1 //===-- ARMISelLowering.cpp - ARM DAG Lowering Implementation -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the interfaces that ARM uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
15 #include "ARMISelLowering.h"
16 #include "ARMCallingConv.h"
17 #include "ARMConstantPoolValue.h"
18 #include "ARMMachineFunctionInfo.h"
19 #include "ARMPerfectShuffle.h"
20 #include "ARMSubtarget.h"
21 #include "ARMTargetMachine.h"
22 #include "ARMTargetObjectFile.h"
23 #include "MCTargetDesc/ARMAddressingModes.h"
24 #include "llvm/ADT/Statistic.h"
25 #include "llvm/ADT/StringExtras.h"
26 #include "llvm/ADT/StringSwitch.h"
27 #include "llvm/CodeGen/CallingConvLower.h"
28 #include "llvm/CodeGen/IntrinsicLowering.h"
29 #include "llvm/CodeGen/MachineBasicBlock.h"
30 #include "llvm/CodeGen/MachineFrameInfo.h"
31 #include "llvm/CodeGen/MachineFunction.h"
32 #include "llvm/CodeGen/MachineInstrBuilder.h"
33 #include "llvm/CodeGen/MachineJumpTableInfo.h"
34 #include "llvm/CodeGen/MachineModuleInfo.h"
35 #include "llvm/CodeGen/MachineRegisterInfo.h"
36 #include "llvm/CodeGen/SelectionDAG.h"
37 #include "llvm/IR/CallingConv.h"
38 #include "llvm/IR/Constants.h"
39 #include "llvm/IR/Function.h"
40 #include "llvm/IR/GlobalValue.h"
41 #include "llvm/IR/IRBuilder.h"
42 #include "llvm/IR/Instruction.h"
43 #include "llvm/IR/Instructions.h"
44 #include "llvm/IR/IntrinsicInst.h"
45 #include "llvm/IR/Intrinsics.h"
46 #include "llvm/IR/Type.h"
47 #include "llvm/MC/MCSectionMachO.h"
48 #include "llvm/Support/CommandLine.h"
49 #include "llvm/Support/Debug.h"
50 #include "llvm/Support/ErrorHandling.h"
51 #include "llvm/Support/MathExtras.h"
52 #include "llvm/Support/raw_ostream.h"
53 #include "llvm/Target/TargetOptions.h"
57 #define DEBUG_TYPE "arm-isel"
59 STATISTIC(NumTailCalls, "Number of tail calls");
60 STATISTIC(NumMovwMovt, "Number of GAs materialized with movw + movt");
61 STATISTIC(NumLoopByVals, "Number of loops generated for byval arguments");
64 ARMInterworking("arm-interworking", cl::Hidden,
65 cl::desc("Enable / disable ARM interworking (for debugging only)"),
69 class ARMCCState : public CCState {
71 ARMCCState(CallingConv::ID CC, bool isVarArg, MachineFunction &MF,
72 SmallVectorImpl<CCValAssign> &locs, LLVMContext &C,
74 : CCState(CC, isVarArg, MF, locs, C) {
75 assert(((PC == Call) || (PC == Prologue)) &&
76 "ARMCCState users must specify whether their context is call"
77 "or prologue generation.");
83 // The APCS parameter registers.
84 static const MCPhysReg GPRArgRegs[] = {
85 ARM::R0, ARM::R1, ARM::R2, ARM::R3
88 void ARMTargetLowering::addTypeForNEON(MVT VT, MVT PromotedLdStVT,
89 MVT PromotedBitwiseVT) {
90 if (VT != PromotedLdStVT) {
91 setOperationAction(ISD::LOAD, VT, Promote);
92 AddPromotedToType (ISD::LOAD, VT, PromotedLdStVT);
94 setOperationAction(ISD::STORE, VT, Promote);
95 AddPromotedToType (ISD::STORE, VT, PromotedLdStVT);
98 MVT ElemTy = VT.getVectorElementType();
99 if (ElemTy != MVT::i64 && ElemTy != MVT::f64)
100 setOperationAction(ISD::SETCC, VT, Custom);
101 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
102 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
103 if (ElemTy == MVT::i32) {
104 setOperationAction(ISD::SINT_TO_FP, VT, Custom);
105 setOperationAction(ISD::UINT_TO_FP, VT, Custom);
106 setOperationAction(ISD::FP_TO_SINT, VT, Custom);
107 setOperationAction(ISD::FP_TO_UINT, VT, Custom);
109 setOperationAction(ISD::SINT_TO_FP, VT, Expand);
110 setOperationAction(ISD::UINT_TO_FP, VT, Expand);
111 setOperationAction(ISD::FP_TO_SINT, VT, Expand);
112 setOperationAction(ISD::FP_TO_UINT, VT, Expand);
114 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
115 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
116 setOperationAction(ISD::CONCAT_VECTORS, VT, Legal);
117 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Legal);
118 setOperationAction(ISD::SELECT, VT, Expand);
119 setOperationAction(ISD::SELECT_CC, VT, Expand);
120 setOperationAction(ISD::VSELECT, VT, Expand);
121 setOperationAction(ISD::SIGN_EXTEND_INREG, VT, Expand);
122 if (VT.isInteger()) {
123 setOperationAction(ISD::SHL, VT, Custom);
124 setOperationAction(ISD::SRA, VT, Custom);
125 setOperationAction(ISD::SRL, VT, Custom);
128 // Promote all bit-wise operations.
129 if (VT.isInteger() && VT != PromotedBitwiseVT) {
130 setOperationAction(ISD::AND, VT, Promote);
131 AddPromotedToType (ISD::AND, VT, PromotedBitwiseVT);
132 setOperationAction(ISD::OR, VT, Promote);
133 AddPromotedToType (ISD::OR, VT, PromotedBitwiseVT);
134 setOperationAction(ISD::XOR, VT, Promote);
135 AddPromotedToType (ISD::XOR, VT, PromotedBitwiseVT);
138 // Neon does not support vector divide/remainder operations.
139 setOperationAction(ISD::SDIV, VT, Expand);
140 setOperationAction(ISD::UDIV, VT, Expand);
141 setOperationAction(ISD::FDIV, VT, Expand);
142 setOperationAction(ISD::SREM, VT, Expand);
143 setOperationAction(ISD::UREM, VT, Expand);
144 setOperationAction(ISD::FREM, VT, Expand);
146 if (VT.isInteger()) {
147 setOperationAction(ISD::SABSDIFF, VT, Legal);
148 setOperationAction(ISD::UABSDIFF, VT, Legal);
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);
242 // These libcalls are not available in 32-bit.
243 setLibcallName(RTLIB::SHL_I128, nullptr);
244 setLibcallName(RTLIB::SRL_I128, nullptr);
245 setLibcallName(RTLIB::SRA_I128, nullptr);
247 if (Subtarget->isAAPCS_ABI() && !Subtarget->isTargetMachO() &&
248 !Subtarget->isTargetWindows()) {
249 static const struct {
250 const RTLIB::Libcall Op;
251 const char * const Name;
252 const CallingConv::ID CC;
253 const ISD::CondCode Cond;
255 // Double-precision floating-point arithmetic helper functions
256 // RTABI chapter 4.1.2, Table 2
257 { RTLIB::ADD_F64, "__aeabi_dadd", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
258 { RTLIB::DIV_F64, "__aeabi_ddiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
259 { RTLIB::MUL_F64, "__aeabi_dmul", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
260 { RTLIB::SUB_F64, "__aeabi_dsub", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
262 // Double-precision floating-point comparison helper functions
263 // RTABI chapter 4.1.2, Table 3
264 { RTLIB::OEQ_F64, "__aeabi_dcmpeq", CallingConv::ARM_AAPCS, ISD::SETNE },
265 { RTLIB::UNE_F64, "__aeabi_dcmpeq", CallingConv::ARM_AAPCS, ISD::SETEQ },
266 { RTLIB::OLT_F64, "__aeabi_dcmplt", CallingConv::ARM_AAPCS, ISD::SETNE },
267 { RTLIB::OLE_F64, "__aeabi_dcmple", CallingConv::ARM_AAPCS, ISD::SETNE },
268 { RTLIB::OGE_F64, "__aeabi_dcmpge", CallingConv::ARM_AAPCS, ISD::SETNE },
269 { RTLIB::OGT_F64, "__aeabi_dcmpgt", CallingConv::ARM_AAPCS, ISD::SETNE },
270 { RTLIB::UO_F64, "__aeabi_dcmpun", CallingConv::ARM_AAPCS, ISD::SETNE },
271 { RTLIB::O_F64, "__aeabi_dcmpun", CallingConv::ARM_AAPCS, ISD::SETEQ },
273 // Single-precision floating-point arithmetic helper functions
274 // RTABI chapter 4.1.2, Table 4
275 { RTLIB::ADD_F32, "__aeabi_fadd", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
276 { RTLIB::DIV_F32, "__aeabi_fdiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
277 { RTLIB::MUL_F32, "__aeabi_fmul", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
278 { RTLIB::SUB_F32, "__aeabi_fsub", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
280 // Single-precision floating-point comparison helper functions
281 // RTABI chapter 4.1.2, Table 5
282 { RTLIB::OEQ_F32, "__aeabi_fcmpeq", CallingConv::ARM_AAPCS, ISD::SETNE },
283 { RTLIB::UNE_F32, "__aeabi_fcmpeq", CallingConv::ARM_AAPCS, ISD::SETEQ },
284 { RTLIB::OLT_F32, "__aeabi_fcmplt", CallingConv::ARM_AAPCS, ISD::SETNE },
285 { RTLIB::OLE_F32, "__aeabi_fcmple", CallingConv::ARM_AAPCS, ISD::SETNE },
286 { RTLIB::OGE_F32, "__aeabi_fcmpge", CallingConv::ARM_AAPCS, ISD::SETNE },
287 { RTLIB::OGT_F32, "__aeabi_fcmpgt", CallingConv::ARM_AAPCS, ISD::SETNE },
288 { RTLIB::UO_F32, "__aeabi_fcmpun", CallingConv::ARM_AAPCS, ISD::SETNE },
289 { RTLIB::O_F32, "__aeabi_fcmpun", CallingConv::ARM_AAPCS, ISD::SETEQ },
291 // Floating-point to integer conversions.
292 // RTABI chapter 4.1.2, Table 6
293 { RTLIB::FPTOSINT_F64_I32, "__aeabi_d2iz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
294 { RTLIB::FPTOUINT_F64_I32, "__aeabi_d2uiz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
295 { RTLIB::FPTOSINT_F64_I64, "__aeabi_d2lz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
296 { RTLIB::FPTOUINT_F64_I64, "__aeabi_d2ulz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
297 { RTLIB::FPTOSINT_F32_I32, "__aeabi_f2iz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
298 { RTLIB::FPTOUINT_F32_I32, "__aeabi_f2uiz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
299 { RTLIB::FPTOSINT_F32_I64, "__aeabi_f2lz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
300 { RTLIB::FPTOUINT_F32_I64, "__aeabi_f2ulz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
302 // Conversions between floating types.
303 // RTABI chapter 4.1.2, Table 7
304 { RTLIB::FPROUND_F64_F32, "__aeabi_d2f", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
305 { RTLIB::FPROUND_F64_F16, "__aeabi_d2h", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
306 { RTLIB::FPEXT_F32_F64, "__aeabi_f2d", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
308 // Integer to floating-point conversions.
309 // RTABI chapter 4.1.2, Table 8
310 { RTLIB::SINTTOFP_I32_F64, "__aeabi_i2d", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
311 { RTLIB::UINTTOFP_I32_F64, "__aeabi_ui2d", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
312 { RTLIB::SINTTOFP_I64_F64, "__aeabi_l2d", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
313 { RTLIB::UINTTOFP_I64_F64, "__aeabi_ul2d", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
314 { RTLIB::SINTTOFP_I32_F32, "__aeabi_i2f", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
315 { RTLIB::UINTTOFP_I32_F32, "__aeabi_ui2f", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
316 { RTLIB::SINTTOFP_I64_F32, "__aeabi_l2f", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
317 { RTLIB::UINTTOFP_I64_F32, "__aeabi_ul2f", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
319 // Long long helper functions
320 // RTABI chapter 4.2, Table 9
321 { RTLIB::MUL_I64, "__aeabi_lmul", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
322 { RTLIB::SHL_I64, "__aeabi_llsl", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
323 { RTLIB::SRL_I64, "__aeabi_llsr", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
324 { RTLIB::SRA_I64, "__aeabi_lasr", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
326 // Integer division functions
327 // RTABI chapter 4.3.1
328 { RTLIB::SDIV_I8, "__aeabi_idiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
329 { RTLIB::SDIV_I16, "__aeabi_idiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
330 { RTLIB::SDIV_I32, "__aeabi_idiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
331 { RTLIB::SDIV_I64, "__aeabi_ldivmod", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
332 { RTLIB::UDIV_I8, "__aeabi_uidiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
333 { RTLIB::UDIV_I16, "__aeabi_uidiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
334 { RTLIB::UDIV_I32, "__aeabi_uidiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
335 { RTLIB::UDIV_I64, "__aeabi_uldivmod", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
338 // RTABI chapter 4.3.4
339 { RTLIB::MEMCPY, "__aeabi_memcpy", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
340 { RTLIB::MEMMOVE, "__aeabi_memmove", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
341 { RTLIB::MEMSET, "__aeabi_memset", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
344 for (const auto &LC : LibraryCalls) {
345 setLibcallName(LC.Op, LC.Name);
346 setLibcallCallingConv(LC.Op, LC.CC);
347 if (LC.Cond != ISD::SETCC_INVALID)
348 setCmpLibcallCC(LC.Op, LC.Cond);
352 if (Subtarget->isTargetWindows()) {
353 static const struct {
354 const RTLIB::Libcall Op;
355 const char * const Name;
356 const CallingConv::ID CC;
358 { RTLIB::FPTOSINT_F32_I64, "__stoi64", CallingConv::ARM_AAPCS_VFP },
359 { RTLIB::FPTOSINT_F64_I64, "__dtoi64", CallingConv::ARM_AAPCS_VFP },
360 { RTLIB::FPTOUINT_F32_I64, "__stou64", CallingConv::ARM_AAPCS_VFP },
361 { RTLIB::FPTOUINT_F64_I64, "__dtou64", CallingConv::ARM_AAPCS_VFP },
362 { RTLIB::SINTTOFP_I64_F32, "__i64tos", CallingConv::ARM_AAPCS_VFP },
363 { RTLIB::SINTTOFP_I64_F64, "__i64tod", CallingConv::ARM_AAPCS_VFP },
364 { RTLIB::UINTTOFP_I64_F32, "__u64tos", CallingConv::ARM_AAPCS_VFP },
365 { RTLIB::UINTTOFP_I64_F64, "__u64tod", CallingConv::ARM_AAPCS_VFP },
367 { RTLIB::SDIV_I32, "__rt_sdiv", CallingConv::ARM_AAPCS_VFP },
368 { RTLIB::UDIV_I32, "__rt_udiv", CallingConv::ARM_AAPCS_VFP },
369 { RTLIB::SDIV_I64, "__rt_sdiv64", CallingConv::ARM_AAPCS_VFP },
370 { RTLIB::UDIV_I64, "__rt_udiv64", CallingConv::ARM_AAPCS_VFP },
373 for (const auto &LC : LibraryCalls) {
374 setLibcallName(LC.Op, LC.Name);
375 setLibcallCallingConv(LC.Op, LC.CC);
379 // Use divmod compiler-rt calls for iOS 5.0 and later.
380 if (Subtarget->getTargetTriple().isiOS() &&
381 !Subtarget->getTargetTriple().isOSVersionLT(5, 0)) {
382 setLibcallName(RTLIB::SDIVREM_I32, "__divmodsi4");
383 setLibcallName(RTLIB::UDIVREM_I32, "__udivmodsi4");
386 // The half <-> float conversion functions are always soft-float, but are
387 // needed for some targets which use a hard-float calling convention by
389 if (Subtarget->isAAPCS_ABI()) {
390 setLibcallCallingConv(RTLIB::FPROUND_F32_F16, CallingConv::ARM_AAPCS);
391 setLibcallCallingConv(RTLIB::FPROUND_F64_F16, CallingConv::ARM_AAPCS);
392 setLibcallCallingConv(RTLIB::FPEXT_F16_F32, CallingConv::ARM_AAPCS);
394 setLibcallCallingConv(RTLIB::FPROUND_F32_F16, CallingConv::ARM_APCS);
395 setLibcallCallingConv(RTLIB::FPROUND_F64_F16, CallingConv::ARM_APCS);
396 setLibcallCallingConv(RTLIB::FPEXT_F16_F32, CallingConv::ARM_APCS);
399 if (Subtarget->isThumb1Only())
400 addRegisterClass(MVT::i32, &ARM::tGPRRegClass);
402 addRegisterClass(MVT::i32, &ARM::GPRRegClass);
403 if (!Subtarget->useSoftFloat() && Subtarget->hasVFP2() &&
404 !Subtarget->isThumb1Only()) {
405 addRegisterClass(MVT::f32, &ARM::SPRRegClass);
406 addRegisterClass(MVT::f64, &ARM::DPRRegClass);
409 for (MVT VT : MVT::vector_valuetypes()) {
410 for (MVT InnerVT : MVT::vector_valuetypes()) {
411 setTruncStoreAction(VT, InnerVT, Expand);
412 setLoadExtAction(ISD::SEXTLOAD, VT, InnerVT, Expand);
413 setLoadExtAction(ISD::ZEXTLOAD, VT, InnerVT, Expand);
414 setLoadExtAction(ISD::EXTLOAD, VT, InnerVT, Expand);
417 setOperationAction(ISD::MULHS, VT, Expand);
418 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
419 setOperationAction(ISD::MULHU, VT, Expand);
420 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
422 setOperationAction(ISD::BSWAP, VT, Expand);
425 setOperationAction(ISD::ConstantFP, MVT::f32, Custom);
426 setOperationAction(ISD::ConstantFP, MVT::f64, Custom);
428 setOperationAction(ISD::READ_REGISTER, MVT::i64, Custom);
429 setOperationAction(ISD::WRITE_REGISTER, MVT::i64, Custom);
431 if (Subtarget->hasNEON()) {
432 addDRTypeForNEON(MVT::v2f32);
433 addDRTypeForNEON(MVT::v8i8);
434 addDRTypeForNEON(MVT::v4i16);
435 addDRTypeForNEON(MVT::v2i32);
436 addDRTypeForNEON(MVT::v1i64);
438 addQRTypeForNEON(MVT::v4f32);
439 addQRTypeForNEON(MVT::v2f64);
440 addQRTypeForNEON(MVT::v16i8);
441 addQRTypeForNEON(MVT::v8i16);
442 addQRTypeForNEON(MVT::v4i32);
443 addQRTypeForNEON(MVT::v2i64);
445 // v2f64 is legal so that QR subregs can be extracted as f64 elements, but
446 // neither Neon nor VFP support any arithmetic operations on it.
447 // The same with v4f32. But keep in mind that vadd, vsub, vmul are natively
448 // supported for v4f32.
449 setOperationAction(ISD::FADD, MVT::v2f64, Expand);
450 setOperationAction(ISD::FSUB, MVT::v2f64, Expand);
451 setOperationAction(ISD::FMUL, MVT::v2f64, Expand);
452 // FIXME: Code duplication: FDIV and FREM are expanded always, see
453 // ARMTargetLowering::addTypeForNEON method for details.
454 setOperationAction(ISD::FDIV, MVT::v2f64, Expand);
455 setOperationAction(ISD::FREM, MVT::v2f64, Expand);
456 // FIXME: Create unittest.
457 // In another words, find a way when "copysign" appears in DAG with vector
459 setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand);
460 // FIXME: Code duplication: SETCC has custom operation action, see
461 // ARMTargetLowering::addTypeForNEON method for details.
462 setOperationAction(ISD::SETCC, MVT::v2f64, Expand);
463 // FIXME: Create unittest for FNEG and for FABS.
464 setOperationAction(ISD::FNEG, MVT::v2f64, Expand);
465 setOperationAction(ISD::FABS, MVT::v2f64, Expand);
466 setOperationAction(ISD::FSQRT, MVT::v2f64, Expand);
467 setOperationAction(ISD::FSIN, MVT::v2f64, Expand);
468 setOperationAction(ISD::FCOS, MVT::v2f64, Expand);
469 setOperationAction(ISD::FPOWI, MVT::v2f64, Expand);
470 setOperationAction(ISD::FPOW, MVT::v2f64, Expand);
471 setOperationAction(ISD::FLOG, MVT::v2f64, Expand);
472 setOperationAction(ISD::FLOG2, MVT::v2f64, Expand);
473 setOperationAction(ISD::FLOG10, MVT::v2f64, Expand);
474 setOperationAction(ISD::FEXP, MVT::v2f64, Expand);
475 setOperationAction(ISD::FEXP2, MVT::v2f64, Expand);
476 // FIXME: Create unittest for FCEIL, FTRUNC, FRINT, FNEARBYINT, FFLOOR.
477 setOperationAction(ISD::FCEIL, MVT::v2f64, Expand);
478 setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand);
479 setOperationAction(ISD::FRINT, MVT::v2f64, Expand);
480 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand);
481 setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand);
482 setOperationAction(ISD::FMA, MVT::v2f64, Expand);
484 setOperationAction(ISD::FSQRT, MVT::v4f32, Expand);
485 setOperationAction(ISD::FSIN, MVT::v4f32, Expand);
486 setOperationAction(ISD::FCOS, MVT::v4f32, Expand);
487 setOperationAction(ISD::FPOWI, MVT::v4f32, Expand);
488 setOperationAction(ISD::FPOW, MVT::v4f32, Expand);
489 setOperationAction(ISD::FLOG, MVT::v4f32, Expand);
490 setOperationAction(ISD::FLOG2, MVT::v4f32, Expand);
491 setOperationAction(ISD::FLOG10, MVT::v4f32, Expand);
492 setOperationAction(ISD::FEXP, MVT::v4f32, Expand);
493 setOperationAction(ISD::FEXP2, MVT::v4f32, Expand);
494 setOperationAction(ISD::FCEIL, MVT::v4f32, Expand);
495 setOperationAction(ISD::FTRUNC, MVT::v4f32, Expand);
496 setOperationAction(ISD::FRINT, MVT::v4f32, Expand);
497 setOperationAction(ISD::FNEARBYINT, MVT::v4f32, Expand);
498 setOperationAction(ISD::FFLOOR, MVT::v4f32, Expand);
500 // Mark v2f32 intrinsics.
501 setOperationAction(ISD::FSQRT, MVT::v2f32, Expand);
502 setOperationAction(ISD::FSIN, MVT::v2f32, Expand);
503 setOperationAction(ISD::FCOS, MVT::v2f32, Expand);
504 setOperationAction(ISD::FPOWI, MVT::v2f32, Expand);
505 setOperationAction(ISD::FPOW, MVT::v2f32, Expand);
506 setOperationAction(ISD::FLOG, MVT::v2f32, Expand);
507 setOperationAction(ISD::FLOG2, MVT::v2f32, Expand);
508 setOperationAction(ISD::FLOG10, MVT::v2f32, Expand);
509 setOperationAction(ISD::FEXP, MVT::v2f32, Expand);
510 setOperationAction(ISD::FEXP2, MVT::v2f32, Expand);
511 setOperationAction(ISD::FCEIL, MVT::v2f32, Expand);
512 setOperationAction(ISD::FTRUNC, MVT::v2f32, Expand);
513 setOperationAction(ISD::FRINT, MVT::v2f32, Expand);
514 setOperationAction(ISD::FNEARBYINT, MVT::v2f32, Expand);
515 setOperationAction(ISD::FFLOOR, MVT::v2f32, Expand);
517 // Neon does not support some operations on v1i64 and v2i64 types.
518 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
519 // Custom handling for some quad-vector types to detect VMULL.
520 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
521 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
522 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
523 // Custom handling for some vector types to avoid expensive expansions
524 setOperationAction(ISD::SDIV, MVT::v4i16, Custom);
525 setOperationAction(ISD::SDIV, MVT::v8i8, Custom);
526 setOperationAction(ISD::UDIV, MVT::v4i16, Custom);
527 setOperationAction(ISD::UDIV, MVT::v8i8, Custom);
528 setOperationAction(ISD::SETCC, MVT::v1i64, Expand);
529 setOperationAction(ISD::SETCC, MVT::v2i64, Expand);
530 // Neon does not have single instruction SINT_TO_FP and UINT_TO_FP with
531 // a destination type that is wider than the source, and nor does
532 // it have a FP_TO_[SU]INT instruction with a narrower destination than
534 setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Custom);
535 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
536 setOperationAction(ISD::FP_TO_UINT, MVT::v4i16, Custom);
537 setOperationAction(ISD::FP_TO_SINT, MVT::v4i16, Custom);
539 setOperationAction(ISD::FP_ROUND, MVT::v2f32, Expand);
540 setOperationAction(ISD::FP_EXTEND, MVT::v2f64, Expand);
542 // NEON does not have single instruction CTPOP for vectors with element
543 // types wider than 8-bits. However, custom lowering can leverage the
544 // v8i8/v16i8 vcnt instruction.
545 setOperationAction(ISD::CTPOP, MVT::v2i32, Custom);
546 setOperationAction(ISD::CTPOP, MVT::v4i32, Custom);
547 setOperationAction(ISD::CTPOP, MVT::v4i16, Custom);
548 setOperationAction(ISD::CTPOP, MVT::v8i16, Custom);
550 // NEON does not have single instruction CTTZ for vectors.
551 setOperationAction(ISD::CTTZ, MVT::v8i8, Custom);
552 setOperationAction(ISD::CTTZ, MVT::v4i16, Custom);
553 setOperationAction(ISD::CTTZ, MVT::v2i32, Custom);
554 setOperationAction(ISD::CTTZ, MVT::v1i64, Custom);
556 setOperationAction(ISD::CTTZ, MVT::v16i8, Custom);
557 setOperationAction(ISD::CTTZ, MVT::v8i16, Custom);
558 setOperationAction(ISD::CTTZ, MVT::v4i32, Custom);
559 setOperationAction(ISD::CTTZ, MVT::v2i64, Custom);
561 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v8i8, Custom);
562 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v4i16, Custom);
563 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v2i32, Custom);
564 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v1i64, Custom);
566 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v16i8, Custom);
567 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v8i16, Custom);
568 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v4i32, Custom);
569 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v2i64, Custom);
571 // NEON only has FMA instructions as of VFP4.
572 if (!Subtarget->hasVFP4()) {
573 setOperationAction(ISD::FMA, MVT::v2f32, Expand);
574 setOperationAction(ISD::FMA, MVT::v4f32, Expand);
577 setTargetDAGCombine(ISD::INTRINSIC_VOID);
578 setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN);
579 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
580 setTargetDAGCombine(ISD::SHL);
581 setTargetDAGCombine(ISD::SRL);
582 setTargetDAGCombine(ISD::SRA);
583 setTargetDAGCombine(ISD::SIGN_EXTEND);
584 setTargetDAGCombine(ISD::ZERO_EXTEND);
585 setTargetDAGCombine(ISD::ANY_EXTEND);
586 setTargetDAGCombine(ISD::SELECT_CC);
587 setTargetDAGCombine(ISD::BUILD_VECTOR);
588 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
589 setTargetDAGCombine(ISD::INSERT_VECTOR_ELT);
590 setTargetDAGCombine(ISD::STORE);
591 setTargetDAGCombine(ISD::FP_TO_SINT);
592 setTargetDAGCombine(ISD::FP_TO_UINT);
593 setTargetDAGCombine(ISD::FDIV);
594 setTargetDAGCombine(ISD::LOAD);
596 // It is legal to extload from v4i8 to v4i16 or v4i32.
597 for (MVT Ty : {MVT::v8i8, MVT::v4i8, MVT::v2i8, MVT::v4i16, MVT::v2i16,
599 for (MVT VT : MVT::integer_vector_valuetypes()) {
600 setLoadExtAction(ISD::EXTLOAD, VT, Ty, Legal);
601 setLoadExtAction(ISD::ZEXTLOAD, VT, Ty, Legal);
602 setLoadExtAction(ISD::SEXTLOAD, VT, Ty, Legal);
607 // ARM and Thumb2 support UMLAL/SMLAL.
608 if (!Subtarget->isThumb1Only())
609 setTargetDAGCombine(ISD::ADDC);
611 if (Subtarget->isFPOnlySP()) {
612 // When targeting a floating-point unit with only single-precision
613 // operations, f64 is legal for the few double-precision instructions which
614 // are present However, no double-precision operations other than moves,
615 // loads and stores are provided by the hardware.
616 setOperationAction(ISD::FADD, MVT::f64, Expand);
617 setOperationAction(ISD::FSUB, MVT::f64, Expand);
618 setOperationAction(ISD::FMUL, MVT::f64, Expand);
619 setOperationAction(ISD::FMA, MVT::f64, Expand);
620 setOperationAction(ISD::FDIV, MVT::f64, Expand);
621 setOperationAction(ISD::FREM, MVT::f64, Expand);
622 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
623 setOperationAction(ISD::FGETSIGN, MVT::f64, Expand);
624 setOperationAction(ISD::FNEG, MVT::f64, Expand);
625 setOperationAction(ISD::FABS, MVT::f64, Expand);
626 setOperationAction(ISD::FSQRT, MVT::f64, Expand);
627 setOperationAction(ISD::FSIN, MVT::f64, Expand);
628 setOperationAction(ISD::FCOS, MVT::f64, Expand);
629 setOperationAction(ISD::FPOWI, MVT::f64, Expand);
630 setOperationAction(ISD::FPOW, MVT::f64, Expand);
631 setOperationAction(ISD::FLOG, MVT::f64, Expand);
632 setOperationAction(ISD::FLOG2, MVT::f64, Expand);
633 setOperationAction(ISD::FLOG10, MVT::f64, Expand);
634 setOperationAction(ISD::FEXP, MVT::f64, Expand);
635 setOperationAction(ISD::FEXP2, MVT::f64, Expand);
636 setOperationAction(ISD::FCEIL, MVT::f64, Expand);
637 setOperationAction(ISD::FTRUNC, MVT::f64, Expand);
638 setOperationAction(ISD::FRINT, MVT::f64, Expand);
639 setOperationAction(ISD::FNEARBYINT, MVT::f64, Expand);
640 setOperationAction(ISD::FFLOOR, MVT::f64, Expand);
641 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
642 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
643 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
644 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
645 setOperationAction(ISD::FP_TO_SINT, MVT::f64, Custom);
646 setOperationAction(ISD::FP_TO_UINT, MVT::f64, Custom);
647 setOperationAction(ISD::FP_ROUND, MVT::f32, Custom);
648 setOperationAction(ISD::FP_EXTEND, MVT::f64, Custom);
651 computeRegisterProperties(Subtarget->getRegisterInfo());
653 // ARM does not have floating-point extending loads.
654 for (MVT VT : MVT::fp_valuetypes()) {
655 setLoadExtAction(ISD::EXTLOAD, VT, MVT::f32, Expand);
656 setLoadExtAction(ISD::EXTLOAD, VT, MVT::f16, Expand);
659 // ... or truncating stores
660 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
661 setTruncStoreAction(MVT::f32, MVT::f16, Expand);
662 setTruncStoreAction(MVT::f64, MVT::f16, Expand);
664 // ARM does not have i1 sign extending load.
665 for (MVT VT : MVT::integer_valuetypes())
666 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Promote);
668 // ARM supports all 4 flavors of integer indexed load / store.
669 if (!Subtarget->isThumb1Only()) {
670 for (unsigned im = (unsigned)ISD::PRE_INC;
671 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
672 setIndexedLoadAction(im, MVT::i1, Legal);
673 setIndexedLoadAction(im, MVT::i8, Legal);
674 setIndexedLoadAction(im, MVT::i16, Legal);
675 setIndexedLoadAction(im, MVT::i32, Legal);
676 setIndexedStoreAction(im, MVT::i1, Legal);
677 setIndexedStoreAction(im, MVT::i8, Legal);
678 setIndexedStoreAction(im, MVT::i16, Legal);
679 setIndexedStoreAction(im, MVT::i32, Legal);
683 setOperationAction(ISD::SADDO, MVT::i32, Custom);
684 setOperationAction(ISD::UADDO, MVT::i32, Custom);
685 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
686 setOperationAction(ISD::USUBO, MVT::i32, Custom);
688 // i64 operation support.
689 setOperationAction(ISD::MUL, MVT::i64, Expand);
690 setOperationAction(ISD::MULHU, MVT::i32, Expand);
691 if (Subtarget->isThumb1Only()) {
692 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
693 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
695 if (Subtarget->isThumb1Only() || !Subtarget->hasV6Ops()
696 || (Subtarget->isThumb2() && !Subtarget->hasThumb2DSP()))
697 setOperationAction(ISD::MULHS, MVT::i32, Expand);
699 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
700 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
701 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
702 setOperationAction(ISD::SRL, MVT::i64, Custom);
703 setOperationAction(ISD::SRA, MVT::i64, Custom);
705 if (!Subtarget->isThumb1Only()) {
706 // FIXME: We should do this for Thumb1 as well.
707 setOperationAction(ISD::ADDC, MVT::i32, Custom);
708 setOperationAction(ISD::ADDE, MVT::i32, Custom);
709 setOperationAction(ISD::SUBC, MVT::i32, Custom);
710 setOperationAction(ISD::SUBE, MVT::i32, Custom);
713 // ARM does not have ROTL.
714 setOperationAction(ISD::ROTL, MVT::i32, Expand);
715 setOperationAction(ISD::CTTZ, MVT::i32, Custom);
716 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
717 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
718 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
720 // These just redirect to CTTZ and CTLZ on ARM.
721 setOperationAction(ISD::CTTZ_ZERO_UNDEF , MVT::i32 , Expand);
722 setOperationAction(ISD::CTLZ_ZERO_UNDEF , MVT::i32 , Expand);
724 setOperationAction(ISD::READCYCLECOUNTER, MVT::i64, Custom);
726 // Only ARMv6 has BSWAP.
727 if (!Subtarget->hasV6Ops())
728 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
730 if (!(Subtarget->hasDivide() && Subtarget->isThumb2()) &&
731 !(Subtarget->hasDivideInARMMode() && !Subtarget->isThumb())) {
732 // These are expanded into libcalls if the cpu doesn't have HW divider.
733 setOperationAction(ISD::SDIV, MVT::i32, Expand);
734 setOperationAction(ISD::UDIV, MVT::i32, Expand);
737 // FIXME: Also set divmod for SREM on EABI/androideabi
738 setOperationAction(ISD::SREM, MVT::i32, Expand);
739 setOperationAction(ISD::UREM, MVT::i32, Expand);
740 // Register based DivRem for AEABI (RTABI 4.2)
741 if (Subtarget->isTargetAEABI() || Subtarget->isTargetAndroid()) {
742 setLibcallName(RTLIB::SDIVREM_I8, "__aeabi_idivmod");
743 setLibcallName(RTLIB::SDIVREM_I16, "__aeabi_idivmod");
744 setLibcallName(RTLIB::SDIVREM_I32, "__aeabi_idivmod");
745 setLibcallName(RTLIB::SDIVREM_I64, "__aeabi_ldivmod");
746 setLibcallName(RTLIB::UDIVREM_I8, "__aeabi_uidivmod");
747 setLibcallName(RTLIB::UDIVREM_I16, "__aeabi_uidivmod");
748 setLibcallName(RTLIB::UDIVREM_I32, "__aeabi_uidivmod");
749 setLibcallName(RTLIB::UDIVREM_I64, "__aeabi_uldivmod");
751 setLibcallCallingConv(RTLIB::SDIVREM_I8, CallingConv::ARM_AAPCS);
752 setLibcallCallingConv(RTLIB::SDIVREM_I16, CallingConv::ARM_AAPCS);
753 setLibcallCallingConv(RTLIB::SDIVREM_I32, CallingConv::ARM_AAPCS);
754 setLibcallCallingConv(RTLIB::SDIVREM_I64, CallingConv::ARM_AAPCS);
755 setLibcallCallingConv(RTLIB::UDIVREM_I8, CallingConv::ARM_AAPCS);
756 setLibcallCallingConv(RTLIB::UDIVREM_I16, CallingConv::ARM_AAPCS);
757 setLibcallCallingConv(RTLIB::UDIVREM_I32, CallingConv::ARM_AAPCS);
758 setLibcallCallingConv(RTLIB::UDIVREM_I64, CallingConv::ARM_AAPCS);
760 setOperationAction(ISD::SDIVREM, MVT::i32, Custom);
761 setOperationAction(ISD::UDIVREM, MVT::i32, Custom);
763 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
764 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
767 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
768 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
769 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
770 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
771 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
773 setOperationAction(ISD::TRAP, MVT::Other, Legal);
775 // Use the default implementation.
776 setOperationAction(ISD::VASTART, MVT::Other, Custom);
777 setOperationAction(ISD::VAARG, MVT::Other, Expand);
778 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
779 setOperationAction(ISD::VAEND, MVT::Other, Expand);
780 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
781 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
783 if (!Subtarget->isTargetMachO()) {
784 // Non-MachO platforms may return values in these registers via the
785 // personality function.
786 setExceptionPointerRegister(ARM::R0);
787 setExceptionSelectorRegister(ARM::R1);
790 if (Subtarget->getTargetTriple().isWindowsItaniumEnvironment())
791 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
793 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
795 // ARMv6 Thumb1 (except for CPUs that support dmb / dsb) and earlier use
796 // the default expansion. If we are targeting a single threaded system,
797 // then set them all for expand so we can lower them later into their
799 if (TM.Options.ThreadModel == ThreadModel::Single)
800 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Expand);
801 else if (Subtarget->hasAnyDataBarrier() && !Subtarget->isThumb1Only()) {
802 // ATOMIC_FENCE needs custom lowering; the others should have been expanded
803 // to ldrex/strex loops already.
804 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Custom);
806 // On v8, we have particularly efficient implementations of atomic fences
807 // if they can be combined with nearby atomic loads and stores.
808 if (!Subtarget->hasV8Ops()) {
809 // Automatically insert fences (dmb ish) around ATOMIC_SWAP etc.
810 setInsertFencesForAtomic(true);
813 // If there's anything we can use as a barrier, go through custom lowering
815 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other,
816 Subtarget->hasAnyDataBarrier() ? Custom : Expand);
818 // Set them all for expansion, which will force libcalls.
819 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Expand);
820 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Expand);
821 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Expand);
822 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Expand);
823 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Expand);
824 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Expand);
825 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Expand);
826 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Expand);
827 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i32, Expand);
828 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i32, Expand);
829 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i32, Expand);
830 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i32, Expand);
831 // Mark ATOMIC_LOAD and ATOMIC_STORE custom so we can handle the
832 // Unordered/Monotonic case.
833 setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Custom);
834 setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Custom);
837 setOperationAction(ISD::PREFETCH, MVT::Other, Custom);
839 // Requires SXTB/SXTH, available on v6 and up in both ARM and Thumb modes.
840 if (!Subtarget->hasV6Ops()) {
841 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
842 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
844 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
846 if (!Subtarget->useSoftFloat() && Subtarget->hasVFP2() &&
847 !Subtarget->isThumb1Only()) {
848 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
849 // iff target supports vfp2.
850 setOperationAction(ISD::BITCAST, MVT::i64, Custom);
851 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
854 // We want to custom lower some of our intrinsics.
855 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
856 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
857 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
858 setOperationAction(ISD::EH_SJLJ_SETUP_DISPATCH, MVT::Other, Custom);
859 if (Subtarget->isTargetDarwin())
860 setLibcallName(RTLIB::UNWIND_RESUME, "_Unwind_SjLj_Resume");
862 setOperationAction(ISD::SETCC, MVT::i32, Expand);
863 setOperationAction(ISD::SETCC, MVT::f32, Expand);
864 setOperationAction(ISD::SETCC, MVT::f64, Expand);
865 setOperationAction(ISD::SELECT, MVT::i32, Custom);
866 setOperationAction(ISD::SELECT, MVT::f32, Custom);
867 setOperationAction(ISD::SELECT, MVT::f64, Custom);
868 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
869 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
870 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
872 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
873 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
874 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
875 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
876 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
878 // We don't support sin/cos/fmod/copysign/pow
879 setOperationAction(ISD::FSIN, MVT::f64, Expand);
880 setOperationAction(ISD::FSIN, MVT::f32, Expand);
881 setOperationAction(ISD::FCOS, MVT::f32, Expand);
882 setOperationAction(ISD::FCOS, MVT::f64, Expand);
883 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
884 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
885 setOperationAction(ISD::FREM, MVT::f64, Expand);
886 setOperationAction(ISD::FREM, MVT::f32, Expand);
887 if (!Subtarget->useSoftFloat() && Subtarget->hasVFP2() &&
888 !Subtarget->isThumb1Only()) {
889 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
890 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
892 setOperationAction(ISD::FPOW, MVT::f64, Expand);
893 setOperationAction(ISD::FPOW, MVT::f32, Expand);
895 if (!Subtarget->hasVFP4()) {
896 setOperationAction(ISD::FMA, MVT::f64, Expand);
897 setOperationAction(ISD::FMA, MVT::f32, Expand);
900 // Various VFP goodness
901 if (!Subtarget->useSoftFloat() && !Subtarget->isThumb1Only()) {
902 // FP-ARMv8 adds f64 <-> f16 conversion. Before that it should be expanded.
903 if (!Subtarget->hasFPARMv8() || Subtarget->isFPOnlySP()) {
904 setOperationAction(ISD::FP16_TO_FP, MVT::f64, Expand);
905 setOperationAction(ISD::FP_TO_FP16, MVT::f64, Expand);
908 // fp16 is a special v7 extension that adds f16 <-> f32 conversions.
909 if (!Subtarget->hasFP16()) {
910 setOperationAction(ISD::FP16_TO_FP, MVT::f32, Expand);
911 setOperationAction(ISD::FP_TO_FP16, MVT::f32, Expand);
915 // Combine sin / cos into one node or libcall if possible.
916 if (Subtarget->hasSinCos()) {
917 setLibcallName(RTLIB::SINCOS_F32, "sincosf");
918 setLibcallName(RTLIB::SINCOS_F64, "sincos");
919 if (Subtarget->getTargetTriple().isiOS()) {
920 // For iOS, we don't want to the normal expansion of a libcall to
921 // sincos. We want to issue a libcall to __sincos_stret.
922 setOperationAction(ISD::FSINCOS, MVT::f64, Custom);
923 setOperationAction(ISD::FSINCOS, MVT::f32, Custom);
927 // FP-ARMv8 implements a lot of rounding-like FP operations.
928 if (Subtarget->hasFPARMv8()) {
929 setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
930 setOperationAction(ISD::FCEIL, MVT::f32, Legal);
931 setOperationAction(ISD::FROUND, MVT::f32, Legal);
932 setOperationAction(ISD::FTRUNC, MVT::f32, Legal);
933 setOperationAction(ISD::FNEARBYINT, MVT::f32, Legal);
934 setOperationAction(ISD::FRINT, MVT::f32, Legal);
935 setOperationAction(ISD::FMINNUM, MVT::f32, Legal);
936 setOperationAction(ISD::FMAXNUM, MVT::f32, Legal);
937 if (!Subtarget->isFPOnlySP()) {
938 setOperationAction(ISD::FFLOOR, MVT::f64, Legal);
939 setOperationAction(ISD::FCEIL, MVT::f64, Legal);
940 setOperationAction(ISD::FROUND, MVT::f64, Legal);
941 setOperationAction(ISD::FTRUNC, MVT::f64, Legal);
942 setOperationAction(ISD::FNEARBYINT, MVT::f64, Legal);
943 setOperationAction(ISD::FRINT, MVT::f64, Legal);
944 setOperationAction(ISD::FMINNUM, MVT::f64, Legal);
945 setOperationAction(ISD::FMAXNUM, MVT::f64, Legal);
949 if (Subtarget->hasVFP3()) {
950 setOperationAction(ISD::FMINNAN, MVT::f32, Legal);
951 setOperationAction(ISD::FMAXNAN, MVT::f32, Legal);
952 setOperationAction(ISD::FMINNAN, MVT::f64, Legal);
953 setOperationAction(ISD::FMAXNAN, MVT::f64, Legal);
956 // We have target-specific dag combine patterns for the following nodes:
957 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
958 setTargetDAGCombine(ISD::ADD);
959 setTargetDAGCombine(ISD::SUB);
960 setTargetDAGCombine(ISD::MUL);
961 setTargetDAGCombine(ISD::AND);
962 setTargetDAGCombine(ISD::OR);
963 setTargetDAGCombine(ISD::XOR);
965 if (Subtarget->hasV6Ops())
966 setTargetDAGCombine(ISD::SRL);
968 setStackPointerRegisterToSaveRestore(ARM::SP);
970 if (Subtarget->useSoftFloat() || Subtarget->isThumb1Only() ||
971 !Subtarget->hasVFP2())
972 setSchedulingPreference(Sched::RegPressure);
974 setSchedulingPreference(Sched::Hybrid);
976 //// temporary - rewrite interface to use type
977 MaxStoresPerMemset = 8;
978 MaxStoresPerMemsetOptSize = 4;
979 MaxStoresPerMemcpy = 4; // For @llvm.memcpy -> sequence of stores
980 MaxStoresPerMemcpyOptSize = 2;
981 MaxStoresPerMemmove = 4; // For @llvm.memmove -> sequence of stores
982 MaxStoresPerMemmoveOptSize = 2;
984 // On ARM arguments smaller than 4 bytes are extended, so all arguments
985 // are at least 4 bytes aligned.
986 setMinStackArgumentAlignment(4);
988 // Prefer likely predicted branches to selects on out-of-order cores.
989 PredictableSelectIsExpensive = Subtarget->isLikeA9();
991 setMinFunctionAlignment(Subtarget->isThumb() ? 1 : 2);
994 bool ARMTargetLowering::useSoftFloat() const {
995 return Subtarget->useSoftFloat();
998 // FIXME: It might make sense to define the representative register class as the
999 // nearest super-register that has a non-null superset. For example, DPR_VFP2 is
1000 // a super-register of SPR, and DPR is a superset if DPR_VFP2. Consequently,
1001 // SPR's representative would be DPR_VFP2. This should work well if register
1002 // pressure tracking were modified such that a register use would increment the
1003 // pressure of the register class's representative and all of it's super
1004 // classes' representatives transitively. We have not implemented this because
1005 // of the difficulty prior to coalescing of modeling operand register classes
1006 // due to the common occurrence of cross class copies and subregister insertions
1008 std::pair<const TargetRegisterClass *, uint8_t>
1009 ARMTargetLowering::findRepresentativeClass(const TargetRegisterInfo *TRI,
1011 const TargetRegisterClass *RRC = nullptr;
1013 switch (VT.SimpleTy) {
1015 return TargetLowering::findRepresentativeClass(TRI, VT);
1016 // Use DPR as representative register class for all floating point
1017 // and vector types. Since there are 32 SPR registers and 32 DPR registers so
1018 // the cost is 1 for both f32 and f64.
1019 case MVT::f32: case MVT::f64: case MVT::v8i8: case MVT::v4i16:
1020 case MVT::v2i32: case MVT::v1i64: case MVT::v2f32:
1021 RRC = &ARM::DPRRegClass;
1022 // When NEON is used for SP, only half of the register file is available
1023 // because operations that define both SP and DP results will be constrained
1024 // to the VFP2 class (D0-D15). We currently model this constraint prior to
1025 // coalescing by double-counting the SP regs. See the FIXME above.
1026 if (Subtarget->useNEONForSinglePrecisionFP())
1029 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
1030 case MVT::v4f32: case MVT::v2f64:
1031 RRC = &ARM::DPRRegClass;
1035 RRC = &ARM::DPRRegClass;
1039 RRC = &ARM::DPRRegClass;
1043 return std::make_pair(RRC, Cost);
1046 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
1047 switch ((ARMISD::NodeType)Opcode) {
1048 case ARMISD::FIRST_NUMBER: break;
1049 case ARMISD::Wrapper: return "ARMISD::Wrapper";
1050 case ARMISD::WrapperPIC: return "ARMISD::WrapperPIC";
1051 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
1052 case ARMISD::COPY_STRUCT_BYVAL: return "ARMISD::COPY_STRUCT_BYVAL";
1053 case ARMISD::CALL: return "ARMISD::CALL";
1054 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
1055 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
1056 case ARMISD::tCALL: return "ARMISD::tCALL";
1057 case ARMISD::BRCOND: return "ARMISD::BRCOND";
1058 case ARMISD::BR_JT: return "ARMISD::BR_JT";
1059 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
1060 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
1061 case ARMISD::INTRET_FLAG: return "ARMISD::INTRET_FLAG";
1062 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
1063 case ARMISD::CMP: return "ARMISD::CMP";
1064 case ARMISD::CMN: return "ARMISD::CMN";
1065 case ARMISD::CMPZ: return "ARMISD::CMPZ";
1066 case ARMISD::CMPFP: return "ARMISD::CMPFP";
1067 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
1068 case ARMISD::BCC_i64: return "ARMISD::BCC_i64";
1069 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
1071 case ARMISD::CMOV: return "ARMISD::CMOV";
1073 case ARMISD::RBIT: return "ARMISD::RBIT";
1075 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
1076 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
1077 case ARMISD::RRX: return "ARMISD::RRX";
1079 case ARMISD::ADDC: return "ARMISD::ADDC";
1080 case ARMISD::ADDE: return "ARMISD::ADDE";
1081 case ARMISD::SUBC: return "ARMISD::SUBC";
1082 case ARMISD::SUBE: return "ARMISD::SUBE";
1084 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
1085 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
1087 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
1088 case ARMISD::EH_SJLJ_LONGJMP: return "ARMISD::EH_SJLJ_LONGJMP";
1089 case ARMISD::EH_SJLJ_SETUP_DISPATCH: return "ARMISD::EH_SJLJ_SETUP_DISPATCH";
1091 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN";
1093 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
1095 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
1097 case ARMISD::MEMBARRIER_MCR: return "ARMISD::MEMBARRIER_MCR";
1099 case ARMISD::PRELOAD: return "ARMISD::PRELOAD";
1101 case ARMISD::WIN__CHKSTK: return "ARMISD:::WIN__CHKSTK";
1103 case ARMISD::VCEQ: return "ARMISD::VCEQ";
1104 case ARMISD::VCEQZ: return "ARMISD::VCEQZ";
1105 case ARMISD::VCGE: return "ARMISD::VCGE";
1106 case ARMISD::VCGEZ: return "ARMISD::VCGEZ";
1107 case ARMISD::VCLEZ: return "ARMISD::VCLEZ";
1108 case ARMISD::VCGEU: return "ARMISD::VCGEU";
1109 case ARMISD::VCGT: return "ARMISD::VCGT";
1110 case ARMISD::VCGTZ: return "ARMISD::VCGTZ";
1111 case ARMISD::VCLTZ: return "ARMISD::VCLTZ";
1112 case ARMISD::VCGTU: return "ARMISD::VCGTU";
1113 case ARMISD::VTST: return "ARMISD::VTST";
1115 case ARMISD::VSHL: return "ARMISD::VSHL";
1116 case ARMISD::VSHRs: return "ARMISD::VSHRs";
1117 case ARMISD::VSHRu: return "ARMISD::VSHRu";
1118 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
1119 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
1120 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
1121 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
1122 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
1123 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
1124 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
1125 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
1126 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
1127 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
1128 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
1129 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
1130 case ARMISD::VSLI: return "ARMISD::VSLI";
1131 case ARMISD::VSRI: return "ARMISD::VSRI";
1132 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
1133 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
1134 case ARMISD::VMOVIMM: return "ARMISD::VMOVIMM";
1135 case ARMISD::VMVNIMM: return "ARMISD::VMVNIMM";
1136 case ARMISD::VMOVFPIMM: return "ARMISD::VMOVFPIMM";
1137 case ARMISD::VDUP: return "ARMISD::VDUP";
1138 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
1139 case ARMISD::VEXT: return "ARMISD::VEXT";
1140 case ARMISD::VREV64: return "ARMISD::VREV64";
1141 case ARMISD::VREV32: return "ARMISD::VREV32";
1142 case ARMISD::VREV16: return "ARMISD::VREV16";
1143 case ARMISD::VZIP: return "ARMISD::VZIP";
1144 case ARMISD::VUZP: return "ARMISD::VUZP";
1145 case ARMISD::VTRN: return "ARMISD::VTRN";
1146 case ARMISD::VTBL1: return "ARMISD::VTBL1";
1147 case ARMISD::VTBL2: return "ARMISD::VTBL2";
1148 case ARMISD::VMULLs: return "ARMISD::VMULLs";
1149 case ARMISD::VMULLu: return "ARMISD::VMULLu";
1150 case ARMISD::UMLAL: return "ARMISD::UMLAL";
1151 case ARMISD::SMLAL: return "ARMISD::SMLAL";
1152 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
1153 case ARMISD::BFI: return "ARMISD::BFI";
1154 case ARMISD::VORRIMM: return "ARMISD::VORRIMM";
1155 case ARMISD::VBICIMM: return "ARMISD::VBICIMM";
1156 case ARMISD::VBSL: return "ARMISD::VBSL";
1157 case ARMISD::VLD2DUP: return "ARMISD::VLD2DUP";
1158 case ARMISD::VLD3DUP: return "ARMISD::VLD3DUP";
1159 case ARMISD::VLD4DUP: return "ARMISD::VLD4DUP";
1160 case ARMISD::VLD1_UPD: return "ARMISD::VLD1_UPD";
1161 case ARMISD::VLD2_UPD: return "ARMISD::VLD2_UPD";
1162 case ARMISD::VLD3_UPD: return "ARMISD::VLD3_UPD";
1163 case ARMISD::VLD4_UPD: return "ARMISD::VLD4_UPD";
1164 case ARMISD::VLD2LN_UPD: return "ARMISD::VLD2LN_UPD";
1165 case ARMISD::VLD3LN_UPD: return "ARMISD::VLD3LN_UPD";
1166 case ARMISD::VLD4LN_UPD: return "ARMISD::VLD4LN_UPD";
1167 case ARMISD::VLD2DUP_UPD: return "ARMISD::VLD2DUP_UPD";
1168 case ARMISD::VLD3DUP_UPD: return "ARMISD::VLD3DUP_UPD";
1169 case ARMISD::VLD4DUP_UPD: return "ARMISD::VLD4DUP_UPD";
1170 case ARMISD::VST1_UPD: return "ARMISD::VST1_UPD";
1171 case ARMISD::VST2_UPD: return "ARMISD::VST2_UPD";
1172 case ARMISD::VST3_UPD: return "ARMISD::VST3_UPD";
1173 case ARMISD::VST4_UPD: return "ARMISD::VST4_UPD";
1174 case ARMISD::VST2LN_UPD: return "ARMISD::VST2LN_UPD";
1175 case ARMISD::VST3LN_UPD: return "ARMISD::VST3LN_UPD";
1176 case ARMISD::VST4LN_UPD: return "ARMISD::VST4LN_UPD";
1181 EVT ARMTargetLowering::getSetCCResultType(const DataLayout &DL, LLVMContext &,
1184 return getPointerTy(DL);
1185 return VT.changeVectorElementTypeToInteger();
1188 /// getRegClassFor - Return the register class that should be used for the
1189 /// specified value type.
1190 const TargetRegisterClass *ARMTargetLowering::getRegClassFor(MVT VT) const {
1191 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
1192 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
1193 // load / store 4 to 8 consecutive D registers.
1194 if (Subtarget->hasNEON()) {
1195 if (VT == MVT::v4i64)
1196 return &ARM::QQPRRegClass;
1197 if (VT == MVT::v8i64)
1198 return &ARM::QQQQPRRegClass;
1200 return TargetLowering::getRegClassFor(VT);
1203 // memcpy, and other memory intrinsics, typically tries to use LDM/STM if the
1204 // source/dest is aligned and the copy size is large enough. We therefore want
1205 // to align such objects passed to memory intrinsics.
1206 bool ARMTargetLowering::shouldAlignPointerArgs(CallInst *CI, unsigned &MinSize,
1207 unsigned &PrefAlign) const {
1208 if (!isa<MemIntrinsic>(CI))
1211 // On ARM11 onwards (excluding M class) 8-byte aligned LDM is typically 1
1212 // cycle faster than 4-byte aligned LDM.
1213 PrefAlign = (Subtarget->hasV6Ops() && !Subtarget->isMClass() ? 8 : 4);
1217 // Create a fast isel object.
1219 ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
1220 const TargetLibraryInfo *libInfo) const {
1221 return ARM::createFastISel(funcInfo, libInfo);
1224 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
1225 unsigned NumVals = N->getNumValues();
1227 return Sched::RegPressure;
1229 for (unsigned i = 0; i != NumVals; ++i) {
1230 EVT VT = N->getValueType(i);
1231 if (VT == MVT::Glue || VT == MVT::Other)
1233 if (VT.isFloatingPoint() || VT.isVector())
1237 if (!N->isMachineOpcode())
1238 return Sched::RegPressure;
1240 // Load are scheduled for latency even if there instruction itinerary
1241 // is not available.
1242 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
1243 const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
1245 if (MCID.getNumDefs() == 0)
1246 return Sched::RegPressure;
1247 if (!Itins->isEmpty() &&
1248 Itins->getOperandCycle(MCID.getSchedClass(), 0) > 2)
1251 return Sched::RegPressure;
1254 //===----------------------------------------------------------------------===//
1256 //===----------------------------------------------------------------------===//
1258 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
1259 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
1261 default: llvm_unreachable("Unknown condition code!");
1262 case ISD::SETNE: return ARMCC::NE;
1263 case ISD::SETEQ: return ARMCC::EQ;
1264 case ISD::SETGT: return ARMCC::GT;
1265 case ISD::SETGE: return ARMCC::GE;
1266 case ISD::SETLT: return ARMCC::LT;
1267 case ISD::SETLE: return ARMCC::LE;
1268 case ISD::SETUGT: return ARMCC::HI;
1269 case ISD::SETUGE: return ARMCC::HS;
1270 case ISD::SETULT: return ARMCC::LO;
1271 case ISD::SETULE: return ARMCC::LS;
1275 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
1276 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
1277 ARMCC::CondCodes &CondCode2) {
1278 CondCode2 = ARMCC::AL;
1280 default: llvm_unreachable("Unknown FP condition!");
1282 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
1284 case ISD::SETOGT: CondCode = ARMCC::GT; break;
1286 case ISD::SETOGE: CondCode = ARMCC::GE; break;
1287 case ISD::SETOLT: CondCode = ARMCC::MI; break;
1288 case ISD::SETOLE: CondCode = ARMCC::LS; break;
1289 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
1290 case ISD::SETO: CondCode = ARMCC::VC; break;
1291 case ISD::SETUO: CondCode = ARMCC::VS; break;
1292 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
1293 case ISD::SETUGT: CondCode = ARMCC::HI; break;
1294 case ISD::SETUGE: CondCode = ARMCC::PL; break;
1296 case ISD::SETULT: CondCode = ARMCC::LT; break;
1298 case ISD::SETULE: CondCode = ARMCC::LE; break;
1300 case ISD::SETUNE: CondCode = ARMCC::NE; break;
1304 //===----------------------------------------------------------------------===//
1305 // Calling Convention Implementation
1306 //===----------------------------------------------------------------------===//
1308 #include "ARMGenCallingConv.inc"
1310 /// getEffectiveCallingConv - Get the effective calling convention, taking into
1311 /// account presence of floating point hardware and calling convention
1312 /// limitations, such as support for variadic functions.
1314 ARMTargetLowering::getEffectiveCallingConv(CallingConv::ID CC,
1315 bool isVarArg) const {
1318 llvm_unreachable("Unsupported calling convention");
1319 case CallingConv::ARM_AAPCS:
1320 case CallingConv::ARM_APCS:
1321 case CallingConv::GHC:
1323 case CallingConv::ARM_AAPCS_VFP:
1324 return isVarArg ? CallingConv::ARM_AAPCS : CallingConv::ARM_AAPCS_VFP;
1325 case CallingConv::C:
1326 if (!Subtarget->isAAPCS_ABI())
1327 return CallingConv::ARM_APCS;
1328 else if (Subtarget->hasVFP2() && !Subtarget->isThumb1Only() &&
1329 getTargetMachine().Options.FloatABIType == FloatABI::Hard &&
1331 return CallingConv::ARM_AAPCS_VFP;
1333 return CallingConv::ARM_AAPCS;
1334 case CallingConv::Fast:
1335 if (!Subtarget->isAAPCS_ABI()) {
1336 if (Subtarget->hasVFP2() && !Subtarget->isThumb1Only() && !isVarArg)
1337 return CallingConv::Fast;
1338 return CallingConv::ARM_APCS;
1339 } else if (Subtarget->hasVFP2() && !Subtarget->isThumb1Only() && !isVarArg)
1340 return CallingConv::ARM_AAPCS_VFP;
1342 return CallingConv::ARM_AAPCS;
1346 /// CCAssignFnForNode - Selects the correct CCAssignFn for the given
1347 /// CallingConvention.
1348 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
1350 bool isVarArg) const {
1351 switch (getEffectiveCallingConv(CC, isVarArg)) {
1353 llvm_unreachable("Unsupported calling convention");
1354 case CallingConv::ARM_APCS:
1355 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1356 case CallingConv::ARM_AAPCS:
1357 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1358 case CallingConv::ARM_AAPCS_VFP:
1359 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1360 case CallingConv::Fast:
1361 return (Return ? RetFastCC_ARM_APCS : FastCC_ARM_APCS);
1362 case CallingConv::GHC:
1363 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS_GHC);
1367 /// LowerCallResult - Lower the result values of a call into the
1368 /// appropriate copies out of appropriate physical registers.
1370 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1371 CallingConv::ID CallConv, bool isVarArg,
1372 const SmallVectorImpl<ISD::InputArg> &Ins,
1373 SDLoc dl, SelectionDAG &DAG,
1374 SmallVectorImpl<SDValue> &InVals,
1375 bool isThisReturn, SDValue ThisVal) const {
1377 // Assign locations to each value returned by this call.
1378 SmallVector<CCValAssign, 16> RVLocs;
1379 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
1380 *DAG.getContext(), Call);
1381 CCInfo.AnalyzeCallResult(Ins,
1382 CCAssignFnForNode(CallConv, /* Return*/ true,
1385 // Copy all of the result registers out of their specified physreg.
1386 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1387 CCValAssign VA = RVLocs[i];
1389 // Pass 'this' value directly from the argument to return value, to avoid
1390 // reg unit interference
1391 if (i == 0 && isThisReturn) {
1392 assert(!VA.needsCustom() && VA.getLocVT() == MVT::i32 &&
1393 "unexpected return calling convention register assignment");
1394 InVals.push_back(ThisVal);
1399 if (VA.needsCustom()) {
1400 // Handle f64 or half of a v2f64.
1401 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1403 Chain = Lo.getValue(1);
1404 InFlag = Lo.getValue(2);
1405 VA = RVLocs[++i]; // skip ahead to next loc
1406 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1408 Chain = Hi.getValue(1);
1409 InFlag = Hi.getValue(2);
1410 if (!Subtarget->isLittle())
1412 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1414 if (VA.getLocVT() == MVT::v2f64) {
1415 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1416 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1417 DAG.getConstant(0, dl, MVT::i32));
1419 VA = RVLocs[++i]; // skip ahead to next loc
1420 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1421 Chain = Lo.getValue(1);
1422 InFlag = Lo.getValue(2);
1423 VA = RVLocs[++i]; // skip ahead to next loc
1424 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1425 Chain = Hi.getValue(1);
1426 InFlag = Hi.getValue(2);
1427 if (!Subtarget->isLittle())
1429 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1430 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1431 DAG.getConstant(1, dl, MVT::i32));
1434 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
1436 Chain = Val.getValue(1);
1437 InFlag = Val.getValue(2);
1440 switch (VA.getLocInfo()) {
1441 default: llvm_unreachable("Unknown loc info!");
1442 case CCValAssign::Full: break;
1443 case CCValAssign::BCvt:
1444 Val = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), Val);
1448 InVals.push_back(Val);
1454 /// LowerMemOpCallTo - Store the argument to the stack.
1456 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
1457 SDValue StackPtr, SDValue Arg,
1458 SDLoc dl, SelectionDAG &DAG,
1459 const CCValAssign &VA,
1460 ISD::ArgFlagsTy Flags) const {
1461 unsigned LocMemOffset = VA.getLocMemOffset();
1462 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset, dl);
1463 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(DAG.getDataLayout()),
1465 return DAG.getStore(Chain, dl, Arg, PtrOff,
1466 MachinePointerInfo::getStack(LocMemOffset),
1470 void ARMTargetLowering::PassF64ArgInRegs(SDLoc dl, SelectionDAG &DAG,
1471 SDValue Chain, SDValue &Arg,
1472 RegsToPassVector &RegsToPass,
1473 CCValAssign &VA, CCValAssign &NextVA,
1475 SmallVectorImpl<SDValue> &MemOpChains,
1476 ISD::ArgFlagsTy Flags) const {
1478 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1479 DAG.getVTList(MVT::i32, MVT::i32), Arg);
1480 unsigned id = Subtarget->isLittle() ? 0 : 1;
1481 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd.getValue(id)));
1483 if (NextVA.isRegLoc())
1484 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1-id)));
1486 assert(NextVA.isMemLoc());
1487 if (!StackPtr.getNode())
1488 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP,
1489 getPointerTy(DAG.getDataLayout()));
1491 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1-id),
1497 /// LowerCall - Lowering a call into a callseq_start <-
1498 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
1501 ARMTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
1502 SmallVectorImpl<SDValue> &InVals) const {
1503 SelectionDAG &DAG = CLI.DAG;
1505 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
1506 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
1507 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
1508 SDValue Chain = CLI.Chain;
1509 SDValue Callee = CLI.Callee;
1510 bool &isTailCall = CLI.IsTailCall;
1511 CallingConv::ID CallConv = CLI.CallConv;
1512 bool doesNotRet = CLI.DoesNotReturn;
1513 bool isVarArg = CLI.IsVarArg;
1515 MachineFunction &MF = DAG.getMachineFunction();
1516 bool isStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1517 bool isThisReturn = false;
1518 bool isSibCall = false;
1519 auto Attr = MF.getFunction()->getFnAttribute("disable-tail-calls");
1521 // Disable tail calls if they're not supported.
1522 if (!Subtarget->supportsTailCall() || Attr.getValueAsString() == "true")
1526 // Check if it's really possible to do a tail call.
1527 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1528 isVarArg, isStructRet, MF.getFunction()->hasStructRetAttr(),
1529 Outs, OutVals, Ins, DAG);
1530 if (!isTailCall && CLI.CS && CLI.CS->isMustTailCall())
1531 report_fatal_error("failed to perform tail call elimination on a call "
1532 "site marked musttail");
1533 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1534 // detected sibcalls.
1541 // Analyze operands of the call, assigning locations to each operand.
1542 SmallVector<CCValAssign, 16> ArgLocs;
1543 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs,
1544 *DAG.getContext(), Call);
1545 CCInfo.AnalyzeCallOperands(Outs,
1546 CCAssignFnForNode(CallConv, /* Return*/ false,
1549 // Get a count of how many bytes are to be pushed on the stack.
1550 unsigned NumBytes = CCInfo.getNextStackOffset();
1552 // For tail calls, memory operands are available in our caller's stack.
1556 // Adjust the stack pointer for the new arguments...
1557 // These operations are automatically eliminated by the prolog/epilog pass
1559 Chain = DAG.getCALLSEQ_START(Chain,
1560 DAG.getIntPtrConstant(NumBytes, dl, true), dl);
1563 DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy(DAG.getDataLayout()));
1565 RegsToPassVector RegsToPass;
1566 SmallVector<SDValue, 8> MemOpChains;
1568 // Walk the register/memloc assignments, inserting copies/loads. In the case
1569 // of tail call optimization, arguments are handled later.
1570 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1572 ++i, ++realArgIdx) {
1573 CCValAssign &VA = ArgLocs[i];
1574 SDValue Arg = OutVals[realArgIdx];
1575 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1576 bool isByVal = Flags.isByVal();
1578 // Promote the value if needed.
1579 switch (VA.getLocInfo()) {
1580 default: llvm_unreachable("Unknown loc info!");
1581 case CCValAssign::Full: break;
1582 case CCValAssign::SExt:
1583 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1585 case CCValAssign::ZExt:
1586 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1588 case CCValAssign::AExt:
1589 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1591 case CCValAssign::BCvt:
1592 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1596 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1597 if (VA.needsCustom()) {
1598 if (VA.getLocVT() == MVT::v2f64) {
1599 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1600 DAG.getConstant(0, dl, MVT::i32));
1601 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1602 DAG.getConstant(1, dl, MVT::i32));
1604 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1605 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1607 VA = ArgLocs[++i]; // skip ahead to next loc
1608 if (VA.isRegLoc()) {
1609 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1610 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1612 assert(VA.isMemLoc());
1614 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1615 dl, DAG, VA, Flags));
1618 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1619 StackPtr, MemOpChains, Flags);
1621 } else if (VA.isRegLoc()) {
1622 if (realArgIdx == 0 && Flags.isReturned() && Outs[0].VT == MVT::i32) {
1623 assert(VA.getLocVT() == MVT::i32 &&
1624 "unexpected calling convention register assignment");
1625 assert(!Ins.empty() && Ins[0].VT == MVT::i32 &&
1626 "unexpected use of 'returned'");
1627 isThisReturn = true;
1629 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1630 } else if (isByVal) {
1631 assert(VA.isMemLoc());
1632 unsigned offset = 0;
1634 // True if this byval aggregate will be split between registers
1636 unsigned ByValArgsCount = CCInfo.getInRegsParamsCount();
1637 unsigned CurByValIdx = CCInfo.getInRegsParamsProcessed();
1639 if (CurByValIdx < ByValArgsCount) {
1641 unsigned RegBegin, RegEnd;
1642 CCInfo.getInRegsParamInfo(CurByValIdx, RegBegin, RegEnd);
1645 DAG.getTargetLoweringInfo().getPointerTy(DAG.getDataLayout());
1647 for (i = 0, j = RegBegin; j < RegEnd; i++, j++) {
1648 SDValue Const = DAG.getConstant(4*i, dl, MVT::i32);
1649 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
1650 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
1651 MachinePointerInfo(),
1652 false, false, false,
1653 DAG.InferPtrAlignment(AddArg));
1654 MemOpChains.push_back(Load.getValue(1));
1655 RegsToPass.push_back(std::make_pair(j, Load));
1658 // If parameter size outsides register area, "offset" value
1659 // helps us to calculate stack slot for remained part properly.
1660 offset = RegEnd - RegBegin;
1662 CCInfo.nextInRegsParam();
1665 if (Flags.getByValSize() > 4*offset) {
1666 auto PtrVT = getPointerTy(DAG.getDataLayout());
1667 unsigned LocMemOffset = VA.getLocMemOffset();
1668 SDValue StkPtrOff = DAG.getIntPtrConstant(LocMemOffset, dl);
1669 SDValue Dst = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, StkPtrOff);
1670 SDValue SrcOffset = DAG.getIntPtrConstant(4*offset, dl);
1671 SDValue Src = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, SrcOffset);
1672 SDValue SizeNode = DAG.getConstant(Flags.getByValSize() - 4*offset, dl,
1674 SDValue AlignNode = DAG.getConstant(Flags.getByValAlign(), dl,
1677 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
1678 SDValue Ops[] = { Chain, Dst, Src, SizeNode, AlignNode};
1679 MemOpChains.push_back(DAG.getNode(ARMISD::COPY_STRUCT_BYVAL, dl, VTs,
1682 } else if (!isSibCall) {
1683 assert(VA.isMemLoc());
1685 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1686 dl, DAG, VA, Flags));
1690 if (!MemOpChains.empty())
1691 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains);
1693 // Build a sequence of copy-to-reg nodes chained together with token chain
1694 // and flag operands which copy the outgoing args into the appropriate regs.
1696 // Tail call byval lowering might overwrite argument registers so in case of
1697 // tail call optimization the copies to registers are lowered later.
1699 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1700 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1701 RegsToPass[i].second, InFlag);
1702 InFlag = Chain.getValue(1);
1705 // For tail calls lower the arguments to the 'real' stack slot.
1707 // Force all the incoming stack arguments to be loaded from the stack
1708 // before any new outgoing arguments are stored to the stack, because the
1709 // outgoing stack slots may alias the incoming argument stack slots, and
1710 // the alias isn't otherwise explicit. This is slightly more conservative
1711 // than necessary, because it means that each store effectively depends
1712 // on every argument instead of just those arguments it would clobber.
1714 // Do not flag preceding copytoreg stuff together with the following stuff.
1716 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1717 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1718 RegsToPass[i].second, InFlag);
1719 InFlag = Chain.getValue(1);
1724 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1725 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1726 // node so that legalize doesn't hack it.
1727 bool isDirect = false;
1728 bool isARMFunc = false;
1729 bool isLocalARMFunc = false;
1730 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1731 auto PtrVt = getPointerTy(DAG.getDataLayout());
1733 if (Subtarget->genLongCalls()) {
1734 assert((Subtarget->isTargetWindows() ||
1735 getTargetMachine().getRelocationModel() == Reloc::Static) &&
1736 "long-calls with non-static relocation model!");
1737 // Handle a global address or an external symbol. If it's not one of
1738 // those, the target's already in a register, so we don't need to do
1740 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1741 const GlobalValue *GV = G->getGlobal();
1742 // Create a constant pool entry for the callee address
1743 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1744 ARMConstantPoolValue *CPV =
1745 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 0);
1747 // Get the address of the callee into a register
1748 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVt, 4);
1749 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1750 Callee = DAG.getLoad(PtrVt, dl, DAG.getEntryNode(), CPAddr,
1751 MachinePointerInfo::getConstantPool(), false, false,
1753 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1754 const char *Sym = S->getSymbol();
1756 // Create a constant pool entry for the callee address
1757 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1758 ARMConstantPoolValue *CPV =
1759 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1760 ARMPCLabelIndex, 0);
1761 // Get the address of the callee into a register
1762 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVt, 4);
1763 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1764 Callee = DAG.getLoad(PtrVt, dl, DAG.getEntryNode(), CPAddr,
1765 MachinePointerInfo::getConstantPool(), false, false,
1768 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1769 const GlobalValue *GV = G->getGlobal();
1771 bool isDef = GV->isStrongDefinitionForLinker();
1772 bool isStub = (!isDef && Subtarget->isTargetMachO()) &&
1773 getTargetMachine().getRelocationModel() != Reloc::Static;
1774 isARMFunc = !Subtarget->isThumb() || (isStub && !Subtarget->isMClass());
1775 // ARM call to a local ARM function is predicable.
1776 isLocalARMFunc = !Subtarget->isThumb() && (isDef || !ARMInterworking);
1777 // tBX takes a register source operand.
1778 if (isStub && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1779 assert(Subtarget->isTargetMachO() && "WrapperPIC use on non-MachO?");
1780 Callee = DAG.getNode(
1781 ARMISD::WrapperPIC, dl, PtrVt,
1782 DAG.getTargetGlobalAddress(GV, dl, PtrVt, 0, ARMII::MO_NONLAZY));
1783 Callee = DAG.getLoad(PtrVt, dl, DAG.getEntryNode(), Callee,
1784 MachinePointerInfo::getGOT(), false, false, true, 0);
1785 } else if (Subtarget->isTargetCOFF()) {
1786 assert(Subtarget->isTargetWindows() &&
1787 "Windows is the only supported COFF target");
1788 unsigned TargetFlags = GV->hasDLLImportStorageClass()
1789 ? ARMII::MO_DLLIMPORT
1790 : ARMII::MO_NO_FLAG;
1792 DAG.getTargetGlobalAddress(GV, dl, PtrVt, /*Offset=*/0, TargetFlags);
1793 if (GV->hasDLLImportStorageClass())
1795 DAG.getLoad(PtrVt, dl, DAG.getEntryNode(),
1796 DAG.getNode(ARMISD::Wrapper, dl, PtrVt, Callee),
1797 MachinePointerInfo::getGOT(), false, false, false, 0);
1799 // On ELF targets for PIC code, direct calls should go through the PLT
1800 unsigned OpFlags = 0;
1801 if (Subtarget->isTargetELF() &&
1802 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1803 OpFlags = ARMII::MO_PLT;
1804 Callee = DAG.getTargetGlobalAddress(GV, dl, PtrVt, 0, OpFlags);
1806 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1808 bool isStub = Subtarget->isTargetMachO() &&
1809 getTargetMachine().getRelocationModel() != Reloc::Static;
1810 isARMFunc = !Subtarget->isThumb() || (isStub && !Subtarget->isMClass());
1811 // tBX takes a register source operand.
1812 const char *Sym = S->getSymbol();
1813 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1814 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1815 ARMConstantPoolValue *CPV =
1816 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1817 ARMPCLabelIndex, 4);
1818 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVt, 4);
1819 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1820 Callee = DAG.getLoad(PtrVt, dl, DAG.getEntryNode(), CPAddr,
1821 MachinePointerInfo::getConstantPool(), false, false,
1823 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, dl, MVT::i32);
1824 Callee = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVt, Callee, PICLabel);
1826 unsigned OpFlags = 0;
1827 // On ELF targets for PIC code, direct calls should go through the PLT
1828 if (Subtarget->isTargetELF() &&
1829 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1830 OpFlags = ARMII::MO_PLT;
1831 Callee = DAG.getTargetExternalSymbol(Sym, PtrVt, OpFlags);
1835 // FIXME: handle tail calls differently.
1837 if (Subtarget->isThumb()) {
1838 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1839 CallOpc = ARMISD::CALL_NOLINK;
1841 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1843 if (!isDirect && !Subtarget->hasV5TOps())
1844 CallOpc = ARMISD::CALL_NOLINK;
1845 else if (doesNotRet && isDirect && Subtarget->hasRAS() &&
1846 // Emit regular call when code size is the priority
1847 !MF.getFunction()->optForMinSize())
1848 // "mov lr, pc; b _foo" to avoid confusing the RSP
1849 CallOpc = ARMISD::CALL_NOLINK;
1851 CallOpc = isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL;
1854 std::vector<SDValue> Ops;
1855 Ops.push_back(Chain);
1856 Ops.push_back(Callee);
1858 // Add argument registers to the end of the list so that they are known live
1860 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1861 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1862 RegsToPass[i].second.getValueType()));
1864 // Add a register mask operand representing the call-preserved registers.
1866 const uint32_t *Mask;
1867 const ARMBaseRegisterInfo *ARI = Subtarget->getRegisterInfo();
1869 // For 'this' returns, use the R0-preserving mask if applicable
1870 Mask = ARI->getThisReturnPreservedMask(MF, CallConv);
1872 // Set isThisReturn to false if the calling convention is not one that
1873 // allows 'returned' to be modeled in this way, so LowerCallResult does
1874 // not try to pass 'this' straight through
1875 isThisReturn = false;
1876 Mask = ARI->getCallPreservedMask(MF, CallConv);
1879 Mask = ARI->getCallPreservedMask(MF, CallConv);
1881 assert(Mask && "Missing call preserved mask for calling convention");
1882 Ops.push_back(DAG.getRegisterMask(Mask));
1885 if (InFlag.getNode())
1886 Ops.push_back(InFlag);
1888 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
1890 MF.getFrameInfo()->setHasTailCall();
1891 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, Ops);
1894 // Returns a chain and a flag for retval copy to use.
1895 Chain = DAG.getNode(CallOpc, dl, NodeTys, Ops);
1896 InFlag = Chain.getValue(1);
1898 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, dl, true),
1899 DAG.getIntPtrConstant(0, dl, true), InFlag, dl);
1901 InFlag = Chain.getValue(1);
1903 // Handle result values, copying them out of physregs into vregs that we
1905 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins, dl, DAG,
1906 InVals, isThisReturn,
1907 isThisReturn ? OutVals[0] : SDValue());
1910 /// HandleByVal - Every parameter *after* a byval parameter is passed
1911 /// on the stack. Remember the next parameter register to allocate,
1912 /// and then confiscate the rest of the parameter registers to insure
1914 void ARMTargetLowering::HandleByVal(CCState *State, unsigned &Size,
1915 unsigned Align) const {
1916 assert((State->getCallOrPrologue() == Prologue ||
1917 State->getCallOrPrologue() == Call) &&
1918 "unhandled ParmContext");
1920 // Byval (as with any stack) slots are always at least 4 byte aligned.
1921 Align = std::max(Align, 4U);
1923 unsigned Reg = State->AllocateReg(GPRArgRegs);
1927 unsigned AlignInRegs = Align / 4;
1928 unsigned Waste = (ARM::R4 - Reg) % AlignInRegs;
1929 for (unsigned i = 0; i < Waste; ++i)
1930 Reg = State->AllocateReg(GPRArgRegs);
1935 unsigned Excess = 4 * (ARM::R4 - Reg);
1937 // Special case when NSAA != SP and parameter size greater than size of
1938 // all remained GPR regs. In that case we can't split parameter, we must
1939 // send it to stack. We also must set NCRN to R4, so waste all
1940 // remained registers.
1941 const unsigned NSAAOffset = State->getNextStackOffset();
1942 if (NSAAOffset != 0 && Size > Excess) {
1943 while (State->AllocateReg(GPRArgRegs))
1948 // First register for byval parameter is the first register that wasn't
1949 // allocated before this method call, so it would be "reg".
1950 // If parameter is small enough to be saved in range [reg, r4), then
1951 // the end (first after last) register would be reg + param-size-in-regs,
1952 // else parameter would be splitted between registers and stack,
1953 // end register would be r4 in this case.
1954 unsigned ByValRegBegin = Reg;
1955 unsigned ByValRegEnd = std::min<unsigned>(Reg + Size / 4, ARM::R4);
1956 State->addInRegsParamInfo(ByValRegBegin, ByValRegEnd);
1957 // Note, first register is allocated in the beginning of function already,
1958 // allocate remained amount of registers we need.
1959 for (unsigned i = Reg + 1; i != ByValRegEnd; ++i)
1960 State->AllocateReg(GPRArgRegs);
1961 // A byval parameter that is split between registers and memory needs its
1962 // size truncated here.
1963 // In the case where the entire structure fits in registers, we set the
1964 // size in memory to zero.
1965 Size = std::max<int>(Size - Excess, 0);
1968 /// MatchingStackOffset - Return true if the given stack call argument is
1969 /// already available in the same position (relatively) of the caller's
1970 /// incoming argument stack.
1972 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1973 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1974 const TargetInstrInfo *TII) {
1975 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1977 if (Arg.getOpcode() == ISD::CopyFromReg) {
1978 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1979 if (!TargetRegisterInfo::isVirtualRegister(VR))
1981 MachineInstr *Def = MRI->getVRegDef(VR);
1984 if (!Flags.isByVal()) {
1985 if (!TII->isLoadFromStackSlot(Def, FI))
1990 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
1991 if (Flags.isByVal())
1992 // ByVal argument is passed in as a pointer but it's now being
1993 // dereferenced. e.g.
1994 // define @foo(%struct.X* %A) {
1995 // tail call @bar(%struct.X* byval %A)
1998 SDValue Ptr = Ld->getBasePtr();
1999 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
2002 FI = FINode->getIndex();
2006 assert(FI != INT_MAX);
2007 if (!MFI->isFixedObjectIndex(FI))
2009 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
2012 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2013 /// for tail call optimization. Targets which want to do tail call
2014 /// optimization should implement this function.
2016 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2017 CallingConv::ID CalleeCC,
2019 bool isCalleeStructRet,
2020 bool isCallerStructRet,
2021 const SmallVectorImpl<ISD::OutputArg> &Outs,
2022 const SmallVectorImpl<SDValue> &OutVals,
2023 const SmallVectorImpl<ISD::InputArg> &Ins,
2024 SelectionDAG& DAG) const {
2025 const Function *CallerF = DAG.getMachineFunction().getFunction();
2026 CallingConv::ID CallerCC = CallerF->getCallingConv();
2027 bool CCMatch = CallerCC == CalleeCC;
2029 // Look for obvious safe cases to perform tail call optimization that do not
2030 // require ABI changes. This is what gcc calls sibcall.
2032 // Do not sibcall optimize vararg calls unless the call site is not passing
2034 if (isVarArg && !Outs.empty())
2037 // Exception-handling functions need a special set of instructions to indicate
2038 // a return to the hardware. Tail-calling another function would probably
2040 if (CallerF->hasFnAttribute("interrupt"))
2043 // Also avoid sibcall optimization if either caller or callee uses struct
2044 // return semantics.
2045 if (isCalleeStructRet || isCallerStructRet)
2048 // FIXME: Completely disable sibcall for Thumb1 since ThumbRegisterInfo::
2049 // emitEpilogue is not ready for them. Thumb tail calls also use t2B, as
2050 // the Thumb1 16-bit unconditional branch doesn't have sufficient relocation
2051 // support in the assembler and linker to be used. This would need to be
2052 // fixed to fully support tail calls in Thumb1.
2054 // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take
2055 // LR. This means if we need to reload LR, it takes an extra instructions,
2056 // which outweighs the value of the tail call; but here we don't know yet
2057 // whether LR is going to be used. Probably the right approach is to
2058 // generate the tail call here and turn it back into CALL/RET in
2059 // emitEpilogue if LR is used.
2061 // Thumb1 PIC calls to external symbols use BX, so they can be tail calls,
2062 // but we need to make sure there are enough registers; the only valid
2063 // registers are the 4 used for parameters. We don't currently do this
2065 if (Subtarget->isThumb1Only())
2068 // Externally-defined functions with weak linkage should not be
2069 // tail-called on ARM when the OS does not support dynamic
2070 // pre-emption of symbols, as the AAELF spec requires normal calls
2071 // to undefined weak functions to be replaced with a NOP or jump to the
2072 // next instruction. The behaviour of branch instructions in this
2073 // situation (as used for tail calls) is implementation-defined, so we
2074 // cannot rely on the linker replacing the tail call with a return.
2075 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2076 const GlobalValue *GV = G->getGlobal();
2077 const Triple &TT = getTargetMachine().getTargetTriple();
2078 if (GV->hasExternalWeakLinkage() &&
2079 (!TT.isOSWindows() || TT.isOSBinFormatELF() || TT.isOSBinFormatMachO()))
2083 // If the calling conventions do not match, then we'd better make sure the
2084 // results are returned in the same way as what the caller expects.
2086 SmallVector<CCValAssign, 16> RVLocs1;
2087 ARMCCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(), RVLocs1,
2088 *DAG.getContext(), Call);
2089 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
2091 SmallVector<CCValAssign, 16> RVLocs2;
2092 ARMCCState CCInfo2(CallerCC, false, DAG.getMachineFunction(), RVLocs2,
2093 *DAG.getContext(), Call);
2094 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
2096 if (RVLocs1.size() != RVLocs2.size())
2098 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
2099 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
2101 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
2103 if (RVLocs1[i].isRegLoc()) {
2104 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
2107 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
2113 // If Caller's vararg or byval argument has been split between registers and
2114 // stack, do not perform tail call, since part of the argument is in caller's
2116 const ARMFunctionInfo *AFI_Caller = DAG.getMachineFunction().
2117 getInfo<ARMFunctionInfo>();
2118 if (AFI_Caller->getArgRegsSaveSize())
2121 // If the callee takes no arguments then go on to check the results of the
2123 if (!Outs.empty()) {
2124 // Check if stack adjustment is needed. For now, do not do this if any
2125 // argument is passed on the stack.
2126 SmallVector<CCValAssign, 16> ArgLocs;
2127 ARMCCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
2128 *DAG.getContext(), Call);
2129 CCInfo.AnalyzeCallOperands(Outs,
2130 CCAssignFnForNode(CalleeCC, false, isVarArg));
2131 if (CCInfo.getNextStackOffset()) {
2132 MachineFunction &MF = DAG.getMachineFunction();
2134 // Check if the arguments are already laid out in the right way as
2135 // the caller's fixed stack objects.
2136 MachineFrameInfo *MFI = MF.getFrameInfo();
2137 const MachineRegisterInfo *MRI = &MF.getRegInfo();
2138 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
2139 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
2141 ++i, ++realArgIdx) {
2142 CCValAssign &VA = ArgLocs[i];
2143 EVT RegVT = VA.getLocVT();
2144 SDValue Arg = OutVals[realArgIdx];
2145 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
2146 if (VA.getLocInfo() == CCValAssign::Indirect)
2148 if (VA.needsCustom()) {
2149 // f64 and vector types are split into multiple registers or
2150 // register/stack-slot combinations. The types will not match
2151 // the registers; give up on memory f64 refs until we figure
2152 // out what to do about this.
2155 if (!ArgLocs[++i].isRegLoc())
2157 if (RegVT == MVT::v2f64) {
2158 if (!ArgLocs[++i].isRegLoc())
2160 if (!ArgLocs[++i].isRegLoc())
2163 } else if (!VA.isRegLoc()) {
2164 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
2176 ARMTargetLowering::CanLowerReturn(CallingConv::ID CallConv,
2177 MachineFunction &MF, bool isVarArg,
2178 const SmallVectorImpl<ISD::OutputArg> &Outs,
2179 LLVMContext &Context) const {
2180 SmallVector<CCValAssign, 16> RVLocs;
2181 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, Context);
2182 return CCInfo.CheckReturn(Outs, CCAssignFnForNode(CallConv, /*Return=*/true,
2186 static SDValue LowerInterruptReturn(SmallVectorImpl<SDValue> &RetOps,
2187 SDLoc DL, SelectionDAG &DAG) {
2188 const MachineFunction &MF = DAG.getMachineFunction();
2189 const Function *F = MF.getFunction();
2191 StringRef IntKind = F->getFnAttribute("interrupt").getValueAsString();
2193 // See ARM ARM v7 B1.8.3. On exception entry LR is set to a possibly offset
2194 // version of the "preferred return address". These offsets affect the return
2195 // instruction if this is a return from PL1 without hypervisor extensions.
2196 // IRQ/FIQ: +4 "subs pc, lr, #4"
2197 // SWI: 0 "subs pc, lr, #0"
2198 // ABORT: +4 "subs pc, lr, #4"
2199 // UNDEF: +4/+2 "subs pc, lr, #0"
2200 // UNDEF varies depending on where the exception came from ARM or Thumb
2201 // mode. Alongside GCC, we throw our hands up in disgust and pretend it's 0.
2204 if (IntKind == "" || IntKind == "IRQ" || IntKind == "FIQ" ||
2207 else if (IntKind == "SWI" || IntKind == "UNDEF")
2210 report_fatal_error("Unsupported interrupt attribute. If present, value "
2211 "must be one of: IRQ, FIQ, SWI, ABORT or UNDEF");
2213 RetOps.insert(RetOps.begin() + 1,
2214 DAG.getConstant(LROffset, DL, MVT::i32, false));
2216 return DAG.getNode(ARMISD::INTRET_FLAG, DL, MVT::Other, RetOps);
2220 ARMTargetLowering::LowerReturn(SDValue Chain,
2221 CallingConv::ID CallConv, bool isVarArg,
2222 const SmallVectorImpl<ISD::OutputArg> &Outs,
2223 const SmallVectorImpl<SDValue> &OutVals,
2224 SDLoc dl, SelectionDAG &DAG) const {
2226 // CCValAssign - represent the assignment of the return value to a location.
2227 SmallVector<CCValAssign, 16> RVLocs;
2229 // CCState - Info about the registers and stack slots.
2230 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
2231 *DAG.getContext(), Call);
2233 // Analyze outgoing return values.
2234 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
2238 SmallVector<SDValue, 4> RetOps;
2239 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
2240 bool isLittleEndian = Subtarget->isLittle();
2242 MachineFunction &MF = DAG.getMachineFunction();
2243 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2244 AFI->setReturnRegsCount(RVLocs.size());
2246 // Copy the result values into the output registers.
2247 for (unsigned i = 0, realRVLocIdx = 0;
2249 ++i, ++realRVLocIdx) {
2250 CCValAssign &VA = RVLocs[i];
2251 assert(VA.isRegLoc() && "Can only return in registers!");
2253 SDValue Arg = OutVals[realRVLocIdx];
2255 switch (VA.getLocInfo()) {
2256 default: llvm_unreachable("Unknown loc info!");
2257 case CCValAssign::Full: break;
2258 case CCValAssign::BCvt:
2259 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
2263 if (VA.needsCustom()) {
2264 if (VA.getLocVT() == MVT::v2f64) {
2265 // Extract the first half and return it in two registers.
2266 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
2267 DAG.getConstant(0, dl, MVT::i32));
2268 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
2269 DAG.getVTList(MVT::i32, MVT::i32), Half);
2271 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2272 HalfGPRs.getValue(isLittleEndian ? 0 : 1),
2274 Flag = Chain.getValue(1);
2275 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2276 VA = RVLocs[++i]; // skip ahead to next loc
2277 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2278 HalfGPRs.getValue(isLittleEndian ? 1 : 0),
2280 Flag = Chain.getValue(1);
2281 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2282 VA = RVLocs[++i]; // skip ahead to next loc
2284 // Extract the 2nd half and fall through to handle it as an f64 value.
2285 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
2286 DAG.getConstant(1, dl, MVT::i32));
2288 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
2290 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
2291 DAG.getVTList(MVT::i32, MVT::i32), Arg);
2292 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2293 fmrrd.getValue(isLittleEndian ? 0 : 1),
2295 Flag = Chain.getValue(1);
2296 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2297 VA = RVLocs[++i]; // skip ahead to next loc
2298 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2299 fmrrd.getValue(isLittleEndian ? 1 : 0),
2302 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
2304 // Guarantee that all emitted copies are
2305 // stuck together, avoiding something bad.
2306 Flag = Chain.getValue(1);
2307 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2310 // Update chain and glue.
2313 RetOps.push_back(Flag);
2315 // CPUs which aren't M-class use a special sequence to return from
2316 // exceptions (roughly, any instruction setting pc and cpsr simultaneously,
2317 // though we use "subs pc, lr, #N").
2319 // M-class CPUs actually use a normal return sequence with a special
2320 // (hardware-provided) value in LR, so the normal code path works.
2321 if (DAG.getMachineFunction().getFunction()->hasFnAttribute("interrupt") &&
2322 !Subtarget->isMClass()) {
2323 if (Subtarget->isThumb1Only())
2324 report_fatal_error("interrupt attribute is not supported in Thumb1");
2325 return LowerInterruptReturn(RetOps, dl, DAG);
2328 return DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, RetOps);
2331 bool ARMTargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
2332 if (N->getNumValues() != 1)
2334 if (!N->hasNUsesOfValue(1, 0))
2337 SDValue TCChain = Chain;
2338 SDNode *Copy = *N->use_begin();
2339 if (Copy->getOpcode() == ISD::CopyToReg) {
2340 // If the copy has a glue operand, we conservatively assume it isn't safe to
2341 // perform a tail call.
2342 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
2344 TCChain = Copy->getOperand(0);
2345 } else if (Copy->getOpcode() == ARMISD::VMOVRRD) {
2346 SDNode *VMov = Copy;
2347 // f64 returned in a pair of GPRs.
2348 SmallPtrSet<SDNode*, 2> Copies;
2349 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
2351 if (UI->getOpcode() != ISD::CopyToReg)
2355 if (Copies.size() > 2)
2358 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
2360 SDValue UseChain = UI->getOperand(0);
2361 if (Copies.count(UseChain.getNode()))
2365 // We are at the top of this chain.
2366 // If the copy has a glue operand, we conservatively assume it
2367 // isn't safe to perform a tail call.
2368 if (UI->getOperand(UI->getNumOperands()-1).getValueType() == MVT::Glue)
2374 } else if (Copy->getOpcode() == ISD::BITCAST) {
2375 // f32 returned in a single GPR.
2376 if (!Copy->hasOneUse())
2378 Copy = *Copy->use_begin();
2379 if (Copy->getOpcode() != ISD::CopyToReg || !Copy->hasNUsesOfValue(1, 0))
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);
2390 bool HasRet = false;
2391 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
2393 if (UI->getOpcode() != ARMISD::RET_FLAG &&
2394 UI->getOpcode() != ARMISD::INTRET_FLAG)
2406 bool ARMTargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
2407 if (!Subtarget->supportsTailCall())
2411 CI->getParent()->getParent()->getFnAttribute("disable-tail-calls");
2412 if (!CI->isTailCall() || Attr.getValueAsString() == "true")
2415 return !Subtarget->isThumb1Only();
2418 // Trying to write a 64 bit value so need to split into two 32 bit values first,
2419 // and pass the lower and high parts through.
2420 static SDValue LowerWRITE_REGISTER(SDValue Op, SelectionDAG &DAG) {
2422 SDValue WriteValue = Op->getOperand(2);
2424 // This function is only supposed to be called for i64 type argument.
2425 assert(WriteValue.getValueType() == MVT::i64
2426 && "LowerWRITE_REGISTER called for non-i64 type argument.");
2428 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, WriteValue,
2429 DAG.getConstant(0, DL, MVT::i32));
2430 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, WriteValue,
2431 DAG.getConstant(1, DL, MVT::i32));
2432 SDValue Ops[] = { Op->getOperand(0), Op->getOperand(1), Lo, Hi };
2433 return DAG.getNode(ISD::WRITE_REGISTER, DL, MVT::Other, Ops);
2436 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
2437 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
2438 // one of the above mentioned nodes. It has to be wrapped because otherwise
2439 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
2440 // be used to form addressing mode. These wrapped nodes will be selected
2442 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
2443 EVT PtrVT = Op.getValueType();
2444 // FIXME there is no actual debug info here
2446 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
2448 if (CP->isMachineConstantPoolEntry())
2449 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
2450 CP->getAlignment());
2452 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
2453 CP->getAlignment());
2454 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
2457 unsigned ARMTargetLowering::getJumpTableEncoding() const {
2458 return MachineJumpTableInfo::EK_Inline;
2461 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
2462 SelectionDAG &DAG) const {
2463 MachineFunction &MF = DAG.getMachineFunction();
2464 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2465 unsigned ARMPCLabelIndex = 0;
2467 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2468 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
2469 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2471 if (RelocM == Reloc::Static) {
2472 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
2474 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2475 ARMPCLabelIndex = AFI->createPICLabelUId();
2476 ARMConstantPoolValue *CPV =
2477 ARMConstantPoolConstant::Create(BA, ARMPCLabelIndex,
2478 ARMCP::CPBlockAddress, PCAdj);
2479 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2481 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
2482 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
2483 MachinePointerInfo::getConstantPool(),
2484 false, false, false, 0);
2485 if (RelocM == Reloc::Static)
2487 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, DL, MVT::i32);
2488 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
2491 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
2493 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
2494 SelectionDAG &DAG) const {
2496 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2497 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2498 MachineFunction &MF = DAG.getMachineFunction();
2499 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2500 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2501 ARMConstantPoolValue *CPV =
2502 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2503 ARMCP::CPValue, PCAdj, ARMCP::TLSGD, true);
2504 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2505 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
2506 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
2507 MachinePointerInfo::getConstantPool(),
2508 false, false, false, 0);
2509 SDValue Chain = Argument.getValue(1);
2511 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, dl, MVT::i32);
2512 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
2514 // call __tls_get_addr.
2517 Entry.Node = Argument;
2518 Entry.Ty = (Type *) Type::getInt32Ty(*DAG.getContext());
2519 Args.push_back(Entry);
2521 // FIXME: is there useful debug info available here?
2522 TargetLowering::CallLoweringInfo CLI(DAG);
2523 CLI.setDebugLoc(dl).setChain(Chain)
2524 .setCallee(CallingConv::C, Type::getInt32Ty(*DAG.getContext()),
2525 DAG.getExternalSymbol("__tls_get_addr", PtrVT), std::move(Args),
2528 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
2529 return CallResult.first;
2532 // Lower ISD::GlobalTLSAddress using the "initial exec" or
2533 // "local exec" model.
2535 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
2537 TLSModel::Model model) const {
2538 const GlobalValue *GV = GA->getGlobal();
2541 SDValue Chain = DAG.getEntryNode();
2542 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2543 // Get the Thread Pointer
2544 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2546 if (model == TLSModel::InitialExec) {
2547 MachineFunction &MF = DAG.getMachineFunction();
2548 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2549 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2550 // Initial exec model.
2551 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2552 ARMConstantPoolValue *CPV =
2553 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2554 ARMCP::CPValue, PCAdj, ARMCP::GOTTPOFF,
2556 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2557 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2558 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2559 MachinePointerInfo::getConstantPool(),
2560 false, false, false, 0);
2561 Chain = Offset.getValue(1);
2563 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, dl, MVT::i32);
2564 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
2566 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2567 MachinePointerInfo::getConstantPool(),
2568 false, false, false, 0);
2571 assert(model == TLSModel::LocalExec);
2572 ARMConstantPoolValue *CPV =
2573 ARMConstantPoolConstant::Create(GV, ARMCP::TPOFF);
2574 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2575 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2576 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2577 MachinePointerInfo::getConstantPool(),
2578 false, false, false, 0);
2581 // The address of the thread local variable is the add of the thread
2582 // pointer with the offset of the variable.
2583 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
2587 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
2588 // TODO: implement the "local dynamic" model
2589 assert(Subtarget->isTargetELF() &&
2590 "TLS not implemented for non-ELF targets");
2591 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
2592 if (DAG.getTarget().Options.EmulatedTLS)
2593 return LowerToTLSEmulatedModel(GA, DAG);
2595 TLSModel::Model model = getTargetMachine().getTLSModel(GA->getGlobal());
2598 case TLSModel::GeneralDynamic:
2599 case TLSModel::LocalDynamic:
2600 return LowerToTLSGeneralDynamicModel(GA, DAG);
2601 case TLSModel::InitialExec:
2602 case TLSModel::LocalExec:
2603 return LowerToTLSExecModels(GA, DAG, model);
2605 llvm_unreachable("bogus TLS model");
2608 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
2609 SelectionDAG &DAG) const {
2610 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2612 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2613 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2614 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
2615 ARMConstantPoolValue *CPV =
2616 ARMConstantPoolConstant::Create(GV,
2617 UseGOTOFF ? ARMCP::GOTOFF : ARMCP::GOT);
2618 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2619 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2620 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
2622 MachinePointerInfo::getConstantPool(),
2623 false, false, false, 0);
2624 SDValue Chain = Result.getValue(1);
2625 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
2626 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
2628 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
2629 MachinePointerInfo::getGOT(),
2630 false, false, false, 0);
2634 // If we have T2 ops, we can materialize the address directly via movt/movw
2635 // pair. This is always cheaper.
2636 if (Subtarget->useMovt(DAG.getMachineFunction())) {
2638 // FIXME: Once remat is capable of dealing with instructions with register
2639 // operands, expand this into two nodes.
2640 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2641 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2643 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2644 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2645 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2646 MachinePointerInfo::getConstantPool(),
2647 false, false, false, 0);
2651 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
2652 SelectionDAG &DAG) const {
2653 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2655 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2656 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2658 if (Subtarget->useMovt(DAG.getMachineFunction()))
2661 // FIXME: Once remat is capable of dealing with instructions with register
2662 // operands, expand this into multiple nodes
2664 RelocM == Reloc::PIC_ ? ARMISD::WrapperPIC : ARMISD::Wrapper;
2666 SDValue G = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, ARMII::MO_NONLAZY);
2667 SDValue Result = DAG.getNode(Wrapper, dl, PtrVT, G);
2669 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2670 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
2671 MachinePointerInfo::getGOT(), false, false, false, 0);
2675 SDValue ARMTargetLowering::LowerGlobalAddressWindows(SDValue Op,
2676 SelectionDAG &DAG) const {
2677 assert(Subtarget->isTargetWindows() && "non-Windows COFF is not supported");
2678 assert(Subtarget->useMovt(DAG.getMachineFunction()) &&
2679 "Windows on ARM expects to use movw/movt");
2681 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2682 const ARMII::TOF TargetFlags =
2683 (GV->hasDLLImportStorageClass() ? ARMII::MO_DLLIMPORT : ARMII::MO_NO_FLAG);
2684 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2690 // FIXME: Once remat is capable of dealing with instructions with register
2691 // operands, expand this into two nodes.
2692 Result = DAG.getNode(ARMISD::Wrapper, DL, PtrVT,
2693 DAG.getTargetGlobalAddress(GV, DL, PtrVT, /*Offset=*/0,
2695 if (GV->hasDLLImportStorageClass())
2696 Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), Result,
2697 MachinePointerInfo::getGOT(), false, false, false, 0);
2701 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
2702 SelectionDAG &DAG) const {
2703 assert(Subtarget->isTargetELF() &&
2704 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
2705 MachineFunction &MF = DAG.getMachineFunction();
2706 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2707 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2708 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2710 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2711 ARMConstantPoolValue *CPV =
2712 ARMConstantPoolSymbol::Create(*DAG.getContext(), "_GLOBAL_OFFSET_TABLE_",
2713 ARMPCLabelIndex, PCAdj);
2714 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2715 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2716 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2717 MachinePointerInfo::getConstantPool(),
2718 false, false, false, 0);
2719 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, dl, MVT::i32);
2720 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2724 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
2726 SDValue Val = DAG.getConstant(0, dl, MVT::i32);
2727 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl,
2728 DAG.getVTList(MVT::i32, MVT::Other), Op.getOperand(0),
2729 Op.getOperand(1), Val);
2733 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
2735 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
2736 Op.getOperand(1), DAG.getConstant(0, dl, MVT::i32));
2739 SDValue ARMTargetLowering::LowerEH_SJLJ_SETUP_DISPATCH(SDValue Op,
2740 SelectionDAG &DAG) const {
2742 return DAG.getNode(ARMISD::EH_SJLJ_SETUP_DISPATCH, dl, MVT::Other,
2747 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
2748 const ARMSubtarget *Subtarget) const {
2749 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2752 default: return SDValue(); // Don't custom lower most intrinsics.
2753 case Intrinsic::arm_rbit: {
2754 assert(Op.getOperand(1).getValueType() == MVT::i32 &&
2755 "RBIT intrinsic must have i32 type!");
2756 return DAG.getNode(ARMISD::RBIT, dl, MVT::i32, Op.getOperand(1));
2758 case Intrinsic::arm_thread_pointer: {
2759 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2760 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2762 case Intrinsic::eh_sjlj_lsda: {
2763 MachineFunction &MF = DAG.getMachineFunction();
2764 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2765 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2766 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2767 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2769 unsigned PCAdj = (RelocM != Reloc::PIC_)
2770 ? 0 : (Subtarget->isThumb() ? 4 : 8);
2771 ARMConstantPoolValue *CPV =
2772 ARMConstantPoolConstant::Create(MF.getFunction(), ARMPCLabelIndex,
2773 ARMCP::CPLSDA, PCAdj);
2774 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2775 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2777 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2778 MachinePointerInfo::getConstantPool(),
2779 false, false, false, 0);
2781 if (RelocM == Reloc::PIC_) {
2782 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, dl, MVT::i32);
2783 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2787 case Intrinsic::arm_neon_vmulls:
2788 case Intrinsic::arm_neon_vmullu: {
2789 unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vmulls)
2790 ? ARMISD::VMULLs : ARMISD::VMULLu;
2791 return DAG.getNode(NewOpc, SDLoc(Op), Op.getValueType(),
2792 Op.getOperand(1), Op.getOperand(2));
2797 static SDValue LowerATOMIC_FENCE(SDValue Op, SelectionDAG &DAG,
2798 const ARMSubtarget *Subtarget) {
2799 // FIXME: handle "fence singlethread" more efficiently.
2801 if (!Subtarget->hasDataBarrier()) {
2802 // Some ARMv6 cpus can support data barriers with an mcr instruction.
2803 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2805 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2806 "Unexpected ISD::ATOMIC_FENCE encountered. Should be libcall!");
2807 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2808 DAG.getConstant(0, dl, MVT::i32));
2811 ConstantSDNode *OrdN = cast<ConstantSDNode>(Op.getOperand(1));
2812 AtomicOrdering Ord = static_cast<AtomicOrdering>(OrdN->getZExtValue());
2813 ARM_MB::MemBOpt Domain = ARM_MB::ISH;
2814 if (Subtarget->isMClass()) {
2815 // Only a full system barrier exists in the M-class architectures.
2816 Domain = ARM_MB::SY;
2817 } else if (Subtarget->isSwift() && Ord == Release) {
2818 // Swift happens to implement ISHST barriers in a way that's compatible with
2819 // Release semantics but weaker than ISH so we'd be fools not to use
2820 // it. Beware: other processors probably don't!
2821 Domain = ARM_MB::ISHST;
2824 return DAG.getNode(ISD::INTRINSIC_VOID, dl, MVT::Other, Op.getOperand(0),
2825 DAG.getConstant(Intrinsic::arm_dmb, dl, MVT::i32),
2826 DAG.getConstant(Domain, dl, MVT::i32));
2829 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG,
2830 const ARMSubtarget *Subtarget) {
2831 // ARM pre v5TE and Thumb1 does not have preload instructions.
2832 if (!(Subtarget->isThumb2() ||
2833 (!Subtarget->isThumb1Only() && Subtarget->hasV5TEOps())))
2834 // Just preserve the chain.
2835 return Op.getOperand(0);
2838 unsigned isRead = ~cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() & 1;
2840 (!Subtarget->hasV7Ops() || !Subtarget->hasMPExtension()))
2841 // ARMv7 with MP extension has PLDW.
2842 return Op.getOperand(0);
2844 unsigned isData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
2845 if (Subtarget->isThumb()) {
2847 isRead = ~isRead & 1;
2848 isData = ~isData & 1;
2851 return DAG.getNode(ARMISD::PRELOAD, dl, MVT::Other, Op.getOperand(0),
2852 Op.getOperand(1), DAG.getConstant(isRead, dl, MVT::i32),
2853 DAG.getConstant(isData, dl, MVT::i32));
2856 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
2857 MachineFunction &MF = DAG.getMachineFunction();
2858 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
2860 // vastart just stores the address of the VarArgsFrameIndex slot into the
2861 // memory location argument.
2863 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy(DAG.getDataLayout());
2864 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2865 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
2866 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
2867 MachinePointerInfo(SV), false, false, 0);
2871 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
2872 SDValue &Root, SelectionDAG &DAG,
2874 MachineFunction &MF = DAG.getMachineFunction();
2875 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2877 const TargetRegisterClass *RC;
2878 if (AFI->isThumb1OnlyFunction())
2879 RC = &ARM::tGPRRegClass;
2881 RC = &ARM::GPRRegClass;
2883 // Transform the arguments stored in physical registers into virtual ones.
2884 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2885 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2888 if (NextVA.isMemLoc()) {
2889 MachineFrameInfo *MFI = MF.getFrameInfo();
2890 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true);
2892 // Create load node to retrieve arguments from the stack.
2893 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
2894 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
2895 MachinePointerInfo::getFixedStack(FI),
2896 false, false, false, 0);
2898 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
2899 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2901 if (!Subtarget->isLittle())
2902 std::swap (ArgValue, ArgValue2);
2903 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
2906 // The remaining GPRs hold either the beginning of variable-argument
2907 // data, or the beginning of an aggregate passed by value (usually
2908 // byval). Either way, we allocate stack slots adjacent to the data
2909 // provided by our caller, and store the unallocated registers there.
2910 // If this is a variadic function, the va_list pointer will begin with
2911 // these values; otherwise, this reassembles a (byval) structure that
2912 // was split between registers and memory.
2913 // Return: The frame index registers were stored into.
2915 ARMTargetLowering::StoreByValRegs(CCState &CCInfo, SelectionDAG &DAG,
2916 SDLoc dl, SDValue &Chain,
2917 const Value *OrigArg,
2918 unsigned InRegsParamRecordIdx,
2920 unsigned ArgSize) const {
2921 // Currently, two use-cases possible:
2922 // Case #1. Non-var-args function, and we meet first byval parameter.
2923 // Setup first unallocated register as first byval register;
2924 // eat all remained registers
2925 // (these two actions are performed by HandleByVal method).
2926 // Then, here, we initialize stack frame with
2927 // "store-reg" instructions.
2928 // Case #2. Var-args function, that doesn't contain byval parameters.
2929 // The same: eat all remained unallocated registers,
2930 // initialize stack frame.
2932 MachineFunction &MF = DAG.getMachineFunction();
2933 MachineFrameInfo *MFI = MF.getFrameInfo();
2934 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2935 unsigned RBegin, REnd;
2936 if (InRegsParamRecordIdx < CCInfo.getInRegsParamsCount()) {
2937 CCInfo.getInRegsParamInfo(InRegsParamRecordIdx, RBegin, REnd);
2939 unsigned RBeginIdx = CCInfo.getFirstUnallocated(GPRArgRegs);
2940 RBegin = RBeginIdx == 4 ? (unsigned)ARM::R4 : GPRArgRegs[RBeginIdx];
2945 ArgOffset = -4 * (ARM::R4 - RBegin);
2947 auto PtrVT = getPointerTy(DAG.getDataLayout());
2948 int FrameIndex = MFI->CreateFixedObject(ArgSize, ArgOffset, false);
2949 SDValue FIN = DAG.getFrameIndex(FrameIndex, PtrVT);
2951 SmallVector<SDValue, 4> MemOps;
2952 const TargetRegisterClass *RC =
2953 AFI->isThumb1OnlyFunction() ? &ARM::tGPRRegClass : &ARM::GPRRegClass;
2955 for (unsigned Reg = RBegin, i = 0; Reg < REnd; ++Reg, ++i) {
2956 unsigned VReg = MF.addLiveIn(Reg, RC);
2957 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2959 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2960 MachinePointerInfo(OrigArg, 4 * i), false, false, 0);
2961 MemOps.push_back(Store);
2962 FIN = DAG.getNode(ISD::ADD, dl, PtrVT, FIN, DAG.getConstant(4, dl, PtrVT));
2965 if (!MemOps.empty())
2966 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOps);
2970 // Setup stack frame, the va_list pointer will start from.
2972 ARMTargetLowering::VarArgStyleRegisters(CCState &CCInfo, SelectionDAG &DAG,
2973 SDLoc dl, SDValue &Chain,
2975 unsigned TotalArgRegsSaveSize,
2976 bool ForceMutable) const {
2977 MachineFunction &MF = DAG.getMachineFunction();
2978 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2980 // Try to store any remaining integer argument regs
2981 // to their spots on the stack so that they may be loaded by deferencing
2982 // the result of va_next.
2983 // If there is no regs to be stored, just point address after last
2984 // argument passed via stack.
2985 int FrameIndex = StoreByValRegs(CCInfo, DAG, dl, Chain, nullptr,
2986 CCInfo.getInRegsParamsCount(),
2987 CCInfo.getNextStackOffset(), 4);
2988 AFI->setVarArgsFrameIndex(FrameIndex);
2992 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
2993 CallingConv::ID CallConv, bool isVarArg,
2994 const SmallVectorImpl<ISD::InputArg>
2996 SDLoc dl, SelectionDAG &DAG,
2997 SmallVectorImpl<SDValue> &InVals)
2999 MachineFunction &MF = DAG.getMachineFunction();
3000 MachineFrameInfo *MFI = MF.getFrameInfo();
3002 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
3004 // Assign locations to all of the incoming arguments.
3005 SmallVector<CCValAssign, 16> ArgLocs;
3006 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs,
3007 *DAG.getContext(), Prologue);
3008 CCInfo.AnalyzeFormalArguments(Ins,
3009 CCAssignFnForNode(CallConv, /* Return*/ false,
3012 SmallVector<SDValue, 16> ArgValues;
3014 Function::const_arg_iterator CurOrigArg = MF.getFunction()->arg_begin();
3015 unsigned CurArgIdx = 0;
3017 // Initially ArgRegsSaveSize is zero.
3018 // Then we increase this value each time we meet byval parameter.
3019 // We also increase this value in case of varargs function.
3020 AFI->setArgRegsSaveSize(0);
3022 // Calculate the amount of stack space that we need to allocate to store
3023 // byval and variadic arguments that are passed in registers.
3024 // We need to know this before we allocate the first byval or variadic
3025 // argument, as they will be allocated a stack slot below the CFA (Canonical
3026 // Frame Address, the stack pointer at entry to the function).
3027 unsigned ArgRegBegin = ARM::R4;
3028 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3029 if (CCInfo.getInRegsParamsProcessed() >= CCInfo.getInRegsParamsCount())
3032 CCValAssign &VA = ArgLocs[i];
3033 unsigned Index = VA.getValNo();
3034 ISD::ArgFlagsTy Flags = Ins[Index].Flags;
3035 if (!Flags.isByVal())
3038 assert(VA.isMemLoc() && "unexpected byval pointer in reg");
3039 unsigned RBegin, REnd;
3040 CCInfo.getInRegsParamInfo(CCInfo.getInRegsParamsProcessed(), RBegin, REnd);
3041 ArgRegBegin = std::min(ArgRegBegin, RBegin);
3043 CCInfo.nextInRegsParam();
3045 CCInfo.rewindByValRegsInfo();
3047 int lastInsIndex = -1;
3048 if (isVarArg && MFI->hasVAStart()) {
3049 unsigned RegIdx = CCInfo.getFirstUnallocated(GPRArgRegs);
3050 if (RegIdx != array_lengthof(GPRArgRegs))
3051 ArgRegBegin = std::min(ArgRegBegin, (unsigned)GPRArgRegs[RegIdx]);
3054 unsigned TotalArgRegsSaveSize = 4 * (ARM::R4 - ArgRegBegin);
3055 AFI->setArgRegsSaveSize(TotalArgRegsSaveSize);
3056 auto PtrVT = getPointerTy(DAG.getDataLayout());
3058 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3059 CCValAssign &VA = ArgLocs[i];
3060 if (Ins[VA.getValNo()].isOrigArg()) {
3061 std::advance(CurOrigArg,
3062 Ins[VA.getValNo()].getOrigArgIndex() - CurArgIdx);
3063 CurArgIdx = Ins[VA.getValNo()].getOrigArgIndex();
3065 // Arguments stored in registers.
3066 if (VA.isRegLoc()) {
3067 EVT RegVT = VA.getLocVT();
3069 if (VA.needsCustom()) {
3070 // f64 and vector types are split up into multiple registers or
3071 // combinations of registers and stack slots.
3072 if (VA.getLocVT() == MVT::v2f64) {
3073 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
3075 VA = ArgLocs[++i]; // skip ahead to next loc
3077 if (VA.isMemLoc()) {
3078 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true);
3079 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
3080 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
3081 MachinePointerInfo::getFixedStack(FI),
3082 false, false, false, 0);
3084 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
3087 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
3088 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
3089 ArgValue, ArgValue1,
3090 DAG.getIntPtrConstant(0, dl));
3091 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
3092 ArgValue, ArgValue2,
3093 DAG.getIntPtrConstant(1, dl));
3095 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
3098 const TargetRegisterClass *RC;
3100 if (RegVT == MVT::f32)
3101 RC = &ARM::SPRRegClass;
3102 else if (RegVT == MVT::f64)
3103 RC = &ARM::DPRRegClass;
3104 else if (RegVT == MVT::v2f64)
3105 RC = &ARM::QPRRegClass;
3106 else if (RegVT == MVT::i32)
3107 RC = AFI->isThumb1OnlyFunction() ? &ARM::tGPRRegClass
3108 : &ARM::GPRRegClass;
3110 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
3112 // Transform the arguments in physical registers into virtual ones.
3113 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
3114 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
3117 // If this is an 8 or 16-bit value, it is really passed promoted
3118 // to 32 bits. Insert an assert[sz]ext to capture this, then
3119 // truncate to the right size.
3120 switch (VA.getLocInfo()) {
3121 default: llvm_unreachable("Unknown loc info!");
3122 case CCValAssign::Full: break;
3123 case CCValAssign::BCvt:
3124 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
3126 case CCValAssign::SExt:
3127 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
3128 DAG.getValueType(VA.getValVT()));
3129 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
3131 case CCValAssign::ZExt:
3132 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
3133 DAG.getValueType(VA.getValVT()));
3134 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
3138 InVals.push_back(ArgValue);
3140 } else { // VA.isRegLoc()
3143 assert(VA.isMemLoc());
3144 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
3146 int index = VA.getValNo();
3148 // Some Ins[] entries become multiple ArgLoc[] entries.
3149 // Process them only once.
3150 if (index != lastInsIndex)
3152 ISD::ArgFlagsTy Flags = Ins[index].Flags;
3153 // FIXME: For now, all byval parameter objects are marked mutable.
3154 // This can be changed with more analysis.
3155 // In case of tail call optimization mark all arguments mutable.
3156 // Since they could be overwritten by lowering of arguments in case of
3158 if (Flags.isByVal()) {
3159 assert(Ins[index].isOrigArg() &&
3160 "Byval arguments cannot be implicit");
3161 unsigned CurByValIndex = CCInfo.getInRegsParamsProcessed();
3163 int FrameIndex = StoreByValRegs(CCInfo, DAG, dl, Chain, CurOrigArg,
3164 CurByValIndex, VA.getLocMemOffset(),
3165 Flags.getByValSize());
3166 InVals.push_back(DAG.getFrameIndex(FrameIndex, PtrVT));
3167 CCInfo.nextInRegsParam();
3169 unsigned FIOffset = VA.getLocMemOffset();
3170 int FI = MFI->CreateFixedObject(VA.getLocVT().getSizeInBits()/8,
3173 // Create load nodes to retrieve arguments from the stack.
3174 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
3175 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
3176 MachinePointerInfo::getFixedStack(FI),
3177 false, false, false, 0));
3179 lastInsIndex = index;
3185 if (isVarArg && MFI->hasVAStart())
3186 VarArgStyleRegisters(CCInfo, DAG, dl, Chain,
3187 CCInfo.getNextStackOffset(),
3188 TotalArgRegsSaveSize);
3190 AFI->setArgumentStackSize(CCInfo.getNextStackOffset());
3195 /// isFloatingPointZero - Return true if this is +0.0.
3196 static bool isFloatingPointZero(SDValue Op) {
3197 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
3198 return CFP->getValueAPF().isPosZero();
3199 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
3200 // Maybe this has already been legalized into the constant pool?
3201 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
3202 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
3203 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
3204 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
3205 return CFP->getValueAPF().isPosZero();
3207 } else if (Op->getOpcode() == ISD::BITCAST &&
3208 Op->getValueType(0) == MVT::f64) {
3209 // Handle (ISD::BITCAST (ARMISD::VMOVIMM (ISD::TargetConstant 0)) MVT::f64)
3210 // created by LowerConstantFP().
3211 SDValue BitcastOp = Op->getOperand(0);
3212 if (BitcastOp->getOpcode() == ARMISD::VMOVIMM) {
3213 SDValue MoveOp = BitcastOp->getOperand(0);
3214 if (MoveOp->getOpcode() == ISD::TargetConstant &&
3215 cast<ConstantSDNode>(MoveOp)->getZExtValue() == 0) {
3223 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
3224 /// the given operands.
3226 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
3227 SDValue &ARMcc, SelectionDAG &DAG,
3229 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
3230 unsigned C = RHSC->getZExtValue();
3231 if (!isLegalICmpImmediate(C)) {
3232 // Constant does not fit, try adjusting it by one?
3237 if (C != 0x80000000 && isLegalICmpImmediate(C-1)) {
3238 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
3239 RHS = DAG.getConstant(C - 1, dl, MVT::i32);
3244 if (C != 0 && isLegalICmpImmediate(C-1)) {
3245 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
3246 RHS = DAG.getConstant(C - 1, dl, MVT::i32);
3251 if (C != 0x7fffffff && isLegalICmpImmediate(C+1)) {
3252 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
3253 RHS = DAG.getConstant(C + 1, dl, MVT::i32);
3258 if (C != 0xffffffff && isLegalICmpImmediate(C+1)) {
3259 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
3260 RHS = DAG.getConstant(C + 1, dl, MVT::i32);
3267 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3268 ARMISD::NodeType CompareType;
3271 CompareType = ARMISD::CMP;
3276 CompareType = ARMISD::CMPZ;
3279 ARMcc = DAG.getConstant(CondCode, dl, MVT::i32);
3280 return DAG.getNode(CompareType, dl, MVT::Glue, LHS, RHS);
3283 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
3285 ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
3287 assert(!Subtarget->isFPOnlySP() || RHS.getValueType() != MVT::f64);
3289 if (!isFloatingPointZero(RHS))
3290 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Glue, LHS, RHS);
3292 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Glue, LHS);
3293 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Glue, Cmp);
3296 /// duplicateCmp - Glue values can have only one use, so this function
3297 /// duplicates a comparison node.
3299 ARMTargetLowering::duplicateCmp(SDValue Cmp, SelectionDAG &DAG) const {
3300 unsigned Opc = Cmp.getOpcode();
3302 if (Opc == ARMISD::CMP || Opc == ARMISD::CMPZ)
3303 return DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
3305 assert(Opc == ARMISD::FMSTAT && "unexpected comparison operation");
3306 Cmp = Cmp.getOperand(0);
3307 Opc = Cmp.getOpcode();
3308 if (Opc == ARMISD::CMPFP)
3309 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
3311 assert(Opc == ARMISD::CMPFPw0 && "unexpected operand of FMSTAT");
3312 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0));
3314 return DAG.getNode(ARMISD::FMSTAT, DL, MVT::Glue, Cmp);
3317 std::pair<SDValue, SDValue>
3318 ARMTargetLowering::getARMXALUOOp(SDValue Op, SelectionDAG &DAG,
3319 SDValue &ARMcc) const {
3320 assert(Op.getValueType() == MVT::i32 && "Unsupported value type");
3322 SDValue Value, OverflowCmp;
3323 SDValue LHS = Op.getOperand(0);
3324 SDValue RHS = Op.getOperand(1);
3327 // FIXME: We are currently always generating CMPs because we don't support
3328 // generating CMN through the backend. This is not as good as the natural
3329 // CMP case because it causes a register dependency and cannot be folded
3332 switch (Op.getOpcode()) {
3334 llvm_unreachable("Unknown overflow instruction!");
3336 ARMcc = DAG.getConstant(ARMCC::VC, dl, MVT::i32);
3337 Value = DAG.getNode(ISD::ADD, dl, Op.getValueType(), LHS, RHS);
3338 OverflowCmp = DAG.getNode(ARMISD::CMP, dl, MVT::Glue, Value, LHS);
3341 ARMcc = DAG.getConstant(ARMCC::HS, dl, MVT::i32);
3342 Value = DAG.getNode(ISD::ADD, dl, Op.getValueType(), LHS, RHS);
3343 OverflowCmp = DAG.getNode(ARMISD::CMP, dl, MVT::Glue, Value, LHS);
3346 ARMcc = DAG.getConstant(ARMCC::VC, dl, MVT::i32);
3347 Value = DAG.getNode(ISD::SUB, dl, Op.getValueType(), LHS, RHS);
3348 OverflowCmp = DAG.getNode(ARMISD::CMP, dl, MVT::Glue, LHS, RHS);
3351 ARMcc = DAG.getConstant(ARMCC::HS, dl, MVT::i32);
3352 Value = DAG.getNode(ISD::SUB, dl, Op.getValueType(), LHS, RHS);
3353 OverflowCmp = DAG.getNode(ARMISD::CMP, dl, MVT::Glue, LHS, RHS);
3357 return std::make_pair(Value, OverflowCmp);
3362 ARMTargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) const {
3363 // Let legalize expand this if it isn't a legal type yet.
3364 if (!DAG.getTargetLoweringInfo().isTypeLegal(Op.getValueType()))
3367 SDValue Value, OverflowCmp;
3369 std::tie(Value, OverflowCmp) = getARMXALUOOp(Op, DAG, ARMcc);
3370 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3372 // We use 0 and 1 as false and true values.
3373 SDValue TVal = DAG.getConstant(1, dl, MVT::i32);
3374 SDValue FVal = DAG.getConstant(0, dl, MVT::i32);
3375 EVT VT = Op.getValueType();
3377 SDValue Overflow = DAG.getNode(ARMISD::CMOV, dl, VT, TVal, FVal,
3378 ARMcc, CCR, OverflowCmp);
3380 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
3381 return DAG.getNode(ISD::MERGE_VALUES, dl, VTs, Value, Overflow);
3385 SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
3386 SDValue Cond = Op.getOperand(0);
3387 SDValue SelectTrue = Op.getOperand(1);
3388 SDValue SelectFalse = Op.getOperand(2);
3390 unsigned Opc = Cond.getOpcode();
3392 if (Cond.getResNo() == 1 &&
3393 (Opc == ISD::SADDO || Opc == ISD::UADDO || Opc == ISD::SSUBO ||
3394 Opc == ISD::USUBO)) {
3395 if (!DAG.getTargetLoweringInfo().isTypeLegal(Cond->getValueType(0)))
3398 SDValue Value, OverflowCmp;
3400 std::tie(Value, OverflowCmp) = getARMXALUOOp(Cond, DAG, ARMcc);
3401 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3402 EVT VT = Op.getValueType();
3404 return getCMOV(dl, VT, SelectTrue, SelectFalse, ARMcc, CCR,
3410 // (select (cmov 1, 0, cond), t, f) -> (cmov t, f, cond)
3411 // (select (cmov 0, 1, cond), t, f) -> (cmov f, t, cond)
3413 if (Cond.getOpcode() == ARMISD::CMOV && Cond.hasOneUse()) {
3414 const ConstantSDNode *CMOVTrue =
3415 dyn_cast<ConstantSDNode>(Cond.getOperand(0));
3416 const ConstantSDNode *CMOVFalse =
3417 dyn_cast<ConstantSDNode>(Cond.getOperand(1));
3419 if (CMOVTrue && CMOVFalse) {
3420 unsigned CMOVTrueVal = CMOVTrue->getZExtValue();
3421 unsigned CMOVFalseVal = CMOVFalse->getZExtValue();
3425 if (CMOVTrueVal == 1 && CMOVFalseVal == 0) {
3427 False = SelectFalse;
3428 } else if (CMOVTrueVal == 0 && CMOVFalseVal == 1) {
3433 if (True.getNode() && False.getNode()) {
3434 EVT VT = Op.getValueType();
3435 SDValue ARMcc = Cond.getOperand(2);
3436 SDValue CCR = Cond.getOperand(3);
3437 SDValue Cmp = duplicateCmp(Cond.getOperand(4), DAG);
3438 assert(True.getValueType() == VT);
3439 return getCMOV(dl, VT, True, False, ARMcc, CCR, Cmp, DAG);
3444 // ARM's BooleanContents value is UndefinedBooleanContent. Mask out the
3445 // undefined bits before doing a full-word comparison with zero.
3446 Cond = DAG.getNode(ISD::AND, dl, Cond.getValueType(), Cond,
3447 DAG.getConstant(1, dl, Cond.getValueType()));
3449 return DAG.getSelectCC(dl, Cond,
3450 DAG.getConstant(0, dl, Cond.getValueType()),
3451 SelectTrue, SelectFalse, ISD::SETNE);
3454 static void checkVSELConstraints(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
3455 bool &swpCmpOps, bool &swpVselOps) {
3456 // Start by selecting the GE condition code for opcodes that return true for
3458 if (CC == ISD::SETUGE || CC == ISD::SETOGE || CC == ISD::SETOLE ||
3460 CondCode = ARMCC::GE;
3462 // and GT for opcodes that return false for 'equality'.
3463 else if (CC == ISD::SETUGT || CC == ISD::SETOGT || CC == ISD::SETOLT ||
3465 CondCode = ARMCC::GT;
3467 // Since we are constrained to GE/GT, if the opcode contains 'less', we need
3468 // to swap the compare operands.
3469 if (CC == ISD::SETOLE || CC == ISD::SETULE || CC == ISD::SETOLT ||
3473 // Both GT and GE are ordered comparisons, and return false for 'unordered'.
3474 // If we have an unordered opcode, we need to swap the operands to the VSEL
3475 // instruction (effectively negating the condition).
3477 // This also has the effect of swapping which one of 'less' or 'greater'
3478 // returns true, so we also swap the compare operands. It also switches
3479 // whether we return true for 'equality', so we compensate by picking the
3480 // opposite condition code to our original choice.
3481 if (CC == ISD::SETULE || CC == ISD::SETULT || CC == ISD::SETUGE ||
3482 CC == ISD::SETUGT) {
3483 swpCmpOps = !swpCmpOps;
3484 swpVselOps = !swpVselOps;
3485 CondCode = CondCode == ARMCC::GT ? ARMCC::GE : ARMCC::GT;
3488 // 'ordered' is 'anything but unordered', so use the VS condition code and
3489 // swap the VSEL operands.
3490 if (CC == ISD::SETO) {
3491 CondCode = ARMCC::VS;
3495 // 'unordered or not equal' is 'anything but equal', so use the EQ condition
3496 // code and swap the VSEL operands.
3497 if (CC == ISD::SETUNE) {
3498 CondCode = ARMCC::EQ;
3503 SDValue ARMTargetLowering::getCMOV(SDLoc dl, EVT VT, SDValue FalseVal,
3504 SDValue TrueVal, SDValue ARMcc, SDValue CCR,
3505 SDValue Cmp, SelectionDAG &DAG) const {
3506 if (Subtarget->isFPOnlySP() && VT == MVT::f64) {
3507 FalseVal = DAG.getNode(ARMISD::VMOVRRD, dl,
3508 DAG.getVTList(MVT::i32, MVT::i32), FalseVal);
3509 TrueVal = DAG.getNode(ARMISD::VMOVRRD, dl,
3510 DAG.getVTList(MVT::i32, MVT::i32), TrueVal);
3512 SDValue TrueLow = TrueVal.getValue(0);
3513 SDValue TrueHigh = TrueVal.getValue(1);
3514 SDValue FalseLow = FalseVal.getValue(0);
3515 SDValue FalseHigh = FalseVal.getValue(1);
3517 SDValue Low = DAG.getNode(ARMISD::CMOV, dl, MVT::i32, FalseLow, TrueLow,
3519 SDValue High = DAG.getNode(ARMISD::CMOV, dl, MVT::i32, FalseHigh, TrueHigh,
3520 ARMcc, CCR, duplicateCmp(Cmp, DAG));
3522 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Low, High);
3524 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,
3529 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
3530 EVT VT = Op.getValueType();
3531 SDValue LHS = Op.getOperand(0);
3532 SDValue RHS = Op.getOperand(1);
3533 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
3534 SDValue TrueVal = Op.getOperand(2);
3535 SDValue FalseVal = Op.getOperand(3);
3538 if (Subtarget->isFPOnlySP() && LHS.getValueType() == MVT::f64) {
3539 DAG.getTargetLoweringInfo().softenSetCCOperands(DAG, MVT::f64, LHS, RHS, CC,
3542 // If softenSetCCOperands only returned one value, we should compare it to
3544 if (!RHS.getNode()) {
3545 RHS = DAG.getConstant(0, dl, LHS.getValueType());
3550 if (LHS.getValueType() == MVT::i32) {
3551 // Try to generate VSEL on ARMv8.
3552 // The VSEL instruction can't use all the usual ARM condition
3553 // codes: it only has two bits to select the condition code, so it's
3554 // constrained to use only GE, GT, VS and EQ.
3556 // To implement all the various ISD::SETXXX opcodes, we sometimes need to
3557 // swap the operands of the previous compare instruction (effectively
3558 // inverting the compare condition, swapping 'less' and 'greater') and
3559 // sometimes need to swap the operands to the VSEL (which inverts the
3560 // condition in the sense of firing whenever the previous condition didn't)
3561 if (Subtarget->hasFPARMv8() && (TrueVal.getValueType() == MVT::f32 ||
3562 TrueVal.getValueType() == MVT::f64)) {
3563 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3564 if (CondCode == ARMCC::LT || CondCode == ARMCC::LE ||
3565 CondCode == ARMCC::VC || CondCode == ARMCC::NE) {
3566 CC = ISD::getSetCCInverse(CC, true);
3567 std::swap(TrueVal, FalseVal);
3572 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3573 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3574 return getCMOV(dl, VT, FalseVal, TrueVal, ARMcc, CCR, Cmp, DAG);
3577 ARMCC::CondCodes CondCode, CondCode2;
3578 FPCCToARMCC(CC, CondCode, CondCode2);
3580 // Try to generate VMAXNM/VMINNM on ARMv8.
3581 if (Subtarget->hasFPARMv8() && (TrueVal.getValueType() == MVT::f32 ||
3582 TrueVal.getValueType() == MVT::f64)) {
3583 // We can use VMAXNM/VMINNM for a compare followed by a select with the
3584 // same operands, as follows:
3585 // c = fcmp [?gt, ?ge, ?lt, ?le] a, b
3587 // In NoNaNsFPMath the CC will have been changed from, e.g., 'ogt' to 'gt'.
3588 bool swapSides = false;
3589 if (!getTargetMachine().Options.NoNaNsFPMath) {
3590 // transformability may depend on which way around we compare
3598 // the non-NaN should be RHS
3599 swapSides = DAG.isKnownNeverNaN(LHS) && !DAG.isKnownNeverNaN(RHS);
3605 // the non-NaN should be LHS
3606 swapSides = DAG.isKnownNeverNaN(RHS) && !DAG.isKnownNeverNaN(LHS);
3610 swapSides = swapSides || (LHS == FalseVal && RHS == TrueVal);
3612 CC = ISD::getSetCCSwappedOperands(CC);
3613 std::swap(LHS, RHS);
3615 if (LHS == TrueVal && RHS == FalseVal) {
3616 bool canTransform = true;
3617 // FIXME: FastMathFlags::noSignedZeros() doesn't appear reachable from here
3618 if (!getTargetMachine().Options.UnsafeFPMath &&
3619 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
3620 const ConstantFPSDNode *Zero;
3627 // RHS must not be -0
3628 canTransform = (Zero = dyn_cast<ConstantFPSDNode>(RHS)) &&
3629 !Zero->isNegative();
3634 // LHS must not be -0
3635 canTransform = (Zero = dyn_cast<ConstantFPSDNode>(LHS)) &&
3636 !Zero->isNegative();
3641 // RHS must not be +0
3642 canTransform = (Zero = dyn_cast<ConstantFPSDNode>(RHS)) &&
3648 // LHS must not be +0
3649 canTransform = (Zero = dyn_cast<ConstantFPSDNode>(LHS)) &&
3655 // Note: If one of the elements in a pair is a number and the other
3656 // element is NaN, the corresponding result element is the number.
3657 // This is consistent with the IEEE 754-2008 standard.
3658 // Therefore, a > b ? a : b <=> vmax(a,b), if b is constant and a is NaN
3664 if (!DAG.isKnownNeverNaN(RHS))
3666 return DAG.getNode(ISD::FMAXNUM, dl, VT, LHS, RHS);
3669 if (!DAG.isKnownNeverNaN(LHS))
3673 return DAG.getNode(ISD::FMAXNUM, dl, VT, LHS, RHS);
3676 if (!DAG.isKnownNeverNaN(RHS))
3678 return DAG.getNode(ISD::FMINNUM, dl, VT, LHS, RHS);
3681 if (!DAG.isKnownNeverNaN(LHS))
3685 return DAG.getNode(ISD::FMINNUM, dl, VT, LHS, RHS);
3690 bool swpCmpOps = false;
3691 bool swpVselOps = false;
3692 checkVSELConstraints(CC, CondCode, swpCmpOps, swpVselOps);
3694 if (CondCode == ARMCC::GT || CondCode == ARMCC::GE ||
3695 CondCode == ARMCC::VS || CondCode == ARMCC::EQ) {
3697 std::swap(LHS, RHS);
3699 std::swap(TrueVal, FalseVal);
3703 SDValue ARMcc = DAG.getConstant(CondCode, dl, MVT::i32);
3704 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
3705 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3706 SDValue Result = getCMOV(dl, VT, FalseVal, TrueVal, ARMcc, CCR, Cmp, DAG);
3707 if (CondCode2 != ARMCC::AL) {
3708 SDValue ARMcc2 = DAG.getConstant(CondCode2, dl, MVT::i32);
3709 // FIXME: Needs another CMP because flag can have but one use.
3710 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
3711 Result = getCMOV(dl, VT, Result, TrueVal, ARMcc2, CCR, Cmp2, DAG);
3716 /// canChangeToInt - Given the fp compare operand, return true if it is suitable
3717 /// to morph to an integer compare sequence.
3718 static bool canChangeToInt(SDValue Op, bool &SeenZero,
3719 const ARMSubtarget *Subtarget) {
3720 SDNode *N = Op.getNode();
3721 if (!N->hasOneUse())
3722 // Otherwise it requires moving the value from fp to integer registers.
3724 if (!N->getNumValues())
3726 EVT VT = Op.getValueType();
3727 if (VT != MVT::f32 && !Subtarget->isFPBrccSlow())
3728 // f32 case is generally profitable. f64 case only makes sense when vcmpe +
3729 // vmrs are very slow, e.g. cortex-a8.
3732 if (isFloatingPointZero(Op)) {
3736 return ISD::isNormalLoad(N);
3739 static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) {
3740 if (isFloatingPointZero(Op))
3741 return DAG.getConstant(0, SDLoc(Op), MVT::i32);
3743 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op))
3744 return DAG.getLoad(MVT::i32, SDLoc(Op),
3745 Ld->getChain(), Ld->getBasePtr(), Ld->getPointerInfo(),
3746 Ld->isVolatile(), Ld->isNonTemporal(),
3747 Ld->isInvariant(), Ld->getAlignment());
3749 llvm_unreachable("Unknown VFP cmp argument!");
3752 static void expandf64Toi32(SDValue Op, SelectionDAG &DAG,
3753 SDValue &RetVal1, SDValue &RetVal2) {
3756 if (isFloatingPointZero(Op)) {
3757 RetVal1 = DAG.getConstant(0, dl, MVT::i32);
3758 RetVal2 = DAG.getConstant(0, dl, MVT::i32);
3762 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) {
3763 SDValue Ptr = Ld->getBasePtr();
3764 RetVal1 = DAG.getLoad(MVT::i32, dl,
3765 Ld->getChain(), Ptr,
3766 Ld->getPointerInfo(),
3767 Ld->isVolatile(), Ld->isNonTemporal(),
3768 Ld->isInvariant(), Ld->getAlignment());
3770 EVT PtrType = Ptr.getValueType();
3771 unsigned NewAlign = MinAlign(Ld->getAlignment(), 4);
3772 SDValue NewPtr = DAG.getNode(ISD::ADD, dl,
3773 PtrType, Ptr, DAG.getConstant(4, dl, PtrType));
3774 RetVal2 = DAG.getLoad(MVT::i32, dl,
3775 Ld->getChain(), NewPtr,
3776 Ld->getPointerInfo().getWithOffset(4),
3777 Ld->isVolatile(), Ld->isNonTemporal(),
3778 Ld->isInvariant(), NewAlign);
3782 llvm_unreachable("Unknown VFP cmp argument!");
3785 /// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some
3786 /// f32 and even f64 comparisons to integer ones.
3788 ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const {
3789 SDValue Chain = Op.getOperand(0);
3790 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3791 SDValue LHS = Op.getOperand(2);
3792 SDValue RHS = Op.getOperand(3);
3793 SDValue Dest = Op.getOperand(4);
3796 bool LHSSeenZero = false;
3797 bool LHSOk = canChangeToInt(LHS, LHSSeenZero, Subtarget);
3798 bool RHSSeenZero = false;
3799 bool RHSOk = canChangeToInt(RHS, RHSSeenZero, Subtarget);
3800 if (LHSOk && RHSOk && (LHSSeenZero || RHSSeenZero)) {
3801 // If unsafe fp math optimization is enabled and there are no other uses of
3802 // the CMP operands, and the condition code is EQ or NE, we can optimize it
3803 // to an integer comparison.
3804 if (CC == ISD::SETOEQ)
3806 else if (CC == ISD::SETUNE)
3809 SDValue Mask = DAG.getConstant(0x7fffffff, dl, MVT::i32);
3811 if (LHS.getValueType() == MVT::f32) {
3812 LHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3813 bitcastf32Toi32(LHS, DAG), Mask);
3814 RHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3815 bitcastf32Toi32(RHS, DAG), Mask);
3816 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3817 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3818 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3819 Chain, Dest, ARMcc, CCR, Cmp);
3824 expandf64Toi32(LHS, DAG, LHS1, LHS2);
3825 expandf64Toi32(RHS, DAG, RHS1, RHS2);
3826 LHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, LHS2, Mask);
3827 RHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, RHS2, Mask);
3828 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3829 ARMcc = DAG.getConstant(CondCode, dl, MVT::i32);
3830 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3831 SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest };
3832 return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops);
3838 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
3839 SDValue Chain = Op.getOperand(0);
3840 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3841 SDValue LHS = Op.getOperand(2);
3842 SDValue RHS = Op.getOperand(3);
3843 SDValue Dest = Op.getOperand(4);
3846 if (Subtarget->isFPOnlySP() && LHS.getValueType() == MVT::f64) {
3847 DAG.getTargetLoweringInfo().softenSetCCOperands(DAG, MVT::f64, LHS, RHS, CC,
3850 // If softenSetCCOperands only returned one value, we should compare it to
3852 if (!RHS.getNode()) {
3853 RHS = DAG.getConstant(0, dl, LHS.getValueType());
3858 if (LHS.getValueType() == MVT::i32) {
3860 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3861 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3862 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3863 Chain, Dest, ARMcc, CCR, Cmp);
3866 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3868 if (getTargetMachine().Options.UnsafeFPMath &&
3869 (CC == ISD::SETEQ || CC == ISD::SETOEQ ||
3870 CC == ISD::SETNE || CC == ISD::SETUNE)) {
3871 SDValue Result = OptimizeVFPBrcond(Op, DAG);
3872 if (Result.getNode())
3876 ARMCC::CondCodes CondCode, CondCode2;
3877 FPCCToARMCC(CC, CondCode, CondCode2);
3879 SDValue ARMcc = DAG.getConstant(CondCode, dl, MVT::i32);
3880 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
3881 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3882 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3883 SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp };
3884 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops);
3885 if (CondCode2 != ARMCC::AL) {
3886 ARMcc = DAG.getConstant(CondCode2, dl, MVT::i32);
3887 SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) };
3888 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops);
3893 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
3894 SDValue Chain = Op.getOperand(0);
3895 SDValue Table = Op.getOperand(1);
3896 SDValue Index = Op.getOperand(2);
3899 EVT PTy = getPointerTy(DAG.getDataLayout());
3900 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
3901 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
3902 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI);
3903 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, dl, PTy));
3904 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
3905 if (Subtarget->isThumb2()) {
3906 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
3907 // which does another jump to the destination. This also makes it easier
3908 // to translate it to TBB / TBH later.
3909 // FIXME: This might not work if the function is extremely large.
3910 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
3911 Addr, Op.getOperand(2), JTI);
3913 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
3914 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
3915 MachinePointerInfo::getJumpTable(),
3916 false, false, false, 0);
3917 Chain = Addr.getValue(1);
3918 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
3919 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI);
3921 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
3922 MachinePointerInfo::getJumpTable(),
3923 false, false, false, 0);
3924 Chain = Addr.getValue(1);
3925 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI);
3929 static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3930 EVT VT = Op.getValueType();
3933 if (Op.getValueType().getVectorElementType() == MVT::i32) {
3934 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::f32)
3936 return DAG.UnrollVectorOp(Op.getNode());
3939 assert(Op.getOperand(0).getValueType() == MVT::v4f32 &&
3940 "Invalid type for custom lowering!");
3941 if (VT != MVT::v4i16)
3942 return DAG.UnrollVectorOp(Op.getNode());
3944 Op = DAG.getNode(Op.getOpcode(), dl, MVT::v4i32, Op.getOperand(0));
3945 return DAG.getNode(ISD::TRUNCATE, dl, VT, Op);
3948 SDValue ARMTargetLowering::LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) const {
3949 EVT VT = Op.getValueType();
3951 return LowerVectorFP_TO_INT(Op, DAG);
3952 if (Subtarget->isFPOnlySP() && Op.getOperand(0).getValueType() == MVT::f64) {
3954 if (Op.getOpcode() == ISD::FP_TO_SINT)
3955 LC = RTLIB::getFPTOSINT(Op.getOperand(0).getValueType(),
3958 LC = RTLIB::getFPTOUINT(Op.getOperand(0).getValueType(),
3960 return makeLibCall(DAG, LC, Op.getValueType(), &Op.getOperand(0), 1,
3961 /*isSigned*/ false, SDLoc(Op)).first;
3967 static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3968 EVT VT = Op.getValueType();
3971 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::i32) {
3972 if (VT.getVectorElementType() == MVT::f32)
3974 return DAG.UnrollVectorOp(Op.getNode());
3977 assert(Op.getOperand(0).getValueType() == MVT::v4i16 &&
3978 "Invalid type for custom lowering!");
3979 if (VT != MVT::v4f32)
3980 return DAG.UnrollVectorOp(Op.getNode());
3984 switch (Op.getOpcode()) {
3985 default: llvm_unreachable("Invalid opcode!");
3986 case ISD::SINT_TO_FP:
3987 CastOpc = ISD::SIGN_EXTEND;
3988 Opc = ISD::SINT_TO_FP;
3990 case ISD::UINT_TO_FP:
3991 CastOpc = ISD::ZERO_EXTEND;
3992 Opc = ISD::UINT_TO_FP;
3996 Op = DAG.getNode(CastOpc, dl, MVT::v4i32, Op.getOperand(0));
3997 return DAG.getNode(Opc, dl, VT, Op);
4000 SDValue ARMTargetLowering::LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) const {
4001 EVT VT = Op.getValueType();
4003 return LowerVectorINT_TO_FP(Op, DAG);
4004 if (Subtarget->isFPOnlySP() && Op.getValueType() == MVT::f64) {
4006 if (Op.getOpcode() == ISD::SINT_TO_FP)
4007 LC = RTLIB::getSINTTOFP(Op.getOperand(0).getValueType(),
4010 LC = RTLIB::getUINTTOFP(Op.getOperand(0).getValueType(),
4012 return makeLibCall(DAG, LC, Op.getValueType(), &Op.getOperand(0), 1,
4013 /*isSigned*/ false, SDLoc(Op)).first;
4019 SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
4020 // Implement fcopysign with a fabs and a conditional fneg.
4021 SDValue Tmp0 = Op.getOperand(0);
4022 SDValue Tmp1 = Op.getOperand(1);
4024 EVT VT = Op.getValueType();
4025 EVT SrcVT = Tmp1.getValueType();
4026 bool InGPR = Tmp0.getOpcode() == ISD::BITCAST ||
4027 Tmp0.getOpcode() == ARMISD::VMOVDRR;
4028 bool UseNEON = !InGPR && Subtarget->hasNEON();
4031 // Use VBSL to copy the sign bit.
4032 unsigned EncodedVal = ARM_AM::createNEONModImm(0x6, 0x80);
4033 SDValue Mask = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v2i32,
4034 DAG.getTargetConstant(EncodedVal, dl, MVT::i32));
4035 EVT OpVT = (VT == MVT::f32) ? MVT::v2i32 : MVT::v1i64;
4037 Mask = DAG.getNode(ARMISD::VSHL, dl, OpVT,
4038 DAG.getNode(ISD::BITCAST, dl, OpVT, Mask),
4039 DAG.getConstant(32, dl, MVT::i32));
4040 else /*if (VT == MVT::f32)*/
4041 Tmp0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp0);
4042 if (SrcVT == MVT::f32) {
4043 Tmp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp1);
4045 Tmp1 = DAG.getNode(ARMISD::VSHL, dl, OpVT,
4046 DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1),
4047 DAG.getConstant(32, dl, MVT::i32));
4048 } else if (VT == MVT::f32)
4049 Tmp1 = DAG.getNode(ARMISD::VSHRu, dl, MVT::v1i64,
4050 DAG.getNode(ISD::BITCAST, dl, MVT::v1i64, Tmp1),
4051 DAG.getConstant(32, dl, MVT::i32));
4052 Tmp0 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp0);
4053 Tmp1 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1);
4055 SDValue AllOnes = DAG.getTargetConstant(ARM_AM::createNEONModImm(0xe, 0xff),
4057 AllOnes = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v8i8, AllOnes);
4058 SDValue MaskNot = DAG.getNode(ISD::XOR, dl, OpVT, Mask,
4059 DAG.getNode(ISD::BITCAST, dl, OpVT, AllOnes));
4061 SDValue Res = DAG.getNode(ISD::OR, dl, OpVT,
4062 DAG.getNode(ISD::AND, dl, OpVT, Tmp1, Mask),
4063 DAG.getNode(ISD::AND, dl, OpVT, Tmp0, MaskNot));
4064 if (VT == MVT::f32) {
4065 Res = DAG.getNode(ISD::BITCAST, dl, MVT::v2f32, Res);
4066 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, Res,
4067 DAG.getConstant(0, dl, MVT::i32));
4069 Res = DAG.getNode(ISD::BITCAST, dl, MVT::f64, Res);
4075 // Bitcast operand 1 to i32.
4076 if (SrcVT == MVT::f64)
4077 Tmp1 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
4079 Tmp1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp1);
4081 // Or in the signbit with integer operations.
4082 SDValue Mask1 = DAG.getConstant(0x80000000, dl, MVT::i32);
4083 SDValue Mask2 = DAG.getConstant(0x7fffffff, dl, MVT::i32);
4084 Tmp1 = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp1, Mask1);
4085 if (VT == MVT::f32) {
4086 Tmp0 = DAG.getNode(ISD::AND, dl, MVT::i32,
4087 DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp0), Mask2);
4088 return DAG.getNode(ISD::BITCAST, dl, MVT::f32,
4089 DAG.getNode(ISD::OR, dl, MVT::i32, Tmp0, Tmp1));
4092 // f64: Or the high part with signbit and then combine two parts.
4093 Tmp0 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
4095 SDValue Lo = Tmp0.getValue(0);
4096 SDValue Hi = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp0.getValue(1), Mask2);
4097 Hi = DAG.getNode(ISD::OR, dl, MVT::i32, Hi, Tmp1);
4098 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
4101 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
4102 MachineFunction &MF = DAG.getMachineFunction();
4103 MachineFrameInfo *MFI = MF.getFrameInfo();
4104 MFI->setReturnAddressIsTaken(true);
4106 if (verifyReturnAddressArgumentIsConstant(Op, DAG))
4109 EVT VT = Op.getValueType();
4111 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
4113 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
4114 SDValue Offset = DAG.getConstant(4, dl, MVT::i32);
4115 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
4116 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
4117 MachinePointerInfo(), false, false, false, 0);
4120 // Return LR, which contains the return address. Mark it an implicit live-in.
4121 unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32));
4122 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
4125 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
4126 const ARMBaseRegisterInfo &ARI =
4127 *static_cast<const ARMBaseRegisterInfo*>(RegInfo);
4128 MachineFunction &MF = DAG.getMachineFunction();
4129 MachineFrameInfo *MFI = MF.getFrameInfo();
4130 MFI->setFrameAddressIsTaken(true);
4132 EVT VT = Op.getValueType();
4133 SDLoc dl(Op); // FIXME probably not meaningful
4134 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
4135 unsigned FrameReg = ARI.getFrameRegister(MF);
4136 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
4138 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
4139 MachinePointerInfo(),
4140 false, false, false, 0);
4144 // FIXME? Maybe this could be a TableGen attribute on some registers and
4145 // this table could be generated automatically from RegInfo.
4146 unsigned ARMTargetLowering::getRegisterByName(const char* RegName, EVT VT,
4147 SelectionDAG &DAG) const {
4148 unsigned Reg = StringSwitch<unsigned>(RegName)
4149 .Case("sp", ARM::SP)
4153 report_fatal_error(Twine("Invalid register name \""
4154 + StringRef(RegName) + "\"."));
4157 // Result is 64 bit value so split into two 32 bit values and return as a
4159 static void ExpandREAD_REGISTER(SDNode *N, SmallVectorImpl<SDValue> &Results,
4160 SelectionDAG &DAG) {
4163 // This function is only supposed to be called for i64 type destination.
4164 assert(N->getValueType(0) == MVT::i64
4165 && "ExpandREAD_REGISTER called for non-i64 type result.");
4167 SDValue Read = DAG.getNode(ISD::READ_REGISTER, DL,
4168 DAG.getVTList(MVT::i32, MVT::i32, MVT::Other),
4172 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Read.getValue(0),
4174 Results.push_back(Read.getOperand(0));
4177 /// ExpandBITCAST - If the target supports VFP, this function is called to
4178 /// expand a bit convert where either the source or destination type is i64 to
4179 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
4180 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
4181 /// vectors), since the legalizer won't know what to do with that.
4182 static SDValue ExpandBITCAST(SDNode *N, SelectionDAG &DAG) {
4183 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4185 SDValue Op = N->getOperand(0);
4187 // This function is only supposed to be called for i64 types, either as the
4188 // source or destination of the bit convert.
4189 EVT SrcVT = Op.getValueType();
4190 EVT DstVT = N->getValueType(0);
4191 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
4192 "ExpandBITCAST called for non-i64 type");
4194 // Turn i64->f64 into VMOVDRR.
4195 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
4196 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
4197 DAG.getConstant(0, dl, MVT::i32));
4198 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
4199 DAG.getConstant(1, dl, MVT::i32));
4200 return DAG.getNode(ISD::BITCAST, dl, DstVT,
4201 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
4204 // Turn f64->i64 into VMOVRRD.
4205 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
4207 if (DAG.getDataLayout().isBigEndian() && SrcVT.isVector() &&
4208 SrcVT.getVectorNumElements() > 1)
4209 Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
4210 DAG.getVTList(MVT::i32, MVT::i32),
4211 DAG.getNode(ARMISD::VREV64, dl, SrcVT, Op));
4213 Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
4214 DAG.getVTList(MVT::i32, MVT::i32), Op);
4215 // Merge the pieces into a single i64 value.
4216 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
4222 /// getZeroVector - Returns a vector of specified type with all zero elements.
4223 /// Zero vectors are used to represent vector negation and in those cases
4224 /// will be implemented with the NEON VNEG instruction. However, VNEG does
4225 /// not support i64 elements, so sometimes the zero vectors will need to be
4226 /// explicitly constructed. Regardless, use a canonical VMOV to create the
4228 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, SDLoc dl) {
4229 assert(VT.isVector() && "Expected a vector type");
4230 // The canonical modified immediate encoding of a zero vector is....0!
4231 SDValue EncodedVal = DAG.getTargetConstant(0, dl, MVT::i32);
4232 EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
4233 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal);
4234 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4237 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
4238 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
4239 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
4240 SelectionDAG &DAG) const {
4241 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4242 EVT VT = Op.getValueType();
4243 unsigned VTBits = VT.getSizeInBits();
4245 SDValue ShOpLo = Op.getOperand(0);
4246 SDValue ShOpHi = Op.getOperand(1);
4247 SDValue ShAmt = Op.getOperand(2);
4249 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
4251 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
4253 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
4254 DAG.getConstant(VTBits, dl, MVT::i32), ShAmt);
4255 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
4256 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
4257 DAG.getConstant(VTBits, dl, MVT::i32));
4258 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
4259 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
4260 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
4262 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
4263 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, dl, MVT::i32),
4264 ISD::SETGE, ARMcc, DAG, dl);
4265 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
4266 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc,
4269 SDValue Ops[2] = { Lo, Hi };
4270 return DAG.getMergeValues(Ops, dl);
4273 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
4274 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
4275 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
4276 SelectionDAG &DAG) const {
4277 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4278 EVT VT = Op.getValueType();
4279 unsigned VTBits = VT.getSizeInBits();
4281 SDValue ShOpLo = Op.getOperand(0);
4282 SDValue ShOpHi = Op.getOperand(1);
4283 SDValue ShAmt = Op.getOperand(2);
4286 assert(Op.getOpcode() == ISD::SHL_PARTS);
4287 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
4288 DAG.getConstant(VTBits, dl, MVT::i32), ShAmt);
4289 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
4290 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
4291 DAG.getConstant(VTBits, dl, MVT::i32));
4292 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
4293 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
4295 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
4296 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
4297 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, dl, MVT::i32),
4298 ISD::SETGE, ARMcc, DAG, dl);
4299 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
4300 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc,
4303 SDValue Ops[2] = { Lo, Hi };
4304 return DAG.getMergeValues(Ops, dl);
4307 SDValue ARMTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
4308 SelectionDAG &DAG) const {
4309 // The rounding mode is in bits 23:22 of the FPSCR.
4310 // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0
4311 // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3)
4312 // so that the shift + and get folded into a bitfield extract.
4314 SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::i32,
4315 DAG.getConstant(Intrinsic::arm_get_fpscr, dl,
4317 SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPSCR,
4318 DAG.getConstant(1U << 22, dl, MVT::i32));
4319 SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds,
4320 DAG.getConstant(22, dl, MVT::i32));
4321 return DAG.getNode(ISD::AND, dl, MVT::i32, RMODE,
4322 DAG.getConstant(3, dl, MVT::i32));
4325 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
4326 const ARMSubtarget *ST) {
4328 EVT VT = N->getValueType(0);
4329 if (VT.isVector()) {
4330 assert(ST->hasNEON());
4332 // Compute the least significant set bit: LSB = X & -X
4333 SDValue X = N->getOperand(0);
4334 SDValue NX = DAG.getNode(ISD::SUB, dl, VT, getZeroVector(VT, DAG, dl), X);
4335 SDValue LSB = DAG.getNode(ISD::AND, dl, VT, X, NX);
4337 EVT ElemTy = VT.getVectorElementType();
4339 if (ElemTy == MVT::i8) {
4340 // Compute with: cttz(x) = ctpop(lsb - 1)
4341 SDValue One = DAG.getNode(ARMISD::VMOVIMM, dl, VT,
4342 DAG.getTargetConstant(1, dl, ElemTy));
4343 SDValue Bits = DAG.getNode(ISD::SUB, dl, VT, LSB, One);
4344 return DAG.getNode(ISD::CTPOP, dl, VT, Bits);
4347 if ((ElemTy == MVT::i16 || ElemTy == MVT::i32) &&
4348 (N->getOpcode() == ISD::CTTZ_ZERO_UNDEF)) {
4349 // Compute with: cttz(x) = (width - 1) - ctlz(lsb), if x != 0
4350 unsigned NumBits = ElemTy.getSizeInBits();
4351 SDValue WidthMinus1 =
4352 DAG.getNode(ARMISD::VMOVIMM, dl, VT,
4353 DAG.getTargetConstant(NumBits - 1, dl, ElemTy));
4354 SDValue CTLZ = DAG.getNode(ISD::CTLZ, dl, VT, LSB);
4355 return DAG.getNode(ISD::SUB, dl, VT, WidthMinus1, CTLZ);
4358 // Compute with: cttz(x) = ctpop(lsb - 1)
4360 // Since we can only compute the number of bits in a byte with vcnt.8, we
4361 // have to gather the result with pairwise addition (vpaddl) for i16, i32,
4366 if (ElemTy == MVT::i64) {
4367 // Load constant 0xffff'ffff'ffff'ffff to register.
4368 SDValue FF = DAG.getNode(ARMISD::VMOVIMM, dl, VT,
4369 DAG.getTargetConstant(0x1eff, dl, MVT::i32));
4370 Bits = DAG.getNode(ISD::ADD, dl, VT, LSB, FF);
4372 SDValue One = DAG.getNode(ARMISD::VMOVIMM, dl, VT,
4373 DAG.getTargetConstant(1, dl, ElemTy));
4374 Bits = DAG.getNode(ISD::SUB, dl, VT, LSB, One);
4377 // Count #bits with vcnt.8.
4378 EVT VT8Bit = VT.is64BitVector() ? MVT::v8i8 : MVT::v16i8;
4379 SDValue BitsVT8 = DAG.getNode(ISD::BITCAST, dl, VT8Bit, Bits);
4380 SDValue Cnt8 = DAG.getNode(ISD::CTPOP, dl, VT8Bit, BitsVT8);
4382 // Gather the #bits with vpaddl (pairwise add.)
4383 EVT VT16Bit = VT.is64BitVector() ? MVT::v4i16 : MVT::v8i16;
4384 SDValue Cnt16 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT16Bit,
4385 DAG.getTargetConstant(Intrinsic::arm_neon_vpaddlu, dl, MVT::i32),
4387 if (ElemTy == MVT::i16)
4390 EVT VT32Bit = VT.is64BitVector() ? MVT::v2i32 : MVT::v4i32;
4391 SDValue Cnt32 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT32Bit,
4392 DAG.getTargetConstant(Intrinsic::arm_neon_vpaddlu, dl, MVT::i32),
4394 if (ElemTy == MVT::i32)
4397 assert(ElemTy == MVT::i64);
4398 SDValue Cnt64 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
4399 DAG.getTargetConstant(Intrinsic::arm_neon_vpaddlu, dl, MVT::i32),
4404 if (!ST->hasV6T2Ops())
4407 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
4408 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
4411 /// getCTPOP16BitCounts - Returns a v8i8/v16i8 vector containing the bit-count
4412 /// for each 16-bit element from operand, repeated. The basic idea is to
4413 /// leverage vcnt to get the 8-bit counts, gather and add the results.
4415 /// Trace for v4i16:
4416 /// input = [v0 v1 v2 v3 ] (vi 16-bit element)
4417 /// cast: N0 = [w0 w1 w2 w3 w4 w5 w6 w7] (v0 = [w0 w1], wi 8-bit element)
4418 /// vcnt: N1 = [b0 b1 b2 b3 b4 b5 b6 b7] (bi = bit-count of 8-bit element wi)
4419 /// vrev: N2 = [b1 b0 b3 b2 b5 b4 b7 b6]
4420 /// [b0 b1 b2 b3 b4 b5 b6 b7]
4421 /// +[b1 b0 b3 b2 b5 b4 b7 b6]
4422 /// N3=N1+N2 = [k0 k0 k1 k1 k2 k2 k3 k3] (k0 = b0+b1 = bit-count of 16-bit v0,
4423 /// vuzp: = [k0 k1 k2 k3 k0 k1 k2 k3] each ki is 8-bits)
4424 static SDValue getCTPOP16BitCounts(SDNode *N, SelectionDAG &DAG) {
4425 EVT VT = N->getValueType(0);
4428 EVT VT8Bit = VT.is64BitVector() ? MVT::v8i8 : MVT::v16i8;
4429 SDValue N0 = DAG.getNode(ISD::BITCAST, DL, VT8Bit, N->getOperand(0));
4430 SDValue N1 = DAG.getNode(ISD::CTPOP, DL, VT8Bit, N0);
4431 SDValue N2 = DAG.getNode(ARMISD::VREV16, DL, VT8Bit, N1);
4432 SDValue N3 = DAG.getNode(ISD::ADD, DL, VT8Bit, N1, N2);
4433 return DAG.getNode(ARMISD::VUZP, DL, VT8Bit, N3, N3);
4436 /// lowerCTPOP16BitElements - Returns a v4i16/v8i16 vector containing the
4437 /// bit-count for each 16-bit element from the operand. We need slightly
4438 /// different sequencing for v4i16 and v8i16 to stay within NEON's available
4439 /// 64/128-bit registers.
4441 /// Trace for v4i16:
4442 /// input = [v0 v1 v2 v3 ] (vi 16-bit element)
4443 /// v8i8: BitCounts = [k0 k1 k2 k3 k0 k1 k2 k3 ] (ki is the bit-count of vi)
4444 /// v8i16:Extended = [k0 k1 k2 k3 k0 k1 k2 k3 ]
4445 /// v4i16:Extracted = [k0 k1 k2 k3 ]
4446 static SDValue lowerCTPOP16BitElements(SDNode *N, SelectionDAG &DAG) {
4447 EVT VT = N->getValueType(0);
4450 SDValue BitCounts = getCTPOP16BitCounts(N, DAG);
4451 if (VT.is64BitVector()) {
4452 SDValue Extended = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v8i16, BitCounts);
4453 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i16, Extended,
4454 DAG.getIntPtrConstant(0, DL));
4456 SDValue Extracted = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v8i8,
4457 BitCounts, DAG.getIntPtrConstant(0, DL));
4458 return DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v8i16, Extracted);
4462 /// lowerCTPOP32BitElements - Returns a v2i32/v4i32 vector containing the
4463 /// bit-count for each 32-bit element from the operand. The idea here is
4464 /// to split the vector into 16-bit elements, leverage the 16-bit count
4465 /// routine, and then combine the results.
4467 /// Trace for v2i32 (v4i32 similar with Extracted/Extended exchanged):
4468 /// input = [v0 v1 ] (vi: 32-bit elements)
4469 /// Bitcast = [w0 w1 w2 w3 ] (wi: 16-bit elements, v0 = [w0 w1])
4470 /// Counts16 = [k0 k1 k2 k3 ] (ki: 16-bit elements, bit-count of wi)
4471 /// vrev: N0 = [k1 k0 k3 k2 ]
4473 /// N1 =+[k1 k0 k3 k2 ]
4475 /// N2 =+[k1 k3 k0 k2 ]
4477 /// Extended =+[k1 k3 k0 k2 ]
4479 /// Extracted=+[k1 k3 ]
4481 static SDValue lowerCTPOP32BitElements(SDNode *N, SelectionDAG &DAG) {
4482 EVT VT = N->getValueType(0);
4485 EVT VT16Bit = VT.is64BitVector() ? MVT::v4i16 : MVT::v8i16;
4487 SDValue Bitcast = DAG.getNode(ISD::BITCAST, DL, VT16Bit, N->getOperand(0));
4488 SDValue Counts16 = lowerCTPOP16BitElements(Bitcast.getNode(), DAG);
4489 SDValue N0 = DAG.getNode(ARMISD::VREV32, DL, VT16Bit, Counts16);
4490 SDValue N1 = DAG.getNode(ISD::ADD, DL, VT16Bit, Counts16, N0);
4491 SDValue N2 = DAG.getNode(ARMISD::VUZP, DL, VT16Bit, N1, N1);
4493 if (VT.is64BitVector()) {
4494 SDValue Extended = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v4i32, N2);
4495 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i32, Extended,
4496 DAG.getIntPtrConstant(0, DL));
4498 SDValue Extracted = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i16, N2,
4499 DAG.getIntPtrConstant(0, DL));
4500 return DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v4i32, Extracted);
4504 static SDValue LowerCTPOP(SDNode *N, SelectionDAG &DAG,
4505 const ARMSubtarget *ST) {
4506 EVT VT = N->getValueType(0);
4508 assert(ST->hasNEON() && "Custom ctpop lowering requires NEON.");
4509 assert((VT == MVT::v2i32 || VT == MVT::v4i32 ||
4510 VT == MVT::v4i16 || VT == MVT::v8i16) &&
4511 "Unexpected type for custom ctpop lowering");
4513 if (VT.getVectorElementType() == MVT::i32)
4514 return lowerCTPOP32BitElements(N, DAG);
4516 return lowerCTPOP16BitElements(N, DAG);
4519 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
4520 const ARMSubtarget *ST) {
4521 EVT VT = N->getValueType(0);
4527 // Lower vector shifts on NEON to use VSHL.
4528 assert(ST->hasNEON() && "unexpected vector shift");
4530 // Left shifts translate directly to the vshiftu intrinsic.
4531 if (N->getOpcode() == ISD::SHL)
4532 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
4533 DAG.getConstant(Intrinsic::arm_neon_vshiftu, dl,
4535 N->getOperand(0), N->getOperand(1));
4537 assert((N->getOpcode() == ISD::SRA ||
4538 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
4540 // NEON uses the same intrinsics for both left and right shifts. For
4541 // right shifts, the shift amounts are negative, so negate the vector of
4543 EVT ShiftVT = N->getOperand(1).getValueType();
4544 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
4545 getZeroVector(ShiftVT, DAG, dl),
4547 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
4548 Intrinsic::arm_neon_vshifts :
4549 Intrinsic::arm_neon_vshiftu);
4550 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
4551 DAG.getConstant(vshiftInt, dl, MVT::i32),
4552 N->getOperand(0), NegatedCount);
4555 static SDValue Expand64BitShift(SDNode *N, SelectionDAG &DAG,
4556 const ARMSubtarget *ST) {
4557 EVT VT = N->getValueType(0);
4560 // We can get here for a node like i32 = ISD::SHL i32, i64
4564 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
4565 "Unknown shift to lower!");
4567 // We only lower SRA, SRL of 1 here, all others use generic lowering.
4568 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
4569 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
4572 // If we are in thumb mode, we don't have RRX.
4573 if (ST->isThumb1Only()) return SDValue();
4575 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
4576 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
4577 DAG.getConstant(0, dl, MVT::i32));
4578 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
4579 DAG.getConstant(1, dl, MVT::i32));
4581 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
4582 // captures the result into a carry flag.
4583 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
4584 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Glue), Hi);
4586 // The low part is an ARMISD::RRX operand, which shifts the carry in.
4587 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
4589 // Merge the pieces into a single i64 value.
4590 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
4593 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
4594 SDValue TmpOp0, TmpOp1;
4595 bool Invert = false;
4599 SDValue Op0 = Op.getOperand(0);
4600 SDValue Op1 = Op.getOperand(1);
4601 SDValue CC = Op.getOperand(2);
4602 EVT CmpVT = Op0.getValueType().changeVectorElementTypeToInteger();
4603 EVT VT = Op.getValueType();
4604 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
4607 if (Op1.getValueType().isFloatingPoint()) {
4608 switch (SetCCOpcode) {
4609 default: llvm_unreachable("Illegal FP comparison");
4611 case ISD::SETNE: Invert = true; // Fallthrough
4613 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
4615 case ISD::SETLT: Swap = true; // Fallthrough
4617 case ISD::SETGT: Opc = ARMISD::VCGT; break;
4619 case ISD::SETLE: Swap = true; // Fallthrough
4621 case ISD::SETGE: Opc = ARMISD::VCGE; break;
4622 case ISD::SETUGE: Swap = true; // Fallthrough
4623 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
4624 case ISD::SETUGT: Swap = true; // Fallthrough
4625 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
4626 case ISD::SETUEQ: Invert = true; // Fallthrough
4628 // Expand this to (OLT | OGT).
4632 Op0 = DAG.getNode(ARMISD::VCGT, dl, CmpVT, TmpOp1, TmpOp0);
4633 Op1 = DAG.getNode(ARMISD::VCGT, dl, CmpVT, TmpOp0, TmpOp1);
4635 case ISD::SETUO: Invert = true; // Fallthrough
4637 // Expand this to (OLT | OGE).
4641 Op0 = DAG.getNode(ARMISD::VCGT, dl, CmpVT, TmpOp1, TmpOp0);
4642 Op1 = DAG.getNode(ARMISD::VCGE, dl, CmpVT, TmpOp0, TmpOp1);
4646 // Integer comparisons.
4647 switch (SetCCOpcode) {
4648 default: llvm_unreachable("Illegal integer comparison");
4649 case ISD::SETNE: Invert = true;
4650 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
4651 case ISD::SETLT: Swap = true;
4652 case ISD::SETGT: Opc = ARMISD::VCGT; break;
4653 case ISD::SETLE: Swap = true;
4654 case ISD::SETGE: Opc = ARMISD::VCGE; break;
4655 case ISD::SETULT: Swap = true;
4656 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
4657 case ISD::SETULE: Swap = true;
4658 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
4661 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
4662 if (Opc == ARMISD::VCEQ) {
4665 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
4667 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
4670 // Ignore bitconvert.
4671 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BITCAST)
4672 AndOp = AndOp.getOperand(0);
4674 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
4676 Op0 = DAG.getNode(ISD::BITCAST, dl, CmpVT, AndOp.getOperand(0));
4677 Op1 = DAG.getNode(ISD::BITCAST, dl, CmpVT, AndOp.getOperand(1));
4684 std::swap(Op0, Op1);
4686 // If one of the operands is a constant vector zero, attempt to fold the
4687 // comparison to a specialized compare-against-zero form.
4689 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
4691 else if (ISD::isBuildVectorAllZeros(Op0.getNode())) {
4692 if (Opc == ARMISD::VCGE)
4693 Opc = ARMISD::VCLEZ;
4694 else if (Opc == ARMISD::VCGT)
4695 Opc = ARMISD::VCLTZ;
4700 if (SingleOp.getNode()) {
4703 Result = DAG.getNode(ARMISD::VCEQZ, dl, CmpVT, SingleOp); break;
4705 Result = DAG.getNode(ARMISD::VCGEZ, dl, CmpVT, SingleOp); break;
4707 Result = DAG.getNode(ARMISD::VCLEZ, dl, CmpVT, SingleOp); break;
4709 Result = DAG.getNode(ARMISD::VCGTZ, dl, CmpVT, SingleOp); break;
4711 Result = DAG.getNode(ARMISD::VCLTZ, dl, CmpVT, SingleOp); break;
4713 Result = DAG.getNode(Opc, dl, CmpVT, Op0, Op1);
4716 Result = DAG.getNode(Opc, dl, CmpVT, Op0, Op1);
4719 Result = DAG.getSExtOrTrunc(Result, dl, VT);
4722 Result = DAG.getNOT(dl, Result, VT);
4727 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
4728 /// valid vector constant for a NEON instruction with a "modified immediate"
4729 /// operand (e.g., VMOV). If so, return the encoded value.
4730 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
4731 unsigned SplatBitSize, SelectionDAG &DAG,
4732 SDLoc dl, EVT &VT, bool is128Bits,
4733 NEONModImmType type) {
4734 unsigned OpCmode, Imm;
4736 // SplatBitSize is set to the smallest size that splats the vector, so a
4737 // zero vector will always have SplatBitSize == 8. However, NEON modified
4738 // immediate instructions others than VMOV do not support the 8-bit encoding
4739 // of a zero vector, and the default encoding of zero is supposed to be the
4744 switch (SplatBitSize) {
4746 if (type != VMOVModImm)
4748 // Any 1-byte value is OK. Op=0, Cmode=1110.
4749 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
4752 VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
4756 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
4757 VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
4758 if ((SplatBits & ~0xff) == 0) {
4759 // Value = 0x00nn: Op=x, Cmode=100x.
4764 if ((SplatBits & ~0xff00) == 0) {
4765 // Value = 0xnn00: Op=x, Cmode=101x.
4767 Imm = SplatBits >> 8;
4773 // NEON's 32-bit VMOV supports splat values where:
4774 // * only one byte is nonzero, or
4775 // * the least significant byte is 0xff and the second byte is nonzero, or
4776 // * the least significant 2 bytes are 0xff and the third is nonzero.
4777 VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
4778 if ((SplatBits & ~0xff) == 0) {
4779 // Value = 0x000000nn: Op=x, Cmode=000x.
4784 if ((SplatBits & ~0xff00) == 0) {
4785 // Value = 0x0000nn00: Op=x, Cmode=001x.
4787 Imm = SplatBits >> 8;
4790 if ((SplatBits & ~0xff0000) == 0) {
4791 // Value = 0x00nn0000: Op=x, Cmode=010x.
4793 Imm = SplatBits >> 16;
4796 if ((SplatBits & ~0xff000000) == 0) {
4797 // Value = 0xnn000000: Op=x, Cmode=011x.
4799 Imm = SplatBits >> 24;
4803 // cmode == 0b1100 and cmode == 0b1101 are not supported for VORR or VBIC
4804 if (type == OtherModImm) return SDValue();
4806 if ((SplatBits & ~0xffff) == 0 &&
4807 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
4808 // Value = 0x0000nnff: Op=x, Cmode=1100.
4810 Imm = SplatBits >> 8;
4814 if ((SplatBits & ~0xffffff) == 0 &&
4815 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
4816 // Value = 0x00nnffff: Op=x, Cmode=1101.
4818 Imm = SplatBits >> 16;
4822 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
4823 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
4824 // VMOV.I32. A (very) minor optimization would be to replicate the value
4825 // and fall through here to test for a valid 64-bit splat. But, then the
4826 // caller would also need to check and handle the change in size.
4830 if (type != VMOVModImm)
4832 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
4833 uint64_t BitMask = 0xff;
4835 unsigned ImmMask = 1;
4837 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
4838 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
4841 } else if ((SplatBits & BitMask) != 0) {
4848 if (DAG.getDataLayout().isBigEndian())
4849 // swap higher and lower 32 bit word
4850 Imm = ((Imm & 0xf) << 4) | ((Imm & 0xf0) >> 4);
4852 // Op=1, Cmode=1110.
4854 VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
4859 llvm_unreachable("unexpected size for isNEONModifiedImm");
4862 unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm);
4863 return DAG.getTargetConstant(EncodedVal, dl, MVT::i32);
4866 SDValue ARMTargetLowering::LowerConstantFP(SDValue Op, SelectionDAG &DAG,
4867 const ARMSubtarget *ST) const {
4871 bool IsDouble = Op.getValueType() == MVT::f64;
4872 ConstantFPSDNode *CFP = cast<ConstantFPSDNode>(Op);
4874 // Use the default (constant pool) lowering for double constants when we have
4876 if (IsDouble && Subtarget->isFPOnlySP())
4879 // Try splatting with a VMOV.f32...
4880 APFloat FPVal = CFP->getValueAPF();
4881 int ImmVal = IsDouble ? ARM_AM::getFP64Imm(FPVal) : ARM_AM::getFP32Imm(FPVal);
4884 if (IsDouble || !ST->useNEONForSinglePrecisionFP()) {
4885 // We have code in place to select a valid ConstantFP already, no need to
4890 // It's a float and we are trying to use NEON operations where
4891 // possible. Lower it to a splat followed by an extract.
4893 SDValue NewVal = DAG.getTargetConstant(ImmVal, DL, MVT::i32);
4894 SDValue VecConstant = DAG.getNode(ARMISD::VMOVFPIMM, DL, MVT::v2f32,
4896 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecConstant,
4897 DAG.getConstant(0, DL, MVT::i32));
4900 // The rest of our options are NEON only, make sure that's allowed before
4902 if (!ST->hasNEON() || (!IsDouble && !ST->useNEONForSinglePrecisionFP()))
4906 uint64_t iVal = FPVal.bitcastToAPInt().getZExtValue();
4908 // It wouldn't really be worth bothering for doubles except for one very
4909 // important value, which does happen to match: 0.0. So make sure we don't do
4911 if (IsDouble && (iVal & 0xffffffff) != (iVal >> 32))
4914 // Try a VMOV.i32 (FIXME: i8, i16, or i64 could work too).
4915 SDValue NewVal = isNEONModifiedImm(iVal & 0xffffffffU, 0, 32, DAG, SDLoc(Op),
4916 VMovVT, false, VMOVModImm);
4917 if (NewVal != SDValue()) {
4919 SDValue VecConstant = DAG.getNode(ARMISD::VMOVIMM, DL, VMovVT,
4922 return DAG.getNode(ISD::BITCAST, DL, MVT::f64, VecConstant);
4924 // It's a float: cast and extract a vector element.
4925 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
4927 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
4928 DAG.getConstant(0, DL, MVT::i32));
4931 // Finally, try a VMVN.i32
4932 NewVal = isNEONModifiedImm(~iVal & 0xffffffffU, 0, 32, DAG, SDLoc(Op), VMovVT,
4934 if (NewVal != SDValue()) {
4936 SDValue VecConstant = DAG.getNode(ARMISD::VMVNIMM, DL, VMovVT, NewVal);
4939 return DAG.getNode(ISD::BITCAST, DL, MVT::f64, VecConstant);
4941 // It's a float: cast and extract a vector element.
4942 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
4944 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
4945 DAG.getConstant(0, DL, MVT::i32));
4951 // check if an VEXT instruction can handle the shuffle mask when the
4952 // vector sources of the shuffle are the same.
4953 static bool isSingletonVEXTMask(ArrayRef<int> M, EVT VT, unsigned &Imm) {
4954 unsigned NumElts = VT.getVectorNumElements();
4956 // Assume that the first shuffle index is not UNDEF. Fail if it is.
4962 // If this is a VEXT shuffle, the immediate value is the index of the first
4963 // element. The other shuffle indices must be the successive elements after
4965 unsigned ExpectedElt = Imm;
4966 for (unsigned i = 1; i < NumElts; ++i) {
4967 // Increment the expected index. If it wraps around, just follow it
4968 // back to index zero and keep going.
4970 if (ExpectedElt == NumElts)
4973 if (M[i] < 0) continue; // ignore UNDEF indices
4974 if (ExpectedElt != static_cast<unsigned>(M[i]))
4982 static bool isVEXTMask(ArrayRef<int> M, EVT VT,
4983 bool &ReverseVEXT, unsigned &Imm) {
4984 unsigned NumElts = VT.getVectorNumElements();
4985 ReverseVEXT = false;
4987 // Assume that the first shuffle index is not UNDEF. Fail if it is.
4993 // If this is a VEXT shuffle, the immediate value is the index of the first
4994 // element. The other shuffle indices must be the successive elements after
4996 unsigned ExpectedElt = Imm;
4997 for (unsigned i = 1; i < NumElts; ++i) {
4998 // Increment the expected index. If it wraps around, it may still be
4999 // a VEXT but the source vectors must be swapped.
5001 if (ExpectedElt == NumElts * 2) {
5006 if (M[i] < 0) continue; // ignore UNDEF indices
5007 if (ExpectedElt != static_cast<unsigned>(M[i]))
5011 // Adjust the index value if the source operands will be swapped.
5018 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
5019 /// instruction with the specified blocksize. (The order of the elements
5020 /// within each block of the vector is reversed.)
5021 static bool isVREVMask(ArrayRef<int> M, EVT VT, unsigned BlockSize) {
5022 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
5023 "Only possible block sizes for VREV are: 16, 32, 64");
5025 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
5029 unsigned NumElts = VT.getVectorNumElements();
5030 unsigned BlockElts = M[0] + 1;
5031 // If the first shuffle index is UNDEF, be optimistic.
5033 BlockElts = BlockSize / EltSz;
5035 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
5038 for (unsigned i = 0; i < NumElts; ++i) {
5039 if (M[i] < 0) continue; // ignore UNDEF indices
5040 if ((unsigned) M[i] != (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
5047 static bool isVTBLMask(ArrayRef<int> M, EVT VT) {
5048 // We can handle <8 x i8> vector shuffles. If the index in the mask is out of
5049 // range, then 0 is placed into the resulting vector. So pretty much any mask
5050 // of 8 elements can work here.
5051 return VT == MVT::v8i8 && M.size() == 8;
5054 // Checks whether the shuffle mask represents a vector transpose (VTRN) by
5055 // checking that pairs of elements in the shuffle mask represent the same index
5056 // in each vector, incrementing the expected index by 2 at each step.
5057 // e.g. For v1,v2 of type v4i32 a valid shuffle mask is: [0, 4, 2, 6]
5058 // v1={a,b,c,d} => x=shufflevector v1, v2 shufflemask => x={a,e,c,g}
5060 // WhichResult gives the offset for each element in the mask based on which
5061 // of the two results it belongs to.
5063 // The transpose can be represented either as:
5064 // result1 = shufflevector v1, v2, result1_shuffle_mask
5065 // result2 = shufflevector v1, v2, result2_shuffle_mask
5066 // where v1/v2 and the shuffle masks have the same number of elements
5067 // (here WhichResult (see below) indicates which result is being checked)
5070 // results = shufflevector v1, v2, shuffle_mask
5071 // where both results are returned in one vector and the shuffle mask has twice
5072 // as many elements as v1/v2 (here WhichResult will always be 0 if true) here we
5073 // want to check the low half and high half of the shuffle mask as if it were
5075 static bool isVTRNMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
5076 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
5080 unsigned NumElts = VT.getVectorNumElements();
5081 if (M.size() != NumElts && M.size() != NumElts*2)
5084 // If the mask is twice as long as the result then we need to check the upper
5085 // and lower parts of the mask
5086 for (unsigned i = 0; i < M.size(); i += NumElts) {
5087 WhichResult = M[i] == 0 ? 0 : 1;
5088 for (unsigned j = 0; j < NumElts; j += 2) {
5089 if ((M[i+j] >= 0 && (unsigned) M[i+j] != j + WhichResult) ||
5090 (M[i+j+1] >= 0 && (unsigned) M[i+j+1] != j + NumElts + WhichResult))
5095 if (M.size() == NumElts*2)
5101 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
5102 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
5103 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
5104 static bool isVTRN_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
5105 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
5109 unsigned NumElts = VT.getVectorNumElements();
5110 if (M.size() != NumElts && M.size() != NumElts*2)
5113 for (unsigned i = 0; i < M.size(); i += NumElts) {
5114 WhichResult = M[i] == 0 ? 0 : 1;
5115 for (unsigned j = 0; j < NumElts; j += 2) {
5116 if ((M[i+j] >= 0 && (unsigned) M[i+j] != j + WhichResult) ||
5117 (M[i+j+1] >= 0 && (unsigned) M[i+j+1] != j + WhichResult))
5122 if (M.size() == NumElts*2)
5128 // Checks whether the shuffle mask represents a vector unzip (VUZP) by checking
5129 // that the mask elements are either all even and in steps of size 2 or all odd
5130 // and in steps of size 2.
5131 // e.g. For v1,v2 of type v4i32 a valid shuffle mask is: [0, 2, 4, 6]
5132 // v1={a,b,c,d} => x=shufflevector v1, v2 shufflemask => x={a,c,e,g}
5134 // Requires similar checks to that of isVTRNMask with
5135 // respect the how results are returned.
5136 static bool isVUZPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
5137 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
5141 unsigned NumElts = VT.getVectorNumElements();
5142 if (M.size() != NumElts && M.size() != NumElts*2)
5145 for (unsigned i = 0; i < M.size(); i += NumElts) {
5146 WhichResult = M[i] == 0 ? 0 : 1;
5147 for (unsigned j = 0; j < NumElts; ++j) {
5148 if (M[i+j] >= 0 && (unsigned) M[i+j] != 2 * j + WhichResult)
5153 if (M.size() == NumElts*2)
5156 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
5157 if (VT.is64BitVector() && EltSz == 32)
5163 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
5164 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
5165 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
5166 static bool isVUZP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
5167 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
5171 unsigned NumElts = VT.getVectorNumElements();
5172 if (M.size() != NumElts && M.size() != NumElts*2)
5175 unsigned Half = NumElts / 2;
5176 for (unsigned i = 0; i < M.size(); i += NumElts) {
5177 WhichResult = M[i] == 0 ? 0 : 1;
5178 for (unsigned j = 0; j < NumElts; j += Half) {
5179 unsigned Idx = WhichResult;
5180 for (unsigned k = 0; k < Half; ++k) {
5181 int MIdx = M[i + j + k];
5182 if (MIdx >= 0 && (unsigned) MIdx != Idx)
5189 if (M.size() == NumElts*2)
5192 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
5193 if (VT.is64BitVector() && EltSz == 32)
5199 // Checks whether the shuffle mask represents a vector zip (VZIP) by checking
5200 // that pairs of elements of the shufflemask represent the same index in each
5201 // vector incrementing sequentially through the vectors.
5202 // e.g. For v1,v2 of type v4i32 a valid shuffle mask is: [0, 4, 1, 5]
5203 // v1={a,b,c,d} => x=shufflevector v1, v2 shufflemask => x={a,e,b,f}
5205 // Requires similar checks to that of isVTRNMask with respect the how results
5207 static bool isVZIPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
5208 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
5212 unsigned NumElts = VT.getVectorNumElements();
5213 if (M.size() != NumElts && M.size() != NumElts*2)
5216 for (unsigned i = 0; i < M.size(); i += NumElts) {
5217 WhichResult = M[i] == 0 ? 0 : 1;
5218 unsigned Idx = WhichResult * NumElts / 2;
5219 for (unsigned j = 0; j < NumElts; j += 2) {
5220 if ((M[i+j] >= 0 && (unsigned) M[i+j] != Idx) ||
5221 (M[i+j+1] >= 0 && (unsigned) M[i+j+1] != Idx + NumElts))
5227 if (M.size() == NumElts*2)
5230 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
5231 if (VT.is64BitVector() && EltSz == 32)
5237 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
5238 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
5239 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
5240 static bool isVZIP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
5241 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
5245 unsigned NumElts = VT.getVectorNumElements();
5246 if (M.size() != NumElts && M.size() != NumElts*2)
5249 for (unsigned i = 0; i < M.size(); i += NumElts) {
5250 WhichResult = M[i] == 0 ? 0 : 1;
5251 unsigned Idx = WhichResult * NumElts / 2;
5252 for (unsigned j = 0; j < NumElts; j += 2) {
5253 if ((M[i+j] >= 0 && (unsigned) M[i+j] != Idx) ||
5254 (M[i+j+1] >= 0 && (unsigned) M[i+j+1] != Idx))
5260 if (M.size() == NumElts*2)
5263 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
5264 if (VT.is64BitVector() && EltSz == 32)
5270 /// Check if \p ShuffleMask is a NEON two-result shuffle (VZIP, VUZP, VTRN),
5271 /// and return the corresponding ARMISD opcode if it is, or 0 if it isn't.
5272 static unsigned isNEONTwoResultShuffleMask(ArrayRef<int> ShuffleMask, EVT VT,
5273 unsigned &WhichResult,
5276 if (isVTRNMask(ShuffleMask, VT, WhichResult))
5277 return ARMISD::VTRN;
5278 if (isVUZPMask(ShuffleMask, VT, WhichResult))
5279 return ARMISD::VUZP;
5280 if (isVZIPMask(ShuffleMask, VT, WhichResult))
5281 return ARMISD::VZIP;
5284 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
5285 return ARMISD::VTRN;
5286 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
5287 return ARMISD::VUZP;
5288 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
5289 return ARMISD::VZIP;
5294 /// \return true if this is a reverse operation on an vector.
5295 static bool isReverseMask(ArrayRef<int> M, EVT VT) {
5296 unsigned NumElts = VT.getVectorNumElements();
5297 // Make sure the mask has the right size.
5298 if (NumElts != M.size())
5301 // Look for <15, ..., 3, -1, 1, 0>.
5302 for (unsigned i = 0; i != NumElts; ++i)
5303 if (M[i] >= 0 && M[i] != (int) (NumElts - 1 - i))
5309 // If N is an integer constant that can be moved into a register in one
5310 // instruction, return an SDValue of such a constant (will become a MOV
5311 // instruction). Otherwise return null.
5312 static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG,
5313 const ARMSubtarget *ST, SDLoc dl) {
5315 if (!isa<ConstantSDNode>(N))
5317 Val = cast<ConstantSDNode>(N)->getZExtValue();
5319 if (ST->isThumb1Only()) {
5320 if (Val <= 255 || ~Val <= 255)
5321 return DAG.getConstant(Val, dl, MVT::i32);
5323 if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1)
5324 return DAG.getConstant(Val, dl, MVT::i32);
5329 // If this is a case we can't handle, return null and let the default
5330 // expansion code take care of it.
5331 SDValue ARMTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
5332 const ARMSubtarget *ST) const {
5333 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
5335 EVT VT = Op.getValueType();
5337 APInt SplatBits, SplatUndef;
5338 unsigned SplatBitSize;
5340 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
5341 if (SplatBitSize <= 64) {
5342 // Check if an immediate VMOV works.
5344 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
5345 SplatUndef.getZExtValue(), SplatBitSize,
5346 DAG, dl, VmovVT, VT.is128BitVector(),
5348 if (Val.getNode()) {
5349 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
5350 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
5353 // Try an immediate VMVN.
5354 uint64_t NegatedImm = (~SplatBits).getZExtValue();
5355 Val = isNEONModifiedImm(NegatedImm,
5356 SplatUndef.getZExtValue(), SplatBitSize,
5357 DAG, dl, VmovVT, VT.is128BitVector(),
5359 if (Val.getNode()) {
5360 SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
5361 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
5364 // Use vmov.f32 to materialize other v2f32 and v4f32 splats.
5365 if ((VT == MVT::v2f32 || VT == MVT::v4f32) && SplatBitSize == 32) {
5366 int ImmVal = ARM_AM::getFP32Imm(SplatBits);
5368 SDValue Val = DAG.getTargetConstant(ImmVal, dl, MVT::i32);
5369 return DAG.getNode(ARMISD::VMOVFPIMM, dl, VT, Val);
5375 // Scan through the operands to see if only one value is used.
5377 // As an optimisation, even if more than one value is used it may be more
5378 // profitable to splat with one value then change some lanes.
5380 // Heuristically we decide to do this if the vector has a "dominant" value,
5381 // defined as splatted to more than half of the lanes.
5382 unsigned NumElts = VT.getVectorNumElements();
5383 bool isOnlyLowElement = true;
5384 bool usesOnlyOneValue = true;
5385 bool hasDominantValue = false;
5386 bool isConstant = true;
5388 // Map of the number of times a particular SDValue appears in the
5390 DenseMap<SDValue, unsigned> ValueCounts;
5392 for (unsigned i = 0; i < NumElts; ++i) {
5393 SDValue V = Op.getOperand(i);
5394 if (V.getOpcode() == ISD::UNDEF)
5397 isOnlyLowElement = false;
5398 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
5401 ValueCounts.insert(std::make_pair(V, 0));
5402 unsigned &Count = ValueCounts[V];
5404 // Is this value dominant? (takes up more than half of the lanes)
5405 if (++Count > (NumElts / 2)) {
5406 hasDominantValue = true;
5410 if (ValueCounts.size() != 1)
5411 usesOnlyOneValue = false;
5412 if (!Value.getNode() && ValueCounts.size() > 0)
5413 Value = ValueCounts.begin()->first;
5415 if (ValueCounts.size() == 0)
5416 return DAG.getUNDEF(VT);
5418 // Loads are better lowered with insert_vector_elt/ARMISD::BUILD_VECTOR.
5419 // Keep going if we are hitting this case.
5420 if (isOnlyLowElement && !ISD::isNormalLoad(Value.getNode()))
5421 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
5423 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5425 // Use VDUP for non-constant splats. For f32 constant splats, reduce to
5426 // i32 and try again.
5427 if (hasDominantValue && EltSize <= 32) {
5431 // If we are VDUPing a value that comes directly from a vector, that will
5432 // cause an unnecessary move to and from a GPR, where instead we could
5433 // just use VDUPLANE. We can only do this if the lane being extracted
5434 // is at a constant index, as the VDUP from lane instructions only have
5435 // constant-index forms.
5436 if (Value->getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5437 isa<ConstantSDNode>(Value->getOperand(1))) {
5438 // We need to create a new undef vector to use for the VDUPLANE if the
5439 // size of the vector from which we get the value is different than the
5440 // size of the vector that we need to create. We will insert the element
5441 // such that the register coalescer will remove unnecessary copies.
5442 if (VT != Value->getOperand(0).getValueType()) {
5443 ConstantSDNode *constIndex;
5444 constIndex = dyn_cast<ConstantSDNode>(Value->getOperand(1));
5445 assert(constIndex && "The index is not a constant!");
5446 unsigned index = constIndex->getAPIntValue().getLimitedValue() %
5447 VT.getVectorNumElements();
5448 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
5449 DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DAG.getUNDEF(VT),
5450 Value, DAG.getConstant(index, dl, MVT::i32)),
5451 DAG.getConstant(index, dl, MVT::i32));
5453 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
5454 Value->getOperand(0), Value->getOperand(1));
5456 N = DAG.getNode(ARMISD::VDUP, dl, VT, Value);
5458 if (!usesOnlyOneValue) {
5459 // The dominant value was splatted as 'N', but we now have to insert
5460 // all differing elements.
5461 for (unsigned I = 0; I < NumElts; ++I) {
5462 if (Op.getOperand(I) == Value)
5464 SmallVector<SDValue, 3> Ops;
5466 Ops.push_back(Op.getOperand(I));
5467 Ops.push_back(DAG.getConstant(I, dl, MVT::i32));
5468 N = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Ops);
5473 if (VT.getVectorElementType().isFloatingPoint()) {
5474 SmallVector<SDValue, 8> Ops;
5475 for (unsigned i = 0; i < NumElts; ++i)
5476 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, MVT::i32,
5478 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
5479 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, Ops);
5480 Val = LowerBUILD_VECTOR(Val, DAG, ST);
5482 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
5484 if (usesOnlyOneValue) {
5485 SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
5486 if (isConstant && Val.getNode())
5487 return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
5491 // If all elements are constants and the case above didn't get hit, fall back
5492 // to the default expansion, which will generate a load from the constant
5497 // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
5499 SDValue shuffle = ReconstructShuffle(Op, DAG);
5500 if (shuffle != SDValue())
5504 // Vectors with 32- or 64-bit elements can be built by directly assigning
5505 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
5506 // will be legalized.
5507 if (EltSize >= 32) {
5508 // Do the expansion with floating-point types, since that is what the VFP
5509 // registers are defined to use, and since i64 is not legal.
5510 EVT EltVT = EVT::getFloatingPointVT(EltSize);
5511 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
5512 SmallVector<SDValue, 8> Ops;
5513 for (unsigned i = 0; i < NumElts; ++i)
5514 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, EltVT, Op.getOperand(i)));
5515 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, Ops);
5516 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
5519 // If all else fails, just use a sequence of INSERT_VECTOR_ELT when we
5520 // know the default expansion would otherwise fall back on something even
5521 // worse. For a vector with one or two non-undef values, that's
5522 // scalar_to_vector for the elements followed by a shuffle (provided the
5523 // shuffle is valid for the target) and materialization element by element
5524 // on the stack followed by a load for everything else.
5525 if (!isConstant && !usesOnlyOneValue) {
5526 SDValue Vec = DAG.getUNDEF(VT);
5527 for (unsigned i = 0 ; i < NumElts; ++i) {
5528 SDValue V = Op.getOperand(i);
5529 if (V.getOpcode() == ISD::UNDEF)
5531 SDValue LaneIdx = DAG.getConstant(i, dl, MVT::i32);
5532 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Vec, V, LaneIdx);
5540 /// getExtFactor - Determine the adjustment factor for the position when
5541 /// generating an "extract from vector registers" instruction.
5542 static unsigned getExtFactor(SDValue &V) {
5543 EVT EltType = V.getValueType().getVectorElementType();
5544 return EltType.getSizeInBits() / 8;
5547 // Gather data to see if the operation can be modelled as a
5548 // shuffle in combination with VEXTs.
5549 SDValue ARMTargetLowering::ReconstructShuffle(SDValue Op,
5550 SelectionDAG &DAG) const {
5551 assert(Op.getOpcode() == ISD::BUILD_VECTOR && "Unknown opcode!");
5553 EVT VT = Op.getValueType();
5554 unsigned NumElts = VT.getVectorNumElements();
5556 struct ShuffleSourceInfo {
5561 // We may insert some combination of BITCASTs and VEXT nodes to force Vec to
5562 // be compatible with the shuffle we intend to construct. As a result
5563 // ShuffleVec will be some sliding window into the original Vec.
5566 // Code should guarantee that element i in Vec starts at element "WindowBase
5567 // + i * WindowScale in ShuffleVec".
5571 bool operator ==(SDValue OtherVec) { return Vec == OtherVec; }
5572 ShuffleSourceInfo(SDValue Vec)
5573 : Vec(Vec), MinElt(UINT_MAX), MaxElt(0), ShuffleVec(Vec), WindowBase(0),
5577 // First gather all vectors used as an immediate source for this BUILD_VECTOR
5579 SmallVector<ShuffleSourceInfo, 2> Sources;
5580 for (unsigned i = 0; i < NumElts; ++i) {
5581 SDValue V = Op.getOperand(i);
5582 if (V.getOpcode() == ISD::UNDEF)
5584 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
5585 // A shuffle can only come from building a vector from various
5586 // elements of other vectors.
5590 // Add this element source to the list if it's not already there.
5591 SDValue SourceVec = V.getOperand(0);
5592 auto Source = std::find(Sources.begin(), Sources.end(), SourceVec);
5593 if (Source == Sources.end())
5594 Source = Sources.insert(Sources.end(), ShuffleSourceInfo(SourceVec));
5596 // Update the minimum and maximum lane number seen.
5597 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
5598 Source->MinElt = std::min(Source->MinElt, EltNo);
5599 Source->MaxElt = std::max(Source->MaxElt, EltNo);
5602 // Currently only do something sane when at most two source vectors
5604 if (Sources.size() > 2)
5607 // Find out the smallest element size among result and two sources, and use
5608 // it as element size to build the shuffle_vector.
5609 EVT SmallestEltTy = VT.getVectorElementType();
5610 for (auto &Source : Sources) {
5611 EVT SrcEltTy = Source.Vec.getValueType().getVectorElementType();
5612 if (SrcEltTy.bitsLT(SmallestEltTy))
5613 SmallestEltTy = SrcEltTy;
5615 unsigned ResMultiplier =
5616 VT.getVectorElementType().getSizeInBits() / SmallestEltTy.getSizeInBits();
5617 NumElts = VT.getSizeInBits() / SmallestEltTy.getSizeInBits();
5618 EVT ShuffleVT = EVT::getVectorVT(*DAG.getContext(), SmallestEltTy, NumElts);
5620 // If the source vector is too wide or too narrow, we may nevertheless be able
5621 // to construct a compatible shuffle either by concatenating it with UNDEF or
5622 // extracting a suitable range of elements.
5623 for (auto &Src : Sources) {
5624 EVT SrcVT = Src.ShuffleVec.getValueType();
5626 if (SrcVT.getSizeInBits() == VT.getSizeInBits())
5629 // This stage of the search produces a source with the same element type as
5630 // the original, but with a total width matching the BUILD_VECTOR output.
5631 EVT EltVT = SrcVT.getVectorElementType();
5632 unsigned NumSrcElts = VT.getSizeInBits() / EltVT.getSizeInBits();
5633 EVT DestVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumSrcElts);
5635 if (SrcVT.getSizeInBits() < VT.getSizeInBits()) {
5636 if (2 * SrcVT.getSizeInBits() != VT.getSizeInBits())
5638 // We can pad out the smaller vector for free, so if it's part of a
5641 DAG.getNode(ISD::CONCAT_VECTORS, dl, DestVT, Src.ShuffleVec,
5642 DAG.getUNDEF(Src.ShuffleVec.getValueType()));
5646 if (SrcVT.getSizeInBits() != 2 * VT.getSizeInBits())
5649 if (Src.MaxElt - Src.MinElt >= NumSrcElts) {
5650 // Span too large for a VEXT to cope
5654 if (Src.MinElt >= NumSrcElts) {
5655 // The extraction can just take the second half
5657 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec,
5658 DAG.getConstant(NumSrcElts, dl, MVT::i32));
5659 Src.WindowBase = -NumSrcElts;
5660 } else if (Src.MaxElt < NumSrcElts) {
5661 // The extraction can just take the first half
5663 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec,
5664 DAG.getConstant(0, dl, MVT::i32));
5666 // An actual VEXT is needed
5668 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec,
5669 DAG.getConstant(0, dl, MVT::i32));
5671 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec,
5672 DAG.getConstant(NumSrcElts, dl, MVT::i32));
5673 unsigned Imm = Src.MinElt * getExtFactor(VEXTSrc1);
5675 Src.ShuffleVec = DAG.getNode(ARMISD::VEXT, dl, DestVT, VEXTSrc1,
5677 DAG.getConstant(Imm, dl, MVT::i32));
5678 Src.WindowBase = -Src.MinElt;
5682 // Another possible incompatibility occurs from the vector element types. We
5683 // can fix this by bitcasting the source vectors to the same type we intend
5685 for (auto &Src : Sources) {
5686 EVT SrcEltTy = Src.ShuffleVec.getValueType().getVectorElementType();
5687 if (SrcEltTy == SmallestEltTy)
5689 assert(ShuffleVT.getVectorElementType() == SmallestEltTy);
5690 Src.ShuffleVec = DAG.getNode(ISD::BITCAST, dl, ShuffleVT, Src.ShuffleVec);
5691 Src.WindowScale = SrcEltTy.getSizeInBits() / SmallestEltTy.getSizeInBits();
5692 Src.WindowBase *= Src.WindowScale;
5695 // Final sanity check before we try to actually produce a shuffle.
5697 for (auto Src : Sources)
5698 assert(Src.ShuffleVec.getValueType() == ShuffleVT);
5701 // The stars all align, our next step is to produce the mask for the shuffle.
5702 SmallVector<int, 8> Mask(ShuffleVT.getVectorNumElements(), -1);
5703 int BitsPerShuffleLane = ShuffleVT.getVectorElementType().getSizeInBits();
5704 for (unsigned i = 0; i < VT.getVectorNumElements(); ++i) {
5705 SDValue Entry = Op.getOperand(i);
5706 if (Entry.getOpcode() == ISD::UNDEF)
5709 auto Src = std::find(Sources.begin(), Sources.end(), Entry.getOperand(0));
5710 int EltNo = cast<ConstantSDNode>(Entry.getOperand(1))->getSExtValue();
5712 // EXTRACT_VECTOR_ELT performs an implicit any_ext; BUILD_VECTOR an implicit
5713 // trunc. So only std::min(SrcBits, DestBits) actually get defined in this
5715 EVT OrigEltTy = Entry.getOperand(0).getValueType().getVectorElementType();
5716 int BitsDefined = std::min(OrigEltTy.getSizeInBits(),
5717 VT.getVectorElementType().getSizeInBits());
5718 int LanesDefined = BitsDefined / BitsPerShuffleLane;
5720 // This source is expected to fill ResMultiplier lanes of the final shuffle,
5721 // starting at the appropriate offset.
5722 int *LaneMask = &Mask[i * ResMultiplier];
5724 int ExtractBase = EltNo * Src->WindowScale + Src->WindowBase;
5725 ExtractBase += NumElts * (Src - Sources.begin());
5726 for (int j = 0; j < LanesDefined; ++j)
5727 LaneMask[j] = ExtractBase + j;
5730 // Final check before we try to produce nonsense...
5731 if (!isShuffleMaskLegal(Mask, ShuffleVT))
5734 // We can't handle more than two sources. This should have already
5735 // been checked before this point.
5736 assert(Sources.size() <= 2 && "Too many sources!");
5738 SDValue ShuffleOps[] = { DAG.getUNDEF(ShuffleVT), DAG.getUNDEF(ShuffleVT) };
5739 for (unsigned i = 0; i < Sources.size(); ++i)
5740 ShuffleOps[i] = Sources[i].ShuffleVec;
5742 SDValue Shuffle = DAG.getVectorShuffle(ShuffleVT, dl, ShuffleOps[0],
5743 ShuffleOps[1], &Mask[0]);
5744 return DAG.getNode(ISD::BITCAST, dl, VT, Shuffle);
5747 /// isShuffleMaskLegal - Targets can use this to indicate that they only
5748 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
5749 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
5750 /// are assumed to be legal.
5752 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
5754 if (VT.getVectorNumElements() == 4 &&
5755 (VT.is128BitVector() || VT.is64BitVector())) {
5756 unsigned PFIndexes[4];
5757 for (unsigned i = 0; i != 4; ++i) {
5761 PFIndexes[i] = M[i];
5764 // Compute the index in the perfect shuffle table.
5765 unsigned PFTableIndex =
5766 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
5767 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
5768 unsigned Cost = (PFEntry >> 30);
5774 bool ReverseVEXT, isV_UNDEF;
5775 unsigned Imm, WhichResult;
5777 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5778 return (EltSize >= 32 ||
5779 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
5780 isVREVMask(M, VT, 64) ||
5781 isVREVMask(M, VT, 32) ||
5782 isVREVMask(M, VT, 16) ||
5783 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
5784 isVTBLMask(M, VT) ||
5785 isNEONTwoResultShuffleMask(M, VT, WhichResult, isV_UNDEF) ||
5786 ((VT == MVT::v8i16 || VT == MVT::v16i8) && isReverseMask(M, VT)));
5789 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
5790 /// the specified operations to build the shuffle.
5791 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
5792 SDValue RHS, SelectionDAG &DAG,
5794 unsigned OpNum = (PFEntry >> 26) & 0x0F;
5795 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
5796 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
5799 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
5808 OP_VUZPL, // VUZP, left result
5809 OP_VUZPR, // VUZP, right result
5810 OP_VZIPL, // VZIP, left result
5811 OP_VZIPR, // VZIP, right result
5812 OP_VTRNL, // VTRN, left result
5813 OP_VTRNR // VTRN, right result
5816 if (OpNum == OP_COPY) {
5817 if (LHSID == (1*9+2)*9+3) return LHS;
5818 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
5822 SDValue OpLHS, OpRHS;
5823 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
5824 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
5825 EVT VT = OpLHS.getValueType();
5828 default: llvm_unreachable("Unknown shuffle opcode!");
5830 // VREV divides the vector in half and swaps within the half.
5831 if (VT.getVectorElementType() == MVT::i32 ||
5832 VT.getVectorElementType() == MVT::f32)
5833 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
5834 // vrev <4 x i16> -> VREV32
5835 if (VT.getVectorElementType() == MVT::i16)
5836 return DAG.getNode(ARMISD::VREV32, dl, VT, OpLHS);
5837 // vrev <4 x i8> -> VREV16
5838 assert(VT.getVectorElementType() == MVT::i8);
5839 return DAG.getNode(ARMISD::VREV16, dl, VT, OpLHS);
5844 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
5845 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, dl, MVT::i32));
5849 return DAG.getNode(ARMISD::VEXT, dl, VT,
5851 DAG.getConstant(OpNum - OP_VEXT1 + 1, dl, MVT::i32));
5854 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
5855 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
5858 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
5859 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
5862 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
5863 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
5867 static SDValue LowerVECTOR_SHUFFLEv8i8(SDValue Op,
5868 ArrayRef<int> ShuffleMask,
5869 SelectionDAG &DAG) {
5870 // Check to see if we can use the VTBL instruction.
5871 SDValue V1 = Op.getOperand(0);
5872 SDValue V2 = Op.getOperand(1);
5875 SmallVector<SDValue, 8> VTBLMask;
5876 for (ArrayRef<int>::iterator
5877 I = ShuffleMask.begin(), E = ShuffleMask.end(); I != E; ++I)
5878 VTBLMask.push_back(DAG.getConstant(*I, DL, MVT::i32));
5880 if (V2.getNode()->getOpcode() == ISD::UNDEF)
5881 return DAG.getNode(ARMISD::VTBL1, DL, MVT::v8i8, V1,
5882 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8, VTBLMask));
5884 return DAG.getNode(ARMISD::VTBL2, DL, MVT::v8i8, V1, V2,
5885 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8, VTBLMask));
5888 static SDValue LowerReverse_VECTOR_SHUFFLEv16i8_v8i16(SDValue Op,
5889 SelectionDAG &DAG) {
5891 SDValue OpLHS = Op.getOperand(0);
5892 EVT VT = OpLHS.getValueType();
5894 assert((VT == MVT::v8i16 || VT == MVT::v16i8) &&
5895 "Expect an v8i16/v16i8 type");
5896 OpLHS = DAG.getNode(ARMISD::VREV64, DL, VT, OpLHS);
5897 // For a v16i8 type: After the VREV, we have got <8, ...15, 8, ..., 0>. Now,
5898 // extract the first 8 bytes into the top double word and the last 8 bytes
5899 // into the bottom double word. The v8i16 case is similar.
5900 unsigned ExtractNum = (VT == MVT::v16i8) ? 8 : 4;
5901 return DAG.getNode(ARMISD::VEXT, DL, VT, OpLHS, OpLHS,
5902 DAG.getConstant(ExtractNum, DL, MVT::i32));
5905 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
5906 SDValue V1 = Op.getOperand(0);
5907 SDValue V2 = Op.getOperand(1);
5909 EVT VT = Op.getValueType();
5910 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
5912 // Convert shuffles that are directly supported on NEON to target-specific
5913 // DAG nodes, instead of keeping them as shuffles and matching them again
5914 // during code selection. This is more efficient and avoids the possibility
5915 // of inconsistencies between legalization and selection.
5916 // FIXME: floating-point vectors should be canonicalized to integer vectors
5917 // of the same time so that they get CSEd properly.
5918 ArrayRef<int> ShuffleMask = SVN->getMask();
5920 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5921 if (EltSize <= 32) {
5922 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
5923 int Lane = SVN->getSplatIndex();
5924 // If this is undef splat, generate it via "just" vdup, if possible.
5925 if (Lane == -1) Lane = 0;
5927 // Test if V1 is a SCALAR_TO_VECTOR.
5928 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
5929 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
5931 // Test if V1 is a BUILD_VECTOR which is equivalent to a SCALAR_TO_VECTOR
5932 // (and probably will turn into a SCALAR_TO_VECTOR once legalization
5934 if (Lane == 0 && V1.getOpcode() == ISD::BUILD_VECTOR &&
5935 !isa<ConstantSDNode>(V1.getOperand(0))) {
5936 bool IsScalarToVector = true;
5937 for (unsigned i = 1, e = V1.getNumOperands(); i != e; ++i)
5938 if (V1.getOperand(i).getOpcode() != ISD::UNDEF) {
5939 IsScalarToVector = false;
5942 if (IsScalarToVector)
5943 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
5945 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
5946 DAG.getConstant(Lane, dl, MVT::i32));
5951 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
5954 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
5955 DAG.getConstant(Imm, dl, MVT::i32));
5958 if (isVREVMask(ShuffleMask, VT, 64))
5959 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
5960 if (isVREVMask(ShuffleMask, VT, 32))
5961 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
5962 if (isVREVMask(ShuffleMask, VT, 16))
5963 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
5965 if (V2->getOpcode() == ISD::UNDEF &&
5966 isSingletonVEXTMask(ShuffleMask, VT, Imm)) {
5967 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V1,
5968 DAG.getConstant(Imm, dl, MVT::i32));
5971 // Check for Neon shuffles that modify both input vectors in place.
5972 // If both results are used, i.e., if there are two shuffles with the same
5973 // source operands and with masks corresponding to both results of one of
5974 // these operations, DAG memoization will ensure that a single node is
5975 // used for both shuffles.
5976 unsigned WhichResult;
5978 if (unsigned ShuffleOpc = isNEONTwoResultShuffleMask(
5979 ShuffleMask, VT, WhichResult, isV_UNDEF)) {
5982 return DAG.getNode(ShuffleOpc, dl, DAG.getVTList(VT, VT), V1, V2)
5983 .getValue(WhichResult);
5986 // Also check for these shuffles through CONCAT_VECTORS: we canonicalize
5987 // shuffles that produce a result larger than their operands with:
5988 // shuffle(concat(v1, undef), concat(v2, undef))
5990 // shuffle(concat(v1, v2), undef)
5991 // because we can access quad vectors (see PerformVECTOR_SHUFFLECombine).
5993 // This is useful in the general case, but there are special cases where
5994 // native shuffles produce larger results: the two-result ops.
5996 // Look through the concat when lowering them:
5997 // shuffle(concat(v1, v2), undef)
5999 // concat(VZIP(v1, v2):0, :1)
6001 if (V1->getOpcode() == ISD::CONCAT_VECTORS &&
6002 V2->getOpcode() == ISD::UNDEF) {
6003 SDValue SubV1 = V1->getOperand(0);
6004 SDValue SubV2 = V1->getOperand(1);
6005 EVT SubVT = SubV1.getValueType();
6007 // We expect these to have been canonicalized to -1.
6008 assert(std::all_of(ShuffleMask.begin(), ShuffleMask.end(), [&](int i) {
6009 return i < (int)VT.getVectorNumElements();
6010 }) && "Unexpected shuffle index into UNDEF operand!");
6012 if (unsigned ShuffleOpc = isNEONTwoResultShuffleMask(
6013 ShuffleMask, SubVT, WhichResult, isV_UNDEF)) {
6016 assert((WhichResult == 0) &&
6017 "In-place shuffle of concat can only have one result!");
6018 SDValue Res = DAG.getNode(ShuffleOpc, dl, DAG.getVTList(SubVT, SubVT),
6020 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, Res.getValue(0),
6026 // If the shuffle is not directly supported and it has 4 elements, use
6027 // the PerfectShuffle-generated table to synthesize it from other shuffles.
6028 unsigned NumElts = VT.getVectorNumElements();
6030 unsigned PFIndexes[4];
6031 for (unsigned i = 0; i != 4; ++i) {
6032 if (ShuffleMask[i] < 0)
6035 PFIndexes[i] = ShuffleMask[i];
6038 // Compute the index in the perfect shuffle table.
6039 unsigned PFTableIndex =
6040 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
6041 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
6042 unsigned Cost = (PFEntry >> 30);
6045 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
6048 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
6049 if (EltSize >= 32) {
6050 // Do the expansion with floating-point types, since that is what the VFP
6051 // registers are defined to use, and since i64 is not legal.
6052 EVT EltVT = EVT::getFloatingPointVT(EltSize);
6053 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
6054 V1 = DAG.getNode(ISD::BITCAST, dl, VecVT, V1);
6055 V2 = DAG.getNode(ISD::BITCAST, dl, VecVT, V2);
6056 SmallVector<SDValue, 8> Ops;
6057 for (unsigned i = 0; i < NumElts; ++i) {
6058 if (ShuffleMask[i] < 0)
6059 Ops.push_back(DAG.getUNDEF(EltVT));
6061 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
6062 ShuffleMask[i] < (int)NumElts ? V1 : V2,
6063 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
6066 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, Ops);
6067 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
6070 if ((VT == MVT::v8i16 || VT == MVT::v16i8) && isReverseMask(ShuffleMask, VT))
6071 return LowerReverse_VECTOR_SHUFFLEv16i8_v8i16(Op, DAG);
6073 if (VT == MVT::v8i8) {
6074 SDValue NewOp = LowerVECTOR_SHUFFLEv8i8(Op, ShuffleMask, DAG);
6075 if (NewOp.getNode())
6082 static SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
6083 // INSERT_VECTOR_ELT is legal only for immediate indexes.
6084 SDValue Lane = Op.getOperand(2);
6085 if (!isa<ConstantSDNode>(Lane))
6091 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
6092 // EXTRACT_VECTOR_ELT is legal only for immediate indexes.
6093 SDValue Lane = Op.getOperand(1);
6094 if (!isa<ConstantSDNode>(Lane))
6097 SDValue Vec = Op.getOperand(0);
6098 if (Op.getValueType() == MVT::i32 &&
6099 Vec.getValueType().getVectorElementType().getSizeInBits() < 32) {
6101 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
6107 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
6108 // The only time a CONCAT_VECTORS operation can have legal types is when
6109 // two 64-bit vectors are concatenated to a 128-bit vector.
6110 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
6111 "unexpected CONCAT_VECTORS");
6113 SDValue Val = DAG.getUNDEF(MVT::v2f64);
6114 SDValue Op0 = Op.getOperand(0);
6115 SDValue Op1 = Op.getOperand(1);
6116 if (Op0.getOpcode() != ISD::UNDEF)
6117 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
6118 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op0),
6119 DAG.getIntPtrConstant(0, dl));
6120 if (Op1.getOpcode() != ISD::UNDEF)
6121 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
6122 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op1),
6123 DAG.getIntPtrConstant(1, dl));
6124 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Val);
6127 /// isExtendedBUILD_VECTOR - Check if N is a constant BUILD_VECTOR where each
6128 /// element has been zero/sign-extended, depending on the isSigned parameter,
6129 /// from an integer type half its size.
6130 static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG,
6132 // A v2i64 BUILD_VECTOR will have been legalized to a BITCAST from v4i32.
6133 EVT VT = N->getValueType(0);
6134 if (VT == MVT::v2i64 && N->getOpcode() == ISD::BITCAST) {
6135 SDNode *BVN = N->getOperand(0).getNode();
6136 if (BVN->getValueType(0) != MVT::v4i32 ||
6137 BVN->getOpcode() != ISD::BUILD_VECTOR)
6139 unsigned LoElt = DAG.getDataLayout().isBigEndian() ? 1 : 0;
6140 unsigned HiElt = 1 - LoElt;
6141 ConstantSDNode *Lo0 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt));
6142 ConstantSDNode *Hi0 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt));
6143 ConstantSDNode *Lo1 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt+2));
6144 ConstantSDNode *Hi1 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt+2));
6145 if (!Lo0 || !Hi0 || !Lo1 || !Hi1)
6148 if (Hi0->getSExtValue() == Lo0->getSExtValue() >> 32 &&
6149 Hi1->getSExtValue() == Lo1->getSExtValue() >> 32)
6152 if (Hi0->isNullValue() && Hi1->isNullValue())
6158 if (N->getOpcode() != ISD::BUILD_VECTOR)
6161 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
6162 SDNode *Elt = N->getOperand(i).getNode();
6163 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
6164 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
6165 unsigned HalfSize = EltSize / 2;
6167 if (!isIntN(HalfSize, C->getSExtValue()))
6170 if (!isUIntN(HalfSize, C->getZExtValue()))
6181 /// isSignExtended - Check if a node is a vector value that is sign-extended
6182 /// or a constant BUILD_VECTOR with sign-extended elements.
6183 static bool isSignExtended(SDNode *N, SelectionDAG &DAG) {
6184 if (N->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N))
6186 if (isExtendedBUILD_VECTOR(N, DAG, true))
6191 /// isZeroExtended - Check if a node is a vector value that is zero-extended
6192 /// or a constant BUILD_VECTOR with zero-extended elements.
6193 static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) {
6194 if (N->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N))
6196 if (isExtendedBUILD_VECTOR(N, DAG, false))
6201 static EVT getExtensionTo64Bits(const EVT &OrigVT) {
6202 if (OrigVT.getSizeInBits() >= 64)
6205 assert(OrigVT.isSimple() && "Expecting a simple value type");
6207 MVT::SimpleValueType OrigSimpleTy = OrigVT.getSimpleVT().SimpleTy;
6208 switch (OrigSimpleTy) {
6209 default: llvm_unreachable("Unexpected Vector Type");
6218 /// AddRequiredExtensionForVMULL - Add a sign/zero extension to extend the total
6219 /// value size to 64 bits. We need a 64-bit D register as an operand to VMULL.
6220 /// We insert the required extension here to get the vector to fill a D register.
6221 static SDValue AddRequiredExtensionForVMULL(SDValue N, SelectionDAG &DAG,
6224 unsigned ExtOpcode) {
6225 // The vector originally had a size of OrigTy. It was then extended to ExtTy.
6226 // We expect the ExtTy to be 128-bits total. If the OrigTy is less than
6227 // 64-bits we need to insert a new extension so that it will be 64-bits.
6228 assert(ExtTy.is128BitVector() && "Unexpected extension size");
6229 if (OrigTy.getSizeInBits() >= 64)
6232 // Must extend size to at least 64 bits to be used as an operand for VMULL.
6233 EVT NewVT = getExtensionTo64Bits(OrigTy);
6235 return DAG.getNode(ExtOpcode, SDLoc(N), NewVT, N);
6238 /// SkipLoadExtensionForVMULL - return a load of the original vector size that
6239 /// does not do any sign/zero extension. If the original vector is less
6240 /// than 64 bits, an appropriate extension will be added after the load to
6241 /// reach a total size of 64 bits. We have to add the extension separately
6242 /// because ARM does not have a sign/zero extending load for vectors.
6243 static SDValue SkipLoadExtensionForVMULL(LoadSDNode *LD, SelectionDAG& DAG) {
6244 EVT ExtendedTy = getExtensionTo64Bits(LD->getMemoryVT());
6246 // The load already has the right type.
6247 if (ExtendedTy == LD->getMemoryVT())
6248 return DAG.getLoad(LD->getMemoryVT(), SDLoc(LD), LD->getChain(),
6249 LD->getBasePtr(), LD->getPointerInfo(), LD->isVolatile(),
6250 LD->isNonTemporal(), LD->isInvariant(),
6251 LD->getAlignment());
6253 // We need to create a zextload/sextload. We cannot just create a load
6254 // followed by a zext/zext node because LowerMUL is also run during normal
6255 // operation legalization where we can't create illegal types.
6256 return DAG.getExtLoad(LD->getExtensionType(), SDLoc(LD), ExtendedTy,
6257 LD->getChain(), LD->getBasePtr(), LD->getPointerInfo(),
6258 LD->getMemoryVT(), LD->isVolatile(), LD->isInvariant(),
6259 LD->isNonTemporal(), LD->getAlignment());
6262 /// SkipExtensionForVMULL - For a node that is a SIGN_EXTEND, ZERO_EXTEND,
6263 /// extending load, or BUILD_VECTOR with extended elements, return the
6264 /// unextended value. The unextended vector should be 64 bits so that it can
6265 /// be used as an operand to a VMULL instruction. If the original vector size
6266 /// before extension is less than 64 bits we add a an extension to resize
6267 /// the vector to 64 bits.
6268 static SDValue SkipExtensionForVMULL(SDNode *N, SelectionDAG &DAG) {
6269 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
6270 return AddRequiredExtensionForVMULL(N->getOperand(0), DAG,
6271 N->getOperand(0)->getValueType(0),
6275 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
6276 return SkipLoadExtensionForVMULL(LD, DAG);
6278 // Otherwise, the value must be a BUILD_VECTOR. For v2i64, it will
6279 // have been legalized as a BITCAST from v4i32.
6280 if (N->getOpcode() == ISD::BITCAST) {
6281 SDNode *BVN = N->getOperand(0).getNode();
6282 assert(BVN->getOpcode() == ISD::BUILD_VECTOR &&
6283 BVN->getValueType(0) == MVT::v4i32 && "expected v4i32 BUILD_VECTOR");
6284 unsigned LowElt = DAG.getDataLayout().isBigEndian() ? 1 : 0;
6285 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), MVT::v2i32,
6286 BVN->getOperand(LowElt), BVN->getOperand(LowElt+2));
6288 // Construct a new BUILD_VECTOR with elements truncated to half the size.
6289 assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR");
6290 EVT VT = N->getValueType(0);
6291 unsigned EltSize = VT.getVectorElementType().getSizeInBits() / 2;
6292 unsigned NumElts = VT.getVectorNumElements();
6293 MVT TruncVT = MVT::getIntegerVT(EltSize);
6294 SmallVector<SDValue, 8> Ops;
6296 for (unsigned i = 0; i != NumElts; ++i) {
6297 ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i));
6298 const APInt &CInt = C->getAPIntValue();
6299 // Element types smaller than 32 bits are not legal, so use i32 elements.
6300 // The values are implicitly truncated so sext vs. zext doesn't matter.
6301 Ops.push_back(DAG.getConstant(CInt.zextOrTrunc(32), dl, MVT::i32));
6303 return DAG.getNode(ISD::BUILD_VECTOR, dl,
6304 MVT::getVectorVT(TruncVT, NumElts), Ops);
6307 static bool isAddSubSExt(SDNode *N, SelectionDAG &DAG) {
6308 unsigned Opcode = N->getOpcode();
6309 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
6310 SDNode *N0 = N->getOperand(0).getNode();
6311 SDNode *N1 = N->getOperand(1).getNode();
6312 return N0->hasOneUse() && N1->hasOneUse() &&
6313 isSignExtended(N0, DAG) && isSignExtended(N1, DAG);
6318 static bool isAddSubZExt(SDNode *N, SelectionDAG &DAG) {
6319 unsigned Opcode = N->getOpcode();
6320 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
6321 SDNode *N0 = N->getOperand(0).getNode();
6322 SDNode *N1 = N->getOperand(1).getNode();
6323 return N0->hasOneUse() && N1->hasOneUse() &&
6324 isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG);
6329 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
6330 // Multiplications are only custom-lowered for 128-bit vectors so that
6331 // VMULL can be detected. Otherwise v2i64 multiplications are not legal.
6332 EVT VT = Op.getValueType();
6333 assert(VT.is128BitVector() && VT.isInteger() &&
6334 "unexpected type for custom-lowering ISD::MUL");
6335 SDNode *N0 = Op.getOperand(0).getNode();
6336 SDNode *N1 = Op.getOperand(1).getNode();
6337 unsigned NewOpc = 0;
6339 bool isN0SExt = isSignExtended(N0, DAG);
6340 bool isN1SExt = isSignExtended(N1, DAG);
6341 if (isN0SExt && isN1SExt)
6342 NewOpc = ARMISD::VMULLs;
6344 bool isN0ZExt = isZeroExtended(N0, DAG);
6345 bool isN1ZExt = isZeroExtended(N1, DAG);
6346 if (isN0ZExt && isN1ZExt)
6347 NewOpc = ARMISD::VMULLu;
6348 else if (isN1SExt || isN1ZExt) {
6349 // Look for (s/zext A + s/zext B) * (s/zext C). We want to turn these
6350 // into (s/zext A * s/zext C) + (s/zext B * s/zext C)
6351 if (isN1SExt && isAddSubSExt(N0, DAG)) {
6352 NewOpc = ARMISD::VMULLs;
6354 } else if (isN1ZExt && isAddSubZExt(N0, DAG)) {
6355 NewOpc = ARMISD::VMULLu;
6357 } else if (isN0ZExt && isAddSubZExt(N1, DAG)) {
6359 NewOpc = ARMISD::VMULLu;
6365 if (VT == MVT::v2i64)
6366 // Fall through to expand this. It is not legal.
6369 // Other vector multiplications are legal.
6374 // Legalize to a VMULL instruction.
6377 SDValue Op1 = SkipExtensionForVMULL(N1, DAG);
6379 Op0 = SkipExtensionForVMULL(N0, DAG);
6380 assert(Op0.getValueType().is64BitVector() &&
6381 Op1.getValueType().is64BitVector() &&
6382 "unexpected types for extended operands to VMULL");
6383 return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
6386 // Optimizing (zext A + zext B) * C, to (VMULL A, C) + (VMULL B, C) during
6387 // isel lowering to take advantage of no-stall back to back vmul + vmla.
6394 SDValue N00 = SkipExtensionForVMULL(N0->getOperand(0).getNode(), DAG);
6395 SDValue N01 = SkipExtensionForVMULL(N0->getOperand(1).getNode(), DAG);
6396 EVT Op1VT = Op1.getValueType();
6397 return DAG.getNode(N0->getOpcode(), DL, VT,
6398 DAG.getNode(NewOpc, DL, VT,
6399 DAG.getNode(ISD::BITCAST, DL, Op1VT, N00), Op1),
6400 DAG.getNode(NewOpc, DL, VT,
6401 DAG.getNode(ISD::BITCAST, DL, Op1VT, N01), Op1));
6405 LowerSDIV_v4i8(SDValue X, SDValue Y, SDLoc dl, SelectionDAG &DAG) {
6407 // float4 xf = vcvt_f32_s32(vmovl_s16(a.lo));
6408 // float4 yf = vcvt_f32_s32(vmovl_s16(b.lo));
6409 X = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, X);
6410 Y = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, Y);
6411 X = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, X);
6412 Y = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, Y);
6413 // Get reciprocal estimate.
6414 // float4 recip = vrecpeq_f32(yf);
6415 Y = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6416 DAG.getConstant(Intrinsic::arm_neon_vrecpe, dl, MVT::i32),
6418 // Because char has a smaller range than uchar, we can actually get away
6419 // without any newton steps. This requires that we use a weird bias
6420 // of 0xb000, however (again, this has been exhaustively tested).
6421 // float4 result = as_float4(as_int4(xf*recip) + 0xb000);
6422 X = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, X, Y);
6423 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, X);
6424 Y = DAG.getConstant(0xb000, dl, MVT::i32);
6425 Y = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Y, Y, Y, Y);
6426 X = DAG.getNode(ISD::ADD, dl, MVT::v4i32, X, Y);
6427 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, X);
6428 // Convert back to short.
6429 X = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, X);
6430 X = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, X);
6435 LowerSDIV_v4i16(SDValue N0, SDValue N1, SDLoc dl, SelectionDAG &DAG) {
6437 // Convert to float.
6438 // float4 yf = vcvt_f32_s32(vmovl_s16(y));
6439 // float4 xf = vcvt_f32_s32(vmovl_s16(x));
6440 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N0);
6441 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N1);
6442 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
6443 N1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
6445 // Use reciprocal estimate and one refinement step.
6446 // float4 recip = vrecpeq_f32(yf);
6447 // recip *= vrecpsq_f32(yf, recip);
6448 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6449 DAG.getConstant(Intrinsic::arm_neon_vrecpe, dl, MVT::i32),
6451 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6452 DAG.getConstant(Intrinsic::arm_neon_vrecps, dl, MVT::i32),
6454 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
6455 // Because short has a smaller range than ushort, we can actually get away
6456 // with only a single newton step. This requires that we use a weird bias
6457 // of 89, however (again, this has been exhaustively tested).
6458 // float4 result = as_float4(as_int4(xf*recip) + 0x89);
6459 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
6460 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
6461 N1 = DAG.getConstant(0x89, dl, MVT::i32);
6462 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
6463 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
6464 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
6465 // Convert back to integer and return.
6466 // return vmovn_s32(vcvt_s32_f32(result));
6467 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
6468 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
6472 static SDValue LowerSDIV(SDValue Op, SelectionDAG &DAG) {
6473 EVT VT = Op.getValueType();
6474 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
6475 "unexpected type for custom-lowering ISD::SDIV");
6478 SDValue N0 = Op.getOperand(0);
6479 SDValue N1 = Op.getOperand(1);
6482 if (VT == MVT::v8i8) {
6483 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N0);
6484 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N1);
6486 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
6487 DAG.getIntPtrConstant(4, dl));
6488 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
6489 DAG.getIntPtrConstant(4, dl));
6490 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
6491 DAG.getIntPtrConstant(0, dl));
6492 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
6493 DAG.getIntPtrConstant(0, dl));
6495 N0 = LowerSDIV_v4i8(N0, N1, dl, DAG); // v4i16
6496 N2 = LowerSDIV_v4i8(N2, N3, dl, DAG); // v4i16
6498 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
6499 N0 = LowerCONCAT_VECTORS(N0, DAG);
6501 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v8i8, N0);
6504 return LowerSDIV_v4i16(N0, N1, dl, DAG);
6507 static SDValue LowerUDIV(SDValue Op, SelectionDAG &DAG) {
6508 EVT VT = Op.getValueType();
6509 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
6510 "unexpected type for custom-lowering ISD::UDIV");
6513 SDValue N0 = Op.getOperand(0);
6514 SDValue N1 = Op.getOperand(1);
6517 if (VT == MVT::v8i8) {
6518 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N0);
6519 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N1);
6521 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
6522 DAG.getIntPtrConstant(4, dl));
6523 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
6524 DAG.getIntPtrConstant(4, dl));
6525 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
6526 DAG.getIntPtrConstant(0, dl));
6527 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
6528 DAG.getIntPtrConstant(0, dl));
6530 N0 = LowerSDIV_v4i16(N0, N1, dl, DAG); // v4i16
6531 N2 = LowerSDIV_v4i16(N2, N3, dl, DAG); // v4i16
6533 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
6534 N0 = LowerCONCAT_VECTORS(N0, DAG);
6536 N0 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v8i8,
6537 DAG.getConstant(Intrinsic::arm_neon_vqmovnsu, dl,
6543 // v4i16 sdiv ... Convert to float.
6544 // float4 yf = vcvt_f32_s32(vmovl_u16(y));
6545 // float4 xf = vcvt_f32_s32(vmovl_u16(x));
6546 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N0);
6547 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N1);
6548 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
6549 SDValue BN1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
6551 // Use reciprocal estimate and two refinement steps.
6552 // float4 recip = vrecpeq_f32(yf);
6553 // recip *= vrecpsq_f32(yf, recip);
6554 // recip *= vrecpsq_f32(yf, recip);
6555 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6556 DAG.getConstant(Intrinsic::arm_neon_vrecpe, dl, MVT::i32),
6558 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6559 DAG.getConstant(Intrinsic::arm_neon_vrecps, dl, MVT::i32),
6561 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
6562 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6563 DAG.getConstant(Intrinsic::arm_neon_vrecps, dl, MVT::i32),
6565 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
6566 // Simply multiplying by the reciprocal estimate can leave us a few ulps
6567 // too low, so we add 2 ulps (exhaustive testing shows that this is enough,
6568 // and that it will never cause us to return an answer too large).
6569 // float4 result = as_float4(as_int4(xf*recip) + 2);
6570 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
6571 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
6572 N1 = DAG.getConstant(2, dl, MVT::i32);
6573 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
6574 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
6575 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
6576 // Convert back to integer and return.
6577 // return vmovn_u32(vcvt_s32_f32(result));
6578 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
6579 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
6583 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
6584 EVT VT = Op.getNode()->getValueType(0);
6585 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
6588 bool ExtraOp = false;
6589 switch (Op.getOpcode()) {
6590 default: llvm_unreachable("Invalid code");
6591 case ISD::ADDC: Opc = ARMISD::ADDC; break;
6592 case ISD::ADDE: Opc = ARMISD::ADDE; ExtraOp = true; break;
6593 case ISD::SUBC: Opc = ARMISD::SUBC; break;
6594 case ISD::SUBE: Opc = ARMISD::SUBE; ExtraOp = true; break;
6598 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
6600 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
6601 Op.getOperand(1), Op.getOperand(2));
6604 SDValue ARMTargetLowering::LowerFSINCOS(SDValue Op, SelectionDAG &DAG) const {
6605 assert(Subtarget->isTargetDarwin());
6607 // For iOS, we want to call an alternative entry point: __sincos_stret,
6608 // return values are passed via sret.
6610 SDValue Arg = Op.getOperand(0);
6611 EVT ArgVT = Arg.getValueType();
6612 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
6613 auto PtrVT = getPointerTy(DAG.getDataLayout());
6615 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
6617 // Pair of floats / doubles used to pass the result.
6618 StructType *RetTy = StructType::get(ArgTy, ArgTy, nullptr);
6620 // Create stack object for sret.
6621 auto &DL = DAG.getDataLayout();
6622 const uint64_t ByteSize = DL.getTypeAllocSize(RetTy);
6623 const unsigned StackAlign = DL.getPrefTypeAlignment(RetTy);
6624 int FrameIdx = FrameInfo->CreateStackObject(ByteSize, StackAlign, false);
6625 SDValue SRet = DAG.getFrameIndex(FrameIdx, getPointerTy(DL));
6631 Entry.Ty = RetTy->getPointerTo();
6632 Entry.isSExt = false;
6633 Entry.isZExt = false;
6634 Entry.isSRet = true;
6635 Args.push_back(Entry);
6639 Entry.isSExt = false;
6640 Entry.isZExt = false;
6641 Args.push_back(Entry);
6643 const char *LibcallName = (ArgVT == MVT::f64)
6644 ? "__sincos_stret" : "__sincosf_stret";
6645 SDValue Callee = DAG.getExternalSymbol(LibcallName, getPointerTy(DL));
6647 TargetLowering::CallLoweringInfo CLI(DAG);
6648 CLI.setDebugLoc(dl).setChain(DAG.getEntryNode())
6649 .setCallee(CallingConv::C, Type::getVoidTy(*DAG.getContext()), Callee,
6651 .setDiscardResult();
6653 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
6655 SDValue LoadSin = DAG.getLoad(ArgVT, dl, CallResult.second, SRet,
6656 MachinePointerInfo(), false, false, false, 0);
6658 // Address of cos field.
6659 SDValue Add = DAG.getNode(ISD::ADD, dl, PtrVT, SRet,
6660 DAG.getIntPtrConstant(ArgVT.getStoreSize(), dl));
6661 SDValue LoadCos = DAG.getLoad(ArgVT, dl, LoadSin.getValue(1), Add,
6662 MachinePointerInfo(), false, false, false, 0);
6664 SDVTList Tys = DAG.getVTList(ArgVT, ArgVT);
6665 return DAG.getNode(ISD::MERGE_VALUES, dl, Tys,
6666 LoadSin.getValue(0), LoadCos.getValue(0));
6669 static SDValue LowerAtomicLoadStore(SDValue Op, SelectionDAG &DAG) {
6670 // Monotonic load/store is legal for all targets
6671 if (cast<AtomicSDNode>(Op)->getOrdering() <= Monotonic)
6674 // Acquire/Release load/store is not legal for targets without a
6675 // dmb or equivalent available.
6679 static void ReplaceREADCYCLECOUNTER(SDNode *N,
6680 SmallVectorImpl<SDValue> &Results,
6682 const ARMSubtarget *Subtarget) {
6684 SDValue Cycles32, OutChain;
6686 if (Subtarget->hasPerfMon()) {
6687 // Under Power Management extensions, the cycle-count is:
6688 // mrc p15, #0, <Rt>, c9, c13, #0
6689 SDValue Ops[] = { N->getOperand(0), // Chain
6690 DAG.getConstant(Intrinsic::arm_mrc, DL, MVT::i32),
6691 DAG.getConstant(15, DL, MVT::i32),
6692 DAG.getConstant(0, DL, MVT::i32),
6693 DAG.getConstant(9, DL, MVT::i32),
6694 DAG.getConstant(13, DL, MVT::i32),
6695 DAG.getConstant(0, DL, MVT::i32)
6698 Cycles32 = DAG.getNode(ISD::INTRINSIC_W_CHAIN, DL,
6699 DAG.getVTList(MVT::i32, MVT::Other), Ops);
6700 OutChain = Cycles32.getValue(1);
6702 // Intrinsic is defined to return 0 on unsupported platforms. Technically
6703 // there are older ARM CPUs that have implementation-specific ways of
6704 // obtaining this information (FIXME!).
6705 Cycles32 = DAG.getConstant(0, DL, MVT::i32);
6706 OutChain = DAG.getEntryNode();
6710 SDValue Cycles64 = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64,
6711 Cycles32, DAG.getConstant(0, DL, MVT::i32));
6712 Results.push_back(Cycles64);
6713 Results.push_back(OutChain);
6716 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
6717 switch (Op.getOpcode()) {
6718 default: llvm_unreachable("Don't know how to custom lower this!");
6719 case ISD::WRITE_REGISTER: return LowerWRITE_REGISