInstr *MI,
6404 SDNode *Node) const {
6405 const MCInstrDesc *MCID = &MI->getDesc();
6406 if (!MCID->hasPostISelHook()) {
6407 assert(!convertAddSubFlagsOpcode(MI->getOpcode()) &&
6408 "Pseudo flag-setting opcodes must be marked with 'hasPostISelHook'");
6412 // Adjust potentially 's' setting instructions after isel, i.e. ADC, SBC, RSB,
6413 // RSC. Coming out of isel, they have an implicit CPSR def, but the optional
6414 // operand is still set to noreg. If needed, set the optional operand's
6415 // register to CPSR, and remove the redundant implicit def.
6417 // e.g. ADCS (..., CPSR<imp-def>) -> ADC (... opt:CPSR<def>).
6419 // Rename pseudo opcodes.
6420 unsigned NewOpc = convertAddSubFlagsOpcode(MI->getOpcode());
6422 const ARMBaseInstrInfo *TII =
6423 static_cast<const ARMBaseInstrInfo*>(getTargetMachine().getInstrInfo());
6424 MCID = &TII->get(NewOpc);
6426 assert(MCID->getNumOperands() == MI->getDesc().getNumOperands() + 1 &&
6427 "converted opcode should be the same except for cc_out");
6431 // Add the optional cc_out operand
6432 MI->addOperand(MachineOperand::CreateReg(0, /*isDef=*/true));
6434 unsigned ccOutIdx = MCID->getNumOperands() - 1;
6436 // Any ARM instruction that sets the 's' bit should specify an optional
6437 // "cc_out" operand in the last operand position.
6438 if (!MCID->hasOptionalDef() || !MCID->OpInfo[ccOutIdx].isOptionalDef()) {
6439 assert(!NewOpc && "Optional cc_out operand required");
6442 // Look for an implicit def of CPSR added by MachineInstr ctor. Remove it
6443 // since we already have an optional CPSR def.
6444 bool definesCPSR = false;
6445 bool deadCPSR = false;
6446 for (unsigned i = MCID->getNumOperands(), e = MI->getNumOperands();
6448 const MachineOperand &MO = MI->getOperand(i);
6449 if (MO.isReg() && MO.isDef() && MO.getReg() == ARM::CPSR) {
6453 MI->RemoveOperand(i);
6458 assert(!NewOpc && "Optional cc_out operand required");
6461 assert(deadCPSR == !Node->hasAnyUseOfValue(1) && "inconsistent dead flag");
6463 assert(!MI->getOperand(ccOutIdx).getReg() &&
6464 "expect uninitialized optional cc_out operand");
6468 // If this instruction was defined with an optional CPSR def and its dag node
6469 // had a live implicit CPSR def, then activate the optional CPSR def.
6470 MachineOperand &MO = MI->getOperand(ccOutIdx);
6471 MO.setReg(ARM::CPSR);
6475 //===----------------------------------------------------------------------===//
6476 // ARM Optimization Hooks
6477 //===----------------------------------------------------------------------===//
6480 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
6481 TargetLowering::DAGCombinerInfo &DCI) {
6482 SelectionDAG &DAG = DCI.DAG;
6483 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6484 EVT VT = N->getValueType(0);
6485 unsigned Opc = N->getOpcode();
6486 bool isSlctCC = Slct.getOpcode() == ISD::SELECT_CC;
6487 SDValue LHS = isSlctCC ? Slct.getOperand(2) : Slct.getOperand(1);
6488 SDValue RHS = isSlctCC ? Slct.getOperand(3) : Slct.getOperand(2);
6489 ISD::CondCode CC = ISD::SETCC_INVALID;
6492 CC = cast<CondCodeSDNode>(Slct.getOperand(4))->get();
6494 SDValue CCOp = Slct.getOperand(0);
6495 if (CCOp.getOpcode() == ISD::SETCC)
6496 CC = cast<CondCodeSDNode>(CCOp.getOperand(2))->get();
6499 bool DoXform = false;
6501 assert ((Opc == ISD::ADD || (Opc == ISD::SUB && Slct == N->getOperand(1))) &&
6504 if (LHS.getOpcode() == ISD::Constant &&
6505 cast<ConstantSDNode>(LHS)->isNullValue()) {
6507 } else if (CC != ISD::SETCC_INVALID &&
6508 RHS.getOpcode() == ISD::Constant &&
6509 cast<ConstantSDNode>(RHS)->isNullValue()) {
6510 std::swap(LHS, RHS);
6511 SDValue Op0 = Slct.getOperand(0);
6512 EVT OpVT = isSlctCC ? Op0.getValueType() :
6513 Op0.getOperand(0).getValueType();
6514 bool isInt = OpVT.isInteger();
6515 CC = ISD::getSetCCInverse(CC, isInt);
6517 if (!TLI.isCondCodeLegal(CC, OpVT))
6518 return SDValue(); // Inverse operator isn't legal.
6525 SDValue Result = DAG.getNode(Opc, RHS.getDebugLoc(), VT, OtherOp, RHS);
6527 return DAG.getSelectCC(N->getDebugLoc(), OtherOp, Result,
6528 Slct.getOperand(0), Slct.getOperand(1), CC);
6529 SDValue CCOp = Slct.getOperand(0);
6531 CCOp = DAG.getSetCC(Slct.getDebugLoc(), CCOp.getValueType(),
6532 CCOp.getOperand(0), CCOp.getOperand(1), CC);
6533 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
6534 CCOp, OtherOp, Result);
6539 // AddCombineToVPADDL- For pair-wise add on neon, use the vpaddl instruction
6540 // (only after legalization).
6541 static SDValue AddCombineToVPADDL(SDNode *N, SDValue N0, SDValue N1,
6542 TargetLowering::DAGCombinerInfo &DCI,
6543 const ARMSubtarget *Subtarget) {
6545 // Only perform optimization if after legalize, and if NEON is available. We
6546 // also expected both operands to be BUILD_VECTORs.
6547 if (DCI.isBeforeLegalize() || !Subtarget->hasNEON()
6548 || N0.getOpcode() != ISD::BUILD_VECTOR
6549 || N1.getOpcode() != ISD::BUILD_VECTOR)
6552 // Check output type since VPADDL operand elements can only be 8, 16, or 32.
6553 EVT VT = N->getValueType(0);
6554 if (!VT.isInteger() || VT.getVectorElementType() == MVT::i64)
6557 // Check that the vector operands are of the right form.
6558 // N0 and N1 are BUILD_VECTOR nodes with N number of EXTRACT_VECTOR
6559 // operands, where N is the size of the formed vector.
6560 // Each EXTRACT_VECTOR should have the same input vector and odd or even
6561 // index such that we have a pair wise add pattern.
6563 // Grab the vector that all EXTRACT_VECTOR nodes should be referencing.
6564 if (N0->getOperand(0)->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
6566 SDValue Vec = N0->getOperand(0)->getOperand(0);
6567 SDNode *V = Vec.getNode();
6568 unsigned nextIndex = 0;
6570 // For each operands to the ADD which are BUILD_VECTORs,
6571 // check to see if each of their operands are an EXTRACT_VECTOR with
6572 // the same vector and appropriate index.
6573 for (unsigned i = 0, e = N0->getNumOperands(); i != e; ++i) {
6574 if (N0->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT
6575 && N1->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
6577 SDValue ExtVec0 = N0->getOperand(i);
6578 SDValue ExtVec1 = N1->getOperand(i);
6580 // First operand is the vector, verify its the same.
6581 if (V != ExtVec0->getOperand(0).getNode() ||
6582 V != ExtVec1->getOperand(0).getNode())
6585 // Second is the constant, verify its correct.
6586 ConstantSDNode *C0 = dyn_cast<ConstantSDNode>(ExtVec0->getOperand(1));
6587 ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(ExtVec1->getOperand(1));
6589 // For the constant, we want to see all the even or all the odd.
6590 if (!C0 || !C1 || C0->getZExtValue() != nextIndex
6591 || C1->getZExtValue() != nextIndex+1)
6600 // Create VPADDL node.
6601 SelectionDAG &DAG = DCI.DAG;
6602 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6604 // Build operand list.
6605 SmallVector<SDValue, 8> Ops;
6606 Ops.push_back(DAG.getConstant(Intrinsic::arm_neon_vpaddls,
6607 TLI.getPointerTy()));
6609 // Input is the vector.
6612 // Get widened type and narrowed type.
6614 unsigned numElem = VT.getVectorNumElements();
6615 switch (VT.getVectorElementType().getSimpleVT().SimpleTy) {
6616 case MVT::i8: widenType = MVT::getVectorVT(MVT::i16, numElem); break;
6617 case MVT::i16: widenType = MVT::getVectorVT(MVT::i32, numElem); break;
6618 case MVT::i32: widenType = MVT::getVectorVT(MVT::i64, numElem); break;
6620 assert(0 && "Invalid vector element type for padd optimization.");
6623 SDValue tmp = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(),
6624 widenType, &Ops[0], Ops.size());
6625 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, tmp);
6628 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
6629 /// operands N0 and N1. This is a helper for PerformADDCombine that is
6630 /// called with the default operands, and if that fails, with commuted
6632 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
6633 TargetLowering::DAGCombinerInfo &DCI,
6634 const ARMSubtarget *Subtarget){
6636 // Attempt to create vpaddl for this add.
6637 SDValue Result = AddCombineToVPADDL(N, N0, N1, DCI, Subtarget);
6638 if (Result.getNode())
6641 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
6642 if (N0.getOpcode() == ISD::SELECT && N0.getNode()->hasOneUse()) {
6643 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
6644 if (Result.getNode()) return Result;
6649 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
6651 static SDValue PerformADDCombine(SDNode *N,
6652 TargetLowering::DAGCombinerInfo &DCI,
6653 const ARMSubtarget *Subtarget) {
6654 SDValue N0 = N->getOperand(0);
6655 SDValue N1 = N->getOperand(1);
6657 // First try with the default operand order.
6658 SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI, Subtarget);
6659 if (Result.getNode())
6662 // If that didn't work, try again with the operands commuted.
6663 return PerformADDCombineWithOperands(N, N1, N0, DCI, Subtarget);
6666 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
6668 static SDValue PerformSUBCombine(SDNode *N,
6669 TargetLowering::DAGCombinerInfo &DCI) {
6670 SDValue N0 = N->getOperand(0);
6671 SDValue N1 = N->getOperand(1);
6673 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
6674 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
6675 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
6676 if (Result.getNode()) return Result;
6682 /// PerformVMULCombine
6683 /// Distribute (A + B) * C to (A * C) + (B * C) to take advantage of the
6684 /// special multiplier accumulator forwarding.
6690 static SDValue PerformVMULCombine(SDNode *N,
6691 TargetLowering::DAGCombinerInfo &DCI,
6692 const ARMSubtarget *Subtarget) {
6693 if (!Subtarget->hasVMLxForwarding())
6696 SelectionDAG &DAG = DCI.DAG;
6697 SDValue N0 = N->getOperand(0);
6698 SDValue N1 = N->getOperand(1);
6699 unsigned Opcode = N0.getOpcode();
6700 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
6701 Opcode != ISD::FADD && Opcode != ISD::FSUB) {
6702 Opcode = N1.getOpcode();
6703 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
6704 Opcode != ISD::FADD && Opcode != ISD::FSUB)
6709 EVT VT = N->getValueType(0);
6710 DebugLoc DL = N->getDebugLoc();
6711 SDValue N00 = N0->getOperand(0);
6712 SDValue N01 = N0->getOperand(1);
6713 return DAG.getNode(Opcode, DL, VT,
6714 DAG.getNode(ISD::MUL, DL, VT, N00, N1),
6715 DAG.getNode(ISD::MUL, DL, VT, N01, N1));
6718 static SDValue PerformMULCombine(SDNode *N,
6719 TargetLowering::DAGCombinerInfo &DCI,
6720 const ARMSubtarget *Subtarget) {
6721 SelectionDAG &DAG = DCI.DAG;
6723 if (Subtarget->isThumb1Only())
6726 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
6729 EVT VT = N->getValueType(0);
6730 if (VT.is64BitVector() || VT.is128BitVector())
6731 return PerformVMULCombine(N, DCI, Subtarget);
6735 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
6739 uint64_t MulAmt = C->getZExtValue();
6740 unsigned ShiftAmt = CountTrailingZeros_64(MulAmt);
6741 ShiftAmt = ShiftAmt & (32 - 1);
6742 SDValue V = N->getOperand(0);
6743 DebugLoc DL = N->getDebugLoc();
6746 MulAmt >>= ShiftAmt;
6747 if (isPowerOf2_32(MulAmt - 1)) {
6748 // (mul x, 2^N + 1) => (add (shl x, N), x)
6749 Res = DAG.getNode(ISD::ADD, DL, VT,
6750 V, DAG.getNode(ISD::SHL, DL, VT,
6751 V, DAG.getConstant(Log2_32(MulAmt-1),
6753 } else if (isPowerOf2_32(MulAmt + 1)) {
6754 // (mul x, 2^N - 1) => (sub (shl x, N), x)
6755 Res = DAG.getNode(ISD::SUB, DL, VT,
6756 DAG.getNode(ISD::SHL, DL, VT,
6757 V, DAG.getConstant(Log2_32(MulAmt+1),
6764 Res = DAG.getNode(ISD::SHL, DL, VT, Res,
6765 DAG.getConstant(ShiftAmt, MVT::i32));
6767 // Do not add new nodes to DAG combiner worklist.
6768 DCI.CombineTo(N, Res, false);
6772 static SDValue PerformANDCombine(SDNode *N,
6773 TargetLowering::DAGCombinerInfo &DCI) {
6775 // Attempt to use immediate-form VBIC
6776 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
6777 DebugLoc dl = N->getDebugLoc();
6778 EVT VT = N->getValueType(0);
6779 SelectionDAG &DAG = DCI.DAG;
6781 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
6784 APInt SplatBits, SplatUndef;
6785 unsigned SplatBitSize;
6788 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
6789 if (SplatBitSize <= 64) {
6791 SDValue Val = isNEONModifiedImm((~SplatBits).getZExtValue(),
6792 SplatUndef.getZExtValue(), SplatBitSize,
6793 DAG, VbicVT, VT.is128BitVector(),
6795 if (Val.getNode()) {
6797 DAG.getNode(ISD::BITCAST, dl, VbicVT, N->getOperand(0));
6798 SDValue Vbic = DAG.getNode(ARMISD::VBICIMM, dl, VbicVT, Input, Val);
6799 return DAG.getNode(ISD::BITCAST, dl, VT, Vbic);
6807 /// PerformORCombine - Target-specific dag combine xforms for ISD::OR
6808 static SDValue PerformORCombine(SDNode *N,
6809 TargetLowering::DAGCombinerInfo &DCI,
6810 const ARMSubtarget *Subtarget) {
6811 // Attempt to use immediate-form VORR
6812 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
6813 DebugLoc dl = N->getDebugLoc();
6814 EVT VT = N->getValueType(0);
6815 SelectionDAG &DAG = DCI.DAG;
6817 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
6820 APInt SplatBits, SplatUndef;
6821 unsigned SplatBitSize;
6823 if (BVN && Subtarget->hasNEON() &&
6824 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
6825 if (SplatBitSize <= 64) {
6827 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
6828 SplatUndef.getZExtValue(), SplatBitSize,
6829 DAG, VorrVT, VT.is128BitVector(),
6831 if (Val.getNode()) {
6833 DAG.getNode(ISD::BITCAST, dl, VorrVT, N->getOperand(0));
6834 SDValue Vorr = DAG.getNode(ARMISD::VORRIMM, dl, VorrVT, Input, Val);
6835 return DAG.getNode(ISD::BITCAST, dl, VT, Vorr);
6840 SDValue N0 = N->getOperand(0);
6841 if (N0.getOpcode() != ISD::AND)
6843 SDValue N1 = N->getOperand(1);
6845 // (or (and B, A), (and C, ~A)) => (VBSL A, B, C) when A is a constant.
6846 if (Subtarget->hasNEON() && N1.getOpcode() == ISD::AND && VT.isVector() &&
6847 DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
6849 unsigned SplatBitSize;
6852 BuildVectorSDNode *BVN0 = dyn_cast<BuildVectorSDNode>(N0->getOperand(1));
6854 if (BVN0 && BVN0->isConstantSplat(SplatBits0, SplatUndef, SplatBitSize,
6855 HasAnyUndefs) && !HasAnyUndefs) {
6856 BuildVectorSDNode *BVN1 = dyn_cast<BuildVectorSDNode>(N1->getOperand(1));
6858 if (BVN1 && BVN1->isConstantSplat(SplatBits1, SplatUndef, SplatBitSize,
6859 HasAnyUndefs) && !HasAnyUndefs &&
6860 SplatBits0 == ~SplatBits1) {
6861 // Canonicalize the vector type to make instruction selection simpler.
6862 EVT CanonicalVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
6863 SDValue Result = DAG.getNode(ARMISD::VBSL, dl, CanonicalVT,
6864 N0->getOperand(1), N0->getOperand(0),
6866 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
6871 // Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when
6874 // BFI is only available on V6T2+
6875 if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops())
6878 DebugLoc DL = N->getDebugLoc();
6879 // 1) or (and A, mask), val => ARMbfi A, val, mask
6880 // iff (val & mask) == val
6882 // 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
6883 // 2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2)
6884 // && mask == ~mask2
6885 // 2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2)
6886 // && ~mask == mask2
6887 // (i.e., copy a bitfield value into another bitfield of the same width)
6892 SDValue N00 = N0.getOperand(0);
6894 // The value and the mask need to be constants so we can verify this is
6895 // actually a bitfield set. If the mask is 0xffff, we can do better
6896 // via a movt instruction, so don't use BFI in that case.
6897 SDValue MaskOp = N0.getOperand(1);
6898 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(MaskOp);
6901 unsigned Mask = MaskC->getZExtValue();
6905 // Case (1): or (and A, mask), val => ARMbfi A, val, mask
6906 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
6908 unsigned Val = N1C->getZExtValue();
6909 if ((Val & ~Mask) != Val)
6912 if (ARM::isBitFieldInvertedMask(Mask)) {
6913 Val >>= CountTrailingZeros_32(~Mask);
6915 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00,
6916 DAG.getConstant(Val, MVT::i32),
6917 DAG.getConstant(Mask, MVT::i32));
6919 // Do not add new nodes to DAG combiner worklist.
6920 DCI.CombineTo(N, Res, false);
6923 } else if (N1.getOpcode() == ISD::AND) {
6924 // case (2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
6925 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
6928 unsigned Mask2 = N11C->getZExtValue();
6930 // Mask and ~Mask2 (or reverse) must be equivalent for the BFI pattern
6932 if (ARM::isBitFieldInvertedMask(Mask) &&
6934 // The pack halfword instruction works better for masks that fit it,
6935 // so use that when it's available.
6936 if (Subtarget->hasT2ExtractPack() &&
6937 (Mask == 0xffff || Mask == 0xffff0000))
6940 unsigned amt = CountTrailingZeros_32(Mask2);
6941 Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0),
6942 DAG.getConstant(amt, MVT::i32));
6943 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00, Res,
6944 DAG.getConstant(Mask, MVT::i32));
6945 // Do not add new nodes to DAG combiner worklist.
6946 DCI.CombineTo(N, Res, false);
6948 } else if (ARM::isBitFieldInvertedMask(~Mask) &&
6950 // The pack halfword instruction works better for masks that fit it,
6951 // so use that when it's available.
6952 if (Subtarget->hasT2ExtractPack() &&
6953 (Mask2 == 0xffff || Mask2 == 0xffff0000))
6956 unsigned lsb = CountTrailingZeros_32(Mask);
6957 Res = DAG.getNode(ISD::SRL, DL, VT, N00,
6958 DAG.getConstant(lsb, MVT::i32));
6959 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res,
6960 DAG.getConstant(Mask2, MVT::i32));
6961 // Do not add new nodes to DAG combiner worklist.
6962 DCI.CombineTo(N, Res, false);
6967 if (DAG.MaskedValueIsZero(N1, MaskC->getAPIntValue()) &&
6968 N00.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N00.getOperand(1)) &&
6969 ARM::isBitFieldInvertedMask(~Mask)) {
6970 // Case (3): or (and (shl A, #shamt), mask), B => ARMbfi B, A, ~mask
6971 // where lsb(mask) == #shamt and masked bits of B are known zero.
6972 SDValue ShAmt = N00.getOperand(1);
6973 unsigned ShAmtC = cast<ConstantSDNode>(ShAmt)->getZExtValue();
6974 unsigned LSB = CountTrailingZeros_32(Mask);
6978 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1, N00.getOperand(0),
6979 DAG.getConstant(~Mask, MVT::i32));
6981 // Do not add new nodes to DAG combiner worklist.
6982 DCI.CombineTo(N, Res, false);
6988 /// PerformBFICombine - (bfi A, (and B, Mask1), Mask2) -> (bfi A, B, Mask2) iff
6989 /// the bits being cleared by the AND are not demanded by the BFI.
