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 },
398 { RTLIB::SDIV_I32, "__rt_sdiv", CallingConv::ARM_AAPCS_VFP },
399 { RTLIB::SDIV_I64, "__rt_sdiv64", CallingConv::ARM_AAPCS_VFP },
402 for (const auto &LC : LibraryCalls) {
403 setLibcallName(LC.Op, LC.Name);
404 setLibcallCallingConv(LC.Op, LC.CC);
408 // Use divmod compiler-rt calls for iOS 5.0 and later.
409 if (Subtarget->isTargetWatchOS() ||
410 (Subtarget->isTargetIOS() &&
411 !Subtarget->getTargetTriple().isOSVersionLT(5, 0))) {
412 setLibcallName(RTLIB::SDIVREM_I32, "__divmodsi4");
413 setLibcallName(RTLIB::UDIVREM_I32, "__udivmodsi4");
416 // The half <-> float conversion functions are always soft-float, but are
417 // needed for some targets which use a hard-float calling convention by
419 if (Subtarget->isAAPCS_ABI()) {
420 setLibcallCallingConv(RTLIB::FPROUND_F32_F16, CallingConv::ARM_AAPCS);
421 setLibcallCallingConv(RTLIB::FPROUND_F64_F16, CallingConv::ARM_AAPCS);
422 setLibcallCallingConv(RTLIB::FPEXT_F16_F32, CallingConv::ARM_AAPCS);
424 setLibcallCallingConv(RTLIB::FPROUND_F32_F16, CallingConv::ARM_APCS);
425 setLibcallCallingConv(RTLIB::FPROUND_F64_F16, CallingConv::ARM_APCS);
426 setLibcallCallingConv(RTLIB::FPEXT_F16_F32, CallingConv::ARM_APCS);
429 // In EABI, these functions have an __aeabi_ prefix, but in GNUEABI they have
430 // a __gnu_ prefix (which is the default).
431 if (Subtarget->isTargetAEABI()) {
432 setLibcallName(RTLIB::FPROUND_F32_F16, "__aeabi_f2h");
433 setLibcallName(RTLIB::FPROUND_F64_F16, "__aeabi_d2h");
434 setLibcallName(RTLIB::FPEXT_F16_F32, "__aeabi_h2f");
437 if (Subtarget->isThumb1Only())
438 addRegisterClass(MVT::i32, &ARM::tGPRRegClass);
440 addRegisterClass(MVT::i32, &ARM::GPRRegClass);
441 if (!Subtarget->useSoftFloat() && Subtarget->hasVFP2() &&
442 !Subtarget->isThumb1Only()) {
443 addRegisterClass(MVT::f32, &ARM::SPRRegClass);
444 addRegisterClass(MVT::f64, &ARM::DPRRegClass);
447 for (MVT VT : MVT::vector_valuetypes()) {
448 for (MVT InnerVT : MVT::vector_valuetypes()) {
449 setTruncStoreAction(VT, InnerVT, Expand);
450 setLoadExtAction(ISD::SEXTLOAD, VT, InnerVT, Expand);
451 setLoadExtAction(ISD::ZEXTLOAD, VT, InnerVT, Expand);
452 setLoadExtAction(ISD::EXTLOAD, VT, InnerVT, Expand);
455 setOperationAction(ISD::MULHS, VT, Expand);
456 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
457 setOperationAction(ISD::MULHU, VT, Expand);
458 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
460 setOperationAction(ISD::BSWAP, VT, Expand);
463 setOperationAction(ISD::ConstantFP, MVT::f32, Custom);
464 setOperationAction(ISD::ConstantFP, MVT::f64, Custom);
466 setOperationAction(ISD::READ_REGISTER, MVT::i64, Custom);
467 setOperationAction(ISD::WRITE_REGISTER, MVT::i64, Custom);
469 if (Subtarget->hasNEON()) {
470 addDRTypeForNEON(MVT::v2f32);
471 addDRTypeForNEON(MVT::v8i8);
472 addDRTypeForNEON(MVT::v4i16);
473 addDRTypeForNEON(MVT::v2i32);
474 addDRTypeForNEON(MVT::v1i64);
476 addQRTypeForNEON(MVT::v4f32);
477 addQRTypeForNEON(MVT::v2f64);
478 addQRTypeForNEON(MVT::v16i8);
479 addQRTypeForNEON(MVT::v8i16);
480 addQRTypeForNEON(MVT::v4i32);
481 addQRTypeForNEON(MVT::v2i64);
483 // v2f64 is legal so that QR subregs can be extracted as f64 elements, but
484 // neither Neon nor VFP support any arithmetic operations on it.
485 // The same with v4f32. But keep in mind that vadd, vsub, vmul are natively
486 // supported for v4f32.
487 setOperationAction(ISD::FADD, MVT::v2f64, Expand);
488 setOperationAction(ISD::FSUB, MVT::v2f64, Expand);
489 setOperationAction(ISD::FMUL, MVT::v2f64, Expand);
490 // FIXME: Code duplication: FDIV and FREM are expanded always, see
491 // ARMTargetLowering::addTypeForNEON method for details.
492 setOperationAction(ISD::FDIV, MVT::v2f64, Expand);
493 setOperationAction(ISD::FREM, MVT::v2f64, Expand);
494 // FIXME: Create unittest.
495 // In another words, find a way when "copysign" appears in DAG with vector
497 setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand);
498 // FIXME: Code duplication: SETCC has custom operation action, see
499 // ARMTargetLowering::addTypeForNEON method for details.
500 setOperationAction(ISD::SETCC, MVT::v2f64, Expand);
501 // FIXME: Create unittest for FNEG and for FABS.
502 setOperationAction(ISD::FNEG, MVT::v2f64, Expand);
503 setOperationAction(ISD::FABS, MVT::v2f64, Expand);
504 setOperationAction(ISD::FSQRT, MVT::v2f64, Expand);
505 setOperationAction(ISD::FSIN, MVT::v2f64, Expand);
506 setOperationAction(ISD::FCOS, MVT::v2f64, Expand);
507 setOperationAction(ISD::FPOWI, MVT::v2f64, Expand);
508 setOperationAction(ISD::FPOW, MVT::v2f64, Expand);
509 setOperationAction(ISD::FLOG, MVT::v2f64, Expand);
510 setOperationAction(ISD::FLOG2, MVT::v2f64, Expand);
511 setOperationAction(ISD::FLOG10, MVT::v2f64, Expand);
512 setOperationAction(ISD::FEXP, MVT::v2f64, Expand);
513 setOperationAction(ISD::FEXP2, MVT::v2f64, Expand);
514 // FIXME: Create unittest for FCEIL, FTRUNC, FRINT, FNEARBYINT, FFLOOR.
515 setOperationAction(ISD::FCEIL, MVT::v2f64, Expand);
516 setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand);
517 setOperationAction(ISD::FRINT, MVT::v2f64, Expand);
518 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand);
519 setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand);
520 setOperationAction(ISD::FMA, MVT::v2f64, Expand);
522 setOperationAction(ISD::FSQRT, MVT::v4f32, Expand);
523 setOperationAction(ISD::FSIN, MVT::v4f32, Expand);
524 setOperationAction(ISD::FCOS, MVT::v4f32, Expand);
525 setOperationAction(ISD::FPOWI, MVT::v4f32, Expand);
526 setOperationAction(ISD::FPOW, MVT::v4f32, Expand);
527 setOperationAction(ISD::FLOG, MVT::v4f32, Expand);
528 setOperationAction(ISD::FLOG2, MVT::v4f32, Expand);
529 setOperationAction(ISD::FLOG10, MVT::v4f32, Expand);
530 setOperationAction(ISD::FEXP, MVT::v4f32, Expand);
531 setOperationAction(ISD::FEXP2, MVT::v4f32, Expand);
532 setOperationAction(ISD::FCEIL, MVT::v4f32, Expand);
533 setOperationAction(ISD::FTRUNC, MVT::v4f32, Expand);
534 setOperationAction(ISD::FRINT, MVT::v4f32, Expand);
535 setOperationAction(ISD::FNEARBYINT, MVT::v4f32, Expand);
536 setOperationAction(ISD::FFLOOR, MVT::v4f32, Expand);
538 // Mark v2f32 intrinsics.
539 setOperationAction(ISD::FSQRT, MVT::v2f32, Expand);
540 setOperationAction(ISD::FSIN, MVT::v2f32, Expand);
541 setOperationAction(ISD::FCOS, MVT::v2f32, Expand);
542 setOperationAction(ISD::FPOWI, MVT::v2f32, Expand);
543 setOperationAction(ISD::FPOW, MVT::v2f32, Expand);
544 setOperationAction(ISD::FLOG, MVT::v2f32, Expand);
545 setOperationAction(ISD::FLOG2, MVT::v2f32, Expand);
546 setOperationAction(ISD::FLOG10, MVT::v2f32, Expand);
547 setOperationAction(ISD::FEXP, MVT::v2f32, Expand);
548 setOperationAction(ISD::FEXP2, MVT::v2f32, Expand);
549 setOperationAction(ISD::FCEIL, MVT::v2f32, Expand);
550 setOperationAction(ISD::FTRUNC, MVT::v2f32, Expand);
551 setOperationAction(ISD::FRINT, MVT::v2f32, Expand);
552 setOperationAction(ISD::FNEARBYINT, MVT::v2f32, Expand);
553 setOperationAction(ISD::FFLOOR, MVT::v2f32, Expand);
555 // Neon does not support some operations on v1i64 and v2i64 types.
556 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
557 // Custom handling for some quad-vector types to detect VMULL.
558 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
559 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
560 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
561 // Custom handling for some vector types to avoid expensive expansions
562 setOperationAction(ISD::SDIV, MVT::v4i16, Custom);
563 setOperationAction(ISD::SDIV, MVT::v8i8, Custom);
564 setOperationAction(ISD::UDIV, MVT::v4i16, Custom);
565 setOperationAction(ISD::UDIV, MVT::v8i8, Custom);
566 setOperationAction(ISD::SETCC, MVT::v1i64, Expand);
567 setOperationAction(ISD::SETCC, MVT::v2i64, Expand);
568 // Neon does not have single instruction SINT_TO_FP and UINT_TO_FP with
569 // a destination type that is wider than the source, and nor does
570 // it have a FP_TO_[SU]INT instruction with a narrower destination than
572 setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Custom);
573 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
574 setOperationAction(ISD::FP_TO_UINT, MVT::v4i16, Custom);
575 setOperationAction(ISD::FP_TO_SINT, MVT::v4i16, Custom);
577 setOperationAction(ISD::FP_ROUND, MVT::v2f32, Expand);
578 setOperationAction(ISD::FP_EXTEND, MVT::v2f64, Expand);
580 // NEON does not have single instruction CTPOP for vectors with element
581 // types wider than 8-bits. However, custom lowering can leverage the
582 // v8i8/v16i8 vcnt instruction.
583 setOperationAction(ISD::CTPOP, MVT::v2i32, Custom);
584 setOperationAction(ISD::CTPOP, MVT::v4i32, Custom);
585 setOperationAction(ISD::CTPOP, MVT::v4i16, Custom);
586 setOperationAction(ISD::CTPOP, MVT::v8i16, Custom);
588 // NEON does not have single instruction CTTZ for vectors.
589 setOperationAction(ISD::CTTZ, MVT::v8i8, Custom);
590 setOperationAction(ISD::CTTZ, MVT::v4i16, Custom);
591 setOperationAction(ISD::CTTZ, MVT::v2i32, Custom);
592 setOperationAction(ISD::CTTZ, MVT::v1i64, Custom);
594 setOperationAction(ISD::CTTZ, MVT::v16i8, Custom);
595 setOperationAction(ISD::CTTZ, MVT::v8i16, Custom);
596 setOperationAction(ISD::CTTZ, MVT::v4i32, Custom);
597 setOperationAction(ISD::CTTZ, MVT::v2i64, Custom);
599 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v8i8, Custom);
600 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v4i16, Custom);
601 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v2i32, Custom);
602 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v1i64, Custom);
604 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v16i8, Custom);
605 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v8i16, Custom);
606 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v4i32, Custom);
607 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v2i64, Custom);
609 // NEON only has FMA instructions as of VFP4.
610 if (!Subtarget->hasVFP4()) {
611 setOperationAction(ISD::FMA, MVT::v2f32, Expand);
612 setOperationAction(ISD::FMA, MVT::v4f32, Expand);
615 setTargetDAGCombine(ISD::INTRINSIC_VOID);
616 setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN);
617 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
618 setTargetDAGCombine(ISD::SHL);
619 setTargetDAGCombine(ISD::SRL);
620 setTargetDAGCombine(ISD::SRA);
621 setTargetDAGCombine(ISD::SIGN_EXTEND);
622 setTargetDAGCombine(ISD::ZERO_EXTEND);
623 setTargetDAGCombine(ISD::ANY_EXTEND);
624 setTargetDAGCombine(ISD::BUILD_VECTOR);
625 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
626 setTargetDAGCombine(ISD::INSERT_VECTOR_ELT);
627 setTargetDAGCombine(ISD::STORE);
628 setTargetDAGCombine(ISD::FP_TO_SINT);
629 setTargetDAGCombine(ISD::FP_TO_UINT);
630 setTargetDAGCombine(ISD::FDIV);
631 setTargetDAGCombine(ISD::LOAD);
633 // It is legal to extload from v4i8 to v4i16 or v4i32.
634 for (MVT Ty : {MVT::v8i8, MVT::v4i8, MVT::v2i8, MVT::v4i16, MVT::v2i16,
636 for (MVT VT : MVT::integer_vector_valuetypes()) {
637 setLoadExtAction(ISD::EXTLOAD, VT, Ty, Legal);
638 setLoadExtAction(ISD::ZEXTLOAD, VT, Ty, Legal);
639 setLoadExtAction(ISD::SEXTLOAD, VT, Ty, Legal);
644 // ARM and Thumb2 support UMLAL/SMLAL.
645 if (!Subtarget->isThumb1Only())
646 setTargetDAGCombine(ISD::ADDC);
648 if (Subtarget->isFPOnlySP()) {
649 // When targeting a floating-point unit with only single-precision
650 // operations, f64 is legal for the few double-precision instructions which
651 // are present However, no double-precision operations other than moves,
652 // loads and stores are provided by the hardware.
653 setOperationAction(ISD::FADD, MVT::f64, Expand);
654 setOperationAction(ISD::FSUB, MVT::f64, Expand);
655 setOperationAction(ISD::FMUL, MVT::f64, Expand);
656 setOperationAction(ISD::FMA, MVT::f64, Expand);
657 setOperationAction(ISD::FDIV, MVT::f64, Expand);
658 setOperationAction(ISD::FREM, MVT::f64, Expand);
659 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
660 setOperationAction(ISD::FGETSIGN, MVT::f64, Expand);
661 setOperationAction(ISD::FNEG, MVT::f64, Expand);
662 setOperationAction(ISD::FABS, MVT::f64, Expand);
663 setOperationAction(ISD::FSQRT, MVT::f64, Expand);
664 setOperationAction(ISD::FSIN, MVT::f64, Expand);
665 setOperationAction(ISD::FCOS, MVT::f64, Expand);
666 setOperationAction(ISD::FPOWI, MVT::f64, Expand);
667 setOperationAction(ISD::FPOW, MVT::f64, Expand);
668 setOperationAction(ISD::FLOG, MVT::f64, Expand);
669 setOperationAction(ISD::FLOG2, MVT::f64, Expand);
670 setOperationAction(ISD::FLOG10, MVT::f64, Expand);
671 setOperationAction(ISD::FEXP, MVT::f64, Expand);
672 setOperationAction(ISD::FEXP2, MVT::f64, Expand);
673 setOperationAction(ISD::FCEIL, MVT::f64, Expand);
674 setOperationAction(ISD::FTRUNC, MVT::f64, Expand);
675 setOperationAction(ISD::FRINT, MVT::f64, Expand);
676 setOperationAction(ISD::FNEARBYINT, MVT::f64, Expand);
677 setOperationAction(ISD::FFLOOR, MVT::f64, Expand);
678 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
679 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
680 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
681 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
682 setOperationAction(ISD::FP_TO_SINT, MVT::f64, Custom);
683 setOperationAction(ISD::FP_TO_UINT, MVT::f64, Custom);
684 setOperationAction(ISD::FP_ROUND, MVT::f32, Custom);
685 setOperationAction(ISD::FP_EXTEND, MVT::f64, Custom);
688 computeRegisterProperties(Subtarget->getRegisterInfo());
690 // ARM does not have floating-point extending loads.
691 for (MVT VT : MVT::fp_valuetypes()) {
692 setLoadExtAction(ISD::EXTLOAD, VT, MVT::f32, Expand);
693 setLoadExtAction(ISD::EXTLOAD, VT, MVT::f16, Expand);
696 // ... or truncating stores
697 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
698 setTruncStoreAction(MVT::f32, MVT::f16, Expand);
699 setTruncStoreAction(MVT::f64, MVT::f16, Expand);
701 // ARM does not have i1 sign extending load.
702 for (MVT VT : MVT::integer_valuetypes())
703 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Promote);
705 // ARM supports all 4 flavors of integer indexed load / store.
706 if (!Subtarget->isThumb1Only()) {
707 for (unsigned im = (unsigned)ISD::PRE_INC;
708 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
709 setIndexedLoadAction(im, MVT::i1, Legal);
710 setIndexedLoadAction(im, MVT::i8, Legal);
711 setIndexedLoadAction(im, MVT::i16, Legal);
712 setIndexedLoadAction(im, MVT::i32, Legal);
713 setIndexedStoreAction(im, MVT::i1, Legal);
714 setIndexedStoreAction(im, MVT::i8, Legal);
715 setIndexedStoreAction(im, MVT::i16, Legal);
716 setIndexedStoreAction(im, MVT::i32, Legal);
720 setOperationAction(ISD::SADDO, MVT::i32, Custom);
721 setOperationAction(ISD::UADDO, MVT::i32, Custom);
722 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
723 setOperationAction(ISD::USUBO, MVT::i32, Custom);
725 // i64 operation support.
726 setOperationAction(ISD::MUL, MVT::i64, Expand);
727 setOperationAction(ISD::MULHU, MVT::i32, Expand);
728 if (Subtarget->isThumb1Only()) {
729 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
730 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
732 if (Subtarget->isThumb1Only() || !Subtarget->hasV6Ops()
733 || (Subtarget->isThumb2() && !Subtarget->hasDSP()))
734 setOperationAction(ISD::MULHS, MVT::i32, Expand);
736 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
737 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
738 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
739 setOperationAction(ISD::SRL, MVT::i64, Custom);
740 setOperationAction(ISD::SRA, MVT::i64, Custom);
742 if (!Subtarget->isThumb1Only()) {
743 // FIXME: We should do this for Thumb1 as well.
744 setOperationAction(ISD::ADDC, MVT::i32, Custom);
745 setOperationAction(ISD::ADDE, MVT::i32, Custom);
746 setOperationAction(ISD::SUBC, MVT::i32, Custom);
747 setOperationAction(ISD::SUBE, MVT::i32, Custom);
750 if (!Subtarget->isThumb1Only())
751 setOperationAction(ISD::BITREVERSE, MVT::i32, Legal);
753 // ARM does not have ROTL.
754 setOperationAction(ISD::ROTL, MVT::i32, Expand);
755 for (MVT VT : MVT::vector_valuetypes()) {
756 setOperationAction(ISD::ROTL, VT, Expand);
757 setOperationAction(ISD::ROTR, VT, Expand);
759 setOperationAction(ISD::CTTZ, MVT::i32, Custom);
760 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
761 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
762 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
764 // These just redirect to CTTZ and CTLZ on ARM.
765 setOperationAction(ISD::CTTZ_ZERO_UNDEF , MVT::i32 , Expand);
766 setOperationAction(ISD::CTLZ_ZERO_UNDEF , MVT::i32 , Expand);
768 // @llvm.readcyclecounter requires the Performance Monitors extension.
769 // Default to the 0 expansion on unsupported platforms.
770 // FIXME: Technically there are older ARM CPUs that have
771 // implementation-specific ways of obtaining this information.
772 if (Subtarget->hasPerfMon())
773 setOperationAction(ISD::READCYCLECOUNTER, MVT::i64, Custom);
775 // Only ARMv6 has BSWAP.
776 if (!Subtarget->hasV6Ops())
777 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
779 if (!(Subtarget->hasDivide() && Subtarget->isThumb2()) &&
780 !(Subtarget->hasDivideInARMMode() && !Subtarget->isThumb())) {
781 // These are expanded into libcalls if the cpu doesn't have HW divider.
782 setOperationAction(ISD::SDIV, MVT::i32, LibCall);
783 setOperationAction(ISD::UDIV, MVT::i32, LibCall);
786 if (Subtarget->isTargetWindows() && !Subtarget->hasDivide()) {
787 setOperationAction(ISD::UDIV, MVT::i32, Custom);
789 setOperationAction(ISD::UDIV, MVT::i64, Custom);
792 setOperationAction(ISD::SREM, MVT::i32, Expand);
793 setOperationAction(ISD::UREM, MVT::i32, Expand);
794 // Register based DivRem for AEABI (RTABI 4.2)
795 if (Subtarget->isTargetAEABI() || Subtarget->isTargetAndroid()) {
796 setOperationAction(ISD::SREM, MVT::i64, Custom);
797 setOperationAction(ISD::UREM, MVT::i64, Custom);
799 setLibcallName(RTLIB::SDIVREM_I8, "__aeabi_idivmod");
800 setLibcallName(RTLIB::SDIVREM_I16, "__aeabi_idivmod");
801 setLibcallName(RTLIB::SDIVREM_I32, "__aeabi_idivmod");
802 setLibcallName(RTLIB::SDIVREM_I64, "__aeabi_ldivmod");
803 setLibcallName(RTLIB::UDIVREM_I8, "__aeabi_uidivmod");
804 setLibcallName(RTLIB::UDIVREM_I16, "__aeabi_uidivmod");
805 setLibcallName(RTLIB::UDIVREM_I32, "__aeabi_uidivmod");
806 setLibcallName(RTLIB::UDIVREM_I64, "__aeabi_uldivmod");
808 setLibcallCallingConv(RTLIB::SDIVREM_I8, CallingConv::ARM_AAPCS);
809 setLibcallCallingConv(RTLIB::SDIVREM_I16, CallingConv::ARM_AAPCS);
810 setLibcallCallingConv(RTLIB::SDIVREM_I32, CallingConv::ARM_AAPCS);
811 setLibcallCallingConv(RTLIB::SDIVREM_I64, CallingConv::ARM_AAPCS);
812 setLibcallCallingConv(RTLIB::UDIVREM_I8, CallingConv::ARM_AAPCS);
813 setLibcallCallingConv(RTLIB::UDIVREM_I16, CallingConv::ARM_AAPCS);
814 setLibcallCallingConv(RTLIB::UDIVREM_I32, CallingConv::ARM_AAPCS);
815 setLibcallCallingConv(RTLIB::UDIVREM_I64, CallingConv::ARM_AAPCS);
817 setOperationAction(ISD::SDIVREM, MVT::i32, Custom);
818 setOperationAction(ISD::UDIVREM, MVT::i32, Custom);
820 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
821 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
824 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
825 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
826 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
827 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
829 setOperationAction(ISD::TRAP, MVT::Other, Legal);
831 // Use the default implementation.
832 setOperationAction(ISD::VASTART, MVT::Other, Custom);
833 setOperationAction(ISD::VAARG, MVT::Other, Expand);
834 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
835 setOperationAction(ISD::VAEND, MVT::Other, Expand);
836 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
837 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
839 if (Subtarget->getTargetTriple().isWindowsItaniumEnvironment())
840 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
842 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
844 // ARMv6 Thumb1 (except for CPUs that support dmb / dsb) and earlier use
845 // the default expansion. If we are targeting a single threaded system,
846 // then set them all for expand so we can lower them later into their
848 if (TM.Options.ThreadModel == ThreadModel::Single)
849 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Expand);
850 else if (Subtarget->hasAnyDataBarrier() && !Subtarget->isThumb1Only()) {
851 // ATOMIC_FENCE needs custom lowering; the others should have been expanded
852 // to ldrex/strex loops already.
853 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Custom);
855 // On v8, we have particularly efficient implementations of atomic fences
856 // if they can be combined with nearby atomic loads and stores.
857 if (!Subtarget->hasV8Ops()) {
858 // Automatically insert fences (dmb ish) around ATOMIC_SWAP etc.
859 setInsertFencesForAtomic(true);
862 // If there's anything we can use as a barrier, go through custom lowering
864 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other,
865 Subtarget->hasAnyDataBarrier() ? Custom : Expand);
867 // Set them all for expansion, which will force libcalls.
868 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Expand);
869 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Expand);
870 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Expand);
871 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Expand);
872 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Expand);
873 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Expand);
874 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Expand);
875 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Expand);
876 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i32, Expand);
877 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i32, Expand);
878 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i32, Expand);
879 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i32, Expand);
880 // Mark ATOMIC_LOAD and ATOMIC_STORE custom so we can handle the
881 // Unordered/Monotonic case.
882 setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Custom);
883 setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Custom);
886 setOperationAction(ISD::PREFETCH, MVT::Other, Custom);
888 // Requires SXTB/SXTH, available on v6 and up in both ARM and Thumb modes.
889 if (!Subtarget->hasV6Ops()) {
890 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
891 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
893 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
895 if (!Subtarget->useSoftFloat() && Subtarget->hasVFP2() &&
896 !Subtarget->isThumb1Only()) {
897 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
898 // iff target supports vfp2.
899 setOperationAction(ISD::BITCAST, MVT::i64, Custom);
900 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
903 // We want to custom lower some of our intrinsics.
904 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
905 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
906 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
907 setOperationAction(ISD::EH_SJLJ_SETUP_DISPATCH, MVT::Other, Custom);
908 if (Subtarget->useSjLjEH())
909 setLibcallName(RTLIB::UNWIND_RESUME, "_Unwind_SjLj_Resume");
911 setOperationAction(ISD::SETCC, MVT::i32, Expand);
912 setOperationAction(ISD::SETCC, MVT::f32, Expand);
913 setOperationAction(ISD::SETCC, MVT::f64, Expand);
914 setOperationAction(ISD::SELECT, MVT::i32, Custom);
915 setOperationAction(ISD::SELECT, MVT::f32, Custom);
916 setOperationAction(ISD::SELECT, MVT::f64, Custom);
917 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
918 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
919 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
921 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
922 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
923 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
924 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
925 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
927 // We don't support sin/cos/fmod/copysign/pow
928 setOperationAction(ISD::FSIN, MVT::f64, Expand);
929 setOperationAction(ISD::FSIN, MVT::f32, Expand);
930 setOperationAction(ISD::FCOS, MVT::f32, Expand);
931 setOperationAction(ISD::FCOS, MVT::f64, Expand);
932 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
933 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
934 setOperationAction(ISD::FREM, MVT::f64, Expand);
935 setOperationAction(ISD::FREM, MVT::f32, Expand);
936 if (!Subtarget->useSoftFloat() && Subtarget->hasVFP2() &&
937 !Subtarget->isThumb1Only()) {
938 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
939 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
941 setOperationAction(ISD::FPOW, MVT::f64, Expand);
942 setOperationAction(ISD::FPOW, MVT::f32, Expand);
944 if (!Subtarget->hasVFP4()) {
945 setOperationAction(ISD::FMA, MVT::f64, Expand);
946 setOperationAction(ISD::FMA, MVT::f32, Expand);
949 // Various VFP goodness
950 if (!Subtarget->useSoftFloat() && !Subtarget->isThumb1Only()) {
951 // FP-ARMv8 adds f64 <-> f16 conversion. Before that it should be expanded.
952 if (!Subtarget->hasFPARMv8() || Subtarget->isFPOnlySP()) {
953 setOperationAction(ISD::FP16_TO_FP, MVT::f64, Expand);
954 setOperationAction(ISD::FP_TO_FP16, MVT::f64, Expand);
957 // fp16 is a special v7 extension that adds f16 <-> f32 conversions.
958 if (!Subtarget->hasFP16()) {
959 setOperationAction(ISD::FP16_TO_FP, MVT::f32, Expand);
960 setOperationAction(ISD::FP_TO_FP16, MVT::f32, Expand);
964 // Combine sin / cos into one node or libcall if possible.
965 if (Subtarget->hasSinCos()) {
966 setLibcallName(RTLIB::SINCOS_F32, "sincosf");
967 setLibcallName(RTLIB::SINCOS_F64, "sincos");
968 if (Subtarget->isTargetWatchOS()) {
969 setLibcallCallingConv(RTLIB::SINCOS_F32, CallingConv::ARM_AAPCS_VFP);
970 setLibcallCallingConv(RTLIB::SINCOS_F64, CallingConv::ARM_AAPCS_VFP);
972 if (Subtarget->isTargetIOS() || Subtarget->isTargetWatchOS()) {
973 // For iOS, we don't want to the normal expansion of a libcall to
974 // sincos. We want to issue a libcall to __sincos_stret.
975 setOperationAction(ISD::FSINCOS, MVT::f64, Custom);
976 setOperationAction(ISD::FSINCOS, MVT::f32, Custom);
980 // FP-ARMv8 implements a lot of rounding-like FP operations.
981 if (Subtarget->hasFPARMv8()) {
982 setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
983 setOperationAction(ISD::FCEIL, MVT::f32, Legal);
984 setOperationAction(ISD::FROUND, MVT::f32, Legal);
985 setOperationAction(ISD::FTRUNC, MVT::f32, Legal);
986 setOperationAction(ISD::FNEARBYINT, MVT::f32, Legal);
987 setOperationAction(ISD::FRINT, MVT::f32, Legal);
988 setOperationAction(ISD::FMINNUM, MVT::f32, Legal);
989 setOperationAction(ISD::FMAXNUM, MVT::f32, Legal);
990 setOperationAction(ISD::FMINNUM, MVT::v2f32, Legal);
991 setOperationAction(ISD::FMAXNUM, MVT::v2f32, Legal);
992 setOperationAction(ISD::FMINNUM, MVT::v4f32, Legal);
993 setOperationAction(ISD::FMAXNUM, MVT::v4f32, Legal);
995 if (!Subtarget->isFPOnlySP()) {
996 setOperationAction(ISD::FFLOOR, MVT::f64, Legal);
997 setOperationAction(ISD::FCEIL, MVT::f64, Legal);
998 setOperationAction(ISD::FROUND, MVT::f64, Legal);
999 setOperationAction(ISD::FTRUNC, MVT::f64, Legal);
1000 setOperationAction(ISD::FNEARBYINT, MVT::f64, Legal);
1001 setOperationAction(ISD::FRINT, MVT::f64, Legal);
1002 setOperationAction(ISD::FMINNUM, MVT::f64, Legal);
1003 setOperationAction(ISD::FMAXNUM, MVT::f64, Legal);
1007 if (Subtarget->hasNEON()) {
1008 // vmin and vmax aren't available in a scalar form, so we use
1009 // a NEON instruction with an undef lane instead.
1010 setOperationAction(ISD::FMINNAN, MVT::f32, Legal);
1011 setOperationAction(ISD::FMAXNAN, MVT::f32, Legal);
1012 setOperationAction(ISD::FMINNAN, MVT::v2f32, Legal);
1013 setOperationAction(ISD::FMAXNAN, MVT::v2f32, Legal);
1014 setOperationAction(ISD::FMINNAN, MVT::v4f32, Legal);
1015 setOperationAction(ISD::FMAXNAN, MVT::v4f32, Legal);
1018 // We have target-specific dag combine patterns for the following nodes:
1019 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
1020 setTargetDAGCombine(ISD::ADD);
1021 setTargetDAGCombine(ISD::SUB);
1022 setTargetDAGCombine(ISD::MUL);
1023 setTargetDAGCombine(ISD::AND);
1024 setTargetDAGCombine(ISD::OR);
1025 setTargetDAGCombine(ISD::XOR);
1027 if (Subtarget->hasV6Ops())
1028 setTargetDAGCombine(ISD::SRL);
1030 setStackPointerRegisterToSaveRestore(ARM::SP);
1032 if (Subtarget->useSoftFloat() || Subtarget->isThumb1Only() ||
1033 !Subtarget->hasVFP2())
1034 setSchedulingPreference(Sched::RegPressure);
1036 setSchedulingPreference(Sched::Hybrid);
1038 //// temporary - rewrite interface to use type
1039 MaxStoresPerMemset = 8;
1040 MaxStoresPerMemsetOptSize = 4;
1041 MaxStoresPerMemcpy = 4; // For @llvm.memcpy -> sequence of stores
1042 MaxStoresPerMemcpyOptSize = 2;
1043 MaxStoresPerMemmove = 4; // For @llvm.memmove -> sequence of stores
1044 MaxStoresPerMemmoveOptSize = 2;
1046 // On ARM arguments smaller than 4 bytes are extended, so all arguments
1047 // are at least 4 bytes aligned.
1048 setMinStackArgumentAlignment(4);
1050 // Prefer likely predicted branches to selects on out-of-order cores.
1051 PredictableSelectIsExpensive = Subtarget->isLikeA9();
1053 setMinFunctionAlignment(Subtarget->isThumb() ? 1 : 2);
1056 bool ARMTargetLowering::useSoftFloat() const {
1057 return Subtarget->useSoftFloat();
1060 // FIXME: It might make sense to define the representative register class as the
1061 // nearest super-register that has a non-null superset. For example, DPR_VFP2 is
1062 // a super-register of SPR, and DPR is a superset if DPR_VFP2. Consequently,
1063 // SPR's representative would be DPR_VFP2. This should work well if register
1064 // pressure tracking were modified such that a register use would increment the
1065 // pressure of the register class's representative and all of it's super
1066 // classes' representatives transitively. We have not implemented this because
1067 // of the difficulty prior to coalescing of modeling operand register classes
1068 // due to the common occurrence of cross class copies and subregister insertions
1070 std::pair<const TargetRegisterClass *, uint8_t>
1071 ARMTargetLowering::findRepresentativeClass(const TargetRegisterInfo *TRI,
1073 const TargetRegisterClass *RRC = nullptr;
1075 switch (VT.SimpleTy) {
1077 return TargetLowering::findRepresentativeClass(TRI, VT);
1078 // Use DPR as representative register class for all floating point
1079 // and vector types. Since there are 32 SPR registers and 32 DPR registers so
1080 // the cost is 1 for both f32 and f64.
1081 case MVT::f32: case MVT::f64: case MVT::v8i8: case MVT::v4i16:
1082 case MVT::v2i32: case MVT::v1i64: case MVT::v2f32:
1083 RRC = &ARM::DPRRegClass;
1084 // When NEON is used for SP, only half of the register file is available
1085 // because operations that define both SP and DP results will be constrained
1086 // to the VFP2 class (D0-D15). We currently model this constraint prior to
1087 // coalescing by double-counting the SP regs. See the FIXME above.
1088 if (Subtarget->useNEONForSinglePrecisionFP())
1091 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
1092 case MVT::v4f32: case MVT::v2f64:
1093 RRC = &ARM::DPRRegClass;
1097 RRC = &ARM::DPRRegClass;
1101 RRC = &ARM::DPRRegClass;
1105 return std::make_pair(RRC, Cost);
1108 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
1109 switch ((ARMISD::NodeType)Opcode) {
1110 case ARMISD::FIRST_NUMBER: break;
1111 case ARMISD::Wrapper: return "ARMISD::Wrapper";
1112 case ARMISD::WrapperPIC: return "ARMISD::WrapperPIC";
1113 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
1114 case ARMISD::COPY_STRUCT_BYVAL: return "ARMISD::COPY_STRUCT_BYVAL";
1115 case ARMISD::CALL: return "ARMISD::CALL";
1116 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
1117 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
1118 case ARMISD::tCALL: return "ARMISD::tCALL";
1119 case ARMISD::BRCOND: return "ARMISD::BRCOND";
1120 case ARMISD::BR_JT: return "ARMISD::BR_JT";
1121 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
1122 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
1123 case ARMISD::INTRET_FLAG: return "ARMISD::INTRET_FLAG";
1124 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
1125 case ARMISD::CMP: return "ARMISD::CMP";
1126 case ARMISD::CMN: return "ARMISD::CMN";
1127 case ARMISD::CMPZ: return "ARMISD::CMPZ";
1128 case ARMISD::CMPFP: return "ARMISD::CMPFP";
1129 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
1130 case ARMISD::BCC_i64: return "ARMISD::BCC_i64";
1131 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
1133 case ARMISD::CMOV: return "ARMISD::CMOV";
1135 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
1136 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
1137 case ARMISD::RRX: return "ARMISD::RRX";
1139 case ARMISD::ADDC: return "ARMISD::ADDC";
1140 case ARMISD::ADDE: return "ARMISD::ADDE";
1141 case ARMISD::SUBC: return "ARMISD::SUBC";
1142 case ARMISD::SUBE: return "ARMISD::SUBE";
1144 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
1145 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
1147 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
1148 case ARMISD::EH_SJLJ_LONGJMP: return "ARMISD::EH_SJLJ_LONGJMP";
1149 case ARMISD::EH_SJLJ_SETUP_DISPATCH: return "ARMISD::EH_SJLJ_SETUP_DISPATCH";
1151 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN";
1153 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
1155 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
1157 case ARMISD::MEMBARRIER_MCR: return "ARMISD::MEMBARRIER_MCR";
1159 case ARMISD::PRELOAD: return "ARMISD::PRELOAD";
1161 case ARMISD::WIN__CHKSTK: return "ARMISD:::WIN__CHKSTK";
1162 case ARMISD::WIN__DBZCHK: return "ARMISD::WIN__DBZCHK";
1164 case ARMISD::VCEQ: return "ARMISD::VCEQ";
1165 case ARMISD::VCEQZ: return "ARMISD::VCEQZ";
1166 case ARMISD::VCGE: return "ARMISD::VCGE";
1167 case ARMISD::VCGEZ: return "ARMISD::VCGEZ";
1168 case ARMISD::VCLEZ: return "ARMISD::VCLEZ";
1169 case ARMISD::VCGEU: return "ARMISD::VCGEU";
1170 case ARMISD::VCGT: return "ARMISD::VCGT";
1171 case ARMISD::VCGTZ: return "ARMISD::VCGTZ";
1172 case ARMISD::VCLTZ: return "ARMISD::VCLTZ";
1173 case ARMISD::VCGTU: return "ARMISD::VCGTU";
1174 case ARMISD::VTST: return "ARMISD::VTST";
1176 case ARMISD::VSHL: return "ARMISD::VSHL";
1177 case ARMISD::VSHRs: return "ARMISD::VSHRs";
1178 case ARMISD::VSHRu: return "ARMISD::VSHRu";
1179 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
1180 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
1181 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
1182 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
1183 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
1184 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
1185 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
1186 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
1187 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
1188 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
1189 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
1190 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
1191 case ARMISD::VSLI: return "ARMISD::VSLI";
1192 case ARMISD::VSRI: return "ARMISD::VSRI";
1193 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
1194 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
1195 case ARMISD::VMOVIMM: return "ARMISD::VMOVIMM";
1196 case ARMISD::VMVNIMM: return "ARMISD::VMVNIMM";
1197 case ARMISD::VMOVFPIMM: return "ARMISD::VMOVFPIMM";
1198 case ARMISD::VDUP: return "ARMISD::VDUP";
1199 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
1200 case ARMISD::VEXT: return "ARMISD::VEXT";
1201 case ARMISD::VREV64: return "ARMISD::VREV64";
1202 case ARMISD::VREV32: return "ARMISD::VREV32";
1203 case ARMISD::VREV16: return "ARMISD::VREV16";
1204 case ARMISD::VZIP: return "ARMISD::VZIP";
1205 case ARMISD::VUZP: return "ARMISD::VUZP";
1206 case ARMISD::VTRN: return "ARMISD::VTRN";
1207 case ARMISD::VTBL1: return "ARMISD::VTBL1";
1208 case ARMISD::VTBL2: return "ARMISD::VTBL2";
1209 case ARMISD::VMULLs: return "ARMISD::VMULLs";
1210 case ARMISD::VMULLu: return "ARMISD::VMULLu";
1211 case ARMISD::UMLAL: return "ARMISD::UMLAL";
1212 case ARMISD::SMLAL: return "ARMISD::SMLAL";
1213 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
1214 case ARMISD::BFI: return "ARMISD::BFI";
1215 case ARMISD::VORRIMM: return "ARMISD::VORRIMM";
1216 case ARMISD::VBICIMM: return "ARMISD::VBICIMM";
1217 case ARMISD::VBSL: return "ARMISD::VBSL";
1218 case ARMISD::MEMCPY: return "ARMISD::MEMCPY";
1219 case ARMISD::VLD2DUP: return "ARMISD::VLD2DUP";
1220 case ARMISD::VLD3DUP: return "ARMISD::VLD3DUP";
1221 case ARMISD::VLD4DUP: return "ARMISD::VLD4DUP";
1222 case ARMISD::VLD1_UPD: return "ARMISD::VLD1_UPD";
1223 case ARMISD::VLD2_UPD: return "ARMISD::VLD2_UPD";
1224 case ARMISD::VLD3_UPD: return "ARMISD::VLD3_UPD";
1225 case ARMISD::VLD4_UPD: return "ARMISD::VLD4_UPD";
1226 case ARMISD::VLD2LN_UPD: return "ARMISD::VLD2LN_UPD";
1227 case ARMISD::VLD3LN_UPD: return "ARMISD::VLD3LN_UPD";
1228 case ARMISD::VLD4LN_UPD: return "ARMISD::VLD4LN_UPD";
1229 case ARMISD::VLD2DUP_UPD: return "ARMISD::VLD2DUP_UPD";
1230 case ARMISD::VLD3DUP_UPD: return "ARMISD::VLD3DUP_UPD";
1231 case ARMISD::VLD4DUP_UPD: return "ARMISD::VLD4DUP_UPD";
1232 case ARMISD::VST1_UPD: return "ARMISD::VST1_UPD";
1233 case ARMISD::VST2_UPD: return "ARMISD::VST2_UPD";
1234 case ARMISD::VST3_UPD: return "ARMISD::VST3_UPD";
1235 case ARMISD::VST4_UPD: return "ARMISD::VST4_UPD";
1236 case ARMISD::VST2LN_UPD: return "ARMISD::VST2LN_UPD";
1237 case ARMISD::VST3LN_UPD: return "ARMISD::VST3LN_UPD";
1238 case ARMISD::VST4LN_UPD: return "ARMISD::VST4LN_UPD";
1243 EVT ARMTargetLowering::getSetCCResultType(const DataLayout &DL, LLVMContext &,
1246 return getPointerTy(DL);
1247 return VT.changeVectorElementTypeToInteger();
1250 /// getRegClassFor - Return the register class that should be used for the
1251 /// specified value type.
1252 const TargetRegisterClass *ARMTargetLowering::getRegClassFor(MVT VT) const {
1253 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
1254 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
1255 // load / store 4 to 8 consecutive D registers.
1256 if (Subtarget->hasNEON()) {
1257 if (VT == MVT::v4i64)
1258 return &ARM::QQPRRegClass;
1259 if (VT == MVT::v8i64)
1260 return &ARM::QQQQPRRegClass;
1262 return TargetLowering::getRegClassFor(VT);
1265 // memcpy, and other memory intrinsics, typically tries to use LDM/STM if the
1266 // source/dest is aligned and the copy size is large enough. We therefore want
1267 // to align such objects passed to memory intrinsics.
1268 bool ARMTargetLowering::shouldAlignPointerArgs(CallInst *CI, unsigned &MinSize,
1269 unsigned &PrefAlign) const {
1270 if (!isa<MemIntrinsic>(CI))
1273 // On ARM11 onwards (excluding M class) 8-byte aligned LDM is typically 1
1274 // cycle faster than 4-byte aligned LDM.
1275 PrefAlign = (Subtarget->hasV6Ops() && !Subtarget->isMClass() ? 8 : 4);
1279 // Create a fast isel object.
1281 ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
1282 const TargetLibraryInfo *libInfo) const {
1283 return ARM::createFastISel(funcInfo, libInfo);
1286 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
1287 unsigned NumVals = N->getNumValues();
1289 return Sched::RegPressure;
1291 for (unsigned i = 0; i != NumVals; ++i) {
1292 EVT VT = N->getValueType(i);
1293 if (VT == MVT::Glue || VT == MVT::Other)
1295 if (VT.isFloatingPoint() || VT.isVector())
1299 if (!N->isMachineOpcode())
1300 return Sched::RegPressure;
1302 // Load are scheduled for latency even if there instruction itinerary
1303 // is not available.
1304 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
1305 const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
1307 if (MCID.getNumDefs() == 0)
1308 return Sched::RegPressure;
1309 if (!Itins->isEmpty() &&
1310 Itins->getOperandCycle(MCID.getSchedClass(), 0) > 2)
1313 return Sched::RegPressure;
1316 //===----------------------------------------------------------------------===//
1318 //===----------------------------------------------------------------------===//
1320 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
1321 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
1323 default: llvm_unreachable("Unknown condition code!");
1324 case ISD::SETNE: return ARMCC::NE;
1325 case ISD::SETEQ: return ARMCC::EQ;
1326 case ISD::SETGT: return ARMCC::GT;
1327 case ISD::SETGE: return ARMCC::GE;
1328 case ISD::SETLT: return ARMCC::LT;
1329 case ISD::SETLE: return ARMCC::LE;
1330 case ISD::SETUGT: return ARMCC::HI;
1331 case ISD::SETUGE: return ARMCC::HS;
1332 case ISD::SETULT: return ARMCC::LO;
1333 case ISD::SETULE: return ARMCC::LS;
1337 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
1338 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
1339 ARMCC::CondCodes &CondCode2) {
1340 CondCode2 = ARMCC::AL;
1342 default: llvm_unreachable("Unknown FP condition!");
1344 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
1346 case ISD::SETOGT: CondCode = ARMCC::GT; break;
1348 case ISD::SETOGE: CondCode = ARMCC::GE; break;
1349 case ISD::SETOLT: CondCode = ARMCC::MI; break;
1350 case ISD::SETOLE: CondCode = ARMCC::LS; break;
1351 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
1352 case ISD::SETO: CondCode = ARMCC::VC; break;
1353 case ISD::SETUO: CondCode = ARMCC::VS; break;
1354 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
1355 case ISD::SETUGT: CondCode = ARMCC::HI; break;
1356 case ISD::SETUGE: CondCode = ARMCC::PL; break;
1358 case ISD::SETULT: CondCode = ARMCC::LT; break;
1360 case ISD::SETULE: CondCode = ARMCC::LE; break;
1362 case ISD::SETUNE: CondCode = ARMCC::NE; break;
1366 //===----------------------------------------------------------------------===//
1367 // Calling Convention Implementation
1368 //===----------------------------------------------------------------------===//
1370 #include "ARMGenCallingConv.inc"
1372 /// getEffectiveCallingConv - Get the effective calling convention, taking into
1373 /// account presence of floating point hardware and calling convention
1374 /// limitations, such as support for variadic functions.
1376 ARMTargetLowering::getEffectiveCallingConv(CallingConv::ID CC,
1377 bool isVarArg) const {
1380 llvm_unreachable("Unsupported calling convention");
1381 case CallingConv::ARM_AAPCS:
1382 case CallingConv::ARM_APCS:
1383 case CallingConv::GHC:
1385 case CallingConv::ARM_AAPCS_VFP:
1386 return isVarArg ? CallingConv::ARM_AAPCS : CallingConv::ARM_AAPCS_VFP;
1387 case CallingConv::C:
1388 if (!Subtarget->isAAPCS_ABI())
1389 return CallingConv::ARM_APCS;
1390 else if (Subtarget->hasVFP2() && !Subtarget->isThumb1Only() &&
1391 getTargetMachine().Options.FloatABIType == FloatABI::Hard &&
1393 return CallingConv::ARM_AAPCS_VFP;
1395 return CallingConv::ARM_AAPCS;
1396 case CallingConv::Fast:
1397 if (!Subtarget->isAAPCS_ABI()) {
1398 if (Subtarget->hasVFP2() && !Subtarget->isThumb1Only() && !isVarArg)
1399 return CallingConv::Fast;
1400 return CallingConv::ARM_APCS;
1401 } else if (Subtarget->hasVFP2() && !Subtarget->isThumb1Only() && !isVarArg)
1402 return CallingConv::ARM_AAPCS_VFP;
1404 return CallingConv::ARM_AAPCS;
1408 /// CCAssignFnForNode - Selects the correct CCAssignFn for the given
1409 /// CallingConvention.
1410 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
1412 bool isVarArg) const {
1413 switch (getEffectiveCallingConv(CC, isVarArg)) {
1415 llvm_unreachable("Unsupported calling convention");
1416 case CallingConv::ARM_APCS:
1417 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1418 case CallingConv::ARM_AAPCS:
1419 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1420 case CallingConv::ARM_AAPCS_VFP:
1421 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1422 case CallingConv::Fast:
1423 return (Return ? RetFastCC_ARM_APCS : FastCC_ARM_APCS);
1424 case CallingConv::GHC:
1425 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS_GHC);
1429 /// LowerCallResult - Lower the result values of a call into the
1430 /// appropriate copies out of appropriate physical registers.
1432 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1433 CallingConv::ID CallConv, bool isVarArg,
1434 const SmallVectorImpl<ISD::InputArg> &Ins,
1435 SDLoc dl, SelectionDAG &DAG,
1436 SmallVectorImpl<SDValue> &InVals,
1437 bool isThisReturn, SDValue ThisVal) const {
1439 // Assign locations to each value returned by this call.
1440 SmallVector<CCValAssign, 16> RVLocs;
1441 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
1442 *DAG.getContext(), Call);
1443 CCInfo.AnalyzeCallResult(Ins,
1444 CCAssignFnForNode(CallConv, /* Return*/ true,
1447 // Copy all of the result registers out of their specified physreg.
1448 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1449 CCValAssign VA = RVLocs[i];
1451 // Pass 'this' value directly from the argument to return value, to avoid
1452 // reg unit interference
1453 if (i == 0 && isThisReturn) {
1454 assert(!VA.needsCustom() && VA.getLocVT() == MVT::i32 &&
1455 "unexpected return calling convention register assignment");
1456 InVals.push_back(ThisVal);
1461 if (VA.needsCustom()) {
1462 // Handle f64 or half of a v2f64.
1463 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1465 Chain = Lo.getValue(1);
1466 InFlag = Lo.getValue(2);
1467 VA = RVLocs[++i]; // skip ahead to next loc
1468 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1470 Chain = Hi.getValue(1);
1471 InFlag = Hi.getValue(2);
1472 if (!Subtarget->isLittle())
1474 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1476 if (VA.getLocVT() == MVT::v2f64) {
1477 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1478 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1479 DAG.getConstant(0, dl, MVT::i32));
1481 VA = RVLocs[++i]; // skip ahead to next loc
1482 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1483 Chain = Lo.getValue(1);
1484 InFlag = Lo.getValue(2);
1485 VA = RVLocs[++i]; // skip ahead to next loc
1486 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1487 Chain = Hi.getValue(1);
1488 InFlag = Hi.getValue(2);
1489 if (!Subtarget->isLittle())
1491 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1492 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1493 DAG.getConstant(1, dl, MVT::i32));
1496 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
1498 Chain = Val.getValue(1);
1499 InFlag = Val.getValue(2);
1502 switch (VA.getLocInfo()) {
1503 default: llvm_unreachable("Unknown loc info!");
1504 case CCValAssign::Full: break;
1505 case CCValAssign::BCvt:
1506 Val = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), Val);
1510 InVals.push_back(Val);
1516 /// LowerMemOpCallTo - Store the argument to the stack.
1518 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
1519 SDValue StackPtr, SDValue Arg,
1520 SDLoc dl, SelectionDAG &DAG,
1521 const CCValAssign &VA,
1522 ISD::ArgFlagsTy Flags) const {
1523 unsigned LocMemOffset = VA.getLocMemOffset();
1524 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset, dl);
1525 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(DAG.getDataLayout()),
1527 return DAG.getStore(
1528 Chain, dl, Arg, PtrOff,
1529 MachinePointerInfo::getStack(DAG.getMachineFunction(), LocMemOffset),
1533 void ARMTargetLowering::PassF64ArgInRegs(SDLoc dl, SelectionDAG &DAG,
1534 SDValue Chain, SDValue &Arg,
1535 RegsToPassVector &RegsToPass,
1536 CCValAssign &VA, CCValAssign &NextVA,
1538 SmallVectorImpl<SDValue> &MemOpChains,
1539 ISD::ArgFlagsTy Flags) const {
1541 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1542 DAG.getVTList(MVT::i32, MVT::i32), Arg);
1543 unsigned id = Subtarget->isLittle() ? 0 : 1;
1544 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd.getValue(id)));
1546 if (NextVA.isRegLoc())
1547 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1-id)));
1549 assert(NextVA.isMemLoc());
1550 if (!StackPtr.getNode())
1551 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP,
1552 getPointerTy(DAG.getDataLayout()));
1554 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1-id),
1560 /// LowerCall - Lowering a call into a callseq_start <-
1561 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
1564 ARMTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
1565 SmallVectorImpl<SDValue> &InVals) const {
1566 SelectionDAG &DAG = CLI.DAG;
1568 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
1569 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
1570 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
1571 SDValue Chain = CLI.Chain;
1572 SDValue Callee = CLI.Callee;
1573 bool &isTailCall = CLI.IsTailCall;
1574 CallingConv::ID CallConv = CLI.CallConv;
1575 bool doesNotRet = CLI.DoesNotReturn;
1576 bool isVarArg = CLI.IsVarArg;
1578 MachineFunction &MF = DAG.getMachineFunction();
1579 bool isStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1580 bool isThisReturn = false;
1581 bool isSibCall = false;
1582 auto Attr = MF.getFunction()->getFnAttribute("disable-tail-calls");
1584 // Disable tail calls if they're not supported.
1585 if (!Subtarget->supportsTailCall() || Attr.getValueAsString() == "true")
1589 // Check if it's really possible to do a tail call.
1590 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1591 isVarArg, isStructRet, MF.getFunction()->hasStructRetAttr(),
1592 Outs, OutVals, Ins, DAG);
1593 if (!isTailCall && CLI.CS && CLI.CS->isMustTailCall())
1594 report_fatal_error("failed to perform tail call elimination on a call "
1595 "site marked musttail");
1596 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1597 // detected sibcalls.
1604 // Analyze operands of the call, assigning locations to each operand.
1605 SmallVector<CCValAssign, 16> ArgLocs;
1606 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs,
1607 *DAG.getContext(), Call);
1608 CCInfo.AnalyzeCallOperands(Outs,
1609 CCAssignFnForNode(CallConv, /* Return*/ false,
1612 // Get a count of how many bytes are to be pushed on the stack.
1613 unsigned NumBytes = CCInfo.getNextStackOffset();
1615 // For tail calls, memory operands are available in our caller's stack.
1619 // Adjust the stack pointer for the new arguments...
1620 // These operations are automatically eliminated by the prolog/epilog pass
1622 Chain = DAG.getCALLSEQ_START(Chain,
1623 DAG.getIntPtrConstant(NumBytes, dl, true), dl);
1626 DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy(DAG.getDataLayout()));
1628 RegsToPassVector RegsToPass;
1629 SmallVector<SDValue, 8> MemOpChains;
1631 // Walk the register/memloc assignments, inserting copies/loads. In the case
1632 // of tail call optimization, arguments are handled later.
