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);
246 // Set the correct calling convention for ARMv7k WatchOS. It's just
247 // AAPCS_VFP for functions as simple as libcalls.
248 if (Subtarget->isTargetWatchOS()) {
249 for (int i = 0; i < RTLIB::UNKNOWN_LIBCALL; ++i)
250 setLibcallCallingConv((RTLIB::Libcall)i, CallingConv::ARM_AAPCS_VFP);
254 // These libcalls are not available in 32-bit.
255 setLibcallName(RTLIB::SHL_I128, nullptr);
256 setLibcallName(RTLIB::SRL_I128, nullptr);
257 setLibcallName(RTLIB::SRA_I128, nullptr);
260 if (Subtarget->isAAPCS_ABI() &&
261 (Subtarget->isTargetAEABI() || Subtarget->isTargetGNUAEABI() ||
262 Subtarget->isTargetAndroid())) {
263 static const struct {
264 const RTLIB::Libcall Op;
265 const char * const Name;
266 const CallingConv::ID CC;
267 const ISD::CondCode Cond;
269 // Double-precision floating-point arithmetic helper functions
270 // RTABI chapter 4.1.2, Table 2
271 { RTLIB::ADD_F64, "__aeabi_dadd", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
272 { RTLIB::DIV_F64, "__aeabi_ddiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
273 { RTLIB::MUL_F64, "__aeabi_dmul", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
274 { RTLIB::SUB_F64, "__aeabi_dsub", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
276 // Double-precision floating-point comparison helper functions
277 // RTABI chapter 4.1.2, Table 3
278 { RTLIB::OEQ_F64, "__aeabi_dcmpeq", CallingConv::ARM_AAPCS, ISD::SETNE },
279 { RTLIB::UNE_F64, "__aeabi_dcmpeq", CallingConv::ARM_AAPCS, ISD::SETEQ },
280 { RTLIB::OLT_F64, "__aeabi_dcmplt", CallingConv::ARM_AAPCS, ISD::SETNE },
281 { RTLIB::OLE_F64, "__aeabi_dcmple", CallingConv::ARM_AAPCS, ISD::SETNE },
282 { RTLIB::OGE_F64, "__aeabi_dcmpge", CallingConv::ARM_AAPCS, ISD::SETNE },
283 { RTLIB::OGT_F64, "__aeabi_dcmpgt", CallingConv::ARM_AAPCS, ISD::SETNE },
284 { RTLIB::UO_F64, "__aeabi_dcmpun", CallingConv::ARM_AAPCS, ISD::SETNE },
285 { RTLIB::O_F64, "__aeabi_dcmpun", CallingConv::ARM_AAPCS, ISD::SETEQ },
287 // Single-precision floating-point arithmetic helper functions
288 // RTABI chapter 4.1.2, Table 4
289 { RTLIB::ADD_F32, "__aeabi_fadd", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
290 { RTLIB::DIV_F32, "__aeabi_fdiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
291 { RTLIB::MUL_F32, "__aeabi_fmul", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
292 { RTLIB::SUB_F32, "__aeabi_fsub", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
294 // Single-precision floating-point comparison helper functions
295 // RTABI chapter 4.1.2, Table 5
296 { RTLIB::OEQ_F32, "__aeabi_fcmpeq", CallingConv::ARM_AAPCS, ISD::SETNE },
297 { RTLIB::UNE_F32, "__aeabi_fcmpeq", CallingConv::ARM_AAPCS, ISD::SETEQ },
298 { RTLIB::OLT_F32, "__aeabi_fcmplt", CallingConv::ARM_AAPCS, ISD::SETNE },
299 { RTLIB::OLE_F32, "__aeabi_fcmple", CallingConv::ARM_AAPCS, ISD::SETNE },
300 { RTLIB::OGE_F32, "__aeabi_fcmpge", CallingConv::ARM_AAPCS, ISD::SETNE },
301 { RTLIB::OGT_F32, "__aeabi_fcmpgt", CallingConv::ARM_AAPCS, ISD::SETNE },
302 { RTLIB::UO_F32, "__aeabi_fcmpun", CallingConv::ARM_AAPCS, ISD::SETNE },
303 { RTLIB::O_F32, "__aeabi_fcmpun", CallingConv::ARM_AAPCS, ISD::SETEQ },
305 // Floating-point to integer conversions.
306 // RTABI chapter 4.1.2, Table 6
307 { RTLIB::FPTOSINT_F64_I32, "__aeabi_d2iz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
308 { RTLIB::FPTOUINT_F64_I32, "__aeabi_d2uiz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
309 { RTLIB::FPTOSINT_F64_I64, "__aeabi_d2lz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
310 { RTLIB::FPTOUINT_F64_I64, "__aeabi_d2ulz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
311 { RTLIB::FPTOSINT_F32_I32, "__aeabi_f2iz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
312 { RTLIB::FPTOUINT_F32_I32, "__aeabi_f2uiz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
313 { RTLIB::FPTOSINT_F32_I64, "__aeabi_f2lz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
314 { RTLIB::FPTOUINT_F32_I64, "__aeabi_f2ulz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
316 // Conversions between floating types.
317 // RTABI chapter 4.1.2, Table 7
318 { RTLIB::FPROUND_F64_F32, "__aeabi_d2f", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
319 { RTLIB::FPROUND_F64_F16, "__aeabi_d2h", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
320 { RTLIB::FPEXT_F32_F64, "__aeabi_f2d", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
322 // Integer to floating-point conversions.
323 // RTABI chapter 4.1.2, Table 8
324 { RTLIB::SINTTOFP_I32_F64, "__aeabi_i2d", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
325 { RTLIB::UINTTOFP_I32_F64, "__aeabi_ui2d", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
326 { RTLIB::SINTTOFP_I64_F64, "__aeabi_l2d", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
327 { RTLIB::UINTTOFP_I64_F64, "__aeabi_ul2d", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
328 { RTLIB::SINTTOFP_I32_F32, "__aeabi_i2f", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
329 { RTLIB::UINTTOFP_I32_F32, "__aeabi_ui2f", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
330 { RTLIB::SINTTOFP_I64_F32, "__aeabi_l2f", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
331 { RTLIB::UINTTOFP_I64_F32, "__aeabi_ul2f", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
333 // Long long helper functions
334 // RTABI chapter 4.2, Table 9
335 { RTLIB::MUL_I64, "__aeabi_lmul", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
336 { RTLIB::SHL_I64, "__aeabi_llsl", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
337 { RTLIB::SRL_I64, "__aeabi_llsr", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
338 { RTLIB::SRA_I64, "__aeabi_lasr", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
340 // Integer division functions
341 // RTABI chapter 4.3.1
342 { RTLIB::SDIV_I8, "__aeabi_idiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
343 { RTLIB::SDIV_I16, "__aeabi_idiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
344 { RTLIB::SDIV_I32, "__aeabi_idiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
345 { RTLIB::SDIV_I64, "__aeabi_ldivmod", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
346 { RTLIB::UDIV_I8, "__aeabi_uidiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
347 { RTLIB::UDIV_I16, "__aeabi_uidiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
348 { RTLIB::UDIV_I32, "__aeabi_uidiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
349 { RTLIB::UDIV_I64, "__aeabi_uldivmod", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
352 for (const auto &LC : LibraryCalls) {
353 setLibcallName(LC.Op, LC.Name);
354 setLibcallCallingConv(LC.Op, LC.CC);
355 if (LC.Cond != ISD::SETCC_INVALID)
356 setCmpLibcallCC(LC.Op, LC.Cond);
359 // EABI dependent RTLIB
360 if (TM.Options.EABIVersion == EABI::EABI4 ||
361 TM.Options.EABIVersion == EABI::EABI5) {
362 static const struct {
363 const RTLIB::Libcall Op;
364 const char *const Name;
365 const CallingConv::ID CC;
366 const ISD::CondCode Cond;
367 } MemOpsLibraryCalls[] = {
369 // RTABI chapter 4.3.4
370 { RTLIB::MEMCPY, "__aeabi_memcpy", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
371 { RTLIB::MEMMOVE, "__aeabi_memmove", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
372 { RTLIB::MEMSET, "__aeabi_memset", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
375 for (const auto &LC : MemOpsLibraryCalls) {
376 setLibcallName(LC.Op, LC.Name);
377 setLibcallCallingConv(LC.Op, LC.CC);
378 if (LC.Cond != ISD::SETCC_INVALID)
379 setCmpLibcallCC(LC.Op, LC.Cond);
384 if (Subtarget->isTargetWindows()) {
385 static const struct {
386 const RTLIB::Libcall Op;
387 const char * const Name;
388 const CallingConv::ID CC;
390 { RTLIB::FPTOSINT_F32_I64, "__stoi64", CallingConv::ARM_AAPCS_VFP },
391 { RTLIB::FPTOSINT_F64_I64, "__dtoi64", CallingConv::ARM_AAPCS_VFP },
392 { RTLIB::FPTOUINT_F32_I64, "__stou64", CallingConv::ARM_AAPCS_VFP },
393 { RTLIB::FPTOUINT_F64_I64, "__dtou64", CallingConv::ARM_AAPCS_VFP },
394 { RTLIB::SINTTOFP_I64_F32, "__i64tos", CallingConv::ARM_AAPCS_VFP },
395 { RTLIB::SINTTOFP_I64_F64, "__i64tod", CallingConv::ARM_AAPCS_VFP },
396 { RTLIB::UINTTOFP_I64_F32, "__u64tos", CallingConv::ARM_AAPCS_VFP },
397 { RTLIB::UINTTOFP_I64_F64, "__u64tod", CallingConv::ARM_AAPCS_VFP },
400 for (const auto &LC : LibraryCalls) {
401 setLibcallName(LC.Op, LC.Name);
402 setLibcallCallingConv(LC.Op, LC.CC);
406 // Use divmod compiler-rt calls for iOS 5.0 and later.
407 if (Subtarget->isTargetWatchOS() ||
408 (Subtarget->isTargetIOS() &&
409 !Subtarget->getTargetTriple().isOSVersionLT(5, 0))) {
410 setLibcallName(RTLIB::SDIVREM_I32, "__divmodsi4");
411 setLibcallName(RTLIB::UDIVREM_I32, "__udivmodsi4");
414 // The half <-> float conversion functions are always soft-float, but are
415 // needed for some targets which use a hard-float calling convention by
417 if (Subtarget->isAAPCS_ABI()) {
418 setLibcallCallingConv(RTLIB::FPROUND_F32_F16, CallingConv::ARM_AAPCS);
419 setLibcallCallingConv(RTLIB::FPROUND_F64_F16, CallingConv::ARM_AAPCS);
420 setLibcallCallingConv(RTLIB::FPEXT_F16_F32, CallingConv::ARM_AAPCS);
422 setLibcallCallingConv(RTLIB::FPROUND_F32_F16, CallingConv::ARM_APCS);
423 setLibcallCallingConv(RTLIB::FPROUND_F64_F16, CallingConv::ARM_APCS);
424 setLibcallCallingConv(RTLIB::FPEXT_F16_F32, CallingConv::ARM_APCS);
427 // In EABI, these functions have an __aeabi_ prefix, but in GNUEABI they have
428 // a __gnu_ prefix (which is the default).
429 if (Subtarget->isTargetAEABI()) {
430 setLibcallName(RTLIB::FPROUND_F32_F16, "__aeabi_f2h");
431 setLibcallName(RTLIB::FPROUND_F64_F16, "__aeabi_d2h");
432 setLibcallName(RTLIB::FPEXT_F16_F32, "__aeabi_h2f");
435 if (Subtarget->isThumb1Only())
436 addRegisterClass(MVT::i32, &ARM::tGPRRegClass);
438 addRegisterClass(MVT::i32, &ARM::GPRRegClass);
439 if (!Subtarget->useSoftFloat() && Subtarget->hasVFP2() &&
440 !Subtarget->isThumb1Only()) {
441 addRegisterClass(MVT::f32, &ARM::SPRRegClass);
442 addRegisterClass(MVT::f64, &ARM::DPRRegClass);
445 for (MVT VT : MVT::vector_valuetypes()) {
446 for (MVT InnerVT : MVT::vector_valuetypes()) {
447 setTruncStoreAction(VT, InnerVT, Expand);
448 setLoadExtAction(ISD::SEXTLOAD, VT, InnerVT, Expand);
449 setLoadExtAction(ISD::ZEXTLOAD, VT, InnerVT, Expand);
450 setLoadExtAction(ISD::EXTLOAD, VT, InnerVT, Expand);
453 setOperationAction(ISD::MULHS, VT, Expand);
454 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
455 setOperationAction(ISD::MULHU, VT, Expand);
456 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
458 setOperationAction(ISD::BSWAP, VT, Expand);
461 setOperationAction(ISD::ConstantFP, MVT::f32, Custom);
462 setOperationAction(ISD::ConstantFP, MVT::f64, Custom);
464 setOperationAction(ISD::READ_REGISTER, MVT::i64, Custom);
465 setOperationAction(ISD::WRITE_REGISTER, MVT::i64, Custom);
467 if (Subtarget->hasNEON()) {
468 addDRTypeForNEON(MVT::v2f32);
469 addDRTypeForNEON(MVT::v8i8);
470 addDRTypeForNEON(MVT::v4i16);
471 addDRTypeForNEON(MVT::v2i32);
472 addDRTypeForNEON(MVT::v1i64);
474 addQRTypeForNEON(MVT::v4f32);
475 addQRTypeForNEON(MVT::v2f64);
476 addQRTypeForNEON(MVT::v16i8);
477 addQRTypeForNEON(MVT::v8i16);
478 addQRTypeForNEON(MVT::v4i32);
479 addQRTypeForNEON(MVT::v2i64);
481 // v2f64 is legal so that QR subregs can be extracted as f64 elements, but
482 // neither Neon nor VFP support any arithmetic operations on it.
483 // The same with v4f32. But keep in mind that vadd, vsub, vmul are natively
484 // supported for v4f32.
485 setOperationAction(ISD::FADD, MVT::v2f64, Expand);
486 setOperationAction(ISD::FSUB, MVT::v2f64, Expand);
487 setOperationAction(ISD::FMUL, MVT::v2f64, Expand);
488 // FIXME: Code duplication: FDIV and FREM are expanded always, see
489 // ARMTargetLowering::addTypeForNEON method for details.
490 setOperationAction(ISD::FDIV, MVT::v2f64, Expand);
491 setOperationAction(ISD::FREM, MVT::v2f64, Expand);
492 // FIXME: Create unittest.
493 // In another words, find a way when "copysign" appears in DAG with vector
495 setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand);
496 // FIXME: Code duplication: SETCC has custom operation action, see
497 // ARMTargetLowering::addTypeForNEON method for details.
498 setOperationAction(ISD::SETCC, MVT::v2f64, Expand);
499 // FIXME: Create unittest for FNEG and for FABS.
500 setOperationAction(ISD::FNEG, MVT::v2f64, Expand);
501 setOperationAction(ISD::FABS, MVT::v2f64, Expand);
502 setOperationAction(ISD::FSQRT, MVT::v2f64, Expand);
503 setOperationAction(ISD::FSIN, MVT::v2f64, Expand);
504 setOperationAction(ISD::FCOS, MVT::v2f64, Expand);
505 setOperationAction(ISD::FPOWI, MVT::v2f64, Expand);
506 setOperationAction(ISD::FPOW, MVT::v2f64, Expand);
507 setOperationAction(ISD::FLOG, MVT::v2f64, Expand);
508 setOperationAction(ISD::FLOG2, MVT::v2f64, Expand);
509 setOperationAction(ISD::FLOG10, MVT::v2f64, Expand);
510 setOperationAction(ISD::FEXP, MVT::v2f64, Expand);
511 setOperationAction(ISD::FEXP2, MVT::v2f64, Expand);
512 // FIXME: Create unittest for FCEIL, FTRUNC, FRINT, FNEARBYINT, FFLOOR.
513 setOperationAction(ISD::FCEIL, MVT::v2f64, Expand);
514 setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand);
515 setOperationAction(ISD::FRINT, MVT::v2f64, Expand);
516 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand);
517 setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand);
518 setOperationAction(ISD::FMA, MVT::v2f64, Expand);
520 setOperationAction(ISD::FSQRT, MVT::v4f32, Expand);
521 setOperationAction(ISD::FSIN, MVT::v4f32, Expand);
522 setOperationAction(ISD::FCOS, MVT::v4f32, Expand);
523 setOperationAction(ISD::FPOWI, MVT::v4f32, Expand);
524 setOperationAction(ISD::FPOW, MVT::v4f32, Expand);
525 setOperationAction(ISD::FLOG, MVT::v4f32, Expand);
526 setOperationAction(ISD::FLOG2, MVT::v4f32, Expand);
527 setOperationAction(ISD::FLOG10, MVT::v4f32, Expand);
528 setOperationAction(ISD::FEXP, MVT::v4f32, Expand);
529 setOperationAction(ISD::FEXP2, MVT::v4f32, Expand);
530 setOperationAction(ISD::FCEIL, MVT::v4f32, Expand);
531 setOperationAction(ISD::FTRUNC, MVT::v4f32, Expand);
532 setOperationAction(ISD::FRINT, MVT::v4f32, Expand);
533 setOperationAction(ISD::FNEARBYINT, MVT::v4f32, Expand);
534 setOperationAction(ISD::FFLOOR, MVT::v4f32, Expand);
536 // Mark v2f32 intrinsics.
537 setOperationAction(ISD::FSQRT, MVT::v2f32, Expand);
538 setOperationAction(ISD::FSIN, MVT::v2f32, Expand);
539 setOperationAction(ISD::FCOS, MVT::v2f32, Expand);
540 setOperationAction(ISD::FPOWI, MVT::v2f32, Expand);
541 setOperationAction(ISD::FPOW, MVT::v2f32, Expand);
542 setOperationAction(ISD::FLOG, MVT::v2f32, Expand);
543 setOperationAction(ISD::FLOG2, MVT::v2f32, Expand);
544 setOperationAction(ISD::FLOG10, MVT::v2f32, Expand);
545 setOperationAction(ISD::FEXP, MVT::v2f32, Expand);
546 setOperationAction(ISD::FEXP2, MVT::v2f32, Expand);
547 setOperationAction(ISD::FCEIL, MVT::v2f32, Expand);
548 setOperationAction(ISD::FTRUNC, MVT::v2f32, Expand);
549 setOperationAction(ISD::FRINT, MVT::v2f32, Expand);
550 setOperationAction(ISD::FNEARBYINT, MVT::v2f32, Expand);
551 setOperationAction(ISD::FFLOOR, MVT::v2f32, Expand);
553 // Neon does not support some operations on v1i64 and v2i64 types.
554 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
555 // Custom handling for some quad-vector types to detect VMULL.
556 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
557 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
558 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
559 // Custom handling for some vector types to avoid expensive expansions
560 setOperationAction(ISD::SDIV, MVT::v4i16, Custom);
561 setOperationAction(ISD::SDIV, MVT::v8i8, Custom);
562 setOperationAction(ISD::UDIV, MVT::v4i16, Custom);
563 setOperationAction(ISD::UDIV, MVT::v8i8, Custom);
564 setOperationAction(ISD::SETCC, MVT::v1i64, Expand);
565 setOperationAction(ISD::SETCC, MVT::v2i64, Expand);
566 // Neon does not have single instruction SINT_TO_FP and UINT_TO_FP with
567 // a destination type that is wider than the source, and nor does
568 // it have a FP_TO_[SU]INT instruction with a narrower destination than
570 setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Custom);
571 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
572 setOperationAction(ISD::FP_TO_UINT, MVT::v4i16, Custom);
573 setOperationAction(ISD::FP_TO_SINT, MVT::v4i16, Custom);
575 setOperationAction(ISD::FP_ROUND, MVT::v2f32, Expand);
576 setOperationAction(ISD::FP_EXTEND, MVT::v2f64, Expand);
578 // NEON does not have single instruction CTPOP for vectors with element
579 // types wider than 8-bits. However, custom lowering can leverage the
580 // v8i8/v16i8 vcnt instruction.
581 setOperationAction(ISD::CTPOP, MVT::v2i32, Custom);
582 setOperationAction(ISD::CTPOP, MVT::v4i32, Custom);
583 setOperationAction(ISD::CTPOP, MVT::v4i16, Custom);
584 setOperationAction(ISD::CTPOP, MVT::v8i16, Custom);
586 // NEON does not have single instruction CTTZ for vectors.
587 setOperationAction(ISD::CTTZ, MVT::v8i8, Custom);
588 setOperationAction(ISD::CTTZ, MVT::v4i16, Custom);
589 setOperationAction(ISD::CTTZ, MVT::v2i32, Custom);
590 setOperationAction(ISD::CTTZ, MVT::v1i64, Custom);
592 setOperationAction(ISD::CTTZ, MVT::v16i8, Custom);
593 setOperationAction(ISD::CTTZ, MVT::v8i16, Custom);
594 setOperationAction(ISD::CTTZ, MVT::v4i32, Custom);
595 setOperationAction(ISD::CTTZ, MVT::v2i64, Custom);
597 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v8i8, Custom);
598 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v4i16, Custom);
599 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v2i32, Custom);
600 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v1i64, Custom);
602 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v16i8, Custom);
603 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v8i16, Custom);
604 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v4i32, Custom);
605 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v2i64, Custom);
607 // NEON only has FMA instructions as of VFP4.
608 if (!Subtarget->hasVFP4()) {
609 setOperationAction(ISD::FMA, MVT::v2f32, Expand);
610 setOperationAction(ISD::FMA, MVT::v4f32, Expand);
613 setTargetDAGCombine(ISD::INTRINSIC_VOID);
614 setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN);
615 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
616 setTargetDAGCombine(ISD::SHL);
617 setTargetDAGCombine(ISD::SRL);
618 setTargetDAGCombine(ISD::SRA);
619 setTargetDAGCombine(ISD::SIGN_EXTEND);
620 setTargetDAGCombine(ISD::ZERO_EXTEND);
621 setTargetDAGCombine(ISD::ANY_EXTEND);
622 setTargetDAGCombine(ISD::BUILD_VECTOR);
623 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
624 setTargetDAGCombine(ISD::INSERT_VECTOR_ELT);
625 setTargetDAGCombine(ISD::STORE);
626 setTargetDAGCombine(ISD::FP_TO_SINT);
627 setTargetDAGCombine(ISD::FP_TO_UINT);
628 setTargetDAGCombine(ISD::FDIV);
629 setTargetDAGCombine(ISD::LOAD);
631 // It is legal to extload from v4i8 to v4i16 or v4i32.
632 for (MVT Ty : {MVT::v8i8, MVT::v4i8, MVT::v2i8, MVT::v4i16, MVT::v2i16,
634 for (MVT VT : MVT::integer_vector_valuetypes()) {
635 setLoadExtAction(ISD::EXTLOAD, VT, Ty, Legal);
636 setLoadExtAction(ISD::ZEXTLOAD, VT, Ty, Legal);
637 setLoadExtAction(ISD::SEXTLOAD, VT, Ty, Legal);
642 // ARM and Thumb2 support UMLAL/SMLAL.
643 if (!Subtarget->isThumb1Only())
644 setTargetDAGCombine(ISD::ADDC);
646 if (Subtarget->isFPOnlySP()) {
647 // When targeting a floating-point unit with only single-precision
648 // operations, f64 is legal for the few double-precision instructions which
649 // are present However, no double-precision operations other than moves,
650 // loads and stores are provided by the hardware.
651 setOperationAction(ISD::FADD, MVT::f64, Expand);
652 setOperationAction(ISD::FSUB, MVT::f64, Expand);
653 setOperationAction(ISD::FMUL, MVT::f64, Expand);
654 setOperationAction(ISD::FMA, MVT::f64, Expand);
655 setOperationAction(ISD::FDIV, MVT::f64, Expand);
656 setOperationAction(ISD::FREM, MVT::f64, Expand);
657 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
658 setOperationAction(ISD::FGETSIGN, MVT::f64, Expand);
659 setOperationAction(ISD::FNEG, MVT::f64, Expand);
660 setOperationAction(ISD::FABS, MVT::f64, Expand);
661 setOperationAction(ISD::FSQRT, MVT::f64, Expand);
662 setOperationAction(ISD::FSIN, MVT::f64, Expand);
663 setOperationAction(ISD::FCOS, MVT::f64, Expand);
664 setOperationAction(ISD::FPOWI, MVT::f64, Expand);
665 setOperationAction(ISD::FPOW, MVT::f64, Expand);
666 setOperationAction(ISD::FLOG, MVT::f64, Expand);
667 setOperationAction(ISD::FLOG2, MVT::f64, Expand);
668 setOperationAction(ISD::FLOG10, MVT::f64, Expand);
669 setOperationAction(ISD::FEXP, MVT::f64, Expand);
670 setOperationAction(ISD::FEXP2, MVT::f64, Expand);
671 setOperationAction(ISD::FCEIL, MVT::f64, Expand);
672 setOperationAction(ISD::FTRUNC, MVT::f64, Expand);
673 setOperationAction(ISD::FRINT, MVT::f64, Expand);
674 setOperationAction(ISD::FNEARBYINT, MVT::f64, Expand);
675 setOperationAction(ISD::FFLOOR, MVT::f64, Expand);
676 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
677 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
678 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
679 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
680 setOperationAction(ISD::FP_TO_SINT, MVT::f64, Custom);
681 setOperationAction(ISD::FP_TO_UINT, MVT::f64, Custom);
682 setOperationAction(ISD::FP_ROUND, MVT::f32, Custom);
683 setOperationAction(ISD::FP_EXTEND, MVT::f64, Custom);
686 computeRegisterProperties(Subtarget->getRegisterInfo());
688 // ARM does not have floating-point extending loads.
689 for (MVT VT : MVT::fp_valuetypes()) {
690 setLoadExtAction(ISD::EXTLOAD, VT, MVT::f32, Expand);
691 setLoadExtAction(ISD::EXTLOAD, VT, MVT::f16, Expand);
694 // ... or truncating stores
695 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
696 setTruncStoreAction(MVT::f32, MVT::f16, Expand);
697 setTruncStoreAction(MVT::f64, MVT::f16, Expand);
699 // ARM does not have i1 sign extending load.
700 for (MVT VT : MVT::integer_valuetypes())
701 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Promote);
703 // ARM supports all 4 flavors of integer indexed load / store.
704 if (!Subtarget->isThumb1Only()) {
705 for (unsigned im = (unsigned)ISD::PRE_INC;
706 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
707 setIndexedLoadAction(im, MVT::i1, Legal);
708 setIndexedLoadAction(im, MVT::i8, Legal);
709 setIndexedLoadAction(im, MVT::i16, Legal);
710 setIndexedLoadAction(im, MVT::i32, Legal);
711 setIndexedStoreAction(im, MVT::i1, Legal);
712 setIndexedStoreAction(im, MVT::i8, Legal);
713 setIndexedStoreAction(im, MVT::i16, Legal);
714 setIndexedStoreAction(im, MVT::i32, Legal);
718 setOperationAction(ISD::SADDO, MVT::i32, Custom);
719 setOperationAction(ISD::UADDO, MVT::i32, Custom);
720 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
721 setOperationAction(ISD::USUBO, MVT::i32, Custom);
723 // i64 operation support.
724 setOperationAction(ISD::MUL, MVT::i64, Expand);
725 setOperationAction(ISD::MULHU, MVT::i32, Expand);
726 if (Subtarget->isThumb1Only()) {
727 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
728 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
730 if (Subtarget->isThumb1Only() || !Subtarget->hasV6Ops()
731 || (Subtarget->isThumb2() && !Subtarget->hasDSP()))
732 setOperationAction(ISD::MULHS, MVT::i32, Expand);
734 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
735 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
736 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
737 setOperationAction(ISD::SRL, MVT::i64, Custom);
738 setOperationAction(ISD::SRA, MVT::i64, Custom);
740 if (!Subtarget->isThumb1Only()) {
741 // FIXME: We should do this for Thumb1 as well.
742 setOperationAction(ISD::ADDC, MVT::i32, Custom);
743 setOperationAction(ISD::ADDE, MVT::i32, Custom);
744 setOperationAction(ISD::SUBC, MVT::i32, Custom);
745 setOperationAction(ISD::SUBE, MVT::i32, Custom);
748 // ARM does not have ROTL.
749 setOperationAction(ISD::ROTL, MVT::i32, Expand);
750 for (MVT VT : MVT::vector_valuetypes()) {
751 setOperationAction(ISD::ROTL, VT, Expand);
752 setOperationAction(ISD::ROTR, VT, Expand);
754 setOperationAction(ISD::CTTZ, MVT::i32, Custom);
755 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
756 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
757 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
759 // These just redirect to CTTZ and CTLZ on ARM.
760 setOperationAction(ISD::CTTZ_ZERO_UNDEF , MVT::i32 , Expand);
761 setOperationAction(ISD::CTLZ_ZERO_UNDEF , MVT::i32 , Expand);
763 // @llvm.readcyclecounter requires the Performance Monitors extension.
764 // Default to the 0 expansion on unsupported platforms.
765 // FIXME: Technically there are older ARM CPUs that have
766 // implementation-specific ways of obtaining this information.
767 if (Subtarget->hasPerfMon())
768 setOperationAction(ISD::READCYCLECOUNTER, MVT::i64, Custom);
770 // Only ARMv6 has BSWAP.
771 if (!Subtarget->hasV6Ops())
772 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
774 if (!(Subtarget->hasDivide() && Subtarget->isThumb2()) &&
775 !(Subtarget->hasDivideInARMMode() && !Subtarget->isThumb())) {
776 // These are expanded into libcalls if the cpu doesn't have HW divider.
777 setOperationAction(ISD::SDIV, MVT::i32, LibCall);
778 setOperationAction(ISD::UDIV, MVT::i32, LibCall);
781 if (Subtarget->isTargetWindows() && !Subtarget->hasDivide()) {
782 setOperationAction(ISD::SDIV, MVT::i32, Custom);
783 setOperationAction(ISD::UDIV, MVT::i32, Custom);
785 setOperationAction(ISD::SDIV, MVT::i64, Custom);
786 setOperationAction(ISD::UDIV, MVT::i64, Custom);
789 setOperationAction(ISD::SREM, MVT::i32, Expand);
790 setOperationAction(ISD::UREM, MVT::i32, Expand);
791 // Register based DivRem for AEABI (RTABI 4.2)
792 if (Subtarget->isTargetAEABI() || Subtarget->isTargetAndroid()) {
793 setOperationAction(ISD::SREM, MVT::i64, Custom);
794 setOperationAction(ISD::UREM, MVT::i64, Custom);
796 setLibcallName(RTLIB::SDIVREM_I8, "__aeabi_idivmod");
797 setLibcallName(RTLIB::SDIVREM_I16, "__aeabi_idivmod");
798 setLibcallName(RTLIB::SDIVREM_I32, "__aeabi_idivmod");
799 setLibcallName(RTLIB::SDIVREM_I64, "__aeabi_ldivmod");
800 setLibcallName(RTLIB::UDIVREM_I8, "__aeabi_uidivmod");
801 setLibcallName(RTLIB::UDIVREM_I16, "__aeabi_uidivmod");
802 setLibcallName(RTLIB::UDIVREM_I32, "__aeabi_uidivmod");
803 setLibcallName(RTLIB::UDIVREM_I64, "__aeabi_uldivmod");
805 setLibcallCallingConv(RTLIB::SDIVREM_I8, CallingConv::ARM_AAPCS);
806 setLibcallCallingConv(RTLIB::SDIVREM_I16, CallingConv::ARM_AAPCS);
807 setLibcallCallingConv(RTLIB::SDIVREM_I32, CallingConv::ARM_AAPCS);
808 setLibcallCallingConv(RTLIB::SDIVREM_I64, CallingConv::ARM_AAPCS);
809 setLibcallCallingConv(RTLIB::UDIVREM_I8, CallingConv::ARM_AAPCS);
810 setLibcallCallingConv(RTLIB::UDIVREM_I16, CallingConv::ARM_AAPCS);
811 setLibcallCallingConv(RTLIB::UDIVREM_I32, CallingConv::ARM_AAPCS);
812 setLibcallCallingConv(RTLIB::UDIVREM_I64, CallingConv::ARM_AAPCS);
814 setOperationAction(ISD::SDIVREM, MVT::i32, Custom);
815 setOperationAction(ISD::UDIVREM, MVT::i32, Custom);
817 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
818 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
821 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
822 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
823 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
824 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
826 setOperationAction(ISD::TRAP, MVT::Other, Legal);
828 // Use the default implementation.
829 setOperationAction(ISD::VASTART, MVT::Other, Custom);
830 setOperationAction(ISD::VAARG, MVT::Other, Expand);
831 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
832 setOperationAction(ISD::VAEND, MVT::Other, Expand);
833 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
834 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
836 if (Subtarget->getTargetTriple().isWindowsItaniumEnvironment())
837 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
839 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
841 // ARMv6 Thumb1 (except for CPUs that support dmb / dsb) and earlier use
842 // the default expansion. If we are targeting a single threaded system,
843 // then set them all for expand so we can lower them later into their
845 if (TM.Options.ThreadModel == ThreadModel::Single)
846 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Expand);
847 else if (Subtarget->hasAnyDataBarrier() && !Subtarget->isThumb1Only()) {
848 // ATOMIC_FENCE needs custom lowering; the others should have been expanded
849 // to ldrex/strex loops already.
850 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Custom);
852 // On v8, we have particularly efficient implementations of atomic fences
853 // if they can be combined with nearby atomic loads and stores.
854 if (!Subtarget->hasV8Ops()) {
855 // Automatically insert fences (dmb ish) around ATOMIC_SWAP etc.
856 setInsertFencesForAtomic(true);
859 // If there's anything we can use as a barrier, go through custom lowering
861 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other,
862 Subtarget->hasAnyDataBarrier() ? Custom : Expand);
864 // Set them all for expansion, which will force libcalls.
865 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Expand);
866 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Expand);
867 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Expand);
868 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Expand);
869 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Expand);
870 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Expand);
871 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Expand);
872 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Expand);
873 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i32, Expand);
874 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i32, Expand);
875 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i32, Expand);
876 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i32, Expand);
877 // Mark ATOMIC_LOAD and ATOMIC_STORE custom so we can handle the
878 // Unordered/Monotonic case.
879 setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Custom);
880 setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Custom);
883 setOperationAction(ISD::PREFETCH, MVT::Other, Custom);
885 // Requires SXTB/SXTH, available on v6 and up in both ARM and Thumb modes.
886 if (!Subtarget->hasV6Ops()) {
887 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
888 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
890 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
892 if (!Subtarget->useSoftFloat() && Subtarget->hasVFP2() &&
893 !Subtarget->isThumb1Only()) {
894 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
895 // iff target supports vfp2.
896 setOperationAction(ISD::BITCAST, MVT::i64, Custom);
897 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
900 // We want to custom lower some of our intrinsics.
901 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
902 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
903 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
904 setOperationAction(ISD::EH_SJLJ_SETUP_DISPATCH, MVT::Other, Custom);
905 if (Subtarget->useSjLjEH())
906 setLibcallName(RTLIB::UNWIND_RESUME, "_Unwind_SjLj_Resume");
908 setOperationAction(ISD::SETCC, MVT::i32, Expand);
909 setOperationAction(ISD::SETCC, MVT::f32, Expand);
910 setOperationAction(ISD::SETCC, MVT::f64, Expand);
911 setOperationAction(ISD::SELECT, MVT::i32, Custom);
912 setOperationAction(ISD::SELECT, MVT::f32, Custom);
913 setOperationAction(ISD::SELECT, MVT::f64, Custom);
914 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
915 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
916 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
918 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
919 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
920 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
921 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
922 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
924 // We don't support sin/cos/fmod/copysign/pow
925 setOperationAction(ISD::FSIN, MVT::f64, Expand);
926 setOperationAction(ISD::FSIN, MVT::f32, Expand);
927 setOperationAction(ISD::FCOS, MVT::f32, Expand);
928 setOperationAction(ISD::FCOS, MVT::f64, Expand);
929 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
930 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
931 setOperationAction(ISD::FREM, MVT::f64, Expand);
932 setOperationAction(ISD::FREM, MVT::f32, Expand);
933 if (!Subtarget->useSoftFloat() && Subtarget->hasVFP2() &&
934 !Subtarget->isThumb1Only()) {
935 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
936 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
938 setOperationAction(ISD::FPOW, MVT::f64, Expand);
939 setOperationAction(ISD::FPOW, MVT::f32, Expand);
941 if (!Subtarget->hasVFP4()) {
942 setOperationAction(ISD::FMA, MVT::f64, Expand);
943 setOperationAction(ISD::FMA, MVT::f32, Expand);
946 // Various VFP goodness
947 if (!Subtarget->useSoftFloat() && !Subtarget->isThumb1Only()) {
948 // FP-ARMv8 adds f64 <-> f16 conversion. Before that it should be expanded.
949 if (!Subtarget->hasFPARMv8() || Subtarget->isFPOnlySP()) {
950 setOperationAction(ISD::FP16_TO_FP, MVT::f64, Expand);
951 setOperationAction(ISD::FP_TO_FP16, MVT::f64, Expand);
954 // fp16 is a special v7 extension that adds f16 <-> f32 conversions.
955 if (!Subtarget->hasFP16()) {
956 setOperationAction(ISD::FP16_TO_FP, MVT::f32, Expand);
957 setOperationAction(ISD::FP_TO_FP16, MVT::f32, Expand);
961 // Combine sin / cos into one node or libcall if possible.
962 if (Subtarget->hasSinCos()) {
963 setLibcallName(RTLIB::SINCOS_F32, "sincosf");
964 setLibcallName(RTLIB::SINCOS_F64, "sincos");
965 if (Subtarget->isTargetWatchOS()) {
966 setLibcallCallingConv(RTLIB::SINCOS_F32, CallingConv::ARM_AAPCS_VFP);
967 setLibcallCallingConv(RTLIB::SINCOS_F64, CallingConv::ARM_AAPCS_VFP);
969 if (Subtarget->isTargetIOS() || Subtarget->isTargetWatchOS()) {
970 // For iOS, we don't want to the normal expansion of a libcall to
971 // sincos. We want to issue a libcall to __sincos_stret.
972 setOperationAction(ISD::FSINCOS, MVT::f64, Custom);
973 setOperationAction(ISD::FSINCOS, MVT::f32, Custom);
977 // FP-ARMv8 implements a lot of rounding-like FP operations.
978 if (Subtarget->hasFPARMv8()) {
979 setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
980 setOperationAction(ISD::FCEIL, MVT::f32, Legal);
981 setOperationAction(ISD::FROUND, MVT::f32, Legal);
982 setOperationAction(ISD::FTRUNC, MVT::f32, Legal);
983 setOperationAction(ISD::FNEARBYINT, MVT::f32, Legal);
984 setOperationAction(ISD::FRINT, MVT::f32, Legal);
985 setOperationAction(ISD::FMINNUM, MVT::f32, Legal);
986 setOperationAction(ISD::FMAXNUM, MVT::f32, Legal);
987 setOperationAction(ISD::FMINNUM, MVT::v2f32, Legal);
988 setOperationAction(ISD::FMAXNUM, MVT::v2f32, Legal);
989 setOperationAction(ISD::FMINNUM, MVT::v4f32, Legal);
990 setOperationAction(ISD::FMAXNUM, MVT::v4f32, Legal);
992 if (!Subtarget->isFPOnlySP()) {
993 setOperationAction(ISD::FFLOOR, MVT::f64, Legal);
994 setOperationAction(ISD::FCEIL, MVT::f64, Legal);
995 setOperationAction(ISD::FROUND, MVT::f64, Legal);
996 setOperationAction(ISD::FTRUNC, MVT::f64, Legal);
997 setOperationAction(ISD::FNEARBYINT, MVT::f64, Legal);
998 setOperationAction(ISD::FRINT, MVT::f64, Legal);
999 setOperationAction(ISD::FMINNUM, MVT::f64, Legal);
1000 setOperationAction(ISD::FMAXNUM, MVT::f64, Legal);
1004 if (Subtarget->hasNEON()) {
1005 // vmin and vmax aren't available in a scalar form, so we use
1006 // a NEON instruction with an undef lane instead.
1007 setOperationAction(ISD::FMINNAN, MVT::f32, Legal);
1008 setOperationAction(ISD::FMAXNAN, MVT::f32, Legal);
1009 setOperationAction(ISD::FMINNAN, MVT::v2f32, Legal);
1010 setOperationAction(ISD::FMAXNAN, MVT::v2f32, Legal);
1011 setOperationAction(ISD::FMINNAN, MVT::v4f32, Legal);
1012 setOperationAction(ISD::FMAXNAN, MVT::v4f32, Legal);
1015 // We have target-specific dag combine patterns for the following nodes:
1016 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
1017 setTargetDAGCombine(ISD::ADD);
1018 setTargetDAGCombine(ISD::SUB);
1019 setTargetDAGCombine(ISD::MUL);
1020 setTargetDAGCombine(ISD::AND);
1021 setTargetDAGCombine(ISD::OR);
1022 setTargetDAGCombine(ISD::XOR);
1024 if (Subtarget->hasV6Ops())
1025 setTargetDAGCombine(ISD::SRL);
1027 setStackPointerRegisterToSaveRestore(ARM::SP);
1029 if (Subtarget->useSoftFloat() || Subtarget->isThumb1Only() ||
1030 !Subtarget->hasVFP2())
1031 setSchedulingPreference(Sched::RegPressure);
1033 setSchedulingPreference(Sched::Hybrid);
1035 //// temporary - rewrite interface to use type
1036 MaxStoresPerMemset = 8;
1037 MaxStoresPerMemsetOptSize = 4;
1038 MaxStoresPerMemcpy = 4; // For @llvm.memcpy -> sequence of stores
1039 MaxStoresPerMemcpyOptSize = 2;
1040 MaxStoresPerMemmove = 4; // For @llvm.memmove -> sequence of stores
1041 MaxStoresPerMemmoveOptSize = 2;
1043 // On ARM arguments smaller than 4 bytes are extended, so all arguments
1044 // are at least 4 bytes aligned.
1045 setMinStackArgumentAlignment(4);
1047 // Prefer likely predicted branches to selects on out-of-order cores.
1048 PredictableSelectIsExpensive = Subtarget->isLikeA9();
1050 setMinFunctionAlignment(Subtarget->isThumb() ? 1 : 2);
1053 bool ARMTargetLowering::useSoftFloat() const {
1054 return Subtarget->useSoftFloat();
1057 // FIXME: It might make sense to define the representative register class as the
1058 // nearest super-register that has a non-null superset. For example, DPR_VFP2 is
1059 // a super-register of SPR, and DPR is a superset if DPR_VFP2. Consequently,
1060 // SPR's representative would be DPR_VFP2. This should work well if register
1061 // pressure tracking were modified such that a register use would increment the
1062 // pressure of the register class's representative and all of it's super
1063 // classes' representatives transitively. We have not implemented this because
1064 // of the difficulty prior to coalescing of modeling operand register classes
1065 // due to the common occurrence of cross class copies and subregister insertions
1067 std::pair<const TargetRegisterClass *, uint8_t>
1068 ARMTargetLowering::findRepresentativeClass(const TargetRegisterInfo *TRI,
1070 const TargetRegisterClass *RRC = nullptr;
1072 switch (VT.SimpleTy) {
1074 return TargetLowering::findRepresentativeClass(TRI, VT);
1075 // Use DPR as representative register class for all floating point
1076 // and vector types. Since there are 32 SPR registers and 32 DPR registers so
1077 // the cost is 1 for both f32 and f64.
1078 case MVT::f32: case MVT::f64: case MVT::v8i8: case MVT::v4i16:
1079 case MVT::v2i32: case MVT::v1i64: case MVT::v2f32:
1080 RRC = &ARM::DPRRegClass;
1081 // When NEON is used for SP, only half of the register file is available
1082 // because operations that define both SP and DP results will be constrained
1083 // to the VFP2 class (D0-D15). We currently model this constraint prior to
1084 // coalescing by double-counting the SP regs. See the FIXME above.
1085 if (Subtarget->useNEONForSinglePrecisionFP())
1088 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
1089 case MVT::v4f32: case MVT::v2f64:
1090 RRC = &ARM::DPRRegClass;
1094 RRC = &ARM::DPRRegClass;
1098 RRC = &ARM::DPRRegClass;
1102 return std::make_pair(RRC, Cost);
1105 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
1106 switch ((ARMISD::NodeType)Opcode) {
1107 case ARMISD::FIRST_NUMBER: break;
1108 case ARMISD::Wrapper: return "ARMISD::Wrapper";
1109 case ARMISD::WrapperPIC: return "ARMISD::WrapperPIC";
1110 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
1111 case ARMISD::COPY_STRUCT_BYVAL: return "ARMISD::COPY_STRUCT_BYVAL";
1112 case ARMISD::CALL: return "ARMISD::CALL";
1113 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
1114 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
1115 case ARMISD::tCALL: return "ARMISD::tCALL";
1116 case ARMISD::BRCOND: return "ARMISD::BRCOND";
1117 case ARMISD::BR_JT: return "ARMISD::BR_JT";
1118 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
1119 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
1120 case ARMISD::INTRET_FLAG: return "ARMISD::INTRET_FLAG";
1121 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
1122 case ARMISD::CMP: return "ARMISD::CMP";
1123 case ARMISD::CMN: return "ARMISD::CMN";
1124 case ARMISD::CMPZ: return "ARMISD::CMPZ";
1125 case ARMISD::CMPFP: return "ARMISD::CMPFP";
1126 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
1127 case ARMISD::BCC_i64: return "ARMISD::BCC_i64";
1128 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
1130 case ARMISD::CMOV: return "ARMISD::CMOV";
1132 case ARMISD::RBIT: return "ARMISD::RBIT";
1134 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
1135 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
1136 case ARMISD::RRX: return "ARMISD::RRX";
1138 case ARMISD::ADDC: return "ARMISD::ADDC";
1139 case ARMISD::ADDE: return "ARMISD::ADDE";
1140 case ARMISD::SUBC: return "ARMISD::SUBC";
1141 case ARMISD::SUBE: return "ARMISD::SUBE";
1143 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
1144 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
1146 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
1147 case ARMISD::EH_SJLJ_LONGJMP: return "ARMISD::EH_SJLJ_LONGJMP";
1148 case ARMISD::EH_SJLJ_SETUP_DISPATCH: return "ARMISD::EH_SJLJ_SETUP_DISPATCH";
1150 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN";
1152 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
1154 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
1156 case ARMISD::MEMBARRIER_MCR: return "ARMISD::MEMBARRIER_MCR";
1158 case ARMISD::PRELOAD: return "ARMISD::PRELOAD";
1160 case ARMISD::WIN__CHKSTK: return "ARMISD:::WIN__CHKSTK";
1161 case ARMISD::WIN__DBZCHK: return "ARMISD::WIN__DBZCHK";
1163 case ARMISD::VCEQ: return "ARMISD::VCEQ";
1164 case ARMISD::VCEQZ: return "ARMISD::VCEQZ";
1165 case ARMISD::VCGE: return "ARMISD::VCGE";
1166 case ARMISD::VCGEZ: return "ARMISD::VCGEZ";
1167 case ARMISD::VCLEZ: return "ARMISD::VCLEZ";
1168 case ARMISD::VCGEU: return "ARMISD::VCGEU";
1169 case ARMISD::VCGT: return "ARMISD::VCGT";
1170 case ARMISD::VCGTZ: return "ARMISD::VCGTZ";
1171 case ARMISD::VCLTZ: return "ARMISD::VCLTZ";
1172 case ARMISD::VCGTU: return "ARMISD::VCGTU";
1173 case ARMISD::VTST: return "ARMISD::VTST";
1175 case ARMISD::VSHL: return "ARMISD::VSHL";
1176 case ARMISD::VSHRs: return "ARMISD::VSHRs";
1177 case ARMISD::VSHRu: return "ARMISD::VSHRu";
1178 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
1179 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
1180 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
1181 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
1182 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
1183 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
1184 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
1185 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
1186 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
1187 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
1188 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
1189 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
1190 case ARMISD::VSLI: return "ARMISD::VSLI";
1191 case ARMISD::VSRI: return "ARMISD::VSRI";
1192 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
1193 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
1194 case ARMISD::VMOVIMM: return "ARMISD::VMOVIMM";
1195 case ARMISD::VMVNIMM: return "ARMISD::VMVNIMM";
1196 case ARMISD::VMOVFPIMM: return "ARMISD::VMOVFPIMM";
1197 case ARMISD::VDUP: return "ARMISD::VDUP";
1198 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
1199 case ARMISD::VEXT: return "ARMISD::VEXT";
1200 case ARMISD::VREV64: return "ARMISD::VREV64";
1201 case ARMISD::VREV32: return "ARMISD::VREV32";
1202 case ARMISD::VREV16: return "ARMISD::VREV16";
1203 case ARMISD::VZIP: return "ARMISD::VZIP";
1204 case ARMISD::VUZP: return "ARMISD::VUZP";
1205 case ARMISD::VTRN: return "ARMISD::VTRN";
1206 case ARMISD::VTBL1: return "ARMISD::VTBL1";
1207 case ARMISD::VTBL2: return "ARMISD::VTBL2";
1208 case ARMISD::VMULLs: return "ARMISD::VMULLs";
1209 case ARMISD::VMULLu: return "ARMISD::VMULLu";
1210 case ARMISD::UMLAL: return "ARMISD::UMLAL";
1211 case ARMISD::SMLAL: return "ARMISD::SMLAL";
1212 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
1213 case ARMISD::BFI: return "ARMISD::BFI";
1214 case ARMISD::VORRIMM: return "ARMISD::VORRIMM";
1215 case ARMISD::VBICIMM: return "ARMISD::VBICIMM";
1216 case ARMISD::VBSL: return "ARMISD::VBSL";
1217 case ARMISD::MEMCPY: return "ARMISD::MEMCPY";
1218 case ARMISD::VLD2DUP: return "ARMISD::VLD2DUP";
1219 case ARMISD::VLD3DUP: return "ARMISD::VLD3DUP";
1220 case ARMISD::VLD4DUP: return "ARMISD::VLD4DUP";
1221 case ARMISD::VLD1_UPD: return "ARMISD::VLD1_UPD";
1222 case ARMISD::VLD2_UPD: return "ARMISD::VLD2_UPD";
1223 case ARMISD::VLD3_UPD: return "ARMISD::VLD3_UPD";
1224 case ARMISD::VLD4_UPD: return "ARMISD::VLD4_UPD";
1225 case ARMISD::VLD2LN_UPD: return "ARMISD::VLD2LN_UPD";
1226 case ARMISD::VLD3LN_UPD: return "ARMISD::VLD3LN_UPD";
1227 case ARMISD::VLD4LN_UPD: return "ARMISD::VLD4LN_UPD";
1228 case ARMISD::VLD2DUP_UPD: return "ARMISD::VLD2DUP_UPD";
1229 case ARMISD::VLD3DUP_UPD: return "ARMISD::VLD3DUP_UPD";
1230 case ARMISD::VLD4DUP_UPD: return "ARMISD::VLD4DUP_UPD";
1231 case ARMISD::VST1_UPD: return "ARMISD::VST1_UPD";
1232 case ARMISD::VST2_UPD: return "ARMISD::VST2_UPD";
1233 case ARMISD::VST3_UPD: return "ARMISD::VST3_UPD";
1234 case ARMISD::VST4_UPD: return "ARMISD::VST4_UPD";
1235 case ARMISD::VST2LN_UPD: return "ARMISD::VST2LN_UPD";
1236 case ARMISD::VST3LN_UPD: return "ARMISD::VST3LN_UPD";
1237 case ARMISD::VST4LN_UPD: return "ARMISD::VST4LN_UPD";
1242 EVT ARMTargetLowering::getSetCCResultType(const DataLayout &DL, LLVMContext &,
1245 return getPointerTy(DL);
1246 return VT.changeVectorElementTypeToInteger();
1249 /// getRegClassFor - Return the register class that should be used for the
1250 /// specified value type.
1251 const TargetRegisterClass *ARMTargetLowering::getRegClassFor(MVT VT) const {
1252 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
1253 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
1254 // load / store 4 to 8 consecutive D registers.
1255 if (Subtarget->hasNEON()) {
1256 if (VT == MVT::v4i64)
1257 return &ARM::QQPRRegClass;
1258 if (VT == MVT::v8i64)
1259 return &ARM::QQQQPRRegClass;
1261 return TargetLowering::getRegClassFor(VT);
1264 // memcpy, and other memory intrinsics, typically tries to use LDM/STM if the
1265 // source/dest is aligned and the copy size is large enough. We therefore want
1266 // to align such objects passed to memory intrinsics.
1267 bool ARMTargetLowering::shouldAlignPointerArgs(CallInst *CI, unsigned &MinSize,
1268 unsigned &PrefAlign) const {
1269 if (!isa<MemIntrinsic>(CI))
1272 // On ARM11 onwards (excluding M class) 8-byte aligned LDM is typically 1
1273 // cycle faster than 4-byte aligned LDM.
1274 PrefAlign = (Subtarget->hasV6Ops() && !Subtarget->isMClass() ? 8 : 4);
1278 // Create a fast isel object.
1280 ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
1281 const TargetLibraryInfo *libInfo) const {
1282 return ARM::createFastISel(funcInfo, libInfo);
1285 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
1286 unsigned NumVals = N->getNumValues();
1288 return Sched::RegPressure;
1290 for (unsigned i = 0; i != NumVals; ++i) {
1291 EVT VT = N->getValueType(i);
1292 if (VT == MVT::Glue || VT == MVT::Other)
1294 if (VT.isFloatingPoint() || VT.isVector())
1298 if (!N->isMachineOpcode())
1299 return Sched::RegPressure;
1301 // Load are scheduled for latency even if there instruction itinerary
1302 // is not available.
1303 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
1304 const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
1306 if (MCID.getNumDefs() == 0)
1307 return Sched::RegPressure;
1308 if (!Itins->isEmpty() &&
1309 Itins->getOperandCycle(MCID.getSchedClass(), 0) > 2)
1312 return Sched::RegPressure;
1315 //===----------------------------------------------------------------------===//
1317 //===----------------------------------------------------------------------===//
1319 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
1320 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
1322 default: llvm_unreachable("Unknown condition code!");
1323 case ISD::SETNE: return ARMCC::NE;
1324 case ISD::SETEQ: return ARMCC::EQ;
1325 case ISD::SETGT: return ARMCC::GT;
1326 case ISD::SETGE: return ARMCC::GE;
1327 case ISD::SETLT: return ARMCC::LT;
1328 case ISD::SETLE: return ARMCC::LE;
1329 case ISD::SETUGT: return ARMCC::HI;
1330 case ISD::SETUGE: return ARMCC::HS;
1331 case ISD::SETULT: return ARMCC::LO;
1332 case ISD::SETULE: return ARMCC::LS;
1336 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
1337 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
1338 ARMCC::CondCodes &CondCode2) {
1339 CondCode2 = ARMCC::AL;
1341 default: llvm_unreachable("Unknown FP condition!");
1343 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
1345 case ISD::SETOGT: CondCode = ARMCC::GT; break;
1347 case ISD::SETOGE: CondCode = ARMCC::GE; break;
1348 case ISD::SETOLT: CondCode = ARMCC::MI; break;
1349 case ISD::SETOLE: CondCode = ARMCC::LS; break;
1350 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
1351 case ISD::SETO: CondCode = ARMCC::VC; break;
1352 case ISD::SETUO: CondCode = ARMCC::VS; break;
1353 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
1354 case ISD::SETUGT: CondCode = ARMCC::HI; break;
1355 case ISD::SETUGE: CondCode = ARMCC::PL; break;
1357 case ISD::SETULT: CondCode = ARMCC::LT; break;
1359 case ISD::SETULE: CondCode = ARMCC::LE; break;
1361 case ISD::SETUNE: CondCode = ARMCC::NE; break;
1365 //===----------------------------------------------------------------------===//
1366 // Calling Convention Implementation
1367 //===----------------------------------------------------------------------===//
1369 #include "ARMGenCallingConv.inc"
1371 /// getEffectiveCallingConv - Get the effective calling convention, taking into
1372 /// account presence of floating point hardware and calling convention
1373 /// limitations, such as support for variadic functions.