6990 static SDValue PerformBFICombine(SDNode *N,
6991 TargetLowering::DAGCombinerInfo &DCI) {
6992 SDValue N1 = N->getOperand(1);
6993 if (N1.getOpcode() == ISD::AND) {
6994 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
6997 unsigned InvMask = cast<ConstantSDNode>(N->getOperand(2))->getZExtValue();
6998 unsigned LSB = CountTrailingZeros_32(~InvMask);
6999 unsigned Width = (32 - CountLeadingZeros_32(~InvMask)) - LSB;
7000 unsigned Mask = (1 << Width)-1;
7001 unsigned Mask2 = N11C->getZExtValue();
7002 if ((Mask & (~Mask2)) == 0)
7003 return DCI.DAG.getNode(ARMISD::BFI, N->getDebugLoc(), N->getValueType(0),
7004 N->getOperand(0), N1.getOperand(0),
7010 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
7011 /// ARMISD::VMOVRRD.
7012 static SDValue PerformVMOVRRDCombine(SDNode *N,
7013 TargetLowering::DAGCombinerInfo &DCI) {
7014 // vmovrrd(vmovdrr x, y) -> x,y
7015 SDValue InDouble = N->getOperand(0);
7016 if (InDouble.getOpcode() == ARMISD::VMOVDRR)
7017 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
7019 // vmovrrd(load f64) -> (load i32), (load i32)
7020 SDNode *InNode = InDouble.getNode();
7021 if (ISD::isNormalLoad(InNode) && InNode->hasOneUse() &&
7022 InNode->getValueType(0) == MVT::f64 &&
7023 InNode->getOperand(1).getOpcode() == ISD::FrameIndex &&
7024 !cast<LoadSDNode>(InNode)->isVolatile()) {
7025 // TODO: Should this be done for non-FrameIndex operands?
7026 LoadSDNode *LD = cast<LoadSDNode>(InNode);
7028 SelectionDAG &DAG = DCI.DAG;
7029 DebugLoc DL = LD->getDebugLoc();
7030 SDValue BasePtr = LD->getBasePtr();
7031 SDValue NewLD1 = DAG.getLoad(MVT::i32, DL, LD->getChain(), BasePtr,
7032 LD->getPointerInfo(), LD->isVolatile(),
7033 LD->isNonTemporal(), LD->getAlignment());
7035 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
7036 DAG.getConstant(4, MVT::i32));
7037 SDValue NewLD2 = DAG.getLoad(MVT::i32, DL, NewLD1.getValue(1), OffsetPtr,
7038 LD->getPointerInfo(), LD->isVolatile(),
7039 LD->isNonTemporal(),
7040 std::min(4U, LD->getAlignment() / 2));
7042 DAG.ReplaceAllUsesOfValueWith(SDValue(LD, 1), NewLD2.getValue(1));
7043 SDValue Result = DCI.CombineTo(N, NewLD1, NewLD2);
7044 DCI.RemoveFromWorklist(LD);
7052 /// PerformVMOVDRRCombine - Target-specific dag combine xforms for
7053 /// ARMISD::VMOVDRR. This is also used for BUILD_VECTORs with 2 operands.
7054 static SDValue PerformVMOVDRRCombine(SDNode *N, SelectionDAG &DAG) {
7055 // N=vmovrrd(X); vmovdrr(N:0, N:1) -> bit_convert(X)
7056 SDValue Op0 = N->getOperand(0);
7057 SDValue Op1 = N->getOperand(1);
7058 if (Op0.getOpcode() == ISD::BITCAST)
7059 Op0 = Op0.getOperand(0);
7060 if (Op1.getOpcode() == ISD::BITCAST)
7061 Op1 = Op1.getOperand(0);
7062 if (Op0.getOpcode() == ARMISD::VMOVRRD &&
7063 Op0.getNode() == Op1.getNode() &&
7064 Op0.getResNo() == 0 && Op1.getResNo() == 1)
7065 return DAG.getNode(ISD::BITCAST, N->getDebugLoc(),
7066 N->getValueType(0), Op0.getOperand(0));
7070 /// PerformSTORECombine - Target-specific dag combine xforms for
7072 static SDValue PerformSTORECombine(SDNode *N,
7073 TargetLowering::DAGCombinerInfo &DCI) {
7074 // Bitcast an i64 store extracted from a vector to f64.
7075 // Otherwise, the i64 value will be legalized to a pair of i32 values.
7076 StoreSDNode *St = cast<StoreSDNode>(N);
7077 SDValue StVal = St->getValue();
7078 if (!ISD::isNormalStore(St) || St->isVolatile())
7081 if (StVal.getNode()->getOpcode() == ARMISD::VMOVDRR &&
7082 StVal.getNode()->hasOneUse() && !St->isVolatile()) {
7083 SelectionDAG &DAG = DCI.DAG;
7084 DebugLoc DL = St->getDebugLoc();
7085 SDValue BasePtr = St->getBasePtr();
7086 SDValue NewST1 = DAG.getStore(St->getChain(), DL,
7087 StVal.getNode()->getOperand(0), BasePtr,
7088 St->getPointerInfo(), St->isVolatile(),
7089 St->isNonTemporal(), St->getAlignment());
7091 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
7092 DAG.getConstant(4, MVT::i32));
7093 return DAG.getStore(NewST1.getValue(0), DL, StVal.getNode()->getOperand(1),
7094 OffsetPtr, St->getPointerInfo(), St->isVolatile(),
7095 St->isNonTemporal(),
7096 std::min(4U, St->getAlignment() / 2));
7099 if (StVal.getValueType() != MVT::i64 ||
7100 StVal.getNode()->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
7103 SelectionDAG &DAG = DCI.DAG;
7104 DebugLoc dl = StVal.getDebugLoc();
7105 SDValue IntVec = StVal.getOperand(0);
7106 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
7107 IntVec.getValueType().getVectorNumElements());
7108 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, IntVec);
7109 SDValue ExtElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
7110 Vec, StVal.getOperand(1));
7111 dl = N->getDebugLoc();
7112 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ExtElt);
7113 // Make the DAGCombiner fold the bitcasts.
7114 DCI.AddToWorklist(Vec.getNode());
7115 DCI.AddToWorklist(ExtElt.getNode());
7116 DCI.AddToWorklist(V.getNode());
7117 return DAG.getStore(St->getChain(), dl, V, St->getBasePtr(),
7118 St->getPointerInfo(), St->isVolatile(),
7119 St->isNonTemporal(), St->getAlignment(),
7123 /// hasNormalLoadOperand - Check if any of the operands of a BUILD_VECTOR node
7124 /// are normal, non-volatile loads. If so, it is profitable to bitcast an
7125 /// i64 vector to have f64 elements, since the value can then be loaded
7126 /// directly into a VFP register.
7127 static bool hasNormalLoadOperand(SDNode *N) {
7128 unsigned NumElts = N->getValueType(0).getVectorNumElements();
7129 for (unsigned i = 0; i < NumElts; ++i) {
7130 SDNode *Elt = N->getOperand(i).getNode();
7131 if (ISD::isNormalLoad(Elt) && !cast<LoadSDNode>(Elt)->isVolatile())
7137 /// PerformBUILD_VECTORCombine - Target-specific dag combine xforms for
7138 /// ISD::BUILD_VECTOR.
7139 static SDValue PerformBUILD_VECTORCombine(SDNode *N,
7140 TargetLowering::DAGCombinerInfo &DCI){
7141 // build_vector(N=ARMISD::VMOVRRD(X), N:1) -> bit_convert(X):
7142 // VMOVRRD is introduced when legalizing i64 types. It forces the i64 value
7143 // into a pair of GPRs, which is fine when the value is used as a scalar,
7144 // but if the i64 value is converted to a vector, we need to undo the VMOVRRD.
7145 SelectionDAG &DAG = DCI.DAG;
7146 if (N->getNumOperands() == 2) {
7147 SDValue RV = PerformVMOVDRRCombine(N, DAG);
7152 // Load i64 elements as f64 values so that type legalization does not split
7153 // them up into i32 values.
7154 EVT VT = N->getValueType(0);
7155 if (VT.getVectorElementType() != MVT::i64 || !hasNormalLoadOperand(N))
7157 DebugLoc dl = N->getDebugLoc();
7158 SmallVector<SDValue, 8> Ops;
7159 unsigned NumElts = VT.getVectorNumElements();
7160 for (unsigned i = 0; i < NumElts; ++i) {
7161 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(i));
7163 // Make the DAGCombiner fold the bitcast.
7164 DCI.AddToWorklist(V.getNode());
7166 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, NumElts);
7167 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, FloatVT, Ops.data(), NumElts);
7168 return DAG.getNode(ISD::BITCAST, dl, VT, BV);
7171 /// PerformInsertEltCombine - Target-specific dag combine xforms for
7172 /// ISD::INSERT_VECTOR_ELT.
7173 static SDValue PerformInsertEltCombine(SDNode *N,
7174 TargetLowering::DAGCombinerInfo &DCI) {
7175 // Bitcast an i64 load inserted into a vector to f64.
7176 // Otherwise, the i64 value will be legalized to a pair of i32 values.
7177 EVT VT = N->getValueType(0);
7178 SDNode *Elt = N->getOperand(1).getNode();
7179 if (VT.getVectorElementType() != MVT::i64 ||
7180 !ISD::isNormalLoad(Elt) || cast<LoadSDNode>(Elt)->isVolatile())
7183 SelectionDAG &DAG = DCI.DAG;
7184 DebugLoc dl = N->getDebugLoc();
7185 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
7186 VT.getVectorNumElements());
7187 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, N->getOperand(0));
7188 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(1));
7189 // Make the DAGCombiner fold the bitcasts.
7190 DCI.AddToWorklist(Vec.getNode());
7191 DCI.AddToWorklist(V.getNode());
7192 SDValue InsElt = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, FloatVT,
7193 Vec, V, N->getOperand(2));
7194 return DAG.getNode(ISD::BITCAST, dl, VT, InsElt);
7197 /// PerformVECTOR_SHUFFLECombine - Target-specific dag combine xforms for
7198 /// ISD::VECTOR_SHUFFLE.
7199 static SDValue PerformVECTOR_SHUFFLECombine(SDNode *N, SelectionDAG &DAG) {
7200 // The LLVM shufflevector instruction does not require the shuffle mask
7201 // length to match the operand vector length, but ISD::VECTOR_SHUFFLE does
7202 // have that requirement. When translating to ISD::VECTOR_SHUFFLE, if the
7203 // operands do not match the mask length, they are extended by concatenating
7204 // them with undef vectors. That is probably the right thing for other
7205 // targets, but for NEON it is better to concatenate two double-register
7206 // size vector operands into a single quad-register size vector. Do that
7207 // transformation here:
7208 // shuffle(concat(v1, undef), concat(v2, undef)) ->
7209 // shuffle(concat(v1, v2), undef)
7210 SDValue Op0 = N->getOperand(0);
7211 SDValue Op1 = N->getOperand(1);
7212 if (Op0.getOpcode() != ISD::CONCAT_VECTORS ||
7213 Op1.getOpcode() != ISD::CONCAT_VECTORS ||
7214 Op0.getNumOperands() != 2 ||
7215 Op1.getNumOperands() != 2)
7217 SDValue Concat0Op1 = Op0.getOperand(1);
7218 SDValue Concat1Op1 = Op1.getOperand(1);
7219 if (Concat0Op1.getOpcode() != ISD::UNDEF ||
7220 Concat1Op1.getOpcode() != ISD::UNDEF)
7222 // Skip the transformation if any of the types are illegal.
