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/ADT/StringSwitch.h"
27 #include "llvm/CodeGen/CallingConvLower.h"
28 #include "llvm/CodeGen/IntrinsicLowering.h"
29 #include "llvm/CodeGen/MachineBasicBlock.h"
30 #include "llvm/CodeGen/MachineFrameInfo.h"
31 #include "llvm/CodeGen/MachineFunction.h"
32 #include "llvm/CodeGen/MachineInstrBuilder.h"
33 #include "llvm/CodeGen/MachineJumpTableInfo.h"
34 #include "llvm/CodeGen/MachineModuleInfo.h"
35 #include "llvm/CodeGen/MachineRegisterInfo.h"
36 #include "llvm/CodeGen/SelectionDAG.h"
37 #include "llvm/IR/CallingConv.h"
38 #include "llvm/IR/Constants.h"
39 #include "llvm/IR/Function.h"
40 #include "llvm/IR/GlobalValue.h"
41 #include "llvm/IR/IRBuilder.h"
42 #include "llvm/IR/Instruction.h"
43 #include "llvm/IR/Instructions.h"
44 #include "llvm/IR/IntrinsicInst.h"
45 #include "llvm/IR/Intrinsics.h"
46 #include "llvm/IR/Type.h"
47 #include "llvm/MC/MCSectionMachO.h"
48 #include "llvm/Support/CommandLine.h"
49 #include "llvm/Support/Debug.h"
50 #include "llvm/Support/ErrorHandling.h"
51 #include "llvm/Support/MathExtras.h"
52 #include "llvm/Support/raw_ostream.h"
53 #include "llvm/Target/TargetOptions.h"
57 #define DEBUG_TYPE "arm-isel"
59 STATISTIC(NumTailCalls, "Number of tail calls");
60 STATISTIC(NumMovwMovt, "Number of GAs materialized with movw + movt");
61 STATISTIC(NumLoopByVals, "Number of loops generated for byval arguments");
64 ARMInterworking("arm-interworking", cl::Hidden,
65 cl::desc("Enable / disable ARM interworking (for debugging only)"),
69 class ARMCCState : public CCState {
71 ARMCCState(CallingConv::ID CC, bool isVarArg, MachineFunction &MF,
72 SmallVectorImpl<CCValAssign> &locs, LLVMContext &C,
74 : CCState(CC, isVarArg, MF, locs, C) {
75 assert(((PC == Call) || (PC == Prologue)) &&
76 "ARMCCState users must specify whether their context is call"
77 "or prologue generation.");
83 // The APCS parameter registers.
84 static const MCPhysReg GPRArgRegs[] = {
85 ARM::R0, ARM::R1, ARM::R2, ARM::R3
88 void ARMTargetLowering::addTypeForNEON(MVT VT, MVT PromotedLdStVT,
89 MVT PromotedBitwiseVT) {
90 if (VT != PromotedLdStVT) {
91 setOperationAction(ISD::LOAD, VT, Promote);
92 AddPromotedToType (ISD::LOAD, VT, PromotedLdStVT);
94 setOperationAction(ISD::STORE, VT, Promote);
95 AddPromotedToType (ISD::STORE, VT, PromotedLdStVT);
98 MVT ElemTy = VT.getVectorElementType();
99 if (ElemTy != MVT::i64 && ElemTy != MVT::f64)
100 setOperationAction(ISD::SETCC, VT, Custom);
101 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
102 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
103 if (ElemTy == MVT::i32) {
104 setOperationAction(ISD::SINT_TO_FP, VT, Custom);
105 setOperationAction(ISD::UINT_TO_FP, VT, Custom);
106 setOperationAction(ISD::FP_TO_SINT, VT, Custom);
107 setOperationAction(ISD::FP_TO_UINT, VT, Custom);
109 setOperationAction(ISD::SINT_TO_FP, VT, Expand);
110 setOperationAction(ISD::UINT_TO_FP, VT, Expand);
111 setOperationAction(ISD::FP_TO_SINT, VT, Expand);
112 setOperationAction(ISD::FP_TO_UINT, VT, Expand);
114 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
115 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
116 setOperationAction(ISD::CONCAT_VECTORS, VT, Legal);
117 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Legal);
118 setOperationAction(ISD::SELECT, VT, Expand);
119 setOperationAction(ISD::SELECT_CC, VT, Expand);
120 setOperationAction(ISD::VSELECT, VT, Expand);
121 setOperationAction(ISD::SIGN_EXTEND_INREG, VT, Expand);
122 if (VT.isInteger()) {
123 setOperationAction(ISD::SHL, VT, Custom);
124 setOperationAction(ISD::SRA, VT, Custom);
125 setOperationAction(ISD::SRL, VT, Custom);
128 // Promote all bit-wise operations.
129 if (VT.isInteger() && VT != PromotedBitwiseVT) {
130 setOperationAction(ISD::AND, VT, Promote);
131 AddPromotedToType (ISD::AND, VT, PromotedBitwiseVT);
132 setOperationAction(ISD::OR, VT, Promote);
133 AddPromotedToType (ISD::OR, VT, PromotedBitwiseVT);
134 setOperationAction(ISD::XOR, VT, Promote);
135 AddPromotedToType (ISD::XOR, VT, PromotedBitwiseVT);
138 // Neon does not support vector divide/remainder operations.
139 setOperationAction(ISD::SDIV, VT, Expand);
140 setOperationAction(ISD::UDIV, VT, Expand);
141 setOperationAction(ISD::FDIV, VT, Expand);
142 setOperationAction(ISD::SREM, VT, Expand);
143 setOperationAction(ISD::UREM, VT, Expand);
144 setOperationAction(ISD::FREM, VT, Expand);
146 if (VT.isInteger()) {
147 setOperationAction(ISD::SABSDIFF, VT, Legal);
148 setOperationAction(ISD::UABSDIFF, VT, Legal);
150 if (!VT.isFloatingPoint() &&
151 VT != MVT::v2i64 && VT != MVT::v1i64)
152 for (unsigned Opcode : {ISD::SMIN, ISD::SMAX, ISD::UMIN, ISD::UMAX})
153 setOperationAction(Opcode, VT, Legal);
157 void ARMTargetLowering::addDRTypeForNEON(MVT VT) {
158 addRegisterClass(VT, &ARM::DPRRegClass);
159 addTypeForNEON(VT, MVT::f64, MVT::v2i32);
162 void ARMTargetLowering::addQRTypeForNEON(MVT VT) {
163 addRegisterClass(VT, &ARM::DPairRegClass);
164 addTypeForNEON(VT, MVT::v2f64, MVT::v4i32);
167 ARMTargetLowering::ARMTargetLowering(const TargetMachine &TM,
168 const ARMSubtarget &STI)
169 : TargetLowering(TM), Subtarget(&STI) {
170 RegInfo = Subtarget->getRegisterInfo();
171 Itins = Subtarget->getInstrItineraryData();
173 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
175 if (Subtarget->isTargetMachO()) {
176 // Uses VFP for Thumb libfuncs if available.
177 if (Subtarget->isThumb() && Subtarget->hasVFP2() &&
178 Subtarget->hasARMOps() && !Subtarget->useSoftFloat()) {
179 static const struct {
180 const RTLIB::Libcall Op;
181 const char * const Name;
182 const ISD::CondCode Cond;
184 // Single-precision floating-point arithmetic.
185 { RTLIB::ADD_F32, "__addsf3vfp", ISD::SETCC_INVALID },
186 { RTLIB::SUB_F32, "__subsf3vfp", ISD::SETCC_INVALID },
187 { RTLIB::MUL_F32, "__mulsf3vfp", ISD::SETCC_INVALID },
188 { RTLIB::DIV_F32, "__divsf3vfp", ISD::SETCC_INVALID },
190 // Double-precision floating-point arithmetic.
191 { RTLIB::ADD_F64, "__adddf3vfp", ISD::SETCC_INVALID },
192 { RTLIB::SUB_F64, "__subdf3vfp", ISD::SETCC_INVALID },
193 { RTLIB::MUL_F64, "__muldf3vfp", ISD::SETCC_INVALID },
194 { RTLIB::DIV_F64, "__divdf3vfp", ISD::SETCC_INVALID },
196 // Single-precision comparisons.
197 { RTLIB::OEQ_F32, "__eqsf2vfp", ISD::SETNE },
198 { RTLIB::UNE_F32, "__nesf2vfp", ISD::SETNE },
199 { RTLIB::OLT_F32, "__ltsf2vfp", ISD::SETNE },
200 { RTLIB::OLE_F32, "__lesf2vfp", ISD::SETNE },
201 { RTLIB::OGE_F32, "__gesf2vfp", ISD::SETNE },
202 { RTLIB::OGT_F32, "__gtsf2vfp", ISD::SETNE },
203 { RTLIB::UO_F32, "__unordsf2vfp", ISD::SETNE },
204 { RTLIB::O_F32, "__unordsf2vfp", ISD::SETEQ },
206 // Double-precision comparisons.
207 { RTLIB::OEQ_F64, "__eqdf2vfp", ISD::SETNE },
208 { RTLIB::UNE_F64, "__nedf2vfp", ISD::SETNE },
209 { RTLIB::OLT_F64, "__ltdf2vfp", ISD::SETNE },
210 { RTLIB::OLE_F64, "__ledf2vfp", ISD::SETNE },
211 { RTLIB::OGE_F64, "__gedf2vfp", ISD::SETNE },
212 { RTLIB::OGT_F64, "__gtdf2vfp", ISD::SETNE },
213 { RTLIB::UO_F64, "__unorddf2vfp", ISD::SETNE },
214 { RTLIB::O_F64, "__unorddf2vfp", ISD::SETEQ },
216 // Floating-point to integer conversions.
217 // i64 conversions are done via library routines even when generating VFP
218 // instructions, so use the same ones.
219 { RTLIB::FPTOSINT_F64_I32, "__fixdfsivfp", ISD::SETCC_INVALID },
220 { RTLIB::FPTOUINT_F64_I32, "__fixunsdfsivfp", ISD::SETCC_INVALID },
221 { RTLIB::FPTOSINT_F32_I32, "__fixsfsivfp", ISD::SETCC_INVALID },
222 { RTLIB::FPTOUINT_F32_I32, "__fixunssfsivfp", ISD::SETCC_INVALID },
224 // Conversions between floating types.
225 { RTLIB::FPROUND_F64_F32, "__truncdfsf2vfp", ISD::SETCC_INVALID },
226 { RTLIB::FPEXT_F32_F64, "__extendsfdf2vfp", ISD::SETCC_INVALID },
228 // Integer to floating-point conversions.
229 // i64 conversions are done via library routines even when generating VFP
230 // instructions, so use the same ones.
231 // FIXME: There appears to be some naming inconsistency in ARM libgcc:
232 // e.g., __floatunsidf vs. __floatunssidfvfp.
233 { RTLIB::SINTTOFP_I32_F64, "__floatsidfvfp", ISD::SETCC_INVALID },
234 { RTLIB::UINTTOFP_I32_F64, "__floatunssidfvfp", ISD::SETCC_INVALID },
235 { RTLIB::SINTTOFP_I32_F32, "__floatsisfvfp", ISD::SETCC_INVALID },
236 { RTLIB::UINTTOFP_I32_F32, "__floatunssisfvfp", ISD::SETCC_INVALID },
239 for (const auto &LC : LibraryCalls) {
240 setLibcallName(LC.Op, LC.Name);
241 if (LC.Cond != ISD::SETCC_INVALID)
242 setCmpLibcallCC(LC.Op, LC.Cond);
247 // These libcalls are not available in 32-bit.
248 setLibcallName(RTLIB::SHL_I128, nullptr);
249 setLibcallName(RTLIB::SRL_I128, nullptr);
250 setLibcallName(RTLIB::SRA_I128, nullptr);
252 if (Subtarget->isAAPCS_ABI() && !Subtarget->isTargetMachO() &&
253 !Subtarget->isTargetWindows()) {
254 static const struct {
255 const RTLIB::Libcall Op;
256 const char * const Name;
257 const CallingConv::ID CC;
258 const ISD::CondCode Cond;
260 // Double-precision floating-point arithmetic helper functions
261 // RTABI chapter 4.1.2, Table 2
262 { RTLIB::ADD_F64, "__aeabi_dadd", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
263 { RTLIB::DIV_F64, "__aeabi_ddiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
264 { RTLIB::MUL_F64, "__aeabi_dmul", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
265 { RTLIB::SUB_F64, "__aeabi_dsub", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
267 // Double-precision floating-point comparison helper functions
268 // RTABI chapter 4.1.2, Table 3
269 { RTLIB::OEQ_F64, "__aeabi_dcmpeq", CallingConv::ARM_AAPCS, ISD::SETNE },
270 { RTLIB::UNE_F64, "__aeabi_dcmpeq", CallingConv::ARM_AAPCS, ISD::SETEQ },
271 { RTLIB::OLT_F64, "__aeabi_dcmplt", CallingConv::ARM_AAPCS, ISD::SETNE },
272 { RTLIB::OLE_F64, "__aeabi_dcmple", CallingConv::ARM_AAPCS, ISD::SETNE },
273 { RTLIB::OGE_F64, "__aeabi_dcmpge", CallingConv::ARM_AAPCS, ISD::SETNE },
274 { RTLIB::OGT_F64, "__aeabi_dcmpgt", CallingConv::ARM_AAPCS, ISD::SETNE },
275 { RTLIB::UO_F64, "__aeabi_dcmpun", CallingConv::ARM_AAPCS, ISD::SETNE },
276 { RTLIB::O_F64, "__aeabi_dcmpun", CallingConv::ARM_AAPCS, ISD::SETEQ },
278 // Single-precision floating-point arithmetic helper functions
279 // RTABI chapter 4.1.2, Table 4
280 { RTLIB::ADD_F32, "__aeabi_fadd", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
281 { RTLIB::DIV_F32, "__aeabi_fdiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
282 { RTLIB::MUL_F32, "__aeabi_fmul", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
283 { RTLIB::SUB_F32, "__aeabi_fsub", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
285 // Single-precision floating-point comparison helper functions
286 // RTABI chapter 4.1.2, Table 5
287 { RTLIB::OEQ_F32, "__aeabi_fcmpeq", CallingConv::ARM_AAPCS, ISD::SETNE },
288 { RTLIB::UNE_F32, "__aeabi_fcmpeq", CallingConv::ARM_AAPCS, ISD::SETEQ },
289 { RTLIB::OLT_F32, "__aeabi_fcmplt", CallingConv::ARM_AAPCS, ISD::SETNE },
290 { RTLIB::OLE_F32, "__aeabi_fcmple", CallingConv::ARM_AAPCS, ISD::SETNE },
291 { RTLIB::OGE_F32, "__aeabi_fcmpge", CallingConv::ARM_AAPCS, ISD::SETNE },
292 { RTLIB::OGT_F32, "__aeabi_fcmpgt", CallingConv::ARM_AAPCS, ISD::SETNE },
293 { RTLIB::UO_F32, "__aeabi_fcmpun", CallingConv::ARM_AAPCS, ISD::SETNE },
294 { RTLIB::O_F32, "__aeabi_fcmpun", CallingConv::ARM_AAPCS, ISD::SETEQ },
296 // Floating-point to integer conversions.
297 // RTABI chapter 4.1.2, Table 6
298 { RTLIB::FPTOSINT_F64_I32, "__aeabi_d2iz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
299 { RTLIB::FPTOUINT_F64_I32, "__aeabi_d2uiz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
300 { RTLIB::FPTOSINT_F64_I64, "__aeabi_d2lz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
301 { RTLIB::FPTOUINT_F64_I64, "__aeabi_d2ulz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
302 { RTLIB::FPTOSINT_F32_I32, "__aeabi_f2iz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
303 { RTLIB::FPTOUINT_F32_I32, "__aeabi_f2uiz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
304 { RTLIB::FPTOSINT_F32_I64, "__aeabi_f2lz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
305 { RTLIB::FPTOUINT_F32_I64, "__aeabi_f2ulz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
307 // Conversions between floating types.
308 // RTABI chapter 4.1.2, Table 7
309 { RTLIB::FPROUND_F64_F32, "__aeabi_d2f", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
310 { RTLIB::FPROUND_F64_F16, "__aeabi_d2h", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
311 { RTLIB::FPEXT_F32_F64, "__aeabi_f2d", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
313 // Integer to floating-point conversions.
314 // RTABI chapter 4.1.2, Table 8
315 { RTLIB::SINTTOFP_I32_F64, "__aeabi_i2d", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
316 { RTLIB::UINTTOFP_I32_F64, "__aeabi_ui2d", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
317 { RTLIB::SINTTOFP_I64_F64, "__aeabi_l2d", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
318 { RTLIB::UINTTOFP_I64_F64, "__aeabi_ul2d", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
319 { RTLIB::SINTTOFP_I32_F32, "__aeabi_i2f", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
320 { RTLIB::UINTTOFP_I32_F32, "__aeabi_ui2f", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
321 { RTLIB::SINTTOFP_I64_F32, "__aeabi_l2f", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
322 { RTLIB::UINTTOFP_I64_F32, "__aeabi_ul2f", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
324 // Long long helper functions
325 // RTABI chapter 4.2, Table 9
326 { RTLIB::MUL_I64, "__aeabi_lmul", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
327 { RTLIB::SHL_I64, "__aeabi_llsl", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
328 { RTLIB::SRL_I64, "__aeabi_llsr", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
329 { RTLIB::SRA_I64, "__aeabi_lasr", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
331 // Integer division functions
332 // RTABI chapter 4.3.1
333 { RTLIB::SDIV_I8, "__aeabi_idiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
334 { RTLIB::SDIV_I16, "__aeabi_idiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
335 { RTLIB::SDIV_I32, "__aeabi_idiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
336 { RTLIB::SDIV_I64, "__aeabi_ldivmod", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
337 { RTLIB::UDIV_I8, "__aeabi_uidiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
338 { RTLIB::UDIV_I16, "__aeabi_uidiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
339 { RTLIB::UDIV_I32, "__aeabi_uidiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
340 { RTLIB::UDIV_I64, "__aeabi_uldivmod", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
343 // RTABI chapter 4.3.4
344 { RTLIB::MEMCPY, "__aeabi_memcpy", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
345 { RTLIB::MEMMOVE, "__aeabi_memmove", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
346 { RTLIB::MEMSET, "__aeabi_memset", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
349 for (const auto &LC : LibraryCalls) {
350 setLibcallName(LC.Op, LC.Name);
351 setLibcallCallingConv(LC.Op, LC.CC);
352 if (LC.Cond != ISD::SETCC_INVALID)
353 setCmpLibcallCC(LC.Op, LC.Cond);
357 if (Subtarget->isTargetWindows()) {
358 static const struct {
359 const RTLIB::Libcall Op;
360 const char * const Name;
361 const CallingConv::ID CC;
363 { RTLIB::FPTOSINT_F32_I64, "__stoi64", CallingConv::ARM_AAPCS_VFP },
364 { RTLIB::FPTOSINT_F64_I64, "__dtoi64", CallingConv::ARM_AAPCS_VFP },
365 { RTLIB::FPTOUINT_F32_I64, "__stou64", CallingConv::ARM_AAPCS_VFP },
366 { RTLIB::FPTOUINT_F64_I64, "__dtou64", CallingConv::ARM_AAPCS_VFP },
367 { RTLIB::SINTTOFP_I64_F32, "__i64tos", CallingConv::ARM_AAPCS_VFP },
368 { RTLIB::SINTTOFP_I64_F64, "__i64tod", CallingConv::ARM_AAPCS_VFP },
369 { RTLIB::UINTTOFP_I64_F32, "__u64tos", CallingConv::ARM_AAPCS_VFP },
370 { RTLIB::UINTTOFP_I64_F64, "__u64tod", CallingConv::ARM_AAPCS_VFP },
373 for (const auto &LC : LibraryCalls) {
374 setLibcallName(LC.Op, LC.Name);
375 setLibcallCallingConv(LC.Op, LC.CC);
379 // Use divmod compiler-rt calls for iOS 5.0 and later.
380 if (Subtarget->getTargetTriple().isiOS() &&
381 !Subtarget->getTargetTriple().isOSVersionLT(5, 0)) {
382 setLibcallName(RTLIB::SDIVREM_I32, "__divmodsi4");
383 setLibcallName(RTLIB::UDIVREM_I32, "__udivmodsi4");
386 // The half <-> float conversion functions are always soft-float, but are
387 // needed for some targets which use a hard-float calling convention by
389 if (Subtarget->isAAPCS_ABI()) {
390 setLibcallCallingConv(RTLIB::FPROUND_F32_F16, CallingConv::ARM_AAPCS);
391 setLibcallCallingConv(RTLIB::FPROUND_F64_F16, CallingConv::ARM_AAPCS);
392 setLibcallCallingConv(RTLIB::FPEXT_F16_F32, CallingConv::ARM_AAPCS);
394 setLibcallCallingConv(RTLIB::FPROUND_F32_F16, CallingConv::ARM_APCS);
395 setLibcallCallingConv(RTLIB::FPROUND_F64_F16, CallingConv::ARM_APCS);
396 setLibcallCallingConv(RTLIB::FPEXT_F16_F32, CallingConv::ARM_APCS);
399 // In EABI, these functions have an __aeabi_ prefix, but in GNUEABI they have
400 // a __gnu_ prefix (which is the default).
401 if (Subtarget->isTargetAEABI()) {
402 setLibcallName(RTLIB::FPROUND_F32_F16, "__aeabi_f2h");
403 setLibcallName(RTLIB::FPROUND_F64_F16, "__aeabi_d2h");
404 setLibcallName(RTLIB::FPEXT_F16_F32, "__aeabi_h2f");
407 if (Subtarget->isThumb1Only())
408 addRegisterClass(MVT::i32, &ARM::tGPRRegClass);
410 addRegisterClass(MVT::i32, &ARM::GPRRegClass);
411 if (!Subtarget->useSoftFloat() && Subtarget->hasVFP2() &&
412 !Subtarget->isThumb1Only()) {
413 addRegisterClass(MVT::f32, &ARM::SPRRegClass);
414 addRegisterClass(MVT::f64, &ARM::DPRRegClass);
417 for (MVT VT : MVT::vector_valuetypes()) {
418 for (MVT InnerVT : MVT::vector_valuetypes()) {
419 setTruncStoreAction(VT, InnerVT, Expand);
420 setLoadExtAction(ISD::SEXTLOAD, VT, InnerVT, Expand);
421 setLoadExtAction(ISD::ZEXTLOAD, VT, InnerVT, Expand);
422 setLoadExtAction(ISD::EXTLOAD, VT, InnerVT, Expand);
425 setOperationAction(ISD::MULHS, VT, Expand);
426 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
427 setOperationAction(ISD::MULHU, VT, Expand);
428 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
430 setOperationAction(ISD::BSWAP, VT, Expand);
433 setOperationAction(ISD::ConstantFP, MVT::f32, Custom);
434 setOperationAction(ISD::ConstantFP, MVT::f64, Custom);
436 setOperationAction(ISD::READ_REGISTER, MVT::i64, Custom);
437 setOperationAction(ISD::WRITE_REGISTER, MVT::i64, Custom);
439 if (Subtarget->hasNEON()) {
440 addDRTypeForNEON(MVT::v2f32);
441 addDRTypeForNEON(MVT::v8i8);
442 addDRTypeForNEON(MVT::v4i16);
443 addDRTypeForNEON(MVT::v2i32);
444 addDRTypeForNEON(MVT::v1i64);
446 addQRTypeForNEON(MVT::v4f32);
447 addQRTypeForNEON(MVT::v2f64);
448 addQRTypeForNEON(MVT::v16i8);
449 addQRTypeForNEON(MVT::v8i16);
450 addQRTypeForNEON(MVT::v4i32);
451 addQRTypeForNEON(MVT::v2i64);
453 // v2f64 is legal so that QR subregs can be extracted as f64 elements, but
454 // neither Neon nor VFP support any arithmetic operations on it.
455 // The same with v4f32. But keep in mind that vadd, vsub, vmul are natively
456 // supported for v4f32.
457 setOperationAction(ISD::FADD, MVT::v2f64, Expand);
458 setOperationAction(ISD::FSUB, MVT::v2f64, Expand);
459 setOperationAction(ISD::FMUL, MVT::v2f64, Expand);
460 // FIXME: Code duplication: FDIV and FREM are expanded always, see
461 // ARMTargetLowering::addTypeForNEON method for details.
462 setOperationAction(ISD::FDIV, MVT::v2f64, Expand);
463 setOperationAction(ISD::FREM, MVT::v2f64, Expand);
464 // FIXME: Create unittest.
465 // In another words, find a way when "copysign" appears in DAG with vector
467 setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand);
468 // FIXME: Code duplication: SETCC has custom operation action, see
469 // ARMTargetLowering::addTypeForNEON method for details.
470 setOperationAction(ISD::SETCC, MVT::v2f64, Expand);
471 // FIXME: Create unittest for FNEG and for FABS.
472 setOperationAction(ISD::FNEG, MVT::v2f64, Expand);
473 setOperationAction(ISD::FABS, MVT::v2f64, Expand);
474 setOperationAction(ISD::FSQRT, MVT::v2f64, Expand);
475 setOperationAction(ISD::FSIN, MVT::v2f64, Expand);
476 setOperationAction(ISD::FCOS, MVT::v2f64, Expand);
477 setOperationAction(ISD::FPOWI, MVT::v2f64, Expand);
478 setOperationAction(ISD::FPOW, MVT::v2f64, Expand);
479 setOperationAction(ISD::FLOG, MVT::v2f64, Expand);
480 setOperationAction(ISD::FLOG2, MVT::v2f64, Expand);
481 setOperationAction(ISD::FLOG10, MVT::v2f64, Expand);
482 setOperationAction(ISD::FEXP, MVT::v2f64, Expand);
483 setOperationAction(ISD::FEXP2, MVT::v2f64, Expand);
484 // FIXME: Create unittest for FCEIL, FTRUNC, FRINT, FNEARBYINT, FFLOOR.
485 setOperationAction(ISD::FCEIL, MVT::v2f64, Expand);
486 setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand);
487 setOperationAction(ISD::FRINT, MVT::v2f64, Expand);
488 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand);
489 setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand);
490 setOperationAction(ISD::FMA, MVT::v2f64, Expand);
492 setOperationAction(ISD::FSQRT, MVT::v4f32, Expand);
493 setOperationAction(ISD::FSIN, MVT::v4f32, Expand);
494 setOperationAction(ISD::FCOS, MVT::v4f32, Expand);
495 setOperationAction(ISD::FPOWI, MVT::v4f32, Expand);
496 setOperationAction(ISD::FPOW, MVT::v4f32, Expand);
497 setOperationAction(ISD::FLOG, MVT::v4f32, Expand);
498 setOperationAction(ISD::FLOG2, MVT::v4f32, Expand);
499 setOperationAction(ISD::FLOG10, MVT::v4f32, Expand);
500 setOperationAction(ISD::FEXP, MVT::v4f32, Expand);
501 setOperationAction(ISD::FEXP2, MVT::v4f32, Expand);
502 setOperationAction(ISD::FCEIL, MVT::v4f32, Expand);
503 setOperationAction(ISD::FTRUNC, MVT::v4f32, Expand);
504 setOperationAction(ISD::FRINT, MVT::v4f32, Expand);
505 setOperationAction(ISD::FNEARBYINT, MVT::v4f32, Expand);
506 setOperationAction(ISD::FFLOOR, MVT::v4f32, Expand);
508 // Mark v2f32 intrinsics.
509 setOperationAction(ISD::FSQRT, MVT::v2f32, Expand);
510 setOperationAction(ISD::FSIN, MVT::v2f32, Expand);
511 setOperationAction(ISD::FCOS, MVT::v2f32, Expand);
512 setOperationAction(ISD::FPOWI, MVT::v2f32, Expand);
513 setOperationAction(ISD::FPOW, MVT::v2f32, Expand);
514 setOperationAction(ISD::FLOG, MVT::v2f32, Expand);
515 setOperationAction(ISD::FLOG2, MVT::v2f32, Expand);
516 setOperationAction(ISD::FLOG10, MVT::v2f32, Expand);
517 setOperationAction(ISD::FEXP, MVT::v2f32, Expand);
518 setOperationAction(ISD::FEXP2, MVT::v2f32, Expand);
519 setOperationAction(ISD::FCEIL, MVT::v2f32, Expand);
520 setOperationAction(ISD::FTRUNC, MVT::v2f32, Expand);
521 setOperationAction(ISD::FRINT, MVT::v2f32, Expand);
522 setOperationAction(ISD::FNEARBYINT, MVT::v2f32, Expand);
523 setOperationAction(ISD::FFLOOR, MVT::v2f32, Expand);
525 // Neon does not support some operations on v1i64 and v2i64 types.
526 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
527 // Custom handling for some quad-vector types to detect VMULL.
528 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
529 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
530 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
531 // Custom handling for some vector types to avoid expensive expansions
532 setOperationAction(ISD::SDIV, MVT::v4i16, Custom);
533 setOperationAction(ISD::SDIV, MVT::v8i8, Custom);
534 setOperationAction(ISD::UDIV, MVT::v4i16, Custom);
535 setOperationAction(ISD::UDIV, MVT::v8i8, Custom);
536 setOperationAction(ISD::SETCC, MVT::v1i64, Expand);
537 setOperationAction(ISD::SETCC, MVT::v2i64, Expand);
538 // Neon does not have single instruction SINT_TO_FP and UINT_TO_FP with
539 // a destination type that is wider than the source, and nor does
540 // it have a FP_TO_[SU]INT instruction with a narrower destination than
542 setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Custom);
543 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
544 setOperationAction(ISD::FP_TO_UINT, MVT::v4i16, Custom);
545 setOperationAction(ISD::FP_TO_SINT, MVT::v4i16, Custom);
547 setOperationAction(ISD::FP_ROUND, MVT::v2f32, Expand);
548 setOperationAction(ISD::FP_EXTEND, MVT::v2f64, Expand);
550 // NEON does not have single instruction CTPOP for vectors with element
551 // types wider than 8-bits. However, custom lowering can leverage the
552 // v8i8/v16i8 vcnt instruction.
553 setOperationAction(ISD::CTPOP, MVT::v2i32, Custom);
554 setOperationAction(ISD::CTPOP, MVT::v4i32, Custom);
555 setOperationAction(ISD::CTPOP, MVT::v4i16, Custom);
556 setOperationAction(ISD::CTPOP, MVT::v8i16, Custom);
558 // NEON does not have single instruction CTTZ for vectors.
559 setOperationAction(ISD::CTTZ, MVT::v8i8, Custom);
560 setOperationAction(ISD::CTTZ, MVT::v4i16, Custom);
561 setOperationAction(ISD::CTTZ, MVT::v2i32, Custom);
562 setOperationAction(ISD::CTTZ, MVT::v1i64, Custom);
564 setOperationAction(ISD::CTTZ, MVT::v16i8, Custom);
565 setOperationAction(ISD::CTTZ, MVT::v8i16, Custom);
566 setOperationAction(ISD::CTTZ, MVT::v4i32, Custom);
567 setOperationAction(ISD::CTTZ, MVT::v2i64, Custom);
569 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v8i8, Custom);
570 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v4i16, Custom);
571 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v2i32, Custom);
572 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v1i64, Custom);
574 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v16i8, Custom);
575 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v8i16, Custom);
576 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v4i32, Custom);
577 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v2i64, Custom);
579 // NEON only has FMA instructions as of VFP4.
580 if (!Subtarget->hasVFP4()) {
581 setOperationAction(ISD::FMA, MVT::v2f32, Expand);
582 setOperationAction(ISD::FMA, MVT::v4f32, Expand);
585 setTargetDAGCombine(ISD::INTRINSIC_VOID);
586 setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN);
587 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
588 setTargetDAGCombine(ISD::SHL);
589 setTargetDAGCombine(ISD::SRL);
590 setTargetDAGCombine(ISD::SRA);
591 setTargetDAGCombine(ISD::SIGN_EXTEND);
592 setTargetDAGCombine(ISD::ZERO_EXTEND);
593 setTargetDAGCombine(ISD::ANY_EXTEND);
594 setTargetDAGCombine(ISD::BUILD_VECTOR);
595 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
596 setTargetDAGCombine(ISD::INSERT_VECTOR_ELT);
597 setTargetDAGCombine(ISD::STORE);
598 setTargetDAGCombine(ISD::FP_TO_SINT);
599 setTargetDAGCombine(ISD::FP_TO_UINT);
600 setTargetDAGCombine(ISD::FDIV);
601 setTargetDAGCombine(ISD::LOAD);
603 // It is legal to extload from v4i8 to v4i16 or v4i32.
604 for (MVT Ty : {MVT::v8i8, MVT::v4i8, MVT::v2i8, MVT::v4i16, MVT::v2i16,
606 for (MVT VT : MVT::integer_vector_valuetypes()) {
607 setLoadExtAction(ISD::EXTLOAD, VT, Ty, Legal);
608 setLoadExtAction(ISD::ZEXTLOAD, VT, Ty, Legal);
609 setLoadExtAction(ISD::SEXTLOAD, VT, Ty, Legal);
614 // ARM and Thumb2 support UMLAL/SMLAL.
615 if (!Subtarget->isThumb1Only())
616 setTargetDAGCombine(ISD::ADDC);
618 if (Subtarget->isFPOnlySP()) {
619 // When targeting a floating-point unit with only single-precision
620 // operations, f64 is legal for the few double-precision instructions which
621 // are present However, no double-precision operations other than moves,
622 // loads and stores are provided by the hardware.
623 setOperationAction(ISD::FADD, MVT::f64, Expand);
624 setOperationAction(ISD::FSUB, MVT::f64, Expand);
625 setOperationAction(ISD::FMUL, MVT::f64, Expand);
626 setOperationAction(ISD::FMA, MVT::f64, Expand);
627 setOperationAction(ISD::FDIV, MVT::f64, Expand);
628 setOperationAction(ISD::FREM, MVT::f64, Expand);
629 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
630 setOperationAction(ISD::FGETSIGN, MVT::f64, Expand);
631 setOperationAction(ISD::FNEG, MVT::f64, Expand);
632 setOperationAction(ISD::FABS, MVT::f64, Expand);
633 setOperationAction(ISD::FSQRT, MVT::f64, Expand);
634 setOperationAction(ISD::FSIN, MVT::f64, Expand);
635 setOperationAction(ISD::FCOS, MVT::f64, Expand);
636 setOperationAction(ISD::FPOWI, MVT::f64, Expand);
637 setOperationAction(ISD::FPOW, MVT::f64, Expand);
638 setOperationAction(ISD::FLOG, MVT::f64, Expand);
639 setOperationAction(ISD::FLOG2, MVT::f64, Expand);
640 setOperationAction(ISD::FLOG10, MVT::f64, Expand);
641 setOperationAction(ISD::FEXP, MVT::f64, Expand);
642 setOperationAction(ISD::FEXP2, MVT::f64, Expand);
643 setOperationAction(ISD::FCEIL, MVT::f64, Expand);
644 setOperationAction(ISD::FTRUNC, MVT::f64, Expand);
645 setOperationAction(ISD::FRINT, MVT::f64, Expand);
646 setOperationAction(ISD::FNEARBYINT, MVT::f64, Expand);
647 setOperationAction(ISD::FFLOOR, MVT::f64, Expand);
648 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
649 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
650 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
651 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
652 setOperationAction(ISD::FP_TO_SINT, MVT::f64, Custom);
653 setOperationAction(ISD::FP_TO_UINT, MVT::f64, Custom);
654 setOperationAction(ISD::FP_ROUND, MVT::f32, Custom);
655 setOperationAction(ISD::FP_EXTEND, MVT::f64, Custom);
658 computeRegisterProperties(Subtarget->getRegisterInfo());
660 // ARM does not have floating-point extending loads.
661 for (MVT VT : MVT::fp_valuetypes()) {
662 setLoadExtAction(ISD::EXTLOAD, VT, MVT::f32, Expand);
663 setLoadExtAction(ISD::EXTLOAD, VT, MVT::f16, Expand);
666 // ... or truncating stores
667 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
668 setTruncStoreAction(MVT::f32, MVT::f16, Expand);
669 setTruncStoreAction(MVT::f64, MVT::f16, Expand);
671 // ARM does not have i1 sign extending load.
672 for (MVT VT : MVT::integer_valuetypes())
673 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Promote);
675 // ARM supports all 4 flavors of integer indexed load / store.
676 if (!Subtarget->isThumb1Only()) {
677 for (unsigned im = (unsigned)ISD::PRE_INC;
678 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
679 setIndexedLoadAction(im, MVT::i1, Legal);
680 setIndexedLoadAction(im, MVT::i8, Legal);
681 setIndexedLoadAction(im, MVT::i16, Legal);
682 setIndexedLoadAction(im, MVT::i32, Legal);
683 setIndexedStoreAction(im, MVT::i1, Legal);
684 setIndexedStoreAction(im, MVT::i8, Legal);
685 setIndexedStoreAction(im, MVT::i16, Legal);
686 setIndexedStoreAction(im, MVT::i32, Legal);
690 setOperationAction(ISD::SADDO, MVT::i32, Custom);
691 setOperationAction(ISD::UADDO, MVT::i32, Custom);
692 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
693 setOperationAction(ISD::USUBO, MVT::i32, Custom);
695 // i64 operation support.
696 setOperationAction(ISD::MUL, MVT::i64, Expand);
697 setOperationAction(ISD::MULHU, MVT::i32, Expand);
698 if (Subtarget->isThumb1Only()) {
699 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
700 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
702 if (Subtarget->isThumb1Only() || !Subtarget->hasV6Ops()
703 || (Subtarget->isThumb2() && !Subtarget->hasDSP()))
704 setOperationAction(ISD::MULHS, MVT::i32, Expand);
706 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
707 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
708 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
709 setOperationAction(ISD::SRL, MVT::i64, Custom);
710 setOperationAction(ISD::SRA, MVT::i64, Custom);
712 if (!Subtarget->isThumb1Only()) {
713 // FIXME: We should do this for Thumb1 as well.
714 setOperationAction(ISD::ADDC, MVT::i32, Custom);
715 setOperationAction(ISD::ADDE, MVT::i32, Custom);
716 setOperationAction(ISD::SUBC, MVT::i32, Custom);
717 setOperationAction(ISD::SUBE, MVT::i32, Custom);
720 // ARM does not have ROTL.
721 setOperationAction(ISD::ROTL, MVT::i32, Expand);
722 for (MVT VT : MVT::vector_valuetypes()) {
723 setOperationAction(ISD::ROTL, VT, Expand);
724 setOperationAction(ISD::ROTR, VT, Expand);
726 setOperationAction(ISD::CTTZ, MVT::i32, Custom);
727 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
728 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
729 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
731 // These just redirect to CTTZ and CTLZ on ARM.
732 setOperationAction(ISD::CTTZ_ZERO_UNDEF , MVT::i32 , Expand);
733 setOperationAction(ISD::CTLZ_ZERO_UNDEF , MVT::i32 , Expand);
735 // @llvm.readcyclecounter requires the Performance Monitors extension.
736 // Default to the 0 expansion on unsupported platforms.
737 // FIXME: Technically there are older ARM CPUs that have
738 // implementation-specific ways of obtaining this information.
739 if (Subtarget->hasPerfMon())
740 setOperationAction(ISD::READCYCLECOUNTER, MVT::i64, Custom);
742 // Only ARMv6 has BSWAP.
743 if (!Subtarget->hasV6Ops())
744 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
746 if (!(Subtarget->hasDivide() && Subtarget->isThumb2()) &&
747 !(Subtarget->hasDivideInARMMode() && !Subtarget->isThumb())) {
748 // These are expanded into libcalls if the cpu doesn't have HW divider.
749 setOperationAction(ISD::SDIV, MVT::i32, LibCall);
750 setOperationAction(ISD::UDIV, MVT::i32, LibCall);
753 if (Subtarget->isTargetWindows() && !Subtarget->hasDivide()) {
754 setOperationAction(ISD::SDIV, MVT::i32, Custom);
755 setOperationAction(ISD::UDIV, MVT::i32, Custom);
757 setOperationAction(ISD::SDIV, MVT::i64, Custom);
758 setOperationAction(ISD::UDIV, MVT::i64, Custom);
761 setOperationAction(ISD::SREM, MVT::i32, Expand);
762 setOperationAction(ISD::UREM, MVT::i32, Expand);
763 // Register based DivRem for AEABI (RTABI 4.2)
764 if (Subtarget->isTargetAEABI() || Subtarget->isTargetAndroid()) {
765 setOperationAction(ISD::SREM, MVT::i64, Custom);
766 setOperationAction(ISD::UREM, MVT::i64, Custom);
768 setLibcallName(RTLIB::SDIVREM_I8, "__aeabi_idivmod");
769 setLibcallName(RTLIB::SDIVREM_I16, "__aeabi_idivmod");
770 setLibcallName(RTLIB::SDIVREM_I32, "__aeabi_idivmod");
771 setLibcallName(RTLIB::SDIVREM_I64, "__aeabi_ldivmod");
772 setLibcallName(RTLIB::UDIVREM_I8, "__aeabi_uidivmod");
773 setLibcallName(RTLIB::UDIVREM_I16, "__aeabi_uidivmod");
774 setLibcallName(RTLIB::UDIVREM_I32, "__aeabi_uidivmod");
775 setLibcallName(RTLIB::UDIVREM_I64, "__aeabi_uldivmod");
777 setLibcallCallingConv(RTLIB::SDIVREM_I8, CallingConv::ARM_AAPCS);
778 setLibcallCallingConv(RTLIB::SDIVREM_I16, CallingConv::ARM_AAPCS);
779 setLibcallCallingConv(RTLIB::SDIVREM_I32, CallingConv::ARM_AAPCS);
780 setLibcallCallingConv(RTLIB::SDIVREM_I64, CallingConv::ARM_AAPCS);
781 setLibcallCallingConv(RTLIB::UDIVREM_I8, CallingConv::ARM_AAPCS);
782 setLibcallCallingConv(RTLIB::UDIVREM_I16, CallingConv::ARM_AAPCS);
783 setLibcallCallingConv(RTLIB::UDIVREM_I32, CallingConv::ARM_AAPCS);
784 setLibcallCallingConv(RTLIB::UDIVREM_I64, CallingConv::ARM_AAPCS);
786 setOperationAction(ISD::SDIVREM, MVT::i32, Custom);
787 setOperationAction(ISD::UDIVREM, MVT::i32, Custom);
789 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
790 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
793 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
794 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
795 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
796 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
798 setOperationAction(ISD::TRAP, MVT::Other, Legal);
800 // Use the default implementation.
801 setOperationAction(ISD::VASTART, MVT::Other, Custom);
802 setOperationAction(ISD::VAARG, MVT::Other, Expand);
803 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
804 setOperationAction(ISD::VAEND, MVT::Other, Expand);
805 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
806 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
808 if (!Subtarget->isTargetMachO()) {
809 // Non-MachO platforms may return values in these registers via the
810 // personality function.
811 setExceptionPointerRegister(ARM::R0);
812 setExceptionSelectorRegister(ARM::R1);
815 if (Subtarget->getTargetTriple().isWindowsItaniumEnvironment())
816 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
818 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
820 // ARMv6 Thumb1 (except for CPUs that support dmb / dsb) and earlier use
821 // the default expansion. If we are targeting a single threaded system,
822 // then set them all for expand so we can lower them later into their
824 if (TM.Options.ThreadModel == ThreadModel::Single)
825 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Expand);
826 else if (Subtarget->hasAnyDataBarrier() && !Subtarget->isThumb1Only()) {
827 // ATOMIC_FENCE needs custom lowering; the others should have been expanded
828 // to ldrex/strex loops already.
829 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Custom);
831 // On v8, we have particularly efficient implementations of atomic fences
832 // if they can be combined with nearby atomic loads and stores.
833 if (!Subtarget->hasV8Ops()) {
834 // Automatically insert fences (dmb ish) around ATOMIC_SWAP etc.
835 setInsertFencesForAtomic(true);
838 // If there's anything we can use as a barrier, go through custom lowering
840 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other,
841 Subtarget->hasAnyDataBarrier() ? Custom : Expand);
843 // Set them all for expansion, which will force libcalls.
844 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Expand);
845 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Expand);
846 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Expand);
847 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Expand);
848 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Expand);
849 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Expand);
850 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Expand);
851 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Expand);
852 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i32, Expand);
853 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i32, Expand);
854 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i32, Expand);
855 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i32, Expand);
856 // Mark ATOMIC_LOAD and ATOMIC_STORE custom so we can handle the
857 // Unordered/Monotonic case.
858 setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Custom);
859 setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Custom);
862 setOperationAction(ISD::PREFETCH, MVT::Other, Custom);
864 // Requires SXTB/SXTH, available on v6 and up in both ARM and Thumb modes.
865 if (!Subtarget->hasV6Ops()) {
866 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
867 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
869 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
871 if (!Subtarget->useSoftFloat() && Subtarget->hasVFP2() &&
872 !Subtarget->isThumb1Only()) {
873 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
874 // iff target supports vfp2.
875 setOperationAction(ISD::BITCAST, MVT::i64, Custom);
876 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
879 // We want to custom lower some of our intrinsics.
880 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
881 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
882 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
883 setOperationAction(ISD::EH_SJLJ_SETUP_DISPATCH, MVT::Other, Custom);
884 if (Subtarget->isTargetDarwin())
885 setLibcallName(RTLIB::UNWIND_RESUME, "_Unwind_SjLj_Resume");
887 setOperationAction(ISD::SETCC, MVT::i32, Expand);
888 setOperationAction(ISD::SETCC, MVT::f32, Expand);
889 setOperationAction(ISD::SETCC, MVT::f64, Expand);
890 setOperationAction(ISD::SELECT, MVT::i32, Custom);
891 setOperationAction(ISD::SELECT, MVT::f32, Custom);
892 setOperationAction(ISD::SELECT, MVT::f64, Custom);
893 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
894 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
895 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
897 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
898 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
899 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
900 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
901 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
903 // We don't support sin/cos/fmod/copysign/pow
904 setOperationAction(ISD::FSIN, MVT::f64, Expand);
905 setOperationAction(ISD::FSIN, MVT::f32, Expand);
906 setOperationAction(ISD::FCOS, MVT::f32, Expand);
907 setOperationAction(ISD::FCOS, MVT::f64, Expand);
908 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
909 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
910 setOperationAction(ISD::FREM, MVT::f64, Expand);
911 setOperationAction(ISD::FREM, MVT::f32, Expand);
912 if (!Subtarget->useSoftFloat() && Subtarget->hasVFP2() &&
913 !Subtarget->isThumb1Only()) {
914 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
915 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
917 setOperationAction(ISD::FPOW, MVT::f64, Expand);
918 setOperationAction(ISD::FPOW, MVT::f32, Expand);
920 if (!Subtarget->hasVFP4()) {
921 setOperationAction(ISD::FMA, MVT::f64, Expand);
922 setOperationAction(ISD::FMA, MVT::f32, Expand);
925 // Various VFP goodness
926 if (!Subtarget->useSoftFloat() && !Subtarget->isThumb1Only()) {
927 // FP-ARMv8 adds f64 <-> f16 conversion. Before that it should be expanded.
928 if (!Subtarget->hasFPARMv8() || Subtarget->isFPOnlySP()) {
929 setOperationAction(ISD::FP16_TO_FP, MVT::f64, Expand);
930 setOperationAction(ISD::FP_TO_FP16, MVT::f64, Expand);
933 // fp16 is a special v7 extension that adds f16 <-> f32 conversions.
934 if (!Subtarget->hasFP16()) {
935 setOperationAction(ISD::FP16_TO_FP, MVT::f32, Expand);
936 setOperationAction(ISD::FP_TO_FP16, MVT::f32, Expand);
940 // Combine sin / cos into one node or libcall if possible.
941 if (Subtarget->hasSinCos()) {
942 setLibcallName(RTLIB::SINCOS_F32, "sincosf");
943 setLibcallName(RTLIB::SINCOS_F64, "sincos");
944 if (Subtarget->getTargetTriple().isiOS()) {
945 // For iOS, we don't want to the normal expansion of a libcall to
946 // sincos. We want to issue a libcall to __sincos_stret.
947 setOperationAction(ISD::FSINCOS, MVT::f64, Custom);
948 setOperationAction(ISD::FSINCOS, MVT::f32, Custom);
952 // FP-ARMv8 implements a lot of rounding-like FP operations.
953 if (Subtarget->hasFPARMv8()) {
954 setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
955 setOperationAction(ISD::FCEIL, MVT::f32, Legal);
956 setOperationAction(ISD::FROUND, MVT::f32, Legal);
957 setOperationAction(ISD::FTRUNC, MVT::f32, Legal);
958 setOperationAction(ISD::FNEARBYINT, MVT::f32, Legal);
959 setOperationAction(ISD::FRINT, MVT::f32, Legal);
960 setOperationAction(ISD::FMINNUM, MVT::f32, Legal);
961 setOperationAction(ISD::FMAXNUM, MVT::f32, Legal);
962 setOperationAction(ISD::FMINNUM, MVT::v2f32, Legal);
963 setOperationAction(ISD::FMAXNUM, MVT::v2f32, Legal);
964 setOperationAction(ISD::FMINNUM, MVT::v4f32, Legal);
965 setOperationAction(ISD::FMAXNUM, MVT::v4f32, Legal);
967 if (!Subtarget->isFPOnlySP()) {
968 setOperationAction(ISD::FFLOOR, MVT::f64, Legal);
969 setOperationAction(ISD::FCEIL, MVT::f64, Legal);
970 setOperationAction(ISD::FROUND, MVT::f64, Legal);
971 setOperationAction(ISD::FTRUNC, MVT::f64, Legal);
972 setOperationAction(ISD::FNEARBYINT, MVT::f64, Legal);
973 setOperationAction(ISD::FRINT, MVT::f64, Legal);
974 setOperationAction(ISD::FMINNUM, MVT::f64, Legal);
975 setOperationAction(ISD::FMAXNUM, MVT::f64, Legal);
979 if (Subtarget->hasNEON()) {
980 // vmin and vmax aren't available in a scalar form, so we use
981 // a NEON instruction with an undef lane instead.
982 setOperationAction(ISD::FMINNAN, MVT::f32, Legal);
983 setOperationAction(ISD::FMAXNAN, MVT::f32, Legal);
984 setOperationAction(ISD::FMINNAN, MVT::v2f32, Legal);
985 setOperationAction(ISD::FMAXNAN, MVT::v2f32, Legal);
986 setOperationAction(ISD::FMINNAN, MVT::v4f32, Legal);
987 setOperationAction(ISD::FMAXNAN, MVT::v4f32, Legal);
990 // We have target-specific dag combine patterns for the following nodes:
991 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
992 setTargetDAGCombine(ISD::ADD);
993 setTargetDAGCombine(ISD::SUB);
994 setTargetDAGCombine(ISD::MUL);
995 setTargetDAGCombine(ISD::AND);
996 setTargetDAGCombine(ISD::OR);
997 setTargetDAGCombine(ISD::XOR);
999 if (Subtarget->hasV6Ops())
1000 setTargetDAGCombine(ISD::SRL);
1002 setStackPointerRegisterToSaveRestore(ARM::SP);
1004 if (Subtarget->useSoftFloat() || Subtarget->isThumb1Only() ||
1005 !Subtarget->hasVFP2())
1006 setSchedulingPreference(Sched::RegPressure);
1008 setSchedulingPreference(Sched::Hybrid);
1010 //// temporary - rewrite interface to use type
1011 MaxStoresPerMemset = 8;
1012 MaxStoresPerMemsetOptSize = 4;
1013 MaxStoresPerMemcpy = 4; // For @llvm.memcpy -> sequence of stores
1014 MaxStoresPerMemcpyOptSize = 2;
1015 MaxStoresPerMemmove = 4; // For @llvm.memmove -> sequence of stores
1016 MaxStoresPerMemmoveOptSize = 2;
1018 // On ARM arguments smaller than 4 bytes are extended, so all arguments
1019 // are at least 4 bytes aligned.
1020 setMinStackArgumentAlignment(4);
1022 // Prefer likely predicted branches to selects on out-of-order cores.
1023 PredictableSelectIsExpensive = Subtarget->isLikeA9();
1025 setMinFunctionAlignment(Subtarget->isThumb() ? 1 : 2);
1028 bool ARMTargetLowering::useSoftFloat() const {
1029 return Subtarget->useSoftFloat();
1032 // FIXME: It might make sense to define the representative register class as the
1033 // nearest super-register that has a non-null superset. For example, DPR_VFP2 is
1034 // a super-register of SPR, and DPR is a superset if DPR_VFP2. Consequently,
1035 // SPR's representative would be DPR_VFP2. This should work well if register
1036 // pressure tracking were modified such that a register use would increment the
1037 // pressure of the register class's representative and all of it's super
1038 // classes' representatives transitively. We have not implemented this because
1039 // of the difficulty prior to coalescing of modeling operand register classes
1040 // due to the common occurrence of cross class copies and subregister insertions
1042 std::pair<const TargetRegisterClass *, uint8_t>
1043 ARMTargetLowering::findRepresentativeClass(const TargetRegisterInfo *TRI,
1045 const TargetRegisterClass *RRC = nullptr;
1047 switch (VT.SimpleTy) {
1049 return TargetLowering::findRepresentativeClass(TRI, VT);
1050 // Use DPR as representative register class for all floating point
1051 // and vector types. Since there are 32 SPR registers and 32 DPR registers so
1052 // the cost is 1 for both f32 and f64.
1053 case MVT::f32: case MVT::f64: case MVT::v8i8: case MVT::v4i16:
1054 case MVT::v2i32: case MVT::v1i64: case MVT::v2f32:
1055 RRC = &ARM::DPRRegClass;
1056 // When NEON is used for SP, only half of the register file is available
1057 // because operations that define both SP and DP results will be constrained
1058 // to the VFP2 class (D0-D15). We currently model this constraint prior to
1059 // coalescing by double-counting the SP regs. See the FIXME above.
1060 if (Subtarget->useNEONForSinglePrecisionFP())
1063 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
1064 case MVT::v4f32: case MVT::v2f64:
1065 RRC = &ARM::DPRRegClass;
1069 RRC = &ARM::DPRRegClass;
1073 RRC = &ARM::DPRRegClass;
1077 return std::make_pair(RRC, Cost);
1080 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
1081 switch ((ARMISD::NodeType)Opcode) {
1082 case ARMISD::FIRST_NUMBER: break;
1083 case ARMISD::Wrapper: return "ARMISD::Wrapper";
1084 case ARMISD::WrapperPIC: return "ARMISD::WrapperPIC";
1085 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
1086 case ARMISD::COPY_STRUCT_BYVAL: return "ARMISD::COPY_STRUCT_BYVAL";
1087 case ARMISD::CALL: return "ARMISD::CALL";
1088 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
1089 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
1090 case ARMISD::tCALL: return "ARMISD::tCALL";
1091 case ARMISD::BRCOND: return "ARMISD::BRCOND";
1092 case ARMISD::BR_JT: return "ARMISD::BR_JT";
1093 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
1094 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
1095 case ARMISD::INTRET_FLAG: return "ARMISD::INTRET_FLAG";
1096 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
1097 case ARMISD::CMP: return "ARMISD::CMP";
1098 case ARMISD::CMN: return "ARMISD::CMN";
1099 case ARMISD::CMPZ: return "ARMISD::CMPZ";
1100 case ARMISD::CMPFP: return "ARMISD::CMPFP";
1101 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
1102 case ARMISD::BCC_i64: return "ARMISD::BCC_i64";
1103 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
1105 case ARMISD::CMOV: return "ARMISD::CMOV";
1107 case ARMISD::RBIT: return "ARMISD::RBIT";
1109 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
1110 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
1111 case ARMISD::RRX: return "ARMISD::RRX";
1113 case ARMISD::ADDC: return "ARMISD::ADDC";
1114 case ARMISD::ADDE: return "ARMISD::ADDE";
1115 case ARMISD::SUBC: return "ARMISD::SUBC";
1116 case ARMISD::SUBE: return "ARMISD::SUBE";
1118 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
1119 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
1121 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
1122 case ARMISD::EH_SJLJ_LONGJMP: return "ARMISD::EH_SJLJ_LONGJMP";
1123 case ARMISD::EH_SJLJ_SETUP_DISPATCH: return "ARMISD::EH_SJLJ_SETUP_DISPATCH";
1125 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN";
1127 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
1129 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
1131 case ARMISD::MEMBARRIER_MCR: return "ARMISD::MEMBARRIER_MCR";
1133 case ARMISD::PRELOAD: return "ARMISD::PRELOAD";
1135 case ARMISD::WIN__CHKSTK: return "ARMISD:::WIN__CHKSTK";
1136 case ARMISD::WIN__DBZCHK: return "ARMISD::WIN__DBZCHK";
1138 case ARMISD::VCEQ: return "ARMISD::VCEQ";
1139 case ARMISD::VCEQZ: return "ARMISD::VCEQZ";
1140 case ARMISD::VCGE: return "ARMISD::VCGE";
1141 case ARMISD::VCGEZ: return "ARMISD::VCGEZ";
1142 case ARMISD::VCLEZ: return "ARMISD::VCLEZ";
1143 case ARMISD::VCGEU: return "ARMISD::VCGEU";
1144 case ARMISD::VCGT: return "ARMISD::VCGT";
1145 case ARMISD::VCGTZ: return "ARMISD::VCGTZ";
1146 case ARMISD::VCLTZ: return "ARMISD::VCLTZ";
1147 case ARMISD::VCGTU: return "ARMISD::VCGTU";
1148 case ARMISD::VTST: return "ARMISD::VTST";
1150 case ARMISD::VSHL: return "ARMISD::VSHL";
1151 case ARMISD::VSHRs: return "ARMISD::VSHRs";
1152 case ARMISD::VSHRu: return "ARMISD::VSHRu";
1153 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
1154 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
1155 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
1156 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
1157 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
1158 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
1159 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
1160 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
1161 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
1162 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
1163 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
1164 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
1165 case ARMISD::VSLI: return "ARMISD::VSLI";
1166 case ARMISD::VSRI: return "ARMISD::VSRI";
1167 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
1168 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
1169 case ARMISD::VMOVIMM: return "ARMISD::VMOVIMM";
1170 case ARMISD::VMVNIMM: return "ARMISD::VMVNIMM";
1171 case ARMISD::VMOVFPIMM: return "ARMISD::VMOVFPIMM";
1172 case ARMISD::VDUP: return "ARMISD::VDUP";
1173 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
1174 case ARMISD::VEXT: return "ARMISD::VEXT";
1175 case ARMISD::VREV64: return "ARMISD::VREV64";
1176 case ARMISD::VREV32: return "ARMISD::VREV32";
1177 case ARMISD::VREV16: return "ARMISD::VREV16";
1178 case ARMISD::VZIP: return "ARMISD::VZIP";
1179 case ARMISD::VUZP: return "ARMISD::VUZP";
1180 case ARMISD::VTRN: return "ARMISD::VTRN";
1181 case ARMISD::VTBL1: return "ARMISD::VTBL1";
1182 case ARMISD::VTBL2: return "ARMISD::VTBL2";
1183 case ARMISD::VMULLs: return "ARMISD::VMULLs";
1184 case ARMISD::VMULLu: return "ARMISD::VMULLu";
1185 case ARMISD::UMLAL: return "ARMISD::UMLAL";
1186 case ARMISD::SMLAL: return "ARMISD::SMLAL";
1187 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
1188 case ARMISD::BFI: return "ARMISD::BFI";
1189 case ARMISD::VORRIMM: return "ARMISD::VORRIMM";
1190 case ARMISD::VBICIMM: return "ARMISD::VBICIMM";
1191 case ARMISD::VBSL: return "ARMISD::VBSL";
1192 case ARMISD::MEMCPY: return "ARMISD::MEMCPY";
1193 case ARMISD::VLD2DUP: return "ARMISD::VLD2DUP";
1194 case ARMISD::VLD3DUP: return "ARMISD::VLD3DUP";
1195 case ARMISD::VLD4DUP: return "ARMISD::VLD4DUP";
1196 case ARMISD::VLD1_UPD: return "ARMISD::VLD1_UPD";
1197 case ARMISD::VLD2_UPD: return "ARMISD::VLD2_UPD";
1198 case ARMISD::VLD3_UPD: return "ARMISD::VLD3_UPD";
1199 case ARMISD::VLD4_UPD: return "ARMISD::VLD4_UPD";
1200 case ARMISD::VLD2LN_UPD: return "ARMISD::VLD2LN_UPD";
1201 case ARMISD::VLD3LN_UPD: return "ARMISD::VLD3LN_UPD";
1202 case ARMISD::VLD4LN_UPD: return "ARMISD::VLD4LN_UPD";
1203 case ARMISD::VLD2DUP_UPD: return "ARMISD::VLD2DUP_UPD";
1204 case ARMISD::VLD3DUP_UPD: return "ARMISD::VLD3DUP_UPD";
1205 case ARMISD::VLD4DUP_UPD: return "ARMISD::VLD4DUP_UPD";
1206 case ARMISD::VST1_UPD: return "ARMISD::VST1_UPD";
1207 case ARMISD::VST2_UPD: return "ARMISD::VST2_UPD";
1208 case ARMISD::VST3_UPD: return "ARMISD::VST3_UPD";
1209 case ARMISD::VST4_UPD: return "ARMISD::VST4_UPD";
1210 case ARMISD::VST2LN_UPD: return "ARMISD::VST2LN_UPD";
1211 case ARMISD::VST3LN_UPD: return "ARMISD::VST3LN_UPD";
1212 case ARMISD::VST4LN_UPD: return "ARMISD::VST4LN_UPD";
1217 EVT ARMTargetLowering::getSetCCResultType(const DataLayout &DL, LLVMContext &,
1220 return getPointerTy(DL);
1221 return VT.changeVectorElementTypeToInteger();
1224 /// getRegClassFor - Return the register class that should be used for the
1225 /// specified value type.
1226 const TargetRegisterClass *ARMTargetLowering::getRegClassFor(MVT VT) const {
1227 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
1228 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
1229 // load / store 4 to 8 consecutive D registers.
1230 if (Subtarget->hasNEON()) {
1231 if (VT == MVT::v4i64)
1232 return &ARM::QQPRRegClass;
1233 if (VT == MVT::v8i64)
1234 return &ARM::QQQQPRRegClass;
1236 return TargetLowering::getRegClassFor(VT);
1239 // memcpy, and other memory intrinsics, typically tries to use LDM/STM if the
1240 // source/dest is aligned and the copy size is large enough. We therefore want
1241 // to align such objects passed to memory intrinsics.
1242 bool ARMTargetLowering::shouldAlignPointerArgs(CallInst *CI, unsigned &MinSize,
1243 unsigned &PrefAlign) const {
1244 if (!isa<MemIntrinsic>(CI))
1247 // On ARM11 onwards (excluding M class) 8-byte aligned LDM is typically 1
1248 // cycle faster than 4-byte aligned LDM.
1249 PrefAlign = (Subtarget->hasV6Ops() && !Subtarget->isMClass() ? 8 : 4);
1253 // Create a fast isel object.
1255 ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
1256 const TargetLibraryInfo *libInfo) const {
1257 return ARM::createFastISel(funcInfo, libInfo);
1260 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
1261 unsigned NumVals = N->getNumValues();
1263 return Sched::RegPressure;
1265 for (unsigned i = 0; i != NumVals; ++i) {
1266 EVT VT = N->getValueType(i);
1267 if (VT == MVT::Glue || VT == MVT::Other)
1269 if (VT.isFloatingPoint() || VT.isVector())
1273 if (!N->isMachineOpcode())
1274 return Sched::RegPressure;
1276 // Load are scheduled for latency even if there instruction itinerary
1277 // is not available.
1278 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
1279 const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
1281 if (MCID.getNumDefs() == 0)
1282 return Sched::RegPressure;
1283 if (!Itins->isEmpty() &&
1284 Itins->getOperandCycle(MCID.getSchedClass(), 0) > 2)
1287 return Sched::RegPressure;
1290 //===----------------------------------------------------------------------===//
1292 //===----------------------------------------------------------------------===//
1294 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
1295 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
1297 default: llvm_unreachable("Unknown condition code!");
1298 case ISD::SETNE: return ARMCC::NE;
1299 case ISD::SETEQ: return ARMCC::EQ;
1300 case ISD::SETGT: return ARMCC::GT;
1301 case ISD::SETGE: return ARMCC::GE;
1302 case ISD::SETLT: return ARMCC::LT;
1303 case ISD::SETLE: return ARMCC::LE;
1304 case ISD::SETUGT: return ARMCC::HI;
1305 case ISD::SETUGE: return ARMCC::HS;
1306 case ISD::SETULT: return ARMCC::LO;
1307 case ISD::SETULE: return ARMCC::LS;
1311 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
1312 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
1313 ARMCC::CondCodes &CondCode2) {
1314 CondCode2 = ARMCC::AL;
1316 default: llvm_unreachable("Unknown FP condition!");
1318 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
1320 case ISD::SETOGT: CondCode = ARMCC::GT; break;
1322 case ISD::SETOGE: CondCode = ARMCC::GE; break;
1323 case ISD::SETOLT: CondCode = ARMCC::MI; break;
1324 case ISD::SETOLE: CondCode = ARMCC::LS; break;
1325 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
1326 case ISD::SETO: CondCode = ARMCC::VC; break;
1327 case ISD::SETUO: CondCode = ARMCC::VS; break;
1328 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
1329 case ISD::SETUGT: CondCode = ARMCC::HI; break;
1330 case ISD::SETUGE: CondCode = ARMCC::PL; break;
1332 case ISD::SETULT: CondCode = ARMCC::LT; break;
1334 case ISD::SETULE: CondCode = ARMCC::LE; break;
1336 case ISD::SETUNE: CondCode = ARMCC::NE; break;
1340 //===----------------------------------------------------------------------===//
1341 // Calling Convention Implementation
1342 //===----------------------------------------------------------------------===//
1344 #include "ARMGenCallingConv.inc"
1346 /// getEffectiveCallingConv - Get the effective calling convention, taking into
1347 /// account presence of floating point hardware and calling convention
1348 /// limitations, such as support for variadic functions.
1350 ARMTargetLowering::getEffectiveCallingConv(CallingConv::ID CC,
1351 bool isVarArg) const {
1354 llvm_unreachable("Unsupported calling convention");
1355 case CallingConv::ARM_AAPCS:
1356 case CallingConv::ARM_APCS:
1357 case CallingConv::GHC:
1359 case CallingConv::ARM_AAPCS_VFP:
1360 return isVarArg ? CallingConv::ARM_AAPCS : CallingConv::ARM_AAPCS_VFP;
1361 case CallingConv::C:
1362 if (!Subtarget->isAAPCS_ABI())
1363 return CallingConv::ARM_APCS;
1364 else if (Subtarget->hasVFP2() && !Subtarget->isThumb1Only() &&
1365 getTargetMachine().Options.FloatABIType == FloatABI::Hard &&
1367 return CallingConv::ARM_AAPCS_VFP;
1369 return CallingConv::ARM_AAPCS;
1370 case CallingConv::Fast:
1371 if (!Subtarget->isAAPCS_ABI()) {
1372 if (Subtarget->hasVFP2() && !Subtarget->isThumb1Only() && !isVarArg)
1373 return CallingConv::Fast;
1374 return CallingConv::ARM_APCS;
1375 } else if (Subtarget->hasVFP2() && !Subtarget->isThumb1Only() && !isVarArg)
1376 return CallingConv::ARM_AAPCS_VFP;
1378 return CallingConv::ARM_AAPCS;
1382 /// CCAssignFnForNode - Selects the correct CCAssignFn for the given
1383 /// CallingConvention.
1384 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
1386 bool isVarArg) const {
1387 switch (getEffectiveCallingConv(CC, isVarArg)) {
1389 llvm_unreachable("Unsupported calling convention");
1390 case CallingConv::ARM_APCS:
1391 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1392 case CallingConv::ARM_AAPCS:
1393 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1394 case CallingConv::ARM_AAPCS_VFP:
1395 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1396 case CallingConv::Fast:
1397 return (Return ? RetFastCC_ARM_APCS : FastCC_ARM_APCS);
1398 case CallingConv::GHC:
1399 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS_GHC);
1403 /// LowerCallResult - Lower the result values of a call into the
1404 /// appropriate copies out of appropriate physical registers.
1406 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1407 CallingConv::ID CallConv, bool isVarArg,
1408 const SmallVectorImpl<ISD::InputArg> &Ins,
1409 SDLoc dl, SelectionDAG &DAG,
1410 SmallVectorImpl<SDValue> &InVals,
1411 bool isThisReturn, SDValue ThisVal) const {
1413 // Assign locations to each value returned by this call.
1414 SmallVector<CCValAssign, 16> RVLocs;
1415 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
1416 *DAG.getContext(), Call);
1417 CCInfo.AnalyzeCallResult(Ins,
1418 CCAssignFnForNode(CallConv, /* Return*/ true,
1421 // Copy all of the result registers out of their specified physreg.
1422 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1423 CCValAssign VA = RVLocs[i];
1425 // Pass 'this' value directly from the argument to return value, to avoid
1426 // reg unit interference
1427 if (i == 0 && isThisReturn) {
1428 assert(!VA.needsCustom() && VA.getLocVT() == MVT::i32 &&
1429 "unexpected return calling convention register assignment");
1430 InVals.push_back(ThisVal);
1435 if (VA.needsCustom()) {
1436 // Handle f64 or half of a v2f64.
1437 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1439 Chain = Lo.getValue(1);
1440 InFlag = Lo.getValue(2);
1441 VA = RVLocs[++i]; // skip ahead to next loc
1442 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1444 Chain = Hi.getValue(1);
1445 InFlag = Hi.getValue(2);
1446 if (!Subtarget->isLittle())
1448 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1450 if (VA.getLocVT() == MVT::v2f64) {
1451 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1452 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1453 DAG.getConstant(0, dl, MVT::i32));
1455 VA = RVLocs[++i]; // skip ahead to next loc
1456 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1457 Chain = Lo.getValue(1);
1458 InFlag = Lo.getValue(2);
1459 VA = RVLocs[++i]; // skip ahead to next loc
1460 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1461 Chain = Hi.getValue(1);
1462 InFlag = Hi.getValue(2);
1463 if (!Subtarget->isLittle())
1465 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1466 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1467 DAG.getConstant(1, dl, MVT::i32));
1470 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
1472 Chain = Val.getValue(1);
1473 InFlag = Val.getValue(2);
1476 switch (VA.getLocInfo()) {
1477 default: llvm_unreachable("Unknown loc info!");
1478 case CCValAssign::Full: break;
1479 case CCValAssign::BCvt:
1480 Val = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), Val);
1484 InVals.push_back(Val);
1490 /// LowerMemOpCallTo - Store the argument to the stack.
1492 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
1493 SDValue StackPtr, SDValue Arg,
1494 SDLoc dl, SelectionDAG &DAG,
1495 const CCValAssign &VA,
1496 ISD::ArgFlagsTy Flags) const {
1497 unsigned LocMemOffset = VA.getLocMemOffset();
1498 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset, dl);
1499 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(DAG.getDataLayout()),
1501 return DAG.getStore(
1502 Chain, dl, Arg, PtrOff,
1503 MachinePointerInfo::getStack(DAG.getMachineFunction(), LocMemOffset),
1507 void ARMTargetLowering::PassF64ArgInRegs(SDLoc dl, SelectionDAG &DAG,
1508 SDValue Chain, SDValue &Arg,
1509 RegsToPassVector &RegsToPass,
1510 CCValAssign &VA, CCValAssign &NextVA,
1512 SmallVectorImpl<SDValue> &MemOpChains,
1513 ISD::ArgFlagsTy Flags) const {
1515 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1516 DAG.getVTList(MVT::i32, MVT::i32), Arg);
1517 unsigned id = Subtarget->isLittle() ? 0 : 1;
1518 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd.getValue(id)));
1520 if (NextVA.isRegLoc())
1521 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1-id)));
1523 assert(NextVA.isMemLoc());
1524 if (!StackPtr.getNode())
1525 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP,
1526 getPointerTy(DAG.getDataLayout()));
1528 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1-id),
1534 /// LowerCall - Lowering a call into a callseq_start <-
1535 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
1538 ARMTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
1539 SmallVectorImpl<SDValue> &InVals) const {
1540 SelectionDAG &DAG = CLI.DAG;
1542 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
1543 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
1544 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
1545 SDValue Chain = CLI.Chain;
1546 SDValue Callee = CLI.Callee;
1547 bool &isTailCall = CLI.IsTailCall;
1548 CallingConv::ID CallConv = CLI.CallConv;
1549 bool doesNotRet = CLI.DoesNotReturn;
1550 bool isVarArg = CLI.IsVarArg;
1552 MachineFunction &MF = DAG.getMachineFunction();
1553 bool isStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1554 bool isThisReturn = false;
1555 bool isSibCall = false;
1556 auto Attr = MF.getFunction()->getFnAttribute("disable-tail-calls");
1558 // Disable tail calls if they're not supported.
1559 if (!Subtarget->supportsTailCall() || Attr.getValueAsString() == "true")
1563 // Check if it's really possible to do a tail call.
1564 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1565 isVarArg, isStructRet, MF.getFunction()->hasStructRetAttr(),
1566 Outs, OutVals, Ins, DAG);
1567 if (!isTailCall && CLI.CS && CLI.CS->isMustTailCall())
1568 report_fatal_error("failed to perform tail call elimination on a call "
1569 "site marked musttail");
1570 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1571 // detected sibcalls.
1578 // Analyze operands of the call, assigning locations to each operand.
1579 SmallVector<CCValAssign, 16> ArgLocs;
1580 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs,
1581 *DAG.getContext(), Call);
1582 CCInfo.AnalyzeCallOperands(Outs,
1583 CCAssignFnForNode(CallConv, /* Return*/ false,
1586 // Get a count of how many bytes are to be pushed on the stack.
1587 unsigned NumBytes = CCInfo.getNextStackOffset();
1589 // For tail calls, memory operands are available in our caller's stack.
1593 // Adjust the stack pointer for the new arguments...
1594 // These operations are automatically eliminated by the prolog/epilog pass
1596 Chain = DAG.getCALLSEQ_START(Chain,
1597 DAG.getIntPtrConstant(NumBytes, dl, true), dl);
1600 DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy(DAG.getDataLayout()));
1602 RegsToPassVector RegsToPass;
1603 SmallVector<SDValue, 8> MemOpChains;
1605 // Walk the register/memloc assignments, inserting copies/loads. In the case
1606 // of tail call optimization, arguments are handled later.
1607 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1609 ++i, ++realArgIdx) {
1610 CCValAssign &VA = ArgLocs[i];
1611 SDValue Arg = OutVals[realArgIdx];
1612 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1613 bool isByVal = Flags.isByVal();
1615 // Promote the value if needed.
1616 switch (VA.getLocInfo()) {
1617 default: llvm_unreachable("Unknown loc info!");
1618 case CCValAssign::Full: break;
1619 case CCValAssign::SExt:
1620 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1622 case CCValAssign::ZExt:
1623 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1625 case CCValAssign::AExt:
1626 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1628 case CCValAssign::BCvt:
1629 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1633 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1634 if (VA.needsCustom()) {
1635 if (VA.getLocVT() == MVT::v2f64) {
1636 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1637 DAG.getConstant(0, dl, MVT::i32));
1638 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1639 DAG.getConstant(1, dl, MVT::i32));
1641 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1642 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1644 VA = ArgLocs[++i]; // skip ahead to next loc
1645 if (VA.isRegLoc()) {
1646 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1647 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1649 assert(VA.isMemLoc());
1651 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1652 dl, DAG, VA, Flags));
1655 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1656 StackPtr, MemOpChains, Flags);
1658 } else if (VA.isRegLoc()) {
1659 if (realArgIdx == 0 && Flags.isReturned() && Outs[0].VT == MVT::i32) {
1660 assert(VA.getLocVT() == MVT::i32 &&
1661 "unexpected calling convention register assignment");
1662 assert(!Ins.empty() && Ins[0].VT == MVT::i32 &&
1663 "unexpected use of 'returned'");
1664 isThisReturn = true;
1666 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1667 } else if (isByVal) {
1668 assert(VA.isMemLoc());
1669 unsigned offset = 0;
1671 // True if this byval aggregate will be split between registers
1673 unsigned ByValArgsCount = CCInfo.getInRegsParamsCount();
1674 unsigned CurByValIdx = CCInfo.getInRegsParamsProcessed();
1676 if (CurByValIdx < ByValArgsCount) {
1678 unsigned RegBegin, RegEnd;
1679 CCInfo.getInRegsParamInfo(CurByValIdx, RegBegin, RegEnd);
1682 DAG.getTargetLoweringInfo().getPointerTy(DAG.getDataLayout());
1684 for (i = 0, j = RegBegin; j < RegEnd; i++, j++) {
1685 SDValue Const = DAG.getConstant(4*i, dl, MVT::i32);
1686 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
1687 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
1688 MachinePointerInfo(),
1689 false, false, false,
1690 DAG.InferPtrAlignment(AddArg));
1691 MemOpChains.push_back(Load.getValue(1));
1692 RegsToPass.push_back(std::make_pair(j, Load));
1695 // If parameter size outsides register area, "offset" value
1696 // helps us to calculate stack slot for remained part properly.
1697 offset = RegEnd - RegBegin;
1699 CCInfo.nextInRegsParam();
1702 if (Flags.getByValSize() > 4*offset) {
1703 auto PtrVT = getPointerTy(DAG.getDataLayout());
1704 unsigned LocMemOffset = VA.getLocMemOffset();
1705 SDValue StkPtrOff = DAG.getIntPtrConstant(LocMemOffset, dl);
1706 SDValue Dst = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, StkPtrOff);
1707 SDValue SrcOffset = DAG.getIntPtrConstant(4*offset, dl);
1708 SDValue Src = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, SrcOffset);
1709 SDValue SizeNode = DAG.getConstant(Flags.getByValSize() - 4*offset, dl,
1711 SDValue AlignNode = DAG.getConstant(Flags.getByValAlign(), dl,
1714 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
1715 SDValue Ops[] = { Chain, Dst, Src, SizeNode, AlignNode};
1716 MemOpChains.push_back(DAG.getNode(ARMISD::COPY_STRUCT_BYVAL, dl, VTs,
1719 } else if (!isSibCall) {
1720 assert(VA.isMemLoc());
1722 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1723 dl, DAG, VA, Flags));
1727 if (!MemOpChains.empty())
1728 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains);
1730 // Build a sequence of copy-to-reg nodes chained together with token chain
1731 // and flag operands which copy the outgoing args into the appropriate regs.
1733 // Tail call byval lowering might overwrite argument registers so in case of
1734 // tail call optimization the copies to registers are lowered later.
1736 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1737 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1738 RegsToPass[i].second, InFlag);
1739 InFlag = Chain.getValue(1);
1742 // For tail calls lower the arguments to the 'real' stack slot.
1744 // Force all the incoming stack arguments to be loaded from the stack
1745 // before any new outgoing arguments are stored to the stack, because the
1746 // outgoing stack slots may alias the incoming argument stack slots, and
1747 // the alias isn't otherwise explicit. This is slightly more conservative
1748 // than necessary, because it means that each store effectively depends
1749 // on every argument instead of just those arguments it would clobber.
1751 // Do not flag preceding copytoreg stuff together with the following stuff.
1753 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1754 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1755 RegsToPass[i].second, InFlag);
1756 InFlag = Chain.getValue(1);
1761 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1762 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1763 // node so that legalize doesn't hack it.
1764 bool isDirect = false;
1765 bool isARMFunc = false;
1766 bool isLocalARMFunc = false;
1767 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1768 auto PtrVt = getPointerTy(DAG.getDataLayout());
1770 if (Subtarget->genLongCalls()) {
1771 assert((Subtarget->isTargetWindows() ||
1772 getTargetMachine().getRelocationModel() == Reloc::Static) &&
1773 "long-calls with non-static relocation model!");
1774 // Handle a global address or an external symbol. If it's not one of
1775 // those, the target's already in a register, so we don't need to do
1777 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1778 const GlobalValue *GV = G->getGlobal();
1779 // Create a constant pool entry for the callee address
1780 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1781 ARMConstantPoolValue *CPV =
1782 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 0);
1784 // Get the address of the callee into a register
1785 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVt, 4);
1786 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1787 Callee = DAG.getLoad(
1788 PtrVt, dl, DAG.getEntryNode(), CPAddr,
1789 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), false,
1791 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1792 const char *Sym = S->getSymbol();
1794 // Create a constant pool entry for the callee address
1795 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1796 ARMConstantPoolValue *CPV =
1797 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1798 ARMPCLabelIndex, 0);
1799 // Get the address of the callee into a register
1800 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVt, 4);
1801 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1802 Callee = DAG.getLoad(
1803 PtrVt, dl, DAG.getEntryNode(), CPAddr,
1804 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), false,
1807 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1808 const GlobalValue *GV = G->getGlobal();
1810 bool isDef = GV->isStrongDefinitionForLinker();
1811 bool isStub = (!isDef && Subtarget->isTargetMachO()) &&
1812 getTargetMachine().getRelocationModel() != Reloc::Static;
1813 isARMFunc = !Subtarget->isThumb() || (isStub && !Subtarget->isMClass());
1814 // ARM call to a local ARM function is predicable.
1815 isLocalARMFunc = !Subtarget->isThumb() && (isDef || !ARMInterworking);
1816 // tBX takes a register source operand.
1817 if (isStub && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1818 assert(Subtarget->isTargetMachO() && "WrapperPIC use on non-MachO?");
1819 Callee = DAG.getNode(
1820 ARMISD::WrapperPIC, dl, PtrVt,
1821 DAG.getTargetGlobalAddress(GV, dl, PtrVt, 0, ARMII::MO_NONLAZY));
1822 Callee = DAG.getLoad(PtrVt, dl, DAG.getEntryNode(), Callee,
1823 MachinePointerInfo::getGOT(DAG.getMachineFunction()),
1824 false, false, true, 0);
1825 } else if (Subtarget->isTargetCOFF()) {
1826 assert(Subtarget->isTargetWindows() &&
1827 "Windows is the only supported COFF target");
1828 unsigned TargetFlags = GV->hasDLLImportStorageClass()
1829 ? ARMII::MO_DLLIMPORT
1830 : ARMII::MO_NO_FLAG;
1832 DAG.getTargetGlobalAddress(GV, dl, PtrVt, /*Offset=*/0, TargetFlags);
1833 if (GV->hasDLLImportStorageClass())
1835 DAG.getLoad(PtrVt, dl, DAG.getEntryNode(),
1836 DAG.getNode(ARMISD::Wrapper, dl, PtrVt, Callee),
1837 MachinePointerInfo::getGOT(DAG.getMachineFunction()),
1838 false, false, false, 0);
1840 // On ELF targets for PIC code, direct calls should go through the PLT
1841 unsigned OpFlags = 0;
1842 if (Subtarget->isTargetELF() &&
1843 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1844 OpFlags = ARMII::MO_PLT;
1845 Callee = DAG.getTargetGlobalAddress(GV, dl, PtrVt, 0, OpFlags);
1847 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1849 bool isStub = Subtarget->isTargetMachO() &&
1850 getTargetMachine().getRelocationModel() != Reloc::Static;
1851 isARMFunc = !Subtarget->isThumb() || (isStub && !Subtarget->isMClass());
1852 // tBX takes a register source operand.
1853 const char *Sym = S->getSymbol();
1854 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1855 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1856 ARMConstantPoolValue *CPV =
1857 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1858 ARMPCLabelIndex, 4);
1859 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVt, 4);
1860 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1861 Callee = DAG.getLoad(
1862 PtrVt, dl, DAG.getEntryNode(), CPAddr,
1863 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), false,
1865 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, dl, MVT::i32);
1866 Callee = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVt, Callee, PICLabel);
1868 unsigned OpFlags = 0;
1869 // On ELF targets for PIC code, direct calls should go through the PLT
1870 if (Subtarget->isTargetELF() &&
1871 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1872 OpFlags = ARMII::MO_PLT;
1873 Callee = DAG.getTargetExternalSymbol(Sym, PtrVt, OpFlags);
1877 // FIXME: handle tail calls differently.
1879 if (Subtarget->isThumb()) {
1880 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1881 CallOpc = ARMISD::CALL_NOLINK;
1883 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1885 if (!isDirect && !Subtarget->hasV5TOps())
1886 CallOpc = ARMISD::CALL_NOLINK;
1887 else if (doesNotRet && isDirect && Subtarget->hasRAS() &&
1888 // Emit regular call when code size is the priority
1889 !MF.getFunction()->optForMinSize())
1890 // "mov lr, pc; b _foo" to avoid confusing the RSP
1891 CallOpc = ARMISD::CALL_NOLINK;
1893 CallOpc = isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL;
1896 std::vector<SDValue> Ops;
1897 Ops.push_back(Chain);
1898 Ops.push_back(Callee);
1900 // Add argument registers to the end of the list so that they are known live
1902 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1903 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1904 RegsToPass[i].second.getValueType()));
1906 // Add a register mask operand representing the call-preserved registers.
1908 const uint32_t *Mask;
1909 const ARMBaseRegisterInfo *ARI = Subtarget->getRegisterInfo();
1911 // For 'this' returns, use the R0-preserving mask if applicable
1912 Mask = ARI->getThisReturnPreservedMask(MF, CallConv);
1914 // Set isThisReturn to false if the calling convention is not one that
1915 // allows 'returned' to be modeled in this way, so LowerCallResult does
1916 // not try to pass 'this' straight through
1917 isThisReturn = false;
1918 Mask = ARI->getCallPreservedMask(MF, CallConv);
1921 Mask = ARI->getCallPreservedMask(MF, CallConv);
1923 assert(Mask && "Missing call preserved mask for calling convention");
1924 Ops.push_back(DAG.getRegisterMask(Mask));
1927 if (InFlag.getNode())
1928 Ops.push_back(InFlag);
1930 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
1932 MF.getFrameInfo()->setHasTailCall();
1933 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, Ops);
1936 // Returns a chain and a flag for retval copy to use.
1937 Chain = DAG.getNode(CallOpc, dl, NodeTys, Ops);
1938 InFlag = Chain.getValue(1);
1940 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, dl, true),
1941 DAG.getIntPtrConstant(0, dl, true), InFlag, dl);
1943 InFlag = Chain.getValue(1);
1945 // Handle result values, copying them out of physregs into vregs that we
1947 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins, dl, DAG,
1948 InVals, isThisReturn,
1949 isThisReturn ? OutVals[0] : SDValue());
1952 /// HandleByVal - Every parameter *after* a byval parameter is passed
1953 /// on the stack. Remember the next parameter register to allocate,
1954 /// and then confiscate the rest of the parameter registers to insure
1956 void ARMTargetLowering::HandleByVal(CCState *State, unsigned &Size,
1957 unsigned Align) const {
1958 assert((State->getCallOrPrologue() == Prologue ||
1959 State->getCallOrPrologue() == Call) &&
1960 "unhandled ParmContext");
1962 // Byval (as with any stack) slots are always at least 4 byte aligned.
1963 Align = std::max(Align, 4U);
1965 unsigned Reg = State->AllocateReg(GPRArgRegs);
1969 unsigned AlignInRegs = Align / 4;
1970 unsigned Waste = (ARM::R4 - Reg) % AlignInRegs;
1971 for (unsigned i = 0; i < Waste; ++i)
1972 Reg = State->AllocateReg(GPRArgRegs);
1977 unsigned Excess = 4 * (ARM::R4 - Reg);
1979 // Special case when NSAA != SP and parameter size greater than size of
1980 // all remained GPR regs. In that case we can't split parameter, we must
1981 // send it to stack. We also must set NCRN to R4, so waste all
1982 // remained registers.
1983 const unsigned NSAAOffset = State->getNextStackOffset();
1984 if (NSAAOffset != 0 && Size > Excess) {
1985 while (State->AllocateReg(GPRArgRegs))
1990 // First register for byval parameter is the first register that wasn't
1991 // allocated before this method call, so it would be "reg".
1992 // If parameter is small enough to be saved in range [reg, r4), then
1993 // the end (first after last) register would be reg + param-size-in-regs,
1994 // else parameter would be splitted between registers and stack,
1995 // end register would be r4 in this case.
1996 unsigned ByValRegBegin = Reg;
1997 unsigned ByValRegEnd = std::min<unsigned>(Reg + Size / 4, ARM::R4);
1998 State->addInRegsParamInfo(ByValRegBegin, ByValRegEnd);
1999 // Note, first register is allocated in the beginning of function already,
2000 // allocate remained amount of registers we need.
2001 for (unsigned i = Reg + 1; i != ByValRegEnd; ++i)
2002 State->AllocateReg(GPRArgRegs);
2003 // A byval parameter that is split between registers and memory needs its
2004 // size truncated here.
2005 // In the case where the entire structure fits in registers, we set the
2006 // size in memory to zero.
2007 Size = std::max<int>(Size - Excess, 0);
2010 /// MatchingStackOffset - Return true if the given stack call argument is
2011 /// already available in the same position (relatively) of the caller's
2012 /// incoming argument stack.
2014 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
2015 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
2016 const TargetInstrInfo *TII) {
2017 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
2019 if (Arg.getOpcode() == ISD::CopyFromReg) {
2020 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
2021 if (!TargetRegisterInfo::isVirtualRegister(VR))
2023 MachineInstr *Def = MRI->getVRegDef(VR);
2026 if (!Flags.isByVal()) {
2027 if (!TII->isLoadFromStackSlot(Def, FI))
2032 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
2033 if (Flags.isByVal())
2034 // ByVal argument is passed in as a pointer but it's now being
2035 // dereferenced. e.g.
2036 // define @foo(%struct.X* %A) {
2037 // tail call @bar(%struct.X* byval %A)
2040 SDValue Ptr = Ld->getBasePtr();
2041 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
2044 FI = FINode->getIndex();
2048 assert(FI != INT_MAX);
2049 if (!MFI->isFixedObjectIndex(FI))
2051 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
2054 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2055 /// for tail call optimization. Targets which want to do tail call
2056 /// optimization should implement this function.
2058 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2059 CallingConv::ID CalleeCC,
2061 bool isCalleeStructRet,
2062 bool isCallerStructRet,
2063 const SmallVectorImpl<ISD::OutputArg> &Outs,
2064 const SmallVectorImpl<SDValue> &OutVals,
2065 const SmallVectorImpl<ISD::InputArg> &Ins,
2066 SelectionDAG& DAG) const {
2067 const Function *CallerF = DAG.getMachineFunction().getFunction();
2068 CallingConv::ID CallerCC = CallerF->getCallingConv();
2069 bool CCMatch = CallerCC == CalleeCC;
2071 assert(Subtarget->supportsTailCall());
2073 // Look for obvious safe cases to perform tail call optimization that do not
2074 // require ABI changes. This is what gcc calls sibcall.
2076 // Do not sibcall optimize vararg calls unless the call site is not passing
2078 if (isVarArg && !Outs.empty())
2081 // Exception-handling functions need a special set of instructions to indicate
2082 // a return to the hardware. Tail-calling another function would probably
2084 if (CallerF->hasFnAttribute("interrupt"))
2087 // Also avoid sibcall optimization if either caller or callee uses struct
2088 // return semantics.
2089 if (isCalleeStructRet || isCallerStructRet)
2092 // Externally-defined functions with weak linkage should not be
2093 // tail-called on ARM when the OS does not support dynamic
2094 // pre-emption of symbols, as the AAELF spec requires normal calls
2095 // to undefined weak functions to be replaced with a NOP or jump to the
2096 // next instruction. The behaviour of branch instructions in this
2097 // situation (as used for tail calls) is implementation-defined, so we
2098 // cannot rely on the linker replacing the tail call with a return.
2099 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2100 const GlobalValue *GV = G->getGlobal();
2101 const Triple &TT = getTargetMachine().getTargetTriple();
2102 if (GV->hasExternalWeakLinkage() &&
2103 (!TT.isOSWindows() || TT.isOSBinFormatELF() || TT.isOSBinFormatMachO()))
2107 // If the calling conventions do not match, then we'd better make sure the
2108 // results are returned in the same way as what the caller expects.
2110 SmallVector<CCValAssign, 16> RVLocs1;
2111 ARMCCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(), RVLocs1,
2112 *DAG.getContext(), Call);
2113 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
2115 SmallVector<CCValAssign, 16> RVLocs2;
2116 ARMCCState CCInfo2(CallerCC, false, DAG.getMachineFunction(), RVLocs2,
2117 *DAG.getContext(), Call);
2118 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
2120 if (RVLocs1.size() != RVLocs2.size())
2122 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
2123 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
2125 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
2127 if (RVLocs1[i].isRegLoc()) {
2128 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
2131 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
2137 // If Caller's vararg or byval argument has been split between registers and
2138 // stack, do not perform tail call, since part of the argument is in caller's
2140 const ARMFunctionInfo *AFI_Caller = DAG.getMachineFunction().
2141 getInfo<ARMFunctionInfo>();
2142 if (AFI_Caller->getArgRegsSaveSize())
2145 // If the callee takes no arguments then go on to check the results of the
2147 if (!Outs.empty()) {
2148 // Check if stack adjustment is needed. For now, do not do this if any
2149 // argument is passed on the stack.
2150 SmallVector<CCValAssign, 16> ArgLocs;
2151 ARMCCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
2152 *DAG.getContext(), Call);
2153 CCInfo.AnalyzeCallOperands(Outs,
2154 CCAssignFnForNode(CalleeCC, false, isVarArg));
2155 if (CCInfo.getNextStackOffset()) {
2156 MachineFunction &MF = DAG.getMachineFunction();
2158 // Check if the arguments are already laid out in the right way as
2159 // the caller's fixed stack objects.
2160 MachineFrameInfo *MFI = MF.getFrameInfo();
2161 const MachineRegisterInfo *MRI = &MF.getRegInfo();
2162 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
2163 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
2165 ++i, ++realArgIdx) {
2166 CCValAssign &VA = ArgLocs[i];
2167 EVT RegVT = VA.getLocVT();
2168 SDValue Arg = OutVals[realArgIdx];
2169 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
2170 if (VA.getLocInfo() == CCValAssign::Indirect)
2172 if (VA.needsCustom()) {
2173 // f64 and vector types are split into multiple registers or
2174 // register/stack-slot combinations. The types will not match
2175 // the registers; give up on memory f64 refs until we figure
2176 // out what to do about this.
2179 if (!ArgLocs[++i].isRegLoc())
2181 if (RegVT == MVT::v2f64) {
2182 if (!ArgLocs[++i].isRegLoc())
2184 if (!ArgLocs[++i].isRegLoc())
2187 } else if (!VA.isRegLoc()) {
2188 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
2200 ARMTargetLowering::CanLowerReturn(CallingConv::ID CallConv,
2201 MachineFunction &MF, bool isVarArg,
2202 const SmallVectorImpl<ISD::OutputArg> &Outs,
2203 LLVMContext &Context) const {
2204 SmallVector<CCValAssign, 16> RVLocs;
2205 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, Context);
2206 return CCInfo.CheckReturn(Outs, CCAssignFnForNode(CallConv, /*Return=*/true,
2210 static SDValue LowerInterruptReturn(SmallVectorImpl<SDValue> &RetOps,
2211 SDLoc DL, SelectionDAG &DAG) {
2212 const MachineFunction &MF = DAG.getMachineFunction();
2213 const Function *F = MF.getFunction();
2215 StringRef IntKind = F->getFnAttribute("interrupt").getValueAsString();
2217 // See ARM ARM v7 B1.8.3. On exception entry LR is set to a possibly offset
2218 // version of the "preferred return address". These offsets affect the return
2219 // instruction if this is a return from PL1 without hypervisor extensions.
2220 // IRQ/FIQ: +4 "subs pc, lr, #4"
2221 // SWI: 0 "subs pc, lr, #0"
2222 // ABORT: +4 "subs pc, lr, #4"
2223 // UNDEF: +4/+2 "subs pc, lr, #0"
2224 // UNDEF varies depending on where the exception came from ARM or Thumb
2225 // mode. Alongside GCC, we throw our hands up in disgust and pretend it's 0.
2228 if (IntKind == "" || IntKind == "IRQ" || IntKind == "FIQ" ||
2231 else if (IntKind == "SWI" || IntKind == "UNDEF")
2234 report_fatal_error("Unsupported interrupt attribute. If present, value "
2235 "must be one of: IRQ, FIQ, SWI, ABORT or UNDEF");
2237 RetOps.insert(RetOps.begin() + 1,
2238 DAG.getConstant(LROffset, DL, MVT::i32, false));
2240 return DAG.getNode(ARMISD::INTRET_FLAG, DL, MVT::Other, RetOps);
2244 ARMTargetLowering::LowerReturn(SDValue Chain,
2245 CallingConv::ID CallConv, bool isVarArg,
2246 const SmallVectorImpl<ISD::OutputArg> &Outs,
2247 const SmallVectorImpl<SDValue> &OutVals,
2248 SDLoc dl, SelectionDAG &DAG) const {
2250 // CCValAssign - represent the assignment of the return value to a location.
2251 SmallVector<CCValAssign, 16> RVLocs;
2253 // CCState - Info about the registers and stack slots.
2254 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
2255 *DAG.getContext(), Call);
2257 // Analyze outgoing return values.
2258 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
2262 SmallVector<SDValue, 4> RetOps;
2263 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
2264 bool isLittleEndian = Subtarget->isLittle();
2266 MachineFunction &MF = DAG.getMachineFunction();
2267 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2268 AFI->setReturnRegsCount(RVLocs.size());
2270 // Copy the result values into the output registers.
2271 for (unsigned i = 0, realRVLocIdx = 0;
2273 ++i, ++realRVLocIdx) {
2274 CCValAssign &VA = RVLocs[i];
2275 assert(VA.isRegLoc() && "Can only return in registers!");
2277 SDValue Arg = OutVals[realRVLocIdx];
2279 switch (VA.getLocInfo()) {
2280 default: llvm_unreachable("Unknown loc info!");
2281 case CCValAssign::Full: break;
2282 case CCValAssign::BCvt:
2283 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
2287 if (VA.needsCustom()) {
2288 if (VA.getLocVT() == MVT::v2f64) {
2289 // Extract the first half and return it in two registers.
2290 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
2291 DAG.getConstant(0, dl, MVT::i32));
2292 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
2293 DAG.getVTList(MVT::i32, MVT::i32), Half);
2295 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2296 HalfGPRs.getValue(isLittleEndian ? 0 : 1),
2298 Flag = Chain.getValue(1);
2299 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2300 VA = RVLocs[++i]; // skip ahead to next loc
2301 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2302 HalfGPRs.getValue(isLittleEndian ? 1 : 0),
2304 Flag = Chain.getValue(1);
2305 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2306 VA = RVLocs[++i]; // skip ahead to next loc
2308 // Extract the 2nd half and fall through to handle it as an f64 value.
2309 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
2310 DAG.getConstant(1, dl, MVT::i32));
2312 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
2314 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
2315 DAG.getVTList(MVT::i32, MVT::i32), Arg);
2316 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2317 fmrrd.getValue(isLittleEndian ? 0 : 1),
2319 Flag = Chain.getValue(1);
2320 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2321 VA = RVLocs[++i]; // skip ahead to next loc
2322 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2323 fmrrd.getValue(isLittleEndian ? 1 : 0),
2326 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
2328 // Guarantee that all emitted copies are
2329 // stuck together, avoiding something bad.
2330 Flag = Chain.getValue(1);
2331 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2334 // Update chain and glue.
2337 RetOps.push_back(Flag);
2339 // CPUs which aren't M-class use a special sequence to return from
2340 // exceptions (roughly, any instruction setting pc and cpsr simultaneously,
2341 // though we use "subs pc, lr, #N").
2343 // M-class CPUs actually use a normal return sequence with a special
2344 // (hardware-provided) value in LR, so the normal code path works.
2345 if (DAG.getMachineFunction().getFunction()->hasFnAttribute("interrupt") &&
2346 !Subtarget->isMClass()) {
2347 if (Subtarget->isThumb1Only())
2348 report_fatal_error("interrupt attribute is not supported in Thumb1");
2349 return LowerInterruptReturn(RetOps, dl, DAG);
2352 return DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, RetOps);
2355 bool ARMTargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
2356 if (N->getNumValues() != 1)
2358 if (!N->hasNUsesOfValue(1, 0))
2361 SDValue TCChain = Chain;
2362 SDNode *Copy = *N->use_begin();
2363 if (Copy->getOpcode() == ISD::CopyToReg) {
2364 // If the copy has a glue operand, we conservatively assume it isn't safe to
2365 // perform a tail call.
2366 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
2368 TCChain = Copy->getOperand(0);
2369 } else if (Copy->getOpcode() == ARMISD::VMOVRRD) {
2370 SDNode *VMov = Copy;
2371 // f64 returned in a pair of GPRs.
2372 SmallPtrSet<SDNode*, 2> Copies;
2373 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
2375 if (UI->getOpcode() != ISD::CopyToReg)
2379 if (Copies.size() > 2)
2382 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
2384 SDValue UseChain = UI->getOperand(0);
2385 if (Copies.count(UseChain.getNode()))
2389 // We are at the top of this chain.
2390 // If the copy has a glue operand, we conservatively assume it
2391 // isn't safe to perform a tail call.
2392 if (UI->getOperand(UI->getNumOperands()-1).getValueType() == MVT::Glue)
2398 } else if (Copy->getOpcode() == ISD::BITCAST) {
2399 // f32 returned in a single GPR.
2400 if (!Copy->hasOneUse())
2402 Copy = *Copy->use_begin();
2403 if (Copy->getOpcode() != ISD::CopyToReg || !Copy->hasNUsesOfValue(1, 0))
2405 // If the copy has a glue operand, we conservatively assume it isn't safe to
2406 // perform a tail call.
2407 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
2409 TCChain = Copy->getOperand(0);
2414 bool HasRet = false;
2415 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
2417 if (UI->getOpcode() != ARMISD::RET_FLAG &&
2418 UI->getOpcode() != ARMISD::INTRET_FLAG)
2430 bool ARMTargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
2431 if (!Subtarget->supportsTailCall())
2435 CI->getParent()->getParent()->getFnAttribute("disable-tail-calls");
2436 if (!CI->isTailCall() || Attr.getValueAsString() == "true")
2442 // Trying to write a 64 bit value so need to split into two 32 bit values first,
2443 // and pass the lower and high parts through.
2444 static SDValue LowerWRITE_REGISTER(SDValue Op, SelectionDAG &DAG) {
2446 SDValue WriteValue = Op->getOperand(2);
2448 // This function is only supposed to be called for i64 type argument.
2449 assert(WriteValue.getValueType() == MVT::i64
2450 && "LowerWRITE_REGISTER called for non-i64 type argument.");
2452 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, WriteValue,
2453 DAG.getConstant(0, DL, MVT::i32));
2454 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, WriteValue,
2455 DAG.getConstant(1, DL, MVT::i32));
2456 SDValue Ops[] = { Op->getOperand(0), Op->getOperand(1), Lo, Hi };
2457 return DAG.getNode(ISD::WRITE_REGISTER, DL, MVT::Other, Ops);
2460 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
2461 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
2462 // one of the above mentioned nodes. It has to be wrapped because otherwise
2463 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
2464 // be used to form addressing mode. These wrapped nodes will be selected
2466 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
2467 EVT PtrVT = Op.getValueType();
2468 // FIXME there is no actual debug info here
2470 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
2472 if (CP->isMachineConstantPoolEntry())
2473 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
2474 CP->getAlignment());
2476 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
2477 CP->getAlignment());
2478 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
2481 unsigned ARMTargetLowering::getJumpTableEncoding() const {
2482 return MachineJumpTableInfo::EK_Inline;
2485 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
2486 SelectionDAG &DAG) const {
2487 MachineFunction &MF = DAG.getMachineFunction();
2488 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2489 unsigned ARMPCLabelIndex = 0;
2491 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2492 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
2493 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2495 if (RelocM == Reloc::Static) {
2496 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
2498 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2499 ARMPCLabelIndex = AFI->createPICLabelUId();
2500 ARMConstantPoolValue *CPV =
2501 ARMConstantPoolConstant::Create(BA, ARMPCLabelIndex,
2502 ARMCP::CPBlockAddress, PCAdj);
2503 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2505 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
2507 DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
2508 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
2509 false, false, false, 0);
2510 if (RelocM == Reloc::Static)
2512 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, DL, MVT::i32);
2513 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
2516 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
2518 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
2519 SelectionDAG &DAG) const {
2521 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2522 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2523 MachineFunction &MF = DAG.getMachineFunction();
2524 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2525 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2526 ARMConstantPoolValue *CPV =
2527 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2528 ARMCP::CPValue, PCAdj, ARMCP::TLSGD, true);
2529 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2530 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
2532 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
2533 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
2534 false, false, false, 0);
2535 SDValue Chain = Argument.getValue(1);
2537 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, dl, MVT::i32);
2538 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
2540 // call __tls_get_addr.
2543 Entry.Node = Argument;
2544 Entry.Ty = (Type *) Type::getInt32Ty(*DAG.getContext());
2545 Args.push_back(Entry);
2547 // FIXME: is there useful debug info available here?
2548 TargetLowering::CallLoweringInfo CLI(DAG);
2549 CLI.setDebugLoc(dl).setChain(Chain)
2550 .setCallee(CallingConv::C, Type::getInt32Ty(*DAG.getContext()),
2551 DAG.getExternalSymbol("__tls_get_addr", PtrVT), std::move(Args),
2554 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
2555 return CallResult.first;
2558 // Lower ISD::GlobalTLSAddress using the "initial exec" or
2559 // "local exec" model.
2561 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
2563 TLSModel::Model model) const {
2564 const GlobalValue *GV = GA->getGlobal();
2567 SDValue Chain = DAG.getEntryNode();
2568 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2569 // Get the Thread Pointer
2570 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2572 if (model == TLSModel::InitialExec) {
2573 MachineFunction &MF = DAG.getMachineFunction();
2574 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2575 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2576 // Initial exec model.
2577 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2578 ARMConstantPoolValue *CPV =
2579 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2580 ARMCP::CPValue, PCAdj, ARMCP::GOTTPOFF,
2582 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2583 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2584 Offset = DAG.getLoad(
2585 PtrVT, dl, Chain, Offset,
2586 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), false,
2588 Chain = Offset.getValue(1);
2590 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, dl, MVT::i32);
2591 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
2593 Offset = DAG.getLoad(
2594 PtrVT, dl, Chain, Offset,
2595 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), false,
2599 assert(model == TLSModel::LocalExec);
2600 ARMConstantPoolValue *CPV =
2601 ARMConstantPoolConstant::Create(GV, ARMCP::TPOFF);
2602 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2603 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2604 Offset = DAG.getLoad(
2605 PtrVT, dl, Chain, Offset,
2606 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), false,
2610 // The address of the thread local variable is the add of the thread
2611 // pointer with the offset of the variable.
2612 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
2616 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
2617 // TODO: implement the "local dynamic" model
2618 assert(Subtarget->isTargetELF() &&
2619 "TLS not implemented for non-ELF targets");
2620 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
2621 if (DAG.getTarget().Options.EmulatedTLS)
2622 return LowerToTLSEmulatedModel(GA, DAG);
2624 TLSModel::Model model = getTargetMachine().getTLSModel(GA->getGlobal());
2627 case TLSModel::GeneralDynamic:
2628 case TLSModel::LocalDynamic:
2629 return LowerToTLSGeneralDynamicModel(GA, DAG);
2630 case TLSModel::InitialExec:
2631 case TLSModel::LocalExec:
2632 return LowerToTLSExecModels(GA, DAG, model);
2634 llvm_unreachable("bogus TLS model");
2637 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
2638 SelectionDAG &DAG) const {
2639 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2641 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2642 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2644 !(GV->hasHiddenVisibility() || GV->hasLocalLinkage());
2646 MachineFunction &MF = DAG.getMachineFunction();
2647 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2648 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2649 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2651 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2652 ARMConstantPoolValue *CPV = ARMConstantPoolConstant::Create(
2653 GV, ARMPCLabelIndex, ARMCP::CPValue, PCAdj,
2654 UseGOT_PREL ? ARMCP::GOT_PREL : ARMCP::no_modifier,
2655 /*AddCurrentAddress=*/UseGOT_PREL);
2656 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2657 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2658 SDValue Result = DAG.getLoad(
2659 PtrVT, dl, DAG.getEntryNode(), CPAddr,
2660 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), false,
2662 SDValue Chain = Result.getValue(1);
2663 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, dl, MVT::i32);
2664 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2666 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
2667 MachinePointerInfo::getGOT(DAG.getMachineFunction()),
2668 false, false, false, 0);
2672 // If we have T2 ops, we can materialize the address directly via movt/movw
2673 // pair. This is always cheaper.
2674 if (Subtarget->useMovt(DAG.getMachineFunction())) {
2676 // FIXME: Once remat is capable of dealing with instructions with register
2677 // operands, expand this into two nodes.
2678 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2679 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2681 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2682 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2684 PtrVT, dl, DAG.getEntryNode(), CPAddr,
2685 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), false,
2690 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
2691 SelectionDAG &DAG) const {
2692 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2694 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2695 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2697 if (Subtarget->useMovt(DAG.getMachineFunction()))
2700 // FIXME: Once remat is capable of dealing with instructions with register
2701 // operands, expand this into multiple nodes
2703 RelocM == Reloc::PIC_ ? ARMISD::WrapperPIC : ARMISD::Wrapper;
2705 SDValue G = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, ARMII::MO_NONLAZY);
2706 SDValue Result = DAG.getNode(Wrapper, dl, PtrVT, G);
2708 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2709 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
2710 MachinePointerInfo::getGOT(DAG.getMachineFunction()),
2711 false, false, false, 0);
2715 SDValue ARMTargetLowering::LowerGlobalAddressWindows(SDValue Op,
2716 SelectionDAG &DAG) const {
2717 assert(Subtarget->isTargetWindows() && "non-Windows COFF is not supported");
2718 assert(Subtarget->useMovt(DAG.getMachineFunction()) &&
2719 "Windows on ARM expects to use movw/movt");
2721 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2722 const ARMII::TOF TargetFlags =
2723 (GV->hasDLLImportStorageClass() ? ARMII::MO_DLLIMPORT : ARMII::MO_NO_FLAG);
2724 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2730 // FIXME: Once remat is capable of dealing with instructions with register
2731 // operands, expand this into two nodes.
2732 Result = DAG.getNode(ARMISD::Wrapper, DL, PtrVT,
2733 DAG.getTargetGlobalAddress(GV, DL, PtrVT, /*Offset=*/0,
2735 if (GV->hasDLLImportStorageClass())
2736 Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), Result,
2737 MachinePointerInfo::getGOT(DAG.getMachineFunction()),
2738 false, false, false, 0);
2743 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
2745 SDValue Val = DAG.getConstant(0, dl, MVT::i32);
2746 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl,
2747 DAG.getVTList(MVT::i32, MVT::Other), Op.getOperand(0),
2748 Op.getOperand(1), Val);
2752 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
2754 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
2755 Op.getOperand(1), DAG.getConstant(0, dl, MVT::i32));
2758 SDValue ARMTargetLowering::LowerEH_SJLJ_SETUP_DISPATCH(SDValue Op,
2759 SelectionDAG &DAG) const {
2761 return DAG.getNode(ARMISD::EH_SJLJ_SETUP_DISPATCH, dl, MVT::Other,
2766 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
2767 const ARMSubtarget *Subtarget) const {
2768 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2771 default: return SDValue(); // Don't custom lower most intrinsics.
2772 case Intrinsic::arm_rbit: {
2773 assert(Op.getOperand(1).getValueType() == MVT::i32 &&
2774 "RBIT intrinsic must have i32 type!");
2775 return DAG.getNode(ARMISD::RBIT, dl, MVT::i32, Op.getOperand(1));
2777 case Intrinsic::arm_thread_pointer: {
2778 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2779 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2781 case Intrinsic::eh_sjlj_lsda: {
2782 MachineFunction &MF = DAG.getMachineFunction();
2783 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2784 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2785 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2786 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2788 unsigned PCAdj = (RelocM != Reloc::PIC_)
2789 ? 0 : (Subtarget->isThumb() ? 4 : 8);
2790 ARMConstantPoolValue *CPV =
2791 ARMConstantPoolConstant::Create(MF.getFunction(), ARMPCLabelIndex,
2792 ARMCP::CPLSDA, PCAdj);
2793 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2794 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2795 SDValue Result = DAG.getLoad(
2796 PtrVT, dl, DAG.getEntryNode(), CPAddr,
2797 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), false,
2800 if (RelocM == Reloc::PIC_) {
2801 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, dl, MVT::i32);
2802 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2806 case Intrinsic::arm_neon_vmulls:
2807 case Intrinsic::arm_neon_vmullu: {
2808 unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vmulls)
2809 ? ARMISD::VMULLs : ARMISD::VMULLu;
2810 return DAG.getNode(NewOpc, SDLoc(Op), Op.getValueType(),
2811 Op.getOperand(1), Op.getOperand(2));
2813 case Intrinsic::arm_neon_vminnm:
2814 case Intrinsic::arm_neon_vmaxnm: {
2815 unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vminnm)
2816 ? ISD::FMINNUM : ISD::FMAXNUM;
2817 return DAG.getNode(NewOpc, SDLoc(Op), Op.getValueType(),
2818 Op.getOperand(1), Op.getOperand(2));
2820 case Intrinsic::arm_neon_vminu:
2821 case Intrinsic::arm_neon_vmaxu: {
2822 if (Op.getValueType().isFloatingPoint())
2824 unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vminu)
2825 ? ISD::UMIN : ISD::UMAX;
2826 return DAG.getNode(NewOpc, SDLoc(Op), Op.getValueType(),
2827 Op.getOperand(1), Op.getOperand(2));
2829 case Intrinsic::arm_neon_vmins:
2830 case Intrinsic::arm_neon_vmaxs: {
2831 // v{min,max}s is overloaded between signed integers and floats.
2832 if (!Op.getValueType().isFloatingPoint()) {
2833 unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vmins)
2834 ? ISD::SMIN : ISD::SMAX;
2835 return DAG.getNode(NewOpc, SDLoc(Op), Op.getValueType(),
2836 Op.getOperand(1), Op.getOperand(2));
2838 unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vmins)
2839 ? ISD::FMINNAN : ISD::FMAXNAN;
2840 return DAG.getNode(NewOpc, SDLoc(Op), Op.getValueType(),
2841 Op.getOperand(1), Op.getOperand(2));
2846 static SDValue LowerATOMIC_FENCE(SDValue Op, SelectionDAG &DAG,
2847 const ARMSubtarget *Subtarget) {
2848 // FIXME: handle "fence singlethread" more efficiently.
2850 if (!Subtarget->hasDataBarrier()) {
2851 // Some ARMv6 cpus can support data barriers with an mcr instruction.
2852 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2854 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2855 "Unexpected ISD::ATOMIC_FENCE encountered. Should be libcall!");
2856 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2857 DAG.getConstant(0, dl, MVT::i32));
2860 ConstantSDNode *OrdN = cast<ConstantSDNode>(Op.getOperand(1));
2861 AtomicOrdering Ord = static_cast<AtomicOrdering>(OrdN->getZExtValue());
2862 ARM_MB::MemBOpt Domain = ARM_MB::ISH;
2863 if (Subtarget->isMClass()) {
2864 // Only a full system barrier exists in the M-class architectures.
2865 Domain = ARM_MB::SY;
2866 } else if (Subtarget->isSwift() && Ord == Release) {
2867 // Swift happens to implement ISHST barriers in a way that's compatible with
2868 // Release semantics but weaker than ISH so we'd be fools not to use
2869 // it. Beware: other processors probably don't!
2870 Domain = ARM_MB::ISHST;
2873 return DAG.getNode(ISD::INTRINSIC_VOID, dl, MVT::Other, Op.getOperand(0),
2874 DAG.getConstant(Intrinsic::arm_dmb, dl, MVT::i32),
2875 DAG.getConstant(Domain, dl, MVT::i32));
2878 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG,
2879 const ARMSubtarget *Subtarget) {
2880 // ARM pre v5TE and Thumb1 does not have preload instructions.
2881 if (!(Subtarget->isThumb2() ||
2882 (!Subtarget->isThumb1Only() && Subtarget->hasV5TEOps())))
2883 // Just preserve the chain.
2884 return Op.getOperand(0);
2887 unsigned isRead = ~cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() & 1;
2889 (!Subtarget->hasV7Ops() || !Subtarget->hasMPExtension()))
2890 // ARMv7 with MP extension has PLDW.
2891 return Op.getOperand(0);
2893 unsigned isData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
2894 if (Subtarget->isThumb()) {
2896 isRead = ~isRead & 1;
2897 isData = ~isData & 1;
2900 return DAG.getNode(ARMISD::PRELOAD, dl, MVT::Other, Op.getOperand(0),
2901 Op.getOperand(1), DAG.getConstant(isRead, dl, MVT::i32),
2902 DAG.getConstant(isData, dl, MVT::i32));
2905 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
2906 MachineFunction &MF = DAG.getMachineFunction();
2907 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
2909 // vastart just stores the address of the VarArgsFrameIndex slot into the
2910 // memory location argument.
2912 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy(DAG.getDataLayout());
2913 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2914 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
2915 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
2916 MachinePointerInfo(SV), false, false, 0);
2920 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
2921 SDValue &Root, SelectionDAG &DAG,
2923 MachineFunction &MF = DAG.getMachineFunction();
2924 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2926 const TargetRegisterClass *RC;
2927 if (AFI->isThumb1OnlyFunction())
2928 RC = &ARM::tGPRRegClass;
2930 RC = &ARM::GPRRegClass;
2932 // Transform the arguments stored in physical registers into virtual ones.
2933 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2934 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2937 if (NextVA.isMemLoc()) {
2938 MachineFrameInfo *MFI = MF.getFrameInfo();
2939 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true);
2941 // Create load node to retrieve arguments from the stack.
2942 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
2943 ArgValue2 = DAG.getLoad(
2944 MVT::i32, dl, Root, FIN,
2945 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI), false,
2948 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
2949 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2951 if (!Subtarget->isLittle())
2952 std::swap (ArgValue, ArgValue2);
2953 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
2956 // The remaining GPRs hold either the beginning of variable-argument
2957 // data, or the beginning of an aggregate passed by value (usually
2958 // byval). Either way, we allocate stack slots adjacent to the data
2959 // provided by our caller, and store the unallocated registers there.
2960 // If this is a variadic function, the va_list pointer will begin with
2961 // these values; otherwise, this reassembles a (byval) structure that
2962 // was split between registers and memory.
2963 // Return: The frame index registers were stored into.
2965 ARMTargetLowering::StoreByValRegs(CCState &CCInfo, SelectionDAG &DAG,
2966 SDLoc dl, SDValue &Chain,
2967 const Value *OrigArg,
2968 unsigned InRegsParamRecordIdx,
2970 unsigned ArgSize) const {
2971 // Currently, two use-cases possible:
2972 // Case #1. Non-var-args function, and we meet first byval parameter.
2973 // Setup first unallocated register as first byval register;
2974 // eat all remained registers
2975 // (these two actions are performed by HandleByVal method).
2976 // Then, here, we initialize stack frame with
2977 // "store-reg" instructions.
2978 // Case #2. Var-args function, that doesn't contain byval parameters.
2979 // The same: eat all remained unallocated registers,
2980 // initialize stack frame.
2982 MachineFunction &MF = DAG.getMachineFunction();
2983 MachineFrameInfo *MFI = MF.getFrameInfo();
2984 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2985 unsigned RBegin, REnd;
2986 if (InRegsParamRecordIdx < CCInfo.getInRegsParamsCount()) {
2987 CCInfo.getInRegsParamInfo(InRegsParamRecordIdx, RBegin, REnd);
2989 unsigned RBeginIdx = CCInfo.getFirstUnallocated(GPRArgRegs);
2990 RBegin = RBeginIdx == 4 ? (unsigned)ARM::R4 : GPRArgRegs[RBeginIdx];
2995 ArgOffset = -4 * (ARM::R4 - RBegin);
2997 auto PtrVT = getPointerTy(DAG.getDataLayout());
2998 int FrameIndex = MFI->CreateFixedObject(ArgSize, ArgOffset, false);
2999 SDValue FIN = DAG.getFrameIndex(FrameIndex, PtrVT);
3001 SmallVector<SDValue, 4> MemOps;
3002 const TargetRegisterClass *RC =
3003 AFI->isThumb1OnlyFunction() ? &ARM::tGPRRegClass : &ARM::GPRRegClass;
3005 for (unsigned Reg = RBegin, i = 0; Reg < REnd; ++Reg, ++i) {
3006 unsigned VReg = MF.addLiveIn(Reg, RC);
3007 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
3009 DAG.getStore(Val.getValue(1), dl, Val, FIN,
3010 MachinePointerInfo(OrigArg, 4 * i), false, false, 0);
3011 MemOps.push_back(Store);
3012 FIN = DAG.getNode(ISD::ADD, dl, PtrVT, FIN, DAG.getConstant(4, dl, PtrVT));
3015 if (!MemOps.empty())
3016 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOps);
3020 // Setup stack frame, the va_list pointer will start from.
3022 ARMTargetLowering::VarArgStyleRegisters(CCState &CCInfo, SelectionDAG &DAG,
3023 SDLoc dl, SDValue &Chain,
3025 unsigned TotalArgRegsSaveSize,
3026 bool ForceMutable) const {
3027 MachineFunction &MF = DAG.getMachineFunction();
3028 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
3030 // Try to store any remaining integer argument regs
3031 // to their spots on the stack so that they may be loaded by deferencing
3032 // the result of va_next.
3033 // If there is no regs to be stored, just point address after last
3034 // argument passed via stack.
3035 int FrameIndex = StoreByValRegs(CCInfo, DAG, dl, Chain, nullptr,
3036 CCInfo.getInRegsParamsCount(),
3037 CCInfo.getNextStackOffset(), 4);
3038 AFI->setVarArgsFrameIndex(FrameIndex);
3042 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
3043 CallingConv::ID CallConv, bool isVarArg,
3044 const SmallVectorImpl<ISD::InputArg>
3046 SDLoc dl, SelectionDAG &DAG,
3047 SmallVectorImpl<SDValue> &InVals)
3049 MachineFunction &MF = DAG.getMachineFunction();
3050 MachineFrameInfo *MFI = MF.getFrameInfo();
3052 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
3054 // Assign locations to all of the incoming arguments.
3055 SmallVector<CCValAssign, 16> ArgLocs;
3056 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs,
3057 *DAG.getContext(), Prologue);
3058 CCInfo.AnalyzeFormalArguments(Ins,
3059 CCAssignFnForNode(CallConv, /* Return*/ false,
3062 SmallVector<SDValue, 16> ArgValues;
3064 Function::const_arg_iterator CurOrigArg = MF.getFunction()->arg_begin();
3065 unsigned CurArgIdx = 0;
3067 // Initially ArgRegsSaveSize is zero.
3068 // Then we increase this value each time we meet byval parameter.
3069 // We also increase this value in case of varargs function.
3070 AFI->setArgRegsSaveSize(0);
3072 // Calculate the amount of stack space that we need to allocate to store
3073 // byval and variadic arguments that are passed in registers.
3074 // We need to know this before we allocate the first byval or variadic
3075 // argument, as they will be allocated a stack slot below the CFA (Canonical
3076 // Frame Address, the stack pointer at entry to the function).
3077 unsigned ArgRegBegin = ARM::R4;
3078 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3079 if (CCInfo.getInRegsParamsProcessed() >= CCInfo.getInRegsParamsCount())
3082 CCValAssign &VA = ArgLocs[i];
3083 unsigned Index = VA.getValNo();
3084 ISD::ArgFlagsTy Flags = Ins[Index].Flags;
3085 if (!Flags.isByVal())
3088 assert(VA.isMemLoc() && "unexpected byval pointer in reg");
3089 unsigned RBegin, REnd;
3090 CCInfo.getInRegsParamInfo(CCInfo.getInRegsParamsProcessed(), RBegin, REnd);
3091 ArgRegBegin = std::min(ArgRegBegin, RBegin);
3093 CCInfo.nextInRegsParam();
3095 CCInfo.rewindByValRegsInfo();
3097 int lastInsIndex = -1;
3098 if (isVarArg && MFI->hasVAStart()) {
3099 unsigned RegIdx = CCInfo.getFirstUnallocated(GPRArgRegs);
3100 if (RegIdx != array_lengthof(GPRArgRegs))
3101 ArgRegBegin = std::min(ArgRegBegin, (unsigned)GPRArgRegs[RegIdx]);
3104 unsigned TotalArgRegsSaveSize = 4 * (ARM::R4 - ArgRegBegin);
3105 AFI->setArgRegsSaveSize(TotalArgRegsSaveSize);
3106 auto PtrVT = getPointerTy(DAG.getDataLayout());
3108 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3109 CCValAssign &VA = ArgLocs[i];
3110 if (Ins[VA.getValNo()].isOrigArg()) {
3111 std::advance(CurOrigArg,
3112 Ins[VA.getValNo()].getOrigArgIndex() - CurArgIdx);
3113 CurArgIdx = Ins[VA.getValNo()].getOrigArgIndex();
3115 // Arguments stored in registers.
3116 if (VA.isRegLoc()) {
3117 EVT RegVT = VA.getLocVT();
3119 if (VA.needsCustom()) {
3120 // f64 and vector types are split up into multiple registers or
3121 // combinations of registers and stack slots.
3122 if (VA.getLocVT() == MVT::v2f64) {
3123 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
3125 VA = ArgLocs[++i]; // skip ahead to next loc
3127 if (VA.isMemLoc()) {
3128 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true);
3129 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
3130 ArgValue2 = DAG.getLoad(
3131 MVT::f64, dl, Chain, FIN,
3132 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI),
3133 false, false, false, 0);
3135 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
3138 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
3139 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
3140 ArgValue, ArgValue1,
3141 DAG.getIntPtrConstant(0, dl));
3142 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
3143 ArgValue, ArgValue2,
3144 DAG.getIntPtrConstant(1, dl));
3146 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
3149 const TargetRegisterClass *RC;
3151 if (RegVT == MVT::f32)
3152 RC = &ARM::SPRRegClass;
3153 else if (RegVT == MVT::f64)
3154 RC = &ARM::DPRRegClass;
3155 else if (RegVT == MVT::v2f64)
3156 RC = &ARM::QPRRegClass;
3157 else if (RegVT == MVT::i32)
3158 RC = AFI->isThumb1OnlyFunction() ? &ARM::tGPRRegClass
3159 : &ARM::GPRRegClass;
3161 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
3163 // Transform the arguments in physical registers into virtual ones.
3164 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
3165 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
3168 // If this is an 8 or 16-bit value, it is really passed promoted
3169 // to 32 bits. Insert an assert[sz]ext to capture this, then
3170 // truncate to the right size.
3171 switch (VA.getLocInfo()) {
3172 default: llvm_unreachable("Unknown loc info!");
3173 case CCValAssign::Full: break;
3174 case CCValAssign::BCvt:
3175 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
3177 case CCValAssign::SExt:
3178 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
3179 DAG.getValueType(VA.getValVT()));
3180 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
3182 case CCValAssign::ZExt:
3183 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
3184 DAG.getValueType(VA.getValVT()));
3185 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
3189 InVals.push_back(ArgValue);
3191 } else { // VA.isRegLoc()
3194 assert(VA.isMemLoc());
3195 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
3197 int index = VA.getValNo();
3199 // Some Ins[] entries become multiple ArgLoc[] entries.
3200 // Process them only once.
3201 if (index != lastInsIndex)
3203 ISD::ArgFlagsTy Flags = Ins[index].Flags;
3204 // FIXME: For now, all byval parameter objects are marked mutable.
3205 // This can be changed with more analysis.
3206 // In case of tail call optimization mark all arguments mutable.
3207 // Since they could be overwritten by lowering of arguments in case of
3209 if (Flags.isByVal()) {
3210 assert(Ins[index].isOrigArg() &&
3211 "Byval arguments cannot be implicit");
3212 unsigned CurByValIndex = CCInfo.getInRegsParamsProcessed();
3214 int FrameIndex = StoreByValRegs(
3215 CCInfo, DAG, dl, Chain, &*CurOrigArg, CurByValIndex,
3216 VA.getLocMemOffset(), Flags.getByValSize());
3217 InVals.push_back(DAG.getFrameIndex(FrameIndex, PtrVT));
3218 CCInfo.nextInRegsParam();
3220 unsigned FIOffset = VA.getLocMemOffset();
3221 int FI = MFI->CreateFixedObject(VA.getLocVT().getSizeInBits()/8,
3224 // Create load nodes to retrieve arguments from the stack.
3225 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
3226 InVals.push_back(DAG.getLoad(
3227 VA.getValVT(), dl, Chain, FIN,
3228 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI),
3229 false, false, false, 0));
3231 lastInsIndex = index;
3237 if (isVarArg && MFI->hasVAStart())
3238 VarArgStyleRegisters(CCInfo, DAG, dl, Chain,
3239 CCInfo.getNextStackOffset(),
3240 TotalArgRegsSaveSize);
3242 AFI->setArgumentStackSize(CCInfo.getNextStackOffset());
3247 /// isFloatingPointZero - Return true if this is +0.0.
3248 static bool isFloatingPointZero(SDValue Op) {
3249 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
3250 return CFP->getValueAPF().isPosZero();
3251 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
3252 // Maybe this has already been legalized into the constant pool?
3253 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
3254 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
3255 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
3256 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
3257 return CFP->getValueAPF().isPosZero();
3259 } else if (Op->getOpcode() == ISD::BITCAST &&
3260 Op->getValueType(0) == MVT::f64) {
3261 // Handle (ISD::BITCAST (ARMISD::VMOVIMM (ISD::TargetConstant 0)) MVT::f64)
3262 // created by LowerConstantFP().
3263 SDValue BitcastOp = Op->getOperand(0);
3264 if (BitcastOp->getOpcode() == ARMISD::VMOVIMM) {
3265 SDValue MoveOp = BitcastOp->getOperand(0);
3266 if (MoveOp->getOpcode() == ISD::TargetConstant &&
3267 cast<ConstantSDNode>(MoveOp)->getZExtValue() == 0) {
3275 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
3276 /// the given operands.
3278 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
3279 SDValue &ARMcc, SelectionDAG &DAG,
3281 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
3282 unsigned C = RHSC->getZExtValue();
3283 if (!isLegalICmpImmediate(C)) {
3284 // Constant does not fit, try adjusting it by one?
3289 if (C != 0x80000000 && isLegalICmpImmediate(C-1)) {
3290 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
3291 RHS = DAG.getConstant(C - 1, dl, MVT::i32);
3296 if (C != 0 && isLegalICmpImmediate(C-1)) {
3297 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
3298 RHS = DAG.getConstant(C - 1, dl, MVT::i32);
3303 if (C != 0x7fffffff && isLegalICmpImmediate(C+1)) {
3304 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
3305 RHS = DAG.getConstant(C + 1, dl, MVT::i32);
3310 if (C != 0xffffffff && isLegalICmpImmediate(C+1)) {
3311 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
3312 RHS = DAG.getConstant(C + 1, dl, MVT::i32);
3319 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3320 ARMISD::NodeType CompareType;
3323 CompareType = ARMISD::CMP;
3328 CompareType = ARMISD::CMPZ;
3331 ARMcc = DAG.getConstant(CondCode, dl, MVT::i32);
3332 return DAG.getNode(CompareType, dl, MVT::Glue, LHS, RHS);
3335 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
3337 ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
3339 assert(!Subtarget->isFPOnlySP() || RHS.getValueType() != MVT::f64);
3341 if (!isFloatingPointZero(RHS))
3342 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Glue, LHS, RHS);
3344 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Glue, LHS);
3345 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Glue, Cmp);
3348 /// duplicateCmp - Glue values can have only one use, so this function
3349 /// duplicates a comparison node.
3351 ARMTargetLowering::duplicateCmp(SDValue Cmp, SelectionDAG &DAG) const {
3352 unsigned Opc = Cmp.getOpcode();
3354 if (Opc == ARMISD::CMP || Opc == ARMISD::CMPZ)
3355 return DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
3357 assert(Opc == ARMISD::FMSTAT && "unexpected comparison operation");
3358 Cmp = Cmp.getOperand(0);
3359 Opc = Cmp.getOpcode();
3360 if (Opc == ARMISD::CMPFP)
3361 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
3363 assert(Opc == ARMISD::CMPFPw0 && "unexpected operand of FMSTAT");
3364 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0));
3366 return DAG.getNode(ARMISD::FMSTAT, DL, MVT::Glue, Cmp);
3369 std::pair<SDValue, SDValue>
3370 ARMTargetLowering::getARMXALUOOp(SDValue Op, SelectionDAG &DAG,
3371 SDValue &ARMcc) const {
3372 assert(Op.getValueType() == MVT::i32 && "Unsupported value type");
3374 SDValue Value, OverflowCmp;
3375 SDValue LHS = Op.getOperand(0);
3376 SDValue RHS = Op.getOperand(1);
3379 // FIXME: We are currently always generating CMPs because we don't support
3380 // generating CMN through the backend. This is not as good as the natural
3381 // CMP case because it causes a register dependency and cannot be folded
3384 switch (Op.getOpcode()) {
3386 llvm_unreachable("Unknown overflow instruction!");
3388 ARMcc = DAG.getConstant(ARMCC::VC, dl, MVT::i32);
3389 Value = DAG.getNode(ISD::ADD, dl, Op.getValueType(), LHS, RHS);
3390 OverflowCmp = DAG.getNode(ARMISD::CMP, dl, MVT::Glue, Value, LHS);
3393 ARMcc = DAG.getConstant(ARMCC::HS, dl, MVT::i32);
3394 Value = DAG.getNode(ISD::ADD, dl, Op.getValueType(), LHS, RHS);
3395 OverflowCmp = DAG.getNode(ARMISD::CMP, dl, MVT::Glue, Value, LHS);
3398 ARMcc = DAG.getConstant(ARMCC::VC, dl, MVT::i32);
3399 Value = DAG.getNode(ISD::SUB, dl, Op.getValueType(), LHS, RHS);
3400 OverflowCmp = DAG.getNode(ARMISD::CMP, dl, MVT::Glue, LHS, RHS);
3403 ARMcc = DAG.getConstant(ARMCC::HS, dl, MVT::i32);
3404 Value = DAG.getNode(ISD::SUB, dl, Op.getValueType(), LHS, RHS);
3405 OverflowCmp = DAG.getNode(ARMISD::CMP, dl, MVT::Glue, LHS, RHS);
3409 return std::make_pair(Value, OverflowCmp);
3414 ARMTargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) const {
3415 // Let legalize expand this if it isn't a legal type yet.
3416 if (!DAG.getTargetLoweringInfo().isTypeLegal(Op.getValueType()))
3419 SDValue Value, OverflowCmp;
3421 std::tie(Value, OverflowCmp) = getARMXALUOOp(Op, DAG, ARMcc);
3422 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3424 // We use 0 and 1 as false and true values.
3425 SDValue TVal = DAG.getConstant(1, dl, MVT::i32);
3426 SDValue FVal = DAG.getConstant(0, dl, MVT::i32);
3427 EVT VT = Op.getValueType();
3429 SDValue Overflow = DAG.getNode(ARMISD::CMOV, dl, VT, TVal, FVal,
3430 ARMcc, CCR, OverflowCmp);
3432 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
3433 return DAG.getNode(ISD::MERGE_VALUES, dl, VTs, Value, Overflow);
3437 SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
3438 SDValue Cond = Op.getOperand(0);
3439 SDValue SelectTrue = Op.getOperand(1);
3440 SDValue SelectFalse = Op.getOperand(2);
3442 unsigned Opc = Cond.getOpcode();
3444 if (Cond.getResNo() == 1 &&
3445 (Opc == ISD::SADDO || Opc == ISD::UADDO || Opc == ISD::SSUBO ||
3446 Opc == ISD::USUBO)) {
3447 if (!DAG.getTargetLoweringInfo().isTypeLegal(Cond->getValueType(0)))
3450 SDValue Value, OverflowCmp;
3452 std::tie(Value, OverflowCmp) = getARMXALUOOp(Cond, DAG, ARMcc);
3453 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3454 EVT VT = Op.getValueType();
3456 return getCMOV(dl, VT, SelectTrue, SelectFalse, ARMcc, CCR,
3462 // (select (cmov 1, 0, cond), t, f) -> (cmov t, f, cond)
3463 // (select (cmov 0, 1, cond), t, f) -> (cmov f, t, cond)
3465 if (Cond.getOpcode() == ARMISD::CMOV && Cond.hasOneUse()) {
3466 const ConstantSDNode *CMOVTrue =
3467 dyn_cast<ConstantSDNode>(Cond.getOperand(0));
3468 const ConstantSDNode *CMOVFalse =
3469 dyn_cast<ConstantSDNode>(Cond.getOperand(1));
3471 if (CMOVTrue && CMOVFalse) {
3472 unsigned CMOVTrueVal = CMOVTrue->getZExtValue();
3473 unsigned CMOVFalseVal = CMOVFalse->getZExtValue();
3477 if (CMOVTrueVal == 1 && CMOVFalseVal == 0) {
3479 False = SelectFalse;
3480 } else if (CMOVTrueVal == 0 && CMOVFalseVal == 1) {
3485 if (True.getNode() && False.getNode()) {
3486 EVT VT = Op.getValueType();
3487 SDValue ARMcc = Cond.getOperand(2);
3488 SDValue CCR = Cond.getOperand(3);
3489 SDValue Cmp = duplicateCmp(Cond.getOperand(4), DAG);
3490 assert(True.getValueType() == VT);
3491 return getCMOV(dl, VT, True, False, ARMcc, CCR, Cmp, DAG);
3496 // ARM's BooleanContents value is UndefinedBooleanContent. Mask out the
3497 // undefined bits before doing a full-word comparison with zero.
3498 Cond = DAG.getNode(ISD::AND, dl, Cond.getValueType(), Cond,
3499 DAG.getConstant(1, dl, Cond.getValueType()));
3501 return DAG.getSelectCC(dl, Cond,
3502 DAG.getConstant(0, dl, Cond.getValueType()),
3503 SelectTrue, SelectFalse, ISD::SETNE);
3506 static void checkVSELConstraints(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
3507 bool &swpCmpOps, bool &swpVselOps) {
3508 // Start by selecting the GE condition code for opcodes that return true for
3510 if (CC == ISD::SETUGE || CC == ISD::SETOGE || CC == ISD::SETOLE ||
3512 CondCode = ARMCC::GE;
3514 // and GT for opcodes that return false for 'equality'.
3515 else if (CC == ISD::SETUGT || CC == ISD::SETOGT || CC == ISD::SETOLT ||
3517 CondCode = ARMCC::GT;
3519 // Since we are constrained to GE/GT, if the opcode contains 'less', we need
3520 // to swap the compare operands.
3521 if (CC == ISD::SETOLE || CC == ISD::SETULE || CC == ISD::SETOLT ||
3525 // Both GT and GE are ordered comparisons, and return false for 'unordered'.
3526 // If we have an unordered opcode, we need to swap the operands to the VSEL
3527 // instruction (effectively negating the condition).
3529 // This also has the effect of swapping which one of 'less' or 'greater'
3530 // returns true, so we also swap the compare operands. It also switches
3531 // whether we return true for 'equality', so we compensate by picking the
3532 // opposite condition code to our original choice.
3533 if (CC == ISD::SETULE || CC == ISD::SETULT || CC == ISD::SETUGE ||
3534 CC == ISD::SETUGT) {
3535 swpCmpOps = !swpCmpOps;
3536 swpVselOps = !swpVselOps;
3537 CondCode = CondCode == ARMCC::GT ? ARMCC::GE : ARMCC::GT;
3540 // 'ordered' is 'anything but unordered', so use the VS condition code and
3541 // swap the VSEL operands.
3542 if (CC == ISD::SETO) {
3543 CondCode = ARMCC::VS;
3547 // 'unordered or not equal' is 'anything but equal', so use the EQ condition
3548 // code and swap the VSEL operands.
3549 if (CC == ISD::SETUNE) {
3550 CondCode = ARMCC::EQ;
3555 SDValue ARMTargetLowering::getCMOV(SDLoc dl, EVT VT, SDValue FalseVal,
3556 SDValue TrueVal, SDValue ARMcc, SDValue CCR,
3557 SDValue Cmp, SelectionDAG &DAG) const {
3558 if (Subtarget->isFPOnlySP() && VT == MVT::f64) {
3559 FalseVal = DAG.getNode(ARMISD::VMOVRRD, dl,
3560 DAG.getVTList(MVT::i32, MVT::i32), FalseVal);
3561 TrueVal = DAG.getNode(ARMISD::VMOVRRD, dl,
3562 DAG.getVTList(MVT::i32, MVT::i32), TrueVal);
3564 SDValue TrueLow = TrueVal.getValue(0);
3565 SDValue TrueHigh = TrueVal.getValue(1);
3566 SDValue FalseLow = FalseVal.getValue(0);
3567 SDValue FalseHigh = FalseVal.getValue(1);
3569 SDValue Low = DAG.getNode(ARMISD::CMOV, dl, MVT::i32, FalseLow, TrueLow,
3571 SDValue High = DAG.getNode(ARMISD::CMOV, dl, MVT::i32, FalseHigh, TrueHigh,
3572 ARMcc, CCR, duplicateCmp(Cmp, DAG));
3574 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Low, High);
3576 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,
3581 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
3582 EVT VT = Op.getValueType();
3583 SDValue LHS = Op.getOperand(0);
3584 SDValue RHS = Op.getOperand(1);
3585 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
3586 SDValue TrueVal = Op.getOperand(2);
3587 SDValue FalseVal = Op.getOperand(3);
3590 if (Subtarget->isFPOnlySP() && LHS.getValueType() == MVT::f64) {
3591 DAG.getTargetLoweringInfo().softenSetCCOperands(DAG, MVT::f64, LHS, RHS, CC,
3594 // If softenSetCCOperands only returned one value, we should compare it to
3596 if (!RHS.getNode()) {
3597 RHS = DAG.getConstant(0, dl, LHS.getValueType());
3602 if (LHS.getValueType() == MVT::i32) {
3603 // Try to generate VSEL on ARMv8.
3604 // The VSEL instruction can't use all the usual ARM condition
3605 // codes: it only has two bits to select the condition code, so it's
3606 // constrained to use only GE, GT, VS and EQ.
3608 // To implement all the various ISD::SETXXX opcodes, we sometimes need to
3609 // swap the operands of the previous compare instruction (effectively
3610 // inverting the compare condition, swapping 'less' and 'greater') and
3611 // sometimes need to swap the operands to the VSEL (which inverts the
3612 // condition in the sense of firing whenever the previous condition didn't)
3613 if (Subtarget->hasFPARMv8() && (TrueVal.getValueType() == MVT::f32 ||
3614 TrueVal.getValueType() == MVT::f64)) {
3615 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3616 if (CondCode == ARMCC::LT || CondCode == ARMCC::LE ||
3617 CondCode == ARMCC::VC || CondCode == ARMCC::NE) {
3618 CC = ISD::getSetCCInverse(CC, true);
3619 std::swap(TrueVal, FalseVal);
3624 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3625 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3626 return getCMOV(dl, VT, FalseVal, TrueVal, ARMcc, CCR, Cmp, DAG);
3629 ARMCC::CondCodes CondCode, CondCode2;
3630 FPCCToARMCC(CC, CondCode, CondCode2);
3632 // Try to generate VMAXNM/VMINNM on ARMv8.
3633 if (Subtarget->hasFPARMv8() && (TrueVal.getValueType() == MVT::f32 ||
3634 TrueVal.getValueType() == MVT::f64)) {
3635 bool swpCmpOps = false;
3636 bool swpVselOps = false;
3637 checkVSELConstraints(CC, CondCode, swpCmpOps, swpVselOps);
3639 if (CondCode == ARMCC::GT || CondCode == ARMCC::GE ||
3640 CondCode == ARMCC::VS || CondCode == ARMCC::EQ) {
3642 std::swap(LHS, RHS);
3644 std::swap(TrueVal, FalseVal);
3648 SDValue ARMcc = DAG.getConstant(CondCode, dl, MVT::i32);
3649 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
3650 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3651 SDValue Result = getCMOV(dl, VT, FalseVal, TrueVal, ARMcc, CCR, Cmp, DAG);
3652 if (CondCode2 != ARMCC::AL) {
3653 SDValue ARMcc2 = DAG.getConstant(CondCode2, dl, MVT::i32);
3654 // FIXME: Needs another CMP because flag can have but one use.
3655 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
3656 Result = getCMOV(dl, VT, Result, TrueVal, ARMcc2, CCR, Cmp2, DAG);
3661 /// canChangeToInt - Given the fp compare operand, return true if it is suitable
3662 /// to morph to an integer compare sequence.
3663 static bool canChangeToInt(SDValue Op, bool &SeenZero,
3664 const ARMSubtarget *Subtarget) {
3665 SDNode *N = Op.getNode();
3666 if (!N->hasOneUse())
3667 // Otherwise it requires moving the value from fp to integer registers.
3669 if (!N->getNumValues())
3671 EVT VT = Op.getValueType();
3672 if (VT != MVT::f32 && !Subtarget->isFPBrccSlow())
3673 // f32 case is generally profitable. f64 case only makes sense when vcmpe +
3674 // vmrs are very slow, e.g. cortex-a8.
3677 if (isFloatingPointZero(Op)) {
3681 return ISD::isNormalLoad(N);
3684 static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) {
3685 if (isFloatingPointZero(Op))
3686 return DAG.getConstant(0, SDLoc(Op), MVT::i32);
3688 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op))
3689 return DAG.getLoad(MVT::i32, SDLoc(Op),
3690 Ld->getChain(), Ld->getBasePtr(), Ld->getPointerInfo(),
3691 Ld->isVolatile(), Ld->isNonTemporal(),
3692 Ld->isInvariant(), Ld->getAlignment());
3694 llvm_unreachable("Unknown VFP cmp argument!");
3697 static void expandf64Toi32(SDValue Op, SelectionDAG &DAG,
3698 SDValue &RetVal1, SDValue &RetVal2) {
3701 if (isFloatingPointZero(Op)) {
3702 RetVal1 = DAG.getConstant(0, dl, MVT::i32);
3703 RetVal2 = DAG.getConstant(0, dl, MVT::i32);
3707 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) {
3708 SDValue Ptr = Ld->getBasePtr();
3709 RetVal1 = DAG.getLoad(MVT::i32, dl,
3710 Ld->getChain(), Ptr,
3711 Ld->getPointerInfo(),
3712 Ld->isVolatile(), Ld->isNonTemporal(),
3713 Ld->isInvariant(), Ld->getAlignment());
3715 EVT PtrType = Ptr.getValueType();
3716 unsigned NewAlign = MinAlign(Ld->getAlignment(), 4);
3717 SDValue NewPtr = DAG.getNode(ISD::ADD, dl,
3718 PtrType, Ptr, DAG.getConstant(4, dl, PtrType));
3719 RetVal2 = DAG.getLoad(MVT::i32, dl,
3720 Ld->getChain(), NewPtr,
3721 Ld->getPointerInfo().getWithOffset(4),
3722 Ld->isVolatile(), Ld->isNonTemporal(),
3723 Ld->isInvariant(), NewAlign);
3727 llvm_unreachable("Unknown VFP cmp argument!");
3730 /// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some
3731 /// f32 and even f64 comparisons to integer ones.
3733 ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const {
3734 SDValue Chain = Op.getOperand(0);
3735 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3736 SDValue LHS = Op.getOperand(2);
3737 SDValue RHS = Op.getOperand(3);
3738 SDValue Dest = Op.getOperand(4);
3741 bool LHSSeenZero = false;
3742 bool LHSOk = canChangeToInt(LHS, LHSSeenZero, Subtarget);
3743 bool RHSSeenZero = false;
3744 bool RHSOk = canChangeToInt(RHS, RHSSeenZero, Subtarget);
3745 if (LHSOk && RHSOk && (LHSSeenZero || RHSSeenZero)) {
3746 // If unsafe fp math optimization is enabled and there are no other uses of
3747 // the CMP operands, and the condition code is EQ or NE, we can optimize it
3748 // to an integer comparison.
3749 if (CC == ISD::SETOEQ)
3751 else if (CC == ISD::SETUNE)
3754 SDValue Mask = DAG.getConstant(0x7fffffff, dl, MVT::i32);
3756 if (LHS.getValueType() == MVT::f32) {
3757 LHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3758 bitcastf32Toi32(LHS, DAG), Mask);
3759 RHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3760 bitcastf32Toi32(RHS, DAG), Mask);
3761 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3762 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3763 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3764 Chain, Dest, ARMcc, CCR, Cmp);
3769 expandf64Toi32(LHS, DAG, LHS1, LHS2);
3770 expandf64Toi32(RHS, DAG, RHS1, RHS2);
3771 LHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, LHS2, Mask);
3772 RHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, RHS2, Mask);
3773 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3774 ARMcc = DAG.getConstant(CondCode, dl, MVT::i32);
3775 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3776 SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest };
3777 return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops);
3783 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
3784 SDValue Chain = Op.getOperand(0);
3785 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3786 SDValue LHS = Op.getOperand(2);
3787 SDValue RHS = Op.getOperand(3);
3788 SDValue Dest = Op.getOperand(4);
3791 if (Subtarget->isFPOnlySP() && LHS.getValueType() == MVT::f64) {
3792 DAG.getTargetLoweringInfo().softenSetCCOperands(DAG, MVT::f64, LHS, RHS, CC,
3795 // If softenSetCCOperands only returned one value, we should compare it to
3797 if (!RHS.getNode()) {
3798 RHS = DAG.getConstant(0, dl, LHS.getValueType());
3803 if (LHS.getValueType() == MVT::i32) {
3805 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3806 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3807 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3808 Chain, Dest, ARMcc, CCR, Cmp);
3811 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3813 if (getTargetMachine().Options.UnsafeFPMath &&
3814 (CC == ISD::SETEQ || CC == ISD::SETOEQ ||
3815 CC == ISD::SETNE || CC == ISD::SETUNE)) {
3816 SDValue Result = OptimizeVFPBrcond(Op, DAG);
3817 if (Result.getNode())
3821 ARMCC::CondCodes CondCode, CondCode2;
3822 FPCCToARMCC(CC, CondCode, CondCode2);
3824 SDValue ARMcc = DAG.getConstant(CondCode, dl, MVT::i32);
3825 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
3826 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3827 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3828 SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp };
3829 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops);
3830 if (CondCode2 != ARMCC::AL) {
3831 ARMcc = DAG.getConstant(CondCode2, dl, MVT::i32);
3832 SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) };
3833 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops);
3838 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
3839 SDValue Chain = Op.getOperand(0);
3840 SDValue Table = Op.getOperand(1);
3841 SDValue Index = Op.getOperand(2);
3844 EVT PTy = getPointerTy(DAG.getDataLayout());
3845 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
3846 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
3847 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI);
3848 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, dl, PTy));
3849 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
3850 if (Subtarget->isThumb2()) {
3851 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
3852 // which does another jump to the destination. This also makes it easier
3853 // to translate it to TBB / TBH later.
3854 // FIXME: This might not work if the function is extremely large.
3855 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
3856 Addr, Op.getOperand(2), JTI);
3858 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
3860 DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
3861 MachinePointerInfo::getJumpTable(DAG.getMachineFunction()),
3862 false, false, false, 0);
3863 Chain = Addr.getValue(1);
3864 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
3865 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI);
3868 DAG.getLoad(PTy, dl, Chain, Addr,
3869 MachinePointerInfo::getJumpTable(DAG.getMachineFunction()),
3870 false, false, false, 0);
3871 Chain = Addr.getValue(1);
3872 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI);
3876 static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3877 EVT VT = Op.getValueType();
3880 if (Op.getValueType().getVectorElementType() == MVT::i32) {
3881 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::f32)
3883 return DAG.UnrollVectorOp(Op.getNode());
3886 assert(Op.getOperand(0).getValueType() == MVT::v4f32 &&
3887 "Invalid type for custom lowering!");
3888 if (VT != MVT::v4i16)
3889 return DAG.UnrollVectorOp(Op.getNode());
3891 Op = DAG.getNode(Op.getOpcode(), dl, MVT::v4i32, Op.getOperand(0));
3892 return DAG.getNode(ISD::TRUNCATE, dl, VT, Op);
3895 SDValue ARMTargetLowering::LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) const {
3896 EVT VT = Op.getValueType();
3898 return LowerVectorFP_TO_INT(Op, DAG);
3899 if (Subtarget->isFPOnlySP() && Op.getOperand(0).getValueType() == MVT::f64) {
3901 if (Op.getOpcode() == ISD::FP_TO_SINT)
3902 LC = RTLIB::getFPTOSINT(Op.getOperand(0).getValueType(),
3905 LC = RTLIB::getFPTOUINT(Op.getOperand(0).getValueType(),
3907 return makeLibCall(DAG, LC, Op.getValueType(), Op.getOperand(0),
3908 /*isSigned*/ false, SDLoc(Op)).first;
3914 static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3915 EVT VT = Op.getValueType();
3918 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::i32) {
3919 if (VT.getVectorElementType() == MVT::f32)
3921 return DAG.UnrollVectorOp(Op.getNode());
3924 assert(Op.getOperand(0).getValueType() == MVT::v4i16 &&
3925 "Invalid type for custom lowering!");
3926 if (VT != MVT::v4f32)
3927 return DAG.UnrollVectorOp(Op.getNode());
3931 switch (Op.getOpcode()) {
3932 default: llvm_unreachable("Invalid opcode!");
3933 case ISD::SINT_TO_FP:
3934 CastOpc = ISD::SIGN_EXTEND;
3935 Opc = ISD::SINT_TO_FP;
3937 case ISD::UINT_TO_FP:
3938 CastOpc = ISD::ZERO_EXTEND;
3939 Opc = ISD::UINT_TO_FP;
3943 Op = DAG.getNode(CastOpc, dl, MVT::v4i32, Op.getOperand(0));
3944 return DAG.getNode(Opc, dl, VT, Op);
3947 SDValue ARMTargetLowering::LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) const {
3948 EVT VT = Op.getValueType();
3950 return LowerVectorINT_TO_FP(Op, DAG);
3951 if (Subtarget->isFPOnlySP() && Op.getValueType() == MVT::f64) {
3953 if (Op.getOpcode() == ISD::SINT_TO_FP)
3954 LC = RTLIB::getSINTTOFP(Op.getOperand(0).getValueType(),
3957 LC = RTLIB::getUINTTOFP(Op.getOperand(0).getValueType(),
3959 return makeLibCall(DAG, LC, Op.getValueType(), Op.getOperand(0),
3960 /*isSigned*/ false, SDLoc(Op)).first;
3966 SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
3967 // Implement fcopysign with a fabs and a conditional fneg.
3968 SDValue Tmp0 = Op.getOperand(0);
3969 SDValue Tmp1 = Op.getOperand(1);
3971 EVT VT = Op.getValueType();
3972 EVT SrcVT = Tmp1.getValueType();
3973 bool InGPR = Tmp0.getOpcode() == ISD::BITCAST ||
3974 Tmp0.getOpcode() == ARMISD::VMOVDRR;
3975 bool UseNEON = !InGPR && Subtarget->hasNEON();
3978 // Use VBSL to copy the sign bit.
3979 unsigned EncodedVal = ARM_AM::createNEONModImm(0x6, 0x80);
3980 SDValue Mask = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v2i32,
3981 DAG.getTargetConstant(EncodedVal, dl, MVT::i32));
3982 EVT OpVT = (VT == MVT::f32) ? MVT::v2i32 : MVT::v1i64;
3984 Mask = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3985 DAG.getNode(ISD::BITCAST, dl, OpVT, Mask),
3986 DAG.getConstant(32, dl, MVT::i32));
3987 else /*if (VT == MVT::f32)*/
3988 Tmp0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp0);
3989 if (SrcVT == MVT::f32) {
3990 Tmp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp1);
3992 Tmp1 = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3993 DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1),
3994 DAG.getConstant(32, dl, MVT::i32));
3995 } else if (VT == MVT::f32)
3996 Tmp1 = DAG.getNode(ARMISD::VSHRu, dl, MVT::v1i64,
3997 DAG.getNode(ISD::BITCAST, dl, MVT::v1i64, Tmp1),
3998 DAG.getConstant(32, dl, MVT::i32));
3999 Tmp0 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp0);
4000 Tmp1 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1);
4002 SDValue AllOnes = DAG.getTargetConstant(ARM_AM::createNEONModImm(0xe, 0xff),
4004 AllOnes = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v8i8, AllOnes);
4005 SDValue MaskNot = DAG.getNode(ISD::XOR, dl, OpVT, Mask,
4006 DAG.getNode(ISD::BITCAST, dl, OpVT, AllOnes));
4008 SDValue Res = DAG.getNode(ISD::OR, dl, OpVT,
4009 DAG.getNode(ISD::AND, dl, OpVT, Tmp1, Mask),
4010 DAG.getNode(ISD::AND, dl, OpVT, Tmp0, MaskNot));
4011 if (VT == MVT::f32) {
4012 Res = DAG.getNode(ISD::BITCAST, dl, MVT::v2f32, Res);
4013 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, Res,
4014 DAG.getConstant(0, dl, MVT::i32));
4016 Res = DAG.getNode(ISD::BITCAST, dl, MVT::f64, Res);
4022 // Bitcast operand 1 to i32.
4023 if (SrcVT == MVT::f64)
4024 Tmp1 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
4026 Tmp1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp1);
4028 // Or in the signbit with integer operations.
4029 SDValue Mask1 = DAG.getConstant(0x80000000, dl, MVT::i32);
4030 SDValue Mask2 = DAG.getConstant(0x7fffffff, dl, MVT::i32);
4031 Tmp1 = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp1, Mask1);
4032 if (VT == MVT::f32) {
4033 Tmp0 = DAG.getNode(ISD::AND, dl, MVT::i32,
4034 DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp0), Mask2);
4035 return DAG.getNode(ISD::BITCAST, dl, MVT::f32,
4036 DAG.getNode(ISD::OR, dl, MVT::i32, Tmp0, Tmp1));
4039 // f64: Or the high part with signbit and then combine two parts.
4040 Tmp0 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
4042 SDValue Lo = Tmp0.getValue(0);
4043 SDValue Hi = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp0.getValue(1), Mask2);
4044 Hi = DAG.getNode(ISD::OR, dl, MVT::i32, Hi, Tmp1);
4045 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
4048 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
4049 MachineFunction &MF = DAG.getMachineFunction();
4050 MachineFrameInfo *MFI = MF.getFrameInfo();
4051 MFI->setReturnAddressIsTaken(true);
4053 if (verifyReturnAddressArgumentIsConstant(Op, DAG))
4056 EVT VT = Op.getValueType();
4058 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
4060 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
4061 SDValue Offset = DAG.getConstant(4, dl, MVT::i32);
4062 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
4063 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
4064 MachinePointerInfo(), false, false, false, 0);
4067 // Return LR, which contains the return address. Mark it an implicit live-in.
4068 unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32));
4069 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
4072 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
4073 const ARMBaseRegisterInfo &ARI =
4074 *static_cast<const ARMBaseRegisterInfo*>(RegInfo);
4075 MachineFunction &MF = DAG.getMachineFunction();
4076 MachineFrameInfo *MFI = MF.getFrameInfo();
4077 MFI->setFrameAddressIsTaken(true);
4079 EVT VT = Op.getValueType();
4080 SDLoc dl(Op); // FIXME probably not meaningful
4081 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
4082 unsigned FrameReg = ARI.getFrameRegister(MF);
4083 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
4085 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
4086 MachinePointerInfo(),
4087 false, false, false, 0);
4091 // FIXME? Maybe this could be a TableGen attribute on some registers and
4092 // this table could be generated automatically from RegInfo.
4093 unsigned ARMTargetLowering::getRegisterByName(const char* RegName, EVT VT,
4094 SelectionDAG &DAG) const {
4095 unsigned Reg = StringSwitch<unsigned>(RegName)
4096 .Case("sp", ARM::SP)
4100 report_fatal_error(Twine("Invalid register name \""
4101 + StringRef(RegName) + "\"."));
4104 // Result is 64 bit value so split into two 32 bit values and return as a
4106 static void ExpandREAD_REGISTER(SDNode *N, SmallVectorImpl<SDValue> &Results,
4107 SelectionDAG &DAG) {
4110 // This function is only supposed to be called for i64 type destination.
4111 assert(N->getValueType(0) == MVT::i64
4112 && "ExpandREAD_REGISTER called for non-i64 type result.");
4114 SDValue Read = DAG.getNode(ISD::READ_REGISTER, DL,
4115 DAG.getVTList(MVT::i32, MVT::i32, MVT::Other),
4119 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Read.getValue(0),
4121 Results.push_back(Read.getOperand(0));
4124 /// ExpandBITCAST - If the target supports VFP, this function is called to
4125 /// expand a bit convert where either the source or destination type is i64 to
4126 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
4127 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
4128 /// vectors), since the legalizer won't know what to do with that.
4129 static SDValue ExpandBITCAST(SDNode *N, SelectionDAG &DAG) {
4130 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4132 SDValue Op = N->getOperand(0);
4134 // This function is only supposed to be called for i64 types, either as the
4135 // source or destination of the bit convert.
4136 EVT SrcVT = Op.getValueType();
4137 EVT DstVT = N->getValueType(0);
4138 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
4139 "ExpandBITCAST called for non-i64 type");
4141 // Turn i64->f64 into VMOVDRR.
4142 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
4143 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
4144 DAG.getConstant(0, dl, MVT::i32));
4145 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
4146 DAG.getConstant(1, dl, MVT::i32));
4147 return DAG.getNode(ISD::BITCAST, dl, DstVT,
4148 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
4151 // Turn f64->i64 into VMOVRRD.
4152 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
4154 if (DAG.getDataLayout().isBigEndian() && SrcVT.isVector() &&
4155 SrcVT.getVectorNumElements() > 1)
4156 Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
4157 DAG.getVTList(MVT::i32, MVT::i32),
4158 DAG.getNode(ARMISD::VREV64, dl, SrcVT, Op));
4160 Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
4161 DAG.getVTList(MVT::i32, MVT::i32), Op);
4162 // Merge the pieces into a single i64 value.
4163 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
4169 /// getZeroVector - Returns a vector of specified type with all zero elements.
4170 /// Zero vectors are used to represent vector negation and in those cases
4171 /// will be implemented with the NEON VNEG instruction. However, VNEG does
4172 /// not support i64 elements, so sometimes the zero vectors will need to be
4173 /// explicitly constructed. Regardless, use a canonical VMOV to create the
4175 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, SDLoc dl) {
4176 assert(VT.isVector() && "Expected a vector type");
4177 // The canonical modified immediate encoding of a zero vector is....0!
4178 SDValue EncodedVal = DAG.getTargetConstant(0, dl, MVT::i32);
4179 EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
4180 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal);
4181 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4184 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
4185 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
4186 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
4187 SelectionDAG &DAG) const {
4188 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4189 EVT VT = Op.getValueType();
4190 unsigned VTBits = VT.getSizeInBits();
4192 SDValue ShOpLo = Op.getOperand(0);
4193 SDValue ShOpHi = Op.getOperand(1);
4194 SDValue ShAmt = Op.getOperand(2);
4196 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
4198 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
4200 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
4201 DAG.getConstant(VTBits, dl, MVT::i32), ShAmt);
4202 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
4203 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
4204 DAG.getConstant(VTBits, dl, MVT::i32));
4205 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
4206 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
4207 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
4209 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
4210 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, dl, MVT::i32),
4211 ISD::SETGE, ARMcc, DAG, dl);
4212 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
4213 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc,
4216 SDValue Ops[2] = { Lo, Hi };
4217 return DAG.getMergeValues(Ops, dl);
4220 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
4221 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
4222 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
4223 SelectionDAG &DAG) const {
4224 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4225 EVT VT = Op.getValueType();
4226 unsigned VTBits = VT.getSizeInBits();
4228 SDValue ShOpLo = Op.getOperand(0);
4229 SDValue ShOpHi = Op.getOperand(1);
4230 SDValue ShAmt = Op.getOperand(2);
4233 assert(Op.getOpcode() == ISD::SHL_PARTS);
4234 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
4235 DAG.getConstant(VTBits, dl, MVT::i32), ShAmt);
4236 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
4237 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
4238 DAG.getConstant(VTBits, dl, MVT::i32));
4239 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
4240 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
4242 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
4243 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
4244 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, dl, MVT::i32),
4245 ISD::SETGE, ARMcc, DAG, dl);
4246 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
4247 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc,
4250 SDValue Ops[2] = { Lo, Hi };
4251 return DAG.getMergeValues(Ops, dl);
4254 SDValue ARMTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
4255 SelectionDAG &DAG) const {
4256 // The rounding mode is in bits 23:22 of the FPSCR.
4257 // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0
4258 // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3)
4259 // so that the shift + and get folded into a bitfield extract.
4261 SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::i32,
4262 DAG.getConstant(Intrinsic::arm_get_fpscr, dl,
4264 SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPSCR,
4265 DAG.getConstant(1U << 22, dl, MVT::i32));
4266 SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds,
4267 DAG.getConstant(22, dl, MVT::i32));
4268 return DAG.getNode(ISD::AND, dl, MVT::i32, RMODE,
4269 DAG.getConstant(3, dl, MVT::i32));
4272 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
4273 const ARMSubtarget *ST) {
4275 EVT VT = N->getValueType(0);
4276 if (VT.isVector()) {
4277 assert(ST->hasNEON());
4279 // Compute the least significant set bit: LSB = X & -X
4280 SDValue X = N->getOperand(0);
4281 SDValue NX = DAG.getNode(ISD::SUB, dl, VT, getZeroVector(VT, DAG, dl), X);
4282 SDValue LSB = DAG.getNode(ISD::AND, dl, VT, X, NX);
4284 EVT ElemTy = VT.getVectorElementType();
4286 if (ElemTy == MVT::i8) {
4287 // Compute with: cttz(x) = ctpop(lsb - 1)
4288 SDValue One = DAG.getNode(ARMISD::VMOVIMM, dl, VT,
4289 DAG.getTargetConstant(1, dl, ElemTy));
4290 SDValue Bits = DAG.getNode(ISD::SUB, dl, VT, LSB, One);
4291 return DAG.getNode(ISD::CTPOP, dl, VT, Bits);
4294 if ((ElemTy == MVT::i16 || ElemTy == MVT::i32) &&
4295 (N->getOpcode() == ISD::CTTZ_ZERO_UNDEF)) {
4296 // Compute with: cttz(x) = (width - 1) - ctlz(lsb), if x != 0
4297 unsigned NumBits = ElemTy.getSizeInBits();
4298 SDValue WidthMinus1 =
4299 DAG.getNode(ARMISD::VMOVIMM, dl, VT,
4300 DAG.getTargetConstant(NumBits - 1, dl, ElemTy));
4301 SDValue CTLZ = DAG.getNode(ISD::CTLZ, dl, VT, LSB);
4302 return DAG.getNode(ISD::SUB, dl, VT, WidthMinus1, CTLZ);
4305 // Compute with: cttz(x) = ctpop(lsb - 1)
4307 // Since we can only compute the number of bits in a byte with vcnt.8, we
4308 // have to gather the result with pairwise addition (vpaddl) for i16, i32,
4313 if (ElemTy == MVT::i64) {
4314 // Load constant 0xffff'ffff'ffff'ffff to register.
4315 SDValue FF = DAG.getNode(ARMISD::VMOVIMM, dl, VT,
4316 DAG.getTargetConstant(0x1eff, dl, MVT::i32));
4317 Bits = DAG.getNode(ISD::ADD, dl, VT, LSB, FF);
4319 SDValue One = DAG.getNode(ARMISD::VMOVIMM, dl, VT,
4320 DAG.getTargetConstant(1, dl, ElemTy));
4321 Bits = DAG.getNode(ISD::SUB, dl, VT, LSB, One);
4324 // Count #bits with vcnt.8.
4325 EVT VT8Bit = VT.is64BitVector() ? MVT::v8i8 : MVT::v16i8;
4326 SDValue BitsVT8 = DAG.getNode(ISD::BITCAST, dl, VT8Bit, Bits);
4327 SDValue Cnt8 = DAG.getNode(ISD::CTPOP, dl, VT8Bit, BitsVT8);
4329 // Gather the #bits with vpaddl (pairwise add.)
4330 EVT VT16Bit = VT.is64BitVector() ? MVT::v4i16 : MVT::v8i16;
4331 SDValue Cnt16 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT16Bit,
4332 DAG.getTargetConstant(Intrinsic::arm_neon_vpaddlu, dl, MVT::i32),
4334 if (ElemTy == MVT::i16)
4337 EVT VT32Bit = VT.is64BitVector() ? MVT::v2i32 : MVT::v4i32;
4338 SDValue Cnt32 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT32Bit,
4339 DAG.getTargetConstant(Intrinsic::arm_neon_vpaddlu, dl, MVT::i32),
4341 if (ElemTy == MVT::i32)
4344 assert(ElemTy == MVT::i64);
4345 SDValue Cnt64 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
4346 DAG.getTargetConstant(Intrinsic::arm_neon_vpaddlu, dl, MVT::i32),
4351 if (!ST->hasV6T2Ops())
4354 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
4355 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
4358 /// getCTPOP16BitCounts - Returns a v8i8/v16i8 vector containing the bit-count
4359 /// for each 16-bit element from operand, repeated. The basic idea is to
4360 /// leverage vcnt to get the 8-bit counts, gather and add the results.
4362 /// Trace for v4i16:
4363 /// input = [v0 v1 v2 v3 ] (vi 16-bit element)
4364 /// cast: N0 = [w0 w1 w2 w3 w4 w5 w6 w7] (v0 = [w0 w1], wi 8-bit element)
4365 /// vcnt: N1 = [b0 b1 b2 b3 b4 b5 b6 b7] (bi = bit-count of 8-bit element wi)
4366 /// vrev: N2 = [b1 b0 b3 b2 b5 b4 b7 b6]
4367 /// [b0 b1 b2 b3 b4 b5 b6 b7]
4368 /// +[b1 b0 b3 b2 b5 b4 b7 b6]
4369 /// N3=N1+N2 = [k0 k0 k1 k1 k2 k2 k3 k3] (k0 = b0+b1 = bit-count of 16-bit v0,
4370 /// vuzp: = [k0 k1 k2 k3 k0 k1 k2 k3] each ki is 8-bits)
4371 static SDValue getCTPOP16BitCounts(SDNode *N, SelectionDAG &DAG) {
4372 EVT VT = N->getValueType(0);
4375 EVT VT8Bit = VT.is64BitVector() ? MVT::v8i8 : MVT::v16i8;
4376 SDValue N0 = DAG.getNode(ISD::BITCAST, DL, VT8Bit, N->getOperand(0));
4377 SDValue N1 = DAG.getNode(ISD::CTPOP, DL, VT8Bit, N0);
4378 SDValue N2 = DAG.getNode(ARMISD::VREV16, DL, VT8Bit, N1);
4379 SDValue N3 = DAG.getNode(ISD::ADD, DL, VT8Bit, N1, N2);
4380 return DAG.getNode(ARMISD::VUZP, DL, VT8Bit, N3, N3);
4383 /// lowerCTPOP16BitElements - Returns a v4i16/v8i16 vector containing the
4384 /// bit-count for each 16-bit element from the operand. We need slightly
4385 /// different sequencing for v4i16 and v8i16 to stay within NEON's available
4386 /// 64/128-bit registers.
4388 /// Trace for v4i16:
4389 /// input = [v0 v1 v2 v3 ] (vi 16-bit element)
4390 /// v8i8: BitCounts = [k0 k1 k2 k3 k0 k1 k2 k3 ] (ki is the bit-count of vi)
4391 /// v8i16:Extended = [k0 k1 k2 k3 k0 k1 k2 k3 ]
4392 /// v4i16:Extracted = [k0 k1 k2 k3 ]
4393 static SDValue lowerCTPOP16BitElements(SDNode *N, SelectionDAG &DAG) {
4394 EVT VT = N->getValueType(0);
4397 SDValue BitCounts = getCTPOP16BitCounts(N, DAG);
4398 if (VT.is64BitVector()) {
4399 SDValue Extended = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v8i16, BitCounts);
4400 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i16, Extended,
4401 DAG.getIntPtrConstant(0, DL));
4403 SDValue Extracted = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v8i8,
4404 BitCounts, DAG.getIntPtrConstant(0, DL));
4405 return DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v8i16, Extracted);
4409 /// lowerCTPOP32BitElements - Returns a v2i32/v4i32 vector containing the
4410 /// bit-count for each 32-bit element from the operand. The idea here is
4411 /// to split the vector into 16-bit elements, leverage the 16-bit count
4412 /// routine, and then combine the results.
4414 /// Trace for v2i32 (v4i32 similar with Extracted/Extended exchanged):
4415 /// input = [v0 v1 ] (vi: 32-bit elements)
4416 /// Bitcast = [w0 w1 w2 w3 ] (wi: 16-bit elements, v0 = [w0 w1])
4417 /// Counts16 = [k0 k1 k2 k3 ] (ki: 16-bit elements, bit-count of wi)
4418 /// vrev: N0 = [k1 k0 k3 k2 ]
4420 /// N1 =+[k1 k0 k3 k2 ]
4422 /// N2 =+[k1 k3 k0 k2 ]
4424 /// Extended =+[k1 k3 k0 k2 ]
4426 /// Extracted=+[k1 k3 ]
4428 static SDValue lowerCTPOP32BitElements(SDNode *N, SelectionDAG &DAG) {
4429 EVT VT = N->getValueType(0);
4432 EVT VT16Bit = VT.is64BitVector() ? MVT::v4i16 : MVT::v8i16;
4434 SDValue Bitcast = DAG.getNode(ISD::BITCAST, DL, VT16Bit, N->getOperand(0));
4435 SDValue Counts16 = lowerCTPOP16BitElements(Bitcast.getNode(), DAG);
4436 SDValue N0 = DAG.getNode(ARMISD::VREV32, DL, VT16Bit, Counts16);
4437 SDValue N1 = DAG.getNode(ISD::ADD, DL, VT16Bit, Counts16, N0);
4438 SDValue N2 = DAG.getNode(ARMISD::VUZP, DL, VT16Bit, N1, N1);
4440 if (VT.is64BitVector()) {
4441 SDValue Extended = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v4i32, N2);
4442 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i32, Extended,
4443 DAG.getIntPtrConstant(0, DL));
4445 SDValue Extracted = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i16, N2,
4446 DAG.getIntPtrConstant(0, DL));
4447 return DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v4i32, Extracted);
4451 static SDValue LowerCTPOP(SDNode *N, SelectionDAG &DAG,
4452 const ARMSubtarget *ST) {
4453 EVT VT = N->getValueType(0);
4455 assert(ST->hasNEON() && "Custom ctpop lowering requires NEON.");
4456 assert((VT == MVT::v2i32 || VT == MVT::v4i32 ||
4457 VT == MVT::v4i16 || VT == MVT::v8i16) &&
4458 "Unexpected type for custom ctpop lowering");
4460 if (VT.getVectorElementType() == MVT::i32)
4461 return lowerCTPOP32BitElements(N, DAG);
4463 return lowerCTPOP16BitElements(N, DAG);
4466 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
4467 const ARMSubtarget *ST) {
4468 EVT VT = N->getValueType(0);
4474 // Lower vector shifts on NEON to use VSHL.
4475 assert(ST->hasNEON() && "unexpected vector shift");
4477 // Left shifts translate directly to the vshiftu intrinsic.
4478 if (N->getOpcode() == ISD::SHL)
4479 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
4480 DAG.getConstant(Intrinsic::arm_neon_vshiftu, dl,
4482 N->getOperand(0), N->getOperand(1));
4484 assert((N->getOpcode() == ISD::SRA ||
4485 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
4487 // NEON uses the same intrinsics for both left and right shifts. For
4488 // right shifts, the shift amounts are negative, so negate the vector of
4490 EVT ShiftVT = N->getOperand(1).getValueType();
4491 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
4492 getZeroVector(ShiftVT, DAG, dl),
4494 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
4495 Intrinsic::arm_neon_vshifts :
4496 Intrinsic::arm_neon_vshiftu);
4497 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
4498 DAG.getConstant(vshiftInt, dl, MVT::i32),
4499 N->getOperand(0), NegatedCount);
4502 static SDValue Expand64BitShift(SDNode *N, SelectionDAG &DAG,
4503 const ARMSubtarget *ST) {
4504 EVT VT = N->getValueType(0);
4507 // We can get here for a node like i32 = ISD::SHL i32, i64
4511 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
4512 "Unknown shift to lower!");
4514 // We only lower SRA, SRL of 1 here, all others use generic lowering.
4515 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
4516 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
4519 // If we are in thumb mode, we don't have RRX.
4520 if (ST->isThumb1Only()) return SDValue();
4522 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
4523 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
4524 DAG.getConstant(0, dl, MVT::i32));
4525 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
4526 DAG.getConstant(1, dl, MVT::i32));
4528 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
4529 // captures the result into a carry flag.
4530 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
4531 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Glue), Hi);
4533 // The low part is an ARMISD::RRX operand, which shifts the carry in.
4534 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
4536 // Merge the pieces into a single i64 value.
4537 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
4540 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
4541 SDValue TmpOp0, TmpOp1;
4542 bool Invert = false;
4546 SDValue Op0 = Op.getOperand(0);
4547 SDValue Op1 = Op.getOperand(1);
4548 SDValue CC = Op.getOperand(2);
4549 EVT CmpVT = Op0.getValueType().changeVectorElementTypeToInteger();
4550 EVT VT = Op.getValueType();
4551 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
4554 if (CmpVT.getVectorElementType() == MVT::i64)
4555 // 64-bit comparisons are not legal. We've marked SETCC as non-Custom,
4556 // but it's possible that our operands are 64-bit but our result is 32-bit.
4557 // Bail in this case.
4560 if (Op1.getValueType().isFloatingPoint()) {
4561 switch (SetCCOpcode) {
4562 default: llvm_unreachable("Illegal FP comparison");
4564 case ISD::SETNE: Invert = true; // Fallthrough
4566 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
4568 case ISD::SETLT: Swap = true; // Fallthrough
4570 case ISD::SETGT: Opc = ARMISD::VCGT; break;
4572 case ISD::SETLE: Swap = true; // Fallthrough
4574 case ISD::SETGE: Opc = ARMISD::VCGE; break;
4575 case ISD::SETUGE: Swap = true; // Fallthrough
4576 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
4577 case ISD::SETUGT: Swap = true; // Fallthrough
4578 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
4579 case ISD::SETUEQ: Invert = true; // Fallthrough
4581 // Expand this to (OLT | OGT).
4585 Op0 = DAG.getNode(ARMISD::VCGT, dl, CmpVT, TmpOp1, TmpOp0);
4586 Op1 = DAG.getNode(ARMISD::VCGT, dl, CmpVT, TmpOp0, TmpOp1);
4588 case ISD::SETUO: Invert = true; // Fallthrough
4590 // Expand this to (OLT | OGE).
4594 Op0 = DAG.getNode(ARMISD::VCGT, dl, CmpVT, TmpOp1, TmpOp0);
4595 Op1 = DAG.getNode(ARMISD::VCGE, dl, CmpVT, TmpOp0, TmpOp1);
4599 // Integer comparisons.
4600 switch (SetCCOpcode) {
4601 default: llvm_unreachable("Illegal integer comparison");
4602 case ISD::SETNE: Invert = true;
4603 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
4604 case ISD::SETLT: Swap = true;
4605 case ISD::SETGT: Opc = ARMISD::VCGT; break;
4606 case ISD::SETLE: Swap = true;
4607 case ISD::SETGE: Opc = ARMISD::VCGE; break;
4608 case ISD::SETULT: Swap = true;
4609 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
4610 case ISD::SETULE: Swap = true;
4611 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
4614 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
4615 if (Opc == ARMISD::VCEQ) {
4618 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
4620 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
4623 // Ignore bitconvert.
4624 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BITCAST)
4625 AndOp = AndOp.getOperand(0);
4627 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
4629 Op0 = DAG.getNode(ISD::BITCAST, dl, CmpVT, AndOp.getOperand(0));
4630 Op1 = DAG.getNode(ISD::BITCAST, dl, CmpVT, AndOp.getOperand(1));
4637 std::swap(Op0, Op1);
4639 // If one of the operands is a constant vector zero, attempt to fold the
4640 // comparison to a specialized compare-against-zero form.
4642 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
4644 else if (ISD::isBuildVectorAllZeros(Op0.getNode())) {
4645 if (Opc == ARMISD::VCGE)
4646 Opc = ARMISD::VCLEZ;
4647 else if (Opc == ARMISD::VCGT)
4648 Opc = ARMISD::VCLTZ;
4653 if (SingleOp.getNode()) {
4656 Result = DAG.getNode(ARMISD::VCEQZ, dl, CmpVT, SingleOp); break;
4658 Result = DAG.getNode(ARMISD::VCGEZ, dl, CmpVT, SingleOp); break;
4660 Result = DAG.getNode(ARMISD::VCLEZ, dl, CmpVT, SingleOp); break;
4662 Result = DAG.getNode(ARMISD::VCGTZ, dl, CmpVT, SingleOp); break;
4664 Result = DAG.getNode(ARMISD::VCLTZ, dl, CmpVT, SingleOp); break;
4666 Result = DAG.getNode(Opc, dl, CmpVT, Op0, Op1);
4669 Result = DAG.getNode(Opc, dl, CmpVT, Op0, Op1);
4672 Result = DAG.getSExtOrTrunc(Result, dl, VT);
4675 Result = DAG.getNOT(dl, Result, VT);
4680 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
4681 /// valid vector constant for a NEON instruction with a "modified immediate"
4682 /// operand (e.g., VMOV). If so, return the encoded value.
4683 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
4684 unsigned SplatBitSize, SelectionDAG &DAG,
4685 SDLoc dl, EVT &VT, bool is128Bits,
4686 NEONModImmType type) {
4687 unsigned OpCmode, Imm;
4689 // SplatBitSize is set to the smallest size that splats the vector, so a
4690 // zero vector will always have SplatBitSize == 8. However, NEON modified
4691 // immediate instructions others than VMOV do not support the 8-bit encoding
4692 // of a zero vector, and the default encoding of zero is supposed to be the
4697 switch (SplatBitSize) {
4699 if (type != VMOVModImm)
4701 // Any 1-byte value is OK. Op=0, Cmode=1110.
4702 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
4705 VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
4709 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
4710 VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
4711 if ((SplatBits & ~0xff) == 0) {
4712 // Value = 0x00nn: Op=x, Cmode=100x.
4717 if ((SplatBits & ~0xff00) == 0) {
4718 // Value = 0xnn00: Op=x, Cmode=101x.
4720 Imm = SplatBits >> 8;
4726 // NEON's 32-bit VMOV supports splat values where:
4727 // * only one byte is nonzero, or
4728 // * the least significant byte is 0xff and the second byte is nonzero, or
4729 // * the least significant 2 bytes are 0xff and the third is nonzero.
4730 VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
4731 if ((SplatBits & ~0xff) == 0) {
4732 // Value = 0x000000nn: Op=x, Cmode=000x.
4737 if ((SplatBits & ~0xff00) == 0) {
4738 // Value = 0x0000nn00: Op=x, Cmode=001x.
4740 Imm = SplatBits >> 8;
4743 if ((SplatBits & ~0xff0000) == 0) {
4744 // Value = 0x00nn0000: Op=x, Cmode=010x.
4746 Imm = SplatBits >> 16;
4749 if ((SplatBits & ~0xff000000) == 0) {
4750 // Value = 0xnn000000: Op=x, Cmode=011x.
4752 Imm = SplatBits >> 24;
4756 // cmode == 0b1100 and cmode == 0b1101 are not supported for VORR or VBIC
4757 if (type == OtherModImm) return SDValue();
4759 if ((SplatBits & ~0xffff) == 0 &&
4760 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
4761 // Value = 0x0000nnff: Op=x, Cmode=1100.
4763 Imm = SplatBits >> 8;
4767 if ((SplatBits & ~0xffffff) == 0 &&
4768 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
4769 // Value = 0x00nnffff: Op=x, Cmode=1101.
4771 Imm = SplatBits >> 16;
4775 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
4776 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
4777 // VMOV.I32. A (very) minor optimization would be to replicate the value
4778 // and fall through here to test for a valid 64-bit splat. But, then the
4779 // caller would also need to check and handle the change in size.
4783 if (type != VMOVModImm)
4785 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
4786 uint64_t BitMask = 0xff;
4788 unsigned ImmMask = 1;
4790 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
4791 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
4794 } else if ((SplatBits & BitMask) != 0) {
4801 if (DAG.getDataLayout().isBigEndian())
4802 // swap higher and lower 32 bit word
4803 Imm = ((Imm & 0xf) << 4) | ((Imm & 0xf0) >> 4);
4805 // Op=1, Cmode=1110.
4807 VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
4812 llvm_unreachable("unexpected size for isNEONModifiedImm");
4815 unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm);
4816 return DAG.getTargetConstant(EncodedVal, dl, MVT::i32);
4819 SDValue ARMTargetLowering::LowerConstantFP(SDValue Op, SelectionDAG &DAG,
4820 const ARMSubtarget *ST) const {
4824 bool IsDouble = Op.getValueType() == MVT::f64;
4825 ConstantFPSDNode *CFP = cast<ConstantFPSDNode>(Op);
4827 // Use the default (constant pool) lowering for double constants when we have
4829 if (IsDouble && Subtarget->isFPOnlySP())
4832 // Try splatting with a VMOV.f32...
4833 APFloat FPVal = CFP->getValueAPF();
4834 int ImmVal = IsDouble ? ARM_AM::getFP64Imm(FPVal) : ARM_AM::getFP32Imm(FPVal);
4837 if (IsDouble || !ST->useNEONForSinglePrecisionFP()) {
4838 // We have code in place to select a valid ConstantFP already, no need to
4843 // It's a float and we are trying to use NEON operations where
4844 // possible. Lower it to a splat followed by an extract.
4846 SDValue NewVal = DAG.getTargetConstant(ImmVal, DL, MVT::i32);
4847 SDValue VecConstant = DAG.getNode(ARMISD::VMOVFPIMM, DL, MVT::v2f32,
4849 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecConstant,
4850 DAG.getConstant(0, DL, MVT::i32));
4853 // The rest of our options are NEON only, make sure that's allowed before
4855 if (!ST->hasNEON() || (!IsDouble && !ST->useNEONForSinglePrecisionFP()))
4859 uint64_t iVal = FPVal.bitcastToAPInt().getZExtValue();
4861 // It wouldn't really be worth bothering for doubles except for one very
4862 // important value, which does happen to match: 0.0. So make sure we don't do
4864 if (IsDouble && (iVal & 0xffffffff) != (iVal >> 32))
4867 // Try a VMOV.i32 (FIXME: i8, i16, or i64 could work too).
4868 SDValue NewVal = isNEONModifiedImm(iVal & 0xffffffffU, 0, 32, DAG, SDLoc(Op),
4869 VMovVT, false, VMOVModImm);
4870 if (NewVal != SDValue()) {
4872 SDValue VecConstant = DAG.getNode(ARMISD::VMOVIMM, DL, VMovVT,
4875 return DAG.getNode(ISD::BITCAST, DL, MVT::f64, VecConstant);
4877 // It's a float: cast and extract a vector element.
4878 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
4880 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
4881 DAG.getConstant(0, DL, MVT::i32));
4884 // Finally, try a VMVN.i32
4885 NewVal = isNEONModifiedImm(~iVal & 0xffffffffU, 0, 32, DAG, SDLoc(Op), VMovVT,
4887 if (NewVal != SDValue()) {
4889 SDValue VecConstant = DAG.getNode(ARMISD::VMVNIMM, DL, VMovVT, NewVal);
4892 return DAG.getNode(ISD::BITCAST, DL, MVT::f64, VecConstant);
4894 // It's a float: cast and extract a vector element.
4895 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
4897 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
4898 DAG.getConstant(0, DL, MVT::i32));
4904 // check if an VEXT instruction can handle the shuffle mask when the
4905 // vector sources of the shuffle are the same.
4906 static bool isSingletonVEXTMask(ArrayRef<int> M, EVT VT, unsigned &Imm) {
4907 unsigned NumElts = VT.getVectorNumElements();
4909 // Assume that the first shuffle index is not UNDEF. Fail if it is.
4915 // If this is a VEXT shuffle, the immediate value is the index of the first
4916 // element. The other shuffle indices must be the successive elements after
4918 unsigned ExpectedElt = Imm;
4919 for (unsigned i = 1; i < NumElts; ++i) {
4920 // Increment the expected index. If it wraps around, just follow it
4921 // back to index zero and keep going.
4923 if (ExpectedElt == NumElts)
4926 if (M[i] < 0) continue; // ignore UNDEF indices
4927 if (ExpectedElt != static_cast<unsigned>(M[i]))
4935 static bool isVEXTMask(ArrayRef<int> M, EVT VT,
4936 bool &ReverseVEXT, unsigned &Imm) {
4937 unsigned NumElts = VT.getVectorNumElements();
4938 ReverseVEXT = false;
4940 // Assume that the first shuffle index is not UNDEF. Fail if it is.
4946 // If this is a VEXT shuffle, the immediate value is the index of the first
4947 // element. The other shuffle indices must be the successive elements after
4949 unsigned ExpectedElt = Imm;
4950 for (unsigned i = 1; i < NumElts; ++i) {
4951 // Increment the expected index. If it wraps around, it may still be
4952 // a VEXT but the source vectors must be swapped.
4954 if (ExpectedElt == NumElts * 2) {
4959 if (M[i] < 0) continue; // ignore UNDEF indices
4960 if (ExpectedElt != static_cast<unsigned>(M[i]))
4964 // Adjust the index value if the source operands will be swapped.
4971 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
4972 /// instruction with the specified blocksize. (The order of the elements
4973 /// within each block of the vector is reversed.)
4974 static bool isVREVMask(ArrayRef<int> M, EVT VT, unsigned BlockSize) {
4975 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
4976 "Only possible block sizes for VREV are: 16, 32, 64");
4978 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4982 unsigned NumElts = VT.getVectorNumElements();
4983 unsigned BlockElts = M[0] + 1;
4984 // If the first shuffle index is UNDEF, be optimistic.
4986 BlockElts = BlockSize / EltSz;
4988 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
4991 for (unsigned i = 0; i < NumElts; ++i) {
4992 if (M[i] < 0) continue; // ignore UNDEF indices
4993 if ((unsigned) M[i] != (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
5000 static bool isVTBLMask(ArrayRef<int> M, EVT VT) {
5001 // We can handle <8 x i8> vector shuffles. If the index in the mask is out of
5002 // range, then 0 is placed into the resulting vector. So pretty much any mask
5003 // of 8 elements can work here.
5004 return VT == MVT::v8i8 && M.size() == 8;
5007 // Checks whether the shuffle mask represents a vector transpose (VTRN) by
5008 // checking that pairs of elements in the shuffle mask represent the same index
5009 // in each vector, incrementing the expected index by 2 at each step.
5010 // e.g. For v1,v2 of type v4i32 a valid shuffle mask is: [0, 4, 2, 6]
5011 // v1={a,b,c,d} => x=shufflevector v1, v2 shufflemask => x={a,e,c,g}
5013 // WhichResult gives the offset for each element in the mask based on which
5014 // of the two results it belongs to.
5016 // The transpose can be represented either as:
5017 // result1 = shufflevector v1, v2, result1_shuffle_mask
5018 // result2 = shufflevector v1, v2, result2_shuffle_mask
5019 // where v1/v2 and the shuffle masks have the same number of elements
5020 // (here WhichResult (see below) indicates which result is being checked)
5023 // results = shufflevector v1, v2, shuffle_mask
5024 // where both results are returned in one vector and the shuffle mask has twice
5025 // as many elements as v1/v2 (here WhichResult will always be 0 if true) here we
5026 // want to check the low half and high half of the shuffle mask as if it were
5028 static bool isVTRNMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
5029 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
5033 unsigned NumElts = VT.getVectorNumElements();
5034 if (M.size() != NumElts && M.size() != NumElts*2)
5037 // If the mask is twice as long as the input vector then we need to check the
5038 // upper and lower parts of the mask with a matching value for WhichResult
5039 // FIXME: A mask with only even values will be rejected in case the first
5040 // element is undefined, e.g. [-1, 4, 2, 6] will be rejected, because only
5041 // M[0] is used to determine WhichResult
5042 for (unsigned i = 0; i < M.size(); i += NumElts) {
5043 if (M.size() == NumElts * 2)
5044 WhichResult = i / NumElts;
5046 WhichResult = M[i] == 0 ? 0 : 1;
5047 for (unsigned j = 0; j < NumElts; j += 2) {
5048 if ((M[i+j] >= 0 && (unsigned) M[i+j] != j + WhichResult) ||
5049 (M[i+j+1] >= 0 && (unsigned) M[i+j+1] != j + NumElts + WhichResult))
5054 if (M.size() == NumElts*2)
5060 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
5061 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
5062 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
5063 static bool isVTRN_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
5064 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
5068 unsigned NumElts = VT.getVectorNumElements();
5069 if (M.size() != NumElts && M.size() != NumElts*2)
5072 for (unsigned i = 0; i < M.size(); i += NumElts) {
5073 if (M.size() == NumElts * 2)
5074 WhichResult = i / NumElts;
5076 WhichResult = M[i] == 0 ? 0 : 1;
5077 for (unsigned j = 0; j < NumElts; j += 2) {
5078 if ((M[i+j] >= 0 && (unsigned) M[i+j] != j + WhichResult) ||
5079 (M[i+j+1] >= 0 && (unsigned) M[i+j+1] != j + WhichResult))
5084 if (M.size() == NumElts*2)
5090 // Checks whether the shuffle mask represents a vector unzip (VUZP) by checking
5091 // that the mask elements are either all even and in steps of size 2 or all odd
5092 // and in steps of size 2.
5093 // e.g. For v1,v2 of type v4i32 a valid shuffle mask is: [0, 2, 4, 6]
5094 // v1={a,b,c,d} => x=shufflevector v1, v2 shufflemask => x={a,c,e,g}
5096 // Requires similar checks to that of isVTRNMask with
5097 // respect the how results are returned.
5098 static bool isVUZPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
5099 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
5103 unsigned NumElts = VT.getVectorNumElements();
5104 if (M.size() != NumElts && M.size() != NumElts*2)
5107 for (unsigned i = 0; i < M.size(); i += NumElts) {
5108 WhichResult = M[i] == 0 ? 0 : 1;
5109 for (unsigned j = 0; j < NumElts; ++j) {
5110 if (M[i+j] >= 0 && (unsigned) M[i+j] != 2 * j + WhichResult)
5115 if (M.size() == NumElts*2)
5118 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
5119 if (VT.is64BitVector() && EltSz == 32)
5125 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
5126 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
5127 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
5128 static bool isVUZP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
5129 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
5133 unsigned NumElts = VT.getVectorNumElements();
5134 if (M.size() != NumElts && M.size() != NumElts*2)
5137 unsigned Half = NumElts / 2;
5138 for (unsigned i = 0; i < M.size(); i += NumElts) {
5139 WhichResult = M[i] == 0 ? 0 : 1;
5140 for (unsigned j = 0; j < NumElts; j += Half) {
5141 unsigned Idx = WhichResult;
5142 for (unsigned k = 0; k < Half; ++k) {
5143 int MIdx = M[i + j + k];
5144 if (MIdx >= 0 && (unsigned) MIdx != Idx)
5151 if (M.size() == NumElts*2)
5154 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
5155 if (VT.is64BitVector() && EltSz == 32)
5161 // Checks whether the shuffle mask represents a vector zip (VZIP) by checking
5162 // that pairs of elements of the shufflemask represent the same index in each
5163 // vector incrementing sequentially through the vectors.
5164 // e.g. For v1,v2 of type v4i32 a valid shuffle mask is: [0, 4, 1, 5]
5165 // v1={a,b,c,d} => x=shufflevector v1, v2 shufflemask => x={a,e,b,f}
5167 // Requires similar checks to that of isVTRNMask with respect the how results
5169 static bool isVZIPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
5170 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
5174 unsigned NumElts = VT.getVectorNumElements();
5175 if (M.size() != NumElts && M.size() != NumElts*2)
5178 for (unsigned i = 0; i < M.size(); i += NumElts) {
5179 WhichResult = M[i] == 0 ? 0 : 1;
5180 unsigned Idx = WhichResult * NumElts / 2;
5181 for (unsigned j = 0; j < NumElts; j += 2) {
5182 if ((M[i+j] >= 0 && (unsigned) M[i+j] != Idx) ||
5183 (M[i+j+1] >= 0 && (unsigned) M[i+j+1] != Idx + NumElts))
5189 if (M.size() == NumElts*2)
5192 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
5193 if (VT.is64BitVector() && EltSz == 32)
5199 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
5200 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
5201 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
5202 static bool isVZIP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
5203 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
5207 unsigned NumElts = VT.getVectorNumElements();
5208 if (M.size() != NumElts && M.size() != NumElts*2)
5211 for (unsigned i = 0; i < M.size(); i += NumElts) {
5212 WhichResult = M[i] == 0 ? 0 : 1;
5213 unsigned Idx = WhichResult * NumElts / 2;
5214 for (unsigned j = 0; j < NumElts; j += 2) {
5215 if ((M[i+j] >= 0 && (unsigned) M[i+j] != Idx) ||
5216 (M[i+j+1] >= 0 && (unsigned) M[i+j+1] != Idx))
5222 if (M.size() == NumElts*2)
5225 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
5226 if (VT.is64BitVector() && EltSz == 32)
5232 /// Check if \p ShuffleMask is a NEON two-result shuffle (VZIP, VUZP, VTRN),
5233 /// and return the corresponding ARMISD opcode if it is, or 0 if it isn't.
5234 static unsigned isNEONTwoResultShuffleMask(ArrayRef<int> ShuffleMask, EVT VT,
5235 unsigned &WhichResult,
5238 if (isVTRNMask(ShuffleMask, VT, WhichResult))
5239 return ARMISD::VTRN;
5240 if (isVUZPMask(ShuffleMask, VT, WhichResult))
5241 return ARMISD::VUZP;
5242 if (isVZIPMask(ShuffleMask, VT, WhichResult))
5243 return ARMISD::VZIP;
5246 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
5247 return ARMISD::VTRN;
5248 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
5249 return ARMISD::VUZP;
5250 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
5251 return ARMISD::VZIP;
5256 /// \return true if this is a reverse operation on an vector.
5257 static bool isReverseMask(ArrayRef<int> M, EVT VT) {
5258 unsigned NumElts = VT.getVectorNumElements();
5259 // Make sure the mask has the right size.
5260 if (NumElts != M.size())
5263 // Look for <15, ..., 3, -1, 1, 0>.
5264 for (unsigned i = 0; i != NumElts; ++i)
5265 if (M[i] >= 0 && M[i] != (int) (NumElts - 1 - i))
5271 // If N is an integer constant that can be moved into a register in one
5272 // instruction, return an SDValue of such a constant (will become a MOV
5273 // instruction). Otherwise return null.
5274 static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG,
5275 const ARMSubtarget *ST, SDLoc dl) {
5277 if (!isa<ConstantSDNode>(N))
5279 Val = cast<ConstantSDNode>(N)->getZExtValue();
5281 if (ST->isThumb1Only()) {
5282 if (Val <= 255 || ~Val <= 255)
5283 return DAG.getConstant(Val, dl, MVT::i32);
5285 if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1)
5286 return DAG.getConstant(Val, dl, MVT::i32);
5291 // If this is a case we can't handle, return null and let the default
5292 // expansion code take care of it.
5293 SDValue ARMTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
5294 const ARMSubtarget *ST) const {
5295 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
5297 EVT VT = Op.getValueType();
5299 APInt SplatBits, SplatUndef;
5300 unsigned SplatBitSize;
5302 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
5303 if (SplatBitSize <= 64) {
5304 // Check if an immediate VMOV works.
5306 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
5307 SplatUndef.getZExtValue(), SplatBitSize,
5308 DAG, dl, VmovVT, VT.is128BitVector(),
5310 if (Val.getNode()) {
5311 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
5312 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
5315 // Try an immediate VMVN.
5316 uint64_t NegatedImm = (~SplatBits).getZExtValue();
5317 Val = isNEONModifiedImm(NegatedImm,
5318 SplatUndef.getZExtValue(), SplatBitSize,
5319 DAG, dl, VmovVT, VT.is128BitVector(),
5321 if (Val.getNode()) {
5322 SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
5323 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
5326 // Use vmov.f32 to materialize other v2f32 and v4f32 splats.
5327 if ((VT == MVT::v2f32 || VT == MVT::v4f32) && SplatBitSize == 32) {
5328 int ImmVal = ARM_AM::getFP32Imm(SplatBits);
5330 SDValue Val = DAG.getTargetConstant(ImmVal, dl, MVT::i32);
5331 return DAG.getNode(ARMISD::VMOVFPIMM, dl, VT, Val);
5337 // Scan through the operands to see if only one value is used.
5339 // As an optimisation, even if more than one value is used it may be more
5340 // profitable to splat with one value then change some lanes.
5342 // Heuristically we decide to do this if the vector has a "dominant" value,
5343 // defined as splatted to more than half of the lanes.
5344 unsigned NumElts = VT.getVectorNumElements();
5345 bool isOnlyLowElement = true;
5346 bool usesOnlyOneValue = true;
5347 bool hasDominantValue = false;
5348 bool isConstant = true;
5350 // Map of the number of times a particular SDValue appears in the
5352 DenseMap<SDValue, unsigned> ValueCounts;
5354 for (unsigned i = 0; i < NumElts; ++i) {
5355 SDValue V = Op.getOperand(i);
5356 if (V.getOpcode() == ISD::UNDEF)
5359 isOnlyLowElement = false;
5360 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
5363 ValueCounts.insert(std::make_pair(V, 0));
5364 unsigned &Count = ValueCounts[V];
5366 // Is this value dominant? (takes up more than half of the lanes)
5367 if (++Count > (NumElts / 2)) {
5368 hasDominantValue = true;
5372 if (ValueCounts.size() != 1)
5373 usesOnlyOneValue = false;
5374 if (!Value.getNode() && ValueCounts.size() > 0)
5375 Value = ValueCounts.begin()->first;
5377 if (ValueCounts.size() == 0)
5378 return DAG.getUNDEF(VT);
5380 // Loads are better lowered with insert_vector_elt/ARMISD::BUILD_VECTOR.
5381 // Keep going if we are hitting this case.
5382 if (isOnlyLowElement && !ISD::isNormalLoad(Value.getNode()))
5383 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
5385 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5387 // Use VDUP for non-constant splats. For f32 constant splats, reduce to
5388 // i32 and try again.
5389 if (hasDominantValue && EltSize <= 32) {
5393 // If we are VDUPing a value that comes directly from a vector, that will
5394 // cause an unnecessary move to and from a GPR, where instead we could
5395 // just use VDUPLANE. We can only do this if the lane being extracted
5396 // is at a constant index, as the VDUP from lane instructions only have
5397 // constant-index forms.
5398 if (Value->getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5399 isa<ConstantSDNode>(Value->getOperand(1))) {
5400 // We need to create a new undef vector to use for the VDUPLANE if the
5401 // size of the vector from which we get the value is different than the
5402 // size of the vector that we need to create. We will insert the element
5403 // such that the register coalescer will remove unnecessary copies.
5404 if (VT != Value->getOperand(0).getValueType()) {
5405 ConstantSDNode *constIndex;
5406 constIndex = dyn_cast<ConstantSDNode>(Value->getOperand(1));
5407 assert(constIndex && "The index is not a constant!");
5408 unsigned index = constIndex->getAPIntValue().getLimitedValue() %
5409 VT.getVectorNumElements();
5410 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
5411 DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DAG.getUNDEF(VT),
5412 Value, DAG.getConstant(index, dl, MVT::i32)),
5413 DAG.getConstant(index, dl, MVT::i32));
5415 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
5416 Value->getOperand(0), Value->getOperand(1));
5418 N = DAG.getNode(ARMISD::VDUP, dl, VT, Value);
5420 if (!usesOnlyOneValue) {
5421 // The dominant value was splatted as 'N', but we now have to insert
5422 // all differing elements.
5423 for (unsigned I = 0; I < NumElts; ++I) {
5424 if (Op.getOperand(I) == Value)
5426 SmallVector<SDValue, 3> Ops;
5428 Ops.push_back(Op.getOperand(I));
5429 Ops.push_back(DAG.getConstant(I, dl, MVT::i32));
5430 N = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Ops);
5435 if (VT.getVectorElementType().isFloatingPoint()) {
5436 SmallVector<SDValue, 8> Ops;
5437 for (unsigned i = 0; i < NumElts; ++i)
5438 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, MVT::i32,
5440 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
5441 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, Ops);
5442 Val = LowerBUILD_VECTOR(Val, DAG, ST);
5444 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
5446 if (usesOnlyOneValue) {
5447 SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
5448 if (isConstant && Val.getNode())
5449 return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
5453 // If all elements are constants and the case above didn't get hit, fall back
5454 // to the default expansion, which will generate a load from the constant
5459 // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
5461 SDValue shuffle = ReconstructShuffle(Op, DAG);
5462 if (shuffle != SDValue())
5466 // Vectors with 32- or 64-bit elements can be built by directly assigning
5467 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
5468 // will be legalized.
5469 if (EltSize >= 32) {
5470 // Do the expansion with floating-point types, since that is what the VFP
5471 // registers are defined to use, and since i64 is not legal.
5472 EVT EltVT = EVT::getFloatingPointVT(EltSize);
5473 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
5474 SmallVector<SDValue, 8> Ops;
5475 for (unsigned i = 0; i < NumElts; ++i)
5476 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, EltVT, Op.getOperand(i)));
5477 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, Ops);
5478 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
5481 // If all else fails, just use a sequence of INSERT_VECTOR_ELT when we
5482 // know the default expansion would otherwise fall back on something even
5483 // worse. For a vector with one or two non-undef values, that's
5484 // scalar_to_vector for the elements followed by a shuffle (provided the
5485 // shuffle is valid for the target) and materialization element by element
5486 // on the stack followed by a load for everything else.
5487 if (!isConstant && !usesOnlyOneValue) {
5488 SDValue Vec = DAG.getUNDEF(VT);
5489 for (unsigned i = 0 ; i < NumElts; ++i) {
5490 SDValue V = Op.getOperand(i);
5491 if (V.getOpcode() == ISD::UNDEF)
5493 SDValue LaneIdx = DAG.getConstant(i, dl, MVT::i32);
5494 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Vec, V, LaneIdx);
5502 // Gather data to see if the operation can be modelled as a
5503 // shuffle in combination with VEXTs.
5504 SDValue ARMTargetLowering::ReconstructShuffle(SDValue Op,
5505 SelectionDAG &DAG) const {
5506 assert(Op.getOpcode() == ISD::BUILD_VECTOR && "Unknown opcode!");
5508 EVT VT = Op.getValueType();
5509 unsigned NumElts = VT.getVectorNumElements();
5511 struct ShuffleSourceInfo {
5516 // We may insert some combination of BITCASTs and VEXT nodes to force Vec to
5517 // be compatible with the shuffle we intend to construct. As a result
5518 // ShuffleVec will be some sliding window into the original Vec.
5521 // Code should guarantee that element i in Vec starts at element "WindowBase
5522 // + i * WindowScale in ShuffleVec".
5526 bool operator ==(SDValue OtherVec) { return Vec == OtherVec; }
5527 ShuffleSourceInfo(SDValue Vec)
5528 : Vec(Vec), MinElt(UINT_MAX), MaxElt(0), ShuffleVec(Vec), WindowBase(0),
5532 // First gather all vectors used as an immediate source for this BUILD_VECTOR
5534 SmallVector<ShuffleSourceInfo, 2> Sources;
5535 for (unsigned i = 0; i < NumElts; ++i) {
5536 SDValue V = Op.getOperand(i);
5537 if (V.getOpcode() == ISD::UNDEF)
5539 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
5540 // A shuffle can only come from building a vector from various
5541 // elements of other vectors.
5543 } else if (!isa<ConstantSDNode>(V.getOperand(1))) {
5544 // Furthermore, shuffles require a constant mask, whereas extractelts
5545 // accept variable indices.
5549 // Add this element source to the list if it's not already there.
5550 SDValue SourceVec = V.getOperand(0);
5551 auto Source = std::find(Sources.begin(), Sources.end(), SourceVec);
5552 if (Source == Sources.end())
5553 Source = Sources.insert(Sources.end(), ShuffleSourceInfo(SourceVec));
5555 // Update the minimum and maximum lane number seen.
5556 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
5557 Source->MinElt = std::min(Source->MinElt, EltNo);
5558 Source->MaxElt = std::max(Source->MaxElt, EltNo);
5561 // Currently only do something sane when at most two source vectors
5563 if (Sources.size() > 2)
5566 // Find out the smallest element size among result and two sources, and use
5567 // it as element size to build the shuffle_vector.
5568 EVT SmallestEltTy = VT.getVectorElementType();
5569 for (auto &Source : Sources) {
5570 EVT SrcEltTy = Source.Vec.getValueType().getVectorElementType();
5571 if (SrcEltTy.bitsLT(SmallestEltTy))
5572 SmallestEltTy = SrcEltTy;
5574 unsigned ResMultiplier =
5575 VT.getVectorElementType().getSizeInBits() / SmallestEltTy.getSizeInBits();
5576 NumElts = VT.getSizeInBits() / SmallestEltTy.getSizeInBits();
5577 EVT ShuffleVT = EVT::getVectorVT(*DAG.getContext(), SmallestEltTy, NumElts);
5579 // If the source vector is too wide or too narrow, we may nevertheless be able
5580 // to construct a compatible shuffle either by concatenating it with UNDEF or
5581 // extracting a suitable range of elements.
5582 for (auto &Src : Sources) {
5583 EVT SrcVT = Src.ShuffleVec.getValueType();
5585 if (SrcVT.getSizeInBits() == VT.getSizeInBits())
5588 // This stage of the search produces a source with the same element type as
5589 // the original, but with a total width matching the BUILD_VECTOR output.
5590 EVT EltVT = SrcVT.getVectorElementType();
5591 unsigned NumSrcElts = VT.getSizeInBits() / EltVT.getSizeInBits();
5592 EVT DestVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumSrcElts);
5594 if (SrcVT.getSizeInBits() < VT.getSizeInBits()) {
5595 if (2 * SrcVT.getSizeInBits() != VT.getSizeInBits())
5597 // We can pad out the smaller vector for free, so if it's part of a
5600 DAG.getNode(ISD::CONCAT_VECTORS, dl, DestVT, Src.ShuffleVec,
5601 DAG.getUNDEF(Src.ShuffleVec.getValueType()));
5605 if (SrcVT.getSizeInBits() != 2 * VT.getSizeInBits())
5608 if (Src.MaxElt - Src.MinElt >= NumSrcElts) {
5609 // Span too large for a VEXT to cope
5613 if (Src.MinElt >= NumSrcElts) {
5614 // The extraction can just take the second half
5616 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec,
5617 DAG.getConstant(NumSrcElts, dl, MVT::i32));
5618 Src.WindowBase = -NumSrcElts;
5619 } else if (Src.MaxElt < NumSrcElts) {
5620 // The extraction can just take the first half
5622 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec,
5623 DAG.getConstant(0, dl, MVT::i32));
5625 // An actual VEXT is needed
5627 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec,
5628 DAG.getConstant(0, dl, MVT::i32));
5630 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec,
5631 DAG.getConstant(NumSrcElts, dl, MVT::i32));
5633 Src.ShuffleVec = DAG.getNode(ARMISD::VEXT, dl, DestVT, VEXTSrc1,
5635 DAG.getConstant(Src.MinElt, dl, MVT::i32));
5636 Src.WindowBase = -Src.MinElt;
5640 // Another possible incompatibility occurs from the vector element types. We
5641 // can fix this by bitcasting the source vectors to the same type we intend
5643 for (auto &Src : Sources) {
5644 EVT SrcEltTy = Src.ShuffleVec.getValueType().getVectorElementType();
5645 if (SrcEltTy == SmallestEltTy)
5647 assert(ShuffleVT.getVectorElementType() == SmallestEltTy);
5648 Src.ShuffleVec = DAG.getNode(ISD::BITCAST, dl, ShuffleVT, Src.ShuffleVec);
5649 Src.WindowScale = SrcEltTy.getSizeInBits() / SmallestEltTy.getSizeInBits();
5650 Src.WindowBase *= Src.WindowScale;
5653 // Final sanity check before we try to actually produce a shuffle.
5655 for (auto Src : Sources)
5656 assert(Src.ShuffleVec.getValueType() == ShuffleVT);
5659 // The stars all align, our next step is to produce the mask for the shuffle.
5660 SmallVector<int, 8> Mask(ShuffleVT.getVectorNumElements(), -1);
5661 int BitsPerShuffleLane = ShuffleVT.getVectorElementType().getSizeInBits();
5662 for (unsigned i = 0; i < VT.getVectorNumElements(); ++i) {
5663 SDValue Entry = Op.getOperand(i);
5664 if (Entry.getOpcode() == ISD::UNDEF)
5667 auto Src = std::find(Sources.begin(), Sources.end(), Entry.getOperand(0));
5668 int EltNo = cast<ConstantSDNode>(Entry.getOperand(1))->getSExtValue();
5670 // EXTRACT_VECTOR_ELT performs an implicit any_ext; BUILD_VECTOR an implicit
5671 // trunc. So only std::min(SrcBits, DestBits) actually get defined in this
5673 EVT OrigEltTy = Entry.getOperand(0).getValueType().getVectorElementType();
5674 int BitsDefined = std::min(OrigEltTy.getSizeInBits(),
5675 VT.getVectorElementType().getSizeInBits());
5676 int LanesDefined = BitsDefined / BitsPerShuffleLane;
5678 // This source is expected to fill ResMultiplier lanes of the final shuffle,
5679 // starting at the appropriate offset.
5680 int *LaneMask = &Mask[i * ResMultiplier];
5682 int ExtractBase = EltNo * Src->WindowScale + Src->WindowBase;
5683 ExtractBase += NumElts * (Src - Sources.begin());
5684 for (int j = 0; j < LanesDefined; ++j)
5685 LaneMask[j] = ExtractBase + j;
5688 // Final check before we try to produce nonsense...
5689 if (!isShuffleMaskLegal(Mask, ShuffleVT))
5692 // We can't handle more than two sources. This should have already
5693 // been checked before this point.
5694 assert(Sources.size() <= 2 && "Too many sources!");
5696 SDValue ShuffleOps[] = { DAG.getUNDEF(ShuffleVT), DAG.getUNDEF(ShuffleVT) };
5697 for (unsigned i = 0; i < Sources.size(); ++i)
5698 ShuffleOps[i] = Sources[i].ShuffleVec;
5700 SDValue Shuffle = DAG.getVectorShuffle(ShuffleVT, dl, ShuffleOps[0],
5701 ShuffleOps[1], &Mask[0]);
5702 return DAG.getNode(ISD::BITCAST, dl, VT, Shuffle);
5705 /// isShuffleMaskLegal - Targets can use this to indicate that they only
5706 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
5707 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
5708 /// are assumed to be legal.
5710 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
5712 if (VT.getVectorNumElements() == 4 &&
5713 (VT.is128BitVector() || VT.is64BitVector())) {
5714 unsigned PFIndexes[4];
5715 for (unsigned i = 0; i != 4; ++i) {
5719 PFIndexes[i] = M[i];
5722 // Compute the index in the perfect shuffle table.
5723 unsigned PFTableIndex =
5724 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
5725 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
5726 unsigned Cost = (PFEntry >> 30);
5732 bool ReverseVEXT, isV_UNDEF;
5733 unsigned Imm, WhichResult;
5735 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5736 return (EltSize >= 32 ||
5737 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
5738 isVREVMask(M, VT, 64) ||
5739 isVREVMask(M, VT, 32) ||
5740 isVREVMask(M, VT, 16) ||
5741 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
5742 isVTBLMask(M, VT) ||
5743 isNEONTwoResultShuffleMask(M, VT, WhichResult, isV_UNDEF) ||
5744 ((VT == MVT::v8i16 || VT == MVT::v16i8) && isReverseMask(M, VT)));
5747 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
5748 /// the specified operations to build the shuffle.
5749 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
5750 SDValue RHS, SelectionDAG &DAG,
5752 unsigned OpNum = (PFEntry >> 26) & 0x0F;
5753 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
5754 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
5757 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
5766 OP_VUZPL, // VUZP, left result
5767 OP_VUZPR, // VUZP, right result
5768 OP_VZIPL, // VZIP, left result
5769 OP_VZIPR, // VZIP, right result
5770 OP_VTRNL, // VTRN, left result
5771 OP_VTRNR // VTRN, right result
5774 if (OpNum == OP_COPY) {
5775 if (LHSID == (1*9+2)*9+3) return LHS;
5776 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
5780 SDValue OpLHS, OpRHS;
5781 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
5782 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
5783 EVT VT = OpLHS.getValueType();
5786 default: llvm_unreachable("Unknown shuffle opcode!");
5788 // VREV divides the vector in half and swaps within the half.
5789 if (VT.getVectorElementType() == MVT::i32 ||
5790 VT.getVectorElementType() == MVT::f32)
5791 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
5792 // vrev <4 x i16> -> VREV32
5793 if (VT.getVectorElementType() == MVT::i16)
5794 return DAG.getNode(ARMISD::VREV32, dl, VT, OpLHS);
5795 // vrev <4 x i8> -> VREV16
5796 assert(VT.getVectorElementType() == MVT::i8);
5797 return DAG.getNode(ARMISD::VREV16, dl, VT, OpLHS);
5802 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
5803 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, dl, MVT::i32));
5807 return DAG.getNode(ARMISD::VEXT, dl, VT,
5809 DAG.getConstant(OpNum - OP_VEXT1 + 1, dl, MVT::i32));
5812 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
5813 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
5816 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
5817 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
5820 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
5821 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
5825 static SDValue LowerVECTOR_SHUFFLEv8i8(SDValue Op,
5826 ArrayRef<int> ShuffleMask,
5827 SelectionDAG &DAG) {
5828 // Check to see if we can use the VTBL instruction.
5829 SDValue V1 = Op.getOperand(0);
5830 SDValue V2 = Op.getOperand(1);
5833 SmallVector<SDValue, 8> VTBLMask;
5834 for (ArrayRef<int>::iterator
5835 I = ShuffleMask.begin(), E = ShuffleMask.end(); I != E; ++I)
5836 VTBLMask.push_back(DAG.getConstant(*I, DL, MVT::i32));
5838 if (V2.getNode()->getOpcode() == ISD::UNDEF)
5839 return DAG.getNode(ARMISD::VTBL1, DL, MVT::v8i8, V1,
5840 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8, VTBLMask));
5842 return DAG.getNode(ARMISD::VTBL2, DL, MVT::v8i8, V1, V2,
5843 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8, VTBLMask));
5846 static SDValue LowerReverse_VECTOR_SHUFFLEv16i8_v8i16(SDValue Op,
5847 SelectionDAG &DAG) {
5849 SDValue OpLHS = Op.getOperand(0);
5850 EVT VT = OpLHS.getValueType();
5852 assert((VT == MVT::v8i16 || VT == MVT::v16i8) &&
5853 "Expect an v8i16/v16i8 type");
5854 OpLHS = DAG.getNode(ARMISD::VREV64, DL, VT, OpLHS);
5855 // For a v16i8 type: After the VREV, we have got <8, ...15, 8, ..., 0>. Now,
5856 // extract the first 8 bytes into the top double word and the last 8 bytes
5857 // into the bottom double word. The v8i16 case is similar.
5858 unsigned ExtractNum = (VT == MVT::v16i8) ? 8 : 4;
5859 return DAG.getNode(ARMISD::VEXT, DL, VT, OpLHS, OpLHS,
5860 DAG.getConstant(ExtractNum, DL, MVT::i32));
5863 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
5864 SDValue V1 = Op.getOperand(0);
5865 SDValue V2 = Op.getOperand(1);
5867 EVT VT = Op.getValueType();
5868 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
5870 // Convert shuffles that are directly supported on NEON to target-specific
5871 // DAG nodes, instead of keeping them as shuffles and matching them again
5872 // during code selection. This is more efficient and avoids the possibility
5873 // of inconsistencies between legalization and selection.
5874 // FIXME: floating-point vectors should be canonicalized to integer vectors
5875 // of the same time so that they get CSEd properly.
5876 ArrayRef<int> ShuffleMask = SVN->getMask();
5878 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5879 if (EltSize <= 32) {
5880 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
5881 int Lane = SVN->getSplatIndex();
5882 // If this is undef splat, generate it via "just" vdup, if possible.
5883 if (Lane == -1) Lane = 0;
5885 // Test if V1 is a SCALAR_TO_VECTOR.
5886 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
5887 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
5889 // Test if V1 is a BUILD_VECTOR which is equivalent to a SCALAR_TO_VECTOR
5890 // (and probably will turn into a SCALAR_TO_VECTOR once legalization
5892 if (Lane == 0 && V1.getOpcode() == ISD::BUILD_VECTOR &&
5893 !isa<ConstantSDNode>(V1.getOperand(0))) {
5894 bool IsScalarToVector = true;
5895 for (unsigned i = 1, e = V1.getNumOperands(); i != e; ++i)
5896 if (V1.getOperand(i).getOpcode() != ISD::UNDEF) {
5897 IsScalarToVector = false;
5900 if (IsScalarToVector)
5901 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
5903 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
5904 DAG.getConstant(Lane, dl, MVT::i32));
5909 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
5912 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
5913 DAG.getConstant(Imm, dl, MVT::i32));
5916 if (isVREVMask(ShuffleMask, VT, 64))
5917 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
5918 if (isVREVMask(ShuffleMask, VT, 32))
5919 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
5920 if (isVREVMask(ShuffleMask, VT, 16))
5921 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
5923 if (V2->getOpcode() == ISD::UNDEF &&
5924 isSingletonVEXTMask(ShuffleMask, VT, Imm)) {
5925 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V1,
5926 DAG.getConstant(Imm, dl, MVT::i32));
5929 // Check for Neon shuffles that modify both input vectors in place.
5930 // If both results are used, i.e., if there are two shuffles with the same
5931 // source operands and with masks corresponding to both results of one of
5932 // these operations, DAG memoization will ensure that a single node is
5933 // used for both shuffles.
5934 unsigned WhichResult;
5936 if (unsigned ShuffleOpc = isNEONTwoResultShuffleMask(
5937 ShuffleMask, VT, WhichResult, isV_UNDEF)) {
5940 return DAG.getNode(ShuffleOpc, dl, DAG.getVTList(VT, VT), V1, V2)
5941 .getValue(WhichResult);
5944 // Also check for these shuffles through CONCAT_VECTORS: we canonicalize
5945 // shuffles that produce a result larger than their operands with:
5946 // shuffle(concat(v1, undef), concat(v2, undef))
5948 // shuffle(concat(v1, v2), undef)
5949 // because we can access quad vectors (see PerformVECTOR_SHUFFLECombine).
5951 // This is useful in the general case, but there are special cases where
5952 // native shuffles produce larger results: the two-result ops.
5954 // Look through the concat when lowering them:
5955 // shuffle(concat(v1, v2), undef)
5957 // concat(VZIP(v1, v2):0, :1)
5959 if (V1->getOpcode() == ISD::CONCAT_VECTORS &&
5960 V2->getOpcode() == ISD::UNDEF) {
5961 SDValue SubV1 = V1->getOperand(0);
5962 SDValue SubV2 = V1->getOperand(1);
5963 EVT SubVT = SubV1.getValueType();
5965 // We expect these to have been canonicalized to -1.
5966 assert(std::all_of(ShuffleMask.begin(), ShuffleMask.end(), [&](int i) {
5967 return i < (int)VT.getVectorNumElements();
5968 }) && "Unexpected shuffle index into UNDEF operand!");
5970 if (unsigned ShuffleOpc = isNEONTwoResultShuffleMask(
5971 ShuffleMask, SubVT, WhichResult, isV_UNDEF)) {
5974 assert((WhichResult == 0) &&
5975 "In-place shuffle of concat can only have one result!");
5976 SDValue Res = DAG.getNode(ShuffleOpc, dl, DAG.getVTList(SubVT, SubVT),
5978 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, Res.getValue(0),
5984 // If the shuffle is not directly supported and it has 4 elements, use
5985 // the PerfectShuffle-generated table to synthesize it from other shuffles.
5986 unsigned NumElts = VT.getVectorNumElements();
5988 unsigned PFIndexes[4];
5989 for (unsigned i = 0; i != 4; ++i) {
5990 if (ShuffleMask[i] < 0)
5993 PFIndexes[i] = ShuffleMask[i];
5996 // Compute the index in the perfect shuffle table.
5997 unsigned PFTableIndex =
5998 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
5999 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
6000 unsigned Cost = (PFEntry >> 30);
6003 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
6006 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
6007 if (EltSize >= 32) {
6008 // Do the expansion with floating-point types, since that is what the VFP
6009 // registers are defined to use, and since i64 is not legal.
6010 EVT EltVT = EVT::getFloatingPointVT(EltSize);
6011 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
6012 V1 = DAG.getNode(ISD::BITCAST, dl, VecVT, V1);
6013 V2 = DAG.getNode(ISD::BITCAST, dl, VecVT, V2);
6014 SmallVector<SDValue, 8> Ops;
6015 for (unsigned i = 0; i < NumElts; ++i) {
6016 if (ShuffleMask[i] < 0)
6017 Ops.push_back(DAG.getUNDEF(EltVT));
6019 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
6020 ShuffleMask[i] < (int)NumElts ? V1 : V2,
6021 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
6024 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, Ops);
6025 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
6028 if ((VT == MVT::v8i16 || VT == MVT::v16i8) && isReverseMask(ShuffleMask, VT))
6029 return LowerReverse_VECTOR_SHUFFLEv16i8_v8i16(Op, DAG);
6031 if (VT == MVT::v8i8) {
6032 SDValue NewOp = LowerVECTOR_SHUFFLEv8i8(Op, ShuffleMask, DAG);
6033 if (NewOp.getNode())
6040 static SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
6041 // INSERT_VECTOR_ELT is legal only for immediate indexes.
6042 SDValue Lane = Op.getOperand(2);
6043 if (!isa<ConstantSDNode>(Lane))
6049 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
6050 // EXTRACT_VECTOR_ELT is legal only for immediate indexes.
6051 SDValue Lane = Op.getOperand(1);
6052 if (!isa<ConstantSDNode>(Lane))
6055 SDValue Vec = Op.getOperand(0);
6056 if (Op.getValueType() == MVT::i32 &&
6057 Vec.getValueType().getVectorElementType().getSizeInBits() < 32) {
6059 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
6065 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
6066 // The only time a CONCAT_VECTORS operation can have legal types is when
6067 // two 64-bit vectors are concatenated to a 128-bit vector.
6068 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
6069 "unexpected CONCAT_VECTORS");
6071 SDValue Val = DAG.getUNDEF(MVT::v2f64);
6072 SDValue Op0 = Op.getOperand(0);
6073 SDValue Op1 = Op.getOperand(1);
6074 if (Op0.getOpcode() != ISD::UNDEF)
6075 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
6076 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op0),
6077 DAG.getIntPtrConstant(0, dl));
6078 if (Op1.getOpcode() != ISD::UNDEF)
6079 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
6080 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op1),
6081 DAG.getIntPtrConstant(1, dl));
6082 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Val);
6085 /// isExtendedBUILD_VECTOR - Check if N is a constant BUILD_VECTOR where each
6086 /// element has been zero/sign-extended, depending on the isSigned parameter,
6087 /// from an integer type half its size.
6088 static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG,
6090 // A v2i64 BUILD_VECTOR will have been legalized to a BITCAST from v4i32.
6091 EVT VT = N->getValueType(0);
6092 if (VT == MVT::v2i64 && N->getOpcode() == ISD::BITCAST) {
6093 SDNode *BVN = N->getOperand(0).getNode();
6094 if (BVN->getValueType(0) != MVT::v4i32 ||
6095 BVN->getOpcode() != ISD::BUILD_VECTOR)
6097 unsigned LoElt = DAG.getDataLayout().isBigEndian() ? 1 : 0;
6098 unsigned HiElt = 1 - LoElt;
6099 ConstantSDNode *Lo0 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt));
6100 ConstantSDNode *Hi0 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt));
6101 ConstantSDNode *Lo1 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt+2));
6102 ConstantSDNode *Hi1 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt+2));
6103 if (!Lo0 || !Hi0 || !Lo1 || !Hi1)
6106 if (Hi0->getSExtValue() == Lo0->getSExtValue() >> 32 &&
6107 Hi1->getSExtValue() == Lo1->getSExtValue() >> 32)
6110 if (Hi0->isNullValue() && Hi1->isNullValue())
6116 if (N->getOpcode() != ISD::BUILD_VECTOR)
6119 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
6120 SDNode *Elt = N->getOperand(i).getNode();
6121 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
6122 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
6123 unsigned HalfSize = EltSize / 2;
6125 if (!isIntN(HalfSize, C->getSExtValue()))
6128 if (!isUIntN(HalfSize, C->getZExtValue()))
6139 /// isSignExtended - Check if a node is a vector value that is sign-extended
6140 /// or a constant BUILD_VECTOR with sign-extended elements.
6141 static bool isSignExtended(SDNode *N, SelectionDAG &DAG) {
6142 if (N->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N))
6144 if (isExtendedBUILD_VECTOR(N, DAG, true))
6149 /// isZeroExtended - Check if a node is a vector value that is zero-extended
6150 /// or a constant BUILD_VECTOR with zero-extended elements.
6151 static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) {
6152 if (N->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N))
6154 if (isExtendedBUILD_VECTOR(N, DAG, false))
6159 static EVT getExtensionTo64Bits(const EVT &OrigVT) {
6160 if (OrigVT.getSizeInBits() >= 64)
6163 assert(OrigVT.isSimple() && "Expecting a simple value type");
6165 MVT::SimpleValueType OrigSimpleTy = OrigVT.getSimpleVT().SimpleTy;
6166 switch (OrigSimpleTy) {
6167 default: llvm_unreachable("Unexpected Vector Type");
6176 /// AddRequiredExtensionForVMULL - Add a sign/zero extension to extend the total
6177 /// value size to 64 bits. We need a 64-bit D register as an operand to VMULL.
6178 /// We insert the required extension here to get the vector to fill a D register.
6179 static SDValue AddRequiredExtensionForVMULL(SDValue N, SelectionDAG &DAG,
6182 unsigned ExtOpcode) {
6183 // The vector originally had a size of OrigTy. It was then extended to ExtTy.
6184 // We expect the ExtTy to be 128-bits total. If the OrigTy is less than
6185 // 64-bits we need to insert a new extension so that it will be 64-bits.
6186 assert(ExtTy.is128BitVector() && "Unexpected extension size");
6187 if (OrigTy.getSizeInBits() >= 64)
6190 // Must extend size to at least 64 bits to be used as an operand for VMULL.
6191 EVT NewVT = getExtensionTo64Bits(OrigTy);
6193 return DAG.getNode(ExtOpcode, SDLoc(N), NewVT, N);
6196 /// SkipLoadExtensionForVMULL - return a load of the original vector size that
6197 /// does not do any sign/zero extension. If the original vector is less
6198 /// than 64 bits, an appropriate extension will be added after the load to
6199 /// reach a total size of 64 bits. We have to add the extension separately
6200 /// because ARM does not have a sign/zero extending load for vectors.
6201 static SDValue SkipLoadExtensionForVMULL(LoadSDNode *LD, SelectionDAG& DAG) {
6202 EVT ExtendedTy = getExtensionTo64Bits(LD->getMemoryVT());
6204 // The load already has the right type.
6205 if (ExtendedTy == LD->getMemoryVT())
6206 return DAG.getLoad(LD->getMemoryVT(), SDLoc(LD), LD->getChain(),
6207 LD->getBasePtr(), LD->getPointerInfo(), LD->isVolatile(),
6208 LD->isNonTemporal(), LD->isInvariant(),
6209 LD->getAlignment());
6211 // We need to create a zextload/sextload. We cannot just create a load
6212 // followed by a zext/zext node because LowerMUL is also run during normal
6213 // operation legalization where we can't create illegal types.
6214 return DAG.getExtLoad(LD->getExtensionType(), SDLoc(LD), ExtendedTy,
6215 LD->getChain(), LD->getBasePtr(), LD->getPointerInfo(),
6216 LD->getMemoryVT(), LD->isVolatile(), LD->isInvariant(),
6217 LD->isNonTemporal(), LD->getAlignment());
6220 /// SkipExtensionForVMULL - For a node that is a SIGN_EXTEND, ZERO_EXTEND,
6221 /// extending load, or BUILD_VECTOR with extended elements, return the
6222 /// unextended value. The unextended vector should be 64 bits so that it can
6223 /// be used as an operand to a VMULL instruction. If the original vector size
6224 /// before extension is less than 64 bits we add a an extension to resize
6225 /// the vector to 64 bits.
6226 static SDValue SkipExtensionForVMULL(SDNode *N, SelectionDAG &DAG) {
6227 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
6228 return AddRequiredExtensionForVMULL(N->getOperand(0), DAG,
6229 N->getOperand(0)->getValueType(0),
6233 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
6234 return SkipLoadExtensionForVMULL(LD, DAG);
6236 // Otherwise, the value must be a BUILD_VECTOR. For v2i64, it will
6237 // have been legalized as a BITCAST from v4i32.
6238 if (N->getOpcode() == ISD::BITCAST) {
6239 SDNode *BVN = N->getOperand(0).getNode();
6240 assert(BVN->getOpcode() == ISD::BUILD_VECTOR &&
6241 BVN->getValueType(0) == MVT::v4i32 && "expected v4i32 BUILD_VECTOR");
6242 unsigned LowElt = DAG.getDataLayout().isBigEndian() ? 1 : 0;
6243 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), MVT::v2i32,
6244 BVN->getOperand(LowElt), BVN->getOperand(LowElt+2));
6246 // Construct a new BUILD_VECTOR with elements truncated to half the size.
6247 assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR");
6248 EVT VT = N->getValueType(0);
6249 unsigned EltSize = VT.getVectorElementType().getSizeInBits() / 2;
6250 unsigned NumElts = VT.getVectorNumElements();
6251 MVT TruncVT = MVT::getIntegerVT(EltSize);
6252 SmallVector<SDValue, 8> Ops;
6254 for (unsigned i = 0; i != NumElts; ++i) {
6255 ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i));
6256 const APInt &CInt = C->getAPIntValue();
6257 // Element types smaller than 32 bits are not legal, so use i32 elements.
6258 // The values are implicitly truncated so sext vs. zext doesn't matter.
6259 Ops.push_back(DAG.getConstant(CInt.zextOrTrunc(32), dl, MVT::i32));
6261 return DAG.getNode(ISD::BUILD_VECTOR, dl,
6262 MVT::getVectorVT(TruncVT, NumElts), Ops);
6265 static bool isAddSubSExt(SDNode *N, SelectionDAG &DAG) {
6266 unsigned Opcode = N->getOpcode();
6267 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
6268 SDNode *N0 = N->getOperand(0).getNode();
6269 SDNode *N1 = N->getOperand(1).getNode();
6270 return N0->hasOneUse() && N1->hasOneUse() &&
6271 isSignExtended(N0, DAG) && isSignExtended(N1, DAG);
6276 static bool isAddSubZExt(SDNode *N, SelectionDAG &DAG) {
6277 unsigned Opcode = N->getOpcode();
6278 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
6279 SDNode *N0 = N->getOperand(0).getNode();
6280 SDNode *N1 = N->getOperand(1).getNode();
6281 return N0->hasOneUse() && N1->hasOneUse() &&
6282 isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG);
6287 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
6288 // Multiplications are only custom-lowered for 128-bit vectors so that
6289 // VMULL can be detected. Otherwise v2i64 multiplications are not legal.
6290 EVT VT = Op.getValueType();
6291 assert(VT.is128BitVector() && VT.isInteger() &&
6292 "unexpected type for custom-lowering ISD::MUL");
6293 SDNode *N0 = Op.getOperand(0).getNode();
6294 SDNode *N1 = Op.getOperand(1).getNode();
6295 unsigned NewOpc = 0;
6297 bool isN0SExt = isSignExtended(N0, DAG);
6298 bool isN1SExt = isSignExtended(N1, DAG);
6299 if (isN0SExt && isN1SExt)
6300 NewOpc = ARMISD::VMULLs;
6302 bool isN0ZExt = isZeroExtended(N0, DAG);
6303 bool isN1ZExt = isZeroExtended(N1, DAG);
6304 if (isN0ZExt && isN1ZExt)
6305 NewOpc = ARMISD::VMULLu;
6306 else if (isN1SExt || isN1ZExt) {
6307 // Look for (s/zext A + s/zext B) * (s/zext C). We want to turn these
6308 // into (s/zext A * s/zext C) + (s/zext B * s/zext C)
6309 if (isN1SExt && isAddSubSExt(N0, DAG)) {
6310 NewOpc = ARMISD::VMULLs;
6312 } else if (isN1ZExt && isAddSubZExt(N0, DAG)) {
6313 NewOpc = ARMISD::VMULLu;
6315 } else if (isN0ZExt && isAddSubZExt(N1, DAG)) {
6317 NewOpc = ARMISD::VMULLu;
6323 if (VT == MVT::v2i64)
6324 // Fall through to expand this. It is not legal.
6327 // Other vector multiplications are legal.
6332 // Legalize to a VMULL instruction.
6335 SDValue Op1 = SkipExtensionForVMULL(N1, DAG);
6337 Op0 = SkipExtensionForVMULL(N0, DAG);
6338 assert(Op0.getValueType().is64BitVector() &&
6339 Op1.getValueType().is64BitVector() &&
6340 "unexpected types for extended operands to VMULL");
6341 return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
6344 // Optimizing (zext A + zext B) * C, to (VMULL A, C) + (VMULL B, C) during
6345 // isel lowering to take advantage of no-stall back to back vmul + vmla.
6352 SDValue N00 = SkipExtensionForVMULL(N0->getOperand(0).getNode(), DAG);
6353 SDValue N01 = SkipExtensionForVMULL(N0->getOperand(1).getNode(), DAG);
6354 EVT Op1VT = Op1.getValueType();
6355 return DAG.getNode(N0->getOpcode(), DL, VT,
6356 DAG.getNode(NewOpc, DL, VT,
6357 DAG.getNode(ISD::BITCAST, DL, Op1VT, N00), Op1),
6358 DAG.getNode(NewOpc, DL, VT,
6359 DAG.getNode(ISD::BITCAST, DL, Op1VT, N01), Op1));
6363 LowerSDIV_v4i8(SDValue X, SDValue Y, SDLoc dl, SelectionDAG &DAG) {
6364 // TODO: Should this propagate fast-math-flags?
6367 // float4 xf = vcvt_f32_s32(vmovl_s16(a.lo));
6368 // float4 yf = vcvt_f32_s32(vmovl_s16(b.lo));
6369 X = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, X);
6370 Y = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, Y);
6371 X = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, X);
6372 Y = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, Y);
6373 // Get reciprocal estimate.
6374 // float4 recip = vrecpeq_f32(yf);
6375 Y = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6376 DAG.getConstant(Intrinsic::arm_neon_vrecpe, dl, MVT::i32),
6378 // Because char has a smaller range than uchar, we can actually get away
6379 // without any newton steps. This requires that we use a weird bias
6380 // of 0xb000, however (again, this has been exhaustively tested).
6381 // float4 result = as_float4(as_int4(xf*recip) + 0xb000);
6382 X = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, X, Y);
6383 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, X);
6384 Y = DAG.getConstant(0xb000, dl, MVT::i32);
6385 Y = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Y, Y, Y, Y);
6386 X = DAG.getNode(ISD::ADD, dl, MVT::v4i32, X, Y);
6387 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, X);
6388 // Convert back to short.
6389 X = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, X);
6390 X = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, X);
6395 LowerSDIV_v4i16(SDValue N0, SDValue N1, SDLoc dl, SelectionDAG &DAG) {
6396 // TODO: Should this propagate fast-math-flags?
6399 // Convert to float.
6400 // float4 yf = vcvt_f32_s32(vmovl_s16(y));
6401 // float4 xf = vcvt_f32_s32(vmovl_s16(x));
6402 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N0);
6403 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N1);
6404 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
6405 N1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
6407 // Use reciprocal estimate and one refinement step.
6408 // float4 recip = vrecpeq_f32(yf);
6409 // recip *= vrecpsq_f32(yf, recip);
6410 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6411 DAG.getConstant(Intrinsic::arm_neon_vrecpe, dl, MVT::i32),
6413 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6414 DAG.getConstant(Intrinsic::arm_neon_vrecps, dl, MVT::i32),
6416 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
6417 // Because short has a smaller range than ushort, we can actually get away
6418 // with only a single newton step. This requires that we use a weird bias
6419 // of 89, however (again, this has been exhaustively tested).
6420 // float4 result = as_float4(as_int4(xf*recip) + 0x89);
6421 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
6422 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
6423 N1 = DAG.getConstant(0x89, dl, MVT::i32);
6424 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
6425 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
6426 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
6427 // Convert back to integer and return.
6428 // return vmovn_s32(vcvt_s32_f32(result));
6429 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
6430 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
6434 static SDValue LowerSDIV(SDValue Op, SelectionDAG &DAG) {
6435 EVT VT = Op.getValueType();
6436 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
6437 "unexpected type for custom-lowering ISD::SDIV");
6440 SDValue N0 = Op.getOperand(0);
6441 SDValue N1 = Op.getOperand(1);
6444 if (VT == MVT::v8i8) {
6445 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N0);
6446 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N1);
6448 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
6449 DAG.getIntPtrConstant(4, dl));
6450 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
6451 DAG.getIntPtrConstant(4, dl));
6452 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
6453 DAG.getIntPtrConstant(0, dl));
6454 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
6455 DAG.getIntPtrConstant(0, dl));
6457 N0 = LowerSDIV_v4i8(N0, N1, dl, DAG); // v4i16
6458 N2 = LowerSDIV_v4i8(N2, N3, dl, DAG); // v4i16
6460 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
6461 N0 = LowerCONCAT_VECTORS(N0, DAG);
6463 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v8i8, N0);
6466 return LowerSDIV_v4i16(N0, N1, dl, DAG);
6469 static SDValue LowerUDIV(SDValue Op, SelectionDAG &DAG) {
6470 // TODO: Should this propagate fast-math-flags?
6471 EVT VT = Op.getValueType();
6472 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
6473 "unexpected type for custom-lowering ISD::UDIV");
6476 SDValue N0 = Op.getOperand(0);
6477 SDValue N1 = Op.getOperand(1);
6480 if (VT == MVT::v8i8) {
6481 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N0);
6482 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N1);
6484 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
6485 DAG.getIntPtrConstant(4, dl));
6486 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
6487 DAG.getIntPtrConstant(4, dl));
6488 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
6489 DAG.getIntPtrConstant(0, dl));
6490 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
6491 DAG.getIntPtrConstant(0, dl));
6493 N0 = LowerSDIV_v4i16(N0, N1, dl, DAG); // v4i16
6494 N2 = LowerSDIV_v4i16(N2, N3, dl, DAG); // v4i16
6496 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
6497 N0 = LowerCONCAT_VECTORS(N0, DAG);
6499 N0 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v8i8,
6500 DAG.getConstant(Intrinsic::arm_neon_vqmovnsu, dl,
6506 // v4i16 sdiv ... Convert to float.
6507 // float4 yf = vcvt_f32_s32(vmovl_u16(y));
6508 // float4 xf = vcvt_f32_s32(vmovl_u16(x));
6509 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N0);
6510 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N1);
6511 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
6512 SDValue BN1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
6514 // Use reciprocal estimate and two refinement steps.
6515 // float4 recip = vrecpeq_f32(yf);
6516 // recip *= vrecpsq_f32(yf, recip);
6517 // recip *= vrecpsq_f32(yf, recip);
6518 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6519 DAG.getConstant(Intrinsic::arm_neon_vrecpe, dl, MVT::i32),
6521 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6522 DAG.getConstant(Intrinsic::arm_neon_vrecps, dl, MVT::i32),
6524 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
6525 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6526 DAG.getConstant(Intrinsic::arm_neon_vrecps, dl, MVT::i32),
6528 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
6529 // Simply multiplying by the reciprocal estimate can leave us a few ulps
6530 // too low, so we add 2 ulps (exhaustive testing shows that this is enough,
6531 // and that it will never cause us to return an answer too large).
6532 // float4 result = as_float4(as_int4(xf*recip) + 2);
6533 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
6534 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
6535 N1 = DAG.getConstant(2, dl, MVT::i32);
6536 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
6537 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
6538 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
6539 // Convert back to integer and return.
6540 // return vmovn_u32(vcvt_s32_f32(result));
6541 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
6542 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
6546 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
6547 EVT VT = Op.getNode()->getValueType(0);
6548 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
6551 bool ExtraOp = false;
6552 switch (Op.getOpcode()) {
6553 default: llvm_unreachable("Invalid code");
6554 case ISD::ADDC: Opc = ARMISD::ADDC; break;
6555 case ISD::ADDE: Opc = ARMISD::ADDE; ExtraOp = true; break;
6556 case ISD::SUBC: Opc = ARMISD::SUBC; break;
6557 case ISD::SUBE: Opc = ARMISD::SUBE; ExtraOp = true; break;
6561 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
6563 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
6564 Op.getOperand(1), Op.getOperand(2));
6567 SDValue ARMTargetLowering::LowerFSINCOS(SDValue Op, SelectionDAG &DAG) const {
6568 assert(Subtarget->isTargetDarwin());
6570 // For iOS, we want to call an alternative entry point: __sincos_stret,
6571 // return values are passed via sret.
6573 SDValue Arg = Op.getOperand(0);
6574 EVT ArgVT = Arg.getValueType();
6575 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
6576 auto PtrVT = getPointerTy(DAG.getDataLayout());
6578 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
6580 // Pair of floats / doubles used to pass the result.
6581 StructType *RetTy = StructType::get(ArgTy, ArgTy, nullptr);
6583 // Create stack object for sret.
6584 auto &DL = DAG.getDataLayout();
6585 const uint64_t ByteSize = DL.getTypeAllocSize(RetTy);
6586 const unsigned StackAlign = DL.getPrefTypeAlignment(RetTy);
6587 int FrameIdx = FrameInfo->CreateStackObject(ByteSize, StackAlign, false);
6588 SDValue SRet = DAG.getFrameIndex(FrameIdx, getPointerTy(DL));
6594 Entry.Ty = RetTy->getPointerTo();
6595 Entry.isSExt = false;
6596 Entry.isZExt = false;
6597 Entry.isSRet = true;
6598 Args.push_back(Entry);
6602 Entry.isSExt = false;
6603 Entry.isZExt = false;
6604 Args.push_back(Entry);
6606 const char *LibcallName =
6607 (ArgVT == MVT::f64) ? "__sincos_stret" : "__sincosf_stret";
6608 SDValue Callee = DAG.getExternalSymbol(LibcallName, getPointerTy(DL));
6610 TargetLowering::CallLoweringInfo CLI(DAG);
6611 CLI.setDebugLoc(dl).setChain(DAG.getEntryNode())
6612 .setCallee(CallingConv::C, Type::getVoidTy(*DAG.getContext()), Callee,
6614 .setDiscardResult();
6616 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
6618 SDValue LoadSin = DAG.getLoad(ArgVT, dl, CallResult.second, SRet,
6619 MachinePointerInfo(), false, false, false, 0);
6621 // Address of cos field.
6622 SDValue Add = DAG.getNode(ISD::ADD, dl, PtrVT, SRet,
6623 DAG.getIntPtrConstant(ArgVT.getStoreSize(), dl));
6624 SDValue LoadCos = DAG.getLoad(ArgVT, dl, LoadSin.getValue(1), Add,
6625 MachinePointerInfo(), false, false, false, 0);
6627 SDVTList Tys = DAG.getVTList(ArgVT, ArgVT);
6628 return DAG.getNode(ISD::MERGE_VALUES, dl, Tys,
6629 LoadSin.getValue(0), LoadCos.getValue(0));
6632 SDValue ARMTargetLowering::LowerWindowsDIVLibCall(SDValue Op, SelectionDAG &DAG,
6634 SDValue &Chain) const {
6635 EVT VT = Op.getValueType();
6636 assert((VT == MVT::i32 || VT == MVT::i64) &&
6637 "unexpected type for custom lowering DIV");
6640 const auto &DL = DAG.getDataLayout();
6641 const auto &TLI = DAG.getTargetLoweringInfo();
6643 const char *Name = nullptr;
6645 Name = (VT == MVT::i32) ? "__rt_sdiv" : "__rt_sdiv64";
6647 Name = (VT == MVT::i32) ? "__rt_udiv" : "__rt_udiv64";
6649 SDValue ES = DAG.getExternalSymbol(Name, TLI.getPointerTy(DL));
6651 ARMTargetLowering::ArgListTy Args;
6653 for (auto AI : {1, 0}) {
6655 Arg.Node = Op.getOperand(AI);
6656 Arg.Ty = Arg.Node.getValueType().getTypeForEVT(*DAG.getContext());
6657 Args.push_back(Arg);
6660 CallLoweringInfo CLI(DAG);
6663 .setCallee(CallingConv::ARM_AAPCS_VFP, VT.getTypeForEVT(*DAG.getContext()),
6664 ES, std::move(Args), 0);
6666 return LowerCallTo(CLI).first;
6669 SDValue ARMTargetLowering::LowerDIV_Windows(SDValue Op, SelectionDAG &DAG,
6670 bool Signed) const {
6671 assert(Op.getValueType() == MVT::i32 &&
6672 "unexpected type for custom lowering DIV");
6675 SDValue DBZCHK = DAG.getNode(ARMISD::WIN__DBZCHK, dl, MVT::Other,
6676 DAG.getEntryNode(), Op.getOperand(1));
6678 return LowerWindowsDIVLibCall(Op, DAG, Signed, DBZCHK);
6681 void ARMTargetLowering::ExpandDIV_Windows(
6682 SDValue Op, SelectionDAG &DAG, bool Signed,
6683 SmallVectorImpl<SDValue> &Results) const {
6684 const auto &DL = DAG.getDataLayout();
6685 const auto &TLI = DAG.getTargetLoweringInfo();
6687 assert(Op.getValueType() == MVT::i64 &&
6688 "unexpected type for custom lowering DIV");
6691 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op.getOperand(1),
6692 DAG.getConstant(0, dl, MVT::i32));
6693 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op.getOperand(1),
6694 DAG.getConstant(1, dl, MVT::i32));
6695 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::i32, Lo, Hi);
6698 DAG.getNode(ARMISD::WIN__DBZCHK, dl, MVT::Other, DAG.getEntryNode(), Or);
6700 SDValue Result = LowerWindowsDIVLibCall(Op, DAG, Signed, DBZCHK);
6702 SDValue Lower = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Result);
6703 SDValue Upper = DAG.getNode(ISD::SRL, dl, MVT::i64, Result,
6704 DAG.getConstant(32, dl, TLI.getPointerTy(DL)));
6705 Upper = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Upper);
6707 Results.push_back(Lower);
6708 Results.push_back(Upper);
6711 static SDValue LowerAtomicLoadStore(SDValue Op, SelectionDAG &DAG) {
6712 // Monotonic load/store is legal for all targets
6713 if (cast<AtomicSDNode>(Op)->getOrdering() <= Monotonic)
6716 // Acquire/Release load/store is not legal for targets without a
6717 // dmb or equivalent available.
6721 static void ReplaceREADCYCLECOUNTER(SDNode *N,
6722 SmallVectorImpl<SDValue> &Results,
6724 const ARMSubtarget *Subtarget) {
6726 // Under Power Management extensions, the cycle-count is:
6727 // mrc p15, #0, <Rt>, c9, c13, #0
6728 SDValue Ops[] = { N->getOperand(0), // Chain
6729 DAG.getConstant(Intrinsic::arm_mrc, DL, MVT::i32),
6730 DAG.getConstant(15, DL, MVT::i32),
6731 DAG.getConstant(0, DL, MVT::i32),
6732 DAG.getConstant(9, DL, MVT::i32),
6733 DAG.getConstant(13, DL, MVT::i32),
6734 DAG.getConstant(0, DL, MVT::i32)
6737 SDValue Cycles32 = DAG.getNode(ISD::INTRINSIC_W_CHAIN, DL,
6738 DAG.getVTList(MVT::i32, MVT::Other), Ops);
6739 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Cycles32,
6740 DAG.getConstant(0, DL, MVT::i32)));
6741 Results.push_back(Cycles32.getValue(1));
6744 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
6745 switch (Op.getOpcode()) {
6746 default: llvm_unreachable("Don't know how to custom lower this!");
6747 case ISD::WRITE_REGISTER: return LowerWRITE_REGISTER(Op, DAG);
6748 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
6749 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
6750 case ISD::GlobalAddress:
6751 switch (Subtarget->getTargetTriple().getObjectFormat()) {
6752 default: llvm_unreachable("unknown object format");
6754 return LowerGlobalAddressWindows(Op, DAG);
6756 return LowerGlobalAddressELF(Op, DAG);
6758 return LowerGlobalAddressDarwin(Op, DAG);
6760 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
6761 case ISD::SELECT: return LowerSELECT(Op, DAG);
6762 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
6763 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
6764 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
6765 case ISD::VASTART: return LowerVASTART(Op, DAG);
6766 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, DAG, Subtarget);
6767 case ISD::PREFETCH: return LowerPREFETCH(Op, DAG, Subtarget);
6768 case ISD::SINT_TO_FP:
6769 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
6770 case ISD::FP_TO_SINT:
6771 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
6772 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
6773 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
6774 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
6775 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
6776 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
6777 case ISD::EH_SJLJ_SETUP_DISPATCH: return LowerEH_SJLJ_SETUP_DISPATCH(Op, DAG);
6778 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
6780 case ISD::BITCAST: return ExpandBITCAST(Op.getNode(), DAG);
6783 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
6784 case ISD::SREM: return LowerREM(Op.getNode(), DAG);
6785 case ISD::UREM: return LowerREM(Op.getNode(), DAG);
6786 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
6787 case ISD::SRL_PARTS:
6788 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
6790 case ISD::CTTZ_ZERO_UNDEF: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
6791 case ISD::CTPOP: return LowerCTPOP(Op.getNode(), DAG, Subtarget);
6792 case ISD::SETCC: return LowerVSETCC(Op, DAG);
6793 case ISD::ConstantFP: return LowerConstantFP(Op, DAG, Subtarget);
6794 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG, Subtarget);
6795 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
6796 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
6797 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
6798 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
6799 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
6800 case ISD::MUL: return LowerMUL(Op, DAG);
6802 if (Subtarget->isTargetWindows())
6803 return LowerDIV_Windows(Op, DAG, /* Signed */ true);
6804 return LowerSDIV(Op, DAG);
6806 if (Subtarget->isTargetWindows())
6807 return LowerDIV_Windows(Op, DAG, /* Signed */ false);
6808 return LowerUDIV(Op, DAG);
6812 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
6817 return LowerXALUO(Op, DAG);
6818 case ISD::ATOMIC_LOAD:
6819 case ISD::ATOMIC_STORE: return LowerAtomicLoadStore(Op, DAG);
6820 case ISD::FSINCOS: return LowerFSINCOS(Op, DAG);
6822 case ISD::UDIVREM: return LowerDivRem(Op, DAG);
6823 case ISD::DYNAMIC_STACKALLOC:
6824 if (Subtarget->getTargetTriple().isWindowsItaniumEnvironment())
6825 return LowerDYNAMIC_STACKALLOC(Op, DAG);
6826 llvm_unreachable("Don't know how to custom lower this!");
6827 case ISD::FP_ROUND: return LowerFP_ROUND(Op, DAG);
6828 case ISD::FP_EXTEND: return LowerFP_EXTEND(Op, DAG);
6829 case ARMISD::WIN__DBZCHK: return SDValue();
6833 /// ReplaceNodeResults - Replace the results of node with an illegal result
6834 /// type with new values built out of custom code.
6835 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
6836 SmallVectorImpl<SDValue> &Results,
6837 SelectionDAG &DAG) const {
6839 switch (N->getOpcode()) {
6841 llvm_unreachable("Don't know how to custom expand this!");
6842 case ISD::READ_REGISTER:
6843 ExpandREAD_REGISTER(N, Results, DAG);
6846 Res = ExpandBITCAST(N, DAG);
6850 Res = Expand64BitShift(N, DAG, Subtarget);
6854 Res = LowerREM(N, DAG);
6856 case ISD::READCYCLECOUNTER:
6857 ReplaceREADCYCLECOUNTER(N, Results, DAG, Subtarget);
6861 assert(Subtarget->isTargetWindows() && "can only expand DIV on Windows");
6862 return ExpandDIV_Windows(SDValue(N, 0), DAG, N->getOpcode() == ISD::SDIV,
6866 Results.push_back(Res);
6869 //===----------------------------------------------------------------------===//
6870 // ARM Scheduler Hooks
6871 //===----------------------------------------------------------------------===//
6873 /// SetupEntryBlockForSjLj - Insert code into the entry block that creates and
6874 /// registers the function context.
6875 void ARMTargetLowering::
6876 SetupEntryBlockForSjLj(MachineInstr *MI, MachineBasicBlock *MBB,
6877 MachineBasicBlock *DispatchBB, int FI) const {
6878 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
6879 DebugLoc dl = MI->getDebugLoc();
6880 MachineFunction *MF = MBB->getParent();
6881 MachineRegisterInfo *MRI = &MF->getRegInfo();
6882 MachineConstantPool *MCP = MF->getConstantPool();
6883 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
6884 const Function *F = MF->getFunction();
6886 bool isThumb = Subtarget->isThumb();
6887 bool isThumb2 = Subtarget->isThumb2();
6889 unsigned PCLabelId = AFI->createPICLabelUId();
6890 unsigned PCAdj = (isThumb || isThumb2) ? 4 : 8;
6891 ARMConstantPoolValue *CPV =
6892 ARMConstantPoolMBB::Create(F->getContext(), DispatchBB, PCLabelId, PCAdj);
6893 unsigned CPI = MCP->getConstantPoolIndex(CPV, 4);
6895 const TargetRegisterClass *TRC = isThumb ? &ARM::tGPRRegClass
6896 : &ARM::GPRRegClass;
6898 // Grab constant pool and fixed stack memory operands.
6899 MachineMemOperand *CPMMO =
6900 MF->getMachineMemOperand(MachinePointerInfo::getConstantPool(*MF),
6901 MachineMemOperand::MOLoad, 4, 4);
6903 MachineMemOperand *FIMMOSt =
6904 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(*MF, FI),
6905 MachineMemOperand::MOStore, 4, 4);
6907 // Load the address of the dispatch MBB into the jump buffer.
6909 // Incoming value: jbuf
6910 // ldr.n r5, LCPI1_1
6913 // str r5, [$jbuf, #+4] ; &jbuf[1]
6914 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6915 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2LDRpci), NewVReg1)
6916 .addConstantPoolIndex(CPI)
6917 .addMemOperand(CPMMO));
6918 // Set the low bit because of thumb mode.
6919 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6921 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2ORRri), NewVReg2)
6922 .addReg(NewVReg1, RegState::Kill)
6924 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6925 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg3)
6926 .addReg(NewVReg2, RegState::Kill)
6928 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2STRi12))
6929 .addReg(NewVReg3, RegState::Kill)
6931 .addImm(36) // &jbuf[1] :: pc
6932 .addMemOperand(FIMMOSt));
6933 } else if (isThumb) {
6934 // Incoming value: jbuf
6935 // ldr.n r1, LCPI1_4
6939 // add r2, $jbuf, #+4 ; &jbuf[1]
6941 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6942 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tLDRpci), NewVReg1)
6943 .addConstantPoolIndex(CPI)
6944 .addMemOperand(CPMMO));
6945 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6946 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg2)
6947 .addReg(NewVReg1, RegState::Kill)
6949 // Set the low bit because of thumb mode.
6950 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6951 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tMOVi8), NewVReg3)
6952 .addReg(ARM::CPSR, RegState::Define)
6954 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6955 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tORR), NewVReg4)
6956 .addReg(ARM::CPSR, RegState::Define)
6957 .addReg(NewVReg2, RegState::Kill)
6958 .addReg(NewVReg3, RegState::Kill));
6959 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6960 BuildMI(*MBB, MI, dl, TII->get(ARM::tADDframe), NewVReg5)
6962 .addImm(36); // &jbuf[1] :: pc
6963 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tSTRi))
6964 .addReg(NewVReg4, RegState::Kill)
6965 .addReg(NewVReg5, RegState::Kill)
6967 .addMemOperand(FIMMOSt));
6969 // Incoming value: jbuf
6972 // str r1, [$jbuf, #+4] ; &jbuf[1]
6973 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6974 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::LDRi12), NewVReg1)
6975 .addConstantPoolIndex(CPI)
6977 .addMemOperand(CPMMO));
6978 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6979 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::PICADD), NewVReg2)
6980 .addReg(NewVReg1, RegState::Kill)
6981 .addImm(PCLabelId));
6982 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::STRi12))
6983 .addReg(NewVReg2, RegState::Kill)
6985 .addImm(36) // &jbuf[1] :: pc
6986 .addMemOperand(FIMMOSt));
6990 void ARMTargetLowering::EmitSjLjDispatchBlock(MachineInstr *MI,
6991 MachineBasicBlock *MBB) const {
6992 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
6993 DebugLoc dl = MI->getDebugLoc();
6994 MachineFunction *MF = MBB->getParent();
6995 MachineRegisterInfo *MRI = &MF->getRegInfo();
6996 MachineFrameInfo *MFI = MF->getFrameInfo();
6997 int FI = MFI->getFunctionContextIndex();
6999 const TargetRegisterClass *TRC = Subtarget->isThumb() ? &ARM::tGPRRegClass
7000 : &ARM::GPRnopcRegClass;
7002 // Get a mapping of the call site numbers to all of the landing pads they're
7004 DenseMap<unsigned, SmallVector<MachineBasicBlock*, 2> > CallSiteNumToLPad;
7005 unsigned MaxCSNum = 0;
7006 MachineModuleInfo &MMI = MF->getMMI();
7007 for (MachineFunction::iterator BB = MF->begin(), E = MF->end(); BB != E;
7009 if (!BB->isEHPad()) continue;
7011 // FIXME: We should assert that the EH_LABEL is the first MI in the landing
7013 for (MachineBasicBlock::iterator
7014 II = BB->begin(), IE = BB->end(); II != IE; ++II) {
7015 if (!II->isEHLabel()) continue;
7017 MCSymbol *Sym = II->getOperand(0).getMCSymbol();
7018 if (!MMI.hasCallSiteLandingPad(Sym)) continue;
7020 SmallVectorImpl<unsigned> &CallSiteIdxs = MMI.getCallSiteLandingPad(Sym);
7021 for (SmallVectorImpl<unsigned>::iterator
7022 CSI = CallSiteIdxs.begin(), CSE = CallSiteIdxs.end();
7023 CSI != CSE; ++CSI) {
7024 CallSiteNumToLPad[*CSI].push_back(&*BB);
7025 MaxCSNum = std::max(MaxCSNum, *CSI);
7031 // Get an ordered list of the machine basic blocks for the jump table.
7032 std::vector<MachineBasicBlock*> LPadList;
7033 SmallPtrSet<MachineBasicBlock*, 64> InvokeBBs;
7034 LPadList.reserve(CallSiteNumToLPad.size());
7035 for (unsigned I = 1; I <= MaxCSNum; ++I) {
7036 SmallVectorImpl<MachineBasicBlock*> &MBBList = CallSiteNumToLPad[I];
7037 for (SmallVectorImpl<MachineBasicBlock*>::iterator
7038 II = MBBList.begin(), IE = MBBList.end(); II != IE; ++II) {
7039 LPadList.push_back(*II);
7040 InvokeBBs.insert((*II)->pred_begin(), (*II)->pred_end());
7044 assert(!LPadList.empty() &&
7045 "No landing pad destinations for the dispatch jump table!");
7047 // Create the jump table and associated information.
7048 MachineJumpTableInfo *JTI =
7049 MF->getOrCreateJumpTableInfo(MachineJumpTableInfo::EK_Inline);
7050 unsigned MJTI = JTI->createJumpTableIndex(LPadList);
7051 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
7053 // Create the MBBs for the dispatch code.
7055 // Shove the dispatch's address into the return slot in the function context.
7056 MachineBasicBlock *DispatchBB = MF->CreateMachineBasicBlock();
7057 DispatchBB->setIsEHPad();
7059 MachineBasicBlock *TrapBB = MF->CreateMachineBasicBlock();
7060 unsigned trap_opcode;
7061 if (Subtarget->isThumb())
7062 trap_opcode = ARM::tTRAP;
7064 trap_opcode = Subtarget->useNaClTrap() ? ARM::TRAPNaCl : ARM::TRAP;
7066 BuildMI(TrapBB, dl, TII->get(trap_opcode));
7067 DispatchBB->addSuccessor(TrapBB);
7069 MachineBasicBlock *DispContBB = MF->CreateMachineBasicBlock();
7070 DispatchBB->addSuccessor(DispContBB);
7073 MF->insert(MF->end(), DispatchBB);
7074 MF->insert(MF->end(), DispContBB);
7075 MF->insert(MF->end(), TrapBB);
7077 // Insert code into the entry block that creates and registers the function
7079 SetupEntryBlockForSjLj(MI, MBB, DispatchBB, FI);
7081 MachineMemOperand *FIMMOLd = MF->getMachineMemOperand(
7082 MachinePointerInfo::getFixedStack(*MF, FI),
7083 MachineMemOperand::MOLoad | MachineMemOperand::MOVolatile, 4, 4);
7085 MachineInstrBuilder MIB;
7086 MIB = BuildMI(DispatchBB, dl, TII->get(ARM::Int_eh_sjlj_dispatchsetup));
7088 const ARMBaseInstrInfo *AII = static_cast<const ARMBaseInstrInfo*>(TII);
7089 const ARMBaseRegisterInfo &RI = AII->getRegisterInfo();
7091 // Add a register mask with no preserved registers. This results in all
7092 // registers being marked as clobbered.
7093 MIB.addRegMask(RI.getNoPreservedMask());
7095 unsigned NumLPads = LPadList.size();
7096 if (Subtarget->isThumb2()) {
7097 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
7098 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2LDRi12), NewVReg1)
7101 .addMemOperand(FIMMOLd));
7103 if (NumLPads < 256) {
7104 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPri))
7106 .addImm(LPadList.size()));
7108 unsigned VReg1 = MRI->createVirtualRegister(TRC);
7109 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVi16), VReg1)
7110 .addImm(NumLPads & 0xFFFF));
7112 unsigned VReg2 = VReg1;
7113 if ((NumLPads & 0xFFFF0000) != 0) {
7114 VReg2 = MRI->createVirtualRegister(TRC);
7115 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVTi16), VReg2)
7117 .addImm(NumLPads >> 16));
7120 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPrr))
7125 BuildMI(DispatchBB, dl, TII->get(ARM::t2Bcc))
7130 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
7131 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::t2LEApcrelJT),NewVReg3)
7132 .addJumpTableIndex(MJTI));
7134 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
7137 BuildMI(DispContBB, dl, TII->get(ARM::t2ADDrs), NewVReg4)
7138 .addReg(NewVReg3, RegState::Kill)
7140 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
7142 BuildMI(DispContBB, dl, TII->get(ARM::t2BR_JT))
7143 .addReg(NewVReg4, RegState::Kill)
7145 .addJumpTableIndex(MJTI);
7146 } else if (Subtarget->isThumb()) {
7147 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
7148 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRspi), NewVReg1)
7151 .addMemOperand(FIMMOLd));
7153 if (NumLPads < 256) {
7154 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPi8))
7158 MachineConstantPool *ConstantPool = MF->getConstantPool();
7159 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
7160 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
7162 // MachineConstantPool wants an explicit alignment.
7163 unsigned Align = MF->getDataLayout().getPrefTypeAlignment(Int32Ty);
7165 Align = MF->getDataLayout().getTypeAllocSize(C->getType());
7166 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
7168 unsigned VReg1 = MRI->createVirtualRegister(TRC);
7169 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRpci))
7170 .addReg(VReg1, RegState::Define)
7171 .addConstantPoolIndex(Idx));
7172 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPr))
7177 BuildMI(DispatchBB, dl, TII->get(ARM::tBcc))
7182 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
7183 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLSLri), NewVReg2)
7184 .addReg(ARM::CPSR, RegState::Define)
7188 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
7189 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLEApcrelJT), NewVReg3)
7190 .addJumpTableIndex(MJTI));
7192 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
7193 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg4)
7194 .addReg(ARM::CPSR, RegState::Define)
7195 .addReg(NewVReg2, RegState::Kill)
7198 MachineMemOperand *JTMMOLd = MF->getMachineMemOperand(
7199 MachinePointerInfo::getJumpTable(*MF), MachineMemOperand::MOLoad, 4, 4);
7201 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
7202 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLDRi), NewVReg5)
7203 .addReg(NewVReg4, RegState::Kill)
7205 .addMemOperand(JTMMOLd));
7207 unsigned NewVReg6 = NewVReg5;
7208 if (RelocM == Reloc::PIC_) {
7209 NewVReg6 = MRI->createVirtualRegister(TRC);
7210 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg6)
7211 .addReg(ARM::CPSR, RegState::Define)
7212 .addReg(NewVReg5, RegState::Kill)
7216 BuildMI(DispContBB, dl, TII->get(ARM::tBR_JTr))
7217 .addReg(NewVReg6, RegState::Kill)
7218 .addJumpTableIndex(MJTI);
7220 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
7221 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRi12), NewVReg1)
7224 .addMemOperand(FIMMOLd));
7226 if (NumLPads < 256) {
7227 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPri))
7230 } else if (Subtarget->hasV6T2Ops() && isUInt<16>(NumLPads)) {
7231 unsigned VReg1 = MRI->createVirtualRegister(TRC);
7232 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVi16), VReg1)
7233 .addImm(NumLPads & 0xFFFF));
7235 unsigned VReg2 = VReg1;
7236 if ((NumLPads & 0xFFFF0000) != 0) {
7237 VReg2 = MRI->createVirtualRegister(TRC);
7238 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVTi16), VReg2)
7240 .addImm(NumLPads >> 16));
7243 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
7247 MachineConstantPool *ConstantPool = MF->getConstantPool();
7248 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
7249 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
7251 // MachineConstantPool wants an explicit alignment.
7252 unsigned Align = MF->getDataLayout().getPrefTypeAlignment(Int32Ty);
7254 Align = MF->getDataLayout().getTypeAllocSize(C->getType());
7255 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
7257 unsigned VReg1 = MRI->createVirtualRegister(TRC);
7258 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRcp))
7259 .addReg(VReg1, RegState::Define)
7260 .addConstantPoolIndex(Idx)
7262 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
7264 .addReg(VReg1, RegState::Kill));
7267 BuildMI(DispatchBB, dl, TII->get(ARM::Bcc))
7272 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
7274 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::MOVsi), NewVReg3)
7276 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
7277 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
7278 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::LEApcrelJT), NewVReg4)
7279 .addJumpTableIndex(MJTI));
7281 MachineMemOperand *JTMMOLd = MF->getMachineMemOperand(
7282 MachinePointerInfo::getJumpTable(*MF), MachineMemOperand::MOLoad, 4, 4);
7283 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
7285 BuildMI(DispContBB, dl, TII->get(ARM::LDRrs), NewVReg5)
7286 .addReg(NewVReg3, RegState::Kill)
7289 .addMemOperand(JTMMOLd));
7291 if (RelocM == Reloc::PIC_) {
7292 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTadd))
7293 .addReg(NewVReg5, RegState::Kill)
7295 .addJumpTableIndex(MJTI);
7297 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTr))
7298 .addReg(NewVReg5, RegState::Kill)
7299 .addJumpTableIndex(MJTI);
7303 // Add the jump table entries as successors to the MBB.
7304 SmallPtrSet<MachineBasicBlock*, 8> SeenMBBs;
7305 for (std::vector<MachineBasicBlock*>::iterator
7306 I = LPadList.begin(), E = LPadList.end(); I != E; ++I) {
7307 MachineBasicBlock *CurMBB = *I;
7308 if (SeenMBBs.insert(CurMBB).second)
7309 DispContBB->addSuccessor(CurMBB);
7312 // N.B. the order the invoke BBs are processed in doesn't matter here.
7313 const MCPhysReg *SavedRegs = RI.getCalleeSavedRegs(MF);
7314 SmallVector<MachineBasicBlock*, 64> MBBLPads;
7315 for (MachineBasicBlock *BB : InvokeBBs) {
7317 // Remove the landing pad successor from the invoke block and replace it
7318 // with the new dispatch block.
7319 SmallVector<MachineBasicBlock*, 4> Successors(BB->succ_begin(),
7321 while (!Successors.empty()) {
7322 MachineBasicBlock *SMBB = Successors.pop_back_val();
7323 if (SMBB->isEHPad()) {
7324 BB->removeSuccessor(SMBB);
7325 MBBLPads.push_back(SMBB);
7329 BB->addSuccessor(DispatchBB);
7331 // Find the invoke call and mark all of the callee-saved registers as
7332 // 'implicit defined' so that they're spilled. This prevents code from
7333 // moving instructions to before the EH block, where they will never be
7335 for (MachineBasicBlock::reverse_iterator
7336 II = BB->rbegin(), IE = BB->rend(); II != IE; ++II) {
7337 if (!II->isCall()) continue;
7339 DenseMap<unsigned, bool> DefRegs;
7340 for (MachineInstr::mop_iterator
7341 OI = II->operands_begin(), OE = II->operands_end();
7343 if (!OI->isReg()) continue;
7344 DefRegs[OI->getReg()] = true;
7347 MachineInstrBuilder MIB(*MF, &*II);
7349 for (unsigned i = 0; SavedRegs[i] != 0; ++i) {
7350 unsigned Reg = SavedRegs[i];
7351 if (Subtarget->isThumb2() &&
7352 !ARM::tGPRRegClass.contains(Reg) &&
7353 !ARM::hGPRRegClass.contains(Reg))
7355 if (Subtarget->isThumb1Only() && !ARM::tGPRRegClass.contains(Reg))
7357 if (!Subtarget->isThumb() && !ARM::GPRRegClass.contains(Reg))
7360 MIB.addReg(Reg, RegState::ImplicitDefine | RegState::Dead);
7367 // Mark all former landing pads as non-landing pads. The dispatch is the only
7369 for (SmallVectorImpl<MachineBasicBlock*>::iterator
7370 I = MBBLPads.begin(), E = MBBLPads.end(); I != E; ++I)
7371 (*I)->setIsEHPad(false);
7373 // The instruction is gone now.
7374 MI->eraseFromParent();
7378 MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) {
7379 for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
7380 E = MBB->succ_end(); I != E; ++I)
7383 llvm_unreachable("Expecting a BB with two successors!");
7386 /// Return the load opcode for a given load size. If load size >= 8,
7387 /// neon opcode will be returned.
7388 static unsigned getLdOpcode(unsigned LdSize, bool IsThumb1, bool IsThumb2) {
7390 return LdSize == 16 ? ARM::VLD1q32wb_fixed
7391 : LdSize == 8 ? ARM::VLD1d32wb_fixed : 0;
7393 return LdSize == 4 ? ARM::tLDRi
7394 : LdSize == 2 ? ARM::tLDRHi
7395 : LdSize == 1 ? ARM::tLDRBi : 0;
7397 return LdSize == 4 ? ARM::t2LDR_POST
7398 : LdSize == 2 ? ARM::t2LDRH_POST
7399 : LdSize == 1 ? ARM::t2LDRB_POST : 0;
7400 return LdSize == 4 ? ARM::LDR_POST_IMM
7401 : LdSize == 2 ? ARM::LDRH_POST
7402 : LdSize == 1 ? ARM::LDRB_POST_IMM : 0;
7405 /// Return the store opcode for a given store size. If store size >= 8,
7406 /// neon opcode will be returned.
7407 static unsigned getStOpcode(unsigned StSize, bool IsThumb1, bool IsThumb2) {
7409 return StSize == 16 ? ARM::VST1q32wb_fixed
7410 : StSize == 8 ? ARM::VST1d32wb_fixed : 0;
7412 return StSize == 4 ? ARM::tSTRi
7413 : StSize == 2 ? ARM::tSTRHi
7414 : StSize == 1 ? ARM::tSTRBi : 0;
7416 return StSize == 4 ? ARM::t2STR_POST
7417 : StSize == 2 ? ARM::t2STRH_POST
7418 : StSize == 1 ? ARM::t2STRB_POST : 0;
7419 return StSize == 4 ? ARM::STR_POST_IMM
7420 : StSize == 2 ? ARM::STRH_POST
7421 : StSize == 1 ? ARM::STRB_POST_IMM : 0;
7424 /// Emit a post-increment load operation with given size. The instructions
7425 /// will be added to BB at Pos.
7426 static void emitPostLd(MachineBasicBlock *BB, MachineInstr *Pos,
7427 const TargetInstrInfo *TII, DebugLoc dl,
7428 unsigned LdSize, unsigned Data, unsigned AddrIn,
7429 unsigned AddrOut, bool IsThumb1, bool IsThumb2) {
7430 unsigned LdOpc = getLdOpcode(LdSize, IsThumb1, IsThumb2);
7431 assert(LdOpc != 0 && "Should have a load opcode");
7433 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
7434 .addReg(AddrOut, RegState::Define).addReg(AddrIn)
7436 } else if (IsThumb1) {
7437 // load + update AddrIn
7438 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
7439 .addReg(AddrIn).addImm(0));
7440 MachineInstrBuilder MIB =
7441 BuildMI(*BB, Pos, dl, TII->get(ARM::tADDi8), AddrOut);
7442 MIB = AddDefaultT1CC(MIB);
7443 MIB.addReg(AddrIn).addImm(LdSize);
7444 AddDefaultPred(MIB);
7445 } else if (IsThumb2) {
7446 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
7447 .addReg(AddrOut, RegState::Define).addReg(AddrIn)
7450 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
7451 .addReg(AddrOut, RegState::Define).addReg(AddrIn)
7452 .addReg(0).addImm(LdSize));
7456 /// Emit a post-increment store operation with given size. The instructions
7457 /// will be added to BB at Pos.
7458 static void emitPostSt(MachineBasicBlock *BB, MachineInstr *Pos,
7459 const TargetInstrInfo *TII, DebugLoc dl,
7460 unsigned StSize, unsigned Data, unsigned AddrIn,
7461 unsigned AddrOut, bool IsThumb1, bool IsThumb2) {
7462 unsigned StOpc = getStOpcode(StSize, IsThumb1, IsThumb2);
7463 assert(StOpc != 0 && "Should have a store opcode");
7465 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc), AddrOut)
7466 .addReg(AddrIn).addImm(0).addReg(Data));
7467 } else if (IsThumb1) {
7468 // store + update AddrIn
7469 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc)).addReg(Data)
7470 .addReg(AddrIn).addImm(0));
7471 MachineInstrBuilder MIB =
7472 BuildMI(*BB, Pos, dl, TII->get(ARM::tADDi8), AddrOut);
7473 MIB = AddDefaultT1CC(MIB);
7474 MIB.addReg(AddrIn).addImm(StSize);
7475 AddDefaultPred(MIB);
7476 } else if (IsThumb2) {
7477 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc), AddrOut)
7478 .addReg(Data).addReg(AddrIn).addImm(StSize));
7480 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc), AddrOut)
7481 .addReg(Data).addReg(AddrIn).addReg(0)
7487 ARMTargetLowering::EmitStructByval(MachineInstr *MI,
7488 MachineBasicBlock *BB) const {
7489 // This pseudo instruction has 3 operands: dst, src, size
7490 // We expand it to a loop if size > Subtarget->getMaxInlineSizeThreshold().
7491 // Otherwise, we will generate unrolled scalar copies.
7492 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
7493 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7494 MachineFunction::iterator It = ++BB->getIterator();
7496 unsigned dest = MI->getOperand(0).getReg();
7497 unsigned src = MI->getOperand(1).getReg();
7498 unsigned SizeVal = MI->getOperand(2).getImm();
7499 unsigned Align = MI->getOperand(3).getImm();
7500 DebugLoc dl = MI->getDebugLoc();
7502 MachineFunction *MF = BB->getParent();
7503 MachineRegisterInfo &MRI = MF->getRegInfo();
7504 unsigned UnitSize = 0;
7505 const TargetRegisterClass *TRC = nullptr;
7506 const TargetRegisterClass *VecTRC = nullptr;
7508 bool IsThumb1 = Subtarget->isThumb1Only();
7509 bool IsThumb2 = Subtarget->isThumb2();
7513 } else if (Align & 2) {
7516 // Check whether we can use NEON instructions.
7517 if (!MF->getFunction()->hasFnAttribute(Attribute::NoImplicitFloat) &&
7518 Subtarget->hasNEON()) {
7519 if ((Align % 16 == 0) && SizeVal >= 16)
7521 else if ((Align % 8 == 0) && SizeVal >= 8)
7524 // Can't use NEON instructions.
7529 // Select the correct opcode and register class for unit size load/store
7530 bool IsNeon = UnitSize >= 8;
7531 TRC = (IsThumb1 || IsThumb2) ? &ARM::tGPRRegClass : &ARM::GPRRegClass;
7533 VecTRC = UnitSize == 16 ? &ARM::DPairRegClass
7534 : UnitSize == 8 ? &ARM::DPRRegClass
7537 unsigned BytesLeft = SizeVal % UnitSize;
7538 unsigned LoopSize = SizeVal - BytesLeft;
7540 if (SizeVal <= Subtarget->getMaxInlineSizeThreshold()) {
7541 // Use LDR and STR to copy.
7542 // [scratch, srcOut] = LDR_POST(srcIn, UnitSize)
7543 // [destOut] = STR_POST(scratch, destIn, UnitSize)
7544 unsigned srcIn = src;
7545 unsigned destIn = dest;
7546 for (unsigned i = 0; i < LoopSize; i+=UnitSize) {
7547 unsigned srcOut = MRI.createVirtualRegister(TRC);
7548 unsigned destOut = MRI.createVirtualRegister(TRC);
7549 unsigned scratch = MRI.createVirtualRegister(IsNeon ? VecTRC : TRC);
7550 emitPostLd(BB, MI, TII, dl, UnitSize, scratch, srcIn, srcOut,
7551 IsThumb1, IsThumb2);
7552 emitPostSt(BB, MI, TII, dl, UnitSize, scratch, destIn, destOut,
7553 IsThumb1, IsThumb2);
7558 // Handle the leftover bytes with LDRB and STRB.
7559 // [scratch, srcOut] = LDRB_POST(srcIn, 1)
7560 // [destOut] = STRB_POST(scratch, destIn, 1)
7561 for (unsigned i = 0; i < BytesLeft; i++) {
7562 unsigned srcOut = MRI.createVirtualRegister(TRC);
7563 unsigned destOut = MRI.createVirtualRegister(TRC);
7564 unsigned scratch = MRI.createVirtualRegister(TRC);
7565 emitPostLd(BB, MI, TII, dl, 1, scratch, srcIn, srcOut,
7566 IsThumb1, IsThumb2);
7567 emitPostSt(BB, MI, TII, dl, 1, scratch, destIn, destOut,
7568 IsThumb1, IsThumb2);
7572 MI->eraseFromParent(); // The instruction is gone now.
7576 // Expand the pseudo op to a loop.
7579 // movw varEnd, # --> with thumb2
7581 // ldrcp varEnd, idx --> without thumb2
7582 // fallthrough --> loopMBB
7584 // PHI varPhi, varEnd, varLoop
7585 // PHI srcPhi, src, srcLoop
7586 // PHI destPhi, dst, destLoop
7587 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
7588 // [destLoop] = STR_POST(scratch, destPhi, UnitSize)
7589 // subs varLoop, varPhi, #UnitSize
7591 // fallthrough --> exitMBB
7593 // epilogue to handle left-over bytes
7594 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
7595 // [destOut] = STRB_POST(scratch, destLoop, 1)
7596 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
7597 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
7598 MF->insert(It, loopMBB);
7599 MF->insert(It, exitMBB);
7601 // Transfer the remainder of BB and its successor edges to exitMBB.
7602 exitMBB->splice(exitMBB->begin(), BB,
7603 std::next(MachineBasicBlock::iterator(MI)), BB->end());
7604 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
7606 // Load an immediate to varEnd.
7607 unsigned varEnd = MRI.createVirtualRegister(TRC);
7608 if (Subtarget->useMovt(*MF)) {
7609 unsigned Vtmp = varEnd;
7610 if ((LoopSize & 0xFFFF0000) != 0)
7611 Vtmp = MRI.createVirtualRegister(TRC);
7612 AddDefaultPred(BuildMI(BB, dl,
7613 TII->get(IsThumb2 ? ARM::t2MOVi16 : ARM::MOVi16),
7614 Vtmp).addImm(LoopSize & 0xFFFF));
7616 if ((LoopSize & 0xFFFF0000) != 0)
7617 AddDefaultPred(BuildMI(BB, dl,
7618 TII->get(IsThumb2 ? ARM::t2MOVTi16 : ARM::MOVTi16),
7621 .addImm(LoopSize >> 16));
7623 MachineConstantPool *ConstantPool = MF->getConstantPool();
7624 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
7625 const Constant *C = ConstantInt::get(Int32Ty, LoopSize);
7627 // MachineConstantPool wants an explicit alignment.
7628 unsigned Align = MF->getDataLayout().getPrefTypeAlignment(Int32Ty);
7630 Align = MF->getDataLayout().getTypeAllocSize(C->getType());
7631 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
7634 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(ARM::tLDRpci)).addReg(
7635 varEnd, RegState::Define).addConstantPoolIndex(Idx));
7637 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(ARM::LDRcp)).addReg(
7638 varEnd, RegState::Define).addConstantPoolIndex(Idx).addImm(0));
7640 BB->addSuccessor(loopMBB);
7642 // Generate the loop body:
7643 // varPhi = PHI(varLoop, varEnd)
7644 // srcPhi = PHI(srcLoop, src)
7645 // destPhi = PHI(destLoop, dst)
7646 MachineBasicBlock *entryBB = BB;
7648 unsigned varLoop = MRI.createVirtualRegister(TRC);
7649 unsigned varPhi = MRI.createVirtualRegister(TRC);
7650 unsigned srcLoop = MRI.createVirtualRegister(TRC);
7651 unsigned srcPhi = MRI.createVirtualRegister(TRC);
7652 unsigned destLoop = MRI.createVirtualRegister(TRC);
7653 unsigned destPhi = MRI.createVirtualRegister(TRC);
7655 BuildMI(*BB, BB->begin(), dl, TII->get(ARM::PHI), varPhi)
7656 .addReg(varLoop).addMBB(loopMBB)
7657 .addReg(varEnd).addMBB(entryBB);
7658 BuildMI(BB, dl, TII->get(ARM::PHI), srcPhi)
7659 .addReg(srcLoop).addMBB(loopMBB)
7660 .addReg(src).addMBB(entryBB);
7661 BuildMI(BB, dl, TII->get(ARM::PHI), destPhi)
7662 .addReg(destLoop).addMBB(loopMBB)
7663 .addReg(dest).addMBB(entryBB);
7665 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
7666 // [destLoop] = STR_POST(scratch, destPhi, UnitSiz)
7667 unsigned scratch = MRI.createVirtualRegister(IsNeon ? VecTRC : TRC);
7668 emitPostLd(BB, BB->end(), TII, dl, UnitSize, scratch, srcPhi, srcLoop,
7669 IsThumb1, IsThumb2);
7670 emitPostSt(BB, BB->end(), TII, dl, UnitSize, scratch, destPhi, destLoop,
7671 IsThumb1, IsThumb2);
7673 // Decrement loop variable by UnitSize.
7675 MachineInstrBuilder MIB =
7676 BuildMI(*BB, BB->end(), dl, TII->get(ARM::tSUBi8), varLoop);
7677 MIB = AddDefaultT1CC(MIB);
7678 MIB.addReg(varPhi).addImm(UnitSize);
7679 AddDefaultPred(MIB);
7681 MachineInstrBuilder MIB =
7682 BuildMI(*BB, BB->end(), dl,
7683 TII->get(IsThumb2 ? ARM::t2SUBri : ARM::SUBri), varLoop);
7684 AddDefaultCC(AddDefaultPred(MIB.addReg(varPhi).addImm(UnitSize)));
7685 MIB->getOperand(5).setReg(ARM::CPSR);
7686 MIB->getOperand(5).setIsDef(true);
7688 BuildMI(*BB, BB->end(), dl,
7689 TII->get(IsThumb1 ? ARM::tBcc : IsThumb2 ? ARM::t2Bcc : ARM::Bcc))
7690 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
7692 // loopMBB can loop back to loopMBB or fall through to exitMBB.
7693 BB->addSuccessor(loopMBB);
7694 BB->addSuccessor(exitMBB);
7696 // Add epilogue to handle BytesLeft.
7698 MachineInstr *StartOfExit = exitMBB->begin();
7700 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
7701 // [destOut] = STRB_POST(scratch, destLoop, 1)
7702 unsigned srcIn = srcLoop;
7703 unsigned destIn = destLoop;
7704 for (unsigned i = 0; i < BytesLeft; i++) {
7705 unsigned srcOut = MRI.createVirtualRegister(TRC);
7706 unsigned destOut = MRI.createVirtualRegister(TRC);
7707 unsigned scratch = MRI.createVirtualRegister(TRC);
7708 emitPostLd(BB, StartOfExit, TII, dl, 1, scratch, srcIn, srcOut,
7709 IsThumb1, IsThumb2);
7710 emitPostSt(BB, StartOfExit, TII, dl, 1, scratch, destIn, destOut,
7711 IsThumb1, IsThumb2);
7716 MI->eraseFromParent(); // The instruction is gone now.
7721 ARMTargetLowering::EmitLowered__chkstk(MachineInstr *MI,
7722 MachineBasicBlock *MBB) const {
7723 const TargetMachine &TM = getTargetMachine();
7724 const TargetInstrInfo &TII = *Subtarget->getInstrInfo();
7725 DebugLoc DL = MI->getDebugLoc();
7727 assert(Subtarget->isTargetWindows() &&
7728 "__chkstk is only supported on Windows");
7729 assert(Subtarget->isThumb2() && "Windows on ARM requires Thumb-2 mode");
7731 // __chkstk takes the number of words to allocate on the stack in R4, and
7732 // returns the stack adjustment in number of bytes in R4. This will not
7733 // clober any other registers (other than the obvious lr).
7735 // Although, technically, IP should be considered a register which may be
7736 // clobbered, the call itself will not touch it. Windows on ARM is a pure
7737 // thumb-2 environment, so there is no interworking required. As a result, we
7738 // do not expect a veneer to be emitted by the linker, clobbering IP.
7740 // Each module receives its own copy of __chkstk, so no import thunk is
7741 // required, again, ensuring that IP is not clobbered.
7743 // Finally, although some linkers may theoretically provide a trampoline for
7744 // out of range calls (which is quite common due to a 32M range limitation of
7745 // branches for Thumb), we can generate the long-call version via
7746 // -mcmodel=large, alleviating the need for the trampoline which may clobber
7749 switch (TM.getCodeModel()) {
7750 case CodeModel::Small:
7751 case CodeModel::Medium:
7752 case CodeModel::Default:
7753 case CodeModel::Kernel:
7754 BuildMI(*MBB, MI, DL, TII.get(ARM::tBL))
7755 .addImm((unsigned)ARMCC::AL).addReg(0)
7756 .addExternalSymbol("__chkstk")
7757 .addReg(ARM::R4, RegState::Implicit | RegState::Kill)
7758 .addReg(ARM::R4, RegState::Implicit | RegState::Define)
7759 .addReg(ARM::R12, RegState::Implicit | RegState::Define | RegState::Dead);
7761 case CodeModel::Large:
7762 case CodeModel::JITDefault: {
7763 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
7764 unsigned Reg = MRI.createVirtualRegister(&ARM::rGPRRegClass);
7766 BuildMI(*MBB, MI, DL, TII.get(ARM::t2MOVi32imm), Reg)
7767 .addExternalSymbol("__chkstk");
7768 BuildMI(*MBB, MI, DL, TII.get(ARM::tBLXr))
7769 .addImm((unsigned)ARMCC::AL).addReg(0)
7770 .addReg(Reg, RegState::Kill)
7771 .addReg(ARM::R4, RegState::Implicit | RegState::Kill)
7772 .addReg(ARM::R4, RegState::Implicit | RegState::Define)
7773 .addReg(ARM::R12, RegState::Implicit | RegState::Define | RegState::Dead);
7778 AddDefaultCC(AddDefaultPred(BuildMI(*MBB, MI, DL, TII.get(ARM::t2SUBrr),
7780 .addReg(ARM::SP).addReg(ARM::R4)));
7782 MI->eraseFromParent();
7787 ARMTargetLowering::EmitLowered__dbzchk(MachineInstr *MI,
7788 MachineBasicBlock *MBB) const {
7789 DebugLoc DL = MI->getDebugLoc();
7790 MachineFunction *MF = MBB->getParent();
7791 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
7793 MachineBasicBlock *ContBB = MF->CreateMachineBasicBlock();
7794 MF->push_back(ContBB);
7795 ContBB->splice(ContBB->begin(), MBB,
7796 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
7797 MBB->addSuccessor(ContBB);
7799 MachineBasicBlock *TrapBB = MF->CreateMachineBasicBlock();
7800 MF->push_back(TrapBB);
7801 BuildMI(TrapBB, DL, TII->get(ARM::t2UDF)).addImm(249);
7802 MBB->addSuccessor(TrapBB);
7804 BuildMI(*MBB, MI, DL, TII->get(ARM::tCBZ))
7805 .addReg(MI->getOperand(0).getReg())
7808 MI->eraseFromParent();
7813 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
7814 MachineBasicBlock *BB) const {
7815 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
7816 DebugLoc dl = MI->getDebugLoc();
7817 bool isThumb2 = Subtarget->isThumb2();
7818 switch (MI->getOpcode()) {
7821 llvm_unreachable("Unexpected instr type to insert");
7823 // The Thumb2 pre-indexed stores have the same MI operands, they just
7824 // define them differently in the .td files from the isel patterns, so
7825 // they need pseudos.
7826 case ARM::t2STR_preidx:
7827 MI->setDesc(TII->get(ARM::t2STR_PRE));
7829 case ARM::t2STRB_preidx:
7830 MI->setDesc(TII->get(ARM::t2STRB_PRE));
7832 case ARM::t2STRH_preidx:
7833 MI->setDesc(TII->get(ARM::t2STRH_PRE));
7836 case ARM::STRi_preidx:
7837 case ARM::STRBi_preidx: {
7838 unsigned NewOpc = MI->getOpcode() == ARM::STRi_preidx ?
7839 ARM::STR_PRE_IMM : ARM::STRB_PRE_IMM;
7840 // Decode the offset.
7841 unsigned Offset = MI->getOperand(4).getImm();
7842 bool isSub = ARM_AM::getAM2Op(Offset) == ARM_AM::sub;
7843 Offset = ARM_AM::getAM2Offset(Offset);
7847 MachineMemOperand *MMO = *MI->memoperands_begin();
7848 BuildMI(*BB, MI, dl, TII->get(NewOpc))
7849 .addOperand(MI->getOperand(0)) // Rn_wb
7850 .addOperand(MI->getOperand(1)) // Rt
7851 .addOperand(MI->getOperand(2)) // Rn
7852 .addImm(Offset) // offset (skip GPR==zero_reg)
7853 .addOperand(MI->getOperand(5)) // pred
7854 .addOperand(MI->getOperand(6))
7855 .addMemOperand(MMO);
7856 MI->eraseFromParent();
7859 case ARM::STRr_preidx:
7860 case ARM::STRBr_preidx:
7861 case ARM::STRH_preidx: {
7863 switch (MI->getOpcode()) {
7864 default: llvm_unreachable("unexpected opcode!");
7865 case ARM::STRr_preidx: NewOpc = ARM::STR_PRE_REG; break;
7866 case ARM::STRBr_preidx: NewOpc = ARM::STRB_PRE_REG; break;
7867 case ARM::STRH_preidx: NewOpc = ARM::STRH_PRE; break;
7869 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(NewOpc));
7870 for (unsigned i = 0; i < MI->getNumOperands(); ++i)
7871 MIB.addOperand(MI->getOperand(i));
7872 MI->eraseFromParent();
7876 case ARM::tMOVCCr_pseudo: {
7877 // To "insert" a SELECT_CC instruction, we actually have to insert the
7878 // diamond control-flow pattern. The incoming instruction knows the
7879 // destination vreg to set, the condition code register to branch on, the
7880 // true/false values to select between, and a branch opcode to use.
7881 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7882 MachineFunction::iterator It = ++BB->getIterator();
7887 // cmpTY ccX, r1, r2
7889 // fallthrough --> copy0MBB
7890 MachineBasicBlock *thisMBB = BB;
7891 MachineFunction *F = BB->getParent();
7892 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
7893 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
7894 F->insert(It, copy0MBB);
7895 F->insert(It, sinkMBB);
7897 // Transfer the remainder of BB and its successor edges to sinkMBB.
7898 sinkMBB->splice(sinkMBB->begin(), BB,
7899 std::next(MachineBasicBlock::iterator(MI)), BB->end());
7900 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
7902 BB->addSuccessor(copy0MBB);
7903 BB->addSuccessor(sinkMBB);
7905 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
7906 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
7909 // %FalseValue = ...
7910 // # fallthrough to sinkMBB
7913 // Update machine-CFG edges
7914 BB->addSuccessor(sinkMBB);
7917 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
7920 BuildMI(*BB, BB->begin(), dl,
7921 TII->get(ARM::PHI), MI->getOperand(0).getReg())
7922 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
7923 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
7925 MI->eraseFromParent(); // The pseudo instruction is gone now.
7930 case ARM::BCCZi64: {
7931 // If there is an unconditional branch to the other successor, remove it.
7932 BB->erase(std::next(MachineBasicBlock::iterator(MI)), BB->end());
7934 // Compare both parts that make up the double comparison separately for
7936 bool RHSisZero = MI->getOpcode() == ARM::BCCZi64;
7938 unsigned LHS1 = MI->getOperand(1).getReg();
7939 unsigned LHS2 = MI->getOperand(2).getReg();
7941 AddDefaultPred(BuildMI(BB, dl,
7942 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7943 .addReg(LHS1).addImm(0));
7944 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7945 .addReg(LHS2).addImm(0)
7946 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
7948 unsigned RHS1 = MI->getOperand(3).getReg();
7949 unsigned RHS2 = MI->getOperand(4).getReg();
7950 AddDefaultPred(BuildMI(BB, dl,
7951 TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
7952 .addReg(LHS1).addReg(RHS1));
7953 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
7954 .addReg(LHS2).addReg(RHS2)
7955 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
7958 MachineBasicBlock *destMBB = MI->getOperand(RHSisZero ? 3 : 5).getMBB();
7959 MachineBasicBlock *exitMBB = OtherSucc(BB, destMBB);
7960 if (MI->getOperand(0).getImm() == ARMCC::NE)
7961 std::swap(destMBB, exitMBB);
7963 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
7964 .addMBB(destMBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
7966 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2B)).addMBB(exitMBB));
7968 BuildMI(BB, dl, TII->get(ARM::B)) .addMBB(exitMBB);
7970 MI->eraseFromParent(); // The pseudo instruction is gone now.
7974 case ARM::Int_eh_sjlj_setjmp:
7975 case ARM::Int_eh_sjlj_setjmp_nofp:
7976 case ARM::tInt_eh_sjlj_setjmp:
7977 case ARM::t2Int_eh_sjlj_setjmp:
7978 case ARM::t2Int_eh_sjlj_setjmp_nofp:
7981 case ARM::Int_eh_sjlj_setup_dispatch:
7982 EmitSjLjDispatchBlock(MI, BB);
7987 // To insert an ABS instruction, we have to insert the
7988 // diamond control-flow pattern. The incoming instruction knows the
7989 // source vreg to test against 0, the destination vreg to set,
7990 // the condition code register to branch on, the
7991 // true/false values to select between, and a branch opcode to use.
7996 // BCC (branch to SinkBB if V0 >= 0)
7997 // RSBBB: V3 = RSBri V2, 0 (compute ABS if V2 < 0)
7998 // SinkBB: V1 = PHI(V2, V3)
7999 const BasicBlock *LLVM_BB = BB->getBasicBlock();
8000 MachineFunction::iterator BBI = ++BB->getIterator();
8001 MachineFunction *Fn = BB->getParent();
8002 MachineBasicBlock *RSBBB = Fn->CreateMachineBasicBlock(LLVM_BB);
8003 MachineBasicBlock *SinkBB = Fn->CreateMachineBasicBlock(LLVM_BB);
8004 Fn->insert(BBI, RSBBB);
8005 Fn->insert(BBI, SinkBB);
8007 unsigned int ABSSrcReg = MI->getOperand(1).getReg();
8008 unsigned int ABSDstReg = MI->getOperand(0).getReg();
8009 bool ABSSrcKIll = MI->getOperand(1).isKill();
8010 bool isThumb2 = Subtarget->isThumb2();
8011 MachineRegisterInfo &MRI = Fn->getRegInfo();
8012 // In Thumb mode S must not be specified if source register is the SP or
8013 // PC and if destination register is the SP, so restrict register class
8014 unsigned NewRsbDstReg =
8015 MRI.createVirtualRegister(isThumb2 ? &ARM::rGPRRegClass : &ARM::GPRRegClass);
8017 // Transfer the remainder of BB and its successor edges to sinkMBB.
8018 SinkBB->splice(SinkBB->begin(), BB,
8019 std::next(MachineBasicBlock::iterator(MI)), BB->end());
8020 SinkBB->transferSuccessorsAndUpdatePHIs(BB);
8022 BB->addSuccessor(RSBBB);
8023 BB->addSuccessor(SinkBB);
8025 // fall through to SinkMBB
8026 RSBBB->addSuccessor(SinkBB);
8028 // insert a cmp at the end of BB
8029 AddDefaultPred(BuildMI(BB, dl,
8030 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
8031 .addReg(ABSSrcReg).addImm(0));
8033 // insert a bcc with opposite CC to ARMCC::MI at the end of BB
8035 TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)).addMBB(SinkBB)
8036 .addImm(ARMCC::getOppositeCondition(ARMCC::MI)).addReg(ARM::CPSR);
8038 // insert rsbri in RSBBB
8039 // Note: BCC and rsbri will be converted into predicated rsbmi
8040 // by if-conversion pass
8041 BuildMI(*RSBBB, RSBBB->begin(), dl,
8042 TII->get(isThumb2 ? ARM::t2RSBri : ARM::RSBri), NewRsbDstReg)
8043 .addReg(ABSSrcReg, ABSSrcKIll ? RegState::Kill : 0)
8044 .addImm(0).addImm((unsigned)ARMCC::AL).addReg(0).addReg(0);
8046 // insert PHI in SinkBB,
8047 // reuse ABSDstReg to not change uses of ABS instruction
8048 BuildMI(*SinkBB, SinkBB->begin(), dl,
8049 TII->get(ARM::PHI), ABSDstReg)
8050 .addReg(NewRsbDstReg).addMBB(RSBBB)
8051 .addReg(ABSSrcReg).addMBB(BB);
8053 // remove ABS instruction
8054 MI->eraseFromParent();
8056 // return last added BB
8059 case ARM::COPY_STRUCT_BYVAL_I32:
8061 return EmitStructByval(MI, BB);
8062 case ARM::WIN__CHKSTK:
8063 return EmitLowered__chkstk(MI, BB);
8064 case ARM::WIN__DBZCHK:
8065 return EmitLowered__dbzchk(MI, BB);
8069 /// \brief Attaches vregs to MEMCPY that it will use as scratch registers
8070 /// when it is expanded into LDM/STM. This is done as a post-isel lowering
8071 /// instead of as a custom inserter because we need the use list from the SDNode.
8072 static void attachMEMCPYScratchRegs(const ARMSubtarget *Subtarget,
8073 MachineInstr *MI, const SDNode *Node) {
8074 bool isThumb1 = Subtarget->isThumb1Only();
8076 DebugLoc DL = MI->getDebugLoc();
8077 MachineFunction *MF = MI->getParent()->getParent();
8078 MachineRegisterInfo &MRI = MF->getRegInfo();
8079 MachineInstrBuilder MIB(*MF, MI);
8081 // If the new dst/src is unused mark it as dead.
8082 if (!Node->hasAnyUseOfValue(0)) {
8083 MI->getOperand(0).setIsDead(true);
8085 if (!Node->hasAnyUseOfValue(1)) {
8086 MI->getOperand(1).setIsDead(true);
8089 // The MEMCPY both defines and kills the scratch registers.
8090 for (unsigned I = 0; I != MI->getOperand(4).getImm(); ++I) {
8091 unsigned TmpReg = MRI.createVirtualRegister(isThumb1 ? &ARM::tGPRRegClass
8092 : &ARM::GPRRegClass);
8093 MIB.addReg(TmpReg, RegState::Define|RegState::Dead);
8097 void ARMTargetLowering::AdjustInstrPostInstrSelection(MachineInstr *MI,
8098 SDNode *Node) const {
8099 if (MI->getOpcode() == ARM::MEMCPY) {
8100 attachMEMCPYScratchRegs(Subtarget, MI, Node);
8104 const MCInstrDesc *MCID = &MI->getDesc();
8105 // Adjust potentially 's' setting instructions after isel, i.e. ADC, SBC, RSB,
8106 // RSC. Coming out of isel, they have an implicit CPSR def, but the optional
8107 // operand is still set to noreg. If needed, set the optional operand's
8108 // register to CPSR, and remove the redundant implicit def.
8110 // e.g. ADCS (..., CPSR<imp-def>) -> ADC (... opt:CPSR<def>).
8112 // Rename pseudo opcodes.
8113 unsigned NewOpc = convertAddSubFlagsOpcode(MI->getOpcode());
8115 const ARMBaseInstrInfo *TII = Subtarget->getInstrInfo();
8116 MCID = &TII->get(NewOpc);
8118 assert(MCID->getNumOperands() == MI->getDesc().getNumOperands() + 1 &&
8119 "converted opcode should be the same except for cc_out");
8123 // Add the optional cc_out operand
8124 MI->addOperand(MachineOperand::CreateReg(0, /*isDef=*/true));
8126 unsigned ccOutIdx = MCID->getNumOperands() - 1;
8128 // Any ARM instruction that sets the 's' bit should specify an optional
8129 // "cc_out" operand in the last operand position.
8130 if (!MI->hasOptionalDef() || !MCID->OpInfo[ccOutIdx].isOptionalDef()) {
8131 assert(!NewOpc && "Optional cc_out operand required");
8134 // Look for an implicit def of CPSR added by MachineInstr ctor. Remove it
8135 // since we already have an optional CPSR def.
8136 bool definesCPSR = false;
8137 bool deadCPSR = false;
8138 for (unsigned i = MCID->getNumOperands(), e = MI->getNumOperands();
8140 const MachineOperand &MO = MI->getOperand(i);
8141 if (MO.isReg() && MO.isDef() && MO.getReg() == ARM::CPSR) {
8145 MI->RemoveOperand(i);
8150 assert(!NewOpc && "Optional cc_out operand required");
8153 assert(deadCPSR == !Node->hasAnyUseOfValue(1) && "inconsistent dead flag");
8155 assert(!MI->getOperand(ccOutIdx).getReg() &&
8156 "expect uninitialized optional cc_out operand");
8160 // If this instruction was defined with an optional CPSR def and its dag node
8161 // had a live implicit CPSR def, then activate the optional CPSR def.
8162 MachineOperand &MO = MI->getOperand(ccOutIdx);
8163 MO.setReg(ARM::CPSR);
8167 //===----------------------------------------------------------------------===//
8168 // ARM Optimization Hooks
8169 //===----------------------------------------------------------------------===//
8171 // Helper function that checks if N is a null or all ones constant.
8172 static inline bool isZeroOrAllOnes(SDValue N, bool AllOnes) {
8173 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N);
8176 return AllOnes ? C->isAllOnesValue() : C->isNullValue();
8179 // Return true if N is conditionally 0 or all ones.
8180 // Detects these expressions where cc is an i1 value:
8182 // (select cc 0, y) [AllOnes=0]
8183 // (select cc y, 0) [AllOnes=0]
8184 // (zext cc) [AllOnes=0]
8185 // (sext cc) [AllOnes=0/1]
8186 // (select cc -1, y) [AllOnes=1]
8187 // (select cc y, -1) [AllOnes=1]
8189 // Invert is set when N is the null/all ones constant when CC is false.
8190 // OtherOp is set to the alternative value of N.
8191 static bool isConditionalZeroOrAllOnes(SDNode *N, bool AllOnes,
8192 SDValue &CC, bool &Invert,
8194 SelectionDAG &DAG) {
8195 switch (N->getOpcode()) {
8196 default: return false;
8198 CC = N->getOperand(0);
8199 SDValue N1 = N->getOperand(1);
8200 SDValue N2 = N->getOperand(2);
8201 if (isZeroOrAllOnes(N1, AllOnes)) {
8206 if (isZeroOrAllOnes(N2, AllOnes)) {
8213 case ISD::ZERO_EXTEND:
8214 // (zext cc) can never be the all ones value.
8218 case ISD::SIGN_EXTEND: {
8220 EVT VT = N->getValueType(0);
8221 CC = N->getOperand(0);
8222 if (CC.getValueType() != MVT::i1)
8226 // When looking for an AllOnes constant, N is an sext, and the 'other'
8228 OtherOp = DAG.getConstant(0, dl, VT);
8229 else if (N->getOpcode() == ISD::ZERO_EXTEND)
8230 // When looking for a 0 constant, N can be zext or sext.
8231 OtherOp = DAG.getConstant(1, dl, VT);
8233 OtherOp = DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), dl,
8240 // Combine a constant select operand into its use:
8242 // (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
8243 // (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
8244 // (and (select cc, -1, c), x) -> (select cc, x, (and, x, c)) [AllOnes=1]
8245 // (or (select cc, 0, c), x) -> (select cc, x, (or, x, c))
8246 // (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c))
8248 // The transform is rejected if the select doesn't have a constant operand that
8249 // is null, or all ones when AllOnes is set.
8251 // Also recognize sext/zext from i1:
8253 // (add (zext cc), x) -> (select cc (add x, 1), x)
8254 // (add (sext cc), x) -> (select cc (add x, -1), x)
8256 // These transformations eventually create predicated instructions.
8258 // @param N The node to transform.
8259 // @param Slct The N operand that is a select.
8260 // @param OtherOp The other N operand (x above).
8261 // @param DCI Context.
8262 // @param AllOnes Require the select constant to be all ones instead of null.
8263 // @returns The new node, or SDValue() on failure.
8265 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
8266 TargetLowering::DAGCombinerInfo &DCI,
8267 bool AllOnes = false) {
8268 SelectionDAG &DAG = DCI.DAG;
8269 EVT VT = N->getValueType(0);
8270 SDValue NonConstantVal;
8273 if (!isConditionalZeroOrAllOnes(Slct.getNode(), AllOnes, CCOp, SwapSelectOps,
8274 NonConstantVal, DAG))
8277 // Slct is now know to be the desired identity constant when CC is true.
8278 SDValue TrueVal = OtherOp;
8279 SDValue FalseVal = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
8280 OtherOp, NonConstantVal);
8281 // Unless SwapSelectOps says CC should be false.
8283 std::swap(TrueVal, FalseVal);
8285 return DAG.getNode(ISD::SELECT, SDLoc(N), VT,
8286 CCOp, TrueVal, FalseVal);
8289 // Attempt combineSelectAndUse on each operand of a commutative operator N.
8291 SDValue combineSelectAndUseCommutative(SDNode *N, bool AllOnes,
8292 TargetLowering::DAGCombinerInfo &DCI) {
8293 SDValue N0 = N->getOperand(0);
8294 SDValue N1 = N->getOperand(1);
8295 if (N0.getNode()->hasOneUse()) {
8296 SDValue Result = combineSelectAndUse(N, N0, N1, DCI, AllOnes);
8297 if (Result.getNode())
8300 if (N1.getNode()->hasOneUse()) {
8301 SDValue Result = combineSelectAndUse(N, N1, N0, DCI, AllOnes);
8302 if (Result.getNode())
8308 // AddCombineToVPADDL- For pair-wise add on neon, use the vpaddl instruction
8309 // (only after legalization).
8310 static SDValue AddCombineToVPADDL(SDNode *N, SDValue N0, SDValue N1,
8311 TargetLowering::DAGCombinerInfo &DCI,
8312 const ARMSubtarget *Subtarget) {
8314 // Only perform optimization if after legalize, and if NEON is available. We
8315 // also expected both operands to be BUILD_VECTORs.
8316 if (DCI.isBeforeLegalize() || !Subtarget->hasNEON()
8317 || N0.getOpcode() != ISD::BUILD_VECTOR
8318 || N1.getOpcode() != ISD::BUILD_VECTOR)
8321 // Check output type since VPADDL operand elements can only be 8, 16, or 32.
8322 EVT VT = N->getValueType(0);
8323 if (!VT.isInteger() || VT.getVectorElementType() == MVT::i64)
8326 // Check that the vector operands are of the right form.
8327 // N0 and N1 are BUILD_VECTOR nodes with N number of EXTRACT_VECTOR
8328 // operands, where N is the size of the formed vector.
8329 // Each EXTRACT_VECTOR should have the same input vector and odd or even
8330 // index such that we have a pair wise add pattern.
8332 // Grab the vector that all EXTRACT_VECTOR nodes should be referencing.
8333 if (N0->getOperand(0)->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
8335 SDValue Vec = N0->getOperand(0)->getOperand(0);
8336 SDNode *V = Vec.getNode();
8337 unsigned nextIndex = 0;
8339 // For each operands to the ADD which are BUILD_VECTORs,
8340 // check to see if each of their operands are an EXTRACT_VECTOR with
8341 // the same vector and appropriate index.
8342 for (unsigned i = 0, e = N0->getNumOperands(); i != e; ++i) {
8343 if (N0->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT
8344 && N1->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
8346 SDValue ExtVec0 = N0->getOperand(i);
8347 SDValue ExtVec1 = N1->getOperand(i);
8349 // First operand is the vector, verify its the same.
8350 if (V != ExtVec0->getOperand(0).getNode() ||
8351 V != ExtVec1->getOperand(0).getNode())
8354 // Second is the constant, verify its correct.
8355 ConstantSDNode *C0 = dyn_cast<ConstantSDNode>(ExtVec0->getOperand(1));
8356 ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(ExtVec1->getOperand(1));
8358 // For the constant, we want to see all the even or all the odd.
8359 if (!C0 || !C1 || C0->getZExtValue() != nextIndex
8360 || C1->getZExtValue() != nextIndex+1)
8369 // Create VPADDL node.
8370 SelectionDAG &DAG = DCI.DAG;
8371 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8375 // Build operand list.
8376 SmallVector<SDValue, 8> Ops;
8377 Ops.push_back(DAG.getConstant(Intrinsic::arm_neon_vpaddls, dl,
8378 TLI.getPointerTy(DAG.getDataLayout())));
8380 // Input is the vector.
8383 // Get widened type and narrowed type.
8385 unsigned numElem = VT.getVectorNumElements();
8387 EVT inputLaneType = Vec.getValueType().getVectorElementType();
8388 switch (inputLaneType.getSimpleVT().SimpleTy) {
8389 case MVT::i8: widenType = MVT::getVectorVT(MVT::i16, numElem); break;
8390 case MVT::i16: widenType = MVT::getVectorVT(MVT::i32, numElem); break;
8391 case MVT::i32: widenType = MVT::getVectorVT(MVT::i64, numElem); break;
8393 llvm_unreachable("Invalid vector element type for padd optimization.");
8396 SDValue tmp = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, widenType, Ops);
8397 unsigned ExtOp = VT.bitsGT(tmp.getValueType()) ? ISD::ANY_EXTEND : ISD::TRUNCATE;
8398 return DAG.getNode(ExtOp, dl, VT, tmp);
8401 static SDValue findMUL_LOHI(SDValue V) {
8402 if (V->getOpcode() == ISD::UMUL_LOHI ||
8403 V->getOpcode() == ISD::SMUL_LOHI)
8408 static SDValue AddCombineTo64bitMLAL(SDNode *AddcNode,
8409 TargetLowering::DAGCombinerInfo &DCI,
8410 const ARMSubtarget *Subtarget) {
8412 if (Subtarget->isThumb1Only()) return SDValue();
8414 // Only perform the checks after legalize when the pattern is available.
8415 if (DCI.isBeforeLegalize()) return SDValue();
8417 // Look for multiply add opportunities.
8418 // The pattern is a ISD::UMUL_LOHI followed by two add nodes, where
8419 // each add nodes consumes a value from ISD::UMUL_LOHI and there is
8420 // a glue link from the first add to the second add.
8421 // If we find this pattern, we can replace the U/SMUL_LOHI, ADDC, and ADDE by
8422 // a S/UMLAL instruction.
8425 // / \ [no multiline comment]
8431 assert(AddcNode->getOpcode() == ISD::ADDC && "Expect an ADDC");
8432 SDValue AddcOp0 = AddcNode->getOperand(0);
8433 SDValue AddcOp1 = AddcNode->getOperand(1);
8435 // Check if the two operands are from the same mul_lohi node.
8436 if (AddcOp0.getNode() == AddcOp1.getNode())
8439 assert(AddcNode->getNumValues() == 2 &&
8440 AddcNode->getValueType(0) == MVT::i32 &&
8441 "Expect ADDC with two result values. First: i32");
8443 // Check that we have a glued ADDC node.
8444 if (AddcNode->getValueType(1) != MVT::Glue)
8447 // Check that the ADDC adds the low result of the S/UMUL_LOHI.
8448 if (AddcOp0->getOpcode() != ISD::UMUL_LOHI &&
8449 AddcOp0->getOpcode() != ISD::SMUL_LOHI &&
8450 AddcOp1->getOpcode() != ISD::UMUL_LOHI &&
8451 AddcOp1->getOpcode() != ISD::SMUL_LOHI)
8454 // Look for the glued ADDE.
8455 SDNode* AddeNode = AddcNode->getGluedUser();
8459 // Make sure it is really an ADDE.
8460 if (AddeNode->getOpcode() != ISD::ADDE)
8463 assert(AddeNode->getNumOperands() == 3 &&
8464 AddeNode->getOperand(2).getValueType() == MVT::Glue &&
8465 "ADDE node has the wrong inputs");
8467 // Check for the triangle shape.
8468 SDValue AddeOp0 = AddeNode->getOperand(0);
8469 SDValue AddeOp1 = AddeNode->getOperand(1);
8471 // Make sure that the ADDE operands are not coming from the same node.
8472 if (AddeOp0.getNode() == AddeOp1.getNode())
8475 // Find the MUL_LOHI node walking up ADDE's operands.
8476 bool IsLeftOperandMUL = false;
8477 SDValue MULOp = findMUL_LOHI(AddeOp0);
8478 if (MULOp == SDValue())
8479 MULOp = findMUL_LOHI(AddeOp1);
8481 IsLeftOperandMUL = true;
8482 if (MULOp == SDValue())
8485 // Figure out the right opcode.
8486 unsigned Opc = MULOp->getOpcode();
8487 unsigned FinalOpc = (Opc == ISD::SMUL_LOHI) ? ARMISD::SMLAL : ARMISD::UMLAL;
8489 // Figure out the high and low input values to the MLAL node.
8490 SDValue* HiAdd = nullptr;
8491 SDValue* LoMul = nullptr;
8492 SDValue* LowAdd = nullptr;
8494 // Ensure that ADDE is from high result of ISD::SMUL_LOHI.
8495 if ((AddeOp0 != MULOp.getValue(1)) && (AddeOp1 != MULOp.getValue(1)))
8498 if (IsLeftOperandMUL)
8504 // Ensure that LoMul and LowAdd are taken from correct ISD::SMUL_LOHI node
8505 // whose low result is fed to the ADDC we are checking.
8507 if (AddcOp0 == MULOp.getValue(0)) {
8511 if (AddcOp1 == MULOp.getValue(0)) {
8519 // Create the merged node.
8520 SelectionDAG &DAG = DCI.DAG;
8522 // Build operand list.
8523 SmallVector<SDValue, 8> Ops;
8524 Ops.push_back(LoMul->getOperand(0));
8525 Ops.push_back(LoMul->getOperand(1));
8526 Ops.push_back(*LowAdd);
8527 Ops.push_back(*HiAdd);
8529 SDValue MLALNode = DAG.getNode(FinalOpc, SDLoc(AddcNode),
8530 DAG.getVTList(MVT::i32, MVT::i32), Ops);
8532 // Replace the ADDs' nodes uses by the MLA node's values.
8533 SDValue HiMLALResult(MLALNode.getNode(), 1);
8534 DAG.ReplaceAllUsesOfValueWith(SDValue(AddeNode, 0), HiMLALResult);
8536 SDValue LoMLALResult(MLALNode.getNode(), 0);
8537 DAG.ReplaceAllUsesOfValueWith(SDValue(AddcNode, 0), LoMLALResult);
8539 // Return original node to notify the driver to stop replacing.
8540 SDValue resNode(AddcNode, 0);
8544 /// PerformADDCCombine - Target-specific dag combine transform from
8545 /// ISD::ADDC, ISD::ADDE, and ISD::MUL_LOHI to MLAL.
8546 static SDValue PerformADDCCombine(SDNode *N,
8547 TargetLowering::DAGCombinerInfo &DCI,
8548 const ARMSubtarget *Subtarget) {
8550 return AddCombineTo64bitMLAL(N, DCI, Subtarget);
8554 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
8555 /// operands N0 and N1. This is a helper for PerformADDCombine that is
8556 /// called with the default operands, and if that fails, with commuted
8558 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
8559 TargetLowering::DAGCombinerInfo &DCI,
8560 const ARMSubtarget *Subtarget){
8562 // Attempt to create vpaddl for this add.
8563 SDValue Result = AddCombineToVPADDL(N, N0, N1, DCI, Subtarget);
8564 if (Result.getNode())
8567 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
8568 if (N0.getNode()->hasOneUse()) {
8569 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
8570 if (Result.getNode()) return Result;
8575 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
8577 static SDValue PerformADDCombine(SDNode *N,
8578 TargetLowering::DAGCombinerInfo &DCI,
8579 const ARMSubtarget *Subtarget) {
8580 SDValue N0 = N->getOperand(0);
8581 SDValue N1 = N->getOperand(1);
8583 // First try with the default operand order.
8584 SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI, Subtarget);
8585 if (Result.getNode())
8588 // If that didn't work, try again with the operands commuted.
8589 return PerformADDCombineWithOperands(N, N1, N0, DCI, Subtarget);
8592 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
8594 static SDValue PerformSUBCombine(SDNode *N,
8595 TargetLowering::DAGCombinerInfo &DCI) {
8596 SDValue N0 = N->getOperand(0);
8597 SDValue N1 = N->getOperand(1);
8599 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
8600 if (N1.getNode()->hasOneUse()) {
8601 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
8602 if (Result.getNode()) return Result;
8608 /// PerformVMULCombine
8609 /// Distribute (A + B) * C to (A * C) + (B * C) to take advantage of the
8610 /// special multiplier accumulator forwarding.
8616 // However, for (A + B) * (A + B),
8623 static SDValue PerformVMULCombine(SDNode *N,
8624 TargetLowering::DAGCombinerInfo &DCI,
8625 const ARMSubtarget *Subtarget) {
8626 if (!Subtarget->hasVMLxForwarding())
8629 SelectionDAG &DAG = DCI.DAG;
8630 SDValue N0 = N->getOperand(0);
8631 SDValue N1 = N->getOperand(1);
8632 unsigned Opcode = N0.getOpcode();
8633 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
8634 Opcode != ISD::FADD && Opcode != ISD::FSUB) {
8635 Opcode = N1.getOpcode();
8636 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
8637 Opcode != ISD::FADD && Opcode != ISD::FSUB)
8645 EVT VT = N->getValueType(0);
8647 SDValue N00 = N0->getOperand(0);
8648 SDValue N01 = N0->getOperand(1);
8649 return DAG.getNode(Opcode, DL, VT,
8650 DAG.getNode(ISD::MUL, DL, VT, N00, N1),
8651 DAG.getNode(ISD::MUL, DL, VT, N01, N1));
8654 static SDValue PerformMULCombine(SDNode *N,
8655 TargetLowering::DAGCombinerInfo &DCI,
8656 const ARMSubtarget *Subtarget) {
8657 SelectionDAG &DAG = DCI.DAG;
8659 if (Subtarget->isThumb1Only())
8662 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
8665 EVT VT = N->getValueType(0);
8666 if (VT.is64BitVector() || VT.is128BitVector())
8667 return PerformVMULCombine(N, DCI, Subtarget);
8671 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
8675 int64_t MulAmt = C->getSExtValue();
8676 unsigned ShiftAmt = countTrailingZeros<uint64_t>(MulAmt);
8678 ShiftAmt = ShiftAmt & (32 - 1);
8679 SDValue V = N->getOperand(0);
8683 MulAmt >>= ShiftAmt;
8686 if (isPowerOf2_32(MulAmt - 1)) {
8687 // (mul x, 2^N + 1) => (add (shl x, N), x)
8688 Res = DAG.getNode(ISD::ADD, DL, VT,
8690 DAG.getNode(ISD::SHL, DL, VT,
8692 DAG.getConstant(Log2_32(MulAmt - 1), DL,
8694 } else if (isPowerOf2_32(MulAmt + 1)) {
8695 // (mul x, 2^N - 1) => (sub (shl x, N), x)
8696 Res = DAG.getNode(ISD::SUB, DL, VT,
8697 DAG.getNode(ISD::SHL, DL, VT,
8699 DAG.getConstant(Log2_32(MulAmt + 1), DL,
8705 uint64_t MulAmtAbs = -MulAmt;
8706 if (isPowerOf2_32(MulAmtAbs + 1)) {
8707 // (mul x, -(2^N - 1)) => (sub x, (shl x, N))
8708 Res = DAG.getNode(ISD::SUB, DL, VT,
8710 DAG.getNode(ISD::SHL, DL, VT,
8712 DAG.getConstant(Log2_32(MulAmtAbs + 1), DL,
8714 } else if (isPowerOf2_32(MulAmtAbs - 1)) {
8715 // (mul x, -(2^N + 1)) => - (add (shl x, N), x)
8716 Res = DAG.getNode(ISD::ADD, DL, VT,
8718 DAG.getNode(ISD::SHL, DL, VT,
8720 DAG.getConstant(Log2_32(MulAmtAbs - 1), DL,
8722 Res = DAG.getNode(ISD::SUB, DL, VT,
8723 DAG.getConstant(0, DL, MVT::i32), Res);
8730 Res = DAG.getNode(ISD::SHL, DL, VT,
8731 Res, DAG.getConstant(ShiftAmt, DL, MVT::i32));
8733 // Do not add new nodes to DAG combiner worklist.
8734 DCI.CombineTo(N, Res, false);
8738 static SDValue PerformANDCombine(SDNode *N,
8739 TargetLowering::DAGCombinerInfo &DCI,
8740 const ARMSubtarget *Subtarget) {
8742 // Attempt to use immediate-form VBIC
8743 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
8745 EVT VT = N->getValueType(0);
8746 SelectionDAG &DAG = DCI.DAG;
8748 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8751 APInt SplatBits, SplatUndef;
8752 unsigned SplatBitSize;
8755 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
8756 if (SplatBitSize <= 64) {
8758 SDValue Val = isNEONModifiedImm((~SplatBits).getZExtValue(),
8759 SplatUndef.getZExtValue(), SplatBitSize,
8760 DAG, dl, VbicVT, VT.is128BitVector(),
8762 if (Val.getNode()) {
8764 DAG.getNode(ISD::BITCAST, dl, VbicVT, N->getOperand(0));
8765 SDValue Vbic = DAG.getNode(ARMISD::VBICIMM, dl, VbicVT, Input, Val);
8766 return DAG.getNode(ISD::BITCAST, dl, VT, Vbic);
8771 if (!Subtarget->isThumb1Only()) {
8772 // fold (and (select cc, -1, c), x) -> (select cc, x, (and, x, c))
8773 SDValue Result = combineSelectAndUseCommutative(N, true, DCI);
8774 if (Result.getNode())
8781 /// PerformORCombine - Target-specific dag combine xforms for ISD::OR
8782 static SDValue PerformORCombine(SDNode *N,
8783 TargetLowering::DAGCombinerInfo &DCI,
8784 const ARMSubtarget *Subtarget) {
8785 // Attempt to use immediate-form VORR
8786 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
8788 EVT VT = N->getValueType(0);
8789 SelectionDAG &DAG = DCI.DAG;
8791 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8794 APInt SplatBits, SplatUndef;
8795 unsigned SplatBitSize;
8797 if (BVN && Subtarget->hasNEON() &&
8798 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
8799 if (SplatBitSize <= 64) {
8801 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
8802 SplatUndef.getZExtValue(), SplatBitSize,
8803 DAG, dl, VorrVT, VT.is128BitVector(),
8805 if (Val.getNode()) {
8807 DAG.getNode(ISD::BITCAST, dl, VorrVT, N->getOperand(0));
8808 SDValue Vorr = DAG.getNode(ARMISD::VORRIMM, dl, VorrVT, Input, Val);
8809 return DAG.getNode(ISD::BITCAST, dl, VT, Vorr);
8814 if (!Subtarget->isThumb1Only()) {
8815 // fold (or (select cc, 0, c), x) -> (select cc, x, (or, x, c))
8816 SDValue Result = combineSelectAndUseCommutative(N, false, DCI);
8817 if (Result.getNode())
8821 // The code below optimizes (or (and X, Y), Z).
8822 // The AND operand needs to have a single user to make these optimizations
8824 SDValue N0 = N->getOperand(0);
8825 if (N0.getOpcode() != ISD::AND || !N0.hasOneUse())
8827 SDValue N1 = N->getOperand(1);
8829 // (or (and B, A), (and C, ~A)) => (VBSL A, B, C) when A is a constant.
8830 if (Subtarget->hasNEON() && N1.getOpcode() == ISD::AND && VT.isVector() &&
8831 DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
8833 unsigned SplatBitSize;
8836 APInt SplatBits0, SplatBits1;
8837 BuildVectorSDNode *BVN0 = dyn_cast<BuildVectorSDNode>(N0->getOperand(1));
8838 BuildVectorSDNode *BVN1 = dyn_cast<BuildVectorSDNode>(N1->getOperand(1));
8839 // Ensure that the second operand of both ands are constants
8840 if (BVN0 && BVN0->isConstantSplat(SplatBits0, SplatUndef, SplatBitSize,
8841 HasAnyUndefs) && !HasAnyUndefs) {
8842 if (BVN1 && BVN1->isConstantSplat(SplatBits1, SplatUndef, SplatBitSize,
8843 HasAnyUndefs) && !HasAnyUndefs) {
8844 // Ensure that the bit width of the constants are the same and that
8845 // the splat arguments are logical inverses as per the pattern we
8846 // are trying to simplify.
8847 if (SplatBits0.getBitWidth() == SplatBits1.getBitWidth() &&
8848 SplatBits0 == ~SplatBits1) {
8849 // Canonicalize the vector type to make instruction selection
8851 EVT CanonicalVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
8852 SDValue Result = DAG.getNode(ARMISD::VBSL, dl, CanonicalVT,
8856 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
8862 // Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when
8865 // BFI is only available on V6T2+
8866 if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops())
8870 // 1) or (and A, mask), val => ARMbfi A, val, mask
8871 // iff (val & mask) == val
8873 // 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
8874 // 2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2)
8875 // && mask == ~mask2
8876 // 2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2)
8877 // && ~mask == mask2
8878 // (i.e., copy a bitfield value into another bitfield of the same width)
8883 SDValue N00 = N0.getOperand(0);
8885 // The value and the mask need to be constants so we can verify this is
8886 // actually a bitfield set. If the mask is 0xffff, we can do better
8887 // via a movt instruction, so don't use BFI in that case.
8888 SDValue MaskOp = N0.getOperand(1);
8889 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(MaskOp);
8892 unsigned Mask = MaskC->getZExtValue();
8896 // Case (1): or (and A, mask), val => ARMbfi A, val, mask
8897 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
8899 unsigned Val = N1C->getZExtValue();
8900 if ((Val & ~Mask) != Val)
8903 if (ARM::isBitFieldInvertedMask(Mask)) {
8904 Val >>= countTrailingZeros(~Mask);
8906 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00,
8907 DAG.getConstant(Val, DL, MVT::i32),
8908 DAG.getConstant(Mask, DL, MVT::i32));
8910 // Do not add new nodes to DAG combiner worklist.
8911 DCI.CombineTo(N, Res, false);
8914 } else if (N1.getOpcode() == ISD::AND) {
8915 // case (2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
8916 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
8919 unsigned Mask2 = N11C->getZExtValue();
8921 // Mask and ~Mask2 (or reverse) must be equivalent for the BFI pattern
8923 if (ARM::isBitFieldInvertedMask(Mask) &&
8925 // The pack halfword instruction works better for masks that fit it,
8926 // so use that when it's available.
8927 if (Subtarget->hasT2ExtractPack() &&
8928 (Mask == 0xffff || Mask == 0xffff0000))
8931 unsigned amt = countTrailingZeros(Mask2);
8932 Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0),
8933 DAG.getConstant(amt, DL, MVT::i32));
8934 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00, Res,
8935 DAG.getConstant(Mask, DL, MVT::i32));
8936 // Do not add new nodes to DAG combiner worklist.
8937 DCI.CombineTo(N, Res, false);
8939 } else if (ARM::isBitFieldInvertedMask(~Mask) &&
8941 // The pack halfword instruction works better for masks that fit it,
8942 // so use that when it's available.
8943 if (Subtarget->hasT2ExtractPack() &&
8944 (Mask2 == 0xffff || Mask2 == 0xffff0000))
8947 unsigned lsb = countTrailingZeros(Mask);
8948 Res = DAG.getNode(ISD::SRL, DL, VT, N00,
8949 DAG.getConstant(lsb, DL, MVT::i32));
8950 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res,
8951 DAG.getConstant(Mask2, DL, MVT::i32));
8952 // Do not add new nodes to DAG combiner worklist.
8953 DCI.CombineTo(N, Res, false);
8958 if (DAG.MaskedValueIsZero(N1, MaskC->getAPIntValue()) &&
8959 N00.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N00.getOperand(1)) &&
8960 ARM::isBitFieldInvertedMask(~Mask)) {
8961 // Case (3): or (and (shl A, #shamt), mask), B => ARMbfi B, A, ~mask
8962 // where lsb(mask) == #shamt and masked bits of B are known zero.
8963 SDValue ShAmt = N00.getOperand(1);
8964 unsigned ShAmtC = cast<ConstantSDNode>(ShAmt)->getZExtValue();
8965 unsigned LSB = countTrailingZeros(Mask);
8969 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1, N00.getOperand(0),
8970 DAG.getConstant(~Mask, DL, MVT::i32));
8972 // Do not add new nodes to DAG combiner worklist.
8973 DCI.CombineTo(N, Res, false);
8979 static SDValue PerformXORCombine(SDNode *N,
8980 TargetLowering::DAGCombinerInfo &DCI,
8981 const ARMSubtarget *Subtarget) {
8982 EVT VT = N->getValueType(0);
8983 SelectionDAG &DAG = DCI.DAG;
8985 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8988 if (!Subtarget->isThumb1Only()) {
8989 // fold (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c))
8990 SDValue Result = combineSelectAndUseCommutative(N, false, DCI);
8991 if (Result.getNode())
8998 /// PerformBFICombine - (bfi A, (and B, Mask1), Mask2) -> (bfi A, B, Mask2) iff
8999 /// the bits being cleared by the AND are not demanded by the BFI.
9000 static SDValue PerformBFICombine(SDNode *N,
9001 TargetLowering::DAGCombinerInfo &DCI) {
9002 SDValue N1 = N->getOperand(1);
9003 if (N1.getOpcode() == ISD::AND) {
9004 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
9007 unsigned InvMask = cast<ConstantSDNode>(N->getOperand(2))->getZExtValue();
9008 unsigned LSB = countTrailingZeros(~InvMask);
9009 unsigned Width = (32 - countLeadingZeros(~InvMask)) - LSB;
9011 static_cast<unsigned>(std::numeric_limits<unsigned>::digits) &&
9012 "undefined behavior");
9013 unsigned Mask = (1u << Width) - 1;
9014 unsigned Mask2 = N11C->getZExtValue();
9015 if ((Mask & (~Mask2)) == 0)
9016 return DCI.DAG.getNode(ARMISD::BFI, SDLoc(N), N->getValueType(0),
9017 N->getOperand(0), N1.getOperand(0),
9023 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
9024 /// ARMISD::VMOVRRD.
9025 static SDValue PerformVMOVRRDCombine(SDNode *N,
9026 TargetLowering::DAGCombinerInfo &DCI,
9027 const ARMSubtarget *Subtarget) {
9028 // vmovrrd(vmovdrr x, y) -> x,y
9029 SDValue InDouble = N->getOperand(0);
9030 if (InDouble.getOpcode() == ARMISD::VMOVDRR && !Subtarget->isFPOnlySP())
9031 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
9033 // vmovrrd(load f64) -> (load i32), (load i32)
9034 SDNode *InNode = InDouble.getNode();
9035 if (ISD::isNormalLoad(InNode) && InNode->hasOneUse() &&
9036 InNode->getValueType(0) == MVT::f64 &&
9037 InNode->getOperand(1).getOpcode() == ISD::FrameIndex &&
9038 !cast<LoadSDNode>(InNode)->isVolatile()) {
9039 // TODO: Should this be done for non-FrameIndex operands?
9040 LoadSDNode *LD = cast<LoadSDNode>(InNode);
9042 SelectionDAG &DAG = DCI.DAG;
9044 SDValue BasePtr = LD->getBasePtr();
9045 SDValue NewLD1 = DAG.getLoad(MVT::i32, DL, LD->getChain(), BasePtr,
9046 LD->getPointerInfo(), LD->isVolatile(),
9047 LD->isNonTemporal(), LD->isInvariant(),
9048 LD->getAlignment());
9050 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
9051 DAG.getConstant(4, DL, MVT::i32));
9052 SDValue NewLD2 = DAG.getLoad(MVT::i32, DL, NewLD1.getValue(1), OffsetPtr,
9053 LD->getPointerInfo(), LD->isVolatile(),
9054 LD->isNonTemporal(), LD->isInvariant(),
9055 std::min(4U, LD->getAlignment() / 2));
9057 DAG.ReplaceAllUsesOfValueWith(SDValue(LD, 1), NewLD2.getValue(1));
9058 if (DCI.DAG.getDataLayout().isBigEndian())
9059 std::swap (NewLD1, NewLD2);
9060 SDValue Result = DCI.CombineTo(N, NewLD1, NewLD2);
9067 /// PerformVMOVDRRCombine - Target-specific dag combine xforms for
9068 /// ARMISD::VMOVDRR. This is also used for BUILD_VECTORs with 2 operands.
9069 static SDValue PerformVMOVDRRCombine(SDNode *N, SelectionDAG &DAG) {
9070 // N=vmovrrd(X); vmovdrr(N:0, N:1) -> bit_convert(X)
9071 SDValue Op0 = N->getOperand(0);
9072 SDValue Op1 = N->getOperand(1);
9073 if (Op0.getOpcode() == ISD::BITCAST)
9074 Op0 = Op0.getOperand(0);
9075 if (Op1.getOpcode() == ISD::BITCAST)
9076 Op1 = Op1.getOperand(0);
9077 if (Op0.getOpcode() == ARMISD::VMOVRRD &&
9078 Op0.getNode() == Op1.getNode() &&
9079 Op0.getResNo() == 0 && Op1.getResNo() == 1)
9080 return DAG.getNode(ISD::BITCAST, SDLoc(N),
9081 N->getValueType(0), Op0.getOperand(0));
9085 /// hasNormalLoadOperand - Check if any of the operands of a BUILD_VECTOR node
9086 /// are normal, non-volatile loads. If so, it is profitable to bitcast an
9087 /// i64 vector to have f64 elements, since the value can then be loaded
9088 /// directly into a VFP register.
9089 static bool hasNormalLoadOperand(SDNode *N) {
9090 unsigned NumElts = N->getValueType(0).getVectorNumElements();
9091 for (unsigned i = 0; i < NumElts; ++i) {
9092 SDNode *Elt = N->getOperand(i).getNode();
9093 if (ISD::isNormalLoad(Elt) && !cast<LoadSDNode>(Elt)->isVolatile())
9099 /// PerformBUILD_VECTORCombine - Target-specific dag combine xforms for
9100 /// ISD::BUILD_VECTOR.
9101 static SDValue PerformBUILD_VECTORCombine(SDNode *N,
9102 TargetLowering::DAGCombinerInfo &DCI,
9103 const ARMSubtarget *Subtarget) {
9104 // build_vector(N=ARMISD::VMOVRRD(X), N:1) -> bit_convert(X):
9105 // VMOVRRD is introduced when legalizing i64 types. It forces the i64 value
9106 // into a pair of GPRs, which is fine when the value is used as a scalar,
9107 // but if the i64 value is converted to a vector, we need to undo the VMOVRRD.
9108 SelectionDAG &DAG = DCI.DAG;
9109 if (N->getNumOperands() == 2) {
9110 SDValue RV = PerformVMOVDRRCombine(N, DAG);
9115 // Load i64 elements as f64 values so that type legalization does not split
9116 // them up into i32 values.
9117 EVT VT = N->getValueType(0);
9118 if (VT.getVectorElementType() != MVT::i64 || !hasNormalLoadOperand(N))
9121 SmallVector<SDValue, 8> Ops;
9122 unsigned NumElts = VT.getVectorNumElements();
9123 for (unsigned i = 0; i < NumElts; ++i) {
9124 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(i));
9126 // Make the DAGCombiner fold the bitcast.
9127 DCI.AddToWorklist(V.getNode());
9129 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, NumElts);
9130 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, FloatVT, Ops);
9131 return DAG.getNode(ISD::BITCAST, dl, VT, BV);
9134 /// \brief Target-specific dag combine xforms for ARMISD::BUILD_VECTOR.
9136 PerformARMBUILD_VECTORCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
9137 // ARMISD::BUILD_VECTOR is introduced when legalizing ISD::BUILD_VECTOR.
9138 // At that time, we may have inserted bitcasts from integer to float.
9139 // If these bitcasts have survived DAGCombine, change the lowering of this
9140 // BUILD_VECTOR in something more vector friendly, i.e., that does not
9141 // force to use floating point types.
9143 // Make sure we can change the type of the vector.
9144 // This is possible iff:
9145 // 1. The vector is only used in a bitcast to a integer type. I.e.,
9146 // 1.1. Vector is used only once.
9147 // 1.2. Use is a bit convert to an integer type.
9148 // 2. The size of its operands are 32-bits (64-bits are not legal).
9149 EVT VT = N->getValueType(0);
9150 EVT EltVT = VT.getVectorElementType();
9152 // Check 1.1. and 2.
9153 if (EltVT.getSizeInBits() != 32 || !N->hasOneUse())
9156 // By construction, the input type must be float.
9157 assert(EltVT == MVT::f32 && "Unexpected type!");
9160 SDNode *Use = *N->use_begin();
9161 if (Use->getOpcode() != ISD::BITCAST ||
9162 Use->getValueType(0).isFloatingPoint())
9165 // Check profitability.
9166 // Model is, if more than half of the relevant operands are bitcast from
9167 // i32, turn the build_vector into a sequence of insert_vector_elt.
9168 // Relevant operands are everything that is not statically
9169 // (i.e., at compile time) bitcasted.
9170 unsigned NumOfBitCastedElts = 0;
9171 unsigned NumElts = VT.getVectorNumElements();
9172 unsigned NumOfRelevantElts = NumElts;
9173 for (unsigned Idx = 0; Idx < NumElts; ++Idx) {
9174 SDValue Elt = N->getOperand(Idx);
9175 if (Elt->getOpcode() == ISD::BITCAST) {
9176 // Assume only bit cast to i32 will go away.
9177 if (Elt->getOperand(0).getValueType() == MVT::i32)
9178 ++NumOfBitCastedElts;
9179 } else if (Elt.getOpcode() == ISD::UNDEF || isa<ConstantSDNode>(Elt))
9180 // Constants are statically casted, thus do not count them as
9181 // relevant operands.
9182 --NumOfRelevantElts;
9185 // Check if more than half of the elements require a non-free bitcast.
9186 if (NumOfBitCastedElts <= NumOfRelevantElts / 2)
9189 SelectionDAG &DAG = DCI.DAG;
9190 // Create the new vector type.
9191 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
9192 // Check if the type is legal.
9193 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9194 if (!TLI.isTypeLegal(VecVT))
9198 // ARMISD::BUILD_VECTOR E1, E2, ..., EN.
9199 // => BITCAST INSERT_VECTOR_ELT
9200 // (INSERT_VECTOR_ELT (...), (BITCAST EN-1), N-1),
9202 SDValue Vec = DAG.getUNDEF(VecVT);
9204 for (unsigned Idx = 0 ; Idx < NumElts; ++Idx) {
9205 SDValue V = N->getOperand(Idx);
9206 if (V.getOpcode() == ISD::UNDEF)
9208 if (V.getOpcode() == ISD::BITCAST &&
9209 V->getOperand(0).getValueType() == MVT::i32)
9210 // Fold obvious case.
9211 V = V.getOperand(0);
9213 V = DAG.getNode(ISD::BITCAST, SDLoc(V), MVT::i32, V);
9214 // Make the DAGCombiner fold the bitcasts.
9215 DCI.AddToWorklist(V.getNode());
9217 SDValue LaneIdx = DAG.getConstant(Idx, dl, MVT::i32);
9218 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VecVT, Vec, V, LaneIdx);
9220 Vec = DAG.getNode(ISD::BITCAST, dl, VT, Vec);
9221 // Make the DAGCombiner fold the bitcasts.
9222 DCI.AddToWorklist(Vec.getNode());
9226 /// PerformInsertEltCombine - Target-specific dag combine xforms for
9227 /// ISD::INSERT_VECTOR_ELT.
9228 static SDValue PerformInsertEltCombine(SDNode *N,
9229 TargetLowering::DAGCombinerInfo &DCI) {
9230 // Bitcast an i64 load inserted into a vector to f64.
9231 // Otherwise, the i64 value will be legalized to a pair of i32 values.
9232 EVT VT = N->getValueType(0);
9233 SDNode *Elt = N->getOperand(1).getNode();
9234 if (VT.getVectorElementType() != MVT::i64 ||
9235 !ISD::isNormalLoad(Elt) || cast<LoadSDNode>(Elt)->isVolatile())
9238 SelectionDAG &DAG = DCI.DAG;
9240 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
9241 VT.getVectorNumElements());
9242 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, N->getOperand(0));
9243 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(1));
9244 // Make the DAGCombiner fold the bitcasts.
9245 DCI.AddToWorklist(Vec.getNode());
9246 DCI.AddToWorklist(V.getNode());
9247 SDValue InsElt = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, FloatVT,
9248 Vec, V, N->getOperand(2));
9249 return DAG.getNode(ISD::BITCAST, dl, VT, InsElt);
9252 /// PerformVECTOR_SHUFFLECombine - Target-specific dag combine xforms for
9253 /// ISD::VECTOR_SHUFFLE.
9254 static SDValue PerformVECTOR_SHUFFLECombine(SDNode *N, SelectionDAG &DAG) {
9255 // The LLVM shufflevector instruction does not require the shuffle mask
9256 // length to match the operand vector length, but ISD::VECTOR_SHUFFLE does
9257 // have that requirement. When translating to ISD::VECTOR_SHUFFLE, if the
9258 // operands do not match the mask length, they are extended by concatenating
9259 // them with undef vectors. That is probably the right thing for other
9260 // targets, but for NEON it is better to concatenate two double-register
9261 // size vector operands into a single quad-register size vector. Do that
9262 // transformation here:
9263 // shuffle(concat(v1, undef), concat(v2, undef)) ->
9264 // shuffle(concat(v1, v2), undef)
9265 SDValue Op0 = N->getOperand(0);
9266 SDValue Op1 = N->getOperand(1);
9267 if (Op0.getOpcode() != ISD::CONCAT_VECTORS ||
9268 Op1.getOpcode() != ISD::CONCAT_VECTORS ||
9269 Op0.getNumOperands() != 2 ||
9270 Op1.getNumOperands() != 2)
9272 SDValue Concat0Op1 = Op0.getOperand(1);
9273 SDValue Concat1Op1 = Op1.getOperand(1);
9274 if (Concat0Op1.getOpcode() != ISD::UNDEF ||
9275 Concat1Op1.getOpcode() != ISD::UNDEF)
9277 // Skip the transformation if any of the types are illegal.
9278 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9279 EVT VT = N->getValueType(0);
9280 if (!TLI.isTypeLegal(VT) ||
9281 !TLI.isTypeLegal(Concat0Op1.getValueType()) ||
9282 !TLI.isTypeLegal(Concat1Op1.getValueType()))
9285 SDValue NewConcat = DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT,
9286 Op0.getOperand(0), Op1.getOperand(0));
9287 // Translate the shuffle mask.
9288 SmallVector<int, 16> NewMask;
9289 unsigned NumElts = VT.getVectorNumElements();
9290 unsigned HalfElts = NumElts/2;
9291 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
9292 for (unsigned n = 0; n < NumElts; ++n) {
9293 int MaskElt = SVN->getMaskElt(n);
9295 if (MaskElt < (int)HalfElts)
9297 else if (MaskElt >= (int)NumElts && MaskElt < (int)(NumElts + HalfElts))
9298 NewElt = HalfElts + MaskElt - NumElts;
9299 NewMask.push_back(NewElt);
9301 return DAG.getVectorShuffle(VT, SDLoc(N), NewConcat,
9302 DAG.getUNDEF(VT), NewMask.data());
9305 /// CombineBaseUpdate - Target-specific DAG combine function for VLDDUP,
9306 /// NEON load/store intrinsics, and generic vector load/stores, to merge
9307 /// base address updates.
9308 /// For generic load/stores, the memory type is assumed to be a vector.
9309 /// The caller is assumed to have checked legality.
9310 static SDValue CombineBaseUpdate(SDNode *N,
9311 TargetLowering::DAGCombinerInfo &DCI) {
9312 SelectionDAG &DAG = DCI.DAG;
9313 const bool isIntrinsic = (N->getOpcode() == ISD::INTRINSIC_VOID ||
9314 N->getOpcode() == ISD::INTRINSIC_W_CHAIN);
9315 const bool isStore = N->getOpcode() == ISD::STORE;
9316 const unsigned AddrOpIdx = ((isIntrinsic || isStore) ? 2 : 1);
9317 SDValue Addr = N->getOperand(AddrOpIdx);
9318 MemSDNode *MemN = cast<MemSDNode>(N);
9321 // Search for a use of the address operand that is an increment.
9322 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(),
9323 UE = Addr.getNode()->use_end(); UI != UE; ++UI) {
9325 if (User->getOpcode() != ISD::ADD ||
9326 UI.getUse().getResNo() != Addr.getResNo())
9329 // Check that the add is independent of the load/store. Otherwise, folding
9330 // it would create a cycle.
9331 if (User->isPredecessorOf(N) || N->isPredecessorOf(User))
9334 // Find the new opcode for the updating load/store.
9335 bool isLoadOp = true;
9336 bool isLaneOp = false;
9337 unsigned NewOpc = 0;
9338 unsigned NumVecs = 0;
9340 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
9342 default: llvm_unreachable("unexpected intrinsic for Neon base update");
9343 case Intrinsic::arm_neon_vld1: NewOpc = ARMISD::VLD1_UPD;
9345 case Intrinsic::arm_neon_vld2: NewOpc = ARMISD::VLD2_UPD;
9347 case Intrinsic::arm_neon_vld3: NewOpc = ARMISD::VLD3_UPD;
9349 case Intrinsic::arm_neon_vld4: NewOpc = ARMISD::VLD4_UPD;
9351 case Intrinsic::arm_neon_vld2lane: NewOpc = ARMISD::VLD2LN_UPD;
9352 NumVecs = 2; isLaneOp = true; break;
9353 case Intrinsic::arm_neon_vld3lane: NewOpc = ARMISD::VLD3LN_UPD;
9354 NumVecs = 3; isLaneOp = true; break;
9355 case Intrinsic::arm_neon_vld4lane: NewOpc = ARMISD::VLD4LN_UPD;
9356 NumVecs = 4; isLaneOp = true; break;
9357 case Intrinsic::arm_neon_vst1: NewOpc = ARMISD::VST1_UPD;
9358 NumVecs = 1; isLoadOp = false; break;
9359 case Intrinsic::arm_neon_vst2: NewOpc = ARMISD::VST2_UPD;
9360 NumVecs = 2; isLoadOp = false; break;
9361 case Intrinsic::arm_neon_vst3: NewOpc = ARMISD::VST3_UPD;
9362 NumVecs = 3; isLoadOp = false; break;
9363 case Intrinsic::arm_neon_vst4: NewOpc = ARMISD::VST4_UPD;
9364 NumVecs = 4; isLoadOp = false; break;
9365 case Intrinsic::arm_neon_vst2lane: NewOpc = ARMISD::VST2LN_UPD;
9366 NumVecs = 2; isLoadOp = false; isLaneOp = true; break;
9367 case Intrinsic::arm_neon_vst3lane: NewOpc = ARMISD::VST3LN_UPD;
9368 NumVecs = 3; isLoadOp = false; isLaneOp = true; break;
9369 case Intrinsic::arm_neon_vst4lane: NewOpc = ARMISD::VST4LN_UPD;
9370 NumVecs = 4; isLoadOp = false; isLaneOp = true; break;
9374 switch (N->getOpcode()) {
9375 default: llvm_unreachable("unexpected opcode for Neon base update");
9376 case ARMISD::VLD2DUP: NewOpc = ARMISD::VLD2DUP_UPD; NumVecs = 2; break;
9377 case ARMISD::VLD3DUP: NewOpc = ARMISD::VLD3DUP_UPD; NumVecs = 3; break;
9378 case ARMISD::VLD4DUP: NewOpc = ARMISD::VLD4DUP_UPD; NumVecs = 4; break;
9379 case ISD::LOAD: NewOpc = ARMISD::VLD1_UPD;
9380 NumVecs = 1; isLaneOp = false; break;
9381 case ISD::STORE: NewOpc = ARMISD::VST1_UPD;
9382 NumVecs = 1; isLaneOp = false; isLoadOp = false; break;
9386 // Find the size of memory referenced by the load/store.
9389 VecTy = N->getValueType(0);
9390 } else if (isIntrinsic) {
9391 VecTy = N->getOperand(AddrOpIdx+1).getValueType();
9393 assert(isStore && "Node has to be a load, a store, or an intrinsic!");
9394 VecTy = N->getOperand(1).getValueType();
9397 unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8;
9399 NumBytes /= VecTy.getVectorNumElements();
9401 // If the increment is a constant, it must match the memory ref size.
9402 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
9403 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
9404 uint64_t IncVal = CInc->getZExtValue();
9405 if (IncVal != NumBytes)
9407 } else if (NumBytes >= 3 * 16) {
9408 // VLD3/4 and VST3/4 for 128-bit vectors are implemented with two
9409 // separate instructions that make it harder to use a non-constant update.
9413 // OK, we found an ADD we can fold into the base update.
9414 // Now, create a _UPD node, taking care of not breaking alignment.
9416 EVT AlignedVecTy = VecTy;
9417 unsigned Alignment = MemN->getAlignment();
9419 // If this is a less-than-standard-aligned load/store, change the type to
9420 // match the standard alignment.
9421 // The alignment is overlooked when selecting _UPD variants; and it's
9422 // easier to introduce bitcasts here than fix that.
9423 // There are 3 ways to get to this base-update combine:
9424 // - intrinsics: they are assumed to be properly aligned (to the standard
9425 // alignment of the memory type), so we don't need to do anything.
9426 // - ARMISD::VLDx nodes: they are only generated from the aforementioned
9427 // intrinsics, so, likewise, there's nothing to do.
9428 // - generic load/store instructions: the alignment is specified as an
9429 // explicit operand, rather than implicitly as the standard alignment
9430 // of the memory type (like the intrisics). We need to change the
9431 // memory type to match the explicit alignment. That way, we don't
9432 // generate non-standard-aligned ARMISD::VLDx nodes.
9433 if (isa<LSBaseSDNode>(N)) {
9436 if (Alignment < VecTy.getScalarSizeInBits() / 8) {
9437 MVT EltTy = MVT::getIntegerVT(Alignment * 8);
9438 assert(NumVecs == 1 && "Unexpected multi-element generic load/store.");
9439 assert(!isLaneOp && "Unexpected generic load/store lane.");
9440 unsigned NumElts = NumBytes / (EltTy.getSizeInBits() / 8);
9441 AlignedVecTy = MVT::getVectorVT(EltTy, NumElts);
9443 // Don't set an explicit alignment on regular load/stores that we want
9444 // to transform to VLD/VST 1_UPD nodes.
9445 // This matches the behavior of regular load/stores, which only get an
9446 // explicit alignment if the MMO alignment is larger than the standard
9447 // alignment of the memory type.
9448 // Intrinsics, however, always get an explicit alignment, set to the
9449 // alignment of the MMO.
9453 // Create the new updating load/store node.
9454 // First, create an SDVTList for the new updating node's results.
9456 unsigned NumResultVecs = (isLoadOp ? NumVecs : 0);
9458 for (n = 0; n < NumResultVecs; ++n)
9459 Tys[n] = AlignedVecTy;
9460 Tys[n++] = MVT::i32;
9461 Tys[n] = MVT::Other;
9462 SDVTList SDTys = DAG.getVTList(makeArrayRef(Tys, NumResultVecs+2));
9464 // Then, gather the new node's operands.
9465 SmallVector<SDValue, 8> Ops;
9466 Ops.push_back(N->getOperand(0)); // incoming chain
9467 Ops.push_back(N->getOperand(AddrOpIdx));
9470 if (StoreSDNode *StN = dyn_cast<StoreSDNode>(N)) {
9471 // Try to match the intrinsic's signature
9472 Ops.push_back(StN->getValue());
9474 // Loads (and of course intrinsics) match the intrinsics' signature,
9475 // so just add all but the alignment operand.
9476 for (unsigned i = AddrOpIdx + 1; i < N->getNumOperands() - 1; ++i)
9477 Ops.push_back(N->getOperand(i));
9480 // For all node types, the alignment operand is always the last one.
9481 Ops.push_back(DAG.getConstant(Alignment, dl, MVT::i32));
9483 // If this is a non-standard-aligned STORE, the penultimate operand is the
9484 // stored value. Bitcast it to the aligned type.
9485 if (AlignedVecTy != VecTy && N->getOpcode() == ISD::STORE) {
9486 SDValue &StVal = Ops[Ops.size()-2];
9487 StVal = DAG.getNode(ISD::BITCAST, dl, AlignedVecTy, StVal);
9490 SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, dl, SDTys,
9492 MemN->getMemOperand());
9495 SmallVector<SDValue, 5> NewResults;
9496 for (unsigned i = 0; i < NumResultVecs; ++i)
9497 NewResults.push_back(SDValue(UpdN.getNode(), i));
9499 // If this is an non-standard-aligned LOAD, the first result is the loaded
9500 // value. Bitcast it to the expected result type.
9501 if (AlignedVecTy != VecTy && N->getOpcode() == ISD::LOAD) {
9502 SDValue &LdVal = NewResults[0];
9503 LdVal = DAG.getNode(ISD::BITCAST, dl, VecTy, LdVal);
9506 NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs+1)); // chain
9507 DCI.CombineTo(N, NewResults);
9508 DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs));
9515 static SDValue PerformVLDCombine(SDNode *N,
9516 TargetLowering::DAGCombinerInfo &DCI) {
9517 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
9520 return CombineBaseUpdate(N, DCI);
9523 /// CombineVLDDUP - For a VDUPLANE node N, check if its source operand is a
9524 /// vldN-lane (N > 1) intrinsic, and if all the other uses of that intrinsic
9525 /// are also VDUPLANEs. If so, combine them to a vldN-dup operation and
9527 static bool CombineVLDDUP(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
9528 SelectionDAG &DAG = DCI.DAG;
9529 EVT VT = N->getValueType(0);
9530 // vldN-dup instructions only support 64-bit vectors for N > 1.
9531 if (!VT.is64BitVector())
9534 // Check if the VDUPLANE operand is a vldN-dup intrinsic.
9535 SDNode *VLD = N->getOperand(0).getNode();
9536 if (VLD->getOpcode() != ISD::INTRINSIC_W_CHAIN)
9538 unsigned NumVecs = 0;
9539 unsigned NewOpc = 0;
9540 unsigned IntNo = cast<ConstantSDNode>(VLD->getOperand(1))->getZExtValue();
9541 if (IntNo == Intrinsic::arm_neon_vld2lane) {
9543 NewOpc = ARMISD::VLD2DUP;
9544 } else if (IntNo == Intrinsic::arm_neon_vld3lane) {
9546 NewOpc = ARMISD::VLD3DUP;
9547 } else if (IntNo == Intrinsic::arm_neon_vld4lane) {
9549 NewOpc = ARMISD::VLD4DUP;
9554 // First check that all the vldN-lane uses are VDUPLANEs and that the lane
9555 // numbers match the load.
9556 unsigned VLDLaneNo =
9557 cast<ConstantSDNode>(VLD->getOperand(NumVecs+3))->getZExtValue();
9558 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
9560 // Ignore uses of the chain result.
9561 if (UI.getUse().getResNo() == NumVecs)
9564 if (User->getOpcode() != ARMISD::VDUPLANE ||
9565 VLDLaneNo != cast<ConstantSDNode>(User->getOperand(1))->getZExtValue())
9569 // Create the vldN-dup node.
9572 for (n = 0; n < NumVecs; ++n)
9574 Tys[n] = MVT::Other;
9575 SDVTList SDTys = DAG.getVTList(makeArrayRef(Tys, NumVecs+1));
9576 SDValue Ops[] = { VLD->getOperand(0), VLD->getOperand(2) };
9577 MemIntrinsicSDNode *VLDMemInt = cast<MemIntrinsicSDNode>(VLD);
9578 SDValue VLDDup = DAG.getMemIntrinsicNode(NewOpc, SDLoc(VLD), SDTys,
9579 Ops, VLDMemInt->getMemoryVT(),
9580 VLDMemInt->getMemOperand());
9583 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
9585 unsigned ResNo = UI.getUse().getResNo();
9586 // Ignore uses of the chain result.
9587 if (ResNo == NumVecs)
9590 DCI.CombineTo(User, SDValue(VLDDup.getNode(), ResNo));
9593 // Now the vldN-lane intrinsic is dead except for its chain result.
9594 // Update uses of the chain.
9595 std::vector<SDValue> VLDDupResults;
9596 for (unsigned n = 0; n < NumVecs; ++n)
9597 VLDDupResults.push_back(SDValue(VLDDup.getNode(), n));
9598 VLDDupResults.push_back(SDValue(VLDDup.getNode(), NumVecs));
9599 DCI.CombineTo(VLD, VLDDupResults);
9604 /// PerformVDUPLANECombine - Target-specific dag combine xforms for
9605 /// ARMISD::VDUPLANE.
9606 static SDValue PerformVDUPLANECombine(SDNode *N,
9607 TargetLowering::DAGCombinerInfo &DCI) {
9608 SDValue Op = N->getOperand(0);
9610 // If the source is a vldN-lane (N > 1) intrinsic, and all the other uses
9611 // of that intrinsic are also VDUPLANEs, combine them to a vldN-dup operation.
9612 if (CombineVLDDUP(N, DCI))
9613 return SDValue(N, 0);
9615 // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is
9616 // redundant. Ignore bit_converts for now; element sizes are checked below.
9617 while (Op.getOpcode() == ISD::BITCAST)
9618 Op = Op.getOperand(0);
9619 if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM)
9622 // Make sure the VMOV element size is not bigger than the VDUPLANE elements.
9623 unsigned EltSize = Op.getValueType().getVectorElementType().getSizeInBits();
9624 // The canonical VMOV for a zero vector uses a 32-bit element size.
9625 unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
9627 if (ARM_AM::decodeNEONModImm(Imm, EltBits) == 0)
9629 EVT VT = N->getValueType(0);
9630 if (EltSize > VT.getVectorElementType().getSizeInBits())
9633 return DCI.DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Op);
9636 static SDValue PerformLOADCombine(SDNode *N,
9637 TargetLowering::DAGCombinerInfo &DCI) {
9638 EVT VT = N->getValueType(0);
9640 // If this is a legal vector load, try to combine it into a VLD1_UPD.
9641 if (ISD::isNormalLoad(N) && VT.isVector() &&
9642 DCI.DAG.getTargetLoweringInfo().isTypeLegal(VT))
9643 return CombineBaseUpdate(N, DCI);
9648 /// PerformSTORECombine - Target-specific dag combine xforms for
9650 static SDValue PerformSTORECombine(SDNode *N,
9651 TargetLowering::DAGCombinerInfo &DCI) {
9652 StoreSDNode *St = cast<StoreSDNode>(N);
9653 if (St->isVolatile())
9656 // Optimize trunc store (of multiple scalars) to shuffle and store. First,
9657 // pack all of the elements in one place. Next, store to memory in fewer
9659 SDValue StVal = St->getValue();
9660 EVT VT = StVal.getValueType();
9661 if (St->isTruncatingStore() && VT.isVector()) {
9662 SelectionDAG &DAG = DCI.DAG;
9663 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9664 EVT StVT = St->getMemoryVT();
9665 unsigned NumElems = VT.getVectorNumElements();
9666 assert(StVT != VT && "Cannot truncate to the same type");
9667 unsigned FromEltSz = VT.getVectorElementType().getSizeInBits();
9668 unsigned ToEltSz = StVT.getVectorElementType().getSizeInBits();
9670 // From, To sizes and ElemCount must be pow of two
9671 if (!isPowerOf2_32(NumElems * FromEltSz * ToEltSz)) return SDValue();
9673 // We are going to use the original vector elt for storing.
9674 // Accumulated smaller vector elements must be a multiple of the store size.
9675 if (0 != (NumElems * FromEltSz) % ToEltSz) return SDValue();
9677 unsigned SizeRatio = FromEltSz / ToEltSz;
9678 assert(SizeRatio * NumElems * ToEltSz == VT.getSizeInBits());
9680 // Create a type on which we perform the shuffle.
9681 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(), StVT.getScalarType(),
9682 NumElems*SizeRatio);
9683 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
9686 SDValue WideVec = DAG.getNode(ISD::BITCAST, DL, WideVecVT, StVal);
9687 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
9688 for (unsigned i = 0; i < NumElems; ++i)
9689 ShuffleVec[i] = DAG.getDataLayout().isBigEndian()
9690 ? (i + 1) * SizeRatio - 1
9693 // Can't shuffle using an illegal type.
9694 if (!TLI.isTypeLegal(WideVecVT)) return SDValue();
9696 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, DL, WideVec,
9697 DAG.getUNDEF(WideVec.getValueType()),
9699 // At this point all of the data is stored at the bottom of the
9700 // register. We now need to save it to mem.
9702 // Find the largest store unit
9703 MVT StoreType = MVT::i8;
9704 for (MVT Tp : MVT::integer_valuetypes()) {
9705 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToEltSz)
9708 // Didn't find a legal store type.
9709 if (!TLI.isTypeLegal(StoreType))
9712 // Bitcast the original vector into a vector of store-size units
9713 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
9714 StoreType, VT.getSizeInBits()/EVT(StoreType).getSizeInBits());
9715 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
9716 SDValue ShuffWide = DAG.getNode(ISD::BITCAST, DL, StoreVecVT, Shuff);
9717 SmallVector<SDValue, 8> Chains;
9718 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits() / 8, DL,
9719 TLI.getPointerTy(DAG.getDataLayout()));
9720 SDValue BasePtr = St->getBasePtr();
9722 // Perform one or more big stores into memory.
9723 unsigned E = (ToEltSz*NumElems)/StoreType.getSizeInBits();
9724 for (unsigned I = 0; I < E; I++) {
9725 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
9726 StoreType, ShuffWide,
9727 DAG.getIntPtrConstant(I, DL));
9728 SDValue Ch = DAG.getStore(St->getChain(), DL, SubVec, BasePtr,
9729 St->getPointerInfo(), St->isVolatile(),
9730 St->isNonTemporal(), St->getAlignment());
9731 BasePtr = DAG.getNode(ISD::ADD, DL, BasePtr.getValueType(), BasePtr,
9733 Chains.push_back(Ch);
9735 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
9738 if (!ISD::isNormalStore(St))
9741 // Split a store of a VMOVDRR into two integer stores to avoid mixing NEON and
9742 // ARM stores of arguments in the same cache line.
9743 if (StVal.getNode()->getOpcode() == ARMISD::VMOVDRR &&
9744 StVal.getNode()->hasOneUse()) {
9745 SelectionDAG &DAG = DCI.DAG;
9746 bool isBigEndian = DAG.getDataLayout().isBigEndian();
9748 SDValue BasePtr = St->getBasePtr();
9749 SDValue NewST1 = DAG.getStore(St->getChain(), DL,
9750 StVal.getNode()->getOperand(isBigEndian ? 1 : 0 ),
9751 BasePtr, St->getPointerInfo(), St->isVolatile(),
9752 St->isNonTemporal(), St->getAlignment());
9754 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
9755 DAG.getConstant(4, DL, MVT::i32));
9756 return DAG.getStore(NewST1.getValue(0), DL,
9757 StVal.getNode()->getOperand(isBigEndian ? 0 : 1),
9758 OffsetPtr, St->getPointerInfo(), St->isVolatile(),
9759 St->isNonTemporal(),
9760 std::min(4U, St->getAlignment() / 2));
9763 if (StVal.getValueType() == MVT::i64 &&
9764 StVal.getNode()->getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
9766 // Bitcast an i64 store extracted from a vector to f64.
9767 // Otherwise, the i64 value will be legalized to a pair of i32 values.
9768 SelectionDAG &DAG = DCI.DAG;
9770 SDValue IntVec = StVal.getOperand(0);
9771 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
9772 IntVec.getValueType().getVectorNumElements());
9773 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, IntVec);
9774 SDValue ExtElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
9775 Vec, StVal.getOperand(1));
9777 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ExtElt);
9778 // Make the DAGCombiner fold the bitcasts.
9779 DCI.AddToWorklist(Vec.getNode());
9780 DCI.AddToWorklist(ExtElt.getNode());
9781 DCI.AddToWorklist(V.getNode());
9782 return DAG.getStore(St->getChain(), dl, V, St->getBasePtr(),
9783 St->getPointerInfo(), St->isVolatile(),
9784 St->isNonTemporal(), St->getAlignment(),
9788 // If this is a legal vector store, try to combine it into a VST1_UPD.
9789 if (ISD::isNormalStore(N) && VT.isVector() &&
9790 DCI.DAG.getTargetLoweringInfo().isTypeLegal(VT))
9791 return CombineBaseUpdate(N, DCI);
9796 /// PerformVCVTCombine - VCVT (floating-point to fixed-point, Advanced SIMD)
9797 /// can replace combinations of VMUL and VCVT (floating-point to integer)
9798 /// when the VMUL has a constant operand that is a power of 2.
9800 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
9801 /// vmul.f32 d16, d17, d16
9802 /// vcvt.s32.f32 d16, d16
9804 /// vcvt.s32.f32 d16, d16, #3
9805 static SDValue PerformVCVTCombine(SDNode *N, SelectionDAG &DAG,
9806 const ARMSubtarget *Subtarget) {
9807 if (!Subtarget->hasNEON())
9810 SDValue Op = N->getOperand(0);
9811 if (!Op.getValueType().isVector() || Op.getOpcode() != ISD::FMUL)
9814 SDValue ConstVec = Op->getOperand(1);
9815 if (!isa<BuildVectorSDNode>(ConstVec))
9818 MVT FloatTy = Op.getSimpleValueType().getVectorElementType();
9819 uint32_t FloatBits = FloatTy.getSizeInBits();
9820 MVT IntTy = N->getSimpleValueType(0).getVectorElementType();
9821 uint32_t IntBits = IntTy.getSizeInBits();
9822 unsigned NumLanes = Op.getValueType().getVectorNumElements();
9823 if (FloatBits != 32 || IntBits > 32 || NumLanes > 4) {
9824 // These instructions only exist converting from f32 to i32. We can handle
9825 // smaller integers by generating an extra truncate, but larger ones would
9826 // be lossy. We also can't handle more then 4 lanes, since these intructions
9827 // only support v2i32/v4i32 types.
9831 BitVector UndefElements;
9832 BuildVectorSDNode *BV = cast<BuildVectorSDNode>(ConstVec);
9833 int32_t C = BV->getConstantFPSplatPow2ToLog2Int(&UndefElements, 33);
9834 if (C == -1 || C == 0 || C > 32)
9838 bool isSigned = N->getOpcode() == ISD::FP_TO_SINT;
9839 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfp2fxs :
9840 Intrinsic::arm_neon_vcvtfp2fxu;
9841 SDValue FixConv = DAG.getNode(
9842 ISD::INTRINSIC_WO_CHAIN, dl, NumLanes == 2 ? MVT::v2i32 : MVT::v4i32,
9843 DAG.getConstant(IntrinsicOpcode, dl, MVT::i32), Op->getOperand(0),
9844 DAG.getConstant(C, dl, MVT::i32));
9846 if (IntBits < FloatBits)
9847 FixConv = DAG.getNode(ISD::TRUNCATE, dl, N->getValueType(0), FixConv);
9852 /// PerformVDIVCombine - VCVT (fixed-point to floating-point, Advanced SIMD)
9853 /// can replace combinations of VCVT (integer to floating-point) and VDIV
9854 /// when the VDIV has a constant operand that is a power of 2.
9856 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
9857 /// vcvt.f32.s32 d16, d16
9858 /// vdiv.f32 d16, d17, d16
9860 /// vcvt.f32.s32 d16, d16, #3
9861 static SDValue PerformVDIVCombine(SDNode *N, SelectionDAG &DAG,
9862 const ARMSubtarget *Subtarget) {
9863 if (!Subtarget->hasNEON())
9866 SDValue Op = N->getOperand(0);
9867 unsigned OpOpcode = Op.getNode()->getOpcode();
9868 if (!N->getValueType(0).isVector() ||
9869 (OpOpcode != ISD::SINT_TO_FP && OpOpcode != ISD::UINT_TO_FP))
9872 SDValue ConstVec = N->getOperand(1);
9873 if (!isa<BuildVectorSDNode>(ConstVec))
9876 MVT FloatTy = N->getSimpleValueType(0).getVectorElementType();
9877 uint32_t FloatBits = FloatTy.getSizeInBits();
9878 MVT IntTy = Op.getOperand(0).getSimpleValueType().getVectorElementType();
9879 uint32_t IntBits = IntTy.getSizeInBits();
9880 unsigned NumLanes = Op.getValueType().getVectorNumElements();
9881 if (FloatBits != 32 || IntBits > 32 || NumLanes > 4) {
9882 // These instructions only exist converting from i32 to f32. We can handle
9883 // smaller integers by generating an extra extend, but larger ones would
9884 // be lossy. We also can't handle more then 4 lanes, since these intructions
9885 // only support v2i32/v4i32 types.
9889 BitVector UndefElements;
9890 BuildVectorSDNode *BV = cast<BuildVectorSDNode>(ConstVec);
9891 int32_t C = BV->getConstantFPSplatPow2ToLog2Int(&UndefElements, 33);
9892 if (C == -1 || C == 0 || C > 32)
9896 bool isSigned = OpOpcode == ISD::SINT_TO_FP;
9897 SDValue ConvInput = Op.getOperand(0);
9898 if (IntBits < FloatBits)
9899 ConvInput = DAG.getNode(isSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND,
9900 dl, NumLanes == 2 ? MVT::v2i32 : MVT::v4i32,
9903 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfxs2fp :
9904 Intrinsic::arm_neon_vcvtfxu2fp;
9905 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl,
9907 DAG.getConstant(IntrinsicOpcode, dl, MVT::i32),
9908 ConvInput, DAG.getConstant(C, dl, MVT::i32));
9911 /// Getvshiftimm - Check if this is a valid build_vector for the immediate
9912 /// operand of a vector shift operation, where all the elements of the
9913 /// build_vector must have the same constant integer value.
9914 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
9915 // Ignore bit_converts.
9916 while (Op.getOpcode() == ISD::BITCAST)
9917 Op = Op.getOperand(0);
9918 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
9919 APInt SplatBits, SplatUndef;
9920 unsigned SplatBitSize;
9922 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
9923 HasAnyUndefs, ElementBits) ||
9924 SplatBitSize > ElementBits)
9926 Cnt = SplatBits.getSExtValue();
9930 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
9931 /// operand of a vector shift left operation. That value must be in the range:
9932 /// 0 <= Value < ElementBits for a left shift; or
9933 /// 0 <= Value <= ElementBits for a long left shift.
9934 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
9935 assert(VT.isVector() && "vector shift count is not a vector type");
9936 int64_t ElementBits = VT.getVectorElementType().getSizeInBits();
9937 if (! getVShiftImm(Op, ElementBits, Cnt))
9939 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
9942 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
9943 /// operand of a vector shift right operation. For a shift opcode, the value
9944 /// is positive, but for an intrinsic the value count must be negative. The
9945 /// absolute value must be in the range:
9946 /// 1 <= |Value| <= ElementBits for a right shift; or
9947 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
9948 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
9950 assert(VT.isVector() && "vector shift count is not a vector type");
9951 int64_t ElementBits = VT.getVectorElementType().getSizeInBits();
9952 if (! getVShiftImm(Op, ElementBits, Cnt))
9955 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
9956 if (Cnt >= -(isNarrow ? ElementBits/2 : ElementBits) && Cnt <= -1) {
9963 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
9964 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
9965 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
9968 // Don't do anything for most intrinsics.
9971 case Intrinsic::arm_neon_vabds:
9972 if (!N->getValueType(0).isInteger())
9974 return DAG.getNode(ISD::SABSDIFF, SDLoc(N), N->getValueType(0),
9975 N->getOperand(1), N->getOperand(2));
9976 case Intrinsic::arm_neon_vabdu:
9977 return DAG.getNode(ISD::UABSDIFF, SDLoc(N), N->getValueType(0),
9978 N->getOperand(1), N->getOperand(2));
9980 // Vector shifts: check for immediate versions and lower them.
9981 // Note: This is done during DAG combining instead of DAG legalizing because
9982 // the build_vectors for 64-bit vector element shift counts are generally
9983 // not legal, and it is hard to see their values after they get legalized to
9984 // loads from a constant pool.
9985 case Intrinsic::arm_neon_vshifts:
9986 case Intrinsic::arm_neon_vshiftu:
9987 case Intrinsic::arm_neon_vrshifts:
9988 case Intrinsic::arm_neon_vrshiftu:
9989 case Intrinsic::arm_neon_vrshiftn:
9990 case Intrinsic::arm_neon_vqshifts:
9991 case Intrinsic::arm_neon_vqshiftu:
9992 case Intrinsic::arm_neon_vqshiftsu:
9993 case Intrinsic::arm_neon_vqshiftns:
9994 case Intrinsic::arm_neon_vqshiftnu:
9995 case Intrinsic::arm_neon_vqshiftnsu:
9996 case Intrinsic::arm_neon_vqrshiftns:
9997 case Intrinsic::arm_neon_vqrshiftnu:
9998 case Intrinsic::arm_neon_vqrshiftnsu: {
9999 EVT VT = N->getOperand(1).getValueType();
10001 unsigned VShiftOpc = 0;
10004 case Intrinsic::arm_neon_vshifts:
10005 case Intrinsic::arm_neon_vshiftu:
10006 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
10007 VShiftOpc = ARMISD::VSHL;
10010 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
10011 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
10012 ARMISD::VSHRs : ARMISD::VSHRu);
10017 case Intrinsic::arm_neon_vrshifts:
10018 case Intrinsic::arm_neon_vrshiftu:
10019 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
10023 case Intrinsic::arm_neon_vqshifts:
10024 case Intrinsic::arm_neon_vqshiftu:
10025 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
10029 case Intrinsic::arm_neon_vqshiftsu:
10030 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
10032 llvm_unreachable("invalid shift count for vqshlu intrinsic");
10034 case Intrinsic::arm_neon_vrshiftn:
10035 case Intrinsic::arm_neon_vqshiftns:
10036 case Intrinsic::arm_neon_vqshiftnu:
10037 case Intrinsic::arm_neon_vqshiftnsu:
10038 case Intrinsic::arm_neon_vqrshiftns:
10039 case Intrinsic::arm_neon_vqrshiftnu:
10040 case Intrinsic::arm_neon_vqrshiftnsu:
10041 // Narrowing shifts require an immediate right shift.
10042 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
10044 llvm_unreachable("invalid shift count for narrowing vector shift "
10048 llvm_unreachable("unhandled vector shift");
10052 case Intrinsic::arm_neon_vshifts:
10053 case Intrinsic::arm_neon_vshiftu:
10054 // Opcode already set above.
10056 case Intrinsic::arm_neon_vrshifts:
10057 VShiftOpc = ARMISD::VRSHRs; break;
10058 case Intrinsic::arm_neon_vrshiftu:
10059 VShiftOpc = ARMISD::VRSHRu; break;
10060 case Intrinsic::arm_neon_vrshiftn:
10061 VShiftOpc = ARMISD::VRSHRN; break;
10062 case Intrinsic::arm_neon_vqshifts:
10063 VShiftOpc = ARMISD::VQSHLs; break;
10064 case Intrinsic::arm_neon_vqshiftu:
10065 VShiftOpc = ARMISD::VQSHLu; break;
10066 case Intrinsic::arm_neon_vqshiftsu:
10067 VShiftOpc = ARMISD::VQSHLsu; break;
10068 case Intrinsic::arm_neon_vqshiftns:
10069 VShiftOpc = ARMISD::VQSHRNs; break;
10070 case Intrinsic::arm_neon_vqshiftnu:
10071 VShiftOpc = ARMISD::VQSHRNu; break;
10072 case Intrinsic::arm_neon_vqshiftnsu:
10073 VShiftOpc = ARMISD::VQSHRNsu; break;
10074 case Intrinsic::arm_neon_vqrshiftns:
10075 VShiftOpc = ARMISD::VQRSHRNs; break;
10076 case Intrinsic::arm_neon_vqrshiftnu:
10077 VShiftOpc = ARMISD::VQRSHRNu; break;
10078 case Intrinsic::arm_neon_vqrshiftnsu:
10079 VShiftOpc = ARMISD::VQRSHRNsu; break;
10083 return DAG.getNode(VShiftOpc, dl, N->getValueType(0),
10084 N->getOperand(1), DAG.getConstant(Cnt, dl, MVT::i32));
10087 case Intrinsic::arm_neon_vshiftins: {
10088 EVT VT = N->getOperand(1).getValueType();
10090 unsigned VShiftOpc = 0;
10092 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
10093 VShiftOpc = ARMISD::VSLI;
10094 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
10095 VShiftOpc = ARMISD::VSRI;
10097 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
10101 return DAG.getNode(VShiftOpc, dl, N->getValueType(0),
10102 N->getOperand(1), N->getOperand(2),
10103 DAG.getConstant(Cnt, dl, MVT::i32));
10106 case Intrinsic::arm_neon_vqrshifts:
10107 case Intrinsic::arm_neon_vqrshiftu:
10108 // No immediate versions of these to check for.
10115 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
10116 /// lowers them. As with the vector shift intrinsics, this is done during DAG
10117 /// combining instead of DAG legalizing because the build_vectors for 64-bit
10118 /// vector element shift counts are generally not legal, and it is hard to see
10119 /// their values after they get legalized to loads from a constant pool.
10120 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
10121 const ARMSubtarget *ST) {
10122 EVT VT = N->getValueType(0);
10123 if (N->getOpcode() == ISD::SRL && VT == MVT::i32 && ST->hasV6Ops()) {
10124 // Canonicalize (srl (bswap x), 16) to (rotr (bswap x), 16) if the high
10125 // 16-bits of x is zero. This optimizes rev + lsr 16 to rev16.
10126 SDValue N1 = N->getOperand(1);
10127 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) {
10128 SDValue N0 = N->getOperand(0);
10129 if (C->getZExtValue() == 16 && N0.getOpcode() == ISD::BSWAP &&
10130 DAG.MaskedValueIsZero(N0.getOperand(0),
10131 APInt::getHighBitsSet(32, 16)))
10132 return DAG.getNode(ISD::ROTR, SDLoc(N), VT, N0, N1);
10136 // Nothing to be done for scalar shifts.
10137 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
10138 if (!VT.isVector() || !TLI.isTypeLegal(VT))
10141 assert(ST->hasNEON() && "unexpected vector shift");
10144 switch (N->getOpcode()) {
10145 default: llvm_unreachable("unexpected shift opcode");
10148 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt)) {
10150 return DAG.getNode(ARMISD::VSHL, dl, VT, N->getOperand(0),
10151 DAG.getConstant(Cnt, dl, MVT::i32));
10157 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
10158 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
10159 ARMISD::VSHRs : ARMISD::VSHRu);
10161 return DAG.getNode(VShiftOpc, dl, VT, N->getOperand(0),
10162 DAG.getConstant(Cnt, dl, MVT::i32));
10168 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
10169 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
10170 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
10171 const ARMSubtarget *ST) {
10172 SDValue N0 = N->getOperand(0);
10174 // Check for sign- and zero-extensions of vector extract operations of 8-
10175 // and 16-bit vector elements. NEON supports these directly. They are
10176 // handled during DAG combining because type legalization will promote them
10177 // to 32-bit types and it is messy to recognize the operations after that.
10178 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
10179 SDValue Vec = N0.getOperand(0);
10180 SDValue Lane = N0.getOperand(1);
10181 EVT VT = N->getValueType(0);
10182 EVT EltVT = N0.getValueType();
10183 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
10185 if (VT == MVT::i32 &&
10186 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
10187 TLI.isTypeLegal(Vec.getValueType()) &&
10188 isa<ConstantSDNode>(Lane)) {
10191 switch (N->getOpcode()) {
10192 default: llvm_unreachable("unexpected opcode");
10193 case ISD::SIGN_EXTEND:
10194 Opc = ARMISD::VGETLANEs;
10196 case ISD::ZERO_EXTEND:
10197 case ISD::ANY_EXTEND:
10198 Opc = ARMISD::VGETLANEu;
10201 return DAG.getNode(Opc, SDLoc(N), VT, Vec, Lane);
10208 /// PerformCMOVCombine - Target-specific DAG combining for ARMISD::CMOV.
10210 ARMTargetLowering::PerformCMOVCombine(SDNode *N, SelectionDAG &DAG) const {
10211 SDValue Cmp = N->getOperand(4);
10212 if (Cmp.getOpcode() != ARMISD::CMPZ)
10213 // Only looking at EQ and NE cases.
10216 EVT VT = N->getValueType(0);
10218 SDValue LHS = Cmp.getOperand(0);
10219 SDValue RHS = Cmp.getOperand(1);
10220 SDValue FalseVal = N->getOperand(0);
10221 SDValue TrueVal = N->getOperand(1);
10222 SDValue ARMcc = N->getOperand(2);
10223 ARMCC::CondCodes CC =
10224 (ARMCC::CondCodes)cast<ConstantSDNode>(ARMcc)->getZExtValue();
10242 /// FIXME: Turn this into a target neutral optimization?
10244 if (CC == ARMCC::NE && FalseVal == RHS && FalseVal != LHS) {
10245 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, TrueVal, ARMcc,
10246 N->getOperand(3), Cmp);
10247 } else if (CC == ARMCC::EQ && TrueVal == RHS) {
10249 SDValue NewCmp = getARMCmp(LHS, RHS, ISD::SETNE, ARMcc, DAG, dl);
10250 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, FalseVal, ARMcc,
10251 N->getOperand(3), NewCmp);
10254 if (Res.getNode()) {
10255 APInt KnownZero, KnownOne;
10256 DAG.computeKnownBits(SDValue(N,0), KnownZero, KnownOne);
10257 // Capture demanded bits information that would be otherwise lost.
10258 if (KnownZero == 0xfffffffe)
10259 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
10260 DAG.getValueType(MVT::i1));
10261 else if (KnownZero == 0xffffff00)
10262 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
10263 DAG.getValueType(MVT::i8));
10264 else if (KnownZero == 0xffff0000)
10265 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
10266 DAG.getValueType(MVT::i16));
10272 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
10273 DAGCombinerInfo &DCI) const {
10274 switch (N->getOpcode()) {
10276 case ISD::ADDC: return PerformADDCCombine(N, DCI, Subtarget);
10277 case ISD::ADD: return PerformADDCombine(N, DCI, Subtarget);
10278 case ISD::SUB: return PerformSUBCombine(N, DCI);
10279 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
10280 case ISD::OR: return PerformORCombine(N, DCI, Subtarget);
10281 case ISD::XOR: return PerformXORCombine(N, DCI, Subtarget);
10282 case ISD::AND: return PerformANDCombine(N, DCI, Subtarget);
10283 case ARMISD::BFI: return PerformBFICombine(N, DCI);
10284 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI, Subtarget);
10285 case ARMISD::VMOVDRR: return PerformVMOVDRRCombine(N, DCI.DAG);
10286 case ISD::STORE: return PerformSTORECombine(N, DCI);
10287 case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DCI, Subtarget);
10288 case ISD::INSERT_VECTOR_ELT: return PerformInsertEltCombine(N, DCI);
10289 case ISD::VECTOR_SHUFFLE: return PerformVECTOR_SHUFFLECombine(N, DCI.DAG);
10290 case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI);
10291 case ISD::FP_TO_SINT:
10292 case ISD::FP_TO_UINT:
10293 return PerformVCVTCombine(N, DCI.DAG, Subtarget);
10295 return PerformVDIVCombine(N, DCI.DAG, Subtarget);
10296 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
10299 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
10300 case ISD::SIGN_EXTEND:
10301 case ISD::ZERO_EXTEND:
10302 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
10303 case ARMISD::CMOV: return PerformCMOVCombine(N, DCI.DAG);
10304 case ISD::LOAD: return PerformLOADCombine(N, DCI);
10305 case ARMISD::VLD2DUP:
10306 case ARMISD::VLD3DUP:
10307 case ARMISD::VLD4DUP:
10308 return PerformVLDCombine(N, DCI);
10309 case ARMISD::BUILD_VECTOR:
10310 return PerformARMBUILD_VECTORCombine(N, DCI);
10311 case ISD::INTRINSIC_VOID:
10312 case ISD::INTRINSIC_W_CHAIN:
10313 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
10314 case Intrinsic::arm_neon_vld1:
10315 case Intrinsic::arm_neon_vld2:
10316 case Intrinsic::arm_neon_vld3:
10317 case Intrinsic::arm_neon_vld4:
10318 case Intrinsic::arm_neon_vld2lane:
10319 case Intrinsic::arm_neon_vld3lane:
10320 case Intrinsic::arm_neon_vld4lane:
10321 case Intrinsic::arm_neon_vst1:
10322 case Intrinsic::arm_neon_vst2:
10323 case Intrinsic::arm_neon_vst3:
10324 case Intrinsic::arm_neon_vst4:
10325 case Intrinsic::arm_neon_vst2lane:
10326 case Intrinsic::arm_neon_vst3lane:
10327 case Intrinsic::arm_neon_vst4lane:
10328 return PerformVLDCombine(N, DCI);
10336 bool ARMTargetLowering::isDesirableToTransformToIntegerOp(unsigned Opc,
10338 return (VT == MVT::f32) && (Opc == ISD::LOAD || Opc == ISD::STORE);
10341 bool ARMTargetLowering::allowsMisalignedMemoryAccesses(EVT VT,
10344 bool *Fast) const {
10345 // The AllowsUnaliged flag models the SCTLR.A setting in ARM cpus
10346 bool AllowsUnaligned = Subtarget->allowsUnalignedMem();
10348 switch (VT.getSimpleVT().SimpleTy) {
10354 // Unaligned access can use (for example) LRDB, LRDH, LDR
10355 if (AllowsUnaligned) {
10357 *Fast = Subtarget->hasV7Ops();
10364 // For any little-endian targets with neon, we can support unaligned ld/st
10365 // of D and Q (e.g. {D0,D1}) registers by using vld1.i8/vst1.i8.
10366 // A big-endian target may also explicitly support unaligned accesses
10367 if (Subtarget->hasNEON() && (AllowsUnaligned || Subtarget->isLittle())) {
10377 static bool memOpAlign(unsigned DstAlign, unsigned SrcAlign,
10378 unsigned AlignCheck) {
10379 return ((SrcAlign == 0 || SrcAlign % AlignCheck == 0) &&
10380 (DstAlign == 0 || DstAlign % AlignCheck == 0));
10383 EVT ARMTargetLowering::getOptimalMemOpType(uint64_t Size,
10384 unsigned DstAlign, unsigned SrcAlign,
10385 bool IsMemset, bool ZeroMemset,
10387 MachineFunction &MF) const {
10388 const Function *F = MF.getFunction();
10390 // See if we can use NEON instructions for this...
10391 if ((!IsMemset || ZeroMemset) && Subtarget->hasNEON() &&
10392 !F->hasFnAttribute(Attribute::NoImplicitFloat)) {
10395 (memOpAlign(SrcAlign, DstAlign, 16) ||
10396 (allowsMisalignedMemoryAccesses(MVT::v2f64, 0, 1, &Fast) && Fast))) {
10398 } else if (Size >= 8 &&
10399 (memOpAlign(SrcAlign, DstAlign, 8) ||
10400 (allowsMisalignedMemoryAccesses(MVT::f64, 0, 1, &Fast) &&
10406 // Lowering to i32/i16 if the size permits.
10409 else if (Size >= 2)
10412 // Let the target-independent logic figure it out.
10416 bool ARMTargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
10417 if (Val.getOpcode() != ISD::LOAD)
10420 EVT VT1 = Val.getValueType();
10421 if (!VT1.isSimple() || !VT1.isInteger() ||
10422 !VT2.isSimple() || !VT2.isInteger())
10425 switch (VT1.getSimpleVT().SimpleTy) {
10430 // 8-bit and 16-bit loads implicitly zero-extend to 32-bits.
10437 bool ARMTargetLowering::isVectorLoadExtDesirable(SDValue ExtVal) const {
10438 EVT VT = ExtVal.getValueType();
10440 if (!isTypeLegal(VT))
10443 // Don't create a loadext if we can fold the extension into a wide/long
10445 // If there's more than one user instruction, the loadext is desirable no
10446 // matter what. There can be two uses by the same instruction.
10447 if (ExtVal->use_empty() ||
10448 !ExtVal->use_begin()->isOnlyUserOf(ExtVal.getNode()))
10451 SDNode *U = *ExtVal->use_begin();
10452 if ((U->getOpcode() == ISD::ADD || U->getOpcode() == ISD::SUB ||
10453 U->getOpcode() == ISD::SHL || U->getOpcode() == ARMISD::VSHL))
10459 bool ARMTargetLowering::allowTruncateForTailCall(Type *Ty1, Type *Ty2) const {
10460 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
10463 if (!isTypeLegal(EVT::getEVT(Ty1)))
10466 assert(Ty1->getPrimitiveSizeInBits() <= 64 && "i128 is probably not a noop");
10468 // Assuming the caller doesn't have a zeroext or signext return parameter,
10469 // truncation all the way down to i1 is valid.
10474 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
10478 unsigned Scale = 1;
10479 switch (VT.getSimpleVT().SimpleTy) {
10480 default: return false;
10495 if ((V & (Scale - 1)) != 0)
10498 return V == (V & ((1LL << 5) - 1));
10501 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
10502 const ARMSubtarget *Subtarget) {
10503 bool isNeg = false;
10509 switch (VT.getSimpleVT().SimpleTy) {
10510 default: return false;
10515 // + imm12 or - imm8
10517 return V == (V & ((1LL << 8) - 1));
10518 return V == (V & ((1LL << 12) - 1));
10521 // Same as ARM mode. FIXME: NEON?
10522 if (!Subtarget->hasVFP2())
10527 return V == (V & ((1LL << 8) - 1));
10531 /// isLegalAddressImmediate - Return true if the integer value can be used
10532 /// as the offset of the target addressing mode for load / store of the
10534 static bool isLegalAddressImmediate(int64_t V, EVT VT,
10535 const ARMSubtarget *Subtarget) {
10539 if (!VT.isSimple())
10542 if (Subtarget->isThumb1Only())
10543 return isLegalT1AddressImmediate(V, VT);
10544 else if (Subtarget->isThumb2())
10545 return isLegalT2AddressImmediate(V, VT, Subtarget);
10550 switch (VT.getSimpleVT().SimpleTy) {
10551 default: return false;
10556 return V == (V & ((1LL << 12) - 1));
10559 return V == (V & ((1LL << 8) - 1));
10562 if (!Subtarget->hasVFP2()) // FIXME: NEON?
10567 return V == (V & ((1LL << 8) - 1));
10571 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
10573 int Scale = AM.Scale;
10577 switch (VT.getSimpleVT().SimpleTy) {
10578 default: return false;
10586 Scale = Scale & ~1;
10587 return Scale == 2 || Scale == 4 || Scale == 8;
10590 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
10594 // Note, we allow "void" uses (basically, uses that aren't loads or
10595 // stores), because arm allows folding a scale into many arithmetic
10596 // operations. This should be made more precise and revisited later.
10598 // Allow r << imm, but the imm has to be a multiple of two.
10599 if (Scale & 1) return false;
10600 return isPowerOf2_32(Scale);
10604 /// isLegalAddressingMode - Return true if the addressing mode represented
10605 /// by AM is legal for this target, for a load/store of the specified type.
10606 bool ARMTargetLowering::isLegalAddressingMode(const DataLayout &DL,
10607 const AddrMode &AM, Type *Ty,
10608 unsigned AS) const {
10609 EVT VT = getValueType(DL, Ty, true);
10610 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
10613 // Can never fold addr of global into load/store.
10617 switch (AM.Scale) {
10618 case 0: // no scale reg, must be "r+i" or "r", or "i".
10621 if (Subtarget->isThumb1Only())
10625 // ARM doesn't support any R+R*scale+imm addr modes.
10629 if (!VT.isSimple())
10632 if (Subtarget->isThumb2())
10633 return isLegalT2ScaledAddressingMode(AM, VT);
10635 int Scale = AM.Scale;
10636 switch (VT.getSimpleVT().SimpleTy) {
10637 default: return false;
10641 if (Scale < 0) Scale = -Scale;
10645 return isPowerOf2_32(Scale & ~1);
10649 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
10654 // Note, we allow "void" uses (basically, uses that aren't loads or
10655 // stores), because arm allows folding a scale into many arithmetic
10656 // operations. This should be made more precise and revisited later.
10658 // Allow r << imm, but the imm has to be a multiple of two.
10659 if (Scale & 1) return false;
10660 return isPowerOf2_32(Scale);
10666 /// isLegalICmpImmediate - Return true if the specified immediate is legal
10667 /// icmp immediate, that is the target has icmp instructions which can compare
10668 /// a register against the immediate without having to materialize the
10669 /// immediate into a register.
10670 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
10671 // Thumb2 and ARM modes can use cmn for negative immediates.
10672 if (!Subtarget->isThumb())
10673 return ARM_AM::getSOImmVal(std::abs(Imm)) != -1;
10674 if (Subtarget->isThumb2())
10675 return ARM_AM::getT2SOImmVal(std::abs(Imm)) != -1;
10676 // Thumb1 doesn't have cmn, and only 8-bit immediates.
10677 return Imm >= 0 && Imm <= 255;
10680 /// isLegalAddImmediate - Return true if the specified immediate is a legal add
10681 /// *or sub* immediate, that is the target has add or sub instructions which can
10682 /// add a register with the immediate without having to materialize the
10683 /// immediate into a register.
10684 bool ARMTargetLowering::isLegalAddImmediate(int64_t Imm) const {
10685 // Same encoding for add/sub, just flip the sign.
10686 int64_t AbsImm = std::abs(Imm);
10687 if (!Subtarget->isThumb())
10688 return ARM_AM::getSOImmVal(AbsImm) != -1;
10689 if (Subtarget->isThumb2())
10690 return ARM_AM::getT2SOImmVal(AbsImm) != -1;
10691 // Thumb1 only has 8-bit unsigned immediate.
10692 return AbsImm >= 0 && AbsImm <= 255;
10695 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
10696 bool isSEXTLoad, SDValue &Base,
10697 SDValue &Offset, bool &isInc,
10698 SelectionDAG &DAG) {
10699 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
10702 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
10703 // AddressingMode 3
10704 Base = Ptr->getOperand(0);
10705 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
10706 int RHSC = (int)RHS->getZExtValue();
10707 if (RHSC < 0 && RHSC > -256) {
10708 assert(Ptr->getOpcode() == ISD::ADD);
10710 Offset = DAG.getConstant(-RHSC, SDLoc(Ptr), RHS->getValueType(0));
10714 isInc = (Ptr->getOpcode() == ISD::ADD);
10715 Offset = Ptr->getOperand(1);
10717 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
10718 // AddressingMode 2
10719 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
10720 int RHSC = (int)RHS->getZExtValue();
10721 if (RHSC < 0 && RHSC > -0x1000) {
10722 assert(Ptr->getOpcode() == ISD::ADD);
10724 Offset = DAG.getConstant(-RHSC, SDLoc(Ptr), RHS->getValueType(0));
10725 Base = Ptr->getOperand(0);
10730 if (Ptr->getOpcode() == ISD::ADD) {
10732 ARM_AM::ShiftOpc ShOpcVal=
10733 ARM_AM::getShiftOpcForNode(Ptr->getOperand(0).getOpcode());
10734 if (ShOpcVal != ARM_AM::no_shift) {
10735 Base = Ptr->getOperand(1);
10736 Offset = Ptr->getOperand(0);
10738 Base = Ptr->getOperand(0);
10739 Offset = Ptr->getOperand(1);
10744 isInc = (Ptr->getOpcode() == ISD::ADD);
10745 Base = Ptr->getOperand(0);
10746 Offset = Ptr->getOperand(1);
10750 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
10754 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
10755 bool isSEXTLoad, SDValue &Base,
10756 SDValue &Offset, bool &isInc,
10757 SelectionDAG &DAG) {
10758 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
10761 Base = Ptr->getOperand(0);
10762 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
10763 int RHSC = (int)RHS->getZExtValue();
10764 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
10765 assert(Ptr->getOpcode() == ISD::ADD);
10767 Offset = DAG.getConstant(-RHSC, SDLoc(Ptr), RHS->getValueType(0));
10769 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
10770 isInc = Ptr->getOpcode() == ISD::ADD;
10771 Offset = DAG.getConstant(RHSC, SDLoc(Ptr), RHS->getValueType(0));
10779 /// getPreIndexedAddressParts - returns true by value, base pointer and
10780 /// offset pointer and addressing mode by reference if the node's address
10781 /// can be legally represented as pre-indexed load / store address.
10783 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
10785 ISD::MemIndexedMode &AM,
10786 SelectionDAG &DAG) const {
10787 if (Subtarget->isThumb1Only())
10792 bool isSEXTLoad = false;
10793 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
10794 Ptr = LD->getBasePtr();
10795 VT = LD->getMemoryVT();
10796 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
10797 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
10798 Ptr = ST->getBasePtr();
10799 VT = ST->getMemoryVT();
10804 bool isLegal = false;
10805 if (Subtarget->isThumb2())
10806 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
10807 Offset, isInc, DAG);
10809 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
10810 Offset, isInc, DAG);
10814 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
10818 /// getPostIndexedAddressParts - returns true by value, base pointer and
10819 /// offset pointer and addressing mode by reference if this node can be
10820 /// combined with a load / store to form a post-indexed load / store.
10821 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
10824 ISD::MemIndexedMode &AM,
10825 SelectionDAG &DAG) const {
10826 if (Subtarget->isThumb1Only())
10831 bool isSEXTLoad = false;
10832 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
10833 VT = LD->getMemoryVT();
10834 Ptr = LD->getBasePtr();
10835 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
10836 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
10837 VT = ST->getMemoryVT();
10838 Ptr = ST->getBasePtr();
10843 bool isLegal = false;
10844 if (Subtarget->isThumb2())
10845 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
10848 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
10854 // Swap base ptr and offset to catch more post-index load / store when
10855 // it's legal. In Thumb2 mode, offset must be an immediate.
10856 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
10857 !Subtarget->isThumb2())
10858 std::swap(Base, Offset);
10860 // Post-indexed load / store update the base pointer.
10865 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
10869 void ARMTargetLowering::computeKnownBitsForTargetNode(const SDValue Op,
10872 const SelectionDAG &DAG,
10873 unsigned Depth) const {
10874 unsigned BitWidth = KnownOne.getBitWidth();
10875 KnownZero = KnownOne = APInt(BitWidth, 0);
10876 switch (Op.getOpcode()) {
10882 // These nodes' second result is a boolean
10883 if (Op.getResNo() == 0)
10885 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - 1);
10887 case ARMISD::CMOV: {
10888 // Bits are known zero/one if known on the LHS and RHS.
10889 DAG.computeKnownBits(Op.getOperand(0), KnownZero, KnownOne, Depth+1);
10890 if (KnownZero == 0 && KnownOne == 0) return;
10892 APInt KnownZeroRHS, KnownOneRHS;
10893 DAG.computeKnownBits(Op.getOperand(1), KnownZeroRHS, KnownOneRHS, Depth+1);
10894 KnownZero &= KnownZeroRHS;
10895 KnownOne &= KnownOneRHS;
10898 case ISD::INTRINSIC_W_CHAIN: {
10899 ConstantSDNode *CN = cast<ConstantSDNode>(Op->getOperand(1));
10900 Intrinsic::ID IntID = static_cast<Intrinsic::ID>(CN->getZExtValue());
10903 case Intrinsic::arm_ldaex:
10904 case Intrinsic::arm_ldrex: {
10905 EVT VT = cast<MemIntrinsicSDNode>(Op)->getMemoryVT();
10906 unsigned MemBits = VT.getScalarType().getSizeInBits();
10907 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - MemBits);
10915 //===----------------------------------------------------------------------===//
10916 // ARM Inline Assembly Support
10917 //===----------------------------------------------------------------------===//
10919 bool ARMTargetLowering::ExpandInlineAsm(CallInst *CI) const {
10920 // Looking for "rev" which is V6+.
10921 if (!Subtarget->hasV6Ops())
10924 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
10925 std::string AsmStr = IA->getAsmString();
10926 SmallVector<StringRef, 4> AsmPieces;
10927 SplitString(AsmStr, AsmPieces, ";\n");
10929 switch (AsmPieces.size()) {
10930 default: return false;
10932 AsmStr = AsmPieces[0];
10934 SplitString(AsmStr, AsmPieces, " \t,");
10937 if (AsmPieces.size() == 3 &&
10938 AsmPieces[0] == "rev" && AsmPieces[1] == "$0" && AsmPieces[2] == "$1" &&
10939 IA->getConstraintString().compare(0, 4, "=l,l") == 0) {
10940 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
10941 if (Ty && Ty->getBitWidth() == 32)
10942 return IntrinsicLowering::LowerToByteSwap(CI);
10950 /// getConstraintType - Given a constraint letter, return the type of
10951 /// constraint it is for this target.
10952 ARMTargetLowering::ConstraintType
10953 ARMTargetLowering::getConstraintType(StringRef Constraint) const {
10954 if (Constraint.size() == 1) {
10955 switch (Constraint[0]) {
10957 case 'l': return C_RegisterClass;
10958 case 'w': return C_RegisterClass;
10959 case 'h': return C_RegisterClass;
10960 case 'x': return C_RegisterClass;
10961 case 't': return C_RegisterClass;
10962 case 'j': return C_Other; // Constant for movw.
10963 // An address with a single base register. Due to the way we
10964 // currently handle addresses it is the same as an 'r' memory constraint.
10965 case 'Q': return C_Memory;
10967 } else if (Constraint.size() == 2) {
10968 switch (Constraint[0]) {
10970 // All 'U+' constraints are addresses.
10971 case 'U': return C_Memory;
10974 return TargetLowering::getConstraintType(Constraint);
10977 /// Examine constraint type and operand type and determine a weight value.
10978 /// This object must already have been set up with the operand type
10979 /// and the current alternative constraint selected.
10980 TargetLowering::ConstraintWeight
10981 ARMTargetLowering::getSingleConstraintMatchWeight(
10982 AsmOperandInfo &info, const char *constraint) const {
10983 ConstraintWeight weight = CW_Invalid;
10984 Value *CallOperandVal = info.CallOperandVal;
10985 // If we don't have a value, we can't do a match,
10986 // but allow it at the lowest weight.
10987 if (!CallOperandVal)
10989 Type *type = CallOperandVal->getType();
10990 // Look at the constraint type.
10991 switch (*constraint) {
10993 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
10996 if (type->isIntegerTy()) {
10997 if (Subtarget->isThumb())
10998 weight = CW_SpecificReg;
11000 weight = CW_Register;
11004 if (type->isFloatingPointTy())
11005 weight = CW_Register;
11011 typedef std::pair<unsigned, const TargetRegisterClass*> RCPair;
11012 RCPair ARMTargetLowering::getRegForInlineAsmConstraint(
11013 const TargetRegisterInfo *TRI, StringRef Constraint, MVT VT) const {
11014 if (Constraint.size() == 1) {
11015 // GCC ARM Constraint Letters
11016 switch (Constraint[0]) {
11017 case 'l': // Low regs or general regs.
11018 if (Subtarget->isThumb())
11019 return RCPair(0U, &ARM::tGPRRegClass);
11020 return RCPair(0U, &ARM::GPRRegClass);
11021 case 'h': // High regs or no regs.
11022 if (Subtarget->isThumb())
11023 return RCPair(0U, &ARM::hGPRRegClass);
11026 if (Subtarget->isThumb1Only())
11027 return RCPair(0U, &ARM::tGPRRegClass);
11028 return RCPair(0U, &ARM::GPRRegClass);
11030 if (VT == MVT::Other)
11032 if (VT == MVT::f32)
11033 return RCPair(0U, &ARM::SPRRegClass);
11034 if (VT.getSizeInBits() == 64)
11035 return RCPair(0U, &ARM::DPRRegClass);
11036 if (VT.getSizeInBits() == 128)
11037 return RCPair(0U, &ARM::QPRRegClass);
11040 if (VT == MVT::Other)
11042 if (VT == MVT::f32)
11043 return RCPair(0U, &ARM::SPR_8RegClass);
11044 if (VT.getSizeInBits() == 64)
11045 return RCPair(0U, &ARM::DPR_8RegClass);
11046 if (VT.getSizeInBits() == 128)
11047 return RCPair(0U, &ARM::QPR_8RegClass);
11050 if (VT == MVT::f32)
11051 return RCPair(0U, &ARM::SPRRegClass);
11055 if (StringRef("{cc}").equals_lower(Constraint))
11056 return std::make_pair(unsigned(ARM::CPSR), &ARM::CCRRegClass);
11058 return TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
11061 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
11062 /// vector. If it is invalid, don't add anything to Ops.
11063 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
11064 std::string &Constraint,
11065 std::vector<SDValue>&Ops,
11066 SelectionDAG &DAG) const {
11069 // Currently only support length 1 constraints.
11070 if (Constraint.length() != 1) return;
11072 char ConstraintLetter = Constraint[0];
11073 switch (ConstraintLetter) {
11076 case 'I': case 'J': case 'K': case 'L':
11077 case 'M': case 'N': case 'O':
11078 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
11082 int64_t CVal64 = C->getSExtValue();
11083 int CVal = (int) CVal64;
11084 // None of these constraints allow values larger than 32 bits. Check
11085 // that the value fits in an int.
11086 if (CVal != CVal64)
11089 switch (ConstraintLetter) {
11091 // Constant suitable for movw, must be between 0 and
11093 if (Subtarget->hasV6T2Ops())
11094 if (CVal >= 0 && CVal <= 65535)
11098 if (Subtarget->isThumb1Only()) {
11099 // This must be a constant between 0 and 255, for ADD
11101 if (CVal >= 0 && CVal <= 255)
11103 } else if (Subtarget->isThumb2()) {
11104 // A constant that can be used as an immediate value in a
11105 // data-processing instruction.
11106 if (ARM_AM::getT2SOImmVal(CVal) != -1)
11109 // A constant that can be used as an immediate value in a
11110 // data-processing instruction.
11111 if (ARM_AM::getSOImmVal(CVal) != -1)
11117 if (Subtarget->isThumb()) { // FIXME thumb2
11118 // This must be a constant between -255 and -1, for negated ADD
11119 // immediates. This can be used in GCC with an "n" modifier that
11120 // prints the negated value, for use with SUB instructions. It is
11121 // not useful otherwise but is implemented for compatibility.
11122 if (CVal >= -255 && CVal <= -1)
11125 // This must be a constant between -4095 and 4095. It is not clear
11126 // what this constraint is intended for. Implemented for
11127 // compatibility with GCC.
11128 if (CVal >= -4095 && CVal <= 4095)
11134 if (Subtarget->isThumb1Only()) {
11135 // A 32-bit value where only one byte has a nonzero value. Exclude
11136 // zero to match GCC. This constraint is used by GCC internally for
11137 // constants that can be loaded with a move/shift combination.
11138 // It is not useful otherwise but is implemented for compatibility.
11139 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
11141 } else if (Subtarget->isThumb2()) {
11142 // A constant whose bitwise inverse can be used as an immediate
11143 // value in a data-processing instruction. This can be used in GCC
11144 // with a "B" modifier that prints the inverted value, for use with
11145 // BIC and MVN instructions. It is not useful otherwise but is
11146 // implemented for compatibility.
11147 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
11150 // A constant whose bitwise inverse can be used as an immediate
11151 // value in a data-processing instruction. This can be used in GCC
11152 // with a "B" modifier that prints the inverted value, for use with
11153 // BIC and MVN instructions. It is not useful otherwise but is
11154 // implemented for compatibility.
11155 if (ARM_AM::getSOImmVal(~CVal) != -1)
11161 if (Subtarget->isThumb1Only()) {
11162 // This must be a constant between -7 and 7,
11163 // for 3-operand ADD/SUB immediate instructions.
11164 if (CVal >= -7 && CVal < 7)
11166 } else if (Subtarget->isThumb2()) {
11167 // A constant whose negation can be used as an immediate value in a
11168 // data-processing instruction. This can be used in GCC with an "n"
11169 // modifier that prints the negated value, for use with SUB
11170 // instructions. It is not useful otherwise but is implemented for
11172 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
11175 // A constant whose negation can be used as an immediate value in a
11176 // data-processing instruction. This can be used in GCC with an "n"
11177 // modifier that prints the negated value, for use with SUB
11178 // instructions. It is not useful otherwise but is implemented for
11180 if (ARM_AM::getSOImmVal(-CVal) != -1)
11186 if (Subtarget->isThumb()) { // FIXME thumb2
11187 // This must be a multiple of 4 between 0 and 1020, for
11188 // ADD sp + immediate.
11189 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
11192 // A power of two or a constant between 0 and 32. This is used in
11193 // GCC for the shift amount on shifted register operands, but it is
11194 // useful in general for any shift amounts.
11195 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
11201 if (Subtarget->isThumb()) { // FIXME thumb2
11202 // This must be a constant between 0 and 31, for shift amounts.
11203 if (CVal >= 0 && CVal <= 31)
11209 if (Subtarget->isThumb()) { // FIXME thumb2
11210 // This must be a multiple of 4 between -508 and 508, for
11211 // ADD/SUB sp = sp + immediate.
11212 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
11217 Result = DAG.getTargetConstant(CVal, SDLoc(Op), Op.getValueType());
11221 if (Result.getNode()) {
11222 Ops.push_back(Result);
11225 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
11228 static RTLIB::Libcall getDivRemLibcall(
11229 const SDNode *N, MVT::SimpleValueType SVT) {
11230 assert((N->getOpcode() == ISD::SDIVREM || N->getOpcode() == ISD::UDIVREM ||
11231 N->getOpcode() == ISD::SREM || N->getOpcode() == ISD::UREM) &&
11232 "Unhandled Opcode in getDivRemLibcall");
11233 bool isSigned = N->getOpcode() == ISD::SDIVREM ||
11234 N->getOpcode() == ISD::SREM;
11237 default: llvm_unreachable("Unexpected request for libcall!");
11238 case MVT::i8: LC = isSigned ? RTLIB::SDIVREM_I8 : RTLIB::UDIVREM_I8; break;
11239 case MVT::i16: LC = isSigned ? RTLIB::SDIVREM_I16 : RTLIB::UDIVREM_I16; break;
11240 case MVT::i32: LC = isSigned ? RTLIB::SDIVREM_I32 : RTLIB::UDIVREM_I32; break;
11241 case MVT::i64: LC = isSigned ? RTLIB::SDIVREM_I64 : RTLIB::UDIVREM_I64; break;
11246 static TargetLowering::ArgListTy getDivRemArgList(
11247 const SDNode *N, LLVMContext *Context) {
11248 assert((N->getOpcode() == ISD::SDIVREM || N->getOpcode() == ISD::UDIVREM ||
11249 N->getOpcode() == ISD::SREM || N->getOpcode() == ISD::UREM) &&
11250 "Unhandled Opcode in getDivRemArgList");
11251 bool isSigned = N->getOpcode() == ISD::SDIVREM ||
11252 N->getOpcode() == ISD::SREM;
11253 TargetLowering::ArgListTy Args;
11254 TargetLowering::ArgListEntry Entry;
11255 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
11256 EVT ArgVT = N->getOperand(i).getValueType();
11257 Type *ArgTy = ArgVT.getTypeForEVT(*Context);
11258 Entry.Node = N->getOperand(i);
11260 Entry.isSExt = isSigned;
11261 Entry.isZExt = !isSigned;
11262 Args.push_back(Entry);
11267 SDValue ARMTargetLowering::LowerDivRem(SDValue Op, SelectionDAG &DAG) const {
11268 assert((Subtarget->isTargetAEABI() || Subtarget->isTargetAndroid()) &&
11269 "Register-based DivRem lowering only");
11270 unsigned Opcode = Op->getOpcode();
11271 assert((Opcode == ISD::SDIVREM || Opcode == ISD::UDIVREM) &&
11272 "Invalid opcode for Div/Rem lowering");
11273 bool isSigned = (Opcode == ISD::SDIVREM);
11274 EVT VT = Op->getValueType(0);
11275 Type *Ty = VT.getTypeForEVT(*DAG.getContext());
11277 RTLIB::Libcall LC = getDivRemLibcall(Op.getNode(),
11278 VT.getSimpleVT().SimpleTy);
11279 SDValue InChain = DAG.getEntryNode();
11281 TargetLowering::ArgListTy Args = getDivRemArgList(Op.getNode(),
11284 SDValue Callee = DAG.getExternalSymbol(getLibcallName(LC),
11285 getPointerTy(DAG.getDataLayout()));
11287 Type *RetTy = (Type*)StructType::get(Ty, Ty, nullptr);
11290 TargetLowering::CallLoweringInfo CLI(DAG);
11291 CLI.setDebugLoc(dl).setChain(InChain)
11292 .setCallee(getLibcallCallingConv(LC), RetTy, Callee, std::move(Args), 0)
11293 .setInRegister().setSExtResult(isSigned).setZExtResult(!isSigned);
11295 std::pair<SDValue, SDValue> CallInfo = LowerCallTo(CLI);
11296 return CallInfo.first;
11299 // Lowers REM using divmod helpers
11300 // see RTABI section 4.2/4.3
11301 SDValue ARMTargetLowering::LowerREM(SDNode *N, SelectionDAG &DAG) const {
11302 // Build return types (div and rem)
11303 std::vector<Type*> RetTyParams;
11304 Type *RetTyElement;
11306 switch (N->getValueType(0).getSimpleVT().SimpleTy) {
11307 default: llvm_unreachable("Unexpected request for libcall!");
11308 case MVT::i8: RetTyElement = Type::getInt8Ty(*DAG.getContext()); break;
11309 case MVT::i16: RetTyElement = Type::getInt16Ty(*DAG.getContext()); break;
11310 case MVT::i32: RetTyElement = Type::getInt32Ty(*DAG.getContext()); break;
11311 case MVT::i64: RetTyElement = Type::getInt64Ty(*DAG.getContext()); break;
11314 RetTyParams.push_back(RetTyElement);
11315 RetTyParams.push_back(RetTyElement);
11316 ArrayRef<Type*> ret = ArrayRef<Type*>(RetTyParams);
11317 Type *RetTy = StructType::get(*DAG.getContext(), ret);
11319 RTLIB::Libcall LC = getDivRemLibcall(N, N->getValueType(0).getSimpleVT().
11321 SDValue InChain = DAG.getEntryNode();
11322 TargetLowering::ArgListTy Args = getDivRemArgList(N, DAG.getContext());
11323 bool isSigned = N->getOpcode() == ISD::SREM;
11324 SDValue Callee = DAG.getExternalSymbol(getLibcallName(LC),
11325 getPointerTy(DAG.getDataLayout()));
11328 CallLoweringInfo CLI(DAG);
11329 CLI.setChain(InChain)
11330 .setCallee(CallingConv::ARM_AAPCS, RetTy, Callee, std::move(Args), 0)
11331 .setSExtResult(isSigned).setZExtResult(!isSigned).setDebugLoc(SDLoc(N));
11332 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
11334 // Return second (rem) result operand (first contains div)
11335 SDNode *ResNode = CallResult.first.getNode();
11336 assert(ResNode->getNumOperands() == 2 && "divmod should return two operands");
11337 return ResNode->getOperand(1);
11341 ARMTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const {
11342 assert(Subtarget->isTargetWindows() && "unsupported target platform");
11346 SDValue Chain = Op.getOperand(0);
11347 SDValue Size = Op.getOperand(1);
11349 SDValue Words = DAG.getNode(ISD::SRL, DL, MVT::i32, Size,
11350 DAG.getConstant(2, DL, MVT::i32));
11353 Chain = DAG.getCopyToReg(Chain, DL, ARM::R4, Words, Flag);
11354 Flag = Chain.getValue(1);
11356 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
11357 Chain = DAG.getNode(ARMISD::WIN__CHKSTK, DL, NodeTys, Chain, Flag);
11359 SDValue NewSP = DAG.getCopyFromReg(Chain, DL, ARM::SP, MVT::i32);
11360 Chain = NewSP.getValue(1);
11362 SDValue Ops[2] = { NewSP, Chain };
11363 return DAG.getMergeValues(Ops, DL);
11366 SDValue ARMTargetLowering::LowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) const {
11367 assert(Op.getValueType() == MVT::f64 && Subtarget->isFPOnlySP() &&
11368 "Unexpected type for custom-lowering FP_EXTEND");
11371 LC = RTLIB::getFPEXT(Op.getOperand(0).getValueType(), Op.getValueType());
11373 SDValue SrcVal = Op.getOperand(0);
11374 return makeLibCall(DAG, LC, Op.getValueType(), SrcVal, /*isSigned*/ false,
11378 SDValue ARMTargetLowering::LowerFP_ROUND(SDValue Op, SelectionDAG &DAG) const {
11379 assert(Op.getOperand(0).getValueType() == MVT::f64 &&
11380 Subtarget->isFPOnlySP() &&
11381 "Unexpected type for custom-lowering FP_ROUND");
11384 LC = RTLIB::getFPROUND(Op.getOperand(0).getValueType(), Op.getValueType());
11386 SDValue SrcVal = Op.getOperand(0);
11387 return makeLibCall(DAG, LC, Op.getValueType(), SrcVal, /*isSigned*/ false,
11392 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
11393 // The ARM target isn't yet aware of offsets.
11397 bool ARM::isBitFieldInvertedMask(unsigned v) {
11398 if (v == 0xffffffff)
11401 // there can be 1's on either or both "outsides", all the "inside"
11402 // bits must be 0's
11403 return isShiftedMask_32(~v);
11406 /// isFPImmLegal - Returns true if the target can instruction select the
11407 /// specified FP immediate natively. If false, the legalizer will
11408 /// materialize the FP immediate as a load from a constant pool.
11409 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
11410 if (!Subtarget->hasVFP3())
11412 if (VT == MVT::f32)
11413 return ARM_AM::getFP32Imm(Imm) != -1;
11414 if (VT == MVT::f64 && !Subtarget->isFPOnlySP())
11415 return ARM_AM::getFP64Imm(Imm) != -1;
11419 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
11420 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
11421 /// specified in the intrinsic calls.
11422 bool ARMTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
11424 unsigned Intrinsic) const {
11425 switch (Intrinsic) {
11426 case Intrinsic::arm_neon_vld1:
11427 case Intrinsic::arm_neon_vld2:
11428 case Intrinsic::arm_neon_vld3:
11429 case Intrinsic::arm_neon_vld4:
11430 case Intrinsic::arm_neon_vld2lane:
11431 case Intrinsic::arm_neon_vld3lane:
11432 case Intrinsic::arm_neon_vld4lane: {
11433 Info.opc = ISD::INTRINSIC_W_CHAIN;
11434 // Conservatively set memVT to the entire set of vectors loaded.
11435 auto &DL = I.getCalledFunction()->getParent()->getDataLayout();
11436 uint64_t NumElts = DL.getTypeAllocSize(I.getType()) / 8;
11437 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
11438 Info.ptrVal = I.getArgOperand(0);
11440 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
11441 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
11442 Info.vol = false; // volatile loads with NEON intrinsics not supported
11443 Info.readMem = true;
11444 Info.writeMem = false;
11447 case Intrinsic::arm_neon_vst1:
11448 case Intrinsic::arm_neon_vst2:
11449 case Intrinsic::arm_neon_vst3:
11450 case Intrinsic::arm_neon_vst4:
11451 case Intrinsic::arm_neon_vst2lane:
11452 case Intrinsic::arm_neon_vst3lane:
11453 case Intrinsic::arm_neon_vst4lane: {
11454 Info.opc = ISD::INTRINSIC_VOID;
11455 // Conservatively set memVT to the entire set of vectors stored.
11456 auto &DL = I.getCalledFunction()->getParent()->getDataLayout();
11457 unsigned NumElts = 0;
11458 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
11459 Type *ArgTy = I.getArgOperand(ArgI)->getType();
11460 if (!ArgTy->isVectorTy())
11462 NumElts += DL.getTypeAllocSize(ArgTy) / 8;
11464 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
11465 Info.ptrVal = I.getArgOperand(0);
11467 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
11468 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
11469 Info.vol = false; // volatile stores with NEON intrinsics not supported
11470 Info.readMem = false;
11471 Info.writeMem = true;
11474 case Intrinsic::arm_ldaex:
11475 case Intrinsic::arm_ldrex: {
11476 auto &DL = I.getCalledFunction()->getParent()->getDataLayout();
11477 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(0)->getType());
11478 Info.opc = ISD::INTRINSIC_W_CHAIN;
11479 Info.memVT = MVT::getVT(PtrTy->getElementType());
11480 Info.ptrVal = I.getArgOperand(0);
11482 Info.align = DL.getABITypeAlignment(PtrTy->getElementType());
11484 Info.readMem = true;
11485 Info.writeMem = false;
11488 case Intrinsic::arm_stlex:
11489 case Intrinsic::arm_strex: {
11490 auto &DL = I.getCalledFunction()->getParent()->getDataLayout();
11491 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(1)->getType());
11492 Info.opc = ISD::INTRINSIC_W_CHAIN;
11493 Info.memVT = MVT::getVT(PtrTy->getElementType());
11494 Info.ptrVal = I.getArgOperand(1);
11496 Info.align = DL.getABITypeAlignment(PtrTy->getElementType());
11498 Info.readMem = false;
11499 Info.writeMem = true;
11502 case Intrinsic::arm_stlexd:
11503 case Intrinsic::arm_strexd: {
11504 Info.opc = ISD::INTRINSIC_W_CHAIN;
11505 Info.memVT = MVT::i64;
11506 Info.ptrVal = I.getArgOperand(2);
11510 Info.readMem = false;
11511 Info.writeMem = true;
11514 case Intrinsic::arm_ldaexd:
11515 case Intrinsic::arm_ldrexd: {
11516 Info.opc = ISD::INTRINSIC_W_CHAIN;
11517 Info.memVT = MVT::i64;
11518 Info.ptrVal = I.getArgOperand(0);
11522 Info.readMem = true;
11523 Info.writeMem = false;
11533 /// \brief Returns true if it is beneficial to convert a load of a constant
11534 /// to just the constant itself.
11535 bool ARMTargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
11537 assert(Ty->isIntegerTy());
11539 unsigned Bits = Ty->getPrimitiveSizeInBits();
11540 if (Bits == 0 || Bits > 32)
11545 Instruction* ARMTargetLowering::makeDMB(IRBuilder<> &Builder,
11546 ARM_MB::MemBOpt Domain) const {
11547 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
11549 // First, if the target has no DMB, see what fallback we can use.
11550 if (!Subtarget->hasDataBarrier()) {
11551 // Some ARMv6 cpus can support data barriers with an mcr instruction.
11552 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
11554 if (Subtarget->hasV6Ops() && !Subtarget->isThumb()) {
11555 Function *MCR = llvm::Intrinsic::getDeclaration(M, Intrinsic::arm_mcr);
11556 Value* args[6] = {Builder.getInt32(15), Builder.getInt32(0),
11557 Builder.getInt32(0), Builder.getInt32(7),
11558 Builder.getInt32(10), Builder.getInt32(5)};
11559 return Builder.CreateCall(MCR, args);
11561 // Instead of using barriers, atomic accesses on these subtargets use
11563 llvm_unreachable("makeDMB on a target so old that it has no barriers");
11566 Function *DMB = llvm::Intrinsic::getDeclaration(M, Intrinsic::arm_dmb);
11567 // Only a full system barrier exists in the M-class architectures.
11568 Domain = Subtarget->isMClass() ? ARM_MB::SY : Domain;
11569 Constant *CDomain = Builder.getInt32(Domain);
11570 return Builder.CreateCall(DMB, CDomain);
11574 // Based on http://www.cl.cam.ac.uk/~pes20/cpp/cpp0xmappings.html
11575 Instruction* ARMTargetLowering::emitLeadingFence(IRBuilder<> &Builder,
11576 AtomicOrdering Ord, bool IsStore,
11577 bool IsLoad) const {
11578 if (!getInsertFencesForAtomic())
11584 llvm_unreachable("Invalid fence: unordered/non-atomic");
11587 return nullptr; // Nothing to do
11588 case SequentiallyConsistent:
11590 return nullptr; // Nothing to do
11593 case AcquireRelease:
11594 if (Subtarget->isSwift())
11595 return makeDMB(Builder, ARM_MB::ISHST);
11596 // FIXME: add a comment with a link to documentation justifying this.
11598 return makeDMB(Builder, ARM_MB::ISH);
11600 llvm_unreachable("Unknown fence ordering in emitLeadingFence");
11603 Instruction* ARMTargetLowering::emitTrailingFence(IRBuilder<> &Builder,
11604 AtomicOrdering Ord, bool IsStore,
11605 bool IsLoad) const {
11606 if (!getInsertFencesForAtomic())
11612 llvm_unreachable("Invalid fence: unordered/not-atomic");
11615 return nullptr; // Nothing to do
11617 case AcquireRelease:
11618 case SequentiallyConsistent:
11619 return makeDMB(Builder, ARM_MB::ISH);
11621 llvm_unreachable("Unknown fence ordering in emitTrailingFence");
11624 // Loads and stores less than 64-bits are already atomic; ones above that
11625 // are doomed anyway, so defer to the default libcall and blame the OS when
11626 // things go wrong. Cortex M doesn't have ldrexd/strexd though, so don't emit
11627 // anything for those.
11628 bool ARMTargetLowering::shouldExpandAtomicStoreInIR(StoreInst *SI) const {
11629 unsigned Size = SI->getValueOperand()->getType()->getPrimitiveSizeInBits();
11630 return (Size == 64) && !Subtarget->isMClass();
11633 // Loads and stores less than 64-bits are already atomic; ones above that
11634 // are doomed anyway, so defer to the default libcall and blame the OS when
11635 // things go wrong. Cortex M doesn't have ldrexd/strexd though, so don't emit
11636 // anything for those.
11637 // FIXME: ldrd and strd are atomic if the CPU has LPAE (e.g. A15 has that
11638 // guarantee, see DDI0406C ARM architecture reference manual,
11639 // sections A8.8.72-74 LDRD)
11640 TargetLowering::AtomicExpansionKind
11641 ARMTargetLowering::shouldExpandAtomicLoadInIR(LoadInst *LI) const {
11642 unsigned Size = LI->getType()->getPrimitiveSizeInBits();
11643 return ((Size == 64) && !Subtarget->isMClass()) ? AtomicExpansionKind::LLSC
11644 : AtomicExpansionKind::None;
11647 // For the real atomic operations, we have ldrex/strex up to 32 bits,
11648 // and up to 64 bits on the non-M profiles
11649 TargetLowering::AtomicExpansionKind
11650 ARMTargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
11651 unsigned Size = AI->getType()->getPrimitiveSizeInBits();
11652 return (Size <= (Subtarget->isMClass() ? 32U : 64U))
11653 ? AtomicExpansionKind::LLSC
11654 : AtomicExpansionKind::None;
11657 bool ARMTargetLowering::shouldExpandAtomicCmpXchgInIR(
11658 AtomicCmpXchgInst *AI) const {
11662 // This has so far only been implemented for MachO.
11663 bool ARMTargetLowering::useLoadStackGuardNode() const {
11664 return Subtarget->isTargetMachO();
11667 bool ARMTargetLowering::canCombineStoreAndExtract(Type *VectorTy, Value *Idx,
11668 unsigned &Cost) const {
11669 // If we do not have NEON, vector types are not natively supported.
11670 if (!Subtarget->hasNEON())
11673 // Floating point values and vector values map to the same register file.
11674 // Therefore, although we could do a store extract of a vector type, this is
11675 // better to leave at float as we have more freedom in the addressing mode for
11677 if (VectorTy->isFPOrFPVectorTy())
11680 // If the index is unknown at compile time, this is very expensive to lower
11681 // and it is not possible to combine the store with the extract.
11682 if (!isa<ConstantInt>(Idx))
11685 assert(VectorTy->isVectorTy() && "VectorTy is not a vector type");
11686 unsigned BitWidth = cast<VectorType>(VectorTy)->getBitWidth();
11687 // We can do a store + vector extract on any vector that fits perfectly in a D
11689 if (BitWidth == 64 || BitWidth == 128) {
11696 Value *ARMTargetLowering::emitLoadLinked(IRBuilder<> &Builder, Value *Addr,
11697 AtomicOrdering Ord) const {
11698 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
11699 Type *ValTy = cast<PointerType>(Addr->getType())->getElementType();
11700 bool IsAcquire = isAtLeastAcquire(Ord);
11702 // Since i64 isn't legal and intrinsics don't get type-lowered, the ldrexd
11703 // intrinsic must return {i32, i32} and we have to recombine them into a
11704 // single i64 here.
11705 if (ValTy->getPrimitiveSizeInBits() == 64) {
11706 Intrinsic::ID Int =
11707 IsAcquire ? Intrinsic::arm_ldaexd : Intrinsic::arm_ldrexd;
11708 Function *Ldrex = llvm::Intrinsic::getDeclaration(M, Int);
11710 Addr = Builder.CreateBitCast(Addr, Type::getInt8PtrTy(M->getContext()));
11711 Value *LoHi = Builder.CreateCall(Ldrex, Addr, "lohi");
11713 Value *Lo = Builder.CreateExtractValue(LoHi, 0, "lo");
11714 Value *Hi = Builder.CreateExtractValue(LoHi, 1, "hi");
11715 if (!Subtarget->isLittle())
11716 std::swap (Lo, Hi);
11717 Lo = Builder.CreateZExt(Lo, ValTy, "lo64");
11718 Hi = Builder.CreateZExt(Hi, ValTy, "hi64");
11719 return Builder.CreateOr(
11720 Lo, Builder.CreateShl(Hi, ConstantInt::get(ValTy, 32)), "val64");
11723 Type *Tys[] = { Addr->getType() };
11724 Intrinsic::ID Int = IsAcquire ? Intrinsic::arm_ldaex : Intrinsic::arm_ldrex;
11725 Function *Ldrex = llvm::Intrinsic::getDeclaration(M, Int, Tys);
11727 return Builder.CreateTruncOrBitCast(
11728 Builder.CreateCall(Ldrex, Addr),
11729 cast<PointerType>(Addr->getType())->getElementType());
11732 void ARMTargetLowering::emitAtomicCmpXchgNoStoreLLBalance(
11733 IRBuilder<> &Builder) const {
11734 if (!Subtarget->hasV7Ops())
11736 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
11737 Builder.CreateCall(llvm::Intrinsic::getDeclaration(M, Intrinsic::arm_clrex));
11740 Value *ARMTargetLowering::emitStoreConditional(IRBuilder<> &Builder, Value *Val,
11742 AtomicOrdering Ord) const {
11743 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
11744 bool IsRelease = isAtLeastRelease(Ord);
11746 // Since the intrinsics must have legal type, the i64 intrinsics take two
11747 // parameters: "i32, i32". We must marshal Val into the appropriate form
11748 // before the call.
11749 if (Val->getType()->getPrimitiveSizeInBits() == 64) {
11750 Intrinsic::ID Int =
11751 IsRelease ? Intrinsic::arm_stlexd : Intrinsic::arm_strexd;
11752 Function *Strex = Intrinsic::getDeclaration(M, Int);
11753 Type *Int32Ty = Type::getInt32Ty(M->getContext());
11755 Value *Lo = Builder.CreateTrunc(Val, Int32Ty, "lo");
11756 Value *Hi = Builder.CreateTrunc(Builder.CreateLShr(Val, 32), Int32Ty, "hi");
11757 if (!Subtarget->isLittle())
11758 std::swap (Lo, Hi);
11759 Addr = Builder.CreateBitCast(Addr, Type::getInt8PtrTy(M->getContext()));
11760 return Builder.CreateCall(Strex, {Lo, Hi, Addr});
11763 Intrinsic::ID Int = IsRelease ? Intrinsic::arm_stlex : Intrinsic::arm_strex;
11764 Type *Tys[] = { Addr->getType() };
11765 Function *Strex = Intrinsic::getDeclaration(M, Int, Tys);
11767 return Builder.CreateCall(
11768 Strex, {Builder.CreateZExtOrBitCast(
11769 Val, Strex->getFunctionType()->getParamType(0)),
11773 /// \brief Lower an interleaved load into a vldN intrinsic.
11775 /// E.g. Lower an interleaved load (Factor = 2):
11776 /// %wide.vec = load <8 x i32>, <8 x i32>* %ptr, align 4
11777 /// %v0 = shuffle %wide.vec, undef, <0, 2, 4, 6> ; Extract even elements
11778 /// %v1 = shuffle %wide.vec, undef, <1, 3, 5, 7> ; Extract odd elements
11781 /// %vld2 = { <4 x i32>, <4 x i32> } call llvm.arm.neon.vld2(%ptr, 4)
11782 /// %vec0 = extractelement { <4 x i32>, <4 x i32> } %vld2, i32 0
11783 /// %vec1 = extractelement { <4 x i32>, <4 x i32> } %vld2, i32 1
11784 bool ARMTargetLowering::lowerInterleavedLoad(
11785 LoadInst *LI, ArrayRef<ShuffleVectorInst *> Shuffles,
11786 ArrayRef<unsigned> Indices, unsigned Factor) const {
11787 assert(Factor >= 2 && Factor <= getMaxSupportedInterleaveFactor() &&
11788 "Invalid interleave factor");
11789 assert(!Shuffles.empty() && "Empty shufflevector input");
11790 assert(Shuffles.size() == Indices.size() &&
11791 "Unmatched number of shufflevectors and indices");
11793 VectorType *VecTy = Shuffles[0]->getType();
11794 Type *EltTy = VecTy->getVectorElementType();
11796 const DataLayout &DL = LI->getModule()->getDataLayout();
11797 unsigned VecSize = DL.getTypeAllocSizeInBits(VecTy);
11798 bool EltIs64Bits = DL.getTypeAllocSizeInBits(EltTy) == 64;
11800 // Skip if we do not have NEON and skip illegal vector types and vector types
11801 // with i64/f64 elements (vldN doesn't support i64/f64 elements).
11802 if (!Subtarget->hasNEON() || (VecSize != 64 && VecSize != 128) || EltIs64Bits)
11805 // A pointer vector can not be the return type of the ldN intrinsics. Need to
11806 // load integer vectors first and then convert to pointer vectors.
11807 if (EltTy->isPointerTy())
11809 VectorType::get(DL.getIntPtrType(EltTy), VecTy->getVectorNumElements());
11811 static const Intrinsic::ID LoadInts[3] = {Intrinsic::arm_neon_vld2,
11812 Intrinsic::arm_neon_vld3,
11813 Intrinsic::arm_neon_vld4};
11815 IRBuilder<> Builder(LI);
11816 SmallVector<Value *, 2> Ops;
11818 Type *Int8Ptr = Builder.getInt8PtrTy(LI->getPointerAddressSpace());
11819 Ops.push_back(Builder.CreateBitCast(LI->getPointerOperand(), Int8Ptr));
11820 Ops.push_back(Builder.getInt32(LI->getAlignment()));
11822 Type *Tys[] = { VecTy, Int8Ptr };
11823 Function *VldnFunc =
11824 Intrinsic::getDeclaration(LI->getModule(), LoadInts[Factor - 2], Tys);
11825 CallInst *VldN = Builder.CreateCall(VldnFunc, Ops, "vldN");
11827 // Replace uses of each shufflevector with the corresponding vector loaded
11829 for (unsigned i = 0; i < Shuffles.size(); i++) {
11830 ShuffleVectorInst *SV = Shuffles[i];
11831 unsigned Index = Indices[i];
11833 Value *SubVec = Builder.CreateExtractValue(VldN, Index);
11835 // Convert the integer vector to pointer vector if the element is pointer.
11836 if (EltTy->isPointerTy())
11837 SubVec = Builder.CreateIntToPtr(SubVec, SV->getType());
11839 SV->replaceAllUsesWith(SubVec);
11845 /// \brief Get a mask consisting of sequential integers starting from \p Start.
11847 /// I.e. <Start, Start + 1, ..., Start + NumElts - 1>
11848 static Constant *getSequentialMask(IRBuilder<> &Builder, unsigned Start,
11849 unsigned NumElts) {
11850 SmallVector<Constant *, 16> Mask;
11851 for (unsigned i = 0; i < NumElts; i++)
11852 Mask.push_back(Builder.getInt32(Start + i));
11854 return ConstantVector::get(Mask);
11857 /// \brief Lower an interleaved store into a vstN intrinsic.
11859 /// E.g. Lower an interleaved store (Factor = 3):
11860 /// %i.vec = shuffle <8 x i32> %v0, <8 x i32> %v1,
11861 /// <0, 4, 8, 1, 5, 9, 2, 6, 10, 3, 7, 11>
11862 /// store <12 x i32> %i.vec, <12 x i32>* %ptr, align 4
11865 /// %sub.v0 = shuffle <8 x i32> %v0, <8 x i32> v1, <0, 1, 2, 3>
11866 /// %sub.v1 = shuffle <8 x i32> %v0, <8 x i32> v1, <4, 5, 6, 7>
11867 /// %sub.v2 = shuffle <8 x i32> %v0, <8 x i32> v1, <8, 9, 10, 11>
11868 /// call void llvm.arm.neon.vst3(%ptr, %sub.v0, %sub.v1, %sub.v2, 4)
11870 /// Note that the new shufflevectors will be removed and we'll only generate one
11871 /// vst3 instruction in CodeGen.
11872 bool ARMTargetLowering::lowerInterleavedStore(StoreInst *SI,
11873 ShuffleVectorInst *SVI,
11874 unsigned Factor) const {
11875 assert(Factor >= 2 && Factor <= getMaxSupportedInterleaveFactor() &&
11876 "Invalid interleave factor");
11878 VectorType *VecTy = SVI->getType();
11879 assert(VecTy->getVectorNumElements() % Factor == 0 &&
11880 "Invalid interleaved store");
11882 unsigned NumSubElts = VecTy->getVectorNumElements() / Factor;
11883 Type *EltTy = VecTy->getVectorElementType();
11884 VectorType *SubVecTy = VectorType::get(EltTy, NumSubElts);
11886 const DataLayout &DL = SI->getModule()->getDataLayout();
11887 unsigned SubVecSize = DL.getTypeAllocSizeInBits(SubVecTy);
11888 bool EltIs64Bits = DL.getTypeAllocSizeInBits(EltTy) == 64;
11890 // Skip if we do not have NEON and skip illegal vector types and vector types
11891 // with i64/f64 elements (vstN doesn't support i64/f64 elements).
11892 if (!Subtarget->hasNEON() || (SubVecSize != 64 && SubVecSize != 128) ||
11896 Value *Op0 = SVI->getOperand(0);
11897 Value *Op1 = SVI->getOperand(1);
11898 IRBuilder<> Builder(SI);
11900 // StN intrinsics don't support pointer vectors as arguments. Convert pointer
11901 // vectors to integer vectors.
11902 if (EltTy->isPointerTy()) {
11903 Type *IntTy = DL.getIntPtrType(EltTy);
11905 // Convert to the corresponding integer vector.
11907 VectorType::get(IntTy, Op0->getType()->getVectorNumElements());
11908 Op0 = Builder.CreatePtrToInt(Op0, IntVecTy);
11909 Op1 = Builder.CreatePtrToInt(Op1, IntVecTy);
11911 SubVecTy = VectorType::get(IntTy, NumSubElts);
11914 static const Intrinsic::ID StoreInts[3] = {Intrinsic::arm_neon_vst2,
11915 Intrinsic::arm_neon_vst3,
11916 Intrinsic::arm_neon_vst4};
11917 SmallVector<Value *, 6> Ops;
11919 Type *Int8Ptr = Builder.getInt8PtrTy(SI->getPointerAddressSpace());
11920 Ops.push_back(Builder.CreateBitCast(SI->getPointerOperand(), Int8Ptr));
11922 Type *Tys[] = { Int8Ptr, SubVecTy };
11923 Function *VstNFunc = Intrinsic::getDeclaration(
11924 SI->getModule(), StoreInts[Factor - 2], Tys);
11926 // Split the shufflevector operands into sub vectors for the new vstN call.
11927 for (unsigned i = 0; i < Factor; i++)
11928 Ops.push_back(Builder.CreateShuffleVector(
11929 Op0, Op1, getSequentialMask(Builder, NumSubElts * i, NumSubElts)));
11931 Ops.push_back(Builder.getInt32(SI->getAlignment()));
11932 Builder.CreateCall(VstNFunc, Ops);
11944 static bool isHomogeneousAggregate(Type *Ty, HABaseType &Base,
11945 uint64_t &Members) {
11946 if (auto *ST = dyn_cast<StructType>(Ty)) {
11947 for (unsigned i = 0; i < ST->getNumElements(); ++i) {
11948 uint64_t SubMembers = 0;
11949 if (!isHomogeneousAggregate(ST->getElementType(i), Base, SubMembers))
11951 Members += SubMembers;
11953 } else if (auto *AT = dyn_cast<ArrayType>(Ty)) {
11954 uint64_t SubMembers = 0;
11955 if (!isHomogeneousAggregate(AT->getElementType(), Base, SubMembers))
11957 Members += SubMembers * AT->getNumElements();
11958 } else if (Ty->isFloatTy()) {
11959 if (Base != HA_UNKNOWN && Base != HA_FLOAT)
11963 } else if (Ty->isDoubleTy()) {
11964 if (Base != HA_UNKNOWN && Base != HA_DOUBLE)
11968 } else if (auto *VT = dyn_cast<VectorType>(Ty)) {
11975 return VT->getBitWidth() == 64;
11977 return VT->getBitWidth() == 128;
11979 switch (VT->getBitWidth()) {
11992 return (Members > 0 && Members <= 4);
11995 /// \brief Return true if a type is an AAPCS-VFP homogeneous aggregate or one of
11996 /// [N x i32] or [N x i64]. This allows front-ends to skip emitting padding when
11997 /// passing according to AAPCS rules.
11998 bool ARMTargetLowering::functionArgumentNeedsConsecutiveRegisters(
11999 Type *Ty, CallingConv::ID CallConv, bool isVarArg) const {
12000 if (getEffectiveCallingConv(CallConv, isVarArg) !=
12001 CallingConv::ARM_AAPCS_VFP)
12004 HABaseType Base = HA_UNKNOWN;
12005 uint64_t Members = 0;
12006 bool IsHA = isHomogeneousAggregate(Ty, Base, Members);
12007 DEBUG(dbgs() << "isHA: " << IsHA << " "; Ty->dump());
12009 bool IsIntArray = Ty->isArrayTy() && Ty->getArrayElementType()->isIntegerTy();
12010 return IsHA || IsIntArray;