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.isFloatingPoint() &&
147 VT != MVT::v2i64 && VT != MVT::v1i64)
148 for (unsigned Opcode : {ISD::SMIN, ISD::SMAX, ISD::UMIN, ISD::UMAX})
149 setOperationAction(Opcode, VT, Legal);
152 void ARMTargetLowering::addDRTypeForNEON(MVT VT) {
153 addRegisterClass(VT, &ARM::DPRRegClass);
154 addTypeForNEON(VT, MVT::f64, MVT::v2i32);
157 void ARMTargetLowering::addQRTypeForNEON(MVT VT) {
158 addRegisterClass(VT, &ARM::DPairRegClass);
159 addTypeForNEON(VT, MVT::v2f64, MVT::v4i32);
162 ARMTargetLowering::ARMTargetLowering(const TargetMachine &TM,
163 const ARMSubtarget &STI)
164 : TargetLowering(TM), Subtarget(&STI) {
165 RegInfo = Subtarget->getRegisterInfo();
166 Itins = Subtarget->getInstrItineraryData();
168 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
170 if (Subtarget->isTargetMachO()) {
171 // Uses VFP for Thumb libfuncs if available.
172 if (Subtarget->isThumb() && Subtarget->hasVFP2() &&
173 Subtarget->hasARMOps() && !Subtarget->useSoftFloat()) {
174 static const struct {
175 const RTLIB::Libcall Op;
176 const char * const Name;
177 const ISD::CondCode Cond;
179 // Single-precision floating-point arithmetic.
180 { RTLIB::ADD_F32, "__addsf3vfp", ISD::SETCC_INVALID },
181 { RTLIB::SUB_F32, "__subsf3vfp", ISD::SETCC_INVALID },
182 { RTLIB::MUL_F32, "__mulsf3vfp", ISD::SETCC_INVALID },
183 { RTLIB::DIV_F32, "__divsf3vfp", ISD::SETCC_INVALID },
185 // Double-precision floating-point arithmetic.
186 { RTLIB::ADD_F64, "__adddf3vfp", ISD::SETCC_INVALID },
187 { RTLIB::SUB_F64, "__subdf3vfp", ISD::SETCC_INVALID },
188 { RTLIB::MUL_F64, "__muldf3vfp", ISD::SETCC_INVALID },
189 { RTLIB::DIV_F64, "__divdf3vfp", ISD::SETCC_INVALID },
191 // Single-precision comparisons.
192 { RTLIB::OEQ_F32, "__eqsf2vfp", ISD::SETNE },
193 { RTLIB::UNE_F32, "__nesf2vfp", ISD::SETNE },
194 { RTLIB::OLT_F32, "__ltsf2vfp", ISD::SETNE },
195 { RTLIB::OLE_F32, "__lesf2vfp", ISD::SETNE },
196 { RTLIB::OGE_F32, "__gesf2vfp", ISD::SETNE },
197 { RTLIB::OGT_F32, "__gtsf2vfp", ISD::SETNE },
198 { RTLIB::UO_F32, "__unordsf2vfp", ISD::SETNE },
199 { RTLIB::O_F32, "__unordsf2vfp", ISD::SETEQ },
201 // Double-precision comparisons.
202 { RTLIB::OEQ_F64, "__eqdf2vfp", ISD::SETNE },
203 { RTLIB::UNE_F64, "__nedf2vfp", ISD::SETNE },
204 { RTLIB::OLT_F64, "__ltdf2vfp", ISD::SETNE },
205 { RTLIB::OLE_F64, "__ledf2vfp", ISD::SETNE },
206 { RTLIB::OGE_F64, "__gedf2vfp", ISD::SETNE },
207 { RTLIB::OGT_F64, "__gtdf2vfp", ISD::SETNE },
208 { RTLIB::UO_F64, "__unorddf2vfp", ISD::SETNE },
209 { RTLIB::O_F64, "__unorddf2vfp", ISD::SETEQ },
211 // Floating-point to integer conversions.
212 // i64 conversions are done via library routines even when generating VFP
213 // instructions, so use the same ones.
214 { RTLIB::FPTOSINT_F64_I32, "__fixdfsivfp", ISD::SETCC_INVALID },
215 { RTLIB::FPTOUINT_F64_I32, "__fixunsdfsivfp", ISD::SETCC_INVALID },
216 { RTLIB::FPTOSINT_F32_I32, "__fixsfsivfp", ISD::SETCC_INVALID },
217 { RTLIB::FPTOUINT_F32_I32, "__fixunssfsivfp", ISD::SETCC_INVALID },
219 // Conversions between floating types.
220 { RTLIB::FPROUND_F64_F32, "__truncdfsf2vfp", ISD::SETCC_INVALID },
221 { RTLIB::FPEXT_F32_F64, "__extendsfdf2vfp", ISD::SETCC_INVALID },
223 // Integer to floating-point conversions.
224 // i64 conversions are done via library routines even when generating VFP
225 // instructions, so use the same ones.
226 // FIXME: There appears to be some naming inconsistency in ARM libgcc:
227 // e.g., __floatunsidf vs. __floatunssidfvfp.
228 { RTLIB::SINTTOFP_I32_F64, "__floatsidfvfp", ISD::SETCC_INVALID },
229 { RTLIB::UINTTOFP_I32_F64, "__floatunssidfvfp", ISD::SETCC_INVALID },
230 { RTLIB::SINTTOFP_I32_F32, "__floatsisfvfp", ISD::SETCC_INVALID },
231 { RTLIB::UINTTOFP_I32_F32, "__floatunssisfvfp", ISD::SETCC_INVALID },
234 for (const auto &LC : LibraryCalls) {
235 setLibcallName(LC.Op, LC.Name);
236 if (LC.Cond != ISD::SETCC_INVALID)
237 setCmpLibcallCC(LC.Op, LC.Cond);
241 // Set the correct calling convention for ARMv7k WatchOS. It's just
242 // AAPCS_VFP for functions as simple as libcalls.
243 if (Subtarget->isTargetWatchOS()) {
244 for (int i = 0; i < RTLIB::UNKNOWN_LIBCALL; ++i)
245 setLibcallCallingConv((RTLIB::Libcall)i, CallingConv::ARM_AAPCS_VFP);
249 // These libcalls are not available in 32-bit.
250 setLibcallName(RTLIB::SHL_I128, nullptr);
251 setLibcallName(RTLIB::SRL_I128, nullptr);
252 setLibcallName(RTLIB::SRA_I128, nullptr);
255 if (Subtarget->isAAPCS_ABI() &&
256 (Subtarget->isTargetAEABI() || Subtarget->isTargetGNUAEABI() ||
257 Subtarget->isTargetAndroid())) {
258 static const struct {
259 const RTLIB::Libcall Op;
260 const char * const Name;
261 const CallingConv::ID CC;
262 const ISD::CondCode Cond;
264 // Double-precision floating-point arithmetic helper functions
265 // RTABI chapter 4.1.2, Table 2
266 { RTLIB::ADD_F64, "__aeabi_dadd", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
267 { RTLIB::DIV_F64, "__aeabi_ddiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
268 { RTLIB::MUL_F64, "__aeabi_dmul", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
269 { RTLIB::SUB_F64, "__aeabi_dsub", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
271 // Double-precision floating-point comparison helper functions
272 // RTABI chapter 4.1.2, Table 3
273 { RTLIB::OEQ_F64, "__aeabi_dcmpeq", CallingConv::ARM_AAPCS, ISD::SETNE },
274 { RTLIB::UNE_F64, "__aeabi_dcmpeq", CallingConv::ARM_AAPCS, ISD::SETEQ },
275 { RTLIB::OLT_F64, "__aeabi_dcmplt", CallingConv::ARM_AAPCS, ISD::SETNE },
276 { RTLIB::OLE_F64, "__aeabi_dcmple", CallingConv::ARM_AAPCS, ISD::SETNE },
277 { RTLIB::OGE_F64, "__aeabi_dcmpge", CallingConv::ARM_AAPCS, ISD::SETNE },
278 { RTLIB::OGT_F64, "__aeabi_dcmpgt", CallingConv::ARM_AAPCS, ISD::SETNE },
279 { RTLIB::UO_F64, "__aeabi_dcmpun", CallingConv::ARM_AAPCS, ISD::SETNE },
280 { RTLIB::O_F64, "__aeabi_dcmpun", CallingConv::ARM_AAPCS, ISD::SETEQ },
282 // Single-precision floating-point arithmetic helper functions
283 // RTABI chapter 4.1.2, Table 4
284 { RTLIB::ADD_F32, "__aeabi_fadd", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
285 { RTLIB::DIV_F32, "__aeabi_fdiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
286 { RTLIB::MUL_F32, "__aeabi_fmul", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
287 { RTLIB::SUB_F32, "__aeabi_fsub", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
289 // Single-precision floating-point comparison helper functions
290 // RTABI chapter 4.1.2, Table 5
291 { RTLIB::OEQ_F32, "__aeabi_fcmpeq", CallingConv::ARM_AAPCS, ISD::SETNE },
292 { RTLIB::UNE_F32, "__aeabi_fcmpeq", CallingConv::ARM_AAPCS, ISD::SETEQ },
293 { RTLIB::OLT_F32, "__aeabi_fcmplt", CallingConv::ARM_AAPCS, ISD::SETNE },
294 { RTLIB::OLE_F32, "__aeabi_fcmple", CallingConv::ARM_AAPCS, ISD::SETNE },
295 { RTLIB::OGE_F32, "__aeabi_fcmpge", CallingConv::ARM_AAPCS, ISD::SETNE },
296 { RTLIB::OGT_F32, "__aeabi_fcmpgt", CallingConv::ARM_AAPCS, ISD::SETNE },
297 { RTLIB::UO_F32, "__aeabi_fcmpun", CallingConv::ARM_AAPCS, ISD::SETNE },
298 { RTLIB::O_F32, "__aeabi_fcmpun", CallingConv::ARM_AAPCS, ISD::SETEQ },
300 // Floating-point to integer conversions.
301 // RTABI chapter 4.1.2, Table 6
302 { RTLIB::FPTOSINT_F64_I32, "__aeabi_d2iz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
303 { RTLIB::FPTOUINT_F64_I32, "__aeabi_d2uiz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
304 { RTLIB::FPTOSINT_F64_I64, "__aeabi_d2lz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
305 { RTLIB::FPTOUINT_F64_I64, "__aeabi_d2ulz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
306 { RTLIB::FPTOSINT_F32_I32, "__aeabi_f2iz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
307 { RTLIB::FPTOUINT_F32_I32, "__aeabi_f2uiz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
308 { RTLIB::FPTOSINT_F32_I64, "__aeabi_f2lz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
309 { RTLIB::FPTOUINT_F32_I64, "__aeabi_f2ulz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
311 // Conversions between floating types.
312 // RTABI chapter 4.1.2, Table 7
313 { RTLIB::FPROUND_F64_F32, "__aeabi_d2f", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
314 { RTLIB::FPROUND_F64_F16, "__aeabi_d2h", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
315 { RTLIB::FPEXT_F32_F64, "__aeabi_f2d", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
317 // Integer to floating-point conversions.
318 // RTABI chapter 4.1.2, Table 8
319 { RTLIB::SINTTOFP_I32_F64, "__aeabi_i2d", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
320 { RTLIB::UINTTOFP_I32_F64, "__aeabi_ui2d", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
321 { RTLIB::SINTTOFP_I64_F64, "__aeabi_l2d", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
322 { RTLIB::UINTTOFP_I64_F64, "__aeabi_ul2d", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
323 { RTLIB::SINTTOFP_I32_F32, "__aeabi_i2f", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
324 { RTLIB::UINTTOFP_I32_F32, "__aeabi_ui2f", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
325 { RTLIB::SINTTOFP_I64_F32, "__aeabi_l2f", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
326 { RTLIB::UINTTOFP_I64_F32, "__aeabi_ul2f", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
328 // Long long helper functions
329 // RTABI chapter 4.2, Table 9
330 { RTLIB::MUL_I64, "__aeabi_lmul", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
331 { RTLIB::SHL_I64, "__aeabi_llsl", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
332 { RTLIB::SRL_I64, "__aeabi_llsr", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
333 { RTLIB::SRA_I64, "__aeabi_lasr", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
335 // Integer division functions
336 // RTABI chapter 4.3.1
337 { RTLIB::SDIV_I8, "__aeabi_idiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
338 { RTLIB::SDIV_I16, "__aeabi_idiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
339 { RTLIB::SDIV_I32, "__aeabi_idiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
340 { RTLIB::SDIV_I64, "__aeabi_ldivmod", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
341 { RTLIB::UDIV_I8, "__aeabi_uidiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
342 { RTLIB::UDIV_I16, "__aeabi_uidiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
343 { RTLIB::UDIV_I32, "__aeabi_uidiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
344 { RTLIB::UDIV_I64, "__aeabi_uldivmod", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
347 for (const auto &LC : LibraryCalls) {
348 setLibcallName(LC.Op, LC.Name);
349 setLibcallCallingConv(LC.Op, LC.CC);
350 if (LC.Cond != ISD::SETCC_INVALID)
351 setCmpLibcallCC(LC.Op, LC.Cond);
354 // EABI dependent RTLIB
355 if (TM.Options.EABIVersion == EABI::EABI4 ||
356 TM.Options.EABIVersion == EABI::EABI5) {
357 static const struct {
358 const RTLIB::Libcall Op;
359 const char *const Name;
360 const CallingConv::ID CC;
361 const ISD::CondCode Cond;
362 } MemOpsLibraryCalls[] = {
364 // RTABI chapter 4.3.4
365 { RTLIB::MEMCPY, "__aeabi_memcpy", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
366 { RTLIB::MEMMOVE, "__aeabi_memmove", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
367 { RTLIB::MEMSET, "__aeabi_memset", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
370 for (const auto &LC : MemOpsLibraryCalls) {
371 setLibcallName(LC.Op, LC.Name);
372 setLibcallCallingConv(LC.Op, LC.CC);
373 if (LC.Cond != ISD::SETCC_INVALID)
374 setCmpLibcallCC(LC.Op, LC.Cond);
379 if (Subtarget->isTargetWindows()) {
380 static const struct {
381 const RTLIB::Libcall Op;
382 const char * const Name;
383 const CallingConv::ID CC;
385 { RTLIB::FPTOSINT_F32_I64, "__stoi64", CallingConv::ARM_AAPCS_VFP },
386 { RTLIB::FPTOSINT_F64_I64, "__dtoi64", CallingConv::ARM_AAPCS_VFP },
387 { RTLIB::FPTOUINT_F32_I64, "__stou64", CallingConv::ARM_AAPCS_VFP },
388 { RTLIB::FPTOUINT_F64_I64, "__dtou64", CallingConv::ARM_AAPCS_VFP },
389 { RTLIB::SINTTOFP_I64_F32, "__i64tos", CallingConv::ARM_AAPCS_VFP },
390 { RTLIB::SINTTOFP_I64_F64, "__i64tod", CallingConv::ARM_AAPCS_VFP },
391 { RTLIB::UINTTOFP_I64_F32, "__u64tos", CallingConv::ARM_AAPCS_VFP },
392 { RTLIB::UINTTOFP_I64_F64, "__u64tod", CallingConv::ARM_AAPCS_VFP },
393 { RTLIB::SDIV_I32, "__rt_sdiv", CallingConv::ARM_AAPCS_VFP },
394 { RTLIB::UDIV_I32, "__rt_udiv", CallingConv::ARM_AAPCS_VFP },
395 { RTLIB::SDIV_I64, "__rt_sdiv64", CallingConv::ARM_AAPCS_VFP },
396 { RTLIB::UDIV_I64, "__rt_udiv64", CallingConv::ARM_AAPCS_VFP },
399 for (const auto &LC : LibraryCalls) {
400 setLibcallName(LC.Op, LC.Name);
401 setLibcallCallingConv(LC.Op, LC.CC);
405 // Use divmod compiler-rt calls for iOS 5.0 and later.
406 if (Subtarget->isTargetWatchOS() ||
407 (Subtarget->isTargetIOS() &&
408 !Subtarget->getTargetTriple().isOSVersionLT(5, 0))) {
409 setLibcallName(RTLIB::SDIVREM_I32, "__divmodsi4");
410 setLibcallName(RTLIB::UDIVREM_I32, "__udivmodsi4");
413 // The half <-> float conversion functions are always soft-float, but are
414 // needed for some targets which use a hard-float calling convention by
416 if (Subtarget->isAAPCS_ABI()) {
417 setLibcallCallingConv(RTLIB::FPROUND_F32_F16, CallingConv::ARM_AAPCS);
418 setLibcallCallingConv(RTLIB::FPROUND_F64_F16, CallingConv::ARM_AAPCS);
419 setLibcallCallingConv(RTLIB::FPEXT_F16_F32, CallingConv::ARM_AAPCS);
421 setLibcallCallingConv(RTLIB::FPROUND_F32_F16, CallingConv::ARM_APCS);
422 setLibcallCallingConv(RTLIB::FPROUND_F64_F16, CallingConv::ARM_APCS);
423 setLibcallCallingConv(RTLIB::FPEXT_F16_F32, CallingConv::ARM_APCS);
426 // In EABI, these functions have an __aeabi_ prefix, but in GNUEABI they have
427 // a __gnu_ prefix (which is the default).
428 if (Subtarget->isTargetAEABI()) {
429 setLibcallName(RTLIB::FPROUND_F32_F16, "__aeabi_f2h");
430 setLibcallName(RTLIB::FPROUND_F64_F16, "__aeabi_d2h");
431 setLibcallName(RTLIB::FPEXT_F16_F32, "__aeabi_h2f");
434 if (Subtarget->isThumb1Only())
435 addRegisterClass(MVT::i32, &ARM::tGPRRegClass);
437 addRegisterClass(MVT::i32, &ARM::GPRRegClass);
438 if (!Subtarget->useSoftFloat() && Subtarget->hasVFP2() &&
439 !Subtarget->isThumb1Only()) {
440 addRegisterClass(MVT::f32, &ARM::SPRRegClass);
441 addRegisterClass(MVT::f64, &ARM::DPRRegClass);
444 for (MVT VT : MVT::vector_valuetypes()) {
445 for (MVT InnerVT : MVT::vector_valuetypes()) {
446 setTruncStoreAction(VT, InnerVT, Expand);
447 setLoadExtAction(ISD::SEXTLOAD, VT, InnerVT, Expand);
448 setLoadExtAction(ISD::ZEXTLOAD, VT, InnerVT, Expand);
449 setLoadExtAction(ISD::EXTLOAD, VT, InnerVT, Expand);
452 setOperationAction(ISD::MULHS, VT, Expand);
453 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
454 setOperationAction(ISD::MULHU, VT, Expand);
455 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
457 setOperationAction(ISD::BSWAP, VT, Expand);
460 setOperationAction(ISD::ConstantFP, MVT::f32, Custom);
461 setOperationAction(ISD::ConstantFP, MVT::f64, Custom);
463 setOperationAction(ISD::READ_REGISTER, MVT::i64, Custom);
464 setOperationAction(ISD::WRITE_REGISTER, MVT::i64, Custom);
466 if (Subtarget->hasNEON()) {
467 addDRTypeForNEON(MVT::v2f32);
468 addDRTypeForNEON(MVT::v8i8);
469 addDRTypeForNEON(MVT::v4i16);
470 addDRTypeForNEON(MVT::v2i32);
471 addDRTypeForNEON(MVT::v1i64);
473 addQRTypeForNEON(MVT::v4f32);
474 addQRTypeForNEON(MVT::v2f64);
475 addQRTypeForNEON(MVT::v16i8);
476 addQRTypeForNEON(MVT::v8i16);
477 addQRTypeForNEON(MVT::v4i32);
478 addQRTypeForNEON(MVT::v2i64);
480 // v2f64 is legal so that QR subregs can be extracted as f64 elements, but
481 // neither Neon nor VFP support any arithmetic operations on it.
482 // The same with v4f32. But keep in mind that vadd, vsub, vmul are natively
483 // supported for v4f32.
484 setOperationAction(ISD::FADD, MVT::v2f64, Expand);
485 setOperationAction(ISD::FSUB, MVT::v2f64, Expand);
486 setOperationAction(ISD::FMUL, MVT::v2f64, Expand);
487 // FIXME: Code duplication: FDIV and FREM are expanded always, see
488 // ARMTargetLowering::addTypeForNEON method for details.
489 setOperationAction(ISD::FDIV, MVT::v2f64, Expand);
490 setOperationAction(ISD::FREM, MVT::v2f64, Expand);
491 // FIXME: Create unittest.
492 // In another words, find a way when "copysign" appears in DAG with vector
494 setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand);
495 // FIXME: Code duplication: SETCC has custom operation action, see
496 // ARMTargetLowering::addTypeForNEON method for details.
497 setOperationAction(ISD::SETCC, MVT::v2f64, Expand);
498 // FIXME: Create unittest for FNEG and for FABS.
499 setOperationAction(ISD::FNEG, MVT::v2f64, Expand);
500 setOperationAction(ISD::FABS, MVT::v2f64, Expand);
501 setOperationAction(ISD::FSQRT, MVT::v2f64, Expand);
502 setOperationAction(ISD::FSIN, MVT::v2f64, Expand);
503 setOperationAction(ISD::FCOS, MVT::v2f64, Expand);
504 setOperationAction(ISD::FPOWI, MVT::v2f64, Expand);
505 setOperationAction(ISD::FPOW, MVT::v2f64, Expand);
506 setOperationAction(ISD::FLOG, MVT::v2f64, Expand);
507 setOperationAction(ISD::FLOG2, MVT::v2f64, Expand);
508 setOperationAction(ISD::FLOG10, MVT::v2f64, Expand);
509 setOperationAction(ISD::FEXP, MVT::v2f64, Expand);
510 setOperationAction(ISD::FEXP2, MVT::v2f64, Expand);
511 // FIXME: Create unittest for FCEIL, FTRUNC, FRINT, FNEARBYINT, FFLOOR.
512 setOperationAction(ISD::FCEIL, MVT::v2f64, Expand);
513 setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand);
514 setOperationAction(ISD::FRINT, MVT::v2f64, Expand);
515 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand);
516 setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand);
517 setOperationAction(ISD::FMA, MVT::v2f64, Expand);
519 setOperationAction(ISD::FSQRT, MVT::v4f32, Expand);
520 setOperationAction(ISD::FSIN, MVT::v4f32, Expand);
521 setOperationAction(ISD::FCOS, MVT::v4f32, Expand);
522 setOperationAction(ISD::FPOWI, MVT::v4f32, Expand);
523 setOperationAction(ISD::FPOW, MVT::v4f32, Expand);
524 setOperationAction(ISD::FLOG, MVT::v4f32, Expand);
525 setOperationAction(ISD::FLOG2, MVT::v4f32, Expand);
526 setOperationAction(ISD::FLOG10, MVT::v4f32, Expand);
527 setOperationAction(ISD::FEXP, MVT::v4f32, Expand);
528 setOperationAction(ISD::FEXP2, MVT::v4f32, Expand);
529 setOperationAction(ISD::FCEIL, MVT::v4f32, Expand);
530 setOperationAction(ISD::FTRUNC, MVT::v4f32, Expand);
531 setOperationAction(ISD::FRINT, MVT::v4f32, Expand);
532 setOperationAction(ISD::FNEARBYINT, MVT::v4f32, Expand);
533 setOperationAction(ISD::FFLOOR, MVT::v4f32, Expand);
535 // Mark v2f32 intrinsics.
536 setOperationAction(ISD::FSQRT, MVT::v2f32, Expand);
537 setOperationAction(ISD::FSIN, MVT::v2f32, Expand);
538 setOperationAction(ISD::FCOS, MVT::v2f32, Expand);
539 setOperationAction(ISD::FPOWI, MVT::v2f32, Expand);
540 setOperationAction(ISD::FPOW, MVT::v2f32, Expand);
541 setOperationAction(ISD::FLOG, MVT::v2f32, Expand);
542 setOperationAction(ISD::FLOG2, MVT::v2f32, Expand);
543 setOperationAction(ISD::FLOG10, MVT::v2f32, Expand);
544 setOperationAction(ISD::FEXP, MVT::v2f32, Expand);
545 setOperationAction(ISD::FEXP2, MVT::v2f32, Expand);
546 setOperationAction(ISD::FCEIL, MVT::v2f32, Expand);
547 setOperationAction(ISD::FTRUNC, MVT::v2f32, Expand);
548 setOperationAction(ISD::FRINT, MVT::v2f32, Expand);
549 setOperationAction(ISD::FNEARBYINT, MVT::v2f32, Expand);
550 setOperationAction(ISD::FFLOOR, MVT::v2f32, Expand);
552 // Neon does not support some operations on v1i64 and v2i64 types.
553 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
554 // Custom handling for some quad-vector types to detect VMULL.
555 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
556 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
557 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
558 // Custom handling for some vector types to avoid expensive expansions
559 setOperationAction(ISD::SDIV, MVT::v4i16, Custom);
560 setOperationAction(ISD::SDIV, MVT::v8i8, Custom);
561 setOperationAction(ISD::UDIV, MVT::v4i16, Custom);
562 setOperationAction(ISD::UDIV, MVT::v8i8, Custom);
563 setOperationAction(ISD::SETCC, MVT::v1i64, Expand);
564 setOperationAction(ISD::SETCC, MVT::v2i64, Expand);
565 // Neon does not have single instruction SINT_TO_FP and UINT_TO_FP with
566 // a destination type that is wider than the source, and nor does
567 // it have a FP_TO_[SU]INT instruction with a narrower destination than
569 setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Custom);
570 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
571 setOperationAction(ISD::FP_TO_UINT, MVT::v4i16, Custom);
572 setOperationAction(ISD::FP_TO_SINT, MVT::v4i16, Custom);
574 setOperationAction(ISD::FP_ROUND, MVT::v2f32, Expand);
575 setOperationAction(ISD::FP_EXTEND, MVT::v2f64, Expand);
577 // NEON does not have single instruction CTPOP for vectors with element
578 // types wider than 8-bits. However, custom lowering can leverage the
579 // v8i8/v16i8 vcnt instruction.
580 setOperationAction(ISD::CTPOP, MVT::v2i32, Custom);
581 setOperationAction(ISD::CTPOP, MVT::v4i32, Custom);
582 setOperationAction(ISD::CTPOP, MVT::v4i16, Custom);
583 setOperationAction(ISD::CTPOP, MVT::v8i16, Custom);
585 // NEON does not have single instruction CTTZ for vectors.
586 setOperationAction(ISD::CTTZ, MVT::v8i8, Custom);
587 setOperationAction(ISD::CTTZ, MVT::v4i16, Custom);
588 setOperationAction(ISD::CTTZ, MVT::v2i32, Custom);
589 setOperationAction(ISD::CTTZ, MVT::v1i64, Custom);
591 setOperationAction(ISD::CTTZ, MVT::v16i8, Custom);
592 setOperationAction(ISD::CTTZ, MVT::v8i16, Custom);
593 setOperationAction(ISD::CTTZ, MVT::v4i32, Custom);
594 setOperationAction(ISD::CTTZ, MVT::v2i64, Custom);
596 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v8i8, Custom);
597 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v4i16, Custom);
598 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v2i32, Custom);
599 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v1i64, Custom);
601 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v16i8, Custom);
602 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v8i16, Custom);
603 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v4i32, Custom);
604 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v2i64, Custom);
606 // NEON only has FMA instructions as of VFP4.
607 if (!Subtarget->hasVFP4()) {
608 setOperationAction(ISD::FMA, MVT::v2f32, Expand);
609 setOperationAction(ISD::FMA, MVT::v4f32, Expand);
612 setTargetDAGCombine(ISD::INTRINSIC_VOID);
613 setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN);
614 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
615 setTargetDAGCombine(ISD::SHL);
616 setTargetDAGCombine(ISD::SRL);
617 setTargetDAGCombine(ISD::SRA);
618 setTargetDAGCombine(ISD::SIGN_EXTEND);
619 setTargetDAGCombine(ISD::ZERO_EXTEND);
620 setTargetDAGCombine(ISD::ANY_EXTEND);
621 setTargetDAGCombine(ISD::BUILD_VECTOR);
622 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
623 setTargetDAGCombine(ISD::INSERT_VECTOR_ELT);
624 setTargetDAGCombine(ISD::STORE);
625 setTargetDAGCombine(ISD::FP_TO_SINT);
626 setTargetDAGCombine(ISD::FP_TO_UINT);
627 setTargetDAGCombine(ISD::FDIV);
628 setTargetDAGCombine(ISD::LOAD);
630 // It is legal to extload from v4i8 to v4i16 or v4i32.
631 for (MVT Ty : {MVT::v8i8, MVT::v4i8, MVT::v2i8, MVT::v4i16, MVT::v2i16,
633 for (MVT VT : MVT::integer_vector_valuetypes()) {
634 setLoadExtAction(ISD::EXTLOAD, VT, Ty, Legal);
635 setLoadExtAction(ISD::ZEXTLOAD, VT, Ty, Legal);
636 setLoadExtAction(ISD::SEXTLOAD, VT, Ty, Legal);
641 // ARM and Thumb2 support UMLAL/SMLAL.
642 if (!Subtarget->isThumb1Only())
643 setTargetDAGCombine(ISD::ADDC);
645 if (Subtarget->isFPOnlySP()) {
646 // When targeting a floating-point unit with only single-precision
647 // operations, f64 is legal for the few double-precision instructions which
648 // are present However, no double-precision operations other than moves,
649 // loads and stores are provided by the hardware.
650 setOperationAction(ISD::FADD, MVT::f64, Expand);
651 setOperationAction(ISD::FSUB, MVT::f64, Expand);
652 setOperationAction(ISD::FMUL, MVT::f64, Expand);
653 setOperationAction(ISD::FMA, MVT::f64, Expand);
654 setOperationAction(ISD::FDIV, MVT::f64, Expand);
655 setOperationAction(ISD::FREM, MVT::f64, Expand);
656 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
657 setOperationAction(ISD::FGETSIGN, MVT::f64, Expand);
658 setOperationAction(ISD::FNEG, MVT::f64, Expand);
659 setOperationAction(ISD::FABS, MVT::f64, Expand);
660 setOperationAction(ISD::FSQRT, MVT::f64, Expand);
661 setOperationAction(ISD::FSIN, MVT::f64, Expand);
662 setOperationAction(ISD::FCOS, MVT::f64, Expand);
663 setOperationAction(ISD::FPOWI, MVT::f64, Expand);
664 setOperationAction(ISD::FPOW, MVT::f64, Expand);
665 setOperationAction(ISD::FLOG, MVT::f64, Expand);
666 setOperationAction(ISD::FLOG2, MVT::f64, Expand);
667 setOperationAction(ISD::FLOG10, MVT::f64, Expand);
668 setOperationAction(ISD::FEXP, MVT::f64, Expand);
669 setOperationAction(ISD::FEXP2, MVT::f64, Expand);
670 setOperationAction(ISD::FCEIL, MVT::f64, Expand);
671 setOperationAction(ISD::FTRUNC, MVT::f64, Expand);
672 setOperationAction(ISD::FRINT, MVT::f64, Expand);
673 setOperationAction(ISD::FNEARBYINT, MVT::f64, Expand);
674 setOperationAction(ISD::FFLOOR, MVT::f64, Expand);
675 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
676 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
677 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
678 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
679 setOperationAction(ISD::FP_TO_SINT, MVT::f64, Custom);
680 setOperationAction(ISD::FP_TO_UINT, MVT::f64, Custom);
681 setOperationAction(ISD::FP_ROUND, MVT::f32, Custom);
682 setOperationAction(ISD::FP_EXTEND, MVT::f64, Custom);
685 computeRegisterProperties(Subtarget->getRegisterInfo());
687 // ARM does not have floating-point extending loads.
688 for (MVT VT : MVT::fp_valuetypes()) {
689 setLoadExtAction(ISD::EXTLOAD, VT, MVT::f32, Expand);
690 setLoadExtAction(ISD::EXTLOAD, VT, MVT::f16, Expand);
693 // ... or truncating stores
694 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
695 setTruncStoreAction(MVT::f32, MVT::f16, Expand);
696 setTruncStoreAction(MVT::f64, MVT::f16, Expand);
698 // ARM does not have i1 sign extending load.
699 for (MVT VT : MVT::integer_valuetypes())
700 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Promote);
702 // ARM supports all 4 flavors of integer indexed load / store.
703 if (!Subtarget->isThumb1Only()) {
704 for (unsigned im = (unsigned)ISD::PRE_INC;
705 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
706 setIndexedLoadAction(im, MVT::i1, Legal);
707 setIndexedLoadAction(im, MVT::i8, Legal);
708 setIndexedLoadAction(im, MVT::i16, Legal);
709 setIndexedLoadAction(im, MVT::i32, Legal);
710 setIndexedStoreAction(im, MVT::i1, Legal);
711 setIndexedStoreAction(im, MVT::i8, Legal);
712 setIndexedStoreAction(im, MVT::i16, Legal);
713 setIndexedStoreAction(im, MVT::i32, Legal);
717 setOperationAction(ISD::SADDO, MVT::i32, Custom);
718 setOperationAction(ISD::UADDO, MVT::i32, Custom);
719 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
720 setOperationAction(ISD::USUBO, MVT::i32, Custom);
722 // i64 operation support.
723 setOperationAction(ISD::MUL, MVT::i64, Expand);
724 setOperationAction(ISD::MULHU, MVT::i32, Expand);
725 if (Subtarget->isThumb1Only()) {
726 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
727 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
729 if (Subtarget->isThumb1Only() || !Subtarget->hasV6Ops()
730 || (Subtarget->isThumb2() && !Subtarget->hasDSP()))
731 setOperationAction(ISD::MULHS, MVT::i32, Expand);
733 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
734 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
735 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
736 setOperationAction(ISD::SRL, MVT::i64, Custom);
737 setOperationAction(ISD::SRA, MVT::i64, Custom);
739 if (!Subtarget->isThumb1Only()) {
740 // FIXME: We should do this for Thumb1 as well.
741 setOperationAction(ISD::ADDC, MVT::i32, Custom);
742 setOperationAction(ISD::ADDE, MVT::i32, Custom);
743 setOperationAction(ISD::SUBC, MVT::i32, Custom);
744 setOperationAction(ISD::SUBE, MVT::i32, Custom);
747 if (!Subtarget->isThumb1Only())
748 setOperationAction(ISD::BITREVERSE, MVT::i32, Legal);
750 // ARM does not have ROTL.
751 setOperationAction(ISD::ROTL, MVT::i32, Expand);
752 for (MVT VT : MVT::vector_valuetypes()) {
753 setOperationAction(ISD::ROTL, VT, Expand);
754 setOperationAction(ISD::ROTR, VT, Expand);
756 setOperationAction(ISD::CTTZ, MVT::i32, Custom);
757 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
758 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
759 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
761 // These just redirect to CTTZ and CTLZ on ARM.
762 setOperationAction(ISD::CTTZ_ZERO_UNDEF , MVT::i32 , Expand);
763 setOperationAction(ISD::CTLZ_ZERO_UNDEF , MVT::i32 , Expand);
765 // @llvm.readcyclecounter requires the Performance Monitors extension.
766 // Default to the 0 expansion on unsupported platforms.
767 // FIXME: Technically there are older ARM CPUs that have
768 // implementation-specific ways of obtaining this information.
769 if (Subtarget->hasPerfMon())
770 setOperationAction(ISD::READCYCLECOUNTER, MVT::i64, Custom);
772 // Only ARMv6 has BSWAP.
773 if (!Subtarget->hasV6Ops())
774 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
776 if (!(Subtarget->hasDivide() && Subtarget->isThumb2()) &&
777 !(Subtarget->hasDivideInARMMode() && !Subtarget->isThumb())) {
778 // These are expanded into libcalls if the cpu doesn't have HW divider.
779 setOperationAction(ISD::SDIV, MVT::i32, LibCall);
780 setOperationAction(ISD::UDIV, MVT::i32, LibCall);
783 setOperationAction(ISD::SREM, MVT::i32, Expand);
784 setOperationAction(ISD::UREM, MVT::i32, Expand);
785 // Register based DivRem for AEABI (RTABI 4.2)
786 if (Subtarget->isTargetAEABI() || Subtarget->isTargetAndroid()) {
787 setOperationAction(ISD::SREM, MVT::i64, Custom);
788 setOperationAction(ISD::UREM, MVT::i64, Custom);
790 setLibcallName(RTLIB::SDIVREM_I8, "__aeabi_idivmod");
791 setLibcallName(RTLIB::SDIVREM_I16, "__aeabi_idivmod");
792 setLibcallName(RTLIB::SDIVREM_I32, "__aeabi_idivmod");
793 setLibcallName(RTLIB::SDIVREM_I64, "__aeabi_ldivmod");
794 setLibcallName(RTLIB::UDIVREM_I8, "__aeabi_uidivmod");
795 setLibcallName(RTLIB::UDIVREM_I16, "__aeabi_uidivmod");
796 setLibcallName(RTLIB::UDIVREM_I32, "__aeabi_uidivmod");
797 setLibcallName(RTLIB::UDIVREM_I64, "__aeabi_uldivmod");
799 setLibcallCallingConv(RTLIB::SDIVREM_I8, CallingConv::ARM_AAPCS);
800 setLibcallCallingConv(RTLIB::SDIVREM_I16, CallingConv::ARM_AAPCS);
801 setLibcallCallingConv(RTLIB::SDIVREM_I32, CallingConv::ARM_AAPCS);
802 setLibcallCallingConv(RTLIB::SDIVREM_I64, CallingConv::ARM_AAPCS);
803 setLibcallCallingConv(RTLIB::UDIVREM_I8, CallingConv::ARM_AAPCS);
804 setLibcallCallingConv(RTLIB::UDIVREM_I16, CallingConv::ARM_AAPCS);
805 setLibcallCallingConv(RTLIB::UDIVREM_I32, CallingConv::ARM_AAPCS);
806 setLibcallCallingConv(RTLIB::UDIVREM_I64, CallingConv::ARM_AAPCS);
808 setOperationAction(ISD::SDIVREM, MVT::i32, Custom);
809 setOperationAction(ISD::UDIVREM, MVT::i32, Custom);
811 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
812 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
815 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
816 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
817 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
818 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
820 setOperationAction(ISD::TRAP, MVT::Other, Legal);
822 // Use the default implementation.
823 setOperationAction(ISD::VASTART, MVT::Other, Custom);
824 setOperationAction(ISD::VAARG, MVT::Other, Expand);
825 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
826 setOperationAction(ISD::VAEND, MVT::Other, Expand);
827 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
828 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
830 if (Subtarget->getTargetTriple().isWindowsItaniumEnvironment())
831 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
833 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
835 // ARMv6 Thumb1 (except for CPUs that support dmb / dsb) and earlier use
836 // the default expansion. If we are targeting a single threaded system,
837 // then set them all for expand so we can lower them later into their
839 if (TM.Options.ThreadModel == ThreadModel::Single)
840 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Expand);
841 else if (Subtarget->hasAnyDataBarrier() && !Subtarget->isThumb1Only()) {
842 // ATOMIC_FENCE needs custom lowering; the others should have been expanded
843 // to ldrex/strex loops already.
844 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Custom);
846 // On v8, we have particularly efficient implementations of atomic fences
847 // if they can be combined with nearby atomic loads and stores.
848 if (!Subtarget->hasV8Ops()) {
849 // Automatically insert fences (dmb ish) around ATOMIC_SWAP etc.
850 setInsertFencesForAtomic(true);
853 // If there's anything we can use as a barrier, go through custom lowering
855 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other,
856 Subtarget->hasAnyDataBarrier() ? Custom : Expand);
858 // Set them all for expansion, which will force libcalls.
859 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Expand);
860 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Expand);
861 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Expand);
862 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Expand);
863 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Expand);
864 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Expand);
865 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Expand);
866 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Expand);
867 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i32, Expand);
868 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i32, Expand);
869 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i32, Expand);
870 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i32, Expand);
871 // Mark ATOMIC_LOAD and ATOMIC_STORE custom so we can handle the
872 // Unordered/Monotonic case.
873 setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Custom);
874 setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Custom);
877 setOperationAction(ISD::PREFETCH, MVT::Other, Custom);
879 // Requires SXTB/SXTH, available on v6 and up in both ARM and Thumb modes.
880 if (!Subtarget->hasV6Ops()) {
881 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
882 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
884 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
886 if (!Subtarget->useSoftFloat() && Subtarget->hasVFP2() &&
887 !Subtarget->isThumb1Only()) {
888 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
889 // iff target supports vfp2.
890 setOperationAction(ISD::BITCAST, MVT::i64, Custom);
891 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
894 // We want to custom lower some of our intrinsics.
895 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
896 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
897 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
898 setOperationAction(ISD::EH_SJLJ_SETUP_DISPATCH, MVT::Other, Custom);
899 if (Subtarget->useSjLjEH())
900 setLibcallName(RTLIB::UNWIND_RESUME, "_Unwind_SjLj_Resume");
902 setOperationAction(ISD::SETCC, MVT::i32, Expand);
903 setOperationAction(ISD::SETCC, MVT::f32, Expand);
904 setOperationAction(ISD::SETCC, MVT::f64, Expand);
905 setOperationAction(ISD::SELECT, MVT::i32, Custom);
906 setOperationAction(ISD::SELECT, MVT::f32, Custom);
907 setOperationAction(ISD::SELECT, MVT::f64, Custom);
908 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
909 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
910 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
912 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
913 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
914 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
915 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
916 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
918 // We don't support sin/cos/fmod/copysign/pow
919 setOperationAction(ISD::FSIN, MVT::f64, Expand);
920 setOperationAction(ISD::FSIN, MVT::f32, Expand);
921 setOperationAction(ISD::FCOS, MVT::f32, Expand);
922 setOperationAction(ISD::FCOS, MVT::f64, Expand);
923 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
924 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
925 setOperationAction(ISD::FREM, MVT::f64, Expand);
926 setOperationAction(ISD::FREM, MVT::f32, Expand);
927 if (!Subtarget->useSoftFloat() && Subtarget->hasVFP2() &&
928 !Subtarget->isThumb1Only()) {
929 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
930 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
932 setOperationAction(ISD::FPOW, MVT::f64, Expand);
933 setOperationAction(ISD::FPOW, MVT::f32, Expand);
935 if (!Subtarget->hasVFP4()) {
936 setOperationAction(ISD::FMA, MVT::f64, Expand);
937 setOperationAction(ISD::FMA, MVT::f32, Expand);
940 // Various VFP goodness
941 if (!Subtarget->useSoftFloat() && !Subtarget->isThumb1Only()) {
942 // FP-ARMv8 adds f64 <-> f16 conversion. Before that it should be expanded.
943 if (!Subtarget->hasFPARMv8() || Subtarget->isFPOnlySP()) {
944 setOperationAction(ISD::FP16_TO_FP, MVT::f64, Expand);
945 setOperationAction(ISD::FP_TO_FP16, MVT::f64, Expand);
948 // fp16 is a special v7 extension that adds f16 <-> f32 conversions.
949 if (!Subtarget->hasFP16()) {
950 setOperationAction(ISD::FP16_TO_FP, MVT::f32, Expand);
951 setOperationAction(ISD::FP_TO_FP16, MVT::f32, Expand);
955 // Combine sin / cos into one node or libcall if possible.
956 if (Subtarget->hasSinCos()) {
957 setLibcallName(RTLIB::SINCOS_F32, "sincosf");
958 setLibcallName(RTLIB::SINCOS_F64, "sincos");
959 if (Subtarget->isTargetWatchOS()) {
960 setLibcallCallingConv(RTLIB::SINCOS_F32, CallingConv::ARM_AAPCS_VFP);
961 setLibcallCallingConv(RTLIB::SINCOS_F64, CallingConv::ARM_AAPCS_VFP);
963 if (Subtarget->isTargetIOS() || Subtarget->isTargetWatchOS()) {
964 // For iOS, we don't want to the normal expansion of a libcall to
965 // sincos. We want to issue a libcall to __sincos_stret.
966 setOperationAction(ISD::FSINCOS, MVT::f64, Custom);
967 setOperationAction(ISD::FSINCOS, MVT::f32, Custom);
971 // FP-ARMv8 implements a lot of rounding-like FP operations.
972 if (Subtarget->hasFPARMv8()) {
973 setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
974 setOperationAction(ISD::FCEIL, MVT::f32, Legal);
975 setOperationAction(ISD::FROUND, MVT::f32, Legal);
976 setOperationAction(ISD::FTRUNC, MVT::f32, Legal);
977 setOperationAction(ISD::FNEARBYINT, MVT::f32, Legal);
978 setOperationAction(ISD::FRINT, MVT::f32, Legal);
979 setOperationAction(ISD::FMINNUM, MVT::f32, Legal);
980 setOperationAction(ISD::FMAXNUM, MVT::f32, Legal);
981 setOperationAction(ISD::FMINNUM, MVT::v2f32, Legal);
982 setOperationAction(ISD::FMAXNUM, MVT::v2f32, Legal);
983 setOperationAction(ISD::FMINNUM, MVT::v4f32, Legal);
984 setOperationAction(ISD::FMAXNUM, MVT::v4f32, Legal);
986 if (!Subtarget->isFPOnlySP()) {
987 setOperationAction(ISD::FFLOOR, MVT::f64, Legal);
988 setOperationAction(ISD::FCEIL, MVT::f64, Legal);
989 setOperationAction(ISD::FROUND, MVT::f64, Legal);
990 setOperationAction(ISD::FTRUNC, MVT::f64, Legal);
991 setOperationAction(ISD::FNEARBYINT, MVT::f64, Legal);
992 setOperationAction(ISD::FRINT, MVT::f64, Legal);
993 setOperationAction(ISD::FMINNUM, MVT::f64, Legal);
994 setOperationAction(ISD::FMAXNUM, MVT::f64, Legal);
998 if (Subtarget->hasNEON()) {
999 // vmin and vmax aren't available in a scalar form, so we use
1000 // a NEON instruction with an undef lane instead.
1001 setOperationAction(ISD::FMINNAN, MVT::f32, Legal);
1002 setOperationAction(ISD::FMAXNAN, MVT::f32, Legal);
1003 setOperationAction(ISD::FMINNAN, MVT::v2f32, Legal);
1004 setOperationAction(ISD::FMAXNAN, MVT::v2f32, Legal);
1005 setOperationAction(ISD::FMINNAN, MVT::v4f32, Legal);
1006 setOperationAction(ISD::FMAXNAN, MVT::v4f32, Legal);
1009 // We have target-specific dag combine patterns for the following nodes:
1010 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
1011 setTargetDAGCombine(ISD::ADD);
1012 setTargetDAGCombine(ISD::SUB);
1013 setTargetDAGCombine(ISD::MUL);
1014 setTargetDAGCombine(ISD::AND);
1015 setTargetDAGCombine(ISD::OR);
1016 setTargetDAGCombine(ISD::XOR);
1018 if (Subtarget->hasV6Ops())
1019 setTargetDAGCombine(ISD::SRL);
1021 setStackPointerRegisterToSaveRestore(ARM::SP);
1023 if (Subtarget->useSoftFloat() || Subtarget->isThumb1Only() ||
1024 !Subtarget->hasVFP2())
1025 setSchedulingPreference(Sched::RegPressure);
1027 setSchedulingPreference(Sched::Hybrid);
1029 //// temporary - rewrite interface to use type
1030 MaxStoresPerMemset = 8;
1031 MaxStoresPerMemsetOptSize = 4;
1032 MaxStoresPerMemcpy = 4; // For @llvm.memcpy -> sequence of stores
1033 MaxStoresPerMemcpyOptSize = 2;
1034 MaxStoresPerMemmove = 4; // For @llvm.memmove -> sequence of stores
1035 MaxStoresPerMemmoveOptSize = 2;
1037 // On ARM arguments smaller than 4 bytes are extended, so all arguments
1038 // are at least 4 bytes aligned.
1039 setMinStackArgumentAlignment(4);
1041 // Prefer likely predicted branches to selects on out-of-order cores.
1042 PredictableSelectIsExpensive = Subtarget->isLikeA9();
1044 setMinFunctionAlignment(Subtarget->isThumb() ? 1 : 2);
1047 bool ARMTargetLowering::useSoftFloat() const {
1048 return Subtarget->useSoftFloat();
1051 // FIXME: It might make sense to define the representative register class as the
1052 // nearest super-register that has a non-null superset. For example, DPR_VFP2 is
1053 // a super-register of SPR, and DPR is a superset if DPR_VFP2. Consequently,
1054 // SPR's representative would be DPR_VFP2. This should work well if register
1055 // pressure tracking were modified such that a register use would increment the
1056 // pressure of the register class's representative and all of it's super
1057 // classes' representatives transitively. We have not implemented this because
1058 // of the difficulty prior to coalescing of modeling operand register classes
1059 // due to the common occurrence of cross class copies and subregister insertions
1061 std::pair<const TargetRegisterClass *, uint8_t>
1062 ARMTargetLowering::findRepresentativeClass(const TargetRegisterInfo *TRI,
1064 const TargetRegisterClass *RRC = nullptr;
1066 switch (VT.SimpleTy) {
1068 return TargetLowering::findRepresentativeClass(TRI, VT);
1069 // Use DPR as representative register class for all floating point
1070 // and vector types. Since there are 32 SPR registers and 32 DPR registers so
1071 // the cost is 1 for both f32 and f64.
1072 case MVT::f32: case MVT::f64: case MVT::v8i8: case MVT::v4i16:
1073 case MVT::v2i32: case MVT::v1i64: case MVT::v2f32:
1074 RRC = &ARM::DPRRegClass;
1075 // When NEON is used for SP, only half of the register file is available
1076 // because operations that define both SP and DP results will be constrained
1077 // to the VFP2 class (D0-D15). We currently model this constraint prior to
1078 // coalescing by double-counting the SP regs. See the FIXME above.
1079 if (Subtarget->useNEONForSinglePrecisionFP())
1082 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
1083 case MVT::v4f32: case MVT::v2f64:
1084 RRC = &ARM::DPRRegClass;
1088 RRC = &ARM::DPRRegClass;
1092 RRC = &ARM::DPRRegClass;
1096 return std::make_pair(RRC, Cost);
1099 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
1100 switch ((ARMISD::NodeType)Opcode) {
1101 case ARMISD::FIRST_NUMBER: break;
1102 case ARMISD::Wrapper: return "ARMISD::Wrapper";
1103 case ARMISD::WrapperPIC: return "ARMISD::WrapperPIC";
1104 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
1105 case ARMISD::COPY_STRUCT_BYVAL: return "ARMISD::COPY_STRUCT_BYVAL";
1106 case ARMISD::CALL: return "ARMISD::CALL";
1107 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
1108 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
1109 case ARMISD::tCALL: return "ARMISD::tCALL";
1110 case ARMISD::BRCOND: return "ARMISD::BRCOND";
1111 case ARMISD::BR_JT: return "ARMISD::BR_JT";
1112 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
1113 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
1114 case ARMISD::INTRET_FLAG: return "ARMISD::INTRET_FLAG";
1115 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
1116 case ARMISD::CMP: return "ARMISD::CMP";
1117 case ARMISD::CMN: return "ARMISD::CMN";
1118 case ARMISD::CMPZ: return "ARMISD::CMPZ";
1119 case ARMISD::CMPFP: return "ARMISD::CMPFP";
1120 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
1121 case ARMISD::BCC_i64: return "ARMISD::BCC_i64";
1122 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
1124 case ARMISD::CMOV: return "ARMISD::CMOV";
1126 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
1127 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
1128 case ARMISD::RRX: return "ARMISD::RRX";
1130 case ARMISD::ADDC: return "ARMISD::ADDC";
1131 case ARMISD::ADDE: return "ARMISD::ADDE";
1132 case ARMISD::SUBC: return "ARMISD::SUBC";
1133 case ARMISD::SUBE: return "ARMISD::SUBE";
1135 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
1136 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
1138 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
1139 case ARMISD::EH_SJLJ_LONGJMP: return "ARMISD::EH_SJLJ_LONGJMP";
1140 case ARMISD::EH_SJLJ_SETUP_DISPATCH: return "ARMISD::EH_SJLJ_SETUP_DISPATCH";
1142 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN";
1144 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
1146 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
1148 case ARMISD::MEMBARRIER_MCR: return "ARMISD::MEMBARRIER_MCR";
1150 case ARMISD::PRELOAD: return "ARMISD::PRELOAD";
1152 case ARMISD::WIN__CHKSTK: return "ARMISD:::WIN__CHKSTK";
1153 case ARMISD::WIN__DBZCHK: return "ARMISD::WIN__DBZCHK";
1155 case ARMISD::VCEQ: return "ARMISD::VCEQ";
1156 case ARMISD::VCEQZ: return "ARMISD::VCEQZ";
1157 case ARMISD::VCGE: return "ARMISD::VCGE";
1158 case ARMISD::VCGEZ: return "ARMISD::VCGEZ";
1159 case ARMISD::VCLEZ: return "ARMISD::VCLEZ";
1160 case ARMISD::VCGEU: return "ARMISD::VCGEU";
1161 case ARMISD::VCGT: return "ARMISD::VCGT";
1162 case ARMISD::VCGTZ: return "ARMISD::VCGTZ";
1163 case ARMISD::VCLTZ: return "ARMISD::VCLTZ";
1164 case ARMISD::VCGTU: return "ARMISD::VCGTU";
1165 case ARMISD::VTST: return "ARMISD::VTST";
1167 case ARMISD::VSHL: return "ARMISD::VSHL";
1168 case ARMISD::VSHRs: return "ARMISD::VSHRs";
1169 case ARMISD::VSHRu: return "ARMISD::VSHRu";
1170 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
1171 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
1172 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
1173 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
1174 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
1175 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
1176 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
1177 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
1178 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
1179 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
1180 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
1181 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
1182 case ARMISD::VSLI: return "ARMISD::VSLI";
1183 case ARMISD::VSRI: return "ARMISD::VSRI";
1184 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
1185 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
1186 case ARMISD::VMOVIMM: return "ARMISD::VMOVIMM";
1187 case ARMISD::VMVNIMM: return "ARMISD::VMVNIMM";
1188 case ARMISD::VMOVFPIMM: return "ARMISD::VMOVFPIMM";
1189 case ARMISD::VDUP: return "ARMISD::VDUP";
1190 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
1191 case ARMISD::VEXT: return "ARMISD::VEXT";
1192 case ARMISD::VREV64: return "ARMISD::VREV64";
1193 case ARMISD::VREV32: return "ARMISD::VREV32";
1194 case ARMISD::VREV16: return "ARMISD::VREV16";
1195 case ARMISD::VZIP: return "ARMISD::VZIP";
1196 case ARMISD::VUZP: return "ARMISD::VUZP";
1197 case ARMISD::VTRN: return "ARMISD::VTRN";
1198 case ARMISD::VTBL1: return "ARMISD::VTBL1";
1199 case ARMISD::VTBL2: return "ARMISD::VTBL2";
1200 case ARMISD::VMULLs: return "ARMISD::VMULLs";
1201 case ARMISD::VMULLu: return "ARMISD::VMULLu";
1202 case ARMISD::UMLAL: return "ARMISD::UMLAL";
1203 case ARMISD::SMLAL: return "ARMISD::SMLAL";
1204 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
1205 case ARMISD::BFI: return "ARMISD::BFI";
1206 case ARMISD::VORRIMM: return "ARMISD::VORRIMM";
1207 case ARMISD::VBICIMM: return "ARMISD::VBICIMM";
1208 case ARMISD::VBSL: return "ARMISD::VBSL";
1209 case ARMISD::MEMCPY: return "ARMISD::MEMCPY";
1210 case ARMISD::VLD2DUP: return "ARMISD::VLD2DUP";
1211 case ARMISD::VLD3DUP: return "ARMISD::VLD3DUP";
1212 case ARMISD::VLD4DUP: return "ARMISD::VLD4DUP";
1213 case ARMISD::VLD1_UPD: return "ARMISD::VLD1_UPD";
1214 case ARMISD::VLD2_UPD: return "ARMISD::VLD2_UPD";
1215 case ARMISD::VLD3_UPD: return "ARMISD::VLD3_UPD";
1216 case ARMISD::VLD4_UPD: return "ARMISD::VLD4_UPD";
1217 case ARMISD::VLD2LN_UPD: return "ARMISD::VLD2LN_UPD";
1218 case ARMISD::VLD3LN_UPD: return "ARMISD::VLD3LN_UPD";
1219 case ARMISD::VLD4LN_UPD: return "ARMISD::VLD4LN_UPD";
1220 case ARMISD::VLD2DUP_UPD: return "ARMISD::VLD2DUP_UPD";
1221 case ARMISD::VLD3DUP_UPD: return "ARMISD::VLD3DUP_UPD";
1222 case ARMISD::VLD4DUP_UPD: return "ARMISD::VLD4DUP_UPD";
1223 case ARMISD::VST1_UPD: return "ARMISD::VST1_UPD";
1224 case ARMISD::VST2_UPD: return "ARMISD::VST2_UPD";
1225 case ARMISD::VST3_UPD: return "ARMISD::VST3_UPD";
1226 case ARMISD::VST4_UPD: return "ARMISD::VST4_UPD";
1227 case ARMISD::VST2LN_UPD: return "ARMISD::VST2LN_UPD";
1228 case ARMISD::VST3LN_UPD: return "ARMISD::VST3LN_UPD";
1229 case ARMISD::VST4LN_UPD: return "ARMISD::VST4LN_UPD";
1234 EVT ARMTargetLowering::getSetCCResultType(const DataLayout &DL, LLVMContext &,
1237 return getPointerTy(DL);
1238 return VT.changeVectorElementTypeToInteger();
1241 /// getRegClassFor - Return the register class that should be used for the
1242 /// specified value type.
1243 const TargetRegisterClass *ARMTargetLowering::getRegClassFor(MVT VT) const {
1244 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
1245 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
1246 // load / store 4 to 8 consecutive D registers.
1247 if (Subtarget->hasNEON()) {
1248 if (VT == MVT::v4i64)
1249 return &ARM::QQPRRegClass;
1250 if (VT == MVT::v8i64)
1251 return &ARM::QQQQPRRegClass;
1253 return TargetLowering::getRegClassFor(VT);
1256 // memcpy, and other memory intrinsics, typically tries to use LDM/STM if the
1257 // source/dest is aligned and the copy size is large enough. We therefore want
1258 // to align such objects passed to memory intrinsics.
1259 bool ARMTargetLowering::shouldAlignPointerArgs(CallInst *CI, unsigned &MinSize,
1260 unsigned &PrefAlign) const {
1261 if (!isa<MemIntrinsic>(CI))
1264 // On ARM11 onwards (excluding M class) 8-byte aligned LDM is typically 1
1265 // cycle faster than 4-byte aligned LDM.
1266 PrefAlign = (Subtarget->hasV6Ops() && !Subtarget->isMClass() ? 8 : 4);
1270 // Create a fast isel object.
1272 ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
1273 const TargetLibraryInfo *libInfo) const {
1274 return ARM::createFastISel(funcInfo, libInfo);
1277 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
1278 unsigned NumVals = N->getNumValues();
1280 return Sched::RegPressure;
1282 for (unsigned i = 0; i != NumVals; ++i) {
1283 EVT VT = N->getValueType(i);
1284 if (VT == MVT::Glue || VT == MVT::Other)
1286 if (VT.isFloatingPoint() || VT.isVector())
1290 if (!N->isMachineOpcode())
1291 return Sched::RegPressure;
1293 // Load are scheduled for latency even if there instruction itinerary
1294 // is not available.
1295 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
1296 const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
1298 if (MCID.getNumDefs() == 0)
1299 return Sched::RegPressure;
1300 if (!Itins->isEmpty() &&
1301 Itins->getOperandCycle(MCID.getSchedClass(), 0) > 2)
1304 return Sched::RegPressure;
1307 //===----------------------------------------------------------------------===//
1309 //===----------------------------------------------------------------------===//
1311 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
1312 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
1314 default: llvm_unreachable("Unknown condition code!");
1315 case ISD::SETNE: return ARMCC::NE;
1316 case ISD::SETEQ: return ARMCC::EQ;
1317 case ISD::SETGT: return ARMCC::GT;
1318 case ISD::SETGE: return ARMCC::GE;
1319 case ISD::SETLT: return ARMCC::LT;
1320 case ISD::SETLE: return ARMCC::LE;
1321 case ISD::SETUGT: return ARMCC::HI;
1322 case ISD::SETUGE: return ARMCC::HS;
1323 case ISD::SETULT: return ARMCC::LO;
1324 case ISD::SETULE: return ARMCC::LS;
1328 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
1329 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
1330 ARMCC::CondCodes &CondCode2) {
1331 CondCode2 = ARMCC::AL;
1333 default: llvm_unreachable("Unknown FP condition!");
1335 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
1337 case ISD::SETOGT: CondCode = ARMCC::GT; break;
1339 case ISD::SETOGE: CondCode = ARMCC::GE; break;
1340 case ISD::SETOLT: CondCode = ARMCC::MI; break;
1341 case ISD::SETOLE: CondCode = ARMCC::LS; break;
1342 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
1343 case ISD::SETO: CondCode = ARMCC::VC; break;
1344 case ISD::SETUO: CondCode = ARMCC::VS; break;
1345 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
1346 case ISD::SETUGT: CondCode = ARMCC::HI; break;
1347 case ISD::SETUGE: CondCode = ARMCC::PL; break;
1349 case ISD::SETULT: CondCode = ARMCC::LT; break;
1351 case ISD::SETULE: CondCode = ARMCC::LE; break;
1353 case ISD::SETUNE: CondCode = ARMCC::NE; break;
1357 //===----------------------------------------------------------------------===//
1358 // Calling Convention Implementation
1359 //===----------------------------------------------------------------------===//
1361 #include "ARMGenCallingConv.inc"
1363 /// getEffectiveCallingConv - Get the effective calling convention, taking into
1364 /// account presence of floating point hardware and calling convention
1365 /// limitations, such as support for variadic functions.
1367 ARMTargetLowering::getEffectiveCallingConv(CallingConv::ID CC,
1368 bool isVarArg) const {
1371 llvm_unreachable("Unsupported calling convention");
1372 case CallingConv::ARM_AAPCS:
1373 case CallingConv::ARM_APCS:
1374 case CallingConv::GHC:
1376 case CallingConv::ARM_AAPCS_VFP:
1377 return isVarArg ? CallingConv::ARM_AAPCS : CallingConv::ARM_AAPCS_VFP;
1378 case CallingConv::C:
1379 if (!Subtarget->isAAPCS_ABI())
1380 return CallingConv::ARM_APCS;
1381 else if (Subtarget->hasVFP2() && !Subtarget->isThumb1Only() &&
1382 getTargetMachine().Options.FloatABIType == FloatABI::Hard &&
1384 return CallingConv::ARM_AAPCS_VFP;
1386 return CallingConv::ARM_AAPCS;
1387 case CallingConv::Fast:
1388 if (!Subtarget->isAAPCS_ABI()) {
1389 if (Subtarget->hasVFP2() && !Subtarget->isThumb1Only() && !isVarArg)
1390 return CallingConv::Fast;
1391 return CallingConv::ARM_APCS;
1392 } else if (Subtarget->hasVFP2() && !Subtarget->isThumb1Only() && !isVarArg)
1393 return CallingConv::ARM_AAPCS_VFP;
1395 return CallingConv::ARM_AAPCS;
1399 /// CCAssignFnForNode - Selects the correct CCAssignFn for the given
1400 /// CallingConvention.
1401 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
1403 bool isVarArg) const {
1404 switch (getEffectiveCallingConv(CC, isVarArg)) {
1406 llvm_unreachable("Unsupported calling convention");
1407 case CallingConv::ARM_APCS:
1408 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1409 case CallingConv::ARM_AAPCS:
1410 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1411 case CallingConv::ARM_AAPCS_VFP:
1412 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1413 case CallingConv::Fast:
1414 return (Return ? RetFastCC_ARM_APCS : FastCC_ARM_APCS);
1415 case CallingConv::GHC:
1416 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS_GHC);
1420 /// LowerCallResult - Lower the result values of a call into the
1421 /// appropriate copies out of appropriate physical registers.
1423 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1424 CallingConv::ID CallConv, bool isVarArg,
1425 const SmallVectorImpl<ISD::InputArg> &Ins,
1426 SDLoc dl, SelectionDAG &DAG,
1427 SmallVectorImpl<SDValue> &InVals,
1428 bool isThisReturn, SDValue ThisVal) const {
1430 // Assign locations to each value returned by this call.
1431 SmallVector<CCValAssign, 16> RVLocs;
1432 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
1433 *DAG.getContext(), Call);
1434 CCInfo.AnalyzeCallResult(Ins,
1435 CCAssignFnForNode(CallConv, /* Return*/ true,
1438 // Copy all of the result registers out of their specified physreg.
1439 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1440 CCValAssign VA = RVLocs[i];
1442 // Pass 'this' value directly from the argument to return value, to avoid
1443 // reg unit interference
1444 if (i == 0 && isThisReturn) {
1445 assert(!VA.needsCustom() && VA.getLocVT() == MVT::i32 &&
1446 "unexpected return calling convention register assignment");
1447 InVals.push_back(ThisVal);
1452 if (VA.needsCustom()) {
1453 // Handle f64 or half of a v2f64.
1454 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1456 Chain = Lo.getValue(1);
1457 InFlag = Lo.getValue(2);
1458 VA = RVLocs[++i]; // skip ahead to next loc
1459 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1461 Chain = Hi.getValue(1);
1462 InFlag = Hi.getValue(2);
1463 if (!Subtarget->isLittle())
1465 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1467 if (VA.getLocVT() == MVT::v2f64) {
1468 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1469 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1470 DAG.getConstant(0, dl, MVT::i32));
1472 VA = RVLocs[++i]; // skip ahead to next loc
1473 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1474 Chain = Lo.getValue(1);
1475 InFlag = Lo.getValue(2);
1476 VA = RVLocs[++i]; // skip ahead to next loc
1477 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1478 Chain = Hi.getValue(1);
1479 InFlag = Hi.getValue(2);
1480 if (!Subtarget->isLittle())
1482 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1483 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1484 DAG.getConstant(1, dl, MVT::i32));
1487 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
1489 Chain = Val.getValue(1);
1490 InFlag = Val.getValue(2);
1493 switch (VA.getLocInfo()) {
1494 default: llvm_unreachable("Unknown loc info!");
1495 case CCValAssign::Full: break;
1496 case CCValAssign::BCvt:
1497 Val = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), Val);
1501 InVals.push_back(Val);
1507 /// LowerMemOpCallTo - Store the argument to the stack.
1509 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
1510 SDValue StackPtr, SDValue Arg,
1511 SDLoc dl, SelectionDAG &DAG,
1512 const CCValAssign &VA,
1513 ISD::ArgFlagsTy Flags) const {
1514 unsigned LocMemOffset = VA.getLocMemOffset();
1515 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset, dl);
1516 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(DAG.getDataLayout()),
1518 return DAG.getStore(
1519 Chain, dl, Arg, PtrOff,
1520 MachinePointerInfo::getStack(DAG.getMachineFunction(), LocMemOffset),
1524 void ARMTargetLowering::PassF64ArgInRegs(SDLoc dl, SelectionDAG &DAG,
1525 SDValue Chain, SDValue &Arg,
1526 RegsToPassVector &RegsToPass,
1527 CCValAssign &VA, CCValAssign &NextVA,
1529 SmallVectorImpl<SDValue> &MemOpChains,
1530 ISD::ArgFlagsTy Flags) const {
1532 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1533 DAG.getVTList(MVT::i32, MVT::i32), Arg);
1534 unsigned id = Subtarget->isLittle() ? 0 : 1;
1535 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd.getValue(id)));
1537 if (NextVA.isRegLoc())
1538 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1-id)));
1540 assert(NextVA.isMemLoc());
1541 if (!StackPtr.getNode())
1542 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP,
1543 getPointerTy(DAG.getDataLayout()));
1545 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1-id),
1551 /// LowerCall - Lowering a call into a callseq_start <-
1552 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
1555 ARMTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
1556 SmallVectorImpl<SDValue> &InVals) const {
1557 SelectionDAG &DAG = CLI.DAG;
1559 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
1560 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
1561 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
1562 SDValue Chain = CLI.Chain;
1563 SDValue Callee = CLI.Callee;
1564 bool &isTailCall = CLI.IsTailCall;
1565 CallingConv::ID CallConv = CLI.CallConv;
1566 bool doesNotRet = CLI.DoesNotReturn;
1567 bool isVarArg = CLI.IsVarArg;
1569 MachineFunction &MF = DAG.getMachineFunction();
1570 bool isStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1571 bool isThisReturn = false;
1572 bool isSibCall = false;
1573 auto Attr = MF.getFunction()->getFnAttribute("disable-tail-calls");
1575 // Disable tail calls if they're not supported.
1576 if (!Subtarget->supportsTailCall() || Attr.getValueAsString() == "true")
1580 // Check if it's really possible to do a tail call.
1581 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1582 isVarArg, isStructRet, MF.getFunction()->hasStructRetAttr(),
1583 Outs, OutVals, Ins, DAG);
1584 if (!isTailCall && CLI.CS && CLI.CS->isMustTailCall())
1585 report_fatal_error("failed to perform tail call elimination on a call "
1586 "site marked musttail");
1587 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1588 // detected sibcalls.
1595 // Analyze operands of the call, assigning locations to each operand.
1596 SmallVector<CCValAssign, 16> ArgLocs;
1597 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs,
1598 *DAG.getContext(), Call);
1599 CCInfo.AnalyzeCallOperands(Outs,
1600 CCAssignFnForNode(CallConv, /* Return*/ false,
1603 // Get a count of how many bytes are to be pushed on the stack.
1604 unsigned NumBytes = CCInfo.getNextStackOffset();
1606 // For tail calls, memory operands are available in our caller's stack.
1610 // Adjust the stack pointer for the new arguments...
1611 // These operations are automatically eliminated by the prolog/epilog pass
1613 Chain = DAG.getCALLSEQ_START(Chain,
1614 DAG.getIntPtrConstant(NumBytes, dl, true), dl);
1617 DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy(DAG.getDataLayout()));
1619 RegsToPassVector RegsToPass;
1620 SmallVector<SDValue, 8> MemOpChains;
1622 // Walk the register/memloc assignments, inserting copies/loads. In the case
1623 // of tail call optimization, arguments are handled later.
1624 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1626 ++i, ++realArgIdx) {
1627 CCValAssign &VA = ArgLocs[i];
1628 SDValue Arg = OutVals[realArgIdx];
1629 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1630 bool isByVal = Flags.isByVal();
1632 // Promote the value if needed.
1633 switch (VA.getLocInfo()) {
1634 default: llvm_unreachable("Unknown loc info!");
1635 case CCValAssign::Full: break;
1636 case CCValAssign::SExt:
1637 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1639 case CCValAssign::ZExt:
1640 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1642 case CCValAssign::AExt:
1643 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1645 case CCValAssign::BCvt:
1646 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1650 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1651 if (VA.needsCustom()) {
1652 if (VA.getLocVT() == MVT::v2f64) {
1653 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1654 DAG.getConstant(0, dl, MVT::i32));
1655 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1656 DAG.getConstant(1, dl, MVT::i32));
1658 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1659 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1661 VA = ArgLocs[++i]; // skip ahead to next loc
1662 if (VA.isRegLoc()) {
1663 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1664 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1666 assert(VA.isMemLoc());
1668 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1669 dl, DAG, VA, Flags));
1672 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1673 StackPtr, MemOpChains, Flags);
1675 } else if (VA.isRegLoc()) {
1676 if (realArgIdx == 0 && Flags.isReturned() && Outs[0].VT == MVT::i32) {
1677 assert(VA.getLocVT() == MVT::i32 &&
1678 "unexpected calling convention register assignment");
1679 assert(!Ins.empty() && Ins[0].VT == MVT::i32 &&
1680 "unexpected use of 'returned'");
1681 isThisReturn = true;
1683 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1684 } else if (isByVal) {
1685 assert(VA.isMemLoc());
1686 unsigned offset = 0;
1688 // True if this byval aggregate will be split between registers
1690 unsigned ByValArgsCount = CCInfo.getInRegsParamsCount();
1691 unsigned CurByValIdx = CCInfo.getInRegsParamsProcessed();
1693 if (CurByValIdx < ByValArgsCount) {
1695 unsigned RegBegin, RegEnd;
1696 CCInfo.getInRegsParamInfo(CurByValIdx, RegBegin, RegEnd);
1699 DAG.getTargetLoweringInfo().getPointerTy(DAG.getDataLayout());
1701 for (i = 0, j = RegBegin; j < RegEnd; i++, j++) {
1702 SDValue Const = DAG.getConstant(4*i, dl, MVT::i32);
1703 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
1704 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
1705 MachinePointerInfo(),
1706 false, false, false,
1707 DAG.InferPtrAlignment(AddArg));
1708 MemOpChains.push_back(Load.getValue(1));
1709 RegsToPass.push_back(std::make_pair(j, Load));
1712 // If parameter size outsides register area, "offset" value
1713 // helps us to calculate stack slot for remained part properly.
1714 offset = RegEnd - RegBegin;
1716 CCInfo.nextInRegsParam();
1719 if (Flags.getByValSize() > 4*offset) {
1720 auto PtrVT = getPointerTy(DAG.getDataLayout());
1721 unsigned LocMemOffset = VA.getLocMemOffset();
1722 SDValue StkPtrOff = DAG.getIntPtrConstant(LocMemOffset, dl);
1723 SDValue Dst = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, StkPtrOff);
1724 SDValue SrcOffset = DAG.getIntPtrConstant(4*offset, dl);
1725 SDValue Src = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, SrcOffset);
1726 SDValue SizeNode = DAG.getConstant(Flags.getByValSize() - 4*offset, dl,
1728 SDValue AlignNode = DAG.getConstant(Flags.getByValAlign(), dl,
1731 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
1732 SDValue Ops[] = { Chain, Dst, Src, SizeNode, AlignNode};
1733 MemOpChains.push_back(DAG.getNode(ARMISD::COPY_STRUCT_BYVAL, dl, VTs,
1736 } else if (!isSibCall) {
1737 assert(VA.isMemLoc());
1739 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1740 dl, DAG, VA, Flags));
1744 if (!MemOpChains.empty())
1745 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains);
1747 // Build a sequence of copy-to-reg nodes chained together with token chain
1748 // and flag operands which copy the outgoing args into the appropriate regs.
1750 // Tail call byval lowering might overwrite argument registers so in case of
1751 // tail call optimization the copies to registers are lowered later.
1753 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1754 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1755 RegsToPass[i].second, InFlag);
1756 InFlag = Chain.getValue(1);
1759 // For tail calls lower the arguments to the 'real' stack slot.
1761 // Force all the incoming stack arguments to be loaded from the stack
1762 // before any new outgoing arguments are stored to the stack, because the
1763 // outgoing stack slots may alias the incoming argument stack slots, and
1764 // the alias isn't otherwise explicit. This is slightly more conservative
1765 // than necessary, because it means that each store effectively depends
1766 // on every argument instead of just those arguments it would clobber.
1768 // Do not flag preceding copytoreg stuff together with the following stuff.
1770 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1771 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1772 RegsToPass[i].second, InFlag);
1773 InFlag = Chain.getValue(1);
1778 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1779 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1780 // node so that legalize doesn't hack it.
1781 bool isDirect = false;
1782 bool isARMFunc = false;
1783 bool isLocalARMFunc = false;
1784 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1785 auto PtrVt = getPointerTy(DAG.getDataLayout());
1787 if (Subtarget->genLongCalls()) {
1788 assert((Subtarget->isTargetWindows() ||
1789 getTargetMachine().getRelocationModel() == Reloc::Static) &&
1790 "long-calls with non-static relocation model!");
1791 // Handle a global address or an external symbol. If it's not one of
1792 // those, the target's already in a register, so we don't need to do
1794 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1795 const GlobalValue *GV = G->getGlobal();
1796 // Create a constant pool entry for the callee address
1797 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1798 ARMConstantPoolValue *CPV =
1799 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 0);
1801 // Get the address of the callee into a register
1802 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVt, 4);
1803 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1804 Callee = DAG.getLoad(
1805 PtrVt, dl, DAG.getEntryNode(), CPAddr,
1806 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), false,
1808 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1809 const char *Sym = S->getSymbol();
1811 // Create a constant pool entry for the callee address
1812 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1813 ARMConstantPoolValue *CPV =
1814 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1815 ARMPCLabelIndex, 0);
1816 // Get the address of the callee into a register
1817 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVt, 4);
1818 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1819 Callee = DAG.getLoad(
1820 PtrVt, dl, DAG.getEntryNode(), CPAddr,
1821 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), false,
1824 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1825 const GlobalValue *GV = G->getGlobal();
1827 bool isDef = GV->isStrongDefinitionForLinker();
1828 bool isStub = (!isDef && Subtarget->isTargetMachO()) &&
1829 getTargetMachine().getRelocationModel() != Reloc::Static;
1830 isARMFunc = !Subtarget->isThumb() || (isStub && !Subtarget->isMClass());
1831 // ARM call to a local ARM function is predicable.
1832 isLocalARMFunc = !Subtarget->isThumb() && (isDef || !ARMInterworking);
1833 // tBX takes a register source operand.
1834 if (isStub && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1835 assert(Subtarget->isTargetMachO() && "WrapperPIC use on non-MachO?");
1836 Callee = DAG.getNode(
1837 ARMISD::WrapperPIC, dl, PtrVt,
1838 DAG.getTargetGlobalAddress(GV, dl, PtrVt, 0, ARMII::MO_NONLAZY));
1839 Callee = DAG.getLoad(PtrVt, dl, DAG.getEntryNode(), Callee,
1840 MachinePointerInfo::getGOT(DAG.getMachineFunction()),
1841 false, false, true, 0);
1842 } else if (Subtarget->isTargetCOFF()) {
1843 assert(Subtarget->isTargetWindows() &&
1844 "Windows is the only supported COFF target");
1845 unsigned TargetFlags = GV->hasDLLImportStorageClass()
1846 ? ARMII::MO_DLLIMPORT
1847 : ARMII::MO_NO_FLAG;
1849 DAG.getTargetGlobalAddress(GV, dl, PtrVt, /*Offset=*/0, TargetFlags);
1850 if (GV->hasDLLImportStorageClass())
1852 DAG.getLoad(PtrVt, dl, DAG.getEntryNode(),
1853 DAG.getNode(ARMISD::Wrapper, dl, PtrVt, Callee),
1854 MachinePointerInfo::getGOT(DAG.getMachineFunction()),
1855 false, false, false, 0);
1857 // On ELF targets for PIC code, direct calls should go through the PLT
1858 unsigned OpFlags = 0;
1859 if (Subtarget->isTargetELF() &&
1860 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1861 OpFlags = ARMII::MO_PLT;
1862 Callee = DAG.getTargetGlobalAddress(GV, dl, PtrVt, 0, OpFlags);
1864 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1866 bool isStub = Subtarget->isTargetMachO() &&
1867 getTargetMachine().getRelocationModel() != Reloc::Static;
1868 isARMFunc = !Subtarget->isThumb() || (isStub && !Subtarget->isMClass());
1869 // tBX takes a register source operand.
1870 const char *Sym = S->getSymbol();
1871 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1872 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1873 ARMConstantPoolValue *CPV =
1874 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1875 ARMPCLabelIndex, 4);
1876 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVt, 4);
1877 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1878 Callee = DAG.getLoad(
1879 PtrVt, dl, DAG.getEntryNode(), CPAddr,
1880 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), false,
1882 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, dl, MVT::i32);
1883 Callee = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVt, Callee, PICLabel);
1885 unsigned OpFlags = 0;
1886 // On ELF targets for PIC code, direct calls should go through the PLT
1887 if (Subtarget->isTargetELF() &&
1888 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1889 OpFlags = ARMII::MO_PLT;
1890 Callee = DAG.getTargetExternalSymbol(Sym, PtrVt, OpFlags);
1894 // FIXME: handle tail calls differently.
1896 if (Subtarget->isThumb()) {
1897 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1898 CallOpc = ARMISD::CALL_NOLINK;
1900 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1902 if (!isDirect && !Subtarget->hasV5TOps())
1903 CallOpc = ARMISD::CALL_NOLINK;
1904 else if (doesNotRet && isDirect && Subtarget->hasRAS() &&
1905 // Emit regular call when code size is the priority
1906 !MF.getFunction()->optForMinSize())
1907 // "mov lr, pc; b _foo" to avoid confusing the RSP
1908 CallOpc = ARMISD::CALL_NOLINK;
1910 CallOpc = isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL;
1913 std::vector<SDValue> Ops;
1914 Ops.push_back(Chain);
1915 Ops.push_back(Callee);
1917 // Add argument registers to the end of the list so that they are known live
1919 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1920 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1921 RegsToPass[i].second.getValueType()));
1923 // Add a register mask operand representing the call-preserved registers.
1925 const uint32_t *Mask;
1926 const ARMBaseRegisterInfo *ARI = Subtarget->getRegisterInfo();
1928 // For 'this' returns, use the R0-preserving mask if applicable
1929 Mask = ARI->getThisReturnPreservedMask(MF, CallConv);
1931 // Set isThisReturn to false if the calling convention is not one that
1932 // allows 'returned' to be modeled in this way, so LowerCallResult does
1933 // not try to pass 'this' straight through
1934 isThisReturn = false;
1935 Mask = ARI->getCallPreservedMask(MF, CallConv);
1938 Mask = ARI->getCallPreservedMask(MF, CallConv);
1940 assert(Mask && "Missing call preserved mask for calling convention");
1941 Ops.push_back(DAG.getRegisterMask(Mask));
1944 if (InFlag.getNode())
1945 Ops.push_back(InFlag);
1947 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
1949 MF.getFrameInfo()->setHasTailCall();
1950 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, Ops);
1953 // Returns a chain and a flag for retval copy to use.
1954 Chain = DAG.getNode(CallOpc, dl, NodeTys, Ops);
1955 InFlag = Chain.getValue(1);
1957 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, dl, true),
1958 DAG.getIntPtrConstant(0, dl, true), InFlag, dl);
1960 InFlag = Chain.getValue(1);
1962 // Handle result values, copying them out of physregs into vregs that we
1964 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins, dl, DAG,
1965 InVals, isThisReturn,
1966 isThisReturn ? OutVals[0] : SDValue());
1969 /// HandleByVal - Every parameter *after* a byval parameter is passed
1970 /// on the stack. Remember the next parameter register to allocate,
1971 /// and then confiscate the rest of the parameter registers to insure
1973 void ARMTargetLowering::HandleByVal(CCState *State, unsigned &Size,
1974 unsigned Align) const {
1975 assert((State->getCallOrPrologue() == Prologue ||
1976 State->getCallOrPrologue() == Call) &&
1977 "unhandled ParmContext");
1979 // Byval (as with any stack) slots are always at least 4 byte aligned.
1980 Align = std::max(Align, 4U);
1982 unsigned Reg = State->AllocateReg(GPRArgRegs);
1986 unsigned AlignInRegs = Align / 4;
1987 unsigned Waste = (ARM::R4 - Reg) % AlignInRegs;
1988 for (unsigned i = 0; i < Waste; ++i)
1989 Reg = State->AllocateReg(GPRArgRegs);
1994 unsigned Excess = 4 * (ARM::R4 - Reg);
1996 // Special case when NSAA != SP and parameter size greater than size of
1997 // all remained GPR regs. In that case we can't split parameter, we must
1998 // send it to stack. We also must set NCRN to R4, so waste all
1999 // remained registers.
2000 const unsigned NSAAOffset = State->getNextStackOffset();
2001 if (NSAAOffset != 0 && Size > Excess) {
2002 while (State->AllocateReg(GPRArgRegs))
2007 // First register for byval parameter is the first register that wasn't
2008 // allocated before this method call, so it would be "reg".
2009 // If parameter is small enough to be saved in range [reg, r4), then
2010 // the end (first after last) register would be reg + param-size-in-regs,
2011 // else parameter would be splitted between registers and stack,
2012 // end register would be r4 in this case.
2013 unsigned ByValRegBegin = Reg;
2014 unsigned ByValRegEnd = std::min<unsigned>(Reg + Size / 4, ARM::R4);
2015 State->addInRegsParamInfo(ByValRegBegin, ByValRegEnd);
2016 // Note, first register is allocated in the beginning of function already,
2017 // allocate remained amount of registers we need.
2018 for (unsigned i = Reg + 1; i != ByValRegEnd; ++i)
2019 State->AllocateReg(GPRArgRegs);
2020 // A byval parameter that is split between registers and memory needs its
2021 // size truncated here.
2022 // In the case where the entire structure fits in registers, we set the
2023 // size in memory to zero.
2024 Size = std::max<int>(Size - Excess, 0);
2027 /// MatchingStackOffset - Return true if the given stack call argument is
2028 /// already available in the same position (relatively) of the caller's
2029 /// incoming argument stack.
2031 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
2032 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
2033 const TargetInstrInfo *TII) {
2034 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
2036 if (Arg.getOpcode() == ISD::CopyFromReg) {
2037 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
2038 if (!TargetRegisterInfo::isVirtualRegister(VR))
2040 MachineInstr *Def = MRI->getVRegDef(VR);
2043 if (!Flags.isByVal()) {
2044 if (!TII->isLoadFromStackSlot(Def, FI))
2049 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
2050 if (Flags.isByVal())
2051 // ByVal argument is passed in as a pointer but it's now being
2052 // dereferenced. e.g.
2053 // define @foo(%struct.X* %A) {
2054 // tail call @bar(%struct.X* byval %A)
2057 SDValue Ptr = Ld->getBasePtr();
2058 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
2061 FI = FINode->getIndex();
2065 assert(FI != INT_MAX);
2066 if (!MFI->isFixedObjectIndex(FI))
2068 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
2071 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2072 /// for tail call optimization. Targets which want to do tail call
2073 /// optimization should implement this function.
2075 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2076 CallingConv::ID CalleeCC,
2078 bool isCalleeStructRet,
2079 bool isCallerStructRet,
2080 const SmallVectorImpl<ISD::OutputArg> &Outs,
2081 const SmallVectorImpl<SDValue> &OutVals,
2082 const SmallVectorImpl<ISD::InputArg> &Ins,
2083 SelectionDAG& DAG) const {
2084 const Function *CallerF = DAG.getMachineFunction().getFunction();
2085 CallingConv::ID CallerCC = CallerF->getCallingConv();
2086 bool CCMatch = CallerCC == CalleeCC;
2088 assert(Subtarget->supportsTailCall());
2090 // Look for obvious safe cases to perform tail call optimization that do not
2091 // require ABI changes. This is what gcc calls sibcall.
2093 // Do not sibcall optimize vararg calls unless the call site is not passing
2095 if (isVarArg && !Outs.empty())
2098 // Exception-handling functions need a special set of instructions to indicate
2099 // a return to the hardware. Tail-calling another function would probably
2101 if (CallerF->hasFnAttribute("interrupt"))
2104 // Also avoid sibcall optimization if either caller or callee uses struct
2105 // return semantics.
2106 if (isCalleeStructRet || isCallerStructRet)
2109 // Externally-defined functions with weak linkage should not be
2110 // tail-called on ARM when the OS does not support dynamic
2111 // pre-emption of symbols, as the AAELF spec requires normal calls
2112 // to undefined weak functions to be replaced with a NOP or jump to the
2113 // next instruction. The behaviour of branch instructions in this
2114 // situation (as used for tail calls) is implementation-defined, so we
2115 // cannot rely on the linker replacing the tail call with a return.
2116 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2117 const GlobalValue *GV = G->getGlobal();
2118 const Triple &TT = getTargetMachine().getTargetTriple();
2119 if (GV->hasExternalWeakLinkage() &&
2120 (!TT.isOSWindows() || TT.isOSBinFormatELF() || TT.isOSBinFormatMachO()))
2124 // If the calling conventions do not match, then we'd better make sure the
2125 // results are returned in the same way as what the caller expects.
2127 SmallVector<CCValAssign, 16> RVLocs1;
2128 ARMCCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(), RVLocs1,
2129 *DAG.getContext(), Call);
2130 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
2132 SmallVector<CCValAssign, 16> RVLocs2;
2133 ARMCCState CCInfo2(CallerCC, false, DAG.getMachineFunction(), RVLocs2,
2134 *DAG.getContext(), Call);
2135 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
2137 if (RVLocs1.size() != RVLocs2.size())
2139 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
2140 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
2142 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
2144 if (RVLocs1[i].isRegLoc()) {
2145 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
2148 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
2154 // If Caller's vararg or byval argument has been split between registers and
2155 // stack, do not perform tail call, since part of the argument is in caller's
2157 const ARMFunctionInfo *AFI_Caller = DAG.getMachineFunction().
2158 getInfo<ARMFunctionInfo>();
2159 if (AFI_Caller->getArgRegsSaveSize())
2162 // If the callee takes no arguments then go on to check the results of the
2164 if (!Outs.empty()) {
2165 // Check if stack adjustment is needed. For now, do not do this if any
2166 // argument is passed on the stack.
2167 SmallVector<CCValAssign, 16> ArgLocs;
2168 ARMCCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
2169 *DAG.getContext(), Call);
2170 CCInfo.AnalyzeCallOperands(Outs,
2171 CCAssignFnForNode(CalleeCC, false, isVarArg));
2172 if (CCInfo.getNextStackOffset()) {
2173 MachineFunction &MF = DAG.getMachineFunction();
2175 // Check if the arguments are already laid out in the right way as
2176 // the caller's fixed stack objects.
2177 MachineFrameInfo *MFI = MF.getFrameInfo();
2178 const MachineRegisterInfo *MRI = &MF.getRegInfo();
2179 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
2180 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
2182 ++i, ++realArgIdx) {
2183 CCValAssign &VA = ArgLocs[i];
2184 EVT RegVT = VA.getLocVT();
2185 SDValue Arg = OutVals[realArgIdx];
2186 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
2187 if (VA.getLocInfo() == CCValAssign::Indirect)
2189 if (VA.needsCustom()) {
2190 // f64 and vector types are split into multiple registers or
2191 // register/stack-slot combinations. The types will not match
2192 // the registers; give up on memory f64 refs until we figure
2193 // out what to do about this.
2196 if (!ArgLocs[++i].isRegLoc())
2198 if (RegVT == MVT::v2f64) {
2199 if (!ArgLocs[++i].isRegLoc())
2201 if (!ArgLocs[++i].isRegLoc())
2204 } else if (!VA.isRegLoc()) {
2205 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
2217 ARMTargetLowering::CanLowerReturn(CallingConv::ID CallConv,
2218 MachineFunction &MF, bool isVarArg,
2219 const SmallVectorImpl<ISD::OutputArg> &Outs,
2220 LLVMContext &Context) const {
2221 SmallVector<CCValAssign, 16> RVLocs;
2222 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, Context);
2223 return CCInfo.CheckReturn(Outs, CCAssignFnForNode(CallConv, /*Return=*/true,
2227 static SDValue LowerInterruptReturn(SmallVectorImpl<SDValue> &RetOps,
2228 SDLoc DL, SelectionDAG &DAG) {
2229 const MachineFunction &MF = DAG.getMachineFunction();
2230 const Function *F = MF.getFunction();
2232 StringRef IntKind = F->getFnAttribute("interrupt").getValueAsString();
2234 // See ARM ARM v7 B1.8.3. On exception entry LR is set to a possibly offset
2235 // version of the "preferred return address". These offsets affect the return
2236 // instruction if this is a return from PL1 without hypervisor extensions.
2237 // IRQ/FIQ: +4 "subs pc, lr, #4"
2238 // SWI: 0 "subs pc, lr, #0"
2239 // ABORT: +4 "subs pc, lr, #4"
2240 // UNDEF: +4/+2 "subs pc, lr, #0"
2241 // UNDEF varies depending on where the exception came from ARM or Thumb
2242 // mode. Alongside GCC, we throw our hands up in disgust and pretend it's 0.
2245 if (IntKind == "" || IntKind == "IRQ" || IntKind == "FIQ" ||
2248 else if (IntKind == "SWI" || IntKind == "UNDEF")
2251 report_fatal_error("Unsupported interrupt attribute. If present, value "
2252 "must be one of: IRQ, FIQ, SWI, ABORT or UNDEF");
2254 RetOps.insert(RetOps.begin() + 1,
2255 DAG.getConstant(LROffset, DL, MVT::i32, false));
2257 return DAG.getNode(ARMISD::INTRET_FLAG, DL, MVT::Other, RetOps);
2261 ARMTargetLowering::LowerReturn(SDValue Chain,
2262 CallingConv::ID CallConv, bool isVarArg,
2263 const SmallVectorImpl<ISD::OutputArg> &Outs,
2264 const SmallVectorImpl<SDValue> &OutVals,
2265 SDLoc dl, SelectionDAG &DAG) const {
2267 // CCValAssign - represent the assignment of the return value to a location.
2268 SmallVector<CCValAssign, 16> RVLocs;
2270 // CCState - Info about the registers and stack slots.
2271 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
2272 *DAG.getContext(), Call);
2274 // Analyze outgoing return values.
2275 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
2279 SmallVector<SDValue, 4> RetOps;
2280 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
2281 bool isLittleEndian = Subtarget->isLittle();
2283 MachineFunction &MF = DAG.getMachineFunction();
2284 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2285 AFI->setReturnRegsCount(RVLocs.size());
2287 // Copy the result values into the output registers.
2288 for (unsigned i = 0, realRVLocIdx = 0;
2290 ++i, ++realRVLocIdx) {
2291 CCValAssign &VA = RVLocs[i];
2292 assert(VA.isRegLoc() && "Can only return in registers!");
2294 SDValue Arg = OutVals[realRVLocIdx];
2296 switch (VA.getLocInfo()) {
2297 default: llvm_unreachable("Unknown loc info!");
2298 case CCValAssign::Full: break;
2299 case CCValAssign::BCvt:
2300 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
2304 if (VA.needsCustom()) {
2305 if (VA.getLocVT() == MVT::v2f64) {
2306 // Extract the first half and return it in two registers.
2307 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
2308 DAG.getConstant(0, dl, MVT::i32));
2309 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
2310 DAG.getVTList(MVT::i32, MVT::i32), Half);
2312 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2313 HalfGPRs.getValue(isLittleEndian ? 0 : 1),
2315 Flag = Chain.getValue(1);
2316 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2317 VA = RVLocs[++i]; // skip ahead to next loc
2318 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2319 HalfGPRs.getValue(isLittleEndian ? 1 : 0),
2321 Flag = Chain.getValue(1);
2322 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2323 VA = RVLocs[++i]; // skip ahead to next loc
2325 // Extract the 2nd half and fall through to handle it as an f64 value.
2326 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
2327 DAG.getConstant(1, dl, MVT::i32));
2329 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
2331 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
2332 DAG.getVTList(MVT::i32, MVT::i32), Arg);
2333 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2334 fmrrd.getValue(isLittleEndian ? 0 : 1),
2336 Flag = Chain.getValue(1);
2337 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2338 VA = RVLocs[++i]; // skip ahead to next loc
2339 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2340 fmrrd.getValue(isLittleEndian ? 1 : 0),
2343 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
2345 // Guarantee that all emitted copies are
2346 // stuck together, avoiding something bad.
2347 Flag = Chain.getValue(1);
2348 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2351 // Update chain and glue.
2354 RetOps.push_back(Flag);
2356 // CPUs which aren't M-class use a special sequence to return from
2357 // exceptions (roughly, any instruction setting pc and cpsr simultaneously,
2358 // though we use "subs pc, lr, #N").
2360 // M-class CPUs actually use a normal return sequence with a special
2361 // (hardware-provided) value in LR, so the normal code path works.
2362 if (DAG.getMachineFunction().getFunction()->hasFnAttribute("interrupt") &&
2363 !Subtarget->isMClass()) {
2364 if (Subtarget->isThumb1Only())
2365 report_fatal_error("interrupt attribute is not supported in Thumb1");
2366 return LowerInterruptReturn(RetOps, dl, DAG);
2369 return DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, RetOps);
2372 bool ARMTargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
2373 if (N->getNumValues() != 1)
2375 if (!N->hasNUsesOfValue(1, 0))
2378 SDValue TCChain = Chain;
2379 SDNode *Copy = *N->use_begin();
2380 if (Copy->getOpcode() == ISD::CopyToReg) {
2381 // If the copy has a glue operand, we conservatively assume it isn't safe to
2382 // perform a tail call.
2383 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
2385 TCChain = Copy->getOperand(0);
2386 } else if (Copy->getOpcode() == ARMISD::VMOVRRD) {
2387 SDNode *VMov = Copy;
2388 // f64 returned in a pair of GPRs.
2389 SmallPtrSet<SDNode*, 2> Copies;
2390 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
2392 if (UI->getOpcode() != ISD::CopyToReg)
2396 if (Copies.size() > 2)
2399 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
2401 SDValue UseChain = UI->getOperand(0);
2402 if (Copies.count(UseChain.getNode()))
2406 // We are at the top of this chain.
2407 // If the copy has a glue operand, we conservatively assume it
2408 // isn't safe to perform a tail call.
2409 if (UI->getOperand(UI->getNumOperands()-1).getValueType() == MVT::Glue)
2415 } else if (Copy->getOpcode() == ISD::BITCAST) {
2416 // f32 returned in a single GPR.
2417 if (!Copy->hasOneUse())
2419 Copy = *Copy->use_begin();
2420 if (Copy->getOpcode() != ISD::CopyToReg || !Copy->hasNUsesOfValue(1, 0))
2422 // If the copy has a glue operand, we conservatively assume it isn't safe to
2423 // perform a tail call.
2424 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
2426 TCChain = Copy->getOperand(0);
2431 bool HasRet = false;
2432 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
2434 if (UI->getOpcode() != ARMISD::RET_FLAG &&
2435 UI->getOpcode() != ARMISD::INTRET_FLAG)
2447 bool ARMTargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
2448 if (!Subtarget->supportsTailCall())
2452 CI->getParent()->getParent()->getFnAttribute("disable-tail-calls");
2453 if (!CI->isTailCall() || Attr.getValueAsString() == "true")
2459 // Trying to write a 64 bit value so need to split into two 32 bit values first,
2460 // and pass the lower and high parts through.
2461 static SDValue LowerWRITE_REGISTER(SDValue Op, SelectionDAG &DAG) {
2463 SDValue WriteValue = Op->getOperand(2);
2465 // This function is only supposed to be called for i64 type argument.
2466 assert(WriteValue.getValueType() == MVT::i64
2467 && "LowerWRITE_REGISTER called for non-i64 type argument.");
2469 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, WriteValue,
2470 DAG.getConstant(0, DL, MVT::i32));
2471 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, WriteValue,
2472 DAG.getConstant(1, DL, MVT::i32));
2473 SDValue Ops[] = { Op->getOperand(0), Op->getOperand(1), Lo, Hi };
2474 return DAG.getNode(ISD::WRITE_REGISTER, DL, MVT::Other, Ops);
2477 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
2478 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
2479 // one of the above mentioned nodes. It has to be wrapped because otherwise
2480 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
2481 // be used to form addressing mode. These wrapped nodes will be selected
2483 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
2484 EVT PtrVT = Op.getValueType();
2485 // FIXME there is no actual debug info here
2487 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
2489 if (CP->isMachineConstantPoolEntry())
2490 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
2491 CP->getAlignment());
2493 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
2494 CP->getAlignment());
2495 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
2498 unsigned ARMTargetLowering::getJumpTableEncoding() const {
2499 return MachineJumpTableInfo::EK_Inline;
2502 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
2503 SelectionDAG &DAG) const {
2504 MachineFunction &MF = DAG.getMachineFunction();
2505 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2506 unsigned ARMPCLabelIndex = 0;
2508 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2509 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
2510 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2512 if (RelocM == Reloc::Static) {
2513 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
2515 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2516 ARMPCLabelIndex = AFI->createPICLabelUId();
2517 ARMConstantPoolValue *CPV =
2518 ARMConstantPoolConstant::Create(BA, ARMPCLabelIndex,
2519 ARMCP::CPBlockAddress, PCAdj);
2520 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2522 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
2524 DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
2525 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
2526 false, false, false, 0);
2527 if (RelocM == Reloc::Static)
2529 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, DL, MVT::i32);
2530 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
2533 /// \brief Convert a TLS address reference into the correct sequence of loads
2534 /// and calls to compute the variable's address for Darwin, and return an
2535 /// SDValue containing the final node.
2537 /// Darwin only has one TLS scheme which must be capable of dealing with the
2538 /// fully general situation, in the worst case. This means:
2539 /// + "extern __thread" declaration.
2540 /// + Defined in a possibly unknown dynamic library.
2542 /// The general system is that each __thread variable has a [3 x i32] descriptor
2543 /// which contains information used by the runtime to calculate the address. The
2544 /// only part of this the compiler needs to know about is the first word, which
2545 /// contains a function pointer that must be called with the address of the
2546 /// entire descriptor in "r0".
2548 /// Since this descriptor may be in a different unit, in general access must
2549 /// proceed along the usual ARM rules. A common sequence to produce is:
2551 /// movw rT1, :lower16:_var$non_lazy_ptr
2552 /// movt rT1, :upper16:_var$non_lazy_ptr
2556 /// [...address now in r0...]
2558 ARMTargetLowering::LowerGlobalTLSAddressDarwin(SDValue Op,
2559 SelectionDAG &DAG) const {
2560 assert(Subtarget->isTargetDarwin() && "TLS only supported on Darwin");
2563 // First step is to get the address of the actua global symbol. This is where
2564 // the TLS descriptor lives.
2565 SDValue DescAddr = LowerGlobalAddressDarwin(Op, DAG);
2567 // The first entry in the descriptor is a function pointer that we must call
2568 // to obtain the address of the variable.
2569 SDValue Chain = DAG.getEntryNode();
2570 SDValue FuncTLVGet =
2571 DAG.getLoad(MVT::i32, DL, Chain, DescAddr,
2572 MachinePointerInfo::getGOT(DAG.getMachineFunction()),
2573 false, true, true, 4);
2574 Chain = FuncTLVGet.getValue(1);
2576 MachineFunction &F = DAG.getMachineFunction();
2577 MachineFrameInfo *MFI = F.getFrameInfo();
2578 MFI->setAdjustsStack(true);
2580 // TLS calls preserve all registers except those that absolutely must be
2581 // trashed: R0 (it takes an argument), LR (it's a call) and CPSR (let's not be
2584 getTargetMachine().getSubtargetImpl(*F.getFunction())->getRegisterInfo();
2585 auto ARI = static_cast<const ARMRegisterInfo *>(TRI);
2586 const uint32_t *Mask = ARI->getTLSCallPreservedMask(DAG.getMachineFunction());
2588 // Finally, we can make the call. This is just a degenerate version of a
2589 // normal AArch64 call node: r0 takes the address of the descriptor, and
2590 // returns the address of the variable in this thread.
2591 Chain = DAG.getCopyToReg(Chain, DL, ARM::R0, DescAddr, SDValue());
2593 DAG.getNode(ARMISD::CALL, DL, DAG.getVTList(MVT::Other, MVT::Glue),
2594 Chain, FuncTLVGet, DAG.getRegister(ARM::R0, MVT::i32),
2595 DAG.getRegisterMask(Mask), Chain.getValue(1));
2596 return DAG.getCopyFromReg(Chain, DL, ARM::R0, MVT::i32, Chain.getValue(1));
2599 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
2601 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
2602 SelectionDAG &DAG) const {
2604 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2605 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2606 MachineFunction &MF = DAG.getMachineFunction();
2607 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2608 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2609 ARMConstantPoolValue *CPV =
2610 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2611 ARMCP::CPValue, PCAdj, ARMCP::TLSGD, true);
2612 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2613 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
2615 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
2616 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
2617 false, false, false, 0);
2618 SDValue Chain = Argument.getValue(1);
2620 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, dl, MVT::i32);
2621 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
2623 // call __tls_get_addr.
2626 Entry.Node = Argument;
2627 Entry.Ty = (Type *) Type::getInt32Ty(*DAG.getContext());
2628 Args.push_back(Entry);
2630 // FIXME: is there useful debug info available here?
2631 TargetLowering::CallLoweringInfo CLI(DAG);
2632 CLI.setDebugLoc(dl).setChain(Chain)
2633 .setCallee(CallingConv::C, Type::getInt32Ty(*DAG.getContext()),
2634 DAG.getExternalSymbol("__tls_get_addr", PtrVT), std::move(Args),
2637 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
2638 return CallResult.first;
2641 // Lower ISD::GlobalTLSAddress using the "initial exec" or
2642 // "local exec" model.
2644 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
2646 TLSModel::Model model) const {
2647 const GlobalValue *GV = GA->getGlobal();
2650 SDValue Chain = DAG.getEntryNode();
2651 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2652 // Get the Thread Pointer
2653 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2655 if (model == TLSModel::InitialExec) {
2656 MachineFunction &MF = DAG.getMachineFunction();
2657 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2658 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2659 // Initial exec model.
2660 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2661 ARMConstantPoolValue *CPV =
2662 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2663 ARMCP::CPValue, PCAdj, ARMCP::GOTTPOFF,
2665 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2666 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2667 Offset = DAG.getLoad(
2668 PtrVT, dl, Chain, Offset,
2669 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), false,
2671 Chain = Offset.getValue(1);
2673 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, dl, MVT::i32);
2674 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
2676 Offset = DAG.getLoad(
2677 PtrVT, dl, Chain, Offset,
2678 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), false,
2682 assert(model == TLSModel::LocalExec);
2683 ARMConstantPoolValue *CPV =
2684 ARMConstantPoolConstant::Create(GV, ARMCP::TPOFF);
2685 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2686 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2687 Offset = DAG.getLoad(
2688 PtrVT, dl, Chain, Offset,
2689 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), false,
2693 // The address of the thread local variable is the add of the thread
2694 // pointer with the offset of the variable.
2695 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
2699 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
2700 if (Subtarget->isTargetDarwin())
2701 return LowerGlobalTLSAddressDarwin(Op, DAG);
2703 // TODO: implement the "local dynamic" model
2704 assert(Subtarget->isTargetELF() && "Only ELF implemented here");
2705 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
2706 if (DAG.getTarget().Options.EmulatedTLS)
2707 return LowerToTLSEmulatedModel(GA, DAG);
2709 TLSModel::Model model = getTargetMachine().getTLSModel(GA->getGlobal());
2712 case TLSModel::GeneralDynamic:
2713 case TLSModel::LocalDynamic:
2714 return LowerToTLSGeneralDynamicModel(GA, DAG);
2715 case TLSModel::InitialExec:
2716 case TLSModel::LocalExec:
2717 return LowerToTLSExecModels(GA, DAG, model);
2719 llvm_unreachable("bogus TLS model");
2722 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
2723 SelectionDAG &DAG) const {
2724 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2726 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2727 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2729 !(GV->hasHiddenVisibility() || GV->hasLocalLinkage());
2731 MachineFunction &MF = DAG.getMachineFunction();
2732 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2733 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2734 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2736 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2737 ARMConstantPoolValue *CPV = ARMConstantPoolConstant::Create(
2738 GV, ARMPCLabelIndex, ARMCP::CPValue, PCAdj,
2739 UseGOT_PREL ? ARMCP::GOT_PREL : ARMCP::no_modifier,
2740 /*AddCurrentAddress=*/UseGOT_PREL);
2741 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2742 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2743 SDValue Result = DAG.getLoad(
2744 PtrVT, dl, DAG.getEntryNode(), CPAddr,
2745 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), false,
2747 SDValue Chain = Result.getValue(1);
2748 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, dl, MVT::i32);
2749 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2751 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
2752 MachinePointerInfo::getGOT(DAG.getMachineFunction()),
2753 false, false, false, 0);
2757 // If we have T2 ops, we can materialize the address directly via movt/movw
2758 // pair. This is always cheaper.
2759 if (Subtarget->useMovt(DAG.getMachineFunction())) {
2761 // FIXME: Once remat is capable of dealing with instructions with register
2762 // operands, expand this into two nodes.
2763 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2764 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2766 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2767 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2769 PtrVT, dl, DAG.getEntryNode(), CPAddr,
2770 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), false,
2775 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
2776 SelectionDAG &DAG) const {
2777 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2779 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2780 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2782 if (Subtarget->useMovt(DAG.getMachineFunction()))
2785 // FIXME: Once remat is capable of dealing with instructions with register
2786 // operands, expand this into multiple nodes
2788 RelocM == Reloc::PIC_ ? ARMISD::WrapperPIC : ARMISD::Wrapper;
2790 SDValue G = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, ARMII::MO_NONLAZY);
2791 SDValue Result = DAG.getNode(Wrapper, dl, PtrVT, G);
2793 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2794 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
2795 MachinePointerInfo::getGOT(DAG.getMachineFunction()),
2796 false, false, false, 0);
2800 SDValue ARMTargetLowering::LowerGlobalAddressWindows(SDValue Op,
2801 SelectionDAG &DAG) const {
2802 assert(Subtarget->isTargetWindows() && "non-Windows COFF is not supported");
2803 assert(Subtarget->useMovt(DAG.getMachineFunction()) &&
2804 "Windows on ARM expects to use movw/movt");
2806 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2807 const ARMII::TOF TargetFlags =
2808 (GV->hasDLLImportStorageClass() ? ARMII::MO_DLLIMPORT : ARMII::MO_NO_FLAG);
2809 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2815 // FIXME: Once remat is capable of dealing with instructions with register
2816 // operands, expand this into two nodes.
2817 Result = DAG.getNode(ARMISD::Wrapper, DL, PtrVT,
2818 DAG.getTargetGlobalAddress(GV, DL, PtrVT, /*Offset=*/0,
2820 if (GV->hasDLLImportStorageClass())
2821 Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), Result,
2822 MachinePointerInfo::getGOT(DAG.getMachineFunction()),
2823 false, false, false, 0);
2828 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
2830 SDValue Val = DAG.getConstant(0, dl, MVT::i32);
2831 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl,
2832 DAG.getVTList(MVT::i32, MVT::Other), Op.getOperand(0),
2833 Op.getOperand(1), Val);
2837 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
2839 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
2840 Op.getOperand(1), DAG.getConstant(0, dl, MVT::i32));
2843 SDValue ARMTargetLowering::LowerEH_SJLJ_SETUP_DISPATCH(SDValue Op,
2844 SelectionDAG &DAG) const {
2846 return DAG.getNode(ARMISD::EH_SJLJ_SETUP_DISPATCH, dl, MVT::Other,
2851 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
2852 const ARMSubtarget *Subtarget) const {
2853 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2856 default: return SDValue(); // Don't custom lower most intrinsics.
2857 case Intrinsic::arm_rbit: {
2858 assert(Op.getOperand(1).getValueType() == MVT::i32 &&
2859 "RBIT intrinsic must have i32 type!");
2860 return DAG.getNode(ISD::BITREVERSE, dl, MVT::i32, Op.getOperand(1));
2862 case Intrinsic::arm_thread_pointer: {
2863 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2864 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2866 case Intrinsic::eh_sjlj_lsda: {
2867 MachineFunction &MF = DAG.getMachineFunction();
2868 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2869 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2870 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2871 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2873 unsigned PCAdj = (RelocM != Reloc::PIC_)
2874 ? 0 : (Subtarget->isThumb() ? 4 : 8);
2875 ARMConstantPoolValue *CPV =
2876 ARMConstantPoolConstant::Create(MF.getFunction(), ARMPCLabelIndex,
2877 ARMCP::CPLSDA, PCAdj);
2878 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2879 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2880 SDValue Result = DAG.getLoad(
2881 PtrVT, dl, DAG.getEntryNode(), CPAddr,
2882 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), false,
2885 if (RelocM == Reloc::PIC_) {
2886 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, dl, MVT::i32);
2887 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2891 case Intrinsic::arm_neon_vmulls:
2892 case Intrinsic::arm_neon_vmullu: {
2893 unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vmulls)
2894 ? ARMISD::VMULLs : ARMISD::VMULLu;
2895 return DAG.getNode(NewOpc, SDLoc(Op), Op.getValueType(),
2896 Op.getOperand(1), Op.getOperand(2));
2898 case Intrinsic::arm_neon_vminnm:
2899 case Intrinsic::arm_neon_vmaxnm: {
2900 unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vminnm)
2901 ? ISD::FMINNUM : ISD::FMAXNUM;
2902 return DAG.getNode(NewOpc, SDLoc(Op), Op.getValueType(),
2903 Op.getOperand(1), Op.getOperand(2));
2905 case Intrinsic::arm_neon_vminu:
2906 case Intrinsic::arm_neon_vmaxu: {
2907 if (Op.getValueType().isFloatingPoint())
2909 unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vminu)
2910 ? ISD::UMIN : ISD::UMAX;
2911 return DAG.getNode(NewOpc, SDLoc(Op), Op.getValueType(),
2912 Op.getOperand(1), Op.getOperand(2));
2914 case Intrinsic::arm_neon_vmins:
2915 case Intrinsic::arm_neon_vmaxs: {
2916 // v{min,max}s is overloaded between signed integers and floats.
2917 if (!Op.getValueType().isFloatingPoint()) {
2918 unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vmins)
2919 ? ISD::SMIN : ISD::SMAX;
2920 return DAG.getNode(NewOpc, SDLoc(Op), Op.getValueType(),
2921 Op.getOperand(1), Op.getOperand(2));
2923 unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vmins)
2924 ? ISD::FMINNAN : ISD::FMAXNAN;
2925 return DAG.getNode(NewOpc, SDLoc(Op), Op.getValueType(),
2926 Op.getOperand(1), Op.getOperand(2));
2931 static SDValue LowerATOMIC_FENCE(SDValue Op, SelectionDAG &DAG,
2932 const ARMSubtarget *Subtarget) {
2933 // FIXME: handle "fence singlethread" more efficiently.
2935 if (!Subtarget->hasDataBarrier()) {
2936 // Some ARMv6 cpus can support data barriers with an mcr instruction.
2937 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2939 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2940 "Unexpected ISD::ATOMIC_FENCE encountered. Should be libcall!");
2941 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2942 DAG.getConstant(0, dl, MVT::i32));
2945 ConstantSDNode *OrdN = cast<ConstantSDNode>(Op.getOperand(1));
2946 AtomicOrdering Ord = static_cast<AtomicOrdering>(OrdN->getZExtValue());
2947 ARM_MB::MemBOpt Domain = ARM_MB::ISH;
2948 if (Subtarget->isMClass()) {
2949 // Only a full system barrier exists in the M-class architectures.
2950 Domain = ARM_MB::SY;
2951 } else if (Subtarget->isSwift() && Ord == Release) {
2952 // Swift happens to implement ISHST barriers in a way that's compatible with
2953 // Release semantics but weaker than ISH so we'd be fools not to use
2954 // it. Beware: other processors probably don't!
2955 Domain = ARM_MB::ISHST;
2958 return DAG.getNode(ISD::INTRINSIC_VOID, dl, MVT::Other, Op.getOperand(0),
2959 DAG.getConstant(Intrinsic::arm_dmb, dl, MVT::i32),
2960 DAG.getConstant(Domain, dl, MVT::i32));
2963 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG,
2964 const ARMSubtarget *Subtarget) {
2965 // ARM pre v5TE and Thumb1 does not have preload instructions.
2966 if (!(Subtarget->isThumb2() ||
2967 (!Subtarget->isThumb1Only() && Subtarget->hasV5TEOps())))
2968 // Just preserve the chain.
2969 return Op.getOperand(0);
2972 unsigned isRead = ~cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() & 1;
2974 (!Subtarget->hasV7Ops() || !Subtarget->hasMPExtension()))
2975 // ARMv7 with MP extension has PLDW.
2976 return Op.getOperand(0);
2978 unsigned isData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
2979 if (Subtarget->isThumb()) {
2981 isRead = ~isRead & 1;
2982 isData = ~isData & 1;
2985 return DAG.getNode(ARMISD::PRELOAD, dl, MVT::Other, Op.getOperand(0),
2986 Op.getOperand(1), DAG.getConstant(isRead, dl, MVT::i32),
2987 DAG.getConstant(isData, dl, MVT::i32));
2990 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
2991 MachineFunction &MF = DAG.getMachineFunction();
2992 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
2994 // vastart just stores the address of the VarArgsFrameIndex slot into the
2995 // memory location argument.
2997 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy(DAG.getDataLayout());
2998 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2999 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
3000 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
3001 MachinePointerInfo(SV), false, false, 0);
3005 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
3006 SDValue &Root, SelectionDAG &DAG,
3008 MachineFunction &MF = DAG.getMachineFunction();
3009 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
3011 const TargetRegisterClass *RC;
3012 if (AFI->isThumb1OnlyFunction())
3013 RC = &ARM::tGPRRegClass;
3015 RC = &ARM::GPRRegClass;
3017 // Transform the arguments stored in physical registers into virtual ones.
3018 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
3019 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
3022 if (NextVA.isMemLoc()) {
3023 MachineFrameInfo *MFI = MF.getFrameInfo();
3024 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true);
3026 // Create load node to retrieve arguments from the stack.
3027 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
3028 ArgValue2 = DAG.getLoad(
3029 MVT::i32, dl, Root, FIN,
3030 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI), false,
3033 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
3034 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
3036 if (!Subtarget->isLittle())
3037 std::swap (ArgValue, ArgValue2);
3038 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
3041 // The remaining GPRs hold either the beginning of variable-argument
3042 // data, or the beginning of an aggregate passed by value (usually
3043 // byval). Either way, we allocate stack slots adjacent to the data
3044 // provided by our caller, and store the unallocated registers there.
3045 // If this is a variadic function, the va_list pointer will begin with
3046 // these values; otherwise, this reassembles a (byval) structure that
3047 // was split between registers and memory.
3048 // Return: The frame index registers were stored into.
3050 ARMTargetLowering::StoreByValRegs(CCState &CCInfo, SelectionDAG &DAG,
3051 SDLoc dl, SDValue &Chain,
3052 const Value *OrigArg,
3053 unsigned InRegsParamRecordIdx,
3055 unsigned ArgSize) const {
3056 // Currently, two use-cases possible:
3057 // Case #1. Non-var-args function, and we meet first byval parameter.
3058 // Setup first unallocated register as first byval register;
3059 // eat all remained registers
3060 // (these two actions are performed by HandleByVal method).
3061 // Then, here, we initialize stack frame with
3062 // "store-reg" instructions.
3063 // Case #2. Var-args function, that doesn't contain byval parameters.
3064 // The same: eat all remained unallocated registers,
3065 // initialize stack frame.
3067 MachineFunction &MF = DAG.getMachineFunction();
3068 MachineFrameInfo *MFI = MF.getFrameInfo();
3069 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
3070 unsigned RBegin, REnd;
3071 if (InRegsParamRecordIdx < CCInfo.getInRegsParamsCount()) {
3072 CCInfo.getInRegsParamInfo(InRegsParamRecordIdx, RBegin, REnd);
3074 unsigned RBeginIdx = CCInfo.getFirstUnallocated(GPRArgRegs);
3075 RBegin = RBeginIdx == 4 ? (unsigned)ARM::R4 : GPRArgRegs[RBeginIdx];
3080 ArgOffset = -4 * (ARM::R4 - RBegin);
3082 auto PtrVT = getPointerTy(DAG.getDataLayout());
3083 int FrameIndex = MFI->CreateFixedObject(ArgSize, ArgOffset, false);
3084 SDValue FIN = DAG.getFrameIndex(FrameIndex, PtrVT);
3086 SmallVector<SDValue, 4> MemOps;
3087 const TargetRegisterClass *RC =
3088 AFI->isThumb1OnlyFunction() ? &ARM::tGPRRegClass : &ARM::GPRRegClass;
3090 for (unsigned Reg = RBegin, i = 0; Reg < REnd; ++Reg, ++i) {
3091 unsigned VReg = MF.addLiveIn(Reg, RC);
3092 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
3094 DAG.getStore(Val.getValue(1), dl, Val, FIN,
3095 MachinePointerInfo(OrigArg, 4 * i), false, false, 0);
3096 MemOps.push_back(Store);
3097 FIN = DAG.getNode(ISD::ADD, dl, PtrVT, FIN, DAG.getConstant(4, dl, PtrVT));
3100 if (!MemOps.empty())
3101 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOps);
3105 // Setup stack frame, the va_list pointer will start from.
3107 ARMTargetLowering::VarArgStyleRegisters(CCState &CCInfo, SelectionDAG &DAG,
3108 SDLoc dl, SDValue &Chain,
3110 unsigned TotalArgRegsSaveSize,
3111 bool ForceMutable) const {
3112 MachineFunction &MF = DAG.getMachineFunction();
3113 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
3115 // Try to store any remaining integer argument regs
3116 // to their spots on the stack so that they may be loaded by deferencing
3117 // the result of va_next.
3118 // If there is no regs to be stored, just point address after last
3119 // argument passed via stack.
3120 int FrameIndex = StoreByValRegs(CCInfo, DAG, dl, Chain, nullptr,
3121 CCInfo.getInRegsParamsCount(),
3122 CCInfo.getNextStackOffset(), 4);
3123 AFI->setVarArgsFrameIndex(FrameIndex);
3127 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
3128 CallingConv::ID CallConv, bool isVarArg,
3129 const SmallVectorImpl<ISD::InputArg>
3131 SDLoc dl, SelectionDAG &DAG,
3132 SmallVectorImpl<SDValue> &InVals)
3134 MachineFunction &MF = DAG.getMachineFunction();
3135 MachineFrameInfo *MFI = MF.getFrameInfo();
3137 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
3139 // Assign locations to all of the incoming arguments.
3140 SmallVector<CCValAssign, 16> ArgLocs;
3141 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs,
3142 *DAG.getContext(), Prologue);
3143 CCInfo.AnalyzeFormalArguments(Ins,
3144 CCAssignFnForNode(CallConv, /* Return*/ false,
3147 SmallVector<SDValue, 16> ArgValues;
3149 Function::const_arg_iterator CurOrigArg = MF.getFunction()->arg_begin();
3150 unsigned CurArgIdx = 0;
3152 // Initially ArgRegsSaveSize is zero.
3153 // Then we increase this value each time we meet byval parameter.
3154 // We also increase this value in case of varargs function.
3155 AFI->setArgRegsSaveSize(0);
3157 // Calculate the amount of stack space that we need to allocate to store
3158 // byval and variadic arguments that are passed in registers.
3159 // We need to know this before we allocate the first byval or variadic
3160 // argument, as they will be allocated a stack slot below the CFA (Canonical
3161 // Frame Address, the stack pointer at entry to the function).
3162 unsigned ArgRegBegin = ARM::R4;
3163 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3164 if (CCInfo.getInRegsParamsProcessed() >= CCInfo.getInRegsParamsCount())
3167 CCValAssign &VA = ArgLocs[i];
3168 unsigned Index = VA.getValNo();
3169 ISD::ArgFlagsTy Flags = Ins[Index].Flags;
3170 if (!Flags.isByVal())
3173 assert(VA.isMemLoc() && "unexpected byval pointer in reg");
3174 unsigned RBegin, REnd;
3175 CCInfo.getInRegsParamInfo(CCInfo.getInRegsParamsProcessed(), RBegin, REnd);
3176 ArgRegBegin = std::min(ArgRegBegin, RBegin);
3178 CCInfo.nextInRegsParam();
3180 CCInfo.rewindByValRegsInfo();
3182 int lastInsIndex = -1;
3183 if (isVarArg && MFI->hasVAStart()) {
3184 unsigned RegIdx = CCInfo.getFirstUnallocated(GPRArgRegs);
3185 if (RegIdx != array_lengthof(GPRArgRegs))
3186 ArgRegBegin = std::min(ArgRegBegin, (unsigned)GPRArgRegs[RegIdx]);
3189 unsigned TotalArgRegsSaveSize = 4 * (ARM::R4 - ArgRegBegin);
3190 AFI->setArgRegsSaveSize(TotalArgRegsSaveSize);
3191 auto PtrVT = getPointerTy(DAG.getDataLayout());
3193 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3194 CCValAssign &VA = ArgLocs[i];
3195 if (Ins[VA.getValNo()].isOrigArg()) {
3196 std::advance(CurOrigArg,
3197 Ins[VA.getValNo()].getOrigArgIndex() - CurArgIdx);
3198 CurArgIdx = Ins[VA.getValNo()].getOrigArgIndex();
3200 // Arguments stored in registers.
3201 if (VA.isRegLoc()) {
3202 EVT RegVT = VA.getLocVT();
3204 if (VA.needsCustom()) {
3205 // f64 and vector types are split up into multiple registers or
3206 // combinations of registers and stack slots.
3207 if (VA.getLocVT() == MVT::v2f64) {
3208 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
3210 VA = ArgLocs[++i]; // skip ahead to next loc
3212 if (VA.isMemLoc()) {
3213 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true);
3214 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
3215 ArgValue2 = DAG.getLoad(
3216 MVT::f64, dl, Chain, FIN,
3217 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI),
3218 false, false, false, 0);
3220 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
3223 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
3224 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
3225 ArgValue, ArgValue1,
3226 DAG.getIntPtrConstant(0, dl));
3227 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
3228 ArgValue, ArgValue2,
3229 DAG.getIntPtrConstant(1, dl));
3231 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
3234 const TargetRegisterClass *RC;
3236 if (RegVT == MVT::f32)
3237 RC = &ARM::SPRRegClass;
3238 else if (RegVT == MVT::f64)
3239 RC = &ARM::DPRRegClass;
3240 else if (RegVT == MVT::v2f64)
3241 RC = &ARM::QPRRegClass;
3242 else if (RegVT == MVT::i32)
3243 RC = AFI->isThumb1OnlyFunction() ? &ARM::tGPRRegClass
3244 : &ARM::GPRRegClass;
3246 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
3248 // Transform the arguments in physical registers into virtual ones.
3249 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
3250 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
3253 // If this is an 8 or 16-bit value, it is really passed promoted
3254 // to 32 bits. Insert an assert[sz]ext to capture this, then
3255 // truncate to the right size.
3256 switch (VA.getLocInfo()) {
3257 default: llvm_unreachable("Unknown loc info!");
3258 case CCValAssign::Full: break;
3259 case CCValAssign::BCvt:
3260 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
3262 case CCValAssign::SExt:
3263 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
3264 DAG.getValueType(VA.getValVT()));
3265 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
3267 case CCValAssign::ZExt:
3268 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
3269 DAG.getValueType(VA.getValVT()));
3270 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
3274 InVals.push_back(ArgValue);
3276 } else { // VA.isRegLoc()
3279 assert(VA.isMemLoc());
3280 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
3282 int index = VA.getValNo();
3284 // Some Ins[] entries become multiple ArgLoc[] entries.
3285 // Process them only once.
3286 if (index != lastInsIndex)
3288 ISD::ArgFlagsTy Flags = Ins[index].Flags;
3289 // FIXME: For now, all byval parameter objects are marked mutable.
3290 // This can be changed with more analysis.
3291 // In case of tail call optimization mark all arguments mutable.
3292 // Since they could be overwritten by lowering of arguments in case of
3294 if (Flags.isByVal()) {
3295 assert(Ins[index].isOrigArg() &&
3296 "Byval arguments cannot be implicit");
3297 unsigned CurByValIndex = CCInfo.getInRegsParamsProcessed();
3299 int FrameIndex = StoreByValRegs(
3300 CCInfo, DAG, dl, Chain, &*CurOrigArg, CurByValIndex,
3301 VA.getLocMemOffset(), Flags.getByValSize());
3302 InVals.push_back(DAG.getFrameIndex(FrameIndex, PtrVT));
3303 CCInfo.nextInRegsParam();
3305 unsigned FIOffset = VA.getLocMemOffset();
3306 int FI = MFI->CreateFixedObject(VA.getLocVT().getSizeInBits()/8,
3309 // Create load nodes to retrieve arguments from the stack.
3310 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
3311 InVals.push_back(DAG.getLoad(
3312 VA.getValVT(), dl, Chain, FIN,
3313 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI),
3314 false, false, false, 0));
3316 lastInsIndex = index;
3322 if (isVarArg && MFI->hasVAStart())
3323 VarArgStyleRegisters(CCInfo, DAG, dl, Chain,
3324 CCInfo.getNextStackOffset(),
3325 TotalArgRegsSaveSize);
3327 AFI->setArgumentStackSize(CCInfo.getNextStackOffset());
3332 /// isFloatingPointZero - Return true if this is +0.0.
3333 static bool isFloatingPointZero(SDValue Op) {
3334 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
3335 return CFP->getValueAPF().isPosZero();
3336 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
3337 // Maybe this has already been legalized into the constant pool?
3338 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
3339 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
3340 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
3341 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
3342 return CFP->getValueAPF().isPosZero();
3344 } else if (Op->getOpcode() == ISD::BITCAST &&
3345 Op->getValueType(0) == MVT::f64) {
3346 // Handle (ISD::BITCAST (ARMISD::VMOVIMM (ISD::TargetConstant 0)) MVT::f64)
3347 // created by LowerConstantFP().
3348 SDValue BitcastOp = Op->getOperand(0);
3349 if (BitcastOp->getOpcode() == ARMISD::VMOVIMM &&
3350 isNullConstant(BitcastOp->getOperand(0)))
3356 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
3357 /// the given operands.
3359 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
3360 SDValue &ARMcc, SelectionDAG &DAG,
3362 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
3363 unsigned C = RHSC->getZExtValue();
3364 if (!isLegalICmpImmediate(C)) {
3365 // Constant does not fit, try adjusting it by one?
3370 if (C != 0x80000000 && isLegalICmpImmediate(C-1)) {
3371 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
3372 RHS = DAG.getConstant(C - 1, dl, MVT::i32);
3377 if (C != 0 && isLegalICmpImmediate(C-1)) {
3378 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
3379 RHS = DAG.getConstant(C - 1, dl, MVT::i32);
3384 if (C != 0x7fffffff && isLegalICmpImmediate(C+1)) {
3385 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
3386 RHS = DAG.getConstant(C + 1, dl, MVT::i32);
3391 if (C != 0xffffffff && isLegalICmpImmediate(C+1)) {
3392 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
3393 RHS = DAG.getConstant(C + 1, dl, MVT::i32);
3400 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3401 ARMISD::NodeType CompareType;
3404 CompareType = ARMISD::CMP;
3409 CompareType = ARMISD::CMPZ;
3412 ARMcc = DAG.getConstant(CondCode, dl, MVT::i32);
3413 return DAG.getNode(CompareType, dl, MVT::Glue, LHS, RHS);
3416 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
3418 ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
3420 assert(!Subtarget->isFPOnlySP() || RHS.getValueType() != MVT::f64);
3422 if (!isFloatingPointZero(RHS))
3423 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Glue, LHS, RHS);
3425 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Glue, LHS);
3426 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Glue, Cmp);
3429 /// duplicateCmp - Glue values can have only one use, so this function
3430 /// duplicates a comparison node.
3432 ARMTargetLowering::duplicateCmp(SDValue Cmp, SelectionDAG &DAG) const {
3433 unsigned Opc = Cmp.getOpcode();
3435 if (Opc == ARMISD::CMP || Opc == ARMISD::CMPZ)
3436 return DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
3438 assert(Opc == ARMISD::FMSTAT && "unexpected comparison operation");
3439 Cmp = Cmp.getOperand(0);
3440 Opc = Cmp.getOpcode();
3441 if (Opc == ARMISD::CMPFP)
3442 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
3444 assert(Opc == ARMISD::CMPFPw0 && "unexpected operand of FMSTAT");
3445 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0));
3447 return DAG.getNode(ARMISD::FMSTAT, DL, MVT::Glue, Cmp);
3450 std::pair<SDValue, SDValue>
3451 ARMTargetLowering::getARMXALUOOp(SDValue Op, SelectionDAG &DAG,
3452 SDValue &ARMcc) const {
3453 assert(Op.getValueType() == MVT::i32 && "Unsupported value type");
3455 SDValue Value, OverflowCmp;
3456 SDValue LHS = Op.getOperand(0);
3457 SDValue RHS = Op.getOperand(1);
3460 // FIXME: We are currently always generating CMPs because we don't support
3461 // generating CMN through the backend. This is not as good as the natural
3462 // CMP case because it causes a register dependency and cannot be folded
3465 switch (Op.getOpcode()) {
3467 llvm_unreachable("Unknown overflow instruction!");
3469 ARMcc = DAG.getConstant(ARMCC::VC, dl, MVT::i32);
3470 Value = DAG.getNode(ISD::ADD, dl, Op.getValueType(), LHS, RHS);
3471 OverflowCmp = DAG.getNode(ARMISD::CMP, dl, MVT::Glue, Value, LHS);
3474 ARMcc = DAG.getConstant(ARMCC::HS, dl, MVT::i32);
3475 Value = DAG.getNode(ISD::ADD, dl, Op.getValueType(), LHS, RHS);
3476 OverflowCmp = DAG.getNode(ARMISD::CMP, dl, MVT::Glue, Value, LHS);
3479 ARMcc = DAG.getConstant(ARMCC::VC, dl, MVT::i32);
3480 Value = DAG.getNode(ISD::SUB, dl, Op.getValueType(), LHS, RHS);
3481 OverflowCmp = DAG.getNode(ARMISD::CMP, dl, MVT::Glue, LHS, RHS);
3484 ARMcc = DAG.getConstant(ARMCC::HS, dl, MVT::i32);
3485 Value = DAG.getNode(ISD::SUB, dl, Op.getValueType(), LHS, RHS);
3486 OverflowCmp = DAG.getNode(ARMISD::CMP, dl, MVT::Glue, LHS, RHS);
3490 return std::make_pair(Value, OverflowCmp);
3495 ARMTargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) const {
3496 // Let legalize expand this if it isn't a legal type yet.
3497 if (!DAG.getTargetLoweringInfo().isTypeLegal(Op.getValueType()))
3500 SDValue Value, OverflowCmp;
3502 std::tie(Value, OverflowCmp) = getARMXALUOOp(Op, DAG, ARMcc);
3503 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3505 // We use 0 and 1 as false and true values.
3506 SDValue TVal = DAG.getConstant(1, dl, MVT::i32);
3507 SDValue FVal = DAG.getConstant(0, dl, MVT::i32);
3508 EVT VT = Op.getValueType();
3510 SDValue Overflow = DAG.getNode(ARMISD::CMOV, dl, VT, TVal, FVal,
3511 ARMcc, CCR, OverflowCmp);
3513 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
3514 return DAG.getNode(ISD::MERGE_VALUES, dl, VTs, Value, Overflow);
3518 SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
3519 SDValue Cond = Op.getOperand(0);
3520 SDValue SelectTrue = Op.getOperand(1);
3521 SDValue SelectFalse = Op.getOperand(2);
3523 unsigned Opc = Cond.getOpcode();
3525 if (Cond.getResNo() == 1 &&
3526 (Opc == ISD::SADDO || Opc == ISD::UADDO || Opc == ISD::SSUBO ||
3527 Opc == ISD::USUBO)) {
3528 if (!DAG.getTargetLoweringInfo().isTypeLegal(Cond->getValueType(0)))
3531 SDValue Value, OverflowCmp;
3533 std::tie(Value, OverflowCmp) = getARMXALUOOp(Cond, DAG, ARMcc);
3534 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3535 EVT VT = Op.getValueType();
3537 return getCMOV(dl, VT, SelectTrue, SelectFalse, ARMcc, CCR,
3543 // (select (cmov 1, 0, cond), t, f) -> (cmov t, f, cond)
3544 // (select (cmov 0, 1, cond), t, f) -> (cmov f, t, cond)
3546 if (Cond.getOpcode() == ARMISD::CMOV && Cond.hasOneUse()) {
3547 const ConstantSDNode *CMOVTrue =
3548 dyn_cast<ConstantSDNode>(Cond.getOperand(0));
3549 const ConstantSDNode *CMOVFalse =
3550 dyn_cast<ConstantSDNode>(Cond.getOperand(1));
3552 if (CMOVTrue && CMOVFalse) {
3553 unsigned CMOVTrueVal = CMOVTrue->getZExtValue();
3554 unsigned CMOVFalseVal = CMOVFalse->getZExtValue();
3558 if (CMOVTrueVal == 1 && CMOVFalseVal == 0) {
3560 False = SelectFalse;
3561 } else if (CMOVTrueVal == 0 && CMOVFalseVal == 1) {
3566 if (True.getNode() && False.getNode()) {
3567 EVT VT = Op.getValueType();
3568 SDValue ARMcc = Cond.getOperand(2);
3569 SDValue CCR = Cond.getOperand(3);
3570 SDValue Cmp = duplicateCmp(Cond.getOperand(4), DAG);
3571 assert(True.getValueType() == VT);
3572 return getCMOV(dl, VT, True, False, ARMcc, CCR, Cmp, DAG);
3577 // ARM's BooleanContents value is UndefinedBooleanContent. Mask out the
3578 // undefined bits before doing a full-word comparison with zero.
3579 Cond = DAG.getNode(ISD::AND, dl, Cond.getValueType(), Cond,
3580 DAG.getConstant(1, dl, Cond.getValueType()));
3582 return DAG.getSelectCC(dl, Cond,
3583 DAG.getConstant(0, dl, Cond.getValueType()),
3584 SelectTrue, SelectFalse, ISD::SETNE);
3587 static void checkVSELConstraints(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
3588 bool &swpCmpOps, bool &swpVselOps) {
3589 // Start by selecting the GE condition code for opcodes that return true for
3591 if (CC == ISD::SETUGE || CC == ISD::SETOGE || CC == ISD::SETOLE ||
3593 CondCode = ARMCC::GE;
3595 // and GT for opcodes that return false for 'equality'.
3596 else if (CC == ISD::SETUGT || CC == ISD::SETOGT || CC == ISD::SETOLT ||
3598 CondCode = ARMCC::GT;
3600 // Since we are constrained to GE/GT, if the opcode contains 'less', we need
3601 // to swap the compare operands.
3602 if (CC == ISD::SETOLE || CC == ISD::SETULE || CC == ISD::SETOLT ||
3606 // Both GT and GE are ordered comparisons, and return false for 'unordered'.
3607 // If we have an unordered opcode, we need to swap the operands to the VSEL
3608 // instruction (effectively negating the condition).
3610 // This also has the effect of swapping which one of 'less' or 'greater'
3611 // returns true, so we also swap the compare operands. It also switches
3612 // whether we return true for 'equality', so we compensate by picking the
3613 // opposite condition code to our original choice.
3614 if (CC == ISD::SETULE || CC == ISD::SETULT || CC == ISD::SETUGE ||
3615 CC == ISD::SETUGT) {
3616 swpCmpOps = !swpCmpOps;
3617 swpVselOps = !swpVselOps;
3618 CondCode = CondCode == ARMCC::GT ? ARMCC::GE : ARMCC::GT;
3621 // 'ordered' is 'anything but unordered', so use the VS condition code and
3622 // swap the VSEL operands.
3623 if (CC == ISD::SETO) {
3624 CondCode = ARMCC::VS;
3628 // 'unordered or not equal' is 'anything but equal', so use the EQ condition
3629 // code and swap the VSEL operands.
3630 if (CC == ISD::SETUNE) {
3631 CondCode = ARMCC::EQ;
3636 SDValue ARMTargetLowering::getCMOV(SDLoc dl, EVT VT, SDValue FalseVal,
3637 SDValue TrueVal, SDValue ARMcc, SDValue CCR,
3638 SDValue Cmp, SelectionDAG &DAG) const {
3639 if (Subtarget->isFPOnlySP() && VT == MVT::f64) {
3640 FalseVal = DAG.getNode(ARMISD::VMOVRRD, dl,
3641 DAG.getVTList(MVT::i32, MVT::i32), FalseVal);
3642 TrueVal = DAG.getNode(ARMISD::VMOVRRD, dl,
3643 DAG.getVTList(MVT::i32, MVT::i32), TrueVal);
3645 SDValue TrueLow = TrueVal.getValue(0);
3646 SDValue TrueHigh = TrueVal.getValue(1);
3647 SDValue FalseLow = FalseVal.getValue(0);
3648 SDValue FalseHigh = FalseVal.getValue(1);
3650 SDValue Low = DAG.getNode(ARMISD::CMOV, dl, MVT::i32, FalseLow, TrueLow,
3652 SDValue High = DAG.getNode(ARMISD::CMOV, dl, MVT::i32, FalseHigh, TrueHigh,
3653 ARMcc, CCR, duplicateCmp(Cmp, DAG));
3655 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Low, High);
3657 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,
3662 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
3663 EVT VT = Op.getValueType();
3664 SDValue LHS = Op.getOperand(0);
3665 SDValue RHS = Op.getOperand(1);
3666 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
3667 SDValue TrueVal = Op.getOperand(2);
3668 SDValue FalseVal = Op.getOperand(3);
3671 if (Subtarget->isFPOnlySP() && LHS.getValueType() == MVT::f64) {
3672 DAG.getTargetLoweringInfo().softenSetCCOperands(DAG, MVT::f64, LHS, RHS, CC,
3675 // If softenSetCCOperands only returned one value, we should compare it to
3677 if (!RHS.getNode()) {
3678 RHS = DAG.getConstant(0, dl, LHS.getValueType());
3683 if (LHS.getValueType() == MVT::i32) {
3684 // Try to generate VSEL on ARMv8.
3685 // The VSEL instruction can't use all the usual ARM condition
3686 // codes: it only has two bits to select the condition code, so it's
3687 // constrained to use only GE, GT, VS and EQ.
3689 // To implement all the various ISD::SETXXX opcodes, we sometimes need to
3690 // swap the operands of the previous compare instruction (effectively
3691 // inverting the compare condition, swapping 'less' and 'greater') and
3692 // sometimes need to swap the operands to the VSEL (which inverts the
3693 // condition in the sense of firing whenever the previous condition didn't)
3694 if (Subtarget->hasFPARMv8() && (TrueVal.getValueType() == MVT::f32 ||
3695 TrueVal.getValueType() == MVT::f64)) {
3696 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3697 if (CondCode == ARMCC::LT || CondCode == ARMCC::LE ||
3698 CondCode == ARMCC::VC || CondCode == ARMCC::NE) {
3699 CC = ISD::getSetCCInverse(CC, true);
3700 std::swap(TrueVal, FalseVal);
3705 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3706 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3707 return getCMOV(dl, VT, FalseVal, TrueVal, ARMcc, CCR, Cmp, DAG);
3710 ARMCC::CondCodes CondCode, CondCode2;
3711 FPCCToARMCC(CC, CondCode, CondCode2);
3713 // Try to generate VMAXNM/VMINNM on ARMv8.
3714 if (Subtarget->hasFPARMv8() && (TrueVal.getValueType() == MVT::f32 ||
3715 TrueVal.getValueType() == MVT::f64)) {
3716 bool swpCmpOps = false;
3717 bool swpVselOps = false;
3718 checkVSELConstraints(CC, CondCode, swpCmpOps, swpVselOps);
3720 if (CondCode == ARMCC::GT || CondCode == ARMCC::GE ||
3721 CondCode == ARMCC::VS || CondCode == ARMCC::EQ) {
3723 std::swap(LHS, RHS);
3725 std::swap(TrueVal, FalseVal);
3729 SDValue ARMcc = DAG.getConstant(CondCode, dl, MVT::i32);
3730 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
3731 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3732 SDValue Result = getCMOV(dl, VT, FalseVal, TrueVal, ARMcc, CCR, Cmp, DAG);
3733 if (CondCode2 != ARMCC::AL) {
3734 SDValue ARMcc2 = DAG.getConstant(CondCode2, dl, MVT::i32);
3735 // FIXME: Needs another CMP because flag can have but one use.
3736 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
3737 Result = getCMOV(dl, VT, Result, TrueVal, ARMcc2, CCR, Cmp2, DAG);
3742 /// canChangeToInt - Given the fp compare operand, return true if it is suitable
3743 /// to morph to an integer compare sequence.
3744 static bool canChangeToInt(SDValue Op, bool &SeenZero,
3745 const ARMSubtarget *Subtarget) {
3746 SDNode *N = Op.getNode();
3747 if (!N->hasOneUse())
3748 // Otherwise it requires moving the value from fp to integer registers.
3750 if (!N->getNumValues())
3752 EVT VT = Op.getValueType();
3753 if (VT != MVT::f32 && !Subtarget->isFPBrccSlow())
3754 // f32 case is generally profitable. f64 case only makes sense when vcmpe +
3755 // vmrs are very slow, e.g. cortex-a8.
3758 if (isFloatingPointZero(Op)) {
3762 return ISD::isNormalLoad(N);
3765 static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) {
3766 if (isFloatingPointZero(Op))
3767 return DAG.getConstant(0, SDLoc(Op), MVT::i32);
3769 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op))
3770 return DAG.getLoad(MVT::i32, SDLoc(Op),
3771 Ld->getChain(), Ld->getBasePtr(), Ld->getPointerInfo(),
3772 Ld->isVolatile(), Ld->isNonTemporal(),
3773 Ld->isInvariant(), Ld->getAlignment());
3775 llvm_unreachable("Unknown VFP cmp argument!");
3778 static void expandf64Toi32(SDValue Op, SelectionDAG &DAG,
3779 SDValue &RetVal1, SDValue &RetVal2) {
3782 if (isFloatingPointZero(Op)) {
3783 RetVal1 = DAG.getConstant(0, dl, MVT::i32);
3784 RetVal2 = DAG.getConstant(0, dl, MVT::i32);
3788 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) {
3789 SDValue Ptr = Ld->getBasePtr();
3790 RetVal1 = DAG.getLoad(MVT::i32, dl,
3791 Ld->getChain(), Ptr,
3792 Ld->getPointerInfo(),
3793 Ld->isVolatile(), Ld->isNonTemporal(),
3794 Ld->isInvariant(), Ld->getAlignment());
3796 EVT PtrType = Ptr.getValueType();
3797 unsigned NewAlign = MinAlign(Ld->getAlignment(), 4);
3798 SDValue NewPtr = DAG.getNode(ISD::ADD, dl,
3799 PtrType, Ptr, DAG.getConstant(4, dl, PtrType));
3800 RetVal2 = DAG.getLoad(MVT::i32, dl,
3801 Ld->getChain(), NewPtr,
3802 Ld->getPointerInfo().getWithOffset(4),
3803 Ld->isVolatile(), Ld->isNonTemporal(),
3804 Ld->isInvariant(), NewAlign);
3808 llvm_unreachable("Unknown VFP cmp argument!");
3811 /// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some
3812 /// f32 and even f64 comparisons to integer ones.
3814 ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const {
3815 SDValue Chain = Op.getOperand(0);
3816 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3817 SDValue LHS = Op.getOperand(2);
3818 SDValue RHS = Op.getOperand(3);
3819 SDValue Dest = Op.getOperand(4);
3822 bool LHSSeenZero = false;
3823 bool LHSOk = canChangeToInt(LHS, LHSSeenZero, Subtarget);
3824 bool RHSSeenZero = false;
3825 bool RHSOk = canChangeToInt(RHS, RHSSeenZero, Subtarget);
3826 if (LHSOk && RHSOk && (LHSSeenZero || RHSSeenZero)) {
3827 // If unsafe fp math optimization is enabled and there are no other uses of
3828 // the CMP operands, and the condition code is EQ or NE, we can optimize it
3829 // to an integer comparison.
3830 if (CC == ISD::SETOEQ)
3832 else if (CC == ISD::SETUNE)
3835 SDValue Mask = DAG.getConstant(0x7fffffff, dl, MVT::i32);
3837 if (LHS.getValueType() == MVT::f32) {
3838 LHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3839 bitcastf32Toi32(LHS, DAG), Mask);
3840 RHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3841 bitcastf32Toi32(RHS, DAG), Mask);
3842 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3843 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3844 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3845 Chain, Dest, ARMcc, CCR, Cmp);
3850 expandf64Toi32(LHS, DAG, LHS1, LHS2);
3851 expandf64Toi32(RHS, DAG, RHS1, RHS2);
3852 LHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, LHS2, Mask);
3853 RHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, RHS2, Mask);
3854 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3855 ARMcc = DAG.getConstant(CondCode, dl, MVT::i32);
3856 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3857 SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest };
3858 return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops);
3864 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
3865 SDValue Chain = Op.getOperand(0);
3866 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3867 SDValue LHS = Op.getOperand(2);
3868 SDValue RHS = Op.getOperand(3);
3869 SDValue Dest = Op.getOperand(4);
3872 if (Subtarget->isFPOnlySP() && LHS.getValueType() == MVT::f64) {
3873 DAG.getTargetLoweringInfo().softenSetCCOperands(DAG, MVT::f64, LHS, RHS, CC,
3876 // If softenSetCCOperands only returned one value, we should compare it to
3878 if (!RHS.getNode()) {
3879 RHS = DAG.getConstant(0, dl, LHS.getValueType());
3884 if (LHS.getValueType() == MVT::i32) {
3886 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3887 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3888 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3889 Chain, Dest, ARMcc, CCR, Cmp);
3892 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3894 if (getTargetMachine().Options.UnsafeFPMath &&
3895 (CC == ISD::SETEQ || CC == ISD::SETOEQ ||
3896 CC == ISD::SETNE || CC == ISD::SETUNE)) {
3897 SDValue Result = OptimizeVFPBrcond(Op, DAG);
3898 if (Result.getNode())
3902 ARMCC::CondCodes CondCode, CondCode2;
3903 FPCCToARMCC(CC, CondCode, CondCode2);
3905 SDValue ARMcc = DAG.getConstant(CondCode, dl, MVT::i32);
3906 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
3907 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3908 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3909 SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp };
3910 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops);
3911 if (CondCode2 != ARMCC::AL) {
3912 ARMcc = DAG.getConstant(CondCode2, dl, MVT::i32);
3913 SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) };
3914 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops);
3919 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
3920 SDValue Chain = Op.getOperand(0);
3921 SDValue Table = Op.getOperand(1);
3922 SDValue Index = Op.getOperand(2);
3925 EVT PTy = getPointerTy(DAG.getDataLayout());
3926 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
3927 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
3928 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI);
3929 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, dl, PTy));
3930 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
3931 if (Subtarget->isThumb2()) {
3932 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
3933 // which does another jump to the destination. This also makes it easier
3934 // to translate it to TBB / TBH later.
3935 // FIXME: This might not work if the function is extremely large.
3936 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
3937 Addr, Op.getOperand(2), JTI);
3939 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
3941 DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
3942 MachinePointerInfo::getJumpTable(DAG.getMachineFunction()),
3943 false, false, false, 0);
3944 Chain = Addr.getValue(1);
3945 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
3946 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI);
3949 DAG.getLoad(PTy, dl, Chain, Addr,
3950 MachinePointerInfo::getJumpTable(DAG.getMachineFunction()),
3951 false, false, false, 0);
3952 Chain = Addr.getValue(1);
3953 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI);
3957 static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3958 EVT VT = Op.getValueType();
3961 if (Op.getValueType().getVectorElementType() == MVT::i32) {
3962 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::f32)
3964 return DAG.UnrollVectorOp(Op.getNode());
3967 assert(Op.getOperand(0).getValueType() == MVT::v4f32 &&
3968 "Invalid type for custom lowering!");
3969 if (VT != MVT::v4i16)
3970 return DAG.UnrollVectorOp(Op.getNode());
3972 Op = DAG.getNode(Op.getOpcode(), dl, MVT::v4i32, Op.getOperand(0));
3973 return DAG.getNode(ISD::TRUNCATE, dl, VT, Op);
3976 SDValue ARMTargetLowering::LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) const {
3977 EVT VT = Op.getValueType();
3979 return LowerVectorFP_TO_INT(Op, DAG);
3980 if (Subtarget->isFPOnlySP() && Op.getOperand(0).getValueType() == MVT::f64) {
3982 if (Op.getOpcode() == ISD::FP_TO_SINT)
3983 LC = RTLIB::getFPTOSINT(Op.getOperand(0).getValueType(),
3986 LC = RTLIB::getFPTOUINT(Op.getOperand(0).getValueType(),
3988 return makeLibCall(DAG, LC, Op.getValueType(), Op.getOperand(0),
3989 /*isSigned*/ false, SDLoc(Op)).first;
3995 static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3996 EVT VT = Op.getValueType();
3999 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::i32) {
4000 if (VT.getVectorElementType() == MVT::f32)
4002 return DAG.UnrollVectorOp(Op.getNode());
4005 assert(Op.getOperand(0).getValueType() == MVT::v4i16 &&
4006 "Invalid type for custom lowering!");
4007 if (VT != MVT::v4f32)
4008 return DAG.UnrollVectorOp(Op.getNode());
4012 switch (Op.getOpcode()) {
4013 default: llvm_unreachable("Invalid opcode!");
4014 case ISD::SINT_TO_FP:
4015 CastOpc = ISD::SIGN_EXTEND;
4016 Opc = ISD::SINT_TO_FP;
4018 case ISD::UINT_TO_FP:
4019 CastOpc = ISD::ZERO_EXTEND;
4020 Opc = ISD::UINT_TO_FP;
4024 Op = DAG.getNode(CastOpc, dl, MVT::v4i32, Op.getOperand(0));
4025 return DAG.getNode(Opc, dl, VT, Op);
4028 SDValue ARMTargetLowering::LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) const {
4029 EVT VT = Op.getValueType();
4031 return LowerVectorINT_TO_FP(Op, DAG);
4032 if (Subtarget->isFPOnlySP() && Op.getValueType() == MVT::f64) {
4034 if (Op.getOpcode() == ISD::SINT_TO_FP)
4035 LC = RTLIB::getSINTTOFP(Op.getOperand(0).getValueType(),
4038 LC = RTLIB::getUINTTOFP(Op.getOperand(0).getValueType(),
4040 return makeLibCall(DAG, LC, Op.getValueType(), Op.getOperand(0),
4041 /*isSigned*/ false, SDLoc(Op)).first;
4047 SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
4048 // Implement fcopysign with a fabs and a conditional fneg.
4049 SDValue Tmp0 = Op.getOperand(0);
4050 SDValue Tmp1 = Op.getOperand(1);
4052 EVT VT = Op.getValueType();
4053 EVT SrcVT = Tmp1.getValueType();
4054 bool InGPR = Tmp0.getOpcode() == ISD::BITCAST ||
4055 Tmp0.getOpcode() == ARMISD::VMOVDRR;
4056 bool UseNEON = !InGPR && Subtarget->hasNEON();
4059 // Use VBSL to copy the sign bit.
4060 unsigned EncodedVal = ARM_AM::createNEONModImm(0x6, 0x80);
4061 SDValue Mask = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v2i32,
4062 DAG.getTargetConstant(EncodedVal, dl, MVT::i32));
4063 EVT OpVT = (VT == MVT::f32) ? MVT::v2i32 : MVT::v1i64;
4065 Mask = DAG.getNode(ARMISD::VSHL, dl, OpVT,
4066 DAG.getNode(ISD::BITCAST, dl, OpVT, Mask),
4067 DAG.getConstant(32, dl, MVT::i32));
4068 else /*if (VT == MVT::f32)*/
4069 Tmp0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp0);
4070 if (SrcVT == MVT::f32) {
4071 Tmp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp1);
4073 Tmp1 = DAG.getNode(ARMISD::VSHL, dl, OpVT,
4074 DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1),
4075 DAG.getConstant(32, dl, MVT::i32));
4076 } else if (VT == MVT::f32)
4077 Tmp1 = DAG.getNode(ARMISD::VSHRu, dl, MVT::v1i64,
4078 DAG.getNode(ISD::BITCAST, dl, MVT::v1i64, Tmp1),
4079 DAG.getConstant(32, dl, MVT::i32));
4080 Tmp0 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp0);
4081 Tmp1 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1);
4083 SDValue AllOnes = DAG.getTargetConstant(ARM_AM::createNEONModImm(0xe, 0xff),
4085 AllOnes = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v8i8, AllOnes);
4086 SDValue MaskNot = DAG.getNode(ISD::XOR, dl, OpVT, Mask,
4087 DAG.getNode(ISD::BITCAST, dl, OpVT, AllOnes));
4089 SDValue Res = DAG.getNode(ISD::OR, dl, OpVT,
4090 DAG.getNode(ISD::AND, dl, OpVT, Tmp1, Mask),
4091 DAG.getNode(ISD::AND, dl, OpVT, Tmp0, MaskNot));
4092 if (VT == MVT::f32) {
4093 Res = DAG.getNode(ISD::BITCAST, dl, MVT::v2f32, Res);
4094 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, Res,
4095 DAG.getConstant(0, dl, MVT::i32));
4097 Res = DAG.getNode(ISD::BITCAST, dl, MVT::f64, Res);
4103 // Bitcast operand 1 to i32.
4104 if (SrcVT == MVT::f64)
4105 Tmp1 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
4107 Tmp1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp1);
4109 // Or in the signbit with integer operations.
4110 SDValue Mask1 = DAG.getConstant(0x80000000, dl, MVT::i32);
4111 SDValue Mask2 = DAG.getConstant(0x7fffffff, dl, MVT::i32);
4112 Tmp1 = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp1, Mask1);
4113 if (VT == MVT::f32) {
4114 Tmp0 = DAG.getNode(ISD::AND, dl, MVT::i32,
4115 DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp0), Mask2);
4116 return DAG.getNode(ISD::BITCAST, dl, MVT::f32,
4117 DAG.getNode(ISD::OR, dl, MVT::i32, Tmp0, Tmp1));
4120 // f64: Or the high part with signbit and then combine two parts.
4121 Tmp0 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
4123 SDValue Lo = Tmp0.getValue(0);
4124 SDValue Hi = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp0.getValue(1), Mask2);
4125 Hi = DAG.getNode(ISD::OR, dl, MVT::i32, Hi, Tmp1);
4126 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
4129 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
4130 MachineFunction &MF = DAG.getMachineFunction();
4131 MachineFrameInfo *MFI = MF.getFrameInfo();
4132 MFI->setReturnAddressIsTaken(true);
4134 if (verifyReturnAddressArgumentIsConstant(Op, DAG))
4137 EVT VT = Op.getValueType();
4139 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
4141 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
4142 SDValue Offset = DAG.getConstant(4, dl, MVT::i32);
4143 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
4144 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
4145 MachinePointerInfo(), false, false, false, 0);
4148 // Return LR, which contains the return address. Mark it an implicit live-in.
4149 unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32));
4150 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
4153 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
4154 const ARMBaseRegisterInfo &ARI =
4155 *static_cast<const ARMBaseRegisterInfo*>(RegInfo);
4156 MachineFunction &MF = DAG.getMachineFunction();
4157 MachineFrameInfo *MFI = MF.getFrameInfo();
4158 MFI->setFrameAddressIsTaken(true);
4160 EVT VT = Op.getValueType();
4161 SDLoc dl(Op); // FIXME probably not meaningful
4162 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
4163 unsigned FrameReg = ARI.getFrameRegister(MF);
4164 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
4166 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
4167 MachinePointerInfo(),
4168 false, false, false, 0);
4172 // FIXME? Maybe this could be a TableGen attribute on some registers and
4173 // this table could be generated automatically from RegInfo.
4174 unsigned ARMTargetLowering::getRegisterByName(const char* RegName, EVT VT,
4175 SelectionDAG &DAG) const {
4176 unsigned Reg = StringSwitch<unsigned>(RegName)
4177 .Case("sp", ARM::SP)
4181 report_fatal_error(Twine("Invalid register name \""
4182 + StringRef(RegName) + "\"."));
4185 // Result is 64 bit value so split into two 32 bit values and return as a
4187 static void ExpandREAD_REGISTER(SDNode *N, SmallVectorImpl<SDValue> &Results,
4188 SelectionDAG &DAG) {
4191 // This function is only supposed to be called for i64 type destination.
4192 assert(N->getValueType(0) == MVT::i64
4193 && "ExpandREAD_REGISTER called for non-i64 type result.");
4195 SDValue Read = DAG.getNode(ISD::READ_REGISTER, DL,
4196 DAG.getVTList(MVT::i32, MVT::i32, MVT::Other),
4200 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Read.getValue(0),
4202 Results.push_back(Read.getOperand(0));
4205 /// \p BC is a bitcast that is about to be turned into a VMOVDRR.
4206 /// When \p DstVT, the destination type of \p BC, is on the vector
4207 /// register bank and the source of bitcast, \p Op, operates on the same bank,
4208 /// it might be possible to combine them, such that everything stays on the
4209 /// vector register bank.
4210 /// \p return The node that would replace \p BT, if the combine
4212 static SDValue CombineVMOVDRRCandidateWithVecOp(const SDNode *BC,
4213 SelectionDAG &DAG) {
4214 SDValue Op = BC->getOperand(0);
4215 EVT DstVT = BC->getValueType(0);
4217 // The only vector instruction that can produce a scalar (remember,
4218 // since the bitcast was about to be turned into VMOVDRR, the source
4219 // type is i64) from a vector is EXTRACT_VECTOR_ELT.
4220 // Moreover, we can do this combine only if there is one use.
4221 // Finally, if the destination type is not a vector, there is not
4222 // much point on forcing everything on the vector bank.
4223 if (!DstVT.isVector() || Op.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
4227 // If the index is not constant, we will introduce an additional
4228 // multiply that will stick.
4229 // Give up in that case.
4230 ConstantSDNode *Index = dyn_cast<ConstantSDNode>(Op.getOperand(1));
4233 unsigned DstNumElt = DstVT.getVectorNumElements();
4235 // Compute the new index.
4236 const APInt &APIntIndex = Index->getAPIntValue();
4237 APInt NewIndex(APIntIndex.getBitWidth(), DstNumElt);
4238 NewIndex *= APIntIndex;
4239 // Check if the new constant index fits into i32.
4240 if (NewIndex.getBitWidth() > 32)
4243 // vMTy bitcast(i64 extractelt vNi64 src, i32 index) ->
4244 // vMTy extractsubvector vNxMTy (bitcast vNi64 src), i32 index*M)
4246 SDValue ExtractSrc = Op.getOperand(0);
4247 EVT VecVT = EVT::getVectorVT(
4248 *DAG.getContext(), DstVT.getScalarType(),
4249 ExtractSrc.getValueType().getVectorNumElements() * DstNumElt);
4250 SDValue BitCast = DAG.getNode(ISD::BITCAST, dl, VecVT, ExtractSrc);
4251 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DstVT, BitCast,
4252 DAG.getConstant(NewIndex.getZExtValue(), dl, MVT::i32));
4255 /// ExpandBITCAST - If the target supports VFP, this function is called to
4256 /// expand a bit convert where either the source or destination type is i64 to
4257 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
4258 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
4259 /// vectors), since the legalizer won't know what to do with that.
4260 static SDValue ExpandBITCAST(SDNode *N, SelectionDAG &DAG) {
4261 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4263 SDValue Op = N->getOperand(0);
4265 // This function is only supposed to be called for i64 types, either as the
4266 // source or destination of the bit convert.
4267 EVT SrcVT = Op.getValueType();
4268 EVT DstVT = N->getValueType(0);
4269 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
4270 "ExpandBITCAST called for non-i64 type");
4272 // Turn i64->f64 into VMOVDRR.
4273 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
4274 // Do not force values to GPRs (this is what VMOVDRR does for the inputs)
4275 // if we can combine the bitcast with its source.
4276 if (SDValue Val = CombineVMOVDRRCandidateWithVecOp(N, DAG))
4279 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
4280 DAG.getConstant(0, dl, MVT::i32));
4281 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
4282 DAG.getConstant(1, dl, MVT::i32));
4283 return DAG.getNode(ISD::BITCAST, dl, DstVT,
4284 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
4287 // Turn f64->i64 into VMOVRRD.
4288 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
4290 if (DAG.getDataLayout().isBigEndian() && SrcVT.isVector() &&
4291 SrcVT.getVectorNumElements() > 1)
4292 Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
4293 DAG.getVTList(MVT::i32, MVT::i32),
4294 DAG.getNode(ARMISD::VREV64, dl, SrcVT, Op));
4296 Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
4297 DAG.getVTList(MVT::i32, MVT::i32), Op);
4298 // Merge the pieces into a single i64 value.
4299 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
4305 /// getZeroVector - Returns a vector of specified type with all zero elements.
4306 /// Zero vectors are used to represent vector negation and in those cases
4307 /// will be implemented with the NEON VNEG instruction. However, VNEG does
4308 /// not support i64 elements, so sometimes the zero vectors will need to be
4309 /// explicitly constructed. Regardless, use a canonical VMOV to create the
4311 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, SDLoc dl) {
4312 assert(VT.isVector() && "Expected a vector type");
4313 // The canonical modified immediate encoding of a zero vector is....0!
4314 SDValue EncodedVal = DAG.getTargetConstant(0, dl, MVT::i32);
4315 EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
4316 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal);
4317 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4320 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
4321 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
4322 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
4323 SelectionDAG &DAG) const {
4324 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4325 EVT VT = Op.getValueType();
4326 unsigned VTBits = VT.getSizeInBits();
4328 SDValue ShOpLo = Op.getOperand(0);
4329 SDValue ShOpHi = Op.getOperand(1);
4330 SDValue ShAmt = Op.getOperand(2);
4332 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
4334 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
4336 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
4337 DAG.getConstant(VTBits, dl, MVT::i32), ShAmt);
4338 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
4339 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
4340 DAG.getConstant(VTBits, dl, MVT::i32));
4341 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
4342 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
4343 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
4345 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
4346 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, dl, MVT::i32),
4347 ISD::SETGE, ARMcc, DAG, dl);
4348 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
4349 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc,
4352 SDValue Ops[2] = { Lo, Hi };
4353 return DAG.getMergeValues(Ops, dl);
4356 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
4357 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
4358 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
4359 SelectionDAG &DAG) const {
4360 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4361 EVT VT = Op.getValueType();
4362 unsigned VTBits = VT.getSizeInBits();
4364 SDValue ShOpLo = Op.getOperand(0);
4365 SDValue ShOpHi = Op.getOperand(1);
4366 SDValue ShAmt = Op.getOperand(2);
4369 assert(Op.getOpcode() == ISD::SHL_PARTS);
4370 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
4371 DAG.getConstant(VTBits, dl, MVT::i32), ShAmt);
4372 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
4373 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
4374 DAG.getConstant(VTBits, dl, MVT::i32));
4375 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
4376 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
4378 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
4379 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
4380 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, dl, MVT::i32),
4381 ISD::SETGE, ARMcc, DAG, dl);
4382 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
4383 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc,
4386 SDValue Ops[2] = { Lo, Hi };
4387 return DAG.getMergeValues(Ops, dl);
4390 SDValue ARMTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
4391 SelectionDAG &DAG) const {
4392 // The rounding mode is in bits 23:22 of the FPSCR.
4393 // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0
4394 // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3)
4395 // so that the shift + and get folded into a bitfield extract.
4397 SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::i32,
4398 DAG.getConstant(Intrinsic::arm_get_fpscr, dl,
4400 SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPSCR,
4401 DAG.getConstant(1U << 22, dl, MVT::i32));
4402 SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds,
4403 DAG.getConstant(22, dl, MVT::i32));
4404 return DAG.getNode(ISD::AND, dl, MVT::i32, RMODE,
4405 DAG.getConstant(3, dl, MVT::i32));
4408 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
4409 const ARMSubtarget *ST) {
4411 EVT VT = N->getValueType(0);
4412 if (VT.isVector()) {
4413 assert(ST->hasNEON());
4415 // Compute the least significant set bit: LSB = X & -X
4416 SDValue X = N->getOperand(0);
4417 SDValue NX = DAG.getNode(ISD::SUB, dl, VT, getZeroVector(VT, DAG, dl), X);
4418 SDValue LSB = DAG.getNode(ISD::AND, dl, VT, X, NX);
4420 EVT ElemTy = VT.getVectorElementType();
4422 if (ElemTy == MVT::i8) {
4423 // Compute with: cttz(x) = ctpop(lsb - 1)
4424 SDValue One = DAG.getNode(ARMISD::VMOVIMM, dl, VT,
4425 DAG.getTargetConstant(1, dl, ElemTy));
4426 SDValue Bits = DAG.getNode(ISD::SUB, dl, VT, LSB, One);
4427 return DAG.getNode(ISD::CTPOP, dl, VT, Bits);
4430 if ((ElemTy == MVT::i16 || ElemTy == MVT::i32) &&
4431 (N->getOpcode() == ISD::CTTZ_ZERO_UNDEF)) {
4432 // Compute with: cttz(x) = (width - 1) - ctlz(lsb), if x != 0
4433 unsigned NumBits = ElemTy.getSizeInBits();
4434 SDValue WidthMinus1 =
4435 DAG.getNode(ARMISD::VMOVIMM, dl, VT,
4436 DAG.getTargetConstant(NumBits - 1, dl, ElemTy));
4437 SDValue CTLZ = DAG.getNode(ISD::CTLZ, dl, VT, LSB);
4438 return DAG.getNode(ISD::SUB, dl, VT, WidthMinus1, CTLZ);
4441 // Compute with: cttz(x) = ctpop(lsb - 1)
4443 // Since we can only compute the number of bits in a byte with vcnt.8, we
4444 // have to gather the result with pairwise addition (vpaddl) for i16, i32,
4449 if (ElemTy == MVT::i64) {
4450 // Load constant 0xffff'ffff'ffff'ffff to register.
4451 SDValue FF = DAG.getNode(ARMISD::VMOVIMM, dl, VT,
4452 DAG.getTargetConstant(0x1eff, dl, MVT::i32));
4453 Bits = DAG.getNode(ISD::ADD, dl, VT, LSB, FF);
4455 SDValue One = DAG.getNode(ARMISD::VMOVIMM, dl, VT,
4456 DAG.getTargetConstant(1, dl, ElemTy));
4457 Bits = DAG.getNode(ISD::SUB, dl, VT, LSB, One);
4460 // Count #bits with vcnt.8.
4461 EVT VT8Bit = VT.is64BitVector() ? MVT::v8i8 : MVT::v16i8;
4462 SDValue BitsVT8 = DAG.getNode(ISD::BITCAST, dl, VT8Bit, Bits);
4463 SDValue Cnt8 = DAG.getNode(ISD::CTPOP, dl, VT8Bit, BitsVT8);
4465 // Gather the #bits with vpaddl (pairwise add.)
4466 EVT VT16Bit = VT.is64BitVector() ? MVT::v4i16 : MVT::v8i16;
4467 SDValue Cnt16 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT16Bit,
4468 DAG.getTargetConstant(Intrinsic::arm_neon_vpaddlu, dl, MVT::i32),
4470 if (ElemTy == MVT::i16)
4473 EVT VT32Bit = VT.is64BitVector() ? MVT::v2i32 : MVT::v4i32;
4474 SDValue Cnt32 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT32Bit,
4475 DAG.getTargetConstant(Intrinsic::arm_neon_vpaddlu, dl, MVT::i32),
4477 if (ElemTy == MVT::i32)
4480 assert(ElemTy == MVT::i64);
4481 SDValue Cnt64 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
4482 DAG.getTargetConstant(Intrinsic::arm_neon_vpaddlu, dl, MVT::i32),
4487 if (!ST->hasV6T2Ops())
4490 SDValue rbit = DAG.getNode(ISD::BITREVERSE, dl, VT, N->getOperand(0));
4491 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
4494 /// getCTPOP16BitCounts - Returns a v8i8/v16i8 vector containing the bit-count
4495 /// for each 16-bit element from operand, repeated. The basic idea is to
4496 /// leverage vcnt to get the 8-bit counts, gather and add the results.
4498 /// Trace for v4i16:
4499 /// input = [v0 v1 v2 v3 ] (vi 16-bit element)
4500 /// cast: N0 = [w0 w1 w2 w3 w4 w5 w6 w7] (v0 = [w0 w1], wi 8-bit element)
4501 /// vcnt: N1 = [b0 b1 b2 b3 b4 b5 b6 b7] (bi = bit-count of 8-bit element wi)
4502 /// vrev: N2 = [b1 b0 b3 b2 b5 b4 b7 b6]
4503 /// [b0 b1 b2 b3 b4 b5 b6 b7]
4504 /// +[b1 b0 b3 b2 b5 b4 b7 b6]
4505 /// N3=N1+N2 = [k0 k0 k1 k1 k2 k2 k3 k3] (k0 = b0+b1 = bit-count of 16-bit v0,
4506 /// vuzp: = [k0 k1 k2 k3 k0 k1 k2 k3] each ki is 8-bits)
4507 static SDValue getCTPOP16BitCounts(SDNode *N, SelectionDAG &DAG) {
4508 EVT VT = N->getValueType(0);
4511 EVT VT8Bit = VT.is64BitVector() ? MVT::v8i8 : MVT::v16i8;
4512 SDValue N0 = DAG.getNode(ISD::BITCAST, DL, VT8Bit, N->getOperand(0));
4513 SDValue N1 = DAG.getNode(ISD::CTPOP, DL, VT8Bit, N0);
4514 SDValue N2 = DAG.getNode(ARMISD::VREV16, DL, VT8Bit, N1);
4515 SDValue N3 = DAG.getNode(ISD::ADD, DL, VT8Bit, N1, N2);
4516 return DAG.getNode(ARMISD::VUZP, DL, VT8Bit, N3, N3);
4519 /// lowerCTPOP16BitElements - Returns a v4i16/v8i16 vector containing the
4520 /// bit-count for each 16-bit element from the operand. We need slightly
4521 /// different sequencing for v4i16 and v8i16 to stay within NEON's available
4522 /// 64/128-bit registers.
4524 /// Trace for v4i16:
4525 /// input = [v0 v1 v2 v3 ] (vi 16-bit element)
4526 /// v8i8: BitCounts = [k0 k1 k2 k3 k0 k1 k2 k3 ] (ki is the bit-count of vi)
4527 /// v8i16:Extended = [k0 k1 k2 k3 k0 k1 k2 k3 ]
4528 /// v4i16:Extracted = [k0 k1 k2 k3 ]
4529 static SDValue lowerCTPOP16BitElements(SDNode *N, SelectionDAG &DAG) {
4530 EVT VT = N->getValueType(0);
4533 SDValue BitCounts = getCTPOP16BitCounts(N, DAG);
4534 if (VT.is64BitVector()) {
4535 SDValue Extended = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v8i16, BitCounts);
4536 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i16, Extended,
4537 DAG.getIntPtrConstant(0, DL));
4539 SDValue Extracted = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v8i8,
4540 BitCounts, DAG.getIntPtrConstant(0, DL));
4541 return DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v8i16, Extracted);
4545 /// lowerCTPOP32BitElements - Returns a v2i32/v4i32 vector containing the
4546 /// bit-count for each 32-bit element from the operand. The idea here is
4547 /// to split the vector into 16-bit elements, leverage the 16-bit count
4548 /// routine, and then combine the results.
4550 /// Trace for v2i32 (v4i32 similar with Extracted/Extended exchanged):
4551 /// input = [v0 v1 ] (vi: 32-bit elements)
4552 /// Bitcast = [w0 w1 w2 w3 ] (wi: 16-bit elements, v0 = [w0 w1])
4553 /// Counts16 = [k0 k1 k2 k3 ] (ki: 16-bit elements, bit-count of wi)
4554 /// vrev: N0 = [k1 k0 k3 k2 ]
4556 /// N1 =+[k1 k0 k3 k2 ]
4558 /// N2 =+[k1 k3 k0 k2 ]
4560 /// Extended =+[k1 k3 k0 k2 ]
4562 /// Extracted=+[k1 k3 ]
4564 static SDValue lowerCTPOP32BitElements(SDNode *N, SelectionDAG &DAG) {
4565 EVT VT = N->getValueType(0);
4568 EVT VT16Bit = VT.is64BitVector() ? MVT::v4i16 : MVT::v8i16;
4570 SDValue Bitcast = DAG.getNode(ISD::BITCAST, DL, VT16Bit, N->getOperand(0));
4571 SDValue Counts16 = lowerCTPOP16BitElements(Bitcast.getNode(), DAG);
4572 SDValue N0 = DAG.getNode(ARMISD::VREV32, DL, VT16Bit, Counts16);
4573 SDValue N1 = DAG.getNode(ISD::ADD, DL, VT16Bit, Counts16, N0);
4574 SDValue N2 = DAG.getNode(ARMISD::VUZP, DL, VT16Bit, N1, N1);
4576 if (VT.is64BitVector()) {
4577 SDValue Extended = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v4i32, N2);
4578 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i32, Extended,
4579 DAG.getIntPtrConstant(0, DL));
4581 SDValue Extracted = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i16, N2,
4582 DAG.getIntPtrConstant(0, DL));
4583 return DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v4i32, Extracted);
4587 static SDValue LowerCTPOP(SDNode *N, SelectionDAG &DAG,
4588 const ARMSubtarget *ST) {
4589 EVT VT = N->getValueType(0);
4591 assert(ST->hasNEON() && "Custom ctpop lowering requires NEON.");
4592 assert((VT == MVT::v2i32 || VT == MVT::v4i32 ||
4593 VT == MVT::v4i16 || VT == MVT::v8i16) &&
4594 "Unexpected type for custom ctpop lowering");
4596 if (VT.getVectorElementType() == MVT::i32)
4597 return lowerCTPOP32BitElements(N, DAG);
4599 return lowerCTPOP16BitElements(N, DAG);
4602 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
4603 const ARMSubtarget *ST) {
4604 EVT VT = N->getValueType(0);
4610 // Lower vector shifts on NEON to use VSHL.
4611 assert(ST->hasNEON() && "unexpected vector shift");
4613 // Left shifts translate directly to the vshiftu intrinsic.
4614 if (N->getOpcode() == ISD::SHL)
4615 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
4616 DAG.getConstant(Intrinsic::arm_neon_vshiftu, dl,
4618 N->getOperand(0), N->getOperand(1));
4620 assert((N->getOpcode() == ISD::SRA ||
4621 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
4623 // NEON uses the same intrinsics for both left and right shifts. For
4624 // right shifts, the shift amounts are negative, so negate the vector of
4626 EVT ShiftVT = N->getOperand(1).getValueType();
4627 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
4628 getZeroVector(ShiftVT, DAG, dl),
4630 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
4631 Intrinsic::arm_neon_vshifts :
4632 Intrinsic::arm_neon_vshiftu);
4633 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
4634 DAG.getConstant(vshiftInt, dl, MVT::i32),
4635 N->getOperand(0), NegatedCount);
4638 static SDValue Expand64BitShift(SDNode *N, SelectionDAG &DAG,
4639 const ARMSubtarget *ST) {
4640 EVT VT = N->getValueType(0);
4643 // We can get here for a node like i32 = ISD::SHL i32, i64
4647 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
4648 "Unknown shift to lower!");
4650 // We only lower SRA, SRL of 1 here, all others use generic lowering.
4651 if (!isOneConstant(N->getOperand(1)))
4654 // If we are in thumb mode, we don't have RRX.
4655 if (ST->isThumb1Only()) return SDValue();
4657 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
4658 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
4659 DAG.getConstant(0, dl, MVT::i32));
4660 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
4661 DAG.getConstant(1, dl, MVT::i32));
4663 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
4664 // captures the result into a carry flag.
4665 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
4666 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Glue), Hi);
4668 // The low part is an ARMISD::RRX operand, which shifts the carry in.
4669 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
4671 // Merge the pieces into a single i64 value.
4672 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
4675 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
4676 SDValue TmpOp0, TmpOp1;
4677 bool Invert = false;
4681 SDValue Op0 = Op.getOperand(0);
4682 SDValue Op1 = Op.getOperand(1);
4683 SDValue CC = Op.getOperand(2);
4684 EVT CmpVT = Op0.getValueType().changeVectorElementTypeToInteger();
4685 EVT VT = Op.getValueType();
4686 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
4689 if (CmpVT.getVectorElementType() == MVT::i64)
4690 // 64-bit comparisons are not legal. We've marked SETCC as non-Custom,
4691 // but it's possible that our operands are 64-bit but our result is 32-bit.
4692 // Bail in this case.
4695 if (Op1.getValueType().isFloatingPoint()) {
4696 switch (SetCCOpcode) {
4697 default: llvm_unreachable("Illegal FP comparison");
4699 case ISD::SETNE: Invert = true; // Fallthrough
4701 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
4703 case ISD::SETLT: Swap = true; // Fallthrough
4705 case ISD::SETGT: Opc = ARMISD::VCGT; break;
4707 case ISD::SETLE: Swap = true; // Fallthrough
4709 case ISD::SETGE: Opc = ARMISD::VCGE; break;
4710 case ISD::SETUGE: Swap = true; // Fallthrough
4711 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
4712 case ISD::SETUGT: Swap = true; // Fallthrough
4713 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
4714 case ISD::SETUEQ: Invert = true; // Fallthrough
4716 // Expand this to (OLT | OGT).
4720 Op0 = DAG.getNode(ARMISD::VCGT, dl, CmpVT, TmpOp1, TmpOp0);
4721 Op1 = DAG.getNode(ARMISD::VCGT, dl, CmpVT, TmpOp0, TmpOp1);
4723 case ISD::SETUO: Invert = true; // Fallthrough
4725 // Expand this to (OLT | OGE).
4729 Op0 = DAG.getNode(ARMISD::VCGT, dl, CmpVT, TmpOp1, TmpOp0);
4730 Op1 = DAG.getNode(ARMISD::VCGE, dl, CmpVT, TmpOp0, TmpOp1);
4734 // Integer comparisons.
4735 switch (SetCCOpcode) {
4736 default: llvm_unreachable("Illegal integer comparison");
4737 case ISD::SETNE: Invert = true;
4738 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
4739 case ISD::SETLT: Swap = true;
4740 case ISD::SETGT: Opc = ARMISD::VCGT; break;
4741 case ISD::SETLE: Swap = true;
4742 case ISD::SETGE: Opc = ARMISD::VCGE; break;
4743 case ISD::SETULT: Swap = true;
4744 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
4745 case ISD::SETULE: Swap = true;
4746 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
4749 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
4750 if (Opc == ARMISD::VCEQ) {
4753 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
4755 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
4758 // Ignore bitconvert.
4759 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BITCAST)
4760 AndOp = AndOp.getOperand(0);
4762 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
4764 Op0 = DAG.getNode(ISD::BITCAST, dl, CmpVT, AndOp.getOperand(0));
4765 Op1 = DAG.getNode(ISD::BITCAST, dl, CmpVT, AndOp.getOperand(1));
4772 std::swap(Op0, Op1);
4774 // If one of the operands is a constant vector zero, attempt to fold the
4775 // comparison to a specialized compare-against-zero form.
4777 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
4779 else if (ISD::isBuildVectorAllZeros(Op0.getNode())) {
4780 if (Opc == ARMISD::VCGE)
4781 Opc = ARMISD::VCLEZ;
4782 else if (Opc == ARMISD::VCGT)
4783 Opc = ARMISD::VCLTZ;
4788 if (SingleOp.getNode()) {
4791 Result = DAG.getNode(ARMISD::VCEQZ, dl, CmpVT, SingleOp); break;
4793 Result = DAG.getNode(ARMISD::VCGEZ, dl, CmpVT, SingleOp); break;
4795 Result = DAG.getNode(ARMISD::VCLEZ, dl, CmpVT, SingleOp); break;
4797 Result = DAG.getNode(ARMISD::VCGTZ, dl, CmpVT, SingleOp); break;
4799 Result = DAG.getNode(ARMISD::VCLTZ, dl, CmpVT, SingleOp); break;
4801 Result = DAG.getNode(Opc, dl, CmpVT, Op0, Op1);
4804 Result = DAG.getNode(Opc, dl, CmpVT, Op0, Op1);
4807 Result = DAG.getSExtOrTrunc(Result, dl, VT);
4810 Result = DAG.getNOT(dl, Result, VT);
4815 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
4816 /// valid vector constant for a NEON instruction with a "modified immediate"
4817 /// operand (e.g., VMOV). If so, return the encoded value.
4818 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
4819 unsigned SplatBitSize, SelectionDAG &DAG,
4820 SDLoc dl, EVT &VT, bool is128Bits,
4821 NEONModImmType type) {
4822 unsigned OpCmode, Imm;
4824 // SplatBitSize is set to the smallest size that splats the vector, so a
4825 // zero vector will always have SplatBitSize == 8. However, NEON modified
4826 // immediate instructions others than VMOV do not support the 8-bit encoding
4827 // of a zero vector, and the default encoding of zero is supposed to be the
4832 switch (SplatBitSize) {
4834 if (type != VMOVModImm)
4836 // Any 1-byte value is OK. Op=0, Cmode=1110.
4837 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
4840 VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
4844 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
4845 VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
4846 if ((SplatBits & ~0xff) == 0) {
4847 // Value = 0x00nn: Op=x, Cmode=100x.
4852 if ((SplatBits & ~0xff00) == 0) {
4853 // Value = 0xnn00: Op=x, Cmode=101x.
4855 Imm = SplatBits >> 8;
4861 // NEON's 32-bit VMOV supports splat values where:
4862 // * only one byte is nonzero, or
4863 // * the least significant byte is 0xff and the second byte is nonzero, or
4864 // * the least significant 2 bytes are 0xff and the third is nonzero.
4865 VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
4866 if ((SplatBits & ~0xff) == 0) {
4867 // Value = 0x000000nn: Op=x, Cmode=000x.
4872 if ((SplatBits & ~0xff00) == 0) {
4873 // Value = 0x0000nn00: Op=x, Cmode=001x.
4875 Imm = SplatBits >> 8;
4878 if ((SplatBits & ~0xff0000) == 0) {
4879 // Value = 0x00nn0000: Op=x, Cmode=010x.
4881 Imm = SplatBits >> 16;
4884 if ((SplatBits & ~0xff000000) == 0) {
4885 // Value = 0xnn000000: Op=x, Cmode=011x.
4887 Imm = SplatBits >> 24;
4891 // cmode == 0b1100 and cmode == 0b1101 are not supported for VORR or VBIC
4892 if (type == OtherModImm) return SDValue();
4894 if ((SplatBits & ~0xffff) == 0 &&
4895 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
4896 // Value = 0x0000nnff: Op=x, Cmode=1100.
4898 Imm = SplatBits >> 8;
4902 if ((SplatBits & ~0xffffff) == 0 &&
4903 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
4904 // Value = 0x00nnffff: Op=x, Cmode=1101.
4906 Imm = SplatBits >> 16;
4910 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
4911 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
4912 // VMOV.I32. A (very) minor optimization would be to replicate the value
4913 // and fall through here to test for a valid 64-bit splat. But, then the
4914 // caller would also need to check and handle the change in size.
4918 if (type != VMOVModImm)
4920 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
4921 uint64_t BitMask = 0xff;
4923 unsigned ImmMask = 1;
4925 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
4926 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
4929 } else if ((SplatBits & BitMask) != 0) {
4936 if (DAG.getDataLayout().isBigEndian())
4937 // swap higher and lower 32 bit word
4938 Imm = ((Imm & 0xf) << 4) | ((Imm & 0xf0) >> 4);
4940 // Op=1, Cmode=1110.
4942 VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
4947 llvm_unreachable("unexpected size for isNEONModifiedImm");
4950 unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm);
4951 return DAG.getTargetConstant(EncodedVal, dl, MVT::i32);
4954 SDValue ARMTargetLowering::LowerConstantFP(SDValue Op, SelectionDAG &DAG,
4955 const ARMSubtarget *ST) const {
4959 bool IsDouble = Op.getValueType() == MVT::f64;
4960 ConstantFPSDNode *CFP = cast<ConstantFPSDNode>(Op);
4962 // Use the default (constant pool) lowering for double constants when we have
4964 if (IsDouble && Subtarget->isFPOnlySP())
4967 // Try splatting with a VMOV.f32...
4968 APFloat FPVal = CFP->getValueAPF();
4969 int ImmVal = IsDouble ? ARM_AM::getFP64Imm(FPVal) : ARM_AM::getFP32Imm(FPVal);
4972 if (IsDouble || !ST->useNEONForSinglePrecisionFP()) {
4973 // We have code in place to select a valid ConstantFP already, no need to
4978 // It's a float and we are trying to use NEON operations where
4979 // possible. Lower it to a splat followed by an extract.
4981 SDValue NewVal = DAG.getTargetConstant(ImmVal, DL, MVT::i32);
4982 SDValue VecConstant = DAG.getNode(ARMISD::VMOVFPIMM, DL, MVT::v2f32,
4984 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecConstant,
4985 DAG.getConstant(0, DL, MVT::i32));
4988 // The rest of our options are NEON only, make sure that's allowed before
4990 if (!ST->hasNEON() || (!IsDouble && !ST->useNEONForSinglePrecisionFP()))
4994 uint64_t iVal = FPVal.bitcastToAPInt().getZExtValue();
4996 // It wouldn't really be worth bothering for doubles except for one very
4997 // important value, which does happen to match: 0.0. So make sure we don't do
4999 if (IsDouble && (iVal & 0xffffffff) != (iVal >> 32))
5002 // Try a VMOV.i32 (FIXME: i8, i16, or i64 could work too).
5003 SDValue NewVal = isNEONModifiedImm(iVal & 0xffffffffU, 0, 32, DAG, SDLoc(Op),
5004 VMovVT, false, VMOVModImm);
5005 if (NewVal != SDValue()) {
5007 SDValue VecConstant = DAG.getNode(ARMISD::VMOVIMM, DL, VMovVT,
5010 return DAG.getNode(ISD::BITCAST, DL, MVT::f64, VecConstant);
5012 // It's a float: cast and extract a vector element.
5013 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
5015 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
5016 DAG.getConstant(0, DL, MVT::i32));
5019 // Finally, try a VMVN.i32
5020 NewVal = isNEONModifiedImm(~iVal & 0xffffffffU, 0, 32, DAG, SDLoc(Op), VMovVT,
5022 if (NewVal != SDValue()) {
5024 SDValue VecConstant = DAG.getNode(ARMISD::VMVNIMM, DL, VMovVT, NewVal);
5027 return DAG.getNode(ISD::BITCAST, DL, MVT::f64, VecConstant);
5029 // It's a float: cast and extract a vector element.
5030 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
5032 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
5033 DAG.getConstant(0, DL, MVT::i32));
5039 // check if an VEXT instruction can handle the shuffle mask when the
5040 // vector sources of the shuffle are the same.
5041 static bool isSingletonVEXTMask(ArrayRef<int> M, EVT VT, unsigned &Imm) {
5042 unsigned NumElts = VT.getVectorNumElements();
5044 // Assume that the first shuffle index is not UNDEF. Fail if it is.
5050 // If this is a VEXT shuffle, the immediate value is the index of the first
5051 // element. The other shuffle indices must be the successive elements after
5053 unsigned ExpectedElt = Imm;
5054 for (unsigned i = 1; i < NumElts; ++i) {
5055 // Increment the expected index. If it wraps around, just follow it
5056 // back to index zero and keep going.
5058 if (ExpectedElt == NumElts)
5061 if (M[i] < 0) continue; // ignore UNDEF indices
5062 if (ExpectedElt != static_cast<unsigned>(M[i]))
5070 static bool isVEXTMask(ArrayRef<int> M, EVT VT,
5071 bool &ReverseVEXT, unsigned &Imm) {
5072 unsigned NumElts = VT.getVectorNumElements();
5073 ReverseVEXT = false;
5075 // Assume that the first shuffle index is not UNDEF. Fail if it is.
5081 // If this is a VEXT shuffle, the immediate value is the index of the first
5082 // element. The other shuffle indices must be the successive elements after
5084 unsigned ExpectedElt = Imm;
5085 for (unsigned i = 1; i < NumElts; ++i) {
5086 // Increment the expected index. If it wraps around, it may still be
5087 // a VEXT but the source vectors must be swapped.
5089 if (ExpectedElt == NumElts * 2) {
5094 if (M[i] < 0) continue; // ignore UNDEF indices
5095 if (ExpectedElt != static_cast<unsigned>(M[i]))
5099 // Adjust the index value if the source operands will be swapped.
5106 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
5107 /// instruction with the specified blocksize. (The order of the elements
5108 /// within each block of the vector is reversed.)
5109 static bool isVREVMask(ArrayRef<int> M, EVT VT, unsigned BlockSize) {
5110 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
5111 "Only possible block sizes for VREV are: 16, 32, 64");
5113 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
5117 unsigned NumElts = VT.getVectorNumElements();
5118 unsigned BlockElts = M[0] + 1;
5119 // If the first shuffle index is UNDEF, be optimistic.
5121 BlockElts = BlockSize / EltSz;
5123 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
5126 for (unsigned i = 0; i < NumElts; ++i) {
5127 if (M[i] < 0) continue; // ignore UNDEF indices
5128 if ((unsigned) M[i] != (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
5135 static bool isVTBLMask(ArrayRef<int> M, EVT VT) {
5136 // We can handle <8 x i8> vector shuffles. If the index in the mask is out of
5137 // range, then 0 is placed into the resulting vector. So pretty much any mask
5138 // of 8 elements can work here.
5139 return VT == MVT::v8i8 && M.size() == 8;
5142 // Checks whether the shuffle mask represents a vector transpose (VTRN) by
5143 // checking that pairs of elements in the shuffle mask represent the same index
5144 // in each vector, incrementing the expected index by 2 at each step.
5145 // e.g. For v1,v2 of type v4i32 a valid shuffle mask is: [0, 4, 2, 6]
5146 // v1={a,b,c,d} => x=shufflevector v1, v2 shufflemask => x={a,e,c,g}
5148 // WhichResult gives the offset for each element in the mask based on which
5149 // of the two results it belongs to.
5151 // The transpose can be represented either as:
5152 // result1 = shufflevector v1, v2, result1_shuffle_mask
5153 // result2 = shufflevector v1, v2, result2_shuffle_mask
5154 // where v1/v2 and the shuffle masks have the same number of elements
5155 // (here WhichResult (see below) indicates which result is being checked)
5158 // results = shufflevector v1, v2, shuffle_mask
5159 // where both results are returned in one vector and the shuffle mask has twice
5160 // as many elements as v1/v2 (here WhichResult will always be 0 if true) here we
5161 // want to check the low half and high half of the shuffle mask as if it were
5163 static bool isVTRNMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
5164 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
5168 unsigned NumElts = VT.getVectorNumElements();
5169 if (M.size() != NumElts && M.size() != NumElts*2)
5172 // If the mask is twice as long as the input vector then we need to check the
5173 // upper and lower parts of the mask with a matching value for WhichResult
5174 // FIXME: A mask with only even values will be rejected in case the first
5175 // element is undefined, e.g. [-1, 4, 2, 6] will be rejected, because only
5176 // M[0] is used to determine WhichResult
5177 for (unsigned i = 0; i < M.size(); i += NumElts) {
5178 if (M.size() == NumElts * 2)
5179 WhichResult = i / NumElts;
5181 WhichResult = M[i] == 0 ? 0 : 1;
5182 for (unsigned j = 0; j < NumElts; j += 2) {
5183 if ((M[i+j] >= 0 && (unsigned) M[i+j] != j + WhichResult) ||
5184 (M[i+j+1] >= 0 && (unsigned) M[i+j+1] != j + NumElts + WhichResult))
5189 if (M.size() == NumElts*2)
5195 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
5196 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
5197 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
5198 static bool isVTRN_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
5199 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
5203 unsigned NumElts = VT.getVectorNumElements();
5204 if (M.size() != NumElts && M.size() != NumElts*2)
5207 for (unsigned i = 0; i < M.size(); i += NumElts) {
5208 if (M.size() == NumElts * 2)
5209 WhichResult = i / NumElts;
5211 WhichResult = M[i] == 0 ? 0 : 1;
5212 for (unsigned j = 0; j < NumElts; j += 2) {
5213 if ((M[i+j] >= 0 && (unsigned) M[i+j] != j + WhichResult) ||
5214 (M[i+j+1] >= 0 && (unsigned) M[i+j+1] != j + WhichResult))
5219 if (M.size() == NumElts*2)
5225 // Checks whether the shuffle mask represents a vector unzip (VUZP) by checking
5226 // that the mask elements are either all even and in steps of size 2 or all odd
5227 // and in steps of size 2.
5228 // e.g. For v1,v2 of type v4i32 a valid shuffle mask is: [0, 2, 4, 6]
5229 // v1={a,b,c,d} => x=shufflevector v1, v2 shufflemask => x={a,c,e,g}
5231 // Requires similar checks to that of isVTRNMask with
5232 // respect the how results are returned.
5233 static bool isVUZPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
5234 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
5238 unsigned NumElts = VT.getVectorNumElements();
5239 if (M.size() != NumElts && M.size() != NumElts*2)
5242 for (unsigned i = 0; i < M.size(); i += NumElts) {
5243 WhichResult = M[i] == 0 ? 0 : 1;
5244 for (unsigned j = 0; j < NumElts; ++j) {
5245 if (M[i+j] >= 0 && (unsigned) M[i+j] != 2 * j + WhichResult)
5250 if (M.size() == NumElts*2)
5253 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
5254 if (VT.is64BitVector() && EltSz == 32)
5260 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
5261 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
5262 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
5263 static bool isVUZP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
5264 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
5268 unsigned NumElts = VT.getVectorNumElements();
5269 if (M.size() != NumElts && M.size() != NumElts*2)
5272 unsigned Half = NumElts / 2;
5273 for (unsigned i = 0; i < M.size(); i += NumElts) {
5274 WhichResult = M[i] == 0 ? 0 : 1;
5275 for (unsigned j = 0; j < NumElts; j += Half) {
5276 unsigned Idx = WhichResult;
5277 for (unsigned k = 0; k < Half; ++k) {
5278 int MIdx = M[i + j + k];
5279 if (MIdx >= 0 && (unsigned) MIdx != Idx)
5286 if (M.size() == NumElts*2)
5289 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
5290 if (VT.is64BitVector() && EltSz == 32)
5296 // Checks whether the shuffle mask represents a vector zip (VZIP) by checking
5297 // that pairs of elements of the shufflemask represent the same index in each
5298 // vector incrementing sequentially through the vectors.
5299 // e.g. For v1,v2 of type v4i32 a valid shuffle mask is: [0, 4, 1, 5]
5300 // v1={a,b,c,d} => x=shufflevector v1, v2 shufflemask => x={a,e,b,f}
5302 // Requires similar checks to that of isVTRNMask with respect the how results
5304 static bool isVZIPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
5305 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
5309 unsigned NumElts = VT.getVectorNumElements();
5310 if (M.size() != NumElts && M.size() != NumElts*2)
5313 for (unsigned i = 0; i < M.size(); i += NumElts) {
5314 WhichResult = M[i] == 0 ? 0 : 1;
5315 unsigned Idx = WhichResult * NumElts / 2;
5316 for (unsigned j = 0; j < NumElts; j += 2) {
5317 if ((M[i+j] >= 0 && (unsigned) M[i+j] != Idx) ||
5318 (M[i+j+1] >= 0 && (unsigned) M[i+j+1] != Idx + NumElts))
5324 if (M.size() == NumElts*2)
5327 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
5328 if (VT.is64BitVector() && EltSz == 32)
5334 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
5335 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
5336 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
5337 static bool isVZIP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
5338 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
5342 unsigned NumElts = VT.getVectorNumElements();
5343 if (M.size() != NumElts && M.size() != NumElts*2)
5346 for (unsigned i = 0; i < M.size(); i += NumElts) {
5347 WhichResult = M[i] == 0 ? 0 : 1;
5348 unsigned Idx = WhichResult * NumElts / 2;
5349 for (unsigned j = 0; j < NumElts; j += 2) {
5350 if ((M[i+j] >= 0 && (unsigned) M[i+j] != Idx) ||
5351 (M[i+j+1] >= 0 && (unsigned) M[i+j+1] != Idx))
5357 if (M.size() == NumElts*2)
5360 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
5361 if (VT.is64BitVector() && EltSz == 32)
5367 /// Check if \p ShuffleMask is a NEON two-result shuffle (VZIP, VUZP, VTRN),
5368 /// and return the corresponding ARMISD opcode if it is, or 0 if it isn't.
5369 static unsigned isNEONTwoResultShuffleMask(ArrayRef<int> ShuffleMask, EVT VT,
5370 unsigned &WhichResult,
5373 if (isVTRNMask(ShuffleMask, VT, WhichResult))
5374 return ARMISD::VTRN;
5375 if (isVUZPMask(ShuffleMask, VT, WhichResult))
5376 return ARMISD::VUZP;
5377 if (isVZIPMask(ShuffleMask, VT, WhichResult))
5378 return ARMISD::VZIP;
5381 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
5382 return ARMISD::VTRN;
5383 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
5384 return ARMISD::VUZP;
5385 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
5386 return ARMISD::VZIP;
5391 /// \return true if this is a reverse operation on an vector.
5392 static bool isReverseMask(ArrayRef<int> M, EVT VT) {
5393 unsigned NumElts = VT.getVectorNumElements();
5394 // Make sure the mask has the right size.
5395 if (NumElts != M.size())
5398 // Look for <15, ..., 3, -1, 1, 0>.
5399 for (unsigned i = 0; i != NumElts; ++i)
5400 if (M[i] >= 0 && M[i] != (int) (NumElts - 1 - i))
5406 // If N is an integer constant that can be moved into a register in one
5407 // instruction, return an SDValue of such a constant (will become a MOV
5408 // instruction). Otherwise return null.
5409 static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG,
5410 const ARMSubtarget *ST, SDLoc dl) {
5412 if (!isa<ConstantSDNode>(N))
5414 Val = cast<ConstantSDNode>(N)->getZExtValue();
5416 if (ST->isThumb1Only()) {
5417 if (Val <= 255 || ~Val <= 255)
5418 return DAG.getConstant(Val, dl, MVT::i32);
5420 if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1)
5421 return DAG.getConstant(Val, dl, MVT::i32);
5426 // If this is a case we can't handle, return null and let the default
5427 // expansion code take care of it.
5428 SDValue ARMTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
5429 const ARMSubtarget *ST) const {
5430 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
5432 EVT VT = Op.getValueType();
5434 APInt SplatBits, SplatUndef;
5435 unsigned SplatBitSize;
5437 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
5438 if (SplatBitSize <= 64) {
5439 // Check if an immediate VMOV works.
5441 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
5442 SplatUndef.getZExtValue(), SplatBitSize,
5443 DAG, dl, VmovVT, VT.is128BitVector(),
5445 if (Val.getNode()) {
5446 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
5447 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
5450 // Try an immediate VMVN.
5451 uint64_t NegatedImm = (~SplatBits).getZExtValue();
5452 Val = isNEONModifiedImm(NegatedImm,
5453 SplatUndef.getZExtValue(), SplatBitSize,
5454 DAG, dl, VmovVT, VT.is128BitVector(),
5456 if (Val.getNode()) {
5457 SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
5458 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
5461 // Use vmov.f32 to materialize other v2f32 and v4f32 splats.
5462 if ((VT == MVT::v2f32 || VT == MVT::v4f32) && SplatBitSize == 32) {
5463 int ImmVal = ARM_AM::getFP32Imm(SplatBits);
5465 SDValue Val = DAG.getTargetConstant(ImmVal, dl, MVT::i32);
5466 return DAG.getNode(ARMISD::VMOVFPIMM, dl, VT, Val);
5472 // Scan through the operands to see if only one value is used.
5474 // As an optimisation, even if more than one value is used it may be more
5475 // profitable to splat with one value then change some lanes.
5477 // Heuristically we decide to do this if the vector has a "dominant" value,
5478 // defined as splatted to more than half of the lanes.
5479 unsigned NumElts = VT.getVectorNumElements();
5480 bool isOnlyLowElement = true;
5481 bool usesOnlyOneValue = true;
5482 bool hasDominantValue = false;
5483 bool isConstant = true;
5485 // Map of the number of times a particular SDValue appears in the
5487 DenseMap<SDValue, unsigned> ValueCounts;
5489 for (unsigned i = 0; i < NumElts; ++i) {
5490 SDValue V = Op.getOperand(i);
5491 if (V.getOpcode() == ISD::UNDEF)
5494 isOnlyLowElement = false;
5495 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
5498 ValueCounts.insert(std::make_pair(V, 0));
5499 unsigned &Count = ValueCounts[V];
5501 // Is this value dominant? (takes up more than half of the lanes)
5502 if (++Count > (NumElts / 2)) {
5503 hasDominantValue = true;
5507 if (ValueCounts.size() != 1)
5508 usesOnlyOneValue = false;
5509 if (!Value.getNode() && ValueCounts.size() > 0)
5510 Value = ValueCounts.begin()->first;
5512 if (ValueCounts.size() == 0)
5513 return DAG.getUNDEF(VT);
5515 // Loads are better lowered with insert_vector_elt/ARMISD::BUILD_VECTOR.
5516 // Keep going if we are hitting this case.
5517 if (isOnlyLowElement && !ISD::isNormalLoad(Value.getNode()))
5518 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
5520 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5522 // Use VDUP for non-constant splats. For f32 constant splats, reduce to
5523 // i32 and try again.
5524 if (hasDominantValue && EltSize <= 32) {
5528 // If we are VDUPing a value that comes directly from a vector, that will
5529 // cause an unnecessary move to and from a GPR, where instead we could
5530 // just use VDUPLANE. We can only do this if the lane being extracted
5531 // is at a constant index, as the VDUP from lane instructions only have
5532 // constant-index forms.
5533 ConstantSDNode *constIndex;
5534 if (Value->getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5535 (constIndex = dyn_cast<ConstantSDNode>(Value->getOperand(1)))) {
5536 // We need to create a new undef vector to use for the VDUPLANE if the
5537 // size of the vector from which we get the value is different than the
5538 // size of the vector that we need to create. We will insert the element
5539 // such that the register coalescer will remove unnecessary copies.
5540 if (VT != Value->getOperand(0).getValueType()) {
5541 unsigned index = constIndex->getAPIntValue().getLimitedValue() %
5542 VT.getVectorNumElements();
5543 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
5544 DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DAG.getUNDEF(VT),
5545 Value, DAG.getConstant(index, dl, MVT::i32)),
5546 DAG.getConstant(index, dl, MVT::i32));
5548 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
5549 Value->getOperand(0), Value->getOperand(1));
5551 N = DAG.getNode(ARMISD::VDUP, dl, VT, Value);
5553 if (!usesOnlyOneValue) {
5554 // The dominant value was splatted as 'N', but we now have to insert
5555 // all differing elements.
5556 for (unsigned I = 0; I < NumElts; ++I) {
5557 if (Op.getOperand(I) == Value)
5559 SmallVector<SDValue, 3> Ops;
5561 Ops.push_back(Op.getOperand(I));
5562 Ops.push_back(DAG.getConstant(I, dl, MVT::i32));
5563 N = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Ops);
5568 if (VT.getVectorElementType().isFloatingPoint()) {
5569 SmallVector<SDValue, 8> Ops;
5570 for (unsigned i = 0; i < NumElts; ++i)
5571 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, MVT::i32,
5573 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
5574 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, Ops);
5575 Val = LowerBUILD_VECTOR(Val, DAG, ST);
5577 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
5579 if (usesOnlyOneValue) {
5580 SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
5581 if (isConstant && Val.getNode())
5582 return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
5586 // If all elements are constants and the case above didn't get hit, fall back
5587 // to the default expansion, which will generate a load from the constant
5592 // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
5594 SDValue shuffle = ReconstructShuffle(Op, DAG);
5595 if (shuffle != SDValue())
5599 // Vectors with 32- or 64-bit elements can be built by directly assigning
5600 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
5601 // will be legalized.
5602 if (EltSize >= 32) {
5603 // Do the expansion with floating-point types, since that is what the VFP
5604 // registers are defined to use, and since i64 is not legal.
5605 EVT EltVT = EVT::getFloatingPointVT(EltSize);
5606 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
5607 SmallVector<SDValue, 8> Ops;
5608 for (unsigned i = 0; i < NumElts; ++i)
5609 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, EltVT, Op.getOperand(i)));
5610 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, Ops);
5611 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
5614 // If all else fails, just use a sequence of INSERT_VECTOR_ELT when we
5615 // know the default expansion would otherwise fall back on something even
5616 // worse. For a vector with one or two non-undef values, that's
5617 // scalar_to_vector for the elements followed by a shuffle (provided the
5618 // shuffle is valid for the target) and materialization element by element
5619 // on the stack followed by a load for everything else.
5620 if (!isConstant && !usesOnlyOneValue) {
5621 SDValue Vec = DAG.getUNDEF(VT);
5622 for (unsigned i = 0 ; i < NumElts; ++i) {
5623 SDValue V = Op.getOperand(i);
5624 if (V.getOpcode() == ISD::UNDEF)
5626 SDValue LaneIdx = DAG.getConstant(i, dl, MVT::i32);
5627 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Vec, V, LaneIdx);
5635 // Gather data to see if the operation can be modelled as a
5636 // shuffle in combination with VEXTs.
5637 SDValue ARMTargetLowering::ReconstructShuffle(SDValue Op,
5638 SelectionDAG &DAG) const {
5639 assert(Op.getOpcode() == ISD::BUILD_VECTOR && "Unknown opcode!");
5641 EVT VT = Op.getValueType();
5642 unsigned NumElts = VT.getVectorNumElements();
5644 struct ShuffleSourceInfo {
5649 // We may insert some combination of BITCASTs and VEXT nodes to force Vec to
5650 // be compatible with the shuffle we intend to construct. As a result
5651 // ShuffleVec will be some sliding window into the original Vec.
5654 // Code should guarantee that element i in Vec starts at element "WindowBase
5655 // + i * WindowScale in ShuffleVec".
5659 bool operator ==(SDValue OtherVec) { return Vec == OtherVec; }
5660 ShuffleSourceInfo(SDValue Vec)
5661 : Vec(Vec), MinElt(UINT_MAX), MaxElt(0), ShuffleVec(Vec), WindowBase(0),
5665 // First gather all vectors used as an immediate source for this BUILD_VECTOR
5667 SmallVector<ShuffleSourceInfo, 2> Sources;
5668 for (unsigned i = 0; i < NumElts; ++i) {
5669 SDValue V = Op.getOperand(i);
5670 if (V.getOpcode() == ISD::UNDEF)
5672 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
5673 // A shuffle can only come from building a vector from various
5674 // elements of other vectors.
5676 } else if (!isa<ConstantSDNode>(V.getOperand(1))) {
5677 // Furthermore, shuffles require a constant mask, whereas extractelts
5678 // accept variable indices.
5682 // Add this element source to the list if it's not already there.
5683 SDValue SourceVec = V.getOperand(0);
5684 auto Source = std::find(Sources.begin(), Sources.end(), SourceVec);
5685 if (Source == Sources.end())
5686 Source = Sources.insert(Sources.end(), ShuffleSourceInfo(SourceVec));
5688 // Update the minimum and maximum lane number seen.
5689 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
5690 Source->MinElt = std::min(Source->MinElt, EltNo);
5691 Source->MaxElt = std::max(Source->MaxElt, EltNo);
5694 // Currently only do something sane when at most two source vectors
5696 if (Sources.size() > 2)
5699 // Find out the smallest element size among result and two sources, and use
5700 // it as element size to build the shuffle_vector.
5701 EVT SmallestEltTy = VT.getVectorElementType();
5702 for (auto &Source : Sources) {
5703 EVT SrcEltTy = Source.Vec.getValueType().getVectorElementType();
5704 if (SrcEltTy.bitsLT(SmallestEltTy))
5705 SmallestEltTy = SrcEltTy;
5707 unsigned ResMultiplier =
5708 VT.getVectorElementType().getSizeInBits() / SmallestEltTy.getSizeInBits();
5709 NumElts = VT.getSizeInBits() / SmallestEltTy.getSizeInBits();
5710 EVT ShuffleVT = EVT::getVectorVT(*DAG.getContext(), SmallestEltTy, NumElts);
5712 // If the source vector is too wide or too narrow, we may nevertheless be able
5713 // to construct a compatible shuffle either by concatenating it with UNDEF or
5714 // extracting a suitable range of elements.
5715 for (auto &Src : Sources) {
5716 EVT SrcVT = Src.ShuffleVec.getValueType();
5718 if (SrcVT.getSizeInBits() == VT.getSizeInBits())
5721 // This stage of the search produces a source with the same element type as
5722 // the original, but with a total width matching the BUILD_VECTOR output.
5723 EVT EltVT = SrcVT.getVectorElementType();
5724 unsigned NumSrcElts = VT.getSizeInBits() / EltVT.getSizeInBits();
5725 EVT DestVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumSrcElts);
5727 if (SrcVT.getSizeInBits() < VT.getSizeInBits()) {
5728 if (2 * SrcVT.getSizeInBits() != VT.getSizeInBits())
5730 // We can pad out the smaller vector for free, so if it's part of a
5733 DAG.getNode(ISD::CONCAT_VECTORS, dl, DestVT, Src.ShuffleVec,
5734 DAG.getUNDEF(Src.ShuffleVec.getValueType()));
5738 if (SrcVT.getSizeInBits() != 2 * VT.getSizeInBits())
5741 if (Src.MaxElt - Src.MinElt >= NumSrcElts) {
5742 // Span too large for a VEXT to cope
5746 if (Src.MinElt >= NumSrcElts) {
5747 // The extraction can just take the second half
5749 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec,
5750 DAG.getConstant(NumSrcElts, dl, MVT::i32));
5751 Src.WindowBase = -NumSrcElts;
5752 } else if (Src.MaxElt < NumSrcElts) {
5753 // The extraction can just take the first half
5755 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec,
5756 DAG.getConstant(0, dl, MVT::i32));
5758 // An actual VEXT is needed
5760 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec,
5761 DAG.getConstant(0, dl, MVT::i32));
5763 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec,
5764 DAG.getConstant(NumSrcElts, dl, MVT::i32));
5766 Src.ShuffleVec = DAG.getNode(ARMISD::VEXT, dl, DestVT, VEXTSrc1,
5768 DAG.getConstant(Src.MinElt, dl, MVT::i32));
5769 Src.WindowBase = -Src.MinElt;
5773 // Another possible incompatibility occurs from the vector element types. We
5774 // can fix this by bitcasting the source vectors to the same type we intend
5776 for (auto &Src : Sources) {
5777 EVT SrcEltTy = Src.ShuffleVec.getValueType().getVectorElementType();
5778 if (SrcEltTy == SmallestEltTy)
5780 assert(ShuffleVT.getVectorElementType() == SmallestEltTy);
5781 Src.ShuffleVec = DAG.getNode(ISD::BITCAST, dl, ShuffleVT, Src.ShuffleVec);
5782 Src.WindowScale = SrcEltTy.getSizeInBits() / SmallestEltTy.getSizeInBits();
5783 Src.WindowBase *= Src.WindowScale;
5786 // Final sanity check before we try to actually produce a shuffle.
5788 for (auto Src : Sources)
5789 assert(Src.ShuffleVec.getValueType() == ShuffleVT);
5792 // The stars all align, our next step is to produce the mask for the shuffle.
5793 SmallVector<int, 8> Mask(ShuffleVT.getVectorNumElements(), -1);
5794 int BitsPerShuffleLane = ShuffleVT.getVectorElementType().getSizeInBits();
5795 for (unsigned i = 0; i < VT.getVectorNumElements(); ++i) {
5796 SDValue Entry = Op.getOperand(i);
5797 if (Entry.getOpcode() == ISD::UNDEF)
5800 auto Src = std::find(Sources.begin(), Sources.end(), Entry.getOperand(0));
5801 int EltNo = cast<ConstantSDNode>(Entry.getOperand(1))->getSExtValue();
5803 // EXTRACT_VECTOR_ELT performs an implicit any_ext; BUILD_VECTOR an implicit
5804 // trunc. So only std::min(SrcBits, DestBits) actually get defined in this
5806 EVT OrigEltTy = Entry.getOperand(0).getValueType().getVectorElementType();
5807 int BitsDefined = std::min(OrigEltTy.getSizeInBits(),
5808 VT.getVectorElementType().getSizeInBits());
5809 int LanesDefined = BitsDefined / BitsPerShuffleLane;
5811 // This source is expected to fill ResMultiplier lanes of the final shuffle,
5812 // starting at the appropriate offset.
5813 int *LaneMask = &Mask[i * ResMultiplier];
5815 int ExtractBase = EltNo * Src->WindowScale + Src->WindowBase;
5816 ExtractBase += NumElts * (Src - Sources.begin());
5817 for (int j = 0; j < LanesDefined; ++j)
5818 LaneMask[j] = ExtractBase + j;
5821 // Final check before we try to produce nonsense...
5822 if (!isShuffleMaskLegal(Mask, ShuffleVT))
5825 // We can't handle more than two sources. This should have already
5826 // been checked before this point.
5827 assert(Sources.size() <= 2 && "Too many sources!");
5829 SDValue ShuffleOps[] = { DAG.getUNDEF(ShuffleVT), DAG.getUNDEF(ShuffleVT) };
5830 for (unsigned i = 0; i < Sources.size(); ++i)
5831 ShuffleOps[i] = Sources[i].ShuffleVec;
5833 SDValue Shuffle = DAG.getVectorShuffle(ShuffleVT, dl, ShuffleOps[0],
5834 ShuffleOps[1], &Mask[0]);
5835 return DAG.getNode(ISD::BITCAST, dl, VT, Shuffle);
5838 /// isShuffleMaskLegal - Targets can use this to indicate that they only
5839 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
5840 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
5841 /// are assumed to be legal.
5843 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
5845 if (VT.getVectorNumElements() == 4 &&
5846 (VT.is128BitVector() || VT.is64BitVector())) {
5847 unsigned PFIndexes[4];
5848 for (unsigned i = 0; i != 4; ++i) {
5852 PFIndexes[i] = M[i];
5855 // Compute the index in the perfect shuffle table.
5856 unsigned PFTableIndex =
5857 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
5858 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
5859 unsigned Cost = (PFEntry >> 30);
5865 bool ReverseVEXT, isV_UNDEF;
5866 unsigned Imm, WhichResult;
5868 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5869 return (EltSize >= 32 ||
5870 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
5871 isVREVMask(M, VT, 64) ||
5872 isVREVMask(M, VT, 32) ||
5873 isVREVMask(M, VT, 16) ||
5874 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
5875 isVTBLMask(M, VT) ||
5876 isNEONTwoResultShuffleMask(M, VT, WhichResult, isV_UNDEF) ||
5877 ((VT == MVT::v8i16 || VT == MVT::v16i8) && isReverseMask(M, VT)));
5880 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
5881 /// the specified operations to build the shuffle.
5882 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
5883 SDValue RHS, SelectionDAG &DAG,
5885 unsigned OpNum = (PFEntry >> 26) & 0x0F;
5886 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
5887 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
5890 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
5899 OP_VUZPL, // VUZP, left result
5900 OP_VUZPR, // VUZP, right result
5901 OP_VZIPL, // VZIP, left result
5902 OP_VZIPR, // VZIP, right result
5903 OP_VTRNL, // VTRN, left result
5904 OP_VTRNR // VTRN, right result
5907 if (OpNum == OP_COPY) {
5908 if (LHSID == (1*9+2)*9+3) return LHS;
5909 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
5913 SDValue OpLHS, OpRHS;
5914 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
5915 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
5916 EVT VT = OpLHS.getValueType();
5919 default: llvm_unreachable("Unknown shuffle opcode!");
5921 // VREV divides the vector in half and swaps within the half.
5922 if (VT.getVectorElementType() == MVT::i32 ||
5923 VT.getVectorElementType() == MVT::f32)
5924 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
5925 // vrev <4 x i16> -> VREV32
5926 if (VT.getVectorElementType() == MVT::i16)
5927 return DAG.getNode(ARMISD::VREV32, dl, VT, OpLHS);
5928 // vrev <4 x i8> -> VREV16
5929 assert(VT.getVectorElementType() == MVT::i8);
5930 return DAG.getNode(ARMISD::VREV16, dl, VT, OpLHS);
5935 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
5936 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, dl, MVT::i32));
5940 return DAG.getNode(ARMISD::VEXT, dl, VT,
5942 DAG.getConstant(OpNum - OP_VEXT1 + 1, dl, MVT::i32));
5945 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
5946 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
5949 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
5950 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
5953 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
5954 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
5958 static SDValue LowerVECTOR_SHUFFLEv8i8(SDValue Op,
5959 ArrayRef<int> ShuffleMask,
5960 SelectionDAG &DAG) {
5961 // Check to see if we can use the VTBL instruction.
5962 SDValue V1 = Op.getOperand(0);
5963 SDValue V2 = Op.getOperand(1);
5966 SmallVector<SDValue, 8> VTBLMask;
5967 for (ArrayRef<int>::iterator
5968 I = ShuffleMask.begin(), E = ShuffleMask.end(); I != E; ++I)
5969 VTBLMask.push_back(DAG.getConstant(*I, DL, MVT::i32));
5971 if (V2.getNode()->getOpcode() == ISD::UNDEF)
5972 return DAG.getNode(ARMISD::VTBL1, DL, MVT::v8i8, V1,
5973 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8, VTBLMask));
5975 return DAG.getNode(ARMISD::VTBL2, DL, MVT::v8i8, V1, V2,
5976 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8, VTBLMask));
5979 static SDValue LowerReverse_VECTOR_SHUFFLEv16i8_v8i16(SDValue Op,
5980 SelectionDAG &DAG) {
5982 SDValue OpLHS = Op.getOperand(0);
5983 EVT VT = OpLHS.getValueType();
5985 assert((VT == MVT::v8i16 || VT == MVT::v16i8) &&
5986 "Expect an v8i16/v16i8 type");
5987 OpLHS = DAG.getNode(ARMISD::VREV64, DL, VT, OpLHS);
5988 // For a v16i8 type: After the VREV, we have got <8, ...15, 8, ..., 0>. Now,
5989 // extract the first 8 bytes into the top double word and the last 8 bytes
5990 // into the bottom double word. The v8i16 case is similar.
5991 unsigned ExtractNum = (VT == MVT::v16i8) ? 8 : 4;
5992 return DAG.getNode(ARMISD::VEXT, DL, VT, OpLHS, OpLHS,
5993 DAG.getConstant(ExtractNum, DL, MVT::i32));
5996 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
5997 SDValue V1 = Op.getOperand(0);
5998 SDValue V2 = Op.getOperand(1);
6000 EVT VT = Op.getValueType();
6001 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
6003 // Convert shuffles that are directly supported on NEON to target-specific
6004 // DAG nodes, instead of keeping them as shuffles and matching them again
6005 // during code selection. This is more efficient and avoids the possibility
6006 // of inconsistencies between legalization and selection.
6007 // FIXME: floating-point vectors should be canonicalized to integer vectors
6008 // of the same time so that they get CSEd properly.
6009 ArrayRef<int> ShuffleMask = SVN->getMask();
6011 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
6012 if (EltSize <= 32) {
6013 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
6014 int Lane = SVN->getSplatIndex();
6015 // If this is undef splat, generate it via "just" vdup, if possible.
6016 if (Lane == -1) Lane = 0;
6018 // Test if V1 is a SCALAR_TO_VECTOR.
6019 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
6020 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
6022 // Test if V1 is a BUILD_VECTOR which is equivalent to a SCALAR_TO_VECTOR
6023 // (and probably will turn into a SCALAR_TO_VECTOR once legalization
6025 if (Lane == 0 && V1.getOpcode() == ISD::BUILD_VECTOR &&
6026 !isa<ConstantSDNode>(V1.getOperand(0))) {
6027 bool IsScalarToVector = true;
6028 for (unsigned i = 1, e = V1.getNumOperands(); i != e; ++i)
6029 if (V1.getOperand(i).getOpcode() != ISD::UNDEF) {
6030 IsScalarToVector = false;
6033 if (IsScalarToVector)
6034 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
6036 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
6037 DAG.getConstant(Lane, dl, MVT::i32));
6042 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
6045 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
6046 DAG.getConstant(Imm, dl, MVT::i32));
6049 if (isVREVMask(ShuffleMask, VT, 64))
6050 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
6051 if (isVREVMask(ShuffleMask, VT, 32))
6052 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
6053 if (isVREVMask(ShuffleMask, VT, 16))
6054 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
6056 if (V2->getOpcode() == ISD::UNDEF &&
6057 isSingletonVEXTMask(ShuffleMask, VT, Imm)) {
6058 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V1,
6059 DAG.getConstant(Imm, dl, MVT::i32));
6062 // Check for Neon shuffles that modify both input vectors in place.
6063 // If both results are used, i.e., if there are two shuffles with the same
6064 // source operands and with masks corresponding to both results of one of
6065 // these operations, DAG memoization will ensure that a single node is
6066 // used for both shuffles.
6067 unsigned WhichResult;
6069 if (unsigned ShuffleOpc = isNEONTwoResultShuffleMask(
6070 ShuffleMask, VT, WhichResult, isV_UNDEF)) {
6073 return DAG.getNode(ShuffleOpc, dl, DAG.getVTList(VT, VT), V1, V2)
6074 .getValue(WhichResult);
6077 // Also check for these shuffles through CONCAT_VECTORS: we canonicalize
6078 // shuffles that produce a result larger than their operands with:
6079 // shuffle(concat(v1, undef), concat(v2, undef))
6081 // shuffle(concat(v1, v2), undef)
6082 // because we can access quad vectors (see PerformVECTOR_SHUFFLECombine).
6084 // This is useful in the general case, but there are special cases where
6085 // native shuffles produce larger results: the two-result ops.
6087 // Look through the concat when lowering them:
6088 // shuffle(concat(v1, v2), undef)
6090 // concat(VZIP(v1, v2):0, :1)
6092 if (V1->getOpcode() == ISD::CONCAT_VECTORS &&
6093 V2->getOpcode() == ISD::UNDEF) {
6094 SDValue SubV1 = V1->getOperand(0);
6095 SDValue SubV2 = V1->getOperand(1);
6096 EVT SubVT = SubV1.getValueType();
6098 // We expect these to have been canonicalized to -1.
6099 assert(std::all_of(ShuffleMask.begin(), ShuffleMask.end(), [&](int i) {
6100 return i < (int)VT.getVectorNumElements();
6101 }) && "Unexpected shuffle index into UNDEF operand!");
6103 if (unsigned ShuffleOpc = isNEONTwoResultShuffleMask(
6104 ShuffleMask, SubVT, WhichResult, isV_UNDEF)) {
6107 assert((WhichResult == 0) &&
6108 "In-place shuffle of concat can only have one result!");
6109 SDValue Res = DAG.getNode(ShuffleOpc, dl, DAG.getVTList(SubVT, SubVT),
6111 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, Res.getValue(0),
6117 // If the shuffle is not directly supported and it has 4 elements, use
6118 // the PerfectShuffle-generated table to synthesize it from other shuffles.
6119 unsigned NumElts = VT.getVectorNumElements();
6121 unsigned PFIndexes[4];
6122 for (unsigned i = 0; i != 4; ++i) {
6123 if (ShuffleMask[i] < 0)
6126 PFIndexes[i] = ShuffleMask[i];
6129 // Compute the index in the perfect shuffle table.
6130 unsigned PFTableIndex =
6131 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
6132 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
6133 unsigned Cost = (PFEntry >> 30);
6136 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
6139 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
6140 if (EltSize >= 32) {
6141 // Do the expansion with floating-point types, since that is what the VFP
6142 // registers are defined to use, and since i64 is not legal.
6143 EVT EltVT = EVT::getFloatingPointVT(EltSize);
6144 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
6145 V1 = DAG.getNode(ISD::BITCAST, dl, VecVT, V1);
6146 V2 = DAG.getNode(ISD::BITCAST, dl, VecVT, V2);
6147 SmallVector<SDValue, 8> Ops;
6148 for (unsigned i = 0; i < NumElts; ++i) {
6149 if (ShuffleMask[i] < 0)
6150 Ops.push_back(DAG.getUNDEF(EltVT));
6152 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
6153 ShuffleMask[i] < (int)NumElts ? V1 : V2,
6154 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
6157 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, Ops);
6158 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
6161 if ((VT == MVT::v8i16 || VT == MVT::v16i8) && isReverseMask(ShuffleMask, VT))
6162 return LowerReverse_VECTOR_SHUFFLEv16i8_v8i16(Op, DAG);
6164 if (VT == MVT::v8i8) {
6165 SDValue NewOp = LowerVECTOR_SHUFFLEv8i8(Op, ShuffleMask, DAG);
6166 if (NewOp.getNode())
6173 static SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
6174 // INSERT_VECTOR_ELT is legal only for immediate indexes.
6175 SDValue Lane = Op.getOperand(2);
6176 if (!isa<ConstantSDNode>(Lane))
6182 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
6183 // EXTRACT_VECTOR_ELT is legal only for immediate indexes.
6184 SDValue Lane = Op.getOperand(1);
6185 if (!isa<ConstantSDNode>(Lane))
6188 SDValue Vec = Op.getOperand(0);
6189 if (Op.getValueType() == MVT::i32 &&
6190 Vec.getValueType().getVectorElementType().getSizeInBits() < 32) {
6192 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
6198 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
6199 // The only time a CONCAT_VECTORS operation can have legal types is when
6200 // two 64-bit vectors are concatenated to a 128-bit vector.
6201 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
6202 "unexpected CONCAT_VECTORS");
6204 SDValue Val = DAG.getUNDEF(MVT::v2f64);
6205 SDValue Op0 = Op.getOperand(0);
6206 SDValue Op1 = Op.getOperand(1);
6207 if (Op0.getOpcode() != ISD::UNDEF)
6208 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
6209 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op0),
6210 DAG.getIntPtrConstant(0, dl));
6211 if (Op1.getOpcode() != ISD::UNDEF)
6212 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
6213 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op1),
6214 DAG.getIntPtrConstant(1, dl));
6215 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Val);
6218 /// isExtendedBUILD_VECTOR - Check if N is a constant BUILD_VECTOR where each
6219 /// element has been zero/sign-extended, depending on the isSigned parameter,
6220 /// from an integer type half its size.
6221 static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG,
6223 // A v2i64 BUILD_VECTOR will have been legalized to a BITCAST from v4i32.
6224 EVT VT = N->getValueType(0);
6225 if (VT == MVT::v2i64 && N->getOpcode() == ISD::BITCAST) {
6226 SDNode *BVN = N->getOperand(0).getNode();
6227 if (BVN->getValueType(0) != MVT::v4i32 ||
6228 BVN->getOpcode() != ISD::BUILD_VECTOR)
6230 unsigned LoElt = DAG.getDataLayout().isBigEndian() ? 1 : 0;
6231 unsigned HiElt = 1 - LoElt;
6232 ConstantSDNode *Lo0 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt));
6233 ConstantSDNode *Hi0 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt));
6234 ConstantSDNode *Lo1 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt+2));
6235 ConstantSDNode *Hi1 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt+2));
6236 if (!Lo0 || !Hi0 || !Lo1 || !Hi1)
6239 if (Hi0->getSExtValue() == Lo0->getSExtValue() >> 32 &&
6240 Hi1->getSExtValue() == Lo1->getSExtValue() >> 32)
6243 if (Hi0->isNullValue() && Hi1->isNullValue())
6249 if (N->getOpcode() != ISD::BUILD_VECTOR)
6252 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
6253 SDNode *Elt = N->getOperand(i).getNode();
6254 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
6255 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
6256 unsigned HalfSize = EltSize / 2;
6258 if (!isIntN(HalfSize, C->getSExtValue()))
6261 if (!isUIntN(HalfSize, C->getZExtValue()))
6272 /// isSignExtended - Check if a node is a vector value that is sign-extended
6273 /// or a constant BUILD_VECTOR with sign-extended elements.
6274 static bool isSignExtended(SDNode *N, SelectionDAG &DAG) {
6275 if (N->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N))
6277 if (isExtendedBUILD_VECTOR(N, DAG, true))
6282 /// isZeroExtended - Check if a node is a vector value that is zero-extended
6283 /// or a constant BUILD_VECTOR with zero-extended elements.
6284 static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) {
6285 if (N->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N))
6287 if (isExtendedBUILD_VECTOR(N, DAG, false))
6292 static EVT getExtensionTo64Bits(const EVT &OrigVT) {
6293 if (OrigVT.getSizeInBits() >= 64)
6296 assert(OrigVT.isSimple() && "Expecting a simple value type");
6298 MVT::SimpleValueType OrigSimpleTy = OrigVT.getSimpleVT().SimpleTy;
6299 switch (OrigSimpleTy) {
6300 default: llvm_unreachable("Unexpected Vector Type");
6309 /// AddRequiredExtensionForVMULL - Add a sign/zero extension to extend the total
6310 /// value size to 64 bits. We need a 64-bit D register as an operand to VMULL.
6311 /// We insert the required extension here to get the vector to fill a D register.
6312 static SDValue AddRequiredExtensionForVMULL(SDValue N, SelectionDAG &DAG,
6315 unsigned ExtOpcode) {
6316 // The vector originally had a size of OrigTy. It was then extended to ExtTy.
6317 // We expect the ExtTy to be 128-bits total. If the OrigTy is less than
6318 // 64-bits we need to insert a new extension so that it will be 64-bits.
6319 assert(ExtTy.is128BitVector() && "Unexpected extension size");
6320 if (OrigTy.getSizeInBits() >= 64)
6323 // Must extend size to at least 64 bits to be used as an operand for VMULL.
6324 EVT NewVT = getExtensionTo64Bits(OrigTy);
6326 return DAG.getNode(ExtOpcode, SDLoc(N), NewVT, N);
6329 /// SkipLoadExtensionForVMULL - return a load of the original vector size that
6330 /// does not do any sign/zero extension. If the original vector is less
6331 /// than 64 bits, an appropriate extension will be added after the load to
6332 /// reach a total size of 64 bits. We have to add the extension separately
6333 /// because ARM does not have a sign/zero extending load for vectors.
6334 static SDValue SkipLoadExtensionForVMULL(LoadSDNode *LD, SelectionDAG& DAG) {
6335 EVT ExtendedTy = getExtensionTo64Bits(LD->getMemoryVT());
6337 // The load already has the right type.
6338 if (ExtendedTy == LD->getMemoryVT())
6339 return DAG.getLoad(LD->getMemoryVT(), SDLoc(LD), LD->getChain(),
6340 LD->getBasePtr(), LD->getPointerInfo(), LD->isVolatile(),
6341 LD->isNonTemporal(), LD->isInvariant(),
6342 LD->getAlignment());
6344 // We need to create a zextload/sextload. We cannot just create a load
6345 // followed by a zext/zext node because LowerMUL is also run during normal
6346 // operation legalization where we can't create illegal types.
6347 return DAG.getExtLoad(LD->getExtensionType(), SDLoc(LD), ExtendedTy,
6348 LD->getChain(), LD->getBasePtr(), LD->getPointerInfo(),
6349 LD->getMemoryVT(), LD->isVolatile(), LD->isInvariant(),
6350 LD->isNonTemporal(), LD->getAlignment());
6353 /// SkipExtensionForVMULL - For a node that is a SIGN_EXTEND, ZERO_EXTEND,
6354 /// extending load, or BUILD_VECTOR with extended elements, return the
6355 /// unextended value. The unextended vector should be 64 bits so that it can
6356 /// be used as an operand to a VMULL instruction. If the original vector size
6357 /// before extension is less than 64 bits we add a an extension to resize
6358 /// the vector to 64 bits.
6359 static SDValue SkipExtensionForVMULL(SDNode *N, SelectionDAG &DAG) {
6360 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
6361 return AddRequiredExtensionForVMULL(N->getOperand(0), DAG,
6362 N->getOperand(0)->getValueType(0),
6366 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
6367 return SkipLoadExtensionForVMULL(LD, DAG);
6369 // Otherwise, the value must be a BUILD_VECTOR. For v2i64, it will
6370 // have been legalized as a BITCAST from v4i32.
6371 if (N->getOpcode() == ISD::BITCAST) {
6372 SDNode *BVN = N->getOperand(0).getNode();
6373 assert(BVN->getOpcode() == ISD::BUILD_VECTOR &&
6374 BVN->getValueType(0) == MVT::v4i32 && "expected v4i32 BUILD_VECTOR");
6375 unsigned LowElt = DAG.getDataLayout().isBigEndian() ? 1 : 0;
6376 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), MVT::v2i32,
6377 BVN->getOperand(LowElt), BVN->getOperand(LowElt+2));
6379 // Construct a new BUILD_VECTOR with elements truncated to half the size.
6380 assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR");
6381 EVT VT = N->getValueType(0);
6382 unsigned EltSize = VT.getVectorElementType().getSizeInBits() / 2;
6383 unsigned NumElts = VT.getVectorNumElements();
6384 MVT TruncVT = MVT::getIntegerVT(EltSize);
6385 SmallVector<SDValue, 8> Ops;
6387 for (unsigned i = 0; i != NumElts; ++i) {
6388 ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i));
6389 const APInt &CInt = C->getAPIntValue();
6390 // Element types smaller than 32 bits are not legal, so use i32 elements.
6391 // The values are implicitly truncated so sext vs. zext doesn't matter.
6392 Ops.push_back(DAG.getConstant(CInt.zextOrTrunc(32), dl, MVT::i32));
6394 return DAG.getNode(ISD::BUILD_VECTOR, dl,
6395 MVT::getVectorVT(TruncVT, NumElts), Ops);
6398 static bool isAddSubSExt(SDNode *N, SelectionDAG &DAG) {
6399 unsigned Opcode = N->getOpcode();
6400 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
6401 SDNode *N0 = N->getOperand(0).getNode();
6402 SDNode *N1 = N->getOperand(1).getNode();
6403 return N0->hasOneUse() && N1->hasOneUse() &&
6404 isSignExtended(N0, DAG) && isSignExtended(N1, DAG);
6409 static bool isAddSubZExt(SDNode *N, SelectionDAG &DAG) {
6410 unsigned Opcode = N->getOpcode();
6411 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
6412 SDNode *N0 = N->getOperand(0).getNode();
6413 SDNode *N1 = N->getOperand(1).getNode();
6414 return N0->hasOneUse() && N1->hasOneUse() &&
6415 isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG);
6420 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
6421 // Multiplications are only custom-lowered for 128-bit vectors so that
6422 // VMULL can be detected. Otherwise v2i64 multiplications are not legal.
6423 EVT VT = Op.getValueType();
6424 assert(VT.is128BitVector() && VT.isInteger() &&
6425 "unexpected type for custom-lowering ISD::MUL");
6426 SDNode *N0 = Op.getOperand(0).getNode();
6427 SDNode *N1 = Op.getOperand(1).getNode();
6428 unsigned NewOpc = 0;
6430 bool isN0SExt = isSignExtended(N0, DAG);
6431 bool isN1SExt = isSignExtended(N1, DAG);
6432 if (isN0SExt && isN1SExt)
6433 NewOpc = ARMISD::VMULLs;
6435 bool isN0ZExt = isZeroExtended(N0, DAG);
6436 bool isN1ZExt = isZeroExtended(N1, DAG);
6437 if (isN0ZExt && isN1ZExt)
6438 NewOpc = ARMISD::VMULLu;
6439 else if (isN1SExt || isN1ZExt) {
6440 // Look for (s/zext A + s/zext B) * (s/zext C). We want to turn these
6441 // into (s/zext A * s/zext C) + (s/zext B * s/zext C)
6442 if (isN1SExt && isAddSubSExt(N0, DAG)) {
6443 NewOpc = ARMISD::VMULLs;
6445 } else if (isN1ZExt && isAddSubZExt(N0, DAG)) {
6446 NewOpc = ARMISD::VMULLu;
6448 } else if (isN0ZExt && isAddSubZExt(N1, DAG)) {
6450 NewOpc = ARMISD::VMULLu;
6456 if (VT == MVT::v2i64)
6457 // Fall through to expand this. It is not legal.
6460 // Other vector multiplications are legal.
6465 // Legalize to a VMULL instruction.
6468 SDValue Op1 = SkipExtensionForVMULL(N1, DAG);
6470 Op0 = SkipExtensionForVMULL(N0, DAG);
6471 assert(Op0.getValueType().is64BitVector() &&
6472 Op1.getValueType().is64BitVector() &&
6473 "unexpected types for extended operands to VMULL");
6474 return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
6477 // Optimizing (zext A + zext B) * C, to (VMULL A, C) + (VMULL B, C) during
6478 // isel lowering to take advantage of no-stall back to back vmul + vmla.
6485 SDValue N00 = SkipExtensionForVMULL(N0->getOperand(0).getNode(), DAG);
6486 SDValue N01 = SkipExtensionForVMULL(N0->getOperand(1).getNode(), DAG);
6487 EVT Op1VT = Op1.getValueType();
6488 return DAG.getNode(N0->getOpcode(), DL, VT,
6489 DAG.getNode(NewOpc, DL, VT,
6490 DAG.getNode(ISD::BITCAST, DL, Op1VT, N00), Op1),
6491 DAG.getNode(NewOpc, DL, VT,
6492 DAG.getNode(ISD::BITCAST, DL, Op1VT, N01), Op1));
6496 LowerSDIV_v4i8(SDValue X, SDValue Y, SDLoc dl, SelectionDAG &DAG) {
6497 // TODO: Should this propagate fast-math-flags?
6500 // float4 xf = vcvt_f32_s32(vmovl_s16(a.lo));
6501 // float4 yf = vcvt_f32_s32(vmovl_s16(b.lo));
6502 X = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, X);
6503 Y = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, Y);
6504 X = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, X);
6505 Y = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, Y);
6506 // Get reciprocal estimate.
6507 // float4 recip = vrecpeq_f32(yf);
6508 Y = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6509 DAG.getConstant(Intrinsic::arm_neon_vrecpe, dl, MVT::i32),
6511 // Because char has a smaller range than uchar, we can actually get away
6512 // without any newton steps. This requires that we use a weird bias
6513 // of 0xb000, however (again, this has been exhaustively tested).
6514 // float4 result = as_float4(as_int4(xf*recip) + 0xb000);
6515 X = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, X, Y);
6516 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, X);
6517 Y = DAG.getConstant(0xb000, dl, MVT::i32);
6518 Y = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Y, Y, Y, Y);
6519 X = DAG.getNode(ISD::ADD, dl, MVT::v4i32, X, Y);
6520 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, X);
6521 // Convert back to short.
6522 X = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, X);
6523 X = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, X);
6528 LowerSDIV_v4i16(SDValue N0, SDValue N1, SDLoc dl, SelectionDAG &DAG) {
6529 // TODO: Should this propagate fast-math-flags?
6532 // Convert to float.
6533 // float4 yf = vcvt_f32_s32(vmovl_s16(y));
6534 // float4 xf = vcvt_f32_s32(vmovl_s16(x));
6535 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N0);
6536 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N1);
6537 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
6538 N1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
6540 // Use reciprocal estimate and one refinement step.
6541 // float4 recip = vrecpeq_f32(yf);
6542 // recip *= vrecpsq_f32(yf, recip);
6543 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6544 DAG.getConstant(Intrinsic::arm_neon_vrecpe, dl, MVT::i32),
6546 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6547 DAG.getConstant(Intrinsic::arm_neon_vrecps, dl, MVT::i32),
6549 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
6550 // Because short has a smaller range than ushort, we can actually get away
6551 // with only a single newton step. This requires that we use a weird bias
6552 // of 89, however (again, this has been exhaustively tested).
6553 // float4 result = as_float4(as_int4(xf*recip) + 0x89);
6554 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
6555 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
6556 N1 = DAG.getConstant(0x89, dl, MVT::i32);
6557 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
6558 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
6559 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
6560 // Convert back to integer and return.
6561 // return vmovn_s32(vcvt_s32_f32(result));
6562 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
6563 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
6567 static SDValue LowerSDIV(SDValue Op, SelectionDAG &DAG) {
6568 EVT VT = Op.getValueType();
6569 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
6570 "unexpected type for custom-lowering ISD::SDIV");
6573 SDValue N0 = Op.getOperand(0);
6574 SDValue N1 = Op.getOperand(1);
6577 if (VT == MVT::v8i8) {
6578 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N0);
6579 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N1);
6581 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
6582 DAG.getIntPtrConstant(4, dl));
6583 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
6584 DAG.getIntPtrConstant(4, dl));
6585 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
6586 DAG.getIntPtrConstant(0, dl));
6587 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
6588 DAG.getIntPtrConstant(0, dl));
6590 N0 = LowerSDIV_v4i8(N0, N1, dl, DAG); // v4i16
6591 N2 = LowerSDIV_v4i8(N2, N3, dl, DAG); // v4i16
6593 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
6594 N0 = LowerCONCAT_VECTORS(N0, DAG);
6596 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v8i8, N0);
6599 return LowerSDIV_v4i16(N0, N1, dl, DAG);
6602 static SDValue LowerUDIV(SDValue Op, SelectionDAG &DAG) {
6603 // TODO: Should this propagate fast-math-flags?
6604 EVT VT = Op.getValueType();
6605 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
6606 "unexpected type for custom-lowering ISD::UDIV");
6609 SDValue N0 = Op.getOperand(0);
6610 SDValue N1 = Op.getOperand(1);
6613 if (VT == MVT::v8i8) {
6614 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N0);
6615 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N1);
6617 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
6618 DAG.getIntPtrConstant(4, dl));
6619 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
6620 DAG.getIntPtrConstant(4, dl));
6621 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
6622 DAG.getIntPtrConstant(0, dl));
6623 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
6624 DAG.getIntPtrConstant(0, dl));
6626 N0 = LowerSDIV_v4i16(N0, N1, dl, DAG); // v4i16
6627 N2 = LowerSDIV_v4i16(N2, N3, dl, DAG); // v4i16
6629 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
6630 N0 = LowerCONCAT_VECTORS(N0, DAG);
6632 N0 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v8i8,
6633 DAG.getConstant(Intrinsic::arm_neon_vqmovnsu, dl,
6639 // v4i16 sdiv ... Convert to float.
6640 // float4 yf = vcvt_f32_s32(vmovl_u16(y));
6641 // float4 xf = vcvt_f32_s32(vmovl_u16(x));
6642 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N0);
6643 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N1);
6644 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
6645 SDValue BN1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
6647 // Use reciprocal estimate and two refinement steps.
6648 // float4 recip = vrecpeq_f32(yf);
6649 // recip *= vrecpsq_f32(yf, recip);
6650 // recip *= vrecpsq_f32(yf, recip);
6651 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6652 DAG.getConstant(Intrinsic::arm_neon_vrecpe, dl, MVT::i32),
6654 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6655 DAG.getConstant(Intrinsic::arm_neon_vrecps, dl, MVT::i32),
6657 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
6658 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6659 DAG.getConstant(Intrinsic::arm_neon_vrecps, dl, MVT::i32),
6661 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
6662 // Simply multiplying by the reciprocal estimate can leave us a few ulps
6663 // too low, so we add 2 ulps (exhaustive testing shows that this is enough,
6664 // and that it will never cause us to return an answer too large).
6665 // float4 result = as_float4(as_int4(xf*recip) + 2);
6666 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
6667 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
6668 N1 = DAG.getConstant(2, dl, MVT::i32);
6669 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
6670 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
6671 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
6672 // Convert back to integer and return.
6673 // return vmovn_u32(vcvt_s32_f32(result));
6674 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
6675 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
6679 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
6680 EVT VT = Op.getNode()->getValueType(0);
6681 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
6684 bool ExtraOp = false;
6685 switch (Op.getOpcode()) {
6686 default: llvm_unreachable("Invalid code");
6687 case ISD::ADDC: Opc = ARMISD::ADDC; break;
6688 case ISD::ADDE: Opc = ARMISD::ADDE; ExtraOp = true; break;
6689 case ISD::SUBC: Opc = ARMISD::SUBC; break;
6690 case ISD::SUBE: Opc = ARMISD::SUBE; ExtraOp = true; break;
6694 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
6696 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
6697 Op.getOperand(1), Op.getOperand(2));
6700 SDValue ARMTargetLowering::LowerFSINCOS(SDValue Op, SelectionDAG &DAG) const {
6701 assert(Subtarget->isTargetDarwin());
6703 // For iOS, we want to call an alternative entry point: __sincos_stret,
6704 // return values are passed via sret.
6706 SDValue Arg = Op.getOperand(0);
6707 EVT ArgVT = Arg.getValueType();
6708 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
6709 auto PtrVT = getPointerTy(DAG.getDataLayout());
6711 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
6712 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6714 // Pair of floats / doubles used to pass the result.
6715 Type *RetTy = StructType::get(ArgTy, ArgTy, nullptr);
6716 auto &DL = DAG.getDataLayout();
6719 bool ShouldUseSRet = Subtarget->isAPCS_ABI();
6721 if (ShouldUseSRet) {
6722 // Create stack object for sret.
6723 const uint64_t ByteSize = DL.getTypeAllocSize(RetTy);
6724 const unsigned StackAlign = DL.getPrefTypeAlignment(RetTy);
6725 int FrameIdx = FrameInfo->CreateStackObject(ByteSize, StackAlign, false);
6726 SRet = DAG.getFrameIndex(FrameIdx, TLI.getPointerTy(DL));
6730 Entry.Ty = RetTy->getPointerTo();
6731 Entry.isSExt = false;
6732 Entry.isZExt = false;
6733 Entry.isSRet = true;
6734 Args.push_back(Entry);
6735 RetTy = Type::getVoidTy(*DAG.getContext());
6741 Entry.isSExt = false;
6742 Entry.isZExt = false;
6743 Args.push_back(Entry);
6745 const char *LibcallName =
6746 (ArgVT == MVT::f64) ? "__sincos_stret" : "__sincosf_stret";
6748 (ArgVT == MVT::f64) ? RTLIB::SINCOS_F64 : RTLIB::SINCOS_F32;
6749 CallingConv::ID CC = getLibcallCallingConv(LC);
6750 SDValue Callee = DAG.getExternalSymbol(LibcallName, getPointerTy(DL));
6752 TargetLowering::CallLoweringInfo CLI(DAG);
6754 .setChain(DAG.getEntryNode())
6755 .setCallee(CC, RetTy, Callee, std::move(Args), 0)
6756 .setDiscardResult(ShouldUseSRet);
6757 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
6760 return CallResult.first;
6762 SDValue LoadSin = DAG.getLoad(ArgVT, dl, CallResult.second, SRet,
6763 MachinePointerInfo(), false, false, false, 0);
6765 // Address of cos field.
6766 SDValue Add = DAG.getNode(ISD::ADD, dl, PtrVT, SRet,
6767 DAG.getIntPtrConstant(ArgVT.getStoreSize(), dl));
6768 SDValue LoadCos = DAG.getLoad(ArgVT, dl, LoadSin.getValue(1), Add,
6769 MachinePointerInfo(), false, false, false, 0);
6771 SDVTList Tys = DAG.getVTList(ArgVT, ArgVT);
6772 return DAG.getNode(ISD::MERGE_VALUES, dl, Tys,
6773 LoadSin.getValue(0), LoadCos.getValue(0));
6776 SDValue ARMTargetLowering::LowerWindowsDIVLibCall(SDValue Op, SelectionDAG &DAG,
6778 SDValue &Chain) const {
6779 EVT VT = Op.getValueType();
6780 assert((VT == MVT::i32 || VT == MVT::i64) &&
6781 "unexpected type for custom lowering DIV");
6784 const auto &DL = DAG.getDataLayout();
6785 const auto &TLI = DAG.getTargetLoweringInfo();
6787 const char *Name = nullptr;
6789 Name = (VT == MVT::i32) ? "__rt_sdiv" : "__rt_sdiv64";
6791 Name = (VT == MVT::i32) ? "__rt_udiv" : "__rt_udiv64";
6793 SDValue ES = DAG.getExternalSymbol(Name, TLI.getPointerTy(DL));
6795 ARMTargetLowering::ArgListTy Args;
6797 for (auto AI : {1, 0}) {
6799 Arg.Node = Op.getOperand(AI);
6800 Arg.Ty = Arg.Node.getValueType().getTypeForEVT(*DAG.getContext());
6801 Args.push_back(Arg);
6804 CallLoweringInfo CLI(DAG);
6807 .setCallee(CallingConv::ARM_AAPCS_VFP, VT.getTypeForEVT(*DAG.getContext()),
6808 ES, std::move(Args), 0);
6810 return LowerCallTo(CLI).first;
6813 SDValue ARMTargetLowering::LowerDIV_Windows(SDValue Op, SelectionDAG &DAG,
6814 bool Signed) const {
6815 assert(Op.getValueType() == MVT::i32 &&
6816 "unexpected type for custom lowering DIV");
6819 SDValue DBZCHK = DAG.getNode(ARMISD::WIN__DBZCHK, dl, MVT::Other,
6820 DAG.getEntryNode(), Op.getOperand(1));
6822 return LowerWindowsDIVLibCall(Op, DAG, Signed, DBZCHK);
6825 void ARMTargetLowering::ExpandDIV_Windows(
6826 SDValue Op, SelectionDAG &DAG, bool Signed,
6827 SmallVectorImpl<SDValue> &Results) const {
6828 const auto &DL = DAG.getDataLayout();
6829 const auto &TLI = DAG.getTargetLoweringInfo();
6831 assert(Op.getValueType() == MVT::i64 &&
6832 "unexpected type for custom lowering DIV");
6835 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op.getOperand(1),
6836 DAG.getConstant(0, dl, MVT::i32));
6837 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op.getOperand(1),
6838 DAG.getConstant(1, dl, MVT::i32));
6839 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::i32, Lo, Hi);
6842 DAG.getNode(ARMISD::WIN__DBZCHK, dl, MVT::Other, DAG.getEntryNode(), Or);
6844 SDValue Result = LowerWindowsDIVLibCall(Op, DAG, Signed, DBZCHK);
6846 SDValue Lower = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Result);
6847 SDValue Upper = DAG.getNode(ISD::SRL, dl, MVT::i64, Result,
6848 DAG.getConstant(32, dl, TLI.getPointerTy(DL)));
6849 Upper = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Upper);
6851 Results.push_back(Lower);
6852 Results.push_back(Upper);
6855 static SDValue LowerAtomicLoadStore(SDValue Op, SelectionDAG &DAG) {
6856 // Monotonic load/store is legal for all targets
6857 if (cast<AtomicSDNode>(Op)->getOrdering() <= Monotonic)
6860 // Acquire/Release load/store is not legal for targets without a
6861 // dmb or equivalent available.
6865 static void ReplaceREADCYCLECOUNTER(SDNode *N,
6866 SmallVectorImpl<SDValue> &Results,
6868 const ARMSubtarget *Subtarget) {
6870 // Under Power Management extensions, the cycle-count is:
6871 // mrc p15, #0, <Rt>, c9, c13, #0
6872 SDValue Ops[] = { N->getOperand(0), // Chain
6873 DAG.getConstant(Intrinsic::arm_mrc, DL, MVT::i32),
6874 DAG.getConstant(15, DL, MVT::i32),
6875 DAG.getConstant(0, DL, MVT::i32),
6876 DAG.getConstant(9, DL, MVT::i32),
6877 DAG.getConstant(13, DL, MVT::i32),
6878 DAG.getConstant(0, DL, MVT::i32)
6881 SDValue Cycles32 = DAG.getNode(ISD::INTRINSIC_W_CHAIN, DL,
6882 DAG.getVTList(MVT::i32, MVT::Other), Ops);
6883 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Cycles32,
6884 DAG.getConstant(0, DL, MVT::i32)));
6885 Results.push_back(Cycles32.getValue(1));
6888 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
6889 switch (Op.getOpcode()) {
6890 default: llvm_unreachable("Don't know how to custom lower this!");
6891 case ISD::WRITE_REGISTER: return LowerWRITE_REGISTER(Op, DAG);
6892 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
6893 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
6894 case ISD::GlobalAddress:
6895 switch (Subtarget->getTargetTriple().getObjectFormat()) {
6896 default: llvm_unreachable("unknown object format");
6898 return LowerGlobalAddressWindows(Op, DAG);
6900 return LowerGlobalAddressELF(Op, DAG);
6902 return LowerGlobalAddressDarwin(Op, DAG);
6904 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
6905 case ISD::SELECT: return LowerSELECT(Op, DAG);
6906 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
6907 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
6908 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
6909 case ISD::VASTART: return LowerVASTART(Op, DAG);
6910 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, DAG, Subtarget);
6911 case ISD::PREFETCH: return LowerPREFETCH(Op, DAG, Subtarget);
6912 case ISD::SINT_TO_FP:
6913 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
6914 case ISD::FP_TO_SINT:
6915 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
6916 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
6917 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
6918 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
6919 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
6920 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
6921 case ISD::EH_SJLJ_SETUP_DISPATCH: return LowerEH_SJLJ_SETUP_DISPATCH(Op, DAG);
6922 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
6924 case ISD::BITCAST: return ExpandBITCAST(Op.getNode(), DAG);
6927 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
6928 case ISD::SREM: return LowerREM(Op.getNode(), DAG);
6929 case ISD::UREM: return LowerREM(Op.getNode(), DAG);
6930 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
6931 case ISD::SRL_PARTS:
6932 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
6934 case ISD::CTTZ_ZERO_UNDEF: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
6935 case ISD::CTPOP: return LowerCTPOP(Op.getNode(), DAG, Subtarget);
6936 case ISD::SETCC: return LowerVSETCC(Op, DAG);
6937 case ISD::ConstantFP: return LowerConstantFP(Op, DAG, Subtarget);
6938 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG, Subtarget);
6939 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
6940 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
6941 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
6942 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
6943 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
6944 case ISD::MUL: return LowerMUL(Op, DAG);
6945 case ISD::SDIV: return LowerSDIV(Op, DAG);
6946 case ISD::UDIV: return LowerUDIV(Op, DAG);
6950 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
6955 return LowerXALUO(Op, DAG);
6956 case ISD::ATOMIC_LOAD:
6957 case ISD::ATOMIC_STORE: return LowerAtomicLoadStore(Op, DAG);
6958 case ISD::FSINCOS: return LowerFSINCOS(Op, DAG);
6960 case ISD::UDIVREM: return LowerDivRem(Op, DAG);
6961 case ISD::DYNAMIC_STACKALLOC:
6962 if (Subtarget->getTargetTriple().isWindowsItaniumEnvironment())
6963 return LowerDYNAMIC_STACKALLOC(Op, DAG);
6964 llvm_unreachable("Don't know how to custom lower this!");
6965 case ISD::FP_ROUND: return LowerFP_ROUND(Op, DAG);
6966 case ISD::FP_EXTEND: return LowerFP_EXTEND(Op, DAG);
6967 case ARMISD::WIN__DBZCHK: return SDValue();
6971 /// ReplaceNodeResults - Replace the results of node with an illegal result
6972 /// type with new values built out of custom code.
6973 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
6974 SmallVectorImpl<SDValue> &Results,
6975 SelectionDAG &DAG) const {
6977 switch (N->getOpcode()) {
6979 llvm_unreachable("Don't know how to custom expand this!");
6980 case ISD::READ_REGISTER:
6981 ExpandREAD_REGISTER(N, Results, DAG);
6984 Res = ExpandBITCAST(N, DAG);
6988 Res = Expand64BitShift(N, DAG, Subtarget);
6992 Res = LowerREM(N, DAG);
6994 case ISD::READCYCLECOUNTER:
6995 ReplaceREADCYCLECOUNTER(N, Results, DAG, Subtarget);
6999 assert(Subtarget->isTargetWindows() && "can only expand DIV on Windows");
7000 return ExpandDIV_Windows(SDValue(N, 0), DAG, N->getOpcode() == ISD::SDIV,
7004 Results.push_back(Res);
7007 //===----------------------------------------------------------------------===//
7008 // ARM Scheduler Hooks
7009 //===----------------------------------------------------------------------===//
7011 /// SetupEntryBlockForSjLj - Insert code into the entry block that creates and
7012 /// registers the function context.
7013 void ARMTargetLowering::
7014 SetupEntryBlockForSjLj(MachineInstr *MI, MachineBasicBlock *MBB,
7015 MachineBasicBlock *DispatchBB, int FI) const {
7016 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
7017 DebugLoc dl = MI->getDebugLoc();
7018 MachineFunction *MF = MBB->getParent();
7019 MachineRegisterInfo *MRI = &MF->getRegInfo();
7020 MachineConstantPool *MCP = MF->getConstantPool();
7021 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
7022 const Function *F = MF->getFunction();
7024 bool isThumb = Subtarget->isThumb();
7025 bool isThumb2 = Subtarget->isThumb2();
7027 unsigned PCLabelId = AFI->createPICLabelUId();
7028 unsigned PCAdj = (isThumb || isThumb2) ? 4 : 8;
7029 ARMConstantPoolValue *CPV =
7030 ARMConstantPoolMBB::Create(F->getContext(), DispatchBB, PCLabelId, PCAdj);
7031 unsigned CPI = MCP->getConstantPoolIndex(CPV, 4);
7033 const TargetRegisterClass *TRC = isThumb ? &ARM::tGPRRegClass
7034 : &ARM::GPRRegClass;
7036 // Grab constant pool and fixed stack memory operands.
7037 MachineMemOperand *CPMMO =
7038 MF->getMachineMemOperand(MachinePointerInfo::getConstantPool(*MF),
7039 MachineMemOperand::MOLoad, 4, 4);
7041 MachineMemOperand *FIMMOSt =
7042 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(*MF, FI),
7043 MachineMemOperand::MOStore, 4, 4);
7045 // Load the address of the dispatch MBB into the jump buffer.
7047 // Incoming value: jbuf
7048 // ldr.n r5, LCPI1_1
7051 // str r5, [$jbuf, #+4] ; &jbuf[1]
7052 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
7053 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2LDRpci), NewVReg1)
7054 .addConstantPoolIndex(CPI)
7055 .addMemOperand(CPMMO));
7056 // Set the low bit because of thumb mode.
7057 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
7059 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2ORRri), NewVReg2)
7060 .addReg(NewVReg1, RegState::Kill)
7062 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
7063 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg3)
7064 .addReg(NewVReg2, RegState::Kill)
7066 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2STRi12))
7067 .addReg(NewVReg3, RegState::Kill)
7069 .addImm(36) // &jbuf[1] :: pc
7070 .addMemOperand(FIMMOSt));
7071 } else if (isThumb) {
7072 // Incoming value: jbuf
7073 // ldr.n r1, LCPI1_4
7077 // add r2, $jbuf, #+4 ; &jbuf[1]
7079 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
7080 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tLDRpci), NewVReg1)
7081 .addConstantPoolIndex(CPI)
7082 .addMemOperand(CPMMO));
7083 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
7084 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg2)
7085 .addReg(NewVReg1, RegState::Kill)
7087 // Set the low bit because of thumb mode.
7088 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
7089 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tMOVi8), NewVReg3)
7090 .addReg(ARM::CPSR, RegState::Define)
7092 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
7093 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tORR), NewVReg4)
7094 .addReg(ARM::CPSR, RegState::Define)
7095 .addReg(NewVReg2, RegState::Kill)
7096 .addReg(NewVReg3, RegState::Kill));
7097 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
7098 BuildMI(*MBB, MI, dl, TII->get(ARM::tADDframe), NewVReg5)
7100 .addImm(36); // &jbuf[1] :: pc
7101 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tSTRi))
7102 .addReg(NewVReg4, RegState::Kill)
7103 .addReg(NewVReg5, RegState::Kill)
7105 .addMemOperand(FIMMOSt));
7107 // Incoming value: jbuf
7110 // str r1, [$jbuf, #+4] ; &jbuf[1]
7111 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
7112 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::LDRi12), NewVReg1)
7113 .addConstantPoolIndex(CPI)
7115 .addMemOperand(CPMMO));
7116 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
7117 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::PICADD), NewVReg2)
7118 .addReg(NewVReg1, RegState::Kill)
7119 .addImm(PCLabelId));
7120 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::STRi12))
7121 .addReg(NewVReg2, RegState::Kill)
7123 .addImm(36) // &jbuf[1] :: pc
7124 .addMemOperand(FIMMOSt));
7128 void ARMTargetLowering::EmitSjLjDispatchBlock(MachineInstr *MI,
7129 MachineBasicBlock *MBB) const {
7130 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
7131 DebugLoc dl = MI->getDebugLoc();
7132 MachineFunction *MF = MBB->getParent();
7133 MachineRegisterInfo *MRI = &MF->getRegInfo();
7134 MachineFrameInfo *MFI = MF->getFrameInfo();
7135 int FI = MFI->getFunctionContextIndex();
7137 const TargetRegisterClass *TRC = Subtarget->isThumb() ? &ARM::tGPRRegClass
7138 : &ARM::GPRnopcRegClass;
7140 // Get a mapping of the call site numbers to all of the landing pads they're
7142 DenseMap<unsigned, SmallVector<MachineBasicBlock*, 2> > CallSiteNumToLPad;
7143 unsigned MaxCSNum = 0;
7144 MachineModuleInfo &MMI = MF->getMMI();
7145 for (MachineFunction::iterator BB = MF->begin(), E = MF->end(); BB != E;
7147 if (!BB->isEHPad()) continue;
7149 // FIXME: We should assert that the EH_LABEL is the first MI in the landing
7151 for (MachineBasicBlock::iterator
7152 II = BB->begin(), IE = BB->end(); II != IE; ++II) {
7153 if (!II->isEHLabel()) continue;
7155 MCSymbol *Sym = II->getOperand(0).getMCSymbol();
7156 if (!MMI.hasCallSiteLandingPad(Sym)) continue;
7158 SmallVectorImpl<unsigned> &CallSiteIdxs = MMI.getCallSiteLandingPad(Sym);
7159 for (SmallVectorImpl<unsigned>::iterator
7160 CSI = CallSiteIdxs.begin(), CSE = CallSiteIdxs.end();
7161 CSI != CSE; ++CSI) {
7162 CallSiteNumToLPad[*CSI].push_back(&*BB);
7163 MaxCSNum = std::max(MaxCSNum, *CSI);
7169 // Get an ordered list of the machine basic blocks for the jump table.
7170 std::vector<MachineBasicBlock*> LPadList;
7171 SmallPtrSet<MachineBasicBlock*, 64> InvokeBBs;
7172 LPadList.reserve(CallSiteNumToLPad.size());
7173 for (unsigned I = 1; I <= MaxCSNum; ++I) {
7174 SmallVectorImpl<MachineBasicBlock*> &MBBList = CallSiteNumToLPad[I];
7175 for (SmallVectorImpl<MachineBasicBlock*>::iterator
7176 II = MBBList.begin(), IE = MBBList.end(); II != IE; ++II) {
7177 LPadList.push_back(*II);
7178 InvokeBBs.insert((*II)->pred_begin(), (*II)->pred_end());
7182 assert(!LPadList.empty() &&
7183 "No landing pad destinations for the dispatch jump table!");
7185 // Create the jump table and associated information.
7186 MachineJumpTableInfo *JTI =
7187 MF->getOrCreateJumpTableInfo(MachineJumpTableInfo::EK_Inline);
7188 unsigned MJTI = JTI->createJumpTableIndex(LPadList);
7189 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
7191 // Create the MBBs for the dispatch code.
7193 // Shove the dispatch's address into the return slot in the function context.
7194 MachineBasicBlock *DispatchBB = MF->CreateMachineBasicBlock();
7195 DispatchBB->setIsEHPad();
7197 MachineBasicBlock *TrapBB = MF->CreateMachineBasicBlock();
7198 unsigned trap_opcode;
7199 if (Subtarget->isThumb())
7200 trap_opcode = ARM::tTRAP;
7202 trap_opcode = Subtarget->useNaClTrap() ? ARM::TRAPNaCl : ARM::TRAP;
7204 BuildMI(TrapBB, dl, TII->get(trap_opcode));
7205 DispatchBB->addSuccessor(TrapBB);
7207 MachineBasicBlock *DispContBB = MF->CreateMachineBasicBlock();
7208 DispatchBB->addSuccessor(DispContBB);
7211 MF->insert(MF->end(), DispatchBB);
7212 MF->insert(MF->end(), DispContBB);
7213 MF->insert(MF->end(), TrapBB);
7215 // Insert code into the entry block that creates and registers the function
7217 SetupEntryBlockForSjLj(MI, MBB, DispatchBB, FI);
7219 MachineMemOperand *FIMMOLd = MF->getMachineMemOperand(
7220 MachinePointerInfo::getFixedStack(*MF, FI),
7221 MachineMemOperand::MOLoad | MachineMemOperand::MOVolatile, 4, 4);
7223 MachineInstrBuilder MIB;
7224 MIB = BuildMI(DispatchBB, dl, TII->get(ARM::Int_eh_sjlj_dispatchsetup));
7226 const ARMBaseInstrInfo *AII = static_cast<const ARMBaseInstrInfo*>(TII);
7227 const ARMBaseRegisterInfo &RI = AII->getRegisterInfo();
7229 // Add a register mask with no preserved registers. This results in all
7230 // registers being marked as clobbered.
7231 MIB.addRegMask(RI.getNoPreservedMask());
7233 unsigned NumLPads = LPadList.size();
7234 if (Subtarget->isThumb2()) {
7235 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
7236 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2LDRi12), NewVReg1)
7239 .addMemOperand(FIMMOLd));
7241 if (NumLPads < 256) {
7242 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPri))
7244 .addImm(LPadList.size()));
7246 unsigned VReg1 = MRI->createVirtualRegister(TRC);
7247 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVi16), VReg1)
7248 .addImm(NumLPads & 0xFFFF));
7250 unsigned VReg2 = VReg1;
7251 if ((NumLPads & 0xFFFF0000) != 0) {
7252 VReg2 = MRI->createVirtualRegister(TRC);
7253 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVTi16), VReg2)
7255 .addImm(NumLPads >> 16));
7258 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPrr))
7263 BuildMI(DispatchBB, dl, TII->get(ARM::t2Bcc))
7268 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
7269 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::t2LEApcrelJT),NewVReg3)
7270 .addJumpTableIndex(MJTI));
7272 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
7275 BuildMI(DispContBB, dl, TII->get(ARM::t2ADDrs), NewVReg4)
7276 .addReg(NewVReg3, RegState::Kill)
7278 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
7280 BuildMI(DispContBB, dl, TII->get(ARM::t2BR_JT))
7281 .addReg(NewVReg4, RegState::Kill)
7283 .addJumpTableIndex(MJTI);
7284 } else if (Subtarget->isThumb()) {
7285 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
7286 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRspi), NewVReg1)
7289 .addMemOperand(FIMMOLd));
7291 if (NumLPads < 256) {
7292 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPi8))
7296 MachineConstantPool *ConstantPool = MF->getConstantPool();
7297 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
7298 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
7300 // MachineConstantPool wants an explicit alignment.
7301 unsigned Align = MF->getDataLayout().getPrefTypeAlignment(Int32Ty);
7303 Align = MF->getDataLayout().getTypeAllocSize(C->getType());
7304 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
7306 unsigned VReg1 = MRI->createVirtualRegister(TRC);
7307 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRpci))
7308 .addReg(VReg1, RegState::Define)
7309 .addConstantPoolIndex(Idx));
7310 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPr))
7315 BuildMI(DispatchBB, dl, TII->get(ARM::tBcc))
7320 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
7321 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLSLri), NewVReg2)
7322 .addReg(ARM::CPSR, RegState::Define)
7326 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
7327 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLEApcrelJT), NewVReg3)
7328 .addJumpTableIndex(MJTI));
7330 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
7331 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg4)
7332 .addReg(ARM::CPSR, RegState::Define)
7333 .addReg(NewVReg2, RegState::Kill)
7336 MachineMemOperand *JTMMOLd = MF->getMachineMemOperand(
7337 MachinePointerInfo::getJumpTable(*MF), MachineMemOperand::MOLoad, 4, 4);
7339 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
7340 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLDRi), NewVReg5)
7341 .addReg(NewVReg4, RegState::Kill)
7343 .addMemOperand(JTMMOLd));
7345 unsigned NewVReg6 = NewVReg5;
7346 if (RelocM == Reloc::PIC_) {
7347 NewVReg6 = MRI->createVirtualRegister(TRC);
7348 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg6)
7349 .addReg(ARM::CPSR, RegState::Define)
7350 .addReg(NewVReg5, RegState::Kill)
7354 BuildMI(DispContBB, dl, TII->get(ARM::tBR_JTr))
7355 .addReg(NewVReg6, RegState::Kill)
7356 .addJumpTableIndex(MJTI);
7358 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
7359 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRi12), NewVReg1)
7362 .addMemOperand(FIMMOLd));
7364 if (NumLPads < 256) {
7365 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPri))
7368 } else if (Subtarget->hasV6T2Ops() && isUInt<16>(NumLPads)) {
7369 unsigned VReg1 = MRI->createVirtualRegister(TRC);
7370 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVi16), VReg1)
7371 .addImm(NumLPads & 0xFFFF));
7373 unsigned VReg2 = VReg1;
7374 if ((NumLPads & 0xFFFF0000) != 0) {
7375 VReg2 = MRI->createVirtualRegister(TRC);
7376 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVTi16), VReg2)
7378 .addImm(NumLPads >> 16));
7381 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
7385 MachineConstantPool *ConstantPool = MF->getConstantPool();
7386 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
7387 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
7389 // MachineConstantPool wants an explicit alignment.
7390 unsigned Align = MF->getDataLayout().getPrefTypeAlignment(Int32Ty);
7392 Align = MF->getDataLayout().getTypeAllocSize(C->getType());
7393 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
7395 unsigned VReg1 = MRI->createVirtualRegister(TRC);
7396 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRcp))
7397 .addReg(VReg1, RegState::Define)
7398 .addConstantPoolIndex(Idx)
7400 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
7402 .addReg(VReg1, RegState::Kill));
7405 BuildMI(DispatchBB, dl, TII->get(ARM::Bcc))
7410 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
7412 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::MOVsi), NewVReg3)
7414 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
7415 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
7416 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::LEApcrelJT), NewVReg4)
7417 .addJumpTableIndex(MJTI));
7419 MachineMemOperand *JTMMOLd = MF->getMachineMemOperand(
7420 MachinePointerInfo::getJumpTable(*MF), MachineMemOperand::MOLoad, 4, 4);
7421 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
7423 BuildMI(DispContBB, dl, TII->get(ARM::LDRrs), NewVReg5)
7424 .addReg(NewVReg3, RegState::Kill)
7427 .addMemOperand(JTMMOLd));
7429 if (RelocM == Reloc::PIC_) {
7430 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTadd))
7431 .addReg(NewVReg5, RegState::Kill)
7433 .addJumpTableIndex(MJTI);
7435 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTr))
7436 .addReg(NewVReg5, RegState::Kill)
7437 .addJumpTableIndex(MJTI);
7441 // Add the jump table entries as successors to the MBB.
7442 SmallPtrSet<MachineBasicBlock*, 8> SeenMBBs;
7443 for (std::vector<MachineBasicBlock*>::iterator
7444 I = LPadList.begin(), E = LPadList.end(); I != E; ++I) {
7445 MachineBasicBlock *CurMBB = *I;
7446 if (SeenMBBs.insert(CurMBB).second)
7447 DispContBB->addSuccessor(CurMBB);
7450 // N.B. the order the invoke BBs are processed in doesn't matter here.
7451 const MCPhysReg *SavedRegs = RI.getCalleeSavedRegs(MF);
7452 SmallVector<MachineBasicBlock*, 64> MBBLPads;
7453 for (MachineBasicBlock *BB : InvokeBBs) {
7455 // Remove the landing pad successor from the invoke block and replace it
7456 // with the new dispatch block.
7457 SmallVector<MachineBasicBlock*, 4> Successors(BB->succ_begin(),
7459 while (!Successors.empty()) {
7460 MachineBasicBlock *SMBB = Successors.pop_back_val();
7461 if (SMBB->isEHPad()) {
7462 BB->removeSuccessor(SMBB);
7463 MBBLPads.push_back(SMBB);
7467 BB->addSuccessor(DispatchBB, BranchProbability::getZero());
7468 BB->normalizeSuccProbs();
7470 // Find the invoke call and mark all of the callee-saved registers as
7471 // 'implicit defined' so that they're spilled. This prevents code from
7472 // moving instructions to before the EH block, where they will never be
7474 for (MachineBasicBlock::reverse_iterator
7475 II = BB->rbegin(), IE = BB->rend(); II != IE; ++II) {
7476 if (!II->isCall()) continue;
7478 DenseMap<unsigned, bool> DefRegs;
7479 for (MachineInstr::mop_iterator
7480 OI = II->operands_begin(), OE = II->operands_end();
7482 if (!OI->isReg()) continue;
7483 DefRegs[OI->getReg()] = true;
7486 MachineInstrBuilder MIB(*MF, &*II);
7488 for (unsigned i = 0; SavedRegs[i] != 0; ++i) {
7489 unsigned Reg = SavedRegs[i];
7490 if (Subtarget->isThumb2() &&
7491 !ARM::tGPRRegClass.contains(Reg) &&
7492 !ARM::hGPRRegClass.contains(Reg))
7494 if (Subtarget->isThumb1Only() && !ARM::tGPRRegClass.contains(Reg))
7496 if (!Subtarget->isThumb() && !ARM::GPRRegClass.contains(Reg))
7499 MIB.addReg(Reg, RegState::ImplicitDefine | RegState::Dead);
7506 // Mark all former landing pads as non-landing pads. The dispatch is the only
7508 for (SmallVectorImpl<MachineBasicBlock*>::iterator
7509 I = MBBLPads.begin(), E = MBBLPads.end(); I != E; ++I)
7510 (*I)->setIsEHPad(false);
7512 // The instruction is gone now.
7513 MI->eraseFromParent();
7517 MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) {
7518 for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
7519 E = MBB->succ_end(); I != E; ++I)
7522 llvm_unreachable("Expecting a BB with two successors!");
7525 /// Return the load opcode for a given load size. If load size >= 8,
7526 /// neon opcode will be returned.
7527 static unsigned getLdOpcode(unsigned LdSize, bool IsThumb1, bool IsThumb2) {
7529 return LdSize == 16 ? ARM::VLD1q32wb_fixed
7530 : LdSize == 8 ? ARM::VLD1d32wb_fixed : 0;
7532 return LdSize == 4 ? ARM::tLDRi
7533 : LdSize == 2 ? ARM::tLDRHi
7534 : LdSize == 1 ? ARM::tLDRBi : 0;
7536 return LdSize == 4 ? ARM::t2LDR_POST
7537 : LdSize == 2 ? ARM::t2LDRH_POST
7538 : LdSize == 1 ? ARM::t2LDRB_POST : 0;
7539 return LdSize == 4 ? ARM::LDR_POST_IMM
7540 : LdSize == 2 ? ARM::LDRH_POST
7541 : LdSize == 1 ? ARM::LDRB_POST_IMM : 0;
7544 /// Return the store opcode for a given store size. If store size >= 8,
7545 /// neon opcode will be returned.
7546 static unsigned getStOpcode(unsigned StSize, bool IsThumb1, bool IsThumb2) {
7548 return StSize == 16 ? ARM::VST1q32wb_fixed
7549 : StSize == 8 ? ARM::VST1d32wb_fixed : 0;
7551 return StSize == 4 ? ARM::tSTRi
7552 : StSize == 2 ? ARM::tSTRHi
7553 : StSize == 1 ? ARM::tSTRBi : 0;
7555 return StSize == 4 ? ARM::t2STR_POST
7556 : StSize == 2 ? ARM::t2STRH_POST
7557 : StSize == 1 ? ARM::t2STRB_POST : 0;
7558 return StSize == 4 ? ARM::STR_POST_IMM
7559 : StSize == 2 ? ARM::STRH_POST
7560 : StSize == 1 ? ARM::STRB_POST_IMM : 0;
7563 /// Emit a post-increment load operation with given size. The instructions
7564 /// will be added to BB at Pos.
7565 static void emitPostLd(MachineBasicBlock *BB, MachineInstr *Pos,
7566 const TargetInstrInfo *TII, DebugLoc dl,
7567 unsigned LdSize, unsigned Data, unsigned AddrIn,
7568 unsigned AddrOut, bool IsThumb1, bool IsThumb2) {
7569 unsigned LdOpc = getLdOpcode(LdSize, IsThumb1, IsThumb2);
7570 assert(LdOpc != 0 && "Should have a load opcode");
7572 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
7573 .addReg(AddrOut, RegState::Define).addReg(AddrIn)
7575 } else if (IsThumb1) {
7576 // load + update AddrIn
7577 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
7578 .addReg(AddrIn).addImm(0));
7579 MachineInstrBuilder MIB =
7580 BuildMI(*BB, Pos, dl, TII->get(ARM::tADDi8), AddrOut);
7581 MIB = AddDefaultT1CC(MIB);
7582 MIB.addReg(AddrIn).addImm(LdSize);
7583 AddDefaultPred(MIB);
7584 } else if (IsThumb2) {
7585 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
7586 .addReg(AddrOut, RegState::Define).addReg(AddrIn)
7589 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
7590 .addReg(AddrOut, RegState::Define).addReg(AddrIn)
7591 .addReg(0).addImm(LdSize));
7595 /// Emit a post-increment store operation with given size. The instructions
7596 /// will be added to BB at Pos.
7597 static void emitPostSt(MachineBasicBlock *BB, MachineInstr *Pos,
7598 const TargetInstrInfo *TII, DebugLoc dl,
7599 unsigned StSize, unsigned Data, unsigned AddrIn,
7600 unsigned AddrOut, bool IsThumb1, bool IsThumb2) {
7601 unsigned StOpc = getStOpcode(StSize, IsThumb1, IsThumb2);
7602 assert(StOpc != 0 && "Should have a store opcode");
7604 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc), AddrOut)
7605 .addReg(AddrIn).addImm(0).addReg(Data));
7606 } else if (IsThumb1) {
7607 // store + update AddrIn
7608 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc)).addReg(Data)
7609 .addReg(AddrIn).addImm(0));
7610 MachineInstrBuilder MIB =
7611 BuildMI(*BB, Pos, dl, TII->get(ARM::tADDi8), AddrOut);
7612 MIB = AddDefaultT1CC(MIB);
7613 MIB.addReg(AddrIn).addImm(StSize);
7614 AddDefaultPred(MIB);
7615 } else if (IsThumb2) {
7616 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc), AddrOut)
7617 .addReg(Data).addReg(AddrIn).addImm(StSize));
7619 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc), AddrOut)
7620 .addReg(Data).addReg(AddrIn).addReg(0)
7626 ARMTargetLowering::EmitStructByval(MachineInstr *MI,
7627 MachineBasicBlock *BB) const {
7628 // This pseudo instruction has 3 operands: dst, src, size
7629 // We expand it to a loop if size > Subtarget->getMaxInlineSizeThreshold().
7630 // Otherwise, we will generate unrolled scalar copies.
7631 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
7632 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7633 MachineFunction::iterator It = ++BB->getIterator();
7635 unsigned dest = MI->getOperand(0).getReg();
7636 unsigned src = MI->getOperand(1).getReg();
7637 unsigned SizeVal = MI->getOperand(2).getImm();
7638 unsigned Align = MI->getOperand(3).getImm();
7639 DebugLoc dl = MI->getDebugLoc();
7641 MachineFunction *MF = BB->getParent();
7642 MachineRegisterInfo &MRI = MF->getRegInfo();
7643 unsigned UnitSize = 0;
7644 const TargetRegisterClass *TRC = nullptr;
7645 const TargetRegisterClass *VecTRC = nullptr;
7647 bool IsThumb1 = Subtarget->isThumb1Only();
7648 bool IsThumb2 = Subtarget->isThumb2();
7652 } else if (Align & 2) {
7655 // Check whether we can use NEON instructions.
7656 if (!MF->getFunction()->hasFnAttribute(Attribute::NoImplicitFloat) &&
7657 Subtarget->hasNEON()) {
7658 if ((Align % 16 == 0) && SizeVal >= 16)
7660 else if ((Align % 8 == 0) && SizeVal >= 8)
7663 // Can't use NEON instructions.
7668 // Select the correct opcode and register class for unit size load/store
7669 bool IsNeon = UnitSize >= 8;
7670 TRC = (IsThumb1 || IsThumb2) ? &ARM::tGPRRegClass : &ARM::GPRRegClass;
7672 VecTRC = UnitSize == 16 ? &ARM::DPairRegClass
7673 : UnitSize == 8 ? &ARM::DPRRegClass
7676 unsigned BytesLeft = SizeVal % UnitSize;
7677 unsigned LoopSize = SizeVal - BytesLeft;
7679 if (SizeVal <= Subtarget->getMaxInlineSizeThreshold()) {
7680 // Use LDR and STR to copy.
7681 // [scratch, srcOut] = LDR_POST(srcIn, UnitSize)
7682 // [destOut] = STR_POST(scratch, destIn, UnitSize)
7683 unsigned srcIn = src;
7684 unsigned destIn = dest;
7685 for (unsigned i = 0; i < LoopSize; i+=UnitSize) {
7686 unsigned srcOut = MRI.createVirtualRegister(TRC);
7687 unsigned destOut = MRI.createVirtualRegister(TRC);
7688 unsigned scratch = MRI.createVirtualRegister(IsNeon ? VecTRC : TRC);
7689 emitPostLd(BB, MI, TII, dl, UnitSize, scratch, srcIn, srcOut,
7690 IsThumb1, IsThumb2);
7691 emitPostSt(BB, MI, TII, dl, UnitSize, scratch, destIn, destOut,
7692 IsThumb1, IsThumb2);
7697 // Handle the leftover bytes with LDRB and STRB.
7698 // [scratch, srcOut] = LDRB_POST(srcIn, 1)
7699 // [destOut] = STRB_POST(scratch, destIn, 1)
7700 for (unsigned i = 0; i < BytesLeft; i++) {
7701 unsigned srcOut = MRI.createVirtualRegister(TRC);
7702 unsigned destOut = MRI.createVirtualRegister(TRC);
7703 unsigned scratch = MRI.createVirtualRegister(TRC);
7704 emitPostLd(BB, MI, TII, dl, 1, scratch, srcIn, srcOut,
7705 IsThumb1, IsThumb2);
7706 emitPostSt(BB, MI, TII, dl, 1, scratch, destIn, destOut,
7707 IsThumb1, IsThumb2);
7711 MI->eraseFromParent(); // The instruction is gone now.
7715 // Expand the pseudo op to a loop.
7718 // movw varEnd, # --> with thumb2
7720 // ldrcp varEnd, idx --> without thumb2
7721 // fallthrough --> loopMBB
7723 // PHI varPhi, varEnd, varLoop
7724 // PHI srcPhi, src, srcLoop
7725 // PHI destPhi, dst, destLoop
7726 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
7727 // [destLoop] = STR_POST(scratch, destPhi, UnitSize)
7728 // subs varLoop, varPhi, #UnitSize
7730 // fallthrough --> exitMBB
7732 // epilogue to handle left-over bytes
7733 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
7734 // [destOut] = STRB_POST(scratch, destLoop, 1)
7735 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
7736 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
7737 MF->insert(It, loopMBB);
7738 MF->insert(It, exitMBB);
7740 // Transfer the remainder of BB and its successor edges to exitMBB.
7741 exitMBB->splice(exitMBB->begin(), BB,
7742 std::next(MachineBasicBlock::iterator(MI)), BB->end());
7743 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
7745 // Load an immediate to varEnd.
7746 unsigned varEnd = MRI.createVirtualRegister(TRC);
7747 if (Subtarget->useMovt(*MF)) {
7748 unsigned Vtmp = varEnd;
7749 if ((LoopSize & 0xFFFF0000) != 0)
7750 Vtmp = MRI.createVirtualRegister(TRC);
7751 AddDefaultPred(BuildMI(BB, dl,
7752 TII->get(IsThumb2 ? ARM::t2MOVi16 : ARM::MOVi16),
7753 Vtmp).addImm(LoopSize & 0xFFFF));
7755 if ((LoopSize & 0xFFFF0000) != 0)
7756 AddDefaultPred(BuildMI(BB, dl,
7757 TII->get(IsThumb2 ? ARM::t2MOVTi16 : ARM::MOVTi16),
7760 .addImm(LoopSize >> 16));
7762 MachineConstantPool *ConstantPool = MF->getConstantPool();
7763 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
7764 const Constant *C = ConstantInt::get(Int32Ty, LoopSize);
7766 // MachineConstantPool wants an explicit alignment.
7767 unsigned Align = MF->getDataLayout().getPrefTypeAlignment(Int32Ty);
7769 Align = MF->getDataLayout().getTypeAllocSize(C->getType());
7770 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
7773 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(ARM::tLDRpci)).addReg(
7774 varEnd, RegState::Define).addConstantPoolIndex(Idx));
7776 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(ARM::LDRcp)).addReg(
7777 varEnd, RegState::Define).addConstantPoolIndex(Idx).addImm(0));
7779 BB->addSuccessor(loopMBB);
7781 // Generate the loop body:
7782 // varPhi = PHI(varLoop, varEnd)
7783 // srcPhi = PHI(srcLoop, src)
7784 // destPhi = PHI(destLoop, dst)
7785 MachineBasicBlock *entryBB = BB;
7787 unsigned varLoop = MRI.createVirtualRegister(TRC);
7788 unsigned varPhi = MRI.createVirtualRegister(TRC);
7789 unsigned srcLoop = MRI.createVirtualRegister(TRC);
7790 unsigned srcPhi = MRI.createVirtualRegister(TRC);
7791 unsigned destLoop = MRI.createVirtualRegister(TRC);
7792 unsigned destPhi = MRI.createVirtualRegister(TRC);
7794 BuildMI(*BB, BB->begin(), dl, TII->get(ARM::PHI), varPhi)
7795 .addReg(varLoop).addMBB(loopMBB)
7796 .addReg(varEnd).addMBB(entryBB);
7797 BuildMI(BB, dl, TII->get(ARM::PHI), srcPhi)
7798 .addReg(srcLoop).addMBB(loopMBB)
7799 .addReg(src).addMBB(entryBB);
7800 BuildMI(BB, dl, TII->get(ARM::PHI), destPhi)
7801 .addReg(destLoop).addMBB(loopMBB)
7802 .addReg(dest).addMBB(entryBB);
7804 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
7805 // [destLoop] = STR_POST(scratch, destPhi, UnitSiz)
7806 unsigned scratch = MRI.createVirtualRegister(IsNeon ? VecTRC : TRC);
7807 emitPostLd(BB, BB->end(), TII, dl, UnitSize, scratch, srcPhi, srcLoop,
7808 IsThumb1, IsThumb2);
7809 emitPostSt(BB, BB->end(), TII, dl, UnitSize, scratch, destPhi, destLoop,
7810 IsThumb1, IsThumb2);
7812 // Decrement loop variable by UnitSize.
7814 MachineInstrBuilder MIB =
7815 BuildMI(*BB, BB->end(), dl, TII->get(ARM::tSUBi8), varLoop);
7816 MIB = AddDefaultT1CC(MIB);
7817 MIB.addReg(varPhi).addImm(UnitSize);
7818 AddDefaultPred(MIB);
7820 MachineInstrBuilder MIB =
7821 BuildMI(*BB, BB->end(), dl,
7822 TII->get(IsThumb2 ? ARM::t2SUBri : ARM::SUBri), varLoop);
7823 AddDefaultCC(AddDefaultPred(MIB.addReg(varPhi).addImm(UnitSize)));
7824 MIB->getOperand(5).setReg(ARM::CPSR);
7825 MIB->getOperand(5).setIsDef(true);
7827 BuildMI(*BB, BB->end(), dl,
7828 TII->get(IsThumb1 ? ARM::tBcc : IsThumb2 ? ARM::t2Bcc : ARM::Bcc))
7829 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
7831 // loopMBB can loop back to loopMBB or fall through to exitMBB.
7832 BB->addSuccessor(loopMBB);
7833 BB->addSuccessor(exitMBB);
7835 // Add epilogue to handle BytesLeft.
7837 MachineInstr *StartOfExit = exitMBB->begin();
7839 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
7840 // [destOut] = STRB_POST(scratch, destLoop, 1)
7841 unsigned srcIn = srcLoop;
7842 unsigned destIn = destLoop;
7843 for (unsigned i = 0; i < BytesLeft; i++) {
7844 unsigned srcOut = MRI.createVirtualRegister(TRC);
7845 unsigned destOut = MRI.createVirtualRegister(TRC);
7846 unsigned scratch = MRI.createVirtualRegister(TRC);
7847 emitPostLd(BB, StartOfExit, TII, dl, 1, scratch, srcIn, srcOut,
7848 IsThumb1, IsThumb2);
7849 emitPostSt(BB, StartOfExit, TII, dl, 1, scratch, destIn, destOut,
7850 IsThumb1, IsThumb2);
7855 MI->eraseFromParent(); // The instruction is gone now.
7860 ARMTargetLowering::EmitLowered__chkstk(MachineInstr *MI,
7861 MachineBasicBlock *MBB) const {
7862 const TargetMachine &TM = getTargetMachine();
7863 const TargetInstrInfo &TII = *Subtarget->getInstrInfo();
7864 DebugLoc DL = MI->getDebugLoc();
7866 assert(Subtarget->isTargetWindows() &&
7867 "__chkstk is only supported on Windows");
7868 assert(Subtarget->isThumb2() && "Windows on ARM requires Thumb-2 mode");
7870 // __chkstk takes the number of words to allocate on the stack in R4, and
7871 // returns the stack adjustment in number of bytes in R4. This will not
7872 // clober any other registers (other than the obvious lr).
7874 // Although, technically, IP should be considered a register which may be
7875 // clobbered, the call itself will not touch it. Windows on ARM is a pure
7876 // thumb-2 environment, so there is no interworking required. As a result, we
7877 // do not expect a veneer to be emitted by the linker, clobbering IP.
7879 // Each module receives its own copy of __chkstk, so no import thunk is
7880 // required, again, ensuring that IP is not clobbered.
7882 // Finally, although some linkers may theoretically provide a trampoline for
7883 // out of range calls (which is quite common due to a 32M range limitation of
7884 // branches for Thumb), we can generate the long-call version via
7885 // -mcmodel=large, alleviating the need for the trampoline which may clobber
7888 switch (TM.getCodeModel()) {
7889 case CodeModel::Small:
7890 case CodeModel::Medium:
7891 case CodeModel::Default:
7892 case CodeModel::Kernel:
7893 BuildMI(*MBB, MI, DL, TII.get(ARM::tBL))
7894 .addImm((unsigned)ARMCC::AL).addReg(0)
7895 .addExternalSymbol("__chkstk")
7896 .addReg(ARM::R4, RegState::Implicit | RegState::Kill)
7897 .addReg(ARM::R4, RegState::Implicit | RegState::Define)
7898 .addReg(ARM::R12, RegState::Implicit | RegState::Define | RegState::Dead);
7900 case CodeModel::Large:
7901 case CodeModel::JITDefault: {
7902 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
7903 unsigned Reg = MRI.createVirtualRegister(&ARM::rGPRRegClass);
7905 BuildMI(*MBB, MI, DL, TII.get(ARM::t2MOVi32imm), Reg)
7906 .addExternalSymbol("__chkstk");
7907 BuildMI(*MBB, MI, DL, TII.get(ARM::tBLXr))
7908 .addImm((unsigned)ARMCC::AL).addReg(0)
7909 .addReg(Reg, RegState::Kill)
7910 .addReg(ARM::R4, RegState::Implicit | RegState::Kill)
7911 .addReg(ARM::R4, RegState::Implicit | RegState::Define)
7912 .addReg(ARM::R12, RegState::Implicit | RegState::Define | RegState::Dead);
7917 AddDefaultCC(AddDefaultPred(BuildMI(*MBB, MI, DL, TII.get(ARM::t2SUBrr),
7919 .addReg(ARM::SP).addReg(ARM::R4)));
7921 MI->eraseFromParent();
7926 ARMTargetLowering::EmitLowered__dbzchk(MachineInstr *MI,
7927 MachineBasicBlock *MBB) const {
7928 DebugLoc DL = MI->getDebugLoc();
7929 MachineFunction *MF = MBB->getParent();
7930 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
7932 MachineBasicBlock *ContBB = MF->CreateMachineBasicBlock();
7933 MF->push_back(ContBB);
7934 ContBB->splice(ContBB->begin(), MBB,
7935 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
7936 MBB->addSuccessor(ContBB);
7938 MachineBasicBlock *TrapBB = MF->CreateMachineBasicBlock();
7939 MF->push_back(TrapBB);
7940 BuildMI(TrapBB, DL, TII->get(ARM::t2UDF)).addImm(249);
7941 MBB->addSuccessor(TrapBB);
7943 BuildMI(*MBB, MI, DL, TII->get(ARM::tCBZ))
7944 .addReg(MI->getOperand(0).getReg())
7947 MI->eraseFromParent();
7952 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
7953 MachineBasicBlock *BB) const {
7954 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
7955 DebugLoc dl = MI->getDebugLoc();
7956 bool isThumb2 = Subtarget->isThumb2();
7957 switch (MI->getOpcode()) {
7960 llvm_unreachable("Unexpected instr type to insert");
7962 // The Thumb2 pre-indexed stores have the same MI operands, they just
7963 // define them differently in the .td files from the isel patterns, so
7964 // they need pseudos.
7965 case ARM::t2STR_preidx:
7966 MI->setDesc(TII->get(ARM::t2STR_PRE));
7968 case ARM::t2STRB_preidx:
7969 MI->setDesc(TII->get(ARM::t2STRB_PRE));
7971 case ARM::t2STRH_preidx:
7972 MI->setDesc(TII->get(ARM::t2STRH_PRE));
7975 case ARM::STRi_preidx:
7976 case ARM::STRBi_preidx: {
7977 unsigned NewOpc = MI->getOpcode() == ARM::STRi_preidx ?
7978 ARM::STR_PRE_IMM : ARM::STRB_PRE_IMM;
7979 // Decode the offset.
7980 unsigned Offset = MI->getOperand(4).getImm();
7981 bool isSub = ARM_AM::getAM2Op(Offset) == ARM_AM::sub;
7982 Offset = ARM_AM::getAM2Offset(Offset);
7986 MachineMemOperand *MMO = *MI->memoperands_begin();
7987 BuildMI(*BB, MI, dl, TII->get(NewOpc))
7988 .addOperand(MI->getOperand(0)) // Rn_wb
7989 .addOperand(MI->getOperand(1)) // Rt
7990 .addOperand(MI->getOperand(2)) // Rn
7991 .addImm(Offset) // offset (skip GPR==zero_reg)
7992 .addOperand(MI->getOperand(5)) // pred
7993 .addOperand(MI->getOperand(6))
7994 .addMemOperand(MMO);
7995 MI->eraseFromParent();
7998 case ARM::STRr_preidx:
7999 case ARM::STRBr_preidx:
8000 case ARM::STRH_preidx: {
8002 switch (MI->getOpcode()) {
8003 default: llvm_unreachable("unexpected opcode!");
8004 case ARM::STRr_preidx: NewOpc = ARM::STR_PRE_REG; break;
8005 case ARM::STRBr_preidx: NewOpc = ARM::STRB_PRE_REG; break;
8006 case ARM::STRH_preidx: NewOpc = ARM::STRH_PRE; break;
8008 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(NewOpc));
8009 for (unsigned i = 0; i < MI->getNumOperands(); ++i)
8010 MIB.addOperand(MI->getOperand(i));
8011 MI->eraseFromParent();
8015 case ARM::tMOVCCr_pseudo: {
8016 // To "insert" a SELECT_CC instruction, we actually have to insert the
8017 // diamond control-flow pattern. The incoming instruction knows the
8018 // destination vreg to set, the condition code register to branch on, the
8019 // true/false values to select between, and a branch opcode to use.
8020 const BasicBlock *LLVM_BB = BB->getBasicBlock();
8021 MachineFunction::iterator It = ++BB->getIterator();
8026 // cmpTY ccX, r1, r2
8028 // fallthrough --> copy0MBB
8029 MachineBasicBlock *thisMBB = BB;
8030 MachineFunction *F = BB->getParent();
8031 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
8032 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
8033 F->insert(It, copy0MBB);
8034 F->insert(It, sinkMBB);
8036 // Transfer the remainder of BB and its successor edges to sinkMBB.
8037 sinkMBB->splice(sinkMBB->begin(), BB,
8038 std::next(MachineBasicBlock::iterator(MI)), BB->end());
8039 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
8041 BB->addSuccessor(copy0MBB);
8042 BB->addSuccessor(sinkMBB);
8044 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
8045 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
8048 // %FalseValue = ...
8049 // # fallthrough to sinkMBB
8052 // Update machine-CFG edges
8053 BB->addSuccessor(sinkMBB);
8056 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
8059 BuildMI(*BB, BB->begin(), dl,
8060 TII->get(ARM::PHI), MI->getOperand(0).getReg())
8061 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
8062 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
8064 MI->eraseFromParent(); // The pseudo instruction is gone now.
8069 case ARM::BCCZi64: {
8070 // If there is an unconditional branch to the other successor, remove it.
8071 BB->erase(std::next(MachineBasicBlock::iterator(MI)), BB->end());
8073 // Compare both parts that make up the double comparison separately for
8075 bool RHSisZero = MI->getOpcode() == ARM::BCCZi64;
8077 unsigned LHS1 = MI->getOperand(1).getReg();
8078 unsigned LHS2 = MI->getOperand(2).getReg();
8080 AddDefaultPred(BuildMI(BB, dl,
8081 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
8082 .addReg(LHS1).addImm(0));
8083 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
8084 .addReg(LHS2).addImm(0)
8085 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
8087 unsigned RHS1 = MI->getOperand(3).getReg();
8088 unsigned RHS2 = MI->getOperand(4).getReg();
8089 AddDefaultPred(BuildMI(BB, dl,
8090 TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
8091 .addReg(LHS1).addReg(RHS1));
8092 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
8093 .addReg(LHS2).addReg(RHS2)
8094 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
8097 MachineBasicBlock *destMBB = MI->getOperand(RHSisZero ? 3 : 5).getMBB();
8098 MachineBasicBlock *exitMBB = OtherSucc(BB, destMBB);
8099 if (MI->getOperand(0).getImm() == ARMCC::NE)
8100 std::swap(destMBB, exitMBB);
8102 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
8103 .addMBB(destMBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
8105 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2B)).addMBB(exitMBB));
8107 BuildMI(BB, dl, TII->get(ARM::B)) .addMBB(exitMBB);
8109 MI->eraseFromParent(); // The pseudo instruction is gone now.
8113 case ARM::Int_eh_sjlj_setjmp:
8114 case ARM::Int_eh_sjlj_setjmp_nofp:
8115 case ARM::tInt_eh_sjlj_setjmp:
8116 case ARM::t2Int_eh_sjlj_setjmp:
8117 case ARM::t2Int_eh_sjlj_setjmp_nofp:
8120 case ARM::Int_eh_sjlj_setup_dispatch:
8121 EmitSjLjDispatchBlock(MI, BB);
8126 // To insert an ABS instruction, we have to insert the
8127 // diamond control-flow pattern. The incoming instruction knows the
8128 // source vreg to test against 0, the destination vreg to set,
8129 // the condition code register to branch on, the
8130 // true/false values to select between, and a branch opcode to use.
8135 // BCC (branch to SinkBB if V0 >= 0)
8136 // RSBBB: V3 = RSBri V2, 0 (compute ABS if V2 < 0)
8137 // SinkBB: V1 = PHI(V2, V3)
8138 const BasicBlock *LLVM_BB = BB->getBasicBlock();
8139 MachineFunction::iterator BBI = ++BB->getIterator();
8140 MachineFunction *Fn = BB->getParent();
8141 MachineBasicBlock *RSBBB = Fn->CreateMachineBasicBlock(LLVM_BB);
8142 MachineBasicBlock *SinkBB = Fn->CreateMachineBasicBlock(LLVM_BB);
8143 Fn->insert(BBI, RSBBB);
8144 Fn->insert(BBI, SinkBB);
8146 unsigned int ABSSrcReg = MI->getOperand(1).getReg();
8147 unsigned int ABSDstReg = MI->getOperand(0).getReg();
8148 bool ABSSrcKIll = MI->getOperand(1).isKill();
8149 bool isThumb2 = Subtarget->isThumb2();
8150 MachineRegisterInfo &MRI = Fn->getRegInfo();
8151 // In Thumb mode S must not be specified if source register is the SP or
8152 // PC and if destination register is the SP, so restrict register class
8153 unsigned NewRsbDstReg =
8154 MRI.createVirtualRegister(isThumb2 ? &ARM::rGPRRegClass : &ARM::GPRRegClass);
8156 // Transfer the remainder of BB and its successor edges to sinkMBB.
8157 SinkBB->splice(SinkBB->begin(), BB,
8158 std::next(MachineBasicBlock::iterator(MI)), BB->end());
8159 SinkBB->transferSuccessorsAndUpdatePHIs(BB);
8161 BB->addSuccessor(RSBBB);
8162 BB->addSuccessor(SinkBB);
8164 // fall through to SinkMBB
8165 RSBBB->addSuccessor(SinkBB);
8167 // insert a cmp at the end of BB
8168 AddDefaultPred(BuildMI(BB, dl,
8169 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
8170 .addReg(ABSSrcReg).addImm(0));
8172 // insert a bcc with opposite CC to ARMCC::MI at the end of BB
8174 TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)).addMBB(SinkBB)
8175 .addImm(ARMCC::getOppositeCondition(ARMCC::MI)).addReg(ARM::CPSR);
8177 // insert rsbri in RSBBB
8178 // Note: BCC and rsbri will be converted into predicated rsbmi
8179 // by if-conversion pass
8180 BuildMI(*RSBBB, RSBBB->begin(), dl,
8181 TII->get(isThumb2 ? ARM::t2RSBri : ARM::RSBri), NewRsbDstReg)
8182 .addReg(ABSSrcReg, ABSSrcKIll ? RegState::Kill : 0)
8183 .addImm(0).addImm((unsigned)ARMCC::AL).addReg(0).addReg(0);
8185 // insert PHI in SinkBB,
8186 // reuse ABSDstReg to not change uses of ABS instruction
8187 BuildMI(*SinkBB, SinkBB->begin(), dl,
8188 TII->get(ARM::PHI), ABSDstReg)
8189 .addReg(NewRsbDstReg).addMBB(RSBBB)
8190 .addReg(ABSSrcReg).addMBB(BB);
8192 // remove ABS instruction
8193 MI->eraseFromParent();
8195 // return last added BB
8198 case ARM::COPY_STRUCT_BYVAL_I32:
8200 return EmitStructByval(MI, BB);
8201 case ARM::WIN__CHKSTK:
8202 return EmitLowered__chkstk(MI, BB);
8203 case ARM::WIN__DBZCHK:
8204 return EmitLowered__dbzchk(MI, BB);
8208 /// \brief Attaches vregs to MEMCPY that it will use as scratch registers
8209 /// when it is expanded into LDM/STM. This is done as a post-isel lowering
8210 /// instead of as a custom inserter because we need the use list from the SDNode.
8211 static void attachMEMCPYScratchRegs(const ARMSubtarget *Subtarget,
8212 MachineInstr *MI, const SDNode *Node) {
8213 bool isThumb1 = Subtarget->isThumb1Only();
8215 DebugLoc DL = MI->getDebugLoc();
8216 MachineFunction *MF = MI->getParent()->getParent();
8217 MachineRegisterInfo &MRI = MF->getRegInfo();
8218 MachineInstrBuilder MIB(*MF, MI);
8220 // If the new dst/src is unused mark it as dead.
8221 if (!Node->hasAnyUseOfValue(0)) {
8222 MI->getOperand(0).setIsDead(true);
8224 if (!Node->hasAnyUseOfValue(1)) {
8225 MI->getOperand(1).setIsDead(true);
8228 // The MEMCPY both defines and kills the scratch registers.
8229 for (unsigned I = 0; I != MI->getOperand(4).getImm(); ++I) {
8230 unsigned TmpReg = MRI.createVirtualRegister(isThumb1 ? &ARM::tGPRRegClass
8231 : &ARM::GPRRegClass);
8232 MIB.addReg(TmpReg, RegState::Define|RegState::Dead);
8236 void ARMTargetLowering::AdjustInstrPostInstrSelection(MachineInstr *MI,
8237 SDNode *Node) const {
8238 if (MI->getOpcode() == ARM::MEMCPY) {
8239 attachMEMCPYScratchRegs(Subtarget, MI, Node);
8243 const MCInstrDesc *MCID = &MI->getDesc();
8244 // Adjust potentially 's' setting instructions after isel, i.e. ADC, SBC, RSB,
8245 // RSC. Coming out of isel, they have an implicit CPSR def, but the optional
8246 // operand is still set to noreg. If needed, set the optional operand's
8247 // register to CPSR, and remove the redundant implicit def.
8249 // e.g. ADCS (..., CPSR<imp-def>) -> ADC (... opt:CPSR<def>).
8251 // Rename pseudo opcodes.
8252 unsigned NewOpc = convertAddSubFlagsOpcode(MI->getOpcode());
8254 const ARMBaseInstrInfo *TII = Subtarget->getInstrInfo();
8255 MCID = &TII->get(NewOpc);
8257 assert(MCID->getNumOperands() == MI->getDesc().getNumOperands() + 1 &&
8258 "converted opcode should be the same except for cc_out");
8262 // Add the optional cc_out operand
8263 MI->addOperand(MachineOperand::CreateReg(0, /*isDef=*/true));
8265 unsigned ccOutIdx = MCID->getNumOperands() - 1;
8267 // Any ARM instruction that sets the 's' bit should specify an optional
8268 // "cc_out" operand in the last operand position.
8269 if (!MI->hasOptionalDef() || !MCID->OpInfo[ccOutIdx].isOptionalDef()) {
8270 assert(!NewOpc && "Optional cc_out operand required");
8273 // Look for an implicit def of CPSR added by MachineInstr ctor. Remove it
8274 // since we already have an optional CPSR def.
8275 bool definesCPSR = false;
8276 bool deadCPSR = false;
8277 for (unsigned i = MCID->getNumOperands(), e = MI->getNumOperands();
8279 const MachineOperand &MO = MI->getOperand(i);
8280 if (MO.isReg() && MO.isDef() && MO.getReg() == ARM::CPSR) {
8284 MI->RemoveOperand(i);
8289 assert(!NewOpc && "Optional cc_out operand required");
8292 assert(deadCPSR == !Node->hasAnyUseOfValue(1) && "inconsistent dead flag");
8294 assert(!MI->getOperand(ccOutIdx).getReg() &&
8295 "expect uninitialized optional cc_out operand");
8299 // If this instruction was defined with an optional CPSR def and its dag node
8300 // had a live implicit CPSR def, then activate the optional CPSR def.
8301 MachineOperand &MO = MI->getOperand(ccOutIdx);
8302 MO.setReg(ARM::CPSR);
8306 //===----------------------------------------------------------------------===//
8307 // ARM Optimization Hooks
8308 //===----------------------------------------------------------------------===//
8310 // Helper function that checks if N is a null or all ones constant.
8311 static inline bool isZeroOrAllOnes(SDValue N, bool AllOnes) {
8312 return AllOnes ? isAllOnesConstant(N) : isNullConstant(N);
8315 // Return true if N is conditionally 0 or all ones.
8316 // Detects these expressions where cc is an i1 value:
8318 // (select cc 0, y) [AllOnes=0]
8319 // (select cc y, 0) [AllOnes=0]
8320 // (zext cc) [AllOnes=0]
8321 // (sext cc) [AllOnes=0/1]
8322 // (select cc -1, y) [AllOnes=1]
8323 // (select cc y, -1) [AllOnes=1]
8325 // Invert is set when N is the null/all ones constant when CC is false.
8326 // OtherOp is set to the alternative value of N.
8327 static bool isConditionalZeroOrAllOnes(SDNode *N, bool AllOnes,
8328 SDValue &CC, bool &Invert,
8330 SelectionDAG &DAG) {
8331 switch (N->getOpcode()) {
8332 default: return false;
8334 CC = N->getOperand(0);
8335 SDValue N1 = N->getOperand(1);
8336 SDValue N2 = N->getOperand(2);
8337 if (isZeroOrAllOnes(N1, AllOnes)) {
8342 if (isZeroOrAllOnes(N2, AllOnes)) {
8349 case ISD::ZERO_EXTEND:
8350 // (zext cc) can never be the all ones value.
8354 case ISD::SIGN_EXTEND: {
8356 EVT VT = N->getValueType(0);
8357 CC = N->getOperand(0);
8358 if (CC.getValueType() != MVT::i1)
8362 // When looking for an AllOnes constant, N is an sext, and the 'other'
8364 OtherOp = DAG.getConstant(0, dl, VT);
8365 else if (N->getOpcode() == ISD::ZERO_EXTEND)
8366 // When looking for a 0 constant, N can be zext or sext.
8367 OtherOp = DAG.getConstant(1, dl, VT);
8369 OtherOp = DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), dl,
8376 // Combine a constant select operand into its use:
8378 // (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
8379 // (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
8380 // (and (select cc, -1, c), x) -> (select cc, x, (and, x, c)) [AllOnes=1]
8381 // (or (select cc, 0, c), x) -> (select cc, x, (or, x, c))
8382 // (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c))
8384 // The transform is rejected if the select doesn't have a constant operand that
8385 // is null, or all ones when AllOnes is set.
8387 // Also recognize sext/zext from i1:
8389 // (add (zext cc), x) -> (select cc (add x, 1), x)
8390 // (add (sext cc), x) -> (select cc (add x, -1), x)
8392 // These transformations eventually create predicated instructions.
8394 // @param N The node to transform.
8395 // @param Slct The N operand that is a select.
8396 // @param OtherOp The other N operand (x above).
8397 // @param DCI Context.
8398 // @param AllOnes Require the select constant to be all ones instead of null.
8399 // @returns The new node, or SDValue() on failure.
8401 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
8402 TargetLowering::DAGCombinerInfo &DCI,
8403 bool AllOnes = false) {
8404 SelectionDAG &DAG = DCI.DAG;
8405 EVT VT = N->getValueType(0);
8406 SDValue NonConstantVal;
8409 if (!isConditionalZeroOrAllOnes(Slct.getNode(), AllOnes, CCOp, SwapSelectOps,
8410 NonConstantVal, DAG))
8413 // Slct is now know to be the desired identity constant when CC is true.
8414 SDValue TrueVal = OtherOp;
8415 SDValue FalseVal = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
8416 OtherOp, NonConstantVal);
8417 // Unless SwapSelectOps says CC should be false.
8419 std::swap(TrueVal, FalseVal);
8421 return DAG.getNode(ISD::SELECT, SDLoc(N), VT,
8422 CCOp, TrueVal, FalseVal);
8425 // Attempt combineSelectAndUse on each operand of a commutative operator N.
8427 SDValue combineSelectAndUseCommutative(SDNode *N, bool AllOnes,
8428 TargetLowering::DAGCombinerInfo &DCI) {
8429 SDValue N0 = N->getOperand(0);
8430 SDValue N1 = N->getOperand(1);
8431 if (N0.getNode()->hasOneUse()) {
8432 SDValue Result = combineSelectAndUse(N, N0, N1, DCI, AllOnes);
8433 if (Result.getNode())
8436 if (N1.getNode()->hasOneUse()) {
8437 SDValue Result = combineSelectAndUse(N, N1, N0, DCI, AllOnes);
8438 if (Result.getNode())
8444 // AddCombineToVPADDL- For pair-wise add on neon, use the vpaddl instruction
8445 // (only after legalization).
8446 static SDValue AddCombineToVPADDL(SDNode *N, SDValue N0, SDValue N1,
8447 TargetLowering::DAGCombinerInfo &DCI,
8448 const ARMSubtarget *Subtarget) {
8450 // Only perform optimization if after legalize, and if NEON is available. We
8451 // also expected both operands to be BUILD_VECTORs.
8452 if (DCI.isBeforeLegalize() || !Subtarget->hasNEON()
8453 || N0.getOpcode() != ISD::BUILD_VECTOR
8454 || N1.getOpcode() != ISD::BUILD_VECTOR)
8457 // Check output type since VPADDL operand elements can only be 8, 16, or 32.
8458 EVT VT = N->getValueType(0);
8459 if (!VT.isInteger() || VT.getVectorElementType() == MVT::i64)
8462 // Check that the vector operands are of the right form.
8463 // N0 and N1 are BUILD_VECTOR nodes with N number of EXTRACT_VECTOR
8464 // operands, where N is the size of the formed vector.
8465 // Each EXTRACT_VECTOR should have the same input vector and odd or even
8466 // index such that we have a pair wise add pattern.
8468 // Grab the vector that all EXTRACT_VECTOR nodes should be referencing.
8469 if (N0->getOperand(0)->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
8471 SDValue Vec = N0->getOperand(0)->getOperand(0);
8472 SDNode *V = Vec.getNode();
8473 unsigned nextIndex = 0;
8475 // For each operands to the ADD which are BUILD_VECTORs,
8476 // check to see if each of their operands are an EXTRACT_VECTOR with
8477 // the same vector and appropriate index.
8478 for (unsigned i = 0, e = N0->getNumOperands(); i != e; ++i) {
8479 if (N0->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT
8480 && N1->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
8482 SDValue ExtVec0 = N0->getOperand(i);
8483 SDValue ExtVec1 = N1->getOperand(i);
8485 // First operand is the vector, verify its the same.
8486 if (V != ExtVec0->getOperand(0).getNode() ||
8487 V != ExtVec1->getOperand(0).getNode())
8490 // Second is the constant, verify its correct.
8491 ConstantSDNode *C0 = dyn_cast<ConstantSDNode>(ExtVec0->getOperand(1));
8492 ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(ExtVec1->getOperand(1));
8494 // For the constant, we want to see all the even or all the odd.
8495 if (!C0 || !C1 || C0->getZExtValue() != nextIndex
8496 || C1->getZExtValue() != nextIndex+1)
8505 // Create VPADDL node.
8506 SelectionDAG &DAG = DCI.DAG;
8507 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8511 // Build operand list.
8512 SmallVector<SDValue, 8> Ops;
8513 Ops.push_back(DAG.getConstant(Intrinsic::arm_neon_vpaddls, dl,
8514 TLI.getPointerTy(DAG.getDataLayout())));
8516 // Input is the vector.
8519 // Get widened type and narrowed type.
8521 unsigned numElem = VT.getVectorNumElements();
8523 EVT inputLaneType = Vec.getValueType().getVectorElementType();
8524 switch (inputLaneType.getSimpleVT().SimpleTy) {
8525 case MVT::i8: widenType = MVT::getVectorVT(MVT::i16, numElem); break;
8526 case MVT::i16: widenType = MVT::getVectorVT(MVT::i32, numElem); break;
8527 case MVT::i32: widenType = MVT::getVectorVT(MVT::i64, numElem); break;
8529 llvm_unreachable("Invalid vector element type for padd optimization.");
8532 SDValue tmp = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, widenType, Ops);
8533 unsigned ExtOp = VT.bitsGT(tmp.getValueType()) ? ISD::ANY_EXTEND : ISD::TRUNCATE;
8534 return DAG.getNode(ExtOp, dl, VT, tmp);
8537 static SDValue findMUL_LOHI(SDValue V) {
8538 if (V->getOpcode() == ISD::UMUL_LOHI ||
8539 V->getOpcode() == ISD::SMUL_LOHI)
8544 static SDValue AddCombineTo64bitMLAL(SDNode *AddcNode,
8545 TargetLowering::DAGCombinerInfo &DCI,
8546 const ARMSubtarget *Subtarget) {
8548 if (Subtarget->isThumb1Only()) return SDValue();
8550 // Only perform the checks after legalize when the pattern is available.
8551 if (DCI.isBeforeLegalize()) return SDValue();
8553 // Look for multiply add opportunities.
8554 // The pattern is a ISD::UMUL_LOHI followed by two add nodes, where
8555 // each add nodes consumes a value from ISD::UMUL_LOHI and there is
8556 // a glue link from the first add to the second add.
8557 // If we find this pattern, we can replace the U/SMUL_LOHI, ADDC, and ADDE by
8558 // a S/UMLAL instruction.
8561 // / \ [no multiline comment]
8567 assert(AddcNode->getOpcode() == ISD::ADDC && "Expect an ADDC");
8568 SDValue AddcOp0 = AddcNode->getOperand(0);
8569 SDValue AddcOp1 = AddcNode->getOperand(1);
8571 // Check if the two operands are from the same mul_lohi node.
8572 if (AddcOp0.getNode() == AddcOp1.getNode())
8575 assert(AddcNode->getNumValues() == 2 &&
8576 AddcNode->getValueType(0) == MVT::i32 &&
8577 "Expect ADDC with two result values. First: i32");
8579 // Check that we have a glued ADDC node.
8580 if (AddcNode->getValueType(1) != MVT::Glue)
8583 // Check that the ADDC adds the low result of the S/UMUL_LOHI.
8584 if (AddcOp0->getOpcode() != ISD::UMUL_LOHI &&
8585 AddcOp0->getOpcode() != ISD::SMUL_LOHI &&
8586 AddcOp1->getOpcode() != ISD::UMUL_LOHI &&
8587 AddcOp1->getOpcode() != ISD::SMUL_LOHI)
8590 // Look for the glued ADDE.
8591 SDNode* AddeNode = AddcNode->getGluedUser();
8595 // Make sure it is really an ADDE.
8596 if (AddeNode->getOpcode() != ISD::ADDE)
8599 assert(AddeNode->getNumOperands() == 3 &&
8600 AddeNode->getOperand(2).getValueType() == MVT::Glue &&
8601 "ADDE node has the wrong inputs");
8603 // Check for the triangle shape.
8604 SDValue AddeOp0 = AddeNode->getOperand(0);
8605 SDValue AddeOp1 = AddeNode->getOperand(1);
8607 // Make sure that the ADDE operands are not coming from the same node.
8608 if (AddeOp0.getNode() == AddeOp1.getNode())
8611 // Find the MUL_LOHI node walking up ADDE's operands.
8612 bool IsLeftOperandMUL = false;
8613 SDValue MULOp = findMUL_LOHI(AddeOp0);
8614 if (MULOp == SDValue())
8615 MULOp = findMUL_LOHI(AddeOp1);
8617 IsLeftOperandMUL = true;
8618 if (MULOp == SDValue())
8621 // Figure out the right opcode.
8622 unsigned Opc = MULOp->getOpcode();
8623 unsigned FinalOpc = (Opc == ISD::SMUL_LOHI) ? ARMISD::SMLAL : ARMISD::UMLAL;
8625 // Figure out the high and low input values to the MLAL node.
8626 SDValue* HiAdd = nullptr;
8627 SDValue* LoMul = nullptr;
8628 SDValue* LowAdd = nullptr;
8630 // Ensure that ADDE is from high result of ISD::SMUL_LOHI.
8631 if ((AddeOp0 != MULOp.getValue(1)) && (AddeOp1 != MULOp.getValue(1)))
8634 if (IsLeftOperandMUL)
8640 // Ensure that LoMul and LowAdd are taken from correct ISD::SMUL_LOHI node
8641 // whose low result is fed to the ADDC we are checking.
8643 if (AddcOp0 == MULOp.getValue(0)) {
8647 if (AddcOp1 == MULOp.getValue(0)) {
8655 // Create the merged node.
8656 SelectionDAG &DAG = DCI.DAG;
8658 // Build operand list.
8659 SmallVector<SDValue, 8> Ops;
8660 Ops.push_back(LoMul->getOperand(0));
8661 Ops.push_back(LoMul->getOperand(1));
8662 Ops.push_back(*LowAdd);
8663 Ops.push_back(*HiAdd);
8665 SDValue MLALNode = DAG.getNode(FinalOpc, SDLoc(AddcNode),
8666 DAG.getVTList(MVT::i32, MVT::i32), Ops);
8668 // Replace the ADDs' nodes uses by the MLA node's values.
8669 SDValue HiMLALResult(MLALNode.getNode(), 1);
8670 DAG.ReplaceAllUsesOfValueWith(SDValue(AddeNode, 0), HiMLALResult);
8672 SDValue LoMLALResult(MLALNode.getNode(), 0);
8673 DAG.ReplaceAllUsesOfValueWith(SDValue(AddcNode, 0), LoMLALResult);
8675 // Return original node to notify the driver to stop replacing.
8676 SDValue resNode(AddcNode, 0);
8680 /// PerformADDCCombine - Target-specific dag combine transform from
8681 /// ISD::ADDC, ISD::ADDE, and ISD::MUL_LOHI to MLAL.
8682 static SDValue PerformADDCCombine(SDNode *N,
8683 TargetLowering::DAGCombinerInfo &DCI,
8684 const ARMSubtarget *Subtarget) {
8686 return AddCombineTo64bitMLAL(N, DCI, Subtarget);
8690 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
8691 /// operands N0 and N1. This is a helper for PerformADDCombine that is
8692 /// called with the default operands, and if that fails, with commuted
8694 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
8695 TargetLowering::DAGCombinerInfo &DCI,
8696 const ARMSubtarget *Subtarget){
8698 // Attempt to create vpaddl for this add.
8699 SDValue Result = AddCombineToVPADDL(N, N0, N1, DCI, Subtarget);
8700 if (Result.getNode())
8703 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
8704 if (N0.getNode()->hasOneUse()) {
8705 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
8706 if (Result.getNode()) return Result;
8711 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
8713 static SDValue PerformADDCombine(SDNode *N,
8714 TargetLowering::DAGCombinerInfo &DCI,
8715 const ARMSubtarget *Subtarget) {
8716 SDValue N0 = N->getOperand(0);
8717 SDValue N1 = N->getOperand(1);
8719 // First try with the default operand order.
8720 SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI, Subtarget);
8721 if (Result.getNode())
8724 // If that didn't work, try again with the operands commuted.
8725 return PerformADDCombineWithOperands(N, N1, N0, DCI, Subtarget);
8728 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
8730 static SDValue PerformSUBCombine(SDNode *N,
8731 TargetLowering::DAGCombinerInfo &DCI) {
8732 SDValue N0 = N->getOperand(0);
8733 SDValue N1 = N->getOperand(1);
8735 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
8736 if (N1.getNode()->hasOneUse()) {
8737 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
8738 if (Result.getNode()) return Result;
8744 /// PerformVMULCombine
8745 /// Distribute (A + B) * C to (A * C) + (B * C) to take advantage of the
8746 /// special multiplier accumulator forwarding.
8752 // However, for (A + B) * (A + B),
8759 static SDValue PerformVMULCombine(SDNode *N,
8760 TargetLowering::DAGCombinerInfo &DCI,
8761 const ARMSubtarget *Subtarget) {
8762 if (!Subtarget->hasVMLxForwarding())
8765 SelectionDAG &DAG = DCI.DAG;
8766 SDValue N0 = N->getOperand(0);
8767 SDValue N1 = N->getOperand(1);
8768 unsigned Opcode = N0.getOpcode();
8769 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
8770 Opcode != ISD::FADD && Opcode != ISD::FSUB) {
8771 Opcode = N1.getOpcode();
8772 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
8773 Opcode != ISD::FADD && Opcode != ISD::FSUB)
8781 EVT VT = N->getValueType(0);
8783 SDValue N00 = N0->getOperand(0);
8784 SDValue N01 = N0->getOperand(1);
8785 return DAG.getNode(Opcode, DL, VT,
8786 DAG.getNode(ISD::MUL, DL, VT, N00, N1),
8787 DAG.getNode(ISD::MUL, DL, VT, N01, N1));
8790 static SDValue PerformMULCombine(SDNode *N,
8791 TargetLowering::DAGCombinerInfo &DCI,
8792 const ARMSubtarget *Subtarget) {
8793 SelectionDAG &DAG = DCI.DAG;
8795 if (Subtarget->isThumb1Only())
8798 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
8801 EVT VT = N->getValueType(0);
8802 if (VT.is64BitVector() || VT.is128BitVector())
8803 return PerformVMULCombine(N, DCI, Subtarget);
8807 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
8811 int64_t MulAmt = C->getSExtValue();
8812 unsigned ShiftAmt = countTrailingZeros<uint64_t>(MulAmt);
8814 ShiftAmt = ShiftAmt & (32 - 1);
8815 SDValue V = N->getOperand(0);
8819 MulAmt >>= ShiftAmt;
8822 if (isPowerOf2_32(MulAmt - 1)) {
8823 // (mul x, 2^N + 1) => (add (shl x, N), x)
8824 Res = DAG.getNode(ISD::ADD, DL, VT,
8826 DAG.getNode(ISD::SHL, DL, VT,
8828 DAG.getConstant(Log2_32(MulAmt - 1), DL,
8830 } else if (isPowerOf2_32(MulAmt + 1)) {
8831 // (mul x, 2^N - 1) => (sub (shl x, N), x)
8832 Res = DAG.getNode(ISD::SUB, DL, VT,
8833 DAG.getNode(ISD::SHL, DL, VT,
8835 DAG.getConstant(Log2_32(MulAmt + 1), DL,
8841 uint64_t MulAmtAbs = -MulAmt;
8842 if (isPowerOf2_32(MulAmtAbs + 1)) {
8843 // (mul x, -(2^N - 1)) => (sub x, (shl x, N))
8844 Res = DAG.getNode(ISD::SUB, DL, VT,
8846 DAG.getNode(ISD::SHL, DL, VT,
8848 DAG.getConstant(Log2_32(MulAmtAbs + 1), DL,
8850 } else if (isPowerOf2_32(MulAmtAbs - 1)) {
8851 // (mul x, -(2^N + 1)) => - (add (shl x, N), x)
8852 Res = DAG.getNode(ISD::ADD, DL, VT,
8854 DAG.getNode(ISD::SHL, DL, VT,
8856 DAG.getConstant(Log2_32(MulAmtAbs - 1), DL,
8858 Res = DAG.getNode(ISD::SUB, DL, VT,
8859 DAG.getConstant(0, DL, MVT::i32), Res);
8866 Res = DAG.getNode(ISD::SHL, DL, VT,
8867 Res, DAG.getConstant(ShiftAmt, DL, MVT::i32));
8869 // Do not add new nodes to DAG combiner worklist.
8870 DCI.CombineTo(N, Res, false);
8874 static SDValue PerformANDCombine(SDNode *N,
8875 TargetLowering::DAGCombinerInfo &DCI,
8876 const ARMSubtarget *Subtarget) {
8878 // Attempt to use immediate-form VBIC
8879 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
8881 EVT VT = N->getValueType(0);
8882 SelectionDAG &DAG = DCI.DAG;
8884 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8887 APInt SplatBits, SplatUndef;
8888 unsigned SplatBitSize;
8891 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
8892 if (SplatBitSize <= 64) {
8894 SDValue Val = isNEONModifiedImm((~SplatBits).getZExtValue(),
8895 SplatUndef.getZExtValue(), SplatBitSize,
8896 DAG, dl, VbicVT, VT.is128BitVector(),
8898 if (Val.getNode()) {
8900 DAG.getNode(ISD::BITCAST, dl, VbicVT, N->getOperand(0));
8901 SDValue Vbic = DAG.getNode(ARMISD::VBICIMM, dl, VbicVT, Input, Val);
8902 return DAG.getNode(ISD::BITCAST, dl, VT, Vbic);
8907 if (!Subtarget->isThumb1Only()) {
8908 // fold (and (select cc, -1, c), x) -> (select cc, x, (and, x, c))
8909 SDValue Result = combineSelectAndUseCommutative(N, true, DCI);
8910 if (Result.getNode())
8917 /// PerformORCombine - Target-specific dag combine xforms for ISD::OR
8918 static SDValue PerformORCombine(SDNode *N,
8919 TargetLowering::DAGCombinerInfo &DCI,
8920 const ARMSubtarget *Subtarget) {
8921 // Attempt to use immediate-form VORR
8922 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
8924 EVT VT = N->getValueType(0);
8925 SelectionDAG &DAG = DCI.DAG;
8927 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8930 APInt SplatBits, SplatUndef;
8931 unsigned SplatBitSize;
8933 if (BVN && Subtarget->hasNEON() &&
8934 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
8935 if (SplatBitSize <= 64) {
8937 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
8938 SplatUndef.getZExtValue(), SplatBitSize,
8939 DAG, dl, VorrVT, VT.is128BitVector(),
8941 if (Val.getNode()) {
8943 DAG.getNode(ISD::BITCAST, dl, VorrVT, N->getOperand(0));
8944 SDValue Vorr = DAG.getNode(ARMISD::VORRIMM, dl, VorrVT, Input, Val);
8945 return DAG.getNode(ISD::BITCAST, dl, VT, Vorr);
8950 if (!Subtarget->isThumb1Only()) {
8951 // fold (or (select cc, 0, c), x) -> (select cc, x, (or, x, c))
8952 SDValue Result = combineSelectAndUseCommutative(N, false, DCI);
8953 if (Result.getNode())
8957 // The code below optimizes (or (and X, Y), Z).
8958 // The AND operand needs to have a single user to make these optimizations
8960 SDValue N0 = N->getOperand(0);
8961 if (N0.getOpcode() != ISD::AND || !N0.hasOneUse())
8963 SDValue N1 = N->getOperand(1);
8965 // (or (and B, A), (and C, ~A)) => (VBSL A, B, C) when A is a constant.
8966 if (Subtarget->hasNEON() && N1.getOpcode() == ISD::AND && VT.isVector() &&
8967 DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
8969 unsigned SplatBitSize;
8972 APInt SplatBits0, SplatBits1;
8973 BuildVectorSDNode *BVN0 = dyn_cast<BuildVectorSDNode>(N0->getOperand(1));
8974 BuildVectorSDNode *BVN1 = dyn_cast<BuildVectorSDNode>(N1->getOperand(1));
8975 // Ensure that the second operand of both ands are constants
8976 if (BVN0 && BVN0->isConstantSplat(SplatBits0, SplatUndef, SplatBitSize,
8977 HasAnyUndefs) && !HasAnyUndefs) {
8978 if (BVN1 && BVN1->isConstantSplat(SplatBits1, SplatUndef, SplatBitSize,
8979 HasAnyUndefs) && !HasAnyUndefs) {
8980 // Ensure that the bit width of the constants are the same and that
8981 // the splat arguments are logical inverses as per the pattern we
8982 // are trying to simplify.
8983 if (SplatBits0.getBitWidth() == SplatBits1.getBitWidth() &&
8984 SplatBits0 == ~SplatBits1) {
8985 // Canonicalize the vector type to make instruction selection
8987 EVT CanonicalVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
8988 SDValue Result = DAG.getNode(ARMISD::VBSL, dl, CanonicalVT,
8992 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
8998 // Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when
9001 // BFI is only available on V6T2+
9002 if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops())
9006 // 1) or (and A, mask), val => ARMbfi A, val, mask
9007 // iff (val & mask) == val
9009 // 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
9010 // 2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2)
9011 // && mask == ~mask2
9012 // 2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2)
9013 // && ~mask == mask2
9014 // (i.e., copy a bitfield value into another bitfield of the same width)
9019 SDValue N00 = N0.getOperand(0);
9021 // The value and the mask need to be constants so we can verify this is
9022 // actually a bitfield set. If the mask is 0xffff, we can do better
9023 // via a movt instruction, so don't use BFI in that case.
9024 SDValue MaskOp = N0.getOperand(1);
9025 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(MaskOp);
9028 unsigned Mask = MaskC->getZExtValue();
9032 // Case (1): or (and A, mask), val => ARMbfi A, val, mask
9033 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
9035 unsigned Val = N1C->getZExtValue();
9036 if ((Val & ~Mask) != Val)
9039 if (ARM::isBitFieldInvertedMask(Mask)) {
9040 Val >>= countTrailingZeros(~Mask);
9042 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00,
9043 DAG.getConstant(Val, DL, MVT::i32),
9044 DAG.getConstant(Mask, DL, MVT::i32));
9046 // Do not add new nodes to DAG combiner worklist.
9047 DCI.CombineTo(N, Res, false);
9050 } else if (N1.getOpcode() == ISD::AND) {
9051 // case (2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
9052 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
9055 unsigned Mask2 = N11C->getZExtValue();
9057 // Mask and ~Mask2 (or reverse) must be equivalent for the BFI pattern
9059 if (ARM::isBitFieldInvertedMask(Mask) &&
9061 // The pack halfword instruction works better for masks that fit it,
9062 // so use that when it's available.
9063 if (Subtarget->hasT2ExtractPack() &&
9064 (Mask == 0xffff || Mask == 0xffff0000))
9067 unsigned amt = countTrailingZeros(Mask2);
9068 Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0),
9069 DAG.getConstant(amt, DL, MVT::i32));
9070 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00, Res,
9071 DAG.getConstant(Mask, DL, MVT::i32));
9072 // Do not add new nodes to DAG combiner worklist.
9073 DCI.CombineTo(N, Res, false);
9075 } else if (ARM::isBitFieldInvertedMask(~Mask) &&
9077 // The pack halfword instruction works better for masks that fit it,
9078 // so use that when it's available.
9079 if (Subtarget->hasT2ExtractPack() &&
9080 (Mask2 == 0xffff || Mask2 == 0xffff0000))
9083 unsigned lsb = countTrailingZeros(Mask);
9084 Res = DAG.getNode(ISD::SRL, DL, VT, N00,
9085 DAG.getConstant(lsb, DL, MVT::i32));
9086 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res,
9087 DAG.getConstant(Mask2, DL, MVT::i32));
9088 // Do not add new nodes to DAG combiner worklist.
9089 DCI.CombineTo(N, Res, false);
9094 if (DAG.MaskedValueIsZero(N1, MaskC->getAPIntValue()) &&
9095 N00.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N00.getOperand(1)) &&
9096 ARM::isBitFieldInvertedMask(~Mask)) {
9097 // Case (3): or (and (shl A, #shamt), mask), B => ARMbfi B, A, ~mask
9098 // where lsb(mask) == #shamt and masked bits of B are known zero.
9099 SDValue ShAmt = N00.getOperand(1);
9100 unsigned ShAmtC = cast<ConstantSDNode>(ShAmt)->getZExtValue();
9101 unsigned LSB = countTrailingZeros(Mask);
9105 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1, N00.getOperand(0),
9106 DAG.getConstant(~Mask, DL, MVT::i32));
9108 // Do not add new nodes to DAG combiner worklist.
9109 DCI.CombineTo(N, Res, false);
9115 static SDValue PerformXORCombine(SDNode *N,
9116 TargetLowering::DAGCombinerInfo &DCI,
9117 const ARMSubtarget *Subtarget) {
9118 EVT VT = N->getValueType(0);
9119 SelectionDAG &DAG = DCI.DAG;
9121 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
9124 if (!Subtarget->isThumb1Only()) {
9125 // fold (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c))
9126 SDValue Result = combineSelectAndUseCommutative(N, false, DCI);
9127 if (Result.getNode())
9134 // ParseBFI - given a BFI instruction in N, extract the "from" value (Rn) and return it,
9135 // and fill in FromMask and ToMask with (consecutive) bits in "from" to be extracted and
9136 // their position in "to" (Rd).
9137 static SDValue ParseBFI(SDNode *N, APInt &ToMask, APInt &FromMask) {
9138 assert(N->getOpcode() == ARMISD::BFI);
9140 SDValue From = N->getOperand(1);
9141 ToMask = ~cast<ConstantSDNode>(N->getOperand(2))->getAPIntValue();
9142 FromMask = APInt::getLowBitsSet(ToMask.getBitWidth(), ToMask.countPopulation());
9144 // If the Base came from a SHR #C, we can deduce that it is really testing bit
9145 // #C in the base of the SHR.
9146 if (From->getOpcode() == ISD::SRL &&
9147 isa<ConstantSDNode>(From->getOperand(1))) {
9148 APInt Shift = cast<ConstantSDNode>(From->getOperand(1))->getAPIntValue();
9149 assert(Shift.getLimitedValue() < 32 && "Shift too large!");
9150 FromMask <<= Shift.getLimitedValue(31);
9151 From = From->getOperand(0);
9157 // If A and B contain one contiguous set of bits, does A | B == A . B?
9159 // Neither A nor B must be zero.
9160 static bool BitsProperlyConcatenate(const APInt &A, const APInt &B) {
9161 unsigned LastActiveBitInA = A.countTrailingZeros();
9162 unsigned FirstActiveBitInB = B.getBitWidth() - B.countLeadingZeros() - 1;
9163 return LastActiveBitInA - 1 == FirstActiveBitInB;
9166 static SDValue FindBFIToCombineWith(SDNode *N) {
9167 // We have a BFI in N. Follow a possible chain of BFIs and find a BFI it can combine with,
9169 APInt ToMask, FromMask;
9170 SDValue From = ParseBFI(N, ToMask, FromMask);
9171 SDValue To = N->getOperand(0);
9173 // Now check for a compatible BFI to merge with. We can pass through BFIs that
9174 // aren't compatible, but not if they set the same bit in their destination as
9175 // we do (or that of any BFI we're going to combine with).
9177 APInt CombinedToMask = ToMask;
9178 while (V.getOpcode() == ARMISD::BFI) {
9179 APInt NewToMask, NewFromMask;
9180 SDValue NewFrom = ParseBFI(V.getNode(), NewToMask, NewFromMask);
9181 if (NewFrom != From) {
9182 // This BFI has a different base. Keep going.
9183 CombinedToMask |= NewToMask;
9184 V = V.getOperand(0);
9188 // Do the written bits conflict with any we've seen so far?
9189 if ((NewToMask & CombinedToMask).getBoolValue())
9190 // Conflicting bits - bail out because going further is unsafe.
9193 // Are the new bits contiguous when combined with the old bits?
9194 if (BitsProperlyConcatenate(ToMask, NewToMask) &&
9195 BitsProperlyConcatenate(FromMask, NewFromMask))
9197 if (BitsProperlyConcatenate(NewToMask, ToMask) &&
9198 BitsProperlyConcatenate(NewFromMask, FromMask))
9201 // We've seen a write to some bits, so track it.
9202 CombinedToMask |= NewToMask;
9204 V = V.getOperand(0);
9210 static SDValue PerformBFICombine(SDNode *N,
9211 TargetLowering::DAGCombinerInfo &DCI) {
9212 SDValue N1 = N->getOperand(1);
9213 if (N1.getOpcode() == ISD::AND) {
9214 // (bfi A, (and B, Mask1), Mask2) -> (bfi A, B, Mask2) iff
9215 // the bits being cleared by the AND are not demanded by the BFI.
9216 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
9219 unsigned InvMask = cast<ConstantSDNode>(N->getOperand(2))->getZExtValue();
9220 unsigned LSB = countTrailingZeros(~InvMask);
9221 unsigned Width = (32 - countLeadingZeros(~InvMask)) - LSB;
9223 static_cast<unsigned>(std::numeric_limits<unsigned>::digits) &&
9224 "undefined behavior");
9225 unsigned Mask = (1u << Width) - 1;
9226 unsigned Mask2 = N11C->getZExtValue();
9227 if ((Mask & (~Mask2)) == 0)
9228 return DCI.DAG.getNode(ARMISD::BFI, SDLoc(N), N->getValueType(0),
9229 N->getOperand(0), N1.getOperand(0),
9231 } else if (N->getOperand(0).getOpcode() == ARMISD::BFI) {
9232 // We have a BFI of a BFI. Walk up the BFI chain to see how long it goes.
9233 // Keep track of any consecutive bits set that all come from the same base
9234 // value. We can combine these together into a single BFI.
9235 SDValue CombineBFI = FindBFIToCombineWith(N);
9236 if (CombineBFI == SDValue())
9239 // We've found a BFI.
9240 APInt ToMask1, FromMask1;
9241 SDValue From1 = ParseBFI(N, ToMask1, FromMask1);
9243 APInt ToMask2, FromMask2;
9244 SDValue From2 = ParseBFI(CombineBFI.getNode(), ToMask2, FromMask2);
9245 assert(From1 == From2);
9248 // First, unlink CombineBFI.
9249 DCI.DAG.ReplaceAllUsesWith(CombineBFI, CombineBFI.getOperand(0));
9250 // Then create a new BFI, combining the two together.
9251 APInt NewFromMask = FromMask1 | FromMask2;
9252 APInt NewToMask = ToMask1 | ToMask2;
9254 EVT VT = N->getValueType(0);
9257 if (NewFromMask[0] == 0)
9258 From1 = DCI.DAG.getNode(
9259 ISD::SRL, dl, VT, From1,
9260 DCI.DAG.getConstant(NewFromMask.countTrailingZeros(), dl, VT));
9261 return DCI.DAG.getNode(ARMISD::BFI, dl, VT, N->getOperand(0), From1,
9262 DCI.DAG.getConstant(~NewToMask, dl, VT));
9267 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
9268 /// ARMISD::VMOVRRD.
9269 static SDValue PerformVMOVRRDCombine(SDNode *N,
9270 TargetLowering::DAGCombinerInfo &DCI,
9271 const ARMSubtarget *Subtarget) {
9272 // vmovrrd(vmovdrr x, y) -> x,y
9273 SDValue InDouble = N->getOperand(0);
9274 if (InDouble.getOpcode() == ARMISD::VMOVDRR && !Subtarget->isFPOnlySP())
9275 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
9277 // vmovrrd(load f64) -> (load i32), (load i32)
9278 SDNode *InNode = InDouble.getNode();
9279 if (ISD::isNormalLoad(InNode) && InNode->hasOneUse() &&
9280 InNode->getValueType(0) == MVT::f64 &&
9281 InNode->getOperand(1).getOpcode() == ISD::FrameIndex &&
9282 !cast<LoadSDNode>(InNode)->isVolatile()) {
9283 // TODO: Should this be done for non-FrameIndex operands?
9284 LoadSDNode *LD = cast<LoadSDNode>(InNode);
9286 SelectionDAG &DAG = DCI.DAG;
9288 SDValue BasePtr = LD->getBasePtr();
9289 SDValue NewLD1 = DAG.getLoad(MVT::i32, DL, LD->getChain(), BasePtr,
9290 LD->getPointerInfo(), LD->isVolatile(),
9291 LD->isNonTemporal(), LD->isInvariant(),
9292 LD->getAlignment());
9294 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
9295 DAG.getConstant(4, DL, MVT::i32));
9296 SDValue NewLD2 = DAG.getLoad(MVT::i32, DL, NewLD1.getValue(1), OffsetPtr,
9297 LD->getPointerInfo(), LD->isVolatile(),
9298 LD->isNonTemporal(), LD->isInvariant(),
9299 std::min(4U, LD->getAlignment() / 2));
9301 DAG.ReplaceAllUsesOfValueWith(SDValue(LD, 1), NewLD2.getValue(1));
9302 if (DCI.DAG.getDataLayout().isBigEndian())
9303 std::swap (NewLD1, NewLD2);
9304 SDValue Result = DCI.CombineTo(N, NewLD1, NewLD2);
9311 /// PerformVMOVDRRCombine - Target-specific dag combine xforms for
9312 /// ARMISD::VMOVDRR. This is also used for BUILD_VECTORs with 2 operands.