7223 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7224 EVT VT = N->getValueType(0);
7225 if (!TLI.isTypeLegal(VT) ||
7226 !TLI.isTypeLegal(Concat0Op1.getValueType()) ||
7227 !TLI.isTypeLegal(Concat1Op1.getValueType()))
7230 SDValue NewConcat = DAG.getNode(ISD::CONCAT_VECTORS, N->getDebugLoc(), VT,
7231 Op0.getOperand(0), Op1.getOperand(0));
7232 // Translate the shuffle mask.
7233 SmallVector<int, 16> NewMask;
7234 unsigned NumElts = VT.getVectorNumElements();
7235 unsigned HalfElts = NumElts/2;
7236 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
7237 for (unsigned n = 0; n < NumElts; ++n) {
7238 int MaskElt = SVN->getMaskElt(n);
7240 if (MaskElt < (int)HalfElts)
7242 else if (MaskElt >= (int)NumElts && MaskElt < (int)(NumElts + HalfElts))
7243 NewElt = HalfElts + MaskElt - NumElts;
7244 NewMask.push_back(NewElt);
7246 return DAG.getVectorShuffle(VT, N->getDebugLoc(), NewConcat,
7247 DAG.getUNDEF(VT), NewMask.data());
7250 /// CombineBaseUpdate - Target-specific DAG combine function for VLDDUP and
7251 /// NEON load/store intrinsics to merge base address updates.
7252 static SDValue CombineBaseUpdate(SDNode *N,
7253 TargetLowering::DAGCombinerInfo &DCI) {
7254 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
7257 SelectionDAG &DAG = DCI.DAG;
7258 bool isIntrinsic = (N->getOpcode() == ISD::INTRINSIC_VOID ||
7259 N->getOpcode() == ISD::INTRINSIC_W_CHAIN);
7260 unsigned AddrOpIdx = (isIntrinsic ? 2 : 1);
7261 SDValue Addr = N->getOperand(AddrOpIdx);
7263 // Search for a use of the address operand that is an increment.
7264 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(),
7265 UE = Addr.getNode()->use_end(); UI != UE; ++UI) {
7267 if (User->getOpcode() != ISD::ADD ||
7268 UI.getUse().getResNo() != Addr.getResNo())
7271 // Check that the add is independent of the load/store. Otherwise, folding
7272 // it would create a cycle.
7273 if (User->isPredecessorOf(N) || N->isPredecessorOf(User))
7276 // Find the new opcode for the updating load/store.
7278 bool isLaneOp = false;
7279 unsigned NewOpc = 0;
7280 unsigned NumVecs = 0;
7282 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
7284 default: assert(0 && "unexpected intrinsic for Neon base update");
7285 case Intrinsic::arm_neon_vld1: NewOpc = ARMISD::VLD1_UPD;
7287 case Intrinsic::arm_neon_vld2: NewOpc = ARMISD::VLD2_UPD;
7289 case Intrinsic::arm_neon_vld3: NewOpc = ARMISD::VLD3_UPD;
7291 case Intrinsic::arm_neon_vld4: NewOpc = ARMISD::VLD4_UPD;
7293 case Intrinsic::arm_neon_vld2lane: NewOpc = ARMISD::VLD2LN_UPD;
7294 NumVecs = 2; isLaneOp = true; break;
7295 case Intrinsic::arm_neon_vld3lane: NewOpc = ARMISD::VLD3LN_UPD;
7296 NumVecs = 3; isLaneOp = true; break;
7297 case Intrinsic::arm_neon_vld4lane: NewOpc = ARMISD::VLD4LN_UPD;
7298 NumVecs = 4; isLaneOp = true; break;
7299 case Intrinsic::arm_neon_vst1: NewOpc = ARMISD::VST1_UPD;
7300 NumVecs = 1; isLoad = false; break;
7301 case Intrinsic::arm_neon_vst2: NewOpc = ARMISD::VST2_UPD;
7302 NumVecs = 2; isLoad = false; break;
7303 case Intrinsic::arm_neon_vst3: NewOpc = ARMISD::VST3_UPD;
7304 NumVecs = 3; isLoad = false; break;
7305 case Intrinsic::arm_neon_vst4: NewOpc = ARMISD::VST4_UPD;
7306 NumVecs = 4; isLoad = false; break;
7307 case Intrinsic::arm_neon_vst2lane: NewOpc = ARMISD::VST2LN_UPD;
7308 NumVecs = 2; isLoad = false; isLaneOp = true; break;
7309 case Intrinsic::arm_neon_vst3lane: NewOpc = ARMISD::VST3LN_UPD;
7310 NumVecs = 3; isLoad = false; isLaneOp = true; break;
7311 case Intrinsic::arm_neon_vst4lane: NewOpc = ARMISD::VST4LN_UPD;
7312 NumVecs = 4; isLoad = false; isLaneOp = true; break;
7316 switch (N->getOpcode()) {
7317 default: assert(0 && "unexpected opcode for Neon base update");
7318 case ARMISD::VLD2DUP: NewOpc = ARMISD::VLD2DUP_UPD; NumVecs = 2; break;
7319 case ARMISD::VLD3DUP: NewOpc = ARMISD::VLD3DUP_UPD; NumVecs = 3; break;
7320 case ARMISD::VLD4DUP: NewOpc = ARMISD::VLD4DUP_UPD; NumVecs = 4; break;
7324 // Find the size of memory referenced by the load/store.
7327 VecTy = N->getValueType(0);
7329 VecTy = N->getOperand(AddrOpIdx+1).getValueType();
7330 unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8;
7332 NumBytes /= VecTy.getVectorNumElements();
7334 // If the increment is a constant, it must match the memory ref size.
7335 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
7336 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
7337 uint64_t IncVal = CInc->getZExtValue();
7338 if (IncVal != NumBytes)
7340 } else if (NumBytes >= 3 * 16) {
7341 // VLD3/4 and VST3/4 for 128-bit vectors are implemented with two
7342 // separate instructions that make it harder to use a non-constant update.
7346 // Create the new updating load/store node.
7348 unsigned NumResultVecs = (isLoad ? NumVecs : 0);
7350 for (n = 0; n < NumResultVecs; ++n)
7352 Tys[n++] = MVT::i32;
7353 Tys[n] = MVT::Other;
7354 SDVTList SDTys = DAG.getVTList(Tys, NumResultVecs+2);
7355 SmallVector<SDValue, 8> Ops;
7356 Ops.push_back(N->getOperand(0)); // incoming chain
7357 Ops.push_back(N->getOperand(AddrOpIdx));
7359 for (unsigned i = AddrOpIdx + 1; i < N->getNumOperands(); ++i) {
7360 Ops.push_back(N->getOperand(i));
7362 MemIntrinsicSDNode *MemInt = cast<MemIntrinsicSDNode>(N);
7363 SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, N->getDebugLoc(), SDTys,
7364 Ops.data(), Ops.size(),
7365 MemInt->getMemoryVT(),
7366 MemInt->getMemOperand());
7369 std::vector<SDValue> NewResults;
7370 for (unsigned i = 0; i < NumResultVecs; ++i) {
7371 NewResults.push_back(SDValue(UpdN.getNode(), i));
7373 NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs+1)); // chain
7374 DCI.CombineTo(N, NewResults);
7375 DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs));
7382 /// CombineVLDDUP - For a VDUPLANE node N, check if its source operand is a
7383 /// vldN-lane (N > 1) intrinsic, and if all the other uses of that intrinsic
7384 /// are also VDUPLANEs. If so, combine them to a vldN-dup operation and
7386 static bool CombineVLDDUP(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
7387 SelectionDAG &DAG = DCI.DAG;
7388 EVT VT = N->getValueType(0);
7389 // vldN-dup instructions only support 64-bit vectors for N > 1.
7390 if (!VT.is64BitVector())
7393 // Check if the VDUPLANE operand is a vldN-dup intrinsic.
7394 SDNode *VLD = N->getOperand(0).getNode();
7395 if (VLD->getOpcode() != ISD::INTRINSIC_W_CHAIN)
7397 unsigned NumVecs = 0;
7398 unsigned NewOpc = 0;
7399 unsigned IntNo = cast<ConstantSDNode>(VLD->getOperand(1))->getZExtValue();
7400 if (IntNo == Intrinsic::arm_neon_vld2lane) {
7402 NewOpc = ARMISD::VLD2DUP;
7403 } else if (IntNo == Intrinsic::arm_neon_vld3lane) {
7405 NewOpc = ARMISD::VLD3DUP;
7406 } else if (IntNo == Intrinsic::arm_neon_vld4lane) {
7408 NewOpc = ARMISD::VLD4DUP;
7413 // First check that all the vldN-lane uses are VDUPLANEs and that the lane
7414 // numbers match the load.
7415 unsigned VLDLaneNo =
7416 cast<ConstantSDNode>(VLD->getOperand(NumVecs+3))->getZExtValue();
7417 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
7419 // Ignore uses of the chain result.
7420 if (UI.getUse().getResNo() == NumVecs)
7423 if (User->getOpcode() != ARMISD::VDUPLANE ||
7424 VLDLaneNo != cast<ConstantSDNode>(User->getOperand(1))->getZExtValue())
7428 // Create the vldN-dup node.
7431 for (n = 0; n < NumVecs; ++n)
7433 Tys[n] = MVT::Other;
7434 SDVTList SDTys = DAG.getVTList(Tys, NumVecs+1);
7435 SDValue Ops[] = { VLD->getOperand(0), VLD->getOperand(2) };
7436 MemIntrinsicSDNode *VLDMemInt = cast<MemIntrinsicSDNode>(VLD);
7437 SDValue VLDDup = DAG.getMemIntrinsicNode(NewOpc, VLD->getDebugLoc(), SDTys,
7438 Ops, 2, VLDMemInt->getMemoryVT(),
7439 VLDMemInt->getMemOperand());
7442 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
7444 unsigned ResNo = UI.getUse().getResNo();
7445 // Ignore uses of the chain result.
7446 if (ResNo == NumVecs)
7449 DCI.CombineTo(User, SDValue(VLDDup.getNode(), ResNo));
7452 // Now the vldN-lane intrinsic is dead except for its chain result.
7453 // Update uses of the chain.
7454 std::vector<SDValue> VLDDupResults;
7455 for (unsigned n = 0; n < NumVecs; ++n)
7456 VLDDupResults.push_back(SDValue(VLDDup.getNode(), n));
7457 VLDDupResults.push_back(SDValue(VLDDup.getNode(), NumVecs));
7458 DCI.CombineTo(VLD, VLDDupResults);
7463 /// PerformVDUPLANECombine - Target-specific dag combine xforms for
7464 /// ARMISD::VDUPLANE.
7465 static SDValue PerformVDUPLANECombine(SDNode *N,
7466 TargetLowering::DAGCombinerInfo &DCI) {
7467 SDValue Op = N->getOperand(0);
7469 // If the source is a vldN-lane (N > 1) intrinsic, and all the other uses
7470 // of that intrinsic are also VDUPLANEs, combine them to a vldN-dup operation.
7471 if (CombineVLDDUP(N, DCI))
7472 return SDValue(N, 0);
7474 // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is
7475 // redundant. Ignore bit_converts for now; element sizes are checked below.
7476 while (Op.getOpcode() == ISD::BITCAST)
7477 Op = Op.getOperand(0);
7478 if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM)
7481 // Make sure the VMOV element size is not bigger than the VDUPLANE elements.
7482 unsigned EltSize = Op.getValueType().getVectorElementType().getSizeInBits();
7483 // The canonical VMOV for a zero vector uses a 32-bit element size.
7484 unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7486 if (ARM_AM::decodeNEONModImm(Imm, EltBits) == 0)
7488 EVT VT = N->getValueType(0);
7489 if (EltSize > VT.getVectorElementType().getSizeInBits())
7492 return DCI.DAG.getNode(ISD::BITCAST, N->getDebugLoc(), VT, Op);
7495 // isConstVecPow2 - Return true if each vector element is a power of 2, all
7496 // elements are the same constant, C, and Log2(C) ranges from 1 to 32.
7497 static bool isConstVecPow2(SDValue ConstVec, bool isSigned, uint64_t &C)
7501 for (unsigned I = 0, E = ConstVec.getValueType().getVectorNumElements();
7503 ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(ConstVec.getOperand(I));
7508 APFloat APF = C->getValueAPF();
7509 if (APF.convertToInteger(&cN, 64, isSigned, APFloat::rmTowardZero, &isExact)
7510 != APFloat::opOK || !isExact)
7513 c0 = (I == 0) ? cN : c0;
7514 if (!isPowerOf2_64(cN) || c0 != cN || Log2_64(c0) < 1 || Log2_64(c0) > 32)
7521 /// PerformVCVTCombine - VCVT (floating-point to fixed-point, Advanced SIMD)
7522 /// can replace combinations of VMUL and VCVT (floating-point to integer)
7523 /// when the VMUL has a constant operand that is a power of 2.
7525 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
7526 /// vmul.f32 d16, d17, d16
7527 /// vcvt.s32.f32 d16, d16
7529 /// vcvt.s32.f32 d16, d16, #3
7530 static SDValue PerformVCVTCombine(SDNode *N,
7531 TargetLowering::DAGCombinerInfo &DCI,
7532 const ARMSubtarget *Subtarget) {
7533 SelectionDAG &DAG = DCI.DAG;
7534 SDValue Op = N->getOperand(0);
7536 if (!Subtarget->hasNEON() || !Op.getValueType().isVector() ||
7537 Op.getOpcode() != ISD::FMUL)
7541 SDValue N0 = Op->getOperand(0);
7542 SDValue ConstVec = Op->getOperand(1);
7543 bool isSigned = N->getOpcode() == ISD::FP_TO_SINT;
7545 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
7546 !isConstVecPow2(ConstVec, isSigned, C))
7549 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfp2fxs :
7550 Intrinsic::arm_neon_vcvtfp2fxu;
7551 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(),
7553 DAG.getConstant(IntrinsicOpcode, MVT::i32), N0,
7554 DAG.getConstant(Log2_64(C), MVT::i32));
7557 /// PerformVDIVCombine - VCVT (fixed-point to floating-point, Advanced SIMD)
7558 /// can replace combinations of VCVT (integer to floating-point) and VDIV
7559 /// when the VDIV has a constant operand that is a power of 2.
7561 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
7562 /// vcvt.f32.s32 d16, d16
7563 /// vdiv.f32 d16, d17, d16
7565 /// vcvt.f32.s32 d16, d16, #3
7566 static SDValue PerformVDIVCombine(SDNode *N,
7567 TargetLowering::DAGCombinerInfo &DCI,
7568 const ARMSubtarget *Subtarget) {
7569 SelectionDAG &DAG = DCI.DAG;
7570 SDValue Op = N->getOperand(0);
7571 unsigned OpOpcode = Op.getNode()->getOpcode();
7573 if (!Subtarget->hasNEON() || !N->getValueType(0).isVector() ||
7574 (OpOpcode != ISD::SINT_TO_FP && OpOpcode != ISD::UINT_TO_FP))
7578 SDValue ConstVec = N->getOperand(1);
7579 bool isSigned = OpOpcode == ISD::SINT_TO_FP;
7581 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
7582 !isConstVecPow2(ConstVec, isSigned, C))
7585 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfxs2fp :
7586 Intrinsic::arm_neon_vcvtfxu2fp;
7587 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(),
7589 DAG.getConstant(IntrinsicOpcode, MVT::i32),
7590 Op.getOperand(0), DAG.getConstant(Log2_64(C), MVT::i32));
7593 /// Getvshiftimm - Check if this is a valid build_vector for the immediate
7594 /// operand of a vector shift operation, where all the elements of the
7595 /// build_vector must have the same constant integer value.
7596 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
7597 // Ignore bit_converts.
7598 while (Op.getOpcode() == ISD::BITCAST)
7599 Op = Op.getOperand(0);
7600 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
7601 APInt SplatBits, SplatUndef;
7602 unsigned SplatBitSize;
7604 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
7605 HasAnyUndefs, ElementBits) ||
7606 SplatBitSize > ElementBits)
7608 Cnt = SplatBits.getSExtValue();
7612 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
7613 /// operand of a vector shift left operation. That value must be in the range:
7614 /// 0 <= Value < ElementBits for a left shift; or
7615 /// 0 <= Value <= ElementBits for a long left shift.
7616 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
7617 assert(VT.isVector() && "vector shift count is not a vector type");
7618 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
7619 if (! getVShiftImm(Op, ElementBits, Cnt))
7621 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
7624 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
7625 /// operand of a vector shift right operation. For a shift opcode, the value
7626 /// is positive, but for an intrinsic the value count must be negative. The
7627 /// absolute value must be in the range:
7628 /// 1 <= |Value| <= ElementBits for a right shift; or
7629 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
7630 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
7632 assert(VT.isVector() && "vector shift count is not a vector type");
7633 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
7634 if (! getVShiftImm(Op, ElementBits, Cnt))
7638 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
7641 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
7642 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
7643 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
7646 // Don't do anything for most intrinsics.
7649 // Vector shifts: check for immediate versions and lower them.
7650 // Note: This is done during DAG combining instead of DAG legalizing because
7651 // the build_vectors for 64-bit vector element shift counts are generally
7652 // not legal, and it is hard to see their values after they get legalized to
7653 // loads from a constant pool.
7654 case Intrinsic::arm_neon_vshifts:
7655 case Intrinsic::arm_neon_vshiftu:
7656 case Intrinsic::arm_neon_vshiftls:
7657 case Intrinsic::arm_neon_vshiftlu:
7658 case Intrinsic::arm_neon_vshiftn:
7659 case Intrinsic::arm_neon_vrshifts:
7660 case Intrinsic::arm_neon_vrshiftu:
7661 case Intrinsic::arm_neon_vrshiftn:
7662 case Intrinsic::arm_neon_vqshifts:
7663 case Intrinsic::arm_neon_vqshiftu:
7664 case Intrinsic::arm_neon_vqshiftsu:
7665 case Intrinsic::arm_neon_vqshiftns:
7666 case Intrinsic::arm_neon_vqshiftnu:
7667 case Intrinsic::arm_neon_vqshiftnsu:
7668 case Intrinsic::arm_neon_vqrshiftns:
7669 case Intrinsic::arm_neon_vqrshiftnu:
7670 case Intrinsic::arm_neon_vqrshiftnsu: {
7671 EVT VT = N->getOperand(1).getValueType();
7673 unsigned VShiftOpc = 0;
7676 case Intrinsic::arm_neon_vshifts:
7677 case Intrinsic::arm_neon_vshiftu:
7678 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
7679 VShiftOpc = ARMISD::VSHL;
7682 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
7683 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
7684 ARMISD::VSHRs : ARMISD::VSHRu);
7689 case Intrinsic::arm_neon_vshiftls:
7690 case Intrinsic::arm_neon_vshiftlu:
7691 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
7693 llvm_unreachable("invalid shift count for vshll intrinsic");
7695 case Intrinsic::arm_neon_vrshifts:
7696 case Intrinsic::arm_neon_vrshiftu:
7697 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
7701 case Intrinsic::arm_neon_vqshifts:
7702 case Intrinsic::arm_neon_vqshiftu:
7703 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
7707 case Intrinsic::arm_neon_vqshiftsu:
7708 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
7710 llvm_unreachable("invalid shift count for vqshlu intrinsic");
7712 case Intrinsic::arm_neon_vshiftn:
7713 case Intrinsic::arm_neon_vrshiftn:
7714 case Intrinsic::arm_neon_vqshiftns:
7715 case Intrinsic::arm_neon_vqshiftnu:
7716 case Intrinsic::arm_neon_vqshiftnsu:
7717 case Intrinsic::arm_neon_vqrshiftns:
7718 case Intrinsic::arm_neon_vqrshiftnu:
7719 case Intrinsic::arm_neon_vqrshiftnsu:
7720 // Narrowing shifts require an immediate right shift.
7721 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
7723 llvm_unreachable("invalid shift count for narrowing vector shift "
7727 llvm_unreachable("unhandled vector shift");
7731 case Intrinsic::arm_neon_vshifts:
7732 case Intrinsic::arm_neon_vshiftu:
7733 // Opcode already set above.
