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 #include "ARMISelLowering.h"
16 #include "ARMCallingConv.h"
17 #include "ARMConstantPoolValue.h"
18 #include "ARMMachineFunctionInfo.h"
19 #include "ARMPerfectShuffle.h"
20 #include "ARMSubtarget.h"
21 #include "ARMTargetMachine.h"
22 #include "ARMTargetObjectFile.h"
23 #include "MCTargetDesc/ARMAddressingModes.h"
24 #include "llvm/ADT/Statistic.h"
25 #include "llvm/ADT/StringExtras.h"
26 #include "llvm/CodeGen/CallingConvLower.h"
27 #include "llvm/CodeGen/IntrinsicLowering.h"
28 #include "llvm/CodeGen/MachineBasicBlock.h"
29 #include "llvm/CodeGen/MachineFrameInfo.h"
30 #include "llvm/CodeGen/MachineFunction.h"
31 #include "llvm/CodeGen/MachineInstrBuilder.h"
32 #include "llvm/CodeGen/MachineJumpTableInfo.h"
33 #include "llvm/CodeGen/MachineModuleInfo.h"
34 #include "llvm/CodeGen/MachineRegisterInfo.h"
35 #include "llvm/CodeGen/SelectionDAG.h"
36 #include "llvm/IR/CallingConv.h"
37 #include "llvm/IR/Constants.h"
38 #include "llvm/IR/Function.h"
39 #include "llvm/IR/GlobalValue.h"
40 #include "llvm/IR/IRBuilder.h"
41 #include "llvm/IR/Instruction.h"
42 #include "llvm/IR/Instructions.h"
43 #include "llvm/IR/Intrinsics.h"
44 #include "llvm/IR/Type.h"
45 #include "llvm/MC/MCSectionMachO.h"
46 #include "llvm/Support/CommandLine.h"
47 #include "llvm/Support/Debug.h"
48 #include "llvm/Support/ErrorHandling.h"
49 #include "llvm/Support/MathExtras.h"
50 #include "llvm/Target/TargetOptions.h"
54 #define DEBUG_TYPE "arm-isel"
56 STATISTIC(NumTailCalls, "Number of tail calls");
57 STATISTIC(NumMovwMovt, "Number of GAs materialized with movw + movt");
58 STATISTIC(NumLoopByVals, "Number of loops generated for byval arguments");
61 EnableARMLongCalls("arm-long-calls", cl::Hidden,
62 cl::desc("Generate calls via indirect call instructions"),
66 ARMInterworking("arm-interworking", cl::Hidden,
67 cl::desc("Enable / disable ARM interworking (for debugging only)"),
71 class ARMCCState : public CCState {
73 ARMCCState(CallingConv::ID CC, bool isVarArg, MachineFunction &MF,
74 SmallVectorImpl<CCValAssign> &locs, LLVMContext &C,
76 : CCState(CC, isVarArg, MF, locs, C) {
77 assert(((PC == Call) || (PC == Prologue)) &&
78 "ARMCCState users must specify whether their context is call"
79 "or prologue generation.");
85 // The APCS parameter registers.
86 static const MCPhysReg GPRArgRegs[] = {
87 ARM::R0, ARM::R1, ARM::R2, ARM::R3
90 void ARMTargetLowering::addTypeForNEON(MVT VT, MVT PromotedLdStVT,
91 MVT PromotedBitwiseVT) {
92 if (VT != PromotedLdStVT) {
93 setOperationAction(ISD::LOAD, VT, Promote);
94 AddPromotedToType (ISD::LOAD, VT, PromotedLdStVT);
96 setOperationAction(ISD::STORE, VT, Promote);
97 AddPromotedToType (ISD::STORE, VT, PromotedLdStVT);
100 MVT ElemTy = VT.getVectorElementType();
101 if (ElemTy != MVT::i64 && ElemTy != MVT::f64)
102 setOperationAction(ISD::SETCC, VT, Custom);
103 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
104 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
105 if (ElemTy == MVT::i32) {
106 setOperationAction(ISD::SINT_TO_FP, VT, Custom);
107 setOperationAction(ISD::UINT_TO_FP, VT, Custom);
108 setOperationAction(ISD::FP_TO_SINT, VT, Custom);
109 setOperationAction(ISD::FP_TO_UINT, VT, Custom);
111 setOperationAction(ISD::SINT_TO_FP, VT, Expand);
112 setOperationAction(ISD::UINT_TO_FP, VT, Expand);
113 setOperationAction(ISD::FP_TO_SINT, VT, Expand);
114 setOperationAction(ISD::FP_TO_UINT, VT, Expand);
116 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
117 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
118 setOperationAction(ISD::CONCAT_VECTORS, VT, Legal);
119 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Legal);
120 setOperationAction(ISD::SELECT, VT, Expand);
121 setOperationAction(ISD::SELECT_CC, VT, Expand);
122 setOperationAction(ISD::VSELECT, VT, Expand);
123 setOperationAction(ISD::SIGN_EXTEND_INREG, VT, Expand);
124 if (VT.isInteger()) {
125 setOperationAction(ISD::SHL, VT, Custom);
126 setOperationAction(ISD::SRA, VT, Custom);
127 setOperationAction(ISD::SRL, VT, Custom);
130 // Promote all bit-wise operations.
131 if (VT.isInteger() && VT != PromotedBitwiseVT) {
132 setOperationAction(ISD::AND, VT, Promote);
133 AddPromotedToType (ISD::AND, VT, PromotedBitwiseVT);
134 setOperationAction(ISD::OR, VT, Promote);
135 AddPromotedToType (ISD::OR, VT, PromotedBitwiseVT);
136 setOperationAction(ISD::XOR, VT, Promote);
137 AddPromotedToType (ISD::XOR, VT, PromotedBitwiseVT);
140 // Neon does not support vector divide/remainder operations.
141 setOperationAction(ISD::SDIV, VT, Expand);
142 setOperationAction(ISD::UDIV, VT, Expand);
143 setOperationAction(ISD::FDIV, VT, Expand);
144 setOperationAction(ISD::SREM, VT, Expand);
145 setOperationAction(ISD::UREM, VT, Expand);
146 setOperationAction(ISD::FREM, VT, Expand);
149 void ARMTargetLowering::addDRTypeForNEON(MVT VT) {
150 addRegisterClass(VT, &ARM::DPRRegClass);
151 addTypeForNEON(VT, MVT::f64, MVT::v2i32);
154 void ARMTargetLowering::addQRTypeForNEON(MVT VT) {
155 addRegisterClass(VT, &ARM::DPairRegClass);
156 addTypeForNEON(VT, MVT::v2f64, MVT::v4i32);
159 ARMTargetLowering::ARMTargetLowering(const TargetMachine &TM,
160 const ARMSubtarget &STI)
161 : TargetLowering(TM), Subtarget(&STI) {
162 RegInfo = Subtarget->getRegisterInfo();
163 Itins = Subtarget->getInstrItineraryData();
165 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
167 if (Subtarget->isTargetMachO()) {
168 // Uses VFP for Thumb libfuncs if available.
169 if (Subtarget->isThumb() && Subtarget->hasVFP2() &&
170 Subtarget->hasARMOps() && !TM.Options.UseSoftFloat) {
171 // Single-precision floating-point arithmetic.
172 setLibcallName(RTLIB::ADD_F32, "__addsf3vfp");
173 setLibcallName(RTLIB::SUB_F32, "__subsf3vfp");
174 setLibcallName(RTLIB::MUL_F32, "__mulsf3vfp");
175 setLibcallName(RTLIB::DIV_F32, "__divsf3vfp");
177 // Double-precision floating-point arithmetic.
178 setLibcallName(RTLIB::ADD_F64, "__adddf3vfp");
179 setLibcallName(RTLIB::SUB_F64, "__subdf3vfp");
180 setLibcallName(RTLIB::MUL_F64, "__muldf3vfp");
181 setLibcallName(RTLIB::DIV_F64, "__divdf3vfp");
183 // Single-precision comparisons.
184 setLibcallName(RTLIB::OEQ_F32, "__eqsf2vfp");
185 setLibcallName(RTLIB::UNE_F32, "__nesf2vfp");
186 setLibcallName(RTLIB::OLT_F32, "__ltsf2vfp");
187 setLibcallName(RTLIB::OLE_F32, "__lesf2vfp");
188 setLibcallName(RTLIB::OGE_F32, "__gesf2vfp");
189 setLibcallName(RTLIB::OGT_F32, "__gtsf2vfp");
190 setLibcallName(RTLIB::UO_F32, "__unordsf2vfp");
191 setLibcallName(RTLIB::O_F32, "__unordsf2vfp");
193 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
194 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETNE);
195 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
196 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
197 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
198 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
199 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
200 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
202 // Double-precision comparisons.
203 setLibcallName(RTLIB::OEQ_F64, "__eqdf2vfp");
204 setLibcallName(RTLIB::UNE_F64, "__nedf2vfp");
205 setLibcallName(RTLIB::OLT_F64, "__ltdf2vfp");
206 setLibcallName(RTLIB::OLE_F64, "__ledf2vfp");
207 setLibcallName(RTLIB::OGE_F64, "__gedf2vfp");
208 setLibcallName(RTLIB::OGT_F64, "__gtdf2vfp");
209 setLibcallName(RTLIB::UO_F64, "__unorddf2vfp");
210 setLibcallName(RTLIB::O_F64, "__unorddf2vfp");
212 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
213 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETNE);
214 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
215 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
216 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
217 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
218 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
219 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
221 // Floating-point to integer conversions.
222 // i64 conversions are done via library routines even when generating VFP
223 // instructions, so use the same ones.
224 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__fixdfsivfp");
225 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__fixunsdfsivfp");
226 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__fixsfsivfp");
227 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__fixunssfsivfp");
229 // Conversions between floating types.
230 setLibcallName(RTLIB::FPROUND_F64_F32, "__truncdfsf2vfp");
231 setLibcallName(RTLIB::FPEXT_F32_F64, "__extendsfdf2vfp");
233 // Integer to floating-point conversions.
234 // i64 conversions are done via library routines even when generating VFP
235 // instructions, so use the same ones.
236 // FIXME: There appears to be some naming inconsistency in ARM libgcc:
237 // e.g., __floatunsidf vs. __floatunssidfvfp.
238 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__floatsidfvfp");
239 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__floatunssidfvfp");
240 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__floatsisfvfp");
241 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__floatunssisfvfp");
245 // These libcalls are not available in 32-bit.
246 setLibcallName(RTLIB::SHL_I128, nullptr);
247 setLibcallName(RTLIB::SRL_I128, nullptr);
248 setLibcallName(RTLIB::SRA_I128, nullptr);
250 if (Subtarget->isAAPCS_ABI() && !Subtarget->isTargetMachO() &&
251 !Subtarget->isTargetWindows()) {
252 static const struct {
253 const RTLIB::Libcall Op;
254 const char * const Name;
255 const CallingConv::ID CC;
256 const ISD::CondCode Cond;
258 // Double-precision floating-point arithmetic helper functions
259 // RTABI chapter 4.1.2, Table 2
260 { RTLIB::ADD_F64, "__aeabi_dadd", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
261 { RTLIB::DIV_F64, "__aeabi_ddiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
262 { RTLIB::MUL_F64, "__aeabi_dmul", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
263 { RTLIB::SUB_F64, "__aeabi_dsub", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
265 // Double-precision floating-point comparison helper functions
266 // RTABI chapter 4.1.2, Table 3
267 { RTLIB::OEQ_F64, "__aeabi_dcmpeq", CallingConv::ARM_AAPCS, ISD::SETNE },
268 { RTLIB::UNE_F64, "__aeabi_dcmpeq", CallingConv::ARM_AAPCS, ISD::SETEQ },
269 { RTLIB::OLT_F64, "__aeabi_dcmplt", CallingConv::ARM_AAPCS, ISD::SETNE },
270 { RTLIB::OLE_F64, "__aeabi_dcmple", CallingConv::ARM_AAPCS, ISD::SETNE },
271 { RTLIB::OGE_F64, "__aeabi_dcmpge", CallingConv::ARM_AAPCS, ISD::SETNE },
272 { RTLIB::OGT_F64, "__aeabi_dcmpgt", CallingConv::ARM_AAPCS, ISD::SETNE },
273 { RTLIB::UO_F64, "__aeabi_dcmpun", CallingConv::ARM_AAPCS, ISD::SETNE },
274 { RTLIB::O_F64, "__aeabi_dcmpun", CallingConv::ARM_AAPCS, ISD::SETEQ },
276 // Single-precision floating-point arithmetic helper functions
277 // RTABI chapter 4.1.2, Table 4
278 { RTLIB::ADD_F32, "__aeabi_fadd", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
279 { RTLIB::DIV_F32, "__aeabi_fdiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
280 { RTLIB::MUL_F32, "__aeabi_fmul", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
281 { RTLIB::SUB_F32, "__aeabi_fsub", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
283 // Single-precision floating-point comparison helper functions
284 // RTABI chapter 4.1.2, Table 5
285 { RTLIB::OEQ_F32, "__aeabi_fcmpeq", CallingConv::ARM_AAPCS, ISD::SETNE },
286 { RTLIB::UNE_F32, "__aeabi_fcmpeq", CallingConv::ARM_AAPCS, ISD::SETEQ },
287 { RTLIB::OLT_F32, "__aeabi_fcmplt", CallingConv::ARM_AAPCS, ISD::SETNE },
288 { RTLIB::OLE_F32, "__aeabi_fcmple", CallingConv::ARM_AAPCS, ISD::SETNE },
289 { RTLIB::OGE_F32, "__aeabi_fcmpge", CallingConv::ARM_AAPCS, ISD::SETNE },
290 { RTLIB::OGT_F32, "__aeabi_fcmpgt", CallingConv::ARM_AAPCS, ISD::SETNE },
291 { RTLIB::UO_F32, "__aeabi_fcmpun", CallingConv::ARM_AAPCS, ISD::SETNE },
292 { RTLIB::O_F32, "__aeabi_fcmpun", CallingConv::ARM_AAPCS, ISD::SETEQ },
294 // Floating-point to integer conversions.
295 // RTABI chapter 4.1.2, Table 6
296 { RTLIB::FPTOSINT_F64_I32, "__aeabi_d2iz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
297 { RTLIB::FPTOUINT_F64_I32, "__aeabi_d2uiz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
298 { RTLIB::FPTOSINT_F64_I64, "__aeabi_d2lz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
299 { RTLIB::FPTOUINT_F64_I64, "__aeabi_d2ulz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
300 { RTLIB::FPTOSINT_F32_I32, "__aeabi_f2iz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
301 { RTLIB::FPTOUINT_F32_I32, "__aeabi_f2uiz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
302 { RTLIB::FPTOSINT_F32_I64, "__aeabi_f2lz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
303 { RTLIB::FPTOUINT_F32_I64, "__aeabi_f2ulz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
305 // Conversions between floating types.
306 // RTABI chapter 4.1.2, Table 7
307 { RTLIB::FPROUND_F64_F32, "__aeabi_d2f", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
308 { RTLIB::FPROUND_F64_F16, "__aeabi_d2h", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
309 { RTLIB::FPEXT_F32_F64, "__aeabi_f2d", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
311 // Integer to floating-point conversions.
312 // RTABI chapter 4.1.2, Table 8
313 { RTLIB::SINTTOFP_I32_F64, "__aeabi_i2d", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
314 { RTLIB::UINTTOFP_I32_F64, "__aeabi_ui2d", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
315 { RTLIB::SINTTOFP_I64_F64, "__aeabi_l2d", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
316 { RTLIB::UINTTOFP_I64_F64, "__aeabi_ul2d", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
317 { RTLIB::SINTTOFP_I32_F32, "__aeabi_i2f", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
318 { RTLIB::UINTTOFP_I32_F32, "__aeabi_ui2f", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
319 { RTLIB::SINTTOFP_I64_F32, "__aeabi_l2f", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
320 { RTLIB::UINTTOFP_I64_F32, "__aeabi_ul2f", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
322 // Long long helper functions
323 // RTABI chapter 4.2, Table 9
324 { RTLIB::MUL_I64, "__aeabi_lmul", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
325 { RTLIB::SHL_I64, "__aeabi_llsl", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
326 { RTLIB::SRL_I64, "__aeabi_llsr", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
327 { RTLIB::SRA_I64, "__aeabi_lasr", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
329 // Integer division functions
330 // RTABI chapter 4.3.1
331 { RTLIB::SDIV_I8, "__aeabi_idiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
332 { RTLIB::SDIV_I16, "__aeabi_idiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
333 { RTLIB::SDIV_I32, "__aeabi_idiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
334 { RTLIB::SDIV_I64, "__aeabi_ldivmod", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
335 { RTLIB::UDIV_I8, "__aeabi_uidiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
336 { RTLIB::UDIV_I16, "__aeabi_uidiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
337 { RTLIB::UDIV_I32, "__aeabi_uidiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
338 { RTLIB::UDIV_I64, "__aeabi_uldivmod", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
341 // RTABI chapter 4.3.4
342 { RTLIB::MEMCPY, "__aeabi_memcpy", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
343 { RTLIB::MEMMOVE, "__aeabi_memmove", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
344 { RTLIB::MEMSET, "__aeabi_memset", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
347 for (const auto &LC : LibraryCalls) {
348 setLibcallName(LC.Op, LC.Name);
349 setLibcallCallingConv(LC.Op, LC.CC);
350 if (LC.Cond != ISD::SETCC_INVALID)
351 setCmpLibcallCC(LC.Op, LC.Cond);
355 if (Subtarget->isTargetWindows()) {
356 static const struct {
357 const RTLIB::Libcall Op;
358 const char * const Name;
359 const CallingConv::ID CC;
361 { RTLIB::FPTOSINT_F32_I64, "__stoi64", CallingConv::ARM_AAPCS_VFP },
362 { RTLIB::FPTOSINT_F64_I64, "__dtoi64", CallingConv::ARM_AAPCS_VFP },
363 { RTLIB::FPTOUINT_F32_I64, "__stou64", CallingConv::ARM_AAPCS_VFP },
364 { RTLIB::FPTOUINT_F64_I64, "__dtou64", CallingConv::ARM_AAPCS_VFP },
365 { RTLIB::SINTTOFP_I64_F32, "__i64tos", CallingConv::ARM_AAPCS_VFP },
366 { RTLIB::SINTTOFP_I64_F64, "__i64tod", CallingConv::ARM_AAPCS_VFP },
367 { RTLIB::UINTTOFP_I64_F32, "__u64tos", CallingConv::ARM_AAPCS_VFP },
368 { RTLIB::UINTTOFP_I64_F64, "__u64tod", CallingConv::ARM_AAPCS_VFP },
371 for (const auto &LC : LibraryCalls) {
372 setLibcallName(LC.Op, LC.Name);
373 setLibcallCallingConv(LC.Op, LC.CC);
377 // Use divmod compiler-rt calls for iOS 5.0 and later.
378 if (Subtarget->getTargetTriple().isiOS() &&
379 !Subtarget->getTargetTriple().isOSVersionLT(5, 0)) {
380 setLibcallName(RTLIB::SDIVREM_I32, "__divmodsi4");
381 setLibcallName(RTLIB::UDIVREM_I32, "__udivmodsi4");
384 // The half <-> float conversion functions are always soft-float, but are
385 // needed for some targets which use a hard-float calling convention by
387 if (Subtarget->isAAPCS_ABI()) {
388 setLibcallCallingConv(RTLIB::FPROUND_F32_F16, CallingConv::ARM_AAPCS);
389 setLibcallCallingConv(RTLIB::FPROUND_F64_F16, CallingConv::ARM_AAPCS);
390 setLibcallCallingConv(RTLIB::FPEXT_F16_F32, CallingConv::ARM_AAPCS);
392 setLibcallCallingConv(RTLIB::FPROUND_F32_F16, CallingConv::ARM_APCS);
393 setLibcallCallingConv(RTLIB::FPROUND_F64_F16, CallingConv::ARM_APCS);
394 setLibcallCallingConv(RTLIB::FPEXT_F16_F32, CallingConv::ARM_APCS);
397 if (Subtarget->isThumb1Only())
398 addRegisterClass(MVT::i32, &ARM::tGPRRegClass);
400 addRegisterClass(MVT::i32, &ARM::GPRRegClass);
401 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
402 !Subtarget->isThumb1Only()) {
403 addRegisterClass(MVT::f32, &ARM::SPRRegClass);
404 addRegisterClass(MVT::f64, &ARM::DPRRegClass);
407 for (MVT VT : MVT::vector_valuetypes()) {
408 for (MVT InnerVT : MVT::vector_valuetypes()) {
409 setTruncStoreAction(VT, InnerVT, Expand);
410 setLoadExtAction(ISD::SEXTLOAD, VT, InnerVT, Expand);
411 setLoadExtAction(ISD::ZEXTLOAD, VT, InnerVT, Expand);
412 setLoadExtAction(ISD::EXTLOAD, VT, InnerVT, Expand);
415 setOperationAction(ISD::MULHS, VT, Expand);
416 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
417 setOperationAction(ISD::MULHU, VT, Expand);
418 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
420 setOperationAction(ISD::BSWAP, VT, Expand);
423 setOperationAction(ISD::ConstantFP, MVT::f32, Custom);
424 setOperationAction(ISD::ConstantFP, MVT::f64, Custom);
426 if (Subtarget->hasNEON()) {
427 addDRTypeForNEON(MVT::v2f32);
428 addDRTypeForNEON(MVT::v8i8);
429 addDRTypeForNEON(MVT::v4i16);
430 addDRTypeForNEON(MVT::v2i32);
431 addDRTypeForNEON(MVT::v1i64);
433 addQRTypeForNEON(MVT::v4f32);
434 addQRTypeForNEON(MVT::v2f64);
435 addQRTypeForNEON(MVT::v16i8);
436 addQRTypeForNEON(MVT::v8i16);
437 addQRTypeForNEON(MVT::v4i32);
438 addQRTypeForNEON(MVT::v2i64);
440 // v2f64 is legal so that QR subregs can be extracted as f64 elements, but
441 // neither Neon nor VFP support any arithmetic operations on it.
442 // The same with v4f32. But keep in mind that vadd, vsub, vmul are natively
443 // supported for v4f32.
444 setOperationAction(ISD::FADD, MVT::v2f64, Expand);
445 setOperationAction(ISD::FSUB, MVT::v2f64, Expand);
446 setOperationAction(ISD::FMUL, MVT::v2f64, Expand);
447 // FIXME: Code duplication: FDIV and FREM are expanded always, see
448 // ARMTargetLowering::addTypeForNEON method for details.
449 setOperationAction(ISD::FDIV, MVT::v2f64, Expand);
450 setOperationAction(ISD::FREM, MVT::v2f64, Expand);
451 // FIXME: Create unittest.
452 // In another words, find a way when "copysign" appears in DAG with vector
454 setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand);
455 // FIXME: Code duplication: SETCC has custom operation action, see
456 // ARMTargetLowering::addTypeForNEON method for details.
457 setOperationAction(ISD::SETCC, MVT::v2f64, Expand);
458 // FIXME: Create unittest for FNEG and for FABS.
459 setOperationAction(ISD::FNEG, MVT::v2f64, Expand);
460 setOperationAction(ISD::FABS, MVT::v2f64, Expand);
461 setOperationAction(ISD::FSQRT, MVT::v2f64, Expand);
462 setOperationAction(ISD::FSIN, MVT::v2f64, Expand);
463 setOperationAction(ISD::FCOS, MVT::v2f64, Expand);
464 setOperationAction(ISD::FPOWI, MVT::v2f64, Expand);
465 setOperationAction(ISD::FPOW, MVT::v2f64, Expand);
466 setOperationAction(ISD::FLOG, MVT::v2f64, Expand);
467 setOperationAction(ISD::FLOG2, MVT::v2f64, Expand);
468 setOperationAction(ISD::FLOG10, MVT::v2f64, Expand);
469 setOperationAction(ISD::FEXP, MVT::v2f64, Expand);
470 setOperationAction(ISD::FEXP2, MVT::v2f64, Expand);
471 // FIXME: Create unittest for FCEIL, FTRUNC, FRINT, FNEARBYINT, FFLOOR.
472 setOperationAction(ISD::FCEIL, MVT::v2f64, Expand);
473 setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand);
474 setOperationAction(ISD::FRINT, MVT::v2f64, Expand);
475 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand);
476 setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand);
477 setOperationAction(ISD::FMA, MVT::v2f64, Expand);
479 setOperationAction(ISD::FSQRT, MVT::v4f32, Expand);
480 setOperationAction(ISD::FSIN, MVT::v4f32, Expand);
481 setOperationAction(ISD::FCOS, MVT::v4f32, Expand);
482 setOperationAction(ISD::FPOWI, MVT::v4f32, Expand);
483 setOperationAction(ISD::FPOW, MVT::v4f32, Expand);
484 setOperationAction(ISD::FLOG, MVT::v4f32, Expand);
485 setOperationAction(ISD::FLOG2, MVT::v4f32, Expand);
486 setOperationAction(ISD::FLOG10, MVT::v4f32, Expand);
487 setOperationAction(ISD::FEXP, MVT::v4f32, Expand);
488 setOperationAction(ISD::FEXP2, MVT::v4f32, Expand);
489 setOperationAction(ISD::FCEIL, MVT::v4f32, Expand);
490 setOperationAction(ISD::FTRUNC, MVT::v4f32, Expand);
491 setOperationAction(ISD::FRINT, MVT::v4f32, Expand);
492 setOperationAction(ISD::FNEARBYINT, MVT::v4f32, Expand);
493 setOperationAction(ISD::FFLOOR, MVT::v4f32, Expand);
495 // Mark v2f32 intrinsics.
496 setOperationAction(ISD::FSQRT, MVT::v2f32, Expand);
497 setOperationAction(ISD::FSIN, MVT::v2f32, Expand);
498 setOperationAction(ISD::FCOS, MVT::v2f32, Expand);
499 setOperationAction(ISD::FPOWI, MVT::v2f32, Expand);
500 setOperationAction(ISD::FPOW, MVT::v2f32, Expand);
501 setOperationAction(ISD::FLOG, MVT::v2f32, Expand);
502 setOperationAction(ISD::FLOG2, MVT::v2f32, Expand);
503 setOperationAction(ISD::FLOG10, MVT::v2f32, Expand);
504 setOperationAction(ISD::FEXP, MVT::v2f32, Expand);
505 setOperationAction(ISD::FEXP2, MVT::v2f32, Expand);
506 setOperationAction(ISD::FCEIL, MVT::v2f32, Expand);
507 setOperationAction(ISD::FTRUNC, MVT::v2f32, Expand);
508 setOperationAction(ISD::FRINT, MVT::v2f32, Expand);
509 setOperationAction(ISD::FNEARBYINT, MVT::v2f32, Expand);
510 setOperationAction(ISD::FFLOOR, MVT::v2f32, Expand);
512 // Neon does not support some operations on v1i64 and v2i64 types.
513 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
514 // Custom handling for some quad-vector types to detect VMULL.
515 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
516 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
517 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
518 // Custom handling for some vector types to avoid expensive expansions
519 setOperationAction(ISD::SDIV, MVT::v4i16, Custom);
520 setOperationAction(ISD::SDIV, MVT::v8i8, Custom);
521 setOperationAction(ISD::UDIV, MVT::v4i16, Custom);
522 setOperationAction(ISD::UDIV, MVT::v8i8, Custom);
523 setOperationAction(ISD::SETCC, MVT::v1i64, Expand);
524 setOperationAction(ISD::SETCC, MVT::v2i64, Expand);
525 // Neon does not have single instruction SINT_TO_FP and UINT_TO_FP with
526 // a destination type that is wider than the source, and nor does
527 // it have a FP_TO_[SU]INT instruction with a narrower destination than
529 setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Custom);
530 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
531 setOperationAction(ISD::FP_TO_UINT, MVT::v4i16, Custom);
532 setOperationAction(ISD::FP_TO_SINT, MVT::v4i16, Custom);
534 setOperationAction(ISD::FP_ROUND, MVT::v2f32, Expand);
535 setOperationAction(ISD::FP_EXTEND, MVT::v2f64, Expand);
537 // NEON does not have single instruction CTPOP for vectors with element
538 // types wider than 8-bits. However, custom lowering can leverage the
539 // v8i8/v16i8 vcnt instruction.
540 setOperationAction(ISD::CTPOP, MVT::v2i32, Custom);
541 setOperationAction(ISD::CTPOP, MVT::v4i32, Custom);
542 setOperationAction(ISD::CTPOP, MVT::v4i16, Custom);
543 setOperationAction(ISD::CTPOP, MVT::v8i16, Custom);
545 // NEON only has FMA instructions as of VFP4.
546 if (!Subtarget->hasVFP4()) {
547 setOperationAction(ISD::FMA, MVT::v2f32, Expand);
548 setOperationAction(ISD::FMA, MVT::v4f32, Expand);
551 setTargetDAGCombine(ISD::INTRINSIC_VOID);
552 setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN);
553 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
554 setTargetDAGCombine(ISD::SHL);
555 setTargetDAGCombine(ISD::SRL);
556 setTargetDAGCombine(ISD::SRA);
557 setTargetDAGCombine(ISD::SIGN_EXTEND);
558 setTargetDAGCombine(ISD::ZERO_EXTEND);
559 setTargetDAGCombine(ISD::ANY_EXTEND);
560 setTargetDAGCombine(ISD::SELECT_CC);
561 setTargetDAGCombine(ISD::BUILD_VECTOR);
562 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
563 setTargetDAGCombine(ISD::INSERT_VECTOR_ELT);
564 setTargetDAGCombine(ISD::STORE);
565 setTargetDAGCombine(ISD::FP_TO_SINT);
566 setTargetDAGCombine(ISD::FP_TO_UINT);
567 setTargetDAGCombine(ISD::FDIV);
569 // It is legal to extload from v4i8 to v4i16 or v4i32.
570 MVT Tys[6] = {MVT::v8i8, MVT::v4i8, MVT::v2i8,
571 MVT::v4i16, MVT::v2i16,
573 for (unsigned i = 0; i < 6; ++i) {
574 for (MVT VT : MVT::integer_vector_valuetypes()) {
575 setLoadExtAction(ISD::EXTLOAD, VT, Tys[i], Legal);
576 setLoadExtAction(ISD::ZEXTLOAD, VT, Tys[i], Legal);
577 setLoadExtAction(ISD::SEXTLOAD, VT, Tys[i], Legal);
582 // ARM and Thumb2 support UMLAL/SMLAL.
583 if (!Subtarget->isThumb1Only())
584 setTargetDAGCombine(ISD::ADDC);
586 if (Subtarget->isFPOnlySP()) {
587 // When targetting a floating-point unit with only single-precision
588 // operations, f64 is legal for the few double-precision instructions which
589 // are present However, no double-precision operations other than moves,
590 // loads and stores are provided by the hardware.
591 setOperationAction(ISD::FADD, MVT::f64, Expand);
592 setOperationAction(ISD::FSUB, MVT::f64, Expand);
593 setOperationAction(ISD::FMUL, MVT::f64, Expand);
594 setOperationAction(ISD::FMA, MVT::f64, Expand);
595 setOperationAction(ISD::FDIV, MVT::f64, Expand);
596 setOperationAction(ISD::FREM, MVT::f64, Expand);
597 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
598 setOperationAction(ISD::FGETSIGN, MVT::f64, Expand);
599 setOperationAction(ISD::FNEG, MVT::f64, Expand);
600 setOperationAction(ISD::FABS, MVT::f64, Expand);
601 setOperationAction(ISD::FSQRT, MVT::f64, Expand);
602 setOperationAction(ISD::FSIN, MVT::f64, Expand);
603 setOperationAction(ISD::FCOS, MVT::f64, Expand);
604 setOperationAction(ISD::FPOWI, MVT::f64, Expand);
605 setOperationAction(ISD::FPOW, MVT::f64, Expand);
606 setOperationAction(ISD::FLOG, MVT::f64, Expand);
607 setOperationAction(ISD::FLOG2, MVT::f64, Expand);
608 setOperationAction(ISD::FLOG10, MVT::f64, Expand);
609 setOperationAction(ISD::FEXP, MVT::f64, Expand);
610 setOperationAction(ISD::FEXP2, MVT::f64, Expand);
611 setOperationAction(ISD::FCEIL, MVT::f64, Expand);
612 setOperationAction(ISD::FTRUNC, MVT::f64, Expand);
613 setOperationAction(ISD::FRINT, MVT::f64, Expand);
614 setOperationAction(ISD::FNEARBYINT, MVT::f64, Expand);
615 setOperationAction(ISD::FFLOOR, MVT::f64, Expand);
616 setOperationAction(ISD::FP_ROUND, MVT::f32, Custom);
617 setOperationAction(ISD::FP_EXTEND, MVT::f64, Custom);
620 computeRegisterProperties();
622 // ARM does not have floating-point extending loads.
623 for (MVT VT : MVT::fp_valuetypes()) {
624 setLoadExtAction(ISD::EXTLOAD, VT, MVT::f32, Expand);
625 setLoadExtAction(ISD::EXTLOAD, VT, MVT::f16, Expand);
628 // ... or truncating stores
629 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
630 setTruncStoreAction(MVT::f32, MVT::f16, Expand);
631 setTruncStoreAction(MVT::f64, MVT::f16, Expand);
633 // ARM does not have i1 sign extending load.
634 for (MVT VT : MVT::integer_valuetypes())
635 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Promote);
637 // ARM supports all 4 flavors of integer indexed load / store.
638 if (!Subtarget->isThumb1Only()) {
639 for (unsigned im = (unsigned)ISD::PRE_INC;
640 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
641 setIndexedLoadAction(im, MVT::i1, Legal);
642 setIndexedLoadAction(im, MVT::i8, Legal);
643 setIndexedLoadAction(im, MVT::i16, Legal);
644 setIndexedLoadAction(im, MVT::i32, Legal);
645 setIndexedStoreAction(im, MVT::i1, Legal);
646 setIndexedStoreAction(im, MVT::i8, Legal);
647 setIndexedStoreAction(im, MVT::i16, Legal);
648 setIndexedStoreAction(im, MVT::i32, Legal);
652 setOperationAction(ISD::SADDO, MVT::i32, Custom);
653 setOperationAction(ISD::UADDO, MVT::i32, Custom);
654 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
655 setOperationAction(ISD::USUBO, MVT::i32, Custom);
657 // i64 operation support.
658 setOperationAction(ISD::MUL, MVT::i64, Expand);
659 setOperationAction(ISD::MULHU, MVT::i32, Expand);
660 if (Subtarget->isThumb1Only()) {
661 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
662 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
664 if (Subtarget->isThumb1Only() || !Subtarget->hasV6Ops()
665 || (Subtarget->isThumb2() && !Subtarget->hasThumb2DSP()))
666 setOperationAction(ISD::MULHS, MVT::i32, Expand);
668 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
669 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
670 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
671 setOperationAction(ISD::SRL, MVT::i64, Custom);
672 setOperationAction(ISD::SRA, MVT::i64, Custom);
674 if (!Subtarget->isThumb1Only()) {
675 // FIXME: We should do this for Thumb1 as well.
676 setOperationAction(ISD::ADDC, MVT::i32, Custom);
677 setOperationAction(ISD::ADDE, MVT::i32, Custom);
678 setOperationAction(ISD::SUBC, MVT::i32, Custom);
679 setOperationAction(ISD::SUBE, MVT::i32, Custom);
682 // ARM does not have ROTL.
683 setOperationAction(ISD::ROTL, MVT::i32, Expand);
684 setOperationAction(ISD::CTTZ, MVT::i32, Custom);
685 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
686 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
687 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
689 // These just redirect to CTTZ and CTLZ on ARM.
690 setOperationAction(ISD::CTTZ_ZERO_UNDEF , MVT::i32 , Expand);
691 setOperationAction(ISD::CTLZ_ZERO_UNDEF , MVT::i32 , Expand);
693 setOperationAction(ISD::READCYCLECOUNTER, MVT::i64, Custom);
695 // Only ARMv6 has BSWAP.
696 if (!Subtarget->hasV6Ops())
697 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
699 if (!(Subtarget->hasDivide() && Subtarget->isThumb2()) &&
700 !(Subtarget->hasDivideInARMMode() && !Subtarget->isThumb())) {
701 // These are expanded into libcalls if the cpu doesn't have HW divider.
702 setOperationAction(ISD::SDIV, MVT::i32, Expand);
703 setOperationAction(ISD::UDIV, MVT::i32, Expand);
706 // FIXME: Also set divmod for SREM on EABI
707 setOperationAction(ISD::SREM, MVT::i32, Expand);
708 setOperationAction(ISD::UREM, MVT::i32, Expand);
709 // Register based DivRem for AEABI (RTABI 4.2)
710 if (Subtarget->isTargetAEABI()) {
711 setLibcallName(RTLIB::SDIVREM_I8, "__aeabi_idivmod");
712 setLibcallName(RTLIB::SDIVREM_I16, "__aeabi_idivmod");
713 setLibcallName(RTLIB::SDIVREM_I32, "__aeabi_idivmod");
714 setLibcallName(RTLIB::SDIVREM_I64, "__aeabi_ldivmod");
715 setLibcallName(RTLIB::UDIVREM_I8, "__aeabi_uidivmod");
716 setLibcallName(RTLIB::UDIVREM_I16, "__aeabi_uidivmod");
717 setLibcallName(RTLIB::UDIVREM_I32, "__aeabi_uidivmod");
718 setLibcallName(RTLIB::UDIVREM_I64, "__aeabi_uldivmod");
720 setLibcallCallingConv(RTLIB::SDIVREM_I8, CallingConv::ARM_AAPCS);
721 setLibcallCallingConv(RTLIB::SDIVREM_I16, CallingConv::ARM_AAPCS);
722 setLibcallCallingConv(RTLIB::SDIVREM_I32, CallingConv::ARM_AAPCS);
723 setLibcallCallingConv(RTLIB::SDIVREM_I64, CallingConv::ARM_AAPCS);
724 setLibcallCallingConv(RTLIB::UDIVREM_I8, CallingConv::ARM_AAPCS);
725 setLibcallCallingConv(RTLIB::UDIVREM_I16, CallingConv::ARM_AAPCS);
726 setLibcallCallingConv(RTLIB::UDIVREM_I32, CallingConv::ARM_AAPCS);
727 setLibcallCallingConv(RTLIB::UDIVREM_I64, CallingConv::ARM_AAPCS);
729 setOperationAction(ISD::SDIVREM, MVT::i32, Custom);
730 setOperationAction(ISD::UDIVREM, MVT::i32, Custom);
732 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
733 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
736 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
737 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
738 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
739 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
740 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
742 setOperationAction(ISD::TRAP, MVT::Other, Legal);
744 // Use the default implementation.
745 setOperationAction(ISD::VASTART, MVT::Other, Custom);
746 setOperationAction(ISD::VAARG, MVT::Other, Expand);
747 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
748 setOperationAction(ISD::VAEND, MVT::Other, Expand);
749 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
750 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
752 if (!Subtarget->isTargetMachO()) {
753 // Non-MachO platforms may return values in these registers via the
754 // personality function.
755 setExceptionPointerRegister(ARM::R0);
756 setExceptionSelectorRegister(ARM::R1);
759 if (Subtarget->getTargetTriple().isWindowsItaniumEnvironment())
760 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
762 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
764 // ARMv6 Thumb1 (except for CPUs that support dmb / dsb) and earlier use
765 // the default expansion. If we are targeting a single threaded system,
766 // then set them all for expand so we can lower them later into their
768 if (TM.Options.ThreadModel == ThreadModel::Single)
769 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Expand);
770 else if (Subtarget->hasAnyDataBarrier() && !Subtarget->isThumb1Only()) {
771 // ATOMIC_FENCE needs custom lowering; the others should have been expanded
772 // to ldrex/strex loops already.
773 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Custom);
775 // On v8, we have particularly efficient implementations of atomic fences
776 // if they can be combined with nearby atomic loads and stores.
777 if (!Subtarget->hasV8Ops()) {
778 // Automatically insert fences (dmb ish) around ATOMIC_SWAP etc.
779 setInsertFencesForAtomic(true);
782 // If there's anything we can use as a barrier, go through custom lowering
784 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other,
785 Subtarget->hasAnyDataBarrier() ? Custom : Expand);
787 // Set them all for expansion, which will force libcalls.
788 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Expand);
789 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Expand);
790 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Expand);
791 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Expand);
792 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Expand);
793 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Expand);
794 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Expand);
795 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Expand);
796 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i32, Expand);
797 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i32, Expand);
798 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i32, Expand);
799 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i32, Expand);
800 // Mark ATOMIC_LOAD and ATOMIC_STORE custom so we can handle the
801 // Unordered/Monotonic case.
802 setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Custom);
803 setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Custom);
806 setOperationAction(ISD::PREFETCH, MVT::Other, Custom);
808 // Requires SXTB/SXTH, available on v6 and up in both ARM and Thumb modes.
809 if (!Subtarget->hasV6Ops()) {
810 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
811 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
813 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
815 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
816 !Subtarget->isThumb1Only()) {
817 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
818 // iff target supports vfp2.
819 setOperationAction(ISD::BITCAST, MVT::i64, Custom);
820 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
823 // We want to custom lower some of our intrinsics.
824 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
825 if (Subtarget->isTargetDarwin()) {
826 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
827 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
828 setLibcallName(RTLIB::UNWIND_RESUME, "_Unwind_SjLj_Resume");
831 setOperationAction(ISD::SETCC, MVT::i32, Expand);
832 setOperationAction(ISD::SETCC, MVT::f32, Expand);
833 setOperationAction(ISD::SETCC, MVT::f64, Expand);
834 setOperationAction(ISD::SELECT, MVT::i32, Custom);
835 setOperationAction(ISD::SELECT, MVT::f32, Custom);
836 setOperationAction(ISD::SELECT, MVT::f64, Custom);
837 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
838 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
839 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
841 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
842 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
843 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
844 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
845 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
847 // We don't support sin/cos/fmod/copysign/pow
848 setOperationAction(ISD::FSIN, MVT::f64, Expand);
849 setOperationAction(ISD::FSIN, MVT::f32, Expand);
850 setOperationAction(ISD::FCOS, MVT::f32, Expand);
851 setOperationAction(ISD::FCOS, MVT::f64, Expand);
852 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
853 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
854 setOperationAction(ISD::FREM, MVT::f64, Expand);
855 setOperationAction(ISD::FREM, MVT::f32, Expand);
856 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
857 !Subtarget->isThumb1Only()) {
858 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
859 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
861 setOperationAction(ISD::FPOW, MVT::f64, Expand);
862 setOperationAction(ISD::FPOW, MVT::f32, Expand);
864 if (!Subtarget->hasVFP4()) {
865 setOperationAction(ISD::FMA, MVT::f64, Expand);
866 setOperationAction(ISD::FMA, MVT::f32, Expand);
869 // Various VFP goodness
870 if (!TM.Options.UseSoftFloat && !Subtarget->isThumb1Only()) {
871 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
872 if (Subtarget->hasVFP2()) {
873 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
874 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
875 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
876 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
879 // FP-ARMv8 adds f64 <-> f16 conversion. Before that it should be expanded.
880 if (!Subtarget->hasFPARMv8() || Subtarget->isFPOnlySP()) {
881 setOperationAction(ISD::FP16_TO_FP, MVT::f64, Expand);
882 setOperationAction(ISD::FP_TO_FP16, MVT::f64, Expand);
885 // fp16 is a special v7 extension that adds f16 <-> f32 conversions.
886 if (!Subtarget->hasFP16()) {
887 setOperationAction(ISD::FP16_TO_FP, MVT::f32, Expand);
888 setOperationAction(ISD::FP_TO_FP16, MVT::f32, Expand);
892 // Combine sin / cos into one node or libcall if possible.
893 if (Subtarget->hasSinCos()) {
894 setLibcallName(RTLIB::SINCOS_F32, "sincosf");
895 setLibcallName(RTLIB::SINCOS_F64, "sincos");
896 if (Subtarget->getTargetTriple().isiOS()) {
897 // For iOS, we don't want to the normal expansion of a libcall to
898 // sincos. We want to issue a libcall to __sincos_stret.
899 setOperationAction(ISD::FSINCOS, MVT::f64, Custom);
900 setOperationAction(ISD::FSINCOS, MVT::f32, Custom);
904 // FP-ARMv8 implements a lot of rounding-like FP operations.
905 if (Subtarget->hasFPARMv8()) {
906 setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
907 setOperationAction(ISD::FCEIL, MVT::f32, Legal);
908 setOperationAction(ISD::FROUND, MVT::f32, Legal);
909 setOperationAction(ISD::FTRUNC, MVT::f32, Legal);
910 setOperationAction(ISD::FNEARBYINT, MVT::f32, Legal);
911 setOperationAction(ISD::FRINT, MVT::f32, Legal);
912 if (!Subtarget->isFPOnlySP()) {
913 setOperationAction(ISD::FFLOOR, MVT::f64, Legal);
914 setOperationAction(ISD::FCEIL, MVT::f64, Legal);
915 setOperationAction(ISD::FROUND, MVT::f64, Legal);
916 setOperationAction(ISD::FTRUNC, MVT::f64, Legal);
917 setOperationAction(ISD::FNEARBYINT, MVT::f64, Legal);
918 setOperationAction(ISD::FRINT, MVT::f64, Legal);
921 // We have target-specific dag combine patterns for the following nodes:
922 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
923 setTargetDAGCombine(ISD::ADD);
924 setTargetDAGCombine(ISD::SUB);
925 setTargetDAGCombine(ISD::MUL);
926 setTargetDAGCombine(ISD::AND);
927 setTargetDAGCombine(ISD::OR);
928 setTargetDAGCombine(ISD::XOR);
930 if (Subtarget->hasV6Ops())
931 setTargetDAGCombine(ISD::SRL);
933 setStackPointerRegisterToSaveRestore(ARM::SP);
935 if (TM.Options.UseSoftFloat || Subtarget->isThumb1Only() ||
936 !Subtarget->hasVFP2())
937 setSchedulingPreference(Sched::RegPressure);
939 setSchedulingPreference(Sched::Hybrid);
941 //// temporary - rewrite interface to use type
942 MaxStoresPerMemset = 8;
943 MaxStoresPerMemsetOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
944 MaxStoresPerMemcpy = 4; // For @llvm.memcpy -> sequence of stores
945 MaxStoresPerMemcpyOptSize = Subtarget->isTargetDarwin() ? 4 : 2;
946 MaxStoresPerMemmove = 4; // For @llvm.memmove -> sequence of stores
947 MaxStoresPerMemmoveOptSize = Subtarget->isTargetDarwin() ? 4 : 2;
949 // On ARM arguments smaller than 4 bytes are extended, so all arguments
950 // are at least 4 bytes aligned.
951 setMinStackArgumentAlignment(4);
953 // Prefer likely predicted branches to selects on out-of-order cores.
954 PredictableSelectIsExpensive = Subtarget->isLikeA9();
956 setMinFunctionAlignment(Subtarget->isThumb() ? 1 : 2);
959 // FIXME: It might make sense to define the representative register class as the
960 // nearest super-register that has a non-null superset. For example, DPR_VFP2 is
961 // a super-register of SPR, and DPR is a superset if DPR_VFP2. Consequently,
962 // SPR's representative would be DPR_VFP2. This should work well if register
963 // pressure tracking were modified such that a register use would increment the
964 // pressure of the register class's representative and all of it's super
965 // classes' representatives transitively. We have not implemented this because
966 // of the difficulty prior to coalescing of modeling operand register classes
967 // due to the common occurrence of cross class copies and subregister insertions
969 std::pair<const TargetRegisterClass*, uint8_t>
970 ARMTargetLowering::findRepresentativeClass(MVT VT) const{
971 const TargetRegisterClass *RRC = nullptr;
973 switch (VT.SimpleTy) {
975 return TargetLowering::findRepresentativeClass(VT);
976 // Use DPR as representative register class for all floating point
977 // and vector types. Since there are 32 SPR registers and 32 DPR registers so
978 // the cost is 1 for both f32 and f64.
979 case MVT::f32: case MVT::f64: case MVT::v8i8: case MVT::v4i16:
980 case MVT::v2i32: case MVT::v1i64: case MVT::v2f32:
981 RRC = &ARM::DPRRegClass;
982 // When NEON is used for SP, only half of the register file is available
983 // because operations that define both SP and DP results will be constrained
984 // to the VFP2 class (D0-D15). We currently model this constraint prior to
985 // coalescing by double-counting the SP regs. See the FIXME above.