1375 ARMTargetLowering::getEffectiveCallingConv(CallingConv::ID CC,
1376 bool isVarArg) const {
1379 llvm_unreachable("Unsupported calling convention");
1380 case CallingConv::ARM_AAPCS:
1381 case CallingConv::ARM_APCS:
1382 case CallingConv::GHC:
1384 case CallingConv::ARM_AAPCS_VFP:
1385 return isVarArg ? CallingConv::ARM_AAPCS : CallingConv::ARM_AAPCS_VFP;
1386 case CallingConv::C:
1387 if (!Subtarget->isAAPCS_ABI())
1388 return CallingConv::ARM_APCS;
1389 else if (Subtarget->hasVFP2() && !Subtarget->isThumb1Only() &&
1390 getTargetMachine().Options.FloatABIType == FloatABI::Hard &&
1392 return CallingConv::ARM_AAPCS_VFP;
1394 return CallingConv::ARM_AAPCS;
1395 case CallingConv::Fast:
1396 if (!Subtarget->isAAPCS_ABI()) {
1397 if (Subtarget->hasVFP2() && !Subtarget->isThumb1Only() && !isVarArg)
1398 return CallingConv::Fast;
1399 return CallingConv::ARM_APCS;
1400 } else if (Subtarget->hasVFP2() && !Subtarget->isThumb1Only() && !isVarArg)
1401 return CallingConv::ARM_AAPCS_VFP;
1403 return CallingConv::ARM_AAPCS;
1407 /// CCAssignFnForNode - Selects the correct CCAssignFn for the given
1408 /// CallingConvention.
1409 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
1411 bool isVarArg) const {
1412 switch (getEffectiveCallingConv(CC, isVarArg)) {
1414 llvm_unreachable("Unsupported calling convention");
1415 case CallingConv::ARM_APCS:
1416 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1417 case CallingConv::ARM_AAPCS:
1418 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1419 case CallingConv::ARM_AAPCS_VFP:
1420 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1421 case CallingConv::Fast:
1422 return (Return ? RetFastCC_ARM_APCS : FastCC_ARM_APCS);
1423 case CallingConv::GHC:
1424 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS_GHC);
1428 /// LowerCallResult - Lower the result values of a call into the
1429 /// appropriate copies out of appropriate physical registers.
1431 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1432 CallingConv::ID CallConv, bool isVarArg,
1433 const SmallVectorImpl<ISD::InputArg> &Ins,
1434 SDLoc dl, SelectionDAG &DAG,
1435 SmallVectorImpl<SDValue> &InVals,
1436 bool isThisReturn, SDValue ThisVal) const {
1438 // Assign locations to each value returned by this call.
1439 SmallVector<CCValAssign, 16> RVLocs;
1440 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
1441 *DAG.getContext(), Call);
1442 CCInfo.AnalyzeCallResult(Ins,
1443 CCAssignFnForNode(CallConv, /* Return*/ true,
1446 // Copy all of the result registers out of their specified physreg.
1447 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1448 CCValAssign VA = RVLocs[i];
1450 // Pass 'this' value directly from the argument to return value, to avoid
1451 // reg unit interference
1452 if (i == 0 && isThisReturn) {
1453 assert(!VA.needsCustom() && VA.getLocVT() == MVT::i32 &&
1454 "unexpected return calling convention register assignment");
1455 InVals.push_back(ThisVal);
1460 if (VA.needsCustom()) {
1461 // Handle f64 or half of a v2f64.
1462 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1464 Chain = Lo.getValue(1);
1465 InFlag = Lo.getValue(2);
1466 VA = RVLocs[++i]; // skip ahead to next loc
1467 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1469 Chain = Hi.getValue(1);
1470 InFlag = Hi.getValue(2);
1471 if (!Subtarget->isLittle())
1473 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1475 if (VA.getLocVT() == MVT::v2f64) {
1476 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1477 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1478 DAG.getConstant(0, dl, MVT::i32));
1480 VA = RVLocs[++i]; // skip ahead to next loc
1481 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1482 Chain = Lo.getValue(1);
1483 InFlag = Lo.getValue(2);
1484 VA = RVLocs[++i]; // skip ahead to next loc
1485 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1486 Chain = Hi.getValue(1);
1487 InFlag = Hi.getValue(2);
1488 if (!Subtarget->isLittle())
1490 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1491 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1492 DAG.getConstant(1, dl, MVT::i32));
1495 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
1497 Chain = Val.getValue(1);
1498 InFlag = Val.getValue(2);
1501 switch (VA.getLocInfo()) {
1502 default: llvm_unreachable("Unknown loc info!");
1503 case CCValAssign::Full: break;
1504 case CCValAssign::BCvt:
1505 Val = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), Val);
1509 InVals.push_back(Val);
1515 /// LowerMemOpCallTo - Store the argument to the stack.
1517 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
1518 SDValue StackPtr, SDValue Arg,
1519 SDLoc dl, SelectionDAG &DAG,
1520 const CCValAssign &VA,
1521 ISD::ArgFlagsTy Flags) const {
1522 unsigned LocMemOffset = VA.getLocMemOffset();
1523 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset, dl);
1524 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(DAG.getDataLayout()),
1526 return DAG.getStore(
1527 Chain, dl, Arg, PtrOff,
1528 MachinePointerInfo::getStack(DAG.getMachineFunction(), LocMemOffset),
1532 void ARMTargetLowering::PassF64ArgInRegs(SDLoc dl, SelectionDAG &DAG,
1533 SDValue Chain, SDValue &Arg,
1534 RegsToPassVector &RegsToPass,
1535 CCValAssign &VA, CCValAssign &NextVA,
1537 SmallVectorImpl<SDValue> &MemOpChains,
1538 ISD::ArgFlagsTy Flags) const {
1540 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1541 DAG.getVTList(MVT::i32, MVT::i32), Arg);
1542 unsigned id = Subtarget->isLittle() ? 0 : 1;
1543 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd.getValue(id)));
1545 if (NextVA.isRegLoc())
1546 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1-id)));
1548 assert(NextVA.isMemLoc());
1549 if (!StackPtr.getNode())
1550 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP,
1551 getPointerTy(DAG.getDataLayout()));
1553 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1-id),
1559 /// LowerCall - Lowering a call into a callseq_start <-
1560 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
1563 ARMTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
1564 SmallVectorImpl<SDValue> &InVals) const {
1565 SelectionDAG &DAG = CLI.DAG;
1567 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
1568 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
1569 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
1570 SDValue Chain = CLI.Chain;
1571 SDValue Callee = CLI.Callee;
1572 bool &isTailCall = CLI.IsTailCall;
1573 CallingConv::ID CallConv = CLI.CallConv;
1574 bool doesNotRet = CLI.DoesNotReturn;
1575 bool isVarArg = CLI.IsVarArg;
1577 MachineFunction &MF = DAG.getMachineFunction();
1578 bool isStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1579 bool isThisReturn = false;
1580 bool isSibCall = false;
1581 auto Attr = MF.getFunction()->getFnAttribute("disable-tail-calls");
1583 // Disable tail calls if they're not supported.
1584 if (!Subtarget->supportsTailCall() || Attr.getValueAsString() == "true")
1588 // Check if it's really possible to do a tail call.
1589 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1590 isVarArg, isStructRet, MF.getFunction()->hasStructRetAttr(),
1591 Outs, OutVals, Ins, DAG);
1592 if (!isTailCall && CLI.CS && CLI.CS->isMustTailCall())
1593 report_fatal_error("failed to perform tail call elimination on a call "
1594 "site marked musttail");
1595 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1596 // detected sibcalls.
1603 // Analyze operands of the call, assigning locations to each operand.
1604 SmallVector<CCValAssign, 16> ArgLocs;
1605 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs,
1606 *DAG.getContext(), Call);
1607 CCInfo.AnalyzeCallOperands(Outs,
1608 CCAssignFnForNode(CallConv, /* Return*/ false,
1611 // Get a count of how many bytes are to be pushed on the stack.
1612 unsigned NumBytes = CCInfo.getNextStackOffset();
1614 // For tail calls, memory operands are available in our caller's stack.
1618 // Adjust the stack pointer for the new arguments...
1619 // These operations are automatically eliminated by the prolog/epilog pass
1621 Chain = DAG.getCALLSEQ_START(Chain,
1622 DAG.getIntPtrConstant(NumBytes, dl, true), dl);
1625 DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy(DAG.getDataLayout()));
1627 RegsToPassVector RegsToPass;
1628 SmallVector<SDValue, 8> MemOpChains;
1630 // Walk the register/memloc assignments, inserting copies/loads. In the case
1631 // of tail call optimization, arguments are handled later.
1632 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1634 ++i, ++realArgIdx) {
1635 CCValAssign &VA = ArgLocs[i];
1636 SDValue Arg = OutVals[realArgIdx];
1637 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1638 bool isByVal = Flags.isByVal();
1640 // Promote the value if needed.
1641 switch (VA.getLocInfo()) {
1642 default: llvm_unreachable("Unknown loc info!");
1643 case CCValAssign::Full: break;
1644 case CCValAssign::SExt:
1645 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1647 case CCValAssign::ZExt:
1648 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1650 case CCValAssign::AExt:
1651 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1653 case CCValAssign::BCvt:
1654 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1658 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1659 if (VA.needsCustom()) {
1660 if (VA.getLocVT() == MVT::v2f64) {
1661 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1662 DAG.getConstant(0, dl, MVT::i32));
1663 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1664 DAG.getConstant(1, dl, MVT::i32));
1666 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1667 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1669 VA = ArgLocs[++i]; // skip ahead to next loc
1670 if (VA.isRegLoc()) {
1671 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1672 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1674 assert(VA.isMemLoc());
1676 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1677 dl, DAG, VA, Flags));
1680 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1681 StackPtr, MemOpChains, Flags);
1683 } else if (VA.isRegLoc()) {
1684 if (realArgIdx == 0 && Flags.isReturned() && Outs[0].VT == MVT::i32) {
1685 assert(VA.getLocVT() == MVT::i32 &&
1686 "unexpected calling convention register assignment");
1687 assert(!Ins.empty() && Ins[0].VT == MVT::i32 &&
1688 "unexpected use of 'returned'");
1689 isThisReturn = true;
1691 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1692 } else if (isByVal) {
1693 assert(VA.isMemLoc());
1694 unsigned offset = 0;
1696 // True if this byval aggregate will be split between registers
1698 unsigned ByValArgsCount = CCInfo.getInRegsParamsCount();
1699 unsigned CurByValIdx = CCInfo.getInRegsParamsProcessed();
1701 if (CurByValIdx < ByValArgsCount) {
1703 unsigned RegBegin, RegEnd;
1704 CCInfo.getInRegsParamInfo(CurByValIdx, RegBegin, RegEnd);
1707 DAG.getTargetLoweringInfo().getPointerTy(DAG.getDataLayout());
1709 for (i = 0, j = RegBegin; j < RegEnd; i++, j++) {
1710 SDValue Const = DAG.getConstant(4*i, dl, MVT::i32);
1711 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
1712 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
1713 MachinePointerInfo(),
1714 false, false, false,
1715 DAG.InferPtrAlignment(AddArg));
1716 MemOpChains.push_back(Load.getValue(1));
1717 RegsToPass.push_back(std::make_pair(j, Load));
1720 // If parameter size outsides register area, "offset" value
1721 // helps us to calculate stack slot for remained part properly.
1722 offset = RegEnd - RegBegin;
1724 CCInfo.nextInRegsParam();
1727 if (Flags.getByValSize() > 4*offset) {
1728 auto PtrVT = getPointerTy(DAG.getDataLayout());
1729 unsigned LocMemOffset = VA.getLocMemOffset();
1730 SDValue StkPtrOff = DAG.getIntPtrConstant(LocMemOffset, dl);
1731 SDValue Dst = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, StkPtrOff);
1732 SDValue SrcOffset = DAG.getIntPtrConstant(4*offset, dl);
1733 SDValue Src = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, SrcOffset);
1734 SDValue SizeNode = DAG.getConstant(Flags.getByValSize() - 4*offset, dl,
1736 SDValue AlignNode = DAG.getConstant(Flags.getByValAlign(), dl,
1739 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
1740 SDValue Ops[] = { Chain, Dst, Src, SizeNode, AlignNode};
1741 MemOpChains.push_back(DAG.getNode(ARMISD::COPY_STRUCT_BYVAL, dl, VTs,
1744 } else if (!isSibCall) {
1745 assert(VA.isMemLoc());
1747 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1748 dl, DAG, VA, Flags));
1752 if (!MemOpChains.empty())
1753 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains);
1755 // Build a sequence of copy-to-reg nodes chained together with token chain
1756 // and flag operands which copy the outgoing args into the appropriate regs.
1758 // Tail call byval lowering might overwrite argument registers so in case of
1759 // tail call optimization the copies to registers are lowered later.
1761 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1762 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1763 RegsToPass[i].second, InFlag);
1764 InFlag = Chain.getValue(1);
1767 // For tail calls lower the arguments to the 'real' stack slot.
1769 // Force all the incoming stack arguments to be loaded from the stack
1770 // before any new outgoing arguments are stored to the stack, because the
1771 // outgoing stack slots may alias the incoming argument stack slots, and
1772 // the alias isn't otherwise explicit. This is slightly more conservative
1773 // than necessary, because it means that each store effectively depends
1774 // on every argument instead of just those arguments it would clobber.
1776 // Do not flag preceding copytoreg stuff together with the following stuff.
1778 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1779 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1780 RegsToPass[i].second, InFlag);
1781 InFlag = Chain.getValue(1);
1786 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1787 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1788 // node so that legalize doesn't hack it.
1789 bool isDirect = false;
1790 bool isARMFunc = false;
1791 bool isLocalARMFunc = false;
1792 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1793 auto PtrVt = getPointerTy(DAG.getDataLayout());
1795 if (Subtarget->genLongCalls()) {
1796 assert((Subtarget->isTargetWindows() ||
1797 getTargetMachine().getRelocationModel() == Reloc::Static) &&
1798 "long-calls with non-static relocation model!");
1799 // Handle a global address or an external symbol. If it's not one of
1800 // those, the target's already in a register, so we don't need to do
1802 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1803 const GlobalValue *GV = G->getGlobal();
1804 // Create a constant pool entry for the callee address
1805 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1806 ARMConstantPoolValue *CPV =
1807 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 0);
1809 // Get the address of the callee into a register
1810 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVt, 4);
1811 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1812 Callee = DAG.getLoad(
1813 PtrVt, dl, DAG.getEntryNode(), CPAddr,
1814 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), false,
1816 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1817 const char *Sym = S->getSymbol();
1819 // Create a constant pool entry for the callee address
1820 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1821 ARMConstantPoolValue *CPV =
1822 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1823 ARMPCLabelIndex, 0);
1824 // Get the address of the callee into a register
1825 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVt, 4);
1826 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1827 Callee = DAG.getLoad(
1828 PtrVt, dl, DAG.getEntryNode(), CPAddr,
1829 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), false,
1832 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1833 const GlobalValue *GV = G->getGlobal();
1835 bool isDef = GV->isStrongDefinitionForLinker();
1836 bool isStub = (!isDef && Subtarget->isTargetMachO()) &&
1837 getTargetMachine().getRelocationModel() != Reloc::Static;
1838 isARMFunc = !Subtarget->isThumb() || (isStub && !Subtarget->isMClass());
1839 // ARM call to a local ARM function is predicable.
1840 isLocalARMFunc = !Subtarget->isThumb() && (isDef || !ARMInterworking);
1841 // tBX takes a register source operand.
1842 if (isStub && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1843 assert(Subtarget->isTargetMachO() && "WrapperPIC use on non-MachO?");
1844 Callee = DAG.getNode(
1845 ARMISD::WrapperPIC, dl, PtrVt,
1846 DAG.getTargetGlobalAddress(GV, dl, PtrVt, 0, ARMII::MO_NONLAZY));
1847 Callee = DAG.getLoad(PtrVt, dl, DAG.getEntryNode(), Callee,
1848 MachinePointerInfo::getGOT(DAG.getMachineFunction()),
1849 false, false, true, 0);
1850 } else if (Subtarget->isTargetCOFF()) {
1851 assert(Subtarget->isTargetWindows() &&
1852 "Windows is the only supported COFF target");
1853 unsigned TargetFlags = GV->hasDLLImportStorageClass()
1854 ? ARMII::MO_DLLIMPORT
1855 : ARMII::MO_NO_FLAG;
1857 DAG.getTargetGlobalAddress(GV, dl, PtrVt, /*Offset=*/0, TargetFlags);
1858 if (GV->hasDLLImportStorageClass())
1860 DAG.getLoad(PtrVt, dl, DAG.getEntryNode(),
1861 DAG.getNode(ARMISD::Wrapper, dl, PtrVt, Callee),
1862 MachinePointerInfo::getGOT(DAG.getMachineFunction()),
1863 false, false, false, 0);
1865 // On ELF targets for PIC code, direct calls should go through the PLT
1866 unsigned OpFlags = 0;
1867 if (Subtarget->isTargetELF() &&
1868 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1869 OpFlags = ARMII::MO_PLT;
1870 Callee = DAG.getTargetGlobalAddress(GV, dl, PtrVt, 0, OpFlags);
1872 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1874 bool isStub = Subtarget->isTargetMachO() &&
1875 getTargetMachine().getRelocationModel() != Reloc::Static;
1876 isARMFunc = !Subtarget->isThumb() || (isStub && !Subtarget->isMClass());
1877 // tBX takes a register source operand.
1878 const char *Sym = S->getSymbol();
1879 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1880 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1881 ARMConstantPoolValue *CPV =
1882 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1883 ARMPCLabelIndex, 4);
1884 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVt, 4);
1885 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1886 Callee = DAG.getLoad(
1887 PtrVt, dl, DAG.getEntryNode(), CPAddr,
1888 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), false,
1890 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, dl, MVT::i32);
1891 Callee = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVt, Callee, PICLabel);
1893 unsigned OpFlags = 0;
1894 // On ELF targets for PIC code, direct calls should go through the PLT
1895 if (Subtarget->isTargetELF() &&
1896 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1897 OpFlags = ARMII::MO_PLT;
1898 Callee = DAG.getTargetExternalSymbol(Sym, PtrVt, OpFlags);
1902 // FIXME: handle tail calls differently.
1904 if (Subtarget->isThumb()) {
1905 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1906 CallOpc = ARMISD::CALL_NOLINK;
1908 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1910 if (!isDirect && !Subtarget->hasV5TOps())
1911 CallOpc = ARMISD::CALL_NOLINK;
1912 else if (doesNotRet && isDirect && Subtarget->hasRAS() &&
1913 // Emit regular call when code size is the priority
1914 !MF.getFunction()->optForMinSize())
1915 // "mov lr, pc; b _foo" to avoid confusing the RSP
1916 CallOpc = ARMISD::CALL_NOLINK;
1918 CallOpc = isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL;
1921 std::vector<SDValue> Ops;
1922 Ops.push_back(Chain);
1923 Ops.push_back(Callee);
1925 // Add argument registers to the end of the list so that they are known live
1927 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1928 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1929 RegsToPass[i].second.getValueType()));
1931 // Add a register mask operand representing the call-preserved registers.
1933 const uint32_t *Mask;
1934 const ARMBaseRegisterInfo *ARI = Subtarget->getRegisterInfo();
1936 // For 'this' returns, use the R0-preserving mask if applicable
1937 Mask = ARI->getThisReturnPreservedMask(MF, CallConv);
1939 // Set isThisReturn to false if the calling convention is not one that
1940 // allows 'returned' to be modeled in this way, so LowerCallResult does
1941 // not try to pass 'this' straight through
1942 isThisReturn = false;
1943 Mask = ARI->getCallPreservedMask(MF, CallConv);
1946 Mask = ARI->getCallPreservedMask(MF, CallConv);
1948 assert(Mask && "Missing call preserved mask for calling convention");
1949 Ops.push_back(DAG.getRegisterMask(Mask));
1952 if (InFlag.getNode())
1953 Ops.push_back(InFlag);
1955 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
1957 MF.getFrameInfo()->setHasTailCall();
1958 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, Ops);
1961 // Returns a chain and a flag for retval copy to use.
1962 Chain = DAG.getNode(CallOpc, dl, NodeTys, Ops);
1963 InFlag = Chain.getValue(1);
1965 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, dl, true),
1966 DAG.getIntPtrConstant(0, dl, true), InFlag, dl);
1968 InFlag = Chain.getValue(1);
1970 // Handle result values, copying them out of physregs into vregs that we
1972 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins, dl, DAG,
1973 InVals, isThisReturn,
1974 isThisReturn ? OutVals[0] : SDValue());
1977 /// HandleByVal - Every parameter *after* a byval parameter is passed
1978 /// on the stack. Remember the next parameter register to allocate,
1979 /// and then confiscate the rest of the parameter registers to insure
1981 void ARMTargetLowering::HandleByVal(CCState *State, unsigned &Size,
1982 unsigned Align) const {
1983 assert((State->getCallOrPrologue() == Prologue ||
1984 State->getCallOrPrologue() == Call) &&
1985 "unhandled ParmContext");
1987 // Byval (as with any stack) slots are always at least 4 byte aligned.
1988 Align = std::max(Align, 4U);
1990 unsigned Reg = State->AllocateReg(GPRArgRegs);
1994 unsigned AlignInRegs = Align / 4;
1995 unsigned Waste = (ARM::R4 - Reg) % AlignInRegs;
1996 for (unsigned i = 0; i < Waste; ++i)
1997 Reg = State->AllocateReg(GPRArgRegs);
2002 unsigned Excess = 4 * (ARM::R4 - Reg);
2004 // Special case when NSAA != SP and parameter size greater than size of
2005 // all remained GPR regs. In that case we can't split parameter, we must
2006 // send it to stack. We also must set NCRN to R4, so waste all
2007 // remained registers.
2008 const unsigned NSAAOffset = State->getNextStackOffset();
2009 if (NSAAOffset != 0 && Size > Excess) {
2010 while (State->AllocateReg(GPRArgRegs))
2015 // First register for byval parameter is the first register that wasn't
2016 // allocated before this method call, so it would be "reg".
2017 // If parameter is small enough to be saved in range [reg, r4), then
2018 // the end (first after last) register would be reg + param-size-in-regs,
2019 // else parameter would be splitted between registers and stack,
2020 // end register would be r4 in this case.
2021 unsigned ByValRegBegin = Reg;
2022 unsigned ByValRegEnd = std::min<unsigned>(Reg + Size / 4, ARM::R4);
2023 State->addInRegsParamInfo(ByValRegBegin, ByValRegEnd);
2024 // Note, first register is allocated in the beginning of function already,
2025 // allocate remained amount of registers we need.
2026 for (unsigned i = Reg + 1; i != ByValRegEnd; ++i)
2027 State->AllocateReg(GPRArgRegs);
2028 // A byval parameter that is split between registers and memory needs its
2029 // size truncated here.
2030 // In the case where the entire structure fits in registers, we set the
2031 // size in memory to zero.
2032 Size = std::max<int>(Size - Excess, 0);
2035 /// MatchingStackOffset - Return true if the given stack call argument is
2036 /// already available in the same position (relatively) of the caller's
2037 /// incoming argument stack.
2039 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
2040 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
2041 const TargetInstrInfo *TII) {
2042 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
2044 if (Arg.getOpcode() == ISD::CopyFromReg) {
2045 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
2046 if (!TargetRegisterInfo::isVirtualRegister(VR))
2048 MachineInstr *Def = MRI->getVRegDef(VR);
2051 if (!Flags.isByVal()) {
2052 if (!TII->isLoadFromStackSlot(Def, FI))
2057 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
2058 if (Flags.isByVal())
2059 // ByVal argument is passed in as a pointer but it's now being
2060 // dereferenced. e.g.
2061 // define @foo(%struct.X* %A) {
2062 // tail call @bar(%struct.X* byval %A)
2065 SDValue Ptr = Ld->getBasePtr();
2066 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
2069 FI = FINode->getIndex();
2073 assert(FI != INT_MAX);
2074 if (!MFI->isFixedObjectIndex(FI))
2076 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
2079 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2080 /// for tail call optimization. Targets which want to do tail call
2081 /// optimization should implement this function.
2083 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2084 CallingConv::ID CalleeCC,
2086 bool isCalleeStructRet,
2087 bool isCallerStructRet,
2088 const SmallVectorImpl<ISD::OutputArg> &Outs,
2089 const SmallVectorImpl<SDValue> &OutVals,
2090 const SmallVectorImpl<ISD::InputArg> &Ins,
2091 SelectionDAG& DAG) const {
2092 const Function *CallerF = DAG.getMachineFunction().getFunction();
2093 CallingConv::ID CallerCC = CallerF->getCallingConv();
2094 bool CCMatch = CallerCC == CalleeCC;
2096 assert(Subtarget->supportsTailCall());
2098 // Look for obvious safe cases to perform tail call optimization that do not
2099 // require ABI changes. This is what gcc calls sibcall.
2101 // Do not sibcall optimize vararg calls unless the call site is not passing
2103 if (isVarArg && !Outs.empty())
2106 // Exception-handling functions need a special set of instructions to indicate
2107 // a return to the hardware. Tail-calling another function would probably
2109 if (CallerF->hasFnAttribute("interrupt"))
2112 // Also avoid sibcall optimization if either caller or callee uses struct
2113 // return semantics.
2114 if (isCalleeStructRet || isCallerStructRet)
2117 // Externally-defined functions with weak linkage should not be
2118 // tail-called on ARM when the OS does not support dynamic
2119 // pre-emption of symbols, as the AAELF spec requires normal calls
2120 // to undefined weak functions to be replaced with a NOP or jump to the
2121 // next instruction. The behaviour of branch instructions in this
2122 // situation (as used for tail calls) is implementation-defined, so we
2123 // cannot rely on the linker replacing the tail call with a return.
2124 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2125 const GlobalValue *GV = G->getGlobal();
2126 const Triple &TT = getTargetMachine().getTargetTriple();
2127 if (GV->hasExternalWeakLinkage() &&
2128 (!TT.isOSWindows() || TT.isOSBinFormatELF() || TT.isOSBinFormatMachO()))
2132 // If the calling conventions do not match, then we'd better make sure the
2133 // results are returned in the same way as what the caller expects.
2135 SmallVector<CCValAssign, 16> RVLocs1;
2136 ARMCCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(), RVLocs1,
2137 *DAG.getContext(), Call);
2138 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
2140 SmallVector<CCValAssign, 16> RVLocs2;
2141 ARMCCState CCInfo2(CallerCC, false, DAG.getMachineFunction(), RVLocs2,
2142 *DAG.getContext(), Call);
2143 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
2145 if (RVLocs1.size() != RVLocs2.size())
2147 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
2148 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
2150 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
2152 if (RVLocs1[i].isRegLoc()) {
2153 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
2156 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
2162 // If Caller's vararg or byval argument has been split between registers and
2163 // stack, do not perform tail call, since part of the argument is in caller's
2165 const ARMFunctionInfo *AFI_Caller = DAG.getMachineFunction().
2166 getInfo<ARMFunctionInfo>();
2167 if (AFI_Caller->getArgRegsSaveSize())
2170 // If the callee takes no arguments then go on to check the results of the
2172 if (!Outs.empty()) {
2173 // Check if stack adjustment is needed. For now, do not do this if any
2174 // argument is passed on the stack.
2175 SmallVector<CCValAssign, 16> ArgLocs;
2176 ARMCCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
2177 *DAG.getContext(), Call);
2178 CCInfo.AnalyzeCallOperands(Outs,
2179 CCAssignFnForNode(CalleeCC, false, isVarArg));
2180 if (CCInfo.getNextStackOffset()) {
2181 MachineFunction &MF = DAG.getMachineFunction();
2183 // Check if the arguments are already laid out in the right way as
2184 // the caller's fixed stack objects.
2185 MachineFrameInfo *MFI = MF.getFrameInfo();
2186 const MachineRegisterInfo *MRI = &MF.getRegInfo();
2187 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
2188 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
2190 ++i, ++realArgIdx) {
2191 CCValAssign &VA = ArgLocs[i];
2192 EVT RegVT = VA.getLocVT();
2193 SDValue Arg = OutVals[realArgIdx];
2194 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
2195 if (VA.getLocInfo() == CCValAssign::Indirect)
2197 if (VA.needsCustom()) {
2198 // f64 and vector types are split into multiple registers or
2199 // register/stack-slot combinations. The types will not match
2200 // the registers; give up on memory f64 refs until we figure
2201 // out what to do about this.
2204 if (!ArgLocs[++i].isRegLoc())
2206 if (RegVT == MVT::v2f64) {
2207 if (!ArgLocs[++i].isRegLoc())
2209 if (!ArgLocs[++i].isRegLoc())
2212 } else if (!VA.isRegLoc()) {
2213 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
2225 ARMTargetLowering::CanLowerReturn(CallingConv::ID CallConv,
2226 MachineFunction &MF, bool isVarArg,
2227 const SmallVectorImpl<ISD::OutputArg> &Outs,
2228 LLVMContext &Context) const {
2229 SmallVector<CCValAssign, 16> RVLocs;
2230 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, Context);
2231 return CCInfo.CheckReturn(Outs, CCAssignFnForNode(CallConv, /*Return=*/true,
2235 static SDValue LowerInterruptReturn(SmallVectorImpl<SDValue> &RetOps,
2236 SDLoc DL, SelectionDAG &DAG) {
2237 const MachineFunction &MF = DAG.getMachineFunction();
2238 const Function *F = MF.getFunction();
2240 StringRef IntKind = F->getFnAttribute("interrupt").getValueAsString();
2242 // See ARM ARM v7 B1.8.3. On exception entry LR is set to a possibly offset
2243 // version of the "preferred return address". These offsets affect the return
2244 // instruction if this is a return from PL1 without hypervisor extensions.
2245 // IRQ/FIQ: +4 "subs pc, lr, #4"
2246 // SWI: 0 "subs pc, lr, #0"
2247 // ABORT: +4 "subs pc, lr, #4"
2248 // UNDEF: +4/+2 "subs pc, lr, #0"
2249 // UNDEF varies depending on where the exception came from ARM or Thumb
2250 // mode. Alongside GCC, we throw our hands up in disgust and pretend it's 0.
2253 if (IntKind == "" || IntKind == "IRQ" || IntKind == "FIQ" ||
2256 else if (IntKind == "SWI" || IntKind == "UNDEF")
2259 report_fatal_error("Unsupported interrupt attribute. If present, value "
2260 "must be one of: IRQ, FIQ, SWI, ABORT or UNDEF");
2262 RetOps.insert(RetOps.begin() + 1,
2263 DAG.getConstant(LROffset, DL, MVT::i32, false));
2265 return DAG.getNode(ARMISD::INTRET_FLAG, DL, MVT::Other, RetOps);
2269 ARMTargetLowering::LowerReturn(SDValue Chain,
2270 CallingConv::ID CallConv, bool isVarArg,
2271 const SmallVectorImpl<ISD::OutputArg> &Outs,
2272 const SmallVectorImpl<SDValue> &OutVals,
2273 SDLoc dl, SelectionDAG &DAG) const {
2275 // CCValAssign - represent the assignment of the return value to a location.
2276 SmallVector<CCValAssign, 16> RVLocs;
2278 // CCState - Info about the registers and stack slots.
2279 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
2280 *DAG.getContext(), Call);
2282 // Analyze outgoing return values.
2283 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
2287 SmallVector<SDValue, 4> RetOps;
2288 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
2289 bool isLittleEndian = Subtarget->isLittle();
2291 MachineFunction &MF = DAG.getMachineFunction();
2292 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2293 AFI->setReturnRegsCount(RVLocs.size());
2295 // Copy the result values into the output registers.
2296 for (unsigned i = 0, realRVLocIdx = 0;
2298 ++i, ++realRVLocIdx) {
2299 CCValAssign &VA = RVLocs[i];
2300 assert(VA.isRegLoc() && "Can only return in registers!");
2302 SDValue Arg = OutVals[realRVLocIdx];
2304 switch (VA.getLocInfo()) {
2305 default: llvm_unreachable("Unknown loc info!");
2306 case CCValAssign::Full: break;
2307 case CCValAssign::BCvt:
2308 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
2312 if (VA.needsCustom()) {
2313 if (VA.getLocVT() == MVT::v2f64) {
2314 // Extract the first half and return it in two registers.
2315 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
2316 DAG.getConstant(0, dl, MVT::i32));
2317 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
2318 DAG.getVTList(MVT::i32, MVT::i32), Half);
2320 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2321 HalfGPRs.getValue(isLittleEndian ? 0 : 1),
2323 Flag = Chain.getValue(1);
2324 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2325 VA = RVLocs[++i]; // skip ahead to next loc
2326 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2327 HalfGPRs.getValue(isLittleEndian ? 1 : 0),
2329 Flag = Chain.getValue(1);
2330 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2331 VA = RVLocs[++i]; // skip ahead to next loc
2333 // Extract the 2nd half and fall through to handle it as an f64 value.
2334 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
2335 DAG.getConstant(1, dl, MVT::i32));
2337 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
2339 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
2340 DAG.getVTList(MVT::i32, MVT::i32), Arg);
2341 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2342 fmrrd.getValue(isLittleEndian ? 0 : 1),
2344 Flag = Chain.getValue(1);
2345 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2346 VA = RVLocs[++i]; // skip ahead to next loc
2347 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2348 fmrrd.getValue(isLittleEndian ? 1 : 0),
2351 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
2353 // Guarantee that all emitted copies are
2354 // stuck together, avoiding something bad.
2355 Flag = Chain.getValue(1);
2356 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2359 // Update chain and glue.
2362 RetOps.push_back(Flag);
2364 // CPUs which aren't M-class use a special sequence to return from
2365 // exceptions (roughly, any instruction setting pc and cpsr simultaneously,
2366 // though we use "subs pc, lr, #N").
2368 // M-class CPUs actually use a normal return sequence with a special
2369 // (hardware-provided) value in LR, so the normal code path works.
2370 if (DAG.getMachineFunction().getFunction()->hasFnAttribute("interrupt") &&
2371 !Subtarget->isMClass()) {
2372 if (Subtarget->isThumb1Only())
2373 report_fatal_error("interrupt attribute is not supported in Thumb1");
2374 return LowerInterruptReturn(RetOps, dl, DAG);
2377 return DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, RetOps);
2380 bool ARMTargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
2381 if (N->getNumValues() != 1)
2383 if (!N->hasNUsesOfValue(1, 0))
2386 SDValue TCChain = Chain;
2387 SDNode *Copy = *N->use_begin();
2388 if (Copy->getOpcode() == ISD::CopyToReg) {
2389 // If the copy has a glue operand, we conservatively assume it isn't safe to
2390 // perform a tail call.
2391 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
2393 TCChain = Copy->getOperand(0);
2394 } else if (Copy->getOpcode() == ARMISD::VMOVRRD) {
2395 SDNode *VMov = Copy;
2396 // f64 returned in a pair of GPRs.
2397 SmallPtrSet<SDNode*, 2> Copies;
2398 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
2400 if (UI->getOpcode() != ISD::CopyToReg)
2404 if (Copies.size() > 2)
2407 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
2409 SDValue UseChain = UI->getOperand(0);
2410 if (Copies.count(UseChain.getNode()))
2414 // We are at the top of this chain.
2415 // If the copy has a glue operand, we conservatively assume it
2416 // isn't safe to perform a tail call.
2417 if (UI->getOperand(UI->getNumOperands()-1).getValueType() == MVT::Glue)
2423 } else if (Copy->getOpcode() == ISD::BITCAST) {
2424 // f32 returned in a single GPR.
2425 if (!Copy->hasOneUse())
2427 Copy = *Copy->use_begin();
2428 if (Copy->getOpcode() != ISD::CopyToReg || !Copy->hasNUsesOfValue(1, 0))
2430 // If the copy has a glue operand, we conservatively assume it isn't safe to
2431 // perform a tail call.
2432 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
2434 TCChain = Copy->getOperand(0);
2439 bool HasRet = false;
2440 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
2442 if (UI->getOpcode() != ARMISD::RET_FLAG &&
2443 UI->getOpcode() != ARMISD::INTRET_FLAG)
2455 bool ARMTargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
2456 if (!Subtarget->supportsTailCall())
2460 CI->getParent()->getParent()->getFnAttribute("disable-tail-calls");
2461 if (!CI->isTailCall() || Attr.getValueAsString() == "true")
2467 // Trying to write a 64 bit value so need to split into two 32 bit values first,
2468 // and pass the lower and high parts through.
2469 static SDValue LowerWRITE_REGISTER(SDValue Op, SelectionDAG &DAG) {
2471 SDValue WriteValue = Op->getOperand(2);
2473 // This function is only supposed to be called for i64 type argument.
2474 assert(WriteValue.getValueType() == MVT::i64
2475 && "LowerWRITE_REGISTER called for non-i64 type argument.");
2477 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, WriteValue,
2478 DAG.getConstant(0, DL, MVT::i32));
2479 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, WriteValue,
2480 DAG.getConstant(1, DL, MVT::i32));
2481 SDValue Ops[] = { Op->getOperand(0), Op->getOperand(1), Lo, Hi };
2482 return DAG.getNode(ISD::WRITE_REGISTER, DL, MVT::Other, Ops);
2485 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
2486 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
2487 // one of the above mentioned nodes. It has to be wrapped because otherwise
2488 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
2489 // be used to form addressing mode. These wrapped nodes will be selected
2491 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
2492 EVT PtrVT = Op.getValueType();
2493 // FIXME there is no actual debug info here
2495 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
2497 if (CP->isMachineConstantPoolEntry())
2498 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
2499 CP->getAlignment());
2501 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
2502 CP->getAlignment());
2503 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
2506 unsigned ARMTargetLowering::getJumpTableEncoding() const {
2507 return MachineJumpTableInfo::EK_Inline;
2510 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
2511 SelectionDAG &DAG) const {
2512 MachineFunction &MF = DAG.getMachineFunction();
2513 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2514 unsigned ARMPCLabelIndex = 0;
2516 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2517 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
2518 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2520 if (RelocM == Reloc::Static) {
2521 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
2523 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2524 ARMPCLabelIndex = AFI->createPICLabelUId();
2525 ARMConstantPoolValue *CPV =
2526 ARMConstantPoolConstant::Create(BA, ARMPCLabelIndex,
2527 ARMCP::CPBlockAddress, PCAdj);
2528 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2530 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
2532 DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
2533 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
2534 false, false, false, 0);
2535 if (RelocM == Reloc::Static)
2537 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, DL, MVT::i32);
2538 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
2541 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
2543 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
2544 SelectionDAG &DAG) const {
2546 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2547 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2548 MachineFunction &MF = DAG.getMachineFunction();
2549 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2550 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2551 ARMConstantPoolValue *CPV =
2552 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2553 ARMCP::CPValue, PCAdj, ARMCP::TLSGD, true);
2554 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2555 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
2557 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
2558 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
2559 false, false, false, 0);
2560 SDValue Chain = Argument.getValue(1);
2562 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, dl, MVT::i32);
2563 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
2565 // call __tls_get_addr.
2568 Entry.Node = Argument;
2569 Entry.Ty = (Type *) Type::getInt32Ty(*DAG.getContext());
2570 Args.push_back(Entry);
2572 // FIXME: is there useful debug info available here?
2573 TargetLowering::CallLoweringInfo CLI(DAG);
2574 CLI.setDebugLoc(dl).setChain(Chain)
2575 .setCallee(CallingConv::C, Type::getInt32Ty(*DAG.getContext()),
2576 DAG.getExternalSymbol("__tls_get_addr", PtrVT), std::move(Args),
2579 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
2580 return CallResult.first;
2583 // Lower ISD::GlobalTLSAddress using the "initial exec" or
2584 // "local exec" model.
2586 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
2588 TLSModel::Model model) const {
2589 const GlobalValue *GV = GA->getGlobal();
2592 SDValue Chain = DAG.getEntryNode();
2593 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2594 // Get the Thread Pointer
2595 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2597 if (model == TLSModel::InitialExec) {
2598 MachineFunction &MF = DAG.getMachineFunction();
2599 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2600 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2601 // Initial exec model.
2602 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2603 ARMConstantPoolValue *CPV =
2604 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2605 ARMCP::CPValue, PCAdj, ARMCP::GOTTPOFF,
2607 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2608 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2609 Offset = DAG.getLoad(
2610 PtrVT, dl, Chain, Offset,
2611 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), false,
2613 Chain = Offset.getValue(1);
2615 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, dl, MVT::i32);
2616 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
2618 Offset = DAG.getLoad(
2619 PtrVT, dl, Chain, Offset,
2620 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), false,
2624 assert(model == TLSModel::LocalExec);
2625 ARMConstantPoolValue *CPV =
2626 ARMConstantPoolConstant::Create(GV, ARMCP::TPOFF);
2627 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2628 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2629 Offset = DAG.getLoad(
2630 PtrVT, dl, Chain, Offset,
2631 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), false,
2635 // The address of the thread local variable is the add of the thread
2636 // pointer with the offset of the variable.
2637 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
2641 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
2642 // TODO: implement the "local dynamic" model
2643 assert(Subtarget->isTargetELF() &&
2644 "TLS not implemented for non-ELF targets");
2645 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
2646 if (DAG.getTarget().Options.EmulatedTLS)
2647 return LowerToTLSEmulatedModel(GA, DAG);
2649 TLSModel::Model model = getTargetMachine().getTLSModel(GA->getGlobal());
2652 case TLSModel::GeneralDynamic:
2653 case TLSModel::LocalDynamic:
2654 return LowerToTLSGeneralDynamicModel(GA, DAG);
2655 case TLSModel::InitialExec:
2656 case TLSModel::LocalExec:
2657 return LowerToTLSExecModels(GA, DAG, model);
2659 llvm_unreachable("bogus TLS model");
2662 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
2663 SelectionDAG &DAG) const {
2664 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2666 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2667 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2669 !(GV->hasHiddenVisibility() || GV->hasLocalLinkage());
2671 MachineFunction &MF = DAG.getMachineFunction();
2672 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2673 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2674 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2676 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2677 ARMConstantPoolValue *CPV = ARMConstantPoolConstant::Create(
2678 GV, ARMPCLabelIndex, ARMCP::CPValue, PCAdj,
2679 UseGOT_PREL ? ARMCP::GOT_PREL : ARMCP::no_modifier,
2680 /*AddCurrentAddress=*/UseGOT_PREL);
2681 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2682 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2683 SDValue Result = DAG.getLoad(
2684 PtrVT, dl, DAG.getEntryNode(), CPAddr,
2685 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), false,
2687 SDValue Chain = Result.getValue(1);
2688 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, dl, MVT::i32);
2689 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2691 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
2692 MachinePointerInfo::getGOT(DAG.getMachineFunction()),
2693 false, false, false, 0);
2697 // If we have T2 ops, we can materialize the address directly via movt/movw
2698 // pair. This is always cheaper.
2699 if (Subtarget->useMovt(DAG.getMachineFunction())) {
2701 // FIXME: Once remat is capable of dealing with instructions with register
2702 // operands, expand this into two nodes.
2703 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2704 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2706 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2707 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2709 PtrVT, dl, DAG.getEntryNode(), CPAddr,
2710 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), false,
2715 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
2716 SelectionDAG &DAG) const {
2717 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2719 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2720 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2722 if (Subtarget->useMovt(DAG.getMachineFunction()))
2725 // FIXME: Once remat is capable of dealing with instructions with register
2726 // operands, expand this into multiple nodes
2728 RelocM == Reloc::PIC_ ? ARMISD::WrapperPIC : ARMISD::Wrapper;
2730 SDValue G = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, ARMII::MO_NONLAZY);
2731 SDValue Result = DAG.getNode(Wrapper, dl, PtrVT, G);
2733 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2734 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
2735 MachinePointerInfo::getGOT(DAG.getMachineFunction()),
2736 false, false, false, 0);
2740 SDValue ARMTargetLowering::LowerGlobalAddressWindows(SDValue Op,
2741 SelectionDAG &DAG) const {
2742 assert(Subtarget->isTargetWindows() && "non-Windows COFF is not supported");
2743 assert(Subtarget->useMovt(DAG.getMachineFunction()) &&
2744 "Windows on ARM expects to use movw/movt");
2746 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2747 const ARMII::TOF TargetFlags =
2748 (GV->hasDLLImportStorageClass() ? ARMII::MO_DLLIMPORT : ARMII::MO_NO_FLAG);
2749 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2755 // FIXME: Once remat is capable of dealing with instructions with register
2756 // operands, expand this into two nodes.
2757 Result = DAG.getNode(ARMISD::Wrapper, DL, PtrVT,
2758 DAG.getTargetGlobalAddress(GV, DL, PtrVT, /*Offset=*/0,
2760 if (GV->hasDLLImportStorageClass())
2761 Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), Result,
2762 MachinePointerInfo::getGOT(DAG.getMachineFunction()),
2763 false, false, false, 0);
2768 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
2770 SDValue Val = DAG.getConstant(0, dl, MVT::i32);
2771 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl,
2772 DAG.getVTList(MVT::i32, MVT::Other), Op.getOperand(0),
2773 Op.getOperand(1), Val);
2777 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
2779 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
2780 Op.getOperand(1), DAG.getConstant(0, dl, MVT::i32));
2783 SDValue ARMTargetLowering::LowerEH_SJLJ_SETUP_DISPATCH(SDValue Op,
2784 SelectionDAG &DAG) const {
2786 return DAG.getNode(ARMISD::EH_SJLJ_SETUP_DISPATCH, dl, MVT::Other,
2791 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
2792 const ARMSubtarget *Subtarget) const {
2793 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2796 default: return SDValue(); // Don't custom lower most intrinsics.
2797 case Intrinsic::arm_rbit: {
2798 assert(Op.getOperand(1).getValueType() == MVT::i32 &&
2799 "RBIT intrinsic must have i32 type!");
2800 return DAG.getNode(ARMISD::RBIT, dl, MVT::i32, Op.getOperand(1));
2802 case Intrinsic::arm_thread_pointer: {
2803 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2804 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2806 case Intrinsic::eh_sjlj_lsda: {
2807 MachineFunction &MF = DAG.getMachineFunction();
2808 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2809 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2810 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2811 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2813 unsigned PCAdj = (RelocM != Reloc::PIC_)
2814 ? 0 : (Subtarget->isThumb() ? 4 : 8);
2815 ARMConstantPoolValue *CPV =
2816 ARMConstantPoolConstant::Create(MF.getFunction(), ARMPCLabelIndex,
2817 ARMCP::CPLSDA, PCAdj);
2818 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2819 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2820 SDValue Result = DAG.getLoad(
2821 PtrVT, dl, DAG.getEntryNode(), CPAddr,
2822 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), false,
2825 if (RelocM == Reloc::PIC_) {
2826 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, dl, MVT::i32);
2827 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2831 case Intrinsic::arm_neon_vmulls:
2832 case Intrinsic::arm_neon_vmullu: {
2833 unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vmulls)
2834 ? ARMISD::VMULLs : ARMISD::VMULLu;
2835 return DAG.getNode(NewOpc, SDLoc(Op), Op.getValueType(),
2836 Op.getOperand(1), Op.getOperand(2));
2838 case Intrinsic::arm_neon_vminnm:
2839 case Intrinsic::arm_neon_vmaxnm: {
2840 unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vminnm)
2841 ? ISD::FMINNUM : ISD::FMAXNUM;
2842 return DAG.getNode(NewOpc, SDLoc(Op), Op.getValueType(),
2843 Op.getOperand(1), Op.getOperand(2));
2845 case Intrinsic::arm_neon_vminu:
2846 case Intrinsic::arm_neon_vmaxu: {
2847 if (Op.getValueType().isFloatingPoint())
2849 unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vminu)
2850 ? ISD::UMIN : ISD::UMAX;
2851 return DAG.getNode(NewOpc, SDLoc(Op), Op.getValueType(),
2852 Op.getOperand(1), Op.getOperand(2));
2854 case Intrinsic::arm_neon_vmins:
2855 case Intrinsic::arm_neon_vmaxs: {
2856 // v{min,max}s is overloaded between signed integers and floats.
2857 if (!Op.getValueType().isFloatingPoint()) {
2858 unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vmins)
2859 ? ISD::SMIN : ISD::SMAX;
2860 return DAG.getNode(NewOpc, SDLoc(Op), Op.getValueType(),
2861 Op.getOperand(1), Op.getOperand(2));
2863 unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vmins)
2864 ? ISD::FMINNAN : ISD::FMAXNAN;
2865 return DAG.getNode(NewOpc, SDLoc(Op), Op.getValueType(),
2866 Op.getOperand(1), Op.getOperand(2));
2871 static SDValue LowerATOMIC_FENCE(SDValue Op, SelectionDAG &DAG,
2872 const ARMSubtarget *Subtarget) {
2873 // FIXME: handle "fence singlethread" more efficiently.
2875 if (!Subtarget->hasDataBarrier()) {
2876 // Some ARMv6 cpus can support data barriers with an mcr instruction.
2877 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2879 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2880 "Unexpected ISD::ATOMIC_FENCE encountered. Should be libcall!");
2881 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2882 DAG.getConstant(0, dl, MVT::i32));
2885 ConstantSDNode *OrdN = cast<ConstantSDNode>(Op.getOperand(1));
2886 AtomicOrdering Ord = static_cast<AtomicOrdering>(OrdN->getZExtValue());
2887 ARM_MB::MemBOpt Domain = ARM_MB::ISH;
2888 if (Subtarget->isMClass()) {
2889 // Only a full system barrier exists in the M-class architectures.
2890 Domain = ARM_MB::SY;
2891 } else if (Subtarget->isSwift() && Ord == Release) {
2892 // Swift happens to implement ISHST barriers in a way that's compatible with
2893 // Release semantics but weaker than ISH so we'd be fools not to use
2894 // it. Beware: other processors probably don't!
2895 Domain = ARM_MB::ISHST;
2898 return DAG.getNode(ISD::INTRINSIC_VOID, dl, MVT::Other, Op.getOperand(0),
2899 DAG.getConstant(Intrinsic::arm_dmb, dl, MVT::i32),
2900 DAG.getConstant(Domain, dl, MVT::i32));
2903 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG,
2904 const ARMSubtarget *Subtarget) {
2905 // ARM pre v5TE and Thumb1 does not have preload instructions.
2906 if (!(Subtarget->isThumb2() ||
2907 (!Subtarget->isThumb1Only() && Subtarget->hasV5TEOps())))
2908 // Just preserve the chain.
2909 return Op.getOperand(0);
2912 unsigned isRead = ~cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() & 1;
2914 (!Subtarget->hasV7Ops() || !Subtarget->hasMPExtension()))
2915 // ARMv7 with MP extension has PLDW.
2916 return Op.getOperand(0);
2918 unsigned isData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
2919 if (Subtarget->isThumb()) {
2921 isRead = ~isRead & 1;
2922 isData = ~isData & 1;
2925 return DAG.getNode(ARMISD::PRELOAD, dl, MVT::Other, Op.getOperand(0),
2926 Op.getOperand(1), DAG.getConstant(isRead, dl, MVT::i32),
2927 DAG.getConstant(isData, dl, MVT::i32));
2930 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
2931 MachineFunction &MF = DAG.getMachineFunction();
2932 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
2934 // vastart just stores the address of the VarArgsFrameIndex slot into the
2935 // memory location argument.
2937 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy(DAG.getDataLayout());
2938 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2939 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
2940 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
2941 MachinePointerInfo(SV), false, false, 0);
2945 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
2946 SDValue &Root, SelectionDAG &DAG,
2948 MachineFunction &MF = DAG.getMachineFunction();
2949 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2951 const TargetRegisterClass *RC;
2952 if (AFI->isThumb1OnlyFunction())
2953 RC = &ARM::tGPRRegClass;
2955 RC = &ARM::GPRRegClass;
2957 // Transform the arguments stored in physical registers into virtual ones.
2958 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2959 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2962 if (NextVA.isMemLoc()) {
2963 MachineFrameInfo *MFI = MF.getFrameInfo();
2964 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true);
2966 // Create load node to retrieve arguments from the stack.
2967 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
2968 ArgValue2 = DAG.getLoad(
2969 MVT::i32, dl, Root, FIN,
2970 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI), false,
2973 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
2974 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2976 if (!Subtarget->isLittle())
2977 std::swap (ArgValue, ArgValue2);
2978 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
2981 // The remaining GPRs hold either the beginning of variable-argument
2982 // data, or the beginning of an aggregate passed by value (usually
2983 // byval). Either way, we allocate stack slots adjacent to the data
2984 // provided by our caller, and store the unallocated registers there.
2985 // If this is a variadic function, the va_list pointer will begin with
2986 // these values; otherwise, this reassembles a (byval) structure that
2987 // was split between registers and memory.
2988 // Return: The frame index registers were stored into.
2990 ARMTargetLowering::StoreByValRegs(CCState &CCInfo, SelectionDAG &DAG,
2991 SDLoc dl, SDValue &Chain,
2992 const Value *OrigArg,
2993 unsigned InRegsParamRecordIdx,
2995 unsigned ArgSize) const {
2996 // Currently, two use-cases possible:
2997 // Case #1. Non-var-args function, and we meet first byval parameter.
2998 // Setup first unallocated register as first byval register;
2999 // eat all remained registers
3000 // (these two actions are performed by HandleByVal method).
3001 // Then, here, we initialize stack frame with
3002 // "store-reg" instructions.
3003 // Case #2. Var-args function, that doesn't contain byval parameters.
3004 // The same: eat all remained unallocated registers,
3005 // initialize stack frame.
3007 MachineFunction &MF = DAG.getMachineFunction();
3008 MachineFrameInfo *MFI = MF.getFrameInfo();
3009 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
3010 unsigned RBegin, REnd;
3011 if (InRegsParamRecordIdx < CCInfo.getInRegsParamsCount()) {
3012 CCInfo.getInRegsParamInfo(InRegsParamRecordIdx, RBegin, REnd);
3014 unsigned RBeginIdx = CCInfo.getFirstUnallocated(GPRArgRegs);
3015 RBegin = RBeginIdx == 4 ? (unsigned)ARM::R4 : GPRArgRegs[RBeginIdx];
3020 ArgOffset = -4 * (ARM::R4 - RBegin);
3022 auto PtrVT = getPointerTy(DAG.getDataLayout());
3023 int FrameIndex = MFI->CreateFixedObject(ArgSize, ArgOffset, false);
3024 SDValue FIN = DAG.getFrameIndex(FrameIndex, PtrVT);
3026 SmallVector<SDValue, 4> MemOps;
3027 const TargetRegisterClass *RC =
3028 AFI->isThumb1OnlyFunction() ? &ARM::tGPRRegClass : &ARM::GPRRegClass;
3030 for (unsigned Reg = RBegin, i = 0; Reg < REnd; ++Reg, ++i) {
3031 unsigned VReg = MF.addLiveIn(Reg, RC);
3032 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
3034 DAG.getStore(Val.getValue(1), dl, Val, FIN,
3035 MachinePointerInfo(OrigArg, 4 * i), false, false, 0);
3036 MemOps.push_back(Store);
3037 FIN = DAG.getNode(ISD::ADD, dl, PtrVT, FIN, DAG.getConstant(4, dl, PtrVT));
3040 if (!MemOps.empty())
3041 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOps);
3045 // Setup stack frame, the va_list pointer will start from.
3047 ARMTargetLowering::VarArgStyleRegisters(CCState &CCInfo, SelectionDAG &DAG,
3048 SDLoc dl, SDValue &Chain,
3050 unsigned TotalArgRegsSaveSize,
3051 bool ForceMutable) const {
3052 MachineFunction &MF = DAG.getMachineFunction();
3053 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
3055 // Try to store any remaining integer argument regs
3056 // to their spots on the stack so that they may be loaded by deferencing
3057 // the result of va_next.
3058 // If there is no regs to be stored, just point address after last
3059 // argument passed via stack.
3060 int FrameIndex = StoreByValRegs(CCInfo, DAG, dl, Chain, nullptr,
3061 CCInfo.getInRegsParamsCount(),
3062 CCInfo.getNextStackOffset(), 4);
3063 AFI->setVarArgsFrameIndex(FrameIndex);
3067 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
3068 CallingConv::ID CallConv, bool isVarArg,
3069 const SmallVectorImpl<ISD::InputArg>
3071 SDLoc dl, SelectionDAG &DAG,
3072 SmallVectorImpl<SDValue> &InVals)
3074 MachineFunction &MF = DAG.getMachineFunction();
3075 MachineFrameInfo *MFI = MF.getFrameInfo();
3077 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
3079 // Assign locations to all of the incoming arguments.
3080 SmallVector<CCValAssign, 16> ArgLocs;
3081 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs,
3082 *DAG.getContext(), Prologue);
3083 CCInfo.AnalyzeFormalArguments(Ins,
3084 CCAssignFnForNode(CallConv, /* Return*/ false,
3087 SmallVector<SDValue, 16> ArgValues;
3089 Function::const_arg_iterator CurOrigArg = MF.getFunction()->arg_begin();
3090 unsigned CurArgIdx = 0;
3092 // Initially ArgRegsSaveSize is zero.
3093 // Then we increase this value each time we meet byval parameter.
3094 // We also increase this value in case of varargs function.
3095 AFI->setArgRegsSaveSize(0);
3097 // Calculate the amount of stack space that we need to allocate to store
3098 // byval and variadic arguments that are passed in registers.
3099 // We need to know this before we allocate the first byval or variadic
3100 // argument, as they will be allocated a stack slot below the CFA (Canonical
3101 // Frame Address, the stack pointer at entry to the function).
3102 unsigned ArgRegBegin = ARM::R4;
3103 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3104 if (CCInfo.getInRegsParamsProcessed() >= CCInfo.getInRegsParamsCount())
3107 CCValAssign &VA = ArgLocs[i];
3108 unsigned Index = VA.getValNo();
3109 ISD::ArgFlagsTy Flags = Ins[Index].Flags;
3110 if (!Flags.isByVal())
3113 assert(VA.isMemLoc() && "unexpected byval pointer in reg");
3114 unsigned RBegin, REnd;
3115 CCInfo.getInRegsParamInfo(CCInfo.getInRegsParamsProcessed(), RBegin, REnd);
3116 ArgRegBegin = std::min(ArgRegBegin, RBegin);
3118 CCInfo.nextInRegsParam();
3120 CCInfo.rewindByValRegsInfo();
3122 int lastInsIndex = -1;
3123 if (isVarArg && MFI->hasVAStart()) {
3124 unsigned RegIdx = CCInfo.getFirstUnallocated(GPRArgRegs);
3125 if (RegIdx != array_lengthof(GPRArgRegs))
3126 ArgRegBegin = std::min(ArgRegBegin, (unsigned)GPRArgRegs[RegIdx]);
3129 unsigned TotalArgRegsSaveSize = 4 * (ARM::R4 - ArgRegBegin);
3130 AFI->setArgRegsSaveSize(TotalArgRegsSaveSize);
3131 auto PtrVT = getPointerTy(DAG.getDataLayout());
3133 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3134 CCValAssign &VA = ArgLocs[i];
3135 if (Ins[VA.getValNo()].isOrigArg()) {
3136 std::advance(CurOrigArg,
3137 Ins[VA.getValNo()].getOrigArgIndex() - CurArgIdx);
3138 CurArgIdx = Ins[VA.getValNo()].getOrigArgIndex();
3140 // Arguments stored in registers.
3141 if (VA.isRegLoc()) {
3142 EVT RegVT = VA.getLocVT();
3144 if (VA.needsCustom()) {
3145 // f64 and vector types are split up into multiple registers or
3146 // combinations of registers and stack slots.
3147 if (VA.getLocVT() == MVT::v2f64) {
3148 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
3150 VA = ArgLocs[++i]; // skip ahead to next loc
3152 if (VA.isMemLoc()) {
3153 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true);
3154 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
3155 ArgValue2 = DAG.getLoad(
3156 MVT::f64, dl, Chain, FIN,
3157 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI),
3158 false, false, false, 0);
3160 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
3163 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
3164 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
3165 ArgValue, ArgValue1,
3166 DAG.getIntPtrConstant(0, dl));
3167 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
3168 ArgValue, ArgValue2,
3169 DAG.getIntPtrConstant(1, dl));
3171 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
3174 const TargetRegisterClass *RC;
3176 if (RegVT == MVT::f32)
3177 RC = &ARM::SPRRegClass;
3178 else if (RegVT == MVT::f64)
3179 RC = &ARM::DPRRegClass;
3180 else if (RegVT == MVT::v2f64)
3181 RC = &ARM::QPRRegClass;
3182 else if (RegVT == MVT::i32)
3183 RC = AFI->isThumb1OnlyFunction() ? &ARM::tGPRRegClass
3184 : &ARM::GPRRegClass;
3186 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
3188 // Transform the arguments in physical registers into virtual ones.
3189 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
3190 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
3193 // If this is an 8 or 16-bit value, it is really passed promoted
3194 // to 32 bits. Insert an assert[sz]ext to capture this, then
3195 // truncate to the right size.
3196 switch (VA.getLocInfo()) {
3197 default: llvm_unreachable("Unknown loc info!");
3198 case CCValAssign::Full: break;
3199 case CCValAssign::BCvt:
3200 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
3202 case CCValAssign::SExt:
3203 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
3204 DAG.getValueType(VA.getValVT()));
3205 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
3207 case CCValAssign::ZExt:
3208 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
3209 DAG.getValueType(VA.getValVT()));
3210 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
3214 InVals.push_back(ArgValue);
3216 } else { // VA.isRegLoc()
3219 assert(VA.isMemLoc());
3220 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
3222 int index = VA.getValNo();
3224 // Some Ins[] entries become multiple ArgLoc[] entries.
3225 // Process them only once.
3226 if (index != lastInsIndex)
3228 ISD::ArgFlagsTy Flags = Ins[index].Flags;
3229 // FIXME: For now, all byval parameter objects are marked mutable.
3230 // This can be changed with more analysis.
3231 // In case of tail call optimization mark all arguments mutable.
3232 // Since they could be overwritten by lowering of arguments in case of
3234 if (Flags.isByVal()) {
3235 assert(Ins[index].isOrigArg() &&
3236 "Byval arguments cannot be implicit");
3237 unsigned CurByValIndex = CCInfo.getInRegsParamsProcessed();
3239 int FrameIndex = StoreByValRegs(
3240 CCInfo, DAG, dl, Chain, &*CurOrigArg, CurByValIndex,
3241 VA.getLocMemOffset(), Flags.getByValSize());
3242 InVals.push_back(DAG.getFrameIndex(FrameIndex, PtrVT));
3243 CCInfo.nextInRegsParam();
3245 unsigned FIOffset = VA.getLocMemOffset();
3246 int FI = MFI->CreateFixedObject(VA.getLocVT().getSizeInBits()/8,
3249 // Create load nodes to retrieve arguments from the stack.
3250 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
3251 InVals.push_back(DAG.getLoad(
3252 VA.getValVT(), dl, Chain, FIN,
3253 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI),
3254 false, false, false, 0));
3256 lastInsIndex = index;
3262 if (isVarArg && MFI->hasVAStart())
3263 VarArgStyleRegisters(CCInfo, DAG, dl, Chain,
3264 CCInfo.getNextStackOffset(),
3265 TotalArgRegsSaveSize);
3267 AFI->setArgumentStackSize(CCInfo.getNextStackOffset());
3272 /// isFloatingPointZero - Return true if this is +0.0.
3273 static bool isFloatingPointZero(SDValue Op) {
3274 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
3275 return CFP->getValueAPF().isPosZero();
3276 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
3277 // Maybe this has already been legalized into the constant pool?
3278 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
3279 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
3280 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
3281 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
3282 return CFP->getValueAPF().isPosZero();
3284 } else if (Op->getOpcode() == ISD::BITCAST &&
3285 Op->getValueType(0) == MVT::f64) {
3286 // Handle (ISD::BITCAST (ARMISD::VMOVIMM (ISD::TargetConstant 0)) MVT::f64)
3287 // created by LowerConstantFP().
3288 SDValue BitcastOp = Op->getOperand(0);
3289 if (BitcastOp->getOpcode() == ARMISD::VMOVIMM) {
3290 SDValue MoveOp = BitcastOp->getOperand(0);
3291 if (MoveOp->getOpcode() == ISD::TargetConstant &&
3292 cast<ConstantSDNode>(MoveOp)->getZExtValue() == 0) {
3300 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
3301 /// the given operands.
3303 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
3304 SDValue &ARMcc, SelectionDAG &DAG,
3306 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
3307 unsigned C = RHSC->getZExtValue();
3308 if (!isLegalICmpImmediate(C)) {
3309 // Constant does not fit, try adjusting it by one?
3314 if (C != 0x80000000 && isLegalICmpImmediate(C-1)) {
3315 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
3316 RHS = DAG.getConstant(C - 1, dl, MVT::i32);
3321 if (C != 0 && isLegalICmpImmediate(C-1)) {
3322 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
3323 RHS = DAG.getConstant(C - 1, dl, MVT::i32);
3328 if (C != 0x7fffffff && isLegalICmpImmediate(C+1)) {
3329 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
3330 RHS = DAG.getConstant(C + 1, dl, MVT::i32);
3335 if (C != 0xffffffff && isLegalICmpImmediate(C+1)) {
3336 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
3337 RHS = DAG.getConstant(C + 1, dl, MVT::i32);
3344 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3345 ARMISD::NodeType CompareType;
3348 CompareType = ARMISD::CMP;
3353 CompareType = ARMISD::CMPZ;
3356 ARMcc = DAG.getConstant(CondCode, dl, MVT::i32);
3357 return DAG.getNode(CompareType, dl, MVT::Glue, LHS, RHS);
3360 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
3362 ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
3364 assert(!Subtarget->isFPOnlySP() || RHS.getValueType() != MVT::f64);
3366 if (!isFloatingPointZero(RHS))
3367 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Glue, LHS, RHS);
3369 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Glue, LHS);
3370 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Glue, Cmp);
3373 /// duplicateCmp - Glue values can have only one use, so this function
3374 /// duplicates a comparison node.
3376 ARMTargetLowering::duplicateCmp(SDValue Cmp, SelectionDAG &DAG) const {
3377 unsigned Opc = Cmp.getOpcode();
3379 if (Opc == ARMISD::CMP || Opc == ARMISD::CMPZ)
3380 return DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
3382 assert(Opc == ARMISD::FMSTAT && "unexpected comparison operation");
3383 Cmp = Cmp.getOperand(0);
3384 Opc = Cmp.getOpcode();
3385 if (Opc == ARMISD::CMPFP)
3386 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
3388 assert(Opc == ARMISD::CMPFPw0 && "unexpected operand of FMSTAT");
3389 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0));
3391 return DAG.getNode(ARMISD::FMSTAT, DL, MVT::Glue, Cmp);
3394 std::pair<SDValue, SDValue>
3395 ARMTargetLowering::getARMXALUOOp(SDValue Op, SelectionDAG &DAG,
3396 SDValue &ARMcc) const {
3397 assert(Op.getValueType() == MVT::i32 && "Unsupported value type");
3399 SDValue Value, OverflowCmp;
3400 SDValue LHS = Op.getOperand(0);
3401 SDValue RHS = Op.getOperand(1);
3404 // FIXME: We are currently always generating CMPs because we don't support
3405 // generating CMN through the backend. This is not as good as the natural
3406 // CMP case because it causes a register dependency and cannot be folded
3409 switch (Op.getOpcode()) {
3411 llvm_unreachable("Unknown overflow instruction!");
3413 ARMcc = DAG.getConstant(ARMCC::VC, dl, MVT::i32);
3414 Value = DAG.getNode(ISD::ADD, dl, Op.getValueType(), LHS, RHS);
3415 OverflowCmp = DAG.getNode(ARMISD::CMP, dl, MVT::Glue, Value, LHS);
3418 ARMcc = DAG.getConstant(ARMCC::HS, dl, MVT::i32);
3419 Value = DAG.getNode(ISD::ADD, dl, Op.getValueType(), LHS, RHS);
3420 OverflowCmp = DAG.getNode(ARMISD::CMP, dl, MVT::Glue, Value, LHS);
3423 ARMcc = DAG.getConstant(ARMCC::VC, dl, MVT::i32);
3424 Value = DAG.getNode(ISD::SUB, dl, Op.getValueType(), LHS, RHS);
3425 OverflowCmp = DAG.getNode(ARMISD::CMP, dl, MVT::Glue, LHS, RHS);
3428 ARMcc = DAG.getConstant(ARMCC::HS, dl, MVT::i32);
3429 Value = DAG.getNode(ISD::SUB, dl, Op.getValueType(), LHS, RHS);
3430 OverflowCmp = DAG.getNode(ARMISD::CMP, dl, MVT::Glue, LHS, RHS);
3434 return std::make_pair(Value, OverflowCmp);
3439 ARMTargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) const {
3440 // Let legalize expand this if it isn't a legal type yet.
3441 if (!DAG.getTargetLoweringInfo().isTypeLegal(Op.getValueType()))
3444 SDValue Value, OverflowCmp;
3446 std::tie(Value, OverflowCmp) = getARMXALUOOp(Op, DAG, ARMcc);
3447 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3449 // We use 0 and 1 as false and true values.
3450 SDValue TVal = DAG.getConstant(1, dl, MVT::i32);
3451 SDValue FVal = DAG.getConstant(0, dl, MVT::i32);
3452 EVT VT = Op.getValueType();
3454 SDValue Overflow = DAG.getNode(ARMISD::CMOV, dl, VT, TVal, FVal,
3455 ARMcc, CCR, OverflowCmp);
3457 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
3458 return DAG.getNode(ISD::MERGE_VALUES, dl, VTs, Value, Overflow);
3462 SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
3463 SDValue Cond = Op.getOperand(0);
3464 SDValue SelectTrue = Op.getOperand(1);
3465 SDValue SelectFalse = Op.getOperand(2);
3467 unsigned Opc = Cond.getOpcode();
3469 if (Cond.getResNo() == 1 &&
3470 (Opc == ISD::SADDO || Opc == ISD::UADDO || Opc == ISD::SSUBO ||
3471 Opc == ISD::USUBO)) {
3472 if (!DAG.getTargetLoweringInfo().isTypeLegal(Cond->getValueType(0)))
3475 SDValue Value, OverflowCmp;
3477 std::tie(Value, OverflowCmp) = getARMXALUOOp(Cond, DAG, ARMcc);
3478 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3479 EVT VT = Op.getValueType();
3481 return getCMOV(dl, VT, SelectTrue, SelectFalse, ARMcc, CCR,
3487 // (select (cmov 1, 0, cond), t, f) -> (cmov t, f, cond)
3488 // (select (cmov 0, 1, cond), t, f) -> (cmov f, t, cond)
3490 if (Cond.getOpcode() == ARMISD::CMOV && Cond.hasOneUse()) {
3491 const ConstantSDNode *CMOVTrue =
3492 dyn_cast<ConstantSDNode>(Cond.getOperand(0));
3493 const ConstantSDNode *CMOVFalse =
3494 dyn_cast<ConstantSDNode>(Cond.getOperand(1));
3496 if (CMOVTrue && CMOVFalse) {
3497 unsigned CMOVTrueVal = CMOVTrue->getZExtValue();
3498 unsigned CMOVFalseVal = CMOVFalse->getZExtValue();
3502 if (CMOVTrueVal == 1 && CMOVFalseVal == 0) {
3504 False = SelectFalse;
3505 } else if (CMOVTrueVal == 0 && CMOVFalseVal == 1) {
3510 if (True.getNode() && False.getNode()) {
3511 EVT VT = Op.getValueType();
3512 SDValue ARMcc = Cond.getOperand(2);
3513 SDValue CCR = Cond.getOperand(3);
3514 SDValue Cmp = duplicateCmp(Cond.getOperand(4), DAG);
3515 assert(True.getValueType() == VT);
3516 return getCMOV(dl, VT, True, False, ARMcc, CCR, Cmp, DAG);
3521 // ARM's BooleanContents value is UndefinedBooleanContent. Mask out the
3522 // undefined bits before doing a full-word comparison with zero.
3523 Cond = DAG.getNode(ISD::AND, dl, Cond.getValueType(), Cond,
3524 DAG.getConstant(1, dl, Cond.getValueType()));
3526 return DAG.getSelectCC(dl, Cond,
3527 DAG.getConstant(0, dl, Cond.getValueType()),
3528 SelectTrue, SelectFalse, ISD::SETNE);
3531 static void checkVSELConstraints(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
3532 bool &swpCmpOps, bool &swpVselOps) {
3533 // Start by selecting the GE condition code for opcodes that return true for
3535 if (CC == ISD::SETUGE || CC == ISD::SETOGE || CC == ISD::SETOLE ||
3537 CondCode = ARMCC::GE;
3539 // and GT for opcodes that return false for 'equality'.
3540 else if (CC == ISD::SETUGT || CC == ISD::SETOGT || CC == ISD::SETOLT ||
3542 CondCode = ARMCC::GT;
3544 // Since we are constrained to GE/GT, if the opcode contains 'less', we need
3545 // to swap the compare operands.
3546 if (CC == ISD::SETOLE || CC == ISD::SETULE || CC == ISD::SETOLT ||
3550 // Both GT and GE are ordered comparisons, and return false for 'unordered'.
3551 // If we have an unordered opcode, we need to swap the operands to the VSEL
3552 // instruction (effectively negating the condition).
3554 // This also has the effect of swapping which one of 'less' or 'greater'
3555 // returns true, so we also swap the compare operands. It also switches
3556 // whether we return true for 'equality', so we compensate by picking the
3557 // opposite condition code to our original choice.
3558 if (CC == ISD::SETULE || CC == ISD::SETULT || CC == ISD::SETUGE ||
3559 CC == ISD::SETUGT) {
3560 swpCmpOps = !swpCmpOps;
3561 swpVselOps = !swpVselOps;
3562 CondCode = CondCode == ARMCC::GT ? ARMCC::GE : ARMCC::GT;
3565 // 'ordered' is 'anything but unordered', so use the VS condition code and
3566 // swap the VSEL operands.
3567 if (CC == ISD::SETO) {
3568 CondCode = ARMCC::VS;
3572 // 'unordered or not equal' is 'anything but equal', so use the EQ condition
3573 // code and swap the VSEL operands.
3574 if (CC == ISD::SETUNE) {
3575 CondCode = ARMCC::EQ;
3580 SDValue ARMTargetLowering::getCMOV(SDLoc dl, EVT VT, SDValue FalseVal,
3581 SDValue TrueVal, SDValue ARMcc, SDValue CCR,
3582 SDValue Cmp, SelectionDAG &DAG) const {
3583 if (Subtarget->isFPOnlySP() && VT == MVT::f64) {
3584 FalseVal = DAG.getNode(ARMISD::VMOVRRD, dl,
3585 DAG.getVTList(MVT::i32, MVT::i32), FalseVal);
3586 TrueVal = DAG.getNode(ARMISD::VMOVRRD, dl,
3587 DAG.getVTList(MVT::i32, MVT::i32), TrueVal);
3589 SDValue TrueLow = TrueVal.getValue(0);
3590 SDValue TrueHigh = TrueVal.getValue(1);
3591 SDValue FalseLow = FalseVal.getValue(0);
3592 SDValue FalseHigh = FalseVal.getValue(1);
3594 SDValue Low = DAG.getNode(ARMISD::CMOV, dl, MVT::i32, FalseLow, TrueLow,
3596 SDValue High = DAG.getNode(ARMISD::CMOV, dl, MVT::i32, FalseHigh, TrueHigh,
3597 ARMcc, CCR, duplicateCmp(Cmp, DAG));
3599 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Low, High);
3601 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,
3606 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
3607 EVT VT = Op.getValueType();
3608 SDValue LHS = Op.getOperand(0);
3609 SDValue RHS = Op.getOperand(1);
3610 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
3611 SDValue TrueVal = Op.getOperand(2);
3612 SDValue FalseVal = Op.getOperand(3);
3615 if (Subtarget->isFPOnlySP() && LHS.getValueType() == MVT::f64) {
3616 DAG.getTargetLoweringInfo().softenSetCCOperands(DAG, MVT::f64, LHS, RHS, CC,
3619 // If softenSetCCOperands only returned one value, we should compare it to
3621 if (!RHS.getNode()) {
3622 RHS = DAG.getConstant(0, dl, LHS.getValueType());
3627 if (LHS.getValueType() == MVT::i32) {
3628 // Try to generate VSEL on ARMv8.
3629 // The VSEL instruction can't use all the usual ARM condition
3630 // codes: it only has two bits to select the condition code, so it's
3631 // constrained to use only GE, GT, VS and EQ.
3633 // To implement all the various ISD::SETXXX opcodes, we sometimes need to
3634 // swap the operands of the previous compare instruction (effectively
3635 // inverting the compare condition, swapping 'less' and 'greater') and
3636 // sometimes need to swap the operands to the VSEL (which inverts the
3637 // condition in the sense of firing whenever the previous condition didn't)
3638 if (Subtarget->hasFPARMv8() && (TrueVal.getValueType() == MVT::f32 ||
3639 TrueVal.getValueType() == MVT::f64)) {
3640 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3641 if (CondCode == ARMCC::LT || CondCode == ARMCC::LE ||
3642 CondCode == ARMCC::VC || CondCode == ARMCC::NE) {
3643 CC = ISD::getSetCCInverse(CC, true);
3644 std::swap(TrueVal, FalseVal);
3649 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3650 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3651 return getCMOV(dl, VT, FalseVal, TrueVal, ARMcc, CCR, Cmp, DAG);
3654 ARMCC::CondCodes CondCode, CondCode2;
3655 FPCCToARMCC(CC, CondCode, CondCode2);
3657 // Try to generate VMAXNM/VMINNM on ARMv8.
3658 if (Subtarget->hasFPARMv8() && (TrueVal.getValueType() == MVT::f32 ||
3659 TrueVal.getValueType() == MVT::f64)) {
3660 bool swpCmpOps = false;
3661 bool swpVselOps = false;
3662 checkVSELConstraints(CC, CondCode, swpCmpOps, swpVselOps);
3664 if (CondCode == ARMCC::GT || CondCode == ARMCC::GE ||
3665 CondCode == ARMCC::VS || CondCode == ARMCC::EQ) {
3667 std::swap(LHS, RHS);
3669 std::swap(TrueVal, FalseVal);
3673 SDValue ARMcc = DAG.getConstant(CondCode, dl, MVT::i32);
3674 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
3675 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3676 SDValue Result = getCMOV(dl, VT, FalseVal, TrueVal, ARMcc, CCR, Cmp, DAG);
3677 if (CondCode2 != ARMCC::AL) {
3678 SDValue ARMcc2 = DAG.getConstant(CondCode2, dl, MVT::i32);
3679 // FIXME: Needs another CMP because flag can have but one use.
3680 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
3681 Result = getCMOV(dl, VT, Result, TrueVal, ARMcc2, CCR, Cmp2, DAG);
3686 /// canChangeToInt - Given the fp compare operand, return true if it is suitable
3687 /// to morph to an integer compare sequence.
3688 static bool canChangeToInt(SDValue Op, bool &SeenZero,
3689 const ARMSubtarget *Subtarget) {
3690 SDNode *N = Op.getNode();
3691 if (!N->hasOneUse())
3692 // Otherwise it requires moving the value from fp to integer registers.
3694 if (!N->getNumValues())
3696 EVT VT = Op.getValueType();
3697 if (VT != MVT::f32 && !Subtarget->isFPBrccSlow())
3698 // f32 case is generally profitable. f64 case only makes sense when vcmpe +
3699 // vmrs are very slow, e.g. cortex-a8.
3702 if (isFloatingPointZero(Op)) {
3706 return ISD::isNormalLoad(N);
3709 static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) {
3710 if (isFloatingPointZero(Op))
3711 return DAG.getConstant(0, SDLoc(Op), MVT::i32);
3713 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op))
3714 return DAG.getLoad(MVT::i32, SDLoc(Op),
3715 Ld->getChain(), Ld->getBasePtr(), Ld->getPointerInfo(),
3716 Ld->isVolatile(), Ld->isNonTemporal(),
3717 Ld->isInvariant(), Ld->getAlignment());
3719 llvm_unreachable("Unknown VFP cmp argument!");
3722 static void expandf64Toi32(SDValue Op, SelectionDAG &DAG,
3723 SDValue &RetVal1, SDValue &RetVal2) {
3726 if (isFloatingPointZero(Op)) {
3727 RetVal1 = DAG.getConstant(0, dl, MVT::i32);
3728 RetVal2 = DAG.getConstant(0, dl, MVT::i32);
3732 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) {
3733 SDValue Ptr = Ld->getBasePtr();
3734 RetVal1 = DAG.getLoad(MVT::i32, dl,
3735 Ld->getChain(), Ptr,
3736 Ld->getPointerInfo(),
3737 Ld->isVolatile(), Ld->isNonTemporal(),
3738 Ld->isInvariant(), Ld->getAlignment());
3740 EVT PtrType = Ptr.getValueType();
3741 unsigned NewAlign = MinAlign(Ld->getAlignment(), 4);
3742 SDValue NewPtr = DAG.getNode(ISD::ADD, dl,
3743 PtrType, Ptr, DAG.getConstant(4, dl, PtrType));
3744 RetVal2 = DAG.getLoad(MVT::i32, dl,
3745 Ld->getChain(), NewPtr,
3746 Ld->getPointerInfo().getWithOffset(4),
3747 Ld->isVolatile(), Ld->isNonTemporal(),
3748 Ld->isInvariant(), NewAlign);
3752 llvm_unreachable("Unknown VFP cmp argument!");
3755 /// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some
3756 /// f32 and even f64 comparisons to integer ones.
3758 ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const {
3759 SDValue Chain = Op.getOperand(0);
3760 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3761 SDValue LHS = Op.getOperand(2);
3762 SDValue RHS = Op.getOperand(3);
3763 SDValue Dest = Op.getOperand(4);
3766 bool LHSSeenZero = false;
3767 bool LHSOk = canChangeToInt(LHS, LHSSeenZero, Subtarget);
3768 bool RHSSeenZero = false;
3769 bool RHSOk = canChangeToInt(RHS, RHSSeenZero, Subtarget);
3770 if (LHSOk && RHSOk && (LHSSeenZero || RHSSeenZero)) {
3771 // If unsafe fp math optimization is enabled and there are no other uses of
3772 // the CMP operands, and the condition code is EQ or NE, we can optimize it
3773 // to an integer comparison.
3774 if (CC == ISD::SETOEQ)
3776 else if (CC == ISD::SETUNE)
3779 SDValue Mask = DAG.getConstant(0x7fffffff, dl, MVT::i32);
3781 if (LHS.getValueType() == MVT::f32) {
3782 LHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3783 bitcastf32Toi32(LHS, DAG), Mask);
3784 RHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3785 bitcastf32Toi32(RHS, DAG), Mask);
3786 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3787 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3788 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3789 Chain, Dest, ARMcc, CCR, Cmp);
3794 expandf64Toi32(LHS, DAG, LHS1, LHS2);
3795 expandf64Toi32(RHS, DAG, RHS1, RHS2);
3796 LHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, LHS2, Mask);
3797 RHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, RHS2, Mask);
3798 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3799 ARMcc = DAG.getConstant(CondCode, dl, MVT::i32);
3800 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3801 SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest };
3802 return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops);
3808 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
3809 SDValue Chain = Op.getOperand(0);
3810 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3811 SDValue LHS = Op.getOperand(2);
3812 SDValue RHS = Op.getOperand(3);
3813 SDValue Dest = Op.getOperand(4);
3816 if (Subtarget->isFPOnlySP() && LHS.getValueType() == MVT::f64) {
3817 DAG.getTargetLoweringInfo().softenSetCCOperands(DAG, MVT::f64, LHS, RHS, CC,
3820 // If softenSetCCOperands only returned one value, we should compare it to
3822 if (!RHS.getNode()) {
3823 RHS = DAG.getConstant(0, dl, LHS.getValueType());
3828 if (LHS.getValueType() == MVT::i32) {
3830 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3831 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3832 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3833 Chain, Dest, ARMcc, CCR, Cmp);
3836 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3838 if (getTargetMachine().Options.UnsafeFPMath &&
3839 (CC == ISD::SETEQ || CC == ISD::SETOEQ ||
3840 CC == ISD::SETNE || CC == ISD::SETUNE)) {
3841 SDValue Result = OptimizeVFPBrcond(Op, DAG);
3842 if (Result.getNode())
3846 ARMCC::CondCodes CondCode, CondCode2;
3847 FPCCToARMCC(CC, CondCode, CondCode2);
3849 SDValue ARMcc = DAG.getConstant(CondCode, dl, MVT::i32);
3850 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
3851 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3852 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3853 SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp };
3854 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops);
3855 if (CondCode2 != ARMCC::AL) {
3856 ARMcc = DAG.getConstant(CondCode2, dl, MVT::i32);
3857 SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) };
3858 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops);
3863 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
3864 SDValue Chain = Op.getOperand(0);
3865 SDValue Table = Op.getOperand(1);
3866 SDValue Index = Op.getOperand(2);
3869 EVT PTy = getPointerTy(DAG.getDataLayout());
3870 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
3871 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
3872 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI);
3873 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, dl, PTy));
3874 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
3875 if (Subtarget->isThumb2()) {
3876 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
3877 // which does another jump to the destination. This also makes it easier
3878 // to translate it to TBB / TBH later.
3879 // FIXME: This might not work if the function is extremely large.
3880 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
3881 Addr, Op.getOperand(2), JTI);
3883 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
3885 DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
3886 MachinePointerInfo::getJumpTable(DAG.getMachineFunction()),
3887 false, false, false, 0);
3888 Chain = Addr.getValue(1);
3889 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
3890 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI);
3893 DAG.getLoad(PTy, dl, Chain, Addr,
3894 MachinePointerInfo::getJumpTable(DAG.getMachineFunction()),
3895 false, false, false, 0);
3896 Chain = Addr.getValue(1);
3897 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI);
3901 static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3902 EVT VT = Op.getValueType();
3905 if (Op.getValueType().getVectorElementType() == MVT::i32) {
3906 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::f32)
3908 return DAG.UnrollVectorOp(Op.getNode());
3911 assert(Op.getOperand(0).getValueType() == MVT::v4f32 &&
3912 "Invalid type for custom lowering!");
3913 if (VT != MVT::v4i16)
3914 return DAG.UnrollVectorOp(Op.getNode());
3916 Op = DAG.getNode(Op.getOpcode(), dl, MVT::v4i32, Op.getOperand(0));
3917 return DAG.getNode(ISD::TRUNCATE, dl, VT, Op);
3920 SDValue ARMTargetLowering::LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) const {
3921 EVT VT = Op.getValueType();
3923 return LowerVectorFP_TO_INT(Op, DAG);
3924 if (Subtarget->isFPOnlySP() && Op.getOperand(0).getValueType() == MVT::f64) {
3926 if (Op.getOpcode() == ISD::FP_TO_SINT)
3927 LC = RTLIB::getFPTOSINT(Op.getOperand(0).getValueType(),
3930 LC = RTLIB::getFPTOUINT(Op.getOperand(0).getValueType(),
3932 return makeLibCall(DAG, LC, Op.getValueType(), Op.getOperand(0),
3933 /*isSigned*/ false, SDLoc(Op)).first;
3939 static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3940 EVT VT = Op.getValueType();
3943 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::i32) {
3944 if (VT.getVectorElementType() == MVT::f32)
3946 return DAG.UnrollVectorOp(Op.getNode());
3949 assert(Op.getOperand(0).getValueType() == MVT::v4i16 &&
3950 "Invalid type for custom lowering!");
3951 if (VT != MVT::v4f32)
3952 return DAG.UnrollVectorOp(Op.getNode());
3956 switch (Op.getOpcode()) {
3957 default: llvm_unreachable("Invalid opcode!");
3958 case ISD::SINT_TO_FP:
3959 CastOpc = ISD::SIGN_EXTEND;
3960 Opc = ISD::SINT_TO_FP;
3962 case ISD::UINT_TO_FP:
3963 CastOpc = ISD::ZERO_EXTEND;
3964 Opc = ISD::UINT_TO_FP;
3968 Op = DAG.getNode(CastOpc, dl, MVT::v4i32, Op.getOperand(0));
3969 return DAG.getNode(Opc, dl, VT, Op);
3972 SDValue ARMTargetLowering::LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) const {
3973 EVT VT = Op.getValueType();
3975 return LowerVectorINT_TO_FP(Op, DAG);
3976 if (Subtarget->isFPOnlySP() && Op.getValueType() == MVT::f64) {
3978 if (Op.getOpcode() == ISD::SINT_TO_FP)
3979 LC = RTLIB::getSINTTOFP(Op.getOperand(0).getValueType(),
3982 LC = RTLIB::getUINTTOFP(Op.getOperand(0).getValueType(),
3984 return makeLibCall(DAG, LC, Op.getValueType(), Op.getOperand(0),
3985 /*isSigned*/ false, SDLoc(Op)).first;
3991 SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
3992 // Implement fcopysign with a fabs and a conditional fneg.
3993 SDValue Tmp0 = Op.getOperand(0);
3994 SDValue Tmp1 = Op.getOperand(1);
3996 EVT VT = Op.getValueType();
3997 EVT SrcVT = Tmp1.getValueType();
3998 bool InGPR = Tmp0.getOpcode() == ISD::BITCAST ||
3999 Tmp0.getOpcode() == ARMISD::VMOVDRR;
4000 bool UseNEON = !InGPR && Subtarget->hasNEON();
4003 // Use VBSL to copy the sign bit.
4004 unsigned EncodedVal = ARM_AM::createNEONModImm(0x6, 0x80);
4005 SDValue Mask = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v2i32,
4006 DAG.getTargetConstant(EncodedVal, dl, MVT::i32));
4007 EVT OpVT = (VT == MVT::f32) ? MVT::v2i32 : MVT::v1i64;
4009 Mask = DAG.getNode(ARMISD::VSHL, dl, OpVT,
4010 DAG.getNode(ISD::BITCAST, dl, OpVT, Mask),
4011 DAG.getConstant(32, dl, MVT::i32));
4012 else /*if (VT == MVT::f32)*/
4013 Tmp0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp0);
4014 if (SrcVT == MVT::f32) {
4015 Tmp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp1);
4017 Tmp1 = DAG.getNode(ARMISD::VSHL, dl, OpVT,
4018 DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1),
4019 DAG.getConstant(32, dl, MVT::i32));
4020 } else if (VT == MVT::f32)
4021 Tmp1 = DAG.getNode(ARMISD::VSHRu, dl, MVT::v1i64,
4022 DAG.getNode(ISD::BITCAST, dl, MVT::v1i64, Tmp1),
4023 DAG.getConstant(32, dl, MVT::i32));
4024 Tmp0 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp0);
4025 Tmp1 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1);
4027 SDValue AllOnes = DAG.getTargetConstant(ARM_AM::createNEONModImm(0xe, 0xff),
4029 AllOnes = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v8i8, AllOnes);
4030 SDValue MaskNot = DAG.getNode(ISD::XOR, dl, OpVT, Mask,
4031 DAG.getNode(ISD::BITCAST, dl, OpVT, AllOnes));
4033 SDValue Res = DAG.getNode(ISD::OR, dl, OpVT,
4034 DAG.getNode(ISD::AND, dl, OpVT, Tmp1, Mask),
4035 DAG.getNode(ISD::AND, dl, OpVT, Tmp0, MaskNot));
4036 if (VT == MVT::f32) {
4037 Res = DAG.getNode(ISD::BITCAST, dl, MVT::v2f32, Res);
4038 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, Res,
4039 DAG.getConstant(0, dl, MVT::i32));
4041 Res = DAG.getNode(ISD::BITCAST, dl, MVT::f64, Res);
4047 // Bitcast operand 1 to i32.
4048 if (SrcVT == MVT::f64)
4049 Tmp1 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
4051 Tmp1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp1);
4053 // Or in the signbit with integer operations.
4054 SDValue Mask1 = DAG.getConstant(0x80000000, dl, MVT::i32);
4055 SDValue Mask2 = DAG.getConstant(0x7fffffff, dl, MVT::i32);
4056 Tmp1 = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp1, Mask1);
4057 if (VT == MVT::f32) {
4058 Tmp0 = DAG.getNode(ISD::AND, dl, MVT::i32,
4059 DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp0), Mask2);
4060 return DAG.getNode(ISD::BITCAST, dl, MVT::f32,
4061 DAG.getNode(ISD::OR, dl, MVT::i32, Tmp0, Tmp1));
4064 // f64: Or the high part with signbit and then combine two parts.
4065 Tmp0 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
4067 SDValue Lo = Tmp0.getValue(0);
4068 SDValue Hi = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp0.getValue(1), Mask2);
4069 Hi = DAG.getNode(ISD::OR, dl, MVT::i32, Hi, Tmp1);
4070 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
4073 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
4074 MachineFunction &MF = DAG.getMachineFunction();
4075 MachineFrameInfo *MFI = MF.getFrameInfo();
4076 MFI->setReturnAddressIsTaken(true);
4078 if (verifyReturnAddressArgumentIsConstant(Op, DAG))
4081 EVT VT = Op.getValueType();
4083 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
4085 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
4086 SDValue Offset = DAG.getConstant(4, dl, MVT::i32);
4087 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
4088 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
4089 MachinePointerInfo(), false, false, false, 0);
4092 // Return LR, which contains the return address. Mark it an implicit live-in.
4093 unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32));
4094 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
4097 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
4098 const ARMBaseRegisterInfo &ARI =
4099 *static_cast<const ARMBaseRegisterInfo*>(RegInfo);
4100 MachineFunction &MF = DAG.getMachineFunction();
4101 MachineFrameInfo *MFI = MF.getFrameInfo();
4102 MFI->setFrameAddressIsTaken(true);
4104 EVT VT = Op.getValueType();
4105 SDLoc dl(Op); // FIXME probably not meaningful
4106 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
4107 unsigned FrameReg = ARI.getFrameRegister(MF);
4108 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
4110 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
4111 MachinePointerInfo(),
4112 false, false, false, 0);
4116 // FIXME? Maybe this could be a TableGen attribute on some registers and
4117 // this table could be generated automatically from RegInfo.
4118 unsigned ARMTargetLowering::getRegisterByName(const char* RegName, EVT VT,
4119 SelectionDAG &DAG) const {
4120 unsigned Reg = StringSwitch<unsigned>(RegName)
4121 .Case("sp", ARM::SP)
4125 report_fatal_error(Twine("Invalid register name \""
4126 + StringRef(RegName) + "\"."));
4129 // Result is 64 bit value so split into two 32 bit values and return as a
4131 static void ExpandREAD_REGISTER(SDNode *N, SmallVectorImpl<SDValue> &Results,
4132 SelectionDAG &DAG) {
4135 // This function is only supposed to be called for i64 type destination.
4136 assert(N->getValueType(0) == MVT::i64
4137 && "ExpandREAD_REGISTER called for non-i64 type result.");
4139 SDValue Read = DAG.getNode(ISD::READ_REGISTER, DL,
4140 DAG.getVTList(MVT::i32, MVT::i32, MVT::Other),
4144 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Read.getValue(0),
4146 Results.push_back(Read.getOperand(0));
4149 /// ExpandBITCAST - If the target supports VFP, this function is called to
4150 /// expand a bit convert where either the source or destination type is i64 to
4151 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
4152 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
4153 /// vectors), since the legalizer won't know what to do with that.
4154 static SDValue ExpandBITCAST(SDNode *N, SelectionDAG &DAG) {
4155 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4157 SDValue Op = N->getOperand(0);
4159 // This function is only supposed to be called for i64 types, either as the
4160 // source or destination of the bit convert.
4161 EVT SrcVT = Op.getValueType();
4162 EVT DstVT = N->getValueType(0);
4163 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
4164 "ExpandBITCAST called for non-i64 type");
4166 // Turn i64->f64 into VMOVDRR.
4167 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
4168 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
4169 DAG.getConstant(0, dl, MVT::i32));
4170 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
4171 DAG.getConstant(1, dl, MVT::i32));
4172 return DAG.getNode(ISD::BITCAST, dl, DstVT,
4173 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
4176 // Turn f64->i64 into VMOVRRD.
4177 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
4179 if (DAG.getDataLayout().isBigEndian() && SrcVT.isVector() &&
4180 SrcVT.getVectorNumElements() > 1)
4181 Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
4182 DAG.getVTList(MVT::i32, MVT::i32),
4183 DAG.getNode(ARMISD::VREV64, dl, SrcVT, Op));
4185 Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
4186 DAG.getVTList(MVT::i32, MVT::i32), Op);
4187 // Merge the pieces into a single i64 value.
4188 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
4194 /// getZeroVector - Returns a vector of specified type with all zero elements.
4195 /// Zero vectors are used to represent vector negation and in those cases
4196 /// will be implemented with the NEON VNEG instruction. However, VNEG does
4197 /// not support i64 elements, so sometimes the zero vectors will need to be
4198 /// explicitly constructed. Regardless, use a canonical VMOV to create the
4200 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, SDLoc dl) {
4201 assert(VT.isVector() && "Expected a vector type");
4202 // The canonical modified immediate encoding of a zero vector is....0!
4203 SDValue EncodedVal = DAG.getTargetConstant(0, dl, MVT::i32);
4204 EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
4205 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal);
4206 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4209 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
4210 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
4211 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
4212 SelectionDAG &DAG) const {
4213 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4214 EVT VT = Op.getValueType();
4215 unsigned VTBits = VT.getSizeInBits();
4217 SDValue ShOpLo = Op.getOperand(0);
4218 SDValue ShOpHi = Op.getOperand(1);
4219 SDValue ShAmt = Op.getOperand(2);
4221 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
4223 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
4225 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
4226 DAG.getConstant(VTBits, dl, MVT::i32), ShAmt);
4227 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
4228 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
4229 DAG.getConstant(VTBits, dl, MVT::i32));
4230 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
4231 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
4232 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
4234 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
4235 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, dl, MVT::i32),
4236 ISD::SETGE, ARMcc, DAG, dl);
4237 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
4238 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc,
4241 SDValue Ops[2] = { Lo, Hi };
4242 return DAG.getMergeValues(Ops, dl);
4245 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
4246 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
4247 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
4248 SelectionDAG &DAG) const {
4249 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4250 EVT VT = Op.getValueType();
4251 unsigned VTBits = VT.getSizeInBits();
4253 SDValue ShOpLo = Op.getOperand(0);
4254 SDValue ShOpHi = Op.getOperand(1);
4255 SDValue ShAmt = Op.getOperand(2);
4258 assert(Op.getOpcode() == ISD::SHL_PARTS);
4259 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
4260 DAG.getConstant(VTBits, dl, MVT::i32), ShAmt);
4261 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
4262 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
4263 DAG.getConstant(VTBits, dl, MVT::i32));
4264 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
4265 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
4267 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
4268 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
4269 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, dl, MVT::i32),
4270 ISD::SETGE, ARMcc, DAG, dl);
4271 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
4272 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc,
4275 SDValue Ops[2] = { Lo, Hi };
4276 return DAG.getMergeValues(Ops, dl);
4279 SDValue ARMTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
4280 SelectionDAG &DAG) const {
4281 // The rounding mode is in bits 23:22 of the FPSCR.
4282 // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0
4283 // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3)
4284 // so that the shift + and get folded into a bitfield extract.
4286 SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::i32,
4287 DAG.getConstant(Intrinsic::arm_get_fpscr, dl,
4289 SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPSCR,
4290 DAG.getConstant(1U << 22, dl, MVT::i32));
4291 SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds,
4292 DAG.getConstant(22, dl, MVT::i32));
4293 return DAG.getNode(ISD::AND, dl, MVT::i32, RMODE,
4294 DAG.getConstant(3, dl, MVT::i32));
4297 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
4298 const ARMSubtarget *ST) {
4300 EVT VT = N->getValueType(0);
4301 if (VT.isVector()) {
4302 assert(ST->hasNEON());
4304 // Compute the least significant set bit: LSB = X & -X
4305 SDValue X = N->getOperand(0);
4306 SDValue NX = DAG.getNode(ISD::SUB, dl, VT, getZeroVector(VT, DAG, dl), X);
4307 SDValue LSB = DAG.getNode(ISD::AND, dl, VT, X, NX);
4309 EVT ElemTy = VT.getVectorElementType();
4311 if (ElemTy == MVT::i8) {
4312 // Compute with: cttz(x) = ctpop(lsb - 1)
4313 SDValue One = DAG.getNode(ARMISD::VMOVIMM, dl, VT,
4314 DAG.getTargetConstant(1, dl, ElemTy));
4315 SDValue Bits = DAG.getNode(ISD::SUB, dl, VT, LSB, One);
4316 return DAG.getNode(ISD::CTPOP, dl, VT, Bits);
4319 if ((ElemTy == MVT::i16 || ElemTy == MVT::i32) &&
4320 (N->getOpcode() == ISD::CTTZ_ZERO_UNDEF)) {
4321 // Compute with: cttz(x) = (width - 1) - ctlz(lsb), if x != 0
4322 unsigned NumBits = ElemTy.getSizeInBits();
4323 SDValue WidthMinus1 =
4324 DAG.getNode(ARMISD::VMOVIMM, dl, VT,
4325 DAG.getTargetConstant(NumBits - 1, dl, ElemTy));
4326 SDValue CTLZ = DAG.getNode(ISD::CTLZ, dl, VT, LSB);
4327 return DAG.getNode(ISD::SUB, dl, VT, WidthMinus1, CTLZ);
4330 // Compute with: cttz(x) = ctpop(lsb - 1)
4332 // Since we can only compute the number of bits in a byte with vcnt.8, we
4333 // have to gather the result with pairwise addition (vpaddl) for i16, i32,
4338 if (ElemTy == MVT::i64) {
4339 // Load constant 0xffff'ffff'ffff'ffff to register.
4340 SDValue FF = DAG.getNode(ARMISD::VMOVIMM, dl, VT,
4341 DAG.getTargetConstant(0x1eff, dl, MVT::i32));
4342 Bits = DAG.getNode(ISD::ADD, dl, VT, LSB, FF);
4344 SDValue One = DAG.getNode(ARMISD::VMOVIMM, dl, VT,
4345 DAG.getTargetConstant(1, dl, ElemTy));
4346 Bits = DAG.getNode(ISD::SUB, dl, VT, LSB, One);
4349 // Count #bits with vcnt.8.
4350 EVT VT8Bit = VT.is64BitVector() ? MVT::v8i8 : MVT::v16i8;
4351 SDValue BitsVT8 = DAG.getNode(ISD::BITCAST, dl, VT8Bit, Bits);
4352 SDValue Cnt8 = DAG.getNode(ISD::CTPOP, dl, VT8Bit, BitsVT8);
4354 // Gather the #bits with vpaddl (pairwise add.)
4355 EVT VT16Bit = VT.is64BitVector() ? MVT::v4i16 : MVT::v8i16;
4356 SDValue Cnt16 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT16Bit,
4357 DAG.getTargetConstant(Intrinsic::arm_neon_vpaddlu, dl, MVT::i32),
4359 if (ElemTy == MVT::i16)
4362 EVT VT32Bit = VT.is64BitVector() ? MVT::v2i32 : MVT::v4i32;
4363 SDValue Cnt32 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT32Bit,
4364 DAG.getTargetConstant(Intrinsic::arm_neon_vpaddlu, dl, MVT::i32),
4366 if (ElemTy == MVT::i32)
4369 assert(ElemTy == MVT::i64);
4370 SDValue Cnt64 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
4371 DAG.getTargetConstant(Intrinsic::arm_neon_vpaddlu, dl, MVT::i32),
4376 if (!ST->hasV6T2Ops())
4379 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
4380 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
4383 /// getCTPOP16BitCounts - Returns a v8i8/v16i8 vector containing the bit-count
4384 /// for each 16-bit element from operand, repeated. The basic idea is to
4385 /// leverage vcnt to get the 8-bit counts, gather and add the results.
4387 /// Trace for v4i16:
4388 /// input = [v0 v1 v2 v3 ] (vi 16-bit element)
4389 /// cast: N0 = [w0 w1 w2 w3 w4 w5 w6 w7] (v0 = [w0 w1], wi 8-bit element)
4390 /// vcnt: N1 = [b0 b1 b2 b3 b4 b5 b6 b7] (bi = bit-count of 8-bit element wi)
4391 /// vrev: N2 = [b1 b0 b3 b2 b5 b4 b7 b6]
4392 /// [b0 b1 b2 b3 b4 b5 b6 b7]
4393 /// +[b1 b0 b3 b2 b5 b4 b7 b6]
4394 /// N3=N1+N2 = [k0 k0 k1 k1 k2 k2 k3 k3] (k0 = b0+b1 = bit-count of 16-bit v0,
4395 /// vuzp: = [k0 k1 k2 k3 k0 k1 k2 k3] each ki is 8-bits)
4396 static SDValue getCTPOP16BitCounts(SDNode *N, SelectionDAG &DAG) {
4397 EVT VT = N->getValueType(0);
4400 EVT VT8Bit = VT.is64BitVector() ? MVT::v8i8 : MVT::v16i8;
4401 SDValue N0 = DAG.getNode(ISD::BITCAST, DL, VT8Bit, N->getOperand(0));
4402 SDValue N1 = DAG.getNode(ISD::CTPOP, DL, VT8Bit, N0);
4403 SDValue N2 = DAG.getNode(ARMISD::VREV16, DL, VT8Bit, N1);
4404 SDValue N3 = DAG.getNode(ISD::ADD, DL, VT8Bit, N1, N2);
4405 return DAG.getNode(ARMISD::VUZP, DL, VT8Bit, N3, N3);
4408 /// lowerCTPOP16BitElements - Returns a v4i16/v8i16 vector containing the
4409 /// bit-count for each 16-bit element from the operand. We need slightly
4410 /// different sequencing for v4i16 and v8i16 to stay within NEON's available
4411 /// 64/128-bit registers.
4413 /// Trace for v4i16:
4414 /// input = [v0 v1 v2 v3 ] (vi 16-bit element)
4415 /// v8i8: BitCounts = [k0 k1 k2 k3 k0 k1 k2 k3 ] (ki is the bit-count of vi)
4416 /// v8i16:Extended = [k0 k1 k2 k3 k0 k1 k2 k3 ]
4417 /// v4i16:Extracted = [k0 k1 k2 k3 ]
4418 static SDValue lowerCTPOP16BitElements(SDNode *N, SelectionDAG &DAG) {
4419 EVT VT = N->getValueType(0);
4422 SDValue BitCounts = getCTPOP16BitCounts(N, DAG);
4423 if (VT.is64BitVector()) {
4424 SDValue Extended = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v8i16, BitCounts);
4425 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i16, Extended,
4426 DAG.getIntPtrConstant(0, DL));
4428 SDValue Extracted = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v8i8,
4429 BitCounts, DAG.getIntPtrConstant(0, DL));
4430 return DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v8i16, Extracted);
4434 /// lowerCTPOP32BitElements - Returns a v2i32/v4i32 vector containing the
4435 /// bit-count for each 32-bit element from the operand. The idea here is
4436 /// to split the vector into 16-bit elements, leverage the 16-bit count
4437 /// routine, and then combine the results.
4439 /// Trace for v2i32 (v4i32 similar with Extracted/Extended exchanged):
4440 /// input = [v0 v1 ] (vi: 32-bit elements)
4441 /// Bitcast = [w0 w1 w2 w3 ] (wi: 16-bit elements, v0 = [w0 w1])
4442 /// Counts16 = [k0 k1 k2 k3 ] (ki: 16-bit elements, bit-count of wi)
4443 /// vrev: N0 = [k1 k0 k3 k2 ]
4445 /// N1 =+[k1 k0 k3 k2 ]
4447 /// N2 =+[k1 k3 k0 k2 ]
4449 /// Extended =+[k1 k3 k0 k2 ]
4451 /// Extracted=+[k1 k3 ]
4453 static SDValue lowerCTPOP32BitElements(SDNode *N, SelectionDAG &DAG) {
4454 EVT VT = N->getValueType(0);
4457 EVT VT16Bit = VT.is64BitVector() ? MVT::v4i16 : MVT::v8i16;
4459 SDValue Bitcast = DAG.getNode(ISD::BITCAST, DL, VT16Bit, N->getOperand(0));
4460 SDValue Counts16 = lowerCTPOP16BitElements(Bitcast.getNode(), DAG);
4461 SDValue N0 = DAG.getNode(ARMISD::VREV32, DL, VT16Bit, Counts16);
4462 SDValue N1 = DAG.getNode(ISD::ADD, DL, VT16Bit, Counts16, N0);
4463 SDValue N2 = DAG.getNode(ARMISD::VUZP, DL, VT16Bit, N1, N1);
4465 if (VT.is64BitVector()) {
4466 SDValue Extended = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v4i32, N2);
4467 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i32, Extended,
4468 DAG.getIntPtrConstant(0, DL));
4470 SDValue Extracted = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i16, N2,
4471 DAG.getIntPtrConstant(0, DL));
4472 return DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v4i32, Extracted);
4476 static SDValue LowerCTPOP(SDNode *N, SelectionDAG &DAG,
4477 const ARMSubtarget *ST) {
4478 EVT VT = N->getValueType(0);
4480 assert(ST->hasNEON() && "Custom ctpop lowering requires NEON.");
4481 assert((VT == MVT::v2i32 || VT == MVT::v4i32 ||
4482 VT == MVT::v4i16 || VT == MVT::v8i16) &&
4483 "Unexpected type for custom ctpop lowering");
4485 if (VT.getVectorElementType() == MVT::i32)
4486 return lowerCTPOP32BitElements(N, DAG);
4488 return lowerCTPOP16BitElements(N, DAG);
4491 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
4492 const ARMSubtarget *ST) {
4493 EVT VT = N->getValueType(0);
4499 // Lower vector shifts on NEON to use VSHL.
4500 assert(ST->hasNEON() && "unexpected vector shift");
4502 // Left shifts translate directly to the vshiftu intrinsic.
4503 if (N->getOpcode() == ISD::SHL)
4504 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
4505 DAG.getConstant(Intrinsic::arm_neon_vshiftu, dl,
4507 N->getOperand(0), N->getOperand(1));
4509 assert((N->getOpcode() == ISD::SRA ||
4510 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
4512 // NEON uses the same intrinsics for both left and right shifts. For
4513 // right shifts, the shift amounts are negative, so negate the vector of
4515 EVT ShiftVT = N->getOperand(1).getValueType();
4516 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
4517 getZeroVector(ShiftVT, DAG, dl),
4519 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
4520 Intrinsic::arm_neon_vshifts :
4521 Intrinsic::arm_neon_vshiftu);
4522 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
4523 DAG.getConstant(vshiftInt, dl, MVT::i32),
4524 N->getOperand(0), NegatedCount);
4527 static SDValue Expand64BitShift(SDNode *N, SelectionDAG &DAG,
4528 const ARMSubtarget *ST) {
4529 EVT VT = N->getValueType(0);
4532 // We can get here for a node like i32 = ISD::SHL i32, i64
4536 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
4537 "Unknown shift to lower!");
4539 // We only lower SRA, SRL of 1 here, all others use generic lowering.