7735 case Intrinsic::arm_neon_vshiftls:
7736 case Intrinsic::arm_neon_vshiftlu:
7737 if (Cnt == VT.getVectorElementType().getSizeInBits())
7738 VShiftOpc = ARMISD::VSHLLi;
7740 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
7741 ARMISD::VSHLLs : ARMISD::VSHLLu);
7743 case Intrinsic::arm_neon_vshiftn:
7744 VShiftOpc = ARMISD::VSHRN; break;
7745 case Intrinsic::arm_neon_vrshifts:
7746 VShiftOpc = ARMISD::VRSHRs; break;
7747 case Intrinsic::arm_neon_vrshiftu:
7748 VShiftOpc = ARMISD::VRSHRu; break;
7749 case Intrinsic::arm_neon_vrshiftn:
7750 VShiftOpc = ARMISD::VRSHRN; break;
7751 case Intrinsic::arm_neon_vqshifts:
7752 VShiftOpc = ARMISD::VQSHLs; break;
7753 case Intrinsic::arm_neon_vqshiftu:
7754 VShiftOpc = ARMISD::VQSHLu; break;
7755 case Intrinsic::arm_neon_vqshiftsu:
7756 VShiftOpc = ARMISD::VQSHLsu; break;
7757 case Intrinsic::arm_neon_vqshiftns:
7758 VShiftOpc = ARMISD::VQSHRNs; break;
7759 case Intrinsic::arm_neon_vqshiftnu:
7760 VShiftOpc = ARMISD::VQSHRNu; break;
7761 case Intrinsic::arm_neon_vqshiftnsu:
7762 VShiftOpc = ARMISD::VQSHRNsu; break;
7763 case Intrinsic::arm_neon_vqrshiftns:
7764 VShiftOpc = ARMISD::VQRSHRNs; break;
7765 case Intrinsic::arm_neon_vqrshiftnu:
7766 VShiftOpc = ARMISD::VQRSHRNu; break;
7767 case Intrinsic::arm_neon_vqrshiftnsu:
7768 VShiftOpc = ARMISD::VQRSHRNsu; break;
7771 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
7772 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
7775 case Intrinsic::arm_neon_vshiftins: {
7776 EVT VT = N->getOperand(1).getValueType();
7778 unsigned VShiftOpc = 0;
7780 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
7781 VShiftOpc = ARMISD::VSLI;
7782 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
7783 VShiftOpc = ARMISD::VSRI;
7785 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
7788 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
7789 N->getOperand(1), N->getOperand(2),
7790 DAG.getConstant(Cnt, MVT::i32));
7793 case Intrinsic::arm_neon_vqrshifts:
7794 case Intrinsic::arm_neon_vqrshiftu:
7795 // No immediate versions of these to check for.
7802 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
7803 /// lowers them. As with the vector shift intrinsics, this is done during DAG
7804 /// combining instead of DAG legalizing because the build_vectors for 64-bit
7805 /// vector element shift counts are generally not legal, and it is hard to see
7806 /// their values after they get legalized to loads from a constant pool.
7807 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
7808 const ARMSubtarget *ST) {
7809 EVT VT = N->getValueType(0);
7811 // Nothing to be done for scalar shifts.
7812 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7813 if (!VT.isVector() || !TLI.isTypeLegal(VT))
7816 assert(ST->hasNEON() && "unexpected vector shift");
7819 switch (N->getOpcode()) {
7820 default: llvm_unreachable("unexpected shift opcode");
7823 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
7824 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
7825 DAG.getConstant(Cnt, MVT::i32));
7830 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
7831 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
7832 ARMISD::VSHRs : ARMISD::VSHRu);
7833 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
7834 DAG.getConstant(Cnt, MVT::i32));
7840 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
7841 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
7842 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
7843 const ARMSubtarget *ST) {
7844 SDValue N0 = N->getOperand(0);
7846 // Check for sign- and zero-extensions of vector extract operations of 8-
7847 // and 16-bit vector elements. NEON supports these directly. They are
7848 // handled during DAG combining because type legalization will promote them
7849 // to 32-bit types and it is messy to recognize the operations after that.
7850 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
7851 SDValue Vec = N0.getOperand(0);
7852 SDValue Lane = N0.getOperand(1);
7853 EVT VT = N->getValueType(0);
7854 EVT EltVT = N0.getValueType();
7855 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7857 if (VT == MVT::i32 &&
7858 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
7859 TLI.isTypeLegal(Vec.getValueType()) &&
7860 isa<ConstantSDNode>(Lane)) {
7863 switch (N->getOpcode()) {
7864 default: llvm_unreachable("unexpected opcode");
7865 case ISD::SIGN_EXTEND:
7866 Opc = ARMISD::VGETLANEs;
7868 case ISD::ZERO_EXTEND:
7869 case ISD::ANY_EXTEND:
7870 Opc = ARMISD::VGETLANEu;
7873 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
7880 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
7881 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
7882 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
7883 const ARMSubtarget *ST) {
7884 // If the target supports NEON, try to use vmax/vmin instructions for f32
7885 // selects like "x < y ? x : y". Unless the NoNaNsFPMath option is set,
7886 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
7887 // a NaN; only do the transformation when it matches that behavior.
7889 // For now only do this when using NEON for FP operations; if using VFP, it
7890 // is not obvious that the benefit outweighs the cost of switching to the
7892 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
7893 N->getValueType(0) != MVT::f32)
7896 SDValue CondLHS = N->getOperand(0);
7897 SDValue CondRHS = N->getOperand(1);
7898 SDValue LHS = N->getOperand(2);
7899 SDValue RHS = N->getOperand(3);
7900 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
7902 unsigned Opcode = 0;
7904 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
7905 IsReversed = false; // x CC y ? x : y
7906 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
7907 IsReversed = true ; // x CC y ? y : x
7921 // If LHS is NaN, an ordered comparison will be false and the result will
7922 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
7923 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
7924 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
7925 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
7927 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
7928 // will return -0, so vmin can only be used for unsafe math or if one of
7929 // the operands is known to be nonzero.
7930 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
7932 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
7934 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
7943 // If LHS is NaN, an ordered comparison will be false and the result will
7944 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
7945 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
7946 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
7947 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
7949 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
7950 // will return +0, so vmax can only be used for unsafe math or if one of
7951 // the operands is known to be nonzero.
7952 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
7954 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
7956 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
7962 return DAG.getNode(Opcode, N->getDebugLoc(), N->getValueType(0), LHS, RHS);
7965 /// PerformCMOVCombine - Target-specific DAG combining for ARMISD::CMOV.
7967 ARMTargetLowering::PerformCMOVCombine(SDNode *N, SelectionDAG &DAG) const {
7968 SDValue Cmp = N->getOperand(4);
7969 if (Cmp.getOpcode() != ARMISD::CMPZ)
7970 // Only looking at EQ and NE cases.
7973 EVT VT = N->getValueType(0);
7974 DebugLoc dl = N->getDebugLoc();
7975 SDValue LHS = Cmp.getOperand(0);
7976 SDValue RHS = Cmp.getOperand(1);
7977 SDValue FalseVal = N->getOperand(0);
7978 SDValue TrueVal = N->getOperand(1);
7979 SDValue ARMcc = N->getOperand(2);
7980 ARMCC::CondCodes CC =
7981 (ARMCC::CondCodes)cast<ConstantSDNode>(ARMcc)->getZExtValue();
7999 /// FIXME: Turn this into a target neutral optimization?
8001 if (CC == ARMCC::NE && FalseVal == RHS && FalseVal != LHS) {
8002 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, TrueVal, ARMcc,
8003 N->getOperand(3), Cmp);
8004 } else if (CC == ARMCC::EQ && TrueVal == RHS) {
8006 SDValue NewCmp = getARMCmp(LHS, RHS, ISD::SETNE, ARMcc, DAG, dl);
8007 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, FalseVal, ARMcc,
8008 N->getOperand(3), NewCmp);
8011 if (Res.getNode()) {
8012 APInt KnownZero, KnownOne;
8013 APInt Mask = APInt::getAllOnesValue(VT.getScalarType().getSizeInBits());
8014 DAG.ComputeMaskedBits(SDValue(N,0), Mask, KnownZero, KnownOne);
8015 // Capture demanded bits information that would be otherwise lost.
8016 if (KnownZero == 0xfffffffe)
8017 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
8018 DAG.getValueType(MVT::i1));
8019 else if (KnownZero == 0xffffff00)
8020 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
8021 DAG.getValueType(MVT::i8));
8022 else if (KnownZero == 0xffff0000)
8023 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
8024 DAG.getValueType(MVT::i16));
8030 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
8031 DAGCombinerInfo &DCI) const {
8032 switch (N->getOpcode()) {
8034 case ISD::ADD: return PerformADDCombine(N, DCI, Subtarget);
8035 case ISD::SUB: return PerformSUBCombine(N, DCI);
8036 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
8037 case ISD::OR: return PerformORCombine(N, DCI, Subtarget);
8038 case ISD::AND: return PerformANDCombine(N, DCI);
8039 case ARMISD::BFI: return PerformBFICombine(N, DCI);
8040 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
8041 case ARMISD::VMOVDRR: return PerformVMOVDRRCombine(N, DCI.DAG);
8042 case ISD::STORE: return PerformSTORECombine(N, DCI);
8043 case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DCI);
8044 case ISD::INSERT_VECTOR_ELT: return PerformInsertEltCombine(N, DCI);
8045 case ISD::VECTOR_SHUFFLE: return PerformVECTOR_SHUFFLECombine(N, DCI.DAG);
8046 case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI);
8047 case ISD::FP_TO_SINT:
8048 case ISD::FP_TO_UINT: return PerformVCVTCombine(N, DCI, Subtarget);
8049 case ISD::FDIV: return PerformVDIVCombine(N, DCI, Subtarget);
8050 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
8053 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
8054 case ISD::SIGN_EXTEND:
8055 case ISD::ZERO_EXTEND:
8056 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
8057 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
8058 case ARMISD::CMOV: return PerformCMOVCombine(N, DCI.DAG);
8059 case ARMISD::VLD2DUP:
8060 case ARMISD::VLD3DUP:
8061 case ARMISD::VLD4DUP:
8062 return CombineBaseUpdate(N, DCI);
8063 case ISD::INTRINSIC_VOID:
8064 case ISD::INTRINSIC_W_CHAIN:
8065 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
8066 case Intrinsic::arm_neon_vld1:
8067 case Intrinsic::arm_neon_vld2:
8068 case Intrinsic::arm_neon_vld3:
8069 case Intrinsic::arm_neon_vld4:
8070 case Intrinsic::arm_neon_vld2lane:
8071 case Intrinsic::arm_neon_vld3lane:
8072 case Intrinsic::arm_neon_vld4lane:
8073 case Intrinsic::arm_neon_vst1:
8074 case Intrinsic::arm_neon_vst2:
8075 case Intrinsic::arm_neon_vst3:
8076 case Intrinsic::arm_neon_vst4:
8077 case Intrinsic::arm_neon_vst2lane:
8078 case Intrinsic::arm_neon_vst3lane:
8079 case Intrinsic::arm_neon_vst4lane:
8080 return CombineBaseUpdate(N, DCI);
8088 bool ARMTargetLowering::isDesirableToTransformToIntegerOp(unsigned Opc,
8090 return (VT == MVT::f32) && (Opc == ISD::LOAD || Opc == ISD::STORE);
8093 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
8094 if (!Subtarget->allowsUnalignedMem())
8097 switch (VT.getSimpleVT().SimpleTy) {
8104 // FIXME: VLD1 etc with standard alignment is legal.