986 if (Subtarget->useNEONForSinglePrecisionFP())
989 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
990 case MVT::v4f32: case MVT::v2f64:
991 RRC = &ARM::DPRRegClass;
995 RRC = &ARM::DPRRegClass;
999 RRC = &ARM::DPRRegClass;
1003 return std::make_pair(RRC, Cost);
1006 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
1008 default: return nullptr;
1009 case ARMISD::Wrapper: return "ARMISD::Wrapper";
1010 case ARMISD::WrapperPIC: return "ARMISD::WrapperPIC";
1011 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
1012 case ARMISD::CALL: return "ARMISD::CALL";
1013 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
1014 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
1015 case ARMISD::tCALL: return "ARMISD::tCALL";
1016 case ARMISD::BRCOND: return "ARMISD::BRCOND";
1017 case ARMISD::BR_JT: return "ARMISD::BR_JT";
1018 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
1019 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
1020 case ARMISD::INTRET_FLAG: return "ARMISD::INTRET_FLAG";
1021 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
1022 case ARMISD::CMP: return "ARMISD::CMP";
1023 case ARMISD::CMN: return "ARMISD::CMN";
1024 case ARMISD::CMPZ: return "ARMISD::CMPZ";
1025 case ARMISD::CMPFP: return "ARMISD::CMPFP";
1026 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
1027 case ARMISD::BCC_i64: return "ARMISD::BCC_i64";
1028 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
1030 case ARMISD::CMOV: return "ARMISD::CMOV";
1032 case ARMISD::RBIT: return "ARMISD::RBIT";
1034 case ARMISD::FTOSI: return "ARMISD::FTOSI";
1035 case ARMISD::FTOUI: return "ARMISD::FTOUI";
1036 case ARMISD::SITOF: return "ARMISD::SITOF";
1037 case ARMISD::UITOF: return "ARMISD::UITOF";
1039 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
1040 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
1041 case ARMISD::RRX: return "ARMISD::RRX";
1043 case ARMISD::ADDC: return "ARMISD::ADDC";
1044 case ARMISD::ADDE: return "ARMISD::ADDE";
1045 case ARMISD::SUBC: return "ARMISD::SUBC";
1046 case ARMISD::SUBE: return "ARMISD::SUBE";
1048 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
1049 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
1051 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
1052 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
1054 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN";
1056 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
1058 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
1060 case ARMISD::MEMBARRIER_MCR: return "ARMISD::MEMBARRIER_MCR";
1062 case ARMISD::PRELOAD: return "ARMISD::PRELOAD";
1064 case ARMISD::WIN__CHKSTK: return "ARMISD:::WIN__CHKSTK";
1066 case ARMISD::VCEQ: return "ARMISD::VCEQ";
1067 case ARMISD::VCEQZ: return "ARMISD::VCEQZ";
1068 case ARMISD::VCGE: return "ARMISD::VCGE";
1069 case ARMISD::VCGEZ: return "ARMISD::VCGEZ";
1070 case ARMISD::VCLEZ: return "ARMISD::VCLEZ";
1071 case ARMISD::VCGEU: return "ARMISD::VCGEU";
1072 case ARMISD::VCGT: return "ARMISD::VCGT";
1073 case ARMISD::VCGTZ: return "ARMISD::VCGTZ";
1074 case ARMISD::VCLTZ: return "ARMISD::VCLTZ";
1075 case ARMISD::VCGTU: return "ARMISD::VCGTU";
1076 case ARMISD::VTST: return "ARMISD::VTST";
1078 case ARMISD::VSHL: return "ARMISD::VSHL";
1079 case ARMISD::VSHRs: return "ARMISD::VSHRs";
1080 case ARMISD::VSHRu: return "ARMISD::VSHRu";
1081 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
1082 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
1083 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
1084 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
1085 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
1086 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
1087 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
1088 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
1089 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
1090 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
1091 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
1092 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
1093 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
1094 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
1095 case ARMISD::VMOVIMM: return "ARMISD::VMOVIMM";
1096 case ARMISD::VMVNIMM: return "ARMISD::VMVNIMM";
1097 case ARMISD::VMOVFPIMM: return "ARMISD::VMOVFPIMM";
1098 case ARMISD::VDUP: return "ARMISD::VDUP";
1099 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
1100 case ARMISD::VEXT: return "ARMISD::VEXT";
1101 case ARMISD::VREV64: return "ARMISD::VREV64";
1102 case ARMISD::VREV32: return "ARMISD::VREV32";
1103 case ARMISD::VREV16: return "ARMISD::VREV16";
1104 case ARMISD::VZIP: return "ARMISD::VZIP";
1105 case ARMISD::VUZP: return "ARMISD::VUZP";
1106 case ARMISD::VTRN: return "ARMISD::VTRN";
1107 case ARMISD::VTBL1: return "ARMISD::VTBL1";
1108 case ARMISD::VTBL2: return "ARMISD::VTBL2";
1109 case ARMISD::VMULLs: return "ARMISD::VMULLs";
1110 case ARMISD::VMULLu: return "ARMISD::VMULLu";
1111 case ARMISD::UMLAL: return "ARMISD::UMLAL";
1112 case ARMISD::SMLAL: return "ARMISD::SMLAL";
1113 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
1114 case ARMISD::FMAX: return "ARMISD::FMAX";
1115 case ARMISD::FMIN: return "ARMISD::FMIN";
1116 case ARMISD::VMAXNM: return "ARMISD::VMAX";
1117 case ARMISD::VMINNM: return "ARMISD::VMIN";
1118 case ARMISD::BFI: return "ARMISD::BFI";
1119 case ARMISD::VORRIMM: return "ARMISD::VORRIMM";
1120 case ARMISD::VBICIMM: return "ARMISD::VBICIMM";
1121 case ARMISD::VBSL: return "ARMISD::VBSL";
1122 case ARMISD::VLD2DUP: return "ARMISD::VLD2DUP";
1123 case ARMISD::VLD3DUP: return "ARMISD::VLD3DUP";
1124 case ARMISD::VLD4DUP: return "ARMISD::VLD4DUP";
1125 case ARMISD::VLD1_UPD: return "ARMISD::VLD1_UPD";
1126 case ARMISD::VLD2_UPD: return "ARMISD::VLD2_UPD";
1127 case ARMISD::VLD3_UPD: return "ARMISD::VLD3_UPD";
1128 case ARMISD::VLD4_UPD: return "ARMISD::VLD4_UPD";
1129 case ARMISD::VLD2LN_UPD: return "ARMISD::VLD2LN_UPD";
1130 case ARMISD::VLD3LN_UPD: return "ARMISD::VLD3LN_UPD";
1131 case ARMISD::VLD4LN_UPD: return "ARMISD::VLD4LN_UPD";
1132 case ARMISD::VLD2DUP_UPD: return "ARMISD::VLD2DUP_UPD";
1133 case ARMISD::VLD3DUP_UPD: return "ARMISD::VLD3DUP_UPD";
1134 case ARMISD::VLD4DUP_UPD: return "ARMISD::VLD4DUP_UPD";
1135 case ARMISD::VST1_UPD: return "ARMISD::VST1_UPD";
1136 case ARMISD::VST2_UPD: return "ARMISD::VST2_UPD";
1137 case ARMISD::VST3_UPD: return "ARMISD::VST3_UPD";
1138 case ARMISD::VST4_UPD: return "ARMISD::VST4_UPD";
1139 case ARMISD::VST2LN_UPD: return "ARMISD::VST2LN_UPD";
1140 case ARMISD::VST3LN_UPD: return "ARMISD::VST3LN_UPD";
1141 case ARMISD::VST4LN_UPD: return "ARMISD::VST4LN_UPD";
1145 EVT ARMTargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
1146 if (!VT.isVector()) return getPointerTy();
1147 return VT.changeVectorElementTypeToInteger();
1150 /// getRegClassFor - Return the register class that should be used for the
1151 /// specified value type.
1152 const TargetRegisterClass *ARMTargetLowering::getRegClassFor(MVT VT) const {
1153 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
1154 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
1155 // load / store 4 to 8 consecutive D registers.
1156 if (Subtarget->hasNEON()) {
1157 if (VT == MVT::v4i64)
1158 return &ARM::QQPRRegClass;
1159 if (VT == MVT::v8i64)
1160 return &ARM::QQQQPRRegClass;
1162 return TargetLowering::getRegClassFor(VT);
1165 // Create a fast isel object.
1167 ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
1168 const TargetLibraryInfo *libInfo) const {
1169 return ARM::createFastISel(funcInfo, libInfo);
1172 /// getMaximalGlobalOffset - Returns the maximal possible offset which can
1173 /// be used for loads / stores from the global.
1174 unsigned ARMTargetLowering::getMaximalGlobalOffset() const {
1175 return (Subtarget->isThumb1Only() ? 127 : 4095);
1178 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
1179 unsigned NumVals = N->getNumValues();
1181 return Sched::RegPressure;
1183 for (unsigned i = 0; i != NumVals; ++i) {
1184 EVT VT = N->getValueType(i);
1185 if (VT == MVT::Glue || VT == MVT::Other)
1187 if (VT.isFloatingPoint() || VT.isVector())
1191 if (!N->isMachineOpcode())
1192 return Sched::RegPressure;
1194 // Load are scheduled for latency even if there instruction itinerary
1195 // is not available.
1196 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
1197 const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
1199 if (MCID.getNumDefs() == 0)
1200 return Sched::RegPressure;
1201 if (!Itins->isEmpty() &&
1202 Itins->getOperandCycle(MCID.getSchedClass(), 0) > 2)
1205 return Sched::RegPressure;
1208 //===----------------------------------------------------------------------===//
1210 //===----------------------------------------------------------------------===//
1212 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
1213 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
1215 default: llvm_unreachable("Unknown condition code!");
1216 case ISD::SETNE: return ARMCC::NE;
1217 case ISD::SETEQ: return ARMCC::EQ;
1218 case ISD::SETGT: return ARMCC::GT;
1219 case ISD::SETGE: return ARMCC::GE;
1220 case ISD::SETLT: return ARMCC::LT;
1221 case ISD::SETLE: return ARMCC::LE;
1222 case ISD::SETUGT: return ARMCC::HI;
1223 case ISD::SETUGE: return ARMCC::HS;
1224 case ISD::SETULT: return ARMCC::LO;
1225 case ISD::SETULE: return ARMCC::LS;
1229 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
1230 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
1231 ARMCC::CondCodes &CondCode2) {
1232 CondCode2 = ARMCC::AL;
1234 default: llvm_unreachable("Unknown FP condition!");
1236 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
1238 case ISD::SETOGT: CondCode = ARMCC::GT; break;
1240 case ISD::SETOGE: CondCode = ARMCC::GE; break;
1241 case ISD::SETOLT: CondCode = ARMCC::MI; break;
1242 case ISD::SETOLE: CondCode = ARMCC::LS; break;
1243 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
1244 case ISD::SETO: CondCode = ARMCC::VC; break;
1245 case ISD::SETUO: CondCode = ARMCC::VS; break;
1246 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
1247 case ISD::SETUGT: CondCode = ARMCC::HI; break;
1248 case ISD::SETUGE: CondCode = ARMCC::PL; break;
1250 case ISD::SETULT: CondCode = ARMCC::LT; break;
1252 case ISD::SETULE: CondCode = ARMCC::LE; break;
1254 case ISD::SETUNE: CondCode = ARMCC::NE; break;
1258 //===----------------------------------------------------------------------===//
1259 // Calling Convention Implementation
1260 //===----------------------------------------------------------------------===//
1262 #include "ARMGenCallingConv.inc"
1264 /// getEffectiveCallingConv - Get the effective calling convention, taking into
1265 /// account presence of floating point hardware and calling convention
1266 /// limitations, such as support for variadic functions.
1268 ARMTargetLowering::getEffectiveCallingConv(CallingConv::ID CC,
1269 bool isVarArg) const {
1272 llvm_unreachable("Unsupported calling convention");
1273 case CallingConv::ARM_AAPCS:
1274 case CallingConv::ARM_APCS:
1275 case CallingConv::GHC:
1277 case CallingConv::ARM_AAPCS_VFP:
1278 return isVarArg ? CallingConv::ARM_AAPCS : CallingConv::ARM_AAPCS_VFP;
1279 case CallingConv::C:
1280 if (!Subtarget->isAAPCS_ABI())
1281 return CallingConv::ARM_APCS;
1282 else if (Subtarget->hasVFP2() && !Subtarget->isThumb1Only() &&
1283 getTargetMachine().Options.FloatABIType == FloatABI::Hard &&
1285 return CallingConv::ARM_AAPCS_VFP;
1287 return CallingConv::ARM_AAPCS;
1288 case CallingConv::Fast:
1289 if (!Subtarget->isAAPCS_ABI()) {
1290 if (Subtarget->hasVFP2() && !Subtarget->isThumb1Only() && !isVarArg)
1291 return CallingConv::Fast;
1292 return CallingConv::ARM_APCS;
1293 } else if (Subtarget->hasVFP2() && !Subtarget->isThumb1Only() && !isVarArg)
1294 return CallingConv::ARM_AAPCS_VFP;
1296 return CallingConv::ARM_AAPCS;
1300 /// CCAssignFnForNode - Selects the correct CCAssignFn for the given
1301 /// CallingConvention.
1302 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
1304 bool isVarArg) const {
1305 switch (getEffectiveCallingConv(CC, isVarArg)) {
1307 llvm_unreachable("Unsupported calling convention");
1308 case CallingConv::ARM_APCS:
1309 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1310 case CallingConv::ARM_AAPCS:
1311 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1312 case CallingConv::ARM_AAPCS_VFP:
1313 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1314 case CallingConv::Fast:
1315 return (Return ? RetFastCC_ARM_APCS : FastCC_ARM_APCS);
1316 case CallingConv::GHC:
1317 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS_GHC);
1321 /// LowerCallResult - Lower the result values of a call into the
1322 /// appropriate copies out of appropriate physical registers.
1324 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1325 CallingConv::ID CallConv, bool isVarArg,
1326 const SmallVectorImpl<ISD::InputArg> &Ins,
1327 SDLoc dl, SelectionDAG &DAG,
1328 SmallVectorImpl<SDValue> &InVals,
1329 bool isThisReturn, SDValue ThisVal) const {
1331 // Assign locations to each value returned by this call.
1332 SmallVector<CCValAssign, 16> RVLocs;
1333 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
1334 *DAG.getContext(), Call);
1335 CCInfo.AnalyzeCallResult(Ins,
1336 CCAssignFnForNode(CallConv, /* Return*/ true,
1339 // Copy all of the result registers out of their specified physreg.
1340 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1341 CCValAssign VA = RVLocs[i];
1343 // Pass 'this' value directly from the argument to return value, to avoid
1344 // reg unit interference
1345 if (i == 0 && isThisReturn) {
1346 assert(!VA.needsCustom() && VA.getLocVT() == MVT::i32 &&
1347 "unexpected return calling convention register assignment");
1348 InVals.push_back(ThisVal);
1353 if (VA.needsCustom()) {
1354 // Handle f64 or half of a v2f64.
1355 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1357 Chain = Lo.getValue(1);
1358 InFlag = Lo.getValue(2);
1359 VA = RVLocs[++i]; // skip ahead to next loc
1360 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1362 Chain = Hi.getValue(1);
1363 InFlag = Hi.getValue(2);
1364 if (!Subtarget->isLittle())
1366 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1368 if (VA.getLocVT() == MVT::v2f64) {
1369 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1370 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1371 DAG.getConstant(0, MVT::i32));
1373 VA = RVLocs[++i]; // skip ahead to next loc
1374 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1375 Chain = Lo.getValue(1);
1376 InFlag = Lo.getValue(2);
1377 VA = RVLocs[++i]; // skip ahead to next loc
1378 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1379 Chain = Hi.getValue(1);
1380 InFlag = Hi.getValue(2);
1381 if (!Subtarget->isLittle())
1383 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1384 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1385 DAG.getConstant(1, MVT::i32));
1388 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
1390 Chain = Val.getValue(1);
1391 InFlag = Val.getValue(2);
1394 switch (VA.getLocInfo()) {
1395 default: llvm_unreachable("Unknown loc info!");
1396 case CCValAssign::Full: break;
1397 case CCValAssign::BCvt:
1398 Val = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), Val);
1402 InVals.push_back(Val);
1408 /// LowerMemOpCallTo - Store the argument to the stack.
1410 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
1411 SDValue StackPtr, SDValue Arg,
1412 SDLoc dl, SelectionDAG &DAG,
1413 const CCValAssign &VA,
1414 ISD::ArgFlagsTy Flags) const {
1415 unsigned LocMemOffset = VA.getLocMemOffset();
1416 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1417 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1418 return DAG.getStore(Chain, dl, Arg, PtrOff,
1419 MachinePointerInfo::getStack(LocMemOffset),
1423 void ARMTargetLowering::PassF64ArgInRegs(SDLoc dl, SelectionDAG &DAG,
1424 SDValue Chain, SDValue &Arg,
1425 RegsToPassVector &RegsToPass,
1426 CCValAssign &VA, CCValAssign &NextVA,
1428 SmallVectorImpl<SDValue> &MemOpChains,
1429 ISD::ArgFlagsTy Flags) const {
1431 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1432 DAG.getVTList(MVT::i32, MVT::i32), Arg);
1433 unsigned id = Subtarget->isLittle() ? 0 : 1;
1434 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd.getValue(id)));
1436 if (NextVA.isRegLoc())
1437 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1-id)));
1439 assert(NextVA.isMemLoc());
1440 if (!StackPtr.getNode())
1441 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1443 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1-id),
1449 /// LowerCall - Lowering a call into a callseq_start <-
1450 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
1453 ARMTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
1454 SmallVectorImpl<SDValue> &InVals) const {
1455 SelectionDAG &DAG = CLI.DAG;
1457 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
1458 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
1459 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
1460 SDValue Chain = CLI.Chain;
1461 SDValue Callee = CLI.Callee;
1462 bool &isTailCall = CLI.IsTailCall;
1463 CallingConv::ID CallConv = CLI.CallConv;
1464 bool doesNotRet = CLI.DoesNotReturn;
1465 bool isVarArg = CLI.IsVarArg;
1467 MachineFunction &MF = DAG.getMachineFunction();
1468 bool isStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1469 bool isThisReturn = false;
1470 bool isSibCall = false;
1472 // Disable tail calls if they're not supported.
1473 if (!Subtarget->supportsTailCall() || MF.getTarget().Options.DisableTailCalls)
1477 // Check if it's really possible to do a tail call.
1478 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1479 isVarArg, isStructRet, MF.getFunction()->hasStructRetAttr(),
1480 Outs, OutVals, Ins, DAG);
1481 if (!isTailCall && CLI.CS && CLI.CS->isMustTailCall())
1482 report_fatal_error("failed to perform tail call elimination on a call "
1483 "site marked musttail");
1484 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1485 // detected sibcalls.
1492 // Analyze operands of the call, assigning locations to each operand.
1493 SmallVector<CCValAssign, 16> ArgLocs;
1494 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs,
1495 *DAG.getContext(), Call);
1496 CCInfo.AnalyzeCallOperands(Outs,
1497 CCAssignFnForNode(CallConv, /* Return*/ false,
1500 // Get a count of how many bytes are to be pushed on the stack.
1501 unsigned NumBytes = CCInfo.getNextStackOffset();
1503 // For tail calls, memory operands are available in our caller's stack.
1507 // Adjust the stack pointer for the new arguments...
1508 // These operations are automatically eliminated by the prolog/epilog pass
1510 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true),
1513 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1515 RegsToPassVector RegsToPass;
1516 SmallVector<SDValue, 8> MemOpChains;
1518 // Walk the register/memloc assignments, inserting copies/loads. In the case
1519 // of tail call optimization, arguments are handled later.
1520 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1522 ++i, ++realArgIdx) {
1523 CCValAssign &VA = ArgLocs[i];
1524 SDValue Arg = OutVals[realArgIdx];
1525 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1526 bool isByVal = Flags.isByVal();
1528 // Promote the value if needed.
1529 switch (VA.getLocInfo()) {
1530 default: llvm_unreachable("Unknown loc info!");
1531 case CCValAssign::Full: break;
1532 case CCValAssign::SExt:
1533 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1535 case CCValAssign::ZExt:
1536 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1538 case CCValAssign::AExt:
1539 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1541 case CCValAssign::BCvt:
1542 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1546 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1547 if (VA.needsCustom()) {
1548 if (VA.getLocVT() == MVT::v2f64) {
1549 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1550 DAG.getConstant(0, MVT::i32));
1551 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1552 DAG.getConstant(1, MVT::i32));
1554 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1555 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1557 VA = ArgLocs[++i]; // skip ahead to next loc
1558 if (VA.isRegLoc()) {
1559 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1560 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1562 assert(VA.isMemLoc());
1564 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1565 dl, DAG, VA, Flags));
1568 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1569 StackPtr, MemOpChains, Flags);
1571 } else if (VA.isRegLoc()) {
1572 if (realArgIdx == 0 && Flags.isReturned() && Outs[0].VT == MVT::i32) {
1573 assert(VA.getLocVT() == MVT::i32 &&
1574 "unexpected calling convention register assignment");
1575 assert(!Ins.empty() && Ins[0].VT == MVT::i32 &&
1576 "unexpected use of 'returned'");
1577 isThisReturn = true;
1579 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1580 } else if (isByVal) {
1581 assert(VA.isMemLoc());
1582 unsigned offset = 0;
1584 // True if this byval aggregate will be split between registers
1586 unsigned ByValArgsCount = CCInfo.getInRegsParamsCount();
1587 unsigned CurByValIdx = CCInfo.getInRegsParamsProcessed();
1589 if (CurByValIdx < ByValArgsCount) {
1591 unsigned RegBegin, RegEnd;
1592 CCInfo.getInRegsParamInfo(CurByValIdx, RegBegin, RegEnd);
1594 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1596 for (i = 0, j = RegBegin; j < RegEnd; i++, j++) {
1597 SDValue Const = DAG.getConstant(4*i, MVT::i32);
1598 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
1599 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
1600 MachinePointerInfo(),
1601 false, false, false,
1602 DAG.InferPtrAlignment(AddArg));
1603 MemOpChains.push_back(Load.getValue(1));
1604 RegsToPass.push_back(std::make_pair(j, Load));
1607 // If parameter size outsides register area, "offset" value
1608 // helps us to calculate stack slot for remained part properly.
1609 offset = RegEnd - RegBegin;
1611 CCInfo.nextInRegsParam();
1614 if (Flags.getByValSize() > 4*offset) {
1615 unsigned LocMemOffset = VA.getLocMemOffset();
1616 SDValue StkPtrOff = DAG.getIntPtrConstant(LocMemOffset);
1617 SDValue Dst = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr,
1619 SDValue SrcOffset = DAG.getIntPtrConstant(4*offset);
1620 SDValue Src = DAG.getNode(ISD::ADD, dl, getPointerTy(), Arg, SrcOffset);
1621 SDValue SizeNode = DAG.getConstant(Flags.getByValSize() - 4*offset,
1623 SDValue AlignNode = DAG.getConstant(Flags.getByValAlign(), MVT::i32);
1625 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
1626 SDValue Ops[] = { Chain, Dst, Src, SizeNode, AlignNode};
1627 MemOpChains.push_back(DAG.getNode(ARMISD::COPY_STRUCT_BYVAL, dl, VTs,
1630 } else if (!isSibCall) {
1631 assert(VA.isMemLoc());
1633 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1634 dl, DAG, VA, Flags));
1638 if (!MemOpChains.empty())
1639 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains);
1641 // Build a sequence of copy-to-reg nodes chained together with token chain
1642 // and flag operands which copy the outgoing args into the appropriate regs.
1644 // Tail call byval lowering might overwrite argument registers so in case of
1645 // tail call optimization the copies to registers are lowered later.
1647 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1648 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1649 RegsToPass[i].second, InFlag);
1650 InFlag = Chain.getValue(1);
1653 // For tail calls lower the arguments to the 'real' stack slot.
1655 // Force all the incoming stack arguments to be loaded from the stack
1656 // before any new outgoing arguments are stored to the stack, because the
1657 // outgoing stack slots may alias the incoming argument stack slots, and
1658 // the alias isn't otherwise explicit. This is slightly more conservative
1659 // than necessary, because it means that each store effectively depends
1660 // on every argument instead of just those arguments it would clobber.
1662 // Do not flag preceding copytoreg stuff together with the following stuff.
1664 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1665 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1666 RegsToPass[i].second, InFlag);
1667 InFlag = Chain.getValue(1);
1672 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1673 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1674 // node so that legalize doesn't hack it.
1675 bool isDirect = false;
1676 bool isARMFunc = false;
1677 bool isLocalARMFunc = false;
1678 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1680 if (EnableARMLongCalls) {
1681 assert((Subtarget->isTargetWindows() ||
1682 getTargetMachine().getRelocationModel() == Reloc::Static) &&
1683 "long-calls with non-static relocation model!");
1684 // Handle a global address or an external symbol. If it's not one of
1685 // those, the target's already in a register, so we don't need to do
1687 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1688 const GlobalValue *GV = G->getGlobal();
1689 // Create a constant pool entry for the callee address
1690 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1691 ARMConstantPoolValue *CPV =
1692 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 0);
1694 // Get the address of the callee into a register
1695 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1696 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1697 Callee = DAG.getLoad(getPointerTy(), dl,
1698 DAG.getEntryNode(), CPAddr,
1699 MachinePointerInfo::getConstantPool(),
1700 false, false, false, 0);
1701 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1702 const char *Sym = S->getSymbol();
1704 // Create a constant pool entry for the callee address
1705 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1706 ARMConstantPoolValue *CPV =
1707 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1708 ARMPCLabelIndex, 0);
1709 // Get the address of the callee into a register
1710 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1711 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1712 Callee = DAG.getLoad(getPointerTy(), dl,
1713 DAG.getEntryNode(), CPAddr,
1714 MachinePointerInfo::getConstantPool(),
1715 false, false, false, 0);
1717 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1718 const GlobalValue *GV = G->getGlobal();
1720 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1721 bool isStub = (isExt && Subtarget->isTargetMachO()) &&
1722 getTargetMachine().getRelocationModel() != Reloc::Static;
1723 isARMFunc = !Subtarget->isThumb() || (isStub && !Subtarget->isMClass());
1724 // ARM call to a local ARM function is predicable.
1725 isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking);
1726 // tBX takes a register source operand.
1727 if (isStub && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1728 assert(Subtarget->isTargetMachO() && "WrapperPIC use on non-MachO?");
1729 Callee = DAG.getNode(ARMISD::WrapperPIC, dl, getPointerTy(),
1730 DAG.getTargetGlobalAddress(GV, dl, getPointerTy(),
1731 0, ARMII::MO_NONLAZY));
1732 Callee = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Callee,
1733 MachinePointerInfo::getGOT(), false, false, true, 0);
1734 } else if (Subtarget->isTargetCOFF()) {
1735 assert(Subtarget->isTargetWindows() &&
1736 "Windows is the only supported COFF target");
1737 unsigned TargetFlags = GV->hasDLLImportStorageClass()
1738 ? ARMII::MO_DLLIMPORT
1739 : ARMII::MO_NO_FLAG;
1740 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), /*Offset=*/0,
1742 if (GV->hasDLLImportStorageClass())
1743 Callee = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
1744 DAG.getNode(ARMISD::Wrapper, dl, getPointerTy(),
1745 Callee), MachinePointerInfo::getGOT(),
1746 false, false, false, 0);
1748 // On ELF targets for PIC code, direct calls should go through the PLT
1749 unsigned OpFlags = 0;
1750 if (Subtarget->isTargetELF() &&
1751 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1752 OpFlags = ARMII::MO_PLT;
1753 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
1755 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1757 bool isStub = Subtarget->isTargetMachO() &&
1758 getTargetMachine().getRelocationModel() != Reloc::Static;
1759 isARMFunc = !Subtarget->isThumb() || (isStub && !Subtarget->isMClass());
1760 // tBX takes a register source operand.
1761 const char *Sym = S->getSymbol();
1762 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1763 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1764 ARMConstantPoolValue *CPV =
1765 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1766 ARMPCLabelIndex, 4);
1767 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1768 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1769 Callee = DAG.getLoad(getPointerTy(), dl,
1770 DAG.getEntryNode(), CPAddr,
1771 MachinePointerInfo::getConstantPool(),
1772 false, false, false, 0);
1773 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1774 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1775 getPointerTy(), Callee, PICLabel);
1777 unsigned OpFlags = 0;
1778 // On ELF targets for PIC code, direct calls should go through the PLT
1779 if (Subtarget->isTargetELF() &&
1780 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1781 OpFlags = ARMII::MO_PLT;
1782 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlags);
1786 // FIXME: handle tail calls differently.
1788 bool HasMinSizeAttr = MF.getFunction()->getAttributes().hasAttribute(
1789 AttributeSet::FunctionIndex, Attribute::MinSize);
1790 if (Subtarget->isThumb()) {
1791 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1792 CallOpc = ARMISD::CALL_NOLINK;
1794 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1796 if (!isDirect && !Subtarget->hasV5TOps())
1797 CallOpc = ARMISD::CALL_NOLINK;
1798 else if (doesNotRet && isDirect && Subtarget->hasRAS() &&
1799 // Emit regular call when code size is the priority
1801 // "mov lr, pc; b _foo" to avoid confusing the RSP
1802 CallOpc = ARMISD::CALL_NOLINK;
1804 CallOpc = isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL;
1807 std::vector<SDValue> Ops;
1808 Ops.push_back(Chain);
1809 Ops.push_back(Callee);
1811 // Add argument registers to the end of the list so that they are known live
1813 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1814 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1815 RegsToPass[i].second.getValueType()));
1817 // Add a register mask operand representing the call-preserved registers.
1819 const uint32_t *Mask;
1820 const ARMBaseRegisterInfo *ARI = Subtarget->getRegisterInfo();
1822 // For 'this' returns, use the R0-preserving mask if applicable
1823 Mask = ARI->getThisReturnPreservedMask(CallConv);
1825 // Set isThisReturn to false if the calling convention is not one that
1826 // allows 'returned' to be modeled in this way, so LowerCallResult does
1827 // not try to pass 'this' straight through
1828 isThisReturn = false;
1829 Mask = ARI->getCallPreservedMask(CallConv);
1832 Mask = ARI->getCallPreservedMask(CallConv);
1834 assert(Mask && "Missing call preserved mask for calling convention");
1835 Ops.push_back(DAG.getRegisterMask(Mask));
1838 if (InFlag.getNode())
1839 Ops.push_back(InFlag);
1841 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
1843 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, Ops);
1845 // Returns a chain and a flag for retval copy to use.
1846 Chain = DAG.getNode(CallOpc, dl, NodeTys, Ops);
1847 InFlag = Chain.getValue(1);
1849 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1850 DAG.getIntPtrConstant(0, true), InFlag, dl);
1852 InFlag = Chain.getValue(1);
1854 // Handle result values, copying them out of physregs into vregs that we
1856 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins, dl, DAG,
1857 InVals, isThisReturn,
1858 isThisReturn ? OutVals[0] : SDValue());
1861 /// HandleByVal - Every parameter *after* a byval parameter is passed
1862 /// on the stack. Remember the next parameter register to allocate,
1863 /// and then confiscate the rest of the parameter registers to insure
1866 ARMTargetLowering::HandleByVal(
1867 CCState *State, unsigned &size, unsigned Align) const {
1868 unsigned reg = State->AllocateReg(GPRArgRegs, 4);
1869 assert((State->getCallOrPrologue() == Prologue ||
1870 State->getCallOrPrologue() == Call) &&
1871 "unhandled ParmContext");
1873 if ((ARM::R0 <= reg) && (reg <= ARM::R3)) {
1874 if (Subtarget->isAAPCS_ABI() && Align > 4) {
1875 unsigned AlignInRegs = Align / 4;
1876 unsigned Waste = (ARM::R4 - reg) % AlignInRegs;
1877 for (unsigned i = 0; i < Waste; ++i)
1878 reg = State->AllocateReg(GPRArgRegs, 4);
1881 unsigned excess = 4 * (ARM::R4 - reg);
1883 // Special case when NSAA != SP and parameter size greater than size of
1884 // all remained GPR regs. In that case we can't split parameter, we must
1885 // send it to stack. We also must set NCRN to R4, so waste all
1886 // remained registers.
1887 const unsigned NSAAOffset = State->getNextStackOffset();
1888 if (Subtarget->isAAPCS_ABI() && NSAAOffset != 0 && size > excess) {
1889 while (State->AllocateReg(GPRArgRegs, 4))
1894 // First register for byval parameter is the first register that wasn't
1895 // allocated before this method call, so it would be "reg".
1896 // If parameter is small enough to be saved in range [reg, r4), then
1897 // the end (first after last) register would be reg + param-size-in-regs,
1898 // else parameter would be splitted between registers and stack,
1899 // end register would be r4 in this case.
1900 unsigned ByValRegBegin = reg;
1901 unsigned ByValRegEnd = (size < excess) ? reg + size/4 : (unsigned)ARM::R4;
1902 State->addInRegsParamInfo(ByValRegBegin, ByValRegEnd);
1903 // Note, first register is allocated in the beginning of function already,
1904 // allocate remained amount of registers we need.
1905 for (unsigned i = reg+1; i != ByValRegEnd; ++i)
1906 State->AllocateReg(GPRArgRegs, 4);
1907 // A byval parameter that is split between registers and memory needs its
1908 // size truncated here.
1909 // In the case where the entire structure fits in registers, we set the
1910 // size in memory to zero.
1919 /// MatchingStackOffset - Return true if the given stack call argument is
1920 /// already available in the same position (relatively) of the caller's
1921 /// incoming argument stack.
1923 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1924 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1925 const TargetInstrInfo *TII) {
1926 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1928 if (Arg.getOpcode() == ISD::CopyFromReg) {
1929 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1930 if (!TargetRegisterInfo::isVirtualRegister(VR))
1932 MachineInstr *Def = MRI->getVRegDef(VR);
1935 if (!Flags.isByVal()) {
1936 if (!TII->isLoadFromStackSlot(Def, FI))
1941 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
1942 if (Flags.isByVal())
1943 // ByVal argument is passed in as a pointer but it's now being
1944 // dereferenced. e.g.
1945 // define @foo(%struct.X* %A) {
1946 // tail call @bar(%struct.X* byval %A)
1949 SDValue Ptr = Ld->getBasePtr();
1950 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
1953 FI = FINode->getIndex();
1957 assert(FI != INT_MAX);
1958 if (!MFI->isFixedObjectIndex(FI))
1960 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
1963 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
1964 /// for tail call optimization. Targets which want to do tail call
1965 /// optimization should implement this function.
1967 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1968 CallingConv::ID CalleeCC,
1970 bool isCalleeStructRet,
1971 bool isCallerStructRet,
1972 const SmallVectorImpl<ISD::OutputArg> &Outs,
1973 const SmallVectorImpl<SDValue> &OutVals,
1974 const SmallVectorImpl<ISD::InputArg> &Ins,
1975 SelectionDAG& DAG) const {
1976 const Function *CallerF = DAG.getMachineFunction().getFunction();
1977 CallingConv::ID CallerCC = CallerF->getCallingConv();
1978 bool CCMatch = CallerCC == CalleeCC;
1980 // Look for obvious safe cases to perform tail call optimization that do not
1981 // require ABI changes. This is what gcc calls sibcall.
1983 // Do not sibcall optimize vararg calls unless the call site is not passing
1985 if (isVarArg && !Outs.empty())
1988 // Exception-handling functions need a special set of instructions to indicate
1989 // a return to the hardware. Tail-calling another function would probably
1991 if (CallerF->hasFnAttribute("interrupt"))
1994 // Also avoid sibcall optimization if either caller or callee uses struct
1995 // return semantics.
1996 if (isCalleeStructRet || isCallerStructRet)
1999 // FIXME: Completely disable sibcall for Thumb1 since Thumb1RegisterInfo::
2000 // emitEpilogue is not ready for them. Thumb tail calls also use t2B, as
2001 // the Thumb1 16-bit unconditional branch doesn't have sufficient relocation
2002 // support in the assembler and linker to be used. This would need to be
2003 // fixed to fully support tail calls in Thumb1.
2005 // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take
2006 // LR. This means if we need to reload LR, it takes an extra instructions,
2007 // which outweighs the value of the tail call; but here we don't know yet
2008 // whether LR is going to be used. Probably the right approach is to
2009 // generate the tail call here and turn it back into CALL/RET in
2010 // emitEpilogue if LR is used.
2012 // Thumb1 PIC calls to external symbols use BX, so they can be tail calls,
2013 // but we need to make sure there are enough registers; the only valid
2014 // registers are the 4 used for parameters. We don't currently do this
2016 if (Subtarget->isThumb1Only())
2019 // Externally-defined functions with weak linkage should not be
2020 // tail-called on ARM when the OS does not support dynamic
2021 // pre-emption of symbols, as the AAELF spec requires normal calls
2022 // to undefined weak functions to be replaced with a NOP or jump to the
2023 // next instruction. The behaviour of branch instructions in this
2024 // situation (as used for tail calls) is implementation-defined, so we
2025 // cannot rely on the linker replacing the tail call with a return.
2026 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2027 const GlobalValue *GV = G->getGlobal();
2028 const Triple TT(getTargetMachine().getTargetTriple());
2029 if (GV->hasExternalWeakLinkage() &&
2030 (!TT.isOSWindows() || TT.isOSBinFormatELF() || TT.isOSBinFormatMachO()))
2034 // If the calling conventions do not match, then we'd better make sure the
2035 // results are returned in the same way as what the caller expects.
2037 SmallVector<CCValAssign, 16> RVLocs1;
2038 ARMCCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(), RVLocs1,
2039 *DAG.getContext(), Call);
2040 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
2042 SmallVector<CCValAssign, 16> RVLocs2;
2043 ARMCCState CCInfo2(CallerCC, false, DAG.getMachineFunction(), RVLocs2,
2044 *DAG.getContext(), Call);
2045 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
2047 if (RVLocs1.size() != RVLocs2.size())
2049 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
2050 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
2052 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
2054 if (RVLocs1[i].isRegLoc()) {
2055 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
2058 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
2064 // If Caller's vararg or byval argument has been split between registers and
2065 // stack, do not perform tail call, since part of the argument is in caller's
2067 const ARMFunctionInfo *AFI_Caller = DAG.getMachineFunction().
2068 getInfo<ARMFunctionInfo>();
2069 if (AFI_Caller->getArgRegsSaveSize())
2072 // If the callee takes no arguments then go on to check the results of the
2074 if (!Outs.empty()) {
2075 // Check if stack adjustment is needed. For now, do not do this if any
2076 // argument is passed on the stack.
2077 SmallVector<CCValAssign, 16> ArgLocs;
2078 ARMCCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
2079 *DAG.getContext(), Call);
2080 CCInfo.AnalyzeCallOperands(Outs,
2081 CCAssignFnForNode(CalleeCC, false, isVarArg));
2082 if (CCInfo.getNextStackOffset()) {
2083 MachineFunction &MF = DAG.getMachineFunction();
2085 // Check if the arguments are already laid out in the right way as
2086 // the caller's fixed stack objects.
2087 MachineFrameInfo *MFI = MF.getFrameInfo();
2088 const MachineRegisterInfo *MRI = &MF.getRegInfo();
2089 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
2090 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
2092 ++i, ++realArgIdx) {
2093 CCValAssign &VA = ArgLocs[i];
2094 EVT RegVT = VA.getLocVT();
2095 SDValue Arg = OutVals[realArgIdx];
2096 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
2097 if (VA.getLocInfo() == CCValAssign::Indirect)
2099 if (VA.needsCustom()) {
2100 // f64 and vector types are split into multiple registers or
2101 // register/stack-slot combinations. The types will not match
2102 // the registers; give up on memory f64 refs until we figure
2103 // out what to do about this.
2106 if (!ArgLocs[++i].isRegLoc())
2108 if (RegVT == MVT::v2f64) {
2109 if (!ArgLocs[++i].isRegLoc())
2111 if (!ArgLocs[++i].isRegLoc())
2114 } else if (!VA.isRegLoc()) {
2115 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
2127 ARMTargetLowering::CanLowerReturn(CallingConv::ID CallConv,
2128 MachineFunction &MF, bool isVarArg,
2129 const SmallVectorImpl<ISD::OutputArg> &Outs,
2130 LLVMContext &Context) const {
2131 SmallVector<CCValAssign, 16> RVLocs;
2132 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, Context);
2133 return CCInfo.CheckReturn(Outs, CCAssignFnForNode(CallConv, /*Return=*/true,
2137 static SDValue LowerInterruptReturn(SmallVectorImpl<SDValue> &RetOps,
2138 SDLoc DL, SelectionDAG &DAG) {
2139 const MachineFunction &MF = DAG.getMachineFunction();
2140 const Function *F = MF.getFunction();
2142 StringRef IntKind = F->getFnAttribute("interrupt").getValueAsString();
2144 // See ARM ARM v7 B1.8.3. On exception entry LR is set to a possibly offset
2145 // version of the "preferred return address". These offsets affect the return
2146 // instruction if this is a return from PL1 without hypervisor extensions.
2147 // IRQ/FIQ: +4 "subs pc, lr, #4"
2148 // SWI: 0 "subs pc, lr, #0"
2149 // ABORT: +4 "subs pc, lr, #4"
2150 // UNDEF: +4/+2 "subs pc, lr, #0"
2151 // UNDEF varies depending on where the exception came from ARM or Thumb
2152 // mode. Alongside GCC, we throw our hands up in disgust and pretend it's 0.
2155 if (IntKind == "" || IntKind == "IRQ" || IntKind == "FIQ" ||
2158 else if (IntKind == "SWI" || IntKind == "UNDEF")
2161 report_fatal_error("Unsupported interrupt attribute. If present, value "
2162 "must be one of: IRQ, FIQ, SWI, ABORT or UNDEF");
2164 RetOps.insert(RetOps.begin() + 1, DAG.getConstant(LROffset, MVT::i32, false));
2166 return DAG.getNode(ARMISD::INTRET_FLAG, DL, MVT::Other, RetOps);
2170 ARMTargetLowering::LowerReturn(SDValue Chain,
2171 CallingConv::ID CallConv, bool isVarArg,
2172 const SmallVectorImpl<ISD::OutputArg> &Outs,
2173 const SmallVectorImpl<SDValue> &OutVals,
2174 SDLoc dl, SelectionDAG &DAG) const {
2176 // CCValAssign - represent the assignment of the return value to a location.
2177 SmallVector<CCValAssign, 16> RVLocs;
2179 // CCState - Info about the registers and stack slots.
2180 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
2181 *DAG.getContext(), Call);
2183 // Analyze outgoing return values.
2184 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
2188 SmallVector<SDValue, 4> RetOps;
2189 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
2190 bool isLittleEndian = Subtarget->isLittle();
2192 MachineFunction &MF = DAG.getMachineFunction();
2193 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2194 AFI->setReturnRegsCount(RVLocs.size());
2196 // Copy the result values into the output registers.
2197 for (unsigned i = 0, realRVLocIdx = 0;
2199 ++i, ++realRVLocIdx) {
2200 CCValAssign &VA = RVLocs[i];
2201 assert(VA.isRegLoc() && "Can only return in registers!");
2203 SDValue Arg = OutVals[realRVLocIdx];
2205 switch (VA.getLocInfo()) {
2206 default: llvm_unreachable("Unknown loc info!");
2207 case CCValAssign::Full: break;
2208 case CCValAssign::BCvt:
2209 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
2213 if (VA.needsCustom()) {
2214 if (VA.getLocVT() == MVT::v2f64) {
2215 // Extract the first half and return it in two registers.
2216 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
2217 DAG.getConstant(0, MVT::i32));
2218 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
2219 DAG.getVTList(MVT::i32, MVT::i32), Half);
2221 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2222 HalfGPRs.getValue(isLittleEndian ? 0 : 1),
2224 Flag = Chain.getValue(1);
2225 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2226 VA = RVLocs[++i]; // skip ahead to next loc
2227 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2228 HalfGPRs.getValue(isLittleEndian ? 1 : 0),
2230 Flag = Chain.getValue(1);
2231 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2232 VA = RVLocs[++i]; // skip ahead to next loc
2234 // Extract the 2nd half and fall through to handle it as an f64 value.
2235 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
2236 DAG.getConstant(1, MVT::i32));
2238 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
2240 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
2241 DAG.getVTList(MVT::i32, MVT::i32), Arg);
2242 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2243 fmrrd.getValue(isLittleEndian ? 0 : 1),
2245 Flag = Chain.getValue(1);
2246 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2247 VA = RVLocs[++i]; // skip ahead to next loc
2248 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2249 fmrrd.getValue(isLittleEndian ? 1 : 0),
2252 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
2254 // Guarantee that all emitted copies are
2255 // stuck together, avoiding something bad.
2256 Flag = Chain.getValue(1);
2257 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2260 // Update chain and glue.
2263 RetOps.push_back(Flag);
2265 // CPUs which aren't M-class use a special sequence to return from
2266 // exceptions (roughly, any instruction setting pc and cpsr simultaneously,
2267 // though we use "subs pc, lr, #N").
2269 // M-class CPUs actually use a normal return sequence with a special
2270 // (hardware-provided) value in LR, so the normal code path works.
2271 if (DAG.getMachineFunction().getFunction()->hasFnAttribute("interrupt") &&
2272 !Subtarget->isMClass()) {
2273 if (Subtarget->isThumb1Only())
2274 report_fatal_error("interrupt attribute is not supported in Thumb1");
2275 return LowerInterruptReturn(RetOps, dl, DAG);
2278 return DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, RetOps);
2281 bool ARMTargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
2282 if (N->getNumValues() != 1)
2284 if (!N->hasNUsesOfValue(1, 0))
2287 SDValue TCChain = Chain;
2288 SDNode *Copy = *N->use_begin();
2289 if (Copy->getOpcode() == ISD::CopyToReg) {
2290 // If the copy has a glue operand, we conservatively assume it isn't safe to
2291 // perform a tail call.
2292 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
2294 TCChain = Copy->getOperand(0);
2295 } else if (Copy->getOpcode() == ARMISD::VMOVRRD) {
2296 SDNode *VMov = Copy;
2297 // f64 returned in a pair of GPRs.
2298 SmallPtrSet<SDNode*, 2> Copies;
2299 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
2301 if (UI->getOpcode() != ISD::CopyToReg)
2305 if (Copies.size() > 2)
2308 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
2310 SDValue UseChain = UI->getOperand(0);
2311 if (Copies.count(UseChain.getNode()))
2315 // We are at the top of this chain.
2316 // If the copy has a glue operand, we conservatively assume it
2317 // isn't safe to perform a tail call.
2318 if (UI->getOperand(UI->getNumOperands()-1).getValueType() == MVT::Glue)
2324 } else if (Copy->getOpcode() == ISD::BITCAST) {
2325 // f32 returned in a single GPR.
2326 if (!Copy->hasOneUse())
2328 Copy = *Copy->use_begin();
2329 if (Copy->getOpcode() != ISD::CopyToReg || !Copy->hasNUsesOfValue(1, 0))
2331 // If the copy has a glue operand, we conservatively assume it isn't safe to
2332 // perform a tail call.
2333 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
2335 TCChain = Copy->getOperand(0);
2340 bool HasRet = false;
2341 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
2343 if (UI->getOpcode() != ARMISD::RET_FLAG &&
2344 UI->getOpcode() != ARMISD::INTRET_FLAG)
2356 bool ARMTargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
2357 if (!Subtarget->supportsTailCall())
2360 if (!CI->isTailCall() || getTargetMachine().Options.DisableTailCalls)
2363 return !Subtarget->isThumb1Only();
2366 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
2367 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
2368 // one of the above mentioned nodes. It has to be wrapped because otherwise
2369 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
2370 // be used to form addressing mode. These wrapped nodes will be selected
2372 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
2373 EVT PtrVT = Op.getValueType();
2374 // FIXME there is no actual debug info here
2376 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
2378 if (CP->isMachineConstantPoolEntry())
2379 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
2380 CP->getAlignment());
2382 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
2383 CP->getAlignment());
2384 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
2387 unsigned ARMTargetLowering::getJumpTableEncoding() const {
2388 return MachineJumpTableInfo::EK_Inline;
2391 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
2392 SelectionDAG &DAG) const {
2393 MachineFunction &MF = DAG.getMachineFunction();
2394 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2395 unsigned ARMPCLabelIndex = 0;
2397 EVT PtrVT = getPointerTy();
2398 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
2399 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2401 if (RelocM == Reloc::Static) {
2402 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
2404 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2405 ARMPCLabelIndex = AFI->createPICLabelUId();
2406 ARMConstantPoolValue *CPV =
2407 ARMConstantPoolConstant::Create(BA, ARMPCLabelIndex,
2408 ARMCP::CPBlockAddress, PCAdj);
2409 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2411 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
2412 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
2413 MachinePointerInfo::getConstantPool(),
2414 false, false, false, 0);
2415 if (RelocM == Reloc::Static)
2417 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2418 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
2421 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
2423 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
2424 SelectionDAG &DAG) const {
2426 EVT PtrVT = getPointerTy();
2427 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2428 MachineFunction &MF = DAG.getMachineFunction();
2429 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2430 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2431 ARMConstantPoolValue *CPV =
2432 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2433 ARMCP::CPValue, PCAdj, ARMCP::TLSGD, true);
2434 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2435 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
2436 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
2437 MachinePointerInfo::getConstantPool(),
2438 false, false, false, 0);
2439 SDValue Chain = Argument.getValue(1);
2441 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2442 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
2444 // call __tls_get_addr.
2447 Entry.Node = Argument;
2448 Entry.Ty = (Type *) Type::getInt32Ty(*DAG.getContext());
2449 Args.push_back(Entry);
2451 // FIXME: is there useful debug info available here?
2452 TargetLowering::CallLoweringInfo CLI(DAG);
2453 CLI.setDebugLoc(dl).setChain(Chain)
2454 .setCallee(CallingConv::C, Type::getInt32Ty(*DAG.getContext()),
2455 DAG.getExternalSymbol("__tls_get_addr", PtrVT), std::move(Args),
2458 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
2459 return CallResult.first;
2462 // Lower ISD::GlobalTLSAddress using the "initial exec" or
2463 // "local exec" model.
2465 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
2467 TLSModel::Model model) const {
2468 const GlobalValue *GV = GA->getGlobal();
2471 SDValue Chain = DAG.getEntryNode();
2472 EVT PtrVT = getPointerTy();
2473 // Get the Thread Pointer
2474 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2476 if (model == TLSModel::InitialExec) {
2477 MachineFunction &MF = DAG.getMachineFunction();
2478 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2479 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2480 // Initial exec model.
2481 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2482 ARMConstantPoolValue *CPV =
2483 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2484 ARMCP::CPValue, PCAdj, ARMCP::GOTTPOFF,
2486 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2487 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2488 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2489 MachinePointerInfo::getConstantPool(),
2490 false, false, false, 0);
2491 Chain = Offset.getValue(1);
2493 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2494 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
2496 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2497 MachinePointerInfo::getConstantPool(),
2498 false, false, false, 0);
2501 assert(model == TLSModel::LocalExec);
2502 ARMConstantPoolValue *CPV =
2503 ARMConstantPoolConstant::Create(GV, ARMCP::TPOFF);
2504 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2505 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2506 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2507 MachinePointerInfo::getConstantPool(),
2508 false, false, false, 0);
2511 // The address of the thread local variable is the add of the thread
2512 // pointer with the offset of the variable.
2513 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
2517 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
2518 // TODO: implement the "local dynamic" model
2519 assert(Subtarget->isTargetELF() &&
2520 "TLS not implemented for non-ELF targets");
2521 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
2523 TLSModel::Model model = getTargetMachine().getTLSModel(GA->getGlobal());
2526 case TLSModel::GeneralDynamic:
2527 case TLSModel::LocalDynamic:
2528 return LowerToTLSGeneralDynamicModel(GA, DAG);
2529 case TLSModel::InitialExec:
2530 case TLSModel::LocalExec:
2531 return LowerToTLSExecModels(GA, DAG, model);
2533 llvm_unreachable("bogus TLS model");
2536 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
2537 SelectionDAG &DAG) const {
2538 EVT PtrVT = getPointerTy();
2540 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2541 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2542 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
2543 ARMConstantPoolValue *CPV =
2544 ARMConstantPoolConstant::Create(GV,
2545 UseGOTOFF ? ARMCP::GOTOFF : ARMCP::GOT);
2546 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2547 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2548 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
2550 MachinePointerInfo::getConstantPool(),
2551 false, false, false, 0);
2552 SDValue Chain = Result.getValue(1);
2553 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
2554 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
2556 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
2557 MachinePointerInfo::getGOT(),
2558 false, false, false, 0);
2562 // If we have T2 ops, we can materialize the address directly via movt/movw
2563 // pair. This is always cheaper.