9313 static SDValue PerformVMOVDRRCombine(SDNode *N, SelectionDAG &DAG) {
9314 // N=vmovrrd(X); vmovdrr(N:0, N:1) -> bit_convert(X)
9315 SDValue Op0 = N->getOperand(0);
9316 SDValue Op1 = N->getOperand(1);
9317 if (Op0.getOpcode() == ISD::BITCAST)
9318 Op0 = Op0.getOperand(0);
9319 if (Op1.getOpcode() == ISD::BITCAST)
9320 Op1 = Op1.getOperand(0);
9321 if (Op0.getOpcode() == ARMISD::VMOVRRD &&
9322 Op0.getNode() == Op1.getNode() &&
9323 Op0.getResNo() == 0 && Op1.getResNo() == 1)
9324 return DAG.getNode(ISD::BITCAST, SDLoc(N),
9325 N->getValueType(0), Op0.getOperand(0));
9329 /// hasNormalLoadOperand - Check if any of the operands of a BUILD_VECTOR node
9330 /// are normal, non-volatile loads. If so, it is profitable to bitcast an
9331 /// i64 vector to have f64 elements, since the value can then be loaded
9332 /// directly into a VFP register.
9333 static bool hasNormalLoadOperand(SDNode *N) {
9334 unsigned NumElts = N->getValueType(0).getVectorNumElements();
9335 for (unsigned i = 0; i < NumElts; ++i) {
9336 SDNode *Elt = N->getOperand(i).getNode();
9337 if (ISD::isNormalLoad(Elt) && !cast<LoadSDNode>(Elt)->isVolatile())
9343 /// PerformBUILD_VECTORCombine - Target-specific dag combine xforms for
9344 /// ISD::BUILD_VECTOR.
9345 static SDValue PerformBUILD_VECTORCombine(SDNode *N,
9346 TargetLowering::DAGCombinerInfo &DCI,
9347 const ARMSubtarget *Subtarget) {
9348 // build_vector(N=ARMISD::VMOVRRD(X), N:1) -> bit_convert(X):
9349 // VMOVRRD is introduced when legalizing i64 types. It forces the i64 value
9350 // into a pair of GPRs, which is fine when the value is used as a scalar,
9351 // but if the i64 value is converted to a vector, we need to undo the VMOVRRD.
9352 SelectionDAG &DAG = DCI.DAG;
9353 if (N->getNumOperands() == 2) {
9354 SDValue RV = PerformVMOVDRRCombine(N, DAG);
9359 // Load i64 elements as f64 values so that type legalization does not split
9360 // them up into i32 values.
9361 EVT VT = N->getValueType(0);
9362 if (VT.getVectorElementType() != MVT::i64 || !hasNormalLoadOperand(N))
9365 SmallVector<SDValue, 8> Ops;
9366 unsigned NumElts = VT.getVectorNumElements();
9367 for (unsigned i = 0; i < NumElts; ++i) {
9368 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(i));
9370 // Make the DAGCombiner fold the bitcast.
9371 DCI.AddToWorklist(V.getNode());
9373 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, NumElts);
9374 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, FloatVT, Ops);
9375 return DAG.getNode(ISD::BITCAST, dl, VT, BV);
9378 /// \brief Target-specific dag combine xforms for ARMISD::BUILD_VECTOR.
9380 PerformARMBUILD_VECTORCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
9381 // ARMISD::BUILD_VECTOR is introduced when legalizing ISD::BUILD_VECTOR.
9382 // At that time, we may have inserted bitcasts from integer to float.
9383 // If these bitcasts have survived DAGCombine, change the lowering of this
9384 // BUILD_VECTOR in something more vector friendly, i.e., that does not
9385 // force to use floating point types.
9387 // Make sure we can change the type of the vector.
9388 // This is possible iff:
9389 // 1. The vector is only used in a bitcast to a integer type. I.e.,
9390 // 1.1. Vector is used only once.
9391 // 1.2. Use is a bit convert to an integer type.
9392 // 2. The size of its operands are 32-bits (64-bits are not legal).
9393 EVT VT = N->getValueType(0);
9394 EVT EltVT = VT.getVectorElementType();
9396 // Check 1.1. and 2.
9397 if (EltVT.getSizeInBits() != 32 || !N->hasOneUse())
9400 // By construction, the input type must be float.
9401 assert(EltVT == MVT::f32 && "Unexpected type!");
9404 SDNode *Use = *N->use_begin();
9405 if (Use->getOpcode() != ISD::BITCAST ||
9406 Use->getValueType(0).isFloatingPoint())
9409 // Check profitability.
9410 // Model is, if more than half of the relevant operands are bitcast from
9411 // i32, turn the build_vector into a sequence of insert_vector_elt.
9412 // Relevant operands are everything that is not statically
9413 // (i.e., at compile time) bitcasted.
9414 unsigned NumOfBitCastedElts = 0;
9415 unsigned NumElts = VT.getVectorNumElements();
9416 unsigned NumOfRelevantElts = NumElts;
9417 for (unsigned Idx = 0; Idx < NumElts; ++Idx) {
9418 SDValue Elt = N->getOperand(Idx);
9419 if (Elt->getOpcode() == ISD::BITCAST) {
9420 // Assume only bit cast to i32 will go away.
9421 if (Elt->getOperand(0).getValueType() == MVT::i32)
9422 ++NumOfBitCastedElts;
9423 } else if (Elt.getOpcode() == ISD::UNDEF || isa<ConstantSDNode>(Elt))
9424 // Constants are statically casted, thus do not count them as
9425 // relevant operands.
9426 --NumOfRelevantElts;
9429 // Check if more than half of the elements require a non-free bitcast.
9430 if (NumOfBitCastedElts <= NumOfRelevantElts / 2)
9433 SelectionDAG &DAG = DCI.DAG;
9434 // Create the new vector type.
9435 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
9436 // Check if the type is legal.
9437 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9438 if (!TLI.isTypeLegal(VecVT))
9442 // ARMISD::BUILD_VECTOR E1, E2, ..., EN.
9443 // => BITCAST INSERT_VECTOR_ELT
9444 // (INSERT_VECTOR_ELT (...), (BITCAST EN-1), N-1),
9446 SDValue Vec = DAG.getUNDEF(VecVT);
9448 for (unsigned Idx = 0 ; Idx < NumElts; ++Idx) {
9449 SDValue V = N->getOperand(Idx);
9450 if (V.getOpcode() == ISD::UNDEF)
9452 if (V.getOpcode() == ISD::BITCAST &&
9453 V->getOperand(0).getValueType() == MVT::i32)
9454 // Fold obvious case.
9455 V = V.getOperand(0);
9457 V = DAG.getNode(ISD::BITCAST, SDLoc(V), MVT::i32, V);
9458 // Make the DAGCombiner fold the bitcasts.
9459 DCI.AddToWorklist(V.getNode());
9461 SDValue LaneIdx = DAG.getConstant(Idx, dl, MVT::i32);
9462 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VecVT, Vec, V, LaneIdx);
9464 Vec = DAG.getNode(ISD::BITCAST, dl, VT, Vec);
9465 // Make the DAGCombiner fold the bitcasts.
9466 DCI.AddToWorklist(Vec.getNode());
9470 /// PerformInsertEltCombine - Target-specific dag combine xforms for
9471 /// ISD::INSERT_VECTOR_ELT.
9472 static SDValue PerformInsertEltCombine(SDNode *N,
9473 TargetLowering::DAGCombinerInfo &DCI) {
9474 // Bitcast an i64 load inserted into a vector to f64.
9475 // Otherwise, the i64 value will be legalized to a pair of i32 values.
9476 EVT VT = N->getValueType(0);
9477 SDNode *Elt = N->getOperand(1).getNode();
9478 if (VT.getVectorElementType() != MVT::i64 ||
9479 !ISD::isNormalLoad(Elt) || cast<LoadSDNode>(Elt)->isVolatile())
9482 SelectionDAG &DAG = DCI.DAG;
9484 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
9485 VT.getVectorNumElements());
9486 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, N->getOperand(0));
9487 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(1));
9488 // Make the DAGCombiner fold the bitcasts.
9489 DCI.AddToWorklist(Vec.getNode());
9490 DCI.AddToWorklist(V.getNode());
9491 SDValue InsElt = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, FloatVT,
9492 Vec, V, N->getOperand(2));
9493 return DAG.getNode(ISD::BITCAST, dl, VT, InsElt);
9496 /// PerformVECTOR_SHUFFLECombine - Target-specific dag combine xforms for
9497 /// ISD::VECTOR_SHUFFLE.
9498 static SDValue PerformVECTOR_SHUFFLECombine(SDNode *N, SelectionDAG &DAG) {
9499 // The LLVM shufflevector instruction does not require the shuffle mask
9500 // length to match the operand vector length, but ISD::VECTOR_SHUFFLE does
9501 // have that requirement. When translating to ISD::VECTOR_SHUFFLE, if the
9502 // operands do not match the mask length, they are extended by concatenating
9503 // them with undef vectors. That is probably the right thing for other
9504 // targets, but for NEON it is better to concatenate two double-register
9505 // size vector operands into a single quad-register size vector. Do that
9506 // transformation here:
9507 // shuffle(concat(v1, undef), concat(v2, undef)) ->
9508 // shuffle(concat(v1, v2), undef)
9509 SDValue Op0 = N->getOperand(0);
9510 SDValue Op1 = N->getOperand(1);
9511 if (Op0.getOpcode() != ISD::CONCAT_VECTORS ||
9512 Op1.getOpcode() != ISD::CONCAT_VECTORS ||
9513 Op0.getNumOperands() != 2 ||
9514 Op1.getNumOperands() != 2)
9516 SDValue Concat0Op1 = Op0.getOperand(1);
9517 SDValue Concat1Op1 = Op1.getOperand(1);
9518 if (Concat0Op1.getOpcode() != ISD::UNDEF ||
9519 Concat1Op1.getOpcode() != ISD::UNDEF)
9521 // Skip the transformation if any of the types are illegal.
9522 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9523 EVT VT = N->getValueType(0);
9524 if (!TLI.isTypeLegal(VT) ||
9525 !TLI.isTypeLegal(Concat0Op1.getValueType()) ||
9526 !TLI.isTypeLegal(Concat1Op1.getValueType()))
9529 SDValue NewConcat = DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT,
9530 Op0.getOperand(0), Op1.getOperand(0));
9531 // Translate the shuffle mask.
9532 SmallVector<int, 16> NewMask;
9533 unsigned NumElts = VT.getVectorNumElements();
9534 unsigned HalfElts = NumElts/2;
9535 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
9536 for (unsigned n = 0; n < NumElts; ++n) {
9537 int MaskElt = SVN->getMaskElt(n);
9539 if (MaskElt < (int)HalfElts)
9541 else if (MaskElt >= (int)NumElts && MaskElt < (int)(NumElts + HalfElts))
9542 NewElt = HalfElts + MaskElt - NumElts;
9543 NewMask.push_back(NewElt);
9545 return DAG.getVectorShuffle(VT, SDLoc(N), NewConcat,
9546 DAG.getUNDEF(VT), NewMask.data());
9549 /// CombineBaseUpdate - Target-specific DAG combine function for VLDDUP,
9550 /// NEON load/store intrinsics, and generic vector load/stores, to merge
9551 /// base address updates.
9552 /// For generic load/stores, the memory type is assumed to be a vector.
9553 /// The caller is assumed to have checked legality.
9554 static SDValue CombineBaseUpdate(SDNode *N,
9555 TargetLowering::DAGCombinerInfo &DCI) {
9556 SelectionDAG &DAG = DCI.DAG;
9557 const bool isIntrinsic = (N->getOpcode() == ISD::INTRINSIC_VOID ||
9558 N->getOpcode() == ISD::INTRINSIC_W_CHAIN);
9559 const bool isStore = N->getOpcode() == ISD::STORE;
9560 const unsigned AddrOpIdx = ((isIntrinsic || isStore) ? 2 : 1);
9561 SDValue Addr = N->getOperand(AddrOpIdx);
9562 MemSDNode *MemN = cast<MemSDNode>(N);
9565 // Search for a use of the address operand that is an increment.
9566 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(),
9567 UE = Addr.getNode()->use_end(); UI != UE; ++UI) {
9569 if (User->getOpcode() != ISD::ADD ||
9570 UI.getUse().getResNo() != Addr.getResNo())
9573 // Check that the add is independent of the load/store. Otherwise, folding
9574 // it would create a cycle.
9575 if (User->isPredecessorOf(N) || N->isPredecessorOf(User))
9578 // Find the new opcode for the updating load/store.
9579 bool isLoadOp = true;
9580 bool isLaneOp = false;
9581 unsigned NewOpc = 0;
9582 unsigned NumVecs = 0;
9584 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
9586 default: llvm_unreachable("unexpected intrinsic for Neon base update");
9587 case Intrinsic::arm_neon_vld1: NewOpc = ARMISD::VLD1_UPD;
9589 case Intrinsic::arm_neon_vld2: NewOpc = ARMISD::VLD2_UPD;
9591 case Intrinsic::arm_neon_vld3: NewOpc = ARMISD::VLD3_UPD;
9593 case Intrinsic::arm_neon_vld4: NewOpc = ARMISD::VLD4_UPD;
9595 case Intrinsic::arm_neon_vld2lane: NewOpc = ARMISD::VLD2LN_UPD;
9596 NumVecs = 2; isLaneOp = true; break;
9597 case Intrinsic::arm_neon_vld3lane: NewOpc = ARMISD::VLD3LN_UPD;
9598 NumVecs = 3; isLaneOp = true; break;
9599 case Intrinsic::arm_neon_vld4lane: NewOpc = ARMISD::VLD4LN_UPD;
9600 NumVecs = 4; isLaneOp = true; break;
9601 case Intrinsic::arm_neon_vst1: NewOpc = ARMISD::VST1_UPD;
9602 NumVecs = 1; isLoadOp = false; break;
9603 case Intrinsic::arm_neon_vst2: NewOpc = ARMISD::VST2_UPD;
9604 NumVecs = 2; isLoadOp = false; break;
9605 case Intrinsic::arm_neon_vst3: NewOpc = ARMISD::VST3_UPD;
9606 NumVecs = 3; isLoadOp = false; break;
9607 case Intrinsic::arm_neon_vst4: NewOpc = ARMISD::VST4_UPD;
9608 NumVecs = 4; isLoadOp = false; break;
9609 case Intrinsic::arm_neon_vst2lane: NewOpc = ARMISD::VST2LN_UPD;
9610 NumVecs = 2; isLoadOp = false; isLaneOp = true; break;
9611 case Intrinsic::arm_neon_vst3lane: NewOpc = ARMISD::VST3LN_UPD;
9612 NumVecs = 3; isLoadOp = false; isLaneOp = true; break;
9613 case Intrinsic::arm_neon_vst4lane: NewOpc = ARMISD::VST4LN_UPD;
9614 NumVecs = 4; isLoadOp = false; isLaneOp = true; break;
9618 switch (N->getOpcode()) {
9619 default: llvm_unreachable("unexpected opcode for Neon base update");
9620 case ARMISD::VLD2DUP: NewOpc = ARMISD::VLD2DUP_UPD; NumVecs = 2; break;
9621 case ARMISD::VLD3DUP: NewOpc = ARMISD::VLD3DUP_UPD; NumVecs = 3; break;
9622 case ARMISD::VLD4DUP: NewOpc = ARMISD::VLD4DUP_UPD; NumVecs = 4; break;
9623 case ISD::LOAD: NewOpc = ARMISD::VLD1_UPD;
9624 NumVecs = 1; isLaneOp = false; break;
9625 case ISD::STORE: NewOpc = ARMISD::VST1_UPD;
9626 NumVecs = 1; isLaneOp = false; isLoadOp = false; break;
9630 // Find the size of memory referenced by the load/store.
9633 VecTy = N->getValueType(0);
9634 } else if (isIntrinsic) {
9635 VecTy = N->getOperand(AddrOpIdx+1).getValueType();
9637 assert(isStore && "Node has to be a load, a store, or an intrinsic!");
9638 VecTy = N->getOperand(1).getValueType();
9641 unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8;
9643 NumBytes /= VecTy.getVectorNumElements();
9645 // If the increment is a constant, it must match the memory ref size.
9646 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
9647 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
9648 uint64_t IncVal = CInc->getZExtValue();
9649 if (IncVal != NumBytes)
9651 } else if (NumBytes >= 3 * 16) {
9652 // VLD3/4 and VST3/4 for 128-bit vectors are implemented with two
9653 // separate instructions that make it harder to use a non-constant update.
9657 // OK, we found an ADD we can fold into the base update.
9658 // Now, create a _UPD node, taking care of not breaking alignment.
9660 EVT AlignedVecTy = VecTy;
9661 unsigned Alignment = MemN->getAlignment();
9663 // If this is a less-than-standard-aligned load/store, change the type to
9664 // match the standard alignment.
9665 // The alignment is overlooked when selecting _UPD variants; and it's
9666 // easier to introduce bitcasts here than fix that.
9667 // There are 3 ways to get to this base-update combine:
9668 // - intrinsics: they are assumed to be properly aligned (to the standard
9669 // alignment of the memory type), so we don't need to do anything.
9670 // - ARMISD::VLDx nodes: they are only generated from the aforementioned
9671 // intrinsics, so, likewise, there's nothing to do.
9672 // - generic load/store instructions: the alignment is specified as an
9673 // explicit operand, rather than implicitly as the standard alignment
9674 // of the memory type (like the intrisics). We need to change the
9675 // memory type to match the explicit alignment. That way, we don't
9676 // generate non-standard-aligned ARMISD::VLDx nodes.
9677 if (isa<LSBaseSDNode>(N)) {
9680 if (Alignment < VecTy.getScalarSizeInBits() / 8) {
9681 MVT EltTy = MVT::getIntegerVT(Alignment * 8);
9682 assert(NumVecs == 1 && "Unexpected multi-element generic load/store.");
9683 assert(!isLaneOp && "Unexpected generic load/store lane.");
9684 unsigned NumElts = NumBytes / (EltTy.getSizeInBits() / 8);
9685 AlignedVecTy = MVT::getVectorVT(EltTy, NumElts);
9687 // Don't set an explicit alignment on regular load/stores that we want
9688 // to transform to VLD/VST 1_UPD nodes.
9689 // This matches the behavior of regular load/stores, which only get an
9690 // explicit alignment if the MMO alignment is larger than the standard
9691 // alignment of the memory type.
9692 // Intrinsics, however, always get an explicit alignment, set to the
9693 // alignment of the MMO.
9697 // Create the new updating load/store node.
9698 // First, create an SDVTList for the new updating node's results.
9700 unsigned NumResultVecs = (isLoadOp ? NumVecs : 0);
9702 for (n = 0; n < NumResultVecs; ++n)
9703 Tys[n] = AlignedVecTy;
9704 Tys[n++] = MVT::i32;
9705 Tys[n] = MVT::Other;
9706 SDVTList SDTys = DAG.getVTList(makeArrayRef(Tys, NumResultVecs+2));
9708 // Then, gather the new node's operands.
9709 SmallVector<SDValue, 8> Ops;
9710 Ops.push_back(N->getOperand(0)); // incoming chain
9711 Ops.push_back(N->getOperand(AddrOpIdx));
9714 if (StoreSDNode *StN = dyn_cast<StoreSDNode>(N)) {
9715 // Try to match the intrinsic's signature
9716 Ops.push_back(StN->getValue());
9718 // Loads (and of course intrinsics) match the intrinsics' signature,
9719 // so just add all but the alignment operand.
9720 for (unsigned i = AddrOpIdx + 1; i < N->getNumOperands() - 1; ++i)
9721 Ops.push_back(N->getOperand(i));
9724 // For all node types, the alignment operand is always the last one.
9725 Ops.push_back(DAG.getConstant(Alignment, dl, MVT::i32));
9727 // If this is a non-standard-aligned STORE, the penultimate operand is the
9728 // stored value. Bitcast it to the aligned type.
9729 if (AlignedVecTy != VecTy && N->getOpcode() == ISD::STORE) {
9730 SDValue &StVal = Ops[Ops.size()-2];
9731 StVal = DAG.getNode(ISD::BITCAST, dl, AlignedVecTy, StVal);
9734 SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, dl, SDTys,
9736 MemN->getMemOperand());
9739 SmallVector<SDValue, 5> NewResults;
9740 for (unsigned i = 0; i < NumResultVecs; ++i)
9741 NewResults.push_back(SDValue(UpdN.getNode(), i));
9743 // If this is an non-standard-aligned LOAD, the first result is the loaded
9744 // value. Bitcast it to the expected result type.
9745 if (AlignedVecTy != VecTy && N->getOpcode() == ISD::LOAD) {
9746 SDValue &LdVal = NewResults[0];
9747 LdVal = DAG.getNode(ISD::BITCAST, dl, VecTy, LdVal);
9750 NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs+1)); // chain
9751 DCI.CombineTo(N, NewResults);
9752 DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs));
9759 static SDValue PerformVLDCombine(SDNode *N,
9760 TargetLowering::DAGCombinerInfo &DCI) {
9761 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
9764 return CombineBaseUpdate(N, DCI);
9767 /// CombineVLDDUP - For a VDUPLANE node N, check if its source operand is a
9768 /// vldN-lane (N > 1) intrinsic, and if all the other uses of that intrinsic
9769 /// are also VDUPLANEs. If so, combine them to a vldN-dup operation and
9771 static bool CombineVLDDUP(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
9772 SelectionDAG &DAG = DCI.DAG;
9773 EVT VT = N->getValueType(0);
9774 // vldN-dup instructions only support 64-bit vectors for N > 1.
9775 if (!VT.is64BitVector())
9778 // Check if the VDUPLANE operand is a vldN-dup intrinsic.
9779 SDNode *VLD = N->getOperand(0).getNode();
9780 if (VLD->getOpcode() != ISD::INTRINSIC_W_CHAIN)
9782 unsigned NumVecs = 0;
9783 unsigned NewOpc = 0;
9784 unsigned IntNo = cast<ConstantSDNode>(VLD->getOperand(1))->getZExtValue();
9785 if (IntNo == Intrinsic::arm_neon_vld2lane) {
9787 NewOpc = ARMISD::VLD2DUP;
9788 } else if (IntNo == Intrinsic::arm_neon_vld3lane) {
9790 NewOpc = ARMISD::VLD3DUP;
9791 } else if (IntNo == Intrinsic::arm_neon_vld4lane) {
9793 NewOpc = ARMISD::VLD4DUP;
9798 // First check that all the vldN-lane uses are VDUPLANEs and that the lane
9799 // numbers match the load.
9800 unsigned VLDLaneNo =
9801 cast<ConstantSDNode>(VLD->getOperand(NumVecs+3))->getZExtValue();
9802 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
9804 // Ignore uses of the chain result.
9805 if (UI.getUse().getResNo() == NumVecs)
9808 if (User->getOpcode() != ARMISD::VDUPLANE ||
9809 VLDLaneNo != cast<ConstantSDNode>(User->getOperand(1))->getZExtValue())
9813 // Create the vldN-dup node.
9816 for (n = 0; n < NumVecs; ++n)
9818 Tys[n] = MVT::Other;
9819 SDVTList SDTys = DAG.getVTList(makeArrayRef(Tys, NumVecs+1));
9820 SDValue Ops[] = { VLD->getOperand(0), VLD->getOperand(2) };
9821 MemIntrinsicSDNode *VLDMemInt = cast<MemIntrinsicSDNode>(VLD);
9822 SDValue VLDDup = DAG.getMemIntrinsicNode(NewOpc, SDLoc(VLD), SDTys,
9823 Ops, VLDMemInt->getMemoryVT(),
9824 VLDMemInt->getMemOperand());
9827 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
9829 unsigned ResNo = UI.getUse().getResNo();
9830 // Ignore uses of the chain result.
9831 if (ResNo == NumVecs)
9834 DCI.CombineTo(User, SDValue(VLDDup.getNode(), ResNo));
9837 // Now the vldN-lane intrinsic is dead except for its chain result.
9838 // Update uses of the chain.
9839 std::vector<SDValue> VLDDupResults;
9840 for (unsigned n = 0; n < NumVecs; ++n)
9841 VLDDupResults.push_back(SDValue(VLDDup.getNode(), n));
9842 VLDDupResults.push_back(SDValue(VLDDup.getNode(), NumVecs));
9843 DCI.CombineTo(VLD, VLDDupResults);
9848 /// PerformVDUPLANECombine - Target-specific dag combine xforms for
9849 /// ARMISD::VDUPLANE.
9850 static SDValue PerformVDUPLANECombine(SDNode *N,
9851 TargetLowering::DAGCombinerInfo &DCI) {
9852 SDValue Op = N->getOperand(0);
9854 // If the source is a vldN-lane (N > 1) intrinsic, and all the other uses
9855 // of that intrinsic are also VDUPLANEs, combine them to a vldN-dup operation.
9856 if (CombineVLDDUP(N, DCI))
9857 return SDValue(N, 0);
9859 // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is
9860 // redundant. Ignore bit_converts for now; element sizes are checked below.
9861 while (Op.getOpcode() == ISD::BITCAST)
9862 Op = Op.getOperand(0);
9863 if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM)
9866 // Make sure the VMOV element size is not bigger than the VDUPLANE elements.
9867 unsigned EltSize = Op.getValueType().getVectorElementType().getSizeInBits();
9868 // The canonical VMOV for a zero vector uses a 32-bit element size.
9869 unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
9871 if (ARM_AM::decodeNEONModImm(Imm, EltBits) == 0)
9873 EVT VT = N->getValueType(0);
9874 if (EltSize > VT.getVectorElementType().getSizeInBits())
9877 return DCI.DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Op);
9880 static SDValue PerformLOADCombine(SDNode *N,
9881 TargetLowering::DAGCombinerInfo &DCI) {
9882 EVT VT = N->getValueType(0);
9884 // If this is a legal vector load, try to combine it into a VLD1_UPD.
9885 if (ISD::isNormalLoad(N) && VT.isVector() &&
9886 DCI.DAG.getTargetLoweringInfo().isTypeLegal(VT))
9887 return CombineBaseUpdate(N, DCI);
9892 /// PerformSTORECombine - Target-specific dag combine xforms for
9894 static SDValue PerformSTORECombine(SDNode *N,
9895 TargetLowering::DAGCombinerInfo &DCI) {
9896 StoreSDNode *St = cast<StoreSDNode>(N);
9897 if (St->isVolatile())
9900 // Optimize trunc store (of multiple scalars) to shuffle and store. First,
9901 // pack all of the elements in one place. Next, store to memory in fewer
9903 SDValue StVal = St->getValue();
9904 EVT VT = StVal.getValueType();
9905 if (St->isTruncatingStore() && VT.isVector()) {
9906 SelectionDAG &DAG = DCI.DAG;
9907 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9908 EVT StVT = St->getMemoryVT();
9909 unsigned NumElems = VT.getVectorNumElements();
9910 assert(StVT != VT && "Cannot truncate to the same type");
9911 unsigned FromEltSz = VT.getVectorElementType().getSizeInBits();
9912 unsigned ToEltSz = StVT.getVectorElementType().getSizeInBits();
9914 // From, To sizes and ElemCount must be pow of two
9915 if (!isPowerOf2_32(NumElems * FromEltSz * ToEltSz)) return SDValue();
9917 // We are going to use the original vector elt for storing.
9918 // Accumulated smaller vector elements must be a multiple of the store size.
9919 if (0 != (NumElems * FromEltSz) % ToEltSz) return SDValue();
9921 unsigned SizeRatio = FromEltSz / ToEltSz;
9922 assert(SizeRatio * NumElems * ToEltSz == VT.getSizeInBits());
9924 // Create a type on which we perform the shuffle.
9925 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(), StVT.getScalarType(),
9926 NumElems*SizeRatio);
9927 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
9930 SDValue WideVec = DAG.getNode(ISD::BITCAST, DL, WideVecVT, StVal);
9931 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
9932 for (unsigned i = 0; i < NumElems; ++i)
9933 ShuffleVec[i] = DAG.getDataLayout().isBigEndian()
9934 ? (i + 1) * SizeRatio - 1
9937 // Can't shuffle using an illegal type.
9938 if (!TLI.isTypeLegal(WideVecVT)) return SDValue();
9940 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, DL, WideVec,
9941 DAG.getUNDEF(WideVec.getValueType()),
9943 // At this point all of the data is stored at the bottom of the
9944 // register. We now need to save it to mem.
9946 // Find the largest store unit
9947 MVT StoreType = MVT::i8;
9948 for (MVT Tp : MVT::integer_valuetypes()) {
9949 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToEltSz)
9952 // Didn't find a legal store type.
9953 if (!TLI.isTypeLegal(StoreType))
9956 // Bitcast the original vector into a vector of store-size units
9957 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
9958 StoreType, VT.getSizeInBits()/EVT(StoreType).getSizeInBits());
9959 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
9960 SDValue ShuffWide = DAG.getNode(ISD::BITCAST, DL, StoreVecVT, Shuff);
9961 SmallVector<SDValue, 8> Chains;
9962 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits() / 8, DL,
9963 TLI.getPointerTy(DAG.getDataLayout()));
9964 SDValue BasePtr = St->getBasePtr();
9966 // Perform one or more big stores into memory.
9967 unsigned E = (ToEltSz*NumElems)/StoreType.getSizeInBits();
9968 for (unsigned I = 0; I < E; I++) {
9969 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
9970 StoreType, ShuffWide,
9971 DAG.getIntPtrConstant(I, DL));
9972 SDValue Ch = DAG.getStore(St->getChain(), DL, SubVec, BasePtr,
9973 St->getPointerInfo(), St->isVolatile(),
9974 St->isNonTemporal(), St->getAlignment());
9975 BasePtr = DAG.getNode(ISD::ADD, DL, BasePtr.getValueType(), BasePtr,
9977 Chains.push_back(Ch);
9979 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
9982 if (!ISD::isNormalStore(St))
9985 // Split a store of a VMOVDRR into two integer stores to avoid mixing NEON and
9986 // ARM stores of arguments in the same cache line.
9987 if (StVal.getNode()->getOpcode() == ARMISD::VMOVDRR &&
9988 StVal.getNode()->hasOneUse()) {
9989 SelectionDAG &DAG = DCI.DAG;
9990 bool isBigEndian = DAG.getDataLayout().isBigEndian();
9992 SDValue BasePtr = St->getBasePtr();
9993 SDValue NewST1 = DAG.getStore(St->getChain(), DL,
9994 StVal.getNode()->getOperand(isBigEndian ? 1 : 0 ),
9995 BasePtr, St->getPointerInfo(), St->isVolatile(),
9996 St->isNonTemporal(), St->getAlignment());
9998 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
9999 DAG.getConstant(4, DL, MVT::i32));
10000 return DAG.getStore(NewST1.getValue(0), DL,
10001 StVal.getNode()->getOperand(isBigEndian ? 0 : 1),
10002 OffsetPtr, St->getPointerInfo(), St->isVolatile(),
10003 St->isNonTemporal(),
10004 std::min(4U, St->getAlignment() / 2));
10007 if (StVal.getValueType() == MVT::i64 &&
10008 StVal.getNode()->getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
10010 // Bitcast an i64 store extracted from a vector to f64.
10011 // Otherwise, the i64 value will be legalized to a pair of i32 values.
10012 SelectionDAG &DAG = DCI.DAG;
10014 SDValue IntVec = StVal.getOperand(0);
10015 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
10016 IntVec.getValueType().getVectorNumElements());
10017 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, IntVec);
10018 SDValue ExtElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
10019 Vec, StVal.getOperand(1));
10021 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ExtElt);
10022 // Make the DAGCombiner fold the bitcasts.
10023 DCI.AddToWorklist(Vec.getNode());
10024 DCI.AddToWorklist(ExtElt.getNode());
10025 DCI.AddToWorklist(V.getNode());
10026 return DAG.getStore(St->getChain(), dl, V, St->getBasePtr(),
10027 St->getPointerInfo(), St->isVolatile(),
10028 St->isNonTemporal(), St->getAlignment(),
10032 // If this is a legal vector store, try to combine it into a VST1_UPD.
10033 if (ISD::isNormalStore(N) && VT.isVector() &&
10034 DCI.DAG.getTargetLoweringInfo().isTypeLegal(VT))
10035 return CombineBaseUpdate(N, DCI);
10040 /// PerformVCVTCombine - VCVT (floating-point to fixed-point, Advanced SIMD)
10041 /// can replace combinations of VMUL and VCVT (floating-point to integer)
10042 /// when the VMUL has a constant operand that is a power of 2.
10044 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
10045 /// vmul.f32 d16, d17, d16
10046 /// vcvt.s32.f32 d16, d16
10048 /// vcvt.s32.f32 d16, d16, #3
10049 static SDValue PerformVCVTCombine(SDNode *N, SelectionDAG &DAG,
10050 const ARMSubtarget *Subtarget) {
10051 if (!Subtarget->hasNEON())
10054 SDValue Op = N->getOperand(0);
10055 if (!Op.getValueType().isVector() || Op.getOpcode() != ISD::FMUL)
10058 SDValue ConstVec = Op->getOperand(1);
10059 if (!isa<BuildVectorSDNode>(ConstVec))
10062 MVT FloatTy = Op.getSimpleValueType().getVectorElementType();
10063 uint32_t FloatBits = FloatTy.getSizeInBits();
10064 MVT IntTy = N->getSimpleValueType(0).getVectorElementType();
10065 uint32_t IntBits = IntTy.getSizeInBits();
10066 unsigned NumLanes = Op.getValueType().getVectorNumElements();
10067 if (FloatBits != 32 || IntBits > 32 || NumLanes > 4) {
10068 // These instructions only exist converting from f32 to i32. We can handle
10069 // smaller integers by generating an extra truncate, but larger ones would
10070 // be lossy. We also can't handle more then 4 lanes, since these intructions
10071 // only support v2i32/v4i32 types.
10075 BitVector UndefElements;
10076 BuildVectorSDNode *BV = cast<BuildVectorSDNode>(ConstVec);
10077 int32_t C = BV->getConstantFPSplatPow2ToLog2Int(&UndefElements, 33);
10078 if (C == -1 || C == 0 || C > 32)
10082 bool isSigned = N->getOpcode() == ISD::FP_TO_SINT;
10083 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfp2fxs :
10084 Intrinsic::arm_neon_vcvtfp2fxu;
10085 SDValue FixConv = DAG.getNode(
10086 ISD::INTRINSIC_WO_CHAIN, dl, NumLanes == 2 ? MVT::v2i32 : MVT::v4i32,
10087 DAG.getConstant(IntrinsicOpcode, dl, MVT::i32), Op->getOperand(0),
10088 DAG.getConstant(C, dl, MVT::i32));
10090 if (IntBits < FloatBits)
10091 FixConv = DAG.getNode(ISD::TRUNCATE, dl, N->getValueType(0), FixConv);
10096 /// PerformVDIVCombine - VCVT (fixed-point to floating-point, Advanced SIMD)
10097 /// can replace combinations of VCVT (integer to floating-point) and VDIV
10098 /// when the VDIV has a constant operand that is a power of 2.
10100 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
10101 /// vcvt.f32.s32 d16, d16
10102 /// vdiv.f32 d16, d17, d16
10104 /// vcvt.f32.s32 d16, d16, #3
10105 static SDValue PerformVDIVCombine(SDNode *N, SelectionDAG &DAG,
10106 const ARMSubtarget *Subtarget) {
10107 if (!Subtarget->hasNEON())
10110 SDValue Op = N->getOperand(0);
10111 unsigned OpOpcode = Op.getNode()->getOpcode();
10112 if (!N->getValueType(0).isVector() ||
10113 (OpOpcode != ISD::SINT_TO_FP && OpOpcode != ISD::UINT_TO_FP))
10116 SDValue ConstVec = N->getOperand(1);
10117 if (!isa<BuildVectorSDNode>(ConstVec))
10120 MVT FloatTy = N->getSimpleValueType(0).getVectorElementType();
10121 uint32_t FloatBits = FloatTy.getSizeInBits();
10122 MVT IntTy = Op.getOperand(0).getSimpleValueType().getVectorElementType();
10123 uint32_t IntBits = IntTy.getSizeInBits();
10124 unsigned NumLanes = Op.getValueType().getVectorNumElements();
10125 if (FloatBits != 32 || IntBits > 32 || NumLanes > 4) {
10126 // These instructions only exist converting from i32 to f32. We can handle
10127 // smaller integers by generating an extra extend, but larger ones would
10128 // be lossy. We also can't handle more then 4 lanes, since these intructions
10129 // only support v2i32/v4i32 types.
10133 BitVector UndefElements;
10134 BuildVectorSDNode *BV = cast<BuildVectorSDNode>(ConstVec);
10135 int32_t C = BV->getConstantFPSplatPow2ToLog2Int(&UndefElements, 33);
10136 if (C == -1 || C == 0 || C > 32)
10140 bool isSigned = OpOpcode == ISD::SINT_TO_FP;
10141 SDValue ConvInput = Op.getOperand(0);
10142 if (IntBits < FloatBits)
10143 ConvInput = DAG.getNode(isSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND,
10144 dl, NumLanes == 2 ? MVT::v2i32 : MVT::v4i32,
10147 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfxs2fp :
10148 Intrinsic::arm_neon_vcvtfxu2fp;
10149 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl,
10151 DAG.getConstant(IntrinsicOpcode, dl, MVT::i32),
10152 ConvInput, DAG.getConstant(C, dl, MVT::i32));
10155 /// Getvshiftimm - Check if this is a valid build_vector for the immediate
10156 /// operand of a vector shift operation, where all the elements of the
10157 /// build_vector must have the same constant integer value.
10158 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
10159 // Ignore bit_converts.
10160 while (Op.getOpcode() == ISD::BITCAST)
10161 Op = Op.getOperand(0);
10162 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
10163 APInt SplatBits, SplatUndef;
10164 unsigned SplatBitSize;
10166 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
10167 HasAnyUndefs, ElementBits) ||
10168 SplatBitSize > ElementBits)
10170 Cnt = SplatBits.getSExtValue();
10174 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
10175 /// operand of a vector shift left operation. That value must be in the range:
10176 /// 0 <= Value < ElementBits for a left shift; or
10177 /// 0 <= Value <= ElementBits for a long left shift.
10178 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
10179 assert(VT.isVector() && "vector shift count is not a vector type");
10180 int64_t ElementBits = VT.getVectorElementType().getSizeInBits();
10181 if (! getVShiftImm(Op, ElementBits, Cnt))
10183 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
10186 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
10187 /// operand of a vector shift right operation. For a shift opcode, the value
10188 /// is positive, but for an intrinsic the value count must be negative. The
10189 /// absolute value must be in the range:
10190 /// 1 <= |Value| <= ElementBits for a right shift; or
10191 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
10192 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
10194 assert(VT.isVector() && "vector shift count is not a vector type");
10195 int64_t ElementBits = VT.getVectorElementType().getSizeInBits();
10196 if (! getVShiftImm(Op, ElementBits, Cnt))
10199 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
10200 if (Cnt >= -(isNarrow ? ElementBits/2 : ElementBits) && Cnt <= -1) {
10207 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
10208 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
10209 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
10212 // Don't do anything for most intrinsics.
10215 // Vector shifts: check for immediate versions and lower them.
10216 // Note: This is done during DAG combining instead of DAG legalizing because
10217 // the build_vectors for 64-bit vector element shift counts are generally
10218 // not legal, and it is hard to see their values after they get legalized to
10219 // loads from a constant pool.
10220 case Intrinsic::arm_neon_vshifts:
10221 case Intrinsic::arm_neon_vshiftu:
10222 case Intrinsic::arm_neon_vrshifts:
10223 case Intrinsic::arm_neon_vrshiftu:
10224 case Intrinsic::arm_neon_vrshiftn:
10225 case Intrinsic::arm_neon_vqshifts:
10226 case Intrinsic::arm_neon_vqshiftu:
10227 case Intrinsic::arm_neon_vqshiftsu:
10228 case Intrinsic::arm_neon_vqshiftns:
10229 case Intrinsic::arm_neon_vqshiftnu:
10230 case Intrinsic::arm_neon_vqshiftnsu:
10231 case Intrinsic::arm_neon_vqrshiftns:
10232 case Intrinsic::arm_neon_vqrshiftnu:
10233 case Intrinsic::arm_neon_vqrshiftnsu: {
10234 EVT VT = N->getOperand(1).getValueType();
10236 unsigned VShiftOpc = 0;
10239 case Intrinsic::arm_neon_vshifts:
10240 case Intrinsic::arm_neon_vshiftu:
10241 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
10242 VShiftOpc = ARMISD::VSHL;
10245 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
10246 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
10247 ARMISD::VSHRs : ARMISD::VSHRu);
10252 case Intrinsic::arm_neon_vrshifts:
10253 case Intrinsic::arm_neon_vrshiftu:
10254 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
10258 case Intrinsic::arm_neon_vqshifts:
10259 case Intrinsic::arm_neon_vqshiftu:
10260 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
10264 case Intrinsic::arm_neon_vqshiftsu:
10265 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
10267 llvm_unreachable("invalid shift count for vqshlu intrinsic");
10269 case Intrinsic::arm_neon_vrshiftn:
10270 case Intrinsic::arm_neon_vqshiftns:
10271 case Intrinsic::arm_neon_vqshiftnu:
10272 case Intrinsic::arm_neon_vqshiftnsu:
10273 case Intrinsic::arm_neon_vqrshiftns:
10274 case Intrinsic::arm_neon_vqrshiftnu:
10275 case Intrinsic::arm_neon_vqrshiftnsu:
10276 // Narrowing shifts require an immediate right shift.
10277 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
10279 llvm_unreachable("invalid shift count for narrowing vector shift "
10283 llvm_unreachable("unhandled vector shift");
10287 case Intrinsic::arm_neon_vshifts:
10288 case Intrinsic::arm_neon_vshiftu:
10289 // Opcode already set above.
10291 case Intrinsic::arm_neon_vrshifts:
10292 VShiftOpc = ARMISD::VRSHRs; break;
10293 case Intrinsic::arm_neon_vrshiftu:
10294 VShiftOpc = ARMISD::VRSHRu; break;
10295 case Intrinsic::arm_neon_vrshiftn:
10296 VShiftOpc = ARMISD::VRSHRN; break;
10297 case Intrinsic::arm_neon_vqshifts:
10298 VShiftOpc = ARMISD::VQSHLs; break;
10299 case Intrinsic::arm_neon_vqshiftu:
10300 VShiftOpc = ARMISD::VQSHLu; break;
10301 case Intrinsic::arm_neon_vqshiftsu:
10302 VShiftOpc = ARMISD::VQSHLsu; break;
10303 case Intrinsic::arm_neon_vqshiftns:
10304 VShiftOpc = ARMISD::VQSHRNs; break;
10305 case Intrinsic::arm_neon_vqshiftnu:
10306 VShiftOpc = ARMISD::VQSHRNu; break;
10307 case Intrinsic::arm_neon_vqshiftnsu:
10308 VShiftOpc = ARMISD::VQSHRNsu; break;
10309 case Intrinsic::arm_neon_vqrshiftns:
10310 VShiftOpc = ARMISD::VQRSHRNs; break;
10311 case Intrinsic::arm_neon_vqrshiftnu:
10312 VShiftOpc = ARMISD::VQRSHRNu; break;
10313 case Intrinsic::arm_neon_vqrshiftnsu:
10314 VShiftOpc = ARMISD::VQRSHRNsu; break;
10318 return DAG.getNode(VShiftOpc, dl, N->getValueType(0),
10319 N->getOperand(1), DAG.getConstant(Cnt, dl, MVT::i32));
10322 case Intrinsic::arm_neon_vshiftins: {
10323 EVT VT = N->getOperand(1).getValueType();
10325 unsigned VShiftOpc = 0;
10327 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
10328 VShiftOpc = ARMISD::VSLI;
10329 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
10330 VShiftOpc = ARMISD::VSRI;
10332 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
10336 return DAG.getNode(VShiftOpc, dl, N->getValueType(0),
10337 N->getOperand(1), N->getOperand(2),
10338 DAG.getConstant(Cnt, dl, MVT::i32));
10341 case Intrinsic::arm_neon_vqrshifts:
10342 case Intrinsic::arm_neon_vqrshiftu:
10343 // No immediate versions of these to check for.
10350 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
10351 /// lowers them. As with the vector shift intrinsics, this is done during DAG
10352 /// combining instead of DAG legalizing because the build_vectors for 64-bit
10353 /// vector element shift counts are generally not legal, and it is hard to see
10354 /// their values after they get legalized to loads from a constant pool.
10355 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
10356 const ARMSubtarget *ST) {
10357 EVT VT = N->getValueType(0);
10358 if (N->getOpcode() == ISD::SRL && VT == MVT::i32 && ST->hasV6Ops()) {
10359 // Canonicalize (srl (bswap x), 16) to (rotr (bswap x), 16) if the high
10360 // 16-bits of x is zero. This optimizes rev + lsr 16 to rev16.
10361 SDValue N1 = N->getOperand(1);
10362 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) {
10363 SDValue N0 = N->getOperand(0);
10364 if (C->getZExtValue() == 16 && N0.getOpcode() == ISD::BSWAP &&
10365 DAG.MaskedValueIsZero(N0.getOperand(0),
10366 APInt::getHighBitsSet(32, 16)))
10367 return DAG.getNode(ISD::ROTR, SDLoc(N), VT, N0, N1);
10371 // Nothing to be done for scalar shifts.
10372 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
10373 if (!VT.isVector() || !TLI.isTypeLegal(VT))
10376 assert(ST->hasNEON() && "unexpected vector shift");
10379 switch (N->getOpcode()) {
10380 default: llvm_unreachable("unexpected shift opcode");
10383 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt)) {
10385 return DAG.getNode(ARMISD::VSHL, dl, VT, N->getOperand(0),
10386 DAG.getConstant(Cnt, dl, MVT::i32));
10392 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
10393 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
10394 ARMISD::VSHRs : ARMISD::VSHRu);
10396 return DAG.getNode(VShiftOpc, dl, VT, N->getOperand(0),
10397 DAG.getConstant(Cnt, dl, MVT::i32));
10403 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
10404 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
10405 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
10406 const ARMSubtarget *ST) {
10407 SDValue N0 = N->getOperand(0);
10409 // Check for sign- and zero-extensions of vector extract operations of 8-
10410 // and 16-bit vector elements. NEON supports these directly. They are
10411 // handled during DAG combining because type legalization will promote them
10412 // to 32-bit types and it is messy to recognize the operations after that.
10413 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
10414 SDValue Vec = N0.getOperand(0);
10415 SDValue Lane = N0.getOperand(1);
10416 EVT VT = N->getValueType(0);
10417 EVT EltVT = N0.getValueType();
10418 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
10420 if (VT == MVT::i32 &&
10421 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
10422 TLI.isTypeLegal(Vec.getValueType()) &&
10423 isa<ConstantSDNode>(Lane)) {
10426 switch (N->getOpcode()) {
10427 default: llvm_unreachable("unexpected opcode");
10428 case ISD::SIGN_EXTEND:
10429 Opc = ARMISD::VGETLANEs;
10431 case ISD::ZERO_EXTEND:
10432 case ISD::ANY_EXTEND:
10433 Opc = ARMISD::VGETLANEu;
10436 return DAG.getNode(Opc, SDLoc(N), VT, Vec, Lane);
10443 static void computeKnownBits(SelectionDAG &DAG, SDValue Op, APInt &KnownZero,
10445 if (Op.getOpcode() == ARMISD::BFI) {
10446 // Conservatively, we can recurse down the first operand
10447 // and just mask out all affected bits.
10448 computeKnownBits(DAG, Op.getOperand(0), KnownZero, KnownOne);
10450 // The operand to BFI is already a mask suitable for removing the bits it
10452 ConstantSDNode *CI = cast<ConstantSDNode>(Op.getOperand(2));
10453 APInt Mask = CI->getAPIntValue();
10458 if (Op.getOpcode() == ARMISD::CMOV) {
10459 APInt KZ2(KnownZero.getBitWidth(), 0);
10460 APInt KO2(KnownOne.getBitWidth(), 0);
10461 computeKnownBits(DAG, Op.getOperand(1), KnownZero, KnownOne);
10462 computeKnownBits(DAG, Op.getOperand(2), KZ2, KO2);
10468 return DAG.computeKnownBits(Op, KnownZero, KnownOne);
10471 SDValue ARMTargetLowering::PerformCMOVToBFICombine(SDNode *CMOV, SelectionDAG &DAG) const {
10472 // If we have a CMOV, OR and AND combination such as:
10477 // * CN is a single bit;
10478 // * All bits covered by CM are known zero in y
10480 // Then we can convert this into a sequence of BFI instructions. This will
10481 // always be a win if CM is a single bit, will always be no worse than the
10482 // TST&OR sequence if CM is two bits, and for thumb will be no worse if CM is
10483 // three bits (due to the extra IT instruction).
10485 SDValue Op0 = CMOV->getOperand(0);
10486 SDValue Op1 = CMOV->getOperand(1);
10487 auto CCNode = cast<ConstantSDNode>(CMOV->getOperand(2));
10488 auto CC = CCNode->getAPIntValue().getLimitedValue();
10489 SDValue CmpZ = CMOV->getOperand(4);
10491 // The compare must be against zero.
10492 if (!isNullConstant(CmpZ->getOperand(1)))
10495 assert(CmpZ->getOpcode() == ARMISD::CMPZ);
10496 SDValue And = CmpZ->getOperand(0);
10497 if (And->getOpcode() != ISD::AND)
10499 ConstantSDNode *AndC = dyn_cast<ConstantSDNode>(And->getOperand(1));
10500 if (!AndC || !AndC->getAPIntValue().isPowerOf2())
10502 SDValue X = And->getOperand(0);
10504 if (CC == ARMCC::EQ) {
10505 // We're performing an "equal to zero" compare. Swap the operands so we
10506 // canonicalize on a "not equal to zero" compare.
10507 std::swap(Op0, Op1);
10509 assert(CC == ARMCC::NE && "How can a CMPZ node not be EQ or NE?");
10512 if (Op1->getOpcode() != ISD::OR)
10515 ConstantSDNode *OrC = dyn_cast<ConstantSDNode>(Op1->getOperand(1));
10518 SDValue Y = Op1->getOperand(0);
10523 // Now, is it profitable to continue?
10524 APInt OrCI = OrC->getAPIntValue();
10525 unsigned Heuristic = Subtarget->isThumb() ? 3 : 2;
10526 if (OrCI.countPopulation() > Heuristic)
10529 // Lastly, can we determine that the bits defined by OrCI
10531 APInt KnownZero, KnownOne;
10532 computeKnownBits(DAG, Y, KnownZero, KnownOne);
10533 if ((OrCI & KnownZero) != OrCI)
10536 // OK, we can do the combine.
10539 EVT VT = X.getValueType();
10540 unsigned BitInX = AndC->getAPIntValue().logBase2();
10543 // We must shift X first.
10544 X = DAG.getNode(ISD::SRL, dl, VT, X,
10545 DAG.getConstant(BitInX, dl, VT));
10548 for (unsigned BitInY = 0, NumActiveBits = OrCI.getActiveBits();
10549 BitInY < NumActiveBits; ++BitInY) {
10550 if (OrCI[BitInY] == 0)
10552 APInt Mask(VT.getSizeInBits(), 0);
10553 Mask.setBit(BitInY);
10554 V = DAG.getNode(ARMISD::BFI, dl, VT, V, X,
10555 // Confusingly, the operand is an *inverted* mask.
10556 DAG.getConstant(~Mask, dl, VT));
10562 /// PerformCMOVCombine - Target-specific DAG combining for ARMISD::CMOV.
10564 ARMTargetLowering::PerformCMOVCombine(SDNode *N, SelectionDAG &DAG) const {
10565 SDValue Cmp = N->getOperand(4);
10566 if (Cmp.getOpcode() != ARMISD::CMPZ)
10567 // Only looking at EQ and NE cases.
10570 EVT VT = N->getValueType(0);
10572 SDValue LHS = Cmp.getOperand(0);
10573 SDValue RHS = Cmp.getOperand(1);
10574 SDValue FalseVal = N->getOperand(0);
10575 SDValue TrueVal = N->getOperand(1);
10576 SDValue ARMcc = N->getOperand(2);
10577 ARMCC::CondCodes CC =
10578 (ARMCC::CondCodes)cast<ConstantSDNode>(ARMcc)->getZExtValue();
10580 // BFI is only available on V6T2+.
10581 if (!Subtarget->isThumb1Only() && Subtarget->hasV6T2Ops()) {
10582 SDValue R = PerformCMOVToBFICombine(N, DAG);
10603 /// FIXME: Turn this into a target neutral optimization?
10605 if (CC == ARMCC::NE && FalseVal == RHS && FalseVal != LHS) {
10606 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, TrueVal, ARMcc,
10607 N->getOperand(3), Cmp);
10608 } else if (CC == ARMCC::EQ && TrueVal == RHS) {
10610 SDValue NewCmp = getARMCmp(LHS, RHS, ISD::SETNE, ARMcc, DAG, dl);
10611 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, FalseVal, ARMcc,
10612 N->getOperand(3), NewCmp);
10615 if (Res.getNode()) {
10616 APInt KnownZero, KnownOne;
10617 DAG.computeKnownBits(SDValue(N,0), KnownZero, KnownOne);
10618 // Capture demanded bits information that would be otherwise lost.
10619 if (KnownZero == 0xfffffffe)
10620 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
10621 DAG.getValueType(MVT::i1));
10622 else if (KnownZero == 0xffffff00)
10623 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
10624 DAG.getValueType(MVT::i8));
10625 else if (KnownZero == 0xffff0000)
10626 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
10627 DAG.getValueType(MVT::i16));
10633 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
10634 DAGCombinerInfo &DCI) const {
10635 switch (N->getOpcode()) {
10637 case ISD::ADDC: return PerformADDCCombine(N, DCI, Subtarget);
10638 case ISD::ADD: return PerformADDCombine(N, DCI, Subtarget);
10639 case ISD::SUB: return PerformSUBCombine(N, DCI);
10640 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
10641 case ISD::OR: return PerformORCombine(N, DCI, Subtarget);
10642 case ISD::XOR: return PerformXORCombine(N, DCI, Subtarget);
10643 case ISD::AND: return PerformANDCombine(N, DCI, Subtarget);
10644 case ARMISD::BFI: return PerformBFICombine(N, DCI);
10645 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI, Subtarget);
10646 case ARMISD::VMOVDRR: return PerformVMOVDRRCombine(N, DCI.DAG);
10647 case ISD::STORE: return PerformSTORECombine(N, DCI);
10648 case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DCI, Subtarget);
10649 case ISD::INSERT_VECTOR_ELT: return PerformInsertEltCombine(N, DCI);
10650 case ISD::VECTOR_SHUFFLE: return PerformVECTOR_SHUFFLECombine(N, DCI.DAG);
10651 case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI);
10652 case ISD::FP_TO_SINT:
10653 case ISD::FP_TO_UINT:
10654 return PerformVCVTCombine(N, DCI.DAG, Subtarget);
10656 return PerformVDIVCombine(N, DCI.DAG, Subtarget);
10657 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
10660 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
10661 case ISD::SIGN_EXTEND:
10662 case ISD::ZERO_EXTEND:
10663 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
10664 case ARMISD::CMOV: return PerformCMOVCombine(N, DCI.DAG);
10665 case ISD::LOAD: return PerformLOADCombine(N, DCI);
10666 case ARMISD::VLD2DUP:
10667 case ARMISD::VLD3DUP:
10668 case ARMISD::VLD4DUP:
10669 return PerformVLDCombine(N, DCI);
10670 case ARMISD::BUILD_VECTOR:
10671 return PerformARMBUILD_VECTORCombine(N, DCI);
10672 case ISD::INTRINSIC_VOID:
10673 case ISD::INTRINSIC_W_CHAIN:
10674 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
10675 case Intrinsic::arm_neon_vld1:
10676 case Intrinsic::arm_neon_vld2:
10677 case Intrinsic::arm_neon_vld3:
10678 case Intrinsic::arm_neon_vld4:
10679 case Intrinsic::arm_neon_vld2lane:
10680 case Intrinsic::arm_neon_vld3lane:
10681 case Intrinsic::arm_neon_vld4lane:
10682 case Intrinsic::arm_neon_vst1:
10683 case Intrinsic::arm_neon_vst2:
10684 case Intrinsic::arm_neon_vst3:
10685 case Intrinsic::arm_neon_vst4:
10686 case Intrinsic::arm_neon_vst2lane:
10687 case Intrinsic::arm_neon_vst3lane:
10688 case Intrinsic::arm_neon_vst4lane:
10689 return PerformVLDCombine(N, DCI);
10697 bool ARMTargetLowering::isDesirableToTransformToIntegerOp(unsigned Opc,
10699 return (VT == MVT::f32) && (Opc == ISD::LOAD || Opc == ISD::STORE);
10702 bool ARMTargetLowering::allowsMisalignedMemoryAccesses(EVT VT,
10705 bool *Fast) const {
10706 // The AllowsUnaliged flag models the SCTLR.A setting in ARM cpus
10707 bool AllowsUnaligned = Subtarget->allowsUnalignedMem();
10709 switch (VT.getSimpleVT().SimpleTy) {
10715 // Unaligned access can use (for example) LRDB, LRDH, LDR
10716 if (AllowsUnaligned) {
10718 *Fast = Subtarget->hasV7Ops();
10725 // For any little-endian targets with neon, we can support unaligned ld/st
10726 // of D and Q (e.g. {D0,D1}) registers by using vld1.i8/vst1.i8.
10727 // A big-endian target may also explicitly support unaligned accesses
10728 if (Subtarget->hasNEON() && (AllowsUnaligned || Subtarget->isLittle())) {
10738 static bool memOpAlign(unsigned DstAlign, unsigned SrcAlign,
10739 unsigned AlignCheck) {
10740 return ((SrcAlign == 0 || SrcAlign % AlignCheck == 0) &&
10741 (DstAlign == 0 || DstAlign % AlignCheck == 0));
10744 EVT ARMTargetLowering::getOptimalMemOpType(uint64_t Size,
10745 unsigned DstAlign, unsigned SrcAlign,
10746 bool IsMemset, bool ZeroMemset,
10748 MachineFunction &MF) const {
10749 const Function *F = MF.getFunction();
10751 // See if we can use NEON instructions for this...
10752 if ((!IsMemset || ZeroMemset) && Subtarget->hasNEON() &&
10753 !F->hasFnAttribute(Attribute::NoImplicitFloat)) {
10756 (memOpAlign(SrcAlign, DstAlign, 16) ||
10757 (allowsMisalignedMemoryAccesses(MVT::v2f64, 0, 1, &Fast) && Fast))) {
10759 } else if (Size >= 8 &&
10760 (memOpAlign(SrcAlign, DstAlign, 8) ||
10761 (allowsMisalignedMemoryAccesses(MVT::f64, 0, 1, &Fast) &&
10767 // Lowering to i32/i16 if the size permits.
10770 else if (Size >= 2)
10773 // Let the target-independent logic figure it out.
10777 bool ARMTargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
10778 if (Val.getOpcode() != ISD::LOAD)
10781 EVT VT1 = Val.getValueType();
10782 if (!VT1.isSimple() || !VT1.isInteger() ||
10783 !VT2.isSimple() || !VT2.isInteger())
10786 switch (VT1.getSimpleVT().SimpleTy) {
10791 // 8-bit and 16-bit loads implicitly zero-extend to 32-bits.
10798 bool ARMTargetLowering::isVectorLoadExtDesirable(SDValue ExtVal) const {
10799 EVT VT = ExtVal.getValueType();
10801 if (!isTypeLegal(VT))
10804 // Don't create a loadext if we can fold the extension into a wide/long
10806 // If there's more than one user instruction, the loadext is desirable no
10807 // matter what. There can be two uses by the same instruction.
10808 if (ExtVal->use_empty() ||
10809 !ExtVal->use_begin()->isOnlyUserOf(ExtVal.getNode()))
10812 SDNode *U = *ExtVal->use_begin();
10813 if ((U->getOpcode() == ISD::ADD || U->getOpcode() == ISD::SUB ||
10814 U->getOpcode() == ISD::SHL || U->getOpcode() == ARMISD::VSHL))
10820 bool ARMTargetLowering::allowTruncateForTailCall(Type *Ty1, Type *Ty2) const {
10821 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
10824 if (!isTypeLegal(EVT::getEVT(Ty1)))
10827 assert(Ty1->getPrimitiveSizeInBits() <= 64 && "i128 is probably not a noop");
10829 // Assuming the caller doesn't have a zeroext or signext return parameter,
10830 // truncation all the way down to i1 is valid.
10835 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
10839 unsigned Scale = 1;
10840 switch (VT.getSimpleVT().SimpleTy) {
10841 default: return false;
10856 if ((V & (Scale - 1)) != 0)
10859 return V == (V & ((1LL << 5) - 1));
10862 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
10863 const ARMSubtarget *Subtarget) {
10864 bool isNeg = false;
10870 switch (VT.getSimpleVT().SimpleTy) {
10871 default: return false;
10876 // + imm12 or - imm8
10878 return V == (V & ((1LL << 8) - 1));
10879 return V == (V & ((1LL << 12) - 1));
10882 // Same as ARM mode. FIXME: NEON?
10883 if (!Subtarget->hasVFP2())
10888 return V == (V & ((1LL << 8) - 1));
10892 /// isLegalAddressImmediate - Return true if the integer value can be used
10893 /// as the offset of the target addressing mode for load / store of the
10895 static bool isLegalAddressImmediate(int64_t V, EVT VT,
10896 const ARMSubtarget *Subtarget) {
10900 if (!VT.isSimple())
10903 if (Subtarget->isThumb1Only())
10904 return isLegalT1AddressImmediate(V, VT);
10905 else if (Subtarget->isThumb2())
10906 return isLegalT2AddressImmediate(V, VT, Subtarget);
10911 switch (VT.getSimpleVT().SimpleTy) {
10912 default: return false;
10917 return V == (V & ((1LL << 12) - 1));
10920 return V == (V & ((1LL << 8) - 1));
10923 if (!Subtarget->hasVFP2()) // FIXME: NEON?
10928 return V == (V & ((1LL << 8) - 1));
10932 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
10934 int Scale = AM.Scale;
10938 switch (VT.getSimpleVT().SimpleTy) {
10939 default: return false;
10947 Scale = Scale & ~1;
10948 return Scale == 2 || Scale == 4 || Scale == 8;
10951 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
10955 // Note, we allow "void" uses (basically, uses that aren't loads or
10956 // stores), because arm allows folding a scale into many arithmetic
10957 // operations. This should be made more precise and revisited later.
10959 // Allow r << imm, but the imm has to be a multiple of two.
10960 if (Scale & 1) return false;
10961 return isPowerOf2_32(Scale);
10965 /// isLegalAddressingMode - Return true if the addressing mode represented
10966 /// by AM is legal for this target, for a load/store of the specified type.
10967 bool ARMTargetLowering::isLegalAddressingMode(const DataLayout &DL,
10968 const AddrMode &AM, Type *Ty,
10969 unsigned AS) const {
10970 EVT VT = getValueType(DL, Ty, true);
10971 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
10974 // Can never fold addr of global into load/store.
10978 switch (AM.Scale) {
10979 case 0: // no scale reg, must be "r+i" or "r", or "i".
10982 if (Subtarget->isThumb1Only())
10986 // ARM doesn't support any R+R*scale+imm addr modes.
10990 if (!VT.isSimple())
10993 if (Subtarget->isThumb2())
10994 return isLegalT2ScaledAddressingMode(AM, VT);
10996 int Scale = AM.Scale;
10997 switch (VT.getSimpleVT().SimpleTy) {
10998 default: return false;
11002 if (Scale < 0) Scale = -Scale;
11006 return isPowerOf2_32(Scale & ~1);
11010 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
11015 // Note, we allow "void" uses (basically, uses that aren't loads or
11016 // stores), because arm allows folding a scale into many arithmetic
11017 // operations. This should be made more precise and revisited later.
11019 // Allow r << imm, but the imm has to be a multiple of two.
11020 if (Scale & 1) return false;
11021 return isPowerOf2_32(Scale);
11027 /// isLegalICmpImmediate - Return true if the specified immediate is legal
11028 /// icmp immediate, that is the target has icmp instructions which can compare
11029 /// a register against the immediate without having to materialize the
11030 /// immediate into a register.
11031 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
11032 // Thumb2 and ARM modes can use cmn for negative immediates.
11033 if (!Subtarget->isThumb())
11034 return ARM_AM::getSOImmVal(std::abs(Imm)) != -1;
11035 if (Subtarget->isThumb2())
11036 return ARM_AM::getT2SOImmVal(std::abs(Imm)) != -1;
11037 // Thumb1 doesn't have cmn, and only 8-bit immediates.
11038 return Imm >= 0 && Imm <= 255;
11041 /// isLegalAddImmediate - Return true if the specified immediate is a legal add
11042 /// *or sub* immediate, that is the target has add or sub instructions which can
11043 /// add a register with the immediate without having to materialize the
11044 /// immediate into a register.
11045 bool ARMTargetLowering::isLegalAddImmediate(int64_t Imm) const {
11046 // Same encoding for add/sub, just flip the sign.
11047 int64_t AbsImm = std::abs(Imm);
11048 if (!Subtarget->isThumb())
11049 return ARM_AM::getSOImmVal(AbsImm) != -1;
11050 if (Subtarget->isThumb2())
11051 return ARM_AM::getT2SOImmVal(AbsImm) != -1;
11052 // Thumb1 only has 8-bit unsigned immediate.
11053 return AbsImm >= 0 && AbsImm <= 255;
11056 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
11057 bool isSEXTLoad, SDValue &Base,
11058 SDValue &Offset, bool &isInc,
11059 SelectionDAG &DAG) {
11060 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
11063 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
11064 // AddressingMode 3
11065 Base = Ptr->getOperand(0);
11066 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
11067 int RHSC = (int)RHS->getZExtValue();
11068 if (RHSC < 0 && RHSC > -256) {
11069 assert(Ptr->getOpcode() == ISD::ADD);
11071 Offset = DAG.getConstant(-RHSC, SDLoc(Ptr), RHS->getValueType(0));
11075 isInc = (Ptr->getOpcode() == ISD::ADD);
11076 Offset = Ptr->getOperand(1);
11078 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
11079 // AddressingMode 2
11080 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
11081 int RHSC = (int)RHS->getZExtValue();
11082 if (RHSC < 0 && RHSC > -0x1000) {
11083 assert(Ptr->getOpcode() == ISD::ADD);
11085 Offset = DAG.getConstant(-RHSC, SDLoc(Ptr), RHS->getValueType(0));
11086 Base = Ptr->getOperand(0);
11091 if (Ptr->getOpcode() == ISD::ADD) {
11093 ARM_AM::ShiftOpc ShOpcVal=
11094 ARM_AM::getShiftOpcForNode(Ptr->getOperand(0).getOpcode());
11095 if (ShOpcVal != ARM_AM::no_shift) {
11096 Base = Ptr->getOperand(1);
11097 Offset = Ptr->getOperand(0);
11099 Base = Ptr->getOperand(0);
11100 Offset = Ptr->getOperand(1);
11105 isInc = (Ptr->getOpcode() == ISD::ADD);
11106 Base = Ptr->getOperand(0);
11107 Offset = Ptr->getOperand(1);
11111 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
11115 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
11116 bool isSEXTLoad, SDValue &Base,
11117 SDValue &Offset, bool &isInc,
11118 SelectionDAG &DAG) {
11119 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
11122 Base = Ptr->getOperand(0);
11123 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
11124 int RHSC = (int)RHS->getZExtValue();
11125 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
11126 assert(Ptr->getOpcode() == ISD::ADD);
11128 Offset = DAG.getConstant(-RHSC, SDLoc(Ptr), RHS->getValueType(0));
11130 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
11131 isInc = Ptr->getOpcode() == ISD::ADD;
11132 Offset = DAG.getConstant(RHSC, SDLoc(Ptr), RHS->getValueType(0));
11140 /// getPreIndexedAddressParts - returns true by value, base pointer and
11141 /// offset pointer and addressing mode by reference if the node's address
11142 /// can be legally represented as pre-indexed load / store address.
11144 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
11146 ISD::MemIndexedMode &AM,
11147 SelectionDAG &DAG) const {
11148 if (Subtarget->isThumb1Only())
11153 bool isSEXTLoad = false;
11154 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
11155 Ptr = LD->getBasePtr();
11156 VT = LD->getMemoryVT();
11157 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
11158 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
11159 Ptr = ST->getBasePtr();
11160 VT = ST->getMemoryVT();
11165 bool isLegal = false;
11166 if (Subtarget->isThumb2())
11167 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
11168 Offset, isInc, DAG);
11170 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
11171 Offset, isInc, DAG);
11175 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
11179 /// getPostIndexedAddressParts - returns true by value, base pointer and
11180 /// offset pointer and addressing mode by reference if this node can be
11181 /// combined with a load / store to form a post-indexed load / store.
11182 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
11185 ISD::MemIndexedMode &AM,
11186 SelectionDAG &DAG) const {
11187 if (Subtarget->isThumb1Only())
11192 bool isSEXTLoad = false;
11193 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
11194 VT = LD->getMemoryVT();
11195 Ptr = LD->getBasePtr();
11196 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
11197 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
11198 VT = ST->getMemoryVT();
11199 Ptr = ST->getBasePtr();
11204 bool isLegal = false;
11205 if (Subtarget->isThumb2())
11206 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
11209 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
11215 // Swap base ptr and offset to catch more post-index load / store when
11216 // it's legal. In Thumb2 mode, offset must be an immediate.
11217 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
11218 !Subtarget->isThumb2())
11219 std::swap(Base, Offset);
11221 // Post-indexed load / store update the base pointer.
11226 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
11230 void ARMTargetLowering::computeKnownBitsForTargetNode(const SDValue Op,
11233 const SelectionDAG &DAG,
11234 unsigned Depth) const {
11235 unsigned BitWidth = KnownOne.getBitWidth();
11236 KnownZero = KnownOne = APInt(BitWidth, 0);
11237 switch (Op.getOpcode()) {
11243 // These nodes' second result is a boolean
11244 if (Op.getResNo() == 0)
11246 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - 1);
11248 case ARMISD::CMOV: {
11249 // Bits are known zero/one if known on the LHS and RHS.
11250 DAG.computeKnownBits(Op.getOperand(0), KnownZero, KnownOne, Depth+1);
11251 if (KnownZero == 0 && KnownOne == 0) return;
11253 APInt KnownZeroRHS, KnownOneRHS;
11254 DAG.computeKnownBits(Op.getOperand(1), KnownZeroRHS, KnownOneRHS, Depth+1);
11255 KnownZero &= KnownZeroRHS;
11256 KnownOne &= KnownOneRHS;
11259 case ISD::INTRINSIC_W_CHAIN: {
11260 ConstantSDNode *CN = cast<ConstantSDNode>(Op->getOperand(1));
11261 Intrinsic::ID IntID = static_cast<Intrinsic::ID>(CN->getZExtValue());
11264 case Intrinsic::arm_ldaex:
11265 case Intrinsic::arm_ldrex: {
11266 EVT VT = cast<MemIntrinsicSDNode>(Op)->getMemoryVT();
11267 unsigned MemBits = VT.getScalarType().getSizeInBits();
11268 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - MemBits);
11276 //===----------------------------------------------------------------------===//
11277 // ARM Inline Assembly Support
11278 //===----------------------------------------------------------------------===//
11280 bool ARMTargetLowering::ExpandInlineAsm(CallInst *CI) const {
11281 // Looking for "rev" which is V6+.
11282 if (!Subtarget->hasV6Ops())
11285 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
11286 std::string AsmStr = IA->getAsmString();
11287 SmallVector<StringRef, 4> AsmPieces;
11288 SplitString(AsmStr, AsmPieces, ";\n");
11290 switch (AsmPieces.size()) {
11291 default: return false;
11293 AsmStr = AsmPieces[0];
11295 SplitString(AsmStr, AsmPieces, " \t,");
11298 if (AsmPieces.size() == 3 &&
11299 AsmPieces[0] == "rev" && AsmPieces[1] == "$0" && AsmPieces[2] == "$1" &&
11300 IA->getConstraintString().compare(0, 4, "=l,l") == 0) {
11301 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
11302 if (Ty && Ty->getBitWidth() == 32)
11303 return IntrinsicLowering::LowerToByteSwap(CI);
11311 /// getConstraintType - Given a constraint letter, return the type of
11312 /// constraint it is for this target.
11313 ARMTargetLowering::ConstraintType
11314 ARMTargetLowering::getConstraintType(StringRef Constraint) const {
11315 if (Constraint.size() == 1) {
11316 switch (Constraint[0]) {
11318 case 'l': return C_RegisterClass;
11319 case 'w': return C_RegisterClass;
11320 case 'h': return C_RegisterClass;
11321 case 'x': return C_RegisterClass;
11322 case 't': return C_RegisterClass;
11323 case 'j': return C_Other; // Constant for movw.
11324 // An address with a single base register. Due to the way we
11325 // currently handle addresses it is the same as an 'r' memory constraint.
11326 case 'Q': return C_Memory;
11328 } else if (Constraint.size() == 2) {
11329 switch (Constraint[0]) {
11331 // All 'U+' constraints are addresses.
11332 case 'U': return C_Memory;
11335 return TargetLowering::getConstraintType(Constraint);
11338 /// Examine constraint type and operand type and determine a weight value.
11339 /// This object must already have been set up with the operand type
11340 /// and the current alternative constraint selected.
11341 TargetLowering::ConstraintWeight
11342 ARMTargetLowering::getSingleConstraintMatchWeight(
11343 AsmOperandInfo &info, const char *constraint) const {
11344 ConstraintWeight weight = CW_Invalid;
11345 Value *CallOperandVal = info.CallOperandVal;
11346 // If we don't have a value, we can't do a match,
11347 // but allow it at the lowest weight.
11348 if (!CallOperandVal)
11350 Type *type = CallOperandVal->getType();
11351 // Look at the constraint type.
11352 switch (*constraint) {
11354 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
11357 if (type->isIntegerTy()) {
11358 if (Subtarget->isThumb())
11359 weight = CW_SpecificReg;
11361 weight = CW_Register;
11365 if (type->isFloatingPointTy())
11366 weight = CW_Register;
11372 typedef std::pair<unsigned, const TargetRegisterClass*> RCPair;
11373 RCPair ARMTargetLowering::getRegForInlineAsmConstraint(
11374 const TargetRegisterInfo *TRI, StringRef Constraint, MVT VT) const {
11375 if (Constraint.size() == 1) {
11376 // GCC ARM Constraint Letters
11377 switch (Constraint[0]) {
11378 case 'l': // Low regs or general regs.
11379 if (Subtarget->isThumb())
11380 return RCPair(0U, &ARM::tGPRRegClass);
11381 return RCPair(0U, &ARM::GPRRegClass);
11382 case 'h': // High regs or no regs.
11383 if (Subtarget->isThumb())
11384 return RCPair(0U, &ARM::hGPRRegClass);
11387 if (Subtarget->isThumb1Only())
11388 return RCPair(0U, &ARM::tGPRRegClass);
11389 return RCPair(0U, &ARM::GPRRegClass);
11391 if (VT == MVT::Other)
11393 if (VT == MVT::f32)
11394 return RCPair(0U, &ARM::SPRRegClass);
11395 if (VT.getSizeInBits() == 64)
11396 return RCPair(0U, &ARM::DPRRegClass);
11397 if (VT.getSizeInBits() == 128)
11398 return RCPair(0U, &ARM::QPRRegClass);
11401 if (VT == MVT::Other)
11403 if (VT == MVT::f32)
11404 return RCPair(0U, &ARM::SPR_8RegClass);
11405 if (VT.getSizeInBits() == 64)
11406 return RCPair(0U, &ARM::DPR_8RegClass);
11407 if (VT.getSizeInBits() == 128)
11408 return RCPair(0U, &ARM::QPR_8RegClass);
11411 if (VT == MVT::f32)
11412 return RCPair(0U, &ARM::SPRRegClass);
11416 if (StringRef("{cc}").equals_lower(Constraint))
11417 return std::make_pair(unsigned(ARM::CPSR), &ARM::CCRRegClass);
11419 return TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
11422 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
11423 /// vector. If it is invalid, don't add anything to Ops.
11424 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
11425 std::string &Constraint,
11426 std::vector<SDValue>&Ops,
11427 SelectionDAG &DAG) const {
11430 // Currently only support length 1 constraints.
11431 if (Constraint.length() != 1) return;
11433 char ConstraintLetter = Constraint[0];
11434 switch (ConstraintLetter) {
11437 case 'I': case 'J': case 'K': case 'L':
11438 case 'M': case 'N': case 'O':
11439 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
11443 int64_t CVal64 = C->getSExtValue();
11444 int CVal = (int) CVal64;
11445 // None of these constraints allow values larger than 32 bits. Check
11446 // that the value fits in an int.
11447 if (CVal != CVal64)
11450 switch (ConstraintLetter) {
11452 // Constant suitable for movw, must be between 0 and
11454 if (Subtarget->hasV6T2Ops())
11455 if (CVal >= 0 && CVal <= 65535)
11459 if (Subtarget->isThumb1Only()) {
11460 // This must be a constant between 0 and 255, for ADD
11462 if (CVal >= 0 && CVal <= 255)
11464 } else if (Subtarget->isThumb2()) {
11465 // A constant that can be used as an immediate value in a
11466 // data-processing instruction.
11467 if (ARM_AM::getT2SOImmVal(CVal) != -1)
11470 // A constant that can be used as an immediate value in a
11471 // data-processing instruction.
11472 if (ARM_AM::getSOImmVal(CVal) != -1)
11478 if (Subtarget->isThumb()) { // FIXME thumb2
11479 // This must be a constant between -255 and -1, for negated ADD
11480 // immediates. This can be used in GCC with an "n" modifier that
11481 // prints the negated value, for use with SUB instructions. It is
11482 // not useful otherwise but is implemented for compatibility.
11483 if (CVal >= -255 && CVal <= -1)
11486 // This must be a constant between -4095 and 4095. It is not clear
11487 // what this constraint is intended for. Implemented for
11488 // compatibility with GCC.
11489 if (CVal >= -4095 && CVal <= 4095)
11495 if (Subtarget->isThumb1Only()) {
11496 // A 32-bit value where only one byte has a nonzero value. Exclude
11497 // zero to match GCC. This constraint is used by GCC internally for
11498 // constants that can be loaded with a move/shift combination.
11499 // It is not useful otherwise but is implemented for compatibility.
11500 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
11502 } else if (Subtarget->isThumb2()) {
11503 // A constant whose bitwise inverse can be used as an immediate
11504 // value in a data-processing instruction. This can be used in GCC
11505 // with a "B" modifier that prints the inverted value, for use with
11506 // BIC and MVN instructions. It is not useful otherwise but is
11507 // implemented for compatibility.
11508 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
11511 // A constant whose bitwise inverse can be used as an immediate
11512 // value in a data-processing instruction. This can be used in GCC
11513 // with a "B" modifier that prints the inverted value, for use with
11514 // BIC and MVN instructions. It is not useful otherwise but is
11515 // implemented for compatibility.
11516 if (ARM_AM::getSOImmVal(~CVal) != -1)
11522 if (Subtarget->isThumb1Only()) {
11523 // This must be a constant between -7 and 7,
11524 // for 3-operand ADD/SUB immediate instructions.
11525 if (CVal >= -7 && CVal < 7)
11527 } else if (Subtarget->isThumb2()) {
11528 // A constant whose negation can be used as an immediate value in a
11529 // data-processing instruction. This can be used in GCC with an "n"
11530 // modifier that prints the negated value, for use with SUB
11531 // instructions. It is not useful otherwise but is implemented for
11533 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
11536 // A constant whose negation can be used as an immediate value in a
11537 // data-processing instruction. This can be used in GCC with an "n"
11538 // modifier that prints the negated value, for use with SUB
11539 // instructions. It is not useful otherwise but is implemented for
11541 if (ARM_AM::getSOImmVal(-CVal) != -1)
11547 if (Subtarget->isThumb()) { // FIXME thumb2
11548 // This must be a multiple of 4 between 0 and 1020, for
11549 // ADD sp + immediate.
11550 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
11553 // A power of two or a constant between 0 and 32. This is used in
11554 // GCC for the shift amount on shifted register operands, but it is
11555 // useful in general for any shift amounts.
11556 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
11562 if (Subtarget->isThumb()) { // FIXME thumb2
11563 // This must be a constant between 0 and 31, for shift amounts.
11564 if (CVal >= 0 && CVal <= 31)
11570 if (Subtarget->isThumb()) { // FIXME thumb2
11571 // This must be a multiple of 4 between -508 and 508, for
11572 // ADD/SUB sp = sp + immediate.
11573 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
11578 Result = DAG.getTargetConstant(CVal, SDLoc(Op), Op.getValueType());
11582 if (Result.getNode()) {
11583 Ops.push_back(Result);
11586 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
11589 static RTLIB::Libcall getDivRemLibcall(
11590 const SDNode *N, MVT::SimpleValueType SVT) {
11591 assert((N->getOpcode() == ISD::SDIVREM || N->getOpcode() == ISD::UDIVREM ||
11592 N->getOpcode() == ISD::SREM || N->getOpcode() == ISD::UREM) &&
11593 "Unhandled Opcode in getDivRemLibcall");
11594 bool isSigned = N->getOpcode() == ISD::SDIVREM ||
11595 N->getOpcode() == ISD::SREM;
11598 default: llvm_unreachable("Unexpected request for libcall!");
11599 case MVT::i8: LC = isSigned ? RTLIB::SDIVREM_I8 : RTLIB::UDIVREM_I8; break;
11600 case MVT::i16: LC = isSigned ? RTLIB::SDIVREM_I16 : RTLIB::UDIVREM_I16; break;
11601 case MVT::i32: LC = isSigned ? RTLIB::SDIVREM_I32 : RTLIB::UDIVREM_I32; break;
11602 case MVT::i64: LC = isSigned ? RTLIB::SDIVREM_I64 : RTLIB::UDIVREM_I64; break;
11607 static TargetLowering::ArgListTy getDivRemArgList(
11608 const SDNode *N, LLVMContext *Context) {
11609 assert((N->getOpcode() == ISD::SDIVREM || N->getOpcode() == ISD::UDIVREM ||
11610 N->getOpcode() == ISD::SREM || N->getOpcode() == ISD::UREM) &&
11611 "Unhandled Opcode in getDivRemArgList");
11612 bool isSigned = N->getOpcode() == ISD::SDIVREM ||
11613 N->getOpcode() == ISD::SREM;
11614 TargetLowering::ArgListTy Args;
11615 TargetLowering::ArgListEntry Entry;
11616 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
11617 EVT ArgVT = N->getOperand(i).getValueType();
11618 Type *ArgTy = ArgVT.getTypeForEVT(*Context);
11619 Entry.Node = N->getOperand(i);
11621 Entry.isSExt = isSigned;
11622 Entry.isZExt = !isSigned;
11623 Args.push_back(Entry);
11628 SDValue ARMTargetLowering::LowerDivRem(SDValue Op, SelectionDAG &DAG) const {
11629 assert((Subtarget->isTargetAEABI() || Subtarget->isTargetAndroid()) &&
11630 "Register-based DivRem lowering only");
11631 unsigned Opcode = Op->getOpcode();
11632 assert((Opcode == ISD::SDIVREM || Opcode == ISD::UDIVREM) &&
11633 "Invalid opcode for Div/Rem lowering");
11634 bool isSigned = (Opcode == ISD::SDIVREM);
11635 EVT VT = Op->getValueType(0);
11636 Type *Ty = VT.getTypeForEVT(*DAG.getContext());
11638 RTLIB::Libcall LC = getDivRemLibcall(Op.getNode(),
11639 VT.getSimpleVT().SimpleTy);
11640 SDValue InChain = DAG.getEntryNode();
11642 TargetLowering::ArgListTy Args = getDivRemArgList(Op.getNode(),
11645 SDValue Callee = DAG.getExternalSymbol(getLibcallName(LC),
11646 getPointerTy(DAG.getDataLayout()));
11648 Type *RetTy = (Type*)StructType::get(Ty, Ty, nullptr);
11651 TargetLowering::CallLoweringInfo CLI(DAG);
11652 CLI.setDebugLoc(dl).setChain(InChain)
11653 .setCallee(getLibcallCallingConv(LC), RetTy, Callee, std::move(Args), 0)
11654 .setInRegister().setSExtResult(isSigned).setZExtResult(!isSigned);
11656 std::pair<SDValue, SDValue> CallInfo = LowerCallTo(CLI);
11657 return CallInfo.first;
11660 // Lowers REM using divmod helpers
11661 // see RTABI section 4.2/4.3
11662 SDValue ARMTargetLowering::LowerREM(SDNode *N, SelectionDAG &DAG) const {
11663 // Build return types (div and rem)
11664 std::vector<Type*> RetTyParams;
11665 Type *RetTyElement;
11667 switch (N->getValueType(0).getSimpleVT().SimpleTy) {
11668 default: llvm_unreachable("Unexpected request for libcall!");
11669 case MVT::i8: RetTyElement = Type::getInt8Ty(*DAG.getContext()); break;
11670 case MVT::i16: RetTyElement = Type::getInt16Ty(*DAG.getContext()); break;
11671 case MVT::i32: RetTyElement = Type::getInt32Ty(*DAG.getContext()); break;
11672 case MVT::i64: RetTyElement = Type::getInt64Ty(*DAG.getContext()); break;
11675 RetTyParams.push_back(RetTyElement);
11676 RetTyParams.push_back(RetTyElement);
11677 ArrayRef<Type*> ret = ArrayRef<Type*>(RetTyParams);
11678 Type *RetTy = StructType::get(*DAG.getContext(), ret);
11680 RTLIB::Libcall LC = getDivRemLibcall(N, N->getValueType(0).getSimpleVT().
11682 SDValue InChain = DAG.getEntryNode();
11683 TargetLowering::ArgListTy Args = getDivRemArgList(N, DAG.getContext());
11684 bool isSigned = N->getOpcode() == ISD::SREM;
11685 SDValue Callee = DAG.getExternalSymbol(getLibcallName(LC),
11686 getPointerTy(DAG.getDataLayout()));
11689 CallLoweringInfo CLI(DAG);
11690 CLI.setChain(InChain)
11691 .setCallee(CallingConv::ARM_AAPCS, RetTy, Callee, std::move(Args), 0)
11692 .setSExtResult(isSigned).setZExtResult(!isSigned).setDebugLoc(SDLoc(N));
11693 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
11695 // Return second (rem) result operand (first contains div)
11696 SDNode *ResNode = CallResult.first.getNode();
11697 assert(ResNode->getNumOperands() == 2 && "divmod should return two operands");
11698 return ResNode->getOperand(1);
11702 ARMTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const {
11703 assert(Subtarget->isTargetWindows() && "unsupported target platform");
11707 SDValue Chain = Op.getOperand(0);
11708 SDValue Size = Op.getOperand(1);
11710 SDValue Words = DAG.getNode(ISD::SRL, DL, MVT::i32, Size,
11711 DAG.getConstant(2, DL, MVT::i32));
11714 Chain = DAG.getCopyToReg(Chain, DL, ARM::R4, Words, Flag);
11715 Flag = Chain.getValue(1);
11717 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
11718 Chain = DAG.getNode(ARMISD::WIN__CHKSTK, DL, NodeTys, Chain, Flag);
11720 SDValue NewSP = DAG.getCopyFromReg(Chain, DL, ARM::SP, MVT::i32);
11721 Chain = NewSP.getValue(1);
11723 SDValue Ops[2] = { NewSP, Chain };
11724 return DAG.getMergeValues(Ops, DL);
11727 SDValue ARMTargetLowering::LowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) const {
11728 assert(Op.getValueType() == MVT::f64 && Subtarget->isFPOnlySP() &&
11729 "Unexpected type for custom-lowering FP_EXTEND");
11732 LC = RTLIB::getFPEXT(Op.getOperand(0).getValueType(), Op.getValueType());
11734 SDValue SrcVal = Op.getOperand(0);
11735 return makeLibCall(DAG, LC, Op.getValueType(), SrcVal, /*isSigned*/ false,
11739 SDValue ARMTargetLowering::LowerFP_ROUND(SDValue Op, SelectionDAG &DAG) const {
11740 assert(Op.getOperand(0).getValueType() == MVT::f64 &&
11741 Subtarget->isFPOnlySP() &&
11742 "Unexpected type for custom-lowering FP_ROUND");
11745 LC = RTLIB::getFPROUND(Op.getOperand(0).getValueType(), Op.getValueType());
11747 SDValue SrcVal = Op.getOperand(0);
11748 return makeLibCall(DAG, LC, Op.getValueType(), SrcVal, /*isSigned*/ false,
11753 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
11754 // The ARM target isn't yet aware of offsets.
11758 bool ARM::isBitFieldInvertedMask(unsigned v) {
11759 if (v == 0xffffffff)
11762 // there can be 1's on either or both "outsides", all the "inside"
11763 // bits must be 0's
11764 return isShiftedMask_32(~v);
11767 /// isFPImmLegal - Returns true if the target can instruction select the
11768 /// specified FP immediate natively. If false, the legalizer will
11769 /// materialize the FP immediate as a load from a constant pool.
11770 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
11771 if (!Subtarget->hasVFP3())
11773 if (VT == MVT::f32)
11774 return ARM_AM::getFP32Imm(Imm) != -1;
11775 if (VT == MVT::f64 && !Subtarget->isFPOnlySP())
11776 return ARM_AM::getFP64Imm(Imm) != -1;
11780 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
11781 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
11782 /// specified in the intrinsic calls.
11783 bool ARMTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
11785 unsigned Intrinsic) const {
11786 switch (Intrinsic) {
11787 case Intrinsic::arm_neon_vld1:
11788 case Intrinsic::arm_neon_vld2:
11789 case Intrinsic::arm_neon_vld3:
11790 case Intrinsic::arm_neon_vld4:
11791 case Intrinsic::arm_neon_vld2lane:
11792 case Intrinsic::arm_neon_vld3lane:
11793 case Intrinsic::arm_neon_vld4lane: {
11794 Info.opc = ISD::INTRINSIC_W_CHAIN;
11795 // Conservatively set memVT to the entire set of vectors loaded.
11796 auto &DL = I.getCalledFunction()->getParent()->getDataLayout();
11797 uint64_t NumElts = DL.getTypeSizeInBits(I.getType()) / 64;
11798 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
11799 Info.ptrVal = I.getArgOperand(0);
11801 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
11802 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
11803 Info.vol = false; // volatile loads with NEON intrinsics not supported
11804 Info.readMem = true;
11805 Info.writeMem = false;
11808 case Intrinsic::arm_neon_vst1:
11809 case Intrinsic::arm_neon_vst2:
11810 case Intrinsic::arm_neon_vst3:
11811 case Intrinsic::arm_neon_vst4:
11812 case Intrinsic::arm_neon_vst2lane:
11813 case Intrinsic::arm_neon_vst3lane:
11814 case Intrinsic::arm_neon_vst4lane: {
11815 Info.opc = ISD::INTRINSIC_VOID;
11816 // Conservatively set memVT to the entire set of vectors stored.
11817 auto &DL = I.getCalledFunction()->getParent()->getDataLayout();
11818 unsigned NumElts = 0;
11819 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
11820 Type *ArgTy = I.getArgOperand(ArgI)->getType();
11821 if (!ArgTy->isVectorTy())
11823 NumElts += DL.getTypeSizeInBits(ArgTy) / 64;
11825 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
11826 Info.ptrVal = I.getArgOperand(0);
11828 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
11829 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
11830 Info.vol = false; // volatile stores with NEON intrinsics not supported
11831 Info.readMem = false;
11832 Info.writeMem = true;
11835 case Intrinsic::arm_ldaex:
11836 case Intrinsic::arm_ldrex: {
11837 auto &DL = I.getCalledFunction()->getParent()->getDataLayout();
11838 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(0)->getType());
11839 Info.opc = ISD::INTRINSIC_W_CHAIN;
11840 Info.memVT = MVT::getVT(PtrTy->getElementType());
11841 Info.ptrVal = I.getArgOperand(0);
11843 Info.align = DL.getABITypeAlignment(PtrTy->getElementType());
11845 Info.readMem = true;
11846 Info.writeMem = false;
11849 case Intrinsic::arm_stlex:
11850 case Intrinsic::arm_strex: {
11851 auto &DL = I.getCalledFunction()->getParent()->getDataLayout();
11852 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(1)->getType());
11853 Info.opc = ISD::INTRINSIC_W_CHAIN;
11854 Info.memVT = MVT::getVT(PtrTy->getElementType());
11855 Info.ptrVal = I.getArgOperand(1);
11857 Info.align = DL.getABITypeAlignment(PtrTy->getElementType());
11859 Info.readMem = false;
11860 Info.writeMem = true;
11863 case Intrinsic::arm_stlexd:
11864 case Intrinsic::arm_strexd: {
11865 Info.opc = ISD::INTRINSIC_W_CHAIN;
11866 Info.memVT = MVT::i64;
11867 Info.ptrVal = I.getArgOperand(2);
11871 Info.readMem = false;
11872 Info.writeMem = true;
11875 case Intrinsic::arm_ldaexd:
11876 case Intrinsic::arm_ldrexd: {
11877 Info.opc = ISD::INTRINSIC_W_CHAIN;
11878 Info.memVT = MVT::i64;
11879 Info.ptrVal = I.getArgOperand(0);
11883 Info.readMem = true;
11884 Info.writeMem = false;
11894 /// \brief Returns true if it is beneficial to convert a load of a constant
11895 /// to just the constant itself.
11896 bool ARMTargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
11898 assert(Ty->isIntegerTy());
11900 unsigned Bits = Ty->getPrimitiveSizeInBits();
11901 if (Bits == 0 || Bits > 32)
11906 Instruction* ARMTargetLowering::makeDMB(IRBuilder<> &Builder,
11907 ARM_MB::MemBOpt Domain) const {
11908 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
11910 // First, if the target has no DMB, see what fallback we can use.
11911 if (!Subtarget->hasDataBarrier()) {
11912 // Some ARMv6 cpus can support data barriers with an mcr instruction.
11913 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
11915 if (Subtarget->hasV6Ops() && !Subtarget->isThumb()) {
11916 Function *MCR = llvm::Intrinsic::getDeclaration(M, Intrinsic::arm_mcr);
11917 Value* args[6] = {Builder.getInt32(15), Builder.getInt32(0),
11918 Builder.getInt32(0), Builder.getInt32(7),
11919 Builder.getInt32(10), Builder.getInt32(5)};
11920 return Builder.CreateCall(MCR, args);
11922 // Instead of using barriers, atomic accesses on these subtargets use
11924 llvm_unreachable("makeDMB on a target so old that it has no barriers");
11927 Function *DMB = llvm::Intrinsic::getDeclaration(M, Intrinsic::arm_dmb);
11928 // Only a full system barrier exists in the M-class architectures.
11929 Domain = Subtarget->isMClass() ? ARM_MB::SY : Domain;
11930 Constant *CDomain = Builder.getInt32(Domain);
11931 return Builder.CreateCall(DMB, CDomain);
11935 // Based on http://www.cl.cam.ac.uk/~pes20/cpp/cpp0xmappings.html
11936 Instruction* ARMTargetLowering::emitLeadingFence(IRBuilder<> &Builder,
11937 AtomicOrdering Ord, bool IsStore,
11938 bool IsLoad) const {
11939 if (!getInsertFencesForAtomic())
11945 llvm_unreachable("Invalid fence: unordered/non-atomic");
11948 return nullptr; // Nothing to do
11949 case SequentiallyConsistent:
11951 return nullptr; // Nothing to do
11954 case AcquireRelease:
11955 if (Subtarget->isSwift())
11956 return makeDMB(Builder, ARM_MB::ISHST);
11957 // FIXME: add a comment with a link to documentation justifying this.
11959 return makeDMB(Builder, ARM_MB::ISH);
11961 llvm_unreachable("Unknown fence ordering in emitLeadingFence");
11964 Instruction* ARMTargetLowering::emitTrailingFence(IRBuilder<> &Builder,
11965 AtomicOrdering Ord, bool IsStore,
11966 bool IsLoad) const {
11967 if (!getInsertFencesForAtomic())
11973 llvm_unreachable("Invalid fence: unordered/not-atomic");
11976 return nullptr; // Nothing to do
11978 case AcquireRelease:
11979 case SequentiallyConsistent:
11980 return makeDMB(Builder, ARM_MB::ISH);
11982 llvm_unreachable("Unknown fence ordering in emitTrailingFence");
11985 // Loads and stores less than 64-bits are already atomic; ones above that
11986 // are doomed anyway, so defer to the default libcall and blame the OS when
11987 // things go wrong. Cortex M doesn't have ldrexd/strexd though, so don't emit
11988 // anything for those.
11989 bool ARMTargetLowering::shouldExpandAtomicStoreInIR(StoreInst *SI) const {
11990 unsigned Size = SI->getValueOperand()->getType()->getPrimitiveSizeInBits();
11991 return (Size == 64) && !Subtarget->isMClass();
11994 // Loads and stores less than 64-bits are already atomic; ones above that
11995 // are doomed anyway, so defer to the default libcall and blame the OS when
11996 // things go wrong. Cortex M doesn't have ldrexd/strexd though, so don't emit
11997 // anything for those.
11998 // FIXME: ldrd and strd are atomic if the CPU has LPAE (e.g. A15 has that
11999 // guarantee, see DDI0406C ARM architecture reference manual,
12000 // sections A8.8.72-74 LDRD)
12001 TargetLowering::AtomicExpansionKind
12002 ARMTargetLowering::shouldExpandAtomicLoadInIR(LoadInst *LI) const {
12003 unsigned Size = LI->getType()->getPrimitiveSizeInBits();
12004 return ((Size == 64) && !Subtarget->isMClass()) ? AtomicExpansionKind::LLOnly
12005 : AtomicExpansionKind::None;
12008 // For the real atomic operations, we have ldrex/strex up to 32 bits,
12009 // and up to 64 bits on the non-M profiles
12010 TargetLowering::AtomicExpansionKind
12011 ARMTargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
12012 unsigned Size = AI->getType()->getPrimitiveSizeInBits();
12013 return (Size <= (Subtarget->isMClass() ? 32U : 64U))
12014 ? AtomicExpansionKind::LLSC
12015 : AtomicExpansionKind::None;
12018 bool ARMTargetLowering::shouldExpandAtomicCmpXchgInIR(
12019 AtomicCmpXchgInst *AI) const {
12023 // This has so far only been implemented for MachO.
12024 bool ARMTargetLowering::useLoadStackGuardNode() const {
12025 return Subtarget->isTargetMachO();
12028 bool ARMTargetLowering::canCombineStoreAndExtract(Type *VectorTy, Value *Idx,
12029 unsigned &Cost) const {
12030 // If we do not have NEON, vector types are not natively supported.
12031 if (!Subtarget->hasNEON())
12034 // Floating point values and vector values map to the same register file.
12035 // Therefore, although we could do a store extract of a vector type, this is
12036 // better to leave at float as we have more freedom in the addressing mode for
12038 if (VectorTy->isFPOrFPVectorTy())
12041 // If the index is unknown at compile time, this is very expensive to lower
12042 // and it is not possible to combine the store with the extract.
12043 if (!isa<ConstantInt>(Idx))
12046 assert(VectorTy->isVectorTy() && "VectorTy is not a vector type");
12047 unsigned BitWidth = cast<VectorType>(VectorTy)->getBitWidth();
12048 // We can do a store + vector extract on any vector that fits perfectly in a D
12050 if (BitWidth == 64 || BitWidth == 128) {
12057 bool ARMTargetLowering::isCheapToSpeculateCttz() const {
12058 return Subtarget->hasV6T2Ops();
12061 bool ARMTargetLowering::isCheapToSpeculateCtlz() const {
12062 return Subtarget->hasV6T2Ops();
12065 Value *ARMTargetLowering::emitLoadLinked(IRBuilder<> &Builder, Value *Addr,
12066 AtomicOrdering Ord) const {
12067 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
12068 Type *ValTy = cast<PointerType>(Addr->getType())->getElementType();
12069 bool IsAcquire = isAtLeastAcquire(Ord);
12071 // Since i64 isn't legal and intrinsics don't get type-lowered, the ldrexd
12072 // intrinsic must return {i32, i32} and we have to recombine them into a
12073 // single i64 here.
12074 if (ValTy->getPrimitiveSizeInBits() == 64) {
12075 Intrinsic::ID Int =
12076 IsAcquire ? Intrinsic::arm_ldaexd : Intrinsic::arm_ldrexd;
12077 Function *Ldrex = llvm::Intrinsic::getDeclaration(M, Int);
12079 Addr = Builder.CreateBitCast(Addr, Type::getInt8PtrTy(M->getContext()));
12080 Value *LoHi = Builder.CreateCall(Ldrex, Addr, "lohi");
12082 Value *Lo = Builder.CreateExtractValue(LoHi, 0, "lo");
12083 Value *Hi = Builder.CreateExtractValue(LoHi, 1, "hi");
12084 if (!Subtarget->isLittle())
12085 std::swap (Lo, Hi);
12086 Lo = Builder.CreateZExt(Lo, ValTy, "lo64");
12087 Hi = Builder.CreateZExt(Hi, ValTy, "hi64");
12088 return Builder.CreateOr(
12089 Lo, Builder.CreateShl(Hi, ConstantInt::get(ValTy, 32)), "val64");
12092 Type *Tys[] = { Addr->getType() };
12093 Intrinsic::ID Int = IsAcquire ? Intrinsic::arm_ldaex : Intrinsic::arm_ldrex;
12094 Function *Ldrex = llvm::Intrinsic::getDeclaration(M, Int, Tys);
12096 return Builder.CreateTruncOrBitCast(
12097 Builder.CreateCall(Ldrex, Addr),
12098 cast<PointerType>(Addr->getType())->getElementType());
12101 void ARMTargetLowering::emitAtomicCmpXchgNoStoreLLBalance(
12102 IRBuilder<> &Builder) const {
12103 if (!Subtarget->hasV7Ops())
12105 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
12106 Builder.CreateCall(llvm::Intrinsic::getDeclaration(M, Intrinsic::arm_clrex));
12109 Value *ARMTargetLowering::emitStoreConditional(IRBuilder<> &Builder, Value *Val,
12111 AtomicOrdering Ord) const {
12112 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
12113 bool IsRelease = isAtLeastRelease(Ord);
12115 // Since the intrinsics must have legal type, the i64 intrinsics take two
12116 // parameters: "i32, i32". We must marshal Val into the appropriate form
12117 // before the call.
12118 if (Val->getType()->getPrimitiveSizeInBits() == 64) {
12119 Intrinsic::ID Int =
12120 IsRelease ? Intrinsic::arm_stlexd : Intrinsic::arm_strexd;
12121 Function *Strex = Intrinsic::getDeclaration(M, Int);
12122 Type *Int32Ty = Type::getInt32Ty(M->getContext());
12124 Value *Lo = Builder.CreateTrunc(Val, Int32Ty, "lo");
12125 Value *Hi = Builder.CreateTrunc(Builder.CreateLShr(Val, 32), Int32Ty, "hi");
12126 if (!Subtarget->isLittle())
12127 std::swap (Lo, Hi);
12128 Addr = Builder.CreateBitCast(Addr, Type::getInt8PtrTy(M->getContext()));
12129 return Builder.CreateCall(Strex, {Lo, Hi, Addr});
12132 Intrinsic::ID Int = IsRelease ? Intrinsic::arm_stlex : Intrinsic::arm_strex;
12133 Type *Tys[] = { Addr->getType() };
12134 Function *Strex = Intrinsic::getDeclaration(M, Int, Tys);
12136 return Builder.CreateCall(
12137 Strex, {Builder.CreateZExtOrBitCast(
12138 Val, Strex->getFunctionType()->getParamType(0)),
12142 /// \brief Lower an interleaved load into a vldN intrinsic.
12144 /// E.g. Lower an interleaved load (Factor = 2):
12145 /// %wide.vec = load <8 x i32>, <8 x i32>* %ptr, align 4
12146 /// %v0 = shuffle %wide.vec, undef, <0, 2, 4, 6> ; Extract even elements
12147 /// %v1 = shuffle %wide.vec, undef, <1, 3, 5, 7> ; Extract odd elements
12150 /// %vld2 = { <4 x i32>, <4 x i32> } call llvm.arm.neon.vld2(%ptr, 4)
12151 /// %vec0 = extractelement { <4 x i32>, <4 x i32> } %vld2, i32 0
12152 /// %vec1 = extractelement { <4 x i32>, <4 x i32> } %vld2, i32 1
12153 bool ARMTargetLowering::lowerInterleavedLoad(
12154 LoadInst *LI, ArrayRef<ShuffleVectorInst *> Shuffles,
12155 ArrayRef<unsigned> Indices, unsigned Factor) const {
12156 assert(Factor >= 2 && Factor <= getMaxSupportedInterleaveFactor() &&
12157 "Invalid interleave factor");
12158 assert(!Shuffles.empty() && "Empty shufflevector input");
12159 assert(Shuffles.size() == Indices.size() &&
12160 "Unmatched number of shufflevectors and indices");
12162 VectorType *VecTy = Shuffles[0]->getType();
12163 Type *EltTy = VecTy->getVectorElementType();
12165 const DataLayout &DL = LI->getModule()->getDataLayout();
12166 unsigned VecSize = DL.getTypeSizeInBits(VecTy);
12167 bool EltIs64Bits = DL.getTypeSizeInBits(EltTy) == 64;
12169 // Skip if we do not have NEON and skip illegal vector types and vector types
12170 // with i64/f64 elements (vldN doesn't support i64/f64 elements).
12171 if (!Subtarget->hasNEON() || (VecSize != 64 && VecSize != 128) || EltIs64Bits)
12174 // A pointer vector can not be the return type of the ldN intrinsics. Need to
12175 // load integer vectors first and then convert to pointer vectors.
12176 if (EltTy->isPointerTy())
12178 VectorType::get(DL.getIntPtrType(EltTy), VecTy->getVectorNumElements());
12180 static const Intrinsic::ID LoadInts[3] = {Intrinsic::arm_neon_vld2,
12181 Intrinsic::arm_neon_vld3,
12182 Intrinsic::arm_neon_vld4};
12184 IRBuilder<> Builder(LI);
12185 SmallVector<Value *, 2> Ops;
12187 Type *Int8Ptr = Builder.getInt8PtrTy(LI->getPointerAddressSpace());
12188 Ops.push_back(Builder.CreateBitCast(LI->getPointerOperand(), Int8Ptr));
12189 Ops.push_back(Builder.getInt32(LI->getAlignment()));
12191 Type *Tys[] = { VecTy, Int8Ptr };
12192 Function *VldnFunc =
12193 Intrinsic::getDeclaration(LI->getModule(), LoadInts[Factor - 2], Tys);
12194 CallInst *VldN = Builder.CreateCall(VldnFunc, Ops, "vldN");
12196 // Replace uses of each shufflevector with the corresponding vector loaded
12198 for (unsigned i = 0; i < Shuffles.size(); i++) {
12199 ShuffleVectorInst *SV = Shuffles[i];
12200 unsigned Index = Indices[i];
12202 Value *SubVec = Builder.CreateExtractValue(VldN, Index);
12204 // Convert the integer vector to pointer vector if the element is pointer.
12205 if (EltTy->isPointerTy())
12206 SubVec = Builder.CreateIntToPtr(SubVec, SV->getType());
12208 SV->replaceAllUsesWith(SubVec);
12214 /// \brief Get a mask consisting of sequential integers starting from \p Start.
12216 /// I.e. <Start, Start + 1, ..., Start + NumElts - 1>
12217 static Constant *getSequentialMask(IRBuilder<> &Builder, unsigned Start,
12218 unsigned NumElts) {
12219 SmallVector<Constant *, 16> Mask;
12220 for (unsigned i = 0; i < NumElts; i++)
12221 Mask.push_back(Builder.getInt32(Start + i));
12223 return ConstantVector::get(Mask);
12226 /// \brief Lower an interleaved store into a vstN intrinsic.
12228 /// E.g. Lower an interleaved store (Factor = 3):
12229 /// %i.vec = shuffle <8 x i32> %v0, <8 x i32> %v1,
12230 /// <0, 4, 8, 1, 5, 9, 2, 6, 10, 3, 7, 11>
12231 /// store <12 x i32> %i.vec, <12 x i32>* %ptr, align 4
12234 /// %sub.v0 = shuffle <8 x i32> %v0, <8 x i32> v1, <0, 1, 2, 3>
12235 /// %sub.v1 = shuffle <8 x i32> %v0, <8 x i32> v1, <4, 5, 6, 7>
12236 /// %sub.v2 = shuffle <8 x i32> %v0, <8 x i32> v1, <8, 9, 10, 11>
12237 /// call void llvm.arm.neon.vst3(%ptr, %sub.v0, %sub.v1, %sub.v2, 4)
12239 /// Note that the new shufflevectors will be removed and we'll only generate one
12240 /// vst3 instruction in CodeGen.
12241 bool ARMTargetLowering::lowerInterleavedStore(StoreInst *SI,
12242 ShuffleVectorInst *SVI,
12243 unsigned Factor) const {
12244 assert(Factor >= 2 && Factor <= getMaxSupportedInterleaveFactor() &&
12245 "Invalid interleave factor");
12247 VectorType *VecTy = SVI->getType();
12248 assert(VecTy->getVectorNumElements() % Factor == 0 &&
12249 "Invalid interleaved store");
12251 unsigned NumSubElts = VecTy->getVectorNumElements() / Factor;
12252 Type *EltTy = VecTy->getVectorElementType();
12253 VectorType *SubVecTy = VectorType::get(EltTy, NumSubElts);
12255 const DataLayout &DL = SI->getModule()->getDataLayout();
12256 unsigned SubVecSize = DL.getTypeSizeInBits(SubVecTy);
12257 bool EltIs64Bits = DL.getTypeSizeInBits(EltTy) == 64;
12259 // Skip if we do not have NEON and skip illegal vector types and vector types
12260 // with i64/f64 elements (vstN doesn't support i64/f64 elements).
12261 if (!Subtarget->hasNEON() || (SubVecSize != 64 && SubVecSize != 128) ||
12265 Value *Op0 = SVI->getOperand(0);
12266 Value *Op1 = SVI->getOperand(1);
12267 IRBuilder<> Builder(SI);
12269 // StN intrinsics don't support pointer vectors as arguments. Convert pointer
12270 // vectors to integer vectors.
12271 if (EltTy->isPointerTy()) {
12272 Type *IntTy = DL.getIntPtrType(EltTy);
12274 // Convert to the corresponding integer vector.
12276 VectorType::get(IntTy, Op0->getType()->getVectorNumElements());
12277 Op0 = Builder.CreatePtrToInt(Op0, IntVecTy);
12278 Op1 = Builder.CreatePtrToInt(Op1, IntVecTy);
12280 SubVecTy = VectorType::get(IntTy, NumSubElts);
12283 static const Intrinsic::ID StoreInts[3] = {Intrinsic::arm_neon_vst2,
12284 Intrinsic::arm_neon_vst3,
12285 Intrinsic::arm_neon_vst4};
12286 SmallVector<Value *, 6> Ops;
12288 Type *Int8Ptr = Builder.getInt8PtrTy(SI->getPointerAddressSpace());
12289 Ops.push_back(Builder.CreateBitCast(SI->getPointerOperand(), Int8Ptr));
12291 Type *Tys[] = { Int8Ptr, SubVecTy };
12292 Function *VstNFunc = Intrinsic::getDeclaration(
12293 SI->getModule(), StoreInts[Factor - 2], Tys);
12295 // Split the shufflevector operands into sub vectors for the new vstN call.
12296 for (unsigned i = 0; i < Factor; i++)
12297 Ops.push_back(Builder.CreateShuffleVector(
12298 Op0, Op1, getSequentialMask(Builder, NumSubElts * i, NumSubElts)));
12300 Ops.push_back(Builder.getInt32(SI->getAlignment()));
12301 Builder.CreateCall(VstNFunc, Ops);
12313 static bool isHomogeneousAggregate(Type *Ty, HABaseType &Base,
12314 uint64_t &Members) {
12315 if (auto *ST = dyn_cast<StructType>(Ty)) {
12316 for (unsigned i = 0; i < ST->getNumElements(); ++i) {
12317 uint64_t SubMembers = 0;
12318 if (!isHomogeneousAggregate(ST->getElementType(i), Base, SubMembers))
12320 Members += SubMembers;
12322 } else if (auto *AT = dyn_cast<ArrayType>(Ty)) {
12323 uint64_t SubMembers = 0;
12324 if (!isHomogeneousAggregate(AT->getElementType(), Base, SubMembers))
12326 Members += SubMembers * AT->getNumElements();
12327 } else if (Ty->isFloatTy()) {
12328 if (Base != HA_UNKNOWN && Base != HA_FLOAT)
12332 } else if (Ty->isDoubleTy()) {
12333 if (Base != HA_UNKNOWN && Base != HA_DOUBLE)
12337 } else if (auto *VT = dyn_cast<VectorType>(Ty)) {
12344 return VT->getBitWidth() == 64;
12346 return VT->getBitWidth() == 128;
12348 switch (VT->getBitWidth()) {
12361 return (Members > 0 && Members <= 4);
12364 /// \brief Return true if a type is an AAPCS-VFP homogeneous aggregate or one of
12365 /// [N x i32] or [N x i64]. This allows front-ends to skip emitting padding when
12366 /// passing according to AAPCS rules.
12367 bool ARMTargetLowering::functionArgumentNeedsConsecutiveRegisters(
12368 Type *Ty, CallingConv::ID CallConv, bool isVarArg) const {
12369 if (getEffectiveCallingConv(CallConv, isVarArg) !=
12370 CallingConv::ARM_AAPCS_VFP)
12373 HABaseType Base = HA_UNKNOWN;
12374 uint64_t Members = 0;
12375 bool IsHA = isHomogeneousAggregate(Ty, Base, Members);
12376 DEBUG(dbgs() << "isHA: " << IsHA << " "; Ty->dump());
12378 bool IsIntArray = Ty->isArrayTy() && Ty->getArrayElementType()->isIntegerTy();
12379 return IsHA || IsIntArray;
12382 unsigned ARMTargetLowering::getExceptionPointerRegister(
12383 const Constant *PersonalityFn) const {
12384 // Platforms which do not use SjLj EH may return values in these registers
12385 // via the personality function.
12386 return Subtarget->useSjLjEH() ? ARM::NoRegister : ARM::R0;
12389 unsigned ARMTargetLowering::getExceptionSelectorRegister(
12390 const Constant *PersonalityFn) const {
12391 // Platforms which do not use SjLj EH may return values in these registers
12392 // via the personality function.
12393 return Subtarget->useSjLjEH() ? ARM::NoRegister : ARM::R1;