8108 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
8113 switch (VT.getSimpleVT().SimpleTy) {
8114 default: return false;
8129 if ((V & (Scale - 1)) != 0)
8132 return V == (V & ((1LL << 5) - 1));
8135 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
8136 const ARMSubtarget *Subtarget) {
8143 switch (VT.getSimpleVT().SimpleTy) {
8144 default: return false;
8149 // + imm12 or - imm8
8151 return V == (V & ((1LL << 8) - 1));
8152 return V == (V & ((1LL << 12) - 1));
8155 // Same as ARM mode. FIXME: NEON?
8156 if (!Subtarget->hasVFP2())
8161 return V == (V & ((1LL << 8) - 1));
8165 /// isLegalAddressImmediate - Return true if the integer value can be used
8166 /// as the offset of the target addressing mode for load / store of the
8168 static bool isLegalAddressImmediate(int64_t V, EVT VT,
8169 const ARMSubtarget *Subtarget) {
8176 if (Subtarget->isThumb1Only())
8177 return isLegalT1AddressImmediate(V, VT);
8178 else if (Subtarget->isThumb2())
8179 return isLegalT2AddressImmediate(V, VT, Subtarget);
8184 switch (VT.getSimpleVT().SimpleTy) {
8185 default: return false;
8190 return V == (V & ((1LL << 12) - 1));
8193 return V == (V & ((1LL << 8) - 1));
8196 if (!Subtarget->hasVFP2()) // FIXME: NEON?
8201 return V == (V & ((1LL << 8) - 1));
8205 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
8207 int Scale = AM.Scale;
8211 switch (VT.getSimpleVT().SimpleTy) {
8212 default: return false;
8221 return Scale == 2 || Scale == 4 || Scale == 8;
8224 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
8228 // Note, we allow "void" uses (basically, uses that aren't loads or
8229 // stores), because arm allows folding a scale into many arithmetic
8230 // operations. This should be made more precise and revisited later.
8232 // Allow r << imm, but the imm has to be a multiple of two.
8233 if (Scale & 1) return false;
8234 return isPowerOf2_32(Scale);
8238 /// isLegalAddressingMode - Return true if the addressing mode represented
8239 /// by AM is legal for this target, for a load/store of the specified type.
8240 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
8242 EVT VT = getValueType(Ty, true);
8243 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
8246 // Can never fold addr of global into load/store.
8251 case 0: // no scale reg, must be "r+i" or "r", or "i".
8254 if (Subtarget->isThumb1Only())
8258 // ARM doesn't support any R+R*scale+imm addr modes.
8265 if (Subtarget->isThumb2())
8266 return isLegalT2ScaledAddressingMode(AM, VT);
8268 int Scale = AM.Scale;
8269 switch (VT.getSimpleVT().SimpleTy) {
8270 default: return false;
8274 if (Scale < 0) Scale = -Scale;
8278 return isPowerOf2_32(Scale & ~1);
8282 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
8287 // Note, we allow "void" uses (basically, uses that aren't loads or
8288 // stores), because arm allows folding a scale into many arithmetic
8289 // operations. This should be made more precise and revisited later.
8291 // Allow r << imm, but the imm has to be a multiple of two.
8292 if (Scale & 1) return false;
8293 return isPowerOf2_32(Scale);
8300 /// isLegalICmpImmediate - Return true if the specified immediate is legal
8301 /// icmp immediate, that is the target has icmp instructions which can compare
8302 /// a register against the immediate without having to materialize the
8303 /// immediate into a register.
8304 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
8305 if (!Subtarget->isThumb())
8306 return ARM_AM::getSOImmVal(Imm) != -1;
8307 if (Subtarget->isThumb2())
8308 return ARM_AM::getT2SOImmVal(Imm) != -1;
8309 return Imm >= 0 && Imm <= 255;
8312 /// isLegalAddImmediate - Return true if the specified immediate is legal
8313 /// add immediate, that is the target has add instructions which can add
8314 /// a register with the immediate without having to materialize the
8315 /// immediate into a register.
8316 bool ARMTargetLowering::isLegalAddImmediate(int64_t Imm) const {
8317 return ARM_AM::getSOImmVal(Imm) != -1;
8320 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
8321 bool isSEXTLoad, SDValue &Base,
8322 SDValue &Offset, bool &isInc,
8323 SelectionDAG &DAG) {
8324 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
8327 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
8329 Base = Ptr->getOperand(0);
8330 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
8331 int RHSC = (int)RHS->getZExtValue();
8332 if (RHSC < 0 && RHSC > -256) {
8333 assert(Ptr->getOpcode() == ISD::ADD);
8335 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
8339 isInc = (Ptr->getOpcode() == ISD::ADD);
8340 Offset = Ptr->getOperand(1);
8342 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
8344 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
8345 int RHSC = (int)RHS->getZExtValue();
8346 if (RHSC < 0 && RHSC > -0x1000) {
8347 assert(Ptr->getOpcode() == ISD::ADD);
8349 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
8350 Base = Ptr->getOperand(0);
8355 if (Ptr->getOpcode() == ISD::ADD) {
8357 ARM_AM::ShiftOpc ShOpcVal=
8358 ARM_AM::getShiftOpcForNode(Ptr->getOperand(0).getOpcode());
8359 if (ShOpcVal != ARM_AM::no_shift) {
8360 Base = Ptr->getOperand(1);
8361 Offset = Ptr->getOperand(0);
8363 Base = Ptr->getOperand(0);
8364 Offset = Ptr->getOperand(1);
8369 isInc = (Ptr->getOpcode() == ISD::ADD);
8370 Base = Ptr->getOperand(0);
8371 Offset = Ptr->getOperand(1);
8375 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
8379 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
8380 bool isSEXTLoad, SDValue &Base,
8381 SDValue &Offset, bool &isInc,
8382 SelectionDAG &DAG) {
8383 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
8386 Base = Ptr->getOperand(0);
8387 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
8388 int RHSC = (int)RHS->getZExtValue();
8389 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
8390 assert(Ptr->getOpcode() == ISD::ADD);
8392 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
8394 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
8395 isInc = Ptr->getOpcode() == ISD::ADD;
8396 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
8404 /// getPreIndexedAddressParts - returns true by value, base pointer and
8405 /// offset pointer and addressing mode by reference if the node's address
8406 /// can be legally represented as pre-indexed load / store address.
8408 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
8410 ISD::MemIndexedMode &AM,
8411 SelectionDAG &DAG) const {
8412 if (Subtarget->isThumb1Only())
8417 bool isSEXTLoad = false;
8418 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
8419 Ptr = LD->getBasePtr();
8420 VT = LD->getMemoryVT();
8421 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
8422 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
8423 Ptr = ST->getBasePtr();
8424 VT = ST->getMemoryVT();
8429 bool isLegal = false;
8430 if (Subtarget->isThumb2())
8431 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
8432 Offset, isInc, DAG);
8434 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
8435 Offset, isInc, DAG);
8439 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
8443 /// getPostIndexedAddressParts - returns true by value, base pointer and
8444 /// offset pointer and addressing mode by reference if this node can be
8445 /// combined with a load / store to form a post-indexed load / store.
8446 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
8449 ISD::MemIndexedMode &AM,
8450 SelectionDAG &DAG) const {
8451 if (Subtarget->isThumb1Only())
8456 bool isSEXTLoad = false;
8457 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
8458 VT = LD->getMemoryVT();
8459 Ptr = LD->getBasePtr();
8460 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
8461 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
8462 VT = ST->getMemoryVT();
8463 Ptr = ST->getBasePtr();
8468 bool isLegal = false;
8469 if (Subtarget->isThumb2())
8470 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
8473 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
8479 // Swap base ptr and offset to catch more post-index load / store when
8480 // it's legal. In Thumb2 mode, offset must be an immediate.
8481 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
8482 !Subtarget->isThumb2())
8483 std::swap(Base, Offset);
8485 // Post-indexed load / store update the base pointer.
8490 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
8494 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
8498 const SelectionDAG &DAG,
8499 unsigned Depth) const {
8500 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
8501 switch (Op.getOpcode()) {
8503 case ARMISD::CMOV: {
8504 // Bits are known zero/one if known on the LHS and RHS.
8505 DAG.ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
8506 if (KnownZero == 0 && KnownOne == 0) return;
8508 APInt KnownZeroRHS, KnownOneRHS;
8509 DAG.ComputeMaskedBits(Op.getOperand(1), Mask,
8510 KnownZeroRHS, KnownOneRHS, Depth+1);
8511 KnownZero &= KnownZeroRHS;
8512 KnownOne &= KnownOneRHS;
8518 //===----------------------------------------------------------------------===//
8519 // ARM Inline Assembly Support
8520 //===----------------------------------------------------------------------===//
8522 bool ARMTargetLowering::ExpandInlineAsm(CallInst *CI) const {
8523 // Looking for "rev" which is V6+.
8524 if (!Subtarget->hasV6Ops())
8527 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
8528 std::string AsmStr = IA->getAsmString();
8529 SmallVector<StringRef, 4> AsmPieces;
8530 SplitString(AsmStr, AsmPieces, ";\n");
8532 switch (AsmPieces.size()) {
8533 default: return false;
8535 AsmStr = AsmPieces[0];
8537 SplitString(AsmStr, AsmPieces, " \t,");
8540 if (AsmPieces.size() == 3 &&
8541 AsmPieces[0] == "rev" && AsmPieces[1] == "$0" && AsmPieces[2] == "$1" &&
8542 IA->getConstraintString().compare(0, 4, "=l,l") == 0) {
8543 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
8544 if (Ty && Ty->getBitWidth() == 32)
8545 return IntrinsicLowering::LowerToByteSwap(CI);
8553 /// getConstraintType - Given a constraint letter, return the type of
8554 /// constraint it is for this target.
8555 ARMTargetLowering::ConstraintType
8556 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
8557 if (Constraint.size() == 1) {
8558 switch (Constraint[0]) {
8560 case 'l': return C_RegisterClass;
8561 case 'w': return C_RegisterClass;
8562 case 'h': return C_RegisterClass;
8563 case 'x': return C_RegisterClass;
8564 case 't': return C_RegisterClass;
8565 case 'j': return C_Other; // Constant for movw.
8566 // An address with a single base register. Due to the way we
8567 // currently handle addresses it is the same as an 'r' memory constraint.
8568 case 'Q': return C_Memory;
8570 } else if (Constraint.size() == 2) {
8571 switch (Constraint[0]) {
8573 // All 'U+' constraints are addresses.
8574 case 'U': return C_Memory;
8577 return TargetLowering::getConstraintType(Constraint);
8580 /// Examine constraint type and operand type and determine a weight value.