2564 if (Subtarget->useMovt(DAG.getMachineFunction())) {
2566 // FIXME: Once remat is capable of dealing with instructions with register
2567 // operands, expand this into two nodes.
2568 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2569 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2571 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2572 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2573 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2574 MachinePointerInfo::getConstantPool(),
2575 false, false, false, 0);
2579 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
2580 SelectionDAG &DAG) const {
2581 EVT PtrVT = getPointerTy();
2583 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2584 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2586 if (Subtarget->useMovt(DAG.getMachineFunction()))
2589 // FIXME: Once remat is capable of dealing with instructions with register
2590 // operands, expand this into multiple nodes
2592 RelocM == Reloc::PIC_ ? ARMISD::WrapperPIC : ARMISD::Wrapper;
2594 SDValue G = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, ARMII::MO_NONLAZY);
2595 SDValue Result = DAG.getNode(Wrapper, dl, PtrVT, G);
2597 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2598 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
2599 MachinePointerInfo::getGOT(), false, false, false, 0);
2603 SDValue ARMTargetLowering::LowerGlobalAddressWindows(SDValue Op,
2604 SelectionDAG &DAG) const {
2605 assert(Subtarget->isTargetWindows() && "non-Windows COFF is not supported");
2606 assert(Subtarget->useMovt(DAG.getMachineFunction()) &&
2607 "Windows on ARM expects to use movw/movt");
2609 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2610 const ARMII::TOF TargetFlags =
2611 (GV->hasDLLImportStorageClass() ? ARMII::MO_DLLIMPORT : ARMII::MO_NO_FLAG);
2612 EVT PtrVT = getPointerTy();
2618 // FIXME: Once remat is capable of dealing with instructions with register
2619 // operands, expand this into two nodes.
2620 Result = DAG.getNode(ARMISD::Wrapper, DL, PtrVT,
2621 DAG.getTargetGlobalAddress(GV, DL, PtrVT, /*Offset=*/0,
2623 if (GV->hasDLLImportStorageClass())
2624 Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), Result,
2625 MachinePointerInfo::getGOT(), false, false, false, 0);
2629 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
2630 SelectionDAG &DAG) const {
2631 assert(Subtarget->isTargetELF() &&
2632 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
2633 MachineFunction &MF = DAG.getMachineFunction();
2634 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2635 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2636 EVT PtrVT = getPointerTy();
2638 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2639 ARMConstantPoolValue *CPV =
2640 ARMConstantPoolSymbol::Create(*DAG.getContext(), "_GLOBAL_OFFSET_TABLE_",
2641 ARMPCLabelIndex, PCAdj);
2642 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2643 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2644 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2645 MachinePointerInfo::getConstantPool(),
2646 false, false, false, 0);
2647 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2648 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2652 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
2654 SDValue Val = DAG.getConstant(0, MVT::i32);
2655 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl,
2656 DAG.getVTList(MVT::i32, MVT::Other), Op.getOperand(0),
2657 Op.getOperand(1), Val);
2661 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
2663 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
2664 Op.getOperand(1), DAG.getConstant(0, MVT::i32));
2668 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
2669 const ARMSubtarget *Subtarget) const {
2670 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2673 default: return SDValue(); // Don't custom lower most intrinsics.
2674 case Intrinsic::arm_rbit: {
2675 assert(Op.getOperand(1).getValueType() == MVT::i32 &&
2676 "RBIT intrinsic must have i32 type!");
2677 return DAG.getNode(ARMISD::RBIT, dl, MVT::i32, Op.getOperand(1));
2679 case Intrinsic::arm_thread_pointer: {
2680 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2681 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2683 case Intrinsic::eh_sjlj_lsda: {
2684 MachineFunction &MF = DAG.getMachineFunction();
2685 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2686 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2687 EVT PtrVT = getPointerTy();
2688 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2690 unsigned PCAdj = (RelocM != Reloc::PIC_)
2691 ? 0 : (Subtarget->isThumb() ? 4 : 8);
2692 ARMConstantPoolValue *CPV =
2693 ARMConstantPoolConstant::Create(MF.getFunction(), ARMPCLabelIndex,
2694 ARMCP::CPLSDA, PCAdj);
2695 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2696 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2698 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2699 MachinePointerInfo::getConstantPool(),
2700 false, false, false, 0);
2702 if (RelocM == Reloc::PIC_) {
2703 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2704 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2708 case Intrinsic::arm_neon_vmulls:
2709 case Intrinsic::arm_neon_vmullu: {
2710 unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vmulls)
2711 ? ARMISD::VMULLs : ARMISD::VMULLu;
2712 return DAG.getNode(NewOpc, SDLoc(Op), Op.getValueType(),
2713 Op.getOperand(1), Op.getOperand(2));
2718 static SDValue LowerATOMIC_FENCE(SDValue Op, SelectionDAG &DAG,
2719 const ARMSubtarget *Subtarget) {
2720 // FIXME: handle "fence singlethread" more efficiently.
2722 if (!Subtarget->hasDataBarrier()) {
2723 // Some ARMv6 cpus can support data barriers with an mcr instruction.
2724 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2726 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2727 "Unexpected ISD::ATOMIC_FENCE encountered. Should be libcall!");
2728 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2729 DAG.getConstant(0, MVT::i32));
2732 ConstantSDNode *OrdN = cast<ConstantSDNode>(Op.getOperand(1));
2733 AtomicOrdering Ord = static_cast<AtomicOrdering>(OrdN->getZExtValue());
2734 ARM_MB::MemBOpt Domain = ARM_MB::ISH;
2735 if (Subtarget->isMClass()) {
2736 // Only a full system barrier exists in the M-class architectures.
2737 Domain = ARM_MB::SY;
2738 } else if (Subtarget->isSwift() && Ord == Release) {
2739 // Swift happens to implement ISHST barriers in a way that's compatible with
2740 // Release semantics but weaker than ISH so we'd be fools not to use
2741 // it. Beware: other processors probably don't!
2742 Domain = ARM_MB::ISHST;
2745 return DAG.getNode(ISD::INTRINSIC_VOID, dl, MVT::Other, Op.getOperand(0),
2746 DAG.getConstant(Intrinsic::arm_dmb, MVT::i32),
2747 DAG.getConstant(Domain, MVT::i32));
2750 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG,
2751 const ARMSubtarget *Subtarget) {
2752 // ARM pre v5TE and Thumb1 does not have preload instructions.
2753 if (!(Subtarget->isThumb2() ||
2754 (!Subtarget->isThumb1Only() && Subtarget->hasV5TEOps())))
2755 // Just preserve the chain.
2756 return Op.getOperand(0);
2759 unsigned isRead = ~cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() & 1;
2761 (!Subtarget->hasV7Ops() || !Subtarget->hasMPExtension()))
2762 // ARMv7 with MP extension has PLDW.
2763 return Op.getOperand(0);
2765 unsigned isData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
2766 if (Subtarget->isThumb()) {
2768 isRead = ~isRead & 1;
2769 isData = ~isData & 1;
2772 return DAG.getNode(ARMISD::PRELOAD, dl, MVT::Other, Op.getOperand(0),
2773 Op.getOperand(1), DAG.getConstant(isRead, MVT::i32),
2774 DAG.getConstant(isData, MVT::i32));
2777 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
2778 MachineFunction &MF = DAG.getMachineFunction();
2779 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
2781 // vastart just stores the address of the VarArgsFrameIndex slot into the
2782 // memory location argument.
2784 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2785 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2786 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
2787 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
2788 MachinePointerInfo(SV), false, false, 0);
2792 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
2793 SDValue &Root, SelectionDAG &DAG,
2795 MachineFunction &MF = DAG.getMachineFunction();
2796 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2798 const TargetRegisterClass *RC;
2799 if (AFI->isThumb1OnlyFunction())
2800 RC = &ARM::tGPRRegClass;
2802 RC = &ARM::GPRRegClass;
2804 // Transform the arguments stored in physical registers into virtual ones.
2805 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2806 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2809 if (NextVA.isMemLoc()) {
2810 MachineFrameInfo *MFI = MF.getFrameInfo();
2811 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true);
2813 // Create load node to retrieve arguments from the stack.
2814 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2815 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
2816 MachinePointerInfo::getFixedStack(FI),
2817 false, false, false, 0);
2819 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
2820 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2822 if (!Subtarget->isLittle())
2823 std::swap (ArgValue, ArgValue2);
2824 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
2828 ARMTargetLowering::computeRegArea(CCState &CCInfo, MachineFunction &MF,
2829 unsigned InRegsParamRecordIdx,
2831 unsigned &ArgRegsSize,
2832 unsigned &ArgRegsSaveSize)
2835 if (InRegsParamRecordIdx < CCInfo.getInRegsParamsCount()) {
2836 unsigned RBegin, REnd;
2837 CCInfo.getInRegsParamInfo(InRegsParamRecordIdx, RBegin, REnd);
2838 NumGPRs = REnd - RBegin;
2840 unsigned int firstUnalloced;
2841 firstUnalloced = CCInfo.getFirstUnallocated(GPRArgRegs,
2842 sizeof(GPRArgRegs) /
2843 sizeof(GPRArgRegs[0]));
2844 NumGPRs = (firstUnalloced <= 3) ? (4 - firstUnalloced) : 0;
2847 unsigned Align = Subtarget->getFrameLowering()->getStackAlignment();
2848 ArgRegsSize = NumGPRs * 4;
2850 // If parameter is split between stack and GPRs...
2851 if (NumGPRs && Align > 4 &&
2852 (ArgRegsSize < ArgSize ||
2853 InRegsParamRecordIdx >= CCInfo.getInRegsParamsCount())) {
2854 // Add padding for part of param recovered from GPRs. For example,
2855 // if Align == 8, its last byte must be at address K*8 - 1.
2856 // We need to do it, since remained (stack) part of parameter has
2857 // stack alignment, and we need to "attach" "GPRs head" without gaps
2860 // |---- 8 bytes block ----| |---- 8 bytes block ----| |---- 8 bytes...
2861 // [ [padding] [GPRs head] ] [ Tail passed via stack ....
2863 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2865 OffsetToAlignment(ArgRegsSize + AFI->getArgRegsSaveSize(), Align);
2866 ArgRegsSaveSize = ArgRegsSize + Padding;
2868 // We don't need to extend regs save size for byval parameters if they
2869 // are passed via GPRs only.
2870 ArgRegsSaveSize = ArgRegsSize;
2873 // The remaining GPRs hold either the beginning of variable-argument
2874 // data, or the beginning of an aggregate passed by value (usually
2875 // byval). Either way, we allocate stack slots adjacent to the data
2876 // provided by our caller, and store the unallocated registers there.
2877 // If this is a variadic function, the va_list pointer will begin with
2878 // these values; otherwise, this reassembles a (byval) structure that
2879 // was split between registers and memory.
2880 // Return: The frame index registers were stored into.
2882 ARMTargetLowering::StoreByValRegs(CCState &CCInfo, SelectionDAG &DAG,
2883 SDLoc dl, SDValue &Chain,
2884 const Value *OrigArg,
2885 unsigned InRegsParamRecordIdx,
2886 unsigned OffsetFromOrigArg,
2890 unsigned ByValStoreOffset,
2891 unsigned TotalArgRegsSaveSize) const {
2893 // Currently, two use-cases possible:
2894 // Case #1. Non-var-args function, and we meet first byval parameter.
2895 // Setup first unallocated register as first byval register;
2896 // eat all remained registers
2897 // (these two actions are performed by HandleByVal method).
2898 // Then, here, we initialize stack frame with
2899 // "store-reg" instructions.
2900 // Case #2. Var-args function, that doesn't contain byval parameters.
2901 // The same: eat all remained unallocated registers,
2902 // initialize stack frame.
2904 MachineFunction &MF = DAG.getMachineFunction();
2905 MachineFrameInfo *MFI = MF.getFrameInfo();
2906 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2907 unsigned firstRegToSaveIndex, lastRegToSaveIndex;
2908 unsigned RBegin, REnd;
2909 if (InRegsParamRecordIdx < CCInfo.getInRegsParamsCount()) {
2910 CCInfo.getInRegsParamInfo(InRegsParamRecordIdx, RBegin, REnd);
2911 firstRegToSaveIndex = RBegin - ARM::R0;
2912 lastRegToSaveIndex = REnd - ARM::R0;
2914 firstRegToSaveIndex = CCInfo.getFirstUnallocated
2915 (GPRArgRegs, array_lengthof(GPRArgRegs));
2916 lastRegToSaveIndex = 4;
2919 unsigned ArgRegsSize, ArgRegsSaveSize;
2920 computeRegArea(CCInfo, MF, InRegsParamRecordIdx, ArgSize,
2921 ArgRegsSize, ArgRegsSaveSize);
2923 // Store any by-val regs to their spots on the stack so that they may be
2924 // loaded by deferencing the result of formal parameter pointer or va_next.
2925 // Note: once stack area for byval/varargs registers
2926 // was initialized, it can't be initialized again.
2927 if (ArgRegsSaveSize) {
2928 unsigned Padding = ArgRegsSaveSize - ArgRegsSize;
2931 assert(AFI->getStoredByValParamsPadding() == 0 &&
2932 "The only parameter may be padded.");
2933 AFI->setStoredByValParamsPadding(Padding);
2936 int FrameIndex = MFI->CreateFixedObject(ArgRegsSaveSize,
2939 (int64_t)TotalArgRegsSaveSize,
2941 SDValue FIN = DAG.getFrameIndex(FrameIndex, getPointerTy());
2943 MFI->CreateFixedObject(Padding,
2944 ArgOffset + ByValStoreOffset -
2945 (int64_t)ArgRegsSaveSize,
2949 SmallVector<SDValue, 4> MemOps;
2950 for (unsigned i = 0; firstRegToSaveIndex < lastRegToSaveIndex;
2951 ++firstRegToSaveIndex, ++i) {
2952 const TargetRegisterClass *RC;
2953 if (AFI->isThumb1OnlyFunction())
2954 RC = &ARM::tGPRRegClass;
2956 RC = &ARM::GPRRegClass;
2958 unsigned VReg = MF.addLiveIn(GPRArgRegs[firstRegToSaveIndex], RC);
2959 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2961 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2962 MachinePointerInfo(OrigArg, OffsetFromOrigArg + 4*i),
2964 MemOps.push_back(Store);
2965 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
2966 DAG.getConstant(4, getPointerTy()));
2969 AFI->setArgRegsSaveSize(ArgRegsSaveSize + AFI->getArgRegsSaveSize());
2971 if (!MemOps.empty())
2972 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOps);
2976 // We cannot allocate a zero-byte object for the first variadic argument,
2977 // so just make up a size.
2980 // This will point to the next argument passed via stack.
2981 return MFI->CreateFixedObject(
2982 ArgSize, ArgOffset, !ForceMutable);
2986 // Setup stack frame, the va_list pointer will start from.
2988 ARMTargetLowering::VarArgStyleRegisters(CCState &CCInfo, SelectionDAG &DAG,
2989 SDLoc dl, SDValue &Chain,
2991 unsigned TotalArgRegsSaveSize,
2992 bool ForceMutable) const {
2993 MachineFunction &MF = DAG.getMachineFunction();
2994 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2996 // Try to store any remaining integer argument regs
2997 // to their spots on the stack so that they may be loaded by deferencing
2998 // the result of va_next.
2999 // If there is no regs to be stored, just point address after last
3000 // argument passed via stack.
3002 StoreByValRegs(CCInfo, DAG, dl, Chain, nullptr,
3003 CCInfo.getInRegsParamsCount(), 0, ArgOffset, 0, ForceMutable,
3004 0, TotalArgRegsSaveSize);
3006 AFI->setVarArgsFrameIndex(FrameIndex);
3010 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
3011 CallingConv::ID CallConv, bool isVarArg,
3012 const SmallVectorImpl<ISD::InputArg>
3014 SDLoc dl, SelectionDAG &DAG,
3015 SmallVectorImpl<SDValue> &InVals)
3017 MachineFunction &MF = DAG.getMachineFunction();
3018 MachineFrameInfo *MFI = MF.getFrameInfo();
3020 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
3022 // Assign locations to all of the incoming arguments.
3023 SmallVector<CCValAssign, 16> ArgLocs;
3024 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs,
3025 *DAG.getContext(), Prologue);
3026 CCInfo.AnalyzeFormalArguments(Ins,
3027 CCAssignFnForNode(CallConv, /* Return*/ false,
3030 SmallVector<SDValue, 16> ArgValues;
3031 int lastInsIndex = -1;
3033 Function::const_arg_iterator CurOrigArg = MF.getFunction()->arg_begin();
3034 unsigned CurArgIdx = 0;
3036 // Initially ArgRegsSaveSize is zero.
3037 // Then we increase this value each time we meet byval parameter.
3038 // We also increase this value in case of varargs function.
3039 AFI->setArgRegsSaveSize(0);
3041 unsigned ByValStoreOffset = 0;
3042 unsigned TotalArgRegsSaveSize = 0;
3043 unsigned ArgRegsSaveSizeMaxAlign = 4;
3045 // Calculate the amount of stack space that we need to allocate to store
3046 // byval and variadic arguments that are passed in registers.
3047 // We need to know this before we allocate the first byval or variadic
3048 // argument, as they will be allocated a stack slot below the CFA (Canonical
3049 // Frame Address, the stack pointer at entry to the function).
3050 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3051 CCValAssign &VA = ArgLocs[i];
3052 if (VA.isMemLoc()) {
3053 int index = VA.getValNo();
3054 if (index != lastInsIndex) {
3055 ISD::ArgFlagsTy Flags = Ins[index].Flags;
3056 if (Flags.isByVal()) {
3057 unsigned ExtraArgRegsSize;
3058 unsigned ExtraArgRegsSaveSize;
3059 computeRegArea(CCInfo, MF, CCInfo.getInRegsParamsProcessed(),
3060 Flags.getByValSize(),
3061 ExtraArgRegsSize, ExtraArgRegsSaveSize);
3063 TotalArgRegsSaveSize += ExtraArgRegsSaveSize;
3064 if (Flags.getByValAlign() > ArgRegsSaveSizeMaxAlign)
3065 ArgRegsSaveSizeMaxAlign = Flags.getByValAlign();
3066 CCInfo.nextInRegsParam();
3068 lastInsIndex = index;
3072 CCInfo.rewindByValRegsInfo();
3074 if (isVarArg && MFI->hasVAStart()) {
3075 unsigned ExtraArgRegsSize;
3076 unsigned ExtraArgRegsSaveSize;
3077 computeRegArea(CCInfo, MF, CCInfo.getInRegsParamsCount(), 0,
3078 ExtraArgRegsSize, ExtraArgRegsSaveSize);
3079 TotalArgRegsSaveSize += ExtraArgRegsSaveSize;
3081 // If the arg regs save area contains N-byte aligned values, the
3082 // bottom of it must be at least N-byte aligned.
3083 TotalArgRegsSaveSize = RoundUpToAlignment(TotalArgRegsSaveSize, ArgRegsSaveSizeMaxAlign);
3084 TotalArgRegsSaveSize = std::min(TotalArgRegsSaveSize, 16U);
3086 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3087 CCValAssign &VA = ArgLocs[i];
3088 std::advance(CurOrigArg, Ins[VA.getValNo()].OrigArgIndex - CurArgIdx);
3089 CurArgIdx = Ins[VA.getValNo()].OrigArgIndex;
3090 // Arguments stored in registers.
3091 if (VA.isRegLoc()) {
3092 EVT RegVT = VA.getLocVT();
3094 if (VA.needsCustom()) {
3095 // f64 and vector types are split up into multiple registers or
3096 // combinations of registers and stack slots.
3097 if (VA.getLocVT() == MVT::v2f64) {
3098 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
3100 VA = ArgLocs[++i]; // skip ahead to next loc
3102 if (VA.isMemLoc()) {
3103 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true);
3104 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
3105 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
3106 MachinePointerInfo::getFixedStack(FI),
3107 false, false, false, 0);
3109 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
3112 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
3113 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
3114 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
3115 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
3116 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
3118 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
3121 const TargetRegisterClass *RC;
3123 if (RegVT == MVT::f32)
3124 RC = &ARM::SPRRegClass;
3125 else if (RegVT == MVT::f64)
3126 RC = &ARM::DPRRegClass;
3127 else if (RegVT == MVT::v2f64)
3128 RC = &ARM::QPRRegClass;
3129 else if (RegVT == MVT::i32)
3130 RC = AFI->isThumb1OnlyFunction() ? &ARM::tGPRRegClass
3131 : &ARM::GPRRegClass;
3133 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
3135 // Transform the arguments in physical registers into virtual ones.
3136 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
3137 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
3140 // If this is an 8 or 16-bit value, it is really passed promoted
3141 // to 32 bits. Insert an assert[sz]ext to capture this, then
3142 // truncate to the right size.
3143 switch (VA.getLocInfo()) {
3144 default: llvm_unreachable("Unknown loc info!");
3145 case CCValAssign::Full: break;
3146 case CCValAssign::BCvt:
3147 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
3149 case CCValAssign::SExt:
3150 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
3151 DAG.getValueType(VA.getValVT()));
3152 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
3154 case CCValAssign::ZExt:
3155 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
3156 DAG.getValueType(VA.getValVT()));
3157 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
3161 InVals.push_back(ArgValue);
3163 } else { // VA.isRegLoc()
3166 assert(VA.isMemLoc());
3167 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
3169 int index = ArgLocs[i].getValNo();
3171 // Some Ins[] entries become multiple ArgLoc[] entries.
3172 // Process them only once.
3173 if (index != lastInsIndex)
3175 ISD::ArgFlagsTy Flags = Ins[index].Flags;
3176 // FIXME: For now, all byval parameter objects are marked mutable.
3177 // This can be changed with more analysis.
3178 // In case of tail call optimization mark all arguments mutable.
3179 // Since they could be overwritten by lowering of arguments in case of
3181 if (Flags.isByVal()) {
3182 unsigned CurByValIndex = CCInfo.getInRegsParamsProcessed();
3184 ByValStoreOffset = RoundUpToAlignment(ByValStoreOffset, Flags.getByValAlign());
3185 int FrameIndex = StoreByValRegs(
3186 CCInfo, DAG, dl, Chain, CurOrigArg,
3188 Ins[VA.getValNo()].PartOffset,
3189 VA.getLocMemOffset(),
3190 Flags.getByValSize(),
3191 true /*force mutable frames*/,
3193 TotalArgRegsSaveSize);
3194 ByValStoreOffset += Flags.getByValSize();
3195 ByValStoreOffset = std::min(ByValStoreOffset, 16U);
3196 InVals.push_back(DAG.getFrameIndex(FrameIndex, getPointerTy()));
3197 CCInfo.nextInRegsParam();
3199 unsigned FIOffset = VA.getLocMemOffset();
3200 int FI = MFI->CreateFixedObject(VA.getLocVT().getSizeInBits()/8,
3203 // Create load nodes to retrieve arguments from the stack.
3204 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
3205 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
3206 MachinePointerInfo::getFixedStack(FI),
3207 false, false, false, 0));
3209 lastInsIndex = index;
3215 if (isVarArg && MFI->hasVAStart())
3216 VarArgStyleRegisters(CCInfo, DAG, dl, Chain,
3217 CCInfo.getNextStackOffset(),
3218 TotalArgRegsSaveSize);
3220 AFI->setArgumentStackSize(CCInfo.getNextStackOffset());
3225 /// isFloatingPointZero - Return true if this is +0.0.
3226 static bool isFloatingPointZero(SDValue Op) {
3227 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
3228 return CFP->getValueAPF().isPosZero();
3229 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
3230 // Maybe this has already been legalized into the constant pool?
3231 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
3232 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
3233 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
3234 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
3235 return CFP->getValueAPF().isPosZero();
3237 } else if (Op->getOpcode() == ISD::BITCAST &&
3238 Op->getValueType(0) == MVT::f64) {
3239 // Handle (ISD::BITCAST (ARMISD::VMOVIMM (ISD::TargetConstant 0)) MVT::f64)
3240 // created by LowerConstantFP().
3241 SDValue BitcastOp = Op->getOperand(0);
3242 if (BitcastOp->getOpcode() == ARMISD::VMOVIMM) {
3243 SDValue MoveOp = BitcastOp->getOperand(0);
3244 if (MoveOp->getOpcode() == ISD::TargetConstant &&
3245 cast<ConstantSDNode>(MoveOp)->getZExtValue() == 0) {
3253 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
3254 /// the given operands.
3256 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
3257 SDValue &ARMcc, SelectionDAG &DAG,
3259 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
3260 unsigned C = RHSC->getZExtValue();
3261 if (!isLegalICmpImmediate(C)) {
3262 // Constant does not fit, try adjusting it by one?
3267 if (C != 0x80000000 && isLegalICmpImmediate(C-1)) {
3268 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
3269 RHS = DAG.getConstant(C-1, MVT::i32);
3274 if (C != 0 && isLegalICmpImmediate(C-1)) {
3275 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
3276 RHS = DAG.getConstant(C-1, MVT::i32);
3281 if (C != 0x7fffffff && isLegalICmpImmediate(C+1)) {
3282 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
3283 RHS = DAG.getConstant(C+1, MVT::i32);
3288 if (C != 0xffffffff && isLegalICmpImmediate(C+1)) {
3289 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
3290 RHS = DAG.getConstant(C+1, MVT::i32);
3297 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3298 ARMISD::NodeType CompareType;
3301 CompareType = ARMISD::CMP;
3306 CompareType = ARMISD::CMPZ;
3309 ARMcc = DAG.getConstant(CondCode, MVT::i32);
3310 return DAG.getNode(CompareType, dl, MVT::Glue, LHS, RHS);
3313 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
3315 ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
3317 assert(!Subtarget->isFPOnlySP() || RHS.getValueType() != MVT::f64);
3319 if (!isFloatingPointZero(RHS))
3320 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Glue, LHS, RHS);
3322 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Glue, LHS);
3323 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Glue, Cmp);
3326 /// duplicateCmp - Glue values can have only one use, so this function
3327 /// duplicates a comparison node.
3329 ARMTargetLowering::duplicateCmp(SDValue Cmp, SelectionDAG &DAG) const {
3330 unsigned Opc = Cmp.getOpcode();
3332 if (Opc == ARMISD::CMP || Opc == ARMISD::CMPZ)
3333 return DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
3335 assert(Opc == ARMISD::FMSTAT && "unexpected comparison operation");
3336 Cmp = Cmp.getOperand(0);
3337 Opc = Cmp.getOpcode();
3338 if (Opc == ARMISD::CMPFP)
3339 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
3341 assert(Opc == ARMISD::CMPFPw0 && "unexpected operand of FMSTAT");
3342 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0));
3344 return DAG.getNode(ARMISD::FMSTAT, DL, MVT::Glue, Cmp);
3347 std::pair<SDValue, SDValue>
3348 ARMTargetLowering::getARMXALUOOp(SDValue Op, SelectionDAG &DAG,
3349 SDValue &ARMcc) const {
3350 assert(Op.getValueType() == MVT::i32 && "Unsupported value type");
3352 SDValue Value, OverflowCmp;
3353 SDValue LHS = Op.getOperand(0);
3354 SDValue RHS = Op.getOperand(1);
3357 // FIXME: We are currently always generating CMPs because we don't support
3358 // generating CMN through the backend. This is not as good as the natural
3359 // CMP case because it causes a register dependency and cannot be folded
3362 switch (Op.getOpcode()) {
3364 llvm_unreachable("Unknown overflow instruction!");
3366 ARMcc = DAG.getConstant(ARMCC::VC, MVT::i32);
3367 Value = DAG.getNode(ISD::ADD, SDLoc(Op), Op.getValueType(), LHS, RHS);
3368 OverflowCmp = DAG.getNode(ARMISD::CMP, SDLoc(Op), MVT::Glue, Value, LHS);
3371 ARMcc = DAG.getConstant(ARMCC::HS, MVT::i32);
3372 Value = DAG.getNode(ISD::ADD, SDLoc(Op), Op.getValueType(), LHS, RHS);
3373 OverflowCmp = DAG.getNode(ARMISD::CMP, SDLoc(Op), MVT::Glue, Value, LHS);
3376 ARMcc = DAG.getConstant(ARMCC::VC, MVT::i32);
3377 Value = DAG.getNode(ISD::SUB, SDLoc(Op), Op.getValueType(), LHS, RHS);
3378 OverflowCmp = DAG.getNode(ARMISD::CMP, SDLoc(Op), MVT::Glue, LHS, RHS);
3381 ARMcc = DAG.getConstant(ARMCC::HS, MVT::i32);
3382 Value = DAG.getNode(ISD::SUB, SDLoc(Op), Op.getValueType(), LHS, RHS);
3383 OverflowCmp = DAG.getNode(ARMISD::CMP, SDLoc(Op), MVT::Glue, LHS, RHS);
3387 return std::make_pair(Value, OverflowCmp);
3392 ARMTargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) const {
3393 // Let legalize expand this if it isn't a legal type yet.
3394 if (!DAG.getTargetLoweringInfo().isTypeLegal(Op.getValueType()))
3397 SDValue Value, OverflowCmp;
3399 std::tie(Value, OverflowCmp) = getARMXALUOOp(Op, DAG, ARMcc);
3400 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3401 // We use 0 and 1 as false and true values.
3402 SDValue TVal = DAG.getConstant(1, MVT::i32);
3403 SDValue FVal = DAG.getConstant(0, MVT::i32);
3404 EVT VT = Op.getValueType();
3406 SDValue Overflow = DAG.getNode(ARMISD::CMOV, SDLoc(Op), VT, TVal, FVal,
3407 ARMcc, CCR, OverflowCmp);
3409 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
3410 return DAG.getNode(ISD::MERGE_VALUES, SDLoc(Op), VTs, Value, Overflow);
3414 SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
3415 SDValue Cond = Op.getOperand(0);
3416 SDValue SelectTrue = Op.getOperand(1);
3417 SDValue SelectFalse = Op.getOperand(2);
3419 unsigned Opc = Cond.getOpcode();
3421 if (Cond.getResNo() == 1 &&
3422 (Opc == ISD::SADDO || Opc == ISD::UADDO || Opc == ISD::SSUBO ||
3423 Opc == ISD::USUBO)) {
3424 if (!DAG.getTargetLoweringInfo().isTypeLegal(Cond->getValueType(0)))
3427 SDValue Value, OverflowCmp;
3429 std::tie(Value, OverflowCmp) = getARMXALUOOp(Cond, DAG, ARMcc);
3430 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3431 EVT VT = Op.getValueType();
3433 return getCMOV(SDLoc(Op), VT, SelectTrue, SelectFalse, ARMcc, CCR,
3439 // (select (cmov 1, 0, cond), t, f) -> (cmov t, f, cond)
3440 // (select (cmov 0, 1, cond), t, f) -> (cmov f, t, cond)
3442 if (Cond.getOpcode() == ARMISD::CMOV && Cond.hasOneUse()) {
3443 const ConstantSDNode *CMOVTrue =
3444 dyn_cast<ConstantSDNode>(Cond.getOperand(0));
3445 const ConstantSDNode *CMOVFalse =
3446 dyn_cast<ConstantSDNode>(Cond.getOperand(1));
3448 if (CMOVTrue && CMOVFalse) {
3449 unsigned CMOVTrueVal = CMOVTrue->getZExtValue();
3450 unsigned CMOVFalseVal = CMOVFalse->getZExtValue();
3454 if (CMOVTrueVal == 1 && CMOVFalseVal == 0) {
3456 False = SelectFalse;
3457 } else if (CMOVTrueVal == 0 && CMOVFalseVal == 1) {
3462 if (True.getNode() && False.getNode()) {
3463 EVT VT = Op.getValueType();
3464 SDValue ARMcc = Cond.getOperand(2);
3465 SDValue CCR = Cond.getOperand(3);
3466 SDValue Cmp = duplicateCmp(Cond.getOperand(4), DAG);
3467 assert(True.getValueType() == VT);
3468 return getCMOV(dl, VT, True, False, ARMcc, CCR, Cmp, DAG);
3473 // ARM's BooleanContents value is UndefinedBooleanContent. Mask out the
3474 // undefined bits before doing a full-word comparison with zero.
3475 Cond = DAG.getNode(ISD::AND, dl, Cond.getValueType(), Cond,
3476 DAG.getConstant(1, Cond.getValueType()));
3478 return DAG.getSelectCC(dl, Cond,
3479 DAG.getConstant(0, Cond.getValueType()),
3480 SelectTrue, SelectFalse, ISD::SETNE);
3483 static ISD::CondCode getInverseCCForVSEL(ISD::CondCode CC) {
3484 if (CC == ISD::SETNE)
3486 return ISD::getSetCCInverse(CC, true);
3489 static void checkVSELConstraints(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
3490 bool &swpCmpOps, bool &swpVselOps) {
3491 // Start by selecting the GE condition code for opcodes that return true for
3493 if (CC == ISD::SETUGE || CC == ISD::SETOGE || CC == ISD::SETOLE ||
3495 CondCode = ARMCC::GE;
3497 // and GT for opcodes that return false for 'equality'.
3498 else if (CC == ISD::SETUGT || CC == ISD::SETOGT || CC == ISD::SETOLT ||
3500 CondCode = ARMCC::GT;
3502 // Since we are constrained to GE/GT, if the opcode contains 'less', we need
3503 // to swap the compare operands.
3504 if (CC == ISD::SETOLE || CC == ISD::SETULE || CC == ISD::SETOLT ||
3508 // Both GT and GE are ordered comparisons, and return false for 'unordered'.
3509 // If we have an unordered opcode, we need to swap the operands to the VSEL
3510 // instruction (effectively negating the condition).
3512 // This also has the effect of swapping which one of 'less' or 'greater'
3513 // returns true, so we also swap the compare operands. It also switches
3514 // whether we return true for 'equality', so we compensate by picking the
3515 // opposite condition code to our original choice.
3516 if (CC == ISD::SETULE || CC == ISD::SETULT || CC == ISD::SETUGE ||
3517 CC == ISD::SETUGT) {
3518 swpCmpOps = !swpCmpOps;
3519 swpVselOps = !swpVselOps;
3520 CondCode = CondCode == ARMCC::GT ? ARMCC::GE : ARMCC::GT;
3523 // 'ordered' is 'anything but unordered', so use the VS condition code and
3524 // swap the VSEL operands.
3525 if (CC == ISD::SETO) {
3526 CondCode = ARMCC::VS;
3530 // 'unordered or not equal' is 'anything but equal', so use the EQ condition
3531 // code and swap the VSEL operands.
3532 if (CC == ISD::SETUNE) {
3533 CondCode = ARMCC::EQ;
3538 SDValue ARMTargetLowering::getCMOV(SDLoc dl, EVT VT, SDValue FalseVal,
3539 SDValue TrueVal, SDValue ARMcc, SDValue CCR,
3540 SDValue Cmp, SelectionDAG &DAG) const {
3541 if (Subtarget->isFPOnlySP() && VT == MVT::f64) {
3542 FalseVal = DAG.getNode(ARMISD::VMOVRRD, dl,
3543 DAG.getVTList(MVT::i32, MVT::i32), FalseVal);
3544 TrueVal = DAG.getNode(ARMISD::VMOVRRD, dl,
3545 DAG.getVTList(MVT::i32, MVT::i32), TrueVal);
3547 SDValue TrueLow = TrueVal.getValue(0);
3548 SDValue TrueHigh = TrueVal.getValue(1);
3549 SDValue FalseLow = FalseVal.getValue(0);
3550 SDValue FalseHigh = FalseVal.getValue(1);
3552 SDValue Low = DAG.getNode(ARMISD::CMOV, dl, MVT::i32, FalseLow, TrueLow,
3554 SDValue High = DAG.getNode(ARMISD::CMOV, dl, MVT::i32, FalseHigh, TrueHigh,
3555 ARMcc, CCR, duplicateCmp(Cmp, DAG));
3557 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Low, High);
3559 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,
3564 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
3565 EVT VT = Op.getValueType();
3566 SDValue LHS = Op.getOperand(0);
3567 SDValue RHS = Op.getOperand(1);
3568 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
3569 SDValue TrueVal = Op.getOperand(2);
3570 SDValue FalseVal = Op.getOperand(3);
3573 if (Subtarget->isFPOnlySP() && LHS.getValueType() == MVT::f64) {
3574 DAG.getTargetLoweringInfo().softenSetCCOperands(DAG, MVT::f64, LHS, RHS, CC,
3577 // If softenSetCCOperands only returned one value, we should compare it to
3579 if (!RHS.getNode()) {
3580 RHS = DAG.getConstant(0, LHS.getValueType());
3585 if (LHS.getValueType() == MVT::i32) {
3586 // Try to generate VSEL on ARMv8.
3587 // The VSEL instruction can't use all the usual ARM condition
3588 // codes: it only has two bits to select the condition code, so it's
3589 // constrained to use only GE, GT, VS and EQ.
3591 // To implement all the various ISD::SETXXX opcodes, we sometimes need to
3592 // swap the operands of the previous compare instruction (effectively
3593 // inverting the compare condition, swapping 'less' and 'greater') and
3594 // sometimes need to swap the operands to the VSEL (which inverts the
3595 // condition in the sense of firing whenever the previous condition didn't)
3596 if (Subtarget->hasFPARMv8() && (TrueVal.getValueType() == MVT::f32 ||
3597 TrueVal.getValueType() == MVT::f64)) {
3598 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3599 if (CondCode == ARMCC::LT || CondCode == ARMCC::LE ||
3600 CondCode == ARMCC::VC || CondCode == ARMCC::NE) {
3601 CC = getInverseCCForVSEL(CC);
3602 std::swap(TrueVal, FalseVal);
3607 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3608 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3609 return getCMOV(dl, VT, FalseVal, TrueVal, ARMcc, CCR, Cmp, DAG);
3612 ARMCC::CondCodes CondCode, CondCode2;
3613 FPCCToARMCC(CC, CondCode, CondCode2);
3615 // Try to generate VSEL on ARMv8.
3616 if (Subtarget->hasFPARMv8() && (TrueVal.getValueType() == MVT::f32 ||
3617 TrueVal.getValueType() == MVT::f64)) {
3618 // We can select VMAXNM/VMINNM from a compare followed by a select with the
3619 // same operands, as follows:
3620 // c = fcmp [ogt, olt, ugt, ult] a, b
3622 // We only do this in unsafe-fp-math, because signed zeros and NaNs are
3623 // handled differently than the original code sequence.
3624 if (getTargetMachine().Options.UnsafeFPMath) {
3625 if (LHS == TrueVal && RHS == FalseVal) {
3626 if (CC == ISD::SETOGT || CC == ISD::SETUGT)
3627 return DAG.getNode(ARMISD::VMAXNM, dl, VT, TrueVal, FalseVal);
3628 if (CC == ISD::SETOLT || CC == ISD::SETULT)
3629 return DAG.getNode(ARMISD::VMINNM, dl, VT, TrueVal, FalseVal);
3630 } else if (LHS == FalseVal && RHS == TrueVal) {
3631 if (CC == ISD::SETOLT || CC == ISD::SETULT)
3632 return DAG.getNode(ARMISD::VMAXNM, dl, VT, TrueVal, FalseVal);
3633 if (CC == ISD::SETOGT || CC == ISD::SETUGT)
3634 return DAG.getNode(ARMISD::VMINNM, dl, VT, TrueVal, FalseVal);
3638 bool swpCmpOps = false;
3639 bool swpVselOps = false;
3640 checkVSELConstraints(CC, CondCode, swpCmpOps, swpVselOps);
3642 if (CondCode == ARMCC::GT || CondCode == ARMCC::GE ||
3643 CondCode == ARMCC::VS || CondCode == ARMCC::EQ) {
3645 std::swap(LHS, RHS);
3647 std::swap(TrueVal, FalseVal);
3651 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
3652 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
3653 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3654 SDValue Result = getCMOV(dl, VT, FalseVal, TrueVal, ARMcc, CCR, Cmp, DAG);
3655 if (CondCode2 != ARMCC::AL) {
3656 SDValue ARMcc2 = DAG.getConstant(CondCode2, MVT::i32);
3657 // FIXME: Needs another CMP because flag can have but one use.
3658 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
3659 Result = getCMOV(dl, VT, Result, TrueVal, ARMcc2, CCR, Cmp2, DAG);
3664 /// canChangeToInt - Given the fp compare operand, return true if it is suitable
3665 /// to morph to an integer compare sequence.
3666 static bool canChangeToInt(SDValue Op, bool &SeenZero,
3667 const ARMSubtarget *Subtarget) {
3668 SDNode *N = Op.getNode();
3669 if (!N->hasOneUse())
3670 // Otherwise it requires moving the value from fp to integer registers.
3672 if (!N->getNumValues())
3674 EVT VT = Op.getValueType();
3675 if (VT != MVT::f32 && !Subtarget->isFPBrccSlow())
3676 // f32 case is generally profitable. f64 case only makes sense when vcmpe +
3677 // vmrs are very slow, e.g. cortex-a8.
3680 if (isFloatingPointZero(Op)) {
3684 return ISD::isNormalLoad(N);
3687 static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) {
3688 if (isFloatingPointZero(Op))
3689 return DAG.getConstant(0, MVT::i32);
3691 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op))
3692 return DAG.getLoad(MVT::i32, SDLoc(Op),
3693 Ld->getChain(), Ld->getBasePtr(), Ld->getPointerInfo(),
3694 Ld->isVolatile(), Ld->isNonTemporal(),
3695 Ld->isInvariant(), Ld->getAlignment());
3697 llvm_unreachable("Unknown VFP cmp argument!");
3700 static void expandf64Toi32(SDValue Op, SelectionDAG &DAG,
3701 SDValue &RetVal1, SDValue &RetVal2) {
3702 if (isFloatingPointZero(Op)) {
3703 RetVal1 = DAG.getConstant(0, MVT::i32);
3704 RetVal2 = DAG.getConstant(0, MVT::i32);
3708 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) {
3709 SDValue Ptr = Ld->getBasePtr();
3710 RetVal1 = DAG.getLoad(MVT::i32, SDLoc(Op),
3711 Ld->getChain(), Ptr,
3712 Ld->getPointerInfo(),
3713 Ld->isVolatile(), Ld->isNonTemporal(),
3714 Ld->isInvariant(), Ld->getAlignment());
3716 EVT PtrType = Ptr.getValueType();
3717 unsigned NewAlign = MinAlign(Ld->getAlignment(), 4);
3718 SDValue NewPtr = DAG.getNode(ISD::ADD, SDLoc(Op),
3719 PtrType, Ptr, DAG.getConstant(4, PtrType));
3720 RetVal2 = DAG.getLoad(MVT::i32, SDLoc(Op),
3721 Ld->getChain(), NewPtr,
3722 Ld->getPointerInfo().getWithOffset(4),
3723 Ld->isVolatile(), Ld->isNonTemporal(),
3724 Ld->isInvariant(), NewAlign);
3728 llvm_unreachable("Unknown VFP cmp argument!");
3731 /// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some
3732 /// f32 and even f64 comparisons to integer ones.
3734 ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const {
3735 SDValue Chain = Op.getOperand(0);
3736 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3737 SDValue LHS = Op.getOperand(2);
3738 SDValue RHS = Op.getOperand(3);
3739 SDValue Dest = Op.getOperand(4);
3742 bool LHSSeenZero = false;
3743 bool LHSOk = canChangeToInt(LHS, LHSSeenZero, Subtarget);
3744 bool RHSSeenZero = false;
3745 bool RHSOk = canChangeToInt(RHS, RHSSeenZero, Subtarget);
3746 if (LHSOk && RHSOk && (LHSSeenZero || RHSSeenZero)) {
3747 // If unsafe fp math optimization is enabled and there are no other uses of
3748 // the CMP operands, and the condition code is EQ or NE, we can optimize it
3749 // to an integer comparison.
3750 if (CC == ISD::SETOEQ)
3752 else if (CC == ISD::SETUNE)
3755 SDValue Mask = DAG.getConstant(0x7fffffff, MVT::i32);
3757 if (LHS.getValueType() == MVT::f32) {
3758 LHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3759 bitcastf32Toi32(LHS, DAG), Mask);
3760 RHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3761 bitcastf32Toi32(RHS, DAG), Mask);
3762 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3763 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3764 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3765 Chain, Dest, ARMcc, CCR, Cmp);
3770 expandf64Toi32(LHS, DAG, LHS1, LHS2);
3771 expandf64Toi32(RHS, DAG, RHS1, RHS2);
3772 LHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, LHS2, Mask);
3773 RHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, RHS2, Mask);
3774 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3775 ARMcc = DAG.getConstant(CondCode, MVT::i32);
3776 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3777 SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest };
3778 return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops);
3784 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
3785 SDValue Chain = Op.getOperand(0);
3786 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3787 SDValue LHS = Op.getOperand(2);
3788 SDValue RHS = Op.getOperand(3);
3789 SDValue Dest = Op.getOperand(4);
3792 if (Subtarget->isFPOnlySP() && LHS.getValueType() == MVT::f64) {
3793 DAG.getTargetLoweringInfo().softenSetCCOperands(DAG, MVT::f64, LHS, RHS, CC,
3796 // If softenSetCCOperands only returned one value, we should compare it to
3798 if (!RHS.getNode()) {
3799 RHS = DAG.getConstant(0, LHS.getValueType());
3804 if (LHS.getValueType() == MVT::i32) {
3806 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3807 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3808 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3809 Chain, Dest, ARMcc, CCR, Cmp);
3812 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3814 if (getTargetMachine().Options.UnsafeFPMath &&
3815 (CC == ISD::SETEQ || CC == ISD::SETOEQ ||
3816 CC == ISD::SETNE || CC == ISD::SETUNE)) {
3817 SDValue Result = OptimizeVFPBrcond(Op, DAG);
3818 if (Result.getNode())
3822 ARMCC::CondCodes CondCode, CondCode2;
3823 FPCCToARMCC(CC, CondCode, CondCode2);
3825 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
3826 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
3827 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3828 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3829 SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp };
3830 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops);
3831 if (CondCode2 != ARMCC::AL) {
3832 ARMcc = DAG.getConstant(CondCode2, MVT::i32);
3833 SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) };
3834 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops);
3839 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
3840 SDValue Chain = Op.getOperand(0);
3841 SDValue Table = Op.getOperand(1);
3842 SDValue Index = Op.getOperand(2);
3845 EVT PTy = getPointerTy();
3846 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
3847 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
3848 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
3849 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
3850 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
3851 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
3852 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
3853 if (Subtarget->isThumb2()) {
3854 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
3855 // which does another jump to the destination. This also makes it easier
3856 // to translate it to TBB / TBH later.
3857 // FIXME: This might not work if the function is extremely large.
3858 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
3859 Addr, Op.getOperand(2), JTI, UId);
3861 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
3862 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
3863 MachinePointerInfo::getJumpTable(),
3864 false, false, false, 0);
3865 Chain = Addr.getValue(1);
3866 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
3867 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3869 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
3870 MachinePointerInfo::getJumpTable(),
3871 false, false, false, 0);
3872 Chain = Addr.getValue(1);
3873 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3877 static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3878 EVT VT = Op.getValueType();
3881 if (Op.getValueType().getVectorElementType() == MVT::i32) {
3882 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::f32)
3884 return DAG.UnrollVectorOp(Op.getNode());
3887 assert(Op.getOperand(0).getValueType() == MVT::v4f32 &&
3888 "Invalid type for custom lowering!");
3889 if (VT != MVT::v4i16)
3890 return DAG.UnrollVectorOp(Op.getNode());
3892 Op = DAG.getNode(Op.getOpcode(), dl, MVT::v4i32, Op.getOperand(0));
3893 return DAG.getNode(ISD::TRUNCATE, dl, VT, Op);
3896 SDValue ARMTargetLowering::LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) const {
3897 EVT VT = Op.getValueType();
3899 return LowerVectorFP_TO_INT(Op, DAG);
3901 if (Subtarget->isFPOnlySP() && Op.getOperand(0).getValueType() == MVT::f64) {
3903 if (Op.getOpcode() == ISD::FP_TO_SINT)
3904 LC = RTLIB::getFPTOSINT(Op.getOperand(0).getValueType(),
3907 LC = RTLIB::getFPTOUINT(Op.getOperand(0).getValueType(),
3909 return makeLibCall(DAG, LC, Op.getValueType(), &Op.getOperand(0), 1,
3910 /*isSigned*/ false, SDLoc(Op)).first;
3916 switch (Op.getOpcode()) {
3917 default: llvm_unreachable("Invalid opcode!");
3918 case ISD::FP_TO_SINT:
3919 Opc = ARMISD::FTOSI;
3921 case ISD::FP_TO_UINT:
3922 Opc = ARMISD::FTOUI;
3925 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
3926 return DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
3929 static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3930 EVT VT = Op.getValueType();
3933 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::i32) {
3934 if (VT.getVectorElementType() == MVT::f32)
3936 return DAG.UnrollVectorOp(Op.getNode());
3939 assert(Op.getOperand(0).getValueType() == MVT::v4i16 &&
3940 "Invalid type for custom lowering!");
3941 if (VT != MVT::v4f32)
3942 return DAG.UnrollVectorOp(Op.getNode());
3946 switch (Op.getOpcode()) {
3947 default: llvm_unreachable("Invalid opcode!");
3948 case ISD::SINT_TO_FP:
3949 CastOpc = ISD::SIGN_EXTEND;
3950 Opc = ISD::SINT_TO_FP;
3952 case ISD::UINT_TO_FP:
3953 CastOpc = ISD::ZERO_EXTEND;
3954 Opc = ISD::UINT_TO_FP;
3958 Op = DAG.getNode(CastOpc, dl, MVT::v4i32, Op.getOperand(0));
3959 return DAG.getNode(Opc, dl, VT, Op);
3962 SDValue ARMTargetLowering::LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) const {
3963 EVT VT = Op.getValueType();
3965 return LowerVectorINT_TO_FP(Op, DAG);
3967 if (Subtarget->isFPOnlySP() && Op.getValueType() == MVT::f64) {
3969 if (Op.getOpcode() == ISD::SINT_TO_FP)
3970 LC = RTLIB::getSINTTOFP(Op.getOperand(0).getValueType(),
3973 LC = RTLIB::getUINTTOFP(Op.getOperand(0).getValueType(),
3975 return makeLibCall(DAG, LC, Op.getValueType(), &Op.getOperand(0), 1,
3976 /*isSigned*/ false, SDLoc(Op)).first;
3982 switch (Op.getOpcode()) {
3983 default: llvm_unreachable("Invalid opcode!");
3984 case ISD::SINT_TO_FP:
3985 Opc = ARMISD::SITOF;
3987 case ISD::UINT_TO_FP:
3988 Opc = ARMISD::UITOF;
3992 Op = DAG.getNode(ISD::BITCAST, dl, MVT::f32, Op.getOperand(0));
3993 return DAG.getNode(Opc, dl, VT, Op);
3996 SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
3997 // Implement fcopysign with a fabs and a conditional fneg.
3998 SDValue Tmp0 = Op.getOperand(0);
3999 SDValue Tmp1 = Op.getOperand(1);
4001 EVT VT = Op.getValueType();
4002 EVT SrcVT = Tmp1.getValueType();
4003 bool InGPR = Tmp0.getOpcode() == ISD::BITCAST ||
4004 Tmp0.getOpcode() == ARMISD::VMOVDRR;
4005 bool UseNEON = !InGPR && Subtarget->hasNEON();
4008 // Use VBSL to copy the sign bit.
4009 unsigned EncodedVal = ARM_AM::createNEONModImm(0x6, 0x80);
4010 SDValue Mask = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v2i32,
4011 DAG.getTargetConstant(EncodedVal, MVT::i32));
4012 EVT OpVT = (VT == MVT::f32) ? MVT::v2i32 : MVT::v1i64;
4014 Mask = DAG.getNode(ARMISD::VSHL, dl, OpVT,
4015 DAG.getNode(ISD::BITCAST, dl, OpVT, Mask),
4016 DAG.getConstant(32, MVT::i32));
4017 else /*if (VT == MVT::f32)*/
4018 Tmp0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp0);
4019 if (SrcVT == MVT::f32) {
4020 Tmp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp1);
4022 Tmp1 = DAG.getNode(ARMISD::VSHL, dl, OpVT,
4023 DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1),
4024 DAG.getConstant(32, MVT::i32));
4025 } else if (VT == MVT::f32)
4026 Tmp1 = DAG.getNode(ARMISD::VSHRu, dl, MVT::v1i64,
4027 DAG.getNode(ISD::BITCAST, dl, MVT::v1i64, Tmp1),
4028 DAG.getConstant(32, MVT::i32));
4029 Tmp0 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp0);
4030 Tmp1 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1);
4032 SDValue AllOnes = DAG.getTargetConstant(ARM_AM::createNEONModImm(0xe, 0xff),
4034 AllOnes = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v8i8, AllOnes);
4035 SDValue MaskNot = DAG.getNode(ISD::XOR, dl, OpVT, Mask,
4036 DAG.getNode(ISD::BITCAST, dl, OpVT, AllOnes));
4038 SDValue Res = DAG.getNode(ISD::OR, dl, OpVT,
4039 DAG.getNode(ISD::AND, dl, OpVT, Tmp1, Mask),
4040 DAG.getNode(ISD::AND, dl, OpVT, Tmp0, MaskNot));
4041 if (VT == MVT::f32) {
4042 Res = DAG.getNode(ISD::BITCAST, dl, MVT::v2f32, Res);
4043 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, Res,
4044 DAG.getConstant(0, MVT::i32));
4046 Res = DAG.getNode(ISD::BITCAST, dl, MVT::f64, Res);
4052 // Bitcast operand 1 to i32.
4053 if (SrcVT == MVT::f64)
4054 Tmp1 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
4056 Tmp1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp1);
4058 // Or in the signbit with integer operations.
4059 SDValue Mask1 = DAG.getConstant(0x80000000, MVT::i32);
4060 SDValue Mask2 = DAG.getConstant(0x7fffffff, MVT::i32);
4061 Tmp1 = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp1, Mask1);
4062 if (VT == MVT::f32) {
4063 Tmp0 = DAG.getNode(ISD::AND, dl, MVT::i32,
4064 DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp0), Mask2);
4065 return DAG.getNode(ISD::BITCAST, dl, MVT::f32,
4066 DAG.getNode(ISD::OR, dl, MVT::i32, Tmp0, Tmp1));
4069 // f64: Or the high part with signbit and then combine two parts.
4070 Tmp0 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
4072 SDValue Lo = Tmp0.getValue(0);
4073 SDValue Hi = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp0.getValue(1), Mask2);
4074 Hi = DAG.getNode(ISD::OR, dl, MVT::i32, Hi, Tmp1);
4075 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
4078 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
4079 MachineFunction &MF = DAG.getMachineFunction();
4080 MachineFrameInfo *MFI = MF.getFrameInfo();
4081 MFI->setReturnAddressIsTaken(true);
4083 if (verifyReturnAddressArgumentIsConstant(Op, DAG))
4086 EVT VT = Op.getValueType();
4088 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
4090 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
4091 SDValue Offset = DAG.getConstant(4, MVT::i32);
4092 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
4093 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
4094 MachinePointerInfo(), false, false, false, 0);
4097 // Return LR, which contains the return address. Mark it an implicit live-in.
4098 unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32));
4099 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
4102 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
4103 const ARMBaseRegisterInfo &ARI =
4104 *static_cast<const ARMBaseRegisterInfo*>(RegInfo);
4105 MachineFunction &MF = DAG.getMachineFunction();
4106 MachineFrameInfo *MFI = MF.getFrameInfo();
4107 MFI->setFrameAddressIsTaken(true);
4109 EVT VT = Op.getValueType();
4110 SDLoc dl(Op); // FIXME probably not meaningful
4111 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
4112 unsigned FrameReg = ARI.getFrameRegister(MF);
4113 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
4115 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
4116 MachinePointerInfo(),
4117 false, false, false, 0);
4121 // FIXME? Maybe this could be a TableGen attribute on some registers and
4122 // this table could be generated automatically from RegInfo.
4123 unsigned ARMTargetLowering::getRegisterByName(const char* RegName,
4125 unsigned Reg = StringSwitch<unsigned>(RegName)
4126 .Case("sp", ARM::SP)
4130 report_fatal_error("Invalid register name global variable");
4133 /// ExpandBITCAST - If the target supports VFP, this function is called to
4134 /// expand a bit convert where either the source or destination type is i64 to
4135 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
4136 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
4137 /// vectors), since the legalizer won't know what to do with that.
4138 static SDValue ExpandBITCAST(SDNode *N, SelectionDAG &DAG) {
4139 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4141 SDValue Op = N->getOperand(0);
4143 // This function is only supposed to be called for i64 types, either as the
4144 // source or destination of the bit convert.
4145 EVT SrcVT = Op.getValueType();
4146 EVT DstVT = N->getValueType(0);
4147 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
4148 "ExpandBITCAST called for non-i64 type");
4150 // Turn i64->f64 into VMOVDRR.
4151 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
4152 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
4153 DAG.getConstant(0, MVT::i32));
4154 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
4155 DAG.getConstant(1, MVT::i32));
4156 return DAG.getNode(ISD::BITCAST, dl, DstVT,
4157 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
4160 // Turn f64->i64 into VMOVRRD.
4161 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
4163 if (TLI.isBigEndian() && SrcVT.isVector() &&
4164 SrcVT.getVectorNumElements() > 1)
4165 Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
4166 DAG.getVTList(MVT::i32, MVT::i32),
4167 DAG.getNode(ARMISD::VREV64, dl, SrcVT, Op));
4169 Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
4170 DAG.getVTList(MVT::i32, MVT::i32), Op);
4171 // Merge the pieces into a single i64 value.
4172 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
4178 /// getZeroVector - Returns a vector of specified type with all zero elements.
4179 /// Zero vectors are used to represent vector negation and in those cases
4180 /// will be implemented with the NEON VNEG instruction. However, VNEG does
4181 /// not support i64 elements, so sometimes the zero vectors will need to be
4182 /// explicitly constructed. Regardless, use a canonical VMOV to create the
4184 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, SDLoc dl) {
4185 assert(VT.isVector() && "Expected a vector type");
4186 // The canonical modified immediate encoding of a zero vector is....0!
4187 SDValue EncodedVal = DAG.getTargetConstant(0, MVT::i32);
4188 EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
4189 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal);
4190 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4193 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
4194 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
4195 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
4196 SelectionDAG &DAG) const {
4197 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4198 EVT VT = Op.getValueType();
4199 unsigned VTBits = VT.getSizeInBits();
4201 SDValue ShOpLo = Op.getOperand(0);
4202 SDValue ShOpHi = Op.getOperand(1);
4203 SDValue ShAmt = Op.getOperand(2);
4205 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
4207 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
4209 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
4210 DAG.getConstant(VTBits, MVT::i32), ShAmt);
4211 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
4212 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
4213 DAG.getConstant(VTBits, MVT::i32));
4214 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
4215 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
4216 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
4218 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
4219 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
4221 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
4222 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc,
4225 SDValue Ops[2] = { Lo, Hi };
4226 return DAG.getMergeValues(Ops, dl);
4229 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
4230 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
4231 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
4232 SelectionDAG &DAG) const {
4233 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4234 EVT VT = Op.getValueType();
4235 unsigned VTBits = VT.getSizeInBits();
4237 SDValue ShOpLo = Op.getOperand(0);
4238 SDValue ShOpHi = Op.getOperand(1);
4239 SDValue ShAmt = Op.getOperand(2);
4242 assert(Op.getOpcode() == ISD::SHL_PARTS);
4243 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
4244 DAG.getConstant(VTBits, MVT::i32), ShAmt);
4245 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
4246 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
4247 DAG.getConstant(VTBits, MVT::i32));
4248 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
4249 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
4251 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
4252 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
4253 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
4255 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
4256 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc,
4259 SDValue Ops[2] = { Lo, Hi };
4260 return DAG.getMergeValues(Ops, dl);
4263 SDValue ARMTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
4264 SelectionDAG &DAG) const {
4265 // The rounding mode is in bits 23:22 of the FPSCR.
4266 // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0
4267 // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3)
4268 // so that the shift + and get folded into a bitfield extract.
4270 SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::i32,
4271 DAG.getConstant(Intrinsic::arm_get_fpscr,
4273 SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPSCR,
4274 DAG.getConstant(1U << 22, MVT::i32));
4275 SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds,
4276 DAG.getConstant(22, MVT::i32));
4277 return DAG.getNode(ISD::AND, dl, MVT::i32, RMODE,
4278 DAG.getConstant(3, MVT::i32));
4281 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
4282 const ARMSubtarget *ST) {
4283 EVT VT = N->getValueType(0);
4286 if (!ST->hasV6T2Ops())
4289 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
4290 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
4293 /// getCTPOP16BitCounts - Returns a v8i8/v16i8 vector containing the bit-count
4294 /// for each 16-bit element from operand, repeated. The basic idea is to
4295 /// leverage vcnt to get the 8-bit counts, gather and add the results.
4297 /// Trace for v4i16:
4298 /// input = [v0 v1 v2 v3 ] (vi 16-bit element)
4299 /// cast: N0 = [w0 w1 w2 w3 w4 w5 w6 w7] (v0 = [w0 w1], wi 8-bit element)
4300 /// vcnt: N1 = [b0 b1 b2 b3 b4 b5 b6 b7] (bi = bit-count of 8-bit element wi)
4301 /// vrev: N2 = [b1 b0 b3 b2 b5 b4 b7 b6]
4302 /// [b0 b1 b2 b3 b4 b5 b6 b7]
4303 /// +[b1 b0 b3 b2 b5 b4 b7 b6]
4304 /// N3=N1+N2 = [k0 k0 k1 k1 k2 k2 k3 k3] (k0 = b0+b1 = bit-count of 16-bit v0,
4305 /// vuzp: = [k0 k1 k2 k3 k0 k1 k2 k3] each ki is 8-bits)
4306 static SDValue getCTPOP16BitCounts(SDNode *N, SelectionDAG &DAG) {
4307 EVT VT = N->getValueType(0);
4310 EVT VT8Bit = VT.is64BitVector() ? MVT::v8i8 : MVT::v16i8;
4311 SDValue N0 = DAG.getNode(ISD::BITCAST, DL, VT8Bit, N->getOperand(0));
4312 SDValue N1 = DAG.getNode(ISD::CTPOP, DL, VT8Bit, N0);
4313 SDValue N2 = DAG.getNode(ARMISD::VREV16, DL, VT8Bit, N1);
4314 SDValue N3 = DAG.getNode(ISD::ADD, DL, VT8Bit, N1, N2);
4315 return DAG.getNode(ARMISD::VUZP, DL, VT8Bit, N3, N3);
4318 /// lowerCTPOP16BitElements - Returns a v4i16/v8i16 vector containing the
4319 /// bit-count for each 16-bit element from the operand. We need slightly
4320 /// different sequencing for v4i16 and v8i16 to stay within NEON's available
4321 /// 64/128-bit registers.
4323 /// Trace for v4i16:
4324 /// input = [v0 v1 v2 v3 ] (vi 16-bit element)
4325 /// v8i8: BitCounts = [k0 k1 k2 k3 k0 k1 k2 k3 ] (ki is the bit-count of vi)
4326 /// v8i16:Extended = [k0 k1 k2 k3 k0 k1 k2 k3 ]
4327 /// v4i16:Extracted = [k0 k1 k2 k3 ]
4328 static SDValue lowerCTPOP16BitElements(SDNode *N, SelectionDAG &DAG) {
4329 EVT VT = N->getValueType(0);
4332 SDValue BitCounts = getCTPOP16BitCounts(N, DAG);
4333 if (VT.is64BitVector()) {
4334 SDValue Extended = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v8i16, BitCounts);
4335 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i16, Extended,
4336 DAG.getIntPtrConstant(0));
4338 SDValue Extracted = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v8i8,
4339 BitCounts, DAG.getIntPtrConstant(0));
4340 return DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v8i16, Extracted);
4344 /// lowerCTPOP32BitElements - Returns a v2i32/v4i32 vector containing the
4345 /// bit-count for each 32-bit element from the operand. The idea here is
4346 /// to split the vector into 16-bit elements, leverage the 16-bit count
4347 /// routine, and then combine the results.
4349 /// Trace for v2i32 (v4i32 similar with Extracted/Extended exchanged):
4350 /// input = [v0 v1 ] (vi: 32-bit elements)
4351 /// Bitcast = [w0 w1 w2 w3 ] (wi: 16-bit elements, v0 = [w0 w1])
4352 /// Counts16 = [k0 k1 k2 k3 ] (ki: 16-bit elements, bit-count of wi)
4353 /// vrev: N0 = [k1 k0 k3 k2 ]
4355 /// N1 =+[k1 k0 k3 k2 ]
4357 /// N2 =+[k1 k3 k0 k2 ]
4359 /// Extended =+[k1 k3 k0 k2 ]
4361 /// Extracted=+[k1 k3 ]
4363 static SDValue lowerCTPOP32BitElements(SDNode *N, SelectionDAG &DAG) {
4364 EVT VT = N->getValueType(0);
4367 EVT VT16Bit = VT.is64BitVector() ? MVT::v4i16 : MVT::v8i16;
4369 SDValue Bitcast = DAG.getNode(ISD::BITCAST, DL, VT16Bit, N->getOperand(0));
4370 SDValue Counts16 = lowerCTPOP16BitElements(Bitcast.getNode(), DAG);
4371 SDValue N0 = DAG.getNode(ARMISD::VREV32, DL, VT16Bit, Counts16);
4372 SDValue N1 = DAG.getNode(ISD::ADD, DL, VT16Bit, Counts16, N0);
4373 SDValue N2 = DAG.getNode(ARMISD::VUZP, DL, VT16Bit, N1, N1);
4375 if (VT.is64BitVector()) {
4376 SDValue Extended = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v4i32, N2);
4377 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i32, Extended,
4378 DAG.getIntPtrConstant(0));
4380 SDValue Extracted = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i16, N2,
4381 DAG.getIntPtrConstant(0));
4382 return DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v4i32, Extracted);
4386 static SDValue LowerCTPOP(SDNode *N, SelectionDAG &DAG,
4387 const ARMSubtarget *ST) {
4388 EVT VT = N->getValueType(0);
4390 assert(ST->hasNEON() && "Custom ctpop lowering requires NEON.");
4391 assert((VT == MVT::v2i32 || VT == MVT::v4i32 ||
4392 VT == MVT::v4i16 || VT == MVT::v8i16) &&
4393 "Unexpected type for custom ctpop lowering");
4395 if (VT.getVectorElementType() == MVT::i32)
4396 return lowerCTPOP32BitElements(N, DAG);
4398 return lowerCTPOP16BitElements(N, DAG);
4401 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
4402 const ARMSubtarget *ST) {
4403 EVT VT = N->getValueType(0);
4409 // Lower vector shifts on NEON to use VSHL.
4410 assert(ST->hasNEON() && "unexpected vector shift");
4412 // Left shifts translate directly to the vshiftu intrinsic.
4413 if (N->getOpcode() == ISD::SHL)
4414 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
4415 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
4416 N->getOperand(0), N->getOperand(1));
4418 assert((N->getOpcode() == ISD::SRA ||
4419 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
4421 // NEON uses the same intrinsics for both left and right shifts. For
4422 // right shifts, the shift amounts are negative, so negate the vector of
4424 EVT ShiftVT = N->getOperand(1).getValueType();
4425 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
4426 getZeroVector(ShiftVT, DAG, dl),
4428 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
4429 Intrinsic::arm_neon_vshifts :
4430 Intrinsic::arm_neon_vshiftu);
4431 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
4432 DAG.getConstant(vshiftInt, MVT::i32),
4433 N->getOperand(0), NegatedCount);
4436 static SDValue Expand64BitShift(SDNode *N, SelectionDAG &DAG,
4437 const ARMSubtarget *ST) {
4438 EVT VT = N->getValueType(0);
4441 // We can get here for a node like i32 = ISD::SHL i32, i64
4445 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
4446 "Unknown shift to lower!");
4448 // We only lower SRA, SRL of 1 here, all others use generic lowering.
4449 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
4450 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
4453 // If we are in thumb mode, we don't have RRX.
4454 if (ST->isThumb1Only()) return SDValue();
4456 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
4457 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
4458 DAG.getConstant(0, MVT::i32));
4459 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
4460 DAG.getConstant(1, MVT::i32));
4462 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
4463 // captures the result into a carry flag.
4464 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
4465 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Glue), Hi);
4467 // The low part is an ARMISD::RRX operand, which shifts the carry in.
4468 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
4470 // Merge the pieces into a single i64 value.
4471 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
4474 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
4475 SDValue TmpOp0, TmpOp1;
4476 bool Invert = false;
4480 SDValue Op0 = Op.getOperand(0);
4481 SDValue Op1 = Op.getOperand(1);
4482 SDValue CC = Op.getOperand(2);
4483 EVT VT = Op.getValueType();
4484 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
4487 if (Op1.getValueType().isFloatingPoint()) {
4488 switch (SetCCOpcode) {
4489 default: llvm_unreachable("Illegal FP comparison");
4491 case ISD::SETNE: Invert = true; // Fallthrough
4493 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
4495 case ISD::SETLT: Swap = true; // Fallthrough
4497 case ISD::SETGT: Opc = ARMISD::VCGT; break;
4499 case ISD::SETLE: Swap = true; // Fallthrough
4501 case ISD::SETGE: Opc = ARMISD::VCGE; break;
4502 case ISD::SETUGE: Swap = true; // Fallthrough
4503 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
4504 case ISD::SETUGT: Swap = true; // Fallthrough
4505 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
4506 case ISD::SETUEQ: Invert = true; // Fallthrough
4508 // Expand this to (OLT | OGT).
4512 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
4513 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
4515 case ISD::SETUO: Invert = true; // Fallthrough
4517 // Expand this to (OLT | OGE).
4521 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
4522 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
4526 // Integer comparisons.
4527 switch (SetCCOpcode) {
4528 default: llvm_unreachable("Illegal integer comparison");
4529 case ISD::SETNE: Invert = true;
4530 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
4531 case ISD::SETLT: Swap = true;
4532 case ISD::SETGT: Opc = ARMISD::VCGT; break;
4533 case ISD::SETLE: Swap = true;
4534 case ISD::SETGE: Opc = ARMISD::VCGE; break;
4535 case ISD::SETULT: Swap = true;
4536 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
4537 case ISD::SETULE: Swap = true;
4538 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
4541 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
4542 if (Opc == ARMISD::VCEQ) {
4545 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
4547 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
4550 // Ignore bitconvert.
4551 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BITCAST)
4552 AndOp = AndOp.getOperand(0);
4554 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
4556 Op0 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(0));
4557 Op1 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(1));
4564 std::swap(Op0, Op1);
4566 // If one of the operands is a constant vector zero, attempt to fold the
4567 // comparison to a specialized compare-against-zero form.
4569 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
4571 else if (ISD::isBuildVectorAllZeros(Op0.getNode())) {
4572 if (Opc == ARMISD::VCGE)
4573 Opc = ARMISD::VCLEZ;
4574 else if (Opc == ARMISD::VCGT)
4575 Opc = ARMISD::VCLTZ;
4580 if (SingleOp.getNode()) {
4583 Result = DAG.getNode(ARMISD::VCEQZ, dl, VT, SingleOp); break;
4585 Result = DAG.getNode(ARMISD::VCGEZ, dl, VT, SingleOp); break;
4587 Result = DAG.getNode(ARMISD::VCLEZ, dl, VT, SingleOp); break;
4589 Result = DAG.getNode(ARMISD::VCGTZ, dl, VT, SingleOp); break;
4591 Result = DAG.getNode(ARMISD::VCLTZ, dl, VT, SingleOp); break;
4593 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
4596 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
4600 Result = DAG.getNOT(dl, Result, VT);
4605 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
4606 /// valid vector constant for a NEON instruction with a "modified immediate"
4607 /// operand (e.g., VMOV). If so, return the encoded value.
4608 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
4609 unsigned SplatBitSize, SelectionDAG &DAG,
4610 EVT &VT, bool is128Bits, NEONModImmType type) {
4611 unsigned OpCmode, Imm;
4613 // SplatBitSize is set to the smallest size that splats the vector, so a
4614 // zero vector will always have SplatBitSize == 8. However, NEON modified
4615 // immediate instructions others than VMOV do not support the 8-bit encoding
4616 // of a zero vector, and the default encoding of zero is supposed to be the
4621 switch (SplatBitSize) {
4623 if (type != VMOVModImm)
4625 // Any 1-byte value is OK. Op=0, Cmode=1110.
4626 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
4629 VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
4633 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
4634 VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
4635 if ((SplatBits & ~0xff) == 0) {
4636 // Value = 0x00nn: Op=x, Cmode=100x.
4641 if ((SplatBits & ~0xff00) == 0) {
4642 // Value = 0xnn00: Op=x, Cmode=101x.
4644 Imm = SplatBits >> 8;
4650 // NEON's 32-bit VMOV supports splat values where:
4651 // * only one byte is nonzero, or
4652 // * the least significant byte is 0xff and the second byte is nonzero, or
4653 // * the least significant 2 bytes are 0xff and the third is nonzero.
4654 VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
4655 if ((SplatBits & ~0xff) == 0) {
4656 // Value = 0x000000nn: Op=x, Cmode=000x.
4661 if ((SplatBits & ~0xff00) == 0) {
4662 // Value = 0x0000nn00: Op=x, Cmode=001x.
4664 Imm = SplatBits >> 8;
4667 if ((SplatBits & ~0xff0000) == 0) {
4668 // Value = 0x00nn0000: Op=x, Cmode=010x.
4670 Imm = SplatBits >> 16;
4673 if ((SplatBits & ~0xff000000) == 0) {
4674 // Value = 0xnn000000: Op=x, Cmode=011x.
4676 Imm = SplatBits >> 24;
4680 // cmode == 0b1100 and cmode == 0b1101 are not supported for VORR or VBIC
4681 if (type == OtherModImm) return SDValue();
4683 if ((SplatBits & ~0xffff) == 0 &&
4684 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
4685 // Value = 0x0000nnff: Op=x, Cmode=1100.
4687 Imm = SplatBits >> 8;
4691 if ((SplatBits & ~0xffffff) == 0 &&
4692 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
4693 // Value = 0x00nnffff: Op=x, Cmode=1101.
4695 Imm = SplatBits >> 16;
4699 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
4700 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
4701 // VMOV.I32. A (very) minor optimization would be to replicate the value
4702 // and fall through here to test for a valid 64-bit splat. But, then the
4703 // caller would also need to check and handle the change in size.
4707 if (type != VMOVModImm)
4709 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
4710 uint64_t BitMask = 0xff;
4712 unsigned ImmMask = 1;
4714 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
4715 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
4718 } else if ((SplatBits & BitMask) != 0) {
4725 if (DAG.getTargetLoweringInfo().isBigEndian())
4726 // swap higher and lower 32 bit word
4727 Imm = ((Imm & 0xf) << 4) | ((Imm & 0xf0) >> 4);
4729 // Op=1, Cmode=1110.
4731 VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
4736 llvm_unreachable("unexpected size for isNEONModifiedImm");
4739 unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm);
4740 return DAG.getTargetConstant(EncodedVal, MVT::i32);
4743 SDValue ARMTargetLowering::LowerConstantFP(SDValue Op, SelectionDAG &DAG,
4744 const ARMSubtarget *ST) const {
4748 bool IsDouble = Op.getValueType() == MVT::f64;
4749 ConstantFPSDNode *CFP = cast<ConstantFPSDNode>(Op);
4751 // Use the default (constant pool) lowering for double constants when we have
4753 if (IsDouble && Subtarget->isFPOnlySP())
4756 // Try splatting with a VMOV.f32...
4757 APFloat FPVal = CFP->getValueAPF();
4758 int ImmVal = IsDouble ? ARM_AM::getFP64Imm(FPVal) : ARM_AM::getFP32Imm(FPVal);
4761 if (IsDouble || !ST->useNEONForSinglePrecisionFP()) {
4762 // We have code in place to select a valid ConstantFP already, no need to
4767 // It's a float and we are trying to use NEON operations where
4768 // possible. Lower it to a splat followed by an extract.
4770 SDValue NewVal = DAG.getTargetConstant(ImmVal, MVT::i32);
4771 SDValue VecConstant = DAG.getNode(ARMISD::VMOVFPIMM, DL, MVT::v2f32,
4773 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecConstant,
4774 DAG.getConstant(0, MVT::i32));
4777 // The rest of our options are NEON only, make sure that's allowed before
4779 if (!ST->hasNEON() || (!IsDouble && !ST->useNEONForSinglePrecisionFP()))
4783 uint64_t iVal = FPVal.bitcastToAPInt().getZExtValue();
4785 // It wouldn't really be worth bothering for doubles except for one very
4786 // important value, which does happen to match: 0.0. So make sure we don't do
4788 if (IsDouble && (iVal & 0xffffffff) != (iVal >> 32))
4791 // Try a VMOV.i32 (FIXME: i8, i16, or i64 could work too).
4792 SDValue NewVal = isNEONModifiedImm(iVal & 0xffffffffU, 0, 32, DAG, VMovVT,
4794 if (NewVal != SDValue()) {
4796 SDValue VecConstant = DAG.getNode(ARMISD::VMOVIMM, DL, VMovVT,
4799 return DAG.getNode(ISD::BITCAST, DL, MVT::f64, VecConstant);
4801 // It's a float: cast and extract a vector element.
4802 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
4804 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
4805 DAG.getConstant(0, MVT::i32));
4808 // Finally, try a VMVN.i32
4809 NewVal = isNEONModifiedImm(~iVal & 0xffffffffU, 0, 32, DAG, VMovVT,
4811 if (NewVal != SDValue()) {
4813 SDValue VecConstant = DAG.getNode(ARMISD::VMVNIMM, DL, VMovVT, NewVal);
4816 return DAG.getNode(ISD::BITCAST, DL, MVT::f64, VecConstant);
4818 // It's a float: cast and extract a vector element.
4819 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
4821 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
4822 DAG.getConstant(0, MVT::i32));
4828 // check if an VEXT instruction can handle the shuffle mask when the
4829 // vector sources of the shuffle are the same.
4830 static bool isSingletonVEXTMask(ArrayRef<int> M, EVT VT, unsigned &Imm) {
4831 unsigned NumElts = VT.getVectorNumElements();
4833 // Assume that the first shuffle index is not UNDEF. Fail if it is.
4839 // If this is a VEXT shuffle, the immediate value is the index of the first
4840 // element. The other shuffle indices must be the successive elements after
4842 unsigned ExpectedElt = Imm;
4843 for (unsigned i = 1; i < NumElts; ++i) {
4844 // Increment the expected index. If it wraps around, just follow it
4845 // back to index zero and keep going.
4847 if (ExpectedElt == NumElts)
4850 if (M[i] < 0) continue; // ignore UNDEF indices
4851 if (ExpectedElt != static_cast<unsigned>(M[i]))
4859 static bool isVEXTMask(ArrayRef<int> M, EVT VT,
4860 bool &ReverseVEXT, unsigned &Imm) {
4861 unsigned NumElts = VT.getVectorNumElements();
4862 ReverseVEXT = false;
4864 // Assume that the first shuffle index is not UNDEF. Fail if it is.
4870 // If this is a VEXT shuffle, the immediate value is the index of the first
4871 // element. The other shuffle indices must be the successive elements after
4873 unsigned ExpectedElt = Imm;
4874 for (unsigned i = 1; i < NumElts; ++i) {
4875 // Increment the expected index. If it wraps around, it may still be
4876 // a VEXT but the source vectors must be swapped.
4878 if (ExpectedElt == NumElts * 2) {
4883 if (M[i] < 0) continue; // ignore UNDEF indices
4884 if (ExpectedElt != static_cast<unsigned>(M[i]))
4888 // Adjust the index value if the source operands will be swapped.
4895 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
4896 /// instruction with the specified blocksize. (The order of the elements
4897 /// within each block of the vector is reversed.)
4898 static bool isVREVMask(ArrayRef<int> M, EVT VT, unsigned BlockSize) {
4899 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
4900 "Only possible block sizes for VREV are: 16, 32, 64");
4902 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4906 unsigned NumElts = VT.getVectorNumElements();
4907 unsigned BlockElts = M[0] + 1;
4908 // If the first shuffle index is UNDEF, be optimistic.
4910 BlockElts = BlockSize / EltSz;
4912 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
4915 for (unsigned i = 0; i < NumElts; ++i) {
4916 if (M[i] < 0) continue; // ignore UNDEF indices
4917 if ((unsigned) M[i] != (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
4924 static bool isVTBLMask(ArrayRef<int> M, EVT VT) {
4925 // We can handle <8 x i8> vector shuffles. If the index in the mask is out of
4926 // range, then 0 is placed into the resulting vector. So pretty much any mask
4927 // of 8 elements can work here.
4928 return VT == MVT::v8i8 && M.size() == 8;
4931 static bool isVTRNMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4932 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4936 unsigned NumElts = VT.getVectorNumElements();
4937 WhichResult = (M[0] == 0 ? 0 : 1);
4938 for (unsigned i = 0; i < NumElts; i += 2) {
4939 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
4940 (M[i+1] >= 0 && (unsigned) M[i+1] != i + NumElts + WhichResult))
4946 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
4947 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4948 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
4949 static bool isVTRN_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4950 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4954 unsigned NumElts = VT.getVectorNumElements();
4955 WhichResult = (M[0] == 0 ? 0 : 1);
4956 for (unsigned i = 0; i < NumElts; i += 2) {
4957 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
4958 (M[i+1] >= 0 && (unsigned) M[i+1] != i + WhichResult))
4964 static bool isVUZPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4965 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4969 unsigned NumElts = VT.getVectorNumElements();
4970 WhichResult = (M[0] == 0 ? 0 : 1);
4971 for (unsigned i = 0; i != NumElts; ++i) {
4972 if (M[i] < 0) continue; // ignore UNDEF indices
4973 if ((unsigned) M[i] != 2 * i + WhichResult)
4977 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4978 if (VT.is64BitVector() && EltSz == 32)
4984 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
4985 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4986 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
4987 static bool isVUZP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4988 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4992 unsigned Half = VT.getVectorNumElements() / 2;
4993 WhichResult = (M[0] == 0 ? 0 : 1);
4994 for (unsigned j = 0; j != 2; ++j) {
4995 unsigned Idx = WhichResult;
4996 for (unsigned i = 0; i != Half; ++i) {
4997 int MIdx = M[i + j * Half];
4998 if (MIdx >= 0 && (unsigned) MIdx != Idx)
5004 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
5005 if (VT.is64BitVector() && EltSz == 32)
5011 static bool isVZIPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
5012 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
5016 unsigned NumElts = VT.getVectorNumElements();
5017 WhichResult = (M[0] == 0 ? 0 : 1);
5018 unsigned Idx = WhichResult * NumElts / 2;
5019 for (unsigned i = 0; i != NumElts; i += 2) {
5020 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
5021 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx + NumElts))
5026 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
5027 if (VT.is64BitVector() && EltSz == 32)
5033 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
5034 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
5035 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
5036 static bool isVZIP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
5037 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
5041 unsigned NumElts = VT.getVectorNumElements();
5042 WhichResult = (M[0] == 0 ? 0 : 1);
5043 unsigned Idx = WhichResult * NumElts / 2;
5044 for (unsigned i = 0; i != NumElts; i += 2) {
5045 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
5046 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx))
5051 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
5052 if (VT.is64BitVector() && EltSz == 32)
5058 /// \return true if this is a reverse operation on an vector.
5059 static bool isReverseMask(ArrayRef<int> M, EVT VT) {
5060 unsigned NumElts = VT.getVectorNumElements();
5061 // Make sure the mask has the right size.
5062 if (NumElts != M.size())
5065 // Look for <15, ..., 3, -1, 1, 0>.
5066 for (unsigned i = 0; i != NumElts; ++i)
5067 if (M[i] >= 0 && M[i] != (int) (NumElts - 1 - i))
5073 // If N is an integer constant that can be moved into a register in one
5074 // instruction, return an SDValue of such a constant (will become a MOV
5075 // instruction). Otherwise return null.
5076 static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG,
5077 const ARMSubtarget *ST, SDLoc dl) {
5079 if (!isa<ConstantSDNode>(N))
5081 Val = cast<ConstantSDNode>(N)->getZExtValue();
5083 if (ST->isThumb1Only()) {
5084 if (Val <= 255 || ~Val <= 255)
5085 return DAG.getConstant(Val, MVT::i32);
5087 if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1)
5088 return DAG.getConstant(Val, MVT::i32);
5093 // If this is a case we can't handle, return null and let the default
5094 // expansion code take care of it.
5095 SDValue ARMTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
5096 const ARMSubtarget *ST) const {
5097 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
5099 EVT VT = Op.getValueType();
5101 APInt SplatBits, SplatUndef;
5102 unsigned SplatBitSize;
5104 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
5105 if (SplatBitSize <= 64) {
5106 // Check if an immediate VMOV works.
5108 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
5109 SplatUndef.getZExtValue(), SplatBitSize,
5110 DAG, VmovVT, VT.is128BitVector(),
5112 if (Val.getNode()) {
5113 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
5114 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
5117 // Try an immediate VMVN.
5118 uint64_t NegatedImm = (~SplatBits).getZExtValue();
5119 Val = isNEONModifiedImm(NegatedImm,
5120 SplatUndef.getZExtValue(), SplatBitSize,
5121 DAG, VmovVT, VT.is128BitVector(),
5123 if (Val.getNode()) {
5124 SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
5125 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
5128 // Use vmov.f32 to materialize other v2f32 and v4f32 splats.
5129 if ((VT == MVT::v2f32 || VT == MVT::v4f32) && SplatBitSize == 32) {
5130 int ImmVal = ARM_AM::getFP32Imm(SplatBits);
5132 SDValue Val = DAG.getTargetConstant(ImmVal, MVT::i32);
5133 return DAG.getNode(ARMISD::VMOVFPIMM, dl, VT, Val);
5139 // Scan through the operands to see if only one value is used.
5141 // As an optimisation, even if more than one value is used it may be more
5142 // profitable to splat with one value then change some lanes.
5144 // Heuristically we decide to do this if the vector has a "dominant" value,
5145 // defined as splatted to more than half of the lanes.
5146 unsigned NumElts = VT.getVectorNumElements();
5147 bool isOnlyLowElement = true;
5148 bool usesOnlyOneValue = true;
5149 bool hasDominantValue = false;
5150 bool isConstant = true;
5152 // Map of the number of times a particular SDValue appears in the
5154 DenseMap<SDValue, unsigned> ValueCounts;
5156 for (unsigned i = 0; i < NumElts; ++i) {
5157 SDValue V = Op.getOperand(i);
5158 if (V.getOpcode() == ISD::UNDEF)
5161 isOnlyLowElement = false;
5162 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
5165 ValueCounts.insert(std::make_pair(V, 0));
5166 unsigned &Count = ValueCounts[V];
5168 // Is this value dominant? (takes up more than half of the lanes)
5169 if (++Count > (NumElts / 2)) {
5170 hasDominantValue = true;
5174 if (ValueCounts.size() != 1)
5175 usesOnlyOneValue = false;
5176 if (!Value.getNode() && ValueCounts.size() > 0)
5177 Value = ValueCounts.begin()->first;
5179 if (ValueCounts.size() == 0)
5180 return DAG.getUNDEF(VT);
5182 // Loads are better lowered with insert_vector_elt/ARMISD::BUILD_VECTOR.
5183 // Keep going if we are hitting this case.
5184 if (isOnlyLowElement && !ISD::isNormalLoad(Value.getNode()))
5185 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
5187 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5189 // Use VDUP for non-constant splats. For f32 constant splats, reduce to
5190 // i32 and try again.
5191 if (hasDominantValue && EltSize <= 32) {
5195 // If we are VDUPing a value that comes directly from a vector, that will
5196 // cause an unnecessary move to and from a GPR, where instead we could
5197 // just use VDUPLANE. We can only do this if the lane being extracted
5198 // is at a constant index, as the VDUP from lane instructions only have
5199 // constant-index forms.
5200 if (Value->getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5201 isa<ConstantSDNode>(Value->getOperand(1))) {
5202 // We need to create a new undef vector to use for the VDUPLANE if the
5203 // size of the vector from which we get the value is different than the
5204 // size of the vector that we need to create. We will insert the element
5205 // such that the register coalescer will remove unnecessary copies.
5206 if (VT != Value->getOperand(0).getValueType()) {
5207 ConstantSDNode *constIndex;
5208 constIndex = dyn_cast<ConstantSDNode>(Value->getOperand(1));
5209 assert(constIndex && "The index is not a constant!");
5210 unsigned index = constIndex->getAPIntValue().getLimitedValue() %
5211 VT.getVectorNumElements();
5212 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
5213 DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DAG.getUNDEF(VT),
5214 Value, DAG.getConstant(index, MVT::i32)),
5215 DAG.getConstant(index, MVT::i32));
5217 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
5218 Value->getOperand(0), Value->getOperand(1));
5220 N = DAG.getNode(ARMISD::VDUP, dl, VT, Value);
5222 if (!usesOnlyOneValue) {
5223 // The dominant value was splatted as 'N', but we now have to insert
5224 // all differing elements.
5225 for (unsigned I = 0; I < NumElts; ++I) {
5226 if (Op.getOperand(I) == Value)
5228 SmallVector<SDValue, 3> Ops;
5230 Ops.push_back(Op.getOperand(I));
5231 Ops.push_back(DAG.getConstant(I, MVT::i32));
5232 N = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Ops);
5237 if (VT.getVectorElementType().isFloatingPoint()) {
5238 SmallVector<SDValue, 8> Ops;
5239 for (unsigned i = 0; i < NumElts; ++i)
5240 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, MVT::i32,
5242 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
5243 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, Ops);
5244 Val = LowerBUILD_VECTOR(Val, DAG, ST);
5246 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
5248 if (usesOnlyOneValue) {
5249 SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
5250 if (isConstant && Val.getNode())
5251 return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
5255 // If all elements are constants and the case above didn't get hit, fall back
5256 // to the default expansion, which will generate a load from the constant
5261 // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
5263 SDValue shuffle = ReconstructShuffle(Op, DAG);
5264 if (shuffle != SDValue())
5268 // Vectors with 32- or 64-bit elements can be built by directly assigning
5269 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
5270 // will be legalized.
5271 if (EltSize >= 32) {
5272 // Do the expansion with floating-point types, since that is what the VFP
5273 // registers are defined to use, and since i64 is not legal.
5274 EVT EltVT = EVT::getFloatingPointVT(EltSize);
5275 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
5276 SmallVector<SDValue, 8> Ops;
5277 for (unsigned i = 0; i < NumElts; ++i)
5278 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, EltVT, Op.getOperand(i)));
5279 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, Ops);
5280 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
5283 // If all else fails, just use a sequence of INSERT_VECTOR_ELT when we
5284 // know the default expansion would otherwise fall back on something even
5285 // worse. For a vector with one or two non-undef values, that's
5286 // scalar_to_vector for the elements followed by a shuffle (provided the
5287 // shuffle is valid for the target) and materialization element by element
5288 // on the stack followed by a load for everything else.
5289 if (!isConstant && !usesOnlyOneValue) {
5290 SDValue Vec = DAG.getUNDEF(VT);
5291 for (unsigned i = 0 ; i < NumElts; ++i) {
5292 SDValue V = Op.getOperand(i);
5293 if (V.getOpcode() == ISD::UNDEF)
5295 SDValue LaneIdx = DAG.getConstant(i, MVT::i32);
5296 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Vec, V, LaneIdx);
5304 // Gather data to see if the operation can be modelled as a
5305 // shuffle in combination with VEXTs.
5306 SDValue ARMTargetLowering::ReconstructShuffle(SDValue Op,
5307 SelectionDAG &DAG) const {
5309 EVT VT = Op.getValueType();
5310 unsigned NumElts = VT.getVectorNumElements();
5312 SmallVector<SDValue, 2> SourceVecs;
5313 SmallVector<unsigned, 2> MinElts;
5314 SmallVector<unsigned, 2> MaxElts;
5316 for (unsigned i = 0; i < NumElts; ++i) {
5317 SDValue V = Op.getOperand(i);
5318 if (V.getOpcode() == ISD::UNDEF)
5320 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
5321 // A shuffle can only come from building a vector from various
5322 // elements of other vectors.
5324 } else if (V.getOperand(0).getValueType().getVectorElementType() !=
5325 VT.getVectorElementType()) {
5326 // This code doesn't know how to handle shuffles where the vector
5327 // element types do not match (this happens because type legalization
5328 // promotes the return type of EXTRACT_VECTOR_ELT).
5329 // FIXME: It might be appropriate to extend this code to handle
5330 // mismatched types.
5334 // Record this extraction against the appropriate vector if possible...
5335 SDValue SourceVec = V.getOperand(0);
5336 // If the element number isn't a constant, we can't effectively
5337 // analyze what's going on.
5338 if (!isa<ConstantSDNode>(V.getOperand(1)))
5340 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
5341 bool FoundSource = false;
5342 for (unsigned j = 0; j < SourceVecs.size(); ++j) {
5343 if (SourceVecs[j] == SourceVec) {
5344 if (MinElts[j] > EltNo)
5346 if (MaxElts[j] < EltNo)
5353 // Or record a new source if not...
5355 SourceVecs.push_back(SourceVec);
5356 MinElts.push_back(EltNo);
5357 MaxElts.push_back(EltNo);
5361 // Currently only do something sane when at most two source vectors
5363 if (SourceVecs.size() > 2)
5366 SDValue ShuffleSrcs[2] = {DAG.getUNDEF(VT), DAG.getUNDEF(VT) };
5367 int VEXTOffsets[2] = {0, 0};
5369 // This loop extracts the usage patterns of the source vectors
5370 // and prepares appropriate SDValues for a shuffle if possible.
5371 for (unsigned i = 0; i < SourceVecs.size(); ++i) {
5372 if (SourceVecs[i].getValueType() == VT) {
5373 // No VEXT necessary
5374 ShuffleSrcs[i] = SourceVecs[i];
5377 } else if (SourceVecs[i].getValueType().getVectorNumElements() < NumElts) {
5378 // It probably isn't worth padding out a smaller vector just to
5379 // break it down again in a shuffle.
5383 // Since only 64-bit and 128-bit vectors are legal on ARM and
5384 // we've eliminated the other cases...
5385 assert(SourceVecs[i].getValueType().getVectorNumElements() == 2*NumElts &&
5386 "unexpected vector sizes in ReconstructShuffle");
5388 if (MaxElts[i] - MinElts[i] >= NumElts) {
5389 // Span too large for a VEXT to cope
5393 if (MinElts[i] >= NumElts) {
5394 // The extraction can just take the second half
5395 VEXTOffsets[i] = NumElts;
5396 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5398 DAG.getIntPtrConstant(NumElts));
5399 } else if (MaxElts[i] < NumElts) {
5400 // The extraction can just take the first half
5402 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5404 DAG.getIntPtrConstant(0));
5406 // An actual VEXT is needed
5407 VEXTOffsets[i] = MinElts[i];
5408 SDValue VEXTSrc1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5410 DAG.getIntPtrConstant(0));
5411 SDValue VEXTSrc2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5413 DAG.getIntPtrConstant(NumElts));
5414 ShuffleSrcs[i] = DAG.getNode(ARMISD::VEXT, dl, VT, VEXTSrc1, VEXTSrc2,
5415 DAG.getConstant(VEXTOffsets[i], MVT::i32));
5419 SmallVector<int, 8> Mask;
5421 for (unsigned i = 0; i < NumElts; ++i) {
5422 SDValue Entry = Op.getOperand(i);
5423 if (Entry.getOpcode() == ISD::UNDEF) {
5428 SDValue ExtractVec = Entry.getOperand(0);
5429 int ExtractElt = cast<ConstantSDNode>(Op.getOperand(i)
5430 .getOperand(1))->getSExtValue();
5431 if (ExtractVec == SourceVecs[0]) {
5432 Mask.push_back(ExtractElt - VEXTOffsets[0]);
5434 Mask.push_back(ExtractElt + NumElts - VEXTOffsets[1]);
5438 // Final check before we try to produce nonsense...
5439 if (isShuffleMaskLegal(Mask, VT))
5440 return DAG.getVectorShuffle(VT, dl, ShuffleSrcs[0], ShuffleSrcs[1],
5446 /// isShuffleMaskLegal - Targets can use this to indicate that they only
5447 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
5448 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
5449 /// are assumed to be legal.
5451 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
5453 if (VT.getVectorNumElements() == 4 &&
5454 (VT.is128BitVector() || VT.is64BitVector())) {
5455 unsigned PFIndexes[4];
5456 for (unsigned i = 0; i != 4; ++i) {
5460 PFIndexes[i] = M[i];
5463 // Compute the index in the perfect shuffle table.
5464 unsigned PFTableIndex =
5465 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
5466 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
5467 unsigned Cost = (PFEntry >> 30);
5474 unsigned Imm, WhichResult;
5476 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5477 return (EltSize >= 32 ||
5478 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
5479 isVREVMask(M, VT, 64) ||
5480 isVREVMask(M, VT, 32) ||
5481 isVREVMask(M, VT, 16) ||
5482 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
5483 isVTBLMask(M, VT) ||
5484 isVTRNMask(M, VT, WhichResult) ||
5485 isVUZPMask(M, VT, WhichResult) ||
5486 isVZIPMask(M, VT, WhichResult) ||
5487 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
5488 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
5489 isVZIP_v_undef_Mask(M, VT, WhichResult) ||
5490 ((VT == MVT::v8i16 || VT == MVT::v16i8) && isReverseMask(M, VT)));
5493 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
5494 /// the specified operations to build the shuffle.
5495 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
5496 SDValue RHS, SelectionDAG &DAG,
5498 unsigned OpNum = (PFEntry >> 26) & 0x0F;
5499 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
5500 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
5503 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
5512 OP_VUZPL, // VUZP, left result
5513 OP_VUZPR, // VUZP, right result
5514 OP_VZIPL, // VZIP, left result
5515 OP_VZIPR, // VZIP, right result
5516 OP_VTRNL, // VTRN, left result
5517 OP_VTRNR // VTRN, right result
5520 if (OpNum == OP_COPY) {
5521 if (LHSID == (1*9+2)*9+3) return LHS;
5522 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
5526 SDValue OpLHS, OpRHS;
5527 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
5528 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
5529 EVT VT = OpLHS.getValueType();
5532 default: llvm_unreachable("Unknown shuffle opcode!");
5534 // VREV divides the vector in half and swaps within the half.
5535 if (VT.getVectorElementType() == MVT::i32 ||
5536 VT.getVectorElementType() == MVT::f32)
5537 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
5538 // vrev <4 x i16> -> VREV32
5539 if (VT.getVectorElementType() == MVT::i16)
5540 return DAG.getNode(ARMISD::VREV32, dl, VT, OpLHS);
5541 // vrev <4 x i8> -> VREV16
5542 assert(VT.getVectorElementType() == MVT::i8);
5543 return DAG.getNode(ARMISD::VREV16, dl, VT, OpLHS);
5548 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
5549 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
5553 return DAG.getNode(ARMISD::VEXT, dl, VT,
5555 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
5558 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
5559 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
5562 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
5563 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
5566 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
5567 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
5571 static SDValue LowerVECTOR_SHUFFLEv8i8(SDValue Op,
5572 ArrayRef<int> ShuffleMask,
5573 SelectionDAG &DAG) {
5574 // Check to see if we can use the VTBL instruction.
5575 SDValue V1 = Op.getOperand(0);
5576 SDValue V2 = Op.getOperand(1);
5579 SmallVector<SDValue, 8> VTBLMask;
5580 for (ArrayRef<int>::iterator
5581 I = ShuffleMask.begin(), E = ShuffleMask.end(); I != E; ++I)
5582 VTBLMask.push_back(DAG.getConstant(*I, MVT::i32));
5584 if (V2.getNode()->getOpcode() == ISD::UNDEF)
5585 return DAG.getNode(ARMISD::VTBL1, DL, MVT::v8i8, V1,
5586 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8, VTBLMask));
5588 return DAG.getNode(ARMISD::VTBL2, DL, MVT::v8i8, V1, V2,
5589 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8, VTBLMask));
5592 static SDValue LowerReverse_VECTOR_SHUFFLEv16i8_v8i16(SDValue Op,
5593 SelectionDAG &DAG) {
5595 SDValue OpLHS = Op.getOperand(0);
5596 EVT VT = OpLHS.getValueType();
5598 assert((VT == MVT::v8i16 || VT == MVT::v16i8) &&
5599 "Expect an v8i16/v16i8 type");
5600 OpLHS = DAG.getNode(ARMISD::VREV64, DL, VT, OpLHS);
5601 // For a v16i8 type: After the VREV, we have got <8, ...15, 8, ..., 0>. Now,
5602 // extract the first 8 bytes into the top double word and the last 8 bytes
5603 // into the bottom double word. The v8i16 case is similar.
5604 unsigned ExtractNum = (VT == MVT::v16i8) ? 8 : 4;
5605 return DAG.getNode(ARMISD::VEXT, DL, VT, OpLHS, OpLHS,
5606 DAG.getConstant(ExtractNum, MVT::i32));
5609 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
5610 SDValue V1 = Op.getOperand(0);
5611 SDValue V2 = Op.getOperand(1);
5613 EVT VT = Op.getValueType();
5614 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
5616 // Convert shuffles that are directly supported on NEON to target-specific
5617 // DAG nodes, instead of keeping them as shuffles and matching them again
5618 // during code selection. This is more efficient and avoids the possibility
5619 // of inconsistencies between legalization and selection.
5620 // FIXME: floating-point vectors should be canonicalized to integer vectors
5621 // of the same time so that they get CSEd properly.
5622 ArrayRef<int> ShuffleMask = SVN->getMask();
5624 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5625 if (EltSize <= 32) {
5626 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
5627 int Lane = SVN->getSplatIndex();
5628 // If this is undef splat, generate it via "just" vdup, if possible.
5629 if (Lane == -1) Lane = 0;
5631 // Test if V1 is a SCALAR_TO_VECTOR.
5632 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
5633 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
5635 // Test if V1 is a BUILD_VECTOR which is equivalent to a SCALAR_TO_VECTOR
5636 // (and probably will turn into a SCALAR_TO_VECTOR once legalization
5638 if (Lane == 0 && V1.getOpcode() == ISD::BUILD_VECTOR &&
5639 !isa<ConstantSDNode>(V1.getOperand(0))) {
5640 bool IsScalarToVector = true;
5641 for (unsigned i = 1, e = V1.getNumOperands(); i != e; ++i)
5642 if (V1.getOperand(i).getOpcode() != ISD::UNDEF) {
5643 IsScalarToVector = false;
5646 if (IsScalarToVector)
5647 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
5649 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
5650 DAG.getConstant(Lane, MVT::i32));
5655 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
5658 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
5659 DAG.getConstant(Imm, MVT::i32));
5662 if (isVREVMask(ShuffleMask, VT, 64))
5663 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
5664 if (isVREVMask(ShuffleMask, VT, 32))
5665 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
5666 if (isVREVMask(ShuffleMask, VT, 16))
5667 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
5669 if (V2->getOpcode() == ISD::UNDEF &&
5670 isSingletonVEXTMask(ShuffleMask, VT, Imm)) {
5671 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V1,
5672 DAG.getConstant(Imm, MVT::i32));
5675 // Check for Neon shuffles that modify both input vectors in place.
5676 // If both results are used, i.e., if there are two shuffles with the same
5677 // source operands and with masks corresponding to both results of one of
5678 // these operations, DAG memoization will ensure that a single node is
5679 // used for both shuffles.
5680 unsigned WhichResult;
5681 if (isVTRNMask(ShuffleMask, VT, WhichResult))
5682 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
5683 V1, V2).getValue(WhichResult);
5684 if (isVUZPMask(ShuffleMask, VT, WhichResult))
5685 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
5686 V1, V2).getValue(WhichResult);
5687 if (isVZIPMask(ShuffleMask, VT, WhichResult))
5688 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
5689 V1, V2).getValue(WhichResult);
5691 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
5692 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
5693 V1, V1).getValue(WhichResult);
5694 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
5695 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
5696 V1, V1).getValue(WhichResult);
5697 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
5698 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
5699 V1, V1).getValue(WhichResult);
5702 // If the shuffle is not directly supported and it has 4 elements, use
5703 // the PerfectShuffle-generated table to synthesize it from other shuffles.
5704 unsigned NumElts = VT.getVectorNumElements();
5706 unsigned PFIndexes[4];
5707 for (unsigned i = 0; i != 4; ++i) {
5708 if (ShuffleMask[i] < 0)
5711 PFIndexes[i] = ShuffleMask[i];
5714 // Compute the index in the perfect shuffle table.
5715 unsigned PFTableIndex =
5716 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
5717 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
5718 unsigned Cost = (PFEntry >> 30);
5721 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
5724 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
5725 if (EltSize >= 32) {
5726 // Do the expansion with floating-point types, since that is what the VFP
5727 // registers are defined to use, and since i64 is not legal.
5728 EVT EltVT = EVT::getFloatingPointVT(EltSize);
5729 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
5730 V1 = DAG.getNode(ISD::BITCAST, dl, VecVT, V1);
5731 V2 = DAG.getNode(ISD::BITCAST, dl, VecVT, V2);
5732 SmallVector<SDValue, 8> Ops;
5733 for (unsigned i = 0; i < NumElts; ++i) {
5734 if (ShuffleMask[i] < 0)
5735 Ops.push_back(DAG.getUNDEF(EltVT));
5737 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
5738 ShuffleMask[i] < (int)NumElts ? V1 : V2,
5739 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
5742 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, Ops);
5743 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
5746 if ((VT == MVT::v8i16 || VT == MVT::v16i8) && isReverseMask(ShuffleMask, VT))
5747 return LowerReverse_VECTOR_SHUFFLEv16i8_v8i16(Op, DAG);
5749 if (VT == MVT::v8i8) {
5750 SDValue NewOp = LowerVECTOR_SHUFFLEv8i8(Op, ShuffleMask, DAG);
5751 if (NewOp.getNode())
5758 static SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
5759 // INSERT_VECTOR_ELT is legal only for immediate indexes.
5760 SDValue Lane = Op.getOperand(2);
5761 if (!isa<ConstantSDNode>(Lane))
5767 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
5768 // EXTRACT_VECTOR_ELT is legal only for immediate indexes.
5769 SDValue Lane = Op.getOperand(1);
5770 if (!isa<ConstantSDNode>(Lane))
5773 SDValue Vec = Op.getOperand(0);
5774 if (Op.getValueType() == MVT::i32 &&
5775 Vec.getValueType().getVectorElementType().getSizeInBits() < 32) {
5777 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
5783 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
5784 // The only time a CONCAT_VECTORS operation can have legal types is when
5785 // two 64-bit vectors are concatenated to a 128-bit vector.
5786 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
5787 "unexpected CONCAT_VECTORS");
5789 SDValue Val = DAG.getUNDEF(MVT::v2f64);
5790 SDValue Op0 = Op.getOperand(0);
5791 SDValue Op1 = Op.getOperand(1);
5792 if (Op0.getOpcode() != ISD::UNDEF)
5793 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
5794 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op0),
5795 DAG.getIntPtrConstant(0));
5796 if (Op1.getOpcode() != ISD::UNDEF)
5797 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
5798 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op1),
5799 DAG.getIntPtrConstant(1));
5800 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Val);
5803 /// isExtendedBUILD_VECTOR - Check if N is a constant BUILD_VECTOR where each
5804 /// element has been zero/sign-extended, depending on the isSigned parameter,
5805 /// from an integer type half its size.
5806 static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG,
5808 // A v2i64 BUILD_VECTOR will have been legalized to a BITCAST from v4i32.
5809 EVT VT = N->getValueType(0);
5810 if (VT == MVT::v2i64 && N->getOpcode() == ISD::BITCAST) {
5811 SDNode *BVN = N->getOperand(0).getNode();
5812 if (BVN->getValueType(0) != MVT::v4i32 ||
5813 BVN->getOpcode() != ISD::BUILD_VECTOR)
5815 unsigned LoElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
5816 unsigned HiElt = 1 - LoElt;
5817 ConstantSDNode *Lo0 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt));
5818 ConstantSDNode *Hi0 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt));
5819 ConstantSDNode *Lo1 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt+2));
5820 ConstantSDNode *Hi1 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt+2));
5821 if (!Lo0 || !Hi0 || !Lo1 || !Hi1)
5824 if (Hi0->getSExtValue() == Lo0->getSExtValue() >> 32 &&
5825 Hi1->getSExtValue() == Lo1->getSExtValue() >> 32)
5828 if (Hi0->isNullValue() && Hi1->isNullValue())
5834 if (N->getOpcode() != ISD::BUILD_VECTOR)
5837 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
5838 SDNode *Elt = N->getOperand(i).getNode();
5839 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
5840 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5841 unsigned HalfSize = EltSize / 2;
5843 if (!isIntN(HalfSize, C->getSExtValue()))
5846 if (!isUIntN(HalfSize, C->getZExtValue()))
5857 /// isSignExtended - Check if a node is a vector value that is sign-extended
5858 /// or a constant BUILD_VECTOR with sign-extended elements.
5859 static bool isSignExtended(SDNode *N, SelectionDAG &DAG) {
5860 if (N->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N))
5862 if (isExtendedBUILD_VECTOR(N, DAG, true))
5867 /// isZeroExtended - Check if a node is a vector value that is zero-extended
5868 /// or a constant BUILD_VECTOR with zero-extended elements.
5869 static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) {
5870 if (N->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N))
5872 if (isExtendedBUILD_VECTOR(N, DAG, false))
5877 static EVT getExtensionTo64Bits(const EVT &OrigVT) {
5878 if (OrigVT.getSizeInBits() >= 64)
5881 assert(OrigVT.isSimple() && "Expecting a simple value type");
5883 MVT::SimpleValueType OrigSimpleTy = OrigVT.getSimpleVT().SimpleTy;
5884 switch (OrigSimpleTy) {
5885 default: llvm_unreachable("Unexpected Vector Type");
5894 /// AddRequiredExtensionForVMULL - Add a sign/zero extension to extend the total
5895 /// value size to 64 bits. We need a 64-bit D register as an operand to VMULL.
5896 /// We insert the required extension here to get the vector to fill a D register.
5897 static SDValue AddRequiredExtensionForVMULL(SDValue N, SelectionDAG &DAG,
5900 unsigned ExtOpcode) {
5901 // The vector originally had a size of OrigTy. It was then extended to ExtTy.
5902 // We expect the ExtTy to be 128-bits total. If the OrigTy is less than
5903 // 64-bits we need to insert a new extension so that it will be 64-bits.
5904 assert(ExtTy.is128BitVector() && "Unexpected extension size");
5905 if (OrigTy.getSizeInBits() >= 64)
5908 // Must extend size to at least 64 bits to be used as an operand for VMULL.
5909 EVT NewVT = getExtensionTo64Bits(OrigTy);
5911 return DAG.getNode(ExtOpcode, SDLoc(N), NewVT, N);
5914 /// SkipLoadExtensionForVMULL - return a load of the original vector size that
5915 /// does not do any sign/zero extension. If the original vector is less
5916 /// than 64 bits, an appropriate extension will be added after the load to
5917 /// reach a total size of 64 bits. We have to add the extension separately
5918 /// because ARM does not have a sign/zero extending load for vectors.
5919 static SDValue SkipLoadExtensionForVMULL(LoadSDNode *LD, SelectionDAG& DAG) {
5920 EVT ExtendedTy = getExtensionTo64Bits(LD->getMemoryVT());
5922 // The load already has the right type.
5923 if (ExtendedTy == LD->getMemoryVT())
5924 return DAG.getLoad(LD->getMemoryVT(), SDLoc(LD), LD->getChain(),
5925 LD->getBasePtr(), LD->getPointerInfo(), LD->isVolatile(),
5926 LD->isNonTemporal(), LD->isInvariant(),
5927 LD->getAlignment());
5929 // We need to create a zextload/sextload. We cannot just create a load
5930 // followed by a zext/zext node because LowerMUL is also run during normal
5931 // operation legalization where we can't create illegal types.
5932 return DAG.getExtLoad(LD->getExtensionType(), SDLoc(LD), ExtendedTy,
5933 LD->getChain(), LD->getBasePtr(), LD->getPointerInfo(),
5934 LD->getMemoryVT(), LD->isVolatile(), LD->isInvariant(),
5935 LD->isNonTemporal(), LD->getAlignment());
5938 /// SkipExtensionForVMULL - For a node that is a SIGN_EXTEND, ZERO_EXTEND,
5939 /// extending load, or BUILD_VECTOR with extended elements, return the
5940 /// unextended value. The unextended vector should be 64 bits so that it can
5941 /// be used as an operand to a VMULL instruction. If the original vector size
5942 /// before extension is less than 64 bits we add a an extension to resize
5943 /// the vector to 64 bits.
5944 static SDValue SkipExtensionForVMULL(SDNode *N, SelectionDAG &DAG) {
5945 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
5946 return AddRequiredExtensionForVMULL(N->getOperand(0), DAG,
5947 N->getOperand(0)->getValueType(0),
5951 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
5952 return SkipLoadExtensionForVMULL(LD, DAG);
5954 // Otherwise, the value must be a BUILD_VECTOR. For v2i64, it will
5955 // have been legalized as a BITCAST from v4i32.
5956 if (N->getOpcode() == ISD::BITCAST) {
5957 SDNode *BVN = N->getOperand(0).getNode();
5958 assert(BVN->getOpcode() == ISD::BUILD_VECTOR &&
5959 BVN->getValueType(0) == MVT::v4i32 && "expected v4i32 BUILD_VECTOR");
5960 unsigned LowElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
5961 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), MVT::v2i32,
5962 BVN->getOperand(LowElt), BVN->getOperand(LowElt+2));
5964 // Construct a new BUILD_VECTOR with elements truncated to half the size.
5965 assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR");
5966 EVT VT = N->getValueType(0);
5967 unsigned EltSize = VT.getVectorElementType().getSizeInBits() / 2;
5968 unsigned NumElts = VT.getVectorNumElements();
5969 MVT TruncVT = MVT::getIntegerVT(EltSize);
5970 SmallVector<SDValue, 8> Ops;
5971 for (unsigned i = 0; i != NumElts; ++i) {
5972 ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i));
5973 const APInt &CInt = C->getAPIntValue();
5974 // Element types smaller than 32 bits are not legal, so use i32 elements.
5975 // The values are implicitly truncated so sext vs. zext doesn't matter.
5976 Ops.push_back(DAG.getConstant(CInt.zextOrTrunc(32), MVT::i32));
5978 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N),
5979 MVT::getVectorVT(TruncVT, NumElts), Ops);
5982 static bool isAddSubSExt(SDNode *N, SelectionDAG &DAG) {
5983 unsigned Opcode = N->getOpcode();
5984 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
5985 SDNode *N0 = N->getOperand(0).getNode();
5986 SDNode *N1 = N->getOperand(1).getNode();
5987 return N0->hasOneUse() && N1->hasOneUse() &&
5988 isSignExtended(N0, DAG) && isSignExtended(N1, DAG);
5993 static bool isAddSubZExt(SDNode *N, SelectionDAG &DAG) {
5994 unsigned Opcode = N->getOpcode();
5995 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
5996 SDNode *N0 = N->getOperand(0).getNode();
5997 SDNode *N1 = N->getOperand(1).getNode();
5998 return N0->hasOneUse() && N1->hasOneUse() &&
5999 isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG);
6004 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
6005 // Multiplications are only custom-lowered for 128-bit vectors so that
6006 // VMULL can be detected. Otherwise v2i64 multiplications are not legal.
6007 EVT VT = Op.getValueType();
6008 assert(VT.is128BitVector() && VT.isInteger() &&
6009 "unexpected type for custom-lowering ISD::MUL");
6010 SDNode *N0 = Op.getOperand(0).getNode();
6011 SDNode *N1 = Op.getOperand(1).getNode();
6012 unsigned NewOpc = 0;
6014 bool isN0SExt = isSignExtended(N0, DAG);
6015 bool isN1SExt = isSignExtended(N1, DAG);
6016 if (isN0SExt && isN1SExt)
6017 NewOpc = ARMISD::VMULLs;
6019 bool isN0ZExt = isZeroExtended(N0, DAG);
6020 bool isN1ZExt = isZeroExtended(N1, DAG);
6021 if (isN0ZExt && isN1ZExt)
6022 NewOpc = ARMISD::VMULLu;
6023 else if (isN1SExt || isN1ZExt) {
6024 // Look for (s/zext A + s/zext B) * (s/zext C). We want to turn these
6025 // into (s/zext A * s/zext C) + (s/zext B * s/zext C)
6026 if (isN1SExt && isAddSubSExt(N0, DAG)) {
6027 NewOpc = ARMISD::VMULLs;
6029 } else if (isN1ZExt && isAddSubZExt(N0, DAG)) {
6030 NewOpc = ARMISD::VMULLu;
6032 } else if (isN0ZExt && isAddSubZExt(N1, DAG)) {
6034 NewOpc = ARMISD::VMULLu;
6040 if (VT == MVT::v2i64)
6041 // Fall through to expand this. It is not legal.
6044 // Other vector multiplications are legal.
6049 // Legalize to a VMULL instruction.
6052 SDValue Op1 = SkipExtensionForVMULL(N1, DAG);
6054 Op0 = SkipExtensionForVMULL(N0, DAG);
6055 assert(Op0.getValueType().is64BitVector() &&
6056 Op1.getValueType().is64BitVector() &&
6057 "unexpected types for extended operands to VMULL");
6058 return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
6061 // Optimizing (zext A + zext B) * C, to (VMULL A, C) + (VMULL B, C) during
6062 // isel lowering to take advantage of no-stall back to back vmul + vmla.
6069 SDValue N00 = SkipExtensionForVMULL(N0->getOperand(0).getNode(), DAG);
6070 SDValue N01 = SkipExtensionForVMULL(N0->getOperand(1).getNode(), DAG);
6071 EVT Op1VT = Op1.getValueType();
6072 return DAG.getNode(N0->getOpcode(), DL, VT,
6073 DAG.getNode(NewOpc, DL, VT,
6074 DAG.getNode(ISD::BITCAST, DL, Op1VT, N00), Op1),
6075 DAG.getNode(NewOpc, DL, VT,
6076 DAG.getNode(ISD::BITCAST, DL, Op1VT, N01), Op1));
6080 LowerSDIV_v4i8(SDValue X, SDValue Y, SDLoc dl, SelectionDAG &DAG) {
6082 // float4 xf = vcvt_f32_s32(vmovl_s16(a.lo));
6083 // float4 yf = vcvt_f32_s32(vmovl_s16(b.lo));
6084 X = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, X);
6085 Y = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, Y);
6086 X = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, X);
6087 Y = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, Y);
6088 // Get reciprocal estimate.
6089 // float4 recip = vrecpeq_f32(yf);
6090 Y = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6091 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), Y);
6092 // Because char has a smaller range than uchar, we can actually get away
6093 // without any newton steps. This requires that we use a weird bias
6094 // of 0xb000, however (again, this has been exhaustively tested).
6095 // float4 result = as_float4(as_int4(xf*recip) + 0xb000);
6096 X = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, X, Y);
6097 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, X);
6098 Y = DAG.getConstant(0xb000, MVT::i32);
6099 Y = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Y, Y, Y, Y);
6100 X = DAG.getNode(ISD::ADD, dl, MVT::v4i32, X, Y);
6101 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, X);
6102 // Convert back to short.
6103 X = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, X);
6104 X = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, X);
6109 LowerSDIV_v4i16(SDValue N0, SDValue N1, SDLoc dl, SelectionDAG &DAG) {
6111 // Convert to float.
6112 // float4 yf = vcvt_f32_s32(vmovl_s16(y));
6113 // float4 xf = vcvt_f32_s32(vmovl_s16(x));
6114 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N0);
6115 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N1);
6116 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
6117 N1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
6119 // Use reciprocal estimate and one refinement step.
6120 // float4 recip = vrecpeq_f32(yf);
6121 // recip *= vrecpsq_f32(yf, recip);
6122 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6123 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), N1);
6124 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6125 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
6127 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
6128 // Because short has a smaller range than ushort, we can actually get away
6129 // with only a single newton step. This requires that we use a weird bias
6130 // of 89, however (again, this has been exhaustively tested).
6131 // float4 result = as_float4(as_int4(xf*recip) + 0x89);
6132 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
6133 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
6134 N1 = DAG.getConstant(0x89, MVT::i32);
6135 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
6136 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
6137 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
6138 // Convert back to integer and return.
6139 // return vmovn_s32(vcvt_s32_f32(result));
6140 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
6141 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
6145 static SDValue LowerSDIV(SDValue Op, SelectionDAG &DAG) {
6146 EVT VT = Op.getValueType();
6147 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
6148 "unexpected type for custom-lowering ISD::SDIV");
6151 SDValue N0 = Op.getOperand(0);
6152 SDValue N1 = Op.getOperand(1);
6155 if (VT == MVT::v8i8) {
6156 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N0);
6157 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N1);
6159 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
6160 DAG.getIntPtrConstant(4));
6161 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
6162 DAG.getIntPtrConstant(4));
6163 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
6164 DAG.getIntPtrConstant(0));
6165 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
6166 DAG.getIntPtrConstant(0));
6168 N0 = LowerSDIV_v4i8(N0, N1, dl, DAG); // v4i16
6169 N2 = LowerSDIV_v4i8(N2, N3, dl, DAG); // v4i16
6171 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
6172 N0 = LowerCONCAT_VECTORS(N0, DAG);
6174 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v8i8, N0);
6177 return LowerSDIV_v4i16(N0, N1, dl, DAG);
6180 static SDValue LowerUDIV(SDValue Op, SelectionDAG &DAG) {
6181 EVT VT = Op.getValueType();
6182 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
6183 "unexpected type for custom-lowering ISD::UDIV");
6186 SDValue N0 = Op.getOperand(0);
6187 SDValue N1 = Op.getOperand(1);
6190 if (VT == MVT::v8i8) {
6191 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N0);
6192 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N1);
6194 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
6195 DAG.getIntPtrConstant(4));
6196 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
6197 DAG.getIntPtrConstant(4));
6198 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
6199 DAG.getIntPtrConstant(0));
6200 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
6201 DAG.getIntPtrConstant(0));
6203 N0 = LowerSDIV_v4i16(N0, N1, dl, DAG); // v4i16
6204 N2 = LowerSDIV_v4i16(N2, N3, dl, DAG); // v4i16
6206 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
6207 N0 = LowerCONCAT_VECTORS(N0, DAG);
6209 N0 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v8i8,
6210 DAG.getConstant(Intrinsic::arm_neon_vqmovnsu, MVT::i32),
6215 // v4i16 sdiv ... Convert to float.
6216 // float4 yf = vcvt_f32_s32(vmovl_u16(y));
6217 // float4 xf = vcvt_f32_s32(vmovl_u16(x));
6218 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N0);
6219 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N1);
6220 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
6221 SDValue BN1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
6223 // Use reciprocal estimate and two refinement steps.
6224 // float4 recip = vrecpeq_f32(yf);
6225 // recip *= vrecpsq_f32(yf, recip);
6226 // recip *= vrecpsq_f32(yf, recip);
6227 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6228 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), BN1);
6229 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6230 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
6232 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
6233 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6234 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
6236 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
6237 // Simply multiplying by the reciprocal estimate can leave us a few ulps
6238 // too low, so we add 2 ulps (exhaustive testing shows that this is enough,
6239 // and that it will never cause us to return an answer too large).
6240 // float4 result = as_float4(as_int4(xf*recip) + 2);
6241 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
6242 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
6243 N1 = DAG.getConstant(2, MVT::i32);
6244 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
6245 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
6246 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
6247 // Convert back to integer and return.
6248 // return vmovn_u32(vcvt_s32_f32(result));
6249 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
6250 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
6254 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
6255 EVT VT = Op.getNode()->getValueType(0);
6256 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
6259 bool ExtraOp = false;
6260 switch (Op.getOpcode()) {
6261 default: llvm_unreachable("Invalid code");
6262 case ISD::ADDC: Opc = ARMISD::ADDC; break;
6263 case ISD::ADDE: Opc = ARMISD::ADDE; ExtraOp = true; break;
6264 case ISD::SUBC: Opc = ARMISD::SUBC; break;
6265 case ISD::SUBE: Opc = ARMISD::SUBE; ExtraOp = true; break;
6269 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
6271 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
6272 Op.getOperand(1), Op.getOperand(2));
6275 SDValue ARMTargetLowering::LowerFSINCOS(SDValue Op, SelectionDAG &DAG) const {
6276 assert(Subtarget->isTargetDarwin());
6278 // For iOS, we want to call an alternative entry point: __sincos_stret,
6279 // return values are passed via sret.
6281 SDValue Arg = Op.getOperand(0);
6282 EVT ArgVT = Arg.getValueType();
6283 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
6285 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
6286 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6288 // Pair of floats / doubles used to pass the result.
6289 StructType *RetTy = StructType::get(ArgTy, ArgTy, nullptr);
6291 // Create stack object for sret.
6292 const uint64_t ByteSize = TLI.getDataLayout()->getTypeAllocSize(RetTy);
6293 const unsigned StackAlign = TLI.getDataLayout()->getPrefTypeAlignment(RetTy);
6294 int FrameIdx = FrameInfo->CreateStackObject(ByteSize, StackAlign, false);
6295 SDValue SRet = DAG.getFrameIndex(FrameIdx, TLI.getPointerTy());
6301 Entry.Ty = RetTy->getPointerTo();
6302 Entry.isSExt = false;
6303 Entry.isZExt = false;
6304 Entry.isSRet = true;
6305 Args.push_back(Entry);
6309 Entry.isSExt = false;
6310 Entry.isZExt = false;
6311 Args.push_back(Entry);
6313 const char *LibcallName = (ArgVT == MVT::f64)
6314 ? "__sincos_stret" : "__sincosf_stret";
6315 SDValue Callee = DAG.getExternalSymbol(LibcallName, getPointerTy());
6317 TargetLowering::CallLoweringInfo CLI(DAG);
6318 CLI.setDebugLoc(dl).setChain(DAG.getEntryNode())
6319 .setCallee(CallingConv::C, Type::getVoidTy(*DAG.getContext()), Callee,
6321 .setDiscardResult();
6323 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
6325 SDValue LoadSin = DAG.getLoad(ArgVT, dl, CallResult.second, SRet,
6326 MachinePointerInfo(), false, false, false, 0);
6328 // Address of cos field.
6329 SDValue Add = DAG.getNode(ISD::ADD, dl, getPointerTy(), SRet,
6330 DAG.getIntPtrConstant(ArgVT.getStoreSize()));
6331 SDValue LoadCos = DAG.getLoad(ArgVT, dl, LoadSin.getValue(1), Add,
6332 MachinePointerInfo(), false, false, false, 0);
6334 SDVTList Tys = DAG.getVTList(ArgVT, ArgVT);
6335 return DAG.getNode(ISD::MERGE_VALUES, dl, Tys,
6336 LoadSin.getValue(0), LoadCos.getValue(0));
6339 static SDValue LowerAtomicLoadStore(SDValue Op, SelectionDAG &DAG) {
6340 // Monotonic load/store is legal for all targets
6341 if (cast<AtomicSDNode>(Op)->getOrdering() <= Monotonic)
6344 // Acquire/Release load/store is not legal for targets without a
6345 // dmb or equivalent available.
6349 static void ReplaceREADCYCLECOUNTER(SDNode *N,
6350 SmallVectorImpl<SDValue> &Results,
6352 const ARMSubtarget *Subtarget) {
6354 SDValue Cycles32, OutChain;
6356 if (Subtarget->hasPerfMon()) {
6357 // Under Power Management extensions, the cycle-count is:
6358 // mrc p15, #0, <Rt>, c9, c13, #0
6359 SDValue Ops[] = { N->getOperand(0), // Chain
6360 DAG.getConstant(Intrinsic::arm_mrc, MVT::i32),
6361 DAG.getConstant(15, MVT::i32),
6362 DAG.getConstant(0, MVT::i32),
6363 DAG.getConstant(9, MVT::i32),
6364 DAG.getConstant(13, MVT::i32),
6365 DAG.getConstant(0, MVT::i32)
6368 Cycles32 = DAG.getNode(ISD::INTRINSIC_W_CHAIN, DL,
6369 DAG.getVTList(MVT::i32, MVT::Other), Ops);
6370 OutChain = Cycles32.getValue(1);
6372 // Intrinsic is defined to return 0 on unsupported platforms. Technically
6373 // there are older ARM CPUs that have implementation-specific ways of
6374 // obtaining this information (FIXME!).
6375 Cycles32 = DAG.getConstant(0, MVT::i32);
6376 OutChain = DAG.getEntryNode();
6380 SDValue Cycles64 = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64,
6381 Cycles32, DAG.getConstant(0, MVT::i32));
6382 Results.push_back(Cycles64);
6383 Results.push_back(OutChain);
6386 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
6387 switch (Op.getOpcode()) {
6388 default: llvm_unreachable("Don't know how to custom lower this!");
6389 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
6390 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
6391 case ISD::GlobalAddress:
6392 switch (Subtarget->getTargetTriple().getObjectFormat()) {
6393 default: llvm_unreachable("unknown object format");
6395 return LowerGlobalAddressWindows(Op, DAG);
6397 return LowerGlobalAddressELF(Op, DAG);
6399 return LowerGlobalAddressDarwin(Op, DAG);
6401 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
6402 case ISD::SELECT: return LowerSELECT(Op, DAG);
6403 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
6404 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
6405 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
6406 case ISD::VASTART: return LowerVASTART(Op, DAG);
6407 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, DAG, Subtarget);
6408 case ISD::PREFETCH: return LowerPREFETCH(Op, DAG, Subtarget);
6409 case ISD::SINT_TO_FP:
6410 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
6411 case ISD::FP_TO_SINT:
6412 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
6413 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
6414 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
6415 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
6416 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
6417 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
6418 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
6419 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
6421 case ISD::BITCAST: return ExpandBITCAST(Op.getNode(), DAG);
6424 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
6425 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
6426 case ISD::SRL_PARTS:
6427 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
6428 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
6429 case ISD::CTPOP: return LowerCTPOP(Op.getNode(), DAG, Subtarget);
6430 case ISD::SETCC: return LowerVSETCC(Op, DAG);
6431 case ISD::ConstantFP: return LowerConstantFP(Op, DAG, Subtarget);
6432 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG, Subtarget);
6433 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
6434 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
6435 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
6436 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
6437 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
6438 case ISD::MUL: return LowerMUL(Op, DAG);
6439 case ISD::SDIV: return LowerSDIV(Op, DAG);
6440 case ISD::UDIV: return LowerUDIV(Op, DAG);
6444 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
6449 return LowerXALUO(Op, DAG);
6450 case ISD::ATOMIC_LOAD:
6451 case ISD::ATOMIC_STORE: return LowerAtomicLoadStore(Op, DAG);
6452 case ISD::FSINCOS: return LowerFSINCOS(Op, DAG);
6454 case ISD::UDIVREM: return LowerDivRem(Op, DAG);
6455 case ISD::DYNAMIC_STACKALLOC:
6456 if (Subtarget->getTargetTriple().isWindowsItaniumEnvironment())
6457 return LowerDYNAMIC_STACKALLOC(Op, DAG);
6458 llvm_unreachable("Don't know how to custom lower this!");
6459 case ISD::FP_ROUND: return LowerFP_ROUND(Op, DAG);
6460 case ISD::FP_EXTEND: return LowerFP_EXTEND(Op, DAG);
6464 /// ReplaceNodeResults - Replace the results of node with an illegal result
6465 /// type with new values built out of custom code.
6466 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
6467 SmallVectorImpl<SDValue>&Results,
6468 SelectionDAG &DAG) const {
6470 switch (N->getOpcode()) {
6472 llvm_unreachable("Don't know how to custom expand this!");
6474 Res = ExpandBITCAST(N, DAG);
6478 Res = Expand64BitShift(N, DAG, Subtarget);
6480 case ISD::READCYCLECOUNTER:
6481 ReplaceREADCYCLECOUNTER(N, Results, DAG, Subtarget);
6485 Results.push_back(Res);
6488 //===----------------------------------------------------------------------===//
6489 // ARM Scheduler Hooks
6490 //===----------------------------------------------------------------------===//
6492 /// SetupEntryBlockForSjLj - Insert code into the entry block that creates and
6493 /// registers the function context.
6494 void ARMTargetLowering::
6495 SetupEntryBlockForSjLj(MachineInstr *MI, MachineBasicBlock *MBB,
6496 MachineBasicBlock *DispatchBB, int FI) const {
6497 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
6498 DebugLoc dl = MI->getDebugLoc();
6499 MachineFunction *MF = MBB->getParent();
6500 MachineRegisterInfo *MRI = &MF->getRegInfo();
6501 MachineConstantPool *MCP = MF->getConstantPool();
6502 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
6503 const Function *F = MF->getFunction();
6505 bool isThumb = Subtarget->isThumb();
6506 bool isThumb2 = Subtarget->isThumb2();
6508 unsigned PCLabelId = AFI->createPICLabelUId();
6509 unsigned PCAdj = (isThumb || isThumb2) ? 4 : 8;
6510 ARMConstantPoolValue *CPV =
6511 ARMConstantPoolMBB::Create(F->getContext(), DispatchBB, PCLabelId, PCAdj);
6512 unsigned CPI = MCP->getConstantPoolIndex(CPV, 4);
6514 const TargetRegisterClass *TRC = isThumb ? &ARM::tGPRRegClass
6515 : &ARM::GPRRegClass;
6517 // Grab constant pool and fixed stack memory operands.
6518 MachineMemOperand *CPMMO =
6519 MF->getMachineMemOperand(MachinePointerInfo::getConstantPool(),
6520 MachineMemOperand::MOLoad, 4, 4);
6522 MachineMemOperand *FIMMOSt =
6523 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
6524 MachineMemOperand::MOStore, 4, 4);
6526 // Load the address of the dispatch MBB into the jump buffer.
6528 // Incoming value: jbuf
6529 // ldr.n r5, LCPI1_1
6532 // str r5, [$jbuf, #+4] ; &jbuf[1]
6533 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6534 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2LDRpci), NewVReg1)
6535 .addConstantPoolIndex(CPI)
6536 .addMemOperand(CPMMO));
6537 // Set the low bit because of thumb mode.
6538 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6540 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2ORRri), NewVReg2)
6541 .addReg(NewVReg1, RegState::Kill)
6543 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6544 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg3)
6545 .addReg(NewVReg2, RegState::Kill)
6547 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2STRi12))
6548 .addReg(NewVReg3, RegState::Kill)
6550 .addImm(36) // &jbuf[1] :: pc
6551 .addMemOperand(FIMMOSt));
6552 } else if (isThumb) {
6553 // Incoming value: jbuf
6554 // ldr.n r1, LCPI1_4
6558 // add r2, $jbuf, #+4 ; &jbuf[1]
6560 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6561 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tLDRpci), NewVReg1)
6562 .addConstantPoolIndex(CPI)
6563 .addMemOperand(CPMMO));
6564 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6565 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg2)
6566 .addReg(NewVReg1, RegState::Kill)
6568 // Set the low bit because of thumb mode.
6569 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6570 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tMOVi8), NewVReg3)
6571 .addReg(ARM::CPSR, RegState::Define)
6573 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6574 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tORR), NewVReg4)
6575 .addReg(ARM::CPSR, RegState::Define)
6576 .addReg(NewVReg2, RegState::Kill)
6577 .addReg(NewVReg3, RegState::Kill));
6578 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6579 BuildMI(*MBB, MI, dl, TII->get(ARM::tADDframe), NewVReg5)
6581 .addImm(36); // &jbuf[1] :: pc
6582 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tSTRi))
6583 .addReg(NewVReg4, RegState::Kill)
6584 .addReg(NewVReg5, RegState::Kill)
6586 .addMemOperand(FIMMOSt));
6588 // Incoming value: jbuf
6591 // str r1, [$jbuf, #+4] ; &jbuf[1]
6592 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6593 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::LDRi12), NewVReg1)
6594 .addConstantPoolIndex(CPI)
6596 .addMemOperand(CPMMO));
6597 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6598 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::PICADD), NewVReg2)
6599 .addReg(NewVReg1, RegState::Kill)
6600 .addImm(PCLabelId));
6601 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::STRi12))
6602 .addReg(NewVReg2, RegState::Kill)
6604 .addImm(36) // &jbuf[1] :: pc
6605 .addMemOperand(FIMMOSt));
6609 MachineBasicBlock *ARMTargetLowering::
6610 EmitSjLjDispatchBlock(MachineInstr *MI, MachineBasicBlock *MBB) const {
6611 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
6612 DebugLoc dl = MI->getDebugLoc();
6613 MachineFunction *MF = MBB->getParent();
6614 MachineRegisterInfo *MRI = &MF->getRegInfo();
6615 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
6616 MachineFrameInfo *MFI = MF->getFrameInfo();
6617 int FI = MFI->getFunctionContextIndex();
6619 const TargetRegisterClass *TRC = Subtarget->isThumb() ? &ARM::tGPRRegClass
6620 : &ARM::GPRnopcRegClass;
6622 // Get a mapping of the call site numbers to all of the landing pads they're
6624 DenseMap<unsigned, SmallVector<MachineBasicBlock*, 2> > CallSiteNumToLPad;
6625 unsigned MaxCSNum = 0;
6626 MachineModuleInfo &MMI = MF->getMMI();
6627 for (MachineFunction::iterator BB = MF->begin(), E = MF->end(); BB != E;
6629 if (!BB->isLandingPad()) continue;
6631 // FIXME: We should assert that the EH_LABEL is the first MI in the landing
6633 for (MachineBasicBlock::iterator
6634 II = BB->begin(), IE = BB->end(); II != IE; ++II) {
6635 if (!II->isEHLabel()) continue;
6637 MCSymbol *Sym = II->getOperand(0).getMCSymbol();
6638 if (!MMI.hasCallSiteLandingPad(Sym)) continue;
6640 SmallVectorImpl<unsigned> &CallSiteIdxs = MMI.getCallSiteLandingPad(Sym);
6641 for (SmallVectorImpl<unsigned>::iterator
6642 CSI = CallSiteIdxs.begin(), CSE = CallSiteIdxs.end();
6643 CSI != CSE; ++CSI) {
6644 CallSiteNumToLPad[*CSI].push_back(BB);
6645 MaxCSNum = std::max(MaxCSNum, *CSI);
6651 // Get an ordered list of the machine basic blocks for the jump table.
6652 std::vector<MachineBasicBlock*> LPadList;
6653 SmallPtrSet<MachineBasicBlock*, 64> InvokeBBs;
6654 LPadList.reserve(CallSiteNumToLPad.size());
6655 for (unsigned I = 1; I <= MaxCSNum; ++I) {
6656 SmallVectorImpl<MachineBasicBlock*> &MBBList = CallSiteNumToLPad[I];
6657 for (SmallVectorImpl<MachineBasicBlock*>::iterator
6658 II = MBBList.begin(), IE = MBBList.end(); II != IE; ++II) {
6659 LPadList.push_back(*II);
6660 InvokeBBs.insert((*II)->pred_begin(), (*II)->pred_end());
6664 assert(!LPadList.empty() &&
6665 "No landing pad destinations for the dispatch jump table!");
6667 // Create the jump table and associated information.
6668 MachineJumpTableInfo *JTI =
6669 MF->getOrCreateJumpTableInfo(MachineJumpTableInfo::EK_Inline);
6670 unsigned MJTI = JTI->createJumpTableIndex(LPadList);
6671 unsigned UId = AFI->createJumpTableUId();
6672 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
6674 // Create the MBBs for the dispatch code.
6676 // Shove the dispatch's address into the return slot in the function context.
6677 MachineBasicBlock *DispatchBB = MF->CreateMachineBasicBlock();
6678 DispatchBB->setIsLandingPad();
6680 MachineBasicBlock *TrapBB = MF->CreateMachineBasicBlock();
6681 unsigned trap_opcode;
6682 if (Subtarget->isThumb())
6683 trap_opcode = ARM::tTRAP;
6685 trap_opcode = Subtarget->useNaClTrap() ? ARM::TRAPNaCl : ARM::TRAP;
6687 BuildMI(TrapBB, dl, TII->get(trap_opcode));
6688 DispatchBB->addSuccessor(TrapBB);
6690 MachineBasicBlock *DispContBB = MF->CreateMachineBasicBlock();
6691 DispatchBB->addSuccessor(DispContBB);
6694 MF->insert(MF->end(), DispatchBB);
6695 MF->insert(MF->end(), DispContBB);
6696 MF->insert(MF->end(), TrapBB);
6698 // Insert code into the entry block that creates and registers the function
6700 SetupEntryBlockForSjLj(MI, MBB, DispatchBB, FI);
6702 MachineMemOperand *FIMMOLd =
6703 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
6704 MachineMemOperand::MOLoad |
6705 MachineMemOperand::MOVolatile, 4, 4);
6707 MachineInstrBuilder MIB;
6708 MIB = BuildMI(DispatchBB, dl, TII->get(ARM::Int_eh_sjlj_dispatchsetup));
6710 const ARMBaseInstrInfo *AII = static_cast<const ARMBaseInstrInfo*>(TII);
6711 const ARMBaseRegisterInfo &RI = AII->getRegisterInfo();
6713 // Add a register mask with no preserved registers. This results in all
6714 // registers being marked as clobbered.
6715 MIB.addRegMask(RI.getNoPreservedMask());
6717 unsigned NumLPads = LPadList.size();
6718 if (Subtarget->isThumb2()) {
6719 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6720 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2LDRi12), NewVReg1)
6723 .addMemOperand(FIMMOLd));
6725 if (NumLPads < 256) {
6726 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPri))
6728 .addImm(LPadList.size()));
6730 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6731 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVi16), VReg1)
6732 .addImm(NumLPads & 0xFFFF));
6734 unsigned VReg2 = VReg1;
6735 if ((NumLPads & 0xFFFF0000) != 0) {
6736 VReg2 = MRI->createVirtualRegister(TRC);
6737 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVTi16), VReg2)
6739 .addImm(NumLPads >> 16));
6742 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPrr))
6747 BuildMI(DispatchBB, dl, TII->get(ARM::t2Bcc))
6752 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6753 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::t2LEApcrelJT),NewVReg3)
6754 .addJumpTableIndex(MJTI)
6757 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6760 BuildMI(DispContBB, dl, TII->get(ARM::t2ADDrs), NewVReg4)
6761 .addReg(NewVReg3, RegState::Kill)
6763 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
6765 BuildMI(DispContBB, dl, TII->get(ARM::t2BR_JT))
6766 .addReg(NewVReg4, RegState::Kill)
6768 .addJumpTableIndex(MJTI)
6770 } else if (Subtarget->isThumb()) {
6771 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6772 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRspi), NewVReg1)
6775 .addMemOperand(FIMMOLd));
6777 if (NumLPads < 256) {
6778 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPi8))
6782 MachineConstantPool *ConstantPool = MF->getConstantPool();
6783 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
6784 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
6786 // MachineConstantPool wants an explicit alignment.
6787 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
6789 Align = getDataLayout()->getTypeAllocSize(C->getType());
6790 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6792 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6793 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRpci))
6794 .addReg(VReg1, RegState::Define)
6795 .addConstantPoolIndex(Idx));
6796 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPr))
6801 BuildMI(DispatchBB, dl, TII->get(ARM::tBcc))
6806 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6807 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLSLri), NewVReg2)
6808 .addReg(ARM::CPSR, RegState::Define)
6812 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6813 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLEApcrelJT), NewVReg3)
6814 .addJumpTableIndex(MJTI)
6817 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6818 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg4)
6819 .addReg(ARM::CPSR, RegState::Define)
6820 .addReg(NewVReg2, RegState::Kill)
6823 MachineMemOperand *JTMMOLd =
6824 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
6825 MachineMemOperand::MOLoad, 4, 4);
6827 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6828 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLDRi), NewVReg5)
6829 .addReg(NewVReg4, RegState::Kill)
6831 .addMemOperand(JTMMOLd));
6833 unsigned NewVReg6 = NewVReg5;
6834 if (RelocM == Reloc::PIC_) {
6835 NewVReg6 = MRI->createVirtualRegister(TRC);
6836 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg6)
6837 .addReg(ARM::CPSR, RegState::Define)
6838 .addReg(NewVReg5, RegState::Kill)
6842 BuildMI(DispContBB, dl, TII->get(ARM::tBR_JTr))
6843 .addReg(NewVReg6, RegState::Kill)
6844 .addJumpTableIndex(MJTI)
6847 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6848 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRi12), NewVReg1)
6851 .addMemOperand(FIMMOLd));
6853 if (NumLPads < 256) {
6854 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPri))
6857 } else if (Subtarget->hasV6T2Ops() && isUInt<16>(NumLPads)) {
6858 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6859 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVi16), VReg1)
6860 .addImm(NumLPads & 0xFFFF));
6862 unsigned VReg2 = VReg1;
6863 if ((NumLPads & 0xFFFF0000) != 0) {
6864 VReg2 = MRI->createVirtualRegister(TRC);
6865 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVTi16), VReg2)
6867 .addImm(NumLPads >> 16));
6870 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
6874 MachineConstantPool *ConstantPool = MF->getConstantPool();
6875 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
6876 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
6878 // MachineConstantPool wants an explicit alignment.
6879 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
6881 Align = getDataLayout()->getTypeAllocSize(C->getType());
6882 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6884 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6885 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRcp))
6886 .addReg(VReg1, RegState::Define)
6887 .addConstantPoolIndex(Idx)
6889 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
6891 .addReg(VReg1, RegState::Kill));
6894 BuildMI(DispatchBB, dl, TII->get(ARM::Bcc))
6899 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6901 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::MOVsi), NewVReg3)
6903 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
6904 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6905 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::LEApcrelJT), NewVReg4)
6906 .addJumpTableIndex(MJTI)
6909 MachineMemOperand *JTMMOLd =
6910 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
6911 MachineMemOperand::MOLoad, 4, 4);
6912 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6914 BuildMI(DispContBB, dl, TII->get(ARM::LDRrs), NewVReg5)
6915 .addReg(NewVReg3, RegState::Kill)
6918 .addMemOperand(JTMMOLd));
6920 if (RelocM == Reloc::PIC_) {
6921 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTadd))
6922 .addReg(NewVReg5, RegState::Kill)
6924 .addJumpTableIndex(MJTI)
6927 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTr))
6928 .addReg(NewVReg5, RegState::Kill)
6929 .addJumpTableIndex(MJTI)
6934 // Add the jump table entries as successors to the MBB.
6935 SmallPtrSet<MachineBasicBlock*, 8> SeenMBBs;
6936 for (std::vector<MachineBasicBlock*>::iterator
6937 I = LPadList.begin(), E = LPadList.end(); I != E; ++I) {
6938 MachineBasicBlock *CurMBB = *I;
6939 if (SeenMBBs.insert(CurMBB).second)
6940 DispContBB->addSuccessor(CurMBB);
6943 // N.B. the order the invoke BBs are processed in doesn't matter here.
6944 const MCPhysReg *SavedRegs = RI.getCalleeSavedRegs(MF);
6945 SmallVector<MachineBasicBlock*, 64> MBBLPads;
6946 for (MachineBasicBlock *BB : InvokeBBs) {
6948 // Remove the landing pad successor from the invoke block and replace it
6949 // with the new dispatch block.
6950 SmallVector<MachineBasicBlock*, 4> Successors(BB->succ_begin(),
6952 while (!Successors.empty()) {
6953 MachineBasicBlock *SMBB = Successors.pop_back_val();
6954 if (SMBB->isLandingPad()) {
6955 BB->removeSuccessor(SMBB);
6956 MBBLPads.push_back(SMBB);
6960 BB->addSuccessor(DispatchBB);
6962 // Find the invoke call and mark all of the callee-saved registers as
6963 // 'implicit defined' so that they're spilled. This prevents code from
6964 // moving instructions to before the EH block, where they will never be
6966 for (MachineBasicBlock::reverse_iterator
6967 II = BB->rbegin(), IE = BB->rend(); II != IE; ++II) {
6968 if (!II->isCall()) continue;
6970 DenseMap<unsigned, bool> DefRegs;
6971 for (MachineInstr::mop_iterator
6972 OI = II->operands_begin(), OE = II->operands_end();
6974 if (!OI->isReg()) continue;
6975 DefRegs[OI->getReg()] = true;
6978 MachineInstrBuilder MIB(*MF, &*II);
6980 for (unsigned i = 0; SavedRegs[i] != 0; ++i) {
6981 unsigned Reg = SavedRegs[i];
6982 if (Subtarget->isThumb2() &&
6983 !ARM::tGPRRegClass.contains(Reg) &&
6984 !ARM::hGPRRegClass.contains(Reg))
6986 if (Subtarget->isThumb1Only() && !ARM::tGPRRegClass.contains(Reg))
6988 if (!Subtarget->isThumb() && !ARM::GPRRegClass.contains(Reg))
6991 MIB.addReg(Reg, RegState::ImplicitDefine | RegState::Dead);
6998 // Mark all former landing pads as non-landing pads. The dispatch is the only
7000 for (SmallVectorImpl<MachineBasicBlock*>::iterator
7001 I = MBBLPads.begin(), E = MBBLPads.end(); I != E; ++I)
7002 (*I)->setIsLandingPad(false);
7004 // The instruction is gone now.
7005 MI->eraseFromParent();
7011 MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) {
7012 for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
7013 E = MBB->succ_end(); I != E; ++I)
7016 llvm_unreachable("Expecting a BB with two successors!");
7019 /// Return the load opcode for a given load size. If load size >= 8,
7020 /// neon opcode will be returned.
7021 static unsigned getLdOpcode(unsigned LdSize, bool IsThumb1, bool IsThumb2) {
7023 return LdSize == 16 ? ARM::VLD1q32wb_fixed
7024 : LdSize == 8 ? ARM::VLD1d32wb_fixed : 0;
7026 return LdSize == 4 ? ARM::tLDRi
7027 : LdSize == 2 ? ARM::tLDRHi
7028 : LdSize == 1 ? ARM::tLDRBi : 0;
7030 return LdSize == 4 ? ARM::t2LDR_POST
7031 : LdSize == 2 ? ARM::t2LDRH_POST
7032 : LdSize == 1 ? ARM::t2LDRB_POST : 0;
7033 return LdSize == 4 ? ARM::LDR_POST_IMM
7034 : LdSize == 2 ? ARM::LDRH_POST
7035 : LdSize == 1 ? ARM::LDRB_POST_IMM : 0;
7038 /// Return the store opcode for a given store size. If store size >= 8,
7039 /// neon opcode will be returned.