1633 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1635 ++i, ++realArgIdx) {
1636 CCValAssign &VA = ArgLocs[i];
1637 SDValue Arg = OutVals[realArgIdx];
1638 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1639 bool isByVal = Flags.isByVal();
1641 // Promote the value if needed.
1642 switch (VA.getLocInfo()) {
1643 default: llvm_unreachable("Unknown loc info!");
1644 case CCValAssign::Full: break;
1645 case CCValAssign::SExt:
1646 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1648 case CCValAssign::ZExt:
1649 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1651 case CCValAssign::AExt:
1652 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1654 case CCValAssign::BCvt:
1655 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1659 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1660 if (VA.needsCustom()) {
1661 if (VA.getLocVT() == MVT::v2f64) {
1662 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1663 DAG.getConstant(0, dl, MVT::i32));
1664 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1665 DAG.getConstant(1, dl, MVT::i32));
1667 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1668 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1670 VA = ArgLocs[++i]; // skip ahead to next loc
1671 if (VA.isRegLoc()) {
1672 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1673 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1675 assert(VA.isMemLoc());
1677 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1678 dl, DAG, VA, Flags));
1681 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1682 StackPtr, MemOpChains, Flags);
1684 } else if (VA.isRegLoc()) {
1685 if (realArgIdx == 0 && Flags.isReturned() && Outs[0].VT == MVT::i32) {
1686 assert(VA.getLocVT() == MVT::i32 &&
1687 "unexpected calling convention register assignment");
1688 assert(!Ins.empty() && Ins[0].VT == MVT::i32 &&
1689 "unexpected use of 'returned'");
1690 isThisReturn = true;
1692 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1693 } else if (isByVal) {
1694 assert(VA.isMemLoc());
1695 unsigned offset = 0;
1697 // True if this byval aggregate will be split between registers
1699 unsigned ByValArgsCount = CCInfo.getInRegsParamsCount();
1700 unsigned CurByValIdx = CCInfo.getInRegsParamsProcessed();
1702 if (CurByValIdx < ByValArgsCount) {
1704 unsigned RegBegin, RegEnd;
1705 CCInfo.getInRegsParamInfo(CurByValIdx, RegBegin, RegEnd);
1708 DAG.getTargetLoweringInfo().getPointerTy(DAG.getDataLayout());
1710 for (i = 0, j = RegBegin; j < RegEnd; i++, j++) {
1711 SDValue Const = DAG.getConstant(4*i, dl, MVT::i32);
1712 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
1713 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
1714 MachinePointerInfo(),
1715 false, false, false,
1716 DAG.InferPtrAlignment(AddArg));
1717 MemOpChains.push_back(Load.getValue(1));
1718 RegsToPass.push_back(std::make_pair(j, Load));
1721 // If parameter size outsides register area, "offset" value
1722 // helps us to calculate stack slot for remained part properly.
1723 offset = RegEnd - RegBegin;
1725 CCInfo.nextInRegsParam();
1728 if (Flags.getByValSize() > 4*offset) {
1729 auto PtrVT = getPointerTy(DAG.getDataLayout());
1730 unsigned LocMemOffset = VA.getLocMemOffset();
1731 SDValue StkPtrOff = DAG.getIntPtrConstant(LocMemOffset, dl);
1732 SDValue Dst = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, StkPtrOff);
1733 SDValue SrcOffset = DAG.getIntPtrConstant(4*offset, dl);
1734 SDValue Src = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, SrcOffset);
1735 SDValue SizeNode = DAG.getConstant(Flags.getByValSize() - 4*offset, dl,
1737 SDValue AlignNode = DAG.getConstant(Flags.getByValAlign(), dl,
1740 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
1741 SDValue Ops[] = { Chain, Dst, Src, SizeNode, AlignNode};
1742 MemOpChains.push_back(DAG.getNode(ARMISD::COPY_STRUCT_BYVAL, dl, VTs,
1745 } else if (!isSibCall) {
1746 assert(VA.isMemLoc());
1748 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1749 dl, DAG, VA, Flags));
1753 if (!MemOpChains.empty())
1754 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains);
1756 // Build a sequence of copy-to-reg nodes chained together with token chain
1757 // and flag operands which copy the outgoing args into the appropriate regs.
1759 // Tail call byval lowering might overwrite argument registers so in case of
1760 // tail call optimization the copies to registers are lowered later.
1762 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1763 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1764 RegsToPass[i].second, InFlag);
1765 InFlag = Chain.getValue(1);
1768 // For tail calls lower the arguments to the 'real' stack slot.
1770 // Force all the incoming stack arguments to be loaded from the stack
1771 // before any new outgoing arguments are stored to the stack, because the
1772 // outgoing stack slots may alias the incoming argument stack slots, and
1773 // the alias isn't otherwise explicit. This is slightly more conservative
1774 // than necessary, because it means that each store effectively depends
1775 // on every argument instead of just those arguments it would clobber.
1777 // Do not flag preceding copytoreg stuff together with the following stuff.
1779 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1780 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1781 RegsToPass[i].second, InFlag);
1782 InFlag = Chain.getValue(1);
1787 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1788 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1789 // node so that legalize doesn't hack it.
1790 bool isDirect = false;
1791 bool isARMFunc = false;
1792 bool isLocalARMFunc = false;
1793 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1794 auto PtrVt = getPointerTy(DAG.getDataLayout());
1796 if (Subtarget->genLongCalls()) {
1797 assert((Subtarget->isTargetWindows() ||
1798 getTargetMachine().getRelocationModel() == Reloc::Static) &&
1799 "long-calls with non-static relocation model!");
1800 // Handle a global address or an external symbol. If it's not one of
1801 // those, the target's already in a register, so we don't need to do
1803 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1804 const GlobalValue *GV = G->getGlobal();
1805 // Create a constant pool entry for the callee address
1806 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1807 ARMConstantPoolValue *CPV =
1808 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 0);
1810 // Get the address of the callee into a register
1811 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVt, 4);
1812 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1813 Callee = DAG.getLoad(
1814 PtrVt, dl, DAG.getEntryNode(), CPAddr,
1815 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), false,
1817 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1818 const char *Sym = S->getSymbol();
1820 // Create a constant pool entry for the callee address
1821 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1822 ARMConstantPoolValue *CPV =
1823 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1824 ARMPCLabelIndex, 0);
1825 // Get the address of the callee into a register
1826 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVt, 4);
1827 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1828 Callee = DAG.getLoad(
1829 PtrVt, dl, DAG.getEntryNode(), CPAddr,
1830 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), false,
1833 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1834 const GlobalValue *GV = G->getGlobal();
1836 bool isDef = GV->isStrongDefinitionForLinker();
1837 bool isStub = (!isDef && Subtarget->isTargetMachO()) &&
1838 getTargetMachine().getRelocationModel() != Reloc::Static;
1839 isARMFunc = !Subtarget->isThumb() || (isStub && !Subtarget->isMClass());
1840 // ARM call to a local ARM function is predicable.
1841 isLocalARMFunc = !Subtarget->isThumb() && (isDef || !ARMInterworking);
1842 // tBX takes a register source operand.
1843 if (isStub && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1844 assert(Subtarget->isTargetMachO() && "WrapperPIC use on non-MachO?");
1845 Callee = DAG.getNode(
1846 ARMISD::WrapperPIC, dl, PtrVt,
1847 DAG.getTargetGlobalAddress(GV, dl, PtrVt, 0, ARMII::MO_NONLAZY));
1848 Callee = DAG.getLoad(PtrVt, dl, DAG.getEntryNode(), Callee,
1849 MachinePointerInfo::getGOT(DAG.getMachineFunction()),
1850 false, false, true, 0);
1851 } else if (Subtarget->isTargetCOFF()) {
1852 assert(Subtarget->isTargetWindows() &&
1853 "Windows is the only supported COFF target");
1854 unsigned TargetFlags = GV->hasDLLImportStorageClass()
1855 ? ARMII::MO_DLLIMPORT
1856 : ARMII::MO_NO_FLAG;
1858 DAG.getTargetGlobalAddress(GV, dl, PtrVt, /*Offset=*/0, TargetFlags);
1859 if (GV->hasDLLImportStorageClass())
1861 DAG.getLoad(PtrVt, dl, DAG.getEntryNode(),
1862 DAG.getNode(ARMISD::Wrapper, dl, PtrVt, Callee),
1863 MachinePointerInfo::getGOT(DAG.getMachineFunction()),
1864 false, false, false, 0);
1866 // On ELF targets for PIC code, direct calls should go through the PLT
1867 unsigned OpFlags = 0;
1868 if (Subtarget->isTargetELF() &&
1869 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1870 OpFlags = ARMII::MO_PLT;
1871 Callee = DAG.getTargetGlobalAddress(GV, dl, PtrVt, 0, OpFlags);
1873 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1875 bool isStub = Subtarget->isTargetMachO() &&
1876 getTargetMachine().getRelocationModel() != Reloc::Static;
1877 isARMFunc = !Subtarget->isThumb() || (isStub && !Subtarget->isMClass());
1878 // tBX takes a register source operand.
1879 const char *Sym = S->getSymbol();
1880 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1881 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1882 ARMConstantPoolValue *CPV =
1883 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1884 ARMPCLabelIndex, 4);
1885 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVt, 4);
1886 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1887 Callee = DAG.getLoad(
1888 PtrVt, dl, DAG.getEntryNode(), CPAddr,
1889 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), false,
1891 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, dl, MVT::i32);
1892 Callee = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVt, Callee, PICLabel);
1894 unsigned OpFlags = 0;
1895 // On ELF targets for PIC code, direct calls should go through the PLT
1896 if (Subtarget->isTargetELF() &&
1897 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1898 OpFlags = ARMII::MO_PLT;
1899 Callee = DAG.getTargetExternalSymbol(Sym, PtrVt, OpFlags);
1903 // FIXME: handle tail calls differently.
1905 if (Subtarget->isThumb()) {
1906 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1907 CallOpc = ARMISD::CALL_NOLINK;
1909 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1911 if (!isDirect && !Subtarget->hasV5TOps())
1912 CallOpc = ARMISD::CALL_NOLINK;
1913 else if (doesNotRet && isDirect && Subtarget->hasRAS() &&
1914 // Emit regular call when code size is the priority
1915 !MF.getFunction()->optForMinSize())
1916 // "mov lr, pc; b _foo" to avoid confusing the RSP
1917 CallOpc = ARMISD::CALL_NOLINK;
1919 CallOpc = isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL;
1922 std::vector<SDValue> Ops;
1923 Ops.push_back(Chain);
1924 Ops.push_back(Callee);
1926 // Add argument registers to the end of the list so that they are known live
1928 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1929 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1930 RegsToPass[i].second.getValueType()));
1932 // Add a register mask operand representing the call-preserved registers.
1934 const uint32_t *Mask;
1935 const ARMBaseRegisterInfo *ARI = Subtarget->getRegisterInfo();
1937 // For 'this' returns, use the R0-preserving mask if applicable
1938 Mask = ARI->getThisReturnPreservedMask(MF, CallConv);
1940 // Set isThisReturn to false if the calling convention is not one that
1941 // allows 'returned' to be modeled in this way, so LowerCallResult does
1942 // not try to pass 'this' straight through
1943 isThisReturn = false;
1944 Mask = ARI->getCallPreservedMask(MF, CallConv);
1947 Mask = ARI->getCallPreservedMask(MF, CallConv);
1949 assert(Mask && "Missing call preserved mask for calling convention");
1950 Ops.push_back(DAG.getRegisterMask(Mask));
1953 if (InFlag.getNode())
1954 Ops.push_back(InFlag);
1956 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
1958 MF.getFrameInfo()->setHasTailCall();
1959 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, Ops);
1962 // Returns a chain and a flag for retval copy to use.
1963 Chain = DAG.getNode(CallOpc, dl, NodeTys, Ops);
1964 InFlag = Chain.getValue(1);
1966 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, dl, true),
1967 DAG.getIntPtrConstant(0, dl, true), InFlag, dl);
1969 InFlag = Chain.getValue(1);
1971 // Handle result values, copying them out of physregs into vregs that we
1973 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins, dl, DAG,
1974 InVals, isThisReturn,
1975 isThisReturn ? OutVals[0] : SDValue());
1978 /// HandleByVal - Every parameter *after* a byval parameter is passed
1979 /// on the stack. Remember the next parameter register to allocate,
1980 /// and then confiscate the rest of the parameter registers to insure
1982 void ARMTargetLowering::HandleByVal(CCState *State, unsigned &Size,
1983 unsigned Align) const {
1984 assert((State->getCallOrPrologue() == Prologue ||
1985 State->getCallOrPrologue() == Call) &&
1986 "unhandled ParmContext");
1988 // Byval (as with any stack) slots are always at least 4 byte aligned.
1989 Align = std::max(Align, 4U);
1991 unsigned Reg = State->AllocateReg(GPRArgRegs);
1995 unsigned AlignInRegs = Align / 4;
1996 unsigned Waste = (ARM::R4 - Reg) % AlignInRegs;
1997 for (unsigned i = 0; i < Waste; ++i)
1998 Reg = State->AllocateReg(GPRArgRegs);
2003 unsigned Excess = 4 * (ARM::R4 - Reg);
2005 // Special case when NSAA != SP and parameter size greater than size of
2006 // all remained GPR regs. In that case we can't split parameter, we must
2007 // send it to stack. We also must set NCRN to R4, so waste all
2008 // remained registers.
2009 const unsigned NSAAOffset = State->getNextStackOffset();
2010 if (NSAAOffset != 0 && Size > Excess) {
2011 while (State->AllocateReg(GPRArgRegs))
2016 // First register for byval parameter is the first register that wasn't
2017 // allocated before this method call, so it would be "reg".
2018 // If parameter is small enough to be saved in range [reg, r4), then
2019 // the end (first after last) register would be reg + param-size-in-regs,
2020 // else parameter would be splitted between registers and stack,
2021 // end register would be r4 in this case.
2022 unsigned ByValRegBegin = Reg;
2023 unsigned ByValRegEnd = std::min<unsigned>(Reg + Size / 4, ARM::R4);
2024 State->addInRegsParamInfo(ByValRegBegin, ByValRegEnd);
2025 // Note, first register is allocated in the beginning of function already,
2026 // allocate remained amount of registers we need.
2027 for (unsigned i = Reg + 1; i != ByValRegEnd; ++i)
2028 State->AllocateReg(GPRArgRegs);
2029 // A byval parameter that is split between registers and memory needs its
2030 // size truncated here.
2031 // In the case where the entire structure fits in registers, we set the
2032 // size in memory to zero.
2033 Size = std::max<int>(Size - Excess, 0);
2036 /// MatchingStackOffset - Return true if the given stack call argument is
2037 /// already available in the same position (relatively) of the caller's
2038 /// incoming argument stack.
2040 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
2041 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
2042 const TargetInstrInfo *TII) {
2043 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
2045 if (Arg.getOpcode() == ISD::CopyFromReg) {
2046 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
2047 if (!TargetRegisterInfo::isVirtualRegister(VR))
2049 MachineInstr *Def = MRI->getVRegDef(VR);
2052 if (!Flags.isByVal()) {
2053 if (!TII->isLoadFromStackSlot(Def, FI))
2058 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
2059 if (Flags.isByVal())
2060 // ByVal argument is passed in as a pointer but it's now being
2061 // dereferenced. e.g.
2062 // define @foo(%struct.X* %A) {
2063 // tail call @bar(%struct.X* byval %A)
2066 SDValue Ptr = Ld->getBasePtr();
2067 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
2070 FI = FINode->getIndex();
2074 assert(FI != INT_MAX);
2075 if (!MFI->isFixedObjectIndex(FI))
2077 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
2080 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2081 /// for tail call optimization. Targets which want to do tail call
2082 /// optimization should implement this function.
2084 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2085 CallingConv::ID CalleeCC,
2087 bool isCalleeStructRet,
2088 bool isCallerStructRet,
2089 const SmallVectorImpl<ISD::OutputArg> &Outs,
2090 const SmallVectorImpl<SDValue> &OutVals,
2091 const SmallVectorImpl<ISD::InputArg> &Ins,
2092 SelectionDAG& DAG) const {
2093 const Function *CallerF = DAG.getMachineFunction().getFunction();
2094 CallingConv::ID CallerCC = CallerF->getCallingConv();
2095 bool CCMatch = CallerCC == CalleeCC;
2097 assert(Subtarget->supportsTailCall());
2099 // Look for obvious safe cases to perform tail call optimization that do not
2100 // require ABI changes. This is what gcc calls sibcall.
2102 // Do not sibcall optimize vararg calls unless the call site is not passing
2104 if (isVarArg && !Outs.empty())
2107 // Exception-handling functions need a special set of instructions to indicate
2108 // a return to the hardware. Tail-calling another function would probably
2110 if (CallerF->hasFnAttribute("interrupt"))
2113 // Also avoid sibcall optimization if either caller or callee uses struct
2114 // return semantics.
2115 if (isCalleeStructRet || isCallerStructRet)
2118 // Externally-defined functions with weak linkage should not be
2119 // tail-called on ARM when the OS does not support dynamic
2120 // pre-emption of symbols, as the AAELF spec requires normal calls
2121 // to undefined weak functions to be replaced with a NOP or jump to the
2122 // next instruction. The behaviour of branch instructions in this
2123 // situation (as used for tail calls) is implementation-defined, so we
2124 // cannot rely on the linker replacing the tail call with a return.
2125 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2126 const GlobalValue *GV = G->getGlobal();
2127 const Triple &TT = getTargetMachine().getTargetTriple();
2128 if (GV->hasExternalWeakLinkage() &&
2129 (!TT.isOSWindows() || TT.isOSBinFormatELF() || TT.isOSBinFormatMachO()))
2133 // If the calling conventions do not match, then we'd better make sure the
2134 // results are returned in the same way as what the caller expects.
2136 SmallVector<CCValAssign, 16> RVLocs1;
2137 ARMCCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(), RVLocs1,
2138 *DAG.getContext(), Call);
2139 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
2141 SmallVector<CCValAssign, 16> RVLocs2;
2142 ARMCCState CCInfo2(CallerCC, false, DAG.getMachineFunction(), RVLocs2,
2143 *DAG.getContext(), Call);
2144 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
2146 if (RVLocs1.size() != RVLocs2.size())
2148 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
2149 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
2151 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
2153 if (RVLocs1[i].isRegLoc()) {
2154 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
2157 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
2163 // If Caller's vararg or byval argument has been split between registers and
2164 // stack, do not perform tail call, since part of the argument is in caller's
2166 const ARMFunctionInfo *AFI_Caller = DAG.getMachineFunction().
2167 getInfo<ARMFunctionInfo>();
2168 if (AFI_Caller->getArgRegsSaveSize())
2171 // If the callee takes no arguments then go on to check the results of the
2173 if (!Outs.empty()) {
2174 // Check if stack adjustment is needed. For now, do not do this if any
2175 // argument is passed on the stack.
2176 SmallVector<CCValAssign, 16> ArgLocs;
2177 ARMCCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
2178 *DAG.getContext(), Call);
2179 CCInfo.AnalyzeCallOperands(Outs,
2180 CCAssignFnForNode(CalleeCC, false, isVarArg));
2181 if (CCInfo.getNextStackOffset()) {
2182 MachineFunction &MF = DAG.getMachineFunction();
2184 // Check if the arguments are already laid out in the right way as
2185 // the caller's fixed stack objects.
2186 MachineFrameInfo *MFI = MF.getFrameInfo();
2187 const MachineRegisterInfo *MRI = &MF.getRegInfo();
2188 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
2189 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
2191 ++i, ++realArgIdx) {
2192 CCValAssign &VA = ArgLocs[i];
2193 EVT RegVT = VA.getLocVT();
2194 SDValue Arg = OutVals[realArgIdx];
2195 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
2196 if (VA.getLocInfo() == CCValAssign::Indirect)
2198 if (VA.needsCustom()) {
2199 // f64 and vector types are split into multiple registers or
2200 // register/stack-slot combinations. The types will not match
2201 // the registers; give up on memory f64 refs until we figure
2202 // out what to do about this.
2205 if (!ArgLocs[++i].isRegLoc())
2207 if (RegVT == MVT::v2f64) {
2208 if (!ArgLocs[++i].isRegLoc())
2210 if (!ArgLocs[++i].isRegLoc())
2213 } else if (!VA.isRegLoc()) {
2214 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
2226 ARMTargetLowering::CanLowerReturn(CallingConv::ID CallConv,
2227 MachineFunction &MF, bool isVarArg,
2228 const SmallVectorImpl<ISD::OutputArg> &Outs,
2229 LLVMContext &Context) const {
2230 SmallVector<CCValAssign, 16> RVLocs;
2231 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, Context);
2232 return CCInfo.CheckReturn(Outs, CCAssignFnForNode(CallConv, /*Return=*/true,
2236 static SDValue LowerInterruptReturn(SmallVectorImpl<SDValue> &RetOps,
2237 SDLoc DL, SelectionDAG &DAG) {
2238 const MachineFunction &MF = DAG.getMachineFunction();
2239 const Function *F = MF.getFunction();
2241 StringRef IntKind = F->getFnAttribute("interrupt").getValueAsString();
2243 // See ARM ARM v7 B1.8.3. On exception entry LR is set to a possibly offset
2244 // version of the "preferred return address". These offsets affect the return
2245 // instruction if this is a return from PL1 without hypervisor extensions.
2246 // IRQ/FIQ: +4 "subs pc, lr, #4"
2247 // SWI: 0 "subs pc, lr, #0"
2248 // ABORT: +4 "subs pc, lr, #4"
2249 // UNDEF: +4/+2 "subs pc, lr, #0"
2250 // UNDEF varies depending on where the exception came from ARM or Thumb
2251 // mode. Alongside GCC, we throw our hands up in disgust and pretend it's 0.
2254 if (IntKind == "" || IntKind == "IRQ" || IntKind == "FIQ" ||
2257 else if (IntKind == "SWI" || IntKind == "UNDEF")
2260 report_fatal_error("Unsupported interrupt attribute. If present, value "
2261 "must be one of: IRQ, FIQ, SWI, ABORT or UNDEF");
2263 RetOps.insert(RetOps.begin() + 1,
2264 DAG.getConstant(LROffset, DL, MVT::i32, false));
2266 return DAG.getNode(ARMISD::INTRET_FLAG, DL, MVT::Other, RetOps);
2270 ARMTargetLowering::LowerReturn(SDValue Chain,
2271 CallingConv::ID CallConv, bool isVarArg,
2272 const SmallVectorImpl<ISD::OutputArg> &Outs,
2273 const SmallVectorImpl<SDValue> &OutVals,
2274 SDLoc dl, SelectionDAG &DAG) const {
2276 // CCValAssign - represent the assignment of the return value to a location.
2277 SmallVector<CCValAssign, 16> RVLocs;
2279 // CCState - Info about the registers and stack slots.
2280 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
2281 *DAG.getContext(), Call);
2283 // Analyze outgoing return values.
2284 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
2288 SmallVector<SDValue, 4> RetOps;
2289 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
2290 bool isLittleEndian = Subtarget->isLittle();
2292 MachineFunction &MF = DAG.getMachineFunction();
2293 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2294 AFI->setReturnRegsCount(RVLocs.size());
2296 // Copy the result values into the output registers.
2297 for (unsigned i = 0, realRVLocIdx = 0;
2299 ++i, ++realRVLocIdx) {
2300 CCValAssign &VA = RVLocs[i];
2301 assert(VA.isRegLoc() && "Can only return in registers!");
2303 SDValue Arg = OutVals[realRVLocIdx];
2305 switch (VA.getLocInfo()) {
2306 default: llvm_unreachable("Unknown loc info!");
2307 case CCValAssign::Full: break;
2308 case CCValAssign::BCvt:
2309 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
2313 if (VA.needsCustom()) {
2314 if (VA.getLocVT() == MVT::v2f64) {
2315 // Extract the first half and return it in two registers.
2316 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
2317 DAG.getConstant(0, dl, MVT::i32));
2318 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
2319 DAG.getVTList(MVT::i32, MVT::i32), Half);
2321 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2322 HalfGPRs.getValue(isLittleEndian ? 0 : 1),
2324 Flag = Chain.getValue(1);
2325 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2326 VA = RVLocs[++i]; // skip ahead to next loc
2327 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2328 HalfGPRs.getValue(isLittleEndian ? 1 : 0),
2330 Flag = Chain.getValue(1);
2331 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2332 VA = RVLocs[++i]; // skip ahead to next loc
2334 // Extract the 2nd half and fall through to handle it as an f64 value.
2335 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
2336 DAG.getConstant(1, dl, MVT::i32));
2338 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
2340 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
2341 DAG.getVTList(MVT::i32, MVT::i32), Arg);
2342 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2343 fmrrd.getValue(isLittleEndian ? 0 : 1),
2345 Flag = Chain.getValue(1);
2346 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2347 VA = RVLocs[++i]; // skip ahead to next loc
2348 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2349 fmrrd.getValue(isLittleEndian ? 1 : 0),
2352 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
2354 // Guarantee that all emitted copies are
2355 // stuck together, avoiding something bad.
2356 Flag = Chain.getValue(1);
2357 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2360 // Update chain and glue.
2363 RetOps.push_back(Flag);
2365 // CPUs which aren't M-class use a special sequence to return from
2366 // exceptions (roughly, any instruction setting pc and cpsr simultaneously,
2367 // though we use "subs pc, lr, #N").
2369 // M-class CPUs actually use a normal return sequence with a special
2370 // (hardware-provided) value in LR, so the normal code path works.
2371 if (DAG.getMachineFunction().getFunction()->hasFnAttribute("interrupt") &&
2372 !Subtarget->isMClass()) {
2373 if (Subtarget->isThumb1Only())
2374 report_fatal_error("interrupt attribute is not supported in Thumb1");
2375 return LowerInterruptReturn(RetOps, dl, DAG);
2378 return DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, RetOps);
2381 bool ARMTargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
2382 if (N->getNumValues() != 1)
2384 if (!N->hasNUsesOfValue(1, 0))
2387 SDValue TCChain = Chain;
2388 SDNode *Copy = *N->use_begin();
2389 if (Copy->getOpcode() == ISD::CopyToReg) {
2390 // If the copy has a glue operand, we conservatively assume it isn't safe to
2391 // perform a tail call.
2392 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
2394 TCChain = Copy->getOperand(0);
2395 } else if (Copy->getOpcode() == ARMISD::VMOVRRD) {
2396 SDNode *VMov = Copy;
2397 // f64 returned in a pair of GPRs.
2398 SmallPtrSet<SDNode*, 2> Copies;
2399 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
2401 if (UI->getOpcode() != ISD::CopyToReg)
2405 if (Copies.size() > 2)
2408 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
2410 SDValue UseChain = UI->getOperand(0);
2411 if (Copies.count(UseChain.getNode()))
2415 // We are at the top of this chain.
2416 // If the copy has a glue operand, we conservatively assume it
2417 // isn't safe to perform a tail call.
2418 if (UI->getOperand(UI->getNumOperands()-1).getValueType() == MVT::Glue)
2424 } else if (Copy->getOpcode() == ISD::BITCAST) {
2425 // f32 returned in a single GPR.
2426 if (!Copy->hasOneUse())
2428 Copy = *Copy->use_begin();
2429 if (Copy->getOpcode() != ISD::CopyToReg || !Copy->hasNUsesOfValue(1, 0))
2431 // If the copy has a glue operand, we conservatively assume it isn't safe to
2432 // perform a tail call.
2433 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
2435 TCChain = Copy->getOperand(0);
2440 bool HasRet = false;
2441 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
2443 if (UI->getOpcode() != ARMISD::RET_FLAG &&
2444 UI->getOpcode() != ARMISD::INTRET_FLAG)
2456 bool ARMTargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
2457 if (!Subtarget->supportsTailCall())
2461 CI->getParent()->getParent()->getFnAttribute("disable-tail-calls");
2462 if (!CI->isTailCall() || Attr.getValueAsString() == "true")
2468 // Trying to write a 64 bit value so need to split into two 32 bit values first,
2469 // and pass the lower and high parts through.
2470 static SDValue LowerWRITE_REGISTER(SDValue Op, SelectionDAG &DAG) {
2472 SDValue WriteValue = Op->getOperand(2);
2474 // This function is only supposed to be called for i64 type argument.
2475 assert(WriteValue.getValueType() == MVT::i64
2476 && "LowerWRITE_REGISTER called for non-i64 type argument.");
2478 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, WriteValue,
2479 DAG.getConstant(0, DL, MVT::i32));
2480 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, WriteValue,
2481 DAG.getConstant(1, DL, MVT::i32));
2482 SDValue Ops[] = { Op->getOperand(0), Op->getOperand(1), Lo, Hi };
2483 return DAG.getNode(ISD::WRITE_REGISTER, DL, MVT::Other, Ops);
2486 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
2487 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
2488 // one of the above mentioned nodes. It has to be wrapped because otherwise
2489 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
2490 // be used to form addressing mode. These wrapped nodes will be selected
2492 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
2493 EVT PtrVT = Op.getValueType();
2494 // FIXME there is no actual debug info here
2496 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
2498 if (CP->isMachineConstantPoolEntry())
2499 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
2500 CP->getAlignment());
2502 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
2503 CP->getAlignment());
2504 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
2507 unsigned ARMTargetLowering::getJumpTableEncoding() const {
2508 return MachineJumpTableInfo::EK_Inline;
2511 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
2512 SelectionDAG &DAG) const {
2513 MachineFunction &MF = DAG.getMachineFunction();
2514 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2515 unsigned ARMPCLabelIndex = 0;
2517 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2518 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
2519 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2521 if (RelocM == Reloc::Static) {
2522 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
2524 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2525 ARMPCLabelIndex = AFI->createPICLabelUId();
2526 ARMConstantPoolValue *CPV =
2527 ARMConstantPoolConstant::Create(BA, ARMPCLabelIndex,
2528 ARMCP::CPBlockAddress, PCAdj);
2529 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2531 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
2533 DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
2534 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
2535 false, false, false, 0);
2536 if (RelocM == Reloc::Static)
2538 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, DL, MVT::i32);
2539 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
2542 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
2544 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
2545 SelectionDAG &DAG) const {
2547 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2548 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2549 MachineFunction &MF = DAG.getMachineFunction();
2550 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2551 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2552 ARMConstantPoolValue *CPV =
2553 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2554 ARMCP::CPValue, PCAdj, ARMCP::TLSGD, true);
2555 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2556 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
2558 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
2559 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
2560 false, false, false, 0);
2561 SDValue Chain = Argument.getValue(1);
2563 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, dl, MVT::i32);
2564 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
2566 // call __tls_get_addr.
2569 Entry.Node = Argument;
2570 Entry.Ty = (Type *) Type::getInt32Ty(*DAG.getContext());
2571 Args.push_back(Entry);
2573 // FIXME: is there useful debug info available here?
2574 TargetLowering::CallLoweringInfo CLI(DAG);
2575 CLI.setDebugLoc(dl).setChain(Chain)
2576 .setCallee(CallingConv::C, Type::getInt32Ty(*DAG.getContext()),
2577 DAG.getExternalSymbol("__tls_get_addr", PtrVT), std::move(Args),
2580 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
2581 return CallResult.first;
2584 // Lower ISD::GlobalTLSAddress using the "initial exec" or
2585 // "local exec" model.
2587 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
2589 TLSModel::Model model) const {
2590 const GlobalValue *GV = GA->getGlobal();
2593 SDValue Chain = DAG.getEntryNode();
2594 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2595 // Get the Thread Pointer
2596 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2598 if (model == TLSModel::InitialExec) {
2599 MachineFunction &MF = DAG.getMachineFunction();
2600 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2601 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2602 // Initial exec model.
2603 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2604 ARMConstantPoolValue *CPV =
2605 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2606 ARMCP::CPValue, PCAdj, ARMCP::GOTTPOFF,
2608 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2609 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2610 Offset = DAG.getLoad(
2611 PtrVT, dl, Chain, Offset,
2612 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), false,
2614 Chain = Offset.getValue(1);
2616 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, dl, MVT::i32);
2617 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
2619 Offset = DAG.getLoad(
2620 PtrVT, dl, Chain, Offset,
2621 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), false,
2625 assert(model == TLSModel::LocalExec);
2626 ARMConstantPoolValue *CPV =
2627 ARMConstantPoolConstant::Create(GV, ARMCP::TPOFF);
2628 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2629 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2630 Offset = DAG.getLoad(
2631 PtrVT, dl, Chain, Offset,
2632 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), false,
2636 // The address of the thread local variable is the add of the thread
2637 // pointer with the offset of the variable.
2638 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
2642 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
2643 // TODO: implement the "local dynamic" model
2644 assert(Subtarget->isTargetELF() &&
2645 "TLS not implemented for non-ELF targets");
2646 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
2647 if (DAG.getTarget().Options.EmulatedTLS)
2648 return LowerToTLSEmulatedModel(GA, DAG);
2650 TLSModel::Model model = getTargetMachine().getTLSModel(GA->getGlobal());
2653 case TLSModel::GeneralDynamic:
2654 case TLSModel::LocalDynamic:
2655 return LowerToTLSGeneralDynamicModel(GA, DAG);
2656 case TLSModel::InitialExec:
2657 case TLSModel::LocalExec:
2658 return LowerToTLSExecModels(GA, DAG, model);
2660 llvm_unreachable("bogus TLS model");
2663 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
2664 SelectionDAG &DAG) const {
2665 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2667 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2668 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2670 !(GV->hasHiddenVisibility() || GV->hasLocalLinkage());
2672 MachineFunction &MF = DAG.getMachineFunction();
2673 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2674 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2675 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2677 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2678 ARMConstantPoolValue *CPV = ARMConstantPoolConstant::Create(
2679 GV, ARMPCLabelIndex, ARMCP::CPValue, PCAdj,
2680 UseGOT_PREL ? ARMCP::GOT_PREL : ARMCP::no_modifier,
2681 /*AddCurrentAddress=*/UseGOT_PREL);
2682 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2683 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2684 SDValue Result = DAG.getLoad(
2685 PtrVT, dl, DAG.getEntryNode(), CPAddr,
2686 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), false,
2688 SDValue Chain = Result.getValue(1);
2689 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, dl, MVT::i32);
2690 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2692 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
2693 MachinePointerInfo::getGOT(DAG.getMachineFunction()),
2694 false, false, false, 0);
2698 // If we have T2 ops, we can materialize the address directly via movt/movw
2699 // pair. This is always cheaper.
2700 if (Subtarget->useMovt(DAG.getMachineFunction())) {
2702 // FIXME: Once remat is capable of dealing with instructions with register
2703 // operands, expand this into two nodes.
2704 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2705 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2707 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2708 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2710 PtrVT, dl, DAG.getEntryNode(), CPAddr,
2711 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), false,
2716 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
2717 SelectionDAG &DAG) const {
2718 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2720 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2721 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2723 if (Subtarget->useMovt(DAG.getMachineFunction()))
2726 // FIXME: Once remat is capable of dealing with instructions with register
2727 // operands, expand this into multiple nodes
2729 RelocM == Reloc::PIC_ ? ARMISD::WrapperPIC : ARMISD::Wrapper;
2731 SDValue G = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, ARMII::MO_NONLAZY);
2732 SDValue Result = DAG.getNode(Wrapper, dl, PtrVT, G);
2734 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2735 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
2736 MachinePointerInfo::getGOT(DAG.getMachineFunction()),
2737 false, false, false, 0);
2741 SDValue ARMTargetLowering::LowerGlobalAddressWindows(SDValue Op,
2742 SelectionDAG &DAG) const {
2743 assert(Subtarget->isTargetWindows() && "non-Windows COFF is not supported");
2744 assert(Subtarget->useMovt(DAG.getMachineFunction()) &&
2745 "Windows on ARM expects to use movw/movt");
2747 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2748 const ARMII::TOF TargetFlags =
2749 (GV->hasDLLImportStorageClass() ? ARMII::MO_DLLIMPORT : ARMII::MO_NO_FLAG);
2750 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2756 // FIXME: Once remat is capable of dealing with instructions with register
2757 // operands, expand this into two nodes.
2758 Result = DAG.getNode(ARMISD::Wrapper, DL, PtrVT,
2759 DAG.getTargetGlobalAddress(GV, DL, PtrVT, /*Offset=*/0,
2761 if (GV->hasDLLImportStorageClass())
2762 Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), Result,
2763 MachinePointerInfo::getGOT(DAG.getMachineFunction()),
2764 false, false, false, 0);
2769 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
2771 SDValue Val = DAG.getConstant(0, dl, MVT::i32);
2772 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl,
2773 DAG.getVTList(MVT::i32, MVT::Other), Op.getOperand(0),
2774 Op.getOperand(1), Val);
2778 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
2780 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
2781 Op.getOperand(1), DAG.getConstant(0, dl, MVT::i32));
2784 SDValue ARMTargetLowering::LowerEH_SJLJ_SETUP_DISPATCH(SDValue Op,
2785 SelectionDAG &DAG) const {
2787 return DAG.getNode(ARMISD::EH_SJLJ_SETUP_DISPATCH, dl, MVT::Other,
2792 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
2793 const ARMSubtarget *Subtarget) const {
2794 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2797 default: return SDValue(); // Don't custom lower most intrinsics.
2798 case Intrinsic::arm_rbit: {
2799 assert(Op.getOperand(1).getValueType() == MVT::i32 &&
2800 "RBIT intrinsic must have i32 type!");
2801 return DAG.getNode(ISD::BITREVERSE, dl, MVT::i32, Op.getOperand(1));
2803 case Intrinsic::arm_thread_pointer: {
2804 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2805 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2807 case Intrinsic::eh_sjlj_lsda: {
2808 MachineFunction &MF = DAG.getMachineFunction();
2809 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2810 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2811 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2812 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2814 unsigned PCAdj = (RelocM != Reloc::PIC_)
2815 ? 0 : (Subtarget->isThumb() ? 4 : 8);
2816 ARMConstantPoolValue *CPV =
2817 ARMConstantPoolConstant::Create(MF.getFunction(), ARMPCLabelIndex,
2818 ARMCP::CPLSDA, PCAdj);
2819 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2820 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2821 SDValue Result = DAG.getLoad(
2822 PtrVT, dl, DAG.getEntryNode(), CPAddr,
2823 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), false,
2826 if (RelocM == Reloc::PIC_) {
2827 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, dl, MVT::i32);
2828 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2832 case Intrinsic::arm_neon_vmulls:
2833 case Intrinsic::arm_neon_vmullu: {
2834 unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vmulls)
2835 ? ARMISD::VMULLs : ARMISD::VMULLu;
2836 return DAG.getNode(NewOpc, SDLoc(Op), Op.getValueType(),
2837 Op.getOperand(1), Op.getOperand(2));
2839 case Intrinsic::arm_neon_vminnm:
2840 case Intrinsic::arm_neon_vmaxnm: {
2841 unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vminnm)
2842 ? ISD::FMINNUM : ISD::FMAXNUM;
2843 return DAG.getNode(NewOpc, SDLoc(Op), Op.getValueType(),
2844 Op.getOperand(1), Op.getOperand(2));
2846 case Intrinsic::arm_neon_vminu:
2847 case Intrinsic::arm_neon_vmaxu: {
2848 if (Op.getValueType().isFloatingPoint())
2850 unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vminu)
2851 ? ISD::UMIN : ISD::UMAX;
2852 return DAG.getNode(NewOpc, SDLoc(Op), Op.getValueType(),
2853 Op.getOperand(1), Op.getOperand(2));
2855 case Intrinsic::arm_neon_vmins:
2856 case Intrinsic::arm_neon_vmaxs: {
2857 // v{min,max}s is overloaded between signed integers and floats.
2858 if (!Op.getValueType().isFloatingPoint()) {
2859 unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vmins)
2860 ? ISD::SMIN : ISD::SMAX;
2861 return DAG.getNode(NewOpc, SDLoc(Op), Op.getValueType(),
2862 Op.getOperand(1), Op.getOperand(2));
2864 unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vmins)
2865 ? ISD::FMINNAN : ISD::FMAXNAN;
2866 return DAG.getNode(NewOpc, SDLoc(Op), Op.getValueType(),
2867 Op.getOperand(1), Op.getOperand(2));
2872 static SDValue LowerATOMIC_FENCE(SDValue Op, SelectionDAG &DAG,
2873 const ARMSubtarget *Subtarget) {
2874 // FIXME: handle "fence singlethread" more efficiently.
2876 if (!Subtarget->hasDataBarrier()) {
2877 // Some ARMv6 cpus can support data barriers with an mcr instruction.
2878 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2880 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2881 "Unexpected ISD::ATOMIC_FENCE encountered. Should be libcall!");
2882 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2883 DAG.getConstant(0, dl, MVT::i32));
2886 ConstantSDNode *OrdN = cast<ConstantSDNode>(Op.getOperand(1));
2887 AtomicOrdering Ord = static_cast<AtomicOrdering>(OrdN->getZExtValue());
2888 ARM_MB::MemBOpt Domain = ARM_MB::ISH;
2889 if (Subtarget->isMClass()) {
2890 // Only a full system barrier exists in the M-class architectures.
2891 Domain = ARM_MB::SY;
2892 } else if (Subtarget->isSwift() && Ord == Release) {
2893 // Swift happens to implement ISHST barriers in a way that's compatible with
2894 // Release semantics but weaker than ISH so we'd be fools not to use
2895 // it. Beware: other processors probably don't!
2896 Domain = ARM_MB::ISHST;
2899 return DAG.getNode(ISD::INTRINSIC_VOID, dl, MVT::Other, Op.getOperand(0),
2900 DAG.getConstant(Intrinsic::arm_dmb, dl, MVT::i32),
2901 DAG.getConstant(Domain, dl, MVT::i32));
2904 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG,
2905 const ARMSubtarget *Subtarget) {
2906 // ARM pre v5TE and Thumb1 does not have preload instructions.
2907 if (!(Subtarget->isThumb2() ||
2908 (!Subtarget->isThumb1Only() && Subtarget->hasV5TEOps())))
2909 // Just preserve the chain.
2910 return Op.getOperand(0);
2913 unsigned isRead = ~cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() & 1;
2915 (!Subtarget->hasV7Ops() || !Subtarget->hasMPExtension()))
2916 // ARMv7 with MP extension has PLDW.
2917 return Op.getOperand(0);
2919 unsigned isData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
2920 if (Subtarget->isThumb()) {
2922 isRead = ~isRead & 1;
2923 isData = ~isData & 1;
2926 return DAG.getNode(ARMISD::PRELOAD, dl, MVT::Other, Op.getOperand(0),
2927 Op.getOperand(1), DAG.getConstant(isRead, dl, MVT::i32),
2928 DAG.getConstant(isData, dl, MVT::i32));
2931 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
2932 MachineFunction &MF = DAG.getMachineFunction();
2933 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
2935 // vastart just stores the address of the VarArgsFrameIndex slot into the
2936 // memory location argument.
2938 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy(DAG.getDataLayout());
2939 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2940 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
2941 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
2942 MachinePointerInfo(SV), false, false, 0);
2946 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
2947 SDValue &Root, SelectionDAG &DAG,
2949 MachineFunction &MF = DAG.getMachineFunction();
2950 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2952 const TargetRegisterClass *RC;
2953 if (AFI->isThumb1OnlyFunction())
2954 RC = &ARM::tGPRRegClass;
2956 RC = &ARM::GPRRegClass;
2958 // Transform the arguments stored in physical registers into virtual ones.
2959 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2960 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2963 if (NextVA.isMemLoc()) {
2964 MachineFrameInfo *MFI = MF.getFrameInfo();
2965 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true);
2967 // Create load node to retrieve arguments from the stack.
2968 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
2969 ArgValue2 = DAG.getLoad(
2970 MVT::i32, dl, Root, FIN,
2971 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI), false,
2974 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
2975 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2977 if (!Subtarget->isLittle())
2978 std::swap (ArgValue, ArgValue2);
2979 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
2982 // The remaining GPRs hold either the beginning of variable-argument
2983 // data, or the beginning of an aggregate passed by value (usually
2984 // byval). Either way, we allocate stack slots adjacent to the data
2985 // provided by our caller, and store the unallocated registers there.
2986 // If this is a variadic function, the va_list pointer will begin with
2987 // these values; otherwise, this reassembles a (byval) structure that
2988 // was split between registers and memory.
2989 // Return: The frame index registers were stored into.
2991 ARMTargetLowering::StoreByValRegs(CCState &CCInfo, SelectionDAG &DAG,
2992 SDLoc dl, SDValue &Chain,
2993 const Value *OrigArg,
2994 unsigned InRegsParamRecordIdx,
2996 unsigned ArgSize) const {
2997 // Currently, two use-cases possible:
2998 // Case #1. Non-var-args function, and we meet first byval parameter.
2999 // Setup first unallocated register as first byval register;
3000 // eat all remained registers
3001 // (these two actions are performed by HandleByVal method).
3002 // Then, here, we initialize stack frame with
3003 // "store-reg" instructions.
3004 // Case #2. Var-args function, that doesn't contain byval parameters.
3005 // The same: eat all remained unallocated registers,
3006 // initialize stack frame.
3008 MachineFunction &MF = DAG.getMachineFunction();
3009 MachineFrameInfo *MFI = MF.getFrameInfo();
3010 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
3011 unsigned RBegin, REnd;
3012 if (InRegsParamRecordIdx < CCInfo.getInRegsParamsCount()) {
3013 CCInfo.getInRegsParamInfo(InRegsParamRecordIdx, RBegin, REnd);
3015 unsigned RBeginIdx = CCInfo.getFirstUnallocated(GPRArgRegs);
3016 RBegin = RBeginIdx == 4 ? (unsigned)ARM::R4 : GPRArgRegs[RBeginIdx];
3021 ArgOffset = -4 * (ARM::R4 - RBegin);
3023 auto PtrVT = getPointerTy(DAG.getDataLayout());
3024 int FrameIndex = MFI->CreateFixedObject(ArgSize, ArgOffset, false);
3025 SDValue FIN = DAG.getFrameIndex(FrameIndex, PtrVT);
3027 SmallVector<SDValue, 4> MemOps;
3028 const TargetRegisterClass *RC =
3029 AFI->isThumb1OnlyFunction() ? &ARM::tGPRRegClass : &ARM::GPRRegClass;
3031 for (unsigned Reg = RBegin, i = 0; Reg < REnd; ++Reg, ++i) {
3032 unsigned VReg = MF.addLiveIn(Reg, RC);
3033 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
3035 DAG.getStore(Val.getValue(1), dl, Val, FIN,
3036 MachinePointerInfo(OrigArg, 4 * i), false, false, 0);
3037 MemOps.push_back(Store);
3038 FIN = DAG.getNode(ISD::ADD, dl, PtrVT, FIN, DAG.getConstant(4, dl, PtrVT));
3041 if (!MemOps.empty())
3042 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOps);
3046 // Setup stack frame, the va_list pointer will start from.
3048 ARMTargetLowering::VarArgStyleRegisters(CCState &CCInfo, SelectionDAG &DAG,
3049 SDLoc dl, SDValue &Chain,
3051 unsigned TotalArgRegsSaveSize,
3052 bool ForceMutable) const {
3053 MachineFunction &MF = DAG.getMachineFunction();
3054 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
3056 // Try to store any remaining integer argument regs
3057 // to their spots on the stack so that they may be loaded by deferencing
3058 // the result of va_next.
3059 // If there is no regs to be stored, just point address after last
3060 // argument passed via stack.
3061 int FrameIndex = StoreByValRegs(CCInfo, DAG, dl, Chain, nullptr,
3062 CCInfo.getInRegsParamsCount(),
3063 CCInfo.getNextStackOffset(), 4);
3064 AFI->setVarArgsFrameIndex(FrameIndex);
3068 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
3069 CallingConv::ID CallConv, bool isVarArg,
3070 const SmallVectorImpl<ISD::InputArg>
3072 SDLoc dl, SelectionDAG &DAG,
3073 SmallVectorImpl<SDValue> &InVals)
3075 MachineFunction &MF = DAG.getMachineFunction();
3076 MachineFrameInfo *MFI = MF.getFrameInfo();
3078 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
3080 // Assign locations to all of the incoming arguments.
3081 SmallVector<CCValAssign, 16> ArgLocs;
3082 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs,
3083 *DAG.getContext(), Prologue);
3084 CCInfo.AnalyzeFormalArguments(Ins,
3085 CCAssignFnForNode(CallConv, /* Return*/ false,
3088 SmallVector<SDValue, 16> ArgValues;
3090 Function::const_arg_iterator CurOrigArg = MF.getFunction()->arg_begin();
3091 unsigned CurArgIdx = 0;
3093 // Initially ArgRegsSaveSize is zero.
3094 // Then we increase this value each time we meet byval parameter.
3095 // We also increase this value in case of varargs function.
3096 AFI->setArgRegsSaveSize(0);
3098 // Calculate the amount of stack space that we need to allocate to store
3099 // byval and variadic arguments that are passed in registers.
3100 // We need to know this before we allocate the first byval or variadic
3101 // argument, as they will be allocated a stack slot below the CFA (Canonical
3102 // Frame Address, the stack pointer at entry to the function).
3103 unsigned ArgRegBegin = ARM::R4;
3104 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3105 if (CCInfo.getInRegsParamsProcessed() >= CCInfo.getInRegsParamsCount())
3108 CCValAssign &VA = ArgLocs[i];
3109 unsigned Index = VA.getValNo();
3110 ISD::ArgFlagsTy Flags = Ins[Index].Flags;
3111 if (!Flags.isByVal())
3114 assert(VA.isMemLoc() && "unexpected byval pointer in reg");
3115 unsigned RBegin, REnd;
3116 CCInfo.getInRegsParamInfo(CCInfo.getInRegsParamsProcessed(), RBegin, REnd);
3117 ArgRegBegin = std::min(ArgRegBegin, RBegin);
3119 CCInfo.nextInRegsParam();
3121 CCInfo.rewindByValRegsInfo();
3123 int lastInsIndex = -1;
3124 if (isVarArg && MFI->hasVAStart()) {
3125 unsigned RegIdx = CCInfo.getFirstUnallocated(GPRArgRegs);
3126 if (RegIdx != array_lengthof(GPRArgRegs))
3127 ArgRegBegin = std::min(ArgRegBegin, (unsigned)GPRArgRegs[RegIdx]);
3130 unsigned TotalArgRegsSaveSize = 4 * (ARM::R4 - ArgRegBegin);
3131 AFI->setArgRegsSaveSize(TotalArgRegsSaveSize);
3132 auto PtrVT = getPointerTy(DAG.getDataLayout());
3134 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3135 CCValAssign &VA = ArgLocs[i];
3136 if (Ins[VA.getValNo()].isOrigArg()) {
3137 std::advance(CurOrigArg,
3138 Ins[VA.getValNo()].getOrigArgIndex() - CurArgIdx);
3139 CurArgIdx = Ins[VA.getValNo()].getOrigArgIndex();
3141 // Arguments stored in registers.
3142 if (VA.isRegLoc()) {
3143 EVT RegVT = VA.getLocVT();
3145 if (VA.needsCustom()) {
3146 // f64 and vector types are split up into multiple registers or
3147 // combinations of registers and stack slots.
3148 if (VA.getLocVT() == MVT::v2f64) {
3149 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
3151 VA = ArgLocs[++i]; // skip ahead to next loc
3153 if (VA.isMemLoc()) {
3154 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true);
3155 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
3156 ArgValue2 = DAG.getLoad(
3157 MVT::f64, dl, Chain, FIN,
3158 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI),
3159 false, false, false, 0);
3161 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
3164 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
3165 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
3166 ArgValue, ArgValue1,
3167 DAG.getIntPtrConstant(0, dl));
3168 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
3169 ArgValue, ArgValue2,
3170 DAG.getIntPtrConstant(1, dl));
3172 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
3175 const TargetRegisterClass *RC;
3177 if (RegVT == MVT::f32)
3178 RC = &ARM::SPRRegClass;
3179 else if (RegVT == MVT::f64)
3180 RC = &ARM::DPRRegClass;
3181 else if (RegVT == MVT::v2f64)
3182 RC = &ARM::QPRRegClass;
3183 else if (RegVT == MVT::i32)
3184 RC = AFI->isThumb1OnlyFunction() ? &ARM::tGPRRegClass
3185 : &ARM::GPRRegClass;
3187 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
3189 // Transform the arguments in physical registers into virtual ones.
3190 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
3191 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
3194 // If this is an 8 or 16-bit value, it is really passed promoted
3195 // to 32 bits. Insert an assert[sz]ext to capture this, then
3196 // truncate to the right size.
3197 switch (VA.getLocInfo()) {
3198 default: llvm_unreachable("Unknown loc info!");
3199 case CCValAssign::Full: break;
3200 case CCValAssign::BCvt:
3201 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
3203 case CCValAssign::SExt:
3204 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
3205 DAG.getValueType(VA.getValVT()));
3206 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
3208 case CCValAssign::ZExt:
3209 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
3210 DAG.getValueType(VA.getValVT()));
3211 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
3215 InVals.push_back(ArgValue);
3217 } else { // VA.isRegLoc()
3220 assert(VA.isMemLoc());
3221 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
3223 int index = VA.getValNo();
3225 // Some Ins[] entries become multiple ArgLoc[] entries.
3226 // Process them only once.
3227 if (index != lastInsIndex)
3229 ISD::ArgFlagsTy Flags = Ins[index].Flags;
3230 // FIXME: For now, all byval parameter objects are marked mutable.
3231 // This can be changed with more analysis.
3232 // In case of tail call optimization mark all arguments mutable.
3233 // Since they could be overwritten by lowering of arguments in case of
3235 if (Flags.isByVal()) {
3236 assert(Ins[index].isOrigArg() &&
3237 "Byval arguments cannot be implicit");
3238 unsigned CurByValIndex = CCInfo.getInRegsParamsProcessed();
3240 int FrameIndex = StoreByValRegs(
3241 CCInfo, DAG, dl, Chain, &*CurOrigArg, CurByValIndex,
3242 VA.getLocMemOffset(), Flags.getByValSize());
3243 InVals.push_back(DAG.getFrameIndex(FrameIndex, PtrVT));
3244 CCInfo.nextInRegsParam();
3246 unsigned FIOffset = VA.getLocMemOffset();
3247 int FI = MFI->CreateFixedObject(VA.getLocVT().getSizeInBits()/8,
3250 // Create load nodes to retrieve arguments from the stack.
3251 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
3252 InVals.push_back(DAG.getLoad(
3253 VA.getValVT(), dl, Chain, FIN,
3254 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI),
3255 false, false, false, 0));
3257 lastInsIndex = index;
3263 if (isVarArg && MFI->hasVAStart())
3264 VarArgStyleRegisters(CCInfo, DAG, dl, Chain,
3265 CCInfo.getNextStackOffset(),
3266 TotalArgRegsSaveSize);
3268 AFI->setArgumentStackSize(CCInfo.getNextStackOffset());
3273 /// isFloatingPointZero - Return true if this is +0.0.
3274 static bool isFloatingPointZero(SDValue Op) {
3275 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
3276 return CFP->getValueAPF().isPosZero();
3277 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
3278 // Maybe this has already been legalized into the constant pool?