4540 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
4541 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
4544 // If we are in thumb mode, we don't have RRX.
4545 if (ST->isThumb1Only()) return SDValue();
4547 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
4548 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
4549 DAG.getConstant(0, dl, MVT::i32));
4550 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
4551 DAG.getConstant(1, dl, MVT::i32));
4553 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
4554 // captures the result into a carry flag.
4555 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
4556 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Glue), Hi);
4558 // The low part is an ARMISD::RRX operand, which shifts the carry in.
4559 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
4561 // Merge the pieces into a single i64 value.
4562 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
4565 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
4566 SDValue TmpOp0, TmpOp1;
4567 bool Invert = false;
4571 SDValue Op0 = Op.getOperand(0);
4572 SDValue Op1 = Op.getOperand(1);
4573 SDValue CC = Op.getOperand(2);
4574 EVT CmpVT = Op0.getValueType().changeVectorElementTypeToInteger();
4575 EVT VT = Op.getValueType();
4576 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
4579 if (CmpVT.getVectorElementType() == MVT::i64)
4580 // 64-bit comparisons are not legal. We've marked SETCC as non-Custom,
4581 // but it's possible that our operands are 64-bit but our result is 32-bit.
4582 // Bail in this case.
4585 if (Op1.getValueType().isFloatingPoint()) {
4586 switch (SetCCOpcode) {
4587 default: llvm_unreachable("Illegal FP comparison");
4589 case ISD::SETNE: Invert = true; // Fallthrough
4591 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
4593 case ISD::SETLT: Swap = true; // Fallthrough
4595 case ISD::SETGT: Opc = ARMISD::VCGT; break;
4597 case ISD::SETLE: Swap = true; // Fallthrough
4599 case ISD::SETGE: Opc = ARMISD::VCGE; break;
4600 case ISD::SETUGE: Swap = true; // Fallthrough
4601 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
4602 case ISD::SETUGT: Swap = true; // Fallthrough
4603 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
4604 case ISD::SETUEQ: Invert = true; // Fallthrough
4606 // Expand this to (OLT | OGT).
4610 Op0 = DAG.getNode(ARMISD::VCGT, dl, CmpVT, TmpOp1, TmpOp0);
4611 Op1 = DAG.getNode(ARMISD::VCGT, dl, CmpVT, TmpOp0, TmpOp1);
4613 case ISD::SETUO: Invert = true; // Fallthrough
4615 // Expand this to (OLT | OGE).
4619 Op0 = DAG.getNode(ARMISD::VCGT, dl, CmpVT, TmpOp1, TmpOp0);
4620 Op1 = DAG.getNode(ARMISD::VCGE, dl, CmpVT, TmpOp0, TmpOp1);
4624 // Integer comparisons.
4625 switch (SetCCOpcode) {
4626 default: llvm_unreachable("Illegal integer comparison");
4627 case ISD::SETNE: Invert = true;
4628 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
4629 case ISD::SETLT: Swap = true;
4630 case ISD::SETGT: Opc = ARMISD::VCGT; break;
4631 case ISD::SETLE: Swap = true;
4632 case ISD::SETGE: Opc = ARMISD::VCGE; break;
4633 case ISD::SETULT: Swap = true;
4634 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
4635 case ISD::SETULE: Swap = true;
4636 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
4639 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
4640 if (Opc == ARMISD::VCEQ) {
4643 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
4645 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
4648 // Ignore bitconvert.
4649 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BITCAST)
4650 AndOp = AndOp.getOperand(0);
4652 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
4654 Op0 = DAG.getNode(ISD::BITCAST, dl, CmpVT, AndOp.getOperand(0));
4655 Op1 = DAG.getNode(ISD::BITCAST, dl, CmpVT, AndOp.getOperand(1));
4662 std::swap(Op0, Op1);
4664 // If one of the operands is a constant vector zero, attempt to fold the
4665 // comparison to a specialized compare-against-zero form.
4667 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
4669 else if (ISD::isBuildVectorAllZeros(Op0.getNode())) {
4670 if (Opc == ARMISD::VCGE)
4671 Opc = ARMISD::VCLEZ;
4672 else if (Opc == ARMISD::VCGT)
4673 Opc = ARMISD::VCLTZ;
4678 if (SingleOp.getNode()) {
4681 Result = DAG.getNode(ARMISD::VCEQZ, dl, CmpVT, SingleOp); break;
4683 Result = DAG.getNode(ARMISD::VCGEZ, dl, CmpVT, SingleOp); break;
4685 Result = DAG.getNode(ARMISD::VCLEZ, dl, CmpVT, SingleOp); break;
4687 Result = DAG.getNode(ARMISD::VCGTZ, dl, CmpVT, SingleOp); break;
4689 Result = DAG.getNode(ARMISD::VCLTZ, dl, CmpVT, SingleOp); break;
4691 Result = DAG.getNode(Opc, dl, CmpVT, Op0, Op1);
4694 Result = DAG.getNode(Opc, dl, CmpVT, Op0, Op1);
4697 Result = DAG.getSExtOrTrunc(Result, dl, VT);
4700 Result = DAG.getNOT(dl, Result, VT);
4705 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
4706 /// valid vector constant for a NEON instruction with a "modified immediate"
4707 /// operand (e.g., VMOV). If so, return the encoded value.
4708 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
4709 unsigned SplatBitSize, SelectionDAG &DAG,
4710 SDLoc dl, EVT &VT, bool is128Bits,
4711 NEONModImmType type) {
4712 unsigned OpCmode, Imm;
4714 // SplatBitSize is set to the smallest size that splats the vector, so a
4715 // zero vector will always have SplatBitSize == 8. However, NEON modified
4716 // immediate instructions others than VMOV do not support the 8-bit encoding
4717 // of a zero vector, and the default encoding of zero is supposed to be the
4722 switch (SplatBitSize) {
4724 if (type != VMOVModImm)
4726 // Any 1-byte value is OK. Op=0, Cmode=1110.
4727 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
4730 VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
4734 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
4735 VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
4736 if ((SplatBits & ~0xff) == 0) {
4737 // Value = 0x00nn: Op=x, Cmode=100x.
4742 if ((SplatBits & ~0xff00) == 0) {
4743 // Value = 0xnn00: Op=x, Cmode=101x.
4745 Imm = SplatBits >> 8;
4751 // NEON's 32-bit VMOV supports splat values where:
4752 // * only one byte is nonzero, or
4753 // * the least significant byte is 0xff and the second byte is nonzero, or
4754 // * the least significant 2 bytes are 0xff and the third is nonzero.
4755 VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
4756 if ((SplatBits & ~0xff) == 0) {
4757 // Value = 0x000000nn: Op=x, Cmode=000x.
4762 if ((SplatBits & ~0xff00) == 0) {
4763 // Value = 0x0000nn00: Op=x, Cmode=001x.
4765 Imm = SplatBits >> 8;
4768 if ((SplatBits & ~0xff0000) == 0) {
4769 // Value = 0x00nn0000: Op=x, Cmode=010x.
4771 Imm = SplatBits >> 16;
4774 if ((SplatBits & ~0xff000000) == 0) {
4775 // Value = 0xnn000000: Op=x, Cmode=011x.
4777 Imm = SplatBits >> 24;
4781 // cmode == 0b1100 and cmode == 0b1101 are not supported for VORR or VBIC
4782 if (type == OtherModImm) return SDValue();
4784 if ((SplatBits & ~0xffff) == 0 &&
4785 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
4786 // Value = 0x0000nnff: Op=x, Cmode=1100.
4788 Imm = SplatBits >> 8;
4792 if ((SplatBits & ~0xffffff) == 0 &&
4793 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
4794 // Value = 0x00nnffff: Op=x, Cmode=1101.
4796 Imm = SplatBits >> 16;
4800 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
4801 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
4802 // VMOV.I32. A (very) minor optimization would be to replicate the value
4803 // and fall through here to test for a valid 64-bit splat. But, then the
4804 // caller would also need to check and handle the change in size.
4808 if (type != VMOVModImm)
4810 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
4811 uint64_t BitMask = 0xff;
4813 unsigned ImmMask = 1;
4815 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
4816 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
4819 } else if ((SplatBits & BitMask) != 0) {
4826 if (DAG.getDataLayout().isBigEndian())
4827 // swap higher and lower 32 bit word
4828 Imm = ((Imm & 0xf) << 4) | ((Imm & 0xf0) >> 4);
4830 // Op=1, Cmode=1110.
4832 VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
4837 llvm_unreachable("unexpected size for isNEONModifiedImm");
4840 unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm);
4841 return DAG.getTargetConstant(EncodedVal, dl, MVT::i32);
4844 SDValue ARMTargetLowering::LowerConstantFP(SDValue Op, SelectionDAG &DAG,
4845 const ARMSubtarget *ST) const {
4849 bool IsDouble = Op.getValueType() == MVT::f64;
4850 ConstantFPSDNode *CFP = cast<ConstantFPSDNode>(Op);
4852 // Use the default (constant pool) lowering for double constants when we have
4854 if (IsDouble && Subtarget->isFPOnlySP())
4857 // Try splatting with a VMOV.f32...
4858 APFloat FPVal = CFP->getValueAPF();
4859 int ImmVal = IsDouble ? ARM_AM::getFP64Imm(FPVal) : ARM_AM::getFP32Imm(FPVal);
4862 if (IsDouble || !ST->useNEONForSinglePrecisionFP()) {
4863 // We have code in place to select a valid ConstantFP already, no need to
4868 // It's a float and we are trying to use NEON operations where
4869 // possible. Lower it to a splat followed by an extract.
4871 SDValue NewVal = DAG.getTargetConstant(ImmVal, DL, MVT::i32);
4872 SDValue VecConstant = DAG.getNode(ARMISD::VMOVFPIMM, DL, MVT::v2f32,
4874 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecConstant,
4875 DAG.getConstant(0, DL, MVT::i32));
4878 // The rest of our options are NEON only, make sure that's allowed before
4880 if (!ST->hasNEON() || (!IsDouble && !ST->useNEONForSinglePrecisionFP()))
4884 uint64_t iVal = FPVal.bitcastToAPInt().getZExtValue();
4886 // It wouldn't really be worth bothering for doubles except for one very
4887 // important value, which does happen to match: 0.0. So make sure we don't do
4889 if (IsDouble && (iVal & 0xffffffff) != (iVal >> 32))
4892 // Try a VMOV.i32 (FIXME: i8, i16, or i64 could work too).
4893 SDValue NewVal = isNEONModifiedImm(iVal & 0xffffffffU, 0, 32, DAG, SDLoc(Op),
4894 VMovVT, false, VMOVModImm);
4895 if (NewVal != SDValue()) {
4897 SDValue VecConstant = DAG.getNode(ARMISD::VMOVIMM, DL, VMovVT,
4900 return DAG.getNode(ISD::BITCAST, DL, MVT::f64, VecConstant);
4902 // It's a float: cast and extract a vector element.
4903 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
4905 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
4906 DAG.getConstant(0, DL, MVT::i32));
4909 // Finally, try a VMVN.i32
4910 NewVal = isNEONModifiedImm(~iVal & 0xffffffffU, 0, 32, DAG, SDLoc(Op), VMovVT,
4912 if (NewVal != SDValue()) {
4914 SDValue VecConstant = DAG.getNode(ARMISD::VMVNIMM, DL, VMovVT, NewVal);
4917 return DAG.getNode(ISD::BITCAST, DL, MVT::f64, VecConstant);
4919 // It's a float: cast and extract a vector element.
4920 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
4922 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
4923 DAG.getConstant(0, DL, MVT::i32));
4929 // check if an VEXT instruction can handle the shuffle mask when the
4930 // vector sources of the shuffle are the same.
4931 static bool isSingletonVEXTMask(ArrayRef<int> M, EVT VT, unsigned &Imm) {
4932 unsigned NumElts = VT.getVectorNumElements();
4934 // Assume that the first shuffle index is not UNDEF. Fail if it is.
4940 // If this is a VEXT shuffle, the immediate value is the index of the first
4941 // element. The other shuffle indices must be the successive elements after
4943 unsigned ExpectedElt = Imm;
4944 for (unsigned i = 1; i < NumElts; ++i) {
4945 // Increment the expected index. If it wraps around, just follow it
4946 // back to index zero and keep going.
4948 if (ExpectedElt == NumElts)
4951 if (M[i] < 0) continue; // ignore UNDEF indices
4952 if (ExpectedElt != static_cast<unsigned>(M[i]))
4960 static bool isVEXTMask(ArrayRef<int> M, EVT VT,
4961 bool &ReverseVEXT, unsigned &Imm) {
4962 unsigned NumElts = VT.getVectorNumElements();
4963 ReverseVEXT = false;
4965 // Assume that the first shuffle index is not UNDEF. Fail if it is.
4971 // If this is a VEXT shuffle, the immediate value is the index of the first
4972 // element. The other shuffle indices must be the successive elements after
4974 unsigned ExpectedElt = Imm;
4975 for (unsigned i = 1; i < NumElts; ++i) {
4976 // Increment the expected index. If it wraps around, it may still be
4977 // a VEXT but the source vectors must be swapped.
4979 if (ExpectedElt == NumElts * 2) {
4984 if (M[i] < 0) continue; // ignore UNDEF indices
4985 if (ExpectedElt != static_cast<unsigned>(M[i]))
4989 // Adjust the index value if the source operands will be swapped.
4996 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
4997 /// instruction with the specified blocksize. (The order of the elements
4998 /// within each block of the vector is reversed.)
4999 static bool isVREVMask(ArrayRef<int> M, EVT VT, unsigned BlockSize) {
5000 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
5001 "Only possible block sizes for VREV are: 16, 32, 64");
5003 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
5007 unsigned NumElts = VT.getVectorNumElements();
5008 unsigned BlockElts = M[0] + 1;
5009 // If the first shuffle index is UNDEF, be optimistic.
5011 BlockElts = BlockSize / EltSz;
5013 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
5016 for (unsigned i = 0; i < NumElts; ++i) {
5017 if (M[i] < 0) continue; // ignore UNDEF indices
5018 if ((unsigned) M[i] != (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
5025 static bool isVTBLMask(ArrayRef<int> M, EVT VT) {
5026 // We can handle <8 x i8> vector shuffles. If the index in the mask is out of
5027 // range, then 0 is placed into the resulting vector. So pretty much any mask
5028 // of 8 elements can work here.
5029 return VT == MVT::v8i8 && M.size() == 8;
5032 // Checks whether the shuffle mask represents a vector transpose (VTRN) by
5033 // checking that pairs of elements in the shuffle mask represent the same index
5034 // in each vector, incrementing the expected index by 2 at each step.
5035 // e.g. For v1,v2 of type v4i32 a valid shuffle mask is: [0, 4, 2, 6]
5036 // v1={a,b,c,d} => x=shufflevector v1, v2 shufflemask => x={a,e,c,g}
5038 // WhichResult gives the offset for each element in the mask based on which
5039 // of the two results it belongs to.
5041 // The transpose can be represented either as:
5042 // result1 = shufflevector v1, v2, result1_shuffle_mask
5043 // result2 = shufflevector v1, v2, result2_shuffle_mask
5044 // where v1/v2 and the shuffle masks have the same number of elements
5045 // (here WhichResult (see below) indicates which result is being checked)
5048 // results = shufflevector v1, v2, shuffle_mask
5049 // where both results are returned in one vector and the shuffle mask has twice
5050 // as many elements as v1/v2 (here WhichResult will always be 0 if true) here we
5051 // want to check the low half and high half of the shuffle mask as if it were
5053 static bool isVTRNMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
5054 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
5058 unsigned NumElts = VT.getVectorNumElements();
5059 if (M.size() != NumElts && M.size() != NumElts*2)
5062 // If the mask is twice as long as the input vector then we need to check the
5063 // upper and lower parts of the mask with a matching value for WhichResult
5064 // FIXME: A mask with only even values will be rejected in case the first
5065 // element is undefined, e.g. [-1, 4, 2, 6] will be rejected, because only
5066 // M[0] is used to determine WhichResult
5067 for (unsigned i = 0; i < M.size(); i += NumElts) {
5068 if (M.size() == NumElts * 2)
5069 WhichResult = i / NumElts;
5071 WhichResult = M[i] == 0 ? 0 : 1;
5072 for (unsigned j = 0; j < NumElts; j += 2) {
5073 if ((M[i+j] >= 0 && (unsigned) M[i+j] != j + WhichResult) ||
5074 (M[i+j+1] >= 0 && (unsigned) M[i+j+1] != j + NumElts + WhichResult))
5079 if (M.size() == NumElts*2)
5085 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
5086 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
5087 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
5088 static bool isVTRN_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
5089 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
5093 unsigned NumElts = VT.getVectorNumElements();
5094 if (M.size() != NumElts && M.size() != NumElts*2)
5097 for (unsigned i = 0; i < M.size(); i += NumElts) {
5098 if (M.size() == NumElts * 2)
5099 WhichResult = i / NumElts;
5101 WhichResult = M[i] == 0 ? 0 : 1;
5102 for (unsigned j = 0; j < NumElts; j += 2) {
5103 if ((M[i+j] >= 0 && (unsigned) M[i+j] != j + WhichResult) ||
5104 (M[i+j+1] >= 0 && (unsigned) M[i+j+1] != j + WhichResult))
5109 if (M.size() == NumElts*2)
5115 // Checks whether the shuffle mask represents a vector unzip (VUZP) by checking
5116 // that the mask elements are either all even and in steps of size 2 or all odd
5117 // and in steps of size 2.
5118 // e.g. For v1,v2 of type v4i32 a valid shuffle mask is: [0, 2, 4, 6]
5119 // v1={a,b,c,d} => x=shufflevector v1, v2 shufflemask => x={a,c,e,g}
5121 // Requires similar checks to that of isVTRNMask with
5122 // respect the how results are returned.
5123 static bool isVUZPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
5124 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
5128 unsigned NumElts = VT.getVectorNumElements();
5129 if (M.size() != NumElts && M.size() != NumElts*2)
5132 for (unsigned i = 0; i < M.size(); i += NumElts) {
5133 WhichResult = M[i] == 0 ? 0 : 1;
5134 for (unsigned j = 0; j < NumElts; ++j) {
5135 if (M[i+j] >= 0 && (unsigned) M[i+j] != 2 * j + WhichResult)
5140 if (M.size() == NumElts*2)
5143 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
5144 if (VT.is64BitVector() && EltSz == 32)
5150 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
5151 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
5152 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
5153 static bool isVUZP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
5154 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
5158 unsigned NumElts = VT.getVectorNumElements();
5159 if (M.size() != NumElts && M.size() != NumElts*2)
5162 unsigned Half = NumElts / 2;
5163 for (unsigned i = 0; i < M.size(); i += NumElts) {
5164 WhichResult = M[i] == 0 ? 0 : 1;
5165 for (unsigned j = 0; j < NumElts; j += Half) {
5166 unsigned Idx = WhichResult;
5167 for (unsigned k = 0; k < Half; ++k) {
5168 int MIdx = M[i + j + k];
5169 if (MIdx >= 0 && (unsigned) MIdx != Idx)
5176 if (M.size() == NumElts*2)
5179 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
5180 if (VT.is64BitVector() && EltSz == 32)
5186 // Checks whether the shuffle mask represents a vector zip (VZIP) by checking
5187 // that pairs of elements of the shufflemask represent the same index in each
5188 // vector incrementing sequentially through the vectors.
5189 // e.g. For v1,v2 of type v4i32 a valid shuffle mask is: [0, 4, 1, 5]
5190 // v1={a,b,c,d} => x=shufflevector v1, v2 shufflemask => x={a,e,b,f}
5192 // Requires similar checks to that of isVTRNMask with respect the how results
5194 static bool isVZIPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
5195 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
5199 unsigned NumElts = VT.getVectorNumElements();
5200 if (M.size() != NumElts && M.size() != NumElts*2)
5203 for (unsigned i = 0; i < M.size(); i += NumElts) {
5204 WhichResult = M[i] == 0 ? 0 : 1;
5205 unsigned Idx = WhichResult * NumElts / 2;
5206 for (unsigned j = 0; j < NumElts; j += 2) {
5207 if ((M[i+j] >= 0 && (unsigned) M[i+j] != Idx) ||
5208 (M[i+j+1] >= 0 && (unsigned) M[i+j+1] != Idx + NumElts))
5214 if (M.size() == NumElts*2)
5217 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
5218 if (VT.is64BitVector() && EltSz == 32)
5224 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
5225 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
5226 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
5227 static bool isVZIP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
5228 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
5232 unsigned NumElts = VT.getVectorNumElements();
5233 if (M.size() != NumElts && M.size() != NumElts*2)
5236 for (unsigned i = 0; i < M.size(); i += NumElts) {
5237 WhichResult = M[i] == 0 ? 0 : 1;
5238 unsigned Idx = WhichResult * NumElts / 2;
5239 for (unsigned j = 0; j < NumElts; j += 2) {
5240 if ((M[i+j] >= 0 && (unsigned) M[i+j] != Idx) ||
5241 (M[i+j+1] >= 0 && (unsigned) M[i+j+1] != Idx))
5247 if (M.size() == NumElts*2)
5250 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
5251 if (VT.is64BitVector() && EltSz == 32)
5257 /// Check if \p ShuffleMask is a NEON two-result shuffle (VZIP, VUZP, VTRN),
5258 /// and return the corresponding ARMISD opcode if it is, or 0 if it isn't.
5259 static unsigned isNEONTwoResultShuffleMask(ArrayRef<int> ShuffleMask, EVT VT,
5260 unsigned &WhichResult,
5263 if (isVTRNMask(ShuffleMask, VT, WhichResult))
5264 return ARMISD::VTRN;
5265 if (isVUZPMask(ShuffleMask, VT, WhichResult))
5266 return ARMISD::VUZP;
5267 if (isVZIPMask(ShuffleMask, VT, WhichResult))
5268 return ARMISD::VZIP;
5271 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
5272 return ARMISD::VTRN;
5273 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
5274 return ARMISD::VUZP;
5275 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
5276 return ARMISD::VZIP;
5281 /// \return true if this is a reverse operation on an vector.
5282 static bool isReverseMask(ArrayRef<int> M, EVT VT) {
5283 unsigned NumElts = VT.getVectorNumElements();
5284 // Make sure the mask has the right size.
5285 if (NumElts != M.size())
5288 // Look for <15, ..., 3, -1, 1, 0>.
5289 for (unsigned i = 0; i != NumElts; ++i)
5290 if (M[i] >= 0 && M[i] != (int) (NumElts - 1 - i))
5296 // If N is an integer constant that can be moved into a register in one
5297 // instruction, return an SDValue of such a constant (will become a MOV
5298 // instruction). Otherwise return null.
5299 static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG,
5300 const ARMSubtarget *ST, SDLoc dl) {
5302 if (!isa<ConstantSDNode>(N))
5304 Val = cast<ConstantSDNode>(N)->getZExtValue();
5306 if (ST->isThumb1Only()) {
5307 if (Val <= 255 || ~Val <= 255)
5308 return DAG.getConstant(Val, dl, MVT::i32);
5310 if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1)
5311 return DAG.getConstant(Val, dl, MVT::i32);
5316 // If this is a case we can't handle, return null and let the default
5317 // expansion code take care of it.
5318 SDValue ARMTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
5319 const ARMSubtarget *ST) const {
5320 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
5322 EVT VT = Op.getValueType();
5324 APInt SplatBits, SplatUndef;
5325 unsigned SplatBitSize;
5327 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
5328 if (SplatBitSize <= 64) {
5329 // Check if an immediate VMOV works.
5331 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
5332 SplatUndef.getZExtValue(), SplatBitSize,
5333 DAG, dl, VmovVT, VT.is128BitVector(),
5335 if (Val.getNode()) {
5336 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
5337 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
5340 // Try an immediate VMVN.
5341 uint64_t NegatedImm = (~SplatBits).getZExtValue();
5342 Val = isNEONModifiedImm(NegatedImm,
5343 SplatUndef.getZExtValue(), SplatBitSize,
5344 DAG, dl, VmovVT, VT.is128BitVector(),
5346 if (Val.getNode()) {
5347 SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
5348 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
5351 // Use vmov.f32 to materialize other v2f32 and v4f32 splats.
5352 if ((VT == MVT::v2f32 || VT == MVT::v4f32) && SplatBitSize == 32) {
5353 int ImmVal = ARM_AM::getFP32Imm(SplatBits);
5355 SDValue Val = DAG.getTargetConstant(ImmVal, dl, MVT::i32);
5356 return DAG.getNode(ARMISD::VMOVFPIMM, dl, VT, Val);
5362 // Scan through the operands to see if only one value is used.
5364 // As an optimisation, even if more than one value is used it may be more
5365 // profitable to splat with one value then change some lanes.
5367 // Heuristically we decide to do this if the vector has a "dominant" value,
5368 // defined as splatted to more than half of the lanes.
5369 unsigned NumElts = VT.getVectorNumElements();
5370 bool isOnlyLowElement = true;
5371 bool usesOnlyOneValue = true;
5372 bool hasDominantValue = false;
5373 bool isConstant = true;
5375 // Map of the number of times a particular SDValue appears in the
5377 DenseMap<SDValue, unsigned> ValueCounts;
5379 for (unsigned i = 0; i < NumElts; ++i) {
5380 SDValue V = Op.getOperand(i);
5381 if (V.getOpcode() == ISD::UNDEF)
5384 isOnlyLowElement = false;
5385 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
5388 ValueCounts.insert(std::make_pair(V, 0));
5389 unsigned &Count = ValueCounts[V];
5391 // Is this value dominant? (takes up more than half of the lanes)
5392 if (++Count > (NumElts / 2)) {
5393 hasDominantValue = true;
5397 if (ValueCounts.size() != 1)
5398 usesOnlyOneValue = false;
5399 if (!Value.getNode() && ValueCounts.size() > 0)
5400 Value = ValueCounts.begin()->first;
5402 if (ValueCounts.size() == 0)
5403 return DAG.getUNDEF(VT);
5405 // Loads are better lowered with insert_vector_elt/ARMISD::BUILD_VECTOR.
5406 // Keep going if we are hitting this case.
5407 if (isOnlyLowElement && !ISD::isNormalLoad(Value.getNode()))
5408 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
5410 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5412 // Use VDUP for non-constant splats. For f32 constant splats, reduce to
5413 // i32 and try again.
5414 if (hasDominantValue && EltSize <= 32) {
5418 // If we are VDUPing a value that comes directly from a vector, that will
5419 // cause an unnecessary move to and from a GPR, where instead we could
5420 // just use VDUPLANE. We can only do this if the lane being extracted
5421 // is at a constant index, as the VDUP from lane instructions only have
5422 // constant-index forms.
5423 if (Value->getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5424 isa<ConstantSDNode>(Value->getOperand(1))) {
5425 // We need to create a new undef vector to use for the VDUPLANE if the
5426 // size of the vector from which we get the value is different than the
5427 // size of the vector that we need to create. We will insert the element
5428 // such that the register coalescer will remove unnecessary copies.
5429 if (VT != Value->getOperand(0).getValueType()) {
5430 ConstantSDNode *constIndex;
5431 constIndex = dyn_cast<ConstantSDNode>(Value->getOperand(1));
5432 assert(constIndex && "The index is not a constant!");
5433 unsigned index = constIndex->getAPIntValue().getLimitedValue() %
5434 VT.getVectorNumElements();
5435 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
5436 DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DAG.getUNDEF(VT),
5437 Value, DAG.getConstant(index, dl, MVT::i32)),
5438 DAG.getConstant(index, dl, MVT::i32));
5440 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
5441 Value->getOperand(0), Value->getOperand(1));
5443 N = DAG.getNode(ARMISD::VDUP, dl, VT, Value);
5445 if (!usesOnlyOneValue) {
5446 // The dominant value was splatted as 'N', but we now have to insert
5447 // all differing elements.
5448 for (unsigned I = 0; I < NumElts; ++I) {
5449 if (Op.getOperand(I) == Value)
5451 SmallVector<SDValue, 3> Ops;
5453 Ops.push_back(Op.getOperand(I));
5454 Ops.push_back(DAG.getConstant(I, dl, MVT::i32));
5455 N = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Ops);
5460 if (VT.getVectorElementType().isFloatingPoint()) {
5461 SmallVector<SDValue, 8> Ops;
5462 for (unsigned i = 0; i < NumElts; ++i)
5463 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, MVT::i32,
5465 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
5466 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, Ops);
5467 Val = LowerBUILD_VECTOR(Val, DAG, ST);
5469 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
5471 if (usesOnlyOneValue) {
5472 SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
5473 if (isConstant && Val.getNode())
5474 return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
5478 // If all elements are constants and the case above didn't get hit, fall back
5479 // to the default expansion, which will generate a load from the constant
5484 // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
5486 SDValue shuffle = ReconstructShuffle(Op, DAG);
5487 if (shuffle != SDValue())
5491 // Vectors with 32- or 64-bit elements can be built by directly assigning
5492 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
5493 // will be legalized.
5494 if (EltSize >= 32) {
5495 // Do the expansion with floating-point types, since that is what the VFP
5496 // registers are defined to use, and since i64 is not legal.
5497 EVT EltVT = EVT::getFloatingPointVT(EltSize);
5498 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
5499 SmallVector<SDValue, 8> Ops;
5500 for (unsigned i = 0; i < NumElts; ++i)
5501 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, EltVT, Op.getOperand(i)));
5502 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, Ops);
5503 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
5506 // If all else fails, just use a sequence of INSERT_VECTOR_ELT when we
5507 // know the default expansion would otherwise fall back on something even
5508 // worse. For a vector with one or two non-undef values, that's
5509 // scalar_to_vector for the elements followed by a shuffle (provided the
5510 // shuffle is valid for the target) and materialization element by element
5511 // on the stack followed by a load for everything else.
5512 if (!isConstant && !usesOnlyOneValue) {
5513 SDValue Vec = DAG.getUNDEF(VT);
5514 for (unsigned i = 0 ; i < NumElts; ++i) {
5515 SDValue V = Op.getOperand(i);
5516 if (V.getOpcode() == ISD::UNDEF)
5518 SDValue LaneIdx = DAG.getConstant(i, dl, MVT::i32);
5519 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Vec, V, LaneIdx);
5527 // Gather data to see if the operation can be modelled as a
5528 // shuffle in combination with VEXTs.
5529 SDValue ARMTargetLowering::ReconstructShuffle(SDValue Op,
5530 SelectionDAG &DAG) const {
5531 assert(Op.getOpcode() == ISD::BUILD_VECTOR && "Unknown opcode!");
5533 EVT VT = Op.getValueType();
5534 unsigned NumElts = VT.getVectorNumElements();
5536 struct ShuffleSourceInfo {
5541 // We may insert some combination of BITCASTs and VEXT nodes to force Vec to
5542 // be compatible with the shuffle we intend to construct. As a result
5543 // ShuffleVec will be some sliding window into the original Vec.
5546 // Code should guarantee that element i in Vec starts at element "WindowBase
5547 // + i * WindowScale in ShuffleVec".
5551 bool operator ==(SDValue OtherVec) { return Vec == OtherVec; }
5552 ShuffleSourceInfo(SDValue Vec)
5553 : Vec(Vec), MinElt(UINT_MAX), MaxElt(0), ShuffleVec(Vec), WindowBase(0),
5557 // First gather all vectors used as an immediate source for this BUILD_VECTOR
5559 SmallVector<ShuffleSourceInfo, 2> Sources;
5560 for (unsigned i = 0; i < NumElts; ++i) {
5561 SDValue V = Op.getOperand(i);
5562 if (V.getOpcode() == ISD::UNDEF)
5564 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
5565 // A shuffle can only come from building a vector from various
5566 // elements of other vectors.
5568 } else if (!isa<ConstantSDNode>(V.getOperand(1))) {
5569 // Furthermore, shuffles require a constant mask, whereas extractelts
5570 // accept variable indices.
5574 // Add this element source to the list if it's not already there.
5575 SDValue SourceVec = V.getOperand(0);
5576 auto Source = std::find(Sources.begin(), Sources.end(), SourceVec);
5577 if (Source == Sources.end())
5578 Source = Sources.insert(Sources.end(), ShuffleSourceInfo(SourceVec));
5580 // Update the minimum and maximum lane number seen.
5581 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
5582 Source->MinElt = std::min(Source->MinElt, EltNo);
5583 Source->MaxElt = std::max(Source->MaxElt, EltNo);
5586 // Currently only do something sane when at most two source vectors
5588 if (Sources.size() > 2)
5591 // Find out the smallest element size among result and two sources, and use
5592 // it as element size to build the shuffle_vector.
5593 EVT SmallestEltTy = VT.getVectorElementType();
5594 for (auto &Source : Sources) {
5595 EVT SrcEltTy = Source.Vec.getValueType().getVectorElementType();
5596 if (SrcEltTy.bitsLT(SmallestEltTy))
5597 SmallestEltTy = SrcEltTy;
5599 unsigned ResMultiplier =
5600 VT.getVectorElementType().getSizeInBits() / SmallestEltTy.getSizeInBits();
5601 NumElts = VT.getSizeInBits() / SmallestEltTy.getSizeInBits();
5602 EVT ShuffleVT = EVT::getVectorVT(*DAG.getContext(), SmallestEltTy, NumElts);
5604 // If the source vector is too wide or too narrow, we may nevertheless be able
5605 // to construct a compatible shuffle either by concatenating it with UNDEF or
5606 // extracting a suitable range of elements.
5607 for (auto &Src : Sources) {
5608 EVT SrcVT = Src.ShuffleVec.getValueType();
5610 if (SrcVT.getSizeInBits() == VT.getSizeInBits())
5613 // This stage of the search produces a source with the same element type as
5614 // the original, but with a total width matching the BUILD_VECTOR output.
5615 EVT EltVT = SrcVT.getVectorElementType();
5616 unsigned NumSrcElts = VT.getSizeInBits() / EltVT.getSizeInBits();
5617 EVT DestVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumSrcElts);
5619 if (SrcVT.getSizeInBits() < VT.getSizeInBits()) {
5620 if (2 * SrcVT.getSizeInBits() != VT.getSizeInBits())
5622 // We can pad out the smaller vector for free, so if it's part of a
5625 DAG.getNode(ISD::CONCAT_VECTORS, dl, DestVT, Src.ShuffleVec,
5626 DAG.getUNDEF(Src.ShuffleVec.getValueType()));
5630 if (SrcVT.getSizeInBits() != 2 * VT.getSizeInBits())
5633 if (Src.MaxElt - Src.MinElt >= NumSrcElts) {
5634 // Span too large for a VEXT to cope
5638 if (Src.MinElt >= NumSrcElts) {
5639 // The extraction can just take the second half
5641 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec,
5642 DAG.getConstant(NumSrcElts, dl, MVT::i32));
5643 Src.WindowBase = -NumSrcElts;
5644 } else if (Src.MaxElt < NumSrcElts) {
5645 // The extraction can just take the first half
5647 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec,
5648 DAG.getConstant(0, dl, MVT::i32));
5650 // An actual VEXT is needed
5652 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec,
5653 DAG.getConstant(0, dl, MVT::i32));
5655 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec,
5656 DAG.getConstant(NumSrcElts, dl, MVT::i32));
5658 Src.ShuffleVec = DAG.getNode(ARMISD::VEXT, dl, DestVT, VEXTSrc1,
5660 DAG.getConstant(Src.MinElt, dl, MVT::i32));
5661 Src.WindowBase = -Src.MinElt;
5665 // Another possible incompatibility occurs from the vector element types. We
5666 // can fix this by bitcasting the source vectors to the same type we intend
5668 for (auto &Src : Sources) {
5669 EVT SrcEltTy = Src.ShuffleVec.getValueType().getVectorElementType();
5670 if (SrcEltTy == SmallestEltTy)
5672 assert(ShuffleVT.getVectorElementType() == SmallestEltTy);
5673 Src.ShuffleVec = DAG.getNode(ISD::BITCAST, dl, ShuffleVT, Src.ShuffleVec);
5674 Src.WindowScale = SrcEltTy.getSizeInBits() / SmallestEltTy.getSizeInBits();
5675 Src.WindowBase *= Src.WindowScale;
5678 // Final sanity check before we try to actually produce a shuffle.
5680 for (auto Src : Sources)
5681 assert(Src.ShuffleVec.getValueType() == ShuffleVT);
5684 // The stars all align, our next step is to produce the mask for the shuffle.
5685 SmallVector<int, 8> Mask(ShuffleVT.getVectorNumElements(), -1);
5686 int BitsPerShuffleLane = ShuffleVT.getVectorElementType().getSizeInBits();
5687 for (unsigned i = 0; i < VT.getVectorNumElements(); ++i) {
5688 SDValue Entry = Op.getOperand(i);
5689 if (Entry.getOpcode() == ISD::UNDEF)
5692 auto Src = std::find(Sources.begin(), Sources.end(), Entry.getOperand(0));
5693 int EltNo = cast<ConstantSDNode>(Entry.getOperand(1))->getSExtValue();
5695 // EXTRACT_VECTOR_ELT performs an implicit any_ext; BUILD_VECTOR an implicit
5696 // trunc. So only std::min(SrcBits, DestBits) actually get defined in this
5698 EVT OrigEltTy = Entry.getOperand(0).getValueType().getVectorElementType();
5699 int BitsDefined = std::min(OrigEltTy.getSizeInBits(),
5700 VT.getVectorElementType().getSizeInBits());
5701 int LanesDefined = BitsDefined / BitsPerShuffleLane;
5703 // This source is expected to fill ResMultiplier lanes of the final shuffle,
5704 // starting at the appropriate offset.
5705 int *LaneMask = &Mask[i * ResMultiplier];
5707 int ExtractBase = EltNo * Src->WindowScale + Src->WindowBase;
5708 ExtractBase += NumElts * (Src - Sources.begin());
5709 for (int j = 0; j < LanesDefined; ++j)
5710 LaneMask[j] = ExtractBase + j;
5713 // Final check before we try to produce nonsense...
5714 if (!isShuffleMaskLegal(Mask, ShuffleVT))
5717 // We can't handle more than two sources. This should have already
5718 // been checked before this point.
5719 assert(Sources.size() <= 2 && "Too many sources!");
5721 SDValue ShuffleOps[] = { DAG.getUNDEF(ShuffleVT), DAG.getUNDEF(ShuffleVT) };
5722 for (unsigned i = 0; i < Sources.size(); ++i)
5723 ShuffleOps[i] = Sources[i].ShuffleVec;
5725 SDValue Shuffle = DAG.getVectorShuffle(ShuffleVT, dl, ShuffleOps[0],
5726 ShuffleOps[1], &Mask[0]);
5727 return DAG.getNode(ISD::BITCAST, dl, VT, Shuffle);
5730 /// isShuffleMaskLegal - Targets can use this to indicate that they only
5731 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
5732 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
5733 /// are assumed to be legal.
5735 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
5737 if (VT.getVectorNumElements() == 4 &&
5738 (VT.is128BitVector() || VT.is64BitVector())) {
5739 unsigned PFIndexes[4];
5740 for (unsigned i = 0; i != 4; ++i) {
5744 PFIndexes[i] = M[i];
5747 // Compute the index in the perfect shuffle table.
5748 unsigned PFTableIndex =
5749 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
5750 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
5751 unsigned Cost = (PFEntry >> 30);
5757 bool ReverseVEXT, isV_UNDEF;
5758 unsigned Imm, WhichResult;
5760 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5761 return (EltSize >= 32 ||
5762 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
5763 isVREVMask(M, VT, 64) ||
5764 isVREVMask(M, VT, 32) ||
5765 isVREVMask(M, VT, 16) ||
5766 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
5767 isVTBLMask(M, VT) ||
5768 isNEONTwoResultShuffleMask(M, VT, WhichResult, isV_UNDEF) ||
5769 ((VT == MVT::v8i16 || VT == MVT::v16i8) && isReverseMask(M, VT)));
5772 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
5773 /// the specified operations to build the shuffle.
5774 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
5775 SDValue RHS, SelectionDAG &DAG,
5777 unsigned OpNum = (PFEntry >> 26) & 0x0F;
5778 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
5779 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
5782 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
5791 OP_VUZPL, // VUZP, left result
5792 OP_VUZPR, // VUZP, right result
5793 OP_VZIPL, // VZIP, left result
5794 OP_VZIPR, // VZIP, right result
5795 OP_VTRNL, // VTRN, left result
5796 OP_VTRNR // VTRN, right result
5799 if (OpNum == OP_COPY) {
5800 if (LHSID == (1*9+2)*9+3) return LHS;
5801 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
5805 SDValue OpLHS, OpRHS;
5806 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
5807 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
5808 EVT VT = OpLHS.getValueType();
5811 default: llvm_unreachable("Unknown shuffle opcode!");
5813 // VREV divides the vector in half and swaps within the half.
5814 if (VT.getVectorElementType() == MVT::i32 ||
5815 VT.getVectorElementType() == MVT::f32)
5816 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
5817 // vrev <4 x i16> -> VREV32
5818 if (VT.getVectorElementType() == MVT::i16)
5819 return DAG.getNode(ARMISD::VREV32, dl, VT, OpLHS);
5820 // vrev <4 x i8> -> VREV16
5821 assert(VT.getVectorElementType() == MVT::i8);
5822 return DAG.getNode(ARMISD::VREV16, dl, VT, OpLHS);
5827 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
5828 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, dl, MVT::i32));
5832 return DAG.getNode(ARMISD::VEXT, dl, VT,
5834 DAG.getConstant(OpNum - OP_VEXT1 + 1, dl, MVT::i32));
5837 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
5838 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
5841 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
5842 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
5845 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
5846 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
5850 static SDValue LowerVECTOR_SHUFFLEv8i8(SDValue Op,
5851 ArrayRef<int> ShuffleMask,
5852 SelectionDAG &DAG) {
5853 // Check to see if we can use the VTBL instruction.
5854 SDValue V1 = Op.getOperand(0);
5855 SDValue V2 = Op.getOperand(1);
5858 SmallVector<SDValue, 8> VTBLMask;
5859 for (ArrayRef<int>::iterator
5860 I = ShuffleMask.begin(), E = ShuffleMask.end(); I != E; ++I)
5861 VTBLMask.push_back(DAG.getConstant(*I, DL, MVT::i32));
5863 if (V2.getNode()->getOpcode() == ISD::UNDEF)
5864 return DAG.getNode(ARMISD::VTBL1, DL, MVT::v8i8, V1,
5865 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8, VTBLMask));
5867 return DAG.getNode(ARMISD::VTBL2, DL, MVT::v8i8, V1, V2,
5868 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8, VTBLMask));
5871 static SDValue LowerReverse_VECTOR_SHUFFLEv16i8_v8i16(SDValue Op,
5872 SelectionDAG &DAG) {
5874 SDValue OpLHS = Op.getOperand(0);
5875 EVT VT = OpLHS.getValueType();
5877 assert((VT == MVT::v8i16 || VT == MVT::v16i8) &&
5878 "Expect an v8i16/v16i8 type");
5879 OpLHS = DAG.getNode(ARMISD::VREV64, DL, VT, OpLHS);
5880 // For a v16i8 type: After the VREV, we have got <8, ...15, 8, ..., 0>. Now,
5881 // extract the first 8 bytes into the top double word and the last 8 bytes
5882 // into the bottom double word. The v8i16 case is similar.
5883 unsigned ExtractNum = (VT == MVT::v16i8) ? 8 : 4;
5884 return DAG.getNode(ARMISD::VEXT, DL, VT, OpLHS, OpLHS,
5885 DAG.getConstant(ExtractNum, DL, MVT::i32));
5888 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
5889 SDValue V1 = Op.getOperand(0);
5890 SDValue V2 = Op.getOperand(1);
5892 EVT VT = Op.getValueType();
5893 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
5895 // Convert shuffles that are directly supported on NEON to target-specific
5896 // DAG nodes, instead of keeping them as shuffles and matching them again
5897 // during code selection. This is more efficient and avoids the possibility
5898 // of inconsistencies between legalization and selection.
5899 // FIXME: floating-point vectors should be canonicalized to integer vectors
5900 // of the same time so that they get CSEd properly.
5901 ArrayRef<int> ShuffleMask = SVN->getMask();
5903 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5904 if (EltSize <= 32) {
5905 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
5906 int Lane = SVN->getSplatIndex();
5907 // If this is undef splat, generate it via "just" vdup, if possible.
5908 if (Lane == -1) Lane = 0;
5910 // Test if V1 is a SCALAR_TO_VECTOR.
5911 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
5912 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
5914 // Test if V1 is a BUILD_VECTOR which is equivalent to a SCALAR_TO_VECTOR
5915 // (and probably will turn into a SCALAR_TO_VECTOR once legalization
5917 if (Lane == 0 && V1.getOpcode() == ISD::BUILD_VECTOR &&
5918 !isa<ConstantSDNode>(V1.getOperand(0))) {
5919 bool IsScalarToVector = true;
5920 for (unsigned i = 1, e = V1.getNumOperands(); i != e; ++i)
5921 if (V1.getOperand(i).getOpcode() != ISD::UNDEF) {
5922 IsScalarToVector = false;
5925 if (IsScalarToVector)
5926 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
5928 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
5929 DAG.getConstant(Lane, dl, MVT::i32));
5934 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
5937 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
5938 DAG.getConstant(Imm, dl, MVT::i32));
5941 if (isVREVMask(ShuffleMask, VT, 64))
5942 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
5943 if (isVREVMask(ShuffleMask, VT, 32))
5944 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
5945 if (isVREVMask(ShuffleMask, VT, 16))
5946 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
5948 if (V2->getOpcode() == ISD::UNDEF &&
5949 isSingletonVEXTMask(ShuffleMask, VT, Imm)) {
5950 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V1,
5951 DAG.getConstant(Imm, dl, MVT::i32));
5954 // Check for Neon shuffles that modify both input vectors in place.
5955 // If both results are used, i.e., if there are two shuffles with the same
5956 // source operands and with masks corresponding to both results of one of
5957 // these operations, DAG memoization will ensure that a single node is
5958 // used for both shuffles.
5959 unsigned WhichResult;
5961 if (unsigned ShuffleOpc = isNEONTwoResultShuffleMask(
5962 ShuffleMask, VT, WhichResult, isV_UNDEF)) {
5965 return DAG.getNode(ShuffleOpc, dl, DAG.getVTList(VT, VT), V1, V2)
5966 .getValue(WhichResult);
5969 // Also check for these shuffles through CONCAT_VECTORS: we canonicalize
5970 // shuffles that produce a result larger than their operands with:
5971 // shuffle(concat(v1, undef), concat(v2, undef))
5973 // shuffle(concat(v1, v2), undef)
5974 // because we can access quad vectors (see PerformVECTOR_SHUFFLECombine).
5976 // This is useful in the general case, but there are special cases where
5977 // native shuffles produce larger results: the two-result ops.
5979 // Look through the concat when lowering them:
5980 // shuffle(concat(v1, v2), undef)
5982 // concat(VZIP(v1, v2):0, :1)
5984 if (V1->getOpcode() == ISD::CONCAT_VECTORS &&
5985 V2->getOpcode() == ISD::UNDEF) {
5986 SDValue SubV1 = V1->getOperand(0);
5987 SDValue SubV2 = V1->getOperand(1);
5988 EVT SubVT = SubV1.getValueType();
5990 // We expect these to have been canonicalized to -1.
5991 assert(std::all_of(ShuffleMask.begin(), ShuffleMask.end(), [&](int i) {
5992 return i < (int)VT.getVectorNumElements();
5993 }) && "Unexpected shuffle index into UNDEF operand!");
5995 if (unsigned ShuffleOpc = isNEONTwoResultShuffleMask(
5996 ShuffleMask, SubVT, WhichResult, isV_UNDEF)) {
5999 assert((WhichResult == 0) &&
6000 "In-place shuffle of concat can only have one result!");
6001 SDValue Res = DAG.getNode(ShuffleOpc, dl, DAG.getVTList(SubVT, SubVT),
6003 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, Res.getValue(0),
6009 // If the shuffle is not directly supported and it has 4 elements, use
6010 // the PerfectShuffle-generated table to synthesize it from other shuffles.
6011 unsigned NumElts = VT.getVectorNumElements();
6013 unsigned PFIndexes[4];
6014 for (unsigned i = 0; i != 4; ++i) {
6015 if (ShuffleMask[i] < 0)
6018 PFIndexes[i] = ShuffleMask[i];
6021 // Compute the index in the perfect shuffle table.
6022 unsigned PFTableIndex =
6023 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
6024 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
6025 unsigned Cost = (PFEntry >> 30);
6028 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
6031 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
6032 if (EltSize >= 32) {
6033 // Do the expansion with floating-point types, since that is what the VFP
6034 // registers are defined to use, and since i64 is not legal.
6035 EVT EltVT = EVT::getFloatingPointVT(EltSize);
6036 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
6037 V1 = DAG.getNode(ISD::BITCAST, dl, VecVT, V1);
6038 V2 = DAG.getNode(ISD::BITCAST, dl, VecVT, V2);
6039 SmallVector<SDValue, 8> Ops;
6040 for (unsigned i = 0; i < NumElts; ++i) {
6041 if (ShuffleMask[i] < 0)
6042 Ops.push_back(DAG.getUNDEF(EltVT));
6044 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
6045 ShuffleMask[i] < (int)NumElts ? V1 : V2,
6046 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
6049 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, Ops);
6050 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
6053 if ((VT == MVT::v8i16 || VT == MVT::v16i8) && isReverseMask(ShuffleMask, VT))
6054 return LowerReverse_VECTOR_SHUFFLEv16i8_v8i16(Op, DAG);
6056 if (VT == MVT::v8i8) {
6057 SDValue NewOp = LowerVECTOR_SHUFFLEv8i8(Op, ShuffleMask, DAG);
6058 if (NewOp.getNode())
6065 static SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
6066 // INSERT_VECTOR_ELT is legal only for immediate indexes.
6067 SDValue Lane = Op.getOperand(2);
6068 if (!isa<ConstantSDNode>(Lane))
6074 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
6075 // EXTRACT_VECTOR_ELT is legal only for immediate indexes.
6076 SDValue Lane = Op.getOperand(1);
6077 if (!isa<ConstantSDNode>(Lane))
6080 SDValue Vec = Op.getOperand(0);
6081 if (Op.getValueType() == MVT::i32 &&
6082 Vec.getValueType().getVectorElementType().getSizeInBits() < 32) {
6084 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
6090 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
6091 // The only time a CONCAT_VECTORS operation can have legal types is when
6092 // two 64-bit vectors are concatenated to a 128-bit vector.
6093 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
6094 "unexpected CONCAT_VECTORS");
6096 SDValue Val = DAG.getUNDEF(MVT::v2f64);
6097 SDValue Op0 = Op.getOperand(0);
6098 SDValue Op1 = Op.getOperand(1);
6099 if (Op0.getOpcode() != ISD::UNDEF)
6100 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
6101 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op0),
6102 DAG.getIntPtrConstant(0, dl));
6103 if (Op1.getOpcode() != ISD::UNDEF)
6104 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
6105 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op1),
6106 DAG.getIntPtrConstant(1, dl));
6107 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Val);
6110 /// isExtendedBUILD_VECTOR - Check if N is a constant BUILD_VECTOR where each
6111 /// element has been zero/sign-extended, depending on the isSigned parameter,
6112 /// from an integer type half its size.
6113 static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG,
6115 // A v2i64 BUILD_VECTOR will have been legalized to a BITCAST from v4i32.
6116 EVT VT = N->getValueType(0);
6117 if (VT == MVT::v2i64 && N->getOpcode() == ISD::BITCAST) {
6118 SDNode *BVN = N->getOperand(0).getNode();
6119 if (BVN->getValueType(0) != MVT::v4i32 ||
6120 BVN->getOpcode() != ISD::BUILD_VECTOR)
6122 unsigned LoElt = DAG.getDataLayout().isBigEndian() ? 1 : 0;
6123 unsigned HiElt = 1 - LoElt;
6124 ConstantSDNode *Lo0 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt));
6125 ConstantSDNode *Hi0 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt));
6126 ConstantSDNode *Lo1 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt+2));
6127 ConstantSDNode *Hi1 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt+2));
6128 if (!Lo0 || !Hi0 || !Lo1 || !Hi1)
6131 if (Hi0->getSExtValue() == Lo0->getSExtValue() >> 32 &&
6132 Hi1->getSExtValue() == Lo1->getSExtValue() >> 32)
6135 if (Hi0->isNullValue() && Hi1->isNullValue())
6141 if (N->getOpcode() != ISD::BUILD_VECTOR)
6144 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
6145 SDNode *Elt = N->getOperand(i).getNode();
6146 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
6147 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
6148 unsigned HalfSize = EltSize / 2;
6150 if (!isIntN(HalfSize, C->getSExtValue()))
6153 if (!isUIntN(HalfSize, C->getZExtValue()))
6164 /// isSignExtended - Check if a node is a vector value that is sign-extended
6165 /// or a constant BUILD_VECTOR with sign-extended elements.
6166 static bool isSignExtended(SDNode *N, SelectionDAG &DAG) {
6167 if (N->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N))
6169 if (isExtendedBUILD_VECTOR(N, DAG, true))
6174 /// isZeroExtended - Check if a node is a vector value that is zero-extended
6175 /// or a constant BUILD_VECTOR with zero-extended elements.
6176 static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) {
6177 if (N->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N))
6179 if (isExtendedBUILD_VECTOR(N, DAG, false))
6184 static EVT getExtensionTo64Bits(const EVT &OrigVT) {
6185 if (OrigVT.getSizeInBits() >= 64)
6188 assert(OrigVT.isSimple() && "Expecting a simple value type");
6190 MVT::SimpleValueType OrigSimpleTy = OrigVT.getSimpleVT().SimpleTy;
6191 switch (OrigSimpleTy) {
6192 default: llvm_unreachable("Unexpected Vector Type");
6201 /// AddRequiredExtensionForVMULL - Add a sign/zero extension to extend the total
6202 /// value size to 64 bits. We need a 64-bit D register as an operand to VMULL.
6203 /// We insert the required extension here to get the vector to fill a D register.
6204 static SDValue AddRequiredExtensionForVMULL(SDValue N, SelectionDAG &DAG,
6207 unsigned ExtOpcode) {
6208 // The vector originally had a size of OrigTy. It was then extended to ExtTy.
6209 // We expect the ExtTy to be 128-bits total. If the OrigTy is less than
6210 // 64-bits we need to insert a new extension so that it will be 64-bits.
6211 assert(ExtTy.is128BitVector() && "Unexpected extension size");
6212 if (OrigTy.getSizeInBits() >= 64)
6215 // Must extend size to at least 64 bits to be used as an operand for VMULL.
6216 EVT NewVT = getExtensionTo64Bits(OrigTy);
6218 return DAG.getNode(ExtOpcode, SDLoc(N), NewVT, N);
6221 /// SkipLoadExtensionForVMULL - return a load of the original vector size that
6222 /// does not do any sign/zero extension. If the original vector is less
6223 /// than 64 bits, an appropriate extension will be added after the load to
6224 /// reach a total size of 64 bits. We have to add the extension separately
6225 /// because ARM does not have a sign/zero extending load for vectors.
6226 static SDValue SkipLoadExtensionForVMULL(LoadSDNode *LD, SelectionDAG& DAG) {
6227 EVT ExtendedTy = getExtensionTo64Bits(LD->getMemoryVT());
6229 // The load already has the right type.
6230 if (ExtendedTy == LD->getMemoryVT())
6231 return DAG.getLoad(LD->getMemoryVT(), SDLoc(LD), LD->getChain(),
6232 LD->getBasePtr(), LD->getPointerInfo(), LD->isVolatile(),
6233 LD->isNonTemporal(), LD->isInvariant(),
6234 LD->getAlignment());
6236 // We need to create a zextload/sextload. We cannot just create a load
6237 // followed by a zext/zext node because LowerMUL is also run during normal
6238 // operation legalization where we can't create illegal types.
6239 return DAG.getExtLoad(LD->getExtensionType(), SDLoc(LD), ExtendedTy,
6240 LD->getChain(), LD->getBasePtr(), LD->getPointerInfo(),
6241 LD->getMemoryVT(), LD->isVolatile(), LD->isInvariant(),
6242 LD->isNonTemporal(), LD->getAlignment());
6245 /// SkipExtensionForVMULL - For a node that is a SIGN_EXTEND, ZERO_EXTEND,
6246 /// extending load, or BUILD_VECTOR with extended elements, return the
6247 /// unextended value. The unextended vector should be 64 bits so that it can
6248 /// be used as an operand to a VMULL instruction. If the original vector size
6249 /// before extension is less than 64 bits we add a an extension to resize
6250 /// the vector to 64 bits.
6251 static SDValue SkipExtensionForVMULL(SDNode *N, SelectionDAG &DAG) {
6252 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
6253 return AddRequiredExtensionForVMULL(N->getOperand(0), DAG,
6254 N->getOperand(0)->getValueType(0),
6258 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
6259 return SkipLoadExtensionForVMULL(LD, DAG);
6261 // Otherwise, the value must be a BUILD_VECTOR. For v2i64, it will
6262 // have been legalized as a BITCAST from v4i32.
6263 if (N->getOpcode() == ISD::BITCAST) {
6264 SDNode *BVN = N->getOperand(0).getNode();
6265 assert(BVN->getOpcode() == ISD::BUILD_VECTOR &&
6266 BVN->getValueType(0) == MVT::v4i32 && "expected v4i32 BUILD_VECTOR");
6267 unsigned LowElt = DAG.getDataLayout().isBigEndian() ? 1 : 0;
6268 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), MVT::v2i32,
6269 BVN->getOperand(LowElt), BVN->getOperand(LowElt+2));
6271 // Construct a new BUILD_VECTOR with elements truncated to half the size.
6272 assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR");
6273 EVT VT = N->getValueType(0);
6274 unsigned EltSize = VT.getVectorElementType().getSizeInBits() / 2;
6275 unsigned NumElts = VT.getVectorNumElements();
6276 MVT TruncVT = MVT::getIntegerVT(EltSize);
6277 SmallVector<SDValue, 8> Ops;
6279 for (unsigned i = 0; i != NumElts; ++i) {
6280 ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i));
6281 const APInt &CInt = C->getAPIntValue();
6282 // Element types smaller than 32 bits are not legal, so use i32 elements.
6283 // The values are implicitly truncated so sext vs. zext doesn't matter.
6284 Ops.push_back(DAG.getConstant(CInt.zextOrTrunc(32), dl, MVT::i32));
6286 return DAG.getNode(ISD::BUILD_VECTOR, dl,
6287 MVT::getVectorVT(TruncVT, NumElts), Ops);
6290 static bool isAddSubSExt(SDNode *N, SelectionDAG &DAG) {
6291 unsigned Opcode = N->getOpcode();
6292 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
6293 SDNode *N0 = N->getOperand(0).getNode();
6294 SDNode *N1 = N->getOperand(1).getNode();
6295 return N0->hasOneUse() && N1->hasOneUse() &&
6296 isSignExtended(N0, DAG) && isSignExtended(N1, DAG);
6301 static bool isAddSubZExt(SDNode *N, SelectionDAG &DAG) {
6302 unsigned Opcode = N->getOpcode();
6303 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
6304 SDNode *N0 = N->getOperand(0).getNode();
6305 SDNode *N1 = N->getOperand(1).getNode();
6306 return N0->hasOneUse() && N1->hasOneUse() &&
6307 isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG);
6312 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
6313 // Multiplications are only custom-lowered for 128-bit vectors so that
6314 // VMULL can be detected. Otherwise v2i64 multiplications are not legal.
6315 EVT VT = Op.getValueType();
6316 assert(VT.is128BitVector() && VT.isInteger() &&
6317 "unexpected type for custom-lowering ISD::MUL");
6318 SDNode *N0 = Op.getOperand(0).getNode();
6319 SDNode *N1 = Op.getOperand(1).getNode();
6320 unsigned NewOpc = 0;
6322 bool isN0SExt = isSignExtended(N0, DAG);
6323 bool isN1SExt = isSignExtended(N1, DAG);
6324 if (isN0SExt && isN1SExt)
6325 NewOpc = ARMISD::VMULLs;
6327 bool isN0ZExt = isZeroExtended(N0, DAG);
6328 bool isN1ZExt = isZeroExtended(N1, DAG);
6329 if (isN0ZExt && isN1ZExt)
6330 NewOpc = ARMISD::VMULLu;
6331 else if (isN1SExt || isN1ZExt) {
6332 // Look for (s/zext A + s/zext B) * (s/zext C). We want to turn these
6333 // into (s/zext A * s/zext C) + (s/zext B * s/zext C)
6334 if (isN1SExt && isAddSubSExt(N0, DAG)) {
6335 NewOpc = ARMISD::VMULLs;
6337 } else if (isN1ZExt && isAddSubZExt(N0, DAG)) {
6338 NewOpc = ARMISD::VMULLu;
6340 } else if (isN0ZExt && isAddSubZExt(N1, DAG)) {
6342 NewOpc = ARMISD::VMULLu;
6348 if (VT == MVT::v2i64)
6349 // Fall through to expand this. It is not legal.
6352 // Other vector multiplications are legal.
6357 // Legalize to a VMULL instruction.
6360 SDValue Op1 = SkipExtensionForVMULL(N1, DAG);
6362 Op0 = SkipExtensionForVMULL(N0, DAG);
6363 assert(Op0.getValueType().is64BitVector() &&
6364 Op1.getValueType().is64BitVector() &&
6365 "unexpected types for extended operands to VMULL");
6366 return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
6369 // Optimizing (zext A + zext B) * C, to (VMULL A, C) + (VMULL B, C) during
6370 // isel lowering to take advantage of no-stall back to back vmul + vmla.
6377 SDValue N00 = SkipExtensionForVMULL(N0->getOperand(0).getNode(), DAG);
6378 SDValue N01 = SkipExtensionForVMULL(N0->getOperand(1).getNode(), DAG);
6379 EVT Op1VT = Op1.getValueType();
6380 return DAG.getNode(N0->getOpcode(), DL, VT,
6381 DAG.getNode(NewOpc, DL, VT,
6382 DAG.getNode(ISD::BITCAST, DL, Op1VT, N00), Op1),
6383 DAG.getNode(NewOpc, DL, VT,
6384 DAG.getNode(ISD::BITCAST, DL, Op1VT, N01), Op1));
6388 LowerSDIV_v4i8(SDValue X, SDValue Y, SDLoc dl, SelectionDAG &DAG) {
6389 // TODO: Should this propagate fast-math-flags?
6392 // float4 xf = vcvt_f32_s32(vmovl_s16(a.lo));
6393 // float4 yf = vcvt_f32_s32(vmovl_s16(b.lo));
6394 X = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, X);
6395 Y = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, Y);
6396 X = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, X);
6397 Y = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, Y);
6398 // Get reciprocal estimate.
6399 // float4 recip = vrecpeq_f32(yf);
6400 Y = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6401 DAG.getConstant(Intrinsic::arm_neon_vrecpe, dl, MVT::i32),
6403 // Because char has a smaller range than uchar, we can actually get away
6404 // without any newton steps. This requires that we use a weird bias
6405 // of 0xb000, however (again, this has been exhaustively tested).
6406 // float4 result = as_float4(as_int4(xf*recip) + 0xb000);
6407 X = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, X, Y);
6408 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, X);
6409 Y = DAG.getConstant(0xb000, dl, MVT::i32);
6410 Y = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Y, Y, Y, Y);
6411 X = DAG.getNode(ISD::ADD, dl, MVT::v4i32, X, Y);
6412 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, X);
6413 // Convert back to short.
6414 X = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, X);
6415 X = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, X);
6420 LowerSDIV_v4i16(SDValue N0, SDValue N1, SDLoc dl, SelectionDAG &DAG) {
6421 // TODO: Should this propagate fast-math-flags?
6424 // Convert to float.
6425 // float4 yf = vcvt_f32_s32(vmovl_s16(y));
6426 // float4 xf = vcvt_f32_s32(vmovl_s16(x));
6427 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N0);
6428 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N1);
6429 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
6430 N1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
6432 // Use reciprocal estimate and one refinement step.
6433 // float4 recip = vrecpeq_f32(yf);
6434 // recip *= vrecpsq_f32(yf, recip);
6435 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6436 DAG.getConstant(Intrinsic::arm_neon_vrecpe, dl, MVT::i32),
6438 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6439 DAG.getConstant(Intrinsic::arm_neon_vrecps, dl, MVT::i32),
6441 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
6442 // Because short has a smaller range than ushort, we can actually get away
6443 // with only a single newton step. This requires that we use a weird bias
6444 // of 89, however (again, this has been exhaustively tested).
6445 // float4 result = as_float4(as_int4(xf*recip) + 0x89);
6446 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
6447 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
6448 N1 = DAG.getConstant(0x89, dl, MVT::i32);
6449 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
6450 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
6451 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
6452 // Convert back to integer and return.
6453 // return vmovn_s32(vcvt_s32_f32(result));
6454 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
6455 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
6459 static SDValue LowerSDIV(SDValue Op, SelectionDAG &DAG) {
6460 EVT VT = Op.getValueType();
6461 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
6462 "unexpected type for custom-lowering ISD::SDIV");
6465 SDValue N0 = Op.getOperand(0);
6466 SDValue N1 = Op.getOperand(1);
6469 if (VT == MVT::v8i8) {
6470 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N0);
6471 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N1);
6473 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
6474 DAG.getIntPtrConstant(4, dl));
6475 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
6476 DAG.getIntPtrConstant(4, dl));
6477 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
6478 DAG.getIntPtrConstant(0, dl));
6479 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
6480 DAG.getIntPtrConstant(0, dl));
6482 N0 = LowerSDIV_v4i8(N0, N1, dl, DAG); // v4i16
6483 N2 = LowerSDIV_v4i8(N2, N3, dl, DAG); // v4i16
6485 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
6486 N0 = LowerCONCAT_VECTORS(N0, DAG);
6488 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v8i8, N0);
6491 return LowerSDIV_v4i16(N0, N1, dl, DAG);
6494 static SDValue LowerUDIV(SDValue Op, SelectionDAG &DAG) {
6495 // TODO: Should this propagate fast-math-flags?
6496 EVT VT = Op.getValueType();
6497 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
6498 "unexpected type for custom-lowering ISD::UDIV");
6501 SDValue N0 = Op.getOperand(0);
6502 SDValue N1 = Op.getOperand(1);
6505 if (VT == MVT::v8i8) {
6506 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N0);
6507 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N1);
6509 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
6510 DAG.getIntPtrConstant(4, dl));
6511 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
6512 DAG.getIntPtrConstant(4, dl));
6513 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
6514 DAG.getIntPtrConstant(0, dl));
6515 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
6516 DAG.getIntPtrConstant(0, dl));
6518 N0 = LowerSDIV_v4i16(N0, N1, dl, DAG); // v4i16
6519 N2 = LowerSDIV_v4i16(N2, N3, dl, DAG); // v4i16
6521 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
6522 N0 = LowerCONCAT_VECTORS(N0, DAG);
6524 N0 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v8i8,
6525 DAG.getConstant(Intrinsic::arm_neon_vqmovnsu, dl,
6531 // v4i16 sdiv ... Convert to float.
6532 // float4 yf = vcvt_f32_s32(vmovl_u16(y));
6533 // float4 xf = vcvt_f32_s32(vmovl_u16(x));
6534 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N0);
6535 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N1);
6536 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
6537 SDValue BN1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
6539 // Use reciprocal estimate and two refinement steps.
6540 // float4 recip = vrecpeq_f32(yf);
6541 // recip *= vrecpsq_f32(yf, recip);
6542 // recip *= vrecpsq_f32(yf, recip);
6543 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6544 DAG.getConstant(Intrinsic::arm_neon_vrecpe, dl, MVT::i32),
6546 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6547 DAG.getConstant(Intrinsic::arm_neon_vrecps, dl, MVT::i32),
6549 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
6550 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6551 DAG.getConstant(Intrinsic::arm_neon_vrecps, dl, MVT::i32),
6553 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
6554 // Simply multiplying by the reciprocal estimate can leave us a few ulps
6555 // too low, so we add 2 ulps (exhaustive testing shows that this is enough,
6556 // and that it will never cause us to return an answer too large).
6557 // float4 result = as_float4(as_int4(xf*recip) + 2);
6558 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
6559 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
6560 N1 = DAG.getConstant(2, dl, MVT::i32);
6561 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
6562 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
6563 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
6564 // Convert back to integer and return.
6565 // return vmovn_u32(vcvt_s32_f32(result));
6566 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
6567 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
6571 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
6572 EVT VT = Op.getNode()->getValueType(0);
6573 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
6576 bool ExtraOp = false;
6577 switch (Op.getOpcode()) {
6578 default: llvm_unreachable("Invalid code");
6579 case ISD::ADDC: Opc = ARMISD::ADDC; break;
6580 case ISD::ADDE: Opc = ARMISD::ADDE; ExtraOp = true; break;
6581 case ISD::SUBC: Opc = ARMISD::SUBC; break;
6582 case ISD::SUBE: Opc = ARMISD::SUBE; ExtraOp = true; break;
6586 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
6588 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
6589 Op.getOperand(1), Op.getOperand(2));
6592 SDValue ARMTargetLowering::LowerFSINCOS(SDValue Op, SelectionDAG &DAG) const {
6593 assert(Subtarget->isTargetDarwin());
6595 // For iOS, we want to call an alternative entry point: __sincos_stret,
6596 // return values are passed via sret.
6598 SDValue Arg = Op.getOperand(0);
6599 EVT ArgVT = Arg.getValueType();
6600 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
6601 auto PtrVT = getPointerTy(DAG.getDataLayout());
6603 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
6604 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6606 // Pair of floats / doubles used to pass the result.
6607 Type *RetTy = StructType::get(ArgTy, ArgTy, nullptr);
6608 auto &DL = DAG.getDataLayout();
6611 bool ShouldUseSRet = Subtarget->isAPCS_ABI();
6613 if (ShouldUseSRet) {
6614 // Create stack object for sret.
6615 const uint64_t ByteSize = DL.getTypeAllocSize(RetTy);
6616 const unsigned StackAlign = DL.getPrefTypeAlignment(RetTy);
6617 int FrameIdx = FrameInfo->CreateStackObject(ByteSize, StackAlign, false);
6618 SRet = DAG.getFrameIndex(FrameIdx, TLI.getPointerTy(DL));
6622 Entry.Ty = RetTy->getPointerTo();
6623 Entry.isSExt = false;
6624 Entry.isZExt = false;
6625 Entry.isSRet = true;
6626 Args.push_back(Entry);
6627 RetTy = Type::getVoidTy(*DAG.getContext());
6633 Entry.isSExt = false;
6634 Entry.isZExt = false;
6635 Args.push_back(Entry);
6637 const char *LibcallName =
6638 (ArgVT == MVT::f64) ? "__sincos_stret" : "__sincosf_stret";
6640 (ArgVT == MVT::f64) ? RTLIB::SINCOS_F64 : RTLIB::SINCOS_F32;
6641 CallingConv::ID CC = getLibcallCallingConv(LC);
6642 SDValue Callee = DAG.getExternalSymbol(LibcallName, getPointerTy(DL));
6644 TargetLowering::CallLoweringInfo CLI(DAG);
6646 .setChain(DAG.getEntryNode())
6647 .setCallee(CC, RetTy, Callee, std::move(Args), 0)
6648 .setDiscardResult(ShouldUseSRet);
6649 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
6652 return CallResult.first;
6654 SDValue LoadSin = DAG.getLoad(ArgVT, dl, CallResult.second, SRet,
6655 MachinePointerInfo(), false, false, false, 0);
6657 // Address of cos field.
6658 SDValue Add = DAG.getNode(ISD::ADD, dl, PtrVT, SRet,
6659 DAG.getIntPtrConstant(ArgVT.getStoreSize(), dl));
6660 SDValue LoadCos = DAG.getLoad(ArgVT, dl, LoadSin.getValue(1), Add,
6661 MachinePointerInfo(), false, false, false, 0);
6663 SDVTList Tys = DAG.getVTList(ArgVT, ArgVT);
6664 return DAG.getNode(ISD::MERGE_VALUES, dl, Tys,
6665 LoadSin.getValue(0), LoadCos.getValue(0));
6668 SDValue ARMTargetLowering::LowerWindowsDIVLibCall(SDValue Op, SelectionDAG &DAG,
6670 SDValue &Chain) const {
6671 EVT VT = Op.getValueType();
6672 assert((VT == MVT::i32 || VT == MVT::i64) &&
6673 "unexpected type for custom lowering DIV");
6676 const auto &DL = DAG.getDataLayout();
6677 const auto &TLI = DAG.getTargetLoweringInfo();
6679 const char *Name = nullptr;
6681 Name = (VT == MVT::i32) ? "__rt_sdiv" : "__rt_sdiv64";
6683 Name = (VT == MVT::i32) ? "__rt_udiv" : "__rt_udiv64";
6685 SDValue ES = DAG.getExternalSymbol(Name, TLI.getPointerTy(DL));
6687 ARMTargetLowering::ArgListTy Args;
6689 for (auto AI : {1, 0}) {
6691 Arg.Node = Op.getOperand(AI);
6692 Arg.Ty = Arg.Node.getValueType().getTypeForEVT(*DAG.getContext());
6693 Args.push_back(Arg);
6696 CallLoweringInfo CLI(DAG);
6699 .setCallee(CallingConv::ARM_AAPCS_VFP, VT.getTypeForEVT(*DAG.getContext()),
6700 ES, std::move(Args), 0);
6702 return LowerCallTo(CLI).first;
6705 SDValue ARMTargetLowering::LowerDIV_Windows(SDValue Op, SelectionDAG &DAG,
6706 bool Signed) const {
6707 assert(Op.getValueType() == MVT::i32 &&
6708 "unexpected type for custom lowering DIV");
6711 SDValue DBZCHK = DAG.getNode(ARMISD::WIN__DBZCHK, dl, MVT::Other,
6712 DAG.getEntryNode(), Op.getOperand(1));
6714 return LowerWindowsDIVLibCall(Op, DAG, Signed, DBZCHK);
6717 void ARMTargetLowering::ExpandDIV_Windows(
6718 SDValue Op, SelectionDAG &DAG, bool Signed,
6719 SmallVectorImpl<SDValue> &Results) const {
6720 const auto &DL = DAG.getDataLayout();
6721 const auto &TLI = DAG.getTargetLoweringInfo();
6723 assert(Op.getValueType() == MVT::i64 &&
6724 "unexpected type for custom lowering DIV");
6727 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op.getOperand(1),
6728 DAG.getConstant(0, dl, MVT::i32));
6729 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op.getOperand(1),
6730 DAG.getConstant(1, dl, MVT::i32));
6731 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::i32, Lo, Hi);
6734 DAG.getNode(ARMISD::WIN__DBZCHK, dl, MVT::Other, DAG.getEntryNode(), Or);
6736 SDValue Result = LowerWindowsDIVLibCall(Op, DAG, Signed, DBZCHK);
6738 SDValue Lower = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Result);
6739 SDValue Upper = DAG.getNode(ISD::SRL, dl, MVT::i64, Result,
6740 DAG.getConstant(32, dl, TLI.getPointerTy(DL)));
6741 Upper = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Upper);
6743 Results.push_back(Lower);
6744 Results.push_back(Upper);
6747 static SDValue LowerAtomicLoadStore(SDValue Op, SelectionDAG &DAG) {
6748 // Monotonic load/store is legal for all targets
6749 if (cast<AtomicSDNode>(Op)->getOrdering() <= Monotonic)
6752 // Acquire/Release load/store is not legal for targets without a
6753 // dmb or equivalent available.
6757 static void ReplaceREADCYCLECOUNTER(SDNode *N,
6758 SmallVectorImpl<SDValue> &Results,
6760 const ARMSubtarget *Subtarget) {
6762 // Under Power Management extensions, the cycle-count is:
6763 // mrc p15, #0, <Rt>, c9, c13, #0
6764 SDValue Ops[] = { N->getOperand(0), // Chain
6765 DAG.getConstant(Intrinsic::arm_mrc, DL, MVT::i32),
6766 DAG.getConstant(15, DL, MVT::i32),
6767 DAG.getConstant(0, DL, MVT::i32),
6768 DAG.getConstant(9, DL, MVT::i32),
6769 DAG.getConstant(13, DL, MVT::i32),
6770 DAG.getConstant(0, DL, MVT::i32)
6773 SDValue Cycles32 = DAG.getNode(ISD::INTRINSIC_W_CHAIN, DL,
6774 DAG.getVTList(MVT::i32, MVT::Other), Ops);
6775 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Cycles32,
6776 DAG.getConstant(0, DL, MVT::i32)));
6777 Results.push_back(Cycles32.getValue(1));
6780 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
6781 switch (Op.getOpcode()) {
6782 default: llvm_unreachable("Don't know how to custom lower this!");
6783 case ISD::WRITE_REGISTER: return LowerWRITE_REGISTER(Op, DAG);
6784 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
6785 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
6786 case ISD::GlobalAddress:
6787 switch (Subtarget->getTargetTriple().getObjectFormat()) {
6788 default: llvm_unreachable("unknown object format");
6790 return LowerGlobalAddressWindows(Op, DAG);
6792 return LowerGlobalAddressELF(Op, DAG);
6794 return LowerGlobalAddressDarwin(Op, DAG);
6796 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
6797 case ISD::SELECT: return LowerSELECT(Op, DAG);
6798 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
6799 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
6800 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
6801 case ISD::VASTART: return LowerVASTART(Op, DAG);
6802 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, DAG, Subtarget);
6803 case ISD::PREFETCH: return LowerPREFETCH(Op, DAG, Subtarget);
6804 case ISD::SINT_TO_FP:
6805 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
6806 case ISD::FP_TO_SINT:
6807 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
6808 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
6809 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
6810 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
6811 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
6812 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
6813 case ISD::EH_SJLJ_SETUP_DISPATCH: return LowerEH_SJLJ_SETUP_DISPATCH(Op, DAG);
6814 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
6816 case ISD::BITCAST: return ExpandBITCAST(Op.getNode(), DAG);
6819 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
6820 case ISD::SREM: return LowerREM(Op.getNode(), DAG);
6821 case ISD::UREM: return LowerREM(Op.getNode(), DAG);
6822 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
6823 case ISD::SRL_PARTS:
6824 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
6826 case ISD::CTTZ_ZERO_UNDEF: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
6827 case ISD::CTPOP: return LowerCTPOP(Op.getNode(), DAG, Subtarget);
6828 case ISD::SETCC: return LowerVSETCC(Op, DAG);
6829 case ISD::ConstantFP: return LowerConstantFP(Op, DAG, Subtarget);
6830 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG, Subtarget);
6831 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
6832 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
6833 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
6834 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
6835 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
6836 case ISD::MUL: return LowerMUL(Op, DAG);
6838 if (Subtarget->isTargetWindows())
6839 return LowerDIV_Windows(Op, DAG, /* Signed */ true);
6840 return LowerSDIV(Op, DAG);
6842 if (Subtarget->isTargetWindows())
6843 return LowerDIV_Windows(Op, DAG, /* Signed */ false);
6844 return LowerUDIV(Op, DAG);
6848 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
6853 return LowerXALUO(Op, DAG);
6854 case ISD::ATOMIC_LOAD:
6855 case ISD::ATOMIC_STORE: return LowerAtomicLoadStore(Op, DAG);
6856 case ISD::FSINCOS: return LowerFSINCOS(Op, DAG);
6858 case ISD::UDIVREM: return LowerDivRem(Op, DAG);
6859 case ISD::DYNAMIC_STACKALLOC:
6860 if (Subtarget->getTargetTriple().isWindowsItaniumEnvironment())
6861 return LowerDYNAMIC_STACKALLOC(Op, DAG);
6862 llvm_unreachable("Don't know how to custom lower this!");
6863 case ISD::FP_ROUND: return LowerFP_ROUND(Op, DAG);
6864 case ISD::FP_EXTEND: return LowerFP_EXTEND(Op, DAG);
6865 case ARMISD::WIN__DBZCHK: return SDValue();
6869 /// ReplaceNodeResults - Replace the results of node with an illegal result
6870 /// type with new values built out of custom code.
6871 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
6872 SmallVectorImpl<SDValue> &Results,
6873 SelectionDAG &DAG) const {
6875 switch (N->getOpcode()) {
6877 llvm_unreachable("Don't know how to custom expand this!");
6878 case ISD::READ_REGISTER:
6879 ExpandREAD_REGISTER(N, Results, DAG);
6882 Res = ExpandBITCAST(N, DAG);
6886 Res = Expand64BitShift(N, DAG, Subtarget);
6890 Res = LowerREM(N, DAG);
6892 case ISD::READCYCLECOUNTER:
6893 ReplaceREADCYCLECOUNTER(N, Results, DAG, Subtarget);
6897 assert(Subtarget->isTargetWindows() && "can only expand DIV on Windows");
6898 return ExpandDIV_Windows(SDValue(N, 0), DAG, N->getOpcode() == ISD::SDIV,
6902 Results.push_back(Res);
6905 //===----------------------------------------------------------------------===//
6906 // ARM Scheduler Hooks
6907 //===----------------------------------------------------------------------===//
6909 /// SetupEntryBlockForSjLj - Insert code into the entry block that creates and
6910 /// registers the function context.
6911 void ARMTargetLowering::
6912 SetupEntryBlockForSjLj(MachineInstr *MI, MachineBasicBlock *MBB,
6913 MachineBasicBlock *DispatchBB, int FI) const {
6914 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
6915 DebugLoc dl = MI->getDebugLoc();
6916 MachineFunction *MF = MBB->getParent();
6917 MachineRegisterInfo *MRI = &MF->getRegInfo();
6918 MachineConstantPool *MCP = MF->getConstantPool();
6919 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
6920 const Function *F = MF->getFunction();
6922 bool isThumb = Subtarget->isThumb();
6923 bool isThumb2 = Subtarget->isThumb2();
6925 unsigned PCLabelId = AFI->createPICLabelUId();
6926 unsigned PCAdj = (isThumb || isThumb2) ? 4 : 8;
6927 ARMConstantPoolValue *CPV =
6928 ARMConstantPoolMBB::Create(F->getContext(), DispatchBB, PCLabelId, PCAdj);
6929 unsigned CPI = MCP->getConstantPoolIndex(CPV, 4);
6931 const TargetRegisterClass *TRC = isThumb ? &ARM::tGPRRegClass
6932 : &ARM::GPRRegClass;
6934 // Grab constant pool and fixed stack memory operands.
6935 MachineMemOperand *CPMMO =
6936 MF->getMachineMemOperand(MachinePointerInfo::getConstantPool(*MF),
6937 MachineMemOperand::MOLoad, 4, 4);
6939 MachineMemOperand *FIMMOSt =
6940 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(*MF, FI),
6941 MachineMemOperand::MOStore, 4, 4);
6943 // Load the address of the dispatch MBB into the jump buffer.
6945 // Incoming value: jbuf
6946 // ldr.n r5, LCPI1_1
6949 // str r5, [$jbuf, #+4] ; &jbuf[1]
6950 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6951 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2LDRpci), NewVReg1)
6952 .addConstantPoolIndex(CPI)
6953 .addMemOperand(CPMMO));
6954 // Set the low bit because of thumb mode.
6955 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6957 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2ORRri), NewVReg2)
6958 .addReg(NewVReg1, RegState::Kill)
6960 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6961 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg3)
6962 .addReg(NewVReg2, RegState::Kill)
6964 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2STRi12))
6965 .addReg(NewVReg3, RegState::Kill)
6967 .addImm(36) // &jbuf[1] :: pc
6968 .addMemOperand(FIMMOSt));
6969 } else if (isThumb) {
6970 // Incoming value: jbuf
6971 // ldr.n r1, LCPI1_4
6975 // add r2, $jbuf, #+4 ; &jbuf[1]
6977 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6978 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tLDRpci), NewVReg1)
6979 .addConstantPoolIndex(CPI)
6980 .addMemOperand(CPMMO));
6981 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6982 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg2)
6983 .addReg(NewVReg1, RegState::Kill)
6985 // Set the low bit because of thumb mode.
6986 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6987 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tMOVi8), NewVReg3)
6988 .addReg(ARM::CPSR, RegState::Define)
6990 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6991 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tORR), NewVReg4)
6992 .addReg(ARM::CPSR, RegState::Define)
6993 .addReg(NewVReg2, RegState::Kill)
6994 .addReg(NewVReg3, RegState::Kill));
6995 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6996 BuildMI(*MBB, MI, dl, TII->get(ARM::tADDframe), NewVReg5)
6998 .addImm(36); // &jbuf[1] :: pc
6999 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tSTRi))
7000 .addReg(NewVReg4, RegState::Kill)
7001 .addReg(NewVReg5, RegState::Kill)
7003 .addMemOperand(FIMMOSt));
7005 // Incoming value: jbuf
7008 // str r1, [$jbuf, #+4] ; &jbuf[1]
7009 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
7010 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::LDRi12), NewVReg1)
7011 .addConstantPoolIndex(CPI)
7013 .addMemOperand(CPMMO));
7014 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
7015 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::PICADD), NewVReg2)
7016 .addReg(NewVReg1, RegState::Kill)
7017 .addImm(PCLabelId));
7018 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::STRi12))
7019 .addReg(NewVReg2, RegState::Kill)
7021 .addImm(36) // &jbuf[1] :: pc
7022 .addMemOperand(FIMMOSt));
7026 void ARMTargetLowering::EmitSjLjDispatchBlock(MachineInstr *MI,
7027 MachineBasicBlock *MBB) const {
7028 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
7029 DebugLoc dl = MI->getDebugLoc();
7030 MachineFunction *MF = MBB->getParent();
7031 MachineRegisterInfo *MRI = &MF->getRegInfo();
7032 MachineFrameInfo *MFI = MF->getFrameInfo();
7033 int FI = MFI->getFunctionContextIndex();
7035 const TargetRegisterClass *TRC = Subtarget->isThumb() ? &ARM::tGPRRegClass
7036 : &ARM::GPRnopcRegClass;
7038 // Get a mapping of the call site numbers to all of the landing pads they're
7040 DenseMap<unsigned, SmallVector<MachineBasicBlock*, 2> > CallSiteNumToLPad;
7041 unsigned MaxCSNum = 0;
7042 MachineModuleInfo &MMI = MF->getMMI();
7043 for (MachineFunction::iterator BB = MF->begin(), E = MF->end(); BB != E;
7045 if (!BB->isEHPad()) continue;
7047 // FIXME: We should assert that the EH_LABEL is the first MI in the landing
7049 for (MachineBasicBlock::iterator
7050 II = BB->begin(), IE = BB->end(); II != IE; ++II) {
7051 if (!II->isEHLabel()) continue;
7053 MCSymbol *Sym = II->getOperand(0).getMCSymbol();
7054 if (!MMI.hasCallSiteLandingPad(Sym)) continue;
7056 SmallVectorImpl<unsigned> &CallSiteIdxs = MMI.getCallSiteLandingPad(Sym);
7057 for (SmallVectorImpl<unsigned>::iterator
7058 CSI = CallSiteIdxs.begin(), CSE = CallSiteIdxs.end();
7059 CSI != CSE; ++CSI) {
7060 CallSiteNumToLPad[*CSI].push_back(&*BB);
7061 MaxCSNum = std::max(MaxCSNum, *CSI);
7067 // Get an ordered list of the machine basic blocks for the jump table.
7068 std::vector<MachineBasicBlock*> LPadList;
7069 SmallPtrSet<MachineBasicBlock*, 64> InvokeBBs;
7070 LPadList.reserve(CallSiteNumToLPad.size());
7071 for (unsigned I = 1; I <= MaxCSNum; ++I) {
7072 SmallVectorImpl<MachineBasicBlock*> &MBBList = CallSiteNumToLPad[I];
7073 for (SmallVectorImpl<MachineBasicBlock*>::iterator
7074 II = MBBList.begin(), IE = MBBList.end(); II != IE; ++II) {
7075 LPadList.push_back(*II);
7076 InvokeBBs.insert((*II)->pred_begin(), (*II)->pred_end());
7080 assert(!LPadList.empty() &&
7081 "No landing pad destinations for the dispatch jump table!");
7083 // Create the jump table and associated information.
7084 MachineJumpTableInfo *JTI =
7085 MF->getOrCreateJumpTableInfo(MachineJumpTableInfo::EK_Inline);
7086 unsigned MJTI = JTI->createJumpTableIndex(LPadList);
7087 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
7089 // Create the MBBs for the dispatch code.
7091 // Shove the dispatch's address into the return slot in the function context.
7092 MachineBasicBlock *DispatchBB = MF->CreateMachineBasicBlock();
7093 DispatchBB->setIsEHPad();
7095 MachineBasicBlock *TrapBB = MF->CreateMachineBasicBlock();
7096 unsigned trap_opcode;
7097 if (Subtarget->isThumb())
7098 trap_opcode = ARM::tTRAP;
7100 trap_opcode = Subtarget->useNaClTrap() ? ARM::TRAPNaCl : ARM::TRAP;
7102 BuildMI(TrapBB, dl, TII->get(trap_opcode));
7103 DispatchBB->addSuccessor(TrapBB);
7105 MachineBasicBlock *DispContBB = MF->CreateMachineBasicBlock();
7106 DispatchBB->addSuccessor(DispContBB);
7109 MF->insert(MF->end(), DispatchBB);
7110 MF->insert(MF->end(), DispContBB);
7111 MF->insert(MF->end(), TrapBB);
7113 // Insert code into the entry block that creates and registers the function
7115 SetupEntryBlockForSjLj(MI, MBB, DispatchBB, FI);
7117 MachineMemOperand *FIMMOLd = MF->getMachineMemOperand(
7118 MachinePointerInfo::getFixedStack(*MF, FI),
7119 MachineMemOperand::MOLoad | MachineMemOperand::MOVolatile, 4, 4);
7121 MachineInstrBuilder MIB;
7122 MIB = BuildMI(DispatchBB, dl, TII->get(ARM::Int_eh_sjlj_dispatchsetup));
7124 const ARMBaseInstrInfo *AII = static_cast<const ARMBaseInstrInfo*>(TII);
7125 const ARMBaseRegisterInfo &RI = AII->getRegisterInfo();
7127 // Add a register mask with no preserved registers. This results in all
7128 // registers being marked as clobbered.
7129 MIB.addRegMask(RI.getNoPreservedMask());
7131 unsigned NumLPads = LPadList.size();
7132 if (Subtarget->isThumb2()) {
7133 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
7134 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2LDRi12), NewVReg1)
7137 .addMemOperand(FIMMOLd));
7139 if (NumLPads < 256) {
7140 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPri))
7142 .addImm(LPadList.size()));
7144 unsigned VReg1 = MRI->createVirtualRegister(TRC);
7145 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVi16), VReg1)
7146 .addImm(NumLPads & 0xFFFF));
7148 unsigned VReg2 = VReg1;
7149 if ((NumLPads & 0xFFFF0000) != 0) {
7150 VReg2 = MRI->createVirtualRegister(TRC);
7151 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVTi16), VReg2)
7153 .addImm(NumLPads >> 16));
7156 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPrr))
7161 BuildMI(DispatchBB, dl, TII->get(ARM::t2Bcc))
7166 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
7167 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::t2LEApcrelJT),NewVReg3)
7168 .addJumpTableIndex(MJTI));
7170 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
7173 BuildMI(DispContBB, dl, TII->get(ARM::t2ADDrs), NewVReg4)
7174 .addReg(NewVReg3, RegState::Kill)
7176 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
7178 BuildMI(DispContBB, dl, TII->get(ARM::t2BR_JT))
7179 .addReg(NewVReg4, RegState::Kill)
7181 .addJumpTableIndex(MJTI);
7182 } else if (Subtarget->isThumb()) {
7183 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
7184 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRspi), NewVReg1)
7187 .addMemOperand(FIMMOLd));
7189 if (NumLPads < 256) {
7190 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPi8))
7194 MachineConstantPool *ConstantPool = MF->getConstantPool();
7195 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
7196 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
7198 // MachineConstantPool wants an explicit alignment.
7199 unsigned Align = MF->getDataLayout().getPrefTypeAlignment(Int32Ty);
7201 Align = MF->getDataLayout().getTypeAllocSize(C->getType());
7202 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
7204 unsigned VReg1 = MRI->createVirtualRegister(TRC);
7205 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRpci))
7206 .addReg(VReg1, RegState::Define)
7207 .addConstantPoolIndex(Idx));
7208 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPr))
7213 BuildMI(DispatchBB, dl, TII->get(ARM::tBcc))
7218 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
7219 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLSLri), NewVReg2)
7220 .addReg(ARM::CPSR, RegState::Define)
7224 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
7225 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLEApcrelJT), NewVReg3)
7226 .addJumpTableIndex(MJTI));
7228 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
7229 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg4)
7230 .addReg(ARM::CPSR, RegState::Define)
7231 .addReg(NewVReg2, RegState::Kill)
7234 MachineMemOperand *JTMMOLd = MF->getMachineMemOperand(
7235 MachinePointerInfo::getJumpTable(*MF), MachineMemOperand::MOLoad, 4, 4);
7237 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
7238 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLDRi), NewVReg5)
7239 .addReg(NewVReg4, RegState::Kill)
7241 .addMemOperand(JTMMOLd));
7243 unsigned NewVReg6 = NewVReg5;
7244 if (RelocM == Reloc::PIC_) {
7245 NewVReg6 = MRI->createVirtualRegister(TRC);
7246 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg6)
7247 .addReg(ARM::CPSR, RegState::Define)
7248 .addReg(NewVReg5, RegState::Kill)
7252 BuildMI(DispContBB, dl, TII->get(ARM::tBR_JTr))
7253 .addReg(NewVReg6, RegState::Kill)
7254 .addJumpTableIndex(MJTI);
7256 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
7257 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRi12), NewVReg1)
7260 .addMemOperand(FIMMOLd));
7262 if (NumLPads < 256) {
7263 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPri))
7266 } else if (Subtarget->hasV6T2Ops() && isUInt<16>(NumLPads)) {
7267 unsigned VReg1 = MRI->createVirtualRegister(TRC);
7268 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVi16), VReg1)
7269 .addImm(NumLPads & 0xFFFF));
7271 unsigned VReg2 = VReg1;
7272 if ((NumLPads & 0xFFFF0000) != 0) {
7273 VReg2 = MRI->createVirtualRegister(TRC);
7274 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVTi16), VReg2)
7276 .addImm(NumLPads >> 16));
7279 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
7283 MachineConstantPool *ConstantPool = MF->getConstantPool();
7284 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
7285 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
7287 // MachineConstantPool wants an explicit alignment.
7288 unsigned Align = MF->getDataLayout().getPrefTypeAlignment(Int32Ty);
7290 Align = MF->getDataLayout().getTypeAllocSize(C->getType());
7291 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
7293 unsigned VReg1 = MRI->createVirtualRegister(TRC);
7294 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRcp))
7295 .addReg(VReg1, RegState::Define)
7296 .addConstantPoolIndex(Idx)
7298 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
7300 .addReg(VReg1, RegState::Kill));
7303 BuildMI(DispatchBB, dl, TII->get(ARM::Bcc))
7308 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
7310 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::MOVsi), NewVReg3)
7312 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
7313 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
7314 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::LEApcrelJT), NewVReg4)
7315 .addJumpTableIndex(MJTI));
7317 MachineMemOperand *JTMMOLd = MF->getMachineMemOperand(
7318 MachinePointerInfo::getJumpTable(*MF), MachineMemOperand::MOLoad, 4, 4);
7319 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
7321 BuildMI(DispContBB, dl, TII->get(ARM::LDRrs), NewVReg5)
7322 .addReg(NewVReg3, RegState::Kill)
7325 .addMemOperand(JTMMOLd));
7327 if (RelocM == Reloc::PIC_) {
7328 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTadd))
7329 .addReg(NewVReg5, RegState::Kill)
7331 .addJumpTableIndex(MJTI);
7333 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTr))
7334 .addReg(NewVReg5, RegState::Kill)
7335 .addJumpTableIndex(MJTI);
7339 // Add the jump table entries as successors to the MBB.
7340 SmallPtrSet<MachineBasicBlock*, 8> SeenMBBs;
7341 for (std::vector<MachineBasicBlock*>::iterator
7342 I = LPadList.begin(), E = LPadList.end(); I != E; ++I) {
7343 MachineBasicBlock *CurMBB = *I;
7344 if (SeenMBBs.insert(CurMBB).second)
7345 DispContBB->addSuccessor(CurMBB);
7348 // N.B. the order the invoke BBs are processed in doesn't matter here.
7349 const MCPhysReg *SavedRegs = RI.getCalleeSavedRegs(MF);
7350 SmallVector<MachineBasicBlock*, 64> MBBLPads;
7351 for (MachineBasicBlock *BB : InvokeBBs) {
7353 // Remove the landing pad successor from the invoke block and replace it
7354 // with the new dispatch block.
7355 SmallVector<MachineBasicBlock*, 4> Successors(BB->succ_begin(),
7357 while (!Successors.empty()) {
7358 MachineBasicBlock *SMBB = Successors.pop_back_val();
7359 if (SMBB->isEHPad()) {
7360 BB->removeSuccessor(SMBB);
7361 MBBLPads.push_back(SMBB);
7365 BB->addSuccessor(DispatchBB);
7367 // Find the invoke call and mark all of the callee-saved registers as
7368 // 'implicit defined' so that they're spilled. This prevents code from
7369 // moving instructions to before the EH block, where they will never be
7371 for (MachineBasicBlock::reverse_iterator
7372 II = BB->rbegin(), IE = BB->rend(); II != IE; ++II) {
7373 if (!II->isCall()) continue;
7375 DenseMap<unsigned, bool> DefRegs;
7376 for (MachineInstr::mop_iterator
7377 OI = II->operands_begin(), OE = II->operands_end();
7379 if (!OI->isReg()) continue;
7380 DefRegs[OI->getReg()] = true;
7383 MachineInstrBuilder MIB(*MF, &*II);
7385 for (unsigned i = 0; SavedRegs[i] != 0; ++i) {
7386 unsigned Reg = SavedRegs[i];
7387 if (Subtarget->isThumb2() &&
7388 !ARM::tGPRRegClass.contains(Reg) &&
7389 !ARM::hGPRRegClass.contains(Reg))
7391 if (Subtarget->isThumb1Only() && !ARM::tGPRRegClass.contains(Reg))
7393 if (!Subtarget->isThumb() && !ARM::GPRRegClass.contains(Reg))
7396 MIB.addReg(Reg, RegState::ImplicitDefine | RegState::Dead);
7403 // Mark all former landing pads as non-landing pads. The dispatch is the only
7405 for (SmallVectorImpl<MachineBasicBlock*>::iterator
7406 I = MBBLPads.begin(), E = MBBLPads.end(); I != E; ++I)
7407 (*I)->setIsEHPad(false);
7409 // The instruction is gone now.
7410 MI->eraseFromParent();
7414 MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) {
7415 for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
7416 E = MBB->succ_end(); I != E; ++I)
7419 llvm_unreachable("Expecting a BB with two successors!");
7422 /// Return the load opcode for a given load size. If load size >= 8,
7423 /// neon opcode will be returned.
7424 static unsigned getLdOpcode(unsigned LdSize, bool IsThumb1, bool IsThumb2) {
7426 return LdSize == 16 ? ARM::VLD1q32wb_fixed
7427 : LdSize == 8 ? ARM::VLD1d32wb_fixed : 0;
7429 return LdSize == 4 ? ARM::tLDRi
7430 : LdSize == 2 ? ARM::tLDRHi
7431 : LdSize == 1 ? ARM::tLDRBi : 0;
7433 return LdSize == 4 ? ARM::t2LDR_POST
7434 : LdSize == 2 ? ARM::t2LDRH_POST
7435 : LdSize == 1 ? ARM::t2LDRB_POST : 0;
7436 return LdSize == 4 ? ARM::LDR_POST_IMM
7437 : LdSize == 2 ? ARM::LDRH_POST
7438 : LdSize == 1 ? ARM::LDRB_POST_IMM : 0;
7441 /// Return the store opcode for a given store size. If store size >= 8,
7442 /// neon opcode will be returned.
7443 static unsigned getStOpcode(unsigned StSize, bool IsThumb1, bool IsThumb2) {
7445 return StSize == 16 ? ARM::VST1q32wb_fixed
7446 : StSize == 8 ? ARM::VST1d32wb_fixed : 0;
7448 return StSize == 4 ? ARM::tSTRi
7449 : StSize == 2 ? ARM::tSTRHi
7450 : StSize == 1 ? ARM::tSTRBi : 0;
7452 return StSize == 4 ? ARM::t2STR_POST
7453 : StSize == 2 ? ARM::t2STRH_POST
7454 : StSize == 1 ? ARM::t2STRB_POST : 0;
7455 return StSize == 4 ? ARM::STR_POST_IMM
7456 : StSize == 2 ? ARM::STRH_POST
7457 : StSize == 1 ? ARM::STRB_POST_IMM : 0;
7460 /// Emit a post-increment load operation with given size. The instructions
7461 /// will be added to BB at Pos.
7462 static void emitPostLd(MachineBasicBlock *BB, MachineInstr *Pos,
7463 const TargetInstrInfo *TII, DebugLoc dl,
7464 unsigned LdSize, unsigned Data, unsigned AddrIn,
7465 unsigned AddrOut, bool IsThumb1, bool IsThumb2) {
7466 unsigned LdOpc = getLdOpcode(LdSize, IsThumb1, IsThumb2);
7467 assert(LdOpc != 0 && "Should have a load opcode");
7469 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
7470 .addReg(AddrOut, RegState::Define).addReg(AddrIn)
7472 } else if (IsThumb1) {
7473 // load + update AddrIn
7474 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
7475 .addReg(AddrIn).addImm(0));
7476 MachineInstrBuilder MIB =
7477 BuildMI(*BB, Pos, dl, TII->get(ARM::tADDi8), AddrOut);
7478 MIB = AddDefaultT1CC(MIB);
7479 MIB.addReg(AddrIn).addImm(LdSize);
7480 AddDefaultPred(MIB);
7481 } else if (IsThumb2) {
7482 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
7483 .addReg(AddrOut, RegState::Define).addReg(AddrIn)
7486 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
7487 .addReg(AddrOut, RegState::Define).addReg(AddrIn)
7488 .addReg(0).addImm(LdSize));
7492 /// Emit a post-increment store operation with given size. The instructions
7493 /// will be added to BB at Pos.
7494 static void emitPostSt(MachineBasicBlock *BB, MachineInstr *Pos,
7495 const TargetInstrInfo *TII, DebugLoc dl,
7496 unsigned StSize, unsigned Data, unsigned AddrIn,
7497 unsigned AddrOut, bool IsThumb1, bool IsThumb2) {
7498 unsigned StOpc = getStOpcode(StSize, IsThumb1, IsThumb2);
7499 assert(StOpc != 0 && "Should have a store opcode");
7501 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc), AddrOut)
7502 .addReg(AddrIn).addImm(0).addReg(Data));
7503 } else if (IsThumb1) {
7504 // store + update AddrIn
7505 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc)).addReg(Data)
7506 .addReg(AddrIn).addImm(0));
7507 MachineInstrBuilder MIB =
7508 BuildMI(*BB, Pos, dl, TII->get(ARM::tADDi8), AddrOut);
7509 MIB = AddDefaultT1CC(MIB);
7510 MIB.addReg(AddrIn).addImm(StSize);
7511 AddDefaultPred(MIB);
7512 } else if (IsThumb2) {
7513 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc), AddrOut)
7514 .addReg(Data).addReg(AddrIn).addImm(StSize));
7516 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc), AddrOut)
7517 .addReg(Data).addReg(AddrIn).addReg(0)
7523 ARMTargetLowering::EmitStructByval(MachineInstr *MI,
7524 MachineBasicBlock *BB) const {
7525 // This pseudo instruction has 3 operands: dst, src, size
7526 // We expand it to a loop if size > Subtarget->getMaxInlineSizeThreshold().
7527 // Otherwise, we will generate unrolled scalar copies.
7528 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
7529 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7530 MachineFunction::iterator It = ++BB->getIterator();
7532 unsigned dest = MI->getOperand(0).getReg();
7533 unsigned src = MI->getOperand(1).getReg();
7534 unsigned SizeVal = MI->getOperand(2).getImm();
7535 unsigned Align = MI->getOperand(3).getImm();
7536 DebugLoc dl = MI->getDebugLoc();
7538 MachineFunction *MF = BB->getParent();
7539 MachineRegisterInfo &MRI = MF->getRegInfo();
7540 unsigned UnitSize = 0;
7541 const TargetRegisterClass *TRC = nullptr;
7542 const TargetRegisterClass *VecTRC = nullptr;
7544 bool IsThumb1 = Subtarget->isThumb1Only();
7545 bool IsThumb2 = Subtarget->isThumb2();
7549 } else if (Align & 2) {
7552 // Check whether we can use NEON instructions.
7553 if (!MF->getFunction()->hasFnAttribute(Attribute::NoImplicitFloat) &&
7554 Subtarget->hasNEON()) {
7555 if ((Align % 16 == 0) && SizeVal >= 16)
7557 else if ((Align % 8 == 0) && SizeVal >= 8)
7560 // Can't use NEON instructions.
7565 // Select the correct opcode and register class for unit size load/store
7566 bool IsNeon = UnitSize >= 8;
7567 TRC = (IsThumb1 || IsThumb2) ? &ARM::tGPRRegClass : &ARM::GPRRegClass;
7569 VecTRC = UnitSize == 16 ? &ARM::DPairRegClass
7570 : UnitSize == 8 ? &ARM::DPRRegClass
7573 unsigned BytesLeft = SizeVal % UnitSize;
7574 unsigned LoopSize = SizeVal - BytesLeft;
7576 if (SizeVal <= Subtarget->getMaxInlineSizeThreshold()) {
7577 // Use LDR and STR to copy.
7578 // [scratch, srcOut] = LDR_POST(srcIn, UnitSize)
7579 // [destOut] = STR_POST(scratch, destIn, UnitSize)
7580 unsigned srcIn = src;
7581 unsigned destIn = dest;
7582 for (unsigned i = 0; i < LoopSize; i+=UnitSize) {
7583 unsigned srcOut = MRI.createVirtualRegister(TRC);
7584 unsigned destOut = MRI.createVirtualRegister(TRC);
7585 unsigned scratch = MRI.createVirtualRegister(IsNeon ? VecTRC : TRC);
7586 emitPostLd(BB, MI, TII, dl, UnitSize, scratch, srcIn, srcOut,
7587 IsThumb1, IsThumb2);
7588 emitPostSt(BB, MI, TII, dl, UnitSize, scratch, destIn, destOut,
7589 IsThumb1, IsThumb2);
7594 // Handle the leftover bytes with LDRB and STRB.
7595 // [scratch, srcOut] = LDRB_POST(srcIn, 1)
7596 // [destOut] = STRB_POST(scratch, destIn, 1)
7597 for (unsigned i = 0; i < BytesLeft; i++) {
7598 unsigned srcOut = MRI.createVirtualRegister(TRC);
7599 unsigned destOut = MRI.createVirtualRegister(TRC);
7600 unsigned scratch = MRI.createVirtualRegister(TRC);
7601 emitPostLd(BB, MI, TII, dl, 1, scratch, srcIn, srcOut,
7602 IsThumb1, IsThumb2);
7603 emitPostSt(BB, MI, TII, dl, 1, scratch, destIn, destOut,
7604 IsThumb1, IsThumb2);
7608 MI->eraseFromParent(); // The instruction is gone now.
7612 // Expand the pseudo op to a loop.
7615 // movw varEnd, # --> with thumb2
7617 // ldrcp varEnd, idx --> without thumb2
7618 // fallthrough --> loopMBB
7620 // PHI varPhi, varEnd, varLoop
7621 // PHI srcPhi, src, srcLoop
7622 // PHI destPhi, dst, destLoop
7623 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
7624 // [destLoop] = STR_POST(scratch, destPhi, UnitSize)
7625 // subs varLoop, varPhi, #UnitSize
7627 // fallthrough --> exitMBB
7629 // epilogue to handle left-over bytes
7630 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
7631 // [destOut] = STRB_POST(scratch, destLoop, 1)
7632 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
7633 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
7634 MF->insert(It, loopMBB);
7635 MF->insert(It, exitMBB);
7637 // Transfer the remainder of BB and its successor edges to exitMBB.
7638 exitMBB->splice(exitMBB->begin(), BB,
7639 std::next(MachineBasicBlock::iterator(MI)), BB->end());
7640 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
7642 // Load an immediate to varEnd.
7643 unsigned varEnd = MRI.createVirtualRegister(TRC);
7644 if (Subtarget->useMovt(*MF)) {
7645 unsigned Vtmp = varEnd;
7646 if ((LoopSize & 0xFFFF0000) != 0)
7647 Vtmp = MRI.createVirtualRegister(TRC);
7648 AddDefaultPred(BuildMI(BB, dl,
7649 TII->get(IsThumb2 ? ARM::t2MOVi16 : ARM::MOVi16),
7650 Vtmp).addImm(LoopSize & 0xFFFF));
7652 if ((LoopSize & 0xFFFF0000) != 0)
7653 AddDefaultPred(BuildMI(BB, dl,
7654 TII->get(IsThumb2 ? ARM::t2MOVTi16 : ARM::MOVTi16),
7657 .addImm(LoopSize >> 16));
7659 MachineConstantPool *ConstantPool = MF->getConstantPool();
7660 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
7661 const Constant *C = ConstantInt::get(Int32Ty, LoopSize);
7663 // MachineConstantPool wants an explicit alignment.
7664 unsigned Align = MF->getDataLayout().getPrefTypeAlignment(Int32Ty);
7666 Align = MF->getDataLayout().getTypeAllocSize(C->getType());
7667 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
7670 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(ARM::tLDRpci)).addReg(
7671 varEnd, RegState::Define).addConstantPoolIndex(Idx));
7673 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(ARM::LDRcp)).addReg(
7674 varEnd, RegState::Define).addConstantPoolIndex(Idx).addImm(0));
7676 BB->addSuccessor(loopMBB);
7678 // Generate the loop body:
7679 // varPhi = PHI(varLoop, varEnd)
7680 // srcPhi = PHI(srcLoop, src)
7681 // destPhi = PHI(destLoop, dst)
7682 MachineBasicBlock *entryBB = BB;
7684 unsigned varLoop = MRI.createVirtualRegister(TRC);
7685 unsigned varPhi = MRI.createVirtualRegister(TRC);
7686 unsigned srcLoop = MRI.createVirtualRegister(TRC);
7687 unsigned srcPhi = MRI.createVirtualRegister(TRC);
7688 unsigned destLoop = MRI.createVirtualRegister(TRC);
7689 unsigned destPhi = MRI.createVirtualRegister(TRC);
7691 BuildMI(*BB, BB->begin(), dl, TII->get(ARM::PHI), varPhi)
7692 .addReg(varLoop).addMBB(loopMBB)
7693 .addReg(varEnd).addMBB(entryBB);
7694 BuildMI(BB, dl, TII->get(ARM::PHI), srcPhi)
7695 .addReg(srcLoop).addMBB(loopMBB)
7696 .addReg(src).addMBB(entryBB);
7697 BuildMI(BB, dl, TII->get(ARM::PHI), destPhi)
7698 .addReg(destLoop).addMBB(loopMBB)
7699 .addReg(dest).addMBB(entryBB);
7701 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
7702 // [destLoop] = STR_POST(scratch, destPhi, UnitSiz)
7703 unsigned scratch = MRI.createVirtualRegister(IsNeon ? VecTRC : TRC);
7704 emitPostLd(BB, BB->end(), TII, dl, UnitSize, scratch, srcPhi, srcLoop,
7705 IsThumb1, IsThumb2);
7706 emitPostSt(BB, BB->end(), TII, dl, UnitSize, scratch, destPhi, destLoop,
7707 IsThumb1, IsThumb2);
7709 // Decrement loop variable by UnitSize.
7711 MachineInstrBuilder MIB =
7712 BuildMI(*BB, BB->end(), dl, TII->get(ARM::tSUBi8), varLoop);
7713 MIB = AddDefaultT1CC(MIB);
7714 MIB.addReg(varPhi).addImm(UnitSize);
7715 AddDefaultPred(MIB);
7717 MachineInstrBuilder MIB =
7718 BuildMI(*BB, BB->end(), dl,
7719 TII->get(IsThumb2 ? ARM::t2SUBri : ARM::SUBri), varLoop);
7720 AddDefaultCC(AddDefaultPred(MIB.addReg(varPhi).addImm(UnitSize)));
7721 MIB->getOperand(5).setReg(ARM::CPSR);
7722 MIB->getOperand(5).setIsDef(true);
7724 BuildMI(*BB, BB->end(), dl,
7725 TII->get(IsThumb1 ? ARM::tBcc : IsThumb2 ? ARM::t2Bcc : ARM::Bcc))
7726 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
7728 // loopMBB can loop back to loopMBB or fall through to exitMBB.
7729 BB->addSuccessor(loopMBB);
7730 BB->addSuccessor(exitMBB);
7732 // Add epilogue to handle BytesLeft.
7734 MachineInstr *StartOfExit = exitMBB->begin();
7736 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
7737 // [destOut] = STRB_POST(scratch, destLoop, 1)
7738 unsigned srcIn = srcLoop;
7739 unsigned destIn = destLoop;
7740 for (unsigned i = 0; i < BytesLeft; i++) {
7741 unsigned srcOut = MRI.createVirtualRegister(TRC);
7742 unsigned destOut = MRI.createVirtualRegister(TRC);
7743 unsigned scratch = MRI.createVirtualRegister(TRC);
7744 emitPostLd(BB, StartOfExit, TII, dl, 1, scratch, srcIn, srcOut,
7745 IsThumb1, IsThumb2);
7746 emitPostSt(BB, StartOfExit, TII, dl, 1, scratch, destIn, destOut,
7747 IsThumb1, IsThumb2);
7752 MI->eraseFromParent(); // The instruction is gone now.
7757 ARMTargetLowering::EmitLowered__chkstk(MachineInstr *MI,
7758 MachineBasicBlock *MBB) const {
7759 const TargetMachine &TM = getTargetMachine();
7760 const TargetInstrInfo &TII = *Subtarget->getInstrInfo();
7761 DebugLoc DL = MI->getDebugLoc();
7763 assert(Subtarget->isTargetWindows() &&
7764 "__chkstk is only supported on Windows");
7765 assert(Subtarget->isThumb2() && "Windows on ARM requires Thumb-2 mode");
7767 // __chkstk takes the number of words to allocate on the stack in R4, and
7768 // returns the stack adjustment in number of bytes in R4. This will not
7769 // clober any other registers (other than the obvious lr).
7771 // Although, technically, IP should be considered a register which may be
7772 // clobbered, the call itself will not touch it. Windows on ARM is a pure
7773 // thumb-2 environment, so there is no interworking required. As a result, we
7774 // do not expect a veneer to be emitted by the linker, clobbering IP.
7776 // Each module receives its own copy of __chkstk, so no import thunk is
7777 // required, again, ensuring that IP is not clobbered.
7779 // Finally, although some linkers may theoretically provide a trampoline for
7780 // out of range calls (which is quite common due to a 32M range limitation of
7781 // branches for Thumb), we can generate the long-call version via
7782 // -mcmodel=large, alleviating the need for the trampoline which may clobber
7785 switch (TM.getCodeModel()) {
7786 case CodeModel::Small:
7787 case CodeModel::Medium:
7788 case CodeModel::Default:
7789 case CodeModel::Kernel:
7790 BuildMI(*MBB, MI, DL, TII.get(ARM::tBL))
7791 .addImm((unsigned)ARMCC::AL).addReg(0)
7792 .addExternalSymbol("__chkstk")
7793 .addReg(ARM::R4, RegState::Implicit | RegState::Kill)
7794 .addReg(ARM::R4, RegState::Implicit | RegState::Define)
7795 .addReg(ARM::R12, RegState::Implicit | RegState::Define | RegState::Dead);
7797 case CodeModel::Large:
7798 case CodeModel::JITDefault: {
7799 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
7800 unsigned Reg = MRI.createVirtualRegister(&ARM::rGPRRegClass);
7802 BuildMI(*MBB, MI, DL, TII.get(ARM::t2MOVi32imm), Reg)
7803 .addExternalSymbol("__chkstk");
7804 BuildMI(*MBB, MI, DL, TII.get(ARM::tBLXr))
7805 .addImm((unsigned)ARMCC::AL).addReg(0)
7806 .addReg(Reg, RegState::Kill)
7807 .addReg(ARM::R4, RegState::Implicit | RegState::Kill)
7808 .addReg(ARM::R4, RegState::Implicit | RegState::Define)
7809 .addReg(ARM::R12, RegState::Implicit | RegState::Define | RegState::Dead);
7814 AddDefaultCC(AddDefaultPred(BuildMI(*MBB, MI, DL, TII.get(ARM::t2SUBrr),
7816 .addReg(ARM::SP).addReg(ARM::R4)));
7818 MI->eraseFromParent();
7823 ARMTargetLowering::EmitLowered__dbzchk(MachineInstr *MI,
7824 MachineBasicBlock *MBB) const {
7825 DebugLoc DL = MI->getDebugLoc();
7826 MachineFunction *MF = MBB->getParent();
7827 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
7829 MachineBasicBlock *ContBB = MF->CreateMachineBasicBlock();
7830 MF->push_back(ContBB);
7831 ContBB->splice(ContBB->begin(), MBB,
7832 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
7833 MBB->addSuccessor(ContBB);
7835 MachineBasicBlock *TrapBB = MF->CreateMachineBasicBlock();
7836 MF->push_back(TrapBB);
7837 BuildMI(TrapBB, DL, TII->get(ARM::t2UDF)).addImm(249);
7838 MBB->addSuccessor(TrapBB);
7840 BuildMI(*MBB, MI, DL, TII->get(ARM::tCBZ))
7841 .addReg(MI->getOperand(0).getReg())
7844 MI->eraseFromParent();
7849 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
7850 MachineBasicBlock *BB) const {
7851 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
7852 DebugLoc dl = MI->getDebugLoc();
7853 bool isThumb2 = Subtarget->isThumb2();
7854 switch (MI->getOpcode()) {
7857 llvm_unreachable("Unexpected instr type to insert");
7859 // The Thumb2 pre-indexed stores have the same MI operands, they just
7860 // define them differently in the .td files from the isel patterns, so
7861 // they need pseudos.
7862 case ARM::t2STR_preidx:
7863 MI->setDesc(TII->get(ARM::t2STR_PRE));
7865 case ARM::t2STRB_preidx:
7866 MI->setDesc(TII->get(ARM::t2STRB_PRE));
7868 case ARM::t2STRH_preidx:
7869 MI->setDesc(TII->get(ARM::t2STRH_PRE));
7872 case ARM::STRi_preidx:
7873 case ARM::STRBi_preidx: {
7874 unsigned NewOpc = MI->getOpcode() == ARM::STRi_preidx ?
7875 ARM::STR_PRE_IMM : ARM::STRB_PRE_IMM;
7876 // Decode the offset.
7877 unsigned Offset = MI->getOperand(4).getImm();
7878 bool isSub = ARM_AM::getAM2Op(Offset) == ARM_AM::sub;
7879 Offset = ARM_AM::getAM2Offset(Offset);
7883 MachineMemOperand *MMO = *MI->memoperands_begin();
7884 BuildMI(*BB, MI, dl, TII->get(NewOpc))
7885 .addOperand(MI->getOperand(0)) // Rn_wb
7886 .addOperand(MI->getOperand(1)) // Rt
7887 .addOperand(MI->getOperand(2)) // Rn
7888 .addImm(Offset) // offset (skip GPR==zero_reg)
7889 .addOperand(MI->getOperand(5)) // pred
7890 .addOperand(MI->getOperand(6))
7891 .addMemOperand(MMO);
7892 MI->eraseFromParent();
7895 case ARM::STRr_preidx:
7896 case ARM::STRBr_preidx:
7897 case ARM::STRH_preidx: {
7899 switch (MI->getOpcode()) {
7900 default: llvm_unreachable("unexpected opcode!");
7901 case ARM::STRr_preidx: NewOpc = ARM::STR_PRE_REG; break;
7902 case ARM::STRBr_preidx: NewOpc = ARM::STRB_PRE_REG; break;
7903 case ARM::STRH_preidx: NewOpc = ARM::STRH_PRE; break;
7905 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(NewOpc));
7906 for (unsigned i = 0; i < MI->getNumOperands(); ++i)
7907 MIB.addOperand(MI->getOperand(i));
7908 MI->eraseFromParent();
7912 case ARM::tMOVCCr_pseudo: {
7913 // To "insert" a SELECT_CC instruction, we actually have to insert the
7914 // diamond control-flow pattern. The incoming instruction knows the
7915 // destination vreg to set, the condition code register to branch on, the
7916 // true/false values to select between, and a branch opcode to use.
7917 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7918 MachineFunction::iterator It = ++BB->getIterator();
7923 // cmpTY ccX, r1, r2
7925 // fallthrough --> copy0MBB
7926 MachineBasicBlock *thisMBB = BB;
7927 MachineFunction *F = BB->getParent();
7928 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
7929 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
7930 F->insert(It, copy0MBB);
7931 F->insert(It, sinkMBB);
7933 // Transfer the remainder of BB and its successor edges to sinkMBB.
7934 sinkMBB->splice(sinkMBB->begin(), BB,
7935 std::next(MachineBasicBlock::iterator(MI)), BB->end());
7936 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
7938 BB->addSuccessor(copy0MBB);
7939 BB->addSuccessor(sinkMBB);
7941 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
7942 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
7945 // %FalseValue = ...
7946 // # fallthrough to sinkMBB
7949 // Update machine-CFG edges
7950 BB->addSuccessor(sinkMBB);
7953 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
7956 BuildMI(*BB, BB->begin(), dl,
7957 TII->get(ARM::PHI), MI->getOperand(0).getReg())
7958 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
7959 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
7961 MI->eraseFromParent(); // The pseudo instruction is gone now.
7966 case ARM::BCCZi64: {
7967 // If there is an unconditional branch to the other successor, remove it.
7968 BB->erase(std::next(MachineBasicBlock::iterator(MI)), BB->end());
7970 // Compare both parts that make up the double comparison separately for
7972 bool RHSisZero = MI->getOpcode() == ARM::BCCZi64;
7974 unsigned LHS1 = MI->getOperand(1).getReg();
7975 unsigned LHS2 = MI->getOperand(2).getReg();
7977 AddDefaultPred(BuildMI(BB, dl,
7978 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7979 .addReg(LHS1).addImm(0));
7980 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7981 .addReg(LHS2).addImm(0)
7982 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
7984 unsigned RHS1 = MI->getOperand(3).getReg();
7985 unsigned RHS2 = MI->getOperand(4).getReg();
7986 AddDefaultPred(BuildMI(BB, dl,
7987 TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
7988 .addReg(LHS1).addReg(RHS1));
7989 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
7990 .addReg(LHS2).addReg(RHS2)
7991 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
7994 MachineBasicBlock *destMBB = MI->getOperand(RHSisZero ? 3 : 5).getMBB();
7995 MachineBasicBlock *exitMBB = OtherSucc(BB, destMBB);
7996 if (MI->getOperand(0).getImm() == ARMCC::NE)
7997 std::swap(destMBB, exitMBB);
7999 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
8000 .addMBB(destMBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
8002 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2B)).addMBB(exitMBB));
8004 BuildMI(BB, dl, TII->get(ARM::B)) .addMBB(exitMBB);
8006 MI->eraseFromParent(); // The pseudo instruction is gone now.
8010 case ARM::Int_eh_sjlj_setjmp:
8011 case ARM::Int_eh_sjlj_setjmp_nofp:
8012 case ARM::tInt_eh_sjlj_setjmp:
8013 case ARM::t2Int_eh_sjlj_setjmp:
8014 case ARM::t2Int_eh_sjlj_setjmp_nofp:
8017 case ARM::Int_eh_sjlj_setup_dispatch:
8018 EmitSjLjDispatchBlock(MI, BB);
8023 // To insert an ABS instruction, we have to insert the
8024 // diamond control-flow pattern. The incoming instruction knows the
8025 // source vreg to test against 0, the destination vreg to set,
8026 // the condition code register to branch on, the
8027 // true/false values to select between, and a branch opcode to use.
8032 // BCC (branch to SinkBB if V0 >= 0)
8033 // RSBBB: V3 = RSBri V2, 0 (compute ABS if V2 < 0)
8034 // SinkBB: V1 = PHI(V2, V3)
8035 const BasicBlock *LLVM_BB = BB->getBasicBlock();
8036 MachineFunction::iterator BBI = ++BB->getIterator();
8037 MachineFunction *Fn = BB->getParent();
8038 MachineBasicBlock *RSBBB = Fn->CreateMachineBasicBlock(LLVM_BB);
8039 MachineBasicBlock *SinkBB = Fn->CreateMachineBasicBlock(LLVM_BB);
8040 Fn->insert(BBI, RSBBB);
8041 Fn->insert(BBI, SinkBB);
8043 unsigned int ABSSrcReg = MI->getOperand(1).getReg();
8044 unsigned int ABSDstReg = MI->getOperand(0).getReg();
8045 bool ABSSrcKIll = MI->getOperand(1).isKill();
8046 bool isThumb2 = Subtarget->isThumb2();
8047 MachineRegisterInfo &MRI = Fn->getRegInfo();
8048 // In Thumb mode S must not be specified if source register is the SP or
8049 // PC and if destination register is the SP, so restrict register class
8050 unsigned NewRsbDstReg =
8051 MRI.createVirtualRegister(isThumb2 ? &ARM::rGPRRegClass : &ARM::GPRRegClass);
8053 // Transfer the remainder of BB and its successor edges to sinkMBB.
8054 SinkBB->splice(SinkBB->begin(), BB,
8055 std::next(MachineBasicBlock::iterator(MI)), BB->end());
8056 SinkBB->transferSuccessorsAndUpdatePHIs(BB);
8058 BB->addSuccessor(RSBBB);
8059 BB->addSuccessor(SinkBB);
8061 // fall through to SinkMBB
8062 RSBBB->addSuccessor(SinkBB);
8064 // insert a cmp at the end of BB
8065 AddDefaultPred(BuildMI(BB, dl,
8066 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
8067 .addReg(ABSSrcReg).addImm(0));
8069 // insert a bcc with opposite CC to ARMCC::MI at the end of BB
8071 TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)).addMBB(SinkBB)
8072 .addImm(ARMCC::getOppositeCondition(ARMCC::MI)).addReg(ARM::CPSR);
8074 // insert rsbri in RSBBB
8075 // Note: BCC and rsbri will be converted into predicated rsbmi
8076 // by if-conversion pass
8077 BuildMI(*RSBBB, RSBBB->begin(), dl,
8078 TII->get(isThumb2 ? ARM::t2RSBri : ARM::RSBri), NewRsbDstReg)
8079 .addReg(ABSSrcReg, ABSSrcKIll ? RegState::Kill : 0)
8080 .addImm(0).addImm((unsigned)ARMCC::AL).addReg(0).addReg(0);
8082 // insert PHI in SinkBB,
8083 // reuse ABSDstReg to not change uses of ABS instruction
8084 BuildMI(*SinkBB, SinkBB->begin(), dl,
8085 TII->get(ARM::PHI), ABSDstReg)
8086 .addReg(NewRsbDstReg).addMBB(RSBBB)
8087 .addReg(ABSSrcReg).addMBB(BB);
8089 // remove ABS instruction
8090 MI->eraseFromParent();
8092 // return last added BB
8095 case ARM::COPY_STRUCT_BYVAL_I32:
8097 return EmitStructByval(MI, BB);
8098 case ARM::WIN__CHKSTK:
8099 return EmitLowered__chkstk(MI, BB);
8100 case ARM::WIN__DBZCHK:
8101 return EmitLowered__dbzchk(MI, BB);
8105 /// \brief Attaches vregs to MEMCPY that it will use as scratch registers
8106 /// when it is expanded into LDM/STM. This is done as a post-isel lowering
8107 /// instead of as a custom inserter because we need the use list from the SDNode.
8108 static void attachMEMCPYScratchRegs(const ARMSubtarget *Subtarget,
8109 MachineInstr *MI, const SDNode *Node) {
8110 bool isThumb1 = Subtarget->isThumb1Only();
8112 DebugLoc DL = MI->getDebugLoc();
8113 MachineFunction *MF = MI->getParent()->getParent();
8114 MachineRegisterInfo &MRI = MF->getRegInfo();
8115 MachineInstrBuilder MIB(*MF, MI);
8117 // If the new dst/src is unused mark it as dead.
8118 if (!Node->hasAnyUseOfValue(0)) {
8119 MI->getOperand(0).setIsDead(true);
8121 if (!Node->hasAnyUseOfValue(1)) {
8122 MI->getOperand(1).setIsDead(true);
8125 // The MEMCPY both defines and kills the scratch registers.
8126 for (unsigned I = 0; I != MI->getOperand(4).getImm(); ++I) {
8127 unsigned TmpReg = MRI.createVirtualRegister(isThumb1 ? &ARM::tGPRRegClass
8128 : &ARM::GPRRegClass);
8129 MIB.addReg(TmpReg, RegState::Define|RegState::Dead);
8133 void ARMTargetLowering::AdjustInstrPostInstrSelection(MachineInstr *MI,
8134 SDNode *Node) const {
8135 if (MI->getOpcode() == ARM::MEMCPY) {
8136 attachMEMCPYScratchRegs(Subtarget, MI, Node);
8140 const MCInstrDesc *MCID = &MI->getDesc();
8141 // Adjust potentially 's' setting instructions after isel, i.e. ADC, SBC, RSB,
8142 // RSC. Coming out of isel, they have an implicit CPSR def, but the optional
8143 // operand is still set to noreg. If needed, set the optional operand's
8144 // register to CPSR, and remove the redundant implicit def.
8146 // e.g. ADCS (..., CPSR<imp-def>) -> ADC (... opt:CPSR<def>).
8148 // Rename pseudo opcodes.
8149 unsigned NewOpc = convertAddSubFlagsOpcode(MI->getOpcode());
8151 const ARMBaseInstrInfo *TII = Subtarget->getInstrInfo();
8152 MCID = &TII->get(NewOpc);
8154 assert(MCID->getNumOperands() == MI->getDesc().getNumOperands() + 1 &&
8155 "converted opcode should be the same except for cc_out");
8159 // Add the optional cc_out operand
8160 MI->addOperand(MachineOperand::CreateReg(0, /*isDef=*/true));
8162 unsigned ccOutIdx = MCID->getNumOperands() - 1;
8164 // Any ARM instruction that sets the 's' bit should specify an optional
8165 // "cc_out" operand in the last operand position.
8166 if (!MI->hasOptionalDef() || !MCID->OpInfo[ccOutIdx].isOptionalDef()) {
8167 assert(!NewOpc && "Optional cc_out operand required");
8170 // Look for an implicit def of CPSR added by MachineInstr ctor. Remove it
8171 // since we already have an optional CPSR def.
8172 bool definesCPSR = false;
8173 bool deadCPSR = false;
8174 for (unsigned i = MCID->getNumOperands(), e = MI->getNumOperands();
8176 const MachineOperand &MO = MI->getOperand(i);
8177 if (MO.isReg() && MO.isDef() && MO.getReg() == ARM::CPSR) {
8181 MI->RemoveOperand(i);
8186 assert(!NewOpc && "Optional cc_out operand required");
8189 assert(deadCPSR == !Node->hasAnyUseOfValue(1) && "inconsistent dead flag");
8191 assert(!MI->getOperand(ccOutIdx).getReg() &&
8192 "expect uninitialized optional cc_out operand");
8196 // If this instruction was defined with an optional CPSR def and its dag node
8197 // had a live implicit CPSR def, then activate the optional CPSR def.
8198 MachineOperand &MO = MI->getOperand(ccOutIdx);
8199 MO.setReg(ARM::CPSR);
8203 //===----------------------------------------------------------------------===//
8204 // ARM Optimization Hooks
8205 //===----------------------------------------------------------------------===//
8207 // Helper function that checks if N is a null or all ones constant.
8208 static inline bool isZeroOrAllOnes(SDValue N, bool AllOnes) {
8209 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N);
8212 return AllOnes ? C->isAllOnesValue() : C->isNullValue();
8215 // Return true if N is conditionally 0 or all ones.
8216 // Detects these expressions where cc is an i1 value:
8218 // (select cc 0, y) [AllOnes=0]
8219 // (select cc y, 0) [AllOnes=0]
8220 // (zext cc) [AllOnes=0]
8221 // (sext cc) [AllOnes=0/1]
8222 // (select cc -1, y) [AllOnes=1]
8223 // (select cc y, -1) [AllOnes=1]
8225 // Invert is set when N is the null/all ones constant when CC is false.
8226 // OtherOp is set to the alternative value of N.
8227 static bool isConditionalZeroOrAllOnes(SDNode *N, bool AllOnes,
8228 SDValue &CC, bool &Invert,
8230 SelectionDAG &DAG) {
8231 switch (N->getOpcode()) {
8232 default: return false;
8234 CC = N->getOperand(0);
8235 SDValue N1 = N->getOperand(1);
8236 SDValue N2 = N->getOperand(2);
8237 if (isZeroOrAllOnes(N1, AllOnes)) {
8242 if (isZeroOrAllOnes(N2, AllOnes)) {
8249 case ISD::ZERO_EXTEND:
8250 // (zext cc) can never be the all ones value.
8254 case ISD::SIGN_EXTEND: {
8256 EVT VT = N->getValueType(0);
8257 CC = N->getOperand(0);
8258 if (CC.getValueType() != MVT::i1)
8262 // When looking for an AllOnes constant, N is an sext, and the 'other'
8264 OtherOp = DAG.getConstant(0, dl, VT);
8265 else if (N->getOpcode() == ISD::ZERO_EXTEND)
8266 // When looking for a 0 constant, N can be zext or sext.
8267 OtherOp = DAG.getConstant(1, dl, VT);
8269 OtherOp = DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), dl,
8276 // Combine a constant select operand into its use:
8278 // (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
8279 // (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
8280 // (and (select cc, -1, c), x) -> (select cc, x, (and, x, c)) [AllOnes=1]
8281 // (or (select cc, 0, c), x) -> (select cc, x, (or, x, c))
8282 // (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c))
8284 // The transform is rejected if the select doesn't have a constant operand that
8285 // is null, or all ones when AllOnes is set.
8287 // Also recognize sext/zext from i1:
8289 // (add (zext cc), x) -> (select cc (add x, 1), x)
8290 // (add (sext cc), x) -> (select cc (add x, -1), x)
8292 // These transformations eventually create predicated instructions.
8294 // @param N The node to transform.
8295 // @param Slct The N operand that is a select.
8296 // @param OtherOp The other N operand (x above).
8297 // @param DCI Context.
8298 // @param AllOnes Require the select constant to be all ones instead of null.
8299 // @returns The new node, or SDValue() on failure.
8301 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
8302 TargetLowering::DAGCombinerInfo &DCI,
8303 bool AllOnes = false) {
8304 SelectionDAG &DAG = DCI.DAG;
8305 EVT VT = N->getValueType(0);
8306 SDValue NonConstantVal;
8309 if (!isConditionalZeroOrAllOnes(Slct.getNode(), AllOnes, CCOp, SwapSelectOps,
8310 NonConstantVal, DAG))
8313 // Slct is now know to be the desired identity constant when CC is true.
8314 SDValue TrueVal = OtherOp;
8315 SDValue FalseVal = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
8316 OtherOp, NonConstantVal);
8317 // Unless SwapSelectOps says CC should be false.
8319 std::swap(TrueVal, FalseVal);
8321 return DAG.getNode(ISD::SELECT, SDLoc(N), VT,
8322 CCOp, TrueVal, FalseVal);
8325 // Attempt combineSelectAndUse on each operand of a commutative operator N.
8327 SDValue combineSelectAndUseCommutative(SDNode *N, bool AllOnes,
8328 TargetLowering::DAGCombinerInfo &DCI) {
8329 SDValue N0 = N->getOperand(0);
8330 SDValue N1 = N->getOperand(1);
8331 if (N0.getNode()->hasOneUse()) {
8332 SDValue Result = combineSelectAndUse(N, N0, N1, DCI, AllOnes);
8333 if (Result.getNode())
8336 if (N1.getNode()->hasOneUse()) {
8337 SDValue Result = combineSelectAndUse(N, N1, N0, DCI, AllOnes);
8338 if (Result.getNode())
8344 // AddCombineToVPADDL- For pair-wise add on neon, use the vpaddl instruction
8345 // (only after legalization).
8346 static SDValue AddCombineToVPADDL(SDNode *N, SDValue N0, SDValue N1,
8347 TargetLowering::DAGCombinerInfo &DCI,
8348 const ARMSubtarget *Subtarget) {
8350 // Only perform optimization if after legalize, and if NEON is available. We
8351 // also expected both operands to be BUILD_VECTORs.
8352 if (DCI.isBeforeLegalize() || !Subtarget->hasNEON()
8353 || N0.getOpcode() != ISD::BUILD_VECTOR
8354 || N1.getOpcode() != ISD::BUILD_VECTOR)
8357 // Check output type since VPADDL operand elements can only be 8, 16, or 32.
8358 EVT VT = N->getValueType(0);
8359 if (!VT.isInteger() || VT.getVectorElementType() == MVT::i64)
8362 // Check that the vector operands are of the right form.
8363 // N0 and N1 are BUILD_VECTOR nodes with N number of EXTRACT_VECTOR
8364 // operands, where N is the size of the formed vector.
8365 // Each EXTRACT_VECTOR should have the same input vector and odd or even
8366 // index such that we have a pair wise add pattern.
8368 // Grab the vector that all EXTRACT_VECTOR nodes should be referencing.
8369 if (N0->getOperand(0)->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
8371 SDValue Vec = N0->getOperand(0)->getOperand(0);
8372 SDNode *V = Vec.getNode();
8373 unsigned nextIndex = 0;
8375 // For each operands to the ADD which are BUILD_VECTORs,
8376 // check to see if each of their operands are an EXTRACT_VECTOR with
8377 // the same vector and appropriate index.
8378 for (unsigned i = 0, e = N0->getNumOperands(); i != e; ++i) {
8379 if (N0->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT
8380 && N1->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
8382 SDValue ExtVec0 = N0->getOperand(i);
8383 SDValue ExtVec1 = N1->getOperand(i);
8385 // First operand is the vector, verify its the same.
8386 if (V != ExtVec0->getOperand(0).getNode() ||
8387 V != ExtVec1->getOperand(0).getNode())
8390 // Second is the constant, verify its correct.
8391 ConstantSDNode *C0 = dyn_cast<ConstantSDNode>(ExtVec0->getOperand(1));
8392 ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(ExtVec1->getOperand(1));
8394 // For the constant, we want to see all the even or all the odd.
8395 if (!C0 || !C1 || C0->getZExtValue() != nextIndex
8396 || C1->getZExtValue() != nextIndex+1)
8405 // Create VPADDL node.
8406 SelectionDAG &DAG = DCI.DAG;
8407 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8411 // Build operand list.
8412 SmallVector<SDValue, 8> Ops;
8413 Ops.push_back(DAG.getConstant(Intrinsic::arm_neon_vpaddls, dl,
8414 TLI.getPointerTy(DAG.getDataLayout())));
8416 // Input is the vector.
8419 // Get widened type and narrowed type.
8421 unsigned numElem = VT.getVectorNumElements();
8423 EVT inputLaneType = Vec.getValueType().getVectorElementType();
8424 switch (inputLaneType.getSimpleVT().SimpleTy) {
8425 case MVT::i8: widenType = MVT::getVectorVT(MVT::i16, numElem); break;
8426 case MVT::i16: widenType = MVT::getVectorVT(MVT::i32, numElem); break;
8427 case MVT::i32: widenType = MVT::getVectorVT(MVT::i64, numElem); break;
8429 llvm_unreachable("Invalid vector element type for padd optimization.");
8432 SDValue tmp = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, widenType, Ops);
8433 unsigned ExtOp = VT.bitsGT(tmp.getValueType()) ? ISD::ANY_EXTEND : ISD::TRUNCATE;
8434 return DAG.getNode(ExtOp, dl, VT, tmp);
8437 static SDValue findMUL_LOHI(SDValue V) {
8438 if (V->getOpcode() == ISD::UMUL_LOHI ||
8439 V->getOpcode() == ISD::SMUL_LOHI)
8444 static SDValue AddCombineTo64bitMLAL(SDNode *AddcNode,
8445 TargetLowering::DAGCombinerInfo &DCI,
8446 const ARMSubtarget *Subtarget) {
8448 if (Subtarget->isThumb1Only()) return SDValue();
8450 // Only perform the checks after legalize when the pattern is available.
8451 if (DCI.isBeforeLegalize()) return SDValue();
8453 // Look for multiply add opportunities.
8454 // The pattern is a ISD::UMUL_LOHI followed by two add nodes, where
8455 // each add nodes consumes a value from ISD::UMUL_LOHI and there is
8456 // a glue link from the first add to the second add.
8457 // If we find this pattern, we can replace the U/SMUL_LOHI, ADDC, and ADDE by
8458 // a S/UMLAL instruction.
8461 // / \ [no multiline comment]
8467 assert(AddcNode->getOpcode() == ISD::ADDC && "Expect an ADDC");
8468 SDValue AddcOp0 = AddcNode->getOperand(0);
8469 SDValue AddcOp1 = AddcNode->getOperand(1);
8471 // Check if the two operands are from the same mul_lohi node.
8472 if (AddcOp0.getNode() == AddcOp1.getNode())
8475 assert(AddcNode->getNumValues() == 2 &&
8476 AddcNode->getValueType(0) == MVT::i32 &&
8477 "Expect ADDC with two result values. First: i32");
8479 // Check that we have a glued ADDC node.
8480 if (AddcNode->getValueType(1) != MVT::Glue)
8483 // Check that the ADDC adds the low result of the S/UMUL_LOHI.
8484 if (AddcOp0->getOpcode() != ISD::UMUL_LOHI &&
8485 AddcOp0->getOpcode() != ISD::SMUL_LOHI &&
8486 AddcOp1->getOpcode() != ISD::UMUL_LOHI &&
8487 AddcOp1->getOpcode() != ISD::SMUL_LOHI)
8490 // Look for the glued ADDE.
8491 SDNode* AddeNode = AddcNode->getGluedUser();
8495 // Make sure it is really an ADDE.
8496 if (AddeNode->getOpcode() != ISD::ADDE)
8499 assert(AddeNode->getNumOperands() == 3 &&
8500 AddeNode->getOperand(2).getValueType() == MVT::Glue &&
8501 "ADDE node has the wrong inputs");
8503 // Check for the triangle shape.
8504 SDValue AddeOp0 = AddeNode->getOperand(0);
8505 SDValue AddeOp1 = AddeNode->getOperand(1);
8507 // Make sure that the ADDE operands are not coming from the same node.
8508 if (AddeOp0.getNode() == AddeOp1.getNode())
8511 // Find the MUL_LOHI node walking up ADDE's operands.
8512 bool IsLeftOperandMUL = false;
8513 SDValue MULOp = findMUL_LOHI(AddeOp0);
8514 if (MULOp == SDValue())
8515 MULOp = findMUL_LOHI(AddeOp1);
8517 IsLeftOperandMUL = true;
8518 if (MULOp == SDValue())
8521 // Figure out the right opcode.
8522 unsigned Opc = MULOp->getOpcode();
8523 unsigned FinalOpc = (Opc == ISD::SMUL_LOHI) ? ARMISD::SMLAL : ARMISD::UMLAL;
8525 // Figure out the high and low input values to the MLAL node.
8526 SDValue* HiAdd = nullptr;
8527 SDValue* LoMul = nullptr;
8528 SDValue* LowAdd = nullptr;
8530 // Ensure that ADDE is from high result of ISD::SMUL_LOHI.
8531 if ((AddeOp0 != MULOp.getValue(1)) && (AddeOp1 != MULOp.getValue(1)))
8534 if (IsLeftOperandMUL)
8540 // Ensure that LoMul and LowAdd are taken from correct ISD::SMUL_LOHI node
8541 // whose low result is fed to the ADDC we are checking.
8543 if (AddcOp0 == MULOp.getValue(0)) {
8547 if (AddcOp1 == MULOp.getValue(0)) {
8555 // Create the merged node.
8556 SelectionDAG &DAG = DCI.DAG;
8558 // Build operand list.
8559 SmallVector<SDValue, 8> Ops;
8560 Ops.push_back(LoMul->getOperand(0));
8561 Ops.push_back(LoMul->getOperand(1));
8562 Ops.push_back(*LowAdd);
8563 Ops.push_back(*HiAdd);
8565 SDValue MLALNode = DAG.getNode(FinalOpc, SDLoc(AddcNode),
8566 DAG.getVTList(MVT::i32, MVT::i32), Ops);
8568 // Replace the ADDs' nodes uses by the MLA node's values.
8569 SDValue HiMLALResult(MLALNode.getNode(), 1);
8570 DAG.ReplaceAllUsesOfValueWith(SDValue(AddeNode, 0), HiMLALResult);
8572 SDValue LoMLALResult(MLALNode.getNode(), 0);
8573 DAG.ReplaceAllUsesOfValueWith(SDValue(AddcNode, 0), LoMLALResult);
8575 // Return original node to notify the driver to stop replacing.
8576 SDValue resNode(AddcNode, 0);
8580 /// PerformADDCCombine - Target-specific dag combine transform from
8581 /// ISD::ADDC, ISD::ADDE, and ISD::MUL_LOHI to MLAL.
8582 static SDValue PerformADDCCombine(SDNode *N,
8583 TargetLowering::DAGCombinerInfo &DCI,
8584 const ARMSubtarget *Subtarget) {
8586 return AddCombineTo64bitMLAL(N, DCI, Subtarget);
8590 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
8591 /// operands N0 and N1. This is a helper for PerformADDCombine that is
8592 /// called with the default operands, and if that fails, with commuted
8594 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
8595 TargetLowering::DAGCombinerInfo &DCI,
8596 const ARMSubtarget *Subtarget){
8598 // Attempt to create vpaddl for this add.
8599 SDValue Result = AddCombineToVPADDL(N, N0, N1, DCI, Subtarget);
8600 if (Result.getNode())
8603 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
8604 if (N0.getNode()->hasOneUse()) {
8605 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
8606 if (Result.getNode()) return Result;
8611 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
8613 static SDValue PerformADDCombine(SDNode *N,
8614 TargetLowering::DAGCombinerInfo &DCI,
8615 const ARMSubtarget *Subtarget) {
8616 SDValue N0 = N->getOperand(0);
8617 SDValue N1 = N->getOperand(1);
8619 // First try with the default operand order.
8620 SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI, Subtarget);
8621 if (Result.getNode())
8624 // If that didn't work, try again with the operands commuted.
8625 return PerformADDCombineWithOperands(N, N1, N0, DCI, Subtarget);
8628 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
8630 static SDValue PerformSUBCombine(SDNode *N,
8631 TargetLowering::DAGCombinerInfo &DCI) {
8632 SDValue N0 = N->getOperand(0);
8633 SDValue N1 = N->getOperand(1);
8635 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
8636 if (N1.getNode()->hasOneUse()) {
8637 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
8638 if (Result.getNode()) return Result;
8644 /// PerformVMULCombine
8645 /// Distribute (A + B) * C to (A * C) + (B * C) to take advantage of the
8646 /// special multiplier accumulator forwarding.
8652 // However, for (A + B) * (A + B),
8659 static SDValue PerformVMULCombine(SDNode *N,
8660 TargetLowering::DAGCombinerInfo &DCI,
8661 const ARMSubtarget *Subtarget) {
8662 if (!Subtarget->hasVMLxForwarding())
8665 SelectionDAG &DAG = DCI.DAG;
8666 SDValue N0 = N->getOperand(0);
8667 SDValue N1 = N->getOperand(1);
8668 unsigned Opcode = N0.getOpcode();
8669 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
8670 Opcode != ISD::FADD && Opcode != ISD::FSUB) {
8671 Opcode = N1.getOpcode();
8672 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
8673 Opcode != ISD::FADD && Opcode != ISD::FSUB)
8681 EVT VT = N->getValueType(0);
8683 SDValue N00 = N0->getOperand(0);
8684 SDValue N01 = N0->getOperand(1);
8685 return DAG.getNode(Opcode, DL, VT,
8686 DAG.getNode(ISD::MUL, DL, VT, N00, N1),
8687 DAG.getNode(ISD::MUL, DL, VT, N01, N1));
8690 static SDValue PerformMULCombine(SDNode *N,
8691 TargetLowering::DAGCombinerInfo &DCI,
8692 const ARMSubtarget *Subtarget) {
8693 SelectionDAG &DAG = DCI.DAG;
8695 if (Subtarget->isThumb1Only())
8698 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
8701 EVT VT = N->getValueType(0);
8702 if (VT.is64BitVector() || VT.is128BitVector())
8703 return PerformVMULCombine(N, DCI, Subtarget);
8707 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
8711 int64_t MulAmt = C->getSExtValue();
8712 unsigned ShiftAmt = countTrailingZeros<uint64_t>(MulAmt);
8714 ShiftAmt = ShiftAmt & (32 - 1);
8715 SDValue V = N->getOperand(0);
8719 MulAmt >>= ShiftAmt;
8722 if (isPowerOf2_32(MulAmt - 1)) {
8723 // (mul x, 2^N + 1) => (add (shl x, N), x)
8724 Res = DAG.getNode(ISD::ADD, DL, VT,
8726 DAG.getNode(ISD::SHL, DL, VT,
8728 DAG.getConstant(Log2_32(MulAmt - 1), DL,
8730 } else if (isPowerOf2_32(MulAmt + 1)) {
8731 // (mul x, 2^N - 1) => (sub (shl x, N), x)
8732 Res = DAG.getNode(ISD::SUB, DL, VT,
8733 DAG.getNode(ISD::SHL, DL, VT,
8735 DAG.getConstant(Log2_32(MulAmt + 1), DL,
8741 uint64_t MulAmtAbs = -MulAmt;
8742 if (isPowerOf2_32(MulAmtAbs + 1)) {
8743 // (mul x, -(2^N - 1)) => (sub x, (shl x, N))
8744 Res = DAG.getNode(ISD::SUB, DL, VT,
8746 DAG.getNode(ISD::SHL, DL, VT,
8748 DAG.getConstant(Log2_32(MulAmtAbs + 1), DL,
8750 } else if (isPowerOf2_32(MulAmtAbs - 1)) {
8751 // (mul x, -(2^N + 1)) => - (add (shl x, N), x)
8752 Res = DAG.getNode(ISD::ADD, DL, VT,
8754 DAG.getNode(ISD::SHL, DL, VT,
8756 DAG.getConstant(Log2_32(MulAmtAbs - 1), DL,
8758 Res = DAG.getNode(ISD::SUB, DL, VT,
8759 DAG.getConstant(0, DL, MVT::i32), Res);
8766 Res = DAG.getNode(ISD::SHL, DL, VT,
8767 Res, DAG.getConstant(ShiftAmt, DL, MVT::i32));
8769 // Do not add new nodes to DAG combiner worklist.
8770 DCI.CombineTo(N, Res, false);
8774 static SDValue PerformANDCombine(SDNode *N,
8775 TargetLowering::DAGCombinerInfo &DCI,
8776 const ARMSubtarget *Subtarget) {
8778 // Attempt to use immediate-form VBIC
8779 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
8781 EVT VT = N->getValueType(0);
8782 SelectionDAG &DAG = DCI.DAG;
8784 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8787 APInt SplatBits, SplatUndef;
8788 unsigned SplatBitSize;
8791 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
8792 if (SplatBitSize <= 64) {
8794 SDValue Val = isNEONModifiedImm((~SplatBits).getZExtValue(),
8795 SplatUndef.getZExtValue(), SplatBitSize,
8796 DAG, dl, VbicVT, VT.is128BitVector(),
8798 if (Val.getNode()) {
8800 DAG.getNode(ISD::BITCAST, dl, VbicVT, N->getOperand(0));
8801 SDValue Vbic = DAG.getNode(ARMISD::VBICIMM, dl, VbicVT, Input, Val);
8802 return DAG.getNode(ISD::BITCAST, dl, VT, Vbic);
8807 if (!Subtarget->isThumb1Only()) {
8808 // fold (and (select cc, -1, c), x) -> (select cc, x, (and, x, c))
8809 SDValue Result = combineSelectAndUseCommutative(N, true, DCI);
8810 if (Result.getNode())
8817 /// PerformORCombine - Target-specific dag combine xforms for ISD::OR
8818 static SDValue PerformORCombine(SDNode *N,
8819 TargetLowering::DAGCombinerInfo &DCI,
8820 const ARMSubtarget *Subtarget) {
8821 // Attempt to use immediate-form VORR
8822 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
8824 EVT VT = N->getValueType(0);
8825 SelectionDAG &DAG = DCI.DAG;
8827 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8830 APInt SplatBits, SplatUndef;
8831 unsigned SplatBitSize;
8833 if (BVN && Subtarget->hasNEON() &&
8834 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
8835 if (SplatBitSize <= 64) {
8837 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
8838 SplatUndef.getZExtValue(), SplatBitSize,
8839 DAG, dl, VorrVT, VT.is128BitVector(),
8841 if (Val.getNode()) {
8843 DAG.getNode(ISD::BITCAST, dl, VorrVT, N->getOperand(0));
8844 SDValue Vorr = DAG.getNode(ARMISD::VORRIMM, dl, VorrVT, Input, Val);
8845 return DAG.getNode(ISD::BITCAST, dl, VT, Vorr);
8850 if (!Subtarget->isThumb1Only()) {
8851 // fold (or (select cc, 0, c), x) -> (select cc, x, (or, x, c))
8852 SDValue Result = combineSelectAndUseCommutative(N, false, DCI);
8853 if (Result.getNode())
8857 // The code below optimizes (or (and X, Y), Z).
8858 // The AND operand needs to have a single user to make these optimizations
8860 SDValue N0 = N->getOperand(0);
8861 if (N0.getOpcode() != ISD::AND || !N0.hasOneUse())
8863 SDValue N1 = N->getOperand(1);
8865 // (or (and B, A), (and C, ~A)) => (VBSL A, B, C) when A is a constant.
8866 if (Subtarget->hasNEON() && N1.getOpcode() == ISD::AND && VT.isVector() &&
8867 DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
8869 unsigned SplatBitSize;
8872 APInt SplatBits0, SplatBits1;
8873 BuildVectorSDNode *BVN0 = dyn_cast<BuildVectorSDNode>(N0->getOperand(1));
8874 BuildVectorSDNode *BVN1 = dyn_cast<BuildVectorSDNode>(N1->getOperand(1));
8875 // Ensure that the second operand of both ands are constants
8876 if (BVN0 && BVN0->isConstantSplat(SplatBits0, SplatUndef, SplatBitSize,
8877 HasAnyUndefs) && !HasAnyUndefs) {
8878 if (BVN1 && BVN1->isConstantSplat(SplatBits1, SplatUndef, SplatBitSize,
8879 HasAnyUndefs) && !HasAnyUndefs) {
8880 // Ensure that the bit width of the constants are the same and that
8881 // the splat arguments are logical inverses as per the pattern we
8882 // are trying to simplify.
8883 if (SplatBits0.getBitWidth() == SplatBits1.getBitWidth() &&
8884 SplatBits0 == ~SplatBits1) {
8885 // Canonicalize the vector type to make instruction selection
8887 EVT CanonicalVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
8888 SDValue Result = DAG.getNode(ARMISD::VBSL, dl, CanonicalVT,
8892 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
8898 // Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when
8901 // BFI is only available on V6T2+
8902 if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops())
8906 // 1) or (and A, mask), val => ARMbfi A, val, mask
8907 // iff (val & mask) == val
8909 // 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
8910 // 2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2)
8911 // && mask == ~mask2
8912 // 2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2)
8913 // && ~mask == mask2
8914 // (i.e., copy a bitfield value into another bitfield of the same width)
8919 SDValue N00 = N0.getOperand(0);
8921 // The value and the mask need to be constants so we can verify this is
8922 // actually a bitfield set. If the mask is 0xffff, we can do better
8923 // via a movt instruction, so don't use BFI in that case.
8924 SDValue MaskOp = N0.getOperand(1);
8925 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(MaskOp);
8928 unsigned Mask = MaskC->getZExtValue();
8932 // Case (1): or (and A, mask), val => ARMbfi A, val, mask
8933 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
8935 unsigned Val = N1C->getZExtValue();
8936 if ((Val & ~Mask) != Val)
8939 if (ARM::isBitFieldInvertedMask(Mask)) {
8940 Val >>= countTrailingZeros(~Mask);
8942 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00,
8943 DAG.getConstant(Val, DL, MVT::i32),
8944 DAG.getConstant(Mask, DL, MVT::i32));
8946 // Do not add new nodes to DAG combiner worklist.
8947 DCI.CombineTo(N, Res, false);
8950 } else if (N1.getOpcode() == ISD::AND) {
8951 // case (2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
8952 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
8955 unsigned Mask2 = N11C->getZExtValue();
8957 // Mask and ~Mask2 (or reverse) must be equivalent for the BFI pattern
8959 if (ARM::isBitFieldInvertedMask(Mask) &&
8961 // The pack halfword instruction works better for masks that fit it,
8962 // so use that when it's available.
8963 if (Subtarget->hasT2ExtractPack() &&
8964 (Mask == 0xffff || Mask == 0xffff0000))
8967 unsigned amt = countTrailingZeros(Mask2);
8968 Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0),
8969 DAG.getConstant(amt, DL, MVT::i32));
8970 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00, Res,
8971 DAG.getConstant(Mask, DL, MVT::i32));
8972 // Do not add new nodes to DAG combiner worklist.
8973 DCI.CombineTo(N, Res, false);
8975 } else if (ARM::isBitFieldInvertedMask(~Mask) &&
8977 // The pack halfword instruction works better for masks that fit it,
8978 // so use that when it's available.
8979 if (Subtarget->hasT2ExtractPack() &&
8980 (Mask2 == 0xffff || Mask2 == 0xffff0000))
8983 unsigned lsb = countTrailingZeros(Mask);
8984 Res = DAG.getNode(ISD::SRL, DL, VT, N00,
8985 DAG.getConstant(lsb, DL, MVT::i32));
8986 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res,
8987 DAG.getConstant(Mask2, DL, MVT::i32));
8988 // Do not add new nodes to DAG combiner worklist.
8989 DCI.CombineTo(N, Res, false);
8994 if (DAG.MaskedValueIsZero(N1, MaskC->getAPIntValue()) &&
8995 N00.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N00.getOperand(1)) &&
8996 ARM::isBitFieldInvertedMask(~Mask)) {
8997 // Case (3): or (and (shl A, #shamt), mask), B => ARMbfi B, A, ~mask
8998 // where lsb(mask) == #shamt and masked bits of B are known zero.
8999 SDValue ShAmt = N00.getOperand(1);
9000 unsigned ShAmtC = cast<ConstantSDNode>(ShAmt)->getZExtValue();
9001 unsigned LSB = countTrailingZeros(Mask);
9005 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1, N00.getOperand(0),
9006 DAG.getConstant(~Mask, DL, MVT::i32));
9008 // Do not add new nodes to DAG combiner worklist.
9009 DCI.CombineTo(N, Res, false);
9015 static SDValue PerformXORCombine(SDNode *N,
9016 TargetLowering::DAGCombinerInfo &DCI,
9017 const ARMSubtarget *Subtarget) {
9018 EVT VT = N->getValueType(0);
9019 SelectionDAG &DAG = DCI.DAG;
9021 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
9024 if (!Subtarget->isThumb1Only()) {
9025 // fold (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c))
9026 SDValue Result = combineSelectAndUseCommutative(N, false, DCI);
9027 if (Result.getNode())
9034 /// PerformBFICombine - (bfi A, (and B, Mask1), Mask2) -> (bfi A, B, Mask2) iff
9035 /// the bits being cleared by the AND are not demanded by the BFI.
9036 static SDValue PerformBFICombine(SDNode *N,
9037 TargetLowering::DAGCombinerInfo &DCI) {
9038 SDValue N1 = N->getOperand(1);
9039 if (N1.getOpcode() == ISD::AND) {
9040 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
9043 unsigned InvMask = cast<ConstantSDNode>(N->getOperand(2))->getZExtValue();
9044 unsigned LSB = countTrailingZeros(~InvMask);
9045 unsigned Width = (32 - countLeadingZeros(~InvMask)) - LSB;
9047 static_cast<unsigned>(std::numeric_limits<unsigned>::digits) &&
9048 "undefined behavior");
9049 unsigned Mask = (1u << Width) - 1;
9050 unsigned Mask2 = N11C->getZExtValue();
9051 if ((Mask & (~Mask2)) == 0)
9052 return DCI.DAG.getNode(ARMISD::BFI, SDLoc(N), N->getValueType(0),
9053 N->getOperand(0), N1.getOperand(0),
9059 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
9060 /// ARMISD::VMOVRRD.
9061 static SDValue PerformVMOVRRDCombine(SDNode *N,
9062 TargetLowering::DAGCombinerInfo &DCI,
9063 const ARMSubtarget *Subtarget) {
9064 // vmovrrd(vmovdrr x, y) -> x,y
9065 SDValue InDouble = N->getOperand(0);
9066 if (InDouble.getOpcode() == ARMISD::VMOVDRR && !Subtarget->isFPOnlySP())
9067 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
9069 // vmovrrd(load f64) -> (load i32), (load i32)
9070 SDNode *InNode = InDouble.getNode();
9071 if (ISD::isNormalLoad(InNode) && InNode->hasOneUse() &&
9072 InNode->getValueType(0) == MVT::f64 &&
9073 InNode->getOperand(1).getOpcode() == ISD::FrameIndex &&
9074 !cast<LoadSDNode>(InNode)->isVolatile()) {
9075 // TODO: Should this be done for non-FrameIndex operands?
9076 LoadSDNode *LD = cast<LoadSDNode>(InNode);
9078 SelectionDAG &DAG = DCI.DAG;
9080 SDValue BasePtr = LD->getBasePtr();
9081 SDValue NewLD1 = DAG.getLoad(MVT::i32, DL, LD->getChain(), BasePtr,
9082 LD->getPointerInfo(), LD->isVolatile(),
9083 LD->isNonTemporal(), LD->isInvariant(),
9084 LD->getAlignment());
9086 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
9087 DAG.getConstant(4, DL, MVT::i32));
9088 SDValue NewLD2 = DAG.getLoad(MVT::i32, DL, NewLD1.getValue(1), OffsetPtr,
9089 LD->getPointerInfo(), LD->isVolatile(),
9090 LD->isNonTemporal(), LD->isInvariant(),
9091 std::min(4U, LD->getAlignment() / 2));
9093 DAG.ReplaceAllUsesOfValueWith(SDValue(LD, 1), NewLD2.getValue(1));
9094 if (DCI.DAG.getDataLayout().isBigEndian())
9095 std::swap (NewLD1, NewLD2);
9096 SDValue Result = DCI.CombineTo(N, NewLD1, NewLD2);
9103 /// PerformVMOVDRRCombine - Target-specific dag combine xforms for
9104 /// ARMISD::VMOVDRR. This is also used for BUILD_VECTORs with 2 operands.
9105 static SDValue PerformVMOVDRRCombine(SDNode *N, SelectionDAG &DAG) {
9106 // N=vmovrrd(X); vmovdrr(N:0, N:1) -> bit_convert(X)
9107 SDValue Op0 = N->getOperand(0);
9108 SDValue Op1 = N->getOperand(1);
9109 if (Op0.getOpcode() == ISD::BITCAST)
9110 Op0 = Op0.getOperand(0);
9111 if (Op1.getOpcode() == ISD::BITCAST)
9112 Op1 = Op1.getOperand(0);
9113 if (Op0.getOpcode() == ARMISD::VMOVRRD &&
9114 Op0.getNode() == Op1.getNode() &&
9115 Op0.getResNo() == 0 && Op1.getResNo() == 1)
9116 return DAG.getNode(ISD::BITCAST, SDLoc(N),
9117 N->getValueType(0), Op0.getOperand(0));
9121 /// hasNormalLoadOperand - Check if any of the operands of a BUILD_VECTOR node
9122 /// are normal, non-volatile loads. If so, it is profitable to bitcast an
9123 /// i64 vector to have f64 elements, since the value can then be loaded
9124 /// directly into a VFP register.
9125 static bool hasNormalLoadOperand(SDNode *N) {
9126 unsigned NumElts = N->getValueType(0).getVectorNumElements();
9127 for (unsigned i = 0; i < NumElts; ++i) {
9128 SDNode *Elt = N->getOperand(i).getNode();
9129 if (ISD::isNormalLoad(Elt) && !cast<LoadSDNode>(Elt)->isVolatile())
9135 /// PerformBUILD_VECTORCombine - Target-specific dag combine xforms for
9136 /// ISD::BUILD_VECTOR.
9137 static SDValue PerformBUILD_VECTORCombine(SDNode *N,
9138 TargetLowering::DAGCombinerInfo &DCI,
9139 const ARMSubtarget *Subtarget) {
9140 // build_vector(N=ARMISD::VMOVRRD(X), N:1) -> bit_convert(X):
9141 // VMOVRRD is introduced when legalizing i64 types. It forces the i64 value
9142 // into a pair of GPRs, which is fine when the value is used as a scalar,
9143 // but if the i64 value is converted to a vector, we need to undo the VMOVRRD.
9144 SelectionDAG &DAG = DCI.DAG;
9145 if (N->getNumOperands() == 2) {
9146 SDValue RV = PerformVMOVDRRCombine(N, DAG);
9151 // Load i64 elements as f64 values so that type legalization does not split
9152 // them up into i32 values.
9153 EVT VT = N->getValueType(0);
9154 if (VT.getVectorElementType() != MVT::i64 || !hasNormalLoadOperand(N))
9157 SmallVector<SDValue, 8> Ops;
9158 unsigned NumElts = VT.getVectorNumElements();
9159 for (unsigned i = 0; i < NumElts; ++i) {
9160 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(i));
9162 // Make the DAGCombiner fold the bitcast.
9163 DCI.AddToWorklist(V.getNode());
9165 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, NumElts);
9166 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, FloatVT, Ops);
9167 return DAG.getNode(ISD::BITCAST, dl, VT, BV);
9170 /// \brief Target-specific dag combine xforms for ARMISD::BUILD_VECTOR.
9172 PerformARMBUILD_VECTORCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
9173 // ARMISD::BUILD_VECTOR is introduced when legalizing ISD::BUILD_VECTOR.
9174 // At that time, we may have inserted bitcasts from integer to float.
9175 // If these bitcasts have survived DAGCombine, change the lowering of this
9176 // BUILD_VECTOR in something more vector friendly, i.e., that does not
9177 // force to use floating point types.
9179 // Make sure we can change the type of the vector.
9180 // This is possible iff:
9181 // 1. The vector is only used in a bitcast to a integer type. I.e.,
9182 // 1.1. Vector is used only once.
9183 // 1.2. Use is a bit convert to an integer type.
9184 // 2. The size of its operands are 32-bits (64-bits are not legal).
9185 EVT VT = N->getValueType(0);
9186 EVT EltVT = VT.getVectorElementType();
9188 // Check 1.1. and 2.
9189 if (EltVT.getSizeInBits() != 32 || !N->hasOneUse())
9192 // By construction, the input type must be float.
9193 assert(EltVT == MVT::f32 && "Unexpected type!");
9196 SDNode *Use = *N->use_begin();
9197 if (Use->getOpcode() != ISD::BITCAST ||
9198 Use->getValueType(0).isFloatingPoint())
9201 // Check profitability.
9202 // Model is, if more than half of the relevant operands are bitcast from
9203 // i32, turn the build_vector into a sequence of insert_vector_elt.
9204 // Relevant operands are everything that is not statically
9205 // (i.e., at compile time) bitcasted.
9206 unsigned NumOfBitCastedElts = 0;
9207 unsigned NumElts = VT.getVectorNumElements();
9208 unsigned NumOfRelevantElts = NumElts;
9209 for (unsigned Idx = 0; Idx < NumElts; ++Idx) {
9210 SDValue Elt = N->getOperand(Idx);
9211 if (Elt->getOpcode() == ISD::BITCAST) {
9212 // Assume only bit cast to i32 will go away.
9213 if (Elt->getOperand(0).getValueType() == MVT::i32)
9214 ++NumOfBitCastedElts;
9215 } else if (Elt.getOpcode() == ISD::UNDEF || isa<ConstantSDNode>(Elt))
9216 // Constants are statically casted, thus do not count them as
9217 // relevant operands.
9218 --NumOfRelevantElts;
9221 // Check if more than half of the elements require a non-free bitcast.
9222 if (NumOfBitCastedElts <= NumOfRelevantElts / 2)
9225 SelectionDAG &DAG = DCI.DAG;
9226 // Create the new vector type.
9227 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
9228 // Check if the type is legal.
9229 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9230 if (!TLI.isTypeLegal(VecVT))
9234 // ARMISD::BUILD_VECTOR E1, E2, ..., EN.
9235 // => BITCAST INSERT_VECTOR_ELT
9236 // (INSERT_VECTOR_ELT (...), (BITCAST EN-1), N-1),
9238 SDValue Vec = DAG.getUNDEF(VecVT);
9240 for (unsigned Idx = 0 ; Idx < NumElts; ++Idx) {
9241 SDValue V = N->getOperand(Idx);
9242 if (V.getOpcode() == ISD::UNDEF)
9244 if (V.getOpcode() == ISD::BITCAST &&
9245 V->getOperand(0).getValueType() == MVT::i32)
9246 // Fold obvious case.
9247 V = V.getOperand(0);
9249 V = DAG.getNode(ISD::BITCAST, SDLoc(V), MVT::i32, V);
9250 // Make the DAGCombiner fold the bitcasts.
9251 DCI.AddToWorklist(V.getNode());
9253 SDValue LaneIdx = DAG.getConstant(Idx, dl, MVT::i32);
9254 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VecVT, Vec, V, LaneIdx);
9256 Vec = DAG.getNode(ISD::BITCAST, dl, VT, Vec);
9257 // Make the DAGCombiner fold the bitcasts.
9258 DCI.AddToWorklist(Vec.getNode());
9262 /// PerformInsertEltCombine - Target-specific dag combine xforms for
9263 /// ISD::INSERT_VECTOR_ELT.
9264 static SDValue PerformInsertEltCombine(SDNode *N,
9265 TargetLowering::DAGCombinerInfo &DCI) {
9266 // Bitcast an i64 load inserted into a vector to f64.
9267 // Otherwise, the i64 value will be legalized to a pair of i32 values.
9268 EVT VT = N->getValueType(0);
9269 SDNode *Elt = N->getOperand(1).getNode();
9270 if (VT.getVectorElementType() != MVT::i64 ||
9271 !ISD::isNormalLoad(Elt) || cast<LoadSDNode>(Elt)->isVolatile())
9274 SelectionDAG &DAG = DCI.DAG;
9276 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
9277 VT.getVectorNumElements());
9278 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, N->getOperand(0));
9279 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(1));
9280 // Make the DAGCombiner fold the bitcasts.
9281 DCI.AddToWorklist(Vec.getNode());
9282 DCI.AddToWorklist(V.getNode());
9283 SDValue InsElt = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, FloatVT,
9284 Vec, V, N->getOperand(2));
9285 return DAG.getNode(ISD::BITCAST, dl, VT, InsElt);
9288 /// PerformVECTOR_SHUFFLECombine - Target-specific dag combine xforms for
9289 /// ISD::VECTOR_SHUFFLE.
9290 static SDValue PerformVECTOR_SHUFFLECombine(SDNode *N, SelectionDAG &DAG) {
9291 // The LLVM shufflevector instruction does not require the shuffle mask
9292 // length to match the operand vector length, but ISD::VECTOR_SHUFFLE does
9293 // have that requirement. When translating to ISD::VECTOR_SHUFFLE, if the
9294 // operands do not match the mask length, they are extended by concatenating
9295 // them with undef vectors. That is probably the right thing for other
9296 // targets, but for NEON it is better to concatenate two double-register
9297 // size vector operands into a single quad-register size vector. Do that
9298 // transformation here:
9299 // shuffle(concat(v1, undef), concat(v2, undef)) ->
9300 // shuffle(concat(v1, v2), undef)
9301 SDValue Op0 = N->getOperand(0);
9302 SDValue Op1 = N->getOperand(1);
9303 if (Op0.getOpcode() != ISD::CONCAT_VECTORS ||
9304 Op1.getOpcode() != ISD::CONCAT_VECTORS ||
9305 Op0.getNumOperands() != 2 ||
9306 Op1.getNumOperands() != 2)
9308 SDValue Concat0Op1 = Op0.getOperand(1);
9309 SDValue Concat1Op1 = Op1.getOperand(1);
9310 if (Concat0Op1.getOpcode() != ISD::UNDEF ||
9311 Concat1Op1.getOpcode() != ISD::UNDEF)
9313 // Skip the transformation if any of the types are illegal.
9314 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9315 EVT VT = N->getValueType(0);
9316 if (!TLI.isTypeLegal(VT) ||
9317 !TLI.isTypeLegal(Concat0Op1.getValueType()) ||
9318 !TLI.isTypeLegal(Concat1Op1.getValueType()))
9321 SDValue NewConcat = DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT,
9322 Op0.getOperand(0), Op1.getOperand(0));
9323 // Translate the shuffle mask.
9324 SmallVector<int, 16> NewMask;
9325 unsigned NumElts = VT.getVectorNumElements();
9326 unsigned HalfElts = NumElts/2;
9327 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
9328 for (unsigned n = 0; n < NumElts; ++n) {
9329 int MaskElt = SVN->getMaskElt(n);
9331 if (MaskElt < (int)HalfElts)
9333 else if (MaskElt >= (int)NumElts && MaskElt < (int)(NumElts + HalfElts))
9334 NewElt = HalfElts + MaskElt - NumElts;
9335 NewMask.push_back(NewElt);
9337 return DAG.getVectorShuffle(VT, SDLoc(N), NewConcat,
9338 DAG.getUNDEF(VT), NewMask.data());
9341 /// CombineBaseUpdate - Target-specific DAG combine function for VLDDUP,
9342 /// NEON load/store intrinsics, and generic vector load/stores, to merge
9343 /// base address updates.
9344 /// For generic load/stores, the memory type is assumed to be a vector.
9345 /// The caller is assumed to have checked legality.
9346 static SDValue CombineBaseUpdate(SDNode *N,
9347 TargetLowering::DAGCombinerInfo &DCI) {
9348 SelectionDAG &DAG = DCI.DAG;
9349 const bool isIntrinsic = (N->getOpcode() == ISD::INTRINSIC_VOID ||
9350 N->getOpcode() == ISD::INTRINSIC_W_CHAIN);
9351 const bool isStore = N->getOpcode() == ISD::STORE;
9352 const unsigned AddrOpIdx = ((isIntrinsic || isStore) ? 2 : 1);
9353 SDValue Addr = N->getOperand(AddrOpIdx);
9354 MemSDNode *MemN = cast<MemSDNode>(N);
9357 // Search for a use of the address operand that is an increment.
9358 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(),
9359 UE = Addr.getNode()->use_end(); UI != UE; ++UI) {
9361 if (User->getOpcode() != ISD::ADD ||
9362 UI.getUse().getResNo() != Addr.getResNo())
9365 // Check that the add is independent of the load/store. Otherwise, folding
9366 // it would create a cycle.
9367 if (User->isPredecessorOf(N) || N->isPredecessorOf(User))
9370 // Find the new opcode for the updating load/store.
9371 bool isLoadOp = true;
9372 bool isLaneOp = false;
9373 unsigned NewOpc = 0;
9374 unsigned NumVecs = 0;
9376 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
9378 default: llvm_unreachable("unexpected intrinsic for Neon base update");
9379 case Intrinsic::arm_neon_vld1: NewOpc = ARMISD::VLD1_UPD;
9381 case Intrinsic::arm_neon_vld2: NewOpc = ARMISD::VLD2_UPD;
9383 case Intrinsic::arm_neon_vld3: NewOpc = ARMISD::VLD3_UPD;
9385 case Intrinsic::arm_neon_vld4: NewOpc = ARMISD::VLD4_UPD;
9387 case Intrinsic::arm_neon_vld2lane: NewOpc = ARMISD::VLD2LN_UPD;
9388 NumVecs = 2; isLaneOp = true; break;
9389 case Intrinsic::arm_neon_vld3lane: NewOpc = ARMISD::VLD3LN_UPD;
9390 NumVecs = 3; isLaneOp = true; break;
9391 case Intrinsic::arm_neon_vld4lane: NewOpc = ARMISD::VLD4LN_UPD;
9392 NumVecs = 4; isLaneOp = true; break;
9393 case Intrinsic::arm_neon_vst1: NewOpc = ARMISD::VST1_UPD;
9394 NumVecs = 1; isLoadOp = false; break;
9395 case Intrinsic::arm_neon_vst2: NewOpc = ARMISD::VST2_UPD;
9396 NumVecs = 2; isLoadOp = false; break;
9397 case Intrinsic::arm_neon_vst3: NewOpc = ARMISD::VST3_UPD;
9398 NumVecs = 3; isLoadOp = false; break;
9399 case Intrinsic::arm_neon_vst4: NewOpc = ARMISD::VST4_UPD;
9400 NumVecs = 4; isLoadOp = false; break;
9401 case Intrinsic::arm_neon_vst2lane: NewOpc = ARMISD::VST2LN_UPD;
9402 NumVecs = 2; isLoadOp = false; isLaneOp = true; break;
9403 case Intrinsic::arm_neon_vst3lane: NewOpc = ARMISD::VST3LN_UPD;
9404 NumVecs = 3; isLoadOp = false; isLaneOp = true; break;
9405 case Intrinsic::arm_neon_vst4lane: NewOpc = ARMISD::VST4LN_UPD;
9406 NumVecs = 4; isLoadOp = false; isLaneOp = true; break;
9410 switch (N->getOpcode()) {
9411 default: llvm_unreachable("unexpected opcode for Neon base update");
9412 case ARMISD::VLD2DUP: NewOpc = ARMISD::VLD2DUP_UPD; NumVecs = 2; break;
9413 case ARMISD::VLD3DUP: NewOpc = ARMISD::VLD3DUP_UPD; NumVecs = 3; break;
9414 case ARMISD::VLD4DUP: NewOpc = ARMISD::VLD4DUP_UPD; NumVecs = 4; break;
9415 case ISD::LOAD: NewOpc = ARMISD::VLD1_UPD;
9416 NumVecs = 1; isLaneOp = false; break;
9417 case ISD::STORE: NewOpc = ARMISD::VST1_UPD;
9418 NumVecs = 1; isLaneOp = false; isLoadOp = false; break;
9422 // Find the size of memory referenced by the load/store.
9425 VecTy = N->getValueType(0);
9426 } else if (isIntrinsic) {
9427 VecTy = N->getOperand(AddrOpIdx+1).getValueType();
9429 assert(isStore && "Node has to be a load, a store, or an intrinsic!");
9430 VecTy = N->getOperand(1).getValueType();
9433 unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8;
9435 NumBytes /= VecTy.getVectorNumElements();
9437 // If the increment is a constant, it must match the memory ref size.
9438 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
9439 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
9440 uint64_t IncVal = CInc->getZExtValue();
9441 if (IncVal != NumBytes)
9443 } else if (NumBytes >= 3 * 16) {
9444 // VLD3/4 and VST3/4 for 128-bit vectors are implemented with two
9445 // separate instructions that make it harder to use a non-constant update.
9449 // OK, we found an ADD we can fold into the base update.
9450 // Now, create a _UPD node, taking care of not breaking alignment.
9452 EVT AlignedVecTy = VecTy;
9453 unsigned Alignment = MemN->getAlignment();
9455 // If this is a less-than-standard-aligned load/store, change the type to
9456 // match the standard alignment.
9457 // The alignment is overlooked when selecting _UPD variants; and it's
9458 // easier to introduce bitcasts here than fix that.
9459 // There are 3 ways to get to this base-update combine:
9460 // - intrinsics: they are assumed to be properly aligned (to the standard
9461 // alignment of the memory type), so we don't need to do anything.
9462 // - ARMISD::VLDx nodes: they are only generated from the aforementioned
9463 // intrinsics, so, likewise, there's nothing to do.
9464 // - generic load/store instructions: the alignment is specified as an
9465 // explicit operand, rather than implicitly as the standard alignment
9466 // of the memory type (like the intrisics). We need to change the
9467 // memory type to match the explicit alignment. That way, we don't
9468 // generate non-standard-aligned ARMISD::VLDx nodes.
9469 if (isa<LSBaseSDNode>(N)) {
9472 if (Alignment < VecTy.getScalarSizeInBits() / 8) {
9473 MVT EltTy = MVT::getIntegerVT(Alignment * 8);
9474 assert(NumVecs == 1 && "Unexpected multi-element generic load/store.");
9475 assert(!isLaneOp && "Unexpected generic load/store lane.");
9476 unsigned NumElts = NumBytes / (EltTy.getSizeInBits() / 8);
9477 AlignedVecTy = MVT::getVectorVT(EltTy, NumElts);
9479 // Don't set an explicit alignment on regular load/stores that we want
9480 // to transform to VLD/VST 1_UPD nodes.
9481 // This matches the behavior of regular load/stores, which only get an
9482 // explicit alignment if the MMO alignment is larger than the standard
9483 // alignment of the memory type.
9484 // Intrinsics, however, always get an explicit alignment, set to the
9485 // alignment of the MMO.
9489 // Create the new updating load/store node.
9490 // First, create an SDVTList for the new updating node's results.
9492 unsigned NumResultVecs = (isLoadOp ? NumVecs : 0);
9494 for (n = 0; n < NumResultVecs; ++n)
9495 Tys[n] = AlignedVecTy;
9496 Tys[n++] = MVT::i32;
9497 Tys[n] = MVT::Other;
9498 SDVTList SDTys = DAG.getVTList(makeArrayRef(Tys, NumResultVecs+2));
9500 // Then, gather the new node's operands.
9501 SmallVector<SDValue, 8> Ops;
9502 Ops.push_back(N->getOperand(0)); // incoming chain
9503 Ops.push_back(N->getOperand(AddrOpIdx));
9506 if (StoreSDNode *StN = dyn_cast<StoreSDNode>(N)) {
9507 // Try to match the intrinsic's signature
9508 Ops.push_back(StN->getValue());
9510 // Loads (and of course intrinsics) match the intrinsics' signature,
9511 // so just add all but the alignment operand.
9512 for (unsigned i = AddrOpIdx + 1; i < N->getNumOperands() - 1; ++i)
9513 Ops.push_back(N->getOperand(i));
9516 // For all node types, the alignment operand is always the last one.
9517 Ops.push_back(DAG.getConstant(Alignment, dl, MVT::i32));
9519 // If this is a non-standard-aligned STORE, the penultimate operand is the
9520 // stored value. Bitcast it to the aligned type.
9521 if (AlignedVecTy != VecTy && N->getOpcode() == ISD::STORE) {
9522 SDValue &StVal = Ops[Ops.size()-2];
9523 StVal = DAG.getNode(ISD::BITCAST, dl, AlignedVecTy, StVal);
9526 SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, dl, SDTys,
9528 MemN->getMemOperand());
9531 SmallVector<SDValue, 5> NewResults;
9532 for (unsigned i = 0; i < NumResultVecs; ++i)
9533 NewResults.push_back(SDValue(UpdN.getNode(), i));
9535 // If this is an non-standard-aligned LOAD, the first result is the loaded
9536 // value. Bitcast it to the expected result type.
9537 if (AlignedVecTy != VecTy && N->getOpcode() == ISD::LOAD) {
9538 SDValue &LdVal = NewResults[0];
9539 LdVal = DAG.getNode(ISD::BITCAST, dl, VecTy, LdVal);
9542 NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs+1)); // chain
9543 DCI.CombineTo(N, NewResults);
9544 DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs));
9551 static SDValue PerformVLDCombine(SDNode *N,
9552 TargetLowering::DAGCombinerInfo &DCI) {
9553 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
9556 return CombineBaseUpdate(N, DCI);
9559 /// CombineVLDDUP - For a VDUPLANE node N, check if its source operand is a
9560 /// vldN-lane (N > 1) intrinsic, and if all the other uses of that intrinsic
9561 /// are also VDUPLANEs. If so, combine them to a vldN-dup operation and
9563 static bool CombineVLDDUP(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
9564 SelectionDAG &DAG = DCI.DAG;
9565 EVT VT = N->getValueType(0);
9566 // vldN-dup instructions only support 64-bit vectors for N > 1.
9567 if (!VT.is64BitVector())
9570 // Check if the VDUPLANE operand is a vldN-dup intrinsic.
9571 SDNode *VLD = N->getOperand(0).getNode();
9572 if (VLD->getOpcode() != ISD::INTRINSIC_W_CHAIN)
9574 unsigned NumVecs = 0;
9575 unsigned NewOpc = 0;
9576 unsigned IntNo = cast<ConstantSDNode>(VLD->getOperand(1))->getZExtValue();
9577 if (IntNo == Intrinsic::arm_neon_vld2lane) {
9579 NewOpc = ARMISD::VLD2DUP;
9580 } else if (IntNo == Intrinsic::arm_neon_vld3lane) {
9582 NewOpc = ARMISD::VLD3DUP;
9583 } else if (IntNo == Intrinsic::arm_neon_vld4lane) {
9585 NewOpc = ARMISD::VLD4DUP;
9590 // First check that all the vldN-lane uses are VDUPLANEs and that the lane
9591 // numbers match the load.
9592 unsigned VLDLaneNo =
9593 cast<ConstantSDNode>(VLD->getOperand(NumVecs+3))->getZExtValue();
9594 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
9596 // Ignore uses of the chain result.
9597 if (UI.getUse().getResNo() == NumVecs)
9600 if (User->getOpcode() != ARMISD::VDUPLANE ||
9601 VLDLaneNo != cast<ConstantSDNode>(User->getOperand(1))->getZExtValue())
9605 // Create the vldN-dup node.
9608 for (n = 0; n < NumVecs; ++n)
9610 Tys[n] = MVT::Other;
9611 SDVTList SDTys = DAG.getVTList(makeArrayRef(Tys, NumVecs+1));
9612 SDValue Ops[] = { VLD->getOperand(0), VLD->getOperand(2) };
9613 MemIntrinsicSDNode *VLDMemInt = cast<MemIntrinsicSDNode>(VLD);
9614 SDValue VLDDup = DAG.getMemIntrinsicNode(NewOpc, SDLoc(VLD), SDTys,
9615 Ops, VLDMemInt->getMemoryVT(),
9616 VLDMemInt->getMemOperand());
9619 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
9621 unsigned ResNo = UI.getUse().getResNo();
9622 // Ignore uses of the chain result.
9623 if (ResNo == NumVecs)
9626 DCI.CombineTo(User, SDValue(VLDDup.getNode(), ResNo));
9629 // Now the vldN-lane intrinsic is dead except for its chain result.
9630 // Update uses of the chain.
9631 std::vector<SDValue> VLDDupResults;
9632 for (unsigned n = 0; n < NumVecs; ++n)
9633 VLDDupResults.push_back(SDValue(VLDDup.getNode(), n));
9634 VLDDupResults.push_back(SDValue(VLDDup.getNode(), NumVecs));
9635 DCI.CombineTo(VLD, VLDDupResults);
9640 /// PerformVDUPLANECombine - Target-specific dag combine xforms for
9641 /// ARMISD::VDUPLANE.
9642 static SDValue PerformVDUPLANECombine(SDNode *N,
9643 TargetLowering::DAGCombinerInfo &DCI) {
9644 SDValue Op = N->getOperand(0);
9646 // If the source is a vldN-lane (N > 1) intrinsic, and all the other uses
9647 // of that intrinsic are also VDUPLANEs, combine them to a vldN-dup operation.
9648 if (CombineVLDDUP(N, DCI))
9649 return SDValue(N, 0);
9651 // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is
9652 // redundant. Ignore bit_converts for now; element sizes are checked below.
9653 while (Op.getOpcode() == ISD::BITCAST)
9654 Op = Op.getOperand(0);
9655 if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM)
9658 // Make sure the VMOV element size is not bigger than the VDUPLANE elements.
9659 unsigned EltSize = Op.getValueType().getVectorElementType().getSizeInBits();
9660 // The canonical VMOV for a zero vector uses a 32-bit element size.
9661 unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
9663 if (ARM_AM::decodeNEONModImm(Imm, EltBits) == 0)
9665 EVT VT = N->getValueType(0);
9666 if (EltSize > VT.getVectorElementType().getSizeInBits())
9669 return DCI.DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Op);
9672 static SDValue PerformLOADCombine(SDNode *N,
9673 TargetLowering::DAGCombinerInfo &DCI) {
9674 EVT VT = N->getValueType(0);
9676 // If this is a legal vector load, try to combine it into a VLD1_UPD.
9677 if (ISD::isNormalLoad(N) && VT.isVector() &&
9678 DCI.DAG.getTargetLoweringInfo().isTypeLegal(VT))
9679 return CombineBaseUpdate(N, DCI);
9684 /// PerformSTORECombine - Target-specific dag combine xforms for
9686 static SDValue PerformSTORECombine(SDNode *N,
9687 TargetLowering::DAGCombinerInfo &DCI) {
9688 StoreSDNode *St = cast<StoreSDNode>(N);
9689 if (St->isVolatile())
9692 // Optimize trunc store (of multiple scalars) to shuffle and store. First,
9693 // pack all of the elements in one place. Next, store to memory in fewer
9695 SDValue StVal = St->getValue();
9696 EVT VT = StVal.getValueType();
9697 if (St->isTruncatingStore() && VT.isVector()) {
9698 SelectionDAG &DAG = DCI.DAG;
9699 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9700 EVT StVT = St->getMemoryVT();
9701 unsigned NumElems = VT.getVectorNumElements();
9702 assert(StVT != VT && "Cannot truncate to the same type");
9703 unsigned FromEltSz = VT.getVectorElementType().getSizeInBits();
9704 unsigned ToEltSz = StVT.getVectorElementType().getSizeInBits();
9706 // From, To sizes and ElemCount must be pow of two
9707 if (!isPowerOf2_32(NumElems * FromEltSz * ToEltSz)) return SDValue();
9709 // We are going to use the original vector elt for storing.
9710 // Accumulated smaller vector elements must be a multiple of the store size.
9711 if (0 != (NumElems * FromEltSz) % ToEltSz) return SDValue();
9713 unsigned SizeRatio = FromEltSz / ToEltSz;
9714 assert(SizeRatio * NumElems * ToEltSz == VT.getSizeInBits());
9716 // Create a type on which we perform the shuffle.
9717 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(), StVT.getScalarType(),
9718 NumElems*SizeRatio);
9719 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
9722 SDValue WideVec = DAG.getNode(ISD::BITCAST, DL, WideVecVT, StVal);
9723 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
9724 for (unsigned i = 0; i < NumElems; ++i)
9725 ShuffleVec[i] = DAG.getDataLayout().isBigEndian()
9726 ? (i + 1) * SizeRatio - 1
9729 // Can't shuffle using an illegal type.
9730 if (!TLI.isTypeLegal(WideVecVT)) return SDValue();
9732 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, DL, WideVec,
9733 DAG.getUNDEF(WideVec.getValueType()),
9735 // At this point all of the data is stored at the bottom of the
9736 // register. We now need to save it to mem.
9738 // Find the largest store unit
9739 MVT StoreType = MVT::i8;
9740 for (MVT Tp : MVT::integer_valuetypes()) {
9741 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToEltSz)
9744 // Didn't find a legal store type.
9745 if (!TLI.isTypeLegal(StoreType))
9748 // Bitcast the original vector into a vector of store-size units
9749 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
9750 StoreType, VT.getSizeInBits()/EVT(StoreType).getSizeInBits());
9751 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
9752 SDValue ShuffWide = DAG.getNode(ISD::BITCAST, DL, StoreVecVT, Shuff);
9753 SmallVector<SDValue, 8> Chains;
9754 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits() / 8, DL,
9755 TLI.getPointerTy(DAG.getDataLayout()));
9756 SDValue BasePtr = St->getBasePtr();
9758 // Perform one or more big stores into memory.
9759 unsigned E = (ToEltSz*NumElems)/StoreType.getSizeInBits();
9760 for (unsigned I = 0; I < E; I++) {
9761 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
9762 StoreType, ShuffWide,
9763 DAG.getIntPtrConstant(I, DL));
9764 SDValue Ch = DAG.getStore(St->getChain(), DL, SubVec, BasePtr,
9765 St->getPointerInfo(), St->isVolatile(),
9766 St->isNonTemporal(), St->getAlignment());
9767 BasePtr = DAG.getNode(ISD::ADD, DL, BasePtr.getValueType(), BasePtr,
9769 Chains.push_back(Ch);
9771 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
9774 if (!ISD::isNormalStore(St))
9777 // Split a store of a VMOVDRR into two integer stores to avoid mixing NEON and
9778 // ARM stores of arguments in the same cache line.
9779 if (StVal.getNode()->getOpcode() == ARMISD::VMOVDRR &&
9780 StVal.getNode()->hasOneUse()) {
9781 SelectionDAG &DAG = DCI.DAG;
9782 bool isBigEndian = DAG.getDataLayout().isBigEndian();
9784 SDValue BasePtr = St->getBasePtr();
9785 SDValue NewST1 = DAG.getStore(St->getChain(), DL,
9786 StVal.getNode()->getOperand(isBigEndian ? 1 : 0 ),
9787 BasePtr, St->getPointerInfo(), St->isVolatile(),
9788 St->isNonTemporal(), St->getAlignment());
9790 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
9791 DAG.getConstant(4, DL, MVT::i32));
9792 return DAG.getStore(NewST1.getValue(0), DL,
9793 StVal.getNode()->getOperand(isBigEndian ? 0 : 1),
9794 OffsetPtr, St->getPointerInfo(), St->isVolatile(),
9795 St->isNonTemporal(),
9796 std::min(4U, St->getAlignment() / 2));
9799 if (StVal.getValueType() == MVT::i64 &&
9800 StVal.getNode()->getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
9802 // Bitcast an i64 store extracted from a vector to f64.
9803 // Otherwise, the i64 value will be legalized to a pair of i32 values.
9804 SelectionDAG &DAG = DCI.DAG;
9806 SDValue IntVec = StVal.getOperand(0);
9807 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
9808 IntVec.getValueType().getVectorNumElements());
9809 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, IntVec);
9810 SDValue ExtElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
9811 Vec, StVal.getOperand(1));
9813 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ExtElt);
9814 // Make the DAGCombiner fold the bitcasts.
9815 DCI.AddToWorklist(Vec.getNode());
9816 DCI.AddToWorklist(ExtElt.getNode());
9817 DCI.AddToWorklist(V.getNode());
9818 return DAG.getStore(St->getChain(), dl, V, St->getBasePtr(),
9819 St->getPointerInfo(), St->isVolatile(),
9820 St->isNonTemporal(), St->getAlignment(),
9824 // If this is a legal vector store, try to combine it into a VST1_UPD.
9825 if (ISD::isNormalStore(N) && VT.isVector() &&
9826 DCI.DAG.getTargetLoweringInfo().isTypeLegal(VT))
9827 return CombineBaseUpdate(N, DCI);
9832 /// PerformVCVTCombine - VCVT (floating-point to fixed-point, Advanced SIMD)
9833 /// can replace combinations of VMUL and VCVT (floating-point to integer)
9834 /// when the VMUL has a constant operand that is a power of 2.
9836 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
9837 /// vmul.f32 d16, d17, d16
9838 /// vcvt.s32.f32 d16, d16
9840 /// vcvt.s32.f32 d16, d16, #3
9841 static SDValue PerformVCVTCombine(SDNode *N, SelectionDAG &DAG,
9842 const ARMSubtarget *Subtarget) {
9843 if (!Subtarget->hasNEON())
9846 SDValue Op = N->getOperand(0);
9847 if (!Op.getValueType().isVector() || Op.getOpcode() != ISD::FMUL)
9850 SDValue ConstVec = Op->getOperand(1);
9851 if (!isa<BuildVectorSDNode>(ConstVec))
9854 MVT FloatTy = Op.getSimpleValueType().getVectorElementType();
9855 uint32_t FloatBits = FloatTy.getSizeInBits();
9856 MVT IntTy = N->getSimpleValueType(0).getVectorElementType();
9857 uint32_t IntBits = IntTy.getSizeInBits();
9858 unsigned NumLanes = Op.getValueType().getVectorNumElements();
9859 if (FloatBits != 32 || IntBits > 32 || NumLanes > 4) {
9860 // These instructions only exist converting from f32 to i32. We can handle
9861 // smaller integers by generating an extra truncate, but larger ones would
9862 // be lossy. We also can't handle more then 4 lanes, since these intructions
9863 // only support v2i32/v4i32 types.
9867 BitVector UndefElements;
9868 BuildVectorSDNode *BV = cast<BuildVectorSDNode>(ConstVec);
9869 int32_t C = BV->getConstantFPSplatPow2ToLog2Int(&UndefElements, 33);
9870 if (C == -1 || C == 0 || C > 32)
9874 bool isSigned = N->getOpcode() == ISD::FP_TO_SINT;
9875 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfp2fxs :
9876 Intrinsic::arm_neon_vcvtfp2fxu;
9877 SDValue FixConv = DAG.getNode(
9878 ISD::INTRINSIC_WO_CHAIN, dl, NumLanes == 2 ? MVT::v2i32 : MVT::v4i32,
9879 DAG.getConstant(IntrinsicOpcode, dl, MVT::i32), Op->getOperand(0),
9880 DAG.getConstant(C, dl, MVT::i32));
9882 if (IntBits < FloatBits)
9883 FixConv = DAG.getNode(ISD::TRUNCATE, dl, N->getValueType(0), FixConv);
9888 /// PerformVDIVCombine - VCVT (fixed-point to floating-point, Advanced SIMD)
9889 /// can replace combinations of VCVT (integer to floating-point) and VDIV
9890 /// when the VDIV has a constant operand that is a power of 2.
9892 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
9893 /// vcvt.f32.s32 d16, d16
9894 /// vdiv.f32 d16, d17, d16
9896 /// vcvt.f32.s32 d16, d16, #3
9897 static SDValue PerformVDIVCombine(SDNode *N, SelectionDAG &DAG,
9898 const ARMSubtarget *Subtarget) {
9899 if (!Subtarget->hasNEON())
9902 SDValue Op = N->getOperand(0);
9903 unsigned OpOpcode = Op.getNode()->getOpcode();
9904 if (!N->getValueType(0).isVector() ||
9905 (OpOpcode != ISD::SINT_TO_FP && OpOpcode != ISD::UINT_TO_FP))
9908 SDValue ConstVec = N->getOperand(1);
9909 if (!isa<BuildVectorSDNode>(ConstVec))
9912 MVT FloatTy = N->getSimpleValueType(0).getVectorElementType();
9913 uint32_t FloatBits = FloatTy.getSizeInBits();
9914 MVT IntTy = Op.getOperand(0).getSimpleValueType().getVectorElementType();
9915 uint32_t IntBits = IntTy.getSizeInBits();
9916 unsigned NumLanes = Op.getValueType().getVectorNumElements();
9917 if (FloatBits != 32 || IntBits > 32 || NumLanes > 4) {
9918 // These instructions only exist converting from i32 to f32. We can handle
9919 // smaller integers by generating an extra extend, but larger ones would
9920 // be lossy. We also can't handle more then 4 lanes, since these intructions
9921 // only support v2i32/v4i32 types.
9925 BitVector UndefElements;
9926 BuildVectorSDNode *BV = cast<BuildVectorSDNode>(ConstVec);
9927 int32_t C = BV->getConstantFPSplatPow2ToLog2Int(&UndefElements, 33);
9928 if (C == -1 || C == 0 || C > 32)
9932 bool isSigned = OpOpcode == ISD::SINT_TO_FP;
9933 SDValue ConvInput = Op.getOperand(0);
9934 if (IntBits < FloatBits)
9935 ConvInput = DAG.getNode(isSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND,
9936 dl, NumLanes == 2 ? MVT::v2i32 : MVT::v4i32,
9939 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfxs2fp :
9940 Intrinsic::arm_neon_vcvtfxu2fp;
9941 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl,
9943 DAG.getConstant(IntrinsicOpcode, dl, MVT::i32),
9944 ConvInput, DAG.getConstant(C, dl, MVT::i32));
9947 /// Getvshiftimm - Check if this is a valid build_vector for the immediate
9948 /// operand of a vector shift operation, where all the elements of the
9949 /// build_vector must have the same constant integer value.
9950 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
9951 // Ignore bit_converts.
9952 while (Op.getOpcode() == ISD::BITCAST)
9953 Op = Op.getOperand(0);
9954 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
9955 APInt SplatBits, SplatUndef;
9956 unsigned SplatBitSize;
9958 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
9959 HasAnyUndefs, ElementBits) ||
9960 SplatBitSize > ElementBits)
9962 Cnt = SplatBits.getSExtValue();
9966 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
9967 /// operand of a vector shift left operation. That value must be in the range:
9968 /// 0 <= Value < ElementBits for a left shift; or
9969 /// 0 <= Value <= ElementBits for a long left shift.
9970 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
9971 assert(VT.isVector() && "vector shift count is not a vector type");
9972 int64_t ElementBits = VT.getVectorElementType().getSizeInBits();
9973 if (! getVShiftImm(Op, ElementBits, Cnt))
9975 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
9978 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
9979 /// operand of a vector shift right operation. For a shift opcode, the value
9980 /// is positive, but for an intrinsic the value count must be negative. The
9981 /// absolute value must be in the range:
9982 /// 1 <= |Value| <= ElementBits for a right shift; or
9983 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
9984 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
9986 assert(VT.isVector() && "vector shift count is not a vector type");
9987 int64_t ElementBits = VT.getVectorElementType().getSizeInBits();
9988 if (! getVShiftImm(Op, ElementBits, Cnt))
9991 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
9992 if (Cnt >= -(isNarrow ? ElementBits/2 : ElementBits) && Cnt <= -1) {
9999 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
10000 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
10001 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
10004 // Don't do anything for most intrinsics.
10007 case Intrinsic::arm_neon_vabds:
10008 if (!N->getValueType(0).isInteger())
10010 return DAG.getNode(ISD::SABSDIFF, SDLoc(N), N->getValueType(0),
10011 N->getOperand(1), N->getOperand(2));
10012 case Intrinsic::arm_neon_vabdu:
10013 return DAG.getNode(ISD::UABSDIFF, SDLoc(N), N->getValueType(0),
10014 N->getOperand(1), N->getOperand(2));
10016 // Vector shifts: check for immediate versions and lower them.
10017 // Note: This is done during DAG combining instead of DAG legalizing because
10018 // the build_vectors for 64-bit vector element shift counts are generally
10019 // not legal, and it is hard to see their values after they get legalized to
10020 // loads from a constant pool.
10021 case Intrinsic::arm_neon_vshifts:
10022 case Intrinsic::arm_neon_vshiftu:
10023 case Intrinsic::arm_neon_vrshifts:
10024 case Intrinsic::arm_neon_vrshiftu:
10025 case Intrinsic::arm_neon_vrshiftn:
10026 case Intrinsic::arm_neon_vqshifts:
10027 case Intrinsic::arm_neon_vqshiftu:
10028 case Intrinsic::arm_neon_vqshiftsu:
10029 case Intrinsic::arm_neon_vqshiftns:
10030 case Intrinsic::arm_neon_vqshiftnu:
10031 case Intrinsic::arm_neon_vqshiftnsu:
10032 case Intrinsic::arm_neon_vqrshiftns:
10033 case Intrinsic::arm_neon_vqrshiftnu:
10034 case Intrinsic::arm_neon_vqrshiftnsu: {
10035 EVT VT = N->getOperand(1).getValueType();
10037 unsigned VShiftOpc = 0;
10040 case Intrinsic::arm_neon_vshifts:
10041 case Intrinsic::arm_neon_vshiftu:
10042 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
10043 VShiftOpc = ARMISD::VSHL;
10046 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
10047 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
10048 ARMISD::VSHRs : ARMISD::VSHRu);
10053 case Intrinsic::arm_neon_vrshifts:
10054 case Intrinsic::arm_neon_vrshiftu:
10055 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
10059 case Intrinsic::arm_neon_vqshifts:
10060 case Intrinsic::arm_neon_vqshiftu:
10061 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
10065 case Intrinsic::arm_neon_vqshiftsu:
10066 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
10068 llvm_unreachable("invalid shift count for vqshlu intrinsic");
10070 case Intrinsic::arm_neon_vrshiftn:
10071 case Intrinsic::arm_neon_vqshiftns:
10072 case Intrinsic::arm_neon_vqshiftnu:
10073 case Intrinsic::arm_neon_vqshiftnsu:
10074 case Intrinsic::arm_neon_vqrshiftns:
10075 case Intrinsic::arm_neon_vqrshiftnu:
10076 case Intrinsic::arm_neon_vqrshiftnsu:
10077 // Narrowing shifts require an immediate right shift.
10078 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
10080 llvm_unreachable("invalid shift count for narrowing vector shift "
10084 llvm_unreachable("unhandled vector shift");
10088 case Intrinsic::arm_neon_vshifts:
10089 case Intrinsic::arm_neon_vshiftu:
10090 // Opcode already set above.
10092 case Intrinsic::arm_neon_vrshifts:
10093 VShiftOpc = ARMISD::VRSHRs; break;
10094 case Intrinsic::arm_neon_vrshiftu:
10095 VShiftOpc = ARMISD::VRSHRu; break;
10096 case Intrinsic::arm_neon_vrshiftn:
10097 VShiftOpc = ARMISD::VRSHRN; break;
10098 case Intrinsic::arm_neon_vqshifts:
10099 VShiftOpc = ARMISD::VQSHLs; break;
10100 case Intrinsic::arm_neon_vqshiftu:
10101 VShiftOpc = ARMISD::VQSHLu; break;
10102 case Intrinsic::arm_neon_vqshiftsu:
10103 VShiftOpc = ARMISD::VQSHLsu; break;
10104 case Intrinsic::arm_neon_vqshiftns:
10105 VShiftOpc = ARMISD::VQSHRNs; break;
10106 case Intrinsic::arm_neon_vqshiftnu:
10107 VShiftOpc = ARMISD::VQSHRNu; break;
10108 case Intrinsic::arm_neon_vqshiftnsu:
10109 VShiftOpc = ARMISD::VQSHRNsu; break;
10110 case Intrinsic::arm_neon_vqrshiftns:
10111 VShiftOpc = ARMISD::VQRSHRNs; break;
10112 case Intrinsic::arm_neon_vqrshiftnu:
10113 VShiftOpc = ARMISD::VQRSHRNu; break;
10114 case Intrinsic::arm_neon_vqrshiftnsu:
10115 VShiftOpc = ARMISD::VQRSHRNsu; break;
10119 return DAG.getNode(VShiftOpc, dl, N->getValueType(0),
10120 N->getOperand(1), DAG.getConstant(Cnt, dl, MVT::i32));
10123 case Intrinsic::arm_neon_vshiftins: {
10124 EVT VT = N->getOperand(1).getValueType();
10126 unsigned VShiftOpc = 0;
10128 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
10129 VShiftOpc = ARMISD::VSLI;
10130 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
10131 VShiftOpc = ARMISD::VSRI;
10133 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
10137 return DAG.getNode(VShiftOpc, dl, N->getValueType(0),
10138 N->getOperand(1), N->getOperand(2),
10139 DAG.getConstant(Cnt, dl, MVT::i32));
10142 case Intrinsic::arm_neon_vqrshifts:
10143 case Intrinsic::arm_neon_vqrshiftu:
10144 // No immediate versions of these to check for.
10151 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
10152 /// lowers them. As with the vector shift intrinsics, this is done during DAG
10153 /// combining instead of DAG legalizing because the build_vectors for 64-bit
10154 /// vector element shift counts are generally not legal, and it is hard to see
10155 /// their values after they get legalized to loads from a constant pool.
10156 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
10157 const ARMSubtarget *ST) {
10158 EVT VT = N->getValueType(0);
10159 if (N->getOpcode() == ISD::SRL && VT == MVT::i32 && ST->hasV6Ops()) {
10160 // Canonicalize (srl (bswap x), 16) to (rotr (bswap x), 16) if the high
10161 // 16-bits of x is zero. This optimizes rev + lsr 16 to rev16.
10162 SDValue N1 = N->getOperand(1);
10163 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) {
10164 SDValue N0 = N->getOperand(0);
10165 if (C->getZExtValue() == 16 && N0.getOpcode() == ISD::BSWAP &&
10166 DAG.MaskedValueIsZero(N0.getOperand(0),
10167 APInt::getHighBitsSet(32, 16)))
10168 return DAG.getNode(ISD::ROTR, SDLoc(N), VT, N0, N1);
10172 // Nothing to be done for scalar shifts.
10173 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
10174 if (!VT.isVector() || !TLI.isTypeLegal(VT))
10177 assert(ST->hasNEON() && "unexpected vector shift");
10180 switch (N->getOpcode()) {
10181 default: llvm_unreachable("unexpected shift opcode");
10184 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt)) {
10186 return DAG.getNode(ARMISD::VSHL, dl, VT, N->getOperand(0),
10187 DAG.getConstant(Cnt, dl, MVT::i32));
10193 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
10194 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
10195 ARMISD::VSHRs : ARMISD::VSHRu);
10197 return DAG.getNode(VShiftOpc, dl, VT, N->getOperand(0),
10198 DAG.getConstant(Cnt, dl, MVT::i32));
10204 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
10205 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
10206 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
10207 const ARMSubtarget *ST) {
10208 SDValue N0 = N->getOperand(0);
10210 // Check for sign- and zero-extensions of vector extract operations of 8-
10211 // and 16-bit vector elements. NEON supports these directly. They are
10212 // handled during DAG combining because type legalization will promote them
10213 // to 32-bit types and it is messy to recognize the operations after that.
10214 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
10215 SDValue Vec = N0.getOperand(0);
10216 SDValue Lane = N0.getOperand(1);
10217 EVT VT = N->getValueType(0);
10218 EVT EltVT = N0.getValueType();
10219 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
10221 if (VT == MVT::i32 &&
10222 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
10223 TLI.isTypeLegal(Vec.getValueType()) &&
10224 isa<ConstantSDNode>(Lane)) {
10227 switch (N->getOpcode()) {
10228 default: llvm_unreachable("unexpected opcode");
10229 case ISD::SIGN_EXTEND:
10230 Opc = ARMISD::VGETLANEs;
10232 case ISD::ZERO_EXTEND:
10233 case ISD::ANY_EXTEND:
10234 Opc = ARMISD::VGETLANEu;
10237 return DAG.getNode(Opc, SDLoc(N), VT, Vec, Lane);
10244 static void computeKnownBits(SelectionDAG &DAG, SDValue Op, APInt &KnownZero,
10246 if (Op.getOpcode() == ARMISD::BFI) {
10247 // Conservatively, we can recurse down the first operand
10248 // and just mask out all affected bits.
10249 computeKnownBits(DAG, Op.getOperand(0), KnownZero, KnownOne);
10251 // The operand to BFI is already a mask suitable for removing the bits it
10253 ConstantSDNode *CI = cast<ConstantSDNode>(Op.getOperand(2));
10254 APInt Mask = CI->getAPIntValue();
10259 if (Op.getOpcode() == ARMISD::CMOV) {
10260 APInt KZ2(KnownZero.getBitWidth(), 0);
10261 APInt KO2(KnownOne.getBitWidth(), 0);
10262 computeKnownBits(DAG, Op.getOperand(1), KnownZero, KnownOne);
10263 computeKnownBits(DAG, Op.getOperand(2), KZ2, KO2);
10269 return DAG.computeKnownBits(Op, KnownZero, KnownOne);
10272 SDValue ARMTargetLowering::PerformCMOVToBFICombine(SDNode *CMOV, SelectionDAG &DAG) const {
10273 // If we have a CMOV, OR and AND combination such as:
10278 // * CN is a single bit;
10279 // * All bits covered by CM are known zero in y
10281 // Then we can convert this into a sequence of BFI instructions. This will
10282 // always be a win if CM is a single bit, will always be no worse than the
10283 // TST&OR sequence if CM is two bits, and for thumb will be no worse if CM is
10284 // three bits (due to the extra IT instruction).
10286 SDValue Op0 = CMOV->getOperand(0);
10287 SDValue Op1 = CMOV->getOperand(1);
10288 SDValue CmpZ = CMOV->getOperand(4);
10290 assert(CmpZ->getOpcode() == ARMISD::CMPZ);
10291 SDValue And = CmpZ->getOperand(0);
10292 if (And->getOpcode() != ISD::AND)
10294 ConstantSDNode *AndC = dyn_cast<ConstantSDNode>(And->getOperand(1));
10295 if (!AndC || !AndC->getAPIntValue().isPowerOf2())
10297 SDValue X = And->getOperand(0);
10299 if (Op1->getOpcode() != ISD::OR)
10302 ConstantSDNode *OrC = dyn_cast<ConstantSDNode>(Op1->getOperand(1));
10305 SDValue Y = Op1->getOperand(0);
10310 // Now, is it profitable to continue?
10311 APInt OrCI = OrC->getAPIntValue();
10312 unsigned Heuristic = Subtarget->isThumb() ? 3 : 2;
10313 if (OrCI.countPopulation() > Heuristic)
10316 // Lastly, can we determine that the bits defined by OrCI
10318 APInt KnownZero, KnownOne;
10319 computeKnownBits(DAG, Y, KnownZero, KnownOne);
10320 if ((OrCI & KnownZero) != OrCI)
10323 // OK, we can do the combine.
10326 EVT VT = X.getValueType();
10327 unsigned BitInX = AndC->getAPIntValue().logBase2();
10330 // We must shift X first.
10331 X = DAG.getNode(ISD::SRL, dl, VT, X,
10332 DAG.getConstant(BitInX, dl, VT));
10335 for (unsigned BitInY = 0, NumActiveBits = OrCI.getActiveBits();
10336 BitInY < NumActiveBits; ++BitInY) {
10337 if (OrCI[BitInY] == 0)
10339 APInt Mask(VT.getSizeInBits(), 0);
10340 Mask.setBit(BitInY);
10341 V = DAG.getNode(ARMISD::BFI, dl, VT, V, X,
10342 // Confusingly, the operand is an *inverted* mask.
10343 DAG.getConstant(~Mask, dl, VT));
10349 /// PerformCMOVCombine - Target-specific DAG combining for ARMISD::CMOV.
10351 ARMTargetLowering::PerformCMOVCombine(SDNode *N, SelectionDAG &DAG) const {
10352 SDValue Cmp = N->getOperand(4);
10353 if (Cmp.getOpcode() != ARMISD::CMPZ)
10354 // Only looking at EQ and NE cases.
10357 EVT VT = N->getValueType(0);
10359 SDValue LHS = Cmp.getOperand(0);
10360 SDValue RHS = Cmp.getOperand(1);
10361 SDValue FalseVal = N->getOperand(0);
10362 SDValue TrueVal = N->getOperand(1);
10363 SDValue ARMcc = N->getOperand(2);
10364 ARMCC::CondCodes CC =
10365 (ARMCC::CondCodes)cast<ConstantSDNode>(ARMcc)->getZExtValue();
10367 // BFI is only available on V6T2+.
10368 if (!Subtarget->isThumb1Only() && Subtarget->hasV6T2Ops()) {
10369 SDValue R = PerformCMOVToBFICombine(N, DAG);
10390 /// FIXME: Turn this into a target neutral optimization?
10392 if (CC == ARMCC::NE && FalseVal == RHS && FalseVal != LHS) {
10393 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, TrueVal, ARMcc,
10394 N->getOperand(3), Cmp);
10395 } else if (CC == ARMCC::EQ && TrueVal == RHS) {
10397 SDValue NewCmp = getARMCmp(LHS, RHS, ISD::SETNE, ARMcc, DAG, dl);
10398 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, FalseVal, ARMcc,
10399 N->getOperand(3), NewCmp);
10402 if (Res.getNode()) {
10403 APInt KnownZero, KnownOne;
10404 DAG.computeKnownBits(SDValue(N,0), KnownZero, KnownOne);
10405 // Capture demanded bits information that would be otherwise lost.
10406 if (KnownZero == 0xfffffffe)
10407 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
10408 DAG.getValueType(MVT::i1));
10409 else if (KnownZero == 0xffffff00)
10410 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
10411 DAG.getValueType(MVT::i8));
10412 else if (KnownZero == 0xffff0000)
10413 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
10414 DAG.getValueType(MVT::i16));
10420 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
10421 DAGCombinerInfo &DCI) const {
10422 switch (N->getOpcode()) {
10424 case ISD::ADDC: return PerformADDCCombine(N, DCI, Subtarget);
10425 case ISD::ADD: return PerformADDCombine(N, DCI, Subtarget);
10426 case ISD::SUB: return PerformSUBCombine(N, DCI);
10427 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
10428 case ISD::OR: return PerformORCombine(N, DCI, Subtarget);
10429 case ISD::XOR: return PerformXORCombine(N, DCI, Subtarget);
10430 case ISD::AND: return PerformANDCombine(N, DCI, Subtarget);
10431 case ARMISD::BFI: return PerformBFICombine(N, DCI);
10432 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI, Subtarget);
10433 case ARMISD::VMOVDRR: return PerformVMOVDRRCombine(N, DCI.DAG);
10434 case ISD::STORE: return PerformSTORECombine(N, DCI);
10435 case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DCI, Subtarget);
10436 case ISD::INSERT_VECTOR_ELT: return PerformInsertEltCombine(N, DCI);
10437 case ISD::VECTOR_SHUFFLE: return PerformVECTOR_SHUFFLECombine(N, DCI.DAG);
10438 case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI);
10439 case ISD::FP_TO_SINT:
10440 case ISD::FP_TO_UINT:
10441 return PerformVCVTCombine(N, DCI.DAG, Subtarget);
10443 return PerformVDIVCombine(N, DCI.DAG, Subtarget);
10444 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
10447 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
10448 case ISD::SIGN_EXTEND:
10449 case ISD::ZERO_EXTEND:
10450 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
10451 case ARMISD::CMOV: return PerformCMOVCombine(N, DCI.DAG);
10452 case ISD::LOAD: return PerformLOADCombine(N, DCI);
10453 case ARMISD::VLD2DUP:
10454 case ARMISD::VLD3DUP:
10455 case ARMISD::VLD4DUP:
10456 return PerformVLDCombine(N, DCI);
10457 case ARMISD::BUILD_VECTOR:
10458 return PerformARMBUILD_VECTORCombine(N, DCI);
10459 case ISD::INTRINSIC_VOID:
10460 case ISD::INTRINSIC_W_CHAIN:
10461 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
10462 case Intrinsic::arm_neon_vld1:
10463 case Intrinsic::arm_neon_vld2:
10464 case Intrinsic::arm_neon_vld3:
10465 case Intrinsic::arm_neon_vld4:
10466 case Intrinsic::arm_neon_vld2lane:
10467 case Intrinsic::arm_neon_vld3lane:
10468 case Intrinsic::arm_neon_vld4lane:
10469 case Intrinsic::arm_neon_vst1:
10470 case Intrinsic::arm_neon_vst2:
10471 case Intrinsic::arm_neon_vst3:
10472 case Intrinsic::arm_neon_vst4:
10473 case Intrinsic::arm_neon_vst2lane:
10474 case Intrinsic::arm_neon_vst3lane:
10475 case Intrinsic::arm_neon_vst4lane:
10476 return PerformVLDCombine(N, DCI);
10484 bool ARMTargetLowering::isDesirableToTransformToIntegerOp(unsigned Opc,
10486 return (VT == MVT::f32) && (Opc == ISD::LOAD || Opc == ISD::STORE);
10489 bool ARMTargetLowering::allowsMisalignedMemoryAccesses(EVT VT,
10492 bool *Fast) const {
10493 // The AllowsUnaliged flag models the SCTLR.A setting in ARM cpus
10494 bool AllowsUnaligned = Subtarget->allowsUnalignedMem();
10496 switch (VT.getSimpleVT().SimpleTy) {
10502 // Unaligned access can use (for example) LRDB, LRDH, LDR
10503 if (AllowsUnaligned) {
10505 *Fast = Subtarget->hasV7Ops();
10512 // For any little-endian targets with neon, we can support unaligned ld/st
10513 // of D and Q (e.g. {D0,D1}) registers by using vld1.i8/vst1.i8.
10514 // A big-endian target may also explicitly support unaligned accesses
10515 if (Subtarget->hasNEON() && (AllowsUnaligned || Subtarget->isLittle())) {
10525 static bool memOpAlign(unsigned DstAlign, unsigned SrcAlign,
10526 unsigned AlignCheck) {
10527 return ((SrcAlign == 0 || SrcAlign % AlignCheck == 0) &&
10528 (DstAlign == 0 || DstAlign % AlignCheck == 0));
10531 EVT ARMTargetLowering::getOptimalMemOpType(uint64_t Size,
10532 unsigned DstAlign, unsigned SrcAlign,
10533 bool IsMemset, bool ZeroMemset,
10535 MachineFunction &MF) const {
10536 const Function *F = MF.getFunction();
10538 // See if we can use NEON instructions for this...
10539 if ((!IsMemset || ZeroMemset) && Subtarget->hasNEON() &&
10540 !F->hasFnAttribute(Attribute::NoImplicitFloat)) {
10543 (memOpAlign(SrcAlign, DstAlign, 16) ||
10544 (allowsMisalignedMemoryAccesses(MVT::v2f64, 0, 1, &Fast) && Fast))) {
10546 } else if (Size >= 8 &&
10547 (memOpAlign(SrcAlign, DstAlign, 8) ||
10548 (allowsMisalignedMemoryAccesses(MVT::f64, 0, 1, &Fast) &&
10554 // Lowering to i32/i16 if the size permits.
10557 else if (Size >= 2)
10560 // Let the target-independent logic figure it out.