7040 static unsigned getStOpcode(unsigned StSize, bool IsThumb1, bool IsThumb2) {
7042 return StSize == 16 ? ARM::VST1q32wb_fixed
7043 : StSize == 8 ? ARM::VST1d32wb_fixed : 0;
7045 return StSize == 4 ? ARM::tSTRi
7046 : StSize == 2 ? ARM::tSTRHi
7047 : StSize == 1 ? ARM::tSTRBi : 0;
7049 return StSize == 4 ? ARM::t2STR_POST
7050 : StSize == 2 ? ARM::t2STRH_POST
7051 : StSize == 1 ? ARM::t2STRB_POST : 0;
7052 return StSize == 4 ? ARM::STR_POST_IMM
7053 : StSize == 2 ? ARM::STRH_POST
7054 : StSize == 1 ? ARM::STRB_POST_IMM : 0;
7057 /// Emit a post-increment load operation with given size. The instructions
7058 /// will be added to BB at Pos.
7059 static void emitPostLd(MachineBasicBlock *BB, MachineInstr *Pos,
7060 const TargetInstrInfo *TII, DebugLoc dl,
7061 unsigned LdSize, unsigned Data, unsigned AddrIn,
7062 unsigned AddrOut, bool IsThumb1, bool IsThumb2) {
7063 unsigned LdOpc = getLdOpcode(LdSize, IsThumb1, IsThumb2);
7064 assert(LdOpc != 0 && "Should have a load opcode");
7066 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
7067 .addReg(AddrOut, RegState::Define).addReg(AddrIn)
7069 } else if (IsThumb1) {
7070 // load + update AddrIn
7071 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
7072 .addReg(AddrIn).addImm(0));
7073 MachineInstrBuilder MIB =
7074 BuildMI(*BB, Pos, dl, TII->get(ARM::tADDi8), AddrOut);
7075 MIB = AddDefaultT1CC(MIB);
7076 MIB.addReg(AddrIn).addImm(LdSize);
7077 AddDefaultPred(MIB);
7078 } else if (IsThumb2) {
7079 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
7080 .addReg(AddrOut, RegState::Define).addReg(AddrIn)
7083 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
7084 .addReg(AddrOut, RegState::Define).addReg(AddrIn)
7085 .addReg(0).addImm(LdSize));
7089 /// Emit a post-increment store operation with given size. The instructions
7090 /// will be added to BB at Pos.
7091 static void emitPostSt(MachineBasicBlock *BB, MachineInstr *Pos,
7092 const TargetInstrInfo *TII, DebugLoc dl,
7093 unsigned StSize, unsigned Data, unsigned AddrIn,
7094 unsigned AddrOut, bool IsThumb1, bool IsThumb2) {
7095 unsigned StOpc = getStOpcode(StSize, IsThumb1, IsThumb2);
7096 assert(StOpc != 0 && "Should have a store opcode");
7098 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc), AddrOut)
7099 .addReg(AddrIn).addImm(0).addReg(Data));
7100 } else if (IsThumb1) {
7101 // store + update AddrIn
7102 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc)).addReg(Data)
7103 .addReg(AddrIn).addImm(0));
7104 MachineInstrBuilder MIB =
7105 BuildMI(*BB, Pos, dl, TII->get(ARM::tADDi8), AddrOut);
7106 MIB = AddDefaultT1CC(MIB);
7107 MIB.addReg(AddrIn).addImm(StSize);
7108 AddDefaultPred(MIB);
7109 } else if (IsThumb2) {
7110 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc), AddrOut)
7111 .addReg(Data).addReg(AddrIn).addImm(StSize));
7113 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc), AddrOut)
7114 .addReg(Data).addReg(AddrIn).addReg(0)
7120 ARMTargetLowering::EmitStructByval(MachineInstr *MI,
7121 MachineBasicBlock *BB) const {
7122 // This pseudo instruction has 3 operands: dst, src, size
7123 // We expand it to a loop if size > Subtarget->getMaxInlineSizeThreshold().
7124 // Otherwise, we will generate unrolled scalar copies.
7125 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
7126 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7127 MachineFunction::iterator It = BB;
7130 unsigned dest = MI->getOperand(0).getReg();
7131 unsigned src = MI->getOperand(1).getReg();
7132 unsigned SizeVal = MI->getOperand(2).getImm();
7133 unsigned Align = MI->getOperand(3).getImm();
7134 DebugLoc dl = MI->getDebugLoc();
7136 MachineFunction *MF = BB->getParent();
7137 MachineRegisterInfo &MRI = MF->getRegInfo();
7138 unsigned UnitSize = 0;
7139 const TargetRegisterClass *TRC = nullptr;
7140 const TargetRegisterClass *VecTRC = nullptr;
7142 bool IsThumb1 = Subtarget->isThumb1Only();
7143 bool IsThumb2 = Subtarget->isThumb2();
7147 } else if (Align & 2) {
7150 // Check whether we can use NEON instructions.
7151 if (!MF->getFunction()->getAttributes().
7152 hasAttribute(AttributeSet::FunctionIndex,
7153 Attribute::NoImplicitFloat) &&
7154 Subtarget->hasNEON()) {
7155 if ((Align % 16 == 0) && SizeVal >= 16)
7157 else if ((Align % 8 == 0) && SizeVal >= 8)
7160 // Can't use NEON instructions.
7165 // Select the correct opcode and register class for unit size load/store
7166 bool IsNeon = UnitSize >= 8;
7167 TRC = (IsThumb1 || IsThumb2) ? &ARM::tGPRRegClass : &ARM::GPRRegClass;
7169 VecTRC = UnitSize == 16 ? &ARM::DPairRegClass
7170 : UnitSize == 8 ? &ARM::DPRRegClass
7173 unsigned BytesLeft = SizeVal % UnitSize;
7174 unsigned LoopSize = SizeVal - BytesLeft;
7176 if (SizeVal <= Subtarget->getMaxInlineSizeThreshold()) {
7177 // Use LDR and STR to copy.
7178 // [scratch, srcOut] = LDR_POST(srcIn, UnitSize)
7179 // [destOut] = STR_POST(scratch, destIn, UnitSize)
7180 unsigned srcIn = src;
7181 unsigned destIn = dest;
7182 for (unsigned i = 0; i < LoopSize; i+=UnitSize) {
7183 unsigned srcOut = MRI.createVirtualRegister(TRC);
7184 unsigned destOut = MRI.createVirtualRegister(TRC);
7185 unsigned scratch = MRI.createVirtualRegister(IsNeon ? VecTRC : TRC);
7186 emitPostLd(BB, MI, TII, dl, UnitSize, scratch, srcIn, srcOut,
7187 IsThumb1, IsThumb2);
7188 emitPostSt(BB, MI, TII, dl, UnitSize, scratch, destIn, destOut,
7189 IsThumb1, IsThumb2);
7194 // Handle the leftover bytes with LDRB and STRB.
7195 // [scratch, srcOut] = LDRB_POST(srcIn, 1)
7196 // [destOut] = STRB_POST(scratch, destIn, 1)
7197 for (unsigned i = 0; i < BytesLeft; i++) {
7198 unsigned srcOut = MRI.createVirtualRegister(TRC);
7199 unsigned destOut = MRI.createVirtualRegister(TRC);
7200 unsigned scratch = MRI.createVirtualRegister(TRC);
7201 emitPostLd(BB, MI, TII, dl, 1, scratch, srcIn, srcOut,
7202 IsThumb1, IsThumb2);
7203 emitPostSt(BB, MI, TII, dl, 1, scratch, destIn, destOut,
7204 IsThumb1, IsThumb2);
7208 MI->eraseFromParent(); // The instruction is gone now.
7212 // Expand the pseudo op to a loop.
7215 // movw varEnd, # --> with thumb2
7217 // ldrcp varEnd, idx --> without thumb2
7218 // fallthrough --> loopMBB
7220 // PHI varPhi, varEnd, varLoop
7221 // PHI srcPhi, src, srcLoop
7222 // PHI destPhi, dst, destLoop
7223 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
7224 // [destLoop] = STR_POST(scratch, destPhi, UnitSize)
7225 // subs varLoop, varPhi, #UnitSize
7227 // fallthrough --> exitMBB
7229 // epilogue to handle left-over bytes
7230 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
7231 // [destOut] = STRB_POST(scratch, destLoop, 1)
7232 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
7233 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
7234 MF->insert(It, loopMBB);
7235 MF->insert(It, exitMBB);
7237 // Transfer the remainder of BB and its successor edges to exitMBB.
7238 exitMBB->splice(exitMBB->begin(), BB,
7239 std::next(MachineBasicBlock::iterator(MI)), BB->end());
7240 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
7242 // Load an immediate to varEnd.
7243 unsigned varEnd = MRI.createVirtualRegister(TRC);
7245 unsigned Vtmp = varEnd;
7246 if ((LoopSize & 0xFFFF0000) != 0)
7247 Vtmp = MRI.createVirtualRegister(TRC);
7248 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2MOVi16), Vtmp)
7249 .addImm(LoopSize & 0xFFFF));
7251 if ((LoopSize & 0xFFFF0000) != 0)
7252 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2MOVTi16), varEnd)
7253 .addReg(Vtmp).addImm(LoopSize >> 16));
7255 MachineConstantPool *ConstantPool = MF->getConstantPool();
7256 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
7257 const Constant *C = ConstantInt::get(Int32Ty, LoopSize);
7259 // MachineConstantPool wants an explicit alignment.
7260 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
7262 Align = getDataLayout()->getTypeAllocSize(C->getType());
7263 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
7266 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(ARM::tLDRpci)).addReg(
7267 varEnd, RegState::Define).addConstantPoolIndex(Idx));
7269 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(ARM::LDRcp)).addReg(
7270 varEnd, RegState::Define).addConstantPoolIndex(Idx).addImm(0));
7272 BB->addSuccessor(loopMBB);
7274 // Generate the loop body:
7275 // varPhi = PHI(varLoop, varEnd)
7276 // srcPhi = PHI(srcLoop, src)
7277 // destPhi = PHI(destLoop, dst)
7278 MachineBasicBlock *entryBB = BB;
7280 unsigned varLoop = MRI.createVirtualRegister(TRC);
7281 unsigned varPhi = MRI.createVirtualRegister(TRC);
7282 unsigned srcLoop = MRI.createVirtualRegister(TRC);
7283 unsigned srcPhi = MRI.createVirtualRegister(TRC);
7284 unsigned destLoop = MRI.createVirtualRegister(TRC);
7285 unsigned destPhi = MRI.createVirtualRegister(TRC);
7287 BuildMI(*BB, BB->begin(), dl, TII->get(ARM::PHI), varPhi)
7288 .addReg(varLoop).addMBB(loopMBB)
7289 .addReg(varEnd).addMBB(entryBB);
7290 BuildMI(BB, dl, TII->get(ARM::PHI), srcPhi)
7291 .addReg(srcLoop).addMBB(loopMBB)
7292 .addReg(src).addMBB(entryBB);
7293 BuildMI(BB, dl, TII->get(ARM::PHI), destPhi)
7294 .addReg(destLoop).addMBB(loopMBB)
7295 .addReg(dest).addMBB(entryBB);
7297 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
7298 // [destLoop] = STR_POST(scratch, destPhi, UnitSiz)
7299 unsigned scratch = MRI.createVirtualRegister(IsNeon ? VecTRC : TRC);
7300 emitPostLd(BB, BB->end(), TII, dl, UnitSize, scratch, srcPhi, srcLoop,
7301 IsThumb1, IsThumb2);
7302 emitPostSt(BB, BB->end(), TII, dl, UnitSize, scratch, destPhi, destLoop,
7303 IsThumb1, IsThumb2);
7305 // Decrement loop variable by UnitSize.
7307 MachineInstrBuilder MIB =
7308 BuildMI(*BB, BB->end(), dl, TII->get(ARM::tSUBi8), varLoop);
7309 MIB = AddDefaultT1CC(MIB);
7310 MIB.addReg(varPhi).addImm(UnitSize);
7311 AddDefaultPred(MIB);
7313 MachineInstrBuilder MIB =
7314 BuildMI(*BB, BB->end(), dl,
7315 TII->get(IsThumb2 ? ARM::t2SUBri : ARM::SUBri), varLoop);
7316 AddDefaultCC(AddDefaultPred(MIB.addReg(varPhi).addImm(UnitSize)));
7317 MIB->getOperand(5).setReg(ARM::CPSR);
7318 MIB->getOperand(5).setIsDef(true);
7320 BuildMI(*BB, BB->end(), dl,
7321 TII->get(IsThumb1 ? ARM::tBcc : IsThumb2 ? ARM::t2Bcc : ARM::Bcc))
7322 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
7324 // loopMBB can loop back to loopMBB or fall through to exitMBB.
7325 BB->addSuccessor(loopMBB);
7326 BB->addSuccessor(exitMBB);
7328 // Add epilogue to handle BytesLeft.
7330 MachineInstr *StartOfExit = exitMBB->begin();
7332 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
7333 // [destOut] = STRB_POST(scratch, destLoop, 1)
7334 unsigned srcIn = srcLoop;
7335 unsigned destIn = destLoop;
7336 for (unsigned i = 0; i < BytesLeft; i++) {
7337 unsigned srcOut = MRI.createVirtualRegister(TRC);
7338 unsigned destOut = MRI.createVirtualRegister(TRC);
7339 unsigned scratch = MRI.createVirtualRegister(TRC);
7340 emitPostLd(BB, StartOfExit, TII, dl, 1, scratch, srcIn, srcOut,
7341 IsThumb1, IsThumb2);
7342 emitPostSt(BB, StartOfExit, TII, dl, 1, scratch, destIn, destOut,
7343 IsThumb1, IsThumb2);
7348 MI->eraseFromParent(); // The instruction is gone now.
7353 ARMTargetLowering::EmitLowered__chkstk(MachineInstr *MI,
7354 MachineBasicBlock *MBB) const {
7355 const TargetMachine &TM = getTargetMachine();
7356 const TargetInstrInfo &TII = *Subtarget->getInstrInfo();
7357 DebugLoc DL = MI->getDebugLoc();
7359 assert(Subtarget->isTargetWindows() &&
7360 "__chkstk is only supported on Windows");
7361 assert(Subtarget->isThumb2() && "Windows on ARM requires Thumb-2 mode");
7363 // __chkstk takes the number of words to allocate on the stack in R4, and
7364 // returns the stack adjustment in number of bytes in R4. This will not
7365 // clober any other registers (other than the obvious lr).
7367 // Although, technically, IP should be considered a register which may be
7368 // clobbered, the call itself will not touch it. Windows on ARM is a pure
7369 // thumb-2 environment, so there is no interworking required. As a result, we
7370 // do not expect a veneer to be emitted by the linker, clobbering IP.
7372 // Each module receives its own copy of __chkstk, so no import thunk is
7373 // required, again, ensuring that IP is not clobbered.
7375 // Finally, although some linkers may theoretically provide a trampoline for
7376 // out of range calls (which is quite common due to a 32M range limitation of
7377 // branches for Thumb), we can generate the long-call version via
7378 // -mcmodel=large, alleviating the need for the trampoline which may clobber
7381 switch (TM.getCodeModel()) {
7382 case CodeModel::Small:
7383 case CodeModel::Medium:
7384 case CodeModel::Default:
7385 case CodeModel::Kernel:
7386 BuildMI(*MBB, MI, DL, TII.get(ARM::tBL))
7387 .addImm((unsigned)ARMCC::AL).addReg(0)
7388 .addExternalSymbol("__chkstk")
7389 .addReg(ARM::R4, RegState::Implicit | RegState::Kill)
7390 .addReg(ARM::R4, RegState::Implicit | RegState::Define)
7391 .addReg(ARM::R12, RegState::Implicit | RegState::Define | RegState::Dead);
7393 case CodeModel::Large:
7394 case CodeModel::JITDefault: {
7395 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
7396 unsigned Reg = MRI.createVirtualRegister(&ARM::rGPRRegClass);
7398 BuildMI(*MBB, MI, DL, TII.get(ARM::t2MOVi32imm), Reg)
7399 .addExternalSymbol("__chkstk");
7400 BuildMI(*MBB, MI, DL, TII.get(ARM::tBLXr))
7401 .addImm((unsigned)ARMCC::AL).addReg(0)
7402 .addReg(Reg, RegState::Kill)
7403 .addReg(ARM::R4, RegState::Implicit | RegState::Kill)
7404 .addReg(ARM::R4, RegState::Implicit | RegState::Define)
7405 .addReg(ARM::R12, RegState::Implicit | RegState::Define | RegState::Dead);
7410 AddDefaultCC(AddDefaultPred(BuildMI(*MBB, MI, DL, TII.get(ARM::t2SUBrr),
7412 .addReg(ARM::SP).addReg(ARM::R4)));
7414 MI->eraseFromParent();
7419 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
7420 MachineBasicBlock *BB) const {
7421 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
7422 DebugLoc dl = MI->getDebugLoc();
7423 bool isThumb2 = Subtarget->isThumb2();
7424 switch (MI->getOpcode()) {
7427 llvm_unreachable("Unexpected instr type to insert");
7429 // The Thumb2 pre-indexed stores have the same MI operands, they just
7430 // define them differently in the .td files from the isel patterns, so
7431 // they need pseudos.
7432 case ARM::t2STR_preidx:
7433 MI->setDesc(TII->get(ARM::t2STR_PRE));
7435 case ARM::t2STRB_preidx:
7436 MI->setDesc(TII->get(ARM::t2STRB_PRE));
7438 case ARM::t2STRH_preidx:
7439 MI->setDesc(TII->get(ARM::t2STRH_PRE));
7442 case ARM::STRi_preidx:
7443 case ARM::STRBi_preidx: {
7444 unsigned NewOpc = MI->getOpcode() == ARM::STRi_preidx ?
7445 ARM::STR_PRE_IMM : ARM::STRB_PRE_IMM;
7446 // Decode the offset.
7447 unsigned Offset = MI->getOperand(4).getImm();
7448 bool isSub = ARM_AM::getAM2Op(Offset) == ARM_AM::sub;
7449 Offset = ARM_AM::getAM2Offset(Offset);
7453 MachineMemOperand *MMO = *MI->memoperands_begin();
7454 BuildMI(*BB, MI, dl, TII->get(NewOpc))
7455 .addOperand(MI->getOperand(0)) // Rn_wb
7456 .addOperand(MI->getOperand(1)) // Rt
7457 .addOperand(MI->getOperand(2)) // Rn
7458 .addImm(Offset) // offset (skip GPR==zero_reg)
7459 .addOperand(MI->getOperand(5)) // pred
7460 .addOperand(MI->getOperand(6))
7461 .addMemOperand(MMO);
7462 MI->eraseFromParent();
7465 case ARM::STRr_preidx:
7466 case ARM::STRBr_preidx:
7467 case ARM::STRH_preidx: {
7469 switch (MI->getOpcode()) {
7470 default: llvm_unreachable("unexpected opcode!");
7471 case ARM::STRr_preidx: NewOpc = ARM::STR_PRE_REG; break;
7472 case ARM::STRBr_preidx: NewOpc = ARM::STRB_PRE_REG; break;
7473 case ARM::STRH_preidx: NewOpc = ARM::STRH_PRE; break;
7475 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(NewOpc));
7476 for (unsigned i = 0; i < MI->getNumOperands(); ++i)
7477 MIB.addOperand(MI->getOperand(i));
7478 MI->eraseFromParent();
7482 case ARM::tMOVCCr_pseudo: {
7483 // To "insert" a SELECT_CC instruction, we actually have to insert the
7484 // diamond control-flow pattern. The incoming instruction knows the
7485 // destination vreg to set, the condition code register to branch on, the
7486 // true/false values to select between, and a branch opcode to use.
7487 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7488 MachineFunction::iterator It = BB;
7494 // cmpTY ccX, r1, r2
7496 // fallthrough --> copy0MBB
7497 MachineBasicBlock *thisMBB = BB;
7498 MachineFunction *F = BB->getParent();
7499 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
7500 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
7501 F->insert(It, copy0MBB);
7502 F->insert(It, sinkMBB);
7504 // Transfer the remainder of BB and its successor edges to sinkMBB.
7505 sinkMBB->splice(sinkMBB->begin(), BB,
7506 std::next(MachineBasicBlock::iterator(MI)), BB->end());
7507 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
7509 BB->addSuccessor(copy0MBB);
7510 BB->addSuccessor(sinkMBB);
7512 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
7513 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
7516 // %FalseValue = ...
7517 // # fallthrough to sinkMBB
7520 // Update machine-CFG edges
7521 BB->addSuccessor(sinkMBB);
7524 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
7527 BuildMI(*BB, BB->begin(), dl,
7528 TII->get(ARM::PHI), MI->getOperand(0).getReg())
7529 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
7530 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
7532 MI->eraseFromParent(); // The pseudo instruction is gone now.
7537 case ARM::BCCZi64: {
7538 // If there is an unconditional branch to the other successor, remove it.
7539 BB->erase(std::next(MachineBasicBlock::iterator(MI)), BB->end());
7541 // Compare both parts that make up the double comparison separately for
7543 bool RHSisZero = MI->getOpcode() == ARM::BCCZi64;
7545 unsigned LHS1 = MI->getOperand(1).getReg();
7546 unsigned LHS2 = MI->getOperand(2).getReg();
7548 AddDefaultPred(BuildMI(BB, dl,
7549 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7550 .addReg(LHS1).addImm(0));
7551 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7552 .addReg(LHS2).addImm(0)
7553 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
7555 unsigned RHS1 = MI->getOperand(3).getReg();
7556 unsigned RHS2 = MI->getOperand(4).getReg();
7557 AddDefaultPred(BuildMI(BB, dl,
7558 TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
7559 .addReg(LHS1).addReg(RHS1));
7560 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
7561 .addReg(LHS2).addReg(RHS2)
7562 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
7565 MachineBasicBlock *destMBB = MI->getOperand(RHSisZero ? 3 : 5).getMBB();
7566 MachineBasicBlock *exitMBB = OtherSucc(BB, destMBB);
7567 if (MI->getOperand(0).getImm() == ARMCC::NE)
7568 std::swap(destMBB, exitMBB);
7570 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
7571 .addMBB(destMBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
7573 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2B)).addMBB(exitMBB));
7575 BuildMI(BB, dl, TII->get(ARM::B)) .addMBB(exitMBB);
7577 MI->eraseFromParent(); // The pseudo instruction is gone now.
7581 case ARM::Int_eh_sjlj_setjmp:
7582 case ARM::Int_eh_sjlj_setjmp_nofp:
7583 case ARM::tInt_eh_sjlj_setjmp:
7584 case ARM::t2Int_eh_sjlj_setjmp:
7585 case ARM::t2Int_eh_sjlj_setjmp_nofp:
7586 EmitSjLjDispatchBlock(MI, BB);
7591 // To insert an ABS instruction, we have to insert the
7592 // diamond control-flow pattern. The incoming instruction knows the
7593 // source vreg to test against 0, the destination vreg to set,
7594 // the condition code register to branch on, the
7595 // true/false values to select between, and a branch opcode to use.
7600 // BCC (branch to SinkBB if V0 >= 0)
7601 // RSBBB: V3 = RSBri V2, 0 (compute ABS if V2 < 0)
7602 // SinkBB: V1 = PHI(V2, V3)
7603 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7604 MachineFunction::iterator BBI = BB;
7606 MachineFunction *Fn = BB->getParent();
7607 MachineBasicBlock *RSBBB = Fn->CreateMachineBasicBlock(LLVM_BB);
7608 MachineBasicBlock *SinkBB = Fn->CreateMachineBasicBlock(LLVM_BB);
7609 Fn->insert(BBI, RSBBB);
7610 Fn->insert(BBI, SinkBB);
7612 unsigned int ABSSrcReg = MI->getOperand(1).getReg();
7613 unsigned int ABSDstReg = MI->getOperand(0).getReg();
7614 bool isThumb2 = Subtarget->isThumb2();
7615 MachineRegisterInfo &MRI = Fn->getRegInfo();
7616 // In Thumb mode S must not be specified if source register is the SP or
7617 // PC and if destination register is the SP, so restrict register class
7618 unsigned NewRsbDstReg =
7619 MRI.createVirtualRegister(isThumb2 ? &ARM::rGPRRegClass : &ARM::GPRRegClass);
7621 // Transfer the remainder of BB and its successor edges to sinkMBB.
7622 SinkBB->splice(SinkBB->begin(), BB,
7623 std::next(MachineBasicBlock::iterator(MI)), BB->end());
7624 SinkBB->transferSuccessorsAndUpdatePHIs(BB);
7626 BB->addSuccessor(RSBBB);
7627 BB->addSuccessor(SinkBB);
7629 // fall through to SinkMBB
7630 RSBBB->addSuccessor(SinkBB);
7632 // insert a cmp at the end of BB
7633 AddDefaultPred(BuildMI(BB, dl,
7634 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7635 .addReg(ABSSrcReg).addImm(0));
7637 // insert a bcc with opposite CC to ARMCC::MI at the end of BB
7639 TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)).addMBB(SinkBB)
7640 .addImm(ARMCC::getOppositeCondition(ARMCC::MI)).addReg(ARM::CPSR);
7642 // insert rsbri in RSBBB
7643 // Note: BCC and rsbri will be converted into predicated rsbmi
7644 // by if-conversion pass
7645 BuildMI(*RSBBB, RSBBB->begin(), dl,
7646 TII->get(isThumb2 ? ARM::t2RSBri : ARM::RSBri), NewRsbDstReg)
7647 .addReg(ABSSrcReg, RegState::Kill)
7648 .addImm(0).addImm((unsigned)ARMCC::AL).addReg(0).addReg(0);
7650 // insert PHI in SinkBB,
7651 // reuse ABSDstReg to not change uses of ABS instruction
7652 BuildMI(*SinkBB, SinkBB->begin(), dl,
7653 TII->get(ARM::PHI), ABSDstReg)
7654 .addReg(NewRsbDstReg).addMBB(RSBBB)
7655 .addReg(ABSSrcReg).addMBB(BB);
7657 // remove ABS instruction
7658 MI->eraseFromParent();
7660 // return last added BB
7663 case ARM::COPY_STRUCT_BYVAL_I32:
7665 return EmitStructByval(MI, BB);
7666 case ARM::WIN__CHKSTK:
7667 return EmitLowered__chkstk(MI, BB);
7671 void ARMTargetLowering::AdjustInstrPostInstrSelection(MachineInstr *MI,
7672 SDNode *Node) const {
7673 const MCInstrDesc *MCID = &MI->getDesc();
7674 // Adjust potentially 's' setting instructions after isel, i.e. ADC, SBC, RSB,
7675 // RSC. Coming out of isel, they have an implicit CPSR def, but the optional
7676 // operand is still set to noreg. If needed, set the optional operand's
7677 // register to CPSR, and remove the redundant implicit def.
7679 // e.g. ADCS (..., CPSR<imp-def>) -> ADC (... opt:CPSR<def>).
7681 // Rename pseudo opcodes.
7682 unsigned NewOpc = convertAddSubFlagsOpcode(MI->getOpcode());
7684 const ARMBaseInstrInfo *TII = Subtarget->getInstrInfo();
7685 MCID = &TII->get(NewOpc);
7687 assert(MCID->getNumOperands() == MI->getDesc().getNumOperands() + 1 &&
7688 "converted opcode should be the same except for cc_out");
7692 // Add the optional cc_out operand
7693 MI->addOperand(MachineOperand::CreateReg(0, /*isDef=*/true));
7695 unsigned ccOutIdx = MCID->getNumOperands() - 1;
7697 // Any ARM instruction that sets the 's' bit should specify an optional
7698 // "cc_out" operand in the last operand position.
7699 if (!MI->hasOptionalDef() || !MCID->OpInfo[ccOutIdx].isOptionalDef()) {
7700 assert(!NewOpc && "Optional cc_out operand required");
7703 // Look for an implicit def of CPSR added by MachineInstr ctor. Remove it
7704 // since we already have an optional CPSR def.
7705 bool definesCPSR = false;
7706 bool deadCPSR = false;
7707 for (unsigned i = MCID->getNumOperands(), e = MI->getNumOperands();
7709 const MachineOperand &MO = MI->getOperand(i);
7710 if (MO.isReg() && MO.isDef() && MO.getReg() == ARM::CPSR) {
7714 MI->RemoveOperand(i);
7719 assert(!NewOpc && "Optional cc_out operand required");
7722 assert(deadCPSR == !Node->hasAnyUseOfValue(1) && "inconsistent dead flag");
7724 assert(!MI->getOperand(ccOutIdx).getReg() &&
7725 "expect uninitialized optional cc_out operand");
7729 // If this instruction was defined with an optional CPSR def and its dag node
7730 // had a live implicit CPSR def, then activate the optional CPSR def.
7731 MachineOperand &MO = MI->getOperand(ccOutIdx);
7732 MO.setReg(ARM::CPSR);
7736 //===----------------------------------------------------------------------===//
7737 // ARM Optimization Hooks
7738 //===----------------------------------------------------------------------===//
7740 // Helper function that checks if N is a null or all ones constant.
7741 static inline bool isZeroOrAllOnes(SDValue N, bool AllOnes) {
7742 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N);
7745 return AllOnes ? C->isAllOnesValue() : C->isNullValue();
7748 // Return true if N is conditionally 0 or all ones.
7749 // Detects these expressions where cc is an i1 value:
7751 // (select cc 0, y) [AllOnes=0]
7752 // (select cc y, 0) [AllOnes=0]
7753 // (zext cc) [AllOnes=0]
7754 // (sext cc) [AllOnes=0/1]
7755 // (select cc -1, y) [AllOnes=1]
7756 // (select cc y, -1) [AllOnes=1]
7758 // Invert is set when N is the null/all ones constant when CC is false.
7759 // OtherOp is set to the alternative value of N.
7760 static bool isConditionalZeroOrAllOnes(SDNode *N, bool AllOnes,
7761 SDValue &CC, bool &Invert,
7763 SelectionDAG &DAG) {
7764 switch (N->getOpcode()) {
7765 default: return false;
7767 CC = N->getOperand(0);
7768 SDValue N1 = N->getOperand(1);
7769 SDValue N2 = N->getOperand(2);
7770 if (isZeroOrAllOnes(N1, AllOnes)) {
7775 if (isZeroOrAllOnes(N2, AllOnes)) {
7782 case ISD::ZERO_EXTEND:
7783 // (zext cc) can never be the all ones value.
7787 case ISD::SIGN_EXTEND: {
7788 EVT VT = N->getValueType(0);
7789 CC = N->getOperand(0);
7790 if (CC.getValueType() != MVT::i1)
7794 // When looking for an AllOnes constant, N is an sext, and the 'other'
7796 OtherOp = DAG.getConstant(0, VT);
7797 else if (N->getOpcode() == ISD::ZERO_EXTEND)
7798 // When looking for a 0 constant, N can be zext or sext.
7799 OtherOp = DAG.getConstant(1, VT);
7801 OtherOp = DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), VT);
7807 // Combine a constant select operand into its use:
7809 // (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
7810 // (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
7811 // (and (select cc, -1, c), x) -> (select cc, x, (and, x, c)) [AllOnes=1]
7812 // (or (select cc, 0, c), x) -> (select cc, x, (or, x, c))
7813 // (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c))
7815 // The transform is rejected if the select doesn't have a constant operand that
7816 // is null, or all ones when AllOnes is set.
7818 // Also recognize sext/zext from i1:
7820 // (add (zext cc), x) -> (select cc (add x, 1), x)
7821 // (add (sext cc), x) -> (select cc (add x, -1), x)
7823 // These transformations eventually create predicated instructions.
7825 // @param N The node to transform.
7826 // @param Slct The N operand that is a select.
7827 // @param OtherOp The other N operand (x above).
7828 // @param DCI Context.
7829 // @param AllOnes Require the select constant to be all ones instead of null.
7830 // @returns The new node, or SDValue() on failure.
7832 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
7833 TargetLowering::DAGCombinerInfo &DCI,
7834 bool AllOnes = false) {
7835 SelectionDAG &DAG = DCI.DAG;
7836 EVT VT = N->getValueType(0);
7837 SDValue NonConstantVal;
7840 if (!isConditionalZeroOrAllOnes(Slct.getNode(), AllOnes, CCOp, SwapSelectOps,
7841 NonConstantVal, DAG))
7844 // Slct is now know to be the desired identity constant when CC is true.
7845 SDValue TrueVal = OtherOp;
7846 SDValue FalseVal = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
7847 OtherOp, NonConstantVal);
7848 // Unless SwapSelectOps says CC should be false.
7850 std::swap(TrueVal, FalseVal);
7852 return DAG.getNode(ISD::SELECT, SDLoc(N), VT,
7853 CCOp, TrueVal, FalseVal);
7856 // Attempt combineSelectAndUse on each operand of a commutative operator N.
7858 SDValue combineSelectAndUseCommutative(SDNode *N, bool AllOnes,
7859 TargetLowering::DAGCombinerInfo &DCI) {
7860 SDValue N0 = N->getOperand(0);
7861 SDValue N1 = N->getOperand(1);
7862 if (N0.getNode()->hasOneUse()) {
7863 SDValue Result = combineSelectAndUse(N, N0, N1, DCI, AllOnes);
7864 if (Result.getNode())
7867 if (N1.getNode()->hasOneUse()) {
7868 SDValue Result = combineSelectAndUse(N, N1, N0, DCI, AllOnes);
7869 if (Result.getNode())
7875 // AddCombineToVPADDL- For pair-wise add on neon, use the vpaddl instruction
7876 // (only after legalization).
7877 static SDValue AddCombineToVPADDL(SDNode *N, SDValue N0, SDValue N1,
7878 TargetLowering::DAGCombinerInfo &DCI,
7879 const ARMSubtarget *Subtarget) {
7881 // Only perform optimization if after legalize, and if NEON is available. We
7882 // also expected both operands to be BUILD_VECTORs.
7883 if (DCI.isBeforeLegalize() || !Subtarget->hasNEON()
7884 || N0.getOpcode() != ISD::BUILD_VECTOR
7885 || N1.getOpcode() != ISD::BUILD_VECTOR)
7888 // Check output type since VPADDL operand elements can only be 8, 16, or 32.
7889 EVT VT = N->getValueType(0);
7890 if (!VT.isInteger() || VT.getVectorElementType() == MVT::i64)
7893 // Check that the vector operands are of the right form.
7894 // N0 and N1 are BUILD_VECTOR nodes with N number of EXTRACT_VECTOR
7895 // operands, where N is the size of the formed vector.
7896 // Each EXTRACT_VECTOR should have the same input vector and odd or even
7897 // index such that we have a pair wise add pattern.
7899 // Grab the vector that all EXTRACT_VECTOR nodes should be referencing.
7900 if (N0->getOperand(0)->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
7902 SDValue Vec = N0->getOperand(0)->getOperand(0);
7903 SDNode *V = Vec.getNode();
7904 unsigned nextIndex = 0;
7906 // For each operands to the ADD which are BUILD_VECTORs,
7907 // check to see if each of their operands are an EXTRACT_VECTOR with
7908 // the same vector and appropriate index.
7909 for (unsigned i = 0, e = N0->getNumOperands(); i != e; ++i) {
7910 if (N0->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT
7911 && N1->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
7913 SDValue ExtVec0 = N0->getOperand(i);
7914 SDValue ExtVec1 = N1->getOperand(i);
7916 // First operand is the vector, verify its the same.
7917 if (V != ExtVec0->getOperand(0).getNode() ||
7918 V != ExtVec1->getOperand(0).getNode())
7921 // Second is the constant, verify its correct.
7922 ConstantSDNode *C0 = dyn_cast<ConstantSDNode>(ExtVec0->getOperand(1));
7923 ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(ExtVec1->getOperand(1));
7925 // For the constant, we want to see all the even or all the odd.
7926 if (!C0 || !C1 || C0->getZExtValue() != nextIndex
7927 || C1->getZExtValue() != nextIndex+1)
7936 // Create VPADDL node.
7937 SelectionDAG &DAG = DCI.DAG;
7938 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7940 // Build operand list.
7941 SmallVector<SDValue, 8> Ops;
7942 Ops.push_back(DAG.getConstant(Intrinsic::arm_neon_vpaddls,
7943 TLI.getPointerTy()));
7945 // Input is the vector.
7948 // Get widened type and narrowed type.
7950 unsigned numElem = VT.getVectorNumElements();
7952 EVT inputLaneType = Vec.getValueType().getVectorElementType();
7953 switch (inputLaneType.getSimpleVT().SimpleTy) {
7954 case MVT::i8: widenType = MVT::getVectorVT(MVT::i16, numElem); break;
7955 case MVT::i16: widenType = MVT::getVectorVT(MVT::i32, numElem); break;
7956 case MVT::i32: widenType = MVT::getVectorVT(MVT::i64, numElem); break;
7958 llvm_unreachable("Invalid vector element type for padd optimization.");
7961 SDValue tmp = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N), widenType, Ops);
7962 unsigned ExtOp = VT.bitsGT(tmp.getValueType()) ? ISD::ANY_EXTEND : ISD::TRUNCATE;
7963 return DAG.getNode(ExtOp, SDLoc(N), VT, tmp);
7966 static SDValue findMUL_LOHI(SDValue V) {
7967 if (V->getOpcode() == ISD::UMUL_LOHI ||
7968 V->getOpcode() == ISD::SMUL_LOHI)
7973 static SDValue AddCombineTo64bitMLAL(SDNode *AddcNode,
7974 TargetLowering::DAGCombinerInfo &DCI,
7975 const ARMSubtarget *Subtarget) {
7977 if (Subtarget->isThumb1Only()) return SDValue();
7979 // Only perform the checks after legalize when the pattern is available.
7980 if (DCI.isBeforeLegalize()) return SDValue();
7982 // Look for multiply add opportunities.
7983 // The pattern is a ISD::UMUL_LOHI followed by two add nodes, where
7984 // each add nodes consumes a value from ISD::UMUL_LOHI and there is
7985 // a glue link from the first add to the second add.
7986 // If we find this pattern, we can replace the U/SMUL_LOHI, ADDC, and ADDE by
7987 // a S/UMLAL instruction.
7990 // \ / \ [no multiline comment]
7996 assert(AddcNode->getOpcode() == ISD::ADDC && "Expect an ADDC");
7997 SDValue AddcOp0 = AddcNode->getOperand(0);
7998 SDValue AddcOp1 = AddcNode->getOperand(1);
8000 // Check if the two operands are from the same mul_lohi node.
8001 if (AddcOp0.getNode() == AddcOp1.getNode())
8004 assert(AddcNode->getNumValues() == 2 &&
8005 AddcNode->getValueType(0) == MVT::i32 &&
8006 "Expect ADDC with two result values. First: i32");
8008 // Check that we have a glued ADDC node.
8009 if (AddcNode->getValueType(1) != MVT::Glue)
8012 // Check that the ADDC adds the low result of the S/UMUL_LOHI.
8013 if (AddcOp0->getOpcode() != ISD::UMUL_LOHI &&
8014 AddcOp0->getOpcode() != ISD::SMUL_LOHI &&
8015 AddcOp1->getOpcode() != ISD::UMUL_LOHI &&
8016 AddcOp1->getOpcode() != ISD::SMUL_LOHI)
8019 // Look for the glued ADDE.
8020 SDNode* AddeNode = AddcNode->getGluedUser();
8024 // Make sure it is really an ADDE.
8025 if (AddeNode->getOpcode() != ISD::ADDE)
8028 assert(AddeNode->getNumOperands() == 3 &&
8029 AddeNode->getOperand(2).getValueType() == MVT::Glue &&
8030 "ADDE node has the wrong inputs");
8032 // Check for the triangle shape.
8033 SDValue AddeOp0 = AddeNode->getOperand(0);
8034 SDValue AddeOp1 = AddeNode->getOperand(1);
8036 // Make sure that the ADDE operands are not coming from the same node.
8037 if (AddeOp0.getNode() == AddeOp1.getNode())
8040 // Find the MUL_LOHI node walking up ADDE's operands.
8041 bool IsLeftOperandMUL = false;
8042 SDValue MULOp = findMUL_LOHI(AddeOp0);
8043 if (MULOp == SDValue())
8044 MULOp = findMUL_LOHI(AddeOp1);
8046 IsLeftOperandMUL = true;
8047 if (MULOp == SDValue())
8050 // Figure out the right opcode.
8051 unsigned Opc = MULOp->getOpcode();
8052 unsigned FinalOpc = (Opc == ISD::SMUL_LOHI) ? ARMISD::SMLAL : ARMISD::UMLAL;
8054 // Figure out the high and low input values to the MLAL node.
8055 SDValue* HiAdd = nullptr;
8056 SDValue* LoMul = nullptr;
8057 SDValue* LowAdd = nullptr;
8059 // Ensure that ADDE is from high result of ISD::SMUL_LOHI.
8060 if ((AddeOp0 != MULOp.getValue(1)) && (AddeOp1 != MULOp.getValue(1)))
8063 if (IsLeftOperandMUL)
8069 // Ensure that LoMul and LowAdd are taken from correct ISD::SMUL_LOHI node
8070 // whose low result is fed to the ADDC we are checking.
8072 if (AddcOp0 == MULOp.getValue(0)) {
8076 if (AddcOp1 == MULOp.getValue(0)) {
8084 // Create the merged node.
8085 SelectionDAG &DAG = DCI.DAG;
8087 // Build operand list.
8088 SmallVector<SDValue, 8> Ops;
8089 Ops.push_back(LoMul->getOperand(0));
8090 Ops.push_back(LoMul->getOperand(1));
8091 Ops.push_back(*LowAdd);
8092 Ops.push_back(*HiAdd);
8094 SDValue MLALNode = DAG.getNode(FinalOpc, SDLoc(AddcNode),
8095 DAG.getVTList(MVT::i32, MVT::i32), Ops);
8097 // Replace the ADDs' nodes uses by the MLA node's values.
8098 SDValue HiMLALResult(MLALNode.getNode(), 1);
8099 DAG.ReplaceAllUsesOfValueWith(SDValue(AddeNode, 0), HiMLALResult);
8101 SDValue LoMLALResult(MLALNode.getNode(), 0);
8102 DAG.ReplaceAllUsesOfValueWith(SDValue(AddcNode, 0), LoMLALResult);
8104 // Return original node to notify the driver to stop replacing.
8105 SDValue resNode(AddcNode, 0);
8109 /// PerformADDCCombine - Target-specific dag combine transform from
8110 /// ISD::ADDC, ISD::ADDE, and ISD::MUL_LOHI to MLAL.
8111 static SDValue PerformADDCCombine(SDNode *N,
8112 TargetLowering::DAGCombinerInfo &DCI,
8113 const ARMSubtarget *Subtarget) {
8115 return AddCombineTo64bitMLAL(N, DCI, Subtarget);
8119 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
8120 /// operands N0 and N1. This is a helper for PerformADDCombine that is
8121 /// called with the default operands, and if that fails, with commuted
8123 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
8124 TargetLowering::DAGCombinerInfo &DCI,
8125 const ARMSubtarget *Subtarget){
8127 // Attempt to create vpaddl for this add.
8128 SDValue Result = AddCombineToVPADDL(N, N0, N1, DCI, Subtarget);
8129 if (Result.getNode())
8132 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
8133 if (N0.getNode()->hasOneUse()) {
8134 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
8135 if (Result.getNode()) return Result;
8140 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
8142 static SDValue PerformADDCombine(SDNode *N,
8143 TargetLowering::DAGCombinerInfo &DCI,
8144 const ARMSubtarget *Subtarget) {
8145 SDValue N0 = N->getOperand(0);
8146 SDValue N1 = N->getOperand(1);
8148 // First try with the default operand order.
8149 SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI, Subtarget);
8150 if (Result.getNode())
8153 // If that didn't work, try again with the operands commuted.
8154 return PerformADDCombineWithOperands(N, N1, N0, DCI, Subtarget);
8157 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
8159 static SDValue PerformSUBCombine(SDNode *N,
8160 TargetLowering::DAGCombinerInfo &DCI) {
8161 SDValue N0 = N->getOperand(0);
8162 SDValue N1 = N->getOperand(1);
8164 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
8165 if (N1.getNode()->hasOneUse()) {
8166 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
8167 if (Result.getNode()) return Result;
8173 /// PerformVMULCombine
8174 /// Distribute (A + B) * C to (A * C) + (B * C) to take advantage of the
8175 /// special multiplier accumulator forwarding.
8181 // However, for (A + B) * (A + B),
8188 static SDValue PerformVMULCombine(SDNode *N,
8189 TargetLowering::DAGCombinerInfo &DCI,
8190 const ARMSubtarget *Subtarget) {
8191 if (!Subtarget->hasVMLxForwarding())
8194 SelectionDAG &DAG = DCI.DAG;
8195 SDValue N0 = N->getOperand(0);
8196 SDValue N1 = N->getOperand(1);
8197 unsigned Opcode = N0.getOpcode();
8198 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
8199 Opcode != ISD::FADD && Opcode != ISD::FSUB) {
8200 Opcode = N1.getOpcode();
8201 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
8202 Opcode != ISD::FADD && Opcode != ISD::FSUB)
8210 EVT VT = N->getValueType(0);
8212 SDValue N00 = N0->getOperand(0);
8213 SDValue N01 = N0->getOperand(1);
8214 return DAG.getNode(Opcode, DL, VT,
8215 DAG.getNode(ISD::MUL, DL, VT, N00, N1),
8216 DAG.getNode(ISD::MUL, DL, VT, N01, N1));
8219 static SDValue PerformMULCombine(SDNode *N,
8220 TargetLowering::DAGCombinerInfo &DCI,
8221 const ARMSubtarget *Subtarget) {
8222 SelectionDAG &DAG = DCI.DAG;
8224 if (Subtarget->isThumb1Only())
8227 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
8230 EVT VT = N->getValueType(0);
8231 if (VT.is64BitVector() || VT.is128BitVector())
8232 return PerformVMULCombine(N, DCI, Subtarget);
8236 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
8240 int64_t MulAmt = C->getSExtValue();
8241 unsigned ShiftAmt = countTrailingZeros<uint64_t>(MulAmt);
8243 ShiftAmt = ShiftAmt & (32 - 1);
8244 SDValue V = N->getOperand(0);
8248 MulAmt >>= ShiftAmt;
8251 if (isPowerOf2_32(MulAmt - 1)) {
8252 // (mul x, 2^N + 1) => (add (shl x, N), x)
8253 Res = DAG.getNode(ISD::ADD, DL, VT,
8255 DAG.getNode(ISD::SHL, DL, VT,
8257 DAG.getConstant(Log2_32(MulAmt - 1),
8259 } else if (isPowerOf2_32(MulAmt + 1)) {
8260 // (mul x, 2^N - 1) => (sub (shl x, N), x)
8261 Res = DAG.getNode(ISD::SUB, DL, VT,
8262 DAG.getNode(ISD::SHL, DL, VT,
8264 DAG.getConstant(Log2_32(MulAmt + 1),
8270 uint64_t MulAmtAbs = -MulAmt;
8271 if (isPowerOf2_32(MulAmtAbs + 1)) {
8272 // (mul x, -(2^N - 1)) => (sub x, (shl x, N))
8273 Res = DAG.getNode(ISD::SUB, DL, VT,
8275 DAG.getNode(ISD::SHL, DL, VT,
8277 DAG.getConstant(Log2_32(MulAmtAbs + 1),
8279 } else if (isPowerOf2_32(MulAmtAbs - 1)) {
8280 // (mul x, -(2^N + 1)) => - (add (shl x, N), x)
8281 Res = DAG.getNode(ISD::ADD, DL, VT,
8283 DAG.getNode(ISD::SHL, DL, VT,
8285 DAG.getConstant(Log2_32(MulAmtAbs-1),
8287 Res = DAG.getNode(ISD::SUB, DL, VT,
8288 DAG.getConstant(0, MVT::i32),Res);
8295 Res = DAG.getNode(ISD::SHL, DL, VT,
8296 Res, DAG.getConstant(ShiftAmt, MVT::i32));
8298 // Do not add new nodes to DAG combiner worklist.
8299 DCI.CombineTo(N, Res, false);
8303 static SDValue PerformANDCombine(SDNode *N,
8304 TargetLowering::DAGCombinerInfo &DCI,
8305 const ARMSubtarget *Subtarget) {
8307 // Attempt to use immediate-form VBIC
8308 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
8310 EVT VT = N->getValueType(0);
8311 SelectionDAG &DAG = DCI.DAG;
8313 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8316 APInt SplatBits, SplatUndef;
8317 unsigned SplatBitSize;
8320 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
8321 if (SplatBitSize <= 64) {
8323 SDValue Val = isNEONModifiedImm((~SplatBits).getZExtValue(),
8324 SplatUndef.getZExtValue(), SplatBitSize,
8325 DAG, VbicVT, VT.is128BitVector(),
8327 if (Val.getNode()) {
8329 DAG.getNode(ISD::BITCAST, dl, VbicVT, N->getOperand(0));
8330 SDValue Vbic = DAG.getNode(ARMISD::VBICIMM, dl, VbicVT, Input, Val);
8331 return DAG.getNode(ISD::BITCAST, dl, VT, Vbic);
8336 if (!Subtarget->isThumb1Only()) {
8337 // fold (and (select cc, -1, c), x) -> (select cc, x, (and, x, c))
8338 SDValue Result = combineSelectAndUseCommutative(N, true, DCI);
8339 if (Result.getNode())
8346 /// PerformORCombine - Target-specific dag combine xforms for ISD::OR
8347 static SDValue PerformORCombine(SDNode *N,
8348 TargetLowering::DAGCombinerInfo &DCI,
8349 const ARMSubtarget *Subtarget) {
8350 // Attempt to use immediate-form VORR
8351 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
8353 EVT VT = N->getValueType(0);
8354 SelectionDAG &DAG = DCI.DAG;
8356 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8359 APInt SplatBits, SplatUndef;
8360 unsigned SplatBitSize;
8362 if (BVN && Subtarget->hasNEON() &&
8363 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
8364 if (SplatBitSize <= 64) {
8366 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
8367 SplatUndef.getZExtValue(), SplatBitSize,
8368 DAG, VorrVT, VT.is128BitVector(),
8370 if (Val.getNode()) {
8372 DAG.getNode(ISD::BITCAST, dl, VorrVT, N->getOperand(0));
8373 SDValue Vorr = DAG.getNode(ARMISD::VORRIMM, dl, VorrVT, Input, Val);
8374 return DAG.getNode(ISD::BITCAST, dl, VT, Vorr);
8379 if (!Subtarget->isThumb1Only()) {
8380 // fold (or (select cc, 0, c), x) -> (select cc, x, (or, x, c))
8381 SDValue Result = combineSelectAndUseCommutative(N, false, DCI);
8382 if (Result.getNode())
8386 // The code below optimizes (or (and X, Y), Z).