3279 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
3280 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
3281 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
3282 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
3283 return CFP->getValueAPF().isPosZero();
3285 } else if (Op->getOpcode() == ISD::BITCAST &&
3286 Op->getValueType(0) == MVT::f64) {
3287 // Handle (ISD::BITCAST (ARMISD::VMOVIMM (ISD::TargetConstant 0)) MVT::f64)
3288 // created by LowerConstantFP().
3289 SDValue BitcastOp = Op->getOperand(0);
3290 if (BitcastOp->getOpcode() == ARMISD::VMOVIMM) {
3291 SDValue MoveOp = BitcastOp->getOperand(0);
3292 if (MoveOp->getOpcode() == ISD::TargetConstant &&
3293 cast<ConstantSDNode>(MoveOp)->getZExtValue() == 0) {
3301 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
3302 /// the given operands.
3304 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
3305 SDValue &ARMcc, SelectionDAG &DAG,
3307 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
3308 unsigned C = RHSC->getZExtValue();
3309 if (!isLegalICmpImmediate(C)) {
3310 // Constant does not fit, try adjusting it by one?
3315 if (C != 0x80000000 && isLegalICmpImmediate(C-1)) {
3316 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
3317 RHS = DAG.getConstant(C - 1, dl, MVT::i32);
3322 if (C != 0 && isLegalICmpImmediate(C-1)) {
3323 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
3324 RHS = DAG.getConstant(C - 1, dl, MVT::i32);
3329 if (C != 0x7fffffff && isLegalICmpImmediate(C+1)) {
3330 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
3331 RHS = DAG.getConstant(C + 1, dl, MVT::i32);
3336 if (C != 0xffffffff && isLegalICmpImmediate(C+1)) {
3337 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
3338 RHS = DAG.getConstant(C + 1, dl, MVT::i32);
3345 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3346 ARMISD::NodeType CompareType;
3349 CompareType = ARMISD::CMP;
3354 CompareType = ARMISD::CMPZ;
3357 ARMcc = DAG.getConstant(CondCode, dl, MVT::i32);
3358 return DAG.getNode(CompareType, dl, MVT::Glue, LHS, RHS);
3361 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
3363 ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
3365 assert(!Subtarget->isFPOnlySP() || RHS.getValueType() != MVT::f64);
3367 if (!isFloatingPointZero(RHS))
3368 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Glue, LHS, RHS);
3370 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Glue, LHS);
3371 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Glue, Cmp);
3374 /// duplicateCmp - Glue values can have only one use, so this function
3375 /// duplicates a comparison node.
3377 ARMTargetLowering::duplicateCmp(SDValue Cmp, SelectionDAG &DAG) const {
3378 unsigned Opc = Cmp.getOpcode();
3380 if (Opc == ARMISD::CMP || Opc == ARMISD::CMPZ)
3381 return DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
3383 assert(Opc == ARMISD::FMSTAT && "unexpected comparison operation");
3384 Cmp = Cmp.getOperand(0);
3385 Opc = Cmp.getOpcode();
3386 if (Opc == ARMISD::CMPFP)
3387 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
3389 assert(Opc == ARMISD::CMPFPw0 && "unexpected operand of FMSTAT");
3390 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0));
3392 return DAG.getNode(ARMISD::FMSTAT, DL, MVT::Glue, Cmp);
3395 std::pair<SDValue, SDValue>
3396 ARMTargetLowering::getARMXALUOOp(SDValue Op, SelectionDAG &DAG,
3397 SDValue &ARMcc) const {
3398 assert(Op.getValueType() == MVT::i32 && "Unsupported value type");
3400 SDValue Value, OverflowCmp;
3401 SDValue LHS = Op.getOperand(0);
3402 SDValue RHS = Op.getOperand(1);
3405 // FIXME: We are currently always generating CMPs because we don't support
3406 // generating CMN through the backend. This is not as good as the natural
3407 // CMP case because it causes a register dependency and cannot be folded
3410 switch (Op.getOpcode()) {
3412 llvm_unreachable("Unknown overflow instruction!");
3414 ARMcc = DAG.getConstant(ARMCC::VC, dl, MVT::i32);
3415 Value = DAG.getNode(ISD::ADD, dl, Op.getValueType(), LHS, RHS);
3416 OverflowCmp = DAG.getNode(ARMISD::CMP, dl, MVT::Glue, Value, LHS);
3419 ARMcc = DAG.getConstant(ARMCC::HS, dl, MVT::i32);
3420 Value = DAG.getNode(ISD::ADD, dl, Op.getValueType(), LHS, RHS);
3421 OverflowCmp = DAG.getNode(ARMISD::CMP, dl, MVT::Glue, Value, LHS);
3424 ARMcc = DAG.getConstant(ARMCC::VC, dl, MVT::i32);
3425 Value = DAG.getNode(ISD::SUB, dl, Op.getValueType(), LHS, RHS);
3426 OverflowCmp = DAG.getNode(ARMISD::CMP, dl, MVT::Glue, LHS, RHS);
3429 ARMcc = DAG.getConstant(ARMCC::HS, dl, MVT::i32);
3430 Value = DAG.getNode(ISD::SUB, dl, Op.getValueType(), LHS, RHS);
3431 OverflowCmp = DAG.getNode(ARMISD::CMP, dl, MVT::Glue, LHS, RHS);
3435 return std::make_pair(Value, OverflowCmp);
3440 ARMTargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) const {
3441 // Let legalize expand this if it isn't a legal type yet.
3442 if (!DAG.getTargetLoweringInfo().isTypeLegal(Op.getValueType()))
3445 SDValue Value, OverflowCmp;
3447 std::tie(Value, OverflowCmp) = getARMXALUOOp(Op, DAG, ARMcc);
3448 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3450 // We use 0 and 1 as false and true values.
3451 SDValue TVal = DAG.getConstant(1, dl, MVT::i32);
3452 SDValue FVal = DAG.getConstant(0, dl, MVT::i32);
3453 EVT VT = Op.getValueType();
3455 SDValue Overflow = DAG.getNode(ARMISD::CMOV, dl, VT, TVal, FVal,
3456 ARMcc, CCR, OverflowCmp);
3458 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
3459 return DAG.getNode(ISD::MERGE_VALUES, dl, VTs, Value, Overflow);
3463 SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
3464 SDValue Cond = Op.getOperand(0);
3465 SDValue SelectTrue = Op.getOperand(1);
3466 SDValue SelectFalse = Op.getOperand(2);
3468 unsigned Opc = Cond.getOpcode();
3470 if (Cond.getResNo() == 1 &&
3471 (Opc == ISD::SADDO || Opc == ISD::UADDO || Opc == ISD::SSUBO ||
3472 Opc == ISD::USUBO)) {
3473 if (!DAG.getTargetLoweringInfo().isTypeLegal(Cond->getValueType(0)))
3476 SDValue Value, OverflowCmp;
3478 std::tie(Value, OverflowCmp) = getARMXALUOOp(Cond, DAG, ARMcc);
3479 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3480 EVT VT = Op.getValueType();
3482 return getCMOV(dl, VT, SelectTrue, SelectFalse, ARMcc, CCR,
3488 // (select (cmov 1, 0, cond), t, f) -> (cmov t, f, cond)
3489 // (select (cmov 0, 1, cond), t, f) -> (cmov f, t, cond)
3491 if (Cond.getOpcode() == ARMISD::CMOV && Cond.hasOneUse()) {
3492 const ConstantSDNode *CMOVTrue =
3493 dyn_cast<ConstantSDNode>(Cond.getOperand(0));
3494 const ConstantSDNode *CMOVFalse =
3495 dyn_cast<ConstantSDNode>(Cond.getOperand(1));
3497 if (CMOVTrue && CMOVFalse) {
3498 unsigned CMOVTrueVal = CMOVTrue->getZExtValue();
3499 unsigned CMOVFalseVal = CMOVFalse->getZExtValue();
3503 if (CMOVTrueVal == 1 && CMOVFalseVal == 0) {
3505 False = SelectFalse;
3506 } else if (CMOVTrueVal == 0 && CMOVFalseVal == 1) {
3511 if (True.getNode() && False.getNode()) {
3512 EVT VT = Op.getValueType();
3513 SDValue ARMcc = Cond.getOperand(2);
3514 SDValue CCR = Cond.getOperand(3);
3515 SDValue Cmp = duplicateCmp(Cond.getOperand(4), DAG);
3516 assert(True.getValueType() == VT);
3517 return getCMOV(dl, VT, True, False, ARMcc, CCR, Cmp, DAG);
3522 // ARM's BooleanContents value is UndefinedBooleanContent. Mask out the
3523 // undefined bits before doing a full-word comparison with zero.
3524 Cond = DAG.getNode(ISD::AND, dl, Cond.getValueType(), Cond,
3525 DAG.getConstant(1, dl, Cond.getValueType()));
3527 return DAG.getSelectCC(dl, Cond,
3528 DAG.getConstant(0, dl, Cond.getValueType()),
3529 SelectTrue, SelectFalse, ISD::SETNE);
3532 static void checkVSELConstraints(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
3533 bool &swpCmpOps, bool &swpVselOps) {
3534 // Start by selecting the GE condition code for opcodes that return true for
3536 if (CC == ISD::SETUGE || CC == ISD::SETOGE || CC == ISD::SETOLE ||
3538 CondCode = ARMCC::GE;
3540 // and GT for opcodes that return false for 'equality'.
3541 else if (CC == ISD::SETUGT || CC == ISD::SETOGT || CC == ISD::SETOLT ||
3543 CondCode = ARMCC::GT;
3545 // Since we are constrained to GE/GT, if the opcode contains 'less', we need
3546 // to swap the compare operands.
3547 if (CC == ISD::SETOLE || CC == ISD::SETULE || CC == ISD::SETOLT ||
3551 // Both GT and GE are ordered comparisons, and return false for 'unordered'.
3552 // If we have an unordered opcode, we need to swap the operands to the VSEL
3553 // instruction (effectively negating the condition).
3555 // This also has the effect of swapping which one of 'less' or 'greater'
3556 // returns true, so we also swap the compare operands. It also switches
3557 // whether we return true for 'equality', so we compensate by picking the
3558 // opposite condition code to our original choice.
3559 if (CC == ISD::SETULE || CC == ISD::SETULT || CC == ISD::SETUGE ||
3560 CC == ISD::SETUGT) {
3561 swpCmpOps = !swpCmpOps;
3562 swpVselOps = !swpVselOps;
3563 CondCode = CondCode == ARMCC::GT ? ARMCC::GE : ARMCC::GT;
3566 // 'ordered' is 'anything but unordered', so use the VS condition code and
3567 // swap the VSEL operands.
3568 if (CC == ISD::SETO) {
3569 CondCode = ARMCC::VS;
3573 // 'unordered or not equal' is 'anything but equal', so use the EQ condition
3574 // code and swap the VSEL operands.
3575 if (CC == ISD::SETUNE) {
3576 CondCode = ARMCC::EQ;
3581 SDValue ARMTargetLowering::getCMOV(SDLoc dl, EVT VT, SDValue FalseVal,
3582 SDValue TrueVal, SDValue ARMcc, SDValue CCR,
3583 SDValue Cmp, SelectionDAG &DAG) const {
3584 if (Subtarget->isFPOnlySP() && VT == MVT::f64) {
3585 FalseVal = DAG.getNode(ARMISD::VMOVRRD, dl,
3586 DAG.getVTList(MVT::i32, MVT::i32), FalseVal);
3587 TrueVal = DAG.getNode(ARMISD::VMOVRRD, dl,
3588 DAG.getVTList(MVT::i32, MVT::i32), TrueVal);
3590 SDValue TrueLow = TrueVal.getValue(0);
3591 SDValue TrueHigh = TrueVal.getValue(1);
3592 SDValue FalseLow = FalseVal.getValue(0);
3593 SDValue FalseHigh = FalseVal.getValue(1);
3595 SDValue Low = DAG.getNode(ARMISD::CMOV, dl, MVT::i32, FalseLow, TrueLow,
3597 SDValue High = DAG.getNode(ARMISD::CMOV, dl, MVT::i32, FalseHigh, TrueHigh,
3598 ARMcc, CCR, duplicateCmp(Cmp, DAG));
3600 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Low, High);
3602 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,
3607 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
3608 EVT VT = Op.getValueType();
3609 SDValue LHS = Op.getOperand(0);
3610 SDValue RHS = Op.getOperand(1);
3611 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
3612 SDValue TrueVal = Op.getOperand(2);
3613 SDValue FalseVal = Op.getOperand(3);
3616 if (Subtarget->isFPOnlySP() && LHS.getValueType() == MVT::f64) {
3617 DAG.getTargetLoweringInfo().softenSetCCOperands(DAG, MVT::f64, LHS, RHS, CC,
3620 // If softenSetCCOperands only returned one value, we should compare it to
3622 if (!RHS.getNode()) {
3623 RHS = DAG.getConstant(0, dl, LHS.getValueType());
3628 if (LHS.getValueType() == MVT::i32) {
3629 // Try to generate VSEL on ARMv8.
3630 // The VSEL instruction can't use all the usual ARM condition
3631 // codes: it only has two bits to select the condition code, so it's
3632 // constrained to use only GE, GT, VS and EQ.
3634 // To implement all the various ISD::SETXXX opcodes, we sometimes need to
3635 // swap the operands of the previous compare instruction (effectively
3636 // inverting the compare condition, swapping 'less' and 'greater') and
3637 // sometimes need to swap the operands to the VSEL (which inverts the
3638 // condition in the sense of firing whenever the previous condition didn't)
3639 if (Subtarget->hasFPARMv8() && (TrueVal.getValueType() == MVT::f32 ||
3640 TrueVal.getValueType() == MVT::f64)) {
3641 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3642 if (CondCode == ARMCC::LT || CondCode == ARMCC::LE ||
3643 CondCode == ARMCC::VC || CondCode == ARMCC::NE) {
3644 CC = ISD::getSetCCInverse(CC, true);
3645 std::swap(TrueVal, FalseVal);
3650 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3651 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3652 return getCMOV(dl, VT, FalseVal, TrueVal, ARMcc, CCR, Cmp, DAG);
3655 ARMCC::CondCodes CondCode, CondCode2;
3656 FPCCToARMCC(CC, CondCode, CondCode2);
3658 // Try to generate VMAXNM/VMINNM on ARMv8.
3659 if (Subtarget->hasFPARMv8() && (TrueVal.getValueType() == MVT::f32 ||
3660 TrueVal.getValueType() == MVT::f64)) {
3661 bool swpCmpOps = false;
3662 bool swpVselOps = false;
3663 checkVSELConstraints(CC, CondCode, swpCmpOps, swpVselOps);
3665 if (CondCode == ARMCC::GT || CondCode == ARMCC::GE ||
3666 CondCode == ARMCC::VS || CondCode == ARMCC::EQ) {
3668 std::swap(LHS, RHS);
3670 std::swap(TrueVal, FalseVal);
3674 SDValue ARMcc = DAG.getConstant(CondCode, dl, MVT::i32);
3675 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
3676 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3677 SDValue Result = getCMOV(dl, VT, FalseVal, TrueVal, ARMcc, CCR, Cmp, DAG);
3678 if (CondCode2 != ARMCC::AL) {
3679 SDValue ARMcc2 = DAG.getConstant(CondCode2, dl, MVT::i32);
3680 // FIXME: Needs another CMP because flag can have but one use.
3681 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
3682 Result = getCMOV(dl, VT, Result, TrueVal, ARMcc2, CCR, Cmp2, DAG);
3687 /// canChangeToInt - Given the fp compare operand, return true if it is suitable
3688 /// to morph to an integer compare sequence.
3689 static bool canChangeToInt(SDValue Op, bool &SeenZero,
3690 const ARMSubtarget *Subtarget) {
3691 SDNode *N = Op.getNode();
3692 if (!N->hasOneUse())
3693 // Otherwise it requires moving the value from fp to integer registers.
3695 if (!N->getNumValues())
3697 EVT VT = Op.getValueType();
3698 if (VT != MVT::f32 && !Subtarget->isFPBrccSlow())
3699 // f32 case is generally profitable. f64 case only makes sense when vcmpe +
3700 // vmrs are very slow, e.g. cortex-a8.
3703 if (isFloatingPointZero(Op)) {
3707 return ISD::isNormalLoad(N);
3710 static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) {
3711 if (isFloatingPointZero(Op))
3712 return DAG.getConstant(0, SDLoc(Op), MVT::i32);
3714 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op))
3715 return DAG.getLoad(MVT::i32, SDLoc(Op),
3716 Ld->getChain(), Ld->getBasePtr(), Ld->getPointerInfo(),
3717 Ld->isVolatile(), Ld->isNonTemporal(),
3718 Ld->isInvariant(), Ld->getAlignment());
3720 llvm_unreachable("Unknown VFP cmp argument!");
3723 static void expandf64Toi32(SDValue Op, SelectionDAG &DAG,
3724 SDValue &RetVal1, SDValue &RetVal2) {
3727 if (isFloatingPointZero(Op)) {
3728 RetVal1 = DAG.getConstant(0, dl, MVT::i32);
3729 RetVal2 = DAG.getConstant(0, dl, MVT::i32);
3733 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) {
3734 SDValue Ptr = Ld->getBasePtr();
3735 RetVal1 = DAG.getLoad(MVT::i32, dl,
3736 Ld->getChain(), Ptr,
3737 Ld->getPointerInfo(),
3738 Ld->isVolatile(), Ld->isNonTemporal(),
3739 Ld->isInvariant(), Ld->getAlignment());
3741 EVT PtrType = Ptr.getValueType();
3742 unsigned NewAlign = MinAlign(Ld->getAlignment(), 4);
3743 SDValue NewPtr = DAG.getNode(ISD::ADD, dl,
3744 PtrType, Ptr, DAG.getConstant(4, dl, PtrType));
3745 RetVal2 = DAG.getLoad(MVT::i32, dl,
3746 Ld->getChain(), NewPtr,
3747 Ld->getPointerInfo().getWithOffset(4),
3748 Ld->isVolatile(), Ld->isNonTemporal(),
3749 Ld->isInvariant(), NewAlign);
3753 llvm_unreachable("Unknown VFP cmp argument!");
3756 /// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some
3757 /// f32 and even f64 comparisons to integer ones.
3759 ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const {
3760 SDValue Chain = Op.getOperand(0);
3761 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3762 SDValue LHS = Op.getOperand(2);
3763 SDValue RHS = Op.getOperand(3);
3764 SDValue Dest = Op.getOperand(4);
3767 bool LHSSeenZero = false;
3768 bool LHSOk = canChangeToInt(LHS, LHSSeenZero, Subtarget);
3769 bool RHSSeenZero = false;
3770 bool RHSOk = canChangeToInt(RHS, RHSSeenZero, Subtarget);
3771 if (LHSOk && RHSOk && (LHSSeenZero || RHSSeenZero)) {
3772 // If unsafe fp math optimization is enabled and there are no other uses of
3773 // the CMP operands, and the condition code is EQ or NE, we can optimize it
3774 // to an integer comparison.
3775 if (CC == ISD::SETOEQ)
3777 else if (CC == ISD::SETUNE)
3780 SDValue Mask = DAG.getConstant(0x7fffffff, dl, MVT::i32);
3782 if (LHS.getValueType() == MVT::f32) {
3783 LHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3784 bitcastf32Toi32(LHS, DAG), Mask);
3785 RHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3786 bitcastf32Toi32(RHS, DAG), Mask);
3787 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3788 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3789 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3790 Chain, Dest, ARMcc, CCR, Cmp);
3795 expandf64Toi32(LHS, DAG, LHS1, LHS2);
3796 expandf64Toi32(RHS, DAG, RHS1, RHS2);
3797 LHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, LHS2, Mask);
3798 RHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, RHS2, Mask);
3799 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3800 ARMcc = DAG.getConstant(CondCode, dl, MVT::i32);
3801 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3802 SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest };
3803 return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops);
3809 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
3810 SDValue Chain = Op.getOperand(0);
3811 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3812 SDValue LHS = Op.getOperand(2);
3813 SDValue RHS = Op.getOperand(3);
3814 SDValue Dest = Op.getOperand(4);
3817 if (Subtarget->isFPOnlySP() && LHS.getValueType() == MVT::f64) {
3818 DAG.getTargetLoweringInfo().softenSetCCOperands(DAG, MVT::f64, LHS, RHS, CC,
3821 // If softenSetCCOperands only returned one value, we should compare it to
3823 if (!RHS.getNode()) {
3824 RHS = DAG.getConstant(0, dl, LHS.getValueType());
3829 if (LHS.getValueType() == MVT::i32) {
3831 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3832 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3833 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3834 Chain, Dest, ARMcc, CCR, Cmp);
3837 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3839 if (getTargetMachine().Options.UnsafeFPMath &&
3840 (CC == ISD::SETEQ || CC == ISD::SETOEQ ||
3841 CC == ISD::SETNE || CC == ISD::SETUNE)) {
3842 SDValue Result = OptimizeVFPBrcond(Op, DAG);
3843 if (Result.getNode())
3847 ARMCC::CondCodes CondCode, CondCode2;
3848 FPCCToARMCC(CC, CondCode, CondCode2);
3850 SDValue ARMcc = DAG.getConstant(CondCode, dl, MVT::i32);
3851 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
3852 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3853 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3854 SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp };
3855 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops);
3856 if (CondCode2 != ARMCC::AL) {
3857 ARMcc = DAG.getConstant(CondCode2, dl, MVT::i32);
3858 SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) };
3859 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops);
3864 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
3865 SDValue Chain = Op.getOperand(0);
3866 SDValue Table = Op.getOperand(1);
3867 SDValue Index = Op.getOperand(2);
3870 EVT PTy = getPointerTy(DAG.getDataLayout());
3871 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
3872 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
3873 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI);
3874 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, dl, PTy));
3875 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
3876 if (Subtarget->isThumb2()) {
3877 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
3878 // which does another jump to the destination. This also makes it easier
3879 // to translate it to TBB / TBH later.
3880 // FIXME: This might not work if the function is extremely large.
3881 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
3882 Addr, Op.getOperand(2), JTI);
3884 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
3886 DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
3887 MachinePointerInfo::getJumpTable(DAG.getMachineFunction()),
3888 false, false, false, 0);
3889 Chain = Addr.getValue(1);
3890 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
3891 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI);
3894 DAG.getLoad(PTy, dl, Chain, Addr,
3895 MachinePointerInfo::getJumpTable(DAG.getMachineFunction()),
3896 false, false, false, 0);
3897 Chain = Addr.getValue(1);
3898 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI);
3902 static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3903 EVT VT = Op.getValueType();
3906 if (Op.getValueType().getVectorElementType() == MVT::i32) {
3907 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::f32)
3909 return DAG.UnrollVectorOp(Op.getNode());
3912 assert(Op.getOperand(0).getValueType() == MVT::v4f32 &&
3913 "Invalid type for custom lowering!");
3914 if (VT != MVT::v4i16)
3915 return DAG.UnrollVectorOp(Op.getNode());
3917 Op = DAG.getNode(Op.getOpcode(), dl, MVT::v4i32, Op.getOperand(0));
3918 return DAG.getNode(ISD::TRUNCATE, dl, VT, Op);
3921 SDValue ARMTargetLowering::LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) const {
3922 EVT VT = Op.getValueType();
3924 return LowerVectorFP_TO_INT(Op, DAG);
3925 if (Subtarget->isFPOnlySP() && Op.getOperand(0).getValueType() == MVT::f64) {
3927 if (Op.getOpcode() == ISD::FP_TO_SINT)
3928 LC = RTLIB::getFPTOSINT(Op.getOperand(0).getValueType(),
3931 LC = RTLIB::getFPTOUINT(Op.getOperand(0).getValueType(),
3933 return makeLibCall(DAG, LC, Op.getValueType(), Op.getOperand(0),
3934 /*isSigned*/ false, SDLoc(Op)).first;
3940 static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3941 EVT VT = Op.getValueType();
3944 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::i32) {
3945 if (VT.getVectorElementType() == MVT::f32)
3947 return DAG.UnrollVectorOp(Op.getNode());
3950 assert(Op.getOperand(0).getValueType() == MVT::v4i16 &&
3951 "Invalid type for custom lowering!");
3952 if (VT != MVT::v4f32)
3953 return DAG.UnrollVectorOp(Op.getNode());
3957 switch (Op.getOpcode()) {
3958 default: llvm_unreachable("Invalid opcode!");
3959 case ISD::SINT_TO_FP:
3960 CastOpc = ISD::SIGN_EXTEND;
3961 Opc = ISD::SINT_TO_FP;
3963 case ISD::UINT_TO_FP:
3964 CastOpc = ISD::ZERO_EXTEND;
3965 Opc = ISD::UINT_TO_FP;
3969 Op = DAG.getNode(CastOpc, dl, MVT::v4i32, Op.getOperand(0));
3970 return DAG.getNode(Opc, dl, VT, Op);
3973 SDValue ARMTargetLowering::LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) const {
3974 EVT VT = Op.getValueType();
3976 return LowerVectorINT_TO_FP(Op, DAG);
3977 if (Subtarget->isFPOnlySP() && Op.getValueType() == MVT::f64) {
3979 if (Op.getOpcode() == ISD::SINT_TO_FP)
3980 LC = RTLIB::getSINTTOFP(Op.getOperand(0).getValueType(),
3983 LC = RTLIB::getUINTTOFP(Op.getOperand(0).getValueType(),
3985 return makeLibCall(DAG, LC, Op.getValueType(), Op.getOperand(0),
3986 /*isSigned*/ false, SDLoc(Op)).first;
3992 SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
3993 // Implement fcopysign with a fabs and a conditional fneg.
3994 SDValue Tmp0 = Op.getOperand(0);
3995 SDValue Tmp1 = Op.getOperand(1);
3997 EVT VT = Op.getValueType();
3998 EVT SrcVT = Tmp1.getValueType();
3999 bool InGPR = Tmp0.getOpcode() == ISD::BITCAST ||
4000 Tmp0.getOpcode() == ARMISD::VMOVDRR;
4001 bool UseNEON = !InGPR && Subtarget->hasNEON();
4004 // Use VBSL to copy the sign bit.
4005 unsigned EncodedVal = ARM_AM::createNEONModImm(0x6, 0x80);
4006 SDValue Mask = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v2i32,
4007 DAG.getTargetConstant(EncodedVal, dl, MVT::i32));
4008 EVT OpVT = (VT == MVT::f32) ? MVT::v2i32 : MVT::v1i64;
4010 Mask = DAG.getNode(ARMISD::VSHL, dl, OpVT,
4011 DAG.getNode(ISD::BITCAST, dl, OpVT, Mask),
4012 DAG.getConstant(32, dl, MVT::i32));
4013 else /*if (VT == MVT::f32)*/
4014 Tmp0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp0);
4015 if (SrcVT == MVT::f32) {
4016 Tmp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp1);
4018 Tmp1 = DAG.getNode(ARMISD::VSHL, dl, OpVT,
4019 DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1),
4020 DAG.getConstant(32, dl, MVT::i32));
4021 } else if (VT == MVT::f32)
4022 Tmp1 = DAG.getNode(ARMISD::VSHRu, dl, MVT::v1i64,
4023 DAG.getNode(ISD::BITCAST, dl, MVT::v1i64, Tmp1),
4024 DAG.getConstant(32, dl, MVT::i32));
4025 Tmp0 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp0);
4026 Tmp1 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1);
4028 SDValue AllOnes = DAG.getTargetConstant(ARM_AM::createNEONModImm(0xe, 0xff),
4030 AllOnes = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v8i8, AllOnes);
4031 SDValue MaskNot = DAG.getNode(ISD::XOR, dl, OpVT, Mask,
4032 DAG.getNode(ISD::BITCAST, dl, OpVT, AllOnes));
4034 SDValue Res = DAG.getNode(ISD::OR, dl, OpVT,
4035 DAG.getNode(ISD::AND, dl, OpVT, Tmp1, Mask),
4036 DAG.getNode(ISD::AND, dl, OpVT, Tmp0, MaskNot));
4037 if (VT == MVT::f32) {
4038 Res = DAG.getNode(ISD::BITCAST, dl, MVT::v2f32, Res);
4039 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, Res,
4040 DAG.getConstant(0, dl, MVT::i32));
4042 Res = DAG.getNode(ISD::BITCAST, dl, MVT::f64, Res);
4048 // Bitcast operand 1 to i32.
4049 if (SrcVT == MVT::f64)
4050 Tmp1 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
4052 Tmp1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp1);
4054 // Or in the signbit with integer operations.
4055 SDValue Mask1 = DAG.getConstant(0x80000000, dl, MVT::i32);
4056 SDValue Mask2 = DAG.getConstant(0x7fffffff, dl, MVT::i32);
4057 Tmp1 = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp1, Mask1);
4058 if (VT == MVT::f32) {
4059 Tmp0 = DAG.getNode(ISD::AND, dl, MVT::i32,
4060 DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp0), Mask2);
4061 return DAG.getNode(ISD::BITCAST, dl, MVT::f32,
4062 DAG.getNode(ISD::OR, dl, MVT::i32, Tmp0, Tmp1));
4065 // f64: Or the high part with signbit and then combine two parts.
4066 Tmp0 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
4068 SDValue Lo = Tmp0.getValue(0);
4069 SDValue Hi = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp0.getValue(1), Mask2);
4070 Hi = DAG.getNode(ISD::OR, dl, MVT::i32, Hi, Tmp1);
4071 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
4074 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
4075 MachineFunction &MF = DAG.getMachineFunction();
4076 MachineFrameInfo *MFI = MF.getFrameInfo();
4077 MFI->setReturnAddressIsTaken(true);
4079 if (verifyReturnAddressArgumentIsConstant(Op, DAG))
4082 EVT VT = Op.getValueType();
4084 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
4086 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
4087 SDValue Offset = DAG.getConstant(4, dl, MVT::i32);
4088 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
4089 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
4090 MachinePointerInfo(), false, false, false, 0);
4093 // Return LR, which contains the return address. Mark it an implicit live-in.
4094 unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32));
4095 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
4098 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
4099 const ARMBaseRegisterInfo &ARI =
4100 *static_cast<const ARMBaseRegisterInfo*>(RegInfo);
4101 MachineFunction &MF = DAG.getMachineFunction();
4102 MachineFrameInfo *MFI = MF.getFrameInfo();
4103 MFI->setFrameAddressIsTaken(true);
4105 EVT VT = Op.getValueType();
4106 SDLoc dl(Op); // FIXME probably not meaningful
4107 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
4108 unsigned FrameReg = ARI.getFrameRegister(MF);
4109 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
4111 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
4112 MachinePointerInfo(),
4113 false, false, false, 0);
4117 // FIXME? Maybe this could be a TableGen attribute on some registers and
4118 // this table could be generated automatically from RegInfo.
4119 unsigned ARMTargetLowering::getRegisterByName(const char* RegName, EVT VT,
4120 SelectionDAG &DAG) const {
4121 unsigned Reg = StringSwitch<unsigned>(RegName)
4122 .Case("sp", ARM::SP)
4126 report_fatal_error(Twine("Invalid register name \""
4127 + StringRef(RegName) + "\"."));
4130 // Result is 64 bit value so split into two 32 bit values and return as a
4132 static void ExpandREAD_REGISTER(SDNode *N, SmallVectorImpl<SDValue> &Results,
4133 SelectionDAG &DAG) {
4136 // This function is only supposed to be called for i64 type destination.
4137 assert(N->getValueType(0) == MVT::i64
4138 && "ExpandREAD_REGISTER called for non-i64 type result.");
4140 SDValue Read = DAG.getNode(ISD::READ_REGISTER, DL,
4141 DAG.getVTList(MVT::i32, MVT::i32, MVT::Other),
4145 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Read.getValue(0),
4147 Results.push_back(Read.getOperand(0));
4150 /// ExpandBITCAST - If the target supports VFP, this function is called to
4151 /// expand a bit convert where either the source or destination type is i64 to
4152 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
4153 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
4154 /// vectors), since the legalizer won't know what to do with that.
4155 static SDValue ExpandBITCAST(SDNode *N, SelectionDAG &DAG) {
4156 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4158 SDValue Op = N->getOperand(0);
4160 // This function is only supposed to be called for i64 types, either as the
4161 // source or destination of the bit convert.
4162 EVT SrcVT = Op.getValueType();
4163 EVT DstVT = N->getValueType(0);
4164 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
4165 "ExpandBITCAST called for non-i64 type");
4167 // Turn i64->f64 into VMOVDRR.
4168 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
4169 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
4170 DAG.getConstant(0, dl, MVT::i32));
4171 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
4172 DAG.getConstant(1, dl, MVT::i32));
4173 return DAG.getNode(ISD::BITCAST, dl, DstVT,
4174 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
4177 // Turn f64->i64 into VMOVRRD.
4178 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
4180 if (DAG.getDataLayout().isBigEndian() && SrcVT.isVector() &&
4181 SrcVT.getVectorNumElements() > 1)
4182 Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
4183 DAG.getVTList(MVT::i32, MVT::i32),
4184 DAG.getNode(ARMISD::VREV64, dl, SrcVT, Op));
4186 Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
4187 DAG.getVTList(MVT::i32, MVT::i32), Op);
4188 // Merge the pieces into a single i64 value.
4189 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
4195 /// getZeroVector - Returns a vector of specified type with all zero elements.
4196 /// Zero vectors are used to represent vector negation and in those cases
4197 /// will be implemented with the NEON VNEG instruction. However, VNEG does
4198 /// not support i64 elements, so sometimes the zero vectors will need to be
4199 /// explicitly constructed. Regardless, use a canonical VMOV to create the
4201 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, SDLoc dl) {
4202 assert(VT.isVector() && "Expected a vector type");
4203 // The canonical modified immediate encoding of a zero vector is....0!
4204 SDValue EncodedVal = DAG.getTargetConstant(0, dl, MVT::i32);
4205 EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
4206 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal);
4207 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4210 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
4211 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
4212 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
4213 SelectionDAG &DAG) const {
4214 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4215 EVT VT = Op.getValueType();
4216 unsigned VTBits = VT.getSizeInBits();
4218 SDValue ShOpLo = Op.getOperand(0);
4219 SDValue ShOpHi = Op.getOperand(1);
4220 SDValue ShAmt = Op.getOperand(2);
4222 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
4224 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
4226 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
4227 DAG.getConstant(VTBits, dl, MVT::i32), ShAmt);
4228 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
4229 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
4230 DAG.getConstant(VTBits, dl, MVT::i32));
4231 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
4232 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
4233 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
4235 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
4236 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, dl, MVT::i32),
4237 ISD::SETGE, ARMcc, DAG, dl);
4238 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
4239 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc,
4242 SDValue Ops[2] = { Lo, Hi };
4243 return DAG.getMergeValues(Ops, dl);
4246 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
4247 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
4248 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
4249 SelectionDAG &DAG) const {
4250 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4251 EVT VT = Op.getValueType();
4252 unsigned VTBits = VT.getSizeInBits();
4254 SDValue ShOpLo = Op.getOperand(0);
4255 SDValue ShOpHi = Op.getOperand(1);
4256 SDValue ShAmt = Op.getOperand(2);
4259 assert(Op.getOpcode() == ISD::SHL_PARTS);
4260 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
4261 DAG.getConstant(VTBits, dl, MVT::i32), ShAmt);
4262 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
4263 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
4264 DAG.getConstant(VTBits, dl, MVT::i32));
4265 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
4266 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
4268 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
4269 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
4270 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, dl, MVT::i32),
4271 ISD::SETGE, ARMcc, DAG, dl);
4272 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
4273 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc,
4276 SDValue Ops[2] = { Lo, Hi };
4277 return DAG.getMergeValues(Ops, dl);
4280 SDValue ARMTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
4281 SelectionDAG &DAG) const {
4282 // The rounding mode is in bits 23:22 of the FPSCR.
4283 // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0
4284 // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3)
4285 // so that the shift + and get folded into a bitfield extract.
4287 SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::i32,
4288 DAG.getConstant(Intrinsic::arm_get_fpscr, dl,
4290 SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPSCR,
4291 DAG.getConstant(1U << 22, dl, MVT::i32));
4292 SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds,
4293 DAG.getConstant(22, dl, MVT::i32));
4294 return DAG.getNode(ISD::AND, dl, MVT::i32, RMODE,
4295 DAG.getConstant(3, dl, MVT::i32));
4298 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
4299 const ARMSubtarget *ST) {
4301 EVT VT = N->getValueType(0);
4302 if (VT.isVector()) {
4303 assert(ST->hasNEON());
4305 // Compute the least significant set bit: LSB = X & -X
4306 SDValue X = N->getOperand(0);
4307 SDValue NX = DAG.getNode(ISD::SUB, dl, VT, getZeroVector(VT, DAG, dl), X);
4308 SDValue LSB = DAG.getNode(ISD::AND, dl, VT, X, NX);
4310 EVT ElemTy = VT.getVectorElementType();
4312 if (ElemTy == MVT::i8) {
4313 // Compute with: cttz(x) = ctpop(lsb - 1)
4314 SDValue One = DAG.getNode(ARMISD::VMOVIMM, dl, VT,
4315 DAG.getTargetConstant(1, dl, ElemTy));
4316 SDValue Bits = DAG.getNode(ISD::SUB, dl, VT, LSB, One);
4317 return DAG.getNode(ISD::CTPOP, dl, VT, Bits);
4320 if ((ElemTy == MVT::i16 || ElemTy == MVT::i32) &&
4321 (N->getOpcode() == ISD::CTTZ_ZERO_UNDEF)) {
4322 // Compute with: cttz(x) = (width - 1) - ctlz(lsb), if x != 0
4323 unsigned NumBits = ElemTy.getSizeInBits();
4324 SDValue WidthMinus1 =
4325 DAG.getNode(ARMISD::VMOVIMM, dl, VT,
4326 DAG.getTargetConstant(NumBits - 1, dl, ElemTy));
4327 SDValue CTLZ = DAG.getNode(ISD::CTLZ, dl, VT, LSB);
4328 return DAG.getNode(ISD::SUB, dl, VT, WidthMinus1, CTLZ);
4331 // Compute with: cttz(x) = ctpop(lsb - 1)
4333 // Since we can only compute the number of bits in a byte with vcnt.8, we
4334 // have to gather the result with pairwise addition (vpaddl) for i16, i32,
4339 if (ElemTy == MVT::i64) {
4340 // Load constant 0xffff'ffff'ffff'ffff to register.
4341 SDValue FF = DAG.getNode(ARMISD::VMOVIMM, dl, VT,
4342 DAG.getTargetConstant(0x1eff, dl, MVT::i32));
4343 Bits = DAG.getNode(ISD::ADD, dl, VT, LSB, FF);
4345 SDValue One = DAG.getNode(ARMISD::VMOVIMM, dl, VT,
4346 DAG.getTargetConstant(1, dl, ElemTy));
4347 Bits = DAG.getNode(ISD::SUB, dl, VT, LSB, One);
4350 // Count #bits with vcnt.8.
4351 EVT VT8Bit = VT.is64BitVector() ? MVT::v8i8 : MVT::v16i8;
4352 SDValue BitsVT8 = DAG.getNode(ISD::BITCAST, dl, VT8Bit, Bits);
4353 SDValue Cnt8 = DAG.getNode(ISD::CTPOP, dl, VT8Bit, BitsVT8);
4355 // Gather the #bits with vpaddl (pairwise add.)
4356 EVT VT16Bit = VT.is64BitVector() ? MVT::v4i16 : MVT::v8i16;
4357 SDValue Cnt16 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT16Bit,
4358 DAG.getTargetConstant(Intrinsic::arm_neon_vpaddlu, dl, MVT::i32),
4360 if (ElemTy == MVT::i16)
4363 EVT VT32Bit = VT.is64BitVector() ? MVT::v2i32 : MVT::v4i32;
4364 SDValue Cnt32 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT32Bit,
4365 DAG.getTargetConstant(Intrinsic::arm_neon_vpaddlu, dl, MVT::i32),
4367 if (ElemTy == MVT::i32)
4370 assert(ElemTy == MVT::i64);
4371 SDValue Cnt64 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
4372 DAG.getTargetConstant(Intrinsic::arm_neon_vpaddlu, dl, MVT::i32),
4377 if (!ST->hasV6T2Ops())
4380 SDValue rbit = DAG.getNode(ISD::BITREVERSE, dl, VT, N->getOperand(0));
4381 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
4384 /// getCTPOP16BitCounts - Returns a v8i8/v16i8 vector containing the bit-count
4385 /// for each 16-bit element from operand, repeated. The basic idea is to
4386 /// leverage vcnt to get the 8-bit counts, gather and add the results.
4388 /// Trace for v4i16:
4389 /// input = [v0 v1 v2 v3 ] (vi 16-bit element)
4390 /// cast: N0 = [w0 w1 w2 w3 w4 w5 w6 w7] (v0 = [w0 w1], wi 8-bit element)
4391 /// vcnt: N1 = [b0 b1 b2 b3 b4 b5 b6 b7] (bi = bit-count of 8-bit element wi)
4392 /// vrev: N2 = [b1 b0 b3 b2 b5 b4 b7 b6]
4393 /// [b0 b1 b2 b3 b4 b5 b6 b7]
4394 /// +[b1 b0 b3 b2 b5 b4 b7 b6]
4395 /// N3=N1+N2 = [k0 k0 k1 k1 k2 k2 k3 k3] (k0 = b0+b1 = bit-count of 16-bit v0,
4396 /// vuzp: = [k0 k1 k2 k3 k0 k1 k2 k3] each ki is 8-bits)
4397 static SDValue getCTPOP16BitCounts(SDNode *N, SelectionDAG &DAG) {
4398 EVT VT = N->getValueType(0);
4401 EVT VT8Bit = VT.is64BitVector() ? MVT::v8i8 : MVT::v16i8;
4402 SDValue N0 = DAG.getNode(ISD::BITCAST, DL, VT8Bit, N->getOperand(0));
4403 SDValue N1 = DAG.getNode(ISD::CTPOP, DL, VT8Bit, N0);
4404 SDValue N2 = DAG.getNode(ARMISD::VREV16, DL, VT8Bit, N1);
4405 SDValue N3 = DAG.getNode(ISD::ADD, DL, VT8Bit, N1, N2);
4406 return DAG.getNode(ARMISD::VUZP, DL, VT8Bit, N3, N3);
4409 /// lowerCTPOP16BitElements - Returns a v4i16/v8i16 vector containing the
4410 /// bit-count for each 16-bit element from the operand. We need slightly
4411 /// different sequencing for v4i16 and v8i16 to stay within NEON's available
4412 /// 64/128-bit registers.
4414 /// Trace for v4i16:
4415 /// input = [v0 v1 v2 v3 ] (vi 16-bit element)
4416 /// v8i8: BitCounts = [k0 k1 k2 k3 k0 k1 k2 k3 ] (ki is the bit-count of vi)
4417 /// v8i16:Extended = [k0 k1 k2 k3 k0 k1 k2 k3 ]
4418 /// v4i16:Extracted = [k0 k1 k2 k3 ]
4419 static SDValue lowerCTPOP16BitElements(SDNode *N, SelectionDAG &DAG) {
4420 EVT VT = N->getValueType(0);
4423 SDValue BitCounts = getCTPOP16BitCounts(N, DAG);
4424 if (VT.is64BitVector()) {
4425 SDValue Extended = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v8i16, BitCounts);
4426 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i16, Extended,
4427 DAG.getIntPtrConstant(0, DL));
4429 SDValue Extracted = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v8i8,
4430 BitCounts, DAG.getIntPtrConstant(0, DL));
4431 return DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v8i16, Extracted);
4435 /// lowerCTPOP32BitElements - Returns a v2i32/v4i32 vector containing the
4436 /// bit-count for each 32-bit element from the operand. The idea here is
4437 /// to split the vector into 16-bit elements, leverage the 16-bit count
4438 /// routine, and then combine the results.
4440 /// Trace for v2i32 (v4i32 similar with Extracted/Extended exchanged):
4441 /// input = [v0 v1 ] (vi: 32-bit elements)
4442 /// Bitcast = [w0 w1 w2 w3 ] (wi: 16-bit elements, v0 = [w0 w1])
4443 /// Counts16 = [k0 k1 k2 k3 ] (ki: 16-bit elements, bit-count of wi)
4444 /// vrev: N0 = [k1 k0 k3 k2 ]
4446 /// N1 =+[k1 k0 k3 k2 ]
4448 /// N2 =+[k1 k3 k0 k2 ]
4450 /// Extended =+[k1 k3 k0 k2 ]
4452 /// Extracted=+[k1 k3 ]
4454 static SDValue lowerCTPOP32BitElements(SDNode *N, SelectionDAG &DAG) {
4455 EVT VT = N->getValueType(0);
4458 EVT VT16Bit = VT.is64BitVector() ? MVT::v4i16 : MVT::v8i16;
4460 SDValue Bitcast = DAG.getNode(ISD::BITCAST, DL, VT16Bit, N->getOperand(0));
4461 SDValue Counts16 = lowerCTPOP16BitElements(Bitcast.getNode(), DAG);
4462 SDValue N0 = DAG.getNode(ARMISD::VREV32, DL, VT16Bit, Counts16);
4463 SDValue N1 = DAG.getNode(ISD::ADD, DL, VT16Bit, Counts16, N0);
4464 SDValue N2 = DAG.getNode(ARMISD::VUZP, DL, VT16Bit, N1, N1);
4466 if (VT.is64BitVector()) {
4467 SDValue Extended = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v4i32, N2);
4468 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i32, Extended,
4469 DAG.getIntPtrConstant(0, DL));
4471 SDValue Extracted = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i16, N2,
4472 DAG.getIntPtrConstant(0, DL));
4473 return DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v4i32, Extracted);
4477 static SDValue LowerCTPOP(SDNode *N, SelectionDAG &DAG,
4478 const ARMSubtarget *ST) {
4479 EVT VT = N->getValueType(0);
4481 assert(ST->hasNEON() && "Custom ctpop lowering requires NEON.");
4482 assert((VT == MVT::v2i32 || VT == MVT::v4i32 ||
4483 VT == MVT::v4i16 || VT == MVT::v8i16) &&
4484 "Unexpected type for custom ctpop lowering");
4486 if (VT.getVectorElementType() == MVT::i32)
4487 return lowerCTPOP32BitElements(N, DAG);
4489 return lowerCTPOP16BitElements(N, DAG);
4492 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
4493 const ARMSubtarget *ST) {
4494 EVT VT = N->getValueType(0);
4500 // Lower vector shifts on NEON to use VSHL.
4501 assert(ST->hasNEON() && "unexpected vector shift");
4503 // Left shifts translate directly to the vshiftu intrinsic.
4504 if (N->getOpcode() == ISD::SHL)
4505 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
4506 DAG.getConstant(Intrinsic::arm_neon_vshiftu, dl,
4508 N->getOperand(0), N->getOperand(1));
4510 assert((N->getOpcode() == ISD::SRA ||
4511 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
4513 // NEON uses the same intrinsics for both left and right shifts. For
4514 // right shifts, the shift amounts are negative, so negate the vector of
4516 EVT ShiftVT = N->getOperand(1).getValueType();
4517 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
4518 getZeroVector(ShiftVT, DAG, dl),
4520 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
4521 Intrinsic::arm_neon_vshifts :
4522 Intrinsic::arm_neon_vshiftu);
4523 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
4524 DAG.getConstant(vshiftInt, dl, MVT::i32),
4525 N->getOperand(0), NegatedCount);
4528 static SDValue Expand64BitShift(SDNode *N, SelectionDAG &DAG,
4529 const ARMSubtarget *ST) {
4530 EVT VT = N->getValueType(0);
4533 // We can get here for a node like i32 = ISD::SHL i32, i64
4537 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
4538 "Unknown shift to lower!");
4540 // We only lower SRA, SRL of 1 here, all others use generic lowering.
4541 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
4542 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
4545 // If we are in thumb mode, we don't have RRX.
4546 if (ST->isThumb1Only()) return SDValue();
4548 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
4549 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
4550 DAG.getConstant(0, dl, MVT::i32));
4551 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
4552 DAG.getConstant(1, dl, MVT::i32));
4554 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
4555 // captures the result into a carry flag.
4556 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
4557 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Glue), Hi);
4559 // The low part is an ARMISD::RRX operand, which shifts the carry in.
4560 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
4562 // Merge the pieces into a single i64 value.
4563 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
4566 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
4567 SDValue TmpOp0, TmpOp1;
4568 bool Invert = false;
4572 SDValue Op0 = Op.getOperand(0);
4573 SDValue Op1 = Op.getOperand(1);
4574 SDValue CC = Op.getOperand(2);
4575 EVT CmpVT = Op0.getValueType().changeVectorElementTypeToInteger();
4576 EVT VT = Op.getValueType();
4577 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
4580 if (CmpVT.getVectorElementType() == MVT::i64)
4581 // 64-bit comparisons are not legal. We've marked SETCC as non-Custom,
4582 // but it's possible that our operands are 64-bit but our result is 32-bit.
4583 // Bail in this case.
4586 if (Op1.getValueType().isFloatingPoint()) {
4587 switch (SetCCOpcode) {
4588 default: llvm_unreachable("Illegal FP comparison");
4590 case ISD::SETNE: Invert = true; // Fallthrough
4592 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
4594 case ISD::SETLT: Swap = true; // Fallthrough
4596 case ISD::SETGT: Opc = ARMISD::VCGT; break;
4598 case ISD::SETLE: Swap = true; // Fallthrough
4600 case ISD::SETGE: Opc = ARMISD::VCGE; break;
4601 case ISD::SETUGE: Swap = true; // Fallthrough
4602 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
4603 case ISD::SETUGT: Swap = true; // Fallthrough
4604 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
4605 case ISD::SETUEQ: Invert = true; // Fallthrough
4607 // Expand this to (OLT | OGT).
4611 Op0 = DAG.getNode(ARMISD::VCGT, dl, CmpVT, TmpOp1, TmpOp0);
4612 Op1 = DAG.getNode(ARMISD::VCGT, dl, CmpVT, TmpOp0, TmpOp1);
4614 case ISD::SETUO: Invert = true; // Fallthrough
4616 // Expand this to (OLT | OGE).
4620 Op0 = DAG.getNode(ARMISD::VCGT, dl, CmpVT, TmpOp1, TmpOp0);
4621 Op1 = DAG.getNode(ARMISD::VCGE, dl, CmpVT, TmpOp0, TmpOp1);
4625 // Integer comparisons.
4626 switch (SetCCOpcode) {
4627 default: llvm_unreachable("Illegal integer comparison");
4628 case ISD::SETNE: Invert = true;
4629 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
4630 case ISD::SETLT: Swap = true;
4631 case ISD::SETGT: Opc = ARMISD::VCGT; break;
4632 case ISD::SETLE: Swap = true;
4633 case ISD::SETGE: Opc = ARMISD::VCGE; break;
4634 case ISD::SETULT: Swap = true;
4635 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
4636 case ISD::SETULE: Swap = true;
4637 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
4640 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
4641 if (Opc == ARMISD::VCEQ) {
4644 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
4646 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
4649 // Ignore bitconvert.
4650 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BITCAST)
4651 AndOp = AndOp.getOperand(0);
4653 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
4655 Op0 = DAG.getNode(ISD::BITCAST, dl, CmpVT, AndOp.getOperand(0));
4656 Op1 = DAG.getNode(ISD::BITCAST, dl, CmpVT, AndOp.getOperand(1));
4663 std::swap(Op0, Op1);
4665 // If one of the operands is a constant vector zero, attempt to fold the
4666 // comparison to a specialized compare-against-zero form.
4668 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
4670 else if (ISD::isBuildVectorAllZeros(Op0.getNode())) {
4671 if (Opc == ARMISD::VCGE)
4672 Opc = ARMISD::VCLEZ;
4673 else if (Opc == ARMISD::VCGT)
4674 Opc = ARMISD::VCLTZ;
4679 if (SingleOp.getNode()) {
4682 Result = DAG.getNode(ARMISD::VCEQZ, dl, CmpVT, SingleOp); break;
4684 Result = DAG.getNode(ARMISD::VCGEZ, dl, CmpVT, SingleOp); break;
4686 Result = DAG.getNode(ARMISD::VCLEZ, dl, CmpVT, SingleOp); break;
4688 Result = DAG.getNode(ARMISD::VCGTZ, dl, CmpVT, SingleOp); break;
4690 Result = DAG.getNode(ARMISD::VCLTZ, dl, CmpVT, SingleOp); break;
4692 Result = DAG.getNode(Opc, dl, CmpVT, Op0, Op1);
4695 Result = DAG.getNode(Opc, dl, CmpVT, Op0, Op1);
4698 Result = DAG.getSExtOrTrunc(Result, dl, VT);
4701 Result = DAG.getNOT(dl, Result, VT);
4706 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
4707 /// valid vector constant for a NEON instruction with a "modified immediate"
4708 /// operand (e.g., VMOV). If so, return the encoded value.
4709 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
4710 unsigned SplatBitSize, SelectionDAG &DAG,
4711 SDLoc dl, EVT &VT, bool is128Bits,
4712 NEONModImmType type) {
4713 unsigned OpCmode, Imm;
4715 // SplatBitSize is set to the smallest size that splats the vector, so a
4716 // zero vector will always have SplatBitSize == 8. However, NEON modified
4717 // immediate instructions others than VMOV do not support the 8-bit encoding
4718 // of a zero vector, and the default encoding of zero is supposed to be the
4723 switch (SplatBitSize) {
4725 if (type != VMOVModImm)
4727 // Any 1-byte value is OK. Op=0, Cmode=1110.
4728 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
4731 VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
4735 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
4736 VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
4737 if ((SplatBits & ~0xff) == 0) {
4738 // Value = 0x00nn: Op=x, Cmode=100x.
4743 if ((SplatBits & ~0xff00) == 0) {
4744 // Value = 0xnn00: Op=x, Cmode=101x.
4746 Imm = SplatBits >> 8;
4752 // NEON's 32-bit VMOV supports splat values where:
4753 // * only one byte is nonzero, or
4754 // * the least significant byte is 0xff and the second byte is nonzero, or
4755 // * the least significant 2 bytes are 0xff and the third is nonzero.
4756 VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
4757 if ((SplatBits & ~0xff) == 0) {
4758 // Value = 0x000000nn: Op=x, Cmode=000x.
4763 if ((SplatBits & ~0xff00) == 0) {
4764 // Value = 0x0000nn00: Op=x, Cmode=001x.
4766 Imm = SplatBits >> 8;
4769 if ((SplatBits & ~0xff0000) == 0) {
4770 // Value = 0x00nn0000: Op=x, Cmode=010x.
4772 Imm = SplatBits >> 16;
4775 if ((SplatBits & ~0xff000000) == 0) {
4776 // Value = 0xnn000000: Op=x, Cmode=011x.
4778 Imm = SplatBits >> 24;
4782 // cmode == 0b1100 and cmode == 0b1101 are not supported for VORR or VBIC
4783 if (type == OtherModImm) return SDValue();
4785 if ((SplatBits & ~0xffff) == 0 &&
4786 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
4787 // Value = 0x0000nnff: Op=x, Cmode=1100.