10564 bool ARMTargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
10565 if (Val.getOpcode() != ISD::LOAD)
10568 EVT VT1 = Val.getValueType();
10569 if (!VT1.isSimple() || !VT1.isInteger() ||
10570 !VT2.isSimple() || !VT2.isInteger())
10573 switch (VT1.getSimpleVT().SimpleTy) {
10578 // 8-bit and 16-bit loads implicitly zero-extend to 32-bits.
10585 bool ARMTargetLowering::isVectorLoadExtDesirable(SDValue ExtVal) const {
10586 EVT VT = ExtVal.getValueType();
10588 if (!isTypeLegal(VT))
10591 // Don't create a loadext if we can fold the extension into a wide/long
10593 // If there's more than one user instruction, the loadext is desirable no
10594 // matter what. There can be two uses by the same instruction.
10595 if (ExtVal->use_empty() ||
10596 !ExtVal->use_begin()->isOnlyUserOf(ExtVal.getNode()))
10599 SDNode *U = *ExtVal->use_begin();
10600 if ((U->getOpcode() == ISD::ADD || U->getOpcode() == ISD::SUB ||
10601 U->getOpcode() == ISD::SHL || U->getOpcode() == ARMISD::VSHL))
10607 bool ARMTargetLowering::allowTruncateForTailCall(Type *Ty1, Type *Ty2) const {
10608 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
10611 if (!isTypeLegal(EVT::getEVT(Ty1)))
10614 assert(Ty1->getPrimitiveSizeInBits() <= 64 && "i128 is probably not a noop");
10616 // Assuming the caller doesn't have a zeroext or signext return parameter,
10617 // truncation all the way down to i1 is valid.
10622 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
10626 unsigned Scale = 1;
10627 switch (VT.getSimpleVT().SimpleTy) {
10628 default: return false;
10643 if ((V & (Scale - 1)) != 0)
10646 return V == (V & ((1LL << 5) - 1));
10649 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
10650 const ARMSubtarget *Subtarget) {
10651 bool isNeg = false;
10657 switch (VT.getSimpleVT().SimpleTy) {
10658 default: return false;
10663 // + imm12 or - imm8
10665 return V == (V & ((1LL << 8) - 1));
10666 return V == (V & ((1LL << 12) - 1));
10669 // Same as ARM mode. FIXME: NEON?
10670 if (!Subtarget->hasVFP2())
10675 return V == (V & ((1LL << 8) - 1));
10679 /// isLegalAddressImmediate - Return true if the integer value can be used
10680 /// as the offset of the target addressing mode for load / store of the
10682 static bool isLegalAddressImmediate(int64_t V, EVT VT,
10683 const ARMSubtarget *Subtarget) {
10687 if (!VT.isSimple())
10690 if (Subtarget->isThumb1Only())
10691 return isLegalT1AddressImmediate(V, VT);
10692 else if (Subtarget->isThumb2())
10693 return isLegalT2AddressImmediate(V, VT, Subtarget);
10698 switch (VT.getSimpleVT().SimpleTy) {
10699 default: return false;
10704 return V == (V & ((1LL << 12) - 1));
10707 return V == (V & ((1LL << 8) - 1));
10710 if (!Subtarget->hasVFP2()) // FIXME: NEON?
10715 return V == (V & ((1LL << 8) - 1));
10719 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
10721 int Scale = AM.Scale;
10725 switch (VT.getSimpleVT().SimpleTy) {
10726 default: return false;
10734 Scale = Scale & ~1;
10735 return Scale == 2 || Scale == 4 || Scale == 8;
10738 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
10742 // Note, we allow "void" uses (basically, uses that aren't loads or
10743 // stores), because arm allows folding a scale into many arithmetic
10744 // operations. This should be made more precise and revisited later.
10746 // Allow r << imm, but the imm has to be a multiple of two.
10747 if (Scale & 1) return false;
10748 return isPowerOf2_32(Scale);
10752 /// isLegalAddressingMode - Return true if the addressing mode represented
10753 /// by AM is legal for this target, for a load/store of the specified type.
10754 bool ARMTargetLowering::isLegalAddressingMode(const DataLayout &DL,
10755 const AddrMode &AM, Type *Ty,
10756 unsigned AS) const {
10757 EVT VT = getValueType(DL, Ty, true);
10758 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
10761 // Can never fold addr of global into load/store.
10765 switch (AM.Scale) {
10766 case 0: // no scale reg, must be "r+i" or "r", or "i".
10769 if (Subtarget->isThumb1Only())
10773 // ARM doesn't support any R+R*scale+imm addr modes.
10777 if (!VT.isSimple())
10780 if (Subtarget->isThumb2())
10781 return isLegalT2ScaledAddressingMode(AM, VT);
10783 int Scale = AM.Scale;
10784 switch (VT.getSimpleVT().SimpleTy) {
10785 default: return false;
10789 if (Scale < 0) Scale = -Scale;
10793 return isPowerOf2_32(Scale & ~1);
10797 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
10802 // Note, we allow "void" uses (basically, uses that aren't loads or
10803 // stores), because arm allows folding a scale into many arithmetic
10804 // operations. This should be made more precise and revisited later.
10806 // Allow r << imm, but the imm has to be a multiple of two.
10807 if (Scale & 1) return false;
10808 return isPowerOf2_32(Scale);
10814 /// isLegalICmpImmediate - Return true if the specified immediate is legal
10815 /// icmp immediate, that is the target has icmp instructions which can compare
10816 /// a register against the immediate without having to materialize the
10817 /// immediate into a register.
10818 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
10819 // Thumb2 and ARM modes can use cmn for negative immediates.
10820 if (!Subtarget->isThumb())
10821 return ARM_AM::getSOImmVal(std::abs(Imm)) != -1;
10822 if (Subtarget->isThumb2())
10823 return ARM_AM::getT2SOImmVal(std::abs(Imm)) != -1;
10824 // Thumb1 doesn't have cmn, and only 8-bit immediates.
10825 return Imm >= 0 && Imm <= 255;
10828 /// isLegalAddImmediate - Return true if the specified immediate is a legal add
10829 /// *or sub* immediate, that is the target has add or sub instructions which can
10830 /// add a register with the immediate without having to materialize the
10831 /// immediate into a register.
10832 bool ARMTargetLowering::isLegalAddImmediate(int64_t Imm) const {
10833 // Same encoding for add/sub, just flip the sign.
10834 int64_t AbsImm = std::abs(Imm);
10835 if (!Subtarget->isThumb())
10836 return ARM_AM::getSOImmVal(AbsImm) != -1;
10837 if (Subtarget->isThumb2())
10838 return ARM_AM::getT2SOImmVal(AbsImm) != -1;
10839 // Thumb1 only has 8-bit unsigned immediate.
10840 return AbsImm >= 0 && AbsImm <= 255;
10843 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
10844 bool isSEXTLoad, SDValue &Base,
10845 SDValue &Offset, bool &isInc,
10846 SelectionDAG &DAG) {
10847 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
10850 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
10851 // AddressingMode 3
10852 Base = Ptr->getOperand(0);
10853 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
10854 int RHSC = (int)RHS->getZExtValue();
10855 if (RHSC < 0 && RHSC > -256) {
10856 assert(Ptr->getOpcode() == ISD::ADD);
10858 Offset = DAG.getConstant(-RHSC, SDLoc(Ptr), RHS->getValueType(0));
10862 isInc = (Ptr->getOpcode() == ISD::ADD);
10863 Offset = Ptr->getOperand(1);
10865 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
10866 // AddressingMode 2
10867 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
10868 int RHSC = (int)RHS->getZExtValue();
10869 if (RHSC < 0 && RHSC > -0x1000) {
10870 assert(Ptr->getOpcode() == ISD::ADD);
10872 Offset = DAG.getConstant(-RHSC, SDLoc(Ptr), RHS->getValueType(0));
10873 Base = Ptr->getOperand(0);
10878 if (Ptr->getOpcode() == ISD::ADD) {
10880 ARM_AM::ShiftOpc ShOpcVal=
10881 ARM_AM::getShiftOpcForNode(Ptr->getOperand(0).getOpcode());
10882 if (ShOpcVal != ARM_AM::no_shift) {
10883 Base = Ptr->getOperand(1);
10884 Offset = Ptr->getOperand(0);
10886 Base = Ptr->getOperand(0);
10887 Offset = Ptr->getOperand(1);
10892 isInc = (Ptr->getOpcode() == ISD::ADD);
10893 Base = Ptr->getOperand(0);
10894 Offset = Ptr->getOperand(1);
10898 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
10902 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
10903 bool isSEXTLoad, SDValue &Base,
10904 SDValue &Offset, bool &isInc,
10905 SelectionDAG &DAG) {
10906 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
10909 Base = Ptr->getOperand(0);
10910 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
10911 int RHSC = (int)RHS->getZExtValue();
10912 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
10913 assert(Ptr->getOpcode() == ISD::ADD);
10915 Offset = DAG.getConstant(-RHSC, SDLoc(Ptr), RHS->getValueType(0));
10917 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
10918 isInc = Ptr->getOpcode() == ISD::ADD;
10919 Offset = DAG.getConstant(RHSC, SDLoc(Ptr), RHS->getValueType(0));
10927 /// getPreIndexedAddressParts - returns true by value, base pointer and
10928 /// offset pointer and addressing mode by reference if the node's address
10929 /// can be legally represented as pre-indexed load / store address.
10931 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
10933 ISD::MemIndexedMode &AM,
10934 SelectionDAG &DAG) const {
10935 if (Subtarget->isThumb1Only())
10940 bool isSEXTLoad = false;
10941 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
10942 Ptr = LD->getBasePtr();
10943 VT = LD->getMemoryVT();
10944 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
10945 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
10946 Ptr = ST->getBasePtr();
10947 VT = ST->getMemoryVT();
10952 bool isLegal = false;
10953 if (Subtarget->isThumb2())
10954 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
10955 Offset, isInc, DAG);
10957 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
10958 Offset, isInc, DAG);
10962 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
10966 /// getPostIndexedAddressParts - returns true by value, base pointer and
10967 /// offset pointer and addressing mode by reference if this node can be
10968 /// combined with a load / store to form a post-indexed load / store.
10969 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
10972 ISD::MemIndexedMode &AM,
10973 SelectionDAG &DAG) const {
10974 if (Subtarget->isThumb1Only())
10979 bool isSEXTLoad = false;
10980 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
10981 VT = LD->getMemoryVT();
10982 Ptr = LD->getBasePtr();
10983 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
10984 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
10985 VT = ST->getMemoryVT();
10986 Ptr = ST->getBasePtr();
10991 bool isLegal = false;
10992 if (Subtarget->isThumb2())
10993 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
10996 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
11002 // Swap base ptr and offset to catch more post-index load / store when
11003 // it's legal. In Thumb2 mode, offset must be an immediate.
11004 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
11005 !Subtarget->isThumb2())
11006 std::swap(Base, Offset);
11008 // Post-indexed load / store update the base pointer.
11013 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
11017 void ARMTargetLowering::computeKnownBitsForTargetNode(const SDValue Op,
11020 const SelectionDAG &DAG,
11021 unsigned Depth) const {
11022 unsigned BitWidth = KnownOne.getBitWidth();
11023 KnownZero = KnownOne = APInt(BitWidth, 0);
11024 switch (Op.getOpcode()) {
11030 // These nodes' second result is a boolean
11031 if (Op.getResNo() == 0)
11033 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - 1);
11035 case ARMISD::CMOV: {
11036 // Bits are known zero/one if known on the LHS and RHS.
11037 DAG.computeKnownBits(Op.getOperand(0), KnownZero, KnownOne, Depth+1);
11038 if (KnownZero == 0 && KnownOne == 0) return;
11040 APInt KnownZeroRHS, KnownOneRHS;
11041 DAG.computeKnownBits(Op.getOperand(1), KnownZeroRHS, KnownOneRHS, Depth+1);
11042 KnownZero &= KnownZeroRHS;
11043 KnownOne &= KnownOneRHS;
11046 case ISD::INTRINSIC_W_CHAIN: {
11047 ConstantSDNode *CN = cast<ConstantSDNode>(Op->getOperand(1));
11048 Intrinsic::ID IntID = static_cast<Intrinsic::ID>(CN->getZExtValue());
11051 case Intrinsic::arm_ldaex:
11052 case Intrinsic::arm_ldrex: {
11053 EVT VT = cast<MemIntrinsicSDNode>(Op)->getMemoryVT();
11054 unsigned MemBits = VT.getScalarType().getSizeInBits();
11055 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - MemBits);
11063 //===----------------------------------------------------------------------===//
11064 // ARM Inline Assembly Support
11065 //===----------------------------------------------------------------------===//
11067 bool ARMTargetLowering::ExpandInlineAsm(CallInst *CI) const {
11068 // Looking for "rev" which is V6+.
11069 if (!Subtarget->hasV6Ops())
11072 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
11073 std::string AsmStr = IA->getAsmString();
11074 SmallVector<StringRef, 4> AsmPieces;
11075 SplitString(AsmStr, AsmPieces, ";\n");
11077 switch (AsmPieces.size()) {
11078 default: return false;
11080 AsmStr = AsmPieces[0];
11082 SplitString(AsmStr, AsmPieces, " \t,");
11085 if (AsmPieces.size() == 3 &&
11086 AsmPieces[0] == "rev" && AsmPieces[1] == "$0" && AsmPieces[2] == "$1" &&
11087 IA->getConstraintString().compare(0, 4, "=l,l") == 0) {
11088 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
11089 if (Ty && Ty->getBitWidth() == 32)
11090 return IntrinsicLowering::LowerToByteSwap(CI);
11098 /// getConstraintType - Given a constraint letter, return the type of
11099 /// constraint it is for this target.
11100 ARMTargetLowering::ConstraintType
11101 ARMTargetLowering::getConstraintType(StringRef Constraint) const {
11102 if (Constraint.size() == 1) {
11103 switch (Constraint[0]) {
11105 case 'l': return C_RegisterClass;
11106 case 'w': return C_RegisterClass;
11107 case 'h': return C_RegisterClass;
11108 case 'x': return C_RegisterClass;
11109 case 't': return C_RegisterClass;
11110 case 'j': return C_Other; // Constant for movw.
11111 // An address with a single base register. Due to the way we
11112 // currently handle addresses it is the same as an 'r' memory constraint.
11113 case 'Q': return C_Memory;
11115 } else if (Constraint.size() == 2) {
11116 switch (Constraint[0]) {
11118 // All 'U+' constraints are addresses.
11119 case 'U': return C_Memory;
11122 return TargetLowering::getConstraintType(Constraint);
11125 /// Examine constraint type and operand type and determine a weight value.
11126 /// This object must already have been set up with the operand type
11127 /// and the current alternative constraint selected.
11128 TargetLowering::ConstraintWeight
11129 ARMTargetLowering::getSingleConstraintMatchWeight(
11130 AsmOperandInfo &info, const char *constraint) const {
11131 ConstraintWeight weight = CW_Invalid;
11132 Value *CallOperandVal = info.CallOperandVal;
11133 // If we don't have a value, we can't do a match,
11134 // but allow it at the lowest weight.
11135 if (!CallOperandVal)
11137 Type *type = CallOperandVal->getType();
11138 // Look at the constraint type.
11139 switch (*constraint) {
11141 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
11144 if (type->isIntegerTy()) {
11145 if (Subtarget->isThumb())
11146 weight = CW_SpecificReg;
11148 weight = CW_Register;
11152 if (type->isFloatingPointTy())
11153 weight = CW_Register;
11159 typedef std::pair<unsigned, const TargetRegisterClass*> RCPair;
11160 RCPair ARMTargetLowering::getRegForInlineAsmConstraint(
11161 const TargetRegisterInfo *TRI, StringRef Constraint, MVT VT) const {
11162 if (Constraint.size() == 1) {
11163 // GCC ARM Constraint Letters
11164 switch (Constraint[0]) {
11165 case 'l': // Low regs or general regs.
11166 if (Subtarget->isThumb())
11167 return RCPair(0U, &ARM::tGPRRegClass);
11168 return RCPair(0U, &ARM::GPRRegClass);
11169 case 'h': // High regs or no regs.
11170 if (Subtarget->isThumb())
11171 return RCPair(0U, &ARM::hGPRRegClass);
11174 if (Subtarget->isThumb1Only())
11175 return RCPair(0U, &ARM::tGPRRegClass);
11176 return RCPair(0U, &ARM::GPRRegClass);
11178 if (VT == MVT::Other)
11180 if (VT == MVT::f32)
11181 return RCPair(0U, &ARM::SPRRegClass);
11182 if (VT.getSizeInBits() == 64)
11183 return RCPair(0U, &ARM::DPRRegClass);
11184 if (VT.getSizeInBits() == 128)
11185 return RCPair(0U, &ARM::QPRRegClass);
11188 if (VT == MVT::Other)
11190 if (VT == MVT::f32)
11191 return RCPair(0U, &ARM::SPR_8RegClass);
11192 if (VT.getSizeInBits() == 64)
11193 return RCPair(0U, &ARM::DPR_8RegClass);
11194 if (VT.getSizeInBits() == 128)
11195 return RCPair(0U, &ARM::QPR_8RegClass);
11198 if (VT == MVT::f32)
11199 return RCPair(0U, &ARM::SPRRegClass);
11203 if (StringRef("{cc}").equals_lower(Constraint))
11204 return std::make_pair(unsigned(ARM::CPSR), &ARM::CCRRegClass);
11206 return TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
11209 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
11210 /// vector. If it is invalid, don't add anything to Ops.
11211 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
11212 std::string &Constraint,
11213 std::vector<SDValue>&Ops,
11214 SelectionDAG &DAG) const {
11217 // Currently only support length 1 constraints.
11218 if (Constraint.length() != 1) return;
11220 char ConstraintLetter = Constraint[0];
11221 switch (ConstraintLetter) {
11224 case 'I': case 'J': case 'K': case 'L':
11225 case 'M': case 'N': case 'O':
11226 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
11230 int64_t CVal64 = C->getSExtValue();
11231 int CVal = (int) CVal64;
11232 // None of these constraints allow values larger than 32 bits. Check
11233 // that the value fits in an int.
11234 if (CVal != CVal64)
11237 switch (ConstraintLetter) {
11239 // Constant suitable for movw, must be between 0 and
11241 if (Subtarget->hasV6T2Ops())
11242 if (CVal >= 0 && CVal <= 65535)
11246 if (Subtarget->isThumb1Only()) {
11247 // This must be a constant between 0 and 255, for ADD
11249 if (CVal >= 0 && CVal <= 255)
11251 } else if (Subtarget->isThumb2()) {
11252 // A constant that can be used as an immediate value in a
11253 // data-processing instruction.
11254 if (ARM_AM::getT2SOImmVal(CVal) != -1)
11257 // A constant that can be used as an immediate value in a
11258 // data-processing instruction.
11259 if (ARM_AM::getSOImmVal(CVal) != -1)
11265 if (Subtarget->isThumb()) { // FIXME thumb2
11266 // This must be a constant between -255 and -1, for negated ADD
11267 // immediates. This can be used in GCC with an "n" modifier that
11268 // prints the negated value, for use with SUB instructions. It is
11269 // not useful otherwise but is implemented for compatibility.
11270 if (CVal >= -255 && CVal <= -1)
11273 // This must be a constant between -4095 and 4095. It is not clear
11274 // what this constraint is intended for. Implemented for
11275 // compatibility with GCC.
11276 if (CVal >= -4095 && CVal <= 4095)
11282 if (Subtarget->isThumb1Only()) {
11283 // A 32-bit value where only one byte has a nonzero value. Exclude
11284 // zero to match GCC. This constraint is used by GCC internally for
11285 // constants that can be loaded with a move/shift combination.
11286 // It is not useful otherwise but is implemented for compatibility.
11287 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
11289 } else if (Subtarget->isThumb2()) {
11290 // A constant whose bitwise inverse can be used as an immediate
11291 // value in a data-processing instruction. This can be used in GCC
11292 // with a "B" modifier that prints the inverted value, for use with
11293 // BIC and MVN instructions. It is not useful otherwise but is
11294 // implemented for compatibility.
11295 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
11298 // A constant whose bitwise inverse can be used as an immediate
11299 // value in a data-processing instruction. This can be used in GCC
11300 // with a "B" modifier that prints the inverted value, for use with
11301 // BIC and MVN instructions. It is not useful otherwise but is
11302 // implemented for compatibility.
11303 if (ARM_AM::getSOImmVal(~CVal) != -1)
11309 if (Subtarget->isThumb1Only()) {
11310 // This must be a constant between -7 and 7,
11311 // for 3-operand ADD/SUB immediate instructions.
11312 if (CVal >= -7 && CVal < 7)
11314 } else if (Subtarget->isThumb2()) {
11315 // A constant whose negation can be used as an immediate value in a
11316 // data-processing instruction. This can be used in GCC with an "n"
11317 // modifier that prints the negated value, for use with SUB
11318 // instructions. It is not useful otherwise but is implemented for
11320 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
11323 // A constant whose negation can be used as an immediate value in a
11324 // data-processing instruction. This can be used in GCC with an "n"
11325 // modifier that prints the negated value, for use with SUB
11326 // instructions. It is not useful otherwise but is implemented for
11328 if (ARM_AM::getSOImmVal(-CVal) != -1)
11334 if (Subtarget->isThumb()) { // FIXME thumb2
11335 // This must be a multiple of 4 between 0 and 1020, for
11336 // ADD sp + immediate.
11337 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
11340 // A power of two or a constant between 0 and 32. This is used in
11341 // GCC for the shift amount on shifted register operands, but it is
11342 // useful in general for any shift amounts.
11343 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
11349 if (Subtarget->isThumb()) { // FIXME thumb2
11350 // This must be a constant between 0 and 31, for shift amounts.
11351 if (CVal >= 0 && CVal <= 31)
11357 if (Subtarget->isThumb()) { // FIXME thumb2
11358 // This must be a multiple of 4 between -508 and 508, for
11359 // ADD/SUB sp = sp + immediate.
11360 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
11365 Result = DAG.getTargetConstant(CVal, SDLoc(Op), Op.getValueType());
11369 if (Result.getNode()) {
11370 Ops.push_back(Result);
11373 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
11376 static RTLIB::Libcall getDivRemLibcall(
11377 const SDNode *N, MVT::SimpleValueType SVT) {
11378 assert((N->getOpcode() == ISD::SDIVREM || N->getOpcode() == ISD::UDIVREM ||
11379 N->getOpcode() == ISD::SREM || N->getOpcode() == ISD::UREM) &&
11380 "Unhandled Opcode in getDivRemLibcall");
11381 bool isSigned = N->getOpcode() == ISD::SDIVREM ||
11382 N->getOpcode() == ISD::SREM;
11385 default: llvm_unreachable("Unexpected request for libcall!");
11386 case MVT::i8: LC = isSigned ? RTLIB::SDIVREM_I8 : RTLIB::UDIVREM_I8; break;
11387 case MVT::i16: LC = isSigned ? RTLIB::SDIVREM_I16 : RTLIB::UDIVREM_I16; break;
11388 case MVT::i32: LC = isSigned ? RTLIB::SDIVREM_I32 : RTLIB::UDIVREM_I32; break;
11389 case MVT::i64: LC = isSigned ? RTLIB::SDIVREM_I64 : RTLIB::UDIVREM_I64; break;
11394 static TargetLowering::ArgListTy getDivRemArgList(
11395 const SDNode *N, LLVMContext *Context) {
11396 assert((N->getOpcode() == ISD::SDIVREM || N->getOpcode() == ISD::UDIVREM ||
11397 N->getOpcode() == ISD::SREM || N->getOpcode() == ISD::UREM) &&
11398 "Unhandled Opcode in getDivRemArgList");
11399 bool isSigned = N->getOpcode() == ISD::SDIVREM ||
11400 N->getOpcode() == ISD::SREM;
11401 TargetLowering::ArgListTy Args;
11402 TargetLowering::ArgListEntry Entry;
11403 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
11404 EVT ArgVT = N->getOperand(i).getValueType();
11405 Type *ArgTy = ArgVT.getTypeForEVT(*Context);
11406 Entry.Node = N->getOperand(i);
11408 Entry.isSExt = isSigned;
11409 Entry.isZExt = !isSigned;
11410 Args.push_back(Entry);
11415 SDValue ARMTargetLowering::LowerDivRem(SDValue Op, SelectionDAG &DAG) const {
11416 assert((Subtarget->isTargetAEABI() || Subtarget->isTargetAndroid()) &&
11417 "Register-based DivRem lowering only");
11418 unsigned Opcode = Op->getOpcode();
11419 assert((Opcode == ISD::SDIVREM || Opcode == ISD::UDIVREM) &&
11420 "Invalid opcode for Div/Rem lowering");
11421 bool isSigned = (Opcode == ISD::SDIVREM);
11422 EVT VT = Op->getValueType(0);
11423 Type *Ty = VT.getTypeForEVT(*DAG.getContext());
11425 RTLIB::Libcall LC = getDivRemLibcall(Op.getNode(),
11426 VT.getSimpleVT().SimpleTy);
11427 SDValue InChain = DAG.getEntryNode();
11429 TargetLowering::ArgListTy Args = getDivRemArgList(Op.getNode(),
11432 SDValue Callee = DAG.getExternalSymbol(getLibcallName(LC),
11433 getPointerTy(DAG.getDataLayout()));
11435 Type *RetTy = (Type*)StructType::get(Ty, Ty, nullptr);
11438 TargetLowering::CallLoweringInfo CLI(DAG);
11439 CLI.setDebugLoc(dl).setChain(InChain)
11440 .setCallee(getLibcallCallingConv(LC), RetTy, Callee, std::move(Args), 0)
11441 .setInRegister().setSExtResult(isSigned).setZExtResult(!isSigned);
11443 std::pair<SDValue, SDValue> CallInfo = LowerCallTo(CLI);
11444 return CallInfo.first;
11447 // Lowers REM using divmod helpers
11448 // see RTABI section 4.2/4.3
11449 SDValue ARMTargetLowering::LowerREM(SDNode *N, SelectionDAG &DAG) const {
11450 // Build return types (div and rem)
11451 std::vector<Type*> RetTyParams;
11452 Type *RetTyElement;
11454 switch (N->getValueType(0).getSimpleVT().SimpleTy) {
11455 default: llvm_unreachable("Unexpected request for libcall!");
11456 case MVT::i8: RetTyElement = Type::getInt8Ty(*DAG.getContext()); break;
11457 case MVT::i16: RetTyElement = Type::getInt16Ty(*DAG.getContext()); break;
11458 case MVT::i32: RetTyElement = Type::getInt32Ty(*DAG.getContext()); break;
11459 case MVT::i64: RetTyElement = Type::getInt64Ty(*DAG.getContext()); break;
11462 RetTyParams.push_back(RetTyElement);
11463 RetTyParams.push_back(RetTyElement);
11464 ArrayRef<Type*> ret = ArrayRef<Type*>(RetTyParams);
11465 Type *RetTy = StructType::get(*DAG.getContext(), ret);
11467 RTLIB::Libcall LC = getDivRemLibcall(N, N->getValueType(0).getSimpleVT().
11469 SDValue InChain = DAG.getEntryNode();
11470 TargetLowering::ArgListTy Args = getDivRemArgList(N, DAG.getContext());
11471 bool isSigned = N->getOpcode() == ISD::SREM;
11472 SDValue Callee = DAG.getExternalSymbol(getLibcallName(LC),
11473 getPointerTy(DAG.getDataLayout()));
11476 CallLoweringInfo CLI(DAG);
11477 CLI.setChain(InChain)
11478 .setCallee(CallingConv::ARM_AAPCS, RetTy, Callee, std::move(Args), 0)
11479 .setSExtResult(isSigned).setZExtResult(!isSigned).setDebugLoc(SDLoc(N));
11480 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
11482 // Return second (rem) result operand (first contains div)
11483 SDNode *ResNode = CallResult.first.getNode();
11484 assert(ResNode->getNumOperands() == 2 && "divmod should return two operands");
11485 return ResNode->getOperand(1);
11489 ARMTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const {
11490 assert(Subtarget->isTargetWindows() && "unsupported target platform");
11494 SDValue Chain = Op.getOperand(0);
11495 SDValue Size = Op.getOperand(1);
11497 SDValue Words = DAG.getNode(ISD::SRL, DL, MVT::i32, Size,
11498 DAG.getConstant(2, DL, MVT::i32));
11501 Chain = DAG.getCopyToReg(Chain, DL, ARM::R4, Words, Flag);
11502 Flag = Chain.getValue(1);
11504 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
11505 Chain = DAG.getNode(ARMISD::WIN__CHKSTK, DL, NodeTys, Chain, Flag);
11507 SDValue NewSP = DAG.getCopyFromReg(Chain, DL, ARM::SP, MVT::i32);
11508 Chain = NewSP.getValue(1);
11510 SDValue Ops[2] = { NewSP, Chain };
11511 return DAG.getMergeValues(Ops, DL);
11514 SDValue ARMTargetLowering::LowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) const {
11515 assert(Op.getValueType() == MVT::f64 && Subtarget->isFPOnlySP() &&
11516 "Unexpected type for custom-lowering FP_EXTEND");
11519 LC = RTLIB::getFPEXT(Op.getOperand(0).getValueType(), Op.getValueType());
11521 SDValue SrcVal = Op.getOperand(0);
11522 return makeLibCall(DAG, LC, Op.getValueType(), SrcVal, /*isSigned*/ false,
11526 SDValue ARMTargetLowering::LowerFP_ROUND(SDValue Op, SelectionDAG &DAG) const {
11527 assert(Op.getOperand(0).getValueType() == MVT::f64 &&
11528 Subtarget->isFPOnlySP() &&
11529 "Unexpected type for custom-lowering FP_ROUND");
11532 LC = RTLIB::getFPROUND(Op.getOperand(0).getValueType(), Op.getValueType());
11534 SDValue SrcVal = Op.getOperand(0);
11535 return makeLibCall(DAG, LC, Op.getValueType(), SrcVal, /*isSigned*/ false,
11540 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
11541 // The ARM target isn't yet aware of offsets.
11545 bool ARM::isBitFieldInvertedMask(unsigned v) {
11546 if (v == 0xffffffff)
11549 // there can be 1's on either or both "outsides", all the "inside"
11550 // bits must be 0's
11551 return isShiftedMask_32(~v);
11554 /// isFPImmLegal - Returns true if the target can instruction select the
11555 /// specified FP immediate natively. If false, the legalizer will
11556 /// materialize the FP immediate as a load from a constant pool.
11557 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
11558 if (!Subtarget->hasVFP3())
11560 if (VT == MVT::f32)
11561 return ARM_AM::getFP32Imm(Imm) != -1;
11562 if (VT == MVT::f64 && !Subtarget->isFPOnlySP())
11563 return ARM_AM::getFP64Imm(Imm) != -1;
11567 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
11568 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
11569 /// specified in the intrinsic calls.
11570 bool ARMTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
11572 unsigned Intrinsic) const {
11573 switch (Intrinsic) {
11574 case Intrinsic::arm_neon_vld1:
11575 case Intrinsic::arm_neon_vld2:
11576 case Intrinsic::arm_neon_vld3:
11577 case Intrinsic::arm_neon_vld4:
11578 case Intrinsic::arm_neon_vld2lane:
11579 case Intrinsic::arm_neon_vld3lane:
11580 case Intrinsic::arm_neon_vld4lane: {
11581 Info.opc = ISD::INTRINSIC_W_CHAIN;
11582 // Conservatively set memVT to the entire set of vectors loaded.
11583 auto &DL = I.getCalledFunction()->getParent()->getDataLayout();
11584 uint64_t NumElts = DL.getTypeAllocSize(I.getType()) / 8;
11585 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
11586 Info.ptrVal = I.getArgOperand(0);
11588 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
11589 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
11590 Info.vol = false; // volatile loads with NEON intrinsics not supported
11591 Info.readMem = true;
11592 Info.writeMem = false;
11595 case Intrinsic::arm_neon_vst1:
11596 case Intrinsic::arm_neon_vst2:
11597 case Intrinsic::arm_neon_vst3:
11598 case Intrinsic::arm_neon_vst4:
11599 case Intrinsic::arm_neon_vst2lane:
11600 case Intrinsic::arm_neon_vst3lane:
11601 case Intrinsic::arm_neon_vst4lane: {
11602 Info.opc = ISD::INTRINSIC_VOID;
11603 // Conservatively set memVT to the entire set of vectors stored.
11604 auto &DL = I.getCalledFunction()->getParent()->getDataLayout();
11605 unsigned NumElts = 0;
11606 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
11607 Type *ArgTy = I.getArgOperand(ArgI)->getType();
11608 if (!ArgTy->isVectorTy())
11610 NumElts += DL.getTypeAllocSize(ArgTy) / 8;
11612 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
11613 Info.ptrVal = I.getArgOperand(0);
11615 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
11616 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
11617 Info.vol = false; // volatile stores with NEON intrinsics not supported
11618 Info.readMem = false;
11619 Info.writeMem = true;
11622 case Intrinsic::arm_ldaex:
11623 case Intrinsic::arm_ldrex: {
11624 auto &DL = I.getCalledFunction()->getParent()->getDataLayout();
11625 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(0)->getType());
11626 Info.opc = ISD::INTRINSIC_W_CHAIN;
11627 Info.memVT = MVT::getVT(PtrTy->getElementType());
11628 Info.ptrVal = I.getArgOperand(0);
11630 Info.align = DL.getABITypeAlignment(PtrTy->getElementType());
11632 Info.readMem = true;
11633 Info.writeMem = false;
11636 case Intrinsic::arm_stlex:
11637 case Intrinsic::arm_strex: {
11638 auto &DL = I.getCalledFunction()->getParent()->getDataLayout();
11639 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(1)->getType());
11640 Info.opc = ISD::INTRINSIC_W_CHAIN;
11641 Info.memVT = MVT::getVT(PtrTy->getElementType());
11642 Info.ptrVal = I.getArgOperand(1);
11644 Info.align = DL.getABITypeAlignment(PtrTy->getElementType());
11646 Info.readMem = false;
11647 Info.writeMem = true;
11650 case Intrinsic::arm_stlexd:
11651 case Intrinsic::arm_strexd: {
11652 Info.opc = ISD::INTRINSIC_W_CHAIN;
11653 Info.memVT = MVT::i64;
11654 Info.ptrVal = I.getArgOperand(2);
11658 Info.readMem = false;
11659 Info.writeMem = true;
11662 case Intrinsic::arm_ldaexd:
11663 case Intrinsic::arm_ldrexd: {
11664 Info.opc = ISD::INTRINSIC_W_CHAIN;
11665 Info.memVT = MVT::i64;
11666 Info.ptrVal = I.getArgOperand(0);
11670 Info.readMem = true;
11671 Info.writeMem = false;
11681 /// \brief Returns true if it is beneficial to convert a load of a constant
11682 /// to just the constant itself.
11683 bool ARMTargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
11685 assert(Ty->isIntegerTy());
11687 unsigned Bits = Ty->getPrimitiveSizeInBits();
11688 if (Bits == 0 || Bits > 32)
11693 Instruction* ARMTargetLowering::makeDMB(IRBuilder<> &Builder,
11694 ARM_MB::MemBOpt Domain) const {
11695 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
11697 // First, if the target has no DMB, see what fallback we can use.
11698 if (!Subtarget->hasDataBarrier()) {
11699 // Some ARMv6 cpus can support data barriers with an mcr instruction.
11700 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
11702 if (Subtarget->hasV6Ops() && !Subtarget->isThumb()) {
11703 Function *MCR = llvm::Intrinsic::getDeclaration(M, Intrinsic::arm_mcr);
11704 Value* args[6] = {Builder.getInt32(15), Builder.getInt32(0),
11705 Builder.getInt32(0), Builder.getInt32(7),
11706 Builder.getInt32(10), Builder.getInt32(5)};
11707 return Builder.CreateCall(MCR, args);
11709 // Instead of using barriers, atomic accesses on these subtargets use
11711 llvm_unreachable("makeDMB on a target so old that it has no barriers");
11714 Function *DMB = llvm::Intrinsic::getDeclaration(M, Intrinsic::arm_dmb);
11715 // Only a full system barrier exists in the M-class architectures.
11716 Domain = Subtarget->isMClass() ? ARM_MB::SY : Domain;
11717 Constant *CDomain = Builder.getInt32(Domain);
11718 return Builder.CreateCall(DMB, CDomain);
11722 // Based on http://www.cl.cam.ac.uk/~pes20/cpp/cpp0xmappings.html
11723 Instruction* ARMTargetLowering::emitLeadingFence(IRBuilder<> &Builder,
11724 AtomicOrdering Ord, bool IsStore,
11725 bool IsLoad) const {
11726 if (!getInsertFencesForAtomic())
11732 llvm_unreachable("Invalid fence: unordered/non-atomic");
11735 return nullptr; // Nothing to do
11736 case SequentiallyConsistent:
11738 return nullptr; // Nothing to do
11741 case AcquireRelease:
11742 if (Subtarget->isSwift())
11743 return makeDMB(Builder, ARM_MB::ISHST);
11744 // FIXME: add a comment with a link to documentation justifying this.
11746 return makeDMB(Builder, ARM_MB::ISH);
11748 llvm_unreachable("Unknown fence ordering in emitLeadingFence");
11751 Instruction* ARMTargetLowering::emitTrailingFence(IRBuilder<> &Builder,
11752 AtomicOrdering Ord, bool IsStore,
11753 bool IsLoad) const {
11754 if (!getInsertFencesForAtomic())
11760 llvm_unreachable("Invalid fence: unordered/not-atomic");
11763 return nullptr; // Nothing to do
11765 case AcquireRelease:
11766 case SequentiallyConsistent:
11767 return makeDMB(Builder, ARM_MB::ISH);
11769 llvm_unreachable("Unknown fence ordering in emitTrailingFence");
11772 // Loads and stores less than 64-bits are already atomic; ones above that
11773 // are doomed anyway, so defer to the default libcall and blame the OS when
11774 // things go wrong. Cortex M doesn't have ldrexd/strexd though, so don't emit
11775 // anything for those.
11776 bool ARMTargetLowering::shouldExpandAtomicStoreInIR(StoreInst *SI) const {
11777 unsigned Size = SI->getValueOperand()->getType()->getPrimitiveSizeInBits();
11778 return (Size == 64) && !Subtarget->isMClass();
11781 // Loads and stores less than 64-bits are already atomic; ones above that
11782 // are doomed anyway, so defer to the default libcall and blame the OS when
11783 // things go wrong. Cortex M doesn't have ldrexd/strexd though, so don't emit
11784 // anything for those.
11785 // FIXME: ldrd and strd are atomic if the CPU has LPAE (e.g. A15 has that
11786 // guarantee, see DDI0406C ARM architecture reference manual,
11787 // sections A8.8.72-74 LDRD)
11788 TargetLowering::AtomicExpansionKind
11789 ARMTargetLowering::shouldExpandAtomicLoadInIR(LoadInst *LI) const {
11790 unsigned Size = LI->getType()->getPrimitiveSizeInBits();
11791 return ((Size == 64) && !Subtarget->isMClass()) ? AtomicExpansionKind::LLSC
11792 : AtomicExpansionKind::None;
11795 // For the real atomic operations, we have ldrex/strex up to 32 bits,
11796 // and up to 64 bits on the non-M profiles
11797 TargetLowering::AtomicExpansionKind
11798 ARMTargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
11799 unsigned Size = AI->getType()->getPrimitiveSizeInBits();
11800 return (Size <= (Subtarget->isMClass() ? 32U : 64U))
11801 ? AtomicExpansionKind::LLSC
11802 : AtomicExpansionKind::None;
11805 bool ARMTargetLowering::shouldExpandAtomicCmpXchgInIR(
11806 AtomicCmpXchgInst *AI) const {
11810 // This has so far only been implemented for MachO.
11811 bool ARMTargetLowering::useLoadStackGuardNode() const {
11812 return Subtarget->isTargetMachO();
11815 bool ARMTargetLowering::canCombineStoreAndExtract(Type *VectorTy, Value *Idx,
11816 unsigned &Cost) const {
11817 // If we do not have NEON, vector types are not natively supported.
11818 if (!Subtarget->hasNEON())
11821 // Floating point values and vector values map to the same register file.
11822 // Therefore, although we could do a store extract of a vector type, this is
11823 // better to leave at float as we have more freedom in the addressing mode for
11825 if (VectorTy->isFPOrFPVectorTy())
11828 // If the index is unknown at compile time, this is very expensive to lower
11829 // and it is not possible to combine the store with the extract.
11830 if (!isa<ConstantInt>(Idx))
11833 assert(VectorTy->isVectorTy() && "VectorTy is not a vector type");
11834 unsigned BitWidth = cast<VectorType>(VectorTy)->getBitWidth();
11835 // We can do a store + vector extract on any vector that fits perfectly in a D
11837 if (BitWidth == 64 || BitWidth == 128) {
11844 Value *ARMTargetLowering::emitLoadLinked(IRBuilder<> &Builder, Value *Addr,
11845 AtomicOrdering Ord) const {
11846 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
11847 Type *ValTy = cast<PointerType>(Addr->getType())->getElementType();
11848 bool IsAcquire = isAtLeastAcquire(Ord);
11850 // Since i64 isn't legal and intrinsics don't get type-lowered, the ldrexd
11851 // intrinsic must return {i32, i32} and we have to recombine them into a
11852 // single i64 here.
11853 if (ValTy->getPrimitiveSizeInBits() == 64) {
11854 Intrinsic::ID Int =
11855 IsAcquire ? Intrinsic::arm_ldaexd : Intrinsic::arm_ldrexd;
11856 Function *Ldrex = llvm::Intrinsic::getDeclaration(M, Int);
11858 Addr = Builder.CreateBitCast(Addr, Type::getInt8PtrTy(M->getContext()));
11859 Value *LoHi = Builder.CreateCall(Ldrex, Addr, "lohi");
11861 Value *Lo = Builder.CreateExtractValue(LoHi, 0, "lo");
11862 Value *Hi = Builder.CreateExtractValue(LoHi, 1, "hi");
11863 if (!Subtarget->isLittle())
11864 std::swap (Lo, Hi);
11865 Lo = Builder.CreateZExt(Lo, ValTy, "lo64");
11866 Hi = Builder.CreateZExt(Hi, ValTy, "hi64");
11867 return Builder.CreateOr(
11868 Lo, Builder.CreateShl(Hi, ConstantInt::get(ValTy, 32)), "val64");
11871 Type *Tys[] = { Addr->getType() };
11872 Intrinsic::ID Int = IsAcquire ? Intrinsic::arm_ldaex : Intrinsic::arm_ldrex;
11873 Function *Ldrex = llvm::Intrinsic::getDeclaration(M, Int, Tys);
11875 return Builder.CreateTruncOrBitCast(
11876 Builder.CreateCall(Ldrex, Addr),
11877 cast<PointerType>(Addr->getType())->getElementType());
11880 void ARMTargetLowering::emitAtomicCmpXchgNoStoreLLBalance(
11881 IRBuilder<> &Builder) const {
11882 if (!Subtarget->hasV7Ops())
11884 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
11885 Builder.CreateCall(llvm::Intrinsic::getDeclaration(M, Intrinsic::arm_clrex));
11888 Value *ARMTargetLowering::emitStoreConditional(IRBuilder<> &Builder, Value *Val,
11890 AtomicOrdering Ord) const {
11891 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
11892 bool IsRelease = isAtLeastRelease(Ord);
11894 // Since the intrinsics must have legal type, the i64 intrinsics take two
11895 // parameters: "i32, i32". We must marshal Val into the appropriate form
11896 // before the call.
11897 if (Val->getType()->getPrimitiveSizeInBits() == 64) {
11898 Intrinsic::ID Int =
11899 IsRelease ? Intrinsic::arm_stlexd : Intrinsic::arm_strexd;
11900 Function *Strex = Intrinsic::getDeclaration(M, Int);
11901 Type *Int32Ty = Type::getInt32Ty(M->getContext());
11903 Value *Lo = Builder.CreateTrunc(Val, Int32Ty, "lo");
11904 Value *Hi = Builder.CreateTrunc(Builder.CreateLShr(Val, 32), Int32Ty, "hi");
11905 if (!Subtarget->isLittle())
11906 std::swap (Lo, Hi);
11907 Addr = Builder.CreateBitCast(Addr, Type::getInt8PtrTy(M->getContext()));
11908 return Builder.CreateCall(Strex, {Lo, Hi, Addr});
11911 Intrinsic::ID Int = IsRelease ? Intrinsic::arm_stlex : Intrinsic::arm_strex;
11912 Type *Tys[] = { Addr->getType() };
11913 Function *Strex = Intrinsic::getDeclaration(M, Int, Tys);
11915 return Builder.CreateCall(
11916 Strex, {Builder.CreateZExtOrBitCast(
11917 Val, Strex->getFunctionType()->getParamType(0)),
11921 /// \brief Lower an interleaved load into a vldN intrinsic.
11923 /// E.g. Lower an interleaved load (Factor = 2):
11924 /// %wide.vec = load <8 x i32>, <8 x i32>* %ptr, align 4
11925 /// %v0 = shuffle %wide.vec, undef, <0, 2, 4, 6> ; Extract even elements
11926 /// %v1 = shuffle %wide.vec, undef, <1, 3, 5, 7> ; Extract odd elements
11929 /// %vld2 = { <4 x i32>, <4 x i32> } call llvm.arm.neon.vld2(%ptr, 4)
11930 /// %vec0 = extractelement { <4 x i32>, <4 x i32> } %vld2, i32 0
11931 /// %vec1 = extractelement { <4 x i32>, <4 x i32> } %vld2, i32 1
11932 bool ARMTargetLowering::lowerInterleavedLoad(
11933 LoadInst *LI, ArrayRef<ShuffleVectorInst *> Shuffles,
11934 ArrayRef<unsigned> Indices, unsigned Factor) const {
11935 assert(Factor >= 2 && Factor <= getMaxSupportedInterleaveFactor() &&
11936 "Invalid interleave factor");
11937 assert(!Shuffles.empty() && "Empty shufflevector input");
11938 assert(Shuffles.size() == Indices.size() &&
11939 "Unmatched number of shufflevectors and indices");
11941 VectorType *VecTy = Shuffles[0]->getType();
11942 Type *EltTy = VecTy->getVectorElementType();
11944 const DataLayout &DL = LI->getModule()->getDataLayout();
11945 unsigned VecSize = DL.getTypeAllocSizeInBits(VecTy);
11946 bool EltIs64Bits = DL.getTypeAllocSizeInBits(EltTy) == 64;
11948 // Skip if we do not have NEON and skip illegal vector types and vector types
11949 // with i64/f64 elements (vldN doesn't support i64/f64 elements).
11950 if (!Subtarget->hasNEON() || (VecSize != 64 && VecSize != 128) || EltIs64Bits)
11953 // A pointer vector can not be the return type of the ldN intrinsics. Need to
11954 // load integer vectors first and then convert to pointer vectors.
11955 if (EltTy->isPointerTy())
11957 VectorType::get(DL.getIntPtrType(EltTy), VecTy->getVectorNumElements());
11959 static const Intrinsic::ID LoadInts[3] = {Intrinsic::arm_neon_vld2,
11960 Intrinsic::arm_neon_vld3,
11961 Intrinsic::arm_neon_vld4};
11963 IRBuilder<> Builder(LI);
11964 SmallVector<Value *, 2> Ops;
11966 Type *Int8Ptr = Builder.getInt8PtrTy(LI->getPointerAddressSpace());
11967 Ops.push_back(Builder.CreateBitCast(LI->getPointerOperand(), Int8Ptr));
11968 Ops.push_back(Builder.getInt32(LI->getAlignment()));
11970 Type *Tys[] = { VecTy, Int8Ptr };
11971 Function *VldnFunc =
11972 Intrinsic::getDeclaration(LI->getModule(), LoadInts[Factor - 2], Tys);
11973 CallInst *VldN = Builder.CreateCall(VldnFunc, Ops, "vldN");
11975 // Replace uses of each shufflevector with the corresponding vector loaded
11977 for (unsigned i = 0; i < Shuffles.size(); i++) {
11978 ShuffleVectorInst *SV = Shuffles[i];
11979 unsigned Index = Indices[i];
11981 Value *SubVec = Builder.CreateExtractValue(VldN, Index);
11983 // Convert the integer vector to pointer vector if the element is pointer.
11984 if (EltTy->isPointerTy())
11985 SubVec = Builder.CreateIntToPtr(SubVec, SV->getType());
11987 SV->replaceAllUsesWith(SubVec);
11993 /// \brief Get a mask consisting of sequential integers starting from \p Start.
11995 /// I.e. <Start, Start + 1, ..., Start + NumElts - 1>
11996 static Constant *getSequentialMask(IRBuilder<> &Builder, unsigned Start,
11997 unsigned NumElts) {
11998 SmallVector<Constant *, 16> Mask;
11999 for (unsigned i = 0; i < NumElts; i++)
12000 Mask.push_back(Builder.getInt32(Start + i));
12002 return ConstantVector::get(Mask);
12005 /// \brief Lower an interleaved store into a vstN intrinsic.
12007 /// E.g. Lower an interleaved store (Factor = 3):
12008 /// %i.vec = shuffle <8 x i32> %v0, <8 x i32> %v1,
12009 /// <0, 4, 8, 1, 5, 9, 2, 6, 10, 3, 7, 11>
12010 /// store <12 x i32> %i.vec, <12 x i32>* %ptr, align 4
12013 /// %sub.v0 = shuffle <8 x i32> %v0, <8 x i32> v1, <0, 1, 2, 3>
12014 /// %sub.v1 = shuffle <8 x i32> %v0, <8 x i32> v1, <4, 5, 6, 7>
12015 /// %sub.v2 = shuffle <8 x i32> %v0, <8 x i32> v1, <8, 9, 10, 11>
12016 /// call void llvm.arm.neon.vst3(%ptr, %sub.v0, %sub.v1, %sub.v2, 4)
12018 /// Note that the new shufflevectors will be removed and we'll only generate one
12019 /// vst3 instruction in CodeGen.
12020 bool ARMTargetLowering::lowerInterleavedStore(StoreInst *SI,
12021 ShuffleVectorInst *SVI,
12022 unsigned Factor) const {
12023 assert(Factor >= 2 && Factor <= getMaxSupportedInterleaveFactor() &&
12024 "Invalid interleave factor");
12026 VectorType *VecTy = SVI->getType();
12027 assert(VecTy->getVectorNumElements() % Factor == 0 &&
12028 "Invalid interleaved store");
12030 unsigned NumSubElts = VecTy->getVectorNumElements() / Factor;
12031 Type *EltTy = VecTy->getVectorElementType();
12032 VectorType *SubVecTy = VectorType::get(EltTy, NumSubElts);
12034 const DataLayout &DL = SI->getModule()->getDataLayout();
12035 unsigned SubVecSize = DL.getTypeAllocSizeInBits(SubVecTy);
12036 bool EltIs64Bits = DL.getTypeAllocSizeInBits(EltTy) == 64;
12038 // Skip if we do not have NEON and skip illegal vector types and vector types
12039 // with i64/f64 elements (vstN doesn't support i64/f64 elements).
12040 if (!Subtarget->hasNEON() || (SubVecSize != 64 && SubVecSize != 128) ||
12044 Value *Op0 = SVI->getOperand(0);
12045 Value *Op1 = SVI->getOperand(1);
12046 IRBuilder<> Builder(SI);
12048 // StN intrinsics don't support pointer vectors as arguments. Convert pointer
12049 // vectors to integer vectors.
12050 if (EltTy->isPointerTy()) {
12051 Type *IntTy = DL.getIntPtrType(EltTy);
12053 // Convert to the corresponding integer vector.
12055 VectorType::get(IntTy, Op0->getType()->getVectorNumElements());
12056 Op0 = Builder.CreatePtrToInt(Op0, IntVecTy);
12057 Op1 = Builder.CreatePtrToInt(Op1, IntVecTy);
12059 SubVecTy = VectorType::get(IntTy, NumSubElts);
12062 static const Intrinsic::ID StoreInts[3] = {Intrinsic::arm_neon_vst2,
12063 Intrinsic::arm_neon_vst3,
12064 Intrinsic::arm_neon_vst4};
12065 SmallVector<Value *, 6> Ops;
12067 Type *Int8Ptr = Builder.getInt8PtrTy(SI->getPointerAddressSpace());
12068 Ops.push_back(Builder.CreateBitCast(SI->getPointerOperand(), Int8Ptr));
12070 Type *Tys[] = { Int8Ptr, SubVecTy };
12071 Function *VstNFunc = Intrinsic::getDeclaration(
12072 SI->getModule(), StoreInts[Factor - 2], Tys);
12074 // Split the shufflevector operands into sub vectors for the new vstN call.
12075 for (unsigned i = 0; i < Factor; i++)
12076 Ops.push_back(Builder.CreateShuffleVector(
12077 Op0, Op1, getSequentialMask(Builder, NumSubElts * i, NumSubElts)));
12079 Ops.push_back(Builder.getInt32(SI->getAlignment()));
12080 Builder.CreateCall(VstNFunc, Ops);
12092 static bool isHomogeneousAggregate(Type *Ty, HABaseType &Base,
12093 uint64_t &Members) {
12094 if (auto *ST = dyn_cast<StructType>(Ty)) {
12095 for (unsigned i = 0; i < ST->getNumElements(); ++i) {
12096 uint64_t SubMembers = 0;
12097 if (!isHomogeneousAggregate(ST->getElementType(i), Base, SubMembers))
12099 Members += SubMembers;
12101 } else if (auto *AT = dyn_cast<ArrayType>(Ty)) {
12102 uint64_t SubMembers = 0;
12103 if (!isHomogeneousAggregate(AT->getElementType(), Base, SubMembers))
12105 Members += SubMembers * AT->getNumElements();
12106 } else if (Ty->isFloatTy()) {
12107 if (Base != HA_UNKNOWN && Base != HA_FLOAT)
12111 } else if (Ty->isDoubleTy()) {
12112 if (Base != HA_UNKNOWN && Base != HA_DOUBLE)
12116 } else if (auto *VT = dyn_cast<VectorType>(Ty)) {
12123 return VT->getBitWidth() == 64;
12125 return VT->getBitWidth() == 128;
12127 switch (VT->getBitWidth()) {
12140 return (Members > 0 && Members <= 4);
12143 /// \brief Return true if a type is an AAPCS-VFP homogeneous aggregate or one of
12144 /// [N x i32] or [N x i64]. This allows front-ends to skip emitting padding when
12145 /// passing according to AAPCS rules.
12146 bool ARMTargetLowering::functionArgumentNeedsConsecutiveRegisters(
12147 Type *Ty, CallingConv::ID CallConv, bool isVarArg) const {
12148 if (getEffectiveCallingConv(CallConv, isVarArg) !=
12149 CallingConv::ARM_AAPCS_VFP)
12152 HABaseType Base = HA_UNKNOWN;
12153 uint64_t Members = 0;
12154 bool IsHA = isHomogeneousAggregate(Ty, Base, Members);
12155 DEBUG(dbgs() << "isHA: " << IsHA << " "; Ty->dump());
12157 bool IsIntArray = Ty->isArrayTy() && Ty->getArrayElementType()->isIntegerTy();
12158 return IsHA || IsIntArray;
12161 unsigned ARMTargetLowering::getExceptionPointerRegister(
12162 const Constant *PersonalityFn) const {
12163 // Platforms which do not use SjLj EH may return values in these registers
12164 // via the personality function.
12165 return Subtarget->useSjLjEH() ? ARM::NoRegister : ARM::R0;
12168 unsigned ARMTargetLowering::getExceptionSelectorRegister(
12169 const Constant *PersonalityFn) const {
12170 // Platforms which do not use SjLj EH may return values in these registers
12171 // via the personality function.
12172 return Subtarget->useSjLjEH() ? ARM::NoRegister : ARM::R1;