8387 // The AND operand needs to have a single user to make these optimizations
8389 SDValue N0 = N->getOperand(0);
8390 if (N0.getOpcode() != ISD::AND || !N0.hasOneUse())
8392 SDValue N1 = N->getOperand(1);
8394 // (or (and B, A), (and C, ~A)) => (VBSL A, B, C) when A is a constant.
8395 if (Subtarget->hasNEON() && N1.getOpcode() == ISD::AND && VT.isVector() &&
8396 DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
8398 unsigned SplatBitSize;
8401 APInt SplatBits0, SplatBits1;
8402 BuildVectorSDNode *BVN0 = dyn_cast<BuildVectorSDNode>(N0->getOperand(1));
8403 BuildVectorSDNode *BVN1 = dyn_cast<BuildVectorSDNode>(N1->getOperand(1));
8404 // Ensure that the second operand of both ands are constants
8405 if (BVN0 && BVN0->isConstantSplat(SplatBits0, SplatUndef, SplatBitSize,
8406 HasAnyUndefs) && !HasAnyUndefs) {
8407 if (BVN1 && BVN1->isConstantSplat(SplatBits1, SplatUndef, SplatBitSize,
8408 HasAnyUndefs) && !HasAnyUndefs) {
8409 // Ensure that the bit width of the constants are the same and that
8410 // the splat arguments are logical inverses as per the pattern we
8411 // are trying to simplify.
8412 if (SplatBits0.getBitWidth() == SplatBits1.getBitWidth() &&
8413 SplatBits0 == ~SplatBits1) {
8414 // Canonicalize the vector type to make instruction selection
8416 EVT CanonicalVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
8417 SDValue Result = DAG.getNode(ARMISD::VBSL, dl, CanonicalVT,
8421 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
8427 // Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when
8430 // BFI is only available on V6T2+
8431 if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops())
8435 // 1) or (and A, mask), val => ARMbfi A, val, mask
8436 // iff (val & mask) == val
8438 // 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
8439 // 2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2)
8440 // && mask == ~mask2
8441 // 2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2)
8442 // && ~mask == mask2
8443 // (i.e., copy a bitfield value into another bitfield of the same width)
8448 SDValue N00 = N0.getOperand(0);
8450 // The value and the mask need to be constants so we can verify this is
8451 // actually a bitfield set. If the mask is 0xffff, we can do better
8452 // via a movt instruction, so don't use BFI in that case.
8453 SDValue MaskOp = N0.getOperand(1);
8454 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(MaskOp);
8457 unsigned Mask = MaskC->getZExtValue();
8461 // Case (1): or (and A, mask), val => ARMbfi A, val, mask
8462 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
8464 unsigned Val = N1C->getZExtValue();
8465 if ((Val & ~Mask) != Val)
8468 if (ARM::isBitFieldInvertedMask(Mask)) {
8469 Val >>= countTrailingZeros(~Mask);
8471 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00,
8472 DAG.getConstant(Val, MVT::i32),
8473 DAG.getConstant(Mask, MVT::i32));
8475 // Do not add new nodes to DAG combiner worklist.
8476 DCI.CombineTo(N, Res, false);
8479 } else if (N1.getOpcode() == ISD::AND) {
8480 // case (2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
8481 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
8484 unsigned Mask2 = N11C->getZExtValue();
8486 // Mask and ~Mask2 (or reverse) must be equivalent for the BFI pattern
8488 if (ARM::isBitFieldInvertedMask(Mask) &&
8490 // The pack halfword instruction works better for masks that fit it,
8491 // so use that when it's available.
8492 if (Subtarget->hasT2ExtractPack() &&
8493 (Mask == 0xffff || Mask == 0xffff0000))
8496 unsigned amt = countTrailingZeros(Mask2);
8497 Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0),
8498 DAG.getConstant(amt, MVT::i32));
8499 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00, Res,
8500 DAG.getConstant(Mask, MVT::i32));
8501 // Do not add new nodes to DAG combiner worklist.
8502 DCI.CombineTo(N, Res, false);
8504 } else if (ARM::isBitFieldInvertedMask(~Mask) &&
8506 // The pack halfword instruction works better for masks that fit it,
8507 // so use that when it's available.
8508 if (Subtarget->hasT2ExtractPack() &&
8509 (Mask2 == 0xffff || Mask2 == 0xffff0000))
8512 unsigned lsb = countTrailingZeros(Mask);
8513 Res = DAG.getNode(ISD::SRL, DL, VT, N00,
8514 DAG.getConstant(lsb, MVT::i32));
8515 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res,
8516 DAG.getConstant(Mask2, MVT::i32));
8517 // Do not add new nodes to DAG combiner worklist.
8518 DCI.CombineTo(N, Res, false);
8523 if (DAG.MaskedValueIsZero(N1, MaskC->getAPIntValue()) &&
8524 N00.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N00.getOperand(1)) &&
8525 ARM::isBitFieldInvertedMask(~Mask)) {
8526 // Case (3): or (and (shl A, #shamt), mask), B => ARMbfi B, A, ~mask
8527 // where lsb(mask) == #shamt and masked bits of B are known zero.
8528 SDValue ShAmt = N00.getOperand(1);
8529 unsigned ShAmtC = cast<ConstantSDNode>(ShAmt)->getZExtValue();
8530 unsigned LSB = countTrailingZeros(Mask);
8534 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1, N00.getOperand(0),
8535 DAG.getConstant(~Mask, MVT::i32));
8537 // Do not add new nodes to DAG combiner worklist.
8538 DCI.CombineTo(N, Res, false);
8544 static SDValue PerformXORCombine(SDNode *N,
8545 TargetLowering::DAGCombinerInfo &DCI,
8546 const ARMSubtarget *Subtarget) {
8547 EVT VT = N->getValueType(0);
8548 SelectionDAG &DAG = DCI.DAG;
8550 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8553 if (!Subtarget->isThumb1Only()) {
8554 // fold (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c))
8555 SDValue Result = combineSelectAndUseCommutative(N, false, DCI);
8556 if (Result.getNode())
8563 /// PerformBFICombine - (bfi A, (and B, Mask1), Mask2) -> (bfi A, B, Mask2) iff
8564 /// the bits being cleared by the AND are not demanded by the BFI.
8565 static SDValue PerformBFICombine(SDNode *N,
8566 TargetLowering::DAGCombinerInfo &DCI) {
8567 SDValue N1 = N->getOperand(1);
8568 if (N1.getOpcode() == ISD::AND) {
8569 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
8572 unsigned InvMask = cast<ConstantSDNode>(N->getOperand(2))->getZExtValue();
8573 unsigned LSB = countTrailingZeros(~InvMask);
8574 unsigned Width = (32 - countLeadingZeros(~InvMask)) - LSB;
8576 static_cast<unsigned>(std::numeric_limits<unsigned>::digits) &&
8577 "undefined behavior");
8578 unsigned Mask = (1u << Width) - 1;
8579 unsigned Mask2 = N11C->getZExtValue();
8580 if ((Mask & (~Mask2)) == 0)
8581 return DCI.DAG.getNode(ARMISD::BFI, SDLoc(N), N->getValueType(0),
8582 N->getOperand(0), N1.getOperand(0),
8588 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
8589 /// ARMISD::VMOVRRD.
8590 static SDValue PerformVMOVRRDCombine(SDNode *N,
8591 TargetLowering::DAGCombinerInfo &DCI,
8592 const ARMSubtarget *Subtarget) {
8593 // vmovrrd(vmovdrr x, y) -> x,y
8594 SDValue InDouble = N->getOperand(0);
8595 if (InDouble.getOpcode() == ARMISD::VMOVDRR && !Subtarget->isFPOnlySP())
8596 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
8598 // vmovrrd(load f64) -> (load i32), (load i32)
8599 SDNode *InNode = InDouble.getNode();
8600 if (ISD::isNormalLoad(InNode) && InNode->hasOneUse() &&
8601 InNode->getValueType(0) == MVT::f64 &&
8602 InNode->getOperand(1).getOpcode() == ISD::FrameIndex &&
8603 !cast<LoadSDNode>(InNode)->isVolatile()) {
8604 // TODO: Should this be done for non-FrameIndex operands?
8605 LoadSDNode *LD = cast<LoadSDNode>(InNode);
8607 SelectionDAG &DAG = DCI.DAG;
8609 SDValue BasePtr = LD->getBasePtr();
8610 SDValue NewLD1 = DAG.getLoad(MVT::i32, DL, LD->getChain(), BasePtr,
8611 LD->getPointerInfo(), LD->isVolatile(),
8612 LD->isNonTemporal(), LD->isInvariant(),
8613 LD->getAlignment());
8615 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
8616 DAG.getConstant(4, MVT::i32));
8617 SDValue NewLD2 = DAG.getLoad(MVT::i32, DL, NewLD1.getValue(1), OffsetPtr,
8618 LD->getPointerInfo(), LD->isVolatile(),
8619 LD->isNonTemporal(), LD->isInvariant(),
8620 std::min(4U, LD->getAlignment() / 2));
8622 DAG.ReplaceAllUsesOfValueWith(SDValue(LD, 1), NewLD2.getValue(1));
8623 if (DCI.DAG.getTargetLoweringInfo().isBigEndian())
8624 std::swap (NewLD1, NewLD2);
8625 SDValue Result = DCI.CombineTo(N, NewLD1, NewLD2);
8632 /// PerformVMOVDRRCombine - Target-specific dag combine xforms for
8633 /// ARMISD::VMOVDRR. This is also used for BUILD_VECTORs with 2 operands.
8634 static SDValue PerformVMOVDRRCombine(SDNode *N, SelectionDAG &DAG) {
8635 // N=vmovrrd(X); vmovdrr(N:0, N:1) -> bit_convert(X)
8636 SDValue Op0 = N->getOperand(0);
8637 SDValue Op1 = N->getOperand(1);
8638 if (Op0.getOpcode() == ISD::BITCAST)
8639 Op0 = Op0.getOperand(0);
8640 if (Op1.getOpcode() == ISD::BITCAST)
8641 Op1 = Op1.getOperand(0);
8642 if (Op0.getOpcode() == ARMISD::VMOVRRD &&
8643 Op0.getNode() == Op1.getNode() &&
8644 Op0.getResNo() == 0 && Op1.getResNo() == 1)
8645 return DAG.getNode(ISD::BITCAST, SDLoc(N),
8646 N->getValueType(0), Op0.getOperand(0));
8650 /// hasNormalLoadOperand - Check if any of the operands of a BUILD_VECTOR node
8651 /// are normal, non-volatile loads. If so, it is profitable to bitcast an
8652 /// i64 vector to have f64 elements, since the value can then be loaded
8653 /// directly into a VFP register.
8654 static bool hasNormalLoadOperand(SDNode *N) {
8655 unsigned NumElts = N->getValueType(0).getVectorNumElements();
8656 for (unsigned i = 0; i < NumElts; ++i) {
8657 SDNode *Elt = N->getOperand(i).getNode();
8658 if (ISD::isNormalLoad(Elt) && !cast<LoadSDNode>(Elt)->isVolatile())
8664 /// PerformBUILD_VECTORCombine - Target-specific dag combine xforms for
8665 /// ISD::BUILD_VECTOR.
8666 static SDValue PerformBUILD_VECTORCombine(SDNode *N,
8667 TargetLowering::DAGCombinerInfo &DCI,
8668 const ARMSubtarget *Subtarget) {
8669 // build_vector(N=ARMISD::VMOVRRD(X), N:1) -> bit_convert(X):
8670 // VMOVRRD is introduced when legalizing i64 types. It forces the i64 value
8671 // into a pair of GPRs, which is fine when the value is used as a scalar,
8672 // but if the i64 value is converted to a vector, we need to undo the VMOVRRD.
8673 SelectionDAG &DAG = DCI.DAG;
8674 if (N->getNumOperands() == 2) {
8675 SDValue RV = PerformVMOVDRRCombine(N, DAG);
8680 // Load i64 elements as f64 values so that type legalization does not split
8681 // them up into i32 values.
8682 EVT VT = N->getValueType(0);
8683 if (VT.getVectorElementType() != MVT::i64 || !hasNormalLoadOperand(N))
8686 SmallVector<SDValue, 8> Ops;
8687 unsigned NumElts = VT.getVectorNumElements();
8688 for (unsigned i = 0; i < NumElts; ++i) {
8689 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(i));
8691 // Make the DAGCombiner fold the bitcast.
8692 DCI.AddToWorklist(V.getNode());
8694 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, NumElts);
8695 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, FloatVT, Ops);
8696 return DAG.getNode(ISD::BITCAST, dl, VT, BV);
8699 /// \brief Target-specific dag combine xforms for ARMISD::BUILD_VECTOR.
8701 PerformARMBUILD_VECTORCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
8702 // ARMISD::BUILD_VECTOR is introduced when legalizing ISD::BUILD_VECTOR.
8703 // At that time, we may have inserted bitcasts from integer to float.
8704 // If these bitcasts have survived DAGCombine, change the lowering of this
8705 // BUILD_VECTOR in something more vector friendly, i.e., that does not
8706 // force to use floating point types.
8708 // Make sure we can change the type of the vector.
8709 // This is possible iff:
8710 // 1. The vector is only used in a bitcast to a integer type. I.e.,
8711 // 1.1. Vector is used only once.
8712 // 1.2. Use is a bit convert to an integer type.
8713 // 2. The size of its operands are 32-bits (64-bits are not legal).
8714 EVT VT = N->getValueType(0);
8715 EVT EltVT = VT.getVectorElementType();
8717 // Check 1.1. and 2.
8718 if (EltVT.getSizeInBits() != 32 || !N->hasOneUse())
8721 // By construction, the input type must be float.
8722 assert(EltVT == MVT::f32 && "Unexpected type!");
8725 SDNode *Use = *N->use_begin();
8726 if (Use->getOpcode() != ISD::BITCAST ||
8727 Use->getValueType(0).isFloatingPoint())
8730 // Check profitability.
8731 // Model is, if more than half of the relevant operands are bitcast from
8732 // i32, turn the build_vector into a sequence of insert_vector_elt.
8733 // Relevant operands are everything that is not statically
8734 // (i.e., at compile time) bitcasted.
8735 unsigned NumOfBitCastedElts = 0;
8736 unsigned NumElts = VT.getVectorNumElements();
8737 unsigned NumOfRelevantElts = NumElts;
8738 for (unsigned Idx = 0; Idx < NumElts; ++Idx) {
8739 SDValue Elt = N->getOperand(Idx);
8740 if (Elt->getOpcode() == ISD::BITCAST) {
8741 // Assume only bit cast to i32 will go away.
8742 if (Elt->getOperand(0).getValueType() == MVT::i32)
8743 ++NumOfBitCastedElts;
8744 } else if (Elt.getOpcode() == ISD::UNDEF || isa<ConstantSDNode>(Elt))
8745 // Constants are statically casted, thus do not count them as
8746 // relevant operands.
8747 --NumOfRelevantElts;
8750 // Check if more than half of the elements require a non-free bitcast.
8751 if (NumOfBitCastedElts <= NumOfRelevantElts / 2)
8754 SelectionDAG &DAG = DCI.DAG;
8755 // Create the new vector type.
8756 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
8757 // Check if the type is legal.
8758 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8759 if (!TLI.isTypeLegal(VecVT))
8763 // ARMISD::BUILD_VECTOR E1, E2, ..., EN.
8764 // => BITCAST INSERT_VECTOR_ELT
8765 // (INSERT_VECTOR_ELT (...), (BITCAST EN-1), N-1),
8767 SDValue Vec = DAG.getUNDEF(VecVT);
8769 for (unsigned Idx = 0 ; Idx < NumElts; ++Idx) {
8770 SDValue V = N->getOperand(Idx);
8771 if (V.getOpcode() == ISD::UNDEF)
8773 if (V.getOpcode() == ISD::BITCAST &&
8774 V->getOperand(0).getValueType() == MVT::i32)
8775 // Fold obvious case.
8776 V = V.getOperand(0);
8778 V = DAG.getNode(ISD::BITCAST, SDLoc(V), MVT::i32, V);
8779 // Make the DAGCombiner fold the bitcasts.
8780 DCI.AddToWorklist(V.getNode());
8782 SDValue LaneIdx = DAG.getConstant(Idx, MVT::i32);
8783 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VecVT, Vec, V, LaneIdx);
8785 Vec = DAG.getNode(ISD::BITCAST, dl, VT, Vec);
8786 // Make the DAGCombiner fold the bitcasts.
8787 DCI.AddToWorklist(Vec.getNode());
8791 /// PerformInsertEltCombine - Target-specific dag combine xforms for
8792 /// ISD::INSERT_VECTOR_ELT.
8793 static SDValue PerformInsertEltCombine(SDNode *N,
8794 TargetLowering::DAGCombinerInfo &DCI) {
8795 // Bitcast an i64 load inserted into a vector to f64.
8796 // Otherwise, the i64 value will be legalized to a pair of i32 values.
8797 EVT VT = N->getValueType(0);
8798 SDNode *Elt = N->getOperand(1).getNode();
8799 if (VT.getVectorElementType() != MVT::i64 ||
8800 !ISD::isNormalLoad(Elt) || cast<LoadSDNode>(Elt)->isVolatile())
8803 SelectionDAG &DAG = DCI.DAG;
8805 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
8806 VT.getVectorNumElements());
8807 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, N->getOperand(0));
8808 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(1));
8809 // Make the DAGCombiner fold the bitcasts.
8810 DCI.AddToWorklist(Vec.getNode());
8811 DCI.AddToWorklist(V.getNode());
8812 SDValue InsElt = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, FloatVT,
8813 Vec, V, N->getOperand(2));
8814 return DAG.getNode(ISD::BITCAST, dl, VT, InsElt);
8817 /// PerformVECTOR_SHUFFLECombine - Target-specific dag combine xforms for
8818 /// ISD::VECTOR_SHUFFLE.
8819 static SDValue PerformVECTOR_SHUFFLECombine(SDNode *N, SelectionDAG &DAG) {
8820 // The LLVM shufflevector instruction does not require the shuffle mask
8821 // length to match the operand vector length, but ISD::VECTOR_SHUFFLE does
8822 // have that requirement. When translating to ISD::VECTOR_SHUFFLE, if the
8823 // operands do not match the mask length, they are extended by concatenating
8824 // them with undef vectors. That is probably the right thing for other
8825 // targets, but for NEON it is better to concatenate two double-register
8826 // size vector operands into a single quad-register size vector. Do that
8827 // transformation here:
8828 // shuffle(concat(v1, undef), concat(v2, undef)) ->
8829 // shuffle(concat(v1, v2), undef)
8830 SDValue Op0 = N->getOperand(0);
8831 SDValue Op1 = N->getOperand(1);
8832 if (Op0.getOpcode() != ISD::CONCAT_VECTORS ||
8833 Op1.getOpcode() != ISD::CONCAT_VECTORS ||
8834 Op0.getNumOperands() != 2 ||
8835 Op1.getNumOperands() != 2)
8837 SDValue Concat0Op1 = Op0.getOperand(1);
8838 SDValue Concat1Op1 = Op1.getOperand(1);
8839 if (Concat0Op1.getOpcode() != ISD::UNDEF ||
8840 Concat1Op1.getOpcode() != ISD::UNDEF)
8842 // Skip the transformation if any of the types are illegal.
8843 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8844 EVT VT = N->getValueType(0);
8845 if (!TLI.isTypeLegal(VT) ||
8846 !TLI.isTypeLegal(Concat0Op1.getValueType()) ||
8847 !TLI.isTypeLegal(Concat1Op1.getValueType()))
8850 SDValue NewConcat = DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT,
8851 Op0.getOperand(0), Op1.getOperand(0));
8852 // Translate the shuffle mask.
8853 SmallVector<int, 16> NewMask;
8854 unsigned NumElts = VT.getVectorNumElements();
8855 unsigned HalfElts = NumElts/2;
8856 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
8857 for (unsigned n = 0; n < NumElts; ++n) {
8858 int MaskElt = SVN->getMaskElt(n);
8860 if (MaskElt < (int)HalfElts)
8862 else if (MaskElt >= (int)NumElts && MaskElt < (int)(NumElts + HalfElts))
8863 NewElt = HalfElts + MaskElt - NumElts;
8864 NewMask.push_back(NewElt);
8866 return DAG.getVectorShuffle(VT, SDLoc(N), NewConcat,
8867 DAG.getUNDEF(VT), NewMask.data());
8870 /// CombineBaseUpdate - Target-specific DAG combine function for VLDDUP and
8871 /// NEON load/store intrinsics to merge base address updates.
8872 static SDValue CombineBaseUpdate(SDNode *N,
8873 TargetLowering::DAGCombinerInfo &DCI) {
8874 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
8877 SelectionDAG &DAG = DCI.DAG;
8878 bool isIntrinsic = (N->getOpcode() == ISD::INTRINSIC_VOID ||
8879 N->getOpcode() == ISD::INTRINSIC_W_CHAIN);
8880 unsigned AddrOpIdx = (isIntrinsic ? 2 : 1);
8881 SDValue Addr = N->getOperand(AddrOpIdx);
8883 // Search for a use of the address operand that is an increment.
8884 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(),
8885 UE = Addr.getNode()->use_end(); UI != UE; ++UI) {
8887 if (User->getOpcode() != ISD::ADD ||
8888 UI.getUse().getResNo() != Addr.getResNo())
8891 // Check that the add is independent of the load/store. Otherwise, folding
8892 // it would create a cycle.
8893 if (User->isPredecessorOf(N) || N->isPredecessorOf(User))
8896 // Find the new opcode for the updating load/store.
8898 bool isLaneOp = false;
8899 unsigned NewOpc = 0;
8900 unsigned NumVecs = 0;
8902 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
8904 default: llvm_unreachable("unexpected intrinsic for Neon base update");
8905 case Intrinsic::arm_neon_vld1: NewOpc = ARMISD::VLD1_UPD;
8907 case Intrinsic::arm_neon_vld2: NewOpc = ARMISD::VLD2_UPD;
8909 case Intrinsic::arm_neon_vld3: NewOpc = ARMISD::VLD3_UPD;
8911 case Intrinsic::arm_neon_vld4: NewOpc = ARMISD::VLD4_UPD;
8913 case Intrinsic::arm_neon_vld2lane: NewOpc = ARMISD::VLD2LN_UPD;
8914 NumVecs = 2; isLaneOp = true; break;
8915 case Intrinsic::arm_neon_vld3lane: NewOpc = ARMISD::VLD3LN_UPD;
8916 NumVecs = 3; isLaneOp = true; break;
8917 case Intrinsic::arm_neon_vld4lane: NewOpc = ARMISD::VLD4LN_UPD;
8918 NumVecs = 4; isLaneOp = true; break;
8919 case Intrinsic::arm_neon_vst1: NewOpc = ARMISD::VST1_UPD;
8920 NumVecs = 1; isLoad = false; break;
8921 case Intrinsic::arm_neon_vst2: NewOpc = ARMISD::VST2_UPD;
8922 NumVecs = 2; isLoad = false; break;
8923 case Intrinsic::arm_neon_vst3: NewOpc = ARMISD::VST3_UPD;
8924 NumVecs = 3; isLoad = false; break;
8925 case Intrinsic::arm_neon_vst4: NewOpc = ARMISD::VST4_UPD;
8926 NumVecs = 4; isLoad = false; break;
8927 case Intrinsic::arm_neon_vst2lane: NewOpc = ARMISD::VST2LN_UPD;
8928 NumVecs = 2; isLoad = false; isLaneOp = true; break;
8929 case Intrinsic::arm_neon_vst3lane: NewOpc = ARMISD::VST3LN_UPD;
8930 NumVecs = 3; isLoad = false; isLaneOp = true; break;
8931 case Intrinsic::arm_neon_vst4lane: NewOpc = ARMISD::VST4LN_UPD;
8932 NumVecs = 4; isLoad = false; isLaneOp = true; break;
8936 switch (N->getOpcode()) {
8937 default: llvm_unreachable("unexpected opcode for Neon base update");
8938 case ARMISD::VLD2DUP: NewOpc = ARMISD::VLD2DUP_UPD; NumVecs = 2; break;
8939 case ARMISD::VLD3DUP: NewOpc = ARMISD::VLD3DUP_UPD; NumVecs = 3; break;
8940 case ARMISD::VLD4DUP: NewOpc = ARMISD::VLD4DUP_UPD; NumVecs = 4; break;
8944 // Find the size of memory referenced by the load/store.
8947 VecTy = N->getValueType(0);
8949 VecTy = N->getOperand(AddrOpIdx+1).getValueType();
8950 unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8;
8952 NumBytes /= VecTy.getVectorNumElements();
8954 // If the increment is a constant, it must match the memory ref size.
8955 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
8956 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
8957 uint64_t IncVal = CInc->getZExtValue();
8958 if (IncVal != NumBytes)
8960 } else if (NumBytes >= 3 * 16) {
8961 // VLD3/4 and VST3/4 for 128-bit vectors are implemented with two
8962 // separate instructions that make it harder to use a non-constant update.
8966 // Create the new updating load/store node.
8968 unsigned NumResultVecs = (isLoad ? NumVecs : 0);
8970 for (n = 0; n < NumResultVecs; ++n)
8972 Tys[n++] = MVT::i32;
8973 Tys[n] = MVT::Other;
8974 SDVTList SDTys = DAG.getVTList(makeArrayRef(Tys, NumResultVecs+2));
8975 SmallVector<SDValue, 8> Ops;
8976 Ops.push_back(N->getOperand(0)); // incoming chain
8977 Ops.push_back(N->getOperand(AddrOpIdx));
8979 for (unsigned i = AddrOpIdx + 1; i < N->getNumOperands(); ++i) {
8980 Ops.push_back(N->getOperand(i));
8982 MemIntrinsicSDNode *MemInt = cast<MemIntrinsicSDNode>(N);
8983 SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, SDLoc(N), SDTys,
8984 Ops, MemInt->getMemoryVT(),
8985 MemInt->getMemOperand());
8988 std::vector<SDValue> NewResults;
8989 for (unsigned i = 0; i < NumResultVecs; ++i) {
8990 NewResults.push_back(SDValue(UpdN.getNode(), i));
8992 NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs+1)); // chain
8993 DCI.CombineTo(N, NewResults);
8994 DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs));
9001 /// CombineVLDDUP - For a VDUPLANE node N, check if its source operand is a
9002 /// vldN-lane (N > 1) intrinsic, and if all the other uses of that intrinsic
9003 /// are also VDUPLANEs. If so, combine them to a vldN-dup operation and
9005 static bool CombineVLDDUP(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
9006 SelectionDAG &DAG = DCI.DAG;
9007 EVT VT = N->getValueType(0);
9008 // vldN-dup instructions only support 64-bit vectors for N > 1.
9009 if (!VT.is64BitVector())
9012 // Check if the VDUPLANE operand is a vldN-dup intrinsic.
9013 SDNode *VLD = N->getOperand(0).getNode();
9014 if (VLD->getOpcode() != ISD::INTRINSIC_W_CHAIN)
9016 unsigned NumVecs = 0;
9017 unsigned NewOpc = 0;
9018 unsigned IntNo = cast<ConstantSDNode>(VLD->getOperand(1))->getZExtValue();
9019 if (IntNo == Intrinsic::arm_neon_vld2lane) {
9021 NewOpc = ARMISD::VLD2DUP;
9022 } else if (IntNo == Intrinsic::arm_neon_vld3lane) {
9024 NewOpc = ARMISD::VLD3DUP;
9025 } else if (IntNo == Intrinsic::arm_neon_vld4lane) {
9027 NewOpc = ARMISD::VLD4DUP;
9032 // First check that all the vldN-lane uses are VDUPLANEs and that the lane
9033 // numbers match the load.
9034 unsigned VLDLaneNo =
9035 cast<ConstantSDNode>(VLD->getOperand(NumVecs+3))->getZExtValue();
9036 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
9038 // Ignore uses of the chain result.
9039 if (UI.getUse().getResNo() == NumVecs)
9042 if (User->getOpcode() != ARMISD::VDUPLANE ||
9043 VLDLaneNo != cast<ConstantSDNode>(User->getOperand(1))->getZExtValue())
9047 // Create the vldN-dup node.
9050 for (n = 0; n < NumVecs; ++n)
9052 Tys[n] = MVT::Other;
9053 SDVTList SDTys = DAG.getVTList(makeArrayRef(Tys, NumVecs+1));
9054 SDValue Ops[] = { VLD->getOperand(0), VLD->getOperand(2) };
9055 MemIntrinsicSDNode *VLDMemInt = cast<MemIntrinsicSDNode>(VLD);
9056 SDValue VLDDup = DAG.getMemIntrinsicNode(NewOpc, SDLoc(VLD), SDTys,
9057 Ops, VLDMemInt->getMemoryVT(),
9058 VLDMemInt->getMemOperand());
9061 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
9063 unsigned ResNo = UI.getUse().getResNo();
9064 // Ignore uses of the chain result.
9065 if (ResNo == NumVecs)
9068 DCI.CombineTo(User, SDValue(VLDDup.getNode(), ResNo));
9071 // Now the vldN-lane intrinsic is dead except for its chain result.
9072 // Update uses of the chain.
9073 std::vector<SDValue> VLDDupResults;
9074 for (unsigned n = 0; n < NumVecs; ++n)
9075 VLDDupResults.push_back(SDValue(VLDDup.getNode(), n));
9076 VLDDupResults.push_back(SDValue(VLDDup.getNode(), NumVecs));
9077 DCI.CombineTo(VLD, VLDDupResults);
9082 /// PerformVDUPLANECombine - Target-specific dag combine xforms for
9083 /// ARMISD::VDUPLANE.
9084 static SDValue PerformVDUPLANECombine(SDNode *N,
9085 TargetLowering::DAGCombinerInfo &DCI) {
9086 SDValue Op = N->getOperand(0);
9088 // If the source is a vldN-lane (N > 1) intrinsic, and all the other uses
9089 // of that intrinsic are also VDUPLANEs, combine them to a vldN-dup operation.
9090 if (CombineVLDDUP(N, DCI))
9091 return SDValue(N, 0);
9093 // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is
9094 // redundant. Ignore bit_converts for now; element sizes are checked below.
9095 while (Op.getOpcode() == ISD::BITCAST)
9096 Op = Op.getOperand(0);
9097 if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM)
9100 // Make sure the VMOV element size is not bigger than the VDUPLANE elements.
9101 unsigned EltSize = Op.getValueType().getVectorElementType().getSizeInBits();
9102 // The canonical VMOV for a zero vector uses a 32-bit element size.
9103 unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
9105 if (ARM_AM::decodeNEONModImm(Imm, EltBits) == 0)
9107 EVT VT = N->getValueType(0);
9108 if (EltSize > VT.getVectorElementType().getSizeInBits())
9111 return DCI.DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Op);
9114 /// PerformSTORECombine - Target-specific dag combine xforms for
9116 static SDValue PerformSTORECombine(SDNode *N,
9117 TargetLowering::DAGCombinerInfo &DCI) {
9118 StoreSDNode *St = cast<StoreSDNode>(N);
9119 if (St->isVolatile())
9122 // Optimize trunc store (of multiple scalars) to shuffle and store. First,
9123 // pack all of the elements in one place. Next, store to memory in fewer
9125 SDValue StVal = St->getValue();
9126 EVT VT = StVal.getValueType();
9127 if (St->isTruncatingStore() && VT.isVector()) {
9128 SelectionDAG &DAG = DCI.DAG;
9129 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9130 EVT StVT = St->getMemoryVT();
9131 unsigned NumElems = VT.getVectorNumElements();
9132 assert(StVT != VT && "Cannot truncate to the same type");
9133 unsigned FromEltSz = VT.getVectorElementType().getSizeInBits();
9134 unsigned ToEltSz = StVT.getVectorElementType().getSizeInBits();
9136 // From, To sizes and ElemCount must be pow of two
9137 if (!isPowerOf2_32(NumElems * FromEltSz * ToEltSz)) return SDValue();
9139 // We are going to use the original vector elt for storing.
9140 // Accumulated smaller vector elements must be a multiple of the store size.
9141 if (0 != (NumElems * FromEltSz) % ToEltSz) return SDValue();
9143 unsigned SizeRatio = FromEltSz / ToEltSz;
9144 assert(SizeRatio * NumElems * ToEltSz == VT.getSizeInBits());
9146 // Create a type on which we perform the shuffle.
9147 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(), StVT.getScalarType(),
9148 NumElems*SizeRatio);
9149 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
9152 SDValue WideVec = DAG.getNode(ISD::BITCAST, DL, WideVecVT, StVal);
9153 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
9154 for (unsigned i = 0; i < NumElems; ++i)
9155 ShuffleVec[i] = TLI.isBigEndian() ? (i+1) * SizeRatio - 1 : i * SizeRatio;
9157 // Can't shuffle using an illegal type.
9158 if (!TLI.isTypeLegal(WideVecVT)) return SDValue();
9160 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, DL, WideVec,
9161 DAG.getUNDEF(WideVec.getValueType()),
9163 // At this point all of the data is stored at the bottom of the
9164 // register. We now need to save it to mem.
9166 // Find the largest store unit
9167 MVT StoreType = MVT::i8;
9168 for (MVT Tp : MVT::integer_valuetypes()) {
9169 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToEltSz)
9172 // Didn't find a legal store type.
9173 if (!TLI.isTypeLegal(StoreType))
9176 // Bitcast the original vector into a vector of store-size units
9177 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
9178 StoreType, VT.getSizeInBits()/EVT(StoreType).getSizeInBits());
9179 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
9180 SDValue ShuffWide = DAG.getNode(ISD::BITCAST, DL, StoreVecVT, Shuff);
9181 SmallVector<SDValue, 8> Chains;
9182 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits()/8,
9183 TLI.getPointerTy());
9184 SDValue BasePtr = St->getBasePtr();
9186 // Perform one or more big stores into memory.
9187 unsigned E = (ToEltSz*NumElems)/StoreType.getSizeInBits();
9188 for (unsigned I = 0; I < E; I++) {
9189 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
9190 StoreType, ShuffWide,
9191 DAG.getIntPtrConstant(I));
9192 SDValue Ch = DAG.getStore(St->getChain(), DL, SubVec, BasePtr,
9193 St->getPointerInfo(), St->isVolatile(),
9194 St->isNonTemporal(), St->getAlignment());
9195 BasePtr = DAG.getNode(ISD::ADD, DL, BasePtr.getValueType(), BasePtr,
9197 Chains.push_back(Ch);
9199 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
9202 if (!ISD::isNormalStore(St))
9205 // Split a store of a VMOVDRR into two integer stores to avoid mixing NEON and
9206 // ARM stores of arguments in the same cache line.
9207 if (StVal.getNode()->getOpcode() == ARMISD::VMOVDRR &&
9208 StVal.getNode()->hasOneUse()) {
9209 SelectionDAG &DAG = DCI.DAG;
9210 bool isBigEndian = DAG.getTargetLoweringInfo().isBigEndian();
9212 SDValue BasePtr = St->getBasePtr();
9213 SDValue NewST1 = DAG.getStore(St->getChain(), DL,
9214 StVal.getNode()->getOperand(isBigEndian ? 1 : 0 ),
9215 BasePtr, St->getPointerInfo(), St->isVolatile(),
9216 St->isNonTemporal(), St->getAlignment());
9218 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
9219 DAG.getConstant(4, MVT::i32));
9220 return DAG.getStore(NewST1.getValue(0), DL,
9221 StVal.getNode()->getOperand(isBigEndian ? 0 : 1),
9222 OffsetPtr, St->getPointerInfo(), St->isVolatile(),
9223 St->isNonTemporal(),
9224 std::min(4U, St->getAlignment() / 2));
9227 if (StVal.getValueType() == MVT::i64 &&
9228 StVal.getNode()->getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
9230 // Bitcast an i64 store extracted from a vector to f64.
9231 // Otherwise, the i64 value will be legalized to a pair of i32 values.
9232 SelectionDAG &DAG = DCI.DAG;
9234 SDValue IntVec = StVal.getOperand(0);
9235 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
9236 IntVec.getValueType().getVectorNumElements());
9237 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, IntVec);
9238 SDValue ExtElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
9239 Vec, StVal.getOperand(1));
9241 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ExtElt);
9242 // Make the DAGCombiner fold the bitcasts.
9243 DCI.AddToWorklist(Vec.getNode());
9244 DCI.AddToWorklist(ExtElt.getNode());
9245 DCI.AddToWorklist(V.getNode());
9246 return DAG.getStore(St->getChain(), dl, V, St->getBasePtr(),
9247 St->getPointerInfo(), St->isVolatile(),
9248 St->isNonTemporal(), St->getAlignment(),
9255 // isConstVecPow2 - Return true if each vector element is a power of 2, all
9256 // elements are the same constant, C, and Log2(C) ranges from 1 to 32.
9257 static bool isConstVecPow2(SDValue ConstVec, bool isSigned, uint64_t &C)
9261 for (unsigned I = 0, E = ConstVec.getValueType().getVectorNumElements();
9263 ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(ConstVec.getOperand(I));
9268 APFloat APF = C->getValueAPF();
9269 if (APF.convertToInteger(&cN, 64, isSigned, APFloat::rmTowardZero, &isExact)
9270 != APFloat::opOK || !isExact)
9273 c0 = (I == 0) ? cN : c0;
9274 if (!isPowerOf2_64(cN) || c0 != cN || Log2_64(c0) < 1 || Log2_64(c0) > 32)
9281 /// PerformVCVTCombine - VCVT (floating-point to fixed-point, Advanced SIMD)
9282 /// can replace combinations of VMUL and VCVT (floating-point to integer)
9283 /// when the VMUL has a constant operand that is a power of 2.
9285 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
9286 /// vmul.f32 d16, d17, d16
9287 /// vcvt.s32.f32 d16, d16
9289 /// vcvt.s32.f32 d16, d16, #3
9290 static SDValue PerformVCVTCombine(SDNode *N,
9291 TargetLowering::DAGCombinerInfo &DCI,
9292 const ARMSubtarget *Subtarget) {
9293 SelectionDAG &DAG = DCI.DAG;
9294 SDValue Op = N->getOperand(0);
9296 if (!Subtarget->hasNEON() || !Op.getValueType().isVector() ||
9297 Op.getOpcode() != ISD::FMUL)
9301 SDValue N0 = Op->getOperand(0);
9302 SDValue ConstVec = Op->getOperand(1);
9303 bool isSigned = N->getOpcode() == ISD::FP_TO_SINT;
9305 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
9306 !isConstVecPow2(ConstVec, isSigned, C))
9309 MVT FloatTy = Op.getSimpleValueType().getVectorElementType();
9310 MVT IntTy = N->getSimpleValueType(0).getVectorElementType();
9311 unsigned NumLanes = Op.getValueType().getVectorNumElements();
9312 if (FloatTy.getSizeInBits() != 32 || IntTy.getSizeInBits() > 32 ||
9314 // These instructions only exist converting from f32 to i32. We can handle
9315 // smaller integers by generating an extra truncate, but larger ones would
9316 // be lossy. We also can't handle more then 4 lanes, since these intructions
9317 // only support v2i32/v4i32 types.
9321 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfp2fxs :
9322 Intrinsic::arm_neon_vcvtfp2fxu;
9323 SDValue FixConv = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N),
9324 NumLanes == 2 ? MVT::v2i32 : MVT::v4i32,
9325 DAG.getConstant(IntrinsicOpcode, MVT::i32), N0,
9326 DAG.getConstant(Log2_64(C), MVT::i32));
9328 if (IntTy.getSizeInBits() < FloatTy.getSizeInBits())
9329 FixConv = DAG.getNode(ISD::TRUNCATE, SDLoc(N), N->getValueType(0), FixConv);
9334 /// PerformVDIVCombine - VCVT (fixed-point to floating-point, Advanced SIMD)
9335 /// can replace combinations of VCVT (integer to floating-point) and VDIV
9336 /// when the VDIV has a constant operand that is a power of 2.
9338 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
9339 /// vcvt.f32.s32 d16, d16
9340 /// vdiv.f32 d16, d17, d16
9342 /// vcvt.f32.s32 d16, d16, #3
9343 static SDValue PerformVDIVCombine(SDNode *N,
9344 TargetLowering::DAGCombinerInfo &DCI,
9345 const ARMSubtarget *Subtarget) {
9346 SelectionDAG &DAG = DCI.DAG;
9347 SDValue Op = N->getOperand(0);
9348 unsigned OpOpcode = Op.getNode()->getOpcode();
9350 if (!Subtarget->hasNEON() || !N->getValueType(0).isVector() ||
9351 (OpOpcode != ISD::SINT_TO_FP && OpOpcode != ISD::UINT_TO_FP))
9355 SDValue ConstVec = N->getOperand(1);
9356 bool isSigned = OpOpcode == ISD::SINT_TO_FP;
9358 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
9359 !isConstVecPow2(ConstVec, isSigned, C))
9362 MVT FloatTy = N->getSimpleValueType(0).getVectorElementType();
9363 MVT IntTy = Op.getOperand(0).getSimpleValueType().getVectorElementType();
9364 if (FloatTy.getSizeInBits() != 32 || IntTy.getSizeInBits() > 32) {
9365 // These instructions only exist converting from i32 to f32. We can handle
9366 // smaller integers by generating an extra extend, but larger ones would
9371 SDValue ConvInput = Op.getOperand(0);
9372 unsigned NumLanes = Op.getValueType().getVectorNumElements();
9373 if (IntTy.getSizeInBits() < FloatTy.getSizeInBits())
9374 ConvInput = DAG.getNode(isSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND,
9375 SDLoc(N), NumLanes == 2 ? MVT::v2i32 : MVT::v4i32,
9378 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfxs2fp :
9379 Intrinsic::arm_neon_vcvtfxu2fp;
9380 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N),
9382 DAG.getConstant(IntrinsicOpcode, MVT::i32),
9383 ConvInput, DAG.getConstant(Log2_64(C), MVT::i32));
9386 /// Getvshiftimm - Check if this is a valid build_vector for the immediate
9387 /// operand of a vector shift operation, where all the elements of the
9388 /// build_vector must have the same constant integer value.
9389 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
9390 // Ignore bit_converts.
9391 while (Op.getOpcode() == ISD::BITCAST)
9392 Op = Op.getOperand(0);
9393 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
9394 APInt SplatBits, SplatUndef;
9395 unsigned SplatBitSize;
9397 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
9398 HasAnyUndefs, ElementBits) ||
9399 SplatBitSize > ElementBits)
9401 Cnt = SplatBits.getSExtValue();
9405 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
9406 /// operand of a vector shift left operation. That value must be in the range:
9407 /// 0 <= Value < ElementBits for a left shift; or
9408 /// 0 <= Value <= ElementBits for a long left shift.
9409 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
9410 assert(VT.isVector() && "vector shift count is not a vector type");
9411 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
9412 if (! getVShiftImm(Op, ElementBits, Cnt))
9414 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
9417 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
9418 /// operand of a vector shift right operation. For a shift opcode, the value
9419 /// is positive, but for an intrinsic the value count must be negative. The
9420 /// absolute value must be in the range:
9421 /// 1 <= |Value| <= ElementBits for a right shift; or
9422 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
9423 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
9425 assert(VT.isVector() && "vector shift count is not a vector type");
9426 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
9427 if (! getVShiftImm(Op, ElementBits, Cnt))
9431 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
9434 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
9435 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
9436 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
9439 // Don't do anything for most intrinsics.
9442 // Vector shifts: check for immediate versions and lower them.
9443 // Note: This is done during DAG combining instead of DAG legalizing because
9444 // the build_vectors for 64-bit vector element shift counts are generally
9445 // not legal, and it is hard to see their values after they get legalized to
9446 // loads from a constant pool.
9447 case Intrinsic::arm_neon_vshifts:
9448 case Intrinsic::arm_neon_vshiftu:
9449 case Intrinsic::arm_neon_vrshifts:
9450 case Intrinsic::arm_neon_vrshiftu:
9451 case Intrinsic::arm_neon_vrshiftn:
9452 case Intrinsic::arm_neon_vqshifts:
9453 case Intrinsic::arm_neon_vqshiftu:
9454 case Intrinsic::arm_neon_vqshiftsu:
9455 case Intrinsic::arm_neon_vqshiftns:
9456 case Intrinsic::arm_neon_vqshiftnu:
9457 case Intrinsic::arm_neon_vqshiftnsu:
9458 case Intrinsic::arm_neon_vqrshiftns:
9459 case Intrinsic::arm_neon_vqrshiftnu:
9460 case Intrinsic::arm_neon_vqrshiftnsu: {
9461 EVT VT = N->getOperand(1).getValueType();
9463 unsigned VShiftOpc = 0;
9466 case Intrinsic::arm_neon_vshifts:
9467 case Intrinsic::arm_neon_vshiftu:
9468 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
9469 VShiftOpc = ARMISD::VSHL;
9472 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
9473 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
9474 ARMISD::VSHRs : ARMISD::VSHRu);
9479 case Intrinsic::arm_neon_vrshifts:
9480 case Intrinsic::arm_neon_vrshiftu:
9481 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
9485 case Intrinsic::arm_neon_vqshifts:
9486 case Intrinsic::arm_neon_vqshiftu:
9487 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
9491 case Intrinsic::arm_neon_vqshiftsu:
9492 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
9494 llvm_unreachable("invalid shift count for vqshlu intrinsic");
9496 case Intrinsic::arm_neon_vrshiftn:
9497 case Intrinsic::arm_neon_vqshiftns:
9498 case Intrinsic::arm_neon_vqshiftnu:
9499 case Intrinsic::arm_neon_vqshiftnsu:
9500 case Intrinsic::arm_neon_vqrshiftns:
9501 case Intrinsic::arm_neon_vqrshiftnu:
9502 case Intrinsic::arm_neon_vqrshiftnsu:
9503 // Narrowing shifts require an immediate right shift.
9504 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
9506 llvm_unreachable("invalid shift count for narrowing vector shift "
9510 llvm_unreachable("unhandled vector shift");
9514 case Intrinsic::arm_neon_vshifts:
9515 case Intrinsic::arm_neon_vshiftu:
9516 // Opcode already set above.
9518 case Intrinsic::arm_neon_vrshifts:
9519 VShiftOpc = ARMISD::VRSHRs; break;
9520 case Intrinsic::arm_neon_vrshiftu:
9521 VShiftOpc = ARMISD::VRSHRu; break;
9522 case Intrinsic::arm_neon_vrshiftn:
9523 VShiftOpc = ARMISD::VRSHRN; break;
9524 case Intrinsic::arm_neon_vqshifts:
9525 VShiftOpc = ARMISD::VQSHLs; break;
9526 case Intrinsic::arm_neon_vqshiftu:
9527 VShiftOpc = ARMISD::VQSHLu; break;
9528 case Intrinsic::arm_neon_vqshiftsu:
9529 VShiftOpc = ARMISD::VQSHLsu; break;
9530 case Intrinsic::arm_neon_vqshiftns:
9531 VShiftOpc = ARMISD::VQSHRNs; break;
9532 case Intrinsic::arm_neon_vqshiftnu:
9533 VShiftOpc = ARMISD::VQSHRNu; break;
9534 case Intrinsic::arm_neon_vqshiftnsu:
9535 VShiftOpc = ARMISD::VQSHRNsu; break;
9536 case Intrinsic::arm_neon_vqrshiftns:
9537 VShiftOpc = ARMISD::VQRSHRNs; break;
9538 case Intrinsic::arm_neon_vqrshiftnu:
9539 VShiftOpc = ARMISD::VQRSHRNu; break;
9540 case Intrinsic::arm_neon_vqrshiftnsu:
9541 VShiftOpc = ARMISD::VQRSHRNsu; break;
9544 return DAG.getNode(VShiftOpc, SDLoc(N), N->getValueType(0),
9545 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
9548 case Intrinsic::arm_neon_vshiftins: {
9549 EVT VT = N->getOperand(1).getValueType();
9551 unsigned VShiftOpc = 0;
9553 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
9554 VShiftOpc = ARMISD::VSLI;
9555 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
9556 VShiftOpc = ARMISD::VSRI;
9558 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
9561 return DAG.getNode(VShiftOpc, SDLoc(N), N->getValueType(0),
9562 N->getOperand(1), N->getOperand(2),
9563 DAG.getConstant(Cnt, MVT::i32));
9566 case Intrinsic::arm_neon_vqrshifts:
9567 case Intrinsic::arm_neon_vqrshiftu:
9568 // No immediate versions of these to check for.
9575 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
9576 /// lowers them. As with the vector shift intrinsics, this is done during DAG
9577 /// combining instead of DAG legalizing because the build_vectors for 64-bit
9578 /// vector element shift counts are generally not legal, and it is hard to see
9579 /// their values after they get legalized to loads from a constant pool.
9580 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
9581 const ARMSubtarget *ST) {
9582 EVT VT = N->getValueType(0);
9583 if (N->getOpcode() == ISD::SRL && VT == MVT::i32 && ST->hasV6Ops()) {
9584 // Canonicalize (srl (bswap x), 16) to (rotr (bswap x), 16) if the high
9585 // 16-bits of x is zero. This optimizes rev + lsr 16 to rev16.
9586 SDValue N1 = N->getOperand(1);
9587 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) {
9588 SDValue N0 = N->getOperand(0);
9589 if (C->getZExtValue() == 16 && N0.getOpcode() == ISD::BSWAP &&
9590 DAG.MaskedValueIsZero(N0.getOperand(0),
9591 APInt::getHighBitsSet(32, 16)))
9592 return DAG.getNode(ISD::ROTR, SDLoc(N), VT, N0, N1);
9596 // Nothing to be done for scalar shifts.
9597 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9598 if (!VT.isVector() || !TLI.isTypeLegal(VT))
9601 assert(ST->hasNEON() && "unexpected vector shift");
9604 switch (N->getOpcode()) {
9605 default: llvm_unreachable("unexpected shift opcode");
9608 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
9609 return DAG.getNode(ARMISD::VSHL, SDLoc(N), VT, N->getOperand(0),
9610 DAG.getConstant(Cnt, MVT::i32));
9615 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
9616 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
9617 ARMISD::VSHRs : ARMISD::VSHRu);
9618 return DAG.getNode(VShiftOpc, SDLoc(N), VT, N->getOperand(0),
9619 DAG.getConstant(Cnt, MVT::i32));
9625 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
9626 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
9627 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
9628 const ARMSubtarget *ST) {
9629 SDValue N0 = N->getOperand(0);
9631 // Check for sign- and zero-extensions of vector extract operations of 8-
9632 // and 16-bit vector elements. NEON supports these directly. They are
9633 // handled during DAG combining because type legalization will promote them
9634 // to 32-bit types and it is messy to recognize the operations after that.