4789 Imm = SplatBits >> 8;
4793 if ((SplatBits & ~0xffffff) == 0 &&
4794 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
4795 // Value = 0x00nnffff: Op=x, Cmode=1101.
4797 Imm = SplatBits >> 16;
4801 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
4802 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
4803 // VMOV.I32. A (very) minor optimization would be to replicate the value
4804 // and fall through here to test for a valid 64-bit splat. But, then the
4805 // caller would also need to check and handle the change in size.
4809 if (type != VMOVModImm)
4811 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
4812 uint64_t BitMask = 0xff;
4814 unsigned ImmMask = 1;
4816 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
4817 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
4820 } else if ((SplatBits & BitMask) != 0) {
4827 if (DAG.getDataLayout().isBigEndian())
4828 // swap higher and lower 32 bit word
4829 Imm = ((Imm & 0xf) << 4) | ((Imm & 0xf0) >> 4);
4831 // Op=1, Cmode=1110.
4833 VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
4838 llvm_unreachable("unexpected size for isNEONModifiedImm");
4841 unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm);
4842 return DAG.getTargetConstant(EncodedVal, dl, MVT::i32);
4845 SDValue ARMTargetLowering::LowerConstantFP(SDValue Op, SelectionDAG &DAG,
4846 const ARMSubtarget *ST) const {
4850 bool IsDouble = Op.getValueType() == MVT::f64;
4851 ConstantFPSDNode *CFP = cast<ConstantFPSDNode>(Op);
4853 // Use the default (constant pool) lowering for double constants when we have
4855 if (IsDouble && Subtarget->isFPOnlySP())
4858 // Try splatting with a VMOV.f32...
4859 APFloat FPVal = CFP->getValueAPF();
4860 int ImmVal = IsDouble ? ARM_AM::getFP64Imm(FPVal) : ARM_AM::getFP32Imm(FPVal);
4863 if (IsDouble || !ST->useNEONForSinglePrecisionFP()) {
4864 // We have code in place to select a valid ConstantFP already, no need to
4869 // It's a float and we are trying to use NEON operations where
4870 // possible. Lower it to a splat followed by an extract.
4872 SDValue NewVal = DAG.getTargetConstant(ImmVal, DL, MVT::i32);
4873 SDValue VecConstant = DAG.getNode(ARMISD::VMOVFPIMM, DL, MVT::v2f32,
4875 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecConstant,
4876 DAG.getConstant(0, DL, MVT::i32));
4879 // The rest of our options are NEON only, make sure that's allowed before
4881 if (!ST->hasNEON() || (!IsDouble && !ST->useNEONForSinglePrecisionFP()))
4885 uint64_t iVal = FPVal.bitcastToAPInt().getZExtValue();
4887 // It wouldn't really be worth bothering for doubles except for one very
4888 // important value, which does happen to match: 0.0. So make sure we don't do
4890 if (IsDouble && (iVal & 0xffffffff) != (iVal >> 32))
4893 // Try a VMOV.i32 (FIXME: i8, i16, or i64 could work too).
4894 SDValue NewVal = isNEONModifiedImm(iVal & 0xffffffffU, 0, 32, DAG, SDLoc(Op),
4895 VMovVT, false, VMOVModImm);
4896 if (NewVal != SDValue()) {
4898 SDValue VecConstant = DAG.getNode(ARMISD::VMOVIMM, DL, VMovVT,
4901 return DAG.getNode(ISD::BITCAST, DL, MVT::f64, VecConstant);
4903 // It's a float: cast and extract a vector element.
4904 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
4906 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
4907 DAG.getConstant(0, DL, MVT::i32));
4910 // Finally, try a VMVN.i32
4911 NewVal = isNEONModifiedImm(~iVal & 0xffffffffU, 0, 32, DAG, SDLoc(Op), VMovVT,
4913 if (NewVal != SDValue()) {
4915 SDValue VecConstant = DAG.getNode(ARMISD::VMVNIMM, DL, VMovVT, NewVal);
4918 return DAG.getNode(ISD::BITCAST, DL, MVT::f64, VecConstant);
4920 // It's a float: cast and extract a vector element.
4921 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
4923 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
4924 DAG.getConstant(0, DL, MVT::i32));
4930 // check if an VEXT instruction can handle the shuffle mask when the
4931 // vector sources of the shuffle are the same.
4932 static bool isSingletonVEXTMask(ArrayRef<int> M, EVT VT, unsigned &Imm) {
4933 unsigned NumElts = VT.getVectorNumElements();
4935 // Assume that the first shuffle index is not UNDEF. Fail if it is.
4941 // If this is a VEXT shuffle, the immediate value is the index of the first
4942 // element. The other shuffle indices must be the successive elements after
4944 unsigned ExpectedElt = Imm;
4945 for (unsigned i = 1; i < NumElts; ++i) {
4946 // Increment the expected index. If it wraps around, just follow it
4947 // back to index zero and keep going.
4949 if (ExpectedElt == NumElts)
4952 if (M[i] < 0) continue; // ignore UNDEF indices
4953 if (ExpectedElt != static_cast<unsigned>(M[i]))
4961 static bool isVEXTMask(ArrayRef<int> M, EVT VT,
4962 bool &ReverseVEXT, unsigned &Imm) {
4963 unsigned NumElts = VT.getVectorNumElements();
4964 ReverseVEXT = false;
4966 // Assume that the first shuffle index is not UNDEF. Fail if it is.
4972 // If this is a VEXT shuffle, the immediate value is the index of the first
4973 // element. The other shuffle indices must be the successive elements after
4975 unsigned ExpectedElt = Imm;
4976 for (unsigned i = 1; i < NumElts; ++i) {
4977 // Increment the expected index. If it wraps around, it may still be
4978 // a VEXT but the source vectors must be swapped.
4980 if (ExpectedElt == NumElts * 2) {
4985 if (M[i] < 0) continue; // ignore UNDEF indices
4986 if (ExpectedElt != static_cast<unsigned>(M[i]))
4990 // Adjust the index value if the source operands will be swapped.
4997 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
4998 /// instruction with the specified blocksize. (The order of the elements
4999 /// within each block of the vector is reversed.)
5000 static bool isVREVMask(ArrayRef<int> M, EVT VT, unsigned BlockSize) {
5001 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
5002 "Only possible block sizes for VREV are: 16, 32, 64");
5004 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
5008 unsigned NumElts = VT.getVectorNumElements();
5009 unsigned BlockElts = M[0] + 1;
5010 // If the first shuffle index is UNDEF, be optimistic.
5012 BlockElts = BlockSize / EltSz;
5014 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
5017 for (unsigned i = 0; i < NumElts; ++i) {
5018 if (M[i] < 0) continue; // ignore UNDEF indices
5019 if ((unsigned) M[i] != (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
5026 static bool isVTBLMask(ArrayRef<int> M, EVT VT) {
5027 // We can handle <8 x i8> vector shuffles. If the index in the mask is out of
5028 // range, then 0 is placed into the resulting vector. So pretty much any mask
5029 // of 8 elements can work here.
5030 return VT == MVT::v8i8 && M.size() == 8;
5033 // Checks whether the shuffle mask represents a vector transpose (VTRN) by
5034 // checking that pairs of elements in the shuffle mask represent the same index
5035 // in each vector, incrementing the expected index by 2 at each step.
5036 // e.g. For v1,v2 of type v4i32 a valid shuffle mask is: [0, 4, 2, 6]
5037 // v1={a,b,c,d} => x=shufflevector v1, v2 shufflemask => x={a,e,c,g}
5039 // WhichResult gives the offset for each element in the mask based on which
5040 // of the two results it belongs to.
5042 // The transpose can be represented either as:
5043 // result1 = shufflevector v1, v2, result1_shuffle_mask
5044 // result2 = shufflevector v1, v2, result2_shuffle_mask
5045 // where v1/v2 and the shuffle masks have the same number of elements
5046 // (here WhichResult (see below) indicates which result is being checked)
5049 // results = shufflevector v1, v2, shuffle_mask
5050 // where both results are returned in one vector and the shuffle mask has twice
5051 // as many elements as v1/v2 (here WhichResult will always be 0 if true) here we
5052 // want to check the low half and high half of the shuffle mask as if it were
5054 static bool isVTRNMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
5055 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
5059 unsigned NumElts = VT.getVectorNumElements();
5060 if (M.size() != NumElts && M.size() != NumElts*2)
5063 // If the mask is twice as long as the input vector then we need to check the
5064 // upper and lower parts of the mask with a matching value for WhichResult
5065 // FIXME: A mask with only even values will be rejected in case the first
5066 // element is undefined, e.g. [-1, 4, 2, 6] will be rejected, because only
5067 // M[0] is used to determine WhichResult
5068 for (unsigned i = 0; i < M.size(); i += NumElts) {
5069 if (M.size() == NumElts * 2)
5070 WhichResult = i / NumElts;
5072 WhichResult = M[i] == 0 ? 0 : 1;
5073 for (unsigned j = 0; j < NumElts; j += 2) {
5074 if ((M[i+j] >= 0 && (unsigned) M[i+j] != j + WhichResult) ||
5075 (M[i+j+1] >= 0 && (unsigned) M[i+j+1] != j + NumElts + WhichResult))
5080 if (M.size() == NumElts*2)
5086 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
5087 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
5088 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
5089 static bool isVTRN_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
5090 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
5094 unsigned NumElts = VT.getVectorNumElements();
5095 if (M.size() != NumElts && M.size() != NumElts*2)
5098 for (unsigned i = 0; i < M.size(); i += NumElts) {
5099 if (M.size() == NumElts * 2)
5100 WhichResult = i / NumElts;
5102 WhichResult = M[i] == 0 ? 0 : 1;
5103 for (unsigned j = 0; j < NumElts; j += 2) {
5104 if ((M[i+j] >= 0 && (unsigned) M[i+j] != j + WhichResult) ||
5105 (M[i+j+1] >= 0 && (unsigned) M[i+j+1] != j + WhichResult))
5110 if (M.size() == NumElts*2)
5116 // Checks whether the shuffle mask represents a vector unzip (VUZP) by checking
5117 // that the mask elements are either all even and in steps of size 2 or all odd
5118 // and in steps of size 2.
5119 // e.g. For v1,v2 of type v4i32 a valid shuffle mask is: [0, 2, 4, 6]
5120 // v1={a,b,c,d} => x=shufflevector v1, v2 shufflemask => x={a,c,e,g}
5122 // Requires similar checks to that of isVTRNMask with
5123 // respect the how results are returned.
5124 static bool isVUZPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
5125 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
5129 unsigned NumElts = VT.getVectorNumElements();
5130 if (M.size() != NumElts && M.size() != NumElts*2)
5133 for (unsigned i = 0; i < M.size(); i += NumElts) {
5134 WhichResult = M[i] == 0 ? 0 : 1;
5135 for (unsigned j = 0; j < NumElts; ++j) {
5136 if (M[i+j] >= 0 && (unsigned) M[i+j] != 2 * j + WhichResult)
5141 if (M.size() == NumElts*2)
5144 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
5145 if (VT.is64BitVector() && EltSz == 32)
5151 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
5152 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
5153 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
5154 static bool isVUZP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
5155 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
5159 unsigned NumElts = VT.getVectorNumElements();
5160 if (M.size() != NumElts && M.size() != NumElts*2)
5163 unsigned Half = NumElts / 2;
5164 for (unsigned i = 0; i < M.size(); i += NumElts) {
5165 WhichResult = M[i] == 0 ? 0 : 1;
5166 for (unsigned j = 0; j < NumElts; j += Half) {
5167 unsigned Idx = WhichResult;
5168 for (unsigned k = 0; k < Half; ++k) {
5169 int MIdx = M[i + j + k];
5170 if (MIdx >= 0 && (unsigned) MIdx != Idx)
5177 if (M.size() == NumElts*2)
5180 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
5181 if (VT.is64BitVector() && EltSz == 32)
5187 // Checks whether the shuffle mask represents a vector zip (VZIP) by checking
5188 // that pairs of elements of the shufflemask represent the same index in each
5189 // vector incrementing sequentially through the vectors.
5190 // e.g. For v1,v2 of type v4i32 a valid shuffle mask is: [0, 4, 1, 5]
5191 // v1={a,b,c,d} => x=shufflevector v1, v2 shufflemask => x={a,e,b,f}
5193 // Requires similar checks to that of isVTRNMask with respect the how results
5195 static bool isVZIPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
5196 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
5200 unsigned NumElts = VT.getVectorNumElements();
5201 if (M.size() != NumElts && M.size() != NumElts*2)
5204 for (unsigned i = 0; i < M.size(); i += NumElts) {
5205 WhichResult = M[i] == 0 ? 0 : 1;
5206 unsigned Idx = WhichResult * NumElts / 2;
5207 for (unsigned j = 0; j < NumElts; j += 2) {
5208 if ((M[i+j] >= 0 && (unsigned) M[i+j] != Idx) ||
5209 (M[i+j+1] >= 0 && (unsigned) M[i+j+1] != Idx + NumElts))
5215 if (M.size() == NumElts*2)
5218 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
5219 if (VT.is64BitVector() && EltSz == 32)
5225 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
5226 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
5227 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
5228 static bool isVZIP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
5229 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
5233 unsigned NumElts = VT.getVectorNumElements();
5234 if (M.size() != NumElts && M.size() != NumElts*2)
5237 for (unsigned i = 0; i < M.size(); i += NumElts) {
5238 WhichResult = M[i] == 0 ? 0 : 1;
5239 unsigned Idx = WhichResult * NumElts / 2;
5240 for (unsigned j = 0; j < NumElts; j += 2) {
5241 if ((M[i+j] >= 0 && (unsigned) M[i+j] != Idx) ||
5242 (M[i+j+1] >= 0 && (unsigned) M[i+j+1] != Idx))
5248 if (M.size() == NumElts*2)
5251 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
5252 if (VT.is64BitVector() && EltSz == 32)
5258 /// Check if \p ShuffleMask is a NEON two-result shuffle (VZIP, VUZP, VTRN),
5259 /// and return the corresponding ARMISD opcode if it is, or 0 if it isn't.
5260 static unsigned isNEONTwoResultShuffleMask(ArrayRef<int> ShuffleMask, EVT VT,
5261 unsigned &WhichResult,
5264 if (isVTRNMask(ShuffleMask, VT, WhichResult))
5265 return ARMISD::VTRN;
5266 if (isVUZPMask(ShuffleMask, VT, WhichResult))
5267 return ARMISD::VUZP;
5268 if (isVZIPMask(ShuffleMask, VT, WhichResult))
5269 return ARMISD::VZIP;
5272 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
5273 return ARMISD::VTRN;
5274 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
5275 return ARMISD::VUZP;
5276 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
5277 return ARMISD::VZIP;
5282 /// \return true if this is a reverse operation on an vector.
5283 static bool isReverseMask(ArrayRef<int> M, EVT VT) {
5284 unsigned NumElts = VT.getVectorNumElements();
5285 // Make sure the mask has the right size.
5286 if (NumElts != M.size())
5289 // Look for <15, ..., 3, -1, 1, 0>.
5290 for (unsigned i = 0; i != NumElts; ++i)
5291 if (M[i] >= 0 && M[i] != (int) (NumElts - 1 - i))
5297 // If N is an integer constant that can be moved into a register in one
5298 // instruction, return an SDValue of such a constant (will become a MOV
5299 // instruction). Otherwise return null.
5300 static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG,
5301 const ARMSubtarget *ST, SDLoc dl) {
5303 if (!isa<ConstantSDNode>(N))
5305 Val = cast<ConstantSDNode>(N)->getZExtValue();
5307 if (ST->isThumb1Only()) {
5308 if (Val <= 255 || ~Val <= 255)
5309 return DAG.getConstant(Val, dl, MVT::i32);
5311 if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1)
5312 return DAG.getConstant(Val, dl, MVT::i32);
5317 // If this is a case we can't handle, return null and let the default
5318 // expansion code take care of it.
5319 SDValue ARMTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
5320 const ARMSubtarget *ST) const {
5321 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
5323 EVT VT = Op.getValueType();
5325 APInt SplatBits, SplatUndef;
5326 unsigned SplatBitSize;
5328 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
5329 if (SplatBitSize <= 64) {
5330 // Check if an immediate VMOV works.
5332 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
5333 SplatUndef.getZExtValue(), SplatBitSize,
5334 DAG, dl, VmovVT, VT.is128BitVector(),
5336 if (Val.getNode()) {
5337 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
5338 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
5341 // Try an immediate VMVN.
5342 uint64_t NegatedImm = (~SplatBits).getZExtValue();
5343 Val = isNEONModifiedImm(NegatedImm,
5344 SplatUndef.getZExtValue(), SplatBitSize,
5345 DAG, dl, VmovVT, VT.is128BitVector(),
5347 if (Val.getNode()) {
5348 SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
5349 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
5352 // Use vmov.f32 to materialize other v2f32 and v4f32 splats.
5353 if ((VT == MVT::v2f32 || VT == MVT::v4f32) && SplatBitSize == 32) {
5354 int ImmVal = ARM_AM::getFP32Imm(SplatBits);
5356 SDValue Val = DAG.getTargetConstant(ImmVal, dl, MVT::i32);
5357 return DAG.getNode(ARMISD::VMOVFPIMM, dl, VT, Val);
5363 // Scan through the operands to see if only one value is used.
5365 // As an optimisation, even if more than one value is used it may be more
5366 // profitable to splat with one value then change some lanes.
5368 // Heuristically we decide to do this if the vector has a "dominant" value,
5369 // defined as splatted to more than half of the lanes.
5370 unsigned NumElts = VT.getVectorNumElements();
5371 bool isOnlyLowElement = true;
5372 bool usesOnlyOneValue = true;
5373 bool hasDominantValue = false;
5374 bool isConstant = true;
5376 // Map of the number of times a particular SDValue appears in the
5378 DenseMap<SDValue, unsigned> ValueCounts;
5380 for (unsigned i = 0; i < NumElts; ++i) {
5381 SDValue V = Op.getOperand(i);
5382 if (V.getOpcode() == ISD::UNDEF)
5385 isOnlyLowElement = false;
5386 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
5389 ValueCounts.insert(std::make_pair(V, 0));
5390 unsigned &Count = ValueCounts[V];
5392 // Is this value dominant? (takes up more than half of the lanes)
5393 if (++Count > (NumElts / 2)) {
5394 hasDominantValue = true;
5398 if (ValueCounts.size() != 1)
5399 usesOnlyOneValue = false;
5400 if (!Value.getNode() && ValueCounts.size() > 0)
5401 Value = ValueCounts.begin()->first;
5403 if (ValueCounts.size() == 0)
5404 return DAG.getUNDEF(VT);
5406 // Loads are better lowered with insert_vector_elt/ARMISD::BUILD_VECTOR.
5407 // Keep going if we are hitting this case.
5408 if (isOnlyLowElement && !ISD::isNormalLoad(Value.getNode()))
5409 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
5411 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5413 // Use VDUP for non-constant splats. For f32 constant splats, reduce to
5414 // i32 and try again.
5415 if (hasDominantValue && EltSize <= 32) {
5419 // If we are VDUPing a value that comes directly from a vector, that will
5420 // cause an unnecessary move to and from a GPR, where instead we could
5421 // just use VDUPLANE. We can only do this if the lane being extracted
5422 // is at a constant index, as the VDUP from lane instructions only have
5423 // constant-index forms.
5424 if (Value->getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5425 isa<ConstantSDNode>(Value->getOperand(1))) {
5426 // We need to create a new undef vector to use for the VDUPLANE if the
5427 // size of the vector from which we get the value is different than the
5428 // size of the vector that we need to create. We will insert the element
5429 // such that the register coalescer will remove unnecessary copies.
5430 if (VT != Value->getOperand(0).getValueType()) {
5431 ConstantSDNode *constIndex;
5432 constIndex = dyn_cast<ConstantSDNode>(Value->getOperand(1));
5433 assert(constIndex && "The index is not a constant!");
5434 unsigned index = constIndex->getAPIntValue().getLimitedValue() %
5435 VT.getVectorNumElements();
5436 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
5437 DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DAG.getUNDEF(VT),
5438 Value, DAG.getConstant(index, dl, MVT::i32)),
5439 DAG.getConstant(index, dl, MVT::i32));
5441 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
5442 Value->getOperand(0), Value->getOperand(1));
5444 N = DAG.getNode(ARMISD::VDUP, dl, VT, Value);
5446 if (!usesOnlyOneValue) {
5447 // The dominant value was splatted as 'N', but we now have to insert
5448 // all differing elements.
5449 for (unsigned I = 0; I < NumElts; ++I) {
5450 if (Op.getOperand(I) == Value)
5452 SmallVector<SDValue, 3> Ops;
5454 Ops.push_back(Op.getOperand(I));
5455 Ops.push_back(DAG.getConstant(I, dl, MVT::i32));
5456 N = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Ops);
5461 if (VT.getVectorElementType().isFloatingPoint()) {
5462 SmallVector<SDValue, 8> Ops;
5463 for (unsigned i = 0; i < NumElts; ++i)
5464 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, MVT::i32,
5466 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
5467 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, Ops);
5468 Val = LowerBUILD_VECTOR(Val, DAG, ST);
5470 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
5472 if (usesOnlyOneValue) {
5473 SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
5474 if (isConstant && Val.getNode())
5475 return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
5479 // If all elements are constants and the case above didn't get hit, fall back
5480 // to the default expansion, which will generate a load from the constant
5485 // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
5487 SDValue shuffle = ReconstructShuffle(Op, DAG);
5488 if (shuffle != SDValue())
5492 // Vectors with 32- or 64-bit elements can be built by directly assigning
5493 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
5494 // will be legalized.
5495 if (EltSize >= 32) {
5496 // Do the expansion with floating-point types, since that is what the VFP
5497 // registers are defined to use, and since i64 is not legal.
5498 EVT EltVT = EVT::getFloatingPointVT(EltSize);
5499 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
5500 SmallVector<SDValue, 8> Ops;
5501 for (unsigned i = 0; i < NumElts; ++i)
5502 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, EltVT, Op.getOperand(i)));
5503 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, Ops);
5504 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
5507 // If all else fails, just use a sequence of INSERT_VECTOR_ELT when we
5508 // know the default expansion would otherwise fall back on something even
5509 // worse. For a vector with one or two non-undef values, that's
5510 // scalar_to_vector for the elements followed by a shuffle (provided the
5511 // shuffle is valid for the target) and materialization element by element
5512 // on the stack followed by a load for everything else.
5513 if (!isConstant && !usesOnlyOneValue) {
5514 SDValue Vec = DAG.getUNDEF(VT);
5515 for (unsigned i = 0 ; i < NumElts; ++i) {
5516 SDValue V = Op.getOperand(i);
5517 if (V.getOpcode() == ISD::UNDEF)
5519 SDValue LaneIdx = DAG.getConstant(i, dl, MVT::i32);
5520 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Vec, V, LaneIdx);
5528 // Gather data to see if the operation can be modelled as a
5529 // shuffle in combination with VEXTs.
5530 SDValue ARMTargetLowering::ReconstructShuffle(SDValue Op,
5531 SelectionDAG &DAG) const {
5532 assert(Op.getOpcode() == ISD::BUILD_VECTOR && "Unknown opcode!");
5534 EVT VT = Op.getValueType();
5535 unsigned NumElts = VT.getVectorNumElements();
5537 struct ShuffleSourceInfo {
5542 // We may insert some combination of BITCASTs and VEXT nodes to force Vec to
5543 // be compatible with the shuffle we intend to construct. As a result
5544 // ShuffleVec will be some sliding window into the original Vec.
5547 // Code should guarantee that element i in Vec starts at element "WindowBase
5548 // + i * WindowScale in ShuffleVec".
5552 bool operator ==(SDValue OtherVec) { return Vec == OtherVec; }
5553 ShuffleSourceInfo(SDValue Vec)
5554 : Vec(Vec), MinElt(UINT_MAX), MaxElt(0), ShuffleVec(Vec), WindowBase(0),
5558 // First gather all vectors used as an immediate source for this BUILD_VECTOR
5560 SmallVector<ShuffleSourceInfo, 2> Sources;
5561 for (unsigned i = 0; i < NumElts; ++i) {
5562 SDValue V = Op.getOperand(i);
5563 if (V.getOpcode() == ISD::UNDEF)
5565 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
5566 // A shuffle can only come from building a vector from various
5567 // elements of other vectors.
5569 } else if (!isa<ConstantSDNode>(V.getOperand(1))) {
5570 // Furthermore, shuffles require a constant mask, whereas extractelts
5571 // accept variable indices.
5575 // Add this element source to the list if it's not already there.
5576 SDValue SourceVec = V.getOperand(0);
5577 auto Source = std::find(Sources.begin(), Sources.end(), SourceVec);
5578 if (Source == Sources.end())
5579 Source = Sources.insert(Sources.end(), ShuffleSourceInfo(SourceVec));
5581 // Update the minimum and maximum lane number seen.
5582 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
5583 Source->MinElt = std::min(Source->MinElt, EltNo);
5584 Source->MaxElt = std::max(Source->MaxElt, EltNo);
5587 // Currently only do something sane when at most two source vectors
5589 if (Sources.size() > 2)
5592 // Find out the smallest element size among result and two sources, and use
5593 // it as element size to build the shuffle_vector.
5594 EVT SmallestEltTy = VT.getVectorElementType();
5595 for (auto &Source : Sources) {
5596 EVT SrcEltTy = Source.Vec.getValueType().getVectorElementType();
5597 if (SrcEltTy.bitsLT(SmallestEltTy))
5598 SmallestEltTy = SrcEltTy;
5600 unsigned ResMultiplier =
5601 VT.getVectorElementType().getSizeInBits() / SmallestEltTy.getSizeInBits();
5602 NumElts = VT.getSizeInBits() / SmallestEltTy.getSizeInBits();
5603 EVT ShuffleVT = EVT::getVectorVT(*DAG.getContext(), SmallestEltTy, NumElts);
5605 // If the source vector is too wide or too narrow, we may nevertheless be able
5606 // to construct a compatible shuffle either by concatenating it with UNDEF or
5607 // extracting a suitable range of elements.
5608 for (auto &Src : Sources) {
5609 EVT SrcVT = Src.ShuffleVec.getValueType();
5611 if (SrcVT.getSizeInBits() == VT.getSizeInBits())
5614 // This stage of the search produces a source with the same element type as
5615 // the original, but with a total width matching the BUILD_VECTOR output.
5616 EVT EltVT = SrcVT.getVectorElementType();
5617 unsigned NumSrcElts = VT.getSizeInBits() / EltVT.getSizeInBits();
5618 EVT DestVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumSrcElts);
5620 if (SrcVT.getSizeInBits() < VT.getSizeInBits()) {
5621 if (2 * SrcVT.getSizeInBits() != VT.getSizeInBits())
5623 // We can pad out the smaller vector for free, so if it's part of a
5626 DAG.getNode(ISD::CONCAT_VECTORS, dl, DestVT, Src.ShuffleVec,
5627 DAG.getUNDEF(Src.ShuffleVec.getValueType()));
5631 if (SrcVT.getSizeInBits() != 2 * VT.getSizeInBits())
5634 if (Src.MaxElt - Src.MinElt >= NumSrcElts) {
5635 // Span too large for a VEXT to cope
5639 if (Src.MinElt >= NumSrcElts) {
5640 // The extraction can just take the second half
5642 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec,
5643 DAG.getConstant(NumSrcElts, dl, MVT::i32));
5644 Src.WindowBase = -NumSrcElts;
5645 } else if (Src.MaxElt < NumSrcElts) {
5646 // The extraction can just take the first half
5648 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec,
5649 DAG.getConstant(0, dl, MVT::i32));
5651 // An actual VEXT is needed
5653 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec,
5654 DAG.getConstant(0, dl, MVT::i32));
5656 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec,
5657 DAG.getConstant(NumSrcElts, dl, MVT::i32));
5659 Src.ShuffleVec = DAG.getNode(ARMISD::VEXT, dl, DestVT, VEXTSrc1,
5661 DAG.getConstant(Src.MinElt, dl, MVT::i32));
5662 Src.WindowBase = -Src.MinElt;
5666 // Another possible incompatibility occurs from the vector element types. We
5667 // can fix this by bitcasting the source vectors to the same type we intend
5669 for (auto &Src : Sources) {
5670 EVT SrcEltTy = Src.ShuffleVec.getValueType().getVectorElementType();
5671 if (SrcEltTy == SmallestEltTy)
5673 assert(ShuffleVT.getVectorElementType() == SmallestEltTy);
5674 Src.ShuffleVec = DAG.getNode(ISD::BITCAST, dl, ShuffleVT, Src.ShuffleVec);
5675 Src.WindowScale = SrcEltTy.getSizeInBits() / SmallestEltTy.getSizeInBits();
5676 Src.WindowBase *= Src.WindowScale;
5679 // Final sanity check before we try to actually produce a shuffle.
5681 for (auto Src : Sources)
5682 assert(Src.ShuffleVec.getValueType() == ShuffleVT);
5685 // The stars all align, our next step is to produce the mask for the shuffle.
5686 SmallVector<int, 8> Mask(ShuffleVT.getVectorNumElements(), -1);
5687 int BitsPerShuffleLane = ShuffleVT.getVectorElementType().getSizeInBits();
5688 for (unsigned i = 0; i < VT.getVectorNumElements(); ++i) {
5689 SDValue Entry = Op.getOperand(i);
5690 if (Entry.getOpcode() == ISD::UNDEF)
5693 auto Src = std::find(Sources.begin(), Sources.end(), Entry.getOperand(0));
5694 int EltNo = cast<ConstantSDNode>(Entry.getOperand(1))->getSExtValue();
5696 // EXTRACT_VECTOR_ELT performs an implicit any_ext; BUILD_VECTOR an implicit
5697 // trunc. So only std::min(SrcBits, DestBits) actually get defined in this
5699 EVT OrigEltTy = Entry.getOperand(0).getValueType().getVectorElementType();
5700 int BitsDefined = std::min(OrigEltTy.getSizeInBits(),
5701 VT.getVectorElementType().getSizeInBits());
5702 int LanesDefined = BitsDefined / BitsPerShuffleLane;
5704 // This source is expected to fill ResMultiplier lanes of the final shuffle,
5705 // starting at the appropriate offset.
5706 int *LaneMask = &Mask[i * ResMultiplier];
5708 int ExtractBase = EltNo * Src->WindowScale + Src->WindowBase;
5709 ExtractBase += NumElts * (Src - Sources.begin());
5710 for (int j = 0; j < LanesDefined; ++j)
5711 LaneMask[j] = ExtractBase + j;
5714 // Final check before we try to produce nonsense...
5715 if (!isShuffleMaskLegal(Mask, ShuffleVT))
5718 // We can't handle more than two sources. This should have already
5719 // been checked before this point.
5720 assert(Sources.size() <= 2 && "Too many sources!");
5722 SDValue ShuffleOps[] = { DAG.getUNDEF(ShuffleVT), DAG.getUNDEF(ShuffleVT) };
5723 for (unsigned i = 0; i < Sources.size(); ++i)
5724 ShuffleOps[i] = Sources[i].ShuffleVec;
5726 SDValue Shuffle = DAG.getVectorShuffle(ShuffleVT, dl, ShuffleOps[0],
5727 ShuffleOps[1], &Mask[0]);
5728 return DAG.getNode(ISD::BITCAST, dl, VT, Shuffle);
5731 /// isShuffleMaskLegal - Targets can use this to indicate that they only
5732 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
5733 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
5734 /// are assumed to be legal.
5736 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
5738 if (VT.getVectorNumElements() == 4 &&
5739 (VT.is128BitVector() || VT.is64BitVector())) {
5740 unsigned PFIndexes[4];
5741 for (unsigned i = 0; i != 4; ++i) {
5745 PFIndexes[i] = M[i];
5748 // Compute the index in the perfect shuffle table.
5749 unsigned PFTableIndex =
5750 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
5751 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
5752 unsigned Cost = (PFEntry >> 30);
5758 bool ReverseVEXT, isV_UNDEF;
5759 unsigned Imm, WhichResult;
5761 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5762 return (EltSize >= 32 ||
5763 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
5764 isVREVMask(M, VT, 64) ||
5765 isVREVMask(M, VT, 32) ||
5766 isVREVMask(M, VT, 16) ||
5767 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
5768 isVTBLMask(M, VT) ||
5769 isNEONTwoResultShuffleMask(M, VT, WhichResult, isV_UNDEF) ||
5770 ((VT == MVT::v8i16 || VT == MVT::v16i8) && isReverseMask(M, VT)));
5773 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
5774 /// the specified operations to build the shuffle.
5775 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
5776 SDValue RHS, SelectionDAG &DAG,
5778 unsigned OpNum = (PFEntry >> 26) & 0x0F;
5779 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
5780 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
5783 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
5792 OP_VUZPL, // VUZP, left result
5793 OP_VUZPR, // VUZP, right result
5794 OP_VZIPL, // VZIP, left result
5795 OP_VZIPR, // VZIP, right result
5796 OP_VTRNL, // VTRN, left result
5797 OP_VTRNR // VTRN, right result
5800 if (OpNum == OP_COPY) {
5801 if (LHSID == (1*9+2)*9+3) return LHS;
5802 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
5806 SDValue OpLHS, OpRHS;
5807 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
5808 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
5809 EVT VT = OpLHS.getValueType();
5812 default: llvm_unreachable("Unknown shuffle opcode!");
5814 // VREV divides the vector in half and swaps within the half.
5815 if (VT.getVectorElementType() == MVT::i32 ||
5816 VT.getVectorElementType() == MVT::f32)
5817 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
5818 // vrev <4 x i16> -> VREV32
5819 if (VT.getVectorElementType() == MVT::i16)
5820 return DAG.getNode(ARMISD::VREV32, dl, VT, OpLHS);
5821 // vrev <4 x i8> -> VREV16
5822 assert(VT.getVectorElementType() == MVT::i8);
5823 return DAG.getNode(ARMISD::VREV16, dl, VT, OpLHS);
5828 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
5829 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, dl, MVT::i32));
5833 return DAG.getNode(ARMISD::VEXT, dl, VT,
5835 DAG.getConstant(OpNum - OP_VEXT1 + 1, dl, MVT::i32));
5838 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
5839 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
5842 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
5843 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
5846 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
5847 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
5851 static SDValue LowerVECTOR_SHUFFLEv8i8(SDValue Op,
5852 ArrayRef<int> ShuffleMask,
5853 SelectionDAG &DAG) {
5854 // Check to see if we can use the VTBL instruction.
5855 SDValue V1 = Op.getOperand(0);
5856 SDValue V2 = Op.getOperand(1);
5859 SmallVector<SDValue, 8> VTBLMask;
5860 for (ArrayRef<int>::iterator
5861 I = ShuffleMask.begin(), E = ShuffleMask.end(); I != E; ++I)
5862 VTBLMask.push_back(DAG.getConstant(*I, DL, MVT::i32));
5864 if (V2.getNode()->getOpcode() == ISD::UNDEF)
5865 return DAG.getNode(ARMISD::VTBL1, DL, MVT::v8i8, V1,
5866 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8, VTBLMask));
5868 return DAG.getNode(ARMISD::VTBL2, DL, MVT::v8i8, V1, V2,
5869 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8, VTBLMask));
5872 static SDValue LowerReverse_VECTOR_SHUFFLEv16i8_v8i16(SDValue Op,
5873 SelectionDAG &DAG) {
5875 SDValue OpLHS = Op.getOperand(0);
5876 EVT VT = OpLHS.getValueType();
5878 assert((VT == MVT::v8i16 || VT == MVT::v16i8) &&
5879 "Expect an v8i16/v16i8 type");
5880 OpLHS = DAG.getNode(ARMISD::VREV64, DL, VT, OpLHS);
5881 // For a v16i8 type: After the VREV, we have got <8, ...15, 8, ..., 0>. Now,
5882 // extract the first 8 bytes into the top double word and the last 8 bytes
5883 // into the bottom double word. The v8i16 case is similar.
5884 unsigned ExtractNum = (VT == MVT::v16i8) ? 8 : 4;
5885 return DAG.getNode(ARMISD::VEXT, DL, VT, OpLHS, OpLHS,
5886 DAG.getConstant(ExtractNum, DL, MVT::i32));
5889 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
5890 SDValue V1 = Op.getOperand(0);
5891 SDValue V2 = Op.getOperand(1);
5893 EVT VT = Op.getValueType();
5894 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
5896 // Convert shuffles that are directly supported on NEON to target-specific
5897 // DAG nodes, instead of keeping them as shuffles and matching them again
5898 // during code selection. This is more efficient and avoids the possibility
5899 // of inconsistencies between legalization and selection.
5900 // FIXME: floating-point vectors should be canonicalized to integer vectors
5901 // of the same time so that they get CSEd properly.
5902 ArrayRef<int> ShuffleMask = SVN->getMask();
5904 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5905 if (EltSize <= 32) {
5906 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
5907 int Lane = SVN->getSplatIndex();
5908 // If this is undef splat, generate it via "just" vdup, if possible.
5909 if (Lane == -1) Lane = 0;
5911 // Test if V1 is a SCALAR_TO_VECTOR.
5912 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
5913 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
5915 // Test if V1 is a BUILD_VECTOR which is equivalent to a SCALAR_TO_VECTOR
5916 // (and probably will turn into a SCALAR_TO_VECTOR once legalization
5918 if (Lane == 0 && V1.getOpcode() == ISD::BUILD_VECTOR &&
5919 !isa<ConstantSDNode>(V1.getOperand(0))) {
5920 bool IsScalarToVector = true;
5921 for (unsigned i = 1, e = V1.getNumOperands(); i != e; ++i)
5922 if (V1.getOperand(i).getOpcode() != ISD::UNDEF) {
5923 IsScalarToVector = false;
5926 if (IsScalarToVector)
5927 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
5929 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
5930 DAG.getConstant(Lane, dl, MVT::i32));
5935 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
5938 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
5939 DAG.getConstant(Imm, dl, MVT::i32));
5942 if (isVREVMask(ShuffleMask, VT, 64))
5943 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
5944 if (isVREVMask(ShuffleMask, VT, 32))
5945 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
5946 if (isVREVMask(ShuffleMask, VT, 16))
5947 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
5949 if (V2->getOpcode() == ISD::UNDEF &&
5950 isSingletonVEXTMask(ShuffleMask, VT, Imm)) {
5951 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V1,
5952 DAG.getConstant(Imm, dl, MVT::i32));
5955 // Check for Neon shuffles that modify both input vectors in place.
5956 // If both results are used, i.e., if there are two shuffles with the same
5957 // source operands and with masks corresponding to both results of one of
5958 // these operations, DAG memoization will ensure that a single node is
5959 // used for both shuffles.
5960 unsigned WhichResult;
5962 if (unsigned ShuffleOpc = isNEONTwoResultShuffleMask(
5963 ShuffleMask, VT, WhichResult, isV_UNDEF)) {
5966 return DAG.getNode(ShuffleOpc, dl, DAG.getVTList(VT, VT), V1, V2)
5967 .getValue(WhichResult);
5970 // Also check for these shuffles through CONCAT_VECTORS: we canonicalize
5971 // shuffles that produce a result larger than their operands with:
5972 // shuffle(concat(v1, undef), concat(v2, undef))
5974 // shuffle(concat(v1, v2), undef)
5975 // because we can access quad vectors (see PerformVECTOR_SHUFFLECombine).
5977 // This is useful in the general case, but there are special cases where
5978 // native shuffles produce larger results: the two-result ops.
5980 // Look through the concat when lowering them:
5981 // shuffle(concat(v1, v2), undef)
5983 // concat(VZIP(v1, v2):0, :1)
5985 if (V1->getOpcode() == ISD::CONCAT_VECTORS &&
5986 V2->getOpcode() == ISD::UNDEF) {
5987 SDValue SubV1 = V1->getOperand(0);
5988 SDValue SubV2 = V1->getOperand(1);
5989 EVT SubVT = SubV1.getValueType();
5991 // We expect these to have been canonicalized to -1.
5992 assert(std::all_of(ShuffleMask.begin(), ShuffleMask.end(), [&](int i) {
5993 return i < (int)VT.getVectorNumElements();
5994 }) && "Unexpected shuffle index into UNDEF operand!");
5996 if (unsigned ShuffleOpc = isNEONTwoResultShuffleMask(
5997 ShuffleMask, SubVT, WhichResult, isV_UNDEF)) {
6000 assert((WhichResult == 0) &&
6001 "In-place shuffle of concat can only have one result!");
6002 SDValue Res = DAG.getNode(ShuffleOpc, dl, DAG.getVTList(SubVT, SubVT),
6004 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, Res.getValue(0),
6010 // If the shuffle is not directly supported and it has 4 elements, use
6011 // the PerfectShuffle-generated table to synthesize it from other shuffles.
6012 unsigned NumElts = VT.getVectorNumElements();
6014 unsigned PFIndexes[4];
6015 for (unsigned i = 0; i != 4; ++i) {
6016 if (ShuffleMask[i] < 0)
6019 PFIndexes[i] = ShuffleMask[i];
6022 // Compute the index in the perfect shuffle table.
6023 unsigned PFTableIndex =
6024 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
6025 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
6026 unsigned Cost = (PFEntry >> 30);
6029 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
6032 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
6033 if (EltSize >= 32) {
6034 // Do the expansion with floating-point types, since that is what the VFP
6035 // registers are defined to use, and since i64 is not legal.
6036 EVT EltVT = EVT::getFloatingPointVT(EltSize);
6037 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
6038 V1 = DAG.getNode(ISD::BITCAST, dl, VecVT, V1);
6039 V2 = DAG.getNode(ISD::BITCAST, dl, VecVT, V2);
6040 SmallVector<SDValue, 8> Ops;
6041 for (unsigned i = 0; i < NumElts; ++i) {
6042 if (ShuffleMask[i] < 0)
6043 Ops.push_back(DAG.getUNDEF(EltVT));
6045 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
6046 ShuffleMask[i] < (int)NumElts ? V1 : V2,
6047 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
6050 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, Ops);
6051 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
6054 if ((VT == MVT::v8i16 || VT == MVT::v16i8) && isReverseMask(ShuffleMask, VT))
6055 return LowerReverse_VECTOR_SHUFFLEv16i8_v8i16(Op, DAG);
6057 if (VT == MVT::v8i8) {
6058 SDValue NewOp = LowerVECTOR_SHUFFLEv8i8(Op, ShuffleMask, DAG);
6059 if (NewOp.getNode())
6066 static SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
6067 // INSERT_VECTOR_ELT is legal only for immediate indexes.
6068 SDValue Lane = Op.getOperand(2);
6069 if (!isa<ConstantSDNode>(Lane))
6075 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
6076 // EXTRACT_VECTOR_ELT is legal only for immediate indexes.
6077 SDValue Lane = Op.getOperand(1);
6078 if (!isa<ConstantSDNode>(Lane))
6081 SDValue Vec = Op.getOperand(0);
6082 if (Op.getValueType() == MVT::i32 &&
6083 Vec.getValueType().getVectorElementType().getSizeInBits() < 32) {
6085 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
6091 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
6092 // The only time a CONCAT_VECTORS operation can have legal types is when
6093 // two 64-bit vectors are concatenated to a 128-bit vector.
6094 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
6095 "unexpected CONCAT_VECTORS");
6097 SDValue Val = DAG.getUNDEF(MVT::v2f64);
6098 SDValue Op0 = Op.getOperand(0);
6099 SDValue Op1 = Op.getOperand(1);
6100 if (Op0.getOpcode() != ISD::UNDEF)
6101 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
6102 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op0),
6103 DAG.getIntPtrConstant(0, dl));
6104 if (Op1.getOpcode() != ISD::UNDEF)
6105 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
6106 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op1),
6107 DAG.getIntPtrConstant(1, dl));
6108 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Val);
6111 /// isExtendedBUILD_VECTOR - Check if N is a constant BUILD_VECTOR where each
6112 /// element has been zero/sign-extended, depending on the isSigned parameter,
6113 /// from an integer type half its size.
6114 static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG,
6116 // A v2i64 BUILD_VECTOR will have been legalized to a BITCAST from v4i32.
6117 EVT VT = N->getValueType(0);
6118 if (VT == MVT::v2i64 && N->getOpcode() == ISD::BITCAST) {
6119 SDNode *BVN = N->getOperand(0).getNode();
6120 if (BVN->getValueType(0) != MVT::v4i32 ||
6121 BVN->getOpcode() != ISD::BUILD_VECTOR)
6123 unsigned LoElt = DAG.getDataLayout().isBigEndian() ? 1 : 0;
6124 unsigned HiElt = 1 - LoElt;
6125 ConstantSDNode *Lo0 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt));
6126 ConstantSDNode *Hi0 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt));
6127 ConstantSDNode *Lo1 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt+2));
6128 ConstantSDNode *Hi1 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt+2));
6129 if (!Lo0 || !Hi0 || !Lo1 || !Hi1)
6132 if (Hi0->getSExtValue() == Lo0->getSExtValue() >> 32 &&
6133 Hi1->getSExtValue() == Lo1->getSExtValue() >> 32)
6136 if (Hi0->isNullValue() && Hi1->isNullValue())
6142 if (N->getOpcode() != ISD::BUILD_VECTOR)
6145 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
6146 SDNode *Elt = N->getOperand(i).getNode();
6147 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
6148 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
6149 unsigned HalfSize = EltSize / 2;
6151 if (!isIntN(HalfSize, C->getSExtValue()))
6154 if (!isUIntN(HalfSize, C->getZExtValue()))
6165 /// isSignExtended - Check if a node is a vector value that is sign-extended
6166 /// or a constant BUILD_VECTOR with sign-extended elements.
6167 static bool isSignExtended(SDNode *N, SelectionDAG &DAG) {
6168 if (N->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N))
6170 if (isExtendedBUILD_VECTOR(N, DAG, true))
6175 /// isZeroExtended - Check if a node is a vector value that is zero-extended
6176 /// or a constant BUILD_VECTOR with zero-extended elements.
6177 static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) {
6178 if (N->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N))
6180 if (isExtendedBUILD_VECTOR(N, DAG, false))
6185 static EVT getExtensionTo64Bits(const EVT &OrigVT) {
6186 if (OrigVT.getSizeInBits() >= 64)
6189 assert(OrigVT.isSimple() && "Expecting a simple value type");
6191 MVT::SimpleValueType OrigSimpleTy = OrigVT.getSimpleVT().SimpleTy;
6192 switch (OrigSimpleTy) {
6193 default: llvm_unreachable("Unexpected Vector Type");
6202 /// AddRequiredExtensionForVMULL - Add a sign/zero extension to extend the total
6203 /// value size to 64 bits. We need a 64-bit D register as an operand to VMULL.
6204 /// We insert the required extension here to get the vector to fill a D register.
6205 static SDValue AddRequiredExtensionForVMULL(SDValue N, SelectionDAG &DAG,
6208 unsigned ExtOpcode) {
6209 // The vector originally had a size of OrigTy. It was then extended to ExtTy.
6210 // We expect the ExtTy to be 128-bits total. If the OrigTy is less than
6211 // 64-bits we need to insert a new extension so that it will be 64-bits.
6212 assert(ExtTy.is128BitVector() && "Unexpected extension size");
6213 if (OrigTy.getSizeInBits() >= 64)
6216 // Must extend size to at least 64 bits to be used as an operand for VMULL.
6217 EVT NewVT = getExtensionTo64Bits(OrigTy);
6219 return DAG.getNode(ExtOpcode, SDLoc(N), NewVT, N);
6222 /// SkipLoadExtensionForVMULL - return a load of the original vector size that
6223 /// does not do any sign/zero extension. If the original vector is less
6224 /// than 64 bits, an appropriate extension will be added after the load to
6225 /// reach a total size of 64 bits. We have to add the extension separately
6226 /// because ARM does not have a sign/zero extending load for vectors.
6227 static SDValue SkipLoadExtensionForVMULL(LoadSDNode *LD, SelectionDAG& DAG) {
6228 EVT ExtendedTy = getExtensionTo64Bits(LD->getMemoryVT());
6230 // The load already has the right type.
6231 if (ExtendedTy == LD->getMemoryVT())
6232 return DAG.getLoad(LD->getMemoryVT(), SDLoc(LD), LD->getChain(),
6233 LD->getBasePtr(), LD->getPointerInfo(), LD->isVolatile(),
6234 LD->isNonTemporal(), LD->isInvariant(),
6235 LD->getAlignment());
6237 // We need to create a zextload/sextload. We cannot just create a load
6238 // followed by a zext/zext node because LowerMUL is also run during normal
6239 // operation legalization where we can't create illegal types.
6240 return DAG.getExtLoad(LD->getExtensionType(), SDLoc(LD), ExtendedTy,
6241 LD->getChain(), LD->getBasePtr(), LD->getPointerInfo(),
6242 LD->getMemoryVT(), LD->isVolatile(), LD->isInvariant(),
6243 LD->isNonTemporal(), LD->getAlignment());
6246 /// SkipExtensionForVMULL - For a node that is a SIGN_EXTEND, ZERO_EXTEND,
6247 /// extending load, or BUILD_VECTOR with extended elements, return the
6248 /// unextended value. The unextended vector should be 64 bits so that it can
6249 /// be used as an operand to a VMULL instruction. If the original vector size
6250 /// before extension is less than 64 bits we add a an extension to resize
6251 /// the vector to 64 bits.
6252 static SDValue SkipExtensionForVMULL(SDNode *N, SelectionDAG &DAG) {
6253 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
6254 return AddRequiredExtensionForVMULL(N->getOperand(0), DAG,
6255 N->getOperand(0)->getValueType(0),
6259 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
6260 return SkipLoadExtensionForVMULL(LD, DAG);
6262 // Otherwise, the value must be a BUILD_VECTOR. For v2i64, it will
6263 // have been legalized as a BITCAST from v4i32.
6264 if (N->getOpcode() == ISD::BITCAST) {
6265 SDNode *BVN = N->getOperand(0).getNode();
6266 assert(BVN->getOpcode() == ISD::BUILD_VECTOR &&
6267 BVN->getValueType(0) == MVT::v4i32 && "expected v4i32 BUILD_VECTOR");
6268 unsigned LowElt = DAG.getDataLayout().isBigEndian() ? 1 : 0;
6269 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), MVT::v2i32,
6270 BVN->getOperand(LowElt), BVN->getOperand(LowElt+2));
6272 // Construct a new BUILD_VECTOR with elements truncated to half the size.
6273 assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR");
6274 EVT VT = N->getValueType(0);
6275 unsigned EltSize = VT.getVectorElementType().getSizeInBits() / 2;
6276 unsigned NumElts = VT.getVectorNumElements();
6277 MVT TruncVT = MVT::getIntegerVT(EltSize);
6278 SmallVector<SDValue, 8> Ops;
6280 for (unsigned i = 0; i != NumElts; ++i) {
6281 ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i));
6282 const APInt &CInt = C->getAPIntValue();
6283 // Element types smaller than 32 bits are not legal, so use i32 elements.
6284 // The values are implicitly truncated so sext vs. zext doesn't matter.
6285 Ops.push_back(DAG.getConstant(CInt.zextOrTrunc(32), dl, MVT::i32));
6287 return DAG.getNode(ISD::BUILD_VECTOR, dl,
6288 MVT::getVectorVT(TruncVT, NumElts), Ops);
6291 static bool isAddSubSExt(SDNode *N, SelectionDAG &DAG) {
6292 unsigned Opcode = N->getOpcode();
6293 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
6294 SDNode *N0 = N->getOperand(0).getNode();
6295 SDNode *N1 = N->getOperand(1).getNode();
6296 return N0->hasOneUse() && N1->hasOneUse() &&
6297 isSignExtended(N0, DAG) && isSignExtended(N1, DAG);
6302 static bool isAddSubZExt(SDNode *N, SelectionDAG &DAG) {
6303 unsigned Opcode = N->getOpcode();
6304 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
6305 SDNode *N0 = N->getOperand(0).getNode();
6306 SDNode *N1 = N->getOperand(1).getNode();
6307 return N0->hasOneUse() && N1->hasOneUse() &&
6308 isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG);
6313 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
6314 // Multiplications are only custom-lowered for 128-bit vectors so that
6315 // VMULL can be detected. Otherwise v2i64 multiplications are not legal.
6316 EVT VT = Op.getValueType();
6317 assert(VT.is128BitVector() && VT.isInteger() &&
6318 "unexpected type for custom-lowering ISD::MUL");
6319 SDNode *N0 = Op.getOperand(0).getNode();
6320 SDNode *N1 = Op.getOperand(1).getNode();
6321 unsigned NewOpc = 0;
6323 bool isN0SExt = isSignExtended(N0, DAG);
6324 bool isN1SExt = isSignExtended(N1, DAG);
6325 if (isN0SExt && isN1SExt)
6326 NewOpc = ARMISD::VMULLs;
6328 bool isN0ZExt = isZeroExtended(N0, DAG);
6329 bool isN1ZExt = isZeroExtended(N1, DAG);
6330 if (isN0ZExt && isN1ZExt)
6331 NewOpc = ARMISD::VMULLu;
6332 else if (isN1SExt || isN1ZExt) {
6333 // Look for (s/zext A + s/zext B) * (s/zext C). We want to turn these
6334 // into (s/zext A * s/zext C) + (s/zext B * s/zext C)
6335 if (isN1SExt && isAddSubSExt(N0, DAG)) {
6336 NewOpc = ARMISD::VMULLs;
6338 } else if (isN1ZExt && isAddSubZExt(N0, DAG)) {
6339 NewOpc = ARMISD::VMULLu;
6341 } else if (isN0ZExt && isAddSubZExt(N1, DAG)) {
6343 NewOpc = ARMISD::VMULLu;
6349 if (VT == MVT::v2i64)
6350 // Fall through to expand this. It is not legal.
6353 // Other vector multiplications are legal.
6358 // Legalize to a VMULL instruction.
6361 SDValue Op1 = SkipExtensionForVMULL(N1, DAG);
6363 Op0 = SkipExtensionForVMULL(N0, DAG);
6364 assert(Op0.getValueType().is64BitVector() &&
6365 Op1.getValueType().is64BitVector() &&
6366 "unexpected types for extended operands to VMULL");
6367 return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
6370 // Optimizing (zext A + zext B) * C, to (VMULL A, C) + (VMULL B, C) during
6371 // isel lowering to take advantage of no-stall back to back vmul + vmla.
6378 SDValue N00 = SkipExtensionForVMULL(N0->getOperand(0).getNode(), DAG);
6379 SDValue N01 = SkipExtensionForVMULL(N0->getOperand(1).getNode(), DAG);
6380 EVT Op1VT = Op1.getValueType();
6381 return DAG.getNode(N0->getOpcode(), DL, VT,
6382 DAG.getNode(NewOpc, DL, VT,
6383 DAG.getNode(ISD::BITCAST, DL, Op1VT, N00), Op1),
6384 DAG.getNode(NewOpc, DL, VT,
6385 DAG.getNode(ISD::BITCAST, DL, Op1VT, N01), Op1));
6389 LowerSDIV_v4i8(SDValue X, SDValue Y, SDLoc dl, SelectionDAG &DAG) {
6390 // TODO: Should this propagate fast-math-flags?