8581 /// This object must already have been set up with the operand type
8582 /// and the current alternative constraint selected.
8583 TargetLowering::ConstraintWeight
8584 ARMTargetLowering::getSingleConstraintMatchWeight(
8585 AsmOperandInfo &info, const char *constraint) const {
8586 ConstraintWeight weight = CW_Invalid;
8587 Value *CallOperandVal = info.CallOperandVal;
8588 // If we don't have a value, we can't do a match,
8589 // but allow it at the lowest weight.
8590 if (CallOperandVal == NULL)
8592 Type *type = CallOperandVal->getType();
8593 // Look at the constraint type.
8594 switch (*constraint) {
8596 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
8599 if (type->isIntegerTy()) {
8600 if (Subtarget->isThumb())
8601 weight = CW_SpecificReg;
8603 weight = CW_Register;
8607 if (type->isFloatingPointTy())
8608 weight = CW_Register;
8614 typedef std::pair<unsigned, const TargetRegisterClass*> RCPair;
8616 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
8618 if (Constraint.size() == 1) {
8619 // GCC ARM Constraint Letters
8620 switch (Constraint[0]) {
8621 case 'l': // Low regs or general regs.
8622 if (Subtarget->isThumb())
8623 return RCPair(0U, ARM::tGPRRegisterClass);
8625 return RCPair(0U, ARM::GPRRegisterClass);
8626 case 'h': // High regs or no regs.
8627 if (Subtarget->isThumb())
8628 return RCPair(0U, ARM::hGPRRegisterClass);
8631 return RCPair(0U, ARM::GPRRegisterClass);
8634 return RCPair(0U, ARM::SPRRegisterClass);
8635 if (VT.getSizeInBits() == 64)
8636 return RCPair(0U, ARM::DPRRegisterClass);
8637 if (VT.getSizeInBits() == 128)
8638 return RCPair(0U, ARM::QPRRegisterClass);
8642 return RCPair(0U, ARM::SPR_8RegisterClass);
8643 if (VT.getSizeInBits() == 64)
8644 return RCPair(0U, ARM::DPR_8RegisterClass);
8645 if (VT.getSizeInBits() == 128)
8646 return RCPair(0U, ARM::QPR_8RegisterClass);
8650 return RCPair(0U, ARM::SPRRegisterClass);
8654 if (StringRef("{cc}").equals_lower(Constraint))
8655 return std::make_pair(unsigned(ARM::CPSR), ARM::CCRRegisterClass);
8657 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
8660 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
8661 /// vector. If it is invalid, don't add anything to Ops.
8662 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
8663 std::string &Constraint,
8664 std::vector<SDValue>&Ops,
8665 SelectionDAG &DAG) const {
8666 SDValue Result(0, 0);
8668 // Currently only support length 1 constraints.
8669 if (Constraint.length() != 1) return;
8671 char ConstraintLetter = Constraint[0];
8672 switch (ConstraintLetter) {
8675 case 'I': case 'J': case 'K': case 'L':
8676 case 'M': case 'N': case 'O':
8677 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
8681 int64_t CVal64 = C->getSExtValue();
8682 int CVal = (int) CVal64;
8683 // None of these constraints allow values larger than 32 bits. Check
8684 // that the value fits in an int.
8688 switch (ConstraintLetter) {
8690 // Constant suitable for movw, must be between 0 and
8692 if (Subtarget->hasV6T2Ops())
8693 if (CVal >= 0 && CVal <= 65535)
8697 if (Subtarget->isThumb1Only()) {
8698 // This must be a constant between 0 and 255, for ADD
8700 if (CVal >= 0 && CVal <= 255)
8702 } else if (Subtarget->isThumb2()) {
8703 // A constant that can be used as an immediate value in a
8704 // data-processing instruction.
8705 if (ARM_AM::getT2SOImmVal(CVal) != -1)
8708 // A constant that can be used as an immediate value in a
8709 // data-processing instruction.
8710 if (ARM_AM::getSOImmVal(CVal) != -1)
8716 if (Subtarget->isThumb()) { // FIXME thumb2
8717 // This must be a constant between -255 and -1, for negated ADD
8718 // immediates. This can be used in GCC with an "n" modifier that
8719 // prints the negated value, for use with SUB instructions. It is
8720 // not useful otherwise but is implemented for compatibility.
8721 if (CVal >= -255 && CVal <= -1)
8724 // This must be a constant between -4095 and 4095. It is not clear
8725 // what this constraint is intended for. Implemented for
8726 // compatibility with GCC.
8727 if (CVal >= -4095 && CVal <= 4095)
8733 if (Subtarget->isThumb1Only()) {
8734 // A 32-bit value where only one byte has a nonzero value. Exclude
8735 // zero to match GCC. This constraint is used by GCC internally for
8736 // constants that can be loaded with a move/shift combination.
8737 // It is not useful otherwise but is implemented for compatibility.
8738 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
8740 } else if (Subtarget->isThumb2()) {
8741 // A constant whose bitwise inverse can be used as an immediate
8742 // value in a data-processing instruction. This can be used in GCC
8743 // with a "B" modifier that prints the inverted value, for use with
8744 // BIC and MVN instructions. It is not useful otherwise but is
8745 // implemented for compatibility.
8746 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
8749 // A constant whose bitwise inverse can be used as an immediate
8750 // value in a data-processing instruction. This can be used in GCC
8751 // with a "B" modifier that prints the inverted value, for use with
8752 // BIC and MVN instructions. It is not useful otherwise but is
8753 // implemented for compatibility.
8754 if (ARM_AM::getSOImmVal(~CVal) != -1)
8760 if (Subtarget->isThumb1Only()) {
8761 // This must be a constant between -7 and 7,
8762 // for 3-operand ADD/SUB immediate instructions.
8763 if (CVal >= -7 && CVal < 7)
8765 } else if (Subtarget->isThumb2()) {
8766 // A constant whose negation can be used as an immediate value in a
8767 // data-processing instruction. This can be used in GCC with an "n"
8768 // modifier that prints the negated value, for use with SUB
8769 // instructions. It is not useful otherwise but is implemented for
8771 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
8774 // A constant whose negation can be used as an immediate value in a
8775 // data-processing instruction. This can be used in GCC with an "n"
8776 // modifier that prints the negated value, for use with SUB
8777 // instructions. It is not useful otherwise but is implemented for
8779 if (ARM_AM::getSOImmVal(-CVal) != -1)
8785 if (Subtarget->isThumb()) { // FIXME thumb2
8786 // This must be a multiple of 4 between 0 and 1020, for
8787 // ADD sp + immediate.
8788 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
8791 // A power of two or a constant between 0 and 32. This is used in
8792 // GCC for the shift amount on shifted register operands, but it is
8793 // useful in general for any shift amounts.
8794 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
8800 if (Subtarget->isThumb()) { // FIXME thumb2
8801 // This must be a constant between 0 and 31, for shift amounts.
8802 if (CVal >= 0 && CVal <= 31)
8808 if (Subtarget->isThumb()) { // FIXME thumb2
8809 // This must be a multiple of 4 between -508 and 508, for
8810 // ADD/SUB sp = sp + immediate.
8811 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
8816 Result = DAG.getTargetConstant(CVal, Op.getValueType());
8820 if (Result.getNode()) {
8821 Ops.push_back(Result);
8824 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
8828 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
8829 // The ARM target isn't yet aware of offsets.
8833 bool ARM::isBitFieldInvertedMask(unsigned v) {
8834 if (v == 0xffffffff)
8836 // there can be 1's on either or both "outsides", all the "inside"
8838 unsigned int lsb = 0, msb = 31;
8839 while (v & (1 << msb)) --msb;
8840 while (v & (1 << lsb)) ++lsb;
8841 for (unsigned int i = lsb; i <= msb; ++i) {
8848 /// isFPImmLegal - Returns true if the target can instruction select the
8849 /// specified FP immediate natively. If false, the legalizer will
8850 /// materialize the FP immediate as a load from a constant pool.
8851 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
8852 if (!Subtarget->hasVFP3())
8855 return ARM_AM::getFP32Imm(Imm) != -1;
8857 return ARM_AM::getFP64Imm(Imm) != -1;
8861 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
8862 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
8863 /// specified in the intrinsic calls.
8864 bool ARMTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
8866 unsigned Intrinsic) const {
8867 switch (Intrinsic) {
8868 case Intrinsic::arm_neon_vld1:
8869 case Intrinsic::arm_neon_vld2:
8870 case Intrinsic::arm_neon_vld3:
8871 case Intrinsic::arm_neon_vld4:
8872 case Intrinsic::arm_neon_vld2lane:
8873 case Intrinsic::arm_neon_vld3lane:
8874 case Intrinsic::arm_neon_vld4lane: {
8875 Info.opc = ISD::INTRINSIC_W_CHAIN;
8876 // Conservatively set memVT to the entire set of vectors loaded.
8877 uint64_t NumElts = getTargetData()->getTypeAllocSize(I.getType()) / 8;
8878 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
8879 Info.ptrVal = I.getArgOperand(0);
8881 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
8882 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
8883 Info.vol = false; // volatile loads with NEON intrinsics not supported
8884 Info.readMem = true;
8885 Info.writeMem = false;
8888 case Intrinsic::arm_neon_vst1:
8889 case Intrinsic::arm_neon_vst2:
8890 case Intrinsic::arm_neon_vst3:
8891 case Intrinsic::arm_neon_vst4:
8892 case Intrinsic::arm_neon_vst2lane:
8893 case Intrinsic::arm_neon_vst3lane:
8894 case Intrinsic::arm_neon_vst4lane: {
8895 Info.opc = ISD::INTRINSIC_VOID;
8896 // Conservatively set memVT to the entire set of vectors stored.
8897 unsigned NumElts = 0;
8898 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
8899 Type *ArgTy = I.getArgOperand(ArgI)->getType();
8900 if (!ArgTy->isVectorTy())
8902 NumElts += getTargetData()->getTypeAllocSize(ArgTy) / 8;
8904 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
8905 Info.ptrVal = I.getArgOperand(0);
8907 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
8908 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
8909 Info.vol = false; // volatile stores with NEON intrinsics not supported
8910 Info.readMem = false;
8911 Info.writeMem = true;
8914 case Intrinsic::arm_strexd: {
8915 Info.opc = ISD::INTRINSIC_W_CHAIN;
8916 Info.memVT = MVT::i64;
8917 Info.ptrVal = I.getArgOperand(2);
8921 Info.readMem = false;
8922 Info.writeMem = true;
8925 case Intrinsic::arm_ldrexd: {
8926 Info.opc = ISD::INTRINSIC_W_CHAIN;
8927 Info.memVT = MVT::i64;
8928 Info.ptrVal = I.getArgOperand(0);
8932 Info.readMem = true;
8933 Info.writeMem = false;