9635 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
9636 SDValue Vec = N0.getOperand(0);
9637 SDValue Lane = N0.getOperand(1);
9638 EVT VT = N->getValueType(0);
9639 EVT EltVT = N0.getValueType();
9640 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9642 if (VT == MVT::i32 &&
9643 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
9644 TLI.isTypeLegal(Vec.getValueType()) &&
9645 isa<ConstantSDNode>(Lane)) {
9648 switch (N->getOpcode()) {
9649 default: llvm_unreachable("unexpected opcode");
9650 case ISD::SIGN_EXTEND:
9651 Opc = ARMISD::VGETLANEs;
9653 case ISD::ZERO_EXTEND:
9654 case ISD::ANY_EXTEND:
9655 Opc = ARMISD::VGETLANEu;
9658 return DAG.getNode(Opc, SDLoc(N), VT, Vec, Lane);
9665 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
9666 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
9667 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
9668 const ARMSubtarget *ST) {
9669 // If the target supports NEON, try to use vmax/vmin instructions for f32
9670 // selects like "x < y ? x : y". Unless the NoNaNsFPMath option is set,
9671 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
9672 // a NaN; only do the transformation when it matches that behavior.
9674 // For now only do this when using NEON for FP operations; if using VFP, it
9675 // is not obvious that the benefit outweighs the cost of switching to the
9677 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
9678 N->getValueType(0) != MVT::f32)
9681 SDValue CondLHS = N->getOperand(0);
9682 SDValue CondRHS = N->getOperand(1);
9683 SDValue LHS = N->getOperand(2);
9684 SDValue RHS = N->getOperand(3);
9685 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
9687 unsigned Opcode = 0;
9689 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
9690 IsReversed = false; // x CC y ? x : y
9691 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
9692 IsReversed = true ; // x CC y ? y : x
9706 // If LHS is NaN, an ordered comparison will be false and the result will
9707 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
9708 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
9709 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
9710 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
9712 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
9713 // will return -0, so vmin can only be used for unsafe math or if one of
9714 // the operands is known to be nonzero.
9715 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
9716 !DAG.getTarget().Options.UnsafeFPMath &&
9717 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9719 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
9728 // If LHS is NaN, an ordered comparison will be false and the result will
9729 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
9730 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
9731 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
9732 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
9734 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
9735 // will return +0, so vmax can only be used for unsafe math or if one of
9736 // the operands is known to be nonzero.
9737 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
9738 !DAG.getTarget().Options.UnsafeFPMath &&
9739 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9741 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
9747 return DAG.getNode(Opcode, SDLoc(N), N->getValueType(0), LHS, RHS);
9750 /// PerformCMOVCombine - Target-specific DAG combining for ARMISD::CMOV.
9752 ARMTargetLowering::PerformCMOVCombine(SDNode *N, SelectionDAG &DAG) const {
9753 SDValue Cmp = N->getOperand(4);
9754 if (Cmp.getOpcode() != ARMISD::CMPZ)
9755 // Only looking at EQ and NE cases.
9758 EVT VT = N->getValueType(0);
9760 SDValue LHS = Cmp.getOperand(0);
9761 SDValue RHS = Cmp.getOperand(1);
9762 SDValue FalseVal = N->getOperand(0);
9763 SDValue TrueVal = N->getOperand(1);
9764 SDValue ARMcc = N->getOperand(2);
9765 ARMCC::CondCodes CC =
9766 (ARMCC::CondCodes)cast<ConstantSDNode>(ARMcc)->getZExtValue();
9784 /// FIXME: Turn this into a target neutral optimization?
9786 if (CC == ARMCC::NE && FalseVal == RHS && FalseVal != LHS) {
9787 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, TrueVal, ARMcc,
9788 N->getOperand(3), Cmp);
9789 } else if (CC == ARMCC::EQ && TrueVal == RHS) {
9791 SDValue NewCmp = getARMCmp(LHS, RHS, ISD::SETNE, ARMcc, DAG, dl);
9792 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, FalseVal, ARMcc,
9793 N->getOperand(3), NewCmp);
9796 if (Res.getNode()) {
9797 APInt KnownZero, KnownOne;
9798 DAG.computeKnownBits(SDValue(N,0), KnownZero, KnownOne);
9799 // Capture demanded bits information that would be otherwise lost.
9800 if (KnownZero == 0xfffffffe)
9801 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
9802 DAG.getValueType(MVT::i1));
9803 else if (KnownZero == 0xffffff00)
9804 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
9805 DAG.getValueType(MVT::i8));
9806 else if (KnownZero == 0xffff0000)
9807 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
9808 DAG.getValueType(MVT::i16));
9814 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
9815 DAGCombinerInfo &DCI) const {
9816 switch (N->getOpcode()) {
9818 case ISD::ADDC: return PerformADDCCombine(N, DCI, Subtarget);
9819 case ISD::ADD: return PerformADDCombine(N, DCI, Subtarget);
9820 case ISD::SUB: return PerformSUBCombine(N, DCI);
9821 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
9822 case ISD::OR: return PerformORCombine(N, DCI, Subtarget);
9823 case ISD::XOR: return PerformXORCombine(N, DCI, Subtarget);
9824 case ISD::AND: return PerformANDCombine(N, DCI, Subtarget);
9825 case ARMISD::BFI: return PerformBFICombine(N, DCI);
9826 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI, Subtarget);
9827 case ARMISD::VMOVDRR: return PerformVMOVDRRCombine(N, DCI.DAG);
9828 case ISD::STORE: return PerformSTORECombine(N, DCI);
9829 case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DCI, Subtarget);
9830 case ISD::INSERT_VECTOR_ELT: return PerformInsertEltCombine(N, DCI);
9831 case ISD::VECTOR_SHUFFLE: return PerformVECTOR_SHUFFLECombine(N, DCI.DAG);
9832 case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI);
9833 case ISD::FP_TO_SINT:
9834 case ISD::FP_TO_UINT: return PerformVCVTCombine(N, DCI, Subtarget);
9835 case ISD::FDIV: return PerformVDIVCombine(N, DCI, Subtarget);
9836 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
9839 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
9840 case ISD::SIGN_EXTEND:
9841 case ISD::ZERO_EXTEND:
9842 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
9843 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
9844 case ARMISD::CMOV: return PerformCMOVCombine(N, DCI.DAG);
9845 case ARMISD::VLD2DUP:
9846 case ARMISD::VLD3DUP:
9847 case ARMISD::VLD4DUP:
9848 return CombineBaseUpdate(N, DCI);
9849 case ARMISD::BUILD_VECTOR:
9850 return PerformARMBUILD_VECTORCombine(N, DCI);
9851 case ISD::INTRINSIC_VOID:
9852 case ISD::INTRINSIC_W_CHAIN:
9853 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
9854 case Intrinsic::arm_neon_vld1:
9855 case Intrinsic::arm_neon_vld2:
9856 case Intrinsic::arm_neon_vld3:
9857 case Intrinsic::arm_neon_vld4:
9858 case Intrinsic::arm_neon_vld2lane:
9859 case Intrinsic::arm_neon_vld3lane:
9860 case Intrinsic::arm_neon_vld4lane:
9861 case Intrinsic::arm_neon_vst1:
9862 case Intrinsic::arm_neon_vst2:
9863 case Intrinsic::arm_neon_vst3:
9864 case Intrinsic::arm_neon_vst4:
9865 case Intrinsic::arm_neon_vst2lane:
9866 case Intrinsic::arm_neon_vst3lane:
9867 case Intrinsic::arm_neon_vst4lane:
9868 return CombineBaseUpdate(N, DCI);
9876 bool ARMTargetLowering::isDesirableToTransformToIntegerOp(unsigned Opc,
9878 return (VT == MVT::f32) && (Opc == ISD::LOAD || Opc == ISD::STORE);
9881 bool ARMTargetLowering::allowsMisalignedMemoryAccesses(EVT VT,
9885 // The AllowsUnaliged flag models the SCTLR.A setting in ARM cpus
9886 bool AllowsUnaligned = Subtarget->allowsUnalignedMem();
9888 switch (VT.getSimpleVT().SimpleTy) {
9894 // Unaligned access can use (for example) LRDB, LRDH, LDR
9895 if (AllowsUnaligned) {
9897 *Fast = Subtarget->hasV7Ops();
9904 // For any little-endian targets with neon, we can support unaligned ld/st
9905 // of D and Q (e.g. {D0,D1}) registers by using vld1.i8/vst1.i8.
9906 // A big-endian target may also explicitly support unaligned accesses
9907 if (Subtarget->hasNEON() && (AllowsUnaligned || isLittleEndian())) {
9917 static bool memOpAlign(unsigned DstAlign, unsigned SrcAlign,
9918 unsigned AlignCheck) {
9919 return ((SrcAlign == 0 || SrcAlign % AlignCheck == 0) &&
9920 (DstAlign == 0 || DstAlign % AlignCheck == 0));
9923 EVT ARMTargetLowering::getOptimalMemOpType(uint64_t Size,
9924 unsigned DstAlign, unsigned SrcAlign,
9925 bool IsMemset, bool ZeroMemset,
9927 MachineFunction &MF) const {
9928 const Function *F = MF.getFunction();
9930 // See if we can use NEON instructions for this...
9931 if ((!IsMemset || ZeroMemset) &&
9932 Subtarget->hasNEON() &&
9933 !F->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
9934 Attribute::NoImplicitFloat)) {
9937 (memOpAlign(SrcAlign, DstAlign, 16) ||
9938 (allowsMisalignedMemoryAccesses(MVT::v2f64, 0, 1, &Fast) && Fast))) {
9940 } else if (Size >= 8 &&
9941 (memOpAlign(SrcAlign, DstAlign, 8) ||
9942 (allowsMisalignedMemoryAccesses(MVT::f64, 0, 1, &Fast) &&
9948 // Lowering to i32/i16 if the size permits.
9954 // Let the target-independent logic figure it out.
9958 bool ARMTargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
9959 if (Val.getOpcode() != ISD::LOAD)
9962 EVT VT1 = Val.getValueType();
9963 if (!VT1.isSimple() || !VT1.isInteger() ||
9964 !VT2.isSimple() || !VT2.isInteger())
9967 switch (VT1.getSimpleVT().SimpleTy) {
9972 // 8-bit and 16-bit loads implicitly zero-extend to 32-bits.
9979 bool ARMTargetLowering::allowTruncateForTailCall(Type *Ty1, Type *Ty2) const {
9980 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
9983 if (!isTypeLegal(EVT::getEVT(Ty1)))
9986 assert(Ty1->getPrimitiveSizeInBits() <= 64 && "i128 is probably not a noop");
9988 // Assuming the caller doesn't have a zeroext or signext return parameter,
9989 // truncation all the way down to i1 is valid.
9994 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
9999 switch (VT.getSimpleVT().SimpleTy) {
10000 default: return false;
10015 if ((V & (Scale - 1)) != 0)
10018 return V == (V & ((1LL << 5) - 1));
10021 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
10022 const ARMSubtarget *Subtarget) {
10023 bool isNeg = false;
10029 switch (VT.getSimpleVT().SimpleTy) {
10030 default: return false;
10035 // + imm12 or - imm8
10037 return V == (V & ((1LL << 8) - 1));
10038 return V == (V & ((1LL << 12) - 1));
10041 // Same as ARM mode. FIXME: NEON?
10042 if (!Subtarget->hasVFP2())
10047 return V == (V & ((1LL << 8) - 1));
10051 /// isLegalAddressImmediate - Return true if the integer value can be used
10052 /// as the offset of the target addressing mode for load / store of the
10054 static bool isLegalAddressImmediate(int64_t V, EVT VT,
10055 const ARMSubtarget *Subtarget) {
10059 if (!VT.isSimple())
10062 if (Subtarget->isThumb1Only())
10063 return isLegalT1AddressImmediate(V, VT);
10064 else if (Subtarget->isThumb2())
10065 return isLegalT2AddressImmediate(V, VT, Subtarget);
10070 switch (VT.getSimpleVT().SimpleTy) {
10071 default: return false;
10076 return V == (V & ((1LL << 12) - 1));
10079 return V == (V & ((1LL << 8) - 1));
10082 if (!Subtarget->hasVFP2()) // FIXME: NEON?
10087 return V == (V & ((1LL << 8) - 1));
10091 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
10093 int Scale = AM.Scale;
10097 switch (VT.getSimpleVT().SimpleTy) {
10098 default: return false;
10106 Scale = Scale & ~1;
10107 return Scale == 2 || Scale == 4 || Scale == 8;
10110 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
10114 // Note, we allow "void" uses (basically, uses that aren't loads or
10115 // stores), because arm allows folding a scale into many arithmetic
10116 // operations. This should be made more precise and revisited later.
10118 // Allow r << imm, but the imm has to be a multiple of two.
10119 if (Scale & 1) return false;
10120 return isPowerOf2_32(Scale);
10124 /// isLegalAddressingMode - Return true if the addressing mode represented
10125 /// by AM is legal for this target, for a load/store of the specified type.
10126 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
10128 EVT VT = getValueType(Ty, true);
10129 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
10132 // Can never fold addr of global into load/store.
10136 switch (AM.Scale) {
10137 case 0: // no scale reg, must be "r+i" or "r", or "i".
10140 if (Subtarget->isThumb1Only())
10144 // ARM doesn't support any R+R*scale+imm addr modes.
10148 if (!VT.isSimple())
10151 if (Subtarget->isThumb2())
10152 return isLegalT2ScaledAddressingMode(AM, VT);
10154 int Scale = AM.Scale;
10155 switch (VT.getSimpleVT().SimpleTy) {
10156 default: return false;
10160 if (Scale < 0) Scale = -Scale;
10164 return isPowerOf2_32(Scale & ~1);
10168 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
10173 // Note, we allow "void" uses (basically, uses that aren't loads or
10174 // stores), because arm allows folding a scale into many arithmetic
10175 // operations. This should be made more precise and revisited later.
10177 // Allow r << imm, but the imm has to be a multiple of two.
10178 if (Scale & 1) return false;
10179 return isPowerOf2_32(Scale);
10185 /// isLegalICmpImmediate - Return true if the specified immediate is legal
10186 /// icmp immediate, that is the target has icmp instructions which can compare
10187 /// a register against the immediate without having to materialize the
10188 /// immediate into a register.
10189 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
10190 // Thumb2 and ARM modes can use cmn for negative immediates.
10191 if (!Subtarget->isThumb())
10192 return ARM_AM::getSOImmVal(llvm::abs64(Imm)) != -1;
10193 if (Subtarget->isThumb2())
10194 return ARM_AM::getT2SOImmVal(llvm::abs64(Imm)) != -1;
10195 // Thumb1 doesn't have cmn, and only 8-bit immediates.
10196 return Imm >= 0 && Imm <= 255;
10199 /// isLegalAddImmediate - Return true if the specified immediate is a legal add
10200 /// *or sub* immediate, that is the target has add or sub instructions which can
10201 /// add a register with the immediate without having to materialize the
10202 /// immediate into a register.
10203 bool ARMTargetLowering::isLegalAddImmediate(int64_t Imm) const {
10204 // Same encoding for add/sub, just flip the sign.
10205 int64_t AbsImm = llvm::abs64(Imm);
10206 if (!Subtarget->isThumb())
10207 return ARM_AM::getSOImmVal(AbsImm) != -1;
10208 if (Subtarget->isThumb2())
10209 return ARM_AM::getT2SOImmVal(AbsImm) != -1;
10210 // Thumb1 only has 8-bit unsigned immediate.
10211 return AbsImm >= 0 && AbsImm <= 255;
10214 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
10215 bool isSEXTLoad, SDValue &Base,
10216 SDValue &Offset, bool &isInc,
10217 SelectionDAG &DAG) {
10218 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
10221 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
10222 // AddressingMode 3
10223 Base = Ptr->getOperand(0);
10224 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
10225 int RHSC = (int)RHS->getZExtValue();
10226 if (RHSC < 0 && RHSC > -256) {
10227 assert(Ptr->getOpcode() == ISD::ADD);
10229 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
10233 isInc = (Ptr->getOpcode() == ISD::ADD);
10234 Offset = Ptr->getOperand(1);
10236 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
10237 // AddressingMode 2
10238 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
10239 int RHSC = (int)RHS->getZExtValue();
10240 if (RHSC < 0 && RHSC > -0x1000) {
10241 assert(Ptr->getOpcode() == ISD::ADD);
10243 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
10244 Base = Ptr->getOperand(0);
10249 if (Ptr->getOpcode() == ISD::ADD) {
10251 ARM_AM::ShiftOpc ShOpcVal=
10252 ARM_AM::getShiftOpcForNode(Ptr->getOperand(0).getOpcode());
10253 if (ShOpcVal != ARM_AM::no_shift) {
10254 Base = Ptr->getOperand(1);
10255 Offset = Ptr->getOperand(0);
10257 Base = Ptr->getOperand(0);
10258 Offset = Ptr->getOperand(1);
10263 isInc = (Ptr->getOpcode() == ISD::ADD);
10264 Base = Ptr->getOperand(0);
10265 Offset = Ptr->getOperand(1);
10269 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
10273 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
10274 bool isSEXTLoad, SDValue &Base,
10275 SDValue &Offset, bool &isInc,
10276 SelectionDAG &DAG) {
10277 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
10280 Base = Ptr->getOperand(0);
10281 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
10282 int RHSC = (int)RHS->getZExtValue();
10283 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
10284 assert(Ptr->getOpcode() == ISD::ADD);
10286 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
10288 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
10289 isInc = Ptr->getOpcode() == ISD::ADD;
10290 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
10298 /// getPreIndexedAddressParts - returns true by value, base pointer and
10299 /// offset pointer and addressing mode by reference if the node's address
10300 /// can be legally represented as pre-indexed load / store address.
10302 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
10304 ISD::MemIndexedMode &AM,
10305 SelectionDAG &DAG) const {
10306 if (Subtarget->isThumb1Only())
10311 bool isSEXTLoad = false;
10312 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
10313 Ptr = LD->getBasePtr();
10314 VT = LD->getMemoryVT();
10315 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
10316 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
10317 Ptr = ST->getBasePtr();
10318 VT = ST->getMemoryVT();
10323 bool isLegal = false;
10324 if (Subtarget->isThumb2())
10325 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
10326 Offset, isInc, DAG);
10328 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
10329 Offset, isInc, DAG);
10333 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
10337 /// getPostIndexedAddressParts - returns true by value, base pointer and
10338 /// offset pointer and addressing mode by reference if this node can be
10339 /// combined with a load / store to form a post-indexed load / store.
10340 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
10343 ISD::MemIndexedMode &AM,
10344 SelectionDAG &DAG) const {
10345 if (Subtarget->isThumb1Only())
10350 bool isSEXTLoad = false;
10351 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
10352 VT = LD->getMemoryVT();
10353 Ptr = LD->getBasePtr();
10354 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
10355 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
10356 VT = ST->getMemoryVT();
10357 Ptr = ST->getBasePtr();
10362 bool isLegal = false;
10363 if (Subtarget->isThumb2())
10364 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
10367 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
10373 // Swap base ptr and offset to catch more post-index load / store when
10374 // it's legal. In Thumb2 mode, offset must be an immediate.
10375 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
10376 !Subtarget->isThumb2())
10377 std::swap(Base, Offset);
10379 // Post-indexed load / store update the base pointer.
10384 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
10388 void ARMTargetLowering::computeKnownBitsForTargetNode(const SDValue Op,
10391 const SelectionDAG &DAG,
10392 unsigned Depth) const {
10393 unsigned BitWidth = KnownOne.getBitWidth();
10394 KnownZero = KnownOne = APInt(BitWidth, 0);
10395 switch (Op.getOpcode()) {
10401 // These nodes' second result is a boolean
10402 if (Op.getResNo() == 0)
10404 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - 1);
10406 case ARMISD::CMOV: {
10407 // Bits are known zero/one if known on the LHS and RHS.
10408 DAG.computeKnownBits(Op.getOperand(0), KnownZero, KnownOne, Depth+1);
10409 if (KnownZero == 0 && KnownOne == 0) return;
10411 APInt KnownZeroRHS, KnownOneRHS;
10412 DAG.computeKnownBits(Op.getOperand(1), KnownZeroRHS, KnownOneRHS, Depth+1);
10413 KnownZero &= KnownZeroRHS;
10414 KnownOne &= KnownOneRHS;
10417 case ISD::INTRINSIC_W_CHAIN: {
10418 ConstantSDNode *CN = cast<ConstantSDNode>(Op->getOperand(1));
10419 Intrinsic::ID IntID = static_cast<Intrinsic::ID>(CN->getZExtValue());
10422 case Intrinsic::arm_ldaex:
10423 case Intrinsic::arm_ldrex: {
10424 EVT VT = cast<MemIntrinsicSDNode>(Op)->getMemoryVT();
10425 unsigned MemBits = VT.getScalarType().getSizeInBits();
10426 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - MemBits);
10434 //===----------------------------------------------------------------------===//
10435 // ARM Inline Assembly Support
10436 //===----------------------------------------------------------------------===//
10438 bool ARMTargetLowering::ExpandInlineAsm(CallInst *CI) const {
10439 // Looking for "rev" which is V6+.
10440 if (!Subtarget->hasV6Ops())
10443 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
10444 std::string AsmStr = IA->getAsmString();
10445 SmallVector<StringRef, 4> AsmPieces;
10446 SplitString(AsmStr, AsmPieces, ";\n");
10448 switch (AsmPieces.size()) {
10449 default: return false;
10451 AsmStr = AsmPieces[0];
10453 SplitString(AsmStr, AsmPieces, " \t,");
10456 if (AsmPieces.size() == 3 &&
10457 AsmPieces[0] == "rev" && AsmPieces[1] == "$0" && AsmPieces[2] == "$1" &&
10458 IA->getConstraintString().compare(0, 4, "=l,l") == 0) {
10459 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
10460 if (Ty && Ty->getBitWidth() == 32)
10461 return IntrinsicLowering::LowerToByteSwap(CI);
10469 /// getConstraintType - Given a constraint letter, return the type of
10470 /// constraint it is for this target.
10471 ARMTargetLowering::ConstraintType
10472 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
10473 if (Constraint.size() == 1) {
10474 switch (Constraint[0]) {
10476 case 'l': return C_RegisterClass;
10477 case 'w': return C_RegisterClass;
10478 case 'h': return C_RegisterClass;
10479 case 'x': return C_RegisterClass;
10480 case 't': return C_RegisterClass;
10481 case 'j': return C_Other; // Constant for movw.
10482 // An address with a single base register. Due to the way we
10483 // currently handle addresses it is the same as an 'r' memory constraint.
10484 case 'Q': return C_Memory;
10486 } else if (Constraint.size() == 2) {
10487 switch (Constraint[0]) {
10489 // All 'U+' constraints are addresses.
10490 case 'U': return C_Memory;
10493 return TargetLowering::getConstraintType(Constraint);
10496 /// Examine constraint type and operand type and determine a weight value.
10497 /// This object must already have been set up with the operand type
10498 /// and the current alternative constraint selected.
10499 TargetLowering::ConstraintWeight
10500 ARMTargetLowering::getSingleConstraintMatchWeight(
10501 AsmOperandInfo &info, const char *constraint) const {
10502 ConstraintWeight weight = CW_Invalid;
10503 Value *CallOperandVal = info.CallOperandVal;
10504 // If we don't have a value, we can't do a match,
10505 // but allow it at the lowest weight.
10506 if (!CallOperandVal)
10508 Type *type = CallOperandVal->getType();
10509 // Look at the constraint type.
10510 switch (*constraint) {
10512 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
10515 if (type->isIntegerTy()) {
10516 if (Subtarget->isThumb())
10517 weight = CW_SpecificReg;
10519 weight = CW_Register;
10523 if (type->isFloatingPointTy())
10524 weight = CW_Register;
10530 typedef std::pair<unsigned, const TargetRegisterClass*> RCPair;
10532 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
10534 if (Constraint.size() == 1) {
10535 // GCC ARM Constraint Letters
10536 switch (Constraint[0]) {
10537 case 'l': // Low regs or general regs.
10538 if (Subtarget->isThumb())
10539 return RCPair(0U, &ARM::tGPRRegClass);
10540 return RCPair(0U, &ARM::GPRRegClass);
10541 case 'h': // High regs or no regs.
10542 if (Subtarget->isThumb())
10543 return RCPair(0U, &ARM::hGPRRegClass);
10546 if (Subtarget->isThumb1Only())
10547 return RCPair(0U, &ARM::tGPRRegClass);
10548 return RCPair(0U, &ARM::GPRRegClass);
10550 if (VT == MVT::Other)
10552 if (VT == MVT::f32)
10553 return RCPair(0U, &ARM::SPRRegClass);
10554 if (VT.getSizeInBits() == 64)
10555 return RCPair(0U, &ARM::DPRRegClass);
10556 if (VT.getSizeInBits() == 128)
10557 return RCPair(0U, &ARM::QPRRegClass);
10560 if (VT == MVT::Other)
10562 if (VT == MVT::f32)
10563 return RCPair(0U, &ARM::SPR_8RegClass);
10564 if (VT.getSizeInBits() == 64)
10565 return RCPair(0U, &ARM::DPR_8RegClass);
10566 if (VT.getSizeInBits() == 128)
10567 return RCPair(0U, &ARM::QPR_8RegClass);
10570 if (VT == MVT::f32)
10571 return RCPair(0U, &ARM::SPRRegClass);
10575 if (StringRef("{cc}").equals_lower(Constraint))
10576 return std::make_pair(unsigned(ARM::CPSR), &ARM::CCRRegClass);
10578 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
10581 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
10582 /// vector. If it is invalid, don't add anything to Ops.
10583 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
10584 std::string &Constraint,
10585 std::vector<SDValue>&Ops,
10586 SelectionDAG &DAG) const {
10589 // Currently only support length 1 constraints.
10590 if (Constraint.length() != 1) return;
10592 char ConstraintLetter = Constraint[0];
10593 switch (ConstraintLetter) {
10596 case 'I': case 'J': case 'K': case 'L':
10597 case 'M': case 'N': case 'O':
10598 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
10602 int64_t CVal64 = C->getSExtValue();
10603 int CVal = (int) CVal64;
10604 // None of these constraints allow values larger than 32 bits. Check
10605 // that the value fits in an int.
10606 if (CVal != CVal64)
10609 switch (ConstraintLetter) {
10611 // Constant suitable for movw, must be between 0 and
10613 if (Subtarget->hasV6T2Ops())
10614 if (CVal >= 0 && CVal <= 65535)
10618 if (Subtarget->isThumb1Only()) {
10619 // This must be a constant between 0 and 255, for ADD
10621 if (CVal >= 0 && CVal <= 255)
10623 } else if (Subtarget->isThumb2()) {
10624 // A constant that can be used as an immediate value in a
10625 // data-processing instruction.
10626 if (ARM_AM::getT2SOImmVal(CVal) != -1)
10629 // A constant that can be used as an immediate value in a
10630 // data-processing instruction.
10631 if (ARM_AM::getSOImmVal(CVal) != -1)
10637 if (Subtarget->isThumb()) { // FIXME thumb2
10638 // This must be a constant between -255 and -1, for negated ADD
10639 // immediates. This can be used in GCC with an "n" modifier that
10640 // prints the negated value, for use with SUB instructions. It is
10641 // not useful otherwise but is implemented for compatibility.
10642 if (CVal >= -255 && CVal <= -1)
10645 // This must be a constant between -4095 and 4095. It is not clear
10646 // what this constraint is intended for. Implemented for
10647 // compatibility with GCC.
10648 if (CVal >= -4095 && CVal <= 4095)
10654 if (Subtarget->isThumb1Only()) {
10655 // A 32-bit value where only one byte has a nonzero value. Exclude
10656 // zero to match GCC. This constraint is used by GCC internally for
10657 // constants that can be loaded with a move/shift combination.
10658 // It is not useful otherwise but is implemented for compatibility.
10659 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
10661 } else if (Subtarget->isThumb2()) {
10662 // A constant whose bitwise inverse can be used as an immediate
10663 // value in a data-processing instruction. This can be used in GCC
10664 // with a "B" modifier that prints the inverted value, for use with
10665 // BIC and MVN instructions. It is not useful otherwise but is
10666 // implemented for compatibility.
10667 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
10670 // A constant whose bitwise inverse can be used as an immediate
10671 // value in a data-processing instruction. This can be used in GCC
10672 // with a "B" modifier that prints the inverted value, for use with
10673 // BIC and MVN instructions. It is not useful otherwise but is
10674 // implemented for compatibility.
10675 if (ARM_AM::getSOImmVal(~CVal) != -1)
10681 if (Subtarget->isThumb1Only()) {
10682 // This must be a constant between -7 and 7,
10683 // for 3-operand ADD/SUB immediate instructions.
10684 if (CVal >= -7 && CVal < 7)
10686 } else if (Subtarget->isThumb2()) {
10687 // A constant whose negation can be used as an immediate value in a
10688 // data-processing instruction. This can be used in GCC with an "n"
10689 // modifier that prints the negated value, for use with SUB
10690 // instructions. It is not useful otherwise but is implemented for
10692 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
10695 // A constant whose negation can be used as an immediate value in a
10696 // data-processing instruction. This can be used in GCC with an "n"
10697 // modifier that prints the negated value, for use with SUB
10698 // instructions. It is not useful otherwise but is implemented for
10700 if (ARM_AM::getSOImmVal(-CVal) != -1)
10706 if (Subtarget->isThumb()) { // FIXME thumb2
10707 // This must be a multiple of 4 between 0 and 1020, for
10708 // ADD sp + immediate.
10709 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
10712 // A power of two or a constant between 0 and 32. This is used in
10713 // GCC for the shift amount on shifted register operands, but it is
10714 // useful in general for any shift amounts.
10715 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
10721 if (Subtarget->isThumb()) { // FIXME thumb2
10722 // This must be a constant between 0 and 31, for shift amounts.
10723 if (CVal >= 0 && CVal <= 31)
10729 if (Subtarget->isThumb()) { // FIXME thumb2
10730 // This must be a multiple of 4 between -508 and 508, for
10731 // ADD/SUB sp = sp + immediate.
10732 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
10737 Result = DAG.getTargetConstant(CVal, Op.getValueType());
10741 if (Result.getNode()) {
10742 Ops.push_back(Result);
10745 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
10748 SDValue ARMTargetLowering::LowerDivRem(SDValue Op, SelectionDAG &DAG) const {
10749 assert(Subtarget->isTargetAEABI() && "Register-based DivRem lowering only");
10750 unsigned Opcode = Op->getOpcode();
10751 assert((Opcode == ISD::SDIVREM || Opcode == ISD::UDIVREM) &&
10752 "Invalid opcode for Div/Rem lowering");
10753 bool isSigned = (Opcode == ISD::SDIVREM);
10754 EVT VT = Op->getValueType(0);
10755 Type *Ty = VT.getTypeForEVT(*DAG.getContext());
10758 switch (VT.getSimpleVT().SimpleTy) {
10759 default: llvm_unreachable("Unexpected request for libcall!");
10760 case MVT::i8: LC = isSigned ? RTLIB::SDIVREM_I8 : RTLIB::UDIVREM_I8; break;
10761 case MVT::i16: LC = isSigned ? RTLIB::SDIVREM_I16 : RTLIB::UDIVREM_I16; break;
10762 case MVT::i32: LC = isSigned ? RTLIB::SDIVREM_I32 : RTLIB::UDIVREM_I32; break;
10763 case MVT::i64: LC = isSigned ? RTLIB::SDIVREM_I64 : RTLIB::UDIVREM_I64; break;
10766 SDValue InChain = DAG.getEntryNode();
10768 TargetLowering::ArgListTy Args;
10769 TargetLowering::ArgListEntry Entry;
10770 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i) {
10771 EVT ArgVT = Op->getOperand(i).getValueType();
10772 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
10773 Entry.Node = Op->getOperand(i);
10775 Entry.isSExt = isSigned;
10776 Entry.isZExt = !isSigned;
10777 Args.push_back(Entry);
10780 SDValue Callee = DAG.getExternalSymbol(getLibcallName(LC),
10783 Type *RetTy = (Type*)StructType::get(Ty, Ty, nullptr);
10786 TargetLowering::CallLoweringInfo CLI(DAG);
10787 CLI.setDebugLoc(dl).setChain(InChain)
10788 .setCallee(getLibcallCallingConv(LC), RetTy, Callee, std::move(Args), 0)
10789 .setInRegister().setSExtResult(isSigned).setZExtResult(!isSigned);
10791 std::pair<SDValue, SDValue> CallInfo = LowerCallTo(CLI);
10792 return CallInfo.first;
10796 ARMTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const {
10797 assert(Subtarget->isTargetWindows() && "unsupported target platform");
10801 SDValue Chain = Op.getOperand(0);
10802 SDValue Size = Op.getOperand(1);
10804 SDValue Words = DAG.getNode(ISD::SRL, DL, MVT::i32, Size,
10805 DAG.getConstant(2, MVT::i32));
10808 Chain = DAG.getCopyToReg(Chain, DL, ARM::R4, Words, Flag);
10809 Flag = Chain.getValue(1);
10811 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
10812 Chain = DAG.getNode(ARMISD::WIN__CHKSTK, DL, NodeTys, Chain, Flag);
10814 SDValue NewSP = DAG.getCopyFromReg(Chain, DL, ARM::SP, MVT::i32);
10815 Chain = NewSP.getValue(1);
10817 SDValue Ops[2] = { NewSP, Chain };
10818 return DAG.getMergeValues(Ops, DL);
10821 SDValue ARMTargetLowering::LowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) const {
10822 assert(Op.getValueType() == MVT::f64 && Subtarget->isFPOnlySP() &&
10823 "Unexpected type for custom-lowering FP_EXTEND");
10826 LC = RTLIB::getFPEXT(Op.getOperand(0).getValueType(), Op.getValueType());
10828 SDValue SrcVal = Op.getOperand(0);
10829 return makeLibCall(DAG, LC, Op.getValueType(), &SrcVal, 1,
10830 /*isSigned*/ false, SDLoc(Op)).first;
10833 SDValue ARMTargetLowering::LowerFP_ROUND(SDValue Op, SelectionDAG &DAG) const {
10834 assert(Op.getOperand(0).getValueType() == MVT::f64 &&
10835 Subtarget->isFPOnlySP() &&
10836 "Unexpected type for custom-lowering FP_ROUND");
10839 LC = RTLIB::getFPROUND(Op.getOperand(0).getValueType(), Op.getValueType());
10841 SDValue SrcVal = Op.getOperand(0);
10842 return makeLibCall(DAG, LC, Op.getValueType(), &SrcVal, 1,
10843 /*isSigned*/ false, SDLoc(Op)).first;
10847 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
10848 // The ARM target isn't yet aware of offsets.
10852 bool ARM::isBitFieldInvertedMask(unsigned v) {
10853 if (v == 0xffffffff)
10856 // there can be 1's on either or both "outsides", all the "inside"
10857 // bits must be 0's
10858 unsigned TO = CountTrailingOnes_32(v);
10859 unsigned LO = CountLeadingOnes_32(v);
10860 v = (v >> TO) << TO;
10861 v = (v << LO) >> LO;
10865 /// isFPImmLegal - Returns true if the target can instruction select the
10866 /// specified FP immediate natively. If false, the legalizer will
10867 /// materialize the FP immediate as a load from a constant pool.
10868 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
10869 if (!Subtarget->hasVFP3())
10871 if (VT == MVT::f32)
10872 return ARM_AM::getFP32Imm(Imm) != -1;
10873 if (VT == MVT::f64 && !Subtarget->isFPOnlySP())
10874 return ARM_AM::getFP64Imm(Imm) != -1;
10878 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
10879 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
10880 /// specified in the intrinsic calls.
10881 bool ARMTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
10883 unsigned Intrinsic) const {
10884 switch (Intrinsic) {
10885 case Intrinsic::arm_neon_vld1:
10886 case Intrinsic::arm_neon_vld2:
10887 case Intrinsic::arm_neon_vld3:
10888 case Intrinsic::arm_neon_vld4:
10889 case Intrinsic::arm_neon_vld2lane:
10890 case Intrinsic::arm_neon_vld3lane:
10891 case Intrinsic::arm_neon_vld4lane: {
10892 Info.opc = ISD::INTRINSIC_W_CHAIN;
10893 // Conservatively set memVT to the entire set of vectors loaded.
10894 uint64_t NumElts = getDataLayout()->getTypeAllocSize(I.getType()) / 8;
10895 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
10896 Info.ptrVal = I.getArgOperand(0);
10898 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
10899 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
10900 Info.vol = false; // volatile loads with NEON intrinsics not supported
10901 Info.readMem = true;
10902 Info.writeMem = false;
10905 case Intrinsic::arm_neon_vst1:
10906 case Intrinsic::arm_neon_vst2:
10907 case Intrinsic::arm_neon_vst3:
10908 case Intrinsic::arm_neon_vst4:
10909 case Intrinsic::arm_neon_vst2lane:
10910 case Intrinsic::arm_neon_vst3lane:
10911 case Intrinsic::arm_neon_vst4lane: {
10912 Info.opc = ISD::INTRINSIC_VOID;
10913 // Conservatively set memVT to the entire set of vectors stored.
10914 unsigned NumElts = 0;
10915 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
10916 Type *ArgTy = I.getArgOperand(ArgI)->getType();
10917 if (!ArgTy->isVectorTy())
10919 NumElts += getDataLayout()->getTypeAllocSize(ArgTy) / 8;
10921 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
10922 Info.ptrVal = I.getArgOperand(0);
10924 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
10925 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
10926 Info.vol = false; // volatile stores with NEON intrinsics not supported
10927 Info.readMem = false;
10928 Info.writeMem = true;
10931 case Intrinsic::arm_ldaex:
10932 case Intrinsic::arm_ldrex: {
10933 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(0)->getType());
10934 Info.opc = ISD::INTRINSIC_W_CHAIN;
10935 Info.memVT = MVT::getVT(PtrTy->getElementType());
10936 Info.ptrVal = I.getArgOperand(0);
10938 Info.align = getDataLayout()->getABITypeAlignment(PtrTy->getElementType());
10940 Info.readMem = true;
10941 Info.writeMem = false;
10944 case Intrinsic::arm_stlex:
10945 case Intrinsic::arm_strex: {
10946 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(1)->getType());
10947 Info.opc = ISD::INTRINSIC_W_CHAIN;
10948 Info.memVT = MVT::getVT(PtrTy->getElementType());
10949 Info.ptrVal = I.getArgOperand(1);
10951 Info.align = getDataLayout()->getABITypeAlignment(PtrTy->getElementType());
10953 Info.readMem = false;
10954 Info.writeMem = true;
10957 case Intrinsic::arm_stlexd:
10958 case Intrinsic::arm_strexd: {
10959 Info.opc = ISD::INTRINSIC_W_CHAIN;
10960 Info.memVT = MVT::i64;
10961 Info.ptrVal = I.getArgOperand(2);
10965 Info.readMem = false;
10966 Info.writeMem = true;
10969 case Intrinsic::arm_ldaexd:
10970 case Intrinsic::arm_ldrexd: {
10971 Info.opc = ISD::INTRINSIC_W_CHAIN;
10972 Info.memVT = MVT::i64;
10973 Info.ptrVal = I.getArgOperand(0);
10977 Info.readMem = true;
10978 Info.writeMem = false;
10988 /// \brief Returns true if it is beneficial to convert a load of a constant
10989 /// to just the constant itself.
10990 bool ARMTargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
10992 assert(Ty->isIntegerTy());
10994 unsigned Bits = Ty->getPrimitiveSizeInBits();
10995 if (Bits == 0 || Bits > 32)
11000 bool ARMTargetLowering::hasLoadLinkedStoreConditional() const { return true; }
11002 Instruction* ARMTargetLowering::makeDMB(IRBuilder<> &Builder,
11003 ARM_MB::MemBOpt Domain) const {
11004 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
11006 // First, if the target has no DMB, see what fallback we can use.
11007 if (!Subtarget->hasDataBarrier()) {
11008 // Some ARMv6 cpus can support data barriers with an mcr instruction.
11009 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
11011 if (Subtarget->hasV6Ops() && !Subtarget->isThumb()) {
11012 Function *MCR = llvm::Intrinsic::getDeclaration(M, Intrinsic::arm_mcr);
11013 Value* args[6] = {Builder.getInt32(15), Builder.getInt32(0),
11014 Builder.getInt32(0), Builder.getInt32(7),
11015 Builder.getInt32(10), Builder.getInt32(5)};
11016 return Builder.CreateCall(MCR, args);
11018 // Instead of using barriers, atomic accesses on these subtargets use
11020 llvm_unreachable("makeDMB on a target so old that it has no barriers");
11023 Function *DMB = llvm::Intrinsic::getDeclaration(M, Intrinsic::arm_dmb);
11024 // Only a full system barrier exists in the M-class architectures.
11025 Domain = Subtarget->isMClass() ? ARM_MB::SY : Domain;
11026 Constant *CDomain = Builder.getInt32(Domain);
11027 return Builder.CreateCall(DMB, CDomain);
11031 // Based on http://www.cl.cam.ac.uk/~pes20/cpp/cpp0xmappings.html
11032 Instruction* ARMTargetLowering::emitLeadingFence(IRBuilder<> &Builder,
11033 AtomicOrdering Ord, bool IsStore,
11034 bool IsLoad) const {
11035 if (!getInsertFencesForAtomic())
11041 llvm_unreachable("Invalid fence: unordered/non-atomic");
11044 return nullptr; // Nothing to do
11045 case SequentiallyConsistent:
11047 return nullptr; // Nothing to do
11050 case AcquireRelease:
11051 if (Subtarget->isSwift())
11052 return makeDMB(Builder, ARM_MB::ISHST);
11053 // FIXME: add a comment with a link to documentation justifying this.
11055 return makeDMB(Builder, ARM_MB::ISH);
11057 llvm_unreachable("Unknown fence ordering in emitLeadingFence");
11060 Instruction* ARMTargetLowering::emitTrailingFence(IRBuilder<> &Builder,
11061 AtomicOrdering Ord, bool IsStore,
11062 bool IsLoad) const {
11063 if (!getInsertFencesForAtomic())
11069 llvm_unreachable("Invalid fence: unordered/not-atomic");
11072 return nullptr; // Nothing to do
11074 case AcquireRelease:
11075 case SequentiallyConsistent:
11076 return makeDMB(Builder, ARM_MB::ISH);
11078 llvm_unreachable("Unknown fence ordering in emitTrailingFence");
11081 // Loads and stores less than 64-bits are already atomic; ones above that
11082 // are doomed anyway, so defer to the default libcall and blame the OS when
11083 // things go wrong. Cortex M doesn't have ldrexd/strexd though, so don't emit
11084 // anything for those.
11085 bool ARMTargetLowering::shouldExpandAtomicStoreInIR(StoreInst *SI) const {
11086 unsigned Size = SI->getValueOperand()->getType()->getPrimitiveSizeInBits();
11087 return (Size == 64) && !Subtarget->isMClass();
11090 // Loads and stores less than 64-bits are already atomic; ones above that
11091 // are doomed anyway, so defer to the default libcall and blame the OS when
11092 // things go wrong. Cortex M doesn't have ldrexd/strexd though, so don't emit
11093 // anything for those.
11094 // FIXME: ldrd and strd are atomic if the CPU has LPAE (e.g. A15 has that
11095 // guarantee, see DDI0406C ARM architecture reference manual,
11096 // sections A8.8.72-74 LDRD)
11097 bool ARMTargetLowering::shouldExpandAtomicLoadInIR(LoadInst *LI) const {
11098 unsigned Size = LI->getType()->getPrimitiveSizeInBits();
11099 return (Size == 64) && !Subtarget->isMClass();
11102 // For the real atomic operations, we have ldrex/strex up to 32 bits,
11103 // and up to 64 bits on the non-M profiles
11104 bool ARMTargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
11105 unsigned Size = AI->getType()->getPrimitiveSizeInBits();
11106 return Size <= (Subtarget->isMClass() ? 32U : 64U);
11109 // This has so far only been implemented for MachO.
11110 bool ARMTargetLowering::useLoadStackGuardNode() const {
11111 return Subtarget->isTargetMachO();
11114 bool ARMTargetLowering::canCombineStoreAndExtract(Type *VectorTy, Value *Idx,
11115 unsigned &Cost) const {
11116 // If we do not have NEON, vector types are not natively supported.
11117 if (!Subtarget->hasNEON())
11120 // Floating point values and vector values map to the same register file.
11121 // Therefore, althought we could do a store extract of a vector type, this is
11122 // better to leave at float as we have more freedom in the addressing mode for
11124 if (VectorTy->isFPOrFPVectorTy())
11127 // If the index is unknown at compile time, this is very expensive to lower
11128 // and it is not possible to combine the store with the extract.
11129 if (!isa<ConstantInt>(Idx))
11132 assert(VectorTy->isVectorTy() && "VectorTy is not a vector type");
11133 unsigned BitWidth = cast<VectorType>(VectorTy)->getBitWidth();
11134 // We can do a store + vector extract on any vector that fits perfectly in a D
11136 if (BitWidth == 64 || BitWidth == 128) {
11143 Value *ARMTargetLowering::emitLoadLinked(IRBuilder<> &Builder, Value *Addr,
11144 AtomicOrdering Ord) const {
11145 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
11146 Type *ValTy = cast<PointerType>(Addr->getType())->getElementType();
11147 bool IsAcquire = isAtLeastAcquire(Ord);
11149 // Since i64 isn't legal and intrinsics don't get type-lowered, the ldrexd
11150 // intrinsic must return {i32, i32} and we have to recombine them into a
11151 // single i64 here.
11152 if (ValTy->getPrimitiveSizeInBits() == 64) {
11153 Intrinsic::ID Int =
11154 IsAcquire ? Intrinsic::arm_ldaexd : Intrinsic::arm_ldrexd;
11155 Function *Ldrex = llvm::Intrinsic::getDeclaration(M, Int);
11157 Addr = Builder.CreateBitCast(Addr, Type::getInt8PtrTy(M->getContext()));
11158 Value *LoHi = Builder.CreateCall(Ldrex, Addr, "lohi");
11160 Value *Lo = Builder.CreateExtractValue(LoHi, 0, "lo");
11161 Value *Hi = Builder.CreateExtractValue(LoHi, 1, "hi");
11162 if (!Subtarget->isLittle())
11163 std::swap (Lo, Hi);
11164 Lo = Builder.CreateZExt(Lo, ValTy, "lo64");
11165 Hi = Builder.CreateZExt(Hi, ValTy, "hi64");
11166 return Builder.CreateOr(
11167 Lo, Builder.CreateShl(Hi, ConstantInt::get(ValTy, 32)), "val64");
11170 Type *Tys[] = { Addr->getType() };
11171 Intrinsic::ID Int = IsAcquire ? Intrinsic::arm_ldaex : Intrinsic::arm_ldrex;
11172 Function *Ldrex = llvm::Intrinsic::getDeclaration(M, Int, Tys);
11174 return Builder.CreateTruncOrBitCast(
11175 Builder.CreateCall(Ldrex, Addr),
11176 cast<PointerType>(Addr->getType())->getElementType());
11179 Value *ARMTargetLowering::emitStoreConditional(IRBuilder<> &Builder, Value *Val,
11181 AtomicOrdering Ord) const {
11182 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
11183 bool IsRelease = isAtLeastRelease(Ord);
11185 // Since the intrinsics must have legal type, the i64 intrinsics take two
11186 // parameters: "i32, i32". We must marshal Val into the appropriate form
11187 // before the call.
11188 if (Val->getType()->getPrimitiveSizeInBits() == 64) {
11189 Intrinsic::ID Int =
11190 IsRelease ? Intrinsic::arm_stlexd : Intrinsic::arm_strexd;
11191 Function *Strex = Intrinsic::getDeclaration(M, Int);
11192 Type *Int32Ty = Type::getInt32Ty(M->getContext());
11194 Value *Lo = Builder.CreateTrunc(Val, Int32Ty, "lo");
11195 Value *Hi = Builder.CreateTrunc(Builder.CreateLShr(Val, 32), Int32Ty, "hi");
11196 if (!Subtarget->isLittle())
11197 std::swap (Lo, Hi);
11198 Addr = Builder.CreateBitCast(Addr, Type::getInt8PtrTy(M->getContext()));
11199 return Builder.CreateCall3(Strex, Lo, Hi, Addr);
11202 Intrinsic::ID Int = IsRelease ? Intrinsic::arm_stlex : Intrinsic::arm_strex;
11203 Type *Tys[] = { Addr->getType() };
11204 Function *Strex = Intrinsic::getDeclaration(M, Int, Tys);
11206 return Builder.CreateCall2(
11207 Strex, Builder.CreateZExtOrBitCast(
11208 Val, Strex->getFunctionType()->getParamType(0)),
11220 static bool isHomogeneousAggregate(Type *Ty, HABaseType &Base,
11221 uint64_t &Members) {
11222 if (const StructType *ST = dyn_cast<StructType>(Ty)) {
11223 for (unsigned i = 0; i < ST->getNumElements(); ++i) {
11224 uint64_t SubMembers = 0;
11225 if (!isHomogeneousAggregate(ST->getElementType(i), Base, SubMembers))
11227 Members += SubMembers;
11229 } else if (const ArrayType *AT = dyn_cast<ArrayType>(Ty)) {
11230 uint64_t SubMembers = 0;
11231 if (!isHomogeneousAggregate(AT->getElementType(), Base, SubMembers))
11233 Members += SubMembers * AT->getNumElements();
11234 } else if (Ty->isFloatTy()) {
11235 if (Base != HA_UNKNOWN && Base != HA_FLOAT)
11239 } else if (Ty->isDoubleTy()) {
11240 if (Base != HA_UNKNOWN && Base != HA_DOUBLE)
11244 } else if (const VectorType *VT = dyn_cast<VectorType>(Ty)) {
11251 return VT->getBitWidth() == 64;
11253 return VT->getBitWidth() == 128;
11255 switch (VT->getBitWidth()) {
11268 return (Members > 0 && Members <= 4);
11271 /// \brief Return true if a type is an AAPCS-VFP homogeneous aggregate.
11272 bool ARMTargetLowering::functionArgumentNeedsConsecutiveRegisters(
11273 Type *Ty, CallingConv::ID CallConv, bool isVarArg) const {
11274 if (getEffectiveCallingConv(CallConv, isVarArg) !=
11275 CallingConv::ARM_AAPCS_VFP)
11278 HABaseType Base = HA_UNKNOWN;
11279 uint64_t Members = 0;
11280 bool result = isHomogeneousAggregate(Ty, Base, Members);
11281 DEBUG(dbgs() << "isHA: " << result << " "; Ty->dump());