6393 // float4 xf = vcvt_f32_s32(vmovl_s16(a.lo));
6394 // float4 yf = vcvt_f32_s32(vmovl_s16(b.lo));
6395 X = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, X);
6396 Y = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, Y);
6397 X = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, X);
6398 Y = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, Y);
6399 // Get reciprocal estimate.
6400 // float4 recip = vrecpeq_f32(yf);
6401 Y = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6402 DAG.getConstant(Intrinsic::arm_neon_vrecpe, dl, MVT::i32),
6404 // Because char has a smaller range than uchar, we can actually get away
6405 // without any newton steps. This requires that we use a weird bias
6406 // of 0xb000, however (again, this has been exhaustively tested).
6407 // float4 result = as_float4(as_int4(xf*recip) + 0xb000);
6408 X = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, X, Y);
6409 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, X);
6410 Y = DAG.getConstant(0xb000, dl, MVT::i32);
6411 Y = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Y, Y, Y, Y);
6412 X = DAG.getNode(ISD::ADD, dl, MVT::v4i32, X, Y);
6413 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, X);
6414 // Convert back to short.
6415 X = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, X);
6416 X = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, X);
6421 LowerSDIV_v4i16(SDValue N0, SDValue N1, SDLoc dl, SelectionDAG &DAG) {
6422 // TODO: Should this propagate fast-math-flags?
6425 // Convert to float.
6426 // float4 yf = vcvt_f32_s32(vmovl_s16(y));
6427 // float4 xf = vcvt_f32_s32(vmovl_s16(x));
6428 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N0);
6429 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N1);
6430 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
6431 N1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
6433 // Use reciprocal estimate and one refinement step.
6434 // float4 recip = vrecpeq_f32(yf);
6435 // recip *= vrecpsq_f32(yf, recip);
6436 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6437 DAG.getConstant(Intrinsic::arm_neon_vrecpe, dl, MVT::i32),
6439 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6440 DAG.getConstant(Intrinsic::arm_neon_vrecps, dl, MVT::i32),
6442 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
6443 // Because short has a smaller range than ushort, we can actually get away
6444 // with only a single newton step. This requires that we use a weird bias
6445 // of 89, however (again, this has been exhaustively tested).
6446 // float4 result = as_float4(as_int4(xf*recip) + 0x89);
6447 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
6448 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
6449 N1 = DAG.getConstant(0x89, dl, MVT::i32);
6450 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
6451 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
6452 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
6453 // Convert back to integer and return.
6454 // return vmovn_s32(vcvt_s32_f32(result));
6455 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
6456 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
6460 static SDValue LowerSDIV(SDValue Op, SelectionDAG &DAG) {
6461 EVT VT = Op.getValueType();
6462 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
6463 "unexpected type for custom-lowering ISD::SDIV");
6466 SDValue N0 = Op.getOperand(0);
6467 SDValue N1 = Op.getOperand(1);
6470 if (VT == MVT::v8i8) {
6471 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N0);
6472 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N1);
6474 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
6475 DAG.getIntPtrConstant(4, dl));
6476 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
6477 DAG.getIntPtrConstant(4, dl));
6478 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
6479 DAG.getIntPtrConstant(0, dl));
6480 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
6481 DAG.getIntPtrConstant(0, dl));
6483 N0 = LowerSDIV_v4i8(N0, N1, dl, DAG); // v4i16
6484 N2 = LowerSDIV_v4i8(N2, N3, dl, DAG); // v4i16
6486 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
6487 N0 = LowerCONCAT_VECTORS(N0, DAG);
6489 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v8i8, N0);
6492 return LowerSDIV_v4i16(N0, N1, dl, DAG);
6495 static SDValue LowerUDIV(SDValue Op, SelectionDAG &DAG) {
6496 // TODO: Should this propagate fast-math-flags?
6497 EVT VT = Op.getValueType();
6498 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
6499 "unexpected type for custom-lowering ISD::UDIV");
6502 SDValue N0 = Op.getOperand(0);
6503 SDValue N1 = Op.getOperand(1);
6506 if (VT == MVT::v8i8) {
6507 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N0);
6508 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N1);
6510 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
6511 DAG.getIntPtrConstant(4, dl));
6512 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
6513 DAG.getIntPtrConstant(4, dl));
6514 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
6515 DAG.getIntPtrConstant(0, dl));
6516 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
6517 DAG.getIntPtrConstant(0, dl));
6519 N0 = LowerSDIV_v4i16(N0, N1, dl, DAG); // v4i16
6520 N2 = LowerSDIV_v4i16(N2, N3, dl, DAG); // v4i16
6522 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
6523 N0 = LowerCONCAT_VECTORS(N0, DAG);
6525 N0 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v8i8,
6526 DAG.getConstant(Intrinsic::arm_neon_vqmovnsu, dl,
6532 // v4i16 sdiv ... Convert to float.
6533 // float4 yf = vcvt_f32_s32(vmovl_u16(y));
6534 // float4 xf = vcvt_f32_s32(vmovl_u16(x));
6535 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N0);
6536 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N1);
6537 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
6538 SDValue BN1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
6540 // Use reciprocal estimate and two refinement steps.
6541 // float4 recip = vrecpeq_f32(yf);
6542 // recip *= vrecpsq_f32(yf, recip);
6543 // recip *= vrecpsq_f32(yf, recip);
6544 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6545 DAG.getConstant(Intrinsic::arm_neon_vrecpe, dl, MVT::i32),
6547 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6548 DAG.getConstant(Intrinsic::arm_neon_vrecps, dl, MVT::i32),
6550 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
6551 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6552 DAG.getConstant(Intrinsic::arm_neon_vrecps, dl, MVT::i32),
6554 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
6555 // Simply multiplying by the reciprocal estimate can leave us a few ulps
6556 // too low, so we add 2 ulps (exhaustive testing shows that this is enough,
6557 // and that it will never cause us to return an answer too large).
6558 // float4 result = as_float4(as_int4(xf*recip) + 2);
6559 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
6560 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
6561 N1 = DAG.getConstant(2, dl, MVT::i32);
6562 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
6563 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
6564 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
6565 // Convert back to integer and return.
6566 // return vmovn_u32(vcvt_s32_f32(result));
6567 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
6568 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
6572 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
6573 EVT VT = Op.getNode()->getValueType(0);
6574 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
6577 bool ExtraOp = false;
6578 switch (Op.getOpcode()) {
6579 default: llvm_unreachable("Invalid code");
6580 case ISD::ADDC: Opc = ARMISD::ADDC; break;
6581 case ISD::ADDE: Opc = ARMISD::ADDE; ExtraOp = true; break;
6582 case ISD::SUBC: Opc = ARMISD::SUBC; break;
6583 case ISD::SUBE: Opc = ARMISD::SUBE; ExtraOp = true; break;
6587 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
6589 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
6590 Op.getOperand(1), Op.getOperand(2));
6593 SDValue ARMTargetLowering::LowerFSINCOS(SDValue Op, SelectionDAG &DAG) const {
6594 assert(Subtarget->isTargetDarwin());
6596 // For iOS, we want to call an alternative entry point: __sincos_stret,
6597 // return values are passed via sret.
6599 SDValue Arg = Op.getOperand(0);
6600 EVT ArgVT = Arg.getValueType();
6601 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
6602 auto PtrVT = getPointerTy(DAG.getDataLayout());
6604 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
6605 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6607 // Pair of floats / doubles used to pass the result.
6608 Type *RetTy = StructType::get(ArgTy, ArgTy, nullptr);
6609 auto &DL = DAG.getDataLayout();
6612 bool ShouldUseSRet = Subtarget->isAPCS_ABI();
6614 if (ShouldUseSRet) {
6615 // Create stack object for sret.
6616 const uint64_t ByteSize = DL.getTypeAllocSize(RetTy);
6617 const unsigned StackAlign = DL.getPrefTypeAlignment(RetTy);
6618 int FrameIdx = FrameInfo->CreateStackObject(ByteSize, StackAlign, false);
6619 SRet = DAG.getFrameIndex(FrameIdx, TLI.getPointerTy(DL));
6623 Entry.Ty = RetTy->getPointerTo();
6624 Entry.isSExt = false;
6625 Entry.isZExt = false;
6626 Entry.isSRet = true;
6627 Args.push_back(Entry);
6628 RetTy = Type::getVoidTy(*DAG.getContext());
6634 Entry.isSExt = false;
6635 Entry.isZExt = false;
6636 Args.push_back(Entry);
6638 const char *LibcallName =
6639 (ArgVT == MVT::f64) ? "__sincos_stret" : "__sincosf_stret";
6641 (ArgVT == MVT::f64) ? RTLIB::SINCOS_F64 : RTLIB::SINCOS_F32;
6642 CallingConv::ID CC = getLibcallCallingConv(LC);
6643 SDValue Callee = DAG.getExternalSymbol(LibcallName, getPointerTy(DL));
6645 TargetLowering::CallLoweringInfo CLI(DAG);
6647 .setChain(DAG.getEntryNode())
6648 .setCallee(CC, RetTy, Callee, std::move(Args), 0)
6649 .setDiscardResult(ShouldUseSRet);
6650 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
6653 return CallResult.first;
6655 SDValue LoadSin = DAG.getLoad(ArgVT, dl, CallResult.second, SRet,
6656 MachinePointerInfo(), false, false, false, 0);
6658 // Address of cos field.
6659 SDValue Add = DAG.getNode(ISD::ADD, dl, PtrVT, SRet,
6660 DAG.getIntPtrConstant(ArgVT.getStoreSize(), dl));
6661 SDValue LoadCos = DAG.getLoad(ArgVT, dl, LoadSin.getValue(1), Add,
6662 MachinePointerInfo(), false, false, false, 0);
6664 SDVTList Tys = DAG.getVTList(ArgVT, ArgVT);
6665 return DAG.getNode(ISD::MERGE_VALUES, dl, Tys,
6666 LoadSin.getValue(0), LoadCos.getValue(0));
6669 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;
6680 Name = (VT == MVT::i32) ? "__rt_udiv" : "__rt_udiv64";
6682 SDValue ES = DAG.getExternalSymbol(Name, TLI.getPointerTy(DL));
6684 ARMTargetLowering::ArgListTy Args;
6686 for (auto AI : {1, 0}) {
6688 Arg.Node = Op.getOperand(AI);
6689 Arg.Ty = Arg.Node.getValueType().getTypeForEVT(*DAG.getContext());
6690 Args.push_back(Arg);
6693 CallLoweringInfo CLI(DAG);
6696 .setCallee(CallingConv::ARM_AAPCS_VFP, VT.getTypeForEVT(*DAG.getContext()),
6697 ES, std::move(Args), 0);
6699 return LowerCallTo(CLI).first;
6702 SDValue ARMTargetLowering::LowerDIV_Windows(SDValue Op,
6703 SelectionDAG &DAG) const {
6704 assert(Op.getValueType() == MVT::i32 &&
6705 "unexpected type for custom lowering DIV");
6708 SDValue DBZCHK = DAG.getNode(ARMISD::WIN__DBZCHK, dl, MVT::Other,
6709 DAG.getEntryNode(), Op.getOperand(1));
6711 return LowerWindowsDIVLibCall(Op, DAG, DBZCHK);
6714 void ARMTargetLowering::ExpandDIV_Windows(
6715 SDValue Op, SelectionDAG &DAG,
6716 SmallVectorImpl<SDValue> &Results) const {
6717 const auto &DL = DAG.getDataLayout();
6718 const auto &TLI = DAG.getTargetLoweringInfo();
6720 assert(Op.getValueType() == MVT::i64 &&
6721 "unexpected type for custom lowering DIV");
6724 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op.getOperand(1),
6725 DAG.getConstant(0, dl, MVT::i32));
6726 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op.getOperand(1),
6727 DAG.getConstant(1, dl, MVT::i32));
6728 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::i32, Lo, Hi);
6731 DAG.getNode(ARMISD::WIN__DBZCHK, dl, MVT::Other, DAG.getEntryNode(), Or);
6733 SDValue Result = LowerWindowsDIVLibCall(Op, DAG, DBZCHK);
6735 SDValue Lower = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Result);
6736 SDValue Upper = DAG.getNode(ISD::SRL, dl, MVT::i64, Result,
6737 DAG.getConstant(32, dl, TLI.getPointerTy(DL)));
6738 Upper = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Upper);
6740 Results.push_back(Lower);
6741 Results.push_back(Upper);
6744 static SDValue LowerAtomicLoadStore(SDValue Op, SelectionDAG &DAG) {
6745 // Monotonic load/store is legal for all targets
6746 if (cast<AtomicSDNode>(Op)->getOrdering() <= Monotonic)
6749 // Acquire/Release load/store is not legal for targets without a
6750 // dmb or equivalent available.
6754 static void ReplaceREADCYCLECOUNTER(SDNode *N,
6755 SmallVectorImpl<SDValue> &Results,
6757 const ARMSubtarget *Subtarget) {
6759 // Under Power Management extensions, the cycle-count is:
6760 // mrc p15, #0, <Rt>, c9, c13, #0
6761 SDValue Ops[] = { N->getOperand(0), // Chain
6762 DAG.getConstant(Intrinsic::arm_mrc, DL, MVT::i32),
6763 DAG.getConstant(15, DL, MVT::i32),
6764 DAG.getConstant(0, DL, MVT::i32),
6765 DAG.getConstant(9, DL, MVT::i32),
6766 DAG.getConstant(13, DL, MVT::i32),
6767 DAG.getConstant(0, DL, MVT::i32)
6770 SDValue Cycles32 = DAG.getNode(ISD::INTRINSIC_W_CHAIN, DL,
6771 DAG.getVTList(MVT::i32, MVT::Other), Ops);
6772 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Cycles32,
6773 DAG.getConstant(0, DL, MVT::i32)));
6774 Results.push_back(Cycles32.getValue(1));
6777 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
6778 switch (Op.getOpcode()) {
6779 default: llvm_unreachable("Don't know how to custom lower this!");
6780 case ISD::WRITE_REGISTER: return LowerWRITE_REGISTER(Op, DAG);
6781 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
6782 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
6783 case ISD::GlobalAddress:
6784 switch (Subtarget->getTargetTriple().getObjectFormat()) {
6785 default: llvm_unreachable("unknown object format");
6787 return LowerGlobalAddressWindows(Op, DAG);
6789 return LowerGlobalAddressELF(Op, DAG);
6791 return LowerGlobalAddressDarwin(Op, DAG);
6793 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
6794 case ISD::SELECT: return LowerSELECT(Op, DAG);
6795 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
6796 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
6797 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
6798 case ISD::VASTART: return LowerVASTART(Op, DAG);
6799 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, DAG, Subtarget);
6800 case ISD::PREFETCH: return LowerPREFETCH(Op, DAG, Subtarget);
6801 case ISD::SINT_TO_FP:
6802 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
6803 case ISD::FP_TO_SINT:
6804 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
6805 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
6806 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
6807 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
6808 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
6809 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
6810 case ISD::EH_SJLJ_SETUP_DISPATCH: return LowerEH_SJLJ_SETUP_DISPATCH(Op, DAG);
6811 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
6813 case ISD::BITCAST: return ExpandBITCAST(Op.getNode(), DAG);
6816 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
6817 case ISD::SREM: return LowerREM(Op.getNode(), DAG);
6818 case ISD::UREM: return LowerREM(Op.getNode(), DAG);
6819 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
6820 case ISD::SRL_PARTS:
6821 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
6823 case ISD::CTTZ_ZERO_UNDEF: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
6824 case ISD::CTPOP: return LowerCTPOP(Op.getNode(), DAG, Subtarget);
6825 case ISD::SETCC: return LowerVSETCC(Op, DAG);
6826 case ISD::ConstantFP: return LowerConstantFP(Op, DAG, Subtarget);
6827 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG, Subtarget);
6828 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
6829 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
6830 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
6831 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
6832 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
6833 case ISD::MUL: return LowerMUL(Op, DAG);
6834 case ISD::SDIV: return LowerSDIV(Op, DAG);
6836 if (Subtarget->isTargetWindows())
6837 return LowerDIV_Windows(Op, DAG);
6838 return LowerUDIV(Op, DAG);
6842 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
6847 return LowerXALUO(Op, DAG);
6848 case ISD::ATOMIC_LOAD:
6849 case ISD::ATOMIC_STORE: return LowerAtomicLoadStore(Op, DAG);
6850 case ISD::FSINCOS: return LowerFSINCOS(Op, DAG);
6852 case ISD::UDIVREM: return LowerDivRem(Op, DAG);
6853 case ISD::DYNAMIC_STACKALLOC:
6854 if (Subtarget->getTargetTriple().isWindowsItaniumEnvironment())
6855 return LowerDYNAMIC_STACKALLOC(Op, DAG);
6856 llvm_unreachable("Don't know how to custom lower this!");
6857 case ISD::FP_ROUND: return LowerFP_ROUND(Op, DAG);
6858 case ISD::FP_EXTEND: return LowerFP_EXTEND(Op, DAG);
6859 case ARMISD::WIN__DBZCHK: return SDValue();
6863 /// ReplaceNodeResults - Replace the results of node with an illegal result
6864 /// type with new values built out of custom code.
6865 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
6866 SmallVectorImpl<SDValue> &Results,
6867 SelectionDAG &DAG) const {
6869 switch (N->getOpcode()) {
6871 llvm_unreachable("Don't know how to custom expand this!");
6872 case ISD::READ_REGISTER:
6873 ExpandREAD_REGISTER(N, Results, DAG);
6876 Res = ExpandBITCAST(N, DAG);
6880 Res = Expand64BitShift(N, DAG, Subtarget);
6884 Res = LowerREM(N, DAG);
6886 case ISD::READCYCLECOUNTER:
6887 ReplaceREADCYCLECOUNTER(N, Results, DAG, Subtarget);
6890 assert(Subtarget->isTargetWindows() && "can only expand DIV on Windows");
6891 return ExpandDIV_Windows(SDValue(N, 0), DAG, Results);
6894 Results.push_back(Res);
6897 //===----------------------------------------------------------------------===//
6898 // ARM Scheduler Hooks
6899 //===----------------------------------------------------------------------===//
6901 /// SetupEntryBlockForSjLj - Insert code into the entry block that creates and
6902 /// registers the function context.
6903 void ARMTargetLowering::
6904 SetupEntryBlockForSjLj(MachineInstr *MI, MachineBasicBlock *MBB,
6905 MachineBasicBlock *DispatchBB, int FI) const {
6906 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
6907 DebugLoc dl = MI->getDebugLoc();
6908 MachineFunction *MF = MBB->getParent();
6909 MachineRegisterInfo *MRI = &MF->getRegInfo();
6910 MachineConstantPool *MCP = MF->getConstantPool();
6911 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
6912 const Function *F = MF->getFunction();
6914 bool isThumb = Subtarget->isThumb();
6915 bool isThumb2 = Subtarget->isThumb2();
6917 unsigned PCLabelId = AFI->createPICLabelUId();
6918 unsigned PCAdj = (isThumb || isThumb2) ? 4 : 8;
6919 ARMConstantPoolValue *CPV =
6920 ARMConstantPoolMBB::Create(F->getContext(), DispatchBB, PCLabelId, PCAdj);
6921 unsigned CPI = MCP->getConstantPoolIndex(CPV, 4);
6923 const TargetRegisterClass *TRC = isThumb ? &ARM::tGPRRegClass
6924 : &ARM::GPRRegClass;
6926 // Grab constant pool and fixed stack memory operands.
6927 MachineMemOperand *CPMMO =
6928 MF->getMachineMemOperand(MachinePointerInfo::getConstantPool(*MF),
6929 MachineMemOperand::MOLoad, 4, 4);
6931 MachineMemOperand *FIMMOSt =
6932 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(*MF, FI),
6933 MachineMemOperand::MOStore, 4, 4);
6935 // Load the address of the dispatch MBB into the jump buffer.
6937 // Incoming value: jbuf
6938 // ldr.n r5, LCPI1_1
6941 // str r5, [$jbuf, #+4] ; &jbuf[1]
6942 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6943 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2LDRpci), NewVReg1)
6944 .addConstantPoolIndex(CPI)
6945 .addMemOperand(CPMMO));
6946 // Set the low bit because of thumb mode.
6947 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6949 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2ORRri), NewVReg2)
6950 .addReg(NewVReg1, RegState::Kill)
6952 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6953 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg3)
6954 .addReg(NewVReg2, RegState::Kill)
6956 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2STRi12))
6957 .addReg(NewVReg3, RegState::Kill)
6959 .addImm(36) // &jbuf[1] :: pc
6960 .addMemOperand(FIMMOSt));
6961 } else if (isThumb) {
6962 // Incoming value: jbuf
6963 // ldr.n r1, LCPI1_4
6967 // add r2, $jbuf, #+4 ; &jbuf[1]
6969 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6970 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tLDRpci), NewVReg1)
6971 .addConstantPoolIndex(CPI)
6972 .addMemOperand(CPMMO));
6973 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6974 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg2)
6975 .addReg(NewVReg1, RegState::Kill)
6977 // Set the low bit because of thumb mode.
6978 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6979 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tMOVi8), NewVReg3)
6980 .addReg(ARM::CPSR, RegState::Define)
6982 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6983 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tORR), NewVReg4)
6984 .addReg(ARM::CPSR, RegState::Define)
6985 .addReg(NewVReg2, RegState::Kill)
6986 .addReg(NewVReg3, RegState::Kill));
6987 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6988 BuildMI(*MBB, MI, dl, TII->get(ARM::tADDframe), NewVReg5)
6990 .addImm(36); // &jbuf[1] :: pc
6991 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tSTRi))
6992 .addReg(NewVReg4, RegState::Kill)
6993 .addReg(NewVReg5, RegState::Kill)
6995 .addMemOperand(FIMMOSt));
6997 // Incoming value: jbuf
7000 // str r1, [$jbuf, #+4] ; &jbuf[1]
7001 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
7002 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::LDRi12), NewVReg1)
7003 .addConstantPoolIndex(CPI)
7005 .addMemOperand(CPMMO));
7006 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
7007 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::PICADD), NewVReg2)
7008 .addReg(NewVReg1, RegState::Kill)
7009 .addImm(PCLabelId));
7010 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::STRi12))
7011 .addReg(NewVReg2, RegState::Kill)
7013 .addImm(36) // &jbuf[1] :: pc
7014 .addMemOperand(FIMMOSt));
7018 void ARMTargetLowering::EmitSjLjDispatchBlock(MachineInstr *MI,
7019 MachineBasicBlock *MBB) const {
7020 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
7021 DebugLoc dl = MI->getDebugLoc();
7022 MachineFunction *MF = MBB->getParent();
7023 MachineRegisterInfo *MRI = &MF->getRegInfo();
7024 MachineFrameInfo *MFI = MF->getFrameInfo();
7025 int FI = MFI->getFunctionContextIndex();
7027 const TargetRegisterClass *TRC = Subtarget->isThumb() ? &ARM::tGPRRegClass
7028 : &ARM::GPRnopcRegClass;
7030 // Get a mapping of the call site numbers to all of the landing pads they're
7032 DenseMap<unsigned, SmallVector<MachineBasicBlock*, 2> > CallSiteNumToLPad;
7033 unsigned MaxCSNum = 0;
7034 MachineModuleInfo &MMI = MF->getMMI();
7035 for (MachineFunction::iterator BB = MF->begin(), E = MF->end(); BB != E;
7037 if (!BB->isEHPad()) continue;
7039 // FIXME: We should assert that the EH_LABEL is the first MI in the landing
7041 for (MachineBasicBlock::iterator
7042 II = BB->begin(), IE = BB->end(); II != IE; ++II) {
7043 if (!II->isEHLabel()) continue;
7045 MCSymbol *Sym = II->getOperand(0).getMCSymbol();
7046 if (!MMI.hasCallSiteLandingPad(Sym)) continue;
7048 SmallVectorImpl<unsigned> &CallSiteIdxs = MMI.getCallSiteLandingPad(Sym);
7049 for (SmallVectorImpl<unsigned>::iterator
7050 CSI = CallSiteIdxs.begin(), CSE = CallSiteIdxs.end();
7051 CSI != CSE; ++CSI) {
7052 CallSiteNumToLPad[*CSI].push_back(&*BB);
7053 MaxCSNum = std::max(MaxCSNum, *CSI);
7059 // Get an ordered list of the machine basic blocks for the jump table.
7060 std::vector<MachineBasicBlock*> LPadList;
7061 SmallPtrSet<MachineBasicBlock*, 64> InvokeBBs;
7062 LPadList.reserve(CallSiteNumToLPad.size());
7063 for (unsigned I = 1; I <= MaxCSNum; ++I) {
7064 SmallVectorImpl<MachineBasicBlock*> &MBBList = CallSiteNumToLPad[I];
7065 for (SmallVectorImpl<MachineBasicBlock*>::iterator
7066 II = MBBList.begin(), IE = MBBList.end(); II != IE; ++II) {
7067 LPadList.push_back(*II);
7068 InvokeBBs.insert((*II)->pred_begin(), (*II)->pred_end());
7072 assert(!LPadList.empty() &&
7073 "No landing pad destinations for the dispatch jump table!");
7075 // Create the jump table and associated information.
7076 MachineJumpTableInfo *JTI =
7077 MF->getOrCreateJumpTableInfo(MachineJumpTableInfo::EK_Inline);
7078 unsigned MJTI = JTI->createJumpTableIndex(LPadList);
7079 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
7081 // Create the MBBs for the dispatch code.
7083 // Shove the dispatch's address into the return slot in the function context.
7084 MachineBasicBlock *DispatchBB = MF->CreateMachineBasicBlock();
7085 DispatchBB->setIsEHPad();
7087 MachineBasicBlock *TrapBB = MF->CreateMachineBasicBlock();
7088 unsigned trap_opcode;
7089 if (Subtarget->isThumb())
7090 trap_opcode = ARM::tTRAP;
7092 trap_opcode = Subtarget->useNaClTrap() ? ARM::TRAPNaCl : ARM::TRAP;
7094 BuildMI(TrapBB, dl, TII->get(trap_opcode));
7095 DispatchBB->addSuccessor(TrapBB);
7097 MachineBasicBlock *DispContBB = MF->CreateMachineBasicBlock();
7098 DispatchBB->addSuccessor(DispContBB);
7101 MF->insert(MF->end(), DispatchBB);
7102 MF->insert(MF->end(), DispContBB);
7103 MF->insert(MF->end(), TrapBB);
7105 // Insert code into the entry block that creates and registers the function
7107 SetupEntryBlockForSjLj(MI, MBB, DispatchBB, FI);
7109 MachineMemOperand *FIMMOLd = MF->getMachineMemOperand(
7110 MachinePointerInfo::getFixedStack(*MF, FI),
7111 MachineMemOperand::MOLoad | MachineMemOperand::MOVolatile, 4, 4);
7113 MachineInstrBuilder MIB;
7114 MIB = BuildMI(DispatchBB, dl, TII->get(ARM::Int_eh_sjlj_dispatchsetup));
7116 const ARMBaseInstrInfo *AII = static_cast<const ARMBaseInstrInfo*>(TII);
7117 const ARMBaseRegisterInfo &RI = AII->getRegisterInfo();
7119 // Add a register mask with no preserved registers. This results in all
7120 // registers being marked as clobbered.
7121 MIB.addRegMask(RI.getNoPreservedMask());
7123 unsigned NumLPads = LPadList.size();
7124 if (Subtarget->isThumb2()) {
7125 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
7126 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2LDRi12), NewVReg1)
7129 .addMemOperand(FIMMOLd));
7131 if (NumLPads < 256) {
7132 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPri))
7134 .addImm(LPadList.size()));
7136 unsigned VReg1 = MRI->createVirtualRegister(TRC);
7137 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVi16), VReg1)
7138 .addImm(NumLPads & 0xFFFF));
7140 unsigned VReg2 = VReg1;
7141 if ((NumLPads & 0xFFFF0000) != 0) {
7142 VReg2 = MRI->createVirtualRegister(TRC);
7143 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVTi16), VReg2)
7145 .addImm(NumLPads >> 16));
7148 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPrr))
7153 BuildMI(DispatchBB, dl, TII->get(ARM::t2Bcc))
7158 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
7159 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::t2LEApcrelJT),NewVReg3)
7160 .addJumpTableIndex(MJTI));
7162 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
7165 BuildMI(DispContBB, dl, TII->get(ARM::t2ADDrs), NewVReg4)
7166 .addReg(NewVReg3, RegState::Kill)
7168 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
7170 BuildMI(DispContBB, dl, TII->get(ARM::t2BR_JT))
7171 .addReg(NewVReg4, RegState::Kill)
7173 .addJumpTableIndex(MJTI);
7174 } else if (Subtarget->isThumb()) {
7175 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
7176 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRspi), NewVReg1)
7179 .addMemOperand(FIMMOLd));
7181 if (NumLPads < 256) {
7182 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPi8))
7186 MachineConstantPool *ConstantPool = MF->getConstantPool();
7187 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
7188 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
7190 // MachineConstantPool wants an explicit alignment.
7191 unsigned Align = MF->getDataLayout().getPrefTypeAlignment(Int32Ty);
7193 Align = MF->getDataLayout().getTypeAllocSize(C->getType());
7194 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
7196 unsigned VReg1 = MRI->createVirtualRegister(TRC);
7197 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRpci))
7198 .addReg(VReg1, RegState::Define)
7199 .addConstantPoolIndex(Idx));
7200 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPr))
7205 BuildMI(DispatchBB, dl, TII->get(ARM::tBcc))
7210 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
7211 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLSLri), NewVReg2)
7212 .addReg(ARM::CPSR, RegState::Define)
7216 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
7217 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLEApcrelJT), NewVReg3)
7218 .addJumpTableIndex(MJTI));
7220 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
7221 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg4)
7222 .addReg(ARM::CPSR, RegState::Define)
7223 .addReg(NewVReg2, RegState::Kill)
7226 MachineMemOperand *JTMMOLd = MF->getMachineMemOperand(
7227 MachinePointerInfo::getJumpTable(*MF), MachineMemOperand::MOLoad, 4, 4);
7229 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
7230 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLDRi), NewVReg5)
7231 .addReg(NewVReg4, RegState::Kill)
7233 .addMemOperand(JTMMOLd));
7235 unsigned NewVReg6 = NewVReg5;
7236 if (RelocM == Reloc::PIC_) {
7237 NewVReg6 = MRI->createVirtualRegister(TRC);
7238 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg6)
7239 .addReg(ARM::CPSR, RegState::Define)
7240 .addReg(NewVReg5, RegState::Kill)
7244 BuildMI(DispContBB, dl, TII->get(ARM::tBR_JTr))
7245 .addReg(NewVReg6, RegState::Kill)
7246 .addJumpTableIndex(MJTI);
7248 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
7249 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRi12), NewVReg1)
7252 .addMemOperand(FIMMOLd));
7254 if (NumLPads < 256) {
7255 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPri))
7258 } else if (Subtarget->hasV6T2Ops() && isUInt<16>(NumLPads)) {
7259 unsigned VReg1 = MRI->createVirtualRegister(TRC);
7260 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVi16), VReg1)
7261 .addImm(NumLPads & 0xFFFF));
7263 unsigned VReg2 = VReg1;
7264 if ((NumLPads & 0xFFFF0000) != 0) {
7265 VReg2 = MRI->createVirtualRegister(TRC);
7266 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVTi16), VReg2)
7268 .addImm(NumLPads >> 16));
7271 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
7275 MachineConstantPool *ConstantPool = MF->getConstantPool();
7276 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
7277 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
7279 // MachineConstantPool wants an explicit alignment.
7280 unsigned Align = MF->getDataLayout().getPrefTypeAlignment(Int32Ty);
7282 Align = MF->getDataLayout().getTypeAllocSize(C->getType());
7283 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
7285 unsigned VReg1 = MRI->createVirtualRegister(TRC);
7286 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRcp))
7287 .addReg(VReg1, RegState::Define)
7288 .addConstantPoolIndex(Idx)
7290 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
7292 .addReg(VReg1, RegState::Kill));
7295 BuildMI(DispatchBB, dl, TII->get(ARM::Bcc))
7300 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
7302 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::MOVsi), NewVReg3)
7304 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
7305 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
7306 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::LEApcrelJT), NewVReg4)
7307 .addJumpTableIndex(MJTI));
7309 MachineMemOperand *JTMMOLd = MF->getMachineMemOperand(
7310 MachinePointerInfo::getJumpTable(*MF), MachineMemOperand::MOLoad, 4, 4);
7311 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
7313 BuildMI(DispContBB, dl, TII->get(ARM::LDRrs), NewVReg5)
7314 .addReg(NewVReg3, RegState::Kill)
7317 .addMemOperand(JTMMOLd));
7319 if (RelocM == Reloc::PIC_) {
7320 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTadd))
7321 .addReg(NewVReg5, RegState::Kill)
7323 .addJumpTableIndex(MJTI);
7325 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTr))
7326 .addReg(NewVReg5, RegState::Kill)
7327 .addJumpTableIndex(MJTI);
7331 // Add the jump table entries as successors to the MBB.
7332 SmallPtrSet<MachineBasicBlock*, 8> SeenMBBs;
7333 for (std::vector<MachineBasicBlock*>::iterator
7334 I = LPadList.begin(), E = LPadList.end(); I != E; ++I) {
7335 MachineBasicBlock *CurMBB = *I;
7336 if (SeenMBBs.insert(CurMBB).second)
7337 DispContBB->addSuccessor(CurMBB);
7340 // N.B. the order the invoke BBs are processed in doesn't matter here.
7341 const MCPhysReg *SavedRegs = RI.getCalleeSavedRegs(MF);
7342 SmallVector<MachineBasicBlock*, 64> MBBLPads;
7343 for (MachineBasicBlock *BB : InvokeBBs) {
7345 // Remove the landing pad successor from the invoke block and replace it
7346 // with the new dispatch block.
7347 SmallVector<MachineBasicBlock*, 4> Successors(BB->succ_begin(),
7349 while (!Successors.empty()) {
7350 MachineBasicBlock *SMBB = Successors.pop_back_val();
7351 if (SMBB->isEHPad()) {
7352 BB->removeSuccessor(SMBB);
7353 MBBLPads.push_back(SMBB);
7357 BB->addSuccessor(DispatchBB);
7359 // Find the invoke call and mark all of the callee-saved registers as
7360 // 'implicit defined' so that they're spilled. This prevents code from
7361 // moving instructions to before the EH block, where they will never be
7363 for (MachineBasicBlock::reverse_iterator
7364 II = BB->rbegin(), IE = BB->rend(); II != IE; ++II) {
7365 if (!II->isCall()) continue;
7367 DenseMap<unsigned, bool> DefRegs;
7368 for (MachineInstr::mop_iterator
7369 OI = II->operands_begin(), OE = II->operands_end();
7371 if (!OI->isReg()) continue;
7372 DefRegs[OI->getReg()] = true;
7375 MachineInstrBuilder MIB(*MF, &*II);
7377 for (unsigned i = 0; SavedRegs[i] != 0; ++i) {
7378 unsigned Reg = SavedRegs[i];
7379 if (Subtarget->isThumb2() &&
7380 !ARM::tGPRRegClass.contains(Reg) &&
7381 !ARM::hGPRRegClass.contains(Reg))
7383 if (Subtarget->isThumb1Only() && !ARM::tGPRRegClass.contains(Reg))
7385 if (!Subtarget->isThumb() && !ARM::GPRRegClass.contains(Reg))
7388 MIB.addReg(Reg, RegState::ImplicitDefine | RegState::Dead);
7395 // Mark all former landing pads as non-landing pads. The dispatch is the only
7397 for (SmallVectorImpl<MachineBasicBlock*>::iterator
7398 I = MBBLPads.begin(), E = MBBLPads.end(); I != E; ++I)
7399 (*I)->setIsEHPad(false);
7401 // The instruction is gone now.
7402 MI->eraseFromParent();
7406 MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) {
7407 for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
7408 E = MBB->succ_end(); I != E; ++I)
7411 llvm_unreachable("Expecting a BB with two successors!");
7414 /// Return the load opcode for a given load size. If load size >= 8,
7415 /// neon opcode will be returned.
7416 static unsigned getLdOpcode(unsigned LdSize, bool IsThumb1, bool IsThumb2) {
7418 return LdSize == 16 ? ARM::VLD1q32wb_fixed
7419 : LdSize == 8 ? ARM::VLD1d32wb_fixed : 0;
7421 return LdSize == 4 ? ARM::tLDRi
7422 : LdSize == 2 ? ARM::tLDRHi
7423 : LdSize == 1 ? ARM::tLDRBi : 0;
7425 return LdSize == 4 ? ARM::t2LDR_POST
7426 : LdSize == 2 ? ARM::t2LDRH_POST
7427 : LdSize == 1 ? ARM::t2LDRB_POST : 0;
7428 return LdSize == 4 ? ARM::LDR_POST_IMM
7429 : LdSize == 2 ? ARM::LDRH_POST
7430 : LdSize == 1 ? ARM::LDRB_POST_IMM : 0;
7433 /// Return the store opcode for a given store size. If store size >= 8,
7434 /// neon opcode will be returned.
7435 static unsigned getStOpcode(unsigned StSize, bool IsThumb1, bool IsThumb2) {
7437 return StSize == 16 ? ARM::VST1q32wb_fixed
7438 : StSize == 8 ? ARM::VST1d32wb_fixed : 0;
7440 return StSize == 4 ? ARM::tSTRi
7441 : StSize == 2 ? ARM::tSTRHi
7442 : StSize == 1 ? ARM::tSTRBi : 0;
7444 return StSize == 4 ? ARM::t2STR_POST
7445 : StSize == 2 ? ARM::t2STRH_POST
7446 : StSize == 1 ? ARM::t2STRB_POST : 0;
7447 return StSize == 4 ? ARM::STR_POST_IMM
7448 : StSize == 2 ? ARM::STRH_POST
7449 : StSize == 1 ? ARM::STRB_POST_IMM : 0;
7452 /// Emit a post-increment load operation with given size. The instructions
7453 /// will be added to BB at Pos.
7454 static void emitPostLd(MachineBasicBlock *BB, MachineInstr *Pos,
7455 const TargetInstrInfo *TII, DebugLoc dl,
7456 unsigned LdSize, unsigned Data, unsigned AddrIn,
7457 unsigned AddrOut, bool IsThumb1, bool IsThumb2) {
7458 unsigned LdOpc = getLdOpcode(LdSize, IsThumb1, IsThumb2);
7459 assert(LdOpc != 0 && "Should have a load opcode");
7461 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
7462 .addReg(AddrOut, RegState::Define).addReg(AddrIn)
7464 } else if (IsThumb1) {
7465 // load + update AddrIn
7466 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
7467 .addReg(AddrIn).addImm(0));
7468 MachineInstrBuilder MIB =
7469 BuildMI(*BB, Pos, dl, TII->get(ARM::tADDi8), AddrOut);
7470 MIB = AddDefaultT1CC(MIB);
7471 MIB.addReg(AddrIn).addImm(LdSize);
7472 AddDefaultPred(MIB);
7473 } else if (IsThumb2) {
7474 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
7475 .addReg(AddrOut, RegState::Define).addReg(AddrIn)
7478 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
7479 .addReg(AddrOut, RegState::Define).addReg(AddrIn)
7480 .addReg(0).addImm(LdSize));
7484 /// Emit a post-increment store operation with given size. The instructions
7485 /// will be added to BB at Pos.
7486 static void emitPostSt(MachineBasicBlock *BB, MachineInstr *Pos,
7487 const TargetInstrInfo *TII, DebugLoc dl,
7488 unsigned StSize, unsigned Data, unsigned AddrIn,
7489 unsigned AddrOut, bool IsThumb1, bool IsThumb2) {
7490 unsigned StOpc = getStOpcode(StSize, IsThumb1, IsThumb2);
7491 assert(StOpc != 0 && "Should have a store opcode");
7493 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc), AddrOut)
7494 .addReg(AddrIn).addImm(0).addReg(Data));
7495 } else if (IsThumb1) {
7496 // store + update AddrIn
7497 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc)).addReg(Data)
7498 .addReg(AddrIn).addImm(0));
7499 MachineInstrBuilder MIB =
7500 BuildMI(*BB, Pos, dl, TII->get(ARM::tADDi8), AddrOut);
7501 MIB = AddDefaultT1CC(MIB);
7502 MIB.addReg(AddrIn).addImm(StSize);
7503 AddDefaultPred(MIB);
7504 } else if (IsThumb2) {
7505 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc), AddrOut)
7506 .addReg(Data).addReg(AddrIn).addImm(StSize));
7508 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc), AddrOut)
7509 .addReg(Data).addReg(AddrIn).addReg(0)
7515 ARMTargetLowering::EmitStructByval(MachineInstr *MI,
7516 MachineBasicBlock *BB) const {
7517 // This pseudo instruction has 3 operands: dst, src, size
7518 // We expand it to a loop if size > Subtarget->getMaxInlineSizeThreshold().
7519 // Otherwise, we will generate unrolled scalar copies.
7520 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
7521 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7522 MachineFunction::iterator It = ++BB->getIterator();
7524 unsigned dest = MI->getOperand(0).getReg();
7525 unsigned src = MI->getOperand(1).getReg();
7526 unsigned SizeVal = MI->getOperand(2).getImm();
7527 unsigned Align = MI->getOperand(3).getImm();
7528 DebugLoc dl = MI->getDebugLoc();
7530 MachineFunction *MF = BB->getParent();
7531 MachineRegisterInfo &MRI = MF->getRegInfo();
7532 unsigned UnitSize = 0;
7533 const TargetRegisterClass *TRC = nullptr;
7534 const TargetRegisterClass *VecTRC = nullptr;
7536 bool IsThumb1 = Subtarget->isThumb1Only();
7537 bool IsThumb2 = Subtarget->isThumb2();
7541 } else if (Align & 2) {
7544 // Check whether we can use NEON instructions.
7545 if (!MF->getFunction()->hasFnAttribute(Attribute::NoImplicitFloat) &&
7546 Subtarget->hasNEON()) {
7547 if ((Align % 16 == 0) && SizeVal >= 16)
7549 else if ((Align % 8 == 0) && SizeVal >= 8)
7552 // Can't use NEON instructions.
7557 // Select the correct opcode and register class for unit size load/store
7558 bool IsNeon = UnitSize >= 8;
7559 TRC = (IsThumb1 || IsThumb2) ? &ARM::tGPRRegClass : &ARM::GPRRegClass;
7561 VecTRC = UnitSize == 16 ? &ARM::DPairRegClass
7562 : UnitSize == 8 ? &ARM::DPRRegClass
7565 unsigned BytesLeft = SizeVal % UnitSize;
7566 unsigned LoopSize = SizeVal - BytesLeft;
7568 if (SizeVal <= Subtarget->getMaxInlineSizeThreshold()) {
7569 // Use LDR and STR to copy.
7570 // [scratch, srcOut] = LDR_POST(srcIn, UnitSize)
7571 // [destOut] = STR_POST(scratch, destIn, UnitSize)
7572 unsigned srcIn = src;
7573 unsigned destIn = dest;
7574 for (unsigned i = 0; i < LoopSize; i+=UnitSize) {
7575 unsigned srcOut = MRI.createVirtualRegister(TRC);
7576 unsigned destOut = MRI.createVirtualRegister(TRC);
7577 unsigned scratch = MRI.createVirtualRegister(IsNeon ? VecTRC : TRC);
7578 emitPostLd(BB, MI, TII, dl, UnitSize, scratch, srcIn, srcOut,
7579 IsThumb1, IsThumb2);
7580 emitPostSt(BB, MI, TII, dl, UnitSize, scratch, destIn, destOut,
7581 IsThumb1, IsThumb2);
7586 // Handle the leftover bytes with LDRB and STRB.
7587 // [scratch, srcOut] = LDRB_POST(srcIn, 1)
7588 // [destOut] = STRB_POST(scratch, destIn, 1)
7589 for (unsigned i = 0; i < BytesLeft; i++) {
7590 unsigned srcOut = MRI.createVirtualRegister(TRC);
7591 unsigned destOut = MRI.createVirtualRegister(TRC);
7592 unsigned scratch = MRI.createVirtualRegister(TRC);
7593 emitPostLd(BB, MI, TII, dl, 1, scratch, srcIn, srcOut,
7594 IsThumb1, IsThumb2);
7595 emitPostSt(BB, MI, TII, dl, 1, scratch, destIn, destOut,
7596 IsThumb1, IsThumb2);
7600 MI->eraseFromParent(); // The instruction is gone now.
7604 // Expand the pseudo op to a loop.
7607 // movw varEnd, # --> with thumb2
7609 // ldrcp varEnd, idx --> without thumb2
7610 // fallthrough --> loopMBB
7612 // PHI varPhi, varEnd, varLoop
7613 // PHI srcPhi, src, srcLoop
7614 // PHI destPhi, dst, destLoop
7615 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
7616 // [destLoop] = STR_POST(scratch, destPhi, UnitSize)
7617 // subs varLoop, varPhi, #UnitSize
7619 // fallthrough --> exitMBB
7621 // epilogue to handle left-over bytes
7622 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
7623 // [destOut] = STRB_POST(scratch, destLoop, 1)
7624 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
7625 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
7626 MF->insert(It, loopMBB);
7627 MF->insert(It, exitMBB);
7629 // Transfer the remainder of BB and its successor edges to exitMBB.
7630 exitMBB->splice(exitMBB->begin(), BB,
7631 std::next(MachineBasicBlock::iterator(MI)), BB->end());
7632 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
7634 // Load an immediate to varEnd.
7635 unsigned varEnd = MRI.createVirtualRegister(TRC);
7636 if (Subtarget->useMovt(*MF)) {
7637 unsigned Vtmp = varEnd;
7638 if ((LoopSize & 0xFFFF0000) != 0)
7639 Vtmp = MRI.createVirtualRegister(TRC);
7640 AddDefaultPred(BuildMI(BB, dl,
7641 TII->get(IsThumb2 ? ARM::t2MOVi16 : ARM::MOVi16),
7642 Vtmp).addImm(LoopSize & 0xFFFF));
7644 if ((LoopSize & 0xFFFF0000) != 0)
7645 AddDefaultPred(BuildMI(BB, dl,
7646 TII->get(IsThumb2 ? ARM::t2MOVTi16 : ARM::MOVTi16),
7649 .addImm(LoopSize >> 16));
7651 MachineConstantPool *ConstantPool = MF->getConstantPool();
7652 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
7653 const Constant *C = ConstantInt::get(Int32Ty, LoopSize);
7655 // MachineConstantPool wants an explicit alignment.
7656 unsigned Align = MF->getDataLayout().getPrefTypeAlignment(Int32Ty);
7658 Align = MF->getDataLayout().getTypeAllocSize(C->getType());
7659 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
7662 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(ARM::tLDRpci)).addReg(
7663 varEnd, RegState::Define).addConstantPoolIndex(Idx));
7665 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(ARM::LDRcp)).addReg(
7666 varEnd, RegState::Define).addConstantPoolIndex(Idx).addImm(0));
7668 BB->addSuccessor(loopMBB);
7670 // Generate the loop body:
7671 // varPhi = PHI(varLoop, varEnd)
7672 // srcPhi = PHI(srcLoop, src)
7673 // destPhi = PHI(destLoop, dst)
7674 MachineBasicBlock *entryBB = BB;
7676 unsigned varLoop = MRI.createVirtualRegister(TRC);
7677 unsigned varPhi = MRI.createVirtualRegister(TRC);
7678 unsigned srcLoop = MRI.createVirtualRegister(TRC);
7679 unsigned srcPhi = MRI.createVirtualRegister(TRC);
7680 unsigned destLoop = MRI.createVirtualRegister(TRC);
7681 unsigned destPhi = MRI.createVirtualRegister(TRC);
7683 BuildMI(*BB, BB->begin(), dl, TII->get(ARM::PHI), varPhi)
7684 .addReg(varLoop).addMBB(loopMBB)
7685 .addReg(varEnd).addMBB(entryBB);
7686 BuildMI(BB, dl, TII->get(ARM::PHI), srcPhi)
7687 .addReg(srcLoop).addMBB(loopMBB)
7688 .addReg(src).addMBB(entryBB);
7689 BuildMI(BB, dl, TII->get(ARM::PHI), destPhi)
7690 .addReg(destLoop).addMBB(loopMBB)
7691 .addReg(dest).addMBB(entryBB);
7693 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
7694 // [destLoop] = STR_POST(scratch, destPhi, UnitSiz)
7695 unsigned scratch = MRI.createVirtualRegister(IsNeon ? VecTRC : TRC);
7696 emitPostLd(BB, BB->end(), TII, dl, UnitSize, scratch, srcPhi, srcLoop,
7697 IsThumb1, IsThumb2);
7698 emitPostSt(BB, BB->end(), TII, dl, UnitSize, scratch, destPhi, destLoop,
7699 IsThumb1, IsThumb2);
7701 // Decrement loop variable by UnitSize.
7703 MachineInstrBuilder MIB =
7704 BuildMI(*BB, BB->end(), dl, TII->get(ARM::tSUBi8), varLoop);
7705 MIB = AddDefaultT1CC(MIB);
7706 MIB.addReg(varPhi).addImm(UnitSize);
7707 AddDefaultPred(MIB);
7709 MachineInstrBuilder MIB =
7710 BuildMI(*BB, BB->end(), dl,
7711 TII->get(IsThumb2 ? ARM::t2SUBri : ARM::SUBri), varLoop);
7712 AddDefaultCC(AddDefaultPred(MIB.addReg(varPhi).addImm(UnitSize)));
7713 MIB->getOperand(5).setReg(ARM::CPSR);
7714 MIB->getOperand(5).setIsDef(true);
7716 BuildMI(*BB, BB->end(), dl,
7717 TII->get(IsThumb1 ? ARM::tBcc : IsThumb2 ? ARM::t2Bcc : ARM::Bcc))
7718 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
7720 // loopMBB can loop back to loopMBB or fall through to exitMBB.
7721 BB->addSuccessor(loopMBB);
7722 BB->addSuccessor(exitMBB);
7724 // Add epilogue to handle BytesLeft.
7726 MachineInstr *StartOfExit = exitMBB->begin();
7728 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
7729 // [destOut] = STRB_POST(scratch, destLoop, 1)
7730 unsigned srcIn = srcLoop;
7731 unsigned destIn = destLoop;
7732 for (unsigned i = 0; i < BytesLeft; i++) {
7733 unsigned srcOut = MRI.createVirtualRegister(TRC);
7734 unsigned destOut = MRI.createVirtualRegister(TRC);
7735 unsigned scratch = MRI.createVirtualRegister(TRC);
7736 emitPostLd(BB, StartOfExit, TII, dl, 1, scratch, srcIn, srcOut,
7737 IsThumb1, IsThumb2);
7738 emitPostSt(BB, StartOfExit, TII, dl, 1, scratch, destIn, destOut,
7739 IsThumb1, IsThumb2);
7744 MI->eraseFromParent(); // The instruction is gone now.
7749 ARMTargetLowering::EmitLowered__chkstk(MachineInstr *MI,
7750 MachineBasicBlock *MBB) const {
7751 const TargetMachine &TM = getTargetMachine();
7752 const TargetInstrInfo &TII = *Subtarget->getInstrInfo();
7753 DebugLoc DL = MI->getDebugLoc();
7755 assert(Subtarget->isTargetWindows() &&
7756 "__chkstk is only supported on Windows");
7757 assert(Subtarget->isThumb2() && "Windows on ARM requires Thumb-2 mode");
7759 // __chkstk takes the number of words to allocate on the stack in R4, and
7760 // returns the stack adjustment in number of bytes in R4. This will not
7761 // clober any other registers (other than the obvious lr).
7763 // Although, technically, IP should be considered a register which may be
7764 // clobbered, the call itself will not touch it. Windows on ARM is a pure
7765 // thumb-2 environment, so there is no interworking required. As a result, we
7766 // do not expect a veneer to be emitted by the linker, clobbering IP.
7768 // Each module receives its own copy of __chkstk, so no import thunk is
7769 // required, again, ensuring that IP is not clobbered.
7771 // Finally, although some linkers may theoretically provide a trampoline for
7772 // out of range calls (which is quite common due to a 32M range limitation of
7773 // branches for Thumb), we can generate the long-call version via
7774 // -mcmodel=large, alleviating the need for the trampoline which may clobber
7777 switch (TM.getCodeModel()) {
7778 case CodeModel::Small:
7779 case CodeModel::Medium:
7780 case CodeModel::Default:
7781 case CodeModel::Kernel:
7782 BuildMI(*MBB, MI, DL, TII.get(ARM::tBL))
7783 .addImm((unsigned)ARMCC::AL).addReg(0)
7784 .addExternalSymbol("__chkstk")
7785 .addReg(ARM::R4, RegState::Implicit | RegState::Kill)
7786 .addReg(ARM::R4, RegState::Implicit | RegState::Define)
7787 .addReg(ARM::R12, RegState::Implicit | RegState::Define | RegState::Dead);
7789 case CodeModel::Large:
7790 case CodeModel::JITDefault: {
7791 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
7792 unsigned Reg = MRI.createVirtualRegister(&ARM::rGPRRegClass);
7794 BuildMI(*MBB, MI, DL, TII.get(ARM::t2MOVi32imm), Reg)
7795 .addExternalSymbol("__chkstk");
7796 BuildMI(*MBB, MI, DL, TII.get(ARM::tBLXr))
7797 .addImm((unsigned)ARMCC::AL).addReg(0)
7798 .addReg(Reg, RegState::Kill)
7799 .addReg(ARM::R4, RegState::Implicit | RegState::Kill)
7800 .addReg(ARM::R4, RegState::Implicit | RegState::Define)
7801 .addReg(ARM::R12, RegState::Implicit | RegState::Define | RegState::Dead);
7806 AddDefaultCC(AddDefaultPred(BuildMI(*MBB, MI, DL, TII.get(ARM::t2SUBrr),
7808 .addReg(ARM::SP).addReg(ARM::R4)));
7810 MI->eraseFromParent();
7815 ARMTargetLowering::EmitLowered__dbzchk(MachineInstr *MI,
7816 MachineBasicBlock *MBB) const {
7817 DebugLoc DL = MI->getDebugLoc();
7818 MachineFunction *MF = MBB->getParent();
7819 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
7821 MachineBasicBlock *ContBB = MF->CreateMachineBasicBlock();
7822 MF->push_back(ContBB);
7823 ContBB->splice(ContBB->begin(), MBB,
7824 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
7825 MBB->addSuccessor(ContBB);
7827 MachineBasicBlock *TrapBB = MF->CreateMachineBasicBlock();
7828 MF->push_back(TrapBB);
7829 BuildMI(TrapBB, DL, TII->get(ARM::t2UDF)).addImm(249);
7830 MBB->addSuccessor(TrapBB);
7832 BuildMI(*MBB, MI, DL, TII->get(ARM::tCBZ))
7833 .addReg(MI->getOperand(0).getReg())
7836 MI->eraseFromParent();
7841 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
7842 MachineBasicBlock *BB) const {
7843 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
7844 DebugLoc dl = MI->getDebugLoc();
7845 bool isThumb2 = Subtarget->isThumb2();
7846 switch (MI->getOpcode()) {
7849 llvm_unreachable("Unexpected instr type to insert");
7851 // The Thumb2 pre-indexed stores have the same MI operands, they just
7852 // define them differently in the .td files from the isel patterns, so
7853 // they need pseudos.
7854 case ARM::t2STR_preidx:
7855 MI->setDesc(TII->get(ARM::t2STR_PRE));
7857 case ARM::t2STRB_preidx:
7858 MI->setDesc(TII->get(ARM::t2STRB_PRE));
7860 case ARM::t2STRH_preidx:
7861 MI->setDesc(TII->get(ARM::t2STRH_PRE));
7864 case ARM::STRi_preidx:
7865 case ARM::STRBi_preidx: {
7866 unsigned NewOpc = MI->getOpcode() == ARM::STRi_preidx ?
7867 ARM::STR_PRE_IMM : ARM::STRB_PRE_IMM;
7868 // Decode the offset.
7869 unsigned Offset = MI->getOperand(4).getImm();
7870 bool isSub = ARM_AM::getAM2Op(Offset) == ARM_AM::sub;
7871 Offset = ARM_AM::getAM2Offset(Offset);
7875 MachineMemOperand *MMO = *MI->memoperands_begin();
7876 BuildMI(*BB, MI, dl, TII->get(NewOpc))
7877 .addOperand(MI->getOperand(0)) // Rn_wb
7878 .addOperand(MI->getOperand(1)) // Rt
7879 .addOperand(MI->getOperand(2)) // Rn
7880 .addImm(Offset) // offset (skip GPR==zero_reg)
7881 .addOperand(MI->getOperand(5)) // pred
7882 .addOperand(MI->getOperand(6))
7883 .addMemOperand(MMO);
7884 MI->eraseFromParent();
7887 case ARM::STRr_preidx:
7888 case ARM::STRBr_preidx:
7889 case ARM::STRH_preidx: {
7891 switch (MI->getOpcode()) {
7892 default: llvm_unreachable("unexpected opcode!");
7893 case ARM::STRr_preidx: NewOpc = ARM::STR_PRE_REG; break;
7894 case ARM::STRBr_preidx: NewOpc = ARM::STRB_PRE_REG; break;
7895 case ARM::STRH_preidx: NewOpc = ARM::STRH_PRE; break;
7897 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(NewOpc));
7898 for (unsigned i = 0; i < MI->getNumOperands(); ++i)
7899 MIB.addOperand(MI->getOperand(i));
7900 MI->eraseFromParent();
7904 case ARM::tMOVCCr_pseudo: {
7905 // To "insert" a SELECT_CC instruction, we actually have to insert the
7906 // diamond control-flow pattern. The incoming instruction knows the
7907 // destination vreg to set, the condition code register to branch on, the
7908 // true/false values to select between, and a branch opcode to use.
7909 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7910 MachineFunction::iterator It = ++BB->getIterator();
7915 // cmpTY ccX, r1, r2
7917 // fallthrough --> copy0MBB
7918 MachineBasicBlock *thisMBB = BB;
7919 MachineFunction *F = BB->getParent();
7920 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
7921 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
7922 F->insert(It, copy0MBB);
7923 F->insert(It, sinkMBB);
7925 // Transfer the remainder of BB and its successor edges to sinkMBB.
7926 sinkMBB->splice(sinkMBB->begin(), BB,
7927 std::next(MachineBasicBlock::iterator(MI)), BB->end());
7928 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
7930 BB->addSuccessor(copy0MBB);
7931 BB->addSuccessor(sinkMBB);
7933 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
7934 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
7937 // %FalseValue = ...
7938 // # fallthrough to sinkMBB
7941 // Update machine-CFG edges
7942 BB->addSuccessor(sinkMBB);
7945 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
7948 BuildMI(*BB, BB->begin(), dl,
7949 TII->get(ARM::PHI), MI->getOperand(0).getReg())
7950 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
7951 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
7953 MI->eraseFromParent(); // The pseudo instruction is gone now.
7958 case ARM::BCCZi64: {
7959 // If there is an unconditional branch to the other successor, remove it.
7960 BB->erase(std::next(MachineBasicBlock::iterator(MI)), BB->end());
7962 // Compare both parts that make up the double comparison separately for
7964 bool RHSisZero = MI->getOpcode() == ARM::BCCZi64;
7966 unsigned LHS1 = MI->getOperand(1).getReg();
7967 unsigned LHS2 = MI->getOperand(2).getReg();
7969 AddDefaultPred(BuildMI(BB, dl,
7970 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7971 .addReg(LHS1).addImm(0));
7972 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7973 .addReg(LHS2).addImm(0)
7974 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
7976 unsigned RHS1 = MI->getOperand(3).getReg();
7977 unsigned RHS2 = MI->getOperand(4).getReg();
7978 AddDefaultPred(BuildMI(BB, dl,
7979 TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
7980 .addReg(LHS1).addReg(RHS1));
7981 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
7982 .addReg(LHS2).addReg(RHS2)
7983 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
7986 MachineBasicBlock *destMBB = MI->getOperand(RHSisZero ? 3 : 5).getMBB();
7987 MachineBasicBlock *exitMBB = OtherSucc(BB, destMBB);
7988 if (MI->getOperand(0).getImm() == ARMCC::NE)
7989 std::swap(destMBB, exitMBB);
7991 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
7992 .addMBB(destMBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
7994 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2B)).addMBB(exitMBB));
7996 BuildMI(BB, dl, TII->get(ARM::B)) .addMBB(exitMBB);
7998 MI->eraseFromParent(); // The pseudo instruction is gone now.
8002 case ARM::Int_eh_sjlj_setjmp:
8003 case ARM::Int_eh_sjlj_setjmp_nofp:
8004 case ARM::tInt_eh_sjlj_setjmp:
8005 case ARM::t2Int_eh_sjlj_setjmp:
8006 case ARM::t2Int_eh_sjlj_setjmp_nofp:
8009 case ARM::Int_eh_sjlj_setup_dispatch:
8010 EmitSjLjDispatchBlock(MI, BB);
8015 // To insert an ABS instruction, we have to insert the
8016 // diamond control-flow pattern. The incoming instruction knows the
8017 // source vreg to test against 0, the destination vreg to set,
8018 // the condition code register to branch on, the
8019 // true/false values to select between, and a branch opcode to use.
8024 // BCC (branch to SinkBB if V0 >= 0)
8025 // RSBBB: V3 = RSBri V2, 0 (compute ABS if V2 < 0)
8026 // SinkBB: V1 = PHI(V2, V3)
8027 const BasicBlock *LLVM_BB = BB->getBasicBlock();
8028 MachineFunction::iterator BBI = ++BB->getIterator();
8029 MachineFunction *Fn = BB->getParent();
8030 MachineBasicBlock *RSBBB = Fn->CreateMachineBasicBlock(LLVM_BB);
8031 MachineBasicBlock *SinkBB = Fn->CreateMachineBasicBlock(LLVM_BB);
8032 Fn->insert(BBI, RSBBB);
8033 Fn->insert(BBI, SinkBB);
8035 unsigned int ABSSrcReg = MI->getOperand(1).getReg();
8036 unsigned int ABSDstReg = MI->getOperand(0).getReg();
8037 bool ABSSrcKIll = MI->getOperand(1).isKill();
8038 bool isThumb2 = Subtarget->isThumb2();
8039 MachineRegisterInfo &MRI = Fn->getRegInfo();
8040 // In Thumb mode S must not be specified if source register is the SP or
8041 // PC and if destination register is the SP, so restrict register class
8042 unsigned NewRsbDstReg =
8043 MRI.createVirtualRegister(isThumb2 ? &ARM::rGPRRegClass : &ARM::GPRRegClass);
8045 // Transfer the remainder of BB and its successor edges to sinkMBB.
8046 SinkBB->splice(SinkBB->begin(), BB,
8047 std::next(MachineBasicBlock::iterator(MI)), BB->end());
8048 SinkBB->transferSuccessorsAndUpdatePHIs(BB);
8050 BB->addSuccessor(RSBBB);
8051 BB->addSuccessor(SinkBB);
8053 // fall through to SinkMBB
8054 RSBBB->addSuccessor(SinkBB);
8056 // insert a cmp at the end of BB
8057 AddDefaultPred(BuildMI(BB, dl,
8058 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
8059 .addReg(ABSSrcReg).addImm(0));
8061 // insert a bcc with opposite CC to ARMCC::MI at the end of BB
8063 TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)).addMBB(SinkBB)
8064 .addImm(ARMCC::getOppositeCondition(ARMCC::MI)).addReg(ARM::CPSR);
8066 // insert rsbri in RSBBB
8067 // Note: BCC and rsbri will be converted into predicated rsbmi
8068 // by if-conversion pass
8069 BuildMI(*RSBBB, RSBBB->begin(), dl,
8070 TII->get(isThumb2 ? ARM::t2RSBri : ARM::RSBri), NewRsbDstReg)
8071 .addReg(ABSSrcReg, ABSSrcKIll ? RegState::Kill : 0)
8072 .addImm(0).addImm((unsigned)ARMCC::AL).addReg(0).addReg(0);
8074 // insert PHI in SinkBB,
8075 // reuse ABSDstReg to not change uses of ABS instruction
8076 BuildMI(*SinkBB, SinkBB->begin(), dl,
8077 TII->get(ARM::PHI), ABSDstReg)
8078 .addReg(NewRsbDstReg).addMBB(RSBBB)
8079 .addReg(ABSSrcReg).addMBB(BB);
8081 // remove ABS instruction
8082 MI->eraseFromParent();
8084 // return last added BB
8087 case ARM::COPY_STRUCT_BYVAL_I32:
8089 return EmitStructByval(MI, BB);
8090 case ARM::WIN__CHKSTK:
8091 return EmitLowered__chkstk(MI, BB);
8092 case ARM::WIN__DBZCHK:
8093 return EmitLowered__dbzchk(MI, BB);
8097 /// \brief Attaches vregs to MEMCPY that it will use as scratch registers
8098 /// when it is expanded into LDM/STM. This is done as a post-isel lowering
8099 /// instead of as a custom inserter because we need the use list from the SDNode.
8100 static void attachMEMCPYScratchRegs(const ARMSubtarget *Subtarget,
8101 MachineInstr *MI, const SDNode *Node) {
8102 bool isThumb1 = Subtarget->isThumb1Only();
8104 DebugLoc DL = MI->getDebugLoc();
8105 MachineFunction *MF = MI->getParent()->getParent();
8106 MachineRegisterInfo &MRI = MF->getRegInfo();
8107 MachineInstrBuilder MIB(*MF, MI);
8109 // If the new dst/src is unused mark it as dead.
8110 if (!Node->hasAnyUseOfValue(0)) {
8111 MI->getOperand(0).setIsDead(true);
8113 if (!Node->hasAnyUseOfValue(1)) {
8114 MI->getOperand(1).setIsDead(true);
8117 // The MEMCPY both defines and kills the scratch registers.
8118 for (unsigned I = 0; I != MI->getOperand(4).getImm(); ++I) {
8119 unsigned TmpReg = MRI.createVirtualRegister(isThumb1 ? &ARM::tGPRRegClass
8120 : &ARM::GPRRegClass);
8121 MIB.addReg(TmpReg, RegState::Define|RegState::Dead);
8125 void ARMTargetLowering::AdjustInstrPostInstrSelection(MachineInstr *MI,
8126 SDNode *Node) const {
8127 if (MI->getOpcode() == ARM::MEMCPY) {
8128 attachMEMCPYScratchRegs(Subtarget, MI, Node);
8132 const MCInstrDesc *MCID = &MI->getDesc();
8133 // Adjust potentially 's' setting instructions after isel, i.e. ADC, SBC, RSB,
8134 // RSC. Coming out of isel, they have an implicit CPSR def, but the optional
8135 // operand is still set to noreg. If needed, set the optional operand's
8136 // register to CPSR, and remove the redundant implicit def.
8138 // e.g. ADCS (..., CPSR<imp-def>) -> ADC (... opt:CPSR<def>).
8140 // Rename pseudo opcodes.
8141 unsigned NewOpc = convertAddSubFlagsOpcode(MI->getOpcode());
8143 const ARMBaseInstrInfo *TII = Subtarget->getInstrInfo();
8144 MCID = &TII->get(NewOpc);
8146 assert(MCID->getNumOperands() == MI->getDesc().getNumOperands() + 1 &&
8147 "converted opcode should be the same except for cc_out");
8151 // Add the optional cc_out operand
8152 MI->addOperand(MachineOperand::CreateReg(0, /*isDef=*/true));
8154 unsigned ccOutIdx = MCID->getNumOperands() - 1;
8156 // Any ARM instruction that sets the 's' bit should specify an optional
8157 // "cc_out" operand in the last operand position.
8158 if (!MI->hasOptionalDef() || !MCID->OpInfo[ccOutIdx].isOptionalDef()) {
8159 assert(!NewOpc && "Optional cc_out operand required");
8162 // Look for an implicit def of CPSR added by MachineInstr ctor. Remove it
8163 // since we already have an optional CPSR def.
8164 bool definesCPSR = false;
8165 bool deadCPSR = false;
8166 for (unsigned i = MCID->getNumOperands(), e = MI->getNumOperands();
8168 const MachineOperand &MO = MI->getOperand(i);
8169 if (MO.isReg() && MO.isDef() && MO.getReg() == ARM::CPSR) {
8173 MI->RemoveOperand(i);
8178 assert(!NewOpc && "Optional cc_out operand required");
8181 assert(deadCPSR == !Node->hasAnyUseOfValue(1) && "inconsistent dead flag");
8183 assert(!MI->getOperand(ccOutIdx).getReg() &&
8184 "expect uninitialized optional cc_out operand");
8188 // If this instruction was defined with an optional CPSR def and its dag node
8189 // had a live implicit CPSR def, then activate the optional CPSR def.
8190 MachineOperand &MO = MI->getOperand(ccOutIdx);
8191 MO.setReg(ARM::CPSR);
8195 //===----------------------------------------------------------------------===//
8196 // ARM Optimization Hooks
8197 //===----------------------------------------------------------------------===//
8199 // Helper function that checks if N is a null or all ones constant.
8200 static inline bool isZeroOrAllOnes(SDValue N, bool AllOnes) {
8201 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N);
8204 return AllOnes ? C->isAllOnesValue() : C->isNullValue();
8207 // Return true if N is conditionally 0 or all ones.
8208 // Detects these expressions where cc is an i1 value:
8210 // (select cc 0, y) [AllOnes=0]
8211 // (select cc y, 0) [AllOnes=0]
8212 // (zext cc) [AllOnes=0]
8213 // (sext cc) [AllOnes=0/1]
8214 // (select cc -1, y) [AllOnes=1]
8215 // (select cc y, -1) [AllOnes=1]
8217 // Invert is set when N is the null/all ones constant when CC is false.
8218 // OtherOp is set to the alternative value of N.
8219 static bool isConditionalZeroOrAllOnes(SDNode *N, bool AllOnes,
8220 SDValue &CC, bool &Invert,
8222 SelectionDAG &DAG) {
8223 switch (N->getOpcode()) {
8224 default: return false;
8226 CC = N->getOperand(0);
8227 SDValue N1 = N->getOperand(1);
8228 SDValue N2 = N->getOperand(2);
8229 if (isZeroOrAllOnes(N1, AllOnes)) {
8234 if (isZeroOrAllOnes(N2, AllOnes)) {
8241 case ISD::ZERO_EXTEND:
8242 // (zext cc) can never be the all ones value.
8246 case ISD::SIGN_EXTEND: {
8248 EVT VT = N->getValueType(0);
8249 CC = N->getOperand(0);
8250 if (CC.getValueType() != MVT::i1)
8254 // When looking for an AllOnes constant, N is an sext, and the 'other'
8256 OtherOp = DAG.getConstant(0, dl, VT);
8257 else if (N->getOpcode() == ISD::ZERO_EXTEND)
8258 // When looking for a 0 constant, N can be zext or sext.
8259 OtherOp = DAG.getConstant(1, dl, VT);
8261 OtherOp = DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), dl,
8268 // Combine a constant select operand into its use:
8270 // (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
8271 // (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
8272 // (and (select cc, -1, c), x) -> (select cc, x, (and, x, c)) [AllOnes=1]
8273 // (or (select cc, 0, c), x) -> (select cc, x, (or, x, c))
8274 // (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c))
8276 // The transform is rejected if the select doesn't have a constant operand that
8277 // is null, or all ones when AllOnes is set.
8279 // Also recognize sext/zext from i1:
8281 // (add (zext cc), x) -> (select cc (add x, 1), x)
8282 // (add (sext cc), x) -> (select cc (add x, -1), x)
8284 // These transformations eventually create predicated instructions.
8286 // @param N The node to transform.
8287 // @param Slct The N operand that is a select.
8288 // @param OtherOp The other N operand (x above).
8289 // @param DCI Context.
8290 // @param AllOnes Require the select constant to be all ones instead of null.
8291 // @returns The new node, or SDValue() on failure.
8293 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
8294 TargetLowering::DAGCombinerInfo &DCI,
8295 bool AllOnes = false) {
8296 SelectionDAG &DAG = DCI.DAG;
8297 EVT VT = N->getValueType(0);
8298 SDValue NonConstantVal;
8301 if (!isConditionalZeroOrAllOnes(Slct.getNode(), AllOnes, CCOp, SwapSelectOps,
8302 NonConstantVal, DAG))
8305 // Slct is now know to be the desired identity constant when CC is true.
8306 SDValue TrueVal = OtherOp;
8307 SDValue FalseVal = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
8308 OtherOp, NonConstantVal);
8309 // Unless SwapSelectOps says CC should be false.
8311 std::swap(TrueVal, FalseVal);
8313 return DAG.getNode(ISD::SELECT, SDLoc(N), VT,
8314 CCOp, TrueVal, FalseVal);
8317 // Attempt combineSelectAndUse on each operand of a commutative operator N.
8319 SDValue combineSelectAndUseCommutative(SDNode *N, bool AllOnes,
8320 TargetLowering::DAGCombinerInfo &DCI) {
8321 SDValue N0 = N->getOperand(0);
8322 SDValue N1 = N->getOperand(1);
8323 if (N0.getNode()->hasOneUse()) {
8324 SDValue Result = combineSelectAndUse(N, N0, N1, DCI, AllOnes);
8325 if (Result.getNode())
8328 if (N1.getNode()->hasOneUse()) {
8329 SDValue Result = combineSelectAndUse(N, N1, N0, DCI, AllOnes);
8330 if (Result.getNode())
8336 // AddCombineToVPADDL- For pair-wise add on neon, use the vpaddl instruction
8337 // (only after legalization).
8338 static SDValue AddCombineToVPADDL(SDNode *N, SDValue N0, SDValue N1,
8339 TargetLowering::DAGCombinerInfo &DCI,
8340 const ARMSubtarget *Subtarget) {
8342 // Only perform optimization if after legalize, and if NEON is available. We
8343 // also expected both operands to be BUILD_VECTORs.
8344 if (DCI.isBeforeLegalize() || !Subtarget->hasNEON()
8345 || N0.getOpcode() != ISD::BUILD_VECTOR
8346 || N1.getOpcode() != ISD::BUILD_VECTOR)
8349 // Check output type since VPADDL operand elements can only be 8, 16, or 32.
8350 EVT VT = N->getValueType(0);
8351 if (!VT.isInteger() || VT.getVectorElementType() == MVT::i64)
8354 // Check that the vector operands are of the right form.
8355 // N0 and N1 are BUILD_VECTOR nodes with N number of EXTRACT_VECTOR
8356 // operands, where N is the size of the formed vector.
8357 // Each EXTRACT_VECTOR should have the same input vector and odd or even
8358 // index such that we have a pair wise add pattern.
8360 // Grab the vector that all EXTRACT_VECTOR nodes should be referencing.
8361 if (N0->getOperand(0)->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
8363 SDValue Vec = N0->getOperand(0)->getOperand(0);
8364 SDNode *V = Vec.getNode();
8365 unsigned nextIndex = 0;
8367 // For each operands to the ADD which are BUILD_VECTORs,
8368 // check to see if each of their operands are an EXTRACT_VECTOR with
8369 // the same vector and appropriate index.
8370 for (unsigned i = 0, e = N0->getNumOperands(); i != e; ++i) {
8371 if (N0->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT
8372 && N1->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
8374 SDValue ExtVec0 = N0->getOperand(i);
8375 SDValue ExtVec1 = N1->getOperand(i);
8377 // First operand is the vector, verify its the same.
8378 if (V != ExtVec0->getOperand(0).getNode() ||
8379 V != ExtVec1->getOperand(0).getNode())
8382 // Second is the constant, verify its correct.
8383 ConstantSDNode *C0 = dyn_cast<ConstantSDNode>(ExtVec0->getOperand(1));
8384 ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(ExtVec1->getOperand(1));
8386 // For the constant, we want to see all the even or all the odd.
8387 if (!C0 || !C1 || C0->getZExtValue() != nextIndex
8388 || C1->getZExtValue() != nextIndex+1)
8397 // Create VPADDL node.
8398 SelectionDAG &DAG = DCI.DAG;
8399 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8403 // Build operand list.
8404 SmallVector<SDValue, 8> Ops;
8405 Ops.push_back(DAG.getConstant(Intrinsic::arm_neon_vpaddls, dl,
8406 TLI.getPointerTy(DAG.getDataLayout())));
8408 // Input is the vector.
8411 // Get widened type and narrowed type.
8413 unsigned numElem = VT.getVectorNumElements();
8415 EVT inputLaneType = Vec.getValueType().getVectorElementType();
8416 switch (inputLaneType.getSimpleVT().SimpleTy) {
8417 case MVT::i8: widenType = MVT::getVectorVT(MVT::i16, numElem); break;
8418 case MVT::i16: widenType = MVT::getVectorVT(MVT::i32, numElem); break;
8419 case MVT::i32: widenType = MVT::getVectorVT(MVT::i64, numElem); break;
8421 llvm_unreachable("Invalid vector element type for padd optimization.");
8424 SDValue tmp = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, widenType, Ops);
8425 unsigned ExtOp = VT.bitsGT(tmp.getValueType()) ? ISD::ANY_EXTEND : ISD::TRUNCATE;
8426 return DAG.getNode(ExtOp, dl, VT, tmp);
8429 static SDValue findMUL_LOHI(SDValue V) {
8430 if (V->getOpcode() == ISD::UMUL_LOHI ||
8431 V->getOpcode() == ISD::SMUL_LOHI)
8436 static SDValue AddCombineTo64bitMLAL(SDNode *AddcNode,
8437 TargetLowering::DAGCombinerInfo &DCI,
8438 const ARMSubtarget *Subtarget) {
8440 if (Subtarget->isThumb1Only()) return SDValue();
8442 // Only perform the checks after legalize when the pattern is available.
8443 if (DCI.isBeforeLegalize()) return SDValue();
8445 // Look for multiply add opportunities.
8446 // The pattern is a ISD::UMUL_LOHI followed by two add nodes, where
8447 // each add nodes consumes a value from ISD::UMUL_LOHI and there is
8448 // a glue link from the first add to the second add.
8449 // If we find this pattern, we can replace the U/SMUL_LOHI, ADDC, and ADDE by
8450 // a S/UMLAL instruction.
8453 // / \ [no multiline comment]
8459 assert(AddcNode->getOpcode() == ISD::ADDC && "Expect an ADDC");
8460 SDValue AddcOp0 = AddcNode->getOperand(0);
8461 SDValue AddcOp1 = AddcNode->getOperand(1);
8463 // Check if the two operands are from the same mul_lohi node.
8464 if (AddcOp0.getNode() == AddcOp1.getNode())
8467 assert(AddcNode->getNumValues() == 2 &&
8468 AddcNode->getValueType(0) == MVT::i32 &&
8469 "Expect ADDC with two result values. First: i32");
8471 // Check that we have a glued ADDC node.
8472 if (AddcNode->getValueType(1) != MVT::Glue)
8475 // Check that the ADDC adds the low result of the S/UMUL_LOHI.
8476 if (AddcOp0->getOpcode() != ISD::UMUL_LOHI &&
8477 AddcOp0->getOpcode() != ISD::SMUL_LOHI &&
8478 AddcOp1->getOpcode() != ISD::UMUL_LOHI &&
8479 AddcOp1->getOpcode() != ISD::SMUL_LOHI)
8482 // Look for the glued ADDE.
8483 SDNode* AddeNode = AddcNode->getGluedUser();
8487 // Make sure it is really an ADDE.
8488 if (AddeNode->getOpcode() != ISD::ADDE)
8491 assert(AddeNode->getNumOperands() == 3 &&
8492 AddeNode->getOperand(2).getValueType() == MVT::Glue &&
8493 "ADDE node has the wrong inputs");
8495 // Check for the triangle shape.
8496 SDValue AddeOp0 = AddeNode->getOperand(0);
8497 SDValue AddeOp1 = AddeNode->getOperand(1);
8499 // Make sure that the ADDE operands are not coming from the same node.
8500 if (AddeOp0.getNode() == AddeOp1.getNode())
8503 // Find the MUL_LOHI node walking up ADDE's operands.
8504 bool IsLeftOperandMUL = false;
8505 SDValue MULOp = findMUL_LOHI(AddeOp0);
8506 if (MULOp == SDValue())
8507 MULOp = findMUL_LOHI(AddeOp1);
8509 IsLeftOperandMUL = true;
8510 if (MULOp == SDValue())
8513 // Figure out the right opcode.
8514 unsigned Opc = MULOp->getOpcode();
8515 unsigned FinalOpc = (Opc == ISD::SMUL_LOHI) ? ARMISD::SMLAL : ARMISD::UMLAL;
8517 // Figure out the high and low input values to the MLAL node.
8518 SDValue* HiAdd = nullptr;
8519 SDValue* LoMul = nullptr;
8520 SDValue* LowAdd = nullptr;
8522 // Ensure that ADDE is from high result of ISD::SMUL_LOHI.
8523 if ((AddeOp0 != MULOp.getValue(1)) && (AddeOp1 != MULOp.getValue(1)))
8526 if (IsLeftOperandMUL)
8532 // Ensure that LoMul and LowAdd are taken from correct ISD::SMUL_LOHI node
8533 // whose low result is fed to the ADDC we are checking.
8535 if (AddcOp0 == MULOp.getValue(0)) {
8539 if (AddcOp1 == MULOp.getValue(0)) {
8547 // Create the merged node.
8548 SelectionDAG &DAG = DCI.DAG;
8550 // Build operand list.
8551 SmallVector<SDValue, 8> Ops;
8552 Ops.push_back(LoMul->getOperand(0));
8553 Ops.push_back(LoMul->getOperand(1));
8554 Ops.push_back(*LowAdd);
8555 Ops.push_back(*HiAdd);
8557 SDValue MLALNode = DAG.getNode(FinalOpc, SDLoc(AddcNode),
8558 DAG.getVTList(MVT::i32, MVT::i32), Ops);
8560 // Replace the ADDs' nodes uses by the MLA node's values.
8561 SDValue HiMLALResult(MLALNode.getNode(), 1);
8562 DAG.ReplaceAllUsesOfValueWith(SDValue(AddeNode, 0), HiMLALResult);
8564 SDValue LoMLALResult(MLALNode.getNode(), 0);
8565 DAG.ReplaceAllUsesOfValueWith(SDValue(AddcNode, 0), LoMLALResult);
8567 // Return original node to notify the driver to stop replacing.
8568 SDValue resNode(AddcNode, 0);
8572 /// PerformADDCCombine - Target-specific dag combine transform from
8573 /// ISD::ADDC, ISD::ADDE, and ISD::MUL_LOHI to MLAL.
8574 static SDValue PerformADDCCombine(SDNode *N,
8575 TargetLowering::DAGCombinerInfo &DCI,
8576 const ARMSubtarget *Subtarget) {
8578 return AddCombineTo64bitMLAL(N, DCI, Subtarget);
8582 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
8583 /// operands N0 and N1. This is a helper for PerformADDCombine that is
8584 /// called with the default operands, and if that fails, with commuted
8586 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
8587 TargetLowering::DAGCombinerInfo &DCI,
8588 const ARMSubtarget *Subtarget){
8590 // Attempt to create vpaddl for this add.
8591 SDValue Result = AddCombineToVPADDL(N, N0, N1, DCI, Subtarget);
8592 if (Result.getNode())
8595 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
8596 if (N0.getNode()->hasOneUse()) {
8597 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
8598 if (Result.getNode()) return Result;
8603 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
8605 static SDValue PerformADDCombine(SDNode *N,
8606 TargetLowering::DAGCombinerInfo &DCI,
8607 const ARMSubtarget *Subtarget) {
8608 SDValue N0 = N->getOperand(0);
8609 SDValue N1 = N->getOperand(1);
8611 // First try with the default operand order.
8612 SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI, Subtarget);
8613 if (Result.getNode())
8616 // If that didn't work, try again with the operands commuted.
8617 return PerformADDCombineWithOperands(N, N1, N0, DCI, Subtarget);
8620 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
8622 static SDValue PerformSUBCombine(SDNode *N,
8623 TargetLowering::DAGCombinerInfo &DCI) {
8624 SDValue N0 = N->getOperand(0);
8625 SDValue N1 = N->getOperand(1);
8627 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
8628 if (N1.getNode()->hasOneUse()) {
8629 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
8630 if (Result.getNode()) return Result;
8636 /// PerformVMULCombine
8637 /// Distribute (A + B) * C to (A * C) + (B * C) to take advantage of the
8638 /// special multiplier accumulator forwarding.
8644 // However, for (A + B) * (A + B),
8651 static SDValue PerformVMULCombine(SDNode *N,
8652 TargetLowering::DAGCombinerInfo &DCI,
8653 const ARMSubtarget *Subtarget) {
8654 if (!Subtarget->hasVMLxForwarding())
8657 SelectionDAG &DAG = DCI.DAG;
8658 SDValue N0 = N->getOperand(0);
8659 SDValue N1 = N->getOperand(1);
8660 unsigned Opcode = N0.getOpcode();
8661 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
8662 Opcode != ISD::FADD && Opcode != ISD::FSUB) {
8663 Opcode = N1.getOpcode();
8664 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
8665 Opcode != ISD::FADD && Opcode != ISD::FSUB)
8673 EVT VT = N->getValueType(0);
8675 SDValue N00 = N0->getOperand(0);
8676 SDValue N01 = N0->getOperand(1);
8677 return DAG.getNode(Opcode, DL, VT,
8678 DAG.getNode(ISD::MUL, DL, VT, N00, N1),
8679 DAG.getNode(ISD::MUL, DL, VT, N01, N1));
8682 static SDValue PerformMULCombine(SDNode *N,
8683 TargetLowering::DAGCombinerInfo &DCI,
8684 const ARMSubtarget *Subtarget) {
8685 SelectionDAG &DAG = DCI.DAG;
8687 if (Subtarget->isThumb1Only())
8690 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
8693 EVT VT = N->getValueType(0);
8694 if (VT.is64BitVector() || VT.is128BitVector())
8695 return PerformVMULCombine(N, DCI, Subtarget);
8699 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
8703 int64_t MulAmt = C->getSExtValue();
8704 unsigned ShiftAmt = countTrailingZeros<uint64_t>(MulAmt);
8706 ShiftAmt = ShiftAmt & (32 - 1);
8707 SDValue V = N->getOperand(0);
8711 MulAmt >>= ShiftAmt;
8714 if (isPowerOf2_32(MulAmt - 1)) {
8715 // (mul x, 2^N + 1) => (add (shl x, N), x)
8716 Res = DAG.getNode(ISD::ADD, DL, VT,
8718 DAG.getNode(ISD::SHL, DL, VT,
8720 DAG.getConstant(Log2_32(MulAmt - 1), DL,
8722 } else if (isPowerOf2_32(MulAmt + 1)) {
8723 // (mul x, 2^N - 1) => (sub (shl x, N), x)
8724 Res = DAG.getNode(ISD::SUB, DL, VT,
8725 DAG.getNode(ISD::SHL, DL, VT,
8727 DAG.getConstant(Log2_32(MulAmt + 1), DL,
8733 uint64_t MulAmtAbs = -MulAmt;
8734 if (isPowerOf2_32(MulAmtAbs + 1)) {
8735 // (mul x, -(2^N - 1)) => (sub x, (shl x, N))
8736 Res = DAG.getNode(ISD::SUB, DL, VT,
8738 DAG.getNode(ISD::SHL, DL, VT,
8740 DAG.getConstant(Log2_32(MulAmtAbs + 1), DL,
8742 } else if (isPowerOf2_32(MulAmtAbs - 1)) {
8743 // (mul x, -(2^N + 1)) => - (add (shl x, N), x)
8744 Res = DAG.getNode(ISD::ADD, DL, VT,
8746 DAG.getNode(ISD::SHL, DL, VT,
8748 DAG.getConstant(Log2_32(MulAmtAbs - 1), DL,
8750 Res = DAG.getNode(ISD::SUB, DL, VT,
8751 DAG.getConstant(0, DL, MVT::i32), Res);
8758 Res = DAG.getNode(ISD::SHL, DL, VT,
8759 Res, DAG.getConstant(ShiftAmt, DL, MVT::i32));
8761 // Do not add new nodes to DAG combiner worklist.
8762 DCI.CombineTo(N, Res, false);
8766 static SDValue PerformANDCombine(SDNode *N,
8767 TargetLowering::DAGCombinerInfo &DCI,
8768 const ARMSubtarget *Subtarget) {
8770 // Attempt to use immediate-form VBIC
8771 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
8773 EVT VT = N->getValueType(0);
8774 SelectionDAG &DAG = DCI.DAG;
8776 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8779 APInt SplatBits, SplatUndef;
8780 unsigned SplatBitSize;
8783 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
8784 if (SplatBitSize <= 64) {
8786 SDValue Val = isNEONModifiedImm((~SplatBits).getZExtValue(),
8787 SplatUndef.getZExtValue(), SplatBitSize,
8788 DAG, dl, VbicVT, VT.is128BitVector(),
8790 if (Val.getNode()) {
8792 DAG.getNode(ISD::BITCAST, dl, VbicVT, N->getOperand(0));
8793 SDValue Vbic = DAG.getNode(ARMISD::VBICIMM, dl, VbicVT, Input, Val);
8794 return DAG.getNode(ISD::BITCAST, dl, VT, Vbic);
8799 if (!Subtarget->isThumb1Only()) {
8800 // fold (and (select cc, -1, c), x) -> (select cc, x, (and, x, c))
8801 SDValue Result = combineSelectAndUseCommutative(N, true, DCI);
8802 if (Result.getNode())
8809 /// PerformORCombine - Target-specific dag combine xforms for ISD::OR
8810 static SDValue PerformORCombine(SDNode *N,
8811 TargetLowering::DAGCombinerInfo &DCI,
8812 const ARMSubtarget *Subtarget) {
8813 // Attempt to use immediate-form VORR
8814 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
8816 EVT VT = N->getValueType(0);
8817 SelectionDAG &DAG = DCI.DAG;
8819 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8822 APInt SplatBits, SplatUndef;
8823 unsigned SplatBitSize;
8825 if (BVN && Subtarget->hasNEON() &&
8826 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
8827 if (SplatBitSize <= 64) {
8829 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
8830 SplatUndef.getZExtValue(), SplatBitSize,
8831 DAG, dl, VorrVT, VT.is128BitVector(),
8833 if (Val.getNode()) {
8835 DAG.getNode(ISD::BITCAST, dl, VorrVT, N->getOperand(0));
8836 SDValue Vorr = DAG.getNode(ARMISD::VORRIMM, dl, VorrVT, Input, Val);
8837 return DAG.getNode(ISD::BITCAST, dl, VT, Vorr);
8842 if (!Subtarget->isThumb1Only()) {
8843 // fold (or (select cc, 0, c), x) -> (select cc, x, (or, x, c))
8844 SDValue Result = combineSelectAndUseCommutative(N, false, DCI);
8845 if (Result.getNode())
8849 // The code below optimizes (or (and X, Y), Z).
8850 // The AND operand needs to have a single user to make these optimizations
8852 SDValue N0 = N->getOperand(0);
8853 if (N0.getOpcode() != ISD::AND || !N0.hasOneUse())
8855 SDValue N1 = N->getOperand(1);
8857 // (or (and B, A), (and C, ~A)) => (VBSL A, B, C) when A is a constant.
8858 if (Subtarget->hasNEON() && N1.getOpcode() == ISD::AND && VT.isVector() &&
8859 DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
8861 unsigned SplatBitSize;
8864 APInt SplatBits0, SplatBits1;
8865 BuildVectorSDNode *BVN0 = dyn_cast<BuildVectorSDNode>(N0->getOperand(1));
8866 BuildVectorSDNode *BVN1 = dyn_cast<BuildVectorSDNode>(N1->getOperand(1));
8867 // Ensure that the second operand of both ands are constants
8868 if (BVN0 && BVN0->isConstantSplat(SplatBits0, SplatUndef, SplatBitSize,
8869 HasAnyUndefs) && !HasAnyUndefs) {
8870 if (BVN1 && BVN1->isConstantSplat(SplatBits1, SplatUndef, SplatBitSize,
8871 HasAnyUndefs) && !HasAnyUndefs) {
8872 // Ensure that the bit width of the constants are the same and that
8873 // the splat arguments are logical inverses as per the pattern we
8874 // are trying to simplify.
8875 if (SplatBits0.getBitWidth() == SplatBits1.getBitWidth() &&
8876 SplatBits0 == ~SplatBits1) {
8877 // Canonicalize the vector type to make instruction selection
8879 EVT CanonicalVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
8880 SDValue Result = DAG.getNode(ARMISD::VBSL, dl, CanonicalVT,
8884 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
8890 // Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when
8893 // BFI is only available on V6T2+
8894 if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops())
8898 // 1) or (and A, mask), val => ARMbfi A, val, mask
8899 // iff (val & mask) == val
8901 // 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
8902 // 2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2)
8903 // && mask == ~mask2
8904 // 2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2)
8905 // && ~mask == mask2
8906 // (i.e., copy a bitfield value into another bitfield of the same width)
8911 SDValue N00 = N0.getOperand(0);
8913 // The value and the mask need to be constants so we can verify this is
8914 // actually a bitfield set. If the mask is 0xffff, we can do better
8915 // via a movt instruction, so don't use BFI in that case.
8916 SDValue MaskOp = N0.getOperand(1);
8917 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(MaskOp);
8920 unsigned Mask = MaskC->getZExtValue();
8924 // Case (1): or (and A, mask), val => ARMbfi A, val, mask
8925 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
8927 unsigned Val = N1C->getZExtValue();
8928 if ((Val & ~Mask) != Val)
8931 if (ARM::isBitFieldInvertedMask(Mask)) {
8932 Val >>= countTrailingZeros(~Mask);
8934 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00,
8935 DAG.getConstant(Val, DL, MVT::i32),
8936 DAG.getConstant(Mask, DL, MVT::i32));
8938 // Do not add new nodes to DAG combiner worklist.
8939 DCI.CombineTo(N, Res, false);
8942 } else if (N1.getOpcode() == ISD::AND) {
8943 // case (2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
8944 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
8947 unsigned Mask2 = N11C->getZExtValue();
8949 // Mask and ~Mask2 (or reverse) must be equivalent for the BFI pattern
8951 if (ARM::isBitFieldInvertedMask(Mask) &&
8953 // The pack halfword instruction works better for masks that fit it,
8954 // so use that when it's available.
8955 if (Subtarget->hasT2ExtractPack() &&
8956 (Mask == 0xffff || Mask == 0xffff0000))
8959 unsigned amt = countTrailingZeros(Mask2);
8960 Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0),
8961 DAG.getConstant(amt, DL, MVT::i32));
8962 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00, Res,
8963 DAG.getConstant(Mask, DL, MVT::i32));
8964 // Do not add new nodes to DAG combiner worklist.
8965 DCI.CombineTo(N, Res, false);
8967 } else if (ARM::isBitFieldInvertedMask(~Mask) &&
8969 // The pack halfword instruction works better for masks that fit it,
8970 // so use that when it's available.
8971 if (Subtarget->hasT2ExtractPack() &&
8972 (Mask2 == 0xffff || Mask2 == 0xffff0000))
8975 unsigned lsb = countTrailingZeros(Mask);
8976 Res = DAG.getNode(ISD::SRL, DL, VT, N00,
8977 DAG.getConstant(lsb, DL, MVT::i32));
8978 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res,
8979 DAG.getConstant(Mask2, DL, MVT::i32));
8980 // Do not add new nodes to DAG combiner worklist.
8981 DCI.CombineTo(N, Res, false);
8986 if (DAG.MaskedValueIsZero(N1, MaskC->getAPIntValue()) &&
8987 N00.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N00.getOperand(1)) &&
8988 ARM::isBitFieldInvertedMask(~Mask)) {
8989 // Case (3): or (and (shl A, #shamt), mask), B => ARMbfi B, A, ~mask
8990 // where lsb(mask) == #shamt and masked bits of B are known zero.
8991 SDValue ShAmt = N00.getOperand(1);
8992 unsigned ShAmtC = cast<ConstantSDNode>(ShAmt)->getZExtValue();
8993 unsigned LSB = countTrailingZeros(Mask);
8997 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1, N00.getOperand(0),
8998 DAG.getConstant(~Mask, DL, MVT::i32));
9000 // Do not add new nodes to DAG combiner worklist.
9001 DCI.CombineTo(N, Res, false);
9007 static SDValue PerformXORCombine(SDNode *N,
9008 TargetLowering::DAGCombinerInfo &DCI,
9009 const ARMSubtarget *Subtarget) {
9010 EVT VT = N->getValueType(0);
9011 SelectionDAG &DAG = DCI.DAG;
9013 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
9016 if (!Subtarget->isThumb1Only()) {
9017 // fold (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c))
9018 SDValue Result = combineSelectAndUseCommutative(N, false, DCI);
9019 if (Result.getNode())
9026 // ParseBFI - given a BFI instruction in N, extract the "from" value (Rn) and return it,
9027 // and fill in FromMask and ToMask with (consecutive) bits in "from" to be extracted and
9028 // their position in "to" (Rd).
9029 static SDValue ParseBFI(SDNode *N, APInt &ToMask, APInt &FromMask) {
9030 assert(N->getOpcode() == ARMISD::BFI);
9032 SDValue From = N->getOperand(1);
9033 ToMask = ~cast<ConstantSDNode>(N->getOperand(2))->getAPIntValue();
9034 FromMask = APInt::getLowBitsSet(ToMask.getBitWidth(), ToMask.countPopulation());
9036 // If the Base came from a SHR #C, we can deduce that it is really testing bit
9037 // #C in the base of the SHR.
9038 if (From->getOpcode() == ISD::SRL &&
9039 isa<ConstantSDNode>(From->getOperand(1))) {
9040 APInt Shift = cast<ConstantSDNode>(From->getOperand(1))->getAPIntValue();
9041 assert(Shift.getLimitedValue() < 32 && "Shift too large!");
9042 FromMask <<= Shift.getLimitedValue(31);
9043 From = From->getOperand(0);
9049 // If A and B contain one contiguous set of bits, does A | B == A . B?
9051 // Neither A nor B must be zero.
9052 static bool BitsProperlyConcatenate(const APInt &A, const APInt &B) {
9053 unsigned LastActiveBitInA = A.countTrailingZeros();
9054 unsigned FirstActiveBitInB = B.getBitWidth() - B.countLeadingZeros() - 1;
9055 return LastActiveBitInA - 1 == FirstActiveBitInB;
9058 static SDValue FindBFIToCombineWith(SDNode *N) {
9059 // We have a BFI in N. Follow a possible chain of BFIs and find a BFI it can combine with,
9061 APInt ToMask, FromMask;
9062 SDValue From = ParseBFI(N, ToMask, FromMask);
9063 SDValue To = N->getOperand(0);
9065 // Now check for a compatible BFI to merge with. We can pass through BFIs that
9066 // aren't compatible, but not if they set the same bit in their destination as
9067 // we do (or that of any BFI we're going to combine with).
9069 APInt CombinedToMask = ToMask;
9070 while (V.getOpcode() == ARMISD::BFI) {
9071 APInt NewToMask, NewFromMask;
9072 SDValue NewFrom = ParseBFI(V.getNode(), NewToMask, NewFromMask);
9073 if (NewFrom != From) {
9074 // This BFI has a different base. Keep going.
9075 CombinedToMask |= NewToMask;
9076 V = V.getOperand(0);
9080 // Do the written bits conflict with any we've seen so far?
9081 if ((NewToMask & CombinedToMask).getBoolValue())
9082 // Conflicting bits - bail out because going further is unsafe.
9085 // Are the new bits contiguous when combined with the old bits?
9086 if (BitsProperlyConcatenate(ToMask, NewToMask) &&
9087 BitsProperlyConcatenate(FromMask, NewFromMask))
9089 if (BitsProperlyConcatenate(NewToMask, ToMask) &&
9090 BitsProperlyConcatenate(NewFromMask, FromMask))
9093 // We've seen a write to some bits, so track it.
9094 CombinedToMask |= NewToMask;
9096 V = V.getOperand(0);
9102 static SDValue PerformBFICombine(SDNode *N,
9103 TargetLowering::DAGCombinerInfo &DCI) {
9104 SDValue N1 = N->getOperand(1);
9105 if (N1.getOpcode() == ISD::AND) {
9106 // (bfi A, (and B, Mask1), Mask2) -> (bfi A, B, Mask2) iff
9107 // the bits being cleared by the AND are not demanded by the BFI.
9108 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
9111 unsigned InvMask = cast<ConstantSDNode>(N->getOperand(2))->getZExtValue();
9112 unsigned LSB = countTrailingZeros(~InvMask);
9113 unsigned Width = (32 - countLeadingZeros(~InvMask)) - LSB;
9115 static_cast<unsigned>(std::numeric_limits<unsigned>::digits) &&
9116 "undefined behavior");
9117 unsigned Mask = (1u << Width) - 1;
9118 unsigned Mask2 = N11C->getZExtValue();
9119 if ((Mask & (~Mask2)) == 0)
9120 return DCI.DAG.getNode(ARMISD::BFI, SDLoc(N), N->getValueType(0),
9121 N->getOperand(0), N1.getOperand(0),
9123 } else if (N->getOperand(0).getOpcode() == ARMISD::BFI) {
9124 // We have a BFI of a BFI. Walk up the BFI chain to see how long it goes.
9125 // Keep track of any consecutive bits set that all come from the same base
9126 // value. We can combine these together into a single BFI.
9127 SDValue CombineBFI = FindBFIToCombineWith(N);
9128 if (CombineBFI == SDValue())
9131 // We've found a BFI.
9132 APInt ToMask1, FromMask1;
9133 SDValue From1 = ParseBFI(N, ToMask1, FromMask1);
9135 APInt ToMask2, FromMask2;
9136 SDValue From2 = ParseBFI(CombineBFI.getNode(), ToMask2, FromMask2);
9137 assert(From1 == From2);
9140 // First, unlink CombineBFI.
9141 DCI.DAG.ReplaceAllUsesWith(CombineBFI, CombineBFI.getOperand(0));
9142 // Then create a new BFI, combining the two together.
9143 APInt NewFromMask = FromMask1 | FromMask2;
9144 APInt NewToMask = ToMask1 | ToMask2;
9146 EVT VT = N->getValueType(0);
9149 if (NewFromMask[0] == 0)
9150 From1 = DCI.DAG.getNode(
9151 ISD::SRL, dl, VT, From1,
9152 DCI.DAG.getConstant(NewFromMask.countTrailingZeros(), dl, VT));
9153 return DCI.DAG.getNode(ARMISD::BFI, dl, VT, N->getOperand(0), From1,
9154 DCI.DAG.getConstant(~NewToMask, dl, VT));
9159 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
9160 /// ARMISD::VMOVRRD.
9161 static SDValue PerformVMOVRRDCombine(SDNode *N,
9162 TargetLowering::DAGCombinerInfo &DCI,
9163 const ARMSubtarget *Subtarget) {
9164 // vmovrrd(vmovdrr x, y) -> x,y
9165 SDValue InDouble = N->getOperand(0);
9166 if (InDouble.getOpcode() == ARMISD::VMOVDRR && !Subtarget->isFPOnlySP())
9167 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
9169 // vmovrrd(load f64) -> (load i32), (load i32)
9170 SDNode *InNode = InDouble.getNode();
9171 if (ISD::isNormalLoad(InNode) && InNode->hasOneUse() &&
9172 InNode->getValueType(0) == MVT::f64 &&
9173 InNode->getOperand(1).getOpcode() == ISD::FrameIndex &&
9174 !cast<LoadSDNode>(InNode)->isVolatile()) {
9175 // TODO: Should this be done for non-FrameIndex operands?
9176 LoadSDNode *LD = cast<LoadSDNode>(InNode);
9178 SelectionDAG &DAG = DCI.DAG;
9180 SDValue BasePtr = LD->getBasePtr();
9181 SDValue NewLD1 = DAG.getLoad(MVT::i32, DL, LD->getChain(), BasePtr,
9182 LD->getPointerInfo(), LD->isVolatile(),
9183 LD->isNonTemporal(), LD->isInvariant(),
9184 LD->getAlignment());
9186 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
9187 DAG.getConstant(4, DL, MVT::i32));
9188 SDValue NewLD2 = DAG.getLoad(MVT::i32, DL, NewLD1.getValue(1), OffsetPtr,
9189 LD->getPointerInfo(), LD->isVolatile(),
9190 LD->isNonTemporal(), LD->isInvariant(),
9191 std::min(4U, LD->getAlignment() / 2));
9193 DAG.ReplaceAllUsesOfValueWith(SDValue(LD, 1), NewLD2.getValue(1));
9194 if (DCI.DAG.getDataLayout().isBigEndian())
9195 std::swap (NewLD1, NewLD2);
9196 SDValue Result = DCI.CombineTo(N, NewLD1, NewLD2);
9203 /// PerformVMOVDRRCombine - Target-specific dag combine xforms for
9204 /// ARMISD::VMOVDRR. This is also used for BUILD_VECTORs with 2 operands.
9205 static SDValue PerformVMOVDRRCombine(SDNode *N, SelectionDAG &DAG) {
9206 // N=vmovrrd(X); vmovdrr(N:0, N:1) -> bit_convert(X)
9207 SDValue Op0 = N->getOperand(0);
9208 SDValue Op1 = N->getOperand(1);
9209 if (Op0.getOpcode() == ISD::BITCAST)
9210 Op0 = Op0.getOperand(0);
9211 if (Op1.getOpcode() == ISD::BITCAST)
9212 Op1 = Op1.getOperand(0);
9213 if (Op0.getOpcode() == ARMISD::VMOVRRD &&
9214 Op0.getNode() == Op1.getNode() &&
9215 Op0.getResNo() == 0 && Op1.getResNo() == 1)
9216 return DAG.getNode(ISD::BITCAST, SDLoc(N),
9217 N->getValueType(0), Op0.getOperand(0));
9221 /// hasNormalLoadOperand - Check if any of the operands of a BUILD_VECTOR node
9222 /// are normal, non-volatile loads. If so, it is profitable to bitcast an
9223 /// i64 vector to have f64 elements, since the value can then be loaded
9224 /// directly into a VFP register.
9225 static bool hasNormalLoadOperand(SDNode *N) {
9226 unsigned NumElts = N->getValueType(0).getVectorNumElements();
9227 for (unsigned i = 0; i < NumElts; ++i) {
9228 SDNode *Elt = N->getOperand(i).getNode();
9229 if (ISD::isNormalLoad(Elt) && !cast<LoadSDNode>(Elt)->isVolatile())
9235 /// PerformBUILD_VECTORCombine - Target-specific dag combine xforms for
9236 /// ISD::BUILD_VECTOR.
9237 static SDValue PerformBUILD_VECTORCombine(SDNode *N,
9238 TargetLowering::DAGCombinerInfo &DCI,
9239 const ARMSubtarget *Subtarget) {
9240 // build_vector(N=ARMISD::VMOVRRD(X), N:1) -> bit_convert(X):
9241 // VMOVRRD is introduced when legalizing i64 types. It forces the i64 value
9242 // into a pair of GPRs, which is fine when the value is used as a scalar,
9243 // but if the i64 value is converted to a vector, we need to undo the VMOVRRD.
9244 SelectionDAG &DAG = DCI.DAG;
9245 if (N->getNumOperands() == 2) {
9246 SDValue RV = PerformVMOVDRRCombine(N, DAG);
9251 // Load i64 elements as f64 values so that type legalization does not split
9252 // them up into i32 values.
9253 EVT VT = N->getValueType(0);
9254 if (VT.getVectorElementType() != MVT::i64 || !hasNormalLoadOperand(N))
9257 SmallVector<SDValue, 8> Ops;
9258 unsigned NumElts = VT.getVectorNumElements();
9259 for (unsigned i = 0; i < NumElts; ++i) {
9260 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(i));
9262 // Make the DAGCombiner fold the bitcast.
9263 DCI.AddToWorklist(V.getNode());
9265 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, NumElts);
9266 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, FloatVT, Ops);
9267 return DAG.getNode(ISD::BITCAST, dl, VT, BV);
9270 /// \brief Target-specific dag combine xforms for ARMISD::BUILD_VECTOR.
9272 PerformARMBUILD_VECTORCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
9273 // ARMISD::BUILD_VECTOR is introduced when legalizing ISD::BUILD_VECTOR.
9274 // At that time, we may have inserted bitcasts from integer to float.
9275 // If these bitcasts have survived DAGCombine, change the lowering of this
9276 // BUILD_VECTOR in something more vector friendly, i.e., that does not
9277 // force to use floating point types.
9279 // Make sure we can change the type of the vector.
9280 // This is possible iff:
9281 // 1. The vector is only used in a bitcast to a integer type. I.e.,
9282 // 1.1. Vector is used only once.
9283 // 1.2. Use is a bit convert to an integer type.
9284 // 2. The size of its operands are 32-bits (64-bits are not legal).
9285 EVT VT = N->getValueType(0);
9286 EVT EltVT = VT.getVectorElementType();
9288 // Check 1.1. and 2.
9289 if (EltVT.getSizeInBits() != 32 || !N->hasOneUse())
9292 // By construction, the input type must be float.
9293 assert(EltVT == MVT::f32 && "Unexpected type!");
9296 SDNode *Use = *N->use_begin();
9297 if (Use->getOpcode() != ISD::BITCAST ||
9298 Use->getValueType(0).isFloatingPoint())
9301 // Check profitability.
9302 // Model is, if more than half of the relevant operands are bitcast from
9303 // i32, turn the build_vector into a sequence of insert_vector_elt.
9304 // Relevant operands are everything that is not statically
9305 // (i.e., at compile time) bitcasted.
9306 unsigned NumOfBitCastedElts = 0;
9307 unsigned NumElts = VT.getVectorNumElements();
9308 unsigned NumOfRelevantElts = NumElts;
9309 for (unsigned Idx = 0; Idx < NumElts; ++Idx) {
9310 SDValue Elt = N->getOperand(Idx);
9311 if (Elt->getOpcode() == ISD::BITCAST) {
9312 // Assume only bit cast to i32 will go away.
9313 if (Elt->getOperand(0).getValueType() == MVT::i32)
9314 ++NumOfBitCastedElts;
9315 } else if (Elt.getOpcode() == ISD::UNDEF || isa<ConstantSDNode>(Elt))
9316 // Constants are statically casted, thus do not count them as
9317 // relevant operands.
9318 --NumOfRelevantElts;
9321 // Check if more than half of the elements require a non-free bitcast.
9322 if (NumOfBitCastedElts <= NumOfRelevantElts / 2)
9325 SelectionDAG &DAG = DCI.DAG;
9326 // Create the new vector type.
9327 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
9328 // Check if the type is legal.
9329 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9330 if (!TLI.isTypeLegal(VecVT))
9334 // ARMISD::BUILD_VECTOR E1, E2, ..., EN.
9335 // => BITCAST INSERT_VECTOR_ELT
9336 // (INSERT_VECTOR_ELT (...), (BITCAST EN-1), N-1),
9338 SDValue Vec = DAG.getUNDEF(VecVT);
9340 for (unsigned Idx = 0 ; Idx < NumElts; ++Idx) {
9341 SDValue V = N->getOperand(Idx);
9342 if (V.getOpcode() == ISD::UNDEF)
9344 if (V.getOpcode() == ISD::BITCAST &&
9345 V->getOperand(0).getValueType() == MVT::i32)
9346 // Fold obvious case.
9347 V = V.getOperand(0);
9349 V = DAG.getNode(ISD::BITCAST, SDLoc(V), MVT::i32, V);
9350 // Make the DAGCombiner fold the bitcasts.
9351 DCI.AddToWorklist(V.getNode());
9353 SDValue LaneIdx = DAG.getConstant(Idx, dl, MVT::i32);
9354 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VecVT, Vec, V, LaneIdx);
9356 Vec = DAG.getNode(ISD::BITCAST, dl, VT, Vec);
9357 // Make the DAGCombiner fold the bitcasts.
9358 DCI.AddToWorklist(Vec.getNode());
9362 /// PerformInsertEltCombine - Target-specific dag combine xforms for
9363 /// ISD::INSERT_VECTOR_ELT.
9364 static SDValue PerformInsertEltCombine(SDNode *N,
9365 TargetLowering::DAGCombinerInfo &DCI) {
9366 // Bitcast an i64 load inserted into a vector to f64.
9367 // Otherwise, the i64 value will be legalized to a pair of i32 values.
9368 EVT VT = N->getValueType(0);
9369 SDNode *Elt = N->getOperand(1).getNode();
9370 if (VT.getVectorElementType() != MVT::i64 ||
9371 !ISD::isNormalLoad(Elt) || cast<LoadSDNode>(Elt)->isVolatile())
9374 SelectionDAG &DAG = DCI.DAG;
9376 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
9377 VT.getVectorNumElements());
9378 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, N->getOperand(0));
9379 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(1));
9380 // Make the DAGCombiner fold the bitcasts.
9381 DCI.AddToWorklist(Vec.getNode());
9382 DCI.AddToWorklist(V.getNode());
9383 SDValue InsElt = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, FloatVT,
9384 Vec, V, N->getOperand(2));
9385 return DAG.getNode(ISD::BITCAST, dl, VT, InsElt);
9388 /// PerformVECTOR_SHUFFLECombine - Target-specific dag combine xforms for
9389 /// ISD::VECTOR_SHUFFLE.
9390 static SDValue PerformVECTOR_SHUFFLECombine(SDNode *N, SelectionDAG &DAG) {
9391 // The LLVM shufflevector instruction does not require the shuffle mask
9392 // length to match the operand vector length, but ISD::VECTOR_SHUFFLE does
9393 // have that requirement. When translating to ISD::VECTOR_SHUFFLE, if the
9394 // operands do not match the mask length, they are extended by concatenating
9395 // them with undef vectors. That is probably the right thing for other
9396 // targets, but for NEON it is better to concatenate two double-register
9397 // size vector operands into a single quad-register size vector. Do that
9398 // transformation here:
9399 // shuffle(concat(v1, undef), concat(v2, undef)) ->
9400 // shuffle(concat(v1, v2), undef)
9401 SDValue Op0 = N->getOperand(0);
9402 SDValue Op1 = N->getOperand(1);
9403 if (Op0.getOpcode() != ISD::CONCAT_VECTORS ||
9404 Op1.getOpcode() != ISD::CONCAT_VECTORS ||
9405 Op0.getNumOperands() != 2 ||
9406 Op1.getNumOperands() != 2)
9408 SDValue Concat0Op1 = Op0.getOperand(1);
9409 SDValue Concat1Op1 = Op1.getOperand(1);
9410 if (Concat0Op1.getOpcode() != ISD::UNDEF ||
9411 Concat1Op1.getOpcode() != ISD::UNDEF)
9413 // Skip the transformation if any of the types are illegal.
9414 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9415 EVT VT = N->getValueType(0);
9416 if (!TLI.isTypeLegal(VT) ||
9417 !TLI.isTypeLegal(Concat0Op1.getValueType()) ||
9418 !TLI.isTypeLegal(Concat1Op1.getValueType()))
9421 SDValue NewConcat = DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT,
9422 Op0.getOperand(0), Op1.getOperand(0));
9423 // Translate the shuffle mask.
9424 SmallVector<int, 16> NewMask;
9425 unsigned NumElts = VT.getVectorNumElements();
9426 unsigned HalfElts = NumElts/2;
9427 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
9428 for (unsigned n = 0; n < NumElts; ++n) {
9429 int MaskElt = SVN->getMaskElt(n);
9431 if (MaskElt < (int)HalfElts)
9433 else if (MaskElt >= (int)NumElts && MaskElt < (int)(NumElts + HalfElts))
9434 NewElt = HalfElts + MaskElt - NumElts;
9435 NewMask.push_back(NewElt);
9437 return DAG.getVectorShuffle(VT, SDLoc(N), NewConcat,
9438 DAG.getUNDEF(VT), NewMask.data());
9441 /// CombineBaseUpdate - Target-specific DAG combine function for VLDDUP,
9442 /// NEON load/store intrinsics, and generic vector load/stores, to merge
9443 /// base address updates.
9444 /// For generic load/stores, the memory type is assumed to be a vector.
9445 /// The caller is assumed to have checked legality.
9446 static SDValue CombineBaseUpdate(SDNode *N,
9447 TargetLowering::DAGCombinerInfo &DCI) {
9448 SelectionDAG &DAG = DCI.DAG;
9449 const bool isIntrinsic = (N->getOpcode() == ISD::INTRINSIC_VOID ||
9450 N->getOpcode() == ISD::INTRINSIC_W_CHAIN);
9451 const bool isStore = N->getOpcode() == ISD::STORE;
9452 const unsigned AddrOpIdx = ((isIntrinsic || isStore) ? 2 : 1);
9453 SDValue Addr = N->getOperand(AddrOpIdx);
9454 MemSDNode *MemN = cast<MemSDNode>(N);
9457 // Search for a use of the address operand that is an increment.
9458 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(),
9459 UE = Addr.getNode()->use_end(); UI != UE; ++UI) {
9461 if (User->getOpcode() != ISD::ADD ||
9462 UI.getUse().getResNo() != Addr.getResNo())
9465 // Check that the add is independent of the load/store. Otherwise, folding
9466 // it would create a cycle.
9467 if (User->isPredecessorOf(N) || N->isPredecessorOf(User))
9470 // Find the new opcode for the updating load/store.
9471 bool isLoadOp = true;
9472 bool isLaneOp = false;
9473 unsigned NewOpc = 0;
9474 unsigned NumVecs = 0;
9476 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
9478 default: llvm_unreachable("unexpected intrinsic for Neon base update");
9479 case Intrinsic::arm_neon_vld1: NewOpc = ARMISD::VLD1_UPD;
9481 case Intrinsic::arm_neon_vld2: NewOpc = ARMISD::VLD2_UPD;
9483 case Intrinsic::arm_neon_vld3: NewOpc = ARMISD::VLD3_UPD;
9485 case Intrinsic::arm_neon_vld4: NewOpc = ARMISD::VLD4_UPD;
9487 case Intrinsic::arm_neon_vld2lane: NewOpc = ARMISD::VLD2LN_UPD;
9488 NumVecs = 2; isLaneOp = true; break;
9489 case Intrinsic::arm_neon_vld3lane: NewOpc = ARMISD::VLD3LN_UPD;
9490 NumVecs = 3; isLaneOp = true; break;
9491 case Intrinsic::arm_neon_vld4lane: NewOpc = ARMISD::VLD4LN_UPD;
9492 NumVecs = 4; isLaneOp = true; break;
9493 case Intrinsic::arm_neon_vst1: NewOpc = ARMISD::VST1_UPD;
9494 NumVecs = 1; isLoadOp = false; break;
9495 case Intrinsic::arm_neon_vst2: NewOpc = ARMISD::VST2_UPD;
9496 NumVecs = 2; isLoadOp = false; break;
9497 case Intrinsic::arm_neon_vst3: NewOpc = ARMISD::VST3_UPD;
9498 NumVecs = 3; isLoadOp = false; break;
9499 case Intrinsic::arm_neon_vst4: NewOpc = ARMISD::VST4_UPD;
9500 NumVecs = 4; isLoadOp = false; break;
9501 case Intrinsic::arm_neon_vst2lane: NewOpc = ARMISD::VST2LN_UPD;
9502 NumVecs = 2; isLoadOp = false; isLaneOp = true; break;
9503 case Intrinsic::arm_neon_vst3lane: NewOpc = ARMISD::VST3LN_UPD;
9504 NumVecs = 3; isLoadOp = false; isLaneOp = true; break;
9505 case Intrinsic::arm_neon_vst4lane: NewOpc = ARMISD::VST4LN_UPD;
9506 NumVecs = 4; isLoadOp = false; isLaneOp = true; break;
9510 switch (N->getOpcode()) {
9511 default: llvm_unreachable("unexpected opcode for Neon base update");
9512 case ARMISD::VLD2DUP: NewOpc = ARMISD::VLD2DUP_UPD; NumVecs = 2; break;
9513 case ARMISD::VLD3DUP: NewOpc = ARMISD::VLD3DUP_UPD; NumVecs = 3; break;
9514 case ARMISD::VLD4DUP: NewOpc = ARMISD::VLD4DUP_UPD; NumVecs = 4; break;
9515 case ISD::LOAD: NewOpc = ARMISD::VLD1_UPD;
9516 NumVecs = 1; isLaneOp = false; break;
9517 case ISD::STORE: NewOpc = ARMISD::VST1_UPD;
9518 NumVecs = 1; isLaneOp = false; isLoadOp = false; break;
9522 // Find the size of memory referenced by the load/store.
9525 VecTy = N->getValueType(0);
9526 } else if (isIntrinsic) {
9527 VecTy = N->getOperand(AddrOpIdx+1).getValueType();
9529 assert(isStore && "Node has to be a load, a store, or an intrinsic!");
9530 VecTy = N->getOperand(1).getValueType();
9533 unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8;
9535 NumBytes /= VecTy.getVectorNumElements();
9537 // If the increment is a constant, it must match the memory ref size.
9538 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
9539 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
9540 uint64_t IncVal = CInc->getZExtValue();
9541 if (IncVal != NumBytes)
9543 } else if (NumBytes >= 3 * 16) {
9544 // VLD3/4 and VST3/4 for 128-bit vectors are implemented with two
9545 // separate instructions that make it harder to use a non-constant update.
9549 // OK, we found an ADD we can fold into the base update.
9550 // Now, create a _UPD node, taking care of not breaking alignment.
9552 EVT AlignedVecTy = VecTy;
9553 unsigned Alignment = MemN->getAlignment();
9555 // If this is a less-than-standard-aligned load/store, change the type to
9556 // match the standard alignment.
9557 // The alignment is overlooked when selecting _UPD variants; and it's
9558 // easier to introduce bitcasts here than fix that.
9559 // There are 3 ways to get to this base-update combine:
9560 // - intrinsics: they are assumed to be properly aligned (to the standard
9561 // alignment of the memory type), so we don't need to do anything.
9562 // - ARMISD::VLDx nodes: they are only generated from the aforementioned
9563 // intrinsics, so, likewise, there's nothing to do.
9564 // - generic load/store instructions: the alignment is specified as an
9565 // explicit operand, rather than implicitly as the standard alignment
9566 // of the memory type (like the intrisics). We need to change the
9567 // memory type to match the explicit alignment. That way, we don't
9568 // generate non-standard-aligned ARMISD::VLDx nodes.
9569 if (isa<LSBaseSDNode>(N)) {
9572 if (Alignment < VecTy.getScalarSizeInBits() / 8) {
9573 MVT EltTy = MVT::getIntegerVT(Alignment * 8);
9574 assert(NumVecs == 1 && "Unexpected multi-element generic load/store.");
9575 assert(!isLaneOp && "Unexpected generic load/store lane.");
9576 unsigned NumElts = NumBytes / (EltTy.getSizeInBits() / 8);
9577 AlignedVecTy = MVT::getVectorVT(EltTy, NumElts);
9579 // Don't set an explicit alignment on regular load/stores that we want
9580 // to transform to VLD/VST 1_UPD nodes.
9581 // This matches the behavior of regular load/stores, which only get an
9582 // explicit alignment if the MMO alignment is larger than the standard
9583 // alignment of the memory type.
9584 // Intrinsics, however, always get an explicit alignment, set to the
9585 // alignment of the MMO.
9589 // Create the new updating load/store node.
9590 // First, create an SDVTList for the new updating node's results.
9592 unsigned NumResultVecs = (isLoadOp ? NumVecs : 0);
9594 for (n = 0; n < NumResultVecs; ++n)
9595 Tys[n] = AlignedVecTy;
9596 Tys[n++] = MVT::i32;
9597 Tys[n] = MVT::Other;
9598 SDVTList SDTys = DAG.getVTList(makeArrayRef(Tys, NumResultVecs+2));
9600 // Then, gather the new node's operands.
9601 SmallVector<SDValue, 8> Ops;
9602 Ops.push_back(N->getOperand(0)); // incoming chain
9603 Ops.push_back(N->getOperand(AddrOpIdx));
9606 if (StoreSDNode *StN = dyn_cast<StoreSDNode>(N)) {
9607 // Try to match the intrinsic's signature
9608 Ops.push_back(StN->getValue());
9610 // Loads (and of course intrinsics) match the intrinsics' signature,
9611 // so just add all but the alignment operand.
9612 for (unsigned i = AddrOpIdx + 1; i < N->getNumOperands() - 1; ++i)
9613 Ops.push_back(N->getOperand(i));
9616 // For all node types, the alignment operand is always the last one.
9617 Ops.push_back(DAG.getConstant(Alignment, dl, MVT::i32));
9619 // If this is a non-standard-aligned STORE, the penultimate operand is the
9620 // stored value. Bitcast it to the aligned type.
9621 if (AlignedVecTy != VecTy && N->getOpcode() == ISD::STORE) {
9622 SDValue &StVal = Ops[Ops.size()-2];
9623 StVal = DAG.getNode(ISD::BITCAST, dl, AlignedVecTy, StVal);
9626 SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, dl, SDTys,
9628 MemN->getMemOperand());
9631 SmallVector<SDValue, 5> NewResults;
9632 for (unsigned i = 0; i < NumResultVecs; ++i)
9633 NewResults.push_back(SDValue(UpdN.getNode(), i));
9635 // If this is an non-standard-aligned LOAD, the first result is the loaded
9636 // value. Bitcast it to the expected result type.
9637 if (AlignedVecTy != VecTy && N->getOpcode() == ISD::LOAD) {
9638 SDValue &LdVal = NewResults[0];
9639 LdVal = DAG.getNode(ISD::BITCAST, dl, VecTy, LdVal);
9642 NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs+1)); // chain
9643 DCI.CombineTo(N, NewResults);
9644 DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs));
9651 static SDValue PerformVLDCombine(SDNode *N,
9652 TargetLowering::DAGCombinerInfo &DCI) {
9653 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
9656 return CombineBaseUpdate(N, DCI);
9659 /// CombineVLDDUP - For a VDUPLANE node N, check if its source operand is a
9660 /// vldN-lane (N > 1) intrinsic, and if all the other uses of that intrinsic
9661 /// are also VDUPLANEs. If so, combine them to a vldN-dup operation and
9663 static bool CombineVLDDUP(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
9664 SelectionDAG &DAG = DCI.DAG;
9665 EVT VT = N->getValueType(0);
9666 // vldN-dup instructions only support 64-bit vectors for N > 1.
9667 if (!VT.is64BitVector())
9670 // Check if the VDUPLANE operand is a vldN-dup intrinsic.
9671 SDNode *VLD = N->getOperand(0).getNode();
9672 if (VLD->getOpcode() != ISD::INTRINSIC_W_CHAIN)
9674 unsigned NumVecs = 0;
9675 unsigned NewOpc = 0;
9676 unsigned IntNo = cast<ConstantSDNode>(VLD->getOperand(1))->getZExtValue();
9677 if (IntNo == Intrinsic::arm_neon_vld2lane) {
9679 NewOpc = ARMISD::VLD2DUP;
9680 } else if (IntNo == Intrinsic::arm_neon_vld3lane) {
9682 NewOpc = ARMISD::VLD3DUP;
9683 } else if (IntNo == Intrinsic::arm_neon_vld4lane) {
9685 NewOpc = ARMISD::VLD4DUP;
9690 // First check that all the vldN-lane uses are VDUPLANEs and that the lane
9691 // numbers match the load.
9692 unsigned VLDLaneNo =
9693 cast<ConstantSDNode>(VLD->getOperand(NumVecs+3))->getZExtValue();
9694 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
9696 // Ignore uses of the chain result.
9697 if (UI.getUse().getResNo() == NumVecs)
9700 if (User->getOpcode() != ARMISD::VDUPLANE ||
9701 VLDLaneNo != cast<ConstantSDNode>(User->getOperand(1))->getZExtValue())
9705 // Create the vldN-dup node.
9708 for (n = 0; n < NumVecs; ++n)
9710 Tys[n] = MVT::Other;
9711 SDVTList SDTys = DAG.getVTList(makeArrayRef(Tys, NumVecs+1));
9712 SDValue Ops[] = { VLD->getOperand(0), VLD->getOperand(2) };
9713 MemIntrinsicSDNode *VLDMemInt = cast<MemIntrinsicSDNode>(VLD);
9714 SDValue VLDDup = DAG.getMemIntrinsicNode(NewOpc, SDLoc(VLD), SDTys,
9715 Ops, VLDMemInt->getMemoryVT(),
9716 VLDMemInt->getMemOperand());
9719 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
9721 unsigned ResNo = UI.getUse().getResNo();
9722 // Ignore uses of the chain result.
9723 if (ResNo == NumVecs)
9726 DCI.CombineTo(User, SDValue(VLDDup.getNode(), ResNo));
9729 // Now the vldN-lane intrinsic is dead except for its chain result.
9730 // Update uses of the chain.
9731 std::vector<SDValue> VLDDupResults;
9732 for (unsigned n = 0; n < NumVecs; ++n)
9733 VLDDupResults.push_back(SDValue(VLDDup.getNode(), n));
9734 VLDDupResults.push_back(SDValue(VLDDup.getNode(), NumVecs));
9735 DCI.CombineTo(VLD, VLDDupResults);
9740 /// PerformVDUPLANECombine - Target-specific dag combine xforms for
9741 /// ARMISD::VDUPLANE.
9742 static SDValue PerformVDUPLANECombine(SDNode *N,
9743 TargetLowering::DAGCombinerInfo &DCI) {
9744 SDValue Op = N->getOperand(0);
9746 // If the source is a vldN-lane (N > 1) intrinsic, and all the other uses
9747 // of that intrinsic are also VDUPLANEs, combine them to a vldN-dup operation.
9748 if (CombineVLDDUP(N, DCI))
9749 return SDValue(N, 0);
9751 // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is
9752 // redundant. Ignore bit_converts for now; element sizes are checked below.
9753 while (Op.getOpcode() == ISD::BITCAST)
9754 Op = Op.getOperand(0);
9755 if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM)
9758 // Make sure the VMOV element size is not bigger than the VDUPLANE elements.
9759 unsigned EltSize = Op.getValueType().getVectorElementType().getSizeInBits();
9760 // The canonical VMOV for a zero vector uses a 32-bit element size.
9761 unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
9763 if (ARM_AM::decodeNEONModImm(Imm, EltBits) == 0)
9765 EVT VT = N->getValueType(0);
9766 if (EltSize > VT.getVectorElementType().getSizeInBits())
9769 return DCI.DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Op);
9772 static SDValue PerformLOADCombine(SDNode *N,
9773 TargetLowering::DAGCombinerInfo &DCI) {
9774 EVT VT = N->getValueType(0);
9776 // If this is a legal vector load, try to combine it into a VLD1_UPD.
9777 if (ISD::isNormalLoad(N) && VT.isVector() &&
9778 DCI.DAG.getTargetLoweringInfo().isTypeLegal(VT))
9779 return CombineBaseUpdate(N, DCI);
9784 /// PerformSTORECombine - Target-specific dag combine xforms for
9786 static SDValue PerformSTORECombine(SDNode *N,
9787 TargetLowering::DAGCombinerInfo &DCI) {
9788 StoreSDNode *St = cast<StoreSDNode>(N);
9789 if (St->isVolatile())
9792 // Optimize trunc store (of multiple scalars) to shuffle and store. First,
9793 // pack all of the elements in one place. Next, store to memory in fewer
9795 SDValue StVal = St->getValue();
9796 EVT VT = StVal.getValueType();
9797 if (St->isTruncatingStore() && VT.isVector()) {
9798 SelectionDAG &DAG = DCI.DAG;
9799 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9800 EVT StVT = St->getMemoryVT();
9801 unsigned NumElems = VT.getVectorNumElements();
9802 assert(StVT != VT && "Cannot truncate to the same type");
9803 unsigned FromEltSz = VT.getVectorElementType().getSizeInBits();
9804 unsigned ToEltSz = StVT.getVectorElementType().getSizeInBits();
9806 // From, To sizes and ElemCount must be pow of two
9807 if (!isPowerOf2_32(NumElems * FromEltSz * ToEltSz)) return SDValue();
9809 // We are going to use the original vector elt for storing.
9810 // Accumulated smaller vector elements must be a multiple of the store size.
9811 if (0 != (NumElems * FromEltSz) % ToEltSz) return SDValue();
9813 unsigned SizeRatio = FromEltSz / ToEltSz;
9814 assert(SizeRatio * NumElems * ToEltSz == VT.getSizeInBits());
9816 // Create a type on which we perform the shuffle.
9817 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(), StVT.getScalarType(),
9818 NumElems*SizeRatio);
9819 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
9822 SDValue WideVec = DAG.getNode(ISD::BITCAST, DL, WideVecVT, StVal);
9823 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
9824 for (unsigned i = 0; i < NumElems; ++i)
9825 ShuffleVec[i] = DAG.getDataLayout().isBigEndian()
9826 ? (i + 1) * SizeRatio - 1
9829 // Can't shuffle using an illegal type.
9830 if (!TLI.isTypeLegal(WideVecVT)) return SDValue();
9832 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, DL, WideVec,
9833 DAG.getUNDEF(WideVec.getValueType()),
9835 // At this point all of the data is stored at the bottom of the
9836 // register. We now need to save it to mem.
9838 // Find the largest store unit
9839 MVT StoreType = MVT::i8;
9840 for (MVT Tp : MVT::integer_valuetypes()) {
9841 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToEltSz)
9844 // Didn't find a legal store type.
9845 if (!TLI.isTypeLegal(StoreType))
9848 // Bitcast the original vector into a vector of store-size units
9849 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
9850 StoreType, VT.getSizeInBits()/EVT(StoreType).getSizeInBits());
9851 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
9852 SDValue ShuffWide = DAG.getNode(ISD::BITCAST, DL, StoreVecVT, Shuff);
9853 SmallVector<SDValue, 8> Chains;
9854 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits() / 8, DL,
9855 TLI.getPointerTy(DAG.getDataLayout()));
9856 SDValue BasePtr = St->getBasePtr();
9858 // Perform one or more big stores into memory.
9859 unsigned E = (ToEltSz*NumElems)/StoreType.getSizeInBits();
9860 for (unsigned I = 0; I < E; I++) {
9861 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
9862 StoreType, ShuffWide,
9863 DAG.getIntPtrConstant(I, DL));
9864 SDValue Ch = DAG.getStore(St->getChain(), DL, SubVec, BasePtr,
9865 St->getPointerInfo(), St->isVolatile(),
9866 St->isNonTemporal(), St->getAlignment());
9867 BasePtr = DAG.getNode(ISD::ADD, DL, BasePtr.getValueType(), BasePtr,
9869 Chains.push_back(Ch);
9871 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
9874 if (!ISD::isNormalStore(St))
9877 // Split a store of a VMOVDRR into two integer stores to avoid mixing NEON and
9878 // ARM stores of arguments in the same cache line.
9879 if (StVal.getNode()->getOpcode() == ARMISD::VMOVDRR &&
9880 StVal.getNode()->hasOneUse()) {
9881 SelectionDAG &DAG = DCI.DAG;
9882 bool isBigEndian = DAG.getDataLayout().isBigEndian();
9884 SDValue BasePtr = St->getBasePtr();
9885 SDValue NewST1 = DAG.getStore(St->getChain(), DL,
9886 StVal.getNode()->getOperand(isBigEndian ? 1 : 0 ),
9887 BasePtr, St->getPointerInfo(), St->isVolatile(),
9888 St->isNonTemporal(), St->getAlignment());
9890 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
9891 DAG.getConstant(4, DL, MVT::i32));
9892 return DAG.getStore(NewST1.getValue(0), DL,
9893 StVal.getNode()->getOperand(isBigEndian ? 0 : 1),
9894 OffsetPtr, St->getPointerInfo(), St->isVolatile(),
9895 St->isNonTemporal(),
9896 std::min(4U, St->getAlignment() / 2));
9899 if (StVal.getValueType() == MVT::i64 &&
9900 StVal.getNode()->getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
9902 // Bitcast an i64 store extracted from a vector to f64.
9903 // Otherwise, the i64 value will be legalized to a pair of i32 values.
9904 SelectionDAG &DAG = DCI.DAG;
9906 SDValue IntVec = StVal.getOperand(0);
9907 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
9908 IntVec.getValueType().getVectorNumElements());
9909 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, IntVec);
9910 SDValue ExtElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
9911 Vec, StVal.getOperand(1));
9913 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ExtElt);
9914 // Make the DAGCombiner fold the bitcasts.
9915 DCI.AddToWorklist(Vec.getNode());
9916 DCI.AddToWorklist(ExtElt.getNode());
9917 DCI.AddToWorklist(V.getNode());
9918 return DAG.getStore(St->getChain(), dl, V, St->getBasePtr(),
9919 St->getPointerInfo(), St->isVolatile(),
9920 St->isNonTemporal(), St->getAlignment(),
9924 // If this is a legal vector store, try to combine it into a VST1_UPD.
9925 if (ISD::isNormalStore(N) && VT.isVector() &&
9926 DCI.DAG.getTargetLoweringInfo().isTypeLegal(VT))
9927 return CombineBaseUpdate(N, DCI);
9932 /// PerformVCVTCombine - VCVT (floating-point to fixed-point, Advanced SIMD)
9933 /// can replace combinations of VMUL and VCVT (floating-point to integer)
9934 /// when the VMUL has a constant operand that is a power of 2.
9936 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
9937 /// vmul.f32 d16, d17, d16
9938 /// vcvt.s32.f32 d16, d16
9940 /// vcvt.s32.f32 d16, d16, #3
9941 static SDValue PerformVCVTCombine(SDNode *N, SelectionDAG &DAG,
9942 const ARMSubtarget *Subtarget) {
9943 if (!Subtarget->hasNEON())
9946 SDValue Op = N->getOperand(0);
9947 if (!Op.getValueType().isVector() || Op.getOpcode() != ISD::FMUL)
9950 SDValue ConstVec = Op->getOperand(1);
9951 if (!isa<BuildVectorSDNode>(ConstVec))
9954 MVT FloatTy = Op.getSimpleValueType().getVectorElementType();
9955 uint32_t FloatBits = FloatTy.getSizeInBits();
9956 MVT IntTy = N->getSimpleValueType(0).getVectorElementType();
9957 uint32_t IntBits = IntTy.getSizeInBits();
9958 unsigned NumLanes = Op.getValueType().getVectorNumElements();
9959 if (FloatBits != 32 || IntBits > 32 || NumLanes > 4) {
9960 // These instructions only exist converting from f32 to i32. We can handle
9961 // smaller integers by generating an extra truncate, but larger ones would
9962 // be lossy. We also can't handle more then 4 lanes, since these intructions
9963 // only support v2i32/v4i32 types.
9967 BitVector UndefElements;
9968 BuildVectorSDNode *BV = cast<BuildVectorSDNode>(ConstVec);
9969 int32_t C = BV->getConstantFPSplatPow2ToLog2Int(&UndefElements, 33);
9970 if (C == -1 || C == 0 || C > 32)
9974 bool isSigned = N->getOpcode() == ISD::FP_TO_SINT;
9975 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfp2fxs :
9976 Intrinsic::arm_neon_vcvtfp2fxu;
9977 SDValue FixConv = DAG.getNode(
9978 ISD::INTRINSIC_WO_CHAIN, dl, NumLanes == 2 ? MVT::v2i32 : MVT::v4i32,
9979 DAG.getConstant(IntrinsicOpcode, dl, MVT::i32), Op->getOperand(0),
9980 DAG.getConstant(C, dl, MVT::i32));
9982 if (IntBits < FloatBits)
9983 FixConv = DAG.getNode(ISD::TRUNCATE, dl, N->getValueType(0), FixConv);
9988 /// PerformVDIVCombine - VCVT (fixed-point to floating-point, Advanced SIMD)
9989 /// can replace combinations of VCVT (integer to floating-point) and VDIV
9990 /// when the VDIV has a constant operand that is a power of 2.
9992 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
9993 /// vcvt.f32.s32 d16, d16
9994 /// vdiv.f32 d16, d17, d16
9996 /// vcvt.f32.s32 d16, d16, #3
9997 static SDValue PerformVDIVCombine(SDNode *N, SelectionDAG &DAG,
9998 const ARMSubtarget *Subtarget) {
9999 if (!Subtarget->hasNEON())
10002 SDValue Op = N->getOperand(0);
10003 unsigned OpOpcode = Op.getNode()->getOpcode();
10004 if (!N->getValueType(0).isVector() ||
10005 (OpOpcode != ISD::SINT_TO_FP && OpOpcode != ISD::UINT_TO_FP))
10008 SDValue ConstVec = N->getOperand(1);
10009 if (!isa<BuildVectorSDNode>(ConstVec))
10012 MVT FloatTy = N->getSimpleValueType(0).getVectorElementType();
10013 uint32_t FloatBits = FloatTy.getSizeInBits();
10014 MVT IntTy = Op.getOperand(0).getSimpleValueType().getVectorElementType();
10015 uint32_t IntBits = IntTy.getSizeInBits();
10016 unsigned NumLanes = Op.getValueType().getVectorNumElements();
10017 if (FloatBits != 32 || IntBits > 32 || NumLanes > 4) {
10018 // These instructions only exist converting from i32 to f32. We can handle
10019 // smaller integers by generating an extra extend, but larger ones would
10020 // be lossy. We also can't handle more then 4 lanes, since these intructions
10021 // only support v2i32/v4i32 types.
10025 BitVector UndefElements;
10026 BuildVectorSDNode *BV = cast<BuildVectorSDNode>(ConstVec);
10027 int32_t C = BV->getConstantFPSplatPow2ToLog2Int(&UndefElements, 33);
10028 if (C == -1 || C == 0 || C > 32)
10032 bool isSigned = OpOpcode == ISD::SINT_TO_FP;
10033 SDValue ConvInput = Op.getOperand(0);
10034 if (IntBits < FloatBits)
10035 ConvInput = DAG.getNode(isSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND,
10036 dl, NumLanes == 2 ? MVT::v2i32 : MVT::v4i32,
10039 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfxs2fp :
10040 Intrinsic::arm_neon_vcvtfxu2fp;
10041 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl,
10043 DAG.getConstant(IntrinsicOpcode, dl, MVT::i32),
10044 ConvInput, DAG.getConstant(C, dl, MVT::i32));
10047 /// Getvshiftimm - Check if this is a valid build_vector for the immediate
10048 /// operand of a vector shift operation, where all the elements of the
10049 /// build_vector must have the same constant integer value.
10050 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
10051 // Ignore bit_converts.
10052 while (Op.getOpcode() == ISD::BITCAST)
10053 Op = Op.getOperand(0);
10054 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
10055 APInt SplatBits, SplatUndef;
10056 unsigned SplatBitSize;
10058 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
10059 HasAnyUndefs, ElementBits) ||
10060 SplatBitSize > ElementBits)
10062 Cnt = SplatBits.getSExtValue();
10066 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
10067 /// operand of a vector shift left operation. That value must be in the range:
10068 /// 0 <= Value < ElementBits for a left shift; or
10069 /// 0 <= Value <= ElementBits for a long left shift.
10070 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
10071 assert(VT.isVector() && "vector shift count is not a vector type");
10072 int64_t ElementBits = VT.getVectorElementType().getSizeInBits();
10073 if (! getVShiftImm(Op, ElementBits, Cnt))
10075 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
10078 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
10079 /// operand of a vector shift right operation. For a shift opcode, the value
10080 /// is positive, but for an intrinsic the value count must be negative. The
10081 /// absolute value must be in the range:
10082 /// 1 <= |Value| <= ElementBits for a right shift; or
10083 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
10084 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
10086 assert(VT.isVector() && "vector shift count is not a vector type");
10087 int64_t ElementBits = VT.getVectorElementType().getSizeInBits();
10088 if (! getVShiftImm(Op, ElementBits, Cnt))
10091 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
10092 if (Cnt >= -(isNarrow ? ElementBits/2 : ElementBits) && Cnt <= -1) {
10099 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
10100 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
10101 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
10104 // Don't do anything for most intrinsics.
10107 case Intrinsic::arm_neon_vabds:
10108 if (!N->getValueType(0).isInteger())
10110 return DAG.getNode(ISD::SABSDIFF, SDLoc(N), N->getValueType(0),
10111 N->getOperand(1), N->getOperand(2));
10112 case Intrinsic::arm_neon_vabdu:
10113 return DAG.getNode(ISD::UABSDIFF, SDLoc(N), N->getValueType(0),
10114 N->getOperand(1), N->getOperand(2));
10116 // Vector shifts: check for immediate versions and lower them.
10117 // Note: This is done during DAG combining instead of DAG legalizing because
10118 // the build_vectors for 64-bit vector element shift counts are generally
10119 // not legal, and it is hard to see their values after they get legalized to
10120 // loads from a constant pool.
10121 case Intrinsic::arm_neon_vshifts:
10122 case Intrinsic::arm_neon_vshiftu:
10123 case Intrinsic::arm_neon_vrshifts:
10124 case Intrinsic::arm_neon_vrshiftu:
10125 case Intrinsic::arm_neon_vrshiftn:
10126 case Intrinsic::arm_neon_vqshifts:
10127 case Intrinsic::arm_neon_vqshiftu:
10128 case Intrinsic::arm_neon_vqshiftsu:
10129 case Intrinsic::arm_neon_vqshiftns:
10130 case Intrinsic::arm_neon_vqshiftnu:
10131 case Intrinsic::arm_neon_vqshiftnsu:
10132 case Intrinsic::arm_neon_vqrshiftns:
10133 case Intrinsic::arm_neon_vqrshiftnu:
10134 case Intrinsic::arm_neon_vqrshiftnsu: {
10135 EVT VT = N->getOperand(1).getValueType();
10137 unsigned VShiftOpc = 0;
10140 case Intrinsic::arm_neon_vshifts:
10141 case Intrinsic::arm_neon_vshiftu:
10142 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
10143 VShiftOpc = ARMISD::VSHL;
10146 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
10147 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
10148 ARMISD::VSHRs : ARMISD::VSHRu);
10153 case Intrinsic::arm_neon_vrshifts:
10154 case Intrinsic::arm_neon_vrshiftu:
10155 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
10159 case Intrinsic::arm_neon_vqshifts:
10160 case Intrinsic::arm_neon_vqshiftu:
10161 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
10165 case Intrinsic::arm_neon_vqshiftsu:
10166 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
10168 llvm_unreachable("invalid shift count for vqshlu intrinsic");
10170 case Intrinsic::arm_neon_vrshiftn:
10171 case Intrinsic::arm_neon_vqshiftns:
10172 case Intrinsic::arm_neon_vqshiftnu:
10173 case Intrinsic::arm_neon_vqshiftnsu:
10174 case Intrinsic::arm_neon_vqrshiftns:
10175 case Intrinsic::arm_neon_vqrshiftnu:
10176 case Intrinsic::arm_neon_vqrshiftnsu:
10177 // Narrowing shifts require an immediate right shift.
10178 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
10180 llvm_unreachable("invalid shift count for narrowing vector shift "
10184 llvm_unreachable("unhandled vector shift");
10188 case Intrinsic::arm_neon_vshifts:
10189 case Intrinsic::arm_neon_vshiftu:
10190 // Opcode already set above.
10192 case Intrinsic::arm_neon_vrshifts:
10193 VShiftOpc = ARMISD::VRSHRs; break;
10194 case Intrinsic::arm_neon_vrshiftu:
10195 VShiftOpc = ARMISD::VRSHRu; break;
10196 case Intrinsic::arm_neon_vrshiftn:
10197 VShiftOpc = ARMISD::VRSHRN; break;
10198 case Intrinsic::arm_neon_vqshifts:
10199 VShiftOpc = ARMISD::VQSHLs; break;
10200 case Intrinsic::arm_neon_vqshiftu:
10201 VShiftOpc = ARMISD::VQSHLu; break;
10202 case Intrinsic::arm_neon_vqshiftsu:
10203 VShiftOpc = ARMISD::VQSHLsu; break;
10204 case Intrinsic::arm_neon_vqshiftns:
10205 VShiftOpc = ARMISD::VQSHRNs; break;
10206 case Intrinsic::arm_neon_vqshiftnu:
10207 VShiftOpc = ARMISD::VQSHRNu; break;
10208 case Intrinsic::arm_neon_vqshiftnsu:
10209 VShiftOpc = ARMISD::VQSHRNsu; break;
10210 case Intrinsic::arm_neon_vqrshiftns:
10211 VShiftOpc = ARMISD::VQRSHRNs; break;
10212 case Intrinsic::arm_neon_vqrshiftnu:
10213 VShiftOpc = ARMISD::VQRSHRNu; break;
10214 case Intrinsic::arm_neon_vqrshiftnsu:
10215 VShiftOpc = ARMISD::VQRSHRNsu; break;
10219 return DAG.getNode(VShiftOpc, dl, N->getValueType(0),
10220 N->getOperand(1), DAG.getConstant(Cnt, dl, MVT::i32));
10223 case Intrinsic::arm_neon_vshiftins: {
10224 EVT VT = N->getOperand(1).getValueType();
10226 unsigned VShiftOpc = 0;
10228 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
10229 VShiftOpc = ARMISD::VSLI;
10230 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
10231 VShiftOpc = ARMISD::VSRI;
10233 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
10237 return DAG.getNode(VShiftOpc, dl, N->getValueType(0),
10238 N->getOperand(1), N->getOperand(2),
10239 DAG.getConstant(Cnt, dl, MVT::i32));
10242 case Intrinsic::arm_neon_vqrshifts:
10243 case Intrinsic::arm_neon_vqrshiftu:
10244 // No immediate versions of these to check for.
10251 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
10252 /// lowers them. As with the vector shift intrinsics, this is done during DAG
10253 /// combining instead of DAG legalizing because the build_vectors for 64-bit
10254 /// vector element shift counts are generally not legal, and it is hard to see
10255 /// their values after they get legalized to loads from a constant pool.
10256 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
10257 const ARMSubtarget *ST) {
10258 EVT VT = N->getValueType(0);
10259 if (N->getOpcode() == ISD::SRL && VT == MVT::i32 && ST->hasV6Ops()) {
10260 // Canonicalize (srl (bswap x), 16) to (rotr (bswap x), 16) if the high
10261 // 16-bits of x is zero. This optimizes rev + lsr 16 to rev16.
10262 SDValue N1 = N->getOperand(1);
10263 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) {
10264 SDValue N0 = N->getOperand(0);
10265 if (C->getZExtValue() == 16 && N0.getOpcode() == ISD::BSWAP &&
10266 DAG.MaskedValueIsZero(N0.getOperand(0),
10267 APInt::getHighBitsSet(32, 16)))
10268 return DAG.getNode(ISD::ROTR, SDLoc(N), VT, N0, N1);
10272 // Nothing to be done for scalar shifts.
10273 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
10274 if (!VT.isVector() || !TLI.isTypeLegal(VT))
10277 assert(ST->hasNEON() && "unexpected vector shift");
10280 switch (N->getOpcode()) {
10281 default: llvm_unreachable("unexpected shift opcode");
10284 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt)) {
10286 return DAG.getNode(ARMISD::VSHL, dl, VT, N->getOperand(0),
10287 DAG.getConstant(Cnt, dl, MVT::i32));
10293 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
10294 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
10295 ARMISD::VSHRs : ARMISD::VSHRu);
10297 return DAG.getNode(VShiftOpc, dl, VT, N->getOperand(0),
10298 DAG.getConstant(Cnt, dl, MVT::i32));
10304 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
10305 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
10306 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
10307 const ARMSubtarget *ST) {
10308 SDValue N0 = N->getOperand(0);
10310 // Check for sign- and zero-extensions of vector extract operations of 8-
10311 // and 16-bit vector elements. NEON supports these directly. They are
10312 // handled during DAG combining because type legalization will promote them
10313 // to 32-bit types and it is messy to recognize the operations after that.
10314 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
10315 SDValue Vec = N0.getOperand(0);
10316 SDValue Lane = N0.getOperand(1);
10317 EVT VT = N->getValueType(0);
10318 EVT EltVT = N0.getValueType();
10319 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
10321 if (VT == MVT::i32 &&
10322 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
10323 TLI.isTypeLegal(Vec.getValueType()) &&
10324 isa<ConstantSDNode>(Lane)) {
10327 switch (N->getOpcode()) {
10328 default: llvm_unreachable("unexpected opcode");
10329 case ISD::SIGN_EXTEND:
10330 Opc = ARMISD::VGETLANEs;
10332 case ISD::ZERO_EXTEND:
10333 case ISD::ANY_EXTEND:
10334 Opc = ARMISD::VGETLANEu;
10337 return DAG.getNode(Opc, SDLoc(N), VT, Vec, Lane);
10344 static void computeKnownBits(SelectionDAG &DAG, SDValue Op, APInt &KnownZero,
10346 if (Op.getOpcode() == ARMISD::BFI) {
10347 // Conservatively, we can recurse down the first operand
10348 // and just mask out all affected bits.
10349 computeKnownBits(DAG, Op.getOperand(0), KnownZero, KnownOne);
10351 // The operand to BFI is already a mask suitable for removing the bits it
10353 ConstantSDNode *CI = cast<ConstantSDNode>(Op.getOperand(2));
10354 APInt Mask = CI->getAPIntValue();
10359 if (Op.getOpcode() == ARMISD::CMOV) {
10360 APInt KZ2(KnownZero.getBitWidth(), 0);
10361 APInt KO2(KnownOne.getBitWidth(), 0);
10362 computeKnownBits(DAG, Op.getOperand(1), KnownZero, KnownOne);
10363 computeKnownBits(DAG, Op.getOperand(2), KZ2, KO2);
10369 return DAG.computeKnownBits(Op, KnownZero, KnownOne);
10372 SDValue ARMTargetLowering::PerformCMOVToBFICombine(SDNode *CMOV, SelectionDAG &DAG) const {
10373 // If we have a CMOV, OR and AND combination such as:
10378 // * CN is a single bit;
10379 // * All bits covered by CM are known zero in y
10381 // Then we can convert this into a sequence of BFI instructions. This will
10382 // always be a win if CM is a single bit, will always be no worse than the
10383 // TST&OR sequence if CM is two bits, and for thumb will be no worse if CM is
10384 // three bits (due to the extra IT instruction).
10386 SDValue Op0 = CMOV->getOperand(0);
10387 SDValue Op1 = CMOV->getOperand(1);
10388 auto CCNode = cast<ConstantSDNode>(CMOV->getOperand(2));
10389 auto CC = CCNode->getAPIntValue().getLimitedValue();
10390 SDValue CmpZ = CMOV->getOperand(4);
10392 // The compare must be against zero.
10393 SDValue Zero = CmpZ->getOperand(1);
10394 if (!isa<ConstantSDNode>(Zero.getNode()) ||
10395 !cast<ConstantSDNode>(Zero.getNode())->isNullValue())
10398 assert(CmpZ->getOpcode() == ARMISD::CMPZ);
10399 SDValue And = CmpZ->getOperand(0);
10400 if (And->getOpcode() != ISD::AND)
10402 ConstantSDNode *AndC = dyn_cast<ConstantSDNode>(And->getOperand(1));
10403 if (!AndC || !AndC->getAPIntValue().isPowerOf2())
10405 SDValue X = And->getOperand(0);
10407 if (CC == ARMCC::EQ) {
10408 // We're performing an "equal to zero" compare. Swap the operands so we
10409 // canonicalize on a "not equal to zero" compare.
10410 std::swap(Op0, Op1);
10412 assert(CC == ARMCC::NE && "How can a CMPZ node not be EQ or NE?");
10415 if (Op1->getOpcode() != ISD::OR)
10418 ConstantSDNode *OrC = dyn_cast<ConstantSDNode>(Op1->getOperand(1));
10421 SDValue Y = Op1->getOperand(0);
10426 // Now, is it profitable to continue?
10427 APInt OrCI = OrC->getAPIntValue();
10428 unsigned Heuristic = Subtarget->isThumb() ? 3 : 2;
10429 if (OrCI.countPopulation() > Heuristic)
10432 // Lastly, can we determine that the bits defined by OrCI
10434 APInt KnownZero, KnownOne;
10435 computeKnownBits(DAG, Y, KnownZero, KnownOne);
10436 if ((OrCI & KnownZero) != OrCI)
10439 // OK, we can do the combine.
10442 EVT VT = X.getValueType();
10443 unsigned BitInX = AndC->getAPIntValue().logBase2();
10446 // We must shift X first.
10447 X = DAG.getNode(ISD::SRL, dl, VT, X,
10448 DAG.getConstant(BitInX, dl, VT));
10451 for (unsigned BitInY = 0, NumActiveBits = OrCI.getActiveBits();
10452 BitInY < NumActiveBits; ++BitInY) {
10453 if (OrCI[BitInY] == 0)
10455 APInt Mask(VT.getSizeInBits(), 0);
10456 Mask.setBit(BitInY);
10457 V = DAG.getNode(ARMISD::BFI, dl, VT, V, X,
10458 // Confusingly, the operand is an *inverted* mask.
10459 DAG.getConstant(~Mask, dl, VT));
10465 /// PerformCMOVCombine - Target-specific DAG combining for ARMISD::CMOV.
10467 ARMTargetLowering::PerformCMOVCombine(SDNode *N, SelectionDAG &DAG) const {
10468 SDValue Cmp = N->getOperand(4);
10469 if (Cmp.getOpcode() != ARMISD::CMPZ)
10470 // Only looking at EQ and NE cases.
10473 EVT VT = N->getValueType(0);
10475 SDValue LHS = Cmp.getOperand(0);
10476 SDValue RHS = Cmp.getOperand(1);
10477 SDValue FalseVal = N->getOperand(0);
10478 SDValue TrueVal = N->getOperand(1);
10479 SDValue ARMcc = N->getOperand(2);
10480 ARMCC::CondCodes CC =
10481 (ARMCC::CondCodes)cast<ConstantSDNode>(ARMcc)->getZExtValue();
10483 // BFI is only available on V6T2+.
10484 if (!Subtarget->isThumb1Only() && Subtarget->hasV6T2Ops()) {
10485 SDValue R = PerformCMOVToBFICombine(N, DAG);
10506 /// FIXME: Turn this into a target neutral optimization?
10508 if (CC == ARMCC::NE && FalseVal == RHS && FalseVal != LHS) {
10509 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, TrueVal, ARMcc,
10510 N->getOperand(3), Cmp);
10511 } else if (CC == ARMCC::EQ && TrueVal == RHS) {
10513 SDValue NewCmp = getARMCmp(LHS, RHS, ISD::SETNE, ARMcc, DAG, dl);
10514 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, FalseVal, ARMcc,
10515 N->getOperand(3), NewCmp);
10518 if (Res.getNode()) {
10519 APInt KnownZero, KnownOne;
10520 DAG.computeKnownBits(SDValue(N,0), KnownZero, KnownOne);
10521 // Capture demanded bits information that would be otherwise lost.
10522 if (KnownZero == 0xfffffffe)
10523 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
10524 DAG.getValueType(MVT::i1));
10525 else if (KnownZero == 0xffffff00)
10526 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
10527 DAG.getValueType(MVT::i8));
10528 else if (KnownZero == 0xffff0000)
10529 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
10530 DAG.getValueType(MVT::i16));
10536 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
10537 DAGCombinerInfo &DCI) const {
10538 switch (N->getOpcode()) {
10540 case ISD::ADDC: return PerformADDCCombine(N, DCI, Subtarget);
10541 case ISD::ADD: return PerformADDCombine(N, DCI, Subtarget);
10542 case ISD::SUB: return PerformSUBCombine(N, DCI);
10543 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
10544 case ISD::OR: return PerformORCombine(N, DCI, Subtarget);
10545 case ISD::XOR: return PerformXORCombine(N, DCI, Subtarget);
10546 case ISD::AND: return PerformANDCombine(N, DCI, Subtarget);
10547 case ARMISD::BFI: return PerformBFICombine(N, DCI);
10548 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI, Subtarget);
10549 case ARMISD::VMOVDRR: return PerformVMOVDRRCombine(N, DCI.DAG);
10550 case ISD::STORE: return PerformSTORECombine(N, DCI);
10551 case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DCI, Subtarget);
10552 case ISD::INSERT_VECTOR_ELT: return PerformInsertEltCombine(N, DCI);
10553 case ISD::VECTOR_SHUFFLE: return PerformVECTOR_SHUFFLECombine(N, DCI.DAG);
10554 case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI);
10555 case ISD::FP_TO_SINT:
10556 case ISD::FP_TO_UINT:
10557 return PerformVCVTCombine(N, DCI.DAG, Subtarget);
10559 return PerformVDIVCombine(N, DCI.DAG, Subtarget);
10560 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
10563 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
10564 case ISD::SIGN_EXTEND:
10565 case ISD::ZERO_EXTEND:
10566 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
10567 case ARMISD::CMOV: return PerformCMOVCombine(N, DCI.DAG);
10568 case ISD::LOAD: return PerformLOADCombine(N, DCI);
10569 case ARMISD::VLD2DUP:
10570 case ARMISD::VLD3DUP:
10571 case ARMISD::VLD4DUP:
10572 return PerformVLDCombine(N, DCI);
10573 case ARMISD::BUILD_VECTOR:
10574 return PerformARMBUILD_VECTORCombine(N, DCI);
10575 case ISD::INTRINSIC_VOID:
10576 case ISD::INTRINSIC_W_CHAIN:
10577 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
10578 case Intrinsic::arm_neon_vld1:
10579 case Intrinsic::arm_neon_vld2:
10580 case Intrinsic::arm_neon_vld3:
10581 case Intrinsic::arm_neon_vld4:
10582 case Intrinsic::arm_neon_vld2lane:
10583 case Intrinsic::arm_neon_vld3lane:
10584 case Intrinsic::arm_neon_vld4lane:
10585 case Intrinsic::arm_neon_vst1:
10586 case Intrinsic::arm_neon_vst2:
10587 case Intrinsic::arm_neon_vst3:
10588 case Intrinsic::arm_neon_vst4:
10589 case Intrinsic::arm_neon_vst2lane:
10590 case Intrinsic::arm_neon_vst3lane:
10591 case Intrinsic::arm_neon_vst4lane:
10592 return PerformVLDCombine(N, DCI);
10600 bool ARMTargetLowering::isDesirableToTransformToIntegerOp(unsigned Opc,
10602 return (VT == MVT::f32) && (Opc == ISD::LOAD || Opc == ISD::STORE);
10605 bool ARMTargetLowering::allowsMisalignedMemoryAccesses(EVT VT,
10608 bool *Fast) const {
10609 // The AllowsUnaliged flag models the SCTLR.A setting in ARM cpus
10610 bool AllowsUnaligned = Subtarget->allowsUnalignedMem();
10612 switch (VT.getSimpleVT().SimpleTy) {
10618 // Unaligned access can use (for example) LRDB, LRDH, LDR
10619 if (AllowsUnaligned) {
10621 *Fast = Subtarget->hasV7Ops();
10628 // For any little-endian targets with neon, we can support unaligned ld/st
10629 // of D and Q (e.g. {D0,D1}) registers by using vld1.i8/vst1.i8.
10630 // A big-endian target may also explicitly support unaligned accesses
10631 if (Subtarget->hasNEON() && (AllowsUnaligned || Subtarget->isLittle())) {
10641 static bool memOpAlign(unsigned DstAlign, unsigned SrcAlign,
10642 unsigned AlignCheck) {
10643 return ((SrcAlign == 0 || SrcAlign % AlignCheck == 0) &&
10644 (DstAlign == 0 || DstAlign % AlignCheck == 0));
10647 EVT ARMTargetLowering::getOptimalMemOpType(uint64_t Size,
10648 unsigned DstAlign, unsigned SrcAlign,
10649 bool IsMemset, bool ZeroMemset,
10651 MachineFunction &MF) const {
10652 const Function *F = MF.getFunction();
10654 // See if we can use NEON instructions for this...
10655 if ((!IsMemset || ZeroMemset) && Subtarget->hasNEON() &&
10656 !F->hasFnAttribute(Attribute::NoImplicitFloat)) {
10659 (memOpAlign(SrcAlign, DstAlign, 16) ||
10660 (allowsMisalignedMemoryAccesses(MVT::v2f64, 0, 1, &Fast) && Fast))) {
10662 } else if (Size >= 8 &&
10663 (memOpAlign(SrcAlign, DstAlign, 8) ||
10664 (allowsMisalignedMemoryAccesses(MVT::f64, 0, 1, &Fast) &&
10670 // Lowering to i32/i16 if the size permits.
10673 else if (Size >= 2)
10676 // Let the target-independent logic figure it out.
10680 bool ARMTargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
10681 if (Val.getOpcode() != ISD::LOAD)
10684 EVT VT1 = Val.getValueType();
10685 if (!VT1.isSimple() || !VT1.isInteger() ||
10686 !VT2.isSimple() || !VT2.isInteger())
10689 switch (VT1.getSimpleVT().SimpleTy) {
10694 // 8-bit and 16-bit loads implicitly zero-extend to 32-bits.
10701 bool ARMTargetLowering::isVectorLoadExtDesirable(SDValue ExtVal) const {
10702 EVT VT = ExtVal.getValueType();
10704 if (!isTypeLegal(VT))
10707 // Don't create a loadext if we can fold the extension into a wide/long
10709 // If there's more than one user instruction, the loadext is desirable no
10710 // matter what. There can be two uses by the same instruction.
10711 if (ExtVal->use_empty() ||
10712 !ExtVal->use_begin()->isOnlyUserOf(ExtVal.getNode()))
10715 SDNode *U = *ExtVal->use_begin();
10716 if ((U->getOpcode() == ISD::ADD || U->getOpcode() == ISD::SUB ||
10717 U->getOpcode() == ISD::SHL || U->getOpcode() == ARMISD::VSHL))
10723 bool ARMTargetLowering::allowTruncateForTailCall(Type *Ty1, Type *Ty2) const {
10724 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
10727 if (!isTypeLegal(EVT::getEVT(Ty1)))
10730 assert(Ty1->getPrimitiveSizeInBits() <= 64 && "i128 is probably not a noop");
10732 // Assuming the caller doesn't have a zeroext or signext return parameter,
10733 // truncation all the way down to i1 is valid.
10738 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
10742 unsigned Scale = 1;
10743 switch (VT.getSimpleVT().SimpleTy) {
10744 default: return false;
10759 if ((V & (Scale - 1)) != 0)
10762 return V == (V & ((1LL << 5) - 1));
10765 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
10766 const ARMSubtarget *Subtarget) {
10767 bool isNeg = false;
10773 switch (VT.getSimpleVT().SimpleTy) {
10774 default: return false;
10779 // + imm12 or - imm8
10781 return V == (V & ((1LL << 8) - 1));
10782 return V == (V & ((1LL << 12) - 1));
10785 // Same as ARM mode. FIXME: NEON?
10786 if (!Subtarget->hasVFP2())
10791 return V == (V & ((1LL << 8) - 1));
10795 /// isLegalAddressImmediate - Return true if the integer value can be used
10796 /// as the offset of the target addressing mode for load / store of the
10798 static bool isLegalAddressImmediate(int64_t V, EVT VT,
10799 const ARMSubtarget *Subtarget) {
10803 if (!VT.isSimple())
10806 if (Subtarget->isThumb1Only())
10807 return isLegalT1AddressImmediate(V, VT);
10808 else if (Subtarget->isThumb2())
10809 return isLegalT2AddressImmediate(V, VT, Subtarget);
10814 switch (VT.getSimpleVT().SimpleTy) {
10815 default: return false;
10820 return V == (V & ((1LL << 12) - 1));
10823 return V == (V & ((1LL << 8) - 1));
10826 if (!Subtarget->hasVFP2()) // FIXME: NEON?
10831 return V == (V & ((1LL << 8) - 1));
10835 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
10837 int Scale = AM.Scale;
10841 switch (VT.getSimpleVT().SimpleTy) {
10842 default: return false;
10850 Scale = Scale & ~1;
10851 return Scale == 2 || Scale == 4 || Scale == 8;
10854 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
10858 // Note, we allow "void" uses (basically, uses that aren't loads or
10859 // stores), because arm allows folding a scale into many arithmetic
10860 // operations. This should be made more precise and revisited later.
10862 // Allow r << imm, but the imm has to be a multiple of two.
10863 if (Scale & 1) return false;
10864 return isPowerOf2_32(Scale);
10868 /// isLegalAddressingMode - Return true if the addressing mode represented
10869 /// by AM is legal for this target, for a load/store of the specified type.
10870 bool ARMTargetLowering::isLegalAddressingMode(const DataLayout &DL,
10871 const AddrMode &AM, Type *Ty,
10872 unsigned AS) const {
10873 EVT VT = getValueType(DL, Ty, true);
10874 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
10877 // Can never fold addr of global into load/store.
10881 switch (AM.Scale) {
10882 case 0: // no scale reg, must be "r+i" or "r", or "i".
10885 if (Subtarget->isThumb1Only())
10889 // ARM doesn't support any R+R*scale+imm addr modes.
10893 if (!VT.isSimple())
10896 if (Subtarget->isThumb2())
10897 return isLegalT2ScaledAddressingMode(AM, VT);
10899 int Scale = AM.Scale;
10900 switch (VT.getSimpleVT().SimpleTy) {
10901 default: return false;
10905 if (Scale < 0) Scale = -Scale;
10909 return isPowerOf2_32(Scale & ~1);
10913 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
10918 // Note, we allow "void" uses (basically, uses that aren't loads or
10919 // stores), because arm allows folding a scale into many arithmetic
10920 // operations. This should be made more precise and revisited later.
10922 // Allow r << imm, but the imm has to be a multiple of two.
10923 if (Scale & 1) return false;
10924 return isPowerOf2_32(Scale);
10930 /// isLegalICmpImmediate - Return true if the specified immediate is legal
10931 /// icmp immediate, that is the target has icmp instructions which can compare
10932 /// a register against the immediate without having to materialize the
10933 /// immediate into a register.
10934 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
10935 // Thumb2 and ARM modes can use cmn for negative immediates.
10936 if (!Subtarget->isThumb())
10937 return ARM_AM::getSOImmVal(std::abs(Imm)) != -1;
10938 if (Subtarget->isThumb2())
10939 return ARM_AM::getT2SOImmVal(std::abs(Imm)) != -1;
10940 // Thumb1 doesn't have cmn, and only 8-bit immediates.
10941 return Imm >= 0 && Imm <= 255;
10944 /// isLegalAddImmediate - Return true if the specified immediate is a legal add
10945 /// *or sub* immediate, that is the target has add or sub instructions which can
10946 /// add a register with the immediate without having to materialize the
10947 /// immediate into a register.
10948 bool ARMTargetLowering::isLegalAddImmediate(int64_t Imm) const {
10949 // Same encoding for add/sub, just flip the sign.
10950 int64_t AbsImm = std::abs(Imm);
10951 if (!Subtarget->isThumb())
10952 return ARM_AM::getSOImmVal(AbsImm) != -1;
10953 if (Subtarget->isThumb2())
10954 return ARM_AM::getT2SOImmVal(AbsImm) != -1;
10955 // Thumb1 only has 8-bit unsigned immediate.
10956 return AbsImm >= 0 && AbsImm <= 255;
10959 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
10960 bool isSEXTLoad, SDValue &Base,
10961 SDValue &Offset, bool &isInc,
10962 SelectionDAG &DAG) {
10963 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
10966 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
10967 // AddressingMode 3
10968 Base = Ptr->getOperand(0);
10969 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
10970 int RHSC = (int)RHS->getZExtValue();
10971 if (RHSC < 0 && RHSC > -256) {
10972 assert(Ptr->getOpcode() == ISD::ADD);
10974 Offset = DAG.getConstant(-RHSC, SDLoc(Ptr), RHS->getValueType(0));
10978 isInc = (Ptr->getOpcode() == ISD::ADD);
10979 Offset = Ptr->getOperand(1);
10981 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
10982 // AddressingMode 2
10983 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
10984 int RHSC = (int)RHS->getZExtValue();
10985 if (RHSC < 0 && RHSC > -0x1000) {
10986 assert(Ptr->getOpcode() == ISD::ADD);
10988 Offset = DAG.getConstant(-RHSC, SDLoc(Ptr), RHS->getValueType(0));
10989 Base = Ptr->getOperand(0);
10994 if (Ptr->getOpcode() == ISD::ADD) {
10996 ARM_AM::ShiftOpc ShOpcVal=
10997 ARM_AM::getShiftOpcForNode(Ptr->getOperand(0).getOpcode());
10998 if (ShOpcVal != ARM_AM::no_shift) {
10999 Base = Ptr->getOperand(1);
11000 Offset = Ptr->getOperand(0);
11002 Base = Ptr->getOperand(0);
11003 Offset = Ptr->getOperand(1);
11008 isInc = (Ptr->getOpcode() == ISD::ADD);
11009 Base = Ptr->getOperand(0);
11010 Offset = Ptr->getOperand(1);
11014 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
11018 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
11019 bool isSEXTLoad, SDValue &Base,
11020 SDValue &Offset, bool &isInc,
11021 SelectionDAG &DAG) {
11022 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
11025 Base = Ptr->getOperand(0);
11026 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
11027 int RHSC = (int)RHS->getZExtValue();
11028 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
11029 assert(Ptr->getOpcode() == ISD::ADD);
11031 Offset = DAG.getConstant(-RHSC, SDLoc(Ptr), RHS->getValueType(0));
11033 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
11034 isInc = Ptr->getOpcode() == ISD::ADD;
11035 Offset = DAG.getConstant(RHSC, SDLoc(Ptr), RHS->getValueType(0));
11043 /// getPreIndexedAddressParts - returns true by value, base pointer and
11044 /// offset pointer and addressing mode by reference if the node's address
11045 /// can be legally represented as pre-indexed load / store address.
11047 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
11049 ISD::MemIndexedMode &AM,
11050 SelectionDAG &DAG) const {
11051 if (Subtarget->isThumb1Only())
11056 bool isSEXTLoad = false;
11057 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
11058 Ptr = LD->getBasePtr();
11059 VT = LD->getMemoryVT();
11060 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
11061 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
11062 Ptr = ST->getBasePtr();
11063 VT = ST->getMemoryVT();
11068 bool isLegal = false;
11069 if (Subtarget->isThumb2())
11070 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
11071 Offset, isInc, DAG);
11073 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
11074 Offset, isInc, DAG);
11078 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
11082 /// getPostIndexedAddressParts - returns true by value, base pointer and
11083 /// offset pointer and addressing mode by reference if this node can be
11084 /// combined with a load / store to form a post-indexed load / store.
11085 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
11088 ISD::MemIndexedMode &AM,
11089 SelectionDAG &DAG) const {
11090 if (Subtarget->isThumb1Only())
11095 bool isSEXTLoad = false;
11096 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
11097 VT = LD->getMemoryVT();
11098 Ptr = LD->getBasePtr();
11099 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
11100 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
11101 VT = ST->getMemoryVT();
11102 Ptr = ST->getBasePtr();
11107 bool isLegal = false;
11108 if (Subtarget->isThumb2())
11109 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
11112 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
11118 // Swap base ptr and offset to catch more post-index load / store when
11119 // it's legal. In Thumb2 mode, offset must be an immediate.
11120 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
11121 !Subtarget->isThumb2())
11122 std::swap(Base, Offset);
11124 // Post-indexed load / store update the base pointer.
11129 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
11133 void ARMTargetLowering::computeKnownBitsForTargetNode(const SDValue Op,
11136 const SelectionDAG &DAG,
11137 unsigned Depth) const {
11138 unsigned BitWidth = KnownOne.getBitWidth();
11139 KnownZero = KnownOne = APInt(BitWidth, 0);
11140 switch (Op.getOpcode()) {
11146 // These nodes' second result is a boolean
11147 if (Op.getResNo() == 0)
11149 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - 1);
11151 case ARMISD::CMOV: {
11152 // Bits are known zero/one if known on the LHS and RHS.
11153 DAG.computeKnownBits(Op.getOperand(0), KnownZero, KnownOne, Depth+1);
11154 if (KnownZero == 0 && KnownOne == 0) return;
11156 APInt KnownZeroRHS, KnownOneRHS;
11157 DAG.computeKnownBits(Op.getOperand(1), KnownZeroRHS, KnownOneRHS, Depth+1);
11158 KnownZero &= KnownZeroRHS;
11159 KnownOne &= KnownOneRHS;
11162 case ISD::INTRINSIC_W_CHAIN: {
11163 ConstantSDNode *CN = cast<ConstantSDNode>(Op->getOperand(1));
11164 Intrinsic::ID IntID = static_cast<Intrinsic::ID>(CN->getZExtValue());
11167 case Intrinsic::arm_ldaex:
11168 case Intrinsic::arm_ldrex: {
11169 EVT VT = cast<MemIntrinsicSDNode>(Op)->getMemoryVT();
11170 unsigned MemBits = VT.getScalarType().getSizeInBits();
11171 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - MemBits);
11179 //===----------------------------------------------------------------------===//
11180 // ARM Inline Assembly Support
11181 //===----------------------------------------------------------------------===//
11183 bool ARMTargetLowering::ExpandInlineAsm(CallInst *CI) const {
11184 // Looking for "rev" which is V6+.
11185 if (!Subtarget->hasV6Ops())
11188 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
11189 std::string AsmStr = IA->getAsmString();
11190 SmallVector<StringRef, 4> AsmPieces;
11191 SplitString(AsmStr, AsmPieces, ";\n");
11193 switch (AsmPieces.size()) {
11194 default: return false;
11196 AsmStr = AsmPieces[0];
11198 SplitString(AsmStr, AsmPieces, " \t,");
11201 if (AsmPieces.size() == 3 &&
11202 AsmPieces[0] == "rev" && AsmPieces[1] == "$0" && AsmPieces[2] == "$1" &&
11203 IA->getConstraintString().compare(0, 4, "=l,l") == 0) {
11204 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
11205 if (Ty && Ty->getBitWidth() == 32)
11206 return IntrinsicLowering::LowerToByteSwap(CI);
11214 /// getConstraintType - Given a constraint letter, return the type of
11215 /// constraint it is for this target.
11216 ARMTargetLowering::ConstraintType
11217 ARMTargetLowering::getConstraintType(StringRef Constraint) const {
11218 if (Constraint.size() == 1) {
11219 switch (Constraint[0]) {
11221 case 'l': return C_RegisterClass;
11222 case 'w': return C_RegisterClass;
11223 case 'h': return C_RegisterClass;
11224 case 'x': return C_RegisterClass;
11225 case 't': return C_RegisterClass;
11226 case 'j': return C_Other; // Constant for movw.
11227 // An address with a single base register. Due to the way we
11228 // currently handle addresses it is the same as an 'r' memory constraint.
11229 case 'Q': return C_Memory;
11231 } else if (Constraint.size() == 2) {
11232 switch (Constraint[0]) {
11234 // All 'U+' constraints are addresses.
11235 case 'U': return C_Memory;
11238 return TargetLowering::getConstraintType(Constraint);
11241 /// Examine constraint type and operand type and determine a weight value.
11242 /// This object must already have been set up with the operand type
11243 /// and the current alternative constraint selected.
11244 TargetLowering::ConstraintWeight
11245 ARMTargetLowering::getSingleConstraintMatchWeight(
11246 AsmOperandInfo &info, const char *constraint) const {
11247 ConstraintWeight weight = CW_Invalid;
11248 Value *CallOperandVal = info.CallOperandVal;
11249 // If we don't have a value, we can't do a match,
11250 // but allow it at the lowest weight.
11251 if (!CallOperandVal)
11253 Type *type = CallOperandVal->getType();
11254 // Look at the constraint type.
11255 switch (*constraint) {
11257 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
11260 if (type->isIntegerTy()) {
11261 if (Subtarget->isThumb())
11262 weight = CW_SpecificReg;
11264 weight = CW_Register;
11268 if (type->isFloatingPointTy())
11269 weight = CW_Register;
11275 typedef std::pair<unsigned, const TargetRegisterClass*> RCPair;
11276 RCPair ARMTargetLowering::getRegForInlineAsmConstraint(
11277 const TargetRegisterInfo *TRI, StringRef Constraint, MVT VT) const {
11278 if (Constraint.size() == 1) {
11279 // GCC ARM Constraint Letters
11280 switch (Constraint[0]) {
11281 case 'l': // Low regs or general regs.
11282 if (Subtarget->isThumb())
11283 return RCPair(0U, &ARM::tGPRRegClass);
11284 return RCPair(0U, &ARM::GPRRegClass);
11285 case 'h': // High regs or no regs.
11286 if (Subtarget->isThumb())
11287 return RCPair(0U, &ARM::hGPRRegClass);
11290 if (Subtarget->isThumb1Only())
11291 return RCPair(0U, &ARM::tGPRRegClass);
11292 return RCPair(0U, &ARM::GPRRegClass);
11294 if (VT == MVT::Other)
11296 if (VT == MVT::f32)
11297 return RCPair(0U, &ARM::SPRRegClass);
11298 if (VT.getSizeInBits() == 64)
11299 return RCPair(0U, &ARM::DPRRegClass);
11300 if (VT.getSizeInBits() == 128)
11301 return RCPair(0U, &ARM::QPRRegClass);
11304 if (VT == MVT::Other)
11306 if (VT == MVT::f32)
11307 return RCPair(0U, &ARM::SPR_8RegClass);
11308 if (VT.getSizeInBits() == 64)
11309 return RCPair(0U, &ARM::DPR_8RegClass);
11310 if (VT.getSizeInBits() == 128)
11311 return RCPair(0U, &ARM::QPR_8RegClass);
11314 if (VT == MVT::f32)
11315 return RCPair(0U, &ARM::SPRRegClass);
11319 if (StringRef("{cc}").equals_lower(Constraint))
11320 return std::make_pair(unsigned(ARM::CPSR), &ARM::CCRRegClass);
11322 return TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
11325 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
11326 /// vector. If it is invalid, don't add anything to Ops.
11327 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
11328 std::string &Constraint,
11329 std::vector<SDValue>&Ops,
11330 SelectionDAG &DAG) const {
11333 // Currently only support length 1 constraints.
11334 if (Constraint.length() != 1) return;
11336 char ConstraintLetter = Constraint[0];
11337 switch (ConstraintLetter) {
11340 case 'I': case 'J': case 'K': case 'L':
11341 case 'M': case 'N': case 'O':
11342 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
11346 int64_t CVal64 = C->getSExtValue();
11347 int CVal = (int) CVal64;
11348 // None of these constraints allow values larger than 32 bits. Check
11349 // that the value fits in an int.
11350 if (CVal != CVal64)
11353 switch (ConstraintLetter) {
11355 // Constant suitable for movw, must be between 0 and
11357 if (Subtarget->hasV6T2Ops())
11358 if (CVal >= 0 && CVal <= 65535)
11362 if (Subtarget->isThumb1Only()) {
11363 // This must be a constant between 0 and 255, for ADD
11365 if (CVal >= 0 && CVal <= 255)
11367 } else if (Subtarget->isThumb2()) {
11368 // A constant that can be used as an immediate value in a
11369 // data-processing instruction.
11370 if (ARM_AM::getT2SOImmVal(CVal) != -1)
11373 // A constant that can be used as an immediate value in a
11374 // data-processing instruction.
11375 if (ARM_AM::getSOImmVal(CVal) != -1)
11381 if (Subtarget->isThumb()) { // FIXME thumb2
11382 // This must be a constant between -255 and -1, for negated ADD
11383 // immediates. This can be used in GCC with an "n" modifier that
11384 // prints the negated value, for use with SUB instructions. It is
11385 // not useful otherwise but is implemented for compatibility.
11386 if (CVal >= -255 && CVal <= -1)
11389 // This must be a constant between -4095 and 4095. It is not clear
11390 // what this constraint is intended for. Implemented for
11391 // compatibility with GCC.
11392 if (CVal >= -4095 && CVal <= 4095)
11398 if (Subtarget->isThumb1Only()) {
11399 // A 32-bit value where only one byte has a nonzero value. Exclude
11400 // zero to match GCC. This constraint is used by GCC internally for
11401 // constants that can be loaded with a move/shift combination.
11402 // It is not useful otherwise but is implemented for compatibility.
11403 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
11405 } else if (Subtarget->isThumb2()) {
11406 // A constant whose bitwise inverse can be used as an immediate
11407 // value in a data-processing instruction. This can be used in GCC
11408 // with a "B" modifier that prints the inverted value, for use with
11409 // BIC and MVN instructions. It is not useful otherwise but is
11410 // implemented for compatibility.
11411 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
11414 // A constant whose bitwise inverse can be used as an immediate
11415 // value in a data-processing instruction. This can be used in GCC
11416 // with a "B" modifier that prints the inverted value, for use with
11417 // BIC and MVN instructions. It is not useful otherwise but is
11418 // implemented for compatibility.
11419 if (ARM_AM::getSOImmVal(~CVal) != -1)
11425 if (Subtarget->isThumb1Only()) {
11426 // This must be a constant between -7 and 7,
11427 // for 3-operand ADD/SUB immediate instructions.
11428 if (CVal >= -7 && CVal < 7)
11430 } else if (Subtarget->isThumb2()) {
11431 // A constant whose negation can be used as an immediate value in a
11432 // data-processing instruction. This can be used in GCC with an "n"
11433 // modifier that prints the negated value, for use with SUB
11434 // instructions. It is not useful otherwise but is implemented for
11436 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
11439 // A constant whose negation can be used as an immediate value in a
11440 // data-processing instruction. This can be used in GCC with an "n"
11441 // modifier that prints the negated value, for use with SUB
11442 // instructions. It is not useful otherwise but is implemented for
11444 if (ARM_AM::getSOImmVal(-CVal) != -1)
11450 if (Subtarget->isThumb()) { // FIXME thumb2
11451 // This must be a multiple of 4 between 0 and 1020, for
11452 // ADD sp + immediate.
11453 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
11456 // A power of two or a constant between 0 and 32. This is used in
11457 // GCC for the shift amount on shifted register operands, but it is
11458 // useful in general for any shift amounts.
11459 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
11465 if (Subtarget->isThumb()) { // FIXME thumb2
11466 // This must be a constant between 0 and 31, for shift amounts.
11467 if (CVal >= 0 && CVal <= 31)
11473 if (Subtarget->isThumb()) { // FIXME thumb2
11474 // This must be a multiple of 4 between -508 and 508, for
11475 // ADD/SUB sp = sp + immediate.
11476 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
11481 Result = DAG.getTargetConstant(CVal, SDLoc(Op), Op.getValueType());
11485 if (Result.getNode()) {
11486 Ops.push_back(Result);
11489 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
11492 static RTLIB::Libcall getDivRemLibcall(
11493 const SDNode *N, MVT::SimpleValueType SVT) {
11494 assert((N->getOpcode() == ISD::SDIVREM || N->getOpcode() == ISD::UDIVREM ||
11495 N->getOpcode() == ISD::SREM || N->getOpcode() == ISD::UREM) &&
11496 "Unhandled Opcode in getDivRemLibcall");
11497 bool isSigned = N->getOpcode() == ISD::SDIVREM ||
11498 N->getOpcode() == ISD::SREM;
11501 default: llvm_unreachable("Unexpected request for libcall!");
11502 case MVT::i8: LC = isSigned ? RTLIB::SDIVREM_I8 : RTLIB::UDIVREM_I8; break;
11503 case MVT::i16: LC = isSigned ? RTLIB::SDIVREM_I16 : RTLIB::UDIVREM_I16; break;
11504 case MVT::i32: LC = isSigned ? RTLIB::SDIVREM_I32 : RTLIB::UDIVREM_I32; break;
11505 case MVT::i64: LC = isSigned ? RTLIB::SDIVREM_I64 : RTLIB::UDIVREM_I64; break;
11510 static TargetLowering::ArgListTy getDivRemArgList(
11511 const SDNode *N, LLVMContext *Context) {
11512 assert((N->getOpcode() == ISD::SDIVREM || N->getOpcode() == ISD::UDIVREM ||
11513 N->getOpcode() == ISD::SREM || N->getOpcode() == ISD::UREM) &&
11514 "Unhandled Opcode in getDivRemArgList");
11515 bool isSigned = N->getOpcode() == ISD::SDIVREM ||
11516 N->getOpcode() == ISD::SREM;
11517 TargetLowering::ArgListTy Args;
11518 TargetLowering::ArgListEntry Entry;
11519 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
11520 EVT ArgVT = N->getOperand(i).getValueType();
11521 Type *ArgTy = ArgVT.getTypeForEVT(*Context);
11522 Entry.Node = N->getOperand(i);
11524 Entry.isSExt = isSigned;
11525 Entry.isZExt = !isSigned;
11526 Args.push_back(Entry);
11531 SDValue ARMTargetLowering::LowerDivRem(SDValue Op, SelectionDAG &DAG) const {
11532 assert((Subtarget->isTargetAEABI() || Subtarget->isTargetAndroid()) &&
11533 "Register-based DivRem lowering only");
11534 unsigned Opcode = Op->getOpcode();
11535 assert((Opcode == ISD::SDIVREM || Opcode == ISD::UDIVREM) &&
11536 "Invalid opcode for Div/Rem lowering");
11537 bool isSigned = (Opcode == ISD::SDIVREM);
11538 EVT VT = Op->getValueType(0);
11539 Type *Ty = VT.getTypeForEVT(*DAG.getContext());
11541 RTLIB::Libcall LC = getDivRemLibcall(Op.getNode(),
11542 VT.getSimpleVT().SimpleTy);
11543 SDValue InChain = DAG.getEntryNode();
11545 TargetLowering::ArgListTy Args = getDivRemArgList(Op.getNode(),
11548 SDValue Callee = DAG.getExternalSymbol(getLibcallName(LC),
11549 getPointerTy(DAG.getDataLayout()));
11551 Type *RetTy = (Type*)StructType::get(Ty, Ty, nullptr);
11554 TargetLowering::CallLoweringInfo CLI(DAG);
11555 CLI.setDebugLoc(dl).setChain(InChain)
11556 .setCallee(getLibcallCallingConv(LC), RetTy, Callee, std::move(Args), 0)
11557 .setInRegister().setSExtResult(isSigned).setZExtResult(!isSigned);
11559 std::pair<SDValue, SDValue> CallInfo = LowerCallTo(CLI);
11560 return CallInfo.first;
11563 // Lowers REM using divmod helpers
11564 // see RTABI section 4.2/4.3
11565 SDValue ARMTargetLowering::LowerREM(SDNode *N, SelectionDAG &DAG) const {
11566 // Build return types (div and rem)
11567 std::vector<Type*> RetTyParams;
11568 Type *RetTyElement;
11570 switch (N->getValueType(0).getSimpleVT().SimpleTy) {
11571 default: llvm_unreachable("Unexpected request for libcall!");
11572 case MVT::i8: RetTyElement = Type::getInt8Ty(*DAG.getContext()); break;
11573 case MVT::i16: RetTyElement = Type::getInt16Ty(*DAG.getContext()); break;
11574 case MVT::i32: RetTyElement = Type::getInt32Ty(*DAG.getContext()); break;
11575 case MVT::i64: RetTyElement = Type::getInt64Ty(*DAG.getContext()); break;
11578 RetTyParams.push_back(RetTyElement);
11579 RetTyParams.push_back(RetTyElement);
11580 ArrayRef<Type*> ret = ArrayRef<Type*>(RetTyParams);
11581 Type *RetTy = StructType::get(*DAG.getContext(), ret);
11583 RTLIB::Libcall LC = getDivRemLibcall(N, N->getValueType(0).getSimpleVT().
11585 SDValue InChain = DAG.getEntryNode();
11586 TargetLowering::ArgListTy Args = getDivRemArgList(N, DAG.getContext());
11587 bool isSigned = N->getOpcode() == ISD::SREM;
11588 SDValue Callee = DAG.getExternalSymbol(getLibcallName(LC),
11589 getPointerTy(DAG.getDataLayout()));
11592 CallLoweringInfo CLI(DAG);
11593 CLI.setChain(InChain)
11594 .setCallee(CallingConv::ARM_AAPCS, RetTy, Callee, std::move(Args), 0)
11595 .setSExtResult(isSigned).setZExtResult(!isSigned).setDebugLoc(SDLoc(N));
11596 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
11598 // Return second (rem) result operand (first contains div)
11599 SDNode *ResNode = CallResult.first.getNode();
11600 assert(ResNode->getNumOperands() == 2 && "divmod should return two operands");
11601 return ResNode->getOperand(1);
11605 ARMTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const {
11606 assert(Subtarget->isTargetWindows() && "unsupported target platform");
11610 SDValue Chain = Op.getOperand(0);
11611 SDValue Size = Op.getOperand(1);
11613 SDValue Words = DAG.getNode(ISD::SRL, DL, MVT::i32, Size,
11614 DAG.getConstant(2, DL, MVT::i32));
11617 Chain = DAG.getCopyToReg(Chain, DL, ARM::R4, Words, Flag);
11618 Flag = Chain.getValue(1);
11620 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
11621 Chain = DAG.getNode(ARMISD::WIN__CHKSTK, DL, NodeTys, Chain, Flag);
11623 SDValue NewSP = DAG.getCopyFromReg(Chain, DL, ARM::SP, MVT::i32);
11624 Chain = NewSP.getValue(1);
11626 SDValue Ops[2] = { NewSP, Chain };
11627 return DAG.getMergeValues(Ops, DL);
11630 SDValue ARMTargetLowering::LowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) const {
11631 assert(Op.getValueType() == MVT::f64 && Subtarget->isFPOnlySP() &&
11632 "Unexpected type for custom-lowering FP_EXTEND");
11635 LC = RTLIB::getFPEXT(Op.getOperand(0).getValueType(), Op.getValueType());
11637 SDValue SrcVal = Op.getOperand(0);
11638 return makeLibCall(DAG, LC, Op.getValueType(), SrcVal, /*isSigned*/ false,
11642 SDValue ARMTargetLowering::LowerFP_ROUND(SDValue Op, SelectionDAG &DAG) const {
11643 assert(Op.getOperand(0).getValueType() == MVT::f64 &&
11644 Subtarget->isFPOnlySP() &&
11645 "Unexpected type for custom-lowering FP_ROUND");
11648 LC = RTLIB::getFPROUND(Op.getOperand(0).getValueType(), Op.getValueType());
11650 SDValue SrcVal = Op.getOperand(0);
11651 return makeLibCall(DAG, LC, Op.getValueType(), SrcVal, /*isSigned*/ false,
11656 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
11657 // The ARM target isn't yet aware of offsets.
11661 bool ARM::isBitFieldInvertedMask(unsigned v) {
11662 if (v == 0xffffffff)
11665 // there can be 1's on either or both "outsides", all the "inside"
11666 // bits must be 0's
11667 return isShiftedMask_32(~v);
11670 /// isFPImmLegal - Returns true if the target can instruction select the
11671 /// specified FP immediate natively. If false, the legalizer will
11672 /// materialize the FP immediate as a load from a constant pool.
11673 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
11674 if (!Subtarget->hasVFP3())
11676 if (VT == MVT::f32)
11677 return ARM_AM::getFP32Imm(Imm) != -1;
11678 if (VT == MVT::f64 && !Subtarget->isFPOnlySP())
11679 return ARM_AM::getFP64Imm(Imm) != -1;
11683 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
11684 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
11685 /// specified in the intrinsic calls.
11686 bool ARMTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
11688 unsigned Intrinsic) const {
11689 switch (Intrinsic) {
11690 case Intrinsic::arm_neon_vld1:
11691 case Intrinsic::arm_neon_vld2:
11692 case Intrinsic::arm_neon_vld3:
11693 case Intrinsic::arm_neon_vld4:
11694 case Intrinsic::arm_neon_vld2lane:
11695 case Intrinsic::arm_neon_vld3lane:
11696 case Intrinsic::arm_neon_vld4lane: {
11697 Info.opc = ISD::INTRINSIC_W_CHAIN;
11698 // Conservatively set memVT to the entire set of vectors loaded.
11699 auto &DL = I.getCalledFunction()->getParent()->getDataLayout();
11700 uint64_t NumElts = DL.getTypeAllocSize(I.getType()) / 8;
11701 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
11702 Info.ptrVal = I.getArgOperand(0);
11704 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
11705 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
11706 Info.vol = false; // volatile loads with NEON intrinsics not supported
11707 Info.readMem = true;
11708 Info.writeMem = false;
11711 case Intrinsic::arm_neon_vst1:
11712 case Intrinsic::arm_neon_vst2:
11713 case Intrinsic::arm_neon_vst3:
11714 case Intrinsic::arm_neon_vst4:
11715 case Intrinsic::arm_neon_vst2lane:
11716 case Intrinsic::arm_neon_vst3lane:
11717 case Intrinsic::arm_neon_vst4lane: {
11718 Info.opc = ISD::INTRINSIC_VOID;
11719 // Conservatively set memVT to the entire set of vectors stored.
11720 auto &DL = I.getCalledFunction()->getParent()->getDataLayout();
11721 unsigned NumElts = 0;
11722 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
11723 Type *ArgTy = I.getArgOperand(ArgI)->getType();
11724 if (!ArgTy->isVectorTy())
11726 NumElts += DL.getTypeAllocSize(ArgTy) / 8;
11728 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
11729 Info.ptrVal = I.getArgOperand(0);
11731 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
11732 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
11733 Info.vol = false; // volatile stores with NEON intrinsics not supported
11734 Info.readMem = false;
11735 Info.writeMem = true;
11738 case Intrinsic::arm_ldaex:
11739 case Intrinsic::arm_ldrex: {
11740 auto &DL = I.getCalledFunction()->getParent()->getDataLayout();
11741 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(0)->getType());
11742 Info.opc = ISD::INTRINSIC_W_CHAIN;
11743 Info.memVT = MVT::getVT(PtrTy->getElementType());
11744 Info.ptrVal = I.getArgOperand(0);
11746 Info.align = DL.getABITypeAlignment(PtrTy->getElementType());
11748 Info.readMem = true;
11749 Info.writeMem = false;
11752 case Intrinsic::arm_stlex:
11753 case Intrinsic::arm_strex: {
11754 auto &DL = I.getCalledFunction()->getParent()->getDataLayout();
11755 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(1)->getType());
11756 Info.opc = ISD::INTRINSIC_W_CHAIN;
11757 Info.memVT = MVT::getVT(PtrTy->getElementType());
11758 Info.ptrVal = I.getArgOperand(1);
11760 Info.align = DL.getABITypeAlignment(PtrTy->getElementType());
11762 Info.readMem = false;
11763 Info.writeMem = true;
11766 case Intrinsic::arm_stlexd:
11767 case Intrinsic::arm_strexd: {
11768 Info.opc = ISD::INTRINSIC_W_CHAIN;
11769 Info.memVT = MVT::i64;
11770 Info.ptrVal = I.getArgOperand(2);
11774 Info.readMem = false;
11775 Info.writeMem = true;
11778 case Intrinsic::arm_ldaexd:
11779 case Intrinsic::arm_ldrexd: {
11780 Info.opc = ISD::INTRINSIC_W_CHAIN;
11781 Info.memVT = MVT::i64;
11782 Info.ptrVal = I.getArgOperand(0);
11786 Info.readMem = true;
11787 Info.writeMem = false;
11797 /// \brief Returns true if it is beneficial to convert a load of a constant
11798 /// to just the constant itself.
11799 bool ARMTargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
11801 assert(Ty->isIntegerTy());
11803 unsigned Bits = Ty->getPrimitiveSizeInBits();
11804 if (Bits == 0 || Bits > 32)
11809 Instruction* ARMTargetLowering::makeDMB(IRBuilder<> &Builder,
11810 ARM_MB::MemBOpt Domain) const {
11811 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
11813 // First, if the target has no DMB, see what fallback we can use.
11814 if (!Subtarget->hasDataBarrier()) {
11815 // Some ARMv6 cpus can support data barriers with an mcr instruction.
11816 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
11818 if (Subtarget->hasV6Ops() && !Subtarget->isThumb()) {
11819 Function *MCR = llvm::Intrinsic::getDeclaration(M, Intrinsic::arm_mcr);
11820 Value* args[6] = {Builder.getInt32(15), Builder.getInt32(0),
11821 Builder.getInt32(0), Builder.getInt32(7),
11822 Builder.getInt32(10), Builder.getInt32(5)};
11823 return Builder.CreateCall(MCR, args);
11825 // Instead of using barriers, atomic accesses on these subtargets use
11827 llvm_unreachable("makeDMB on a target so old that it has no barriers");
11830 Function *DMB = llvm::Intrinsic::getDeclaration(M, Intrinsic::arm_dmb);
11831 // Only a full system barrier exists in the M-class architectures.
11832 Domain = Subtarget->isMClass() ? ARM_MB::SY : Domain;
11833 Constant *CDomain = Builder.getInt32(Domain);
11834 return Builder.CreateCall(DMB, CDomain);
11838 // Based on http://www.cl.cam.ac.uk/~pes20/cpp/cpp0xmappings.html
11839 Instruction* ARMTargetLowering::emitLeadingFence(IRBuilder<> &Builder,
11840 AtomicOrdering Ord, bool IsStore,
11841 bool IsLoad) const {
11842 if (!getInsertFencesForAtomic())
11848 llvm_unreachable("Invalid fence: unordered/non-atomic");
11851 return nullptr; // Nothing to do
11852 case SequentiallyConsistent:
11854 return nullptr; // Nothing to do
11857 case AcquireRelease:
11858 if (Subtarget->isSwift())
11859 return makeDMB(Builder, ARM_MB::ISHST);
11860 // FIXME: add a comment with a link to documentation justifying this.
11862 return makeDMB(Builder, ARM_MB::ISH);
11864 llvm_unreachable("Unknown fence ordering in emitLeadingFence");
11867 Instruction* ARMTargetLowering::emitTrailingFence(IRBuilder<> &Builder,
11868 AtomicOrdering Ord, bool IsStore,
11869 bool IsLoad) const {
11870 if (!getInsertFencesForAtomic())
11876 llvm_unreachable("Invalid fence: unordered/not-atomic");
11879 return nullptr; // Nothing to do
11881 case AcquireRelease:
11882 case SequentiallyConsistent:
11883 return makeDMB(Builder, ARM_MB::ISH);
11885 llvm_unreachable("Unknown fence ordering in emitTrailingFence");
11888 // Loads and stores less than 64-bits are already atomic; ones above that
11889 // are doomed anyway, so defer to the default libcall and blame the OS when
11890 // things go wrong. Cortex M doesn't have ldrexd/strexd though, so don't emit
11891 // anything for those.
11892 bool ARMTargetLowering::shouldExpandAtomicStoreInIR(StoreInst *SI) const {
11893 unsigned Size = SI->getValueOperand()->getType()->getPrimitiveSizeInBits();
11894 return (Size == 64) && !Subtarget->isMClass();
11897 // Loads and stores less than 64-bits are already atomic; ones above that
11898 // are doomed anyway, so defer to the default libcall and blame the OS when
11899 // things go wrong. Cortex M doesn't have ldrexd/strexd though, so don't emit
11900 // anything for those.
11901 // FIXME: ldrd and strd are atomic if the CPU has LPAE (e.g. A15 has that
11902 // guarantee, see DDI0406C ARM architecture reference manual,
11903 // sections A8.8.72-74 LDRD)
11904 TargetLowering::AtomicExpansionKind
11905 ARMTargetLowering::shouldExpandAtomicLoadInIR(LoadInst *LI) const {
11906 unsigned Size = LI->getType()->getPrimitiveSizeInBits();
11907 return ((Size == 64) && !Subtarget->isMClass()) ? AtomicExpansionKind::LLSC
11908 : AtomicExpansionKind::None;
11911 // For the real atomic operations, we have ldrex/strex up to 32 bits,
11912 // and up to 64 bits on the non-M profiles
11913 TargetLowering::AtomicExpansionKind
11914 ARMTargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
11915 unsigned Size = AI->getType()->getPrimitiveSizeInBits();
11916 return (Size <= (Subtarget->isMClass() ? 32U : 64U))
11917 ? AtomicExpansionKind::LLSC
11918 : AtomicExpansionKind::None;
11921 bool ARMTargetLowering::shouldExpandAtomicCmpXchgInIR(
11922 AtomicCmpXchgInst *AI) const {
11926 // This has so far only been implemented for MachO.
11927 bool ARMTargetLowering::useLoadStackGuardNode() const {
11928 return Subtarget->isTargetMachO();
11931 bool ARMTargetLowering::canCombineStoreAndExtract(Type *VectorTy, Value *Idx,
11932 unsigned &Cost) const {
11933 // If we do not have NEON, vector types are not natively supported.
11934 if (!Subtarget->hasNEON())
11937 // Floating point values and vector values map to the same register file.
11938 // Therefore, although we could do a store extract of a vector type, this is
11939 // better to leave at float as we have more freedom in the addressing mode for
11941 if (VectorTy->isFPOrFPVectorTy())
11944 // If the index is unknown at compile time, this is very expensive to lower
11945 // and it is not possible to combine the store with the extract.
11946 if (!isa<ConstantInt>(Idx))
11949 assert(VectorTy->isVectorTy() && "VectorTy is not a vector type");
11950 unsigned BitWidth = cast<VectorType>(VectorTy)->getBitWidth();
11951 // We can do a store + vector extract on any vector that fits perfectly in a D
11953 if (BitWidth == 64 || BitWidth == 128) {
11960 bool ARMTargetLowering::isCheapToSpeculateCttz() const {
11961 return Subtarget->hasV6T2Ops();
11964 bool ARMTargetLowering::isCheapToSpeculateCtlz() const {
11965 return Subtarget->hasV6T2Ops();
11968 Value *ARMTargetLowering::emitLoadLinked(IRBuilder<> &Builder, Value *Addr,
11969 AtomicOrdering Ord) const {
11970 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
11971 Type *ValTy = cast<PointerType>(Addr->getType())->getElementType();
11972 bool IsAcquire = isAtLeastAcquire(Ord);
11974 // Since i64 isn't legal and intrinsics don't get type-lowered, the ldrexd
11975 // intrinsic must return {i32, i32} and we have to recombine them into a
11976 // single i64 here.
11977 if (ValTy->getPrimitiveSizeInBits() == 64) {
11978 Intrinsic::ID Int =
11979 IsAcquire ? Intrinsic::arm_ldaexd : Intrinsic::arm_ldrexd;
11980 Function *Ldrex = llvm::Intrinsic::getDeclaration(M, Int);
11982 Addr = Builder.CreateBitCast(Addr, Type::getInt8PtrTy(M->getContext()));
11983 Value *LoHi = Builder.CreateCall(Ldrex, Addr, "lohi");
11985 Value *Lo = Builder.CreateExtractValue(LoHi, 0, "lo");
11986 Value *Hi = Builder.CreateExtractValue(LoHi, 1, "hi");
11987 if (!Subtarget->isLittle())
11988 std::swap (Lo, Hi);
11989 Lo = Builder.CreateZExt(Lo, ValTy, "lo64");
11990 Hi = Builder.CreateZExt(Hi, ValTy, "hi64");
11991 return Builder.CreateOr(
11992 Lo, Builder.CreateShl(Hi, ConstantInt::get(ValTy, 32)), "val64");
11995 Type *Tys[] = { Addr->getType() };
11996 Intrinsic::ID Int = IsAcquire ? Intrinsic::arm_ldaex : Intrinsic::arm_ldrex;
11997 Function *Ldrex = llvm::Intrinsic::getDeclaration(M, Int, Tys);
11999 return Builder.CreateTruncOrBitCast(
12000 Builder.CreateCall(Ldrex, Addr),
12001 cast<PointerType>(Addr->getType())->getElementType());
12004 void ARMTargetLowering::emitAtomicCmpXchgNoStoreLLBalance(
12005 IRBuilder<> &Builder) const {
12006 if (!Subtarget->hasV7Ops())
12008 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
12009 Builder.CreateCall(llvm::Intrinsic::getDeclaration(M, Intrinsic::arm_clrex));
12012 Value *ARMTargetLowering::emitStoreConditional(IRBuilder<> &Builder, Value *Val,
12014 AtomicOrdering Ord) const {
12015 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
12016 bool IsRelease = isAtLeastRelease(Ord);
12018 // Since the intrinsics must have legal type, the i64 intrinsics take two
12019 // parameters: "i32, i32". We must marshal Val into the appropriate form
12020 // before the call.
12021 if (Val->getType()->getPrimitiveSizeInBits() == 64) {
12022 Intrinsic::ID Int =
12023 IsRelease ? Intrinsic::arm_stlexd : Intrinsic::arm_strexd;
12024 Function *Strex = Intrinsic::getDeclaration(M, Int);
12025 Type *Int32Ty = Type::getInt32Ty(M->getContext());
12027 Value *Lo = Builder.CreateTrunc(Val, Int32Ty, "lo");
12028 Value *Hi = Builder.CreateTrunc(Builder.CreateLShr(Val, 32), Int32Ty, "hi");
12029 if (!Subtarget->isLittle())
12030 std::swap (Lo, Hi);
12031 Addr = Builder.CreateBitCast(Addr, Type::getInt8PtrTy(M->getContext()));
12032 return Builder.CreateCall(Strex, {Lo, Hi, Addr});
12035 Intrinsic::ID Int = IsRelease ? Intrinsic::arm_stlex : Intrinsic::arm_strex;
12036 Type *Tys[] = { Addr->getType() };
12037 Function *Strex = Intrinsic::getDeclaration(M, Int, Tys);
12039 return Builder.CreateCall(
12040 Strex, {Builder.CreateZExtOrBitCast(
12041 Val, Strex->getFunctionType()->getParamType(0)),
12045 /// \brief Lower an interleaved load into a vldN intrinsic.
12047 /// E.g. Lower an interleaved load (Factor = 2):
12048 /// %wide.vec = load <8 x i32>, <8 x i32>* %ptr, align 4
12049 /// %v0 = shuffle %wide.vec, undef, <0, 2, 4, 6> ; Extract even elements
12050 /// %v1 = shuffle %wide.vec, undef, <1, 3, 5, 7> ; Extract odd elements
12053 /// %vld2 = { <4 x i32>, <4 x i32> } call llvm.arm.neon.vld2(%ptr, 4)
12054 /// %vec0 = extractelement { <4 x i32>, <4 x i32> } %vld2, i32 0
12055 /// %vec1 = extractelement { <4 x i32>, <4 x i32> } %vld2, i32 1
12056 bool ARMTargetLowering::lowerInterleavedLoad(
12057 LoadInst *LI, ArrayRef<ShuffleVectorInst *> Shuffles,
12058 ArrayRef<unsigned> Indices, unsigned Factor) const {
12059 assert(Factor >= 2 && Factor <= getMaxSupportedInterleaveFactor() &&
12060 "Invalid interleave factor");
12061 assert(!Shuffles.empty() && "Empty shufflevector input");
12062 assert(Shuffles.size() == Indices.size() &&
12063 "Unmatched number of shufflevectors and indices");
12065 VectorType *VecTy = Shuffles[0]->getType();
12066 Type *EltTy = VecTy->getVectorElementType();
12068 const DataLayout &DL = LI->getModule()->getDataLayout();
12069 unsigned VecSize = DL.getTypeAllocSizeInBits(VecTy);
12070 bool EltIs64Bits = DL.getTypeAllocSizeInBits(EltTy) == 64;
12072 // Skip if we do not have NEON and skip illegal vector types and vector types
12073 // with i64/f64 elements (vldN doesn't support i64/f64 elements).
12074 if (!Subtarget->hasNEON() || (VecSize != 64 && VecSize != 128) || EltIs64Bits)
12077 // A pointer vector can not be the return type of the ldN intrinsics. Need to
12078 // load integer vectors first and then convert to pointer vectors.
12079 if (EltTy->isPointerTy())
12081 VectorType::get(DL.getIntPtrType(EltTy), VecTy->getVectorNumElements());
12083 static const Intrinsic::ID LoadInts[3] = {Intrinsic::arm_neon_vld2,
12084 Intrinsic::arm_neon_vld3,
12085 Intrinsic::arm_neon_vld4};
12087 IRBuilder<> Builder(LI);
12088 SmallVector<Value *, 2> Ops;
12090 Type *Int8Ptr = Builder.getInt8PtrTy(LI->getPointerAddressSpace());
12091 Ops.push_back(Builder.CreateBitCast(LI->getPointerOperand(), Int8Ptr));
12092 Ops.push_back(Builder.getInt32(LI->getAlignment()));
12094 Type *Tys[] = { VecTy, Int8Ptr };
12095 Function *VldnFunc =
12096 Intrinsic::getDeclaration(LI->getModule(), LoadInts[Factor - 2], Tys);
12097 CallInst *VldN = Builder.CreateCall(VldnFunc, Ops, "vldN");
12099 // Replace uses of each shufflevector with the corresponding vector loaded
12101 for (unsigned i = 0; i < Shuffles.size(); i++) {
12102 ShuffleVectorInst *SV = Shuffles[i];
12103 unsigned Index = Indices[i];
12105 Value *SubVec = Builder.CreateExtractValue(VldN, Index);
12107 // Convert the integer vector to pointer vector if the element is pointer.
12108 if (EltTy->isPointerTy())
12109 SubVec = Builder.CreateIntToPtr(SubVec, SV->getType());
12111 SV->replaceAllUsesWith(SubVec);
12117 /// \brief Get a mask consisting of sequential integers starting from \p Start.
12119 /// I.e. <Start, Start + 1, ..., Start + NumElts - 1>
12120 static Constant *getSequentialMask(IRBuilder<> &Builder, unsigned Start,
12121 unsigned NumElts) {
12122 SmallVector<Constant *, 16> Mask;
12123 for (unsigned i = 0; i < NumElts; i++)
12124 Mask.push_back(Builder.getInt32(Start + i));
12126 return ConstantVector::get(Mask);
12129 /// \brief Lower an interleaved store into a vstN intrinsic.
12131 /// E.g. Lower an interleaved store (Factor = 3):
12132 /// %i.vec = shuffle <8 x i32> %v0, <8 x i32> %v1,
12133 /// <0, 4, 8, 1, 5, 9, 2, 6, 10, 3, 7, 11>
12134 /// store <12 x i32> %i.vec, <12 x i32>* %ptr, align 4
12137 /// %sub.v0 = shuffle <8 x i32> %v0, <8 x i32> v1, <0, 1, 2, 3>
12138 /// %sub.v1 = shuffle <8 x i32> %v0, <8 x i32> v1, <4, 5, 6, 7>
12139 /// %sub.v2 = shuffle <8 x i32> %v0, <8 x i32> v1, <8, 9, 10, 11>
12140 /// call void llvm.arm.neon.vst3(%ptr, %sub.v0, %sub.v1, %sub.v2, 4)
12142 /// Note that the new shufflevectors will be removed and we'll only generate one
12143 /// vst3 instruction in CodeGen.
12144 bool ARMTargetLowering::lowerInterleavedStore(StoreInst *SI,
12145 ShuffleVectorInst *SVI,
12146 unsigned Factor) const {
12147 assert(Factor >= 2 && Factor <= getMaxSupportedInterleaveFactor() &&
12148 "Invalid interleave factor");
12150 VectorType *VecTy = SVI->getType();
12151 assert(VecTy->getVectorNumElements() % Factor == 0 &&
12152 "Invalid interleaved store");
12154 unsigned NumSubElts = VecTy->getVectorNumElements() / Factor;
12155 Type *EltTy = VecTy->getVectorElementType();
12156 VectorType *SubVecTy = VectorType::get(EltTy, NumSubElts);
12158 const DataLayout &DL = SI->getModule()->getDataLayout();
12159 unsigned SubVecSize = DL.getTypeAllocSizeInBits(SubVecTy);
12160 bool EltIs64Bits = DL.getTypeAllocSizeInBits(EltTy) == 64;
12162 // Skip if we do not have NEON and skip illegal vector types and vector types
12163 // with i64/f64 elements (vstN doesn't support i64/f64 elements).
12164 if (!Subtarget->hasNEON() || (SubVecSize != 64 && SubVecSize != 128) ||
12168 Value *Op0 = SVI->getOperand(0);
12169 Value *Op1 = SVI->getOperand(1);
12170 IRBuilder<> Builder(SI);
12172 // StN intrinsics don't support pointer vectors as arguments. Convert pointer
12173 // vectors to integer vectors.
12174 if (EltTy->isPointerTy()) {
12175 Type *IntTy = DL.getIntPtrType(EltTy);
12177 // Convert to the corresponding integer vector.
12179 VectorType::get(IntTy, Op0->getType()->getVectorNumElements());
12180 Op0 = Builder.CreatePtrToInt(Op0, IntVecTy);
12181 Op1 = Builder.CreatePtrToInt(Op1, IntVecTy);
12183 SubVecTy = VectorType::get(IntTy, NumSubElts);
12186 static const Intrinsic::ID StoreInts[3] = {Intrinsic::arm_neon_vst2,
12187 Intrinsic::arm_neon_vst3,
12188 Intrinsic::arm_neon_vst4};
12189 SmallVector<Value *, 6> Ops;
12191 Type *Int8Ptr = Builder.getInt8PtrTy(SI->getPointerAddressSpace());
12192 Ops.push_back(Builder.CreateBitCast(SI->getPointerOperand(), Int8Ptr));
12194 Type *Tys[] = { Int8Ptr, SubVecTy };
12195 Function *VstNFunc = Intrinsic::getDeclaration(
12196 SI->getModule(), StoreInts[Factor - 2], Tys);
12198 // Split the shufflevector operands into sub vectors for the new vstN call.
12199 for (unsigned i = 0; i < Factor; i++)
12200 Ops.push_back(Builder.CreateShuffleVector(
12201 Op0, Op1, getSequentialMask(Builder, NumSubElts * i, NumSubElts)));
12203 Ops.push_back(Builder.getInt32(SI->getAlignment()));
12204 Builder.CreateCall(VstNFunc, Ops);
12216 static bool isHomogeneousAggregate(Type *Ty, HABaseType &Base,
12217 uint64_t &Members) {
12218 if (auto *ST = dyn_cast<StructType>(Ty)) {
12219 for (unsigned i = 0; i < ST->getNumElements(); ++i) {
12220 uint64_t SubMembers = 0;
12221 if (!isHomogeneousAggregate(ST->getElementType(i), Base, SubMembers))
12223 Members += SubMembers;
12225 } else if (auto *AT = dyn_cast<ArrayType>(Ty)) {
12226 uint64_t SubMembers = 0;
12227 if (!isHomogeneousAggregate(AT->getElementType(), Base, SubMembers))
12229 Members += SubMembers * AT->getNumElements();
12230 } else if (Ty->isFloatTy()) {
12231 if (Base != HA_UNKNOWN && Base != HA_FLOAT)
12235 } else if (Ty->isDoubleTy()) {
12236 if (Base != HA_UNKNOWN && Base != HA_DOUBLE)
12240 } else if (auto *VT = dyn_cast<VectorType>(Ty)) {
12247 return VT->getBitWidth() == 64;
12249 return VT->getBitWidth() == 128;
12251 switch (VT->getBitWidth()) {
12264 return (Members > 0 && Members <= 4);
12267 /// \brief Return true if a type is an AAPCS-VFP homogeneous aggregate or one of
12268 /// [N x i32] or [N x i64]. This allows front-ends to skip emitting padding when
12269 /// passing according to AAPCS rules.
12270 bool ARMTargetLowering::functionArgumentNeedsConsecutiveRegisters(
12271 Type *Ty, CallingConv::ID CallConv, bool isVarArg) const {
12272 if (getEffectiveCallingConv(CallConv, isVarArg) !=
12273 CallingConv::ARM_AAPCS_VFP)
12276 HABaseType Base = HA_UNKNOWN;
12277 uint64_t Members = 0;
12278 bool IsHA = isHomogeneousAggregate(Ty, Base, Members);
12279 DEBUG(dbgs() << "isHA: " << IsHA << " "; Ty->dump());
12281 bool IsIntArray = Ty->isArrayTy() && Ty->getArrayElementType()->isIntegerTy();
12282 return IsHA || IsIntArray;
12285 unsigned ARMTargetLowering::getExceptionPointerRegister(
12286 const Constant *PersonalityFn) const {
12287 // Platforms which do not use SjLj EH may return values in these registers
12288 // via the personality function.
12289 return Subtarget->useSjLjEH() ? ARM::NoRegister : ARM::R0;
12292 unsigned ARMTargetLowering::getExceptionSelectorRegister(
12293 const Constant *PersonalityFn) const {
12294 // Platforms which do not use SjLj EH may return values in these registers
12295 // via the personality function.
12296 return Subtarget->useSjLjEH() ? ARM::NoRegister : ARM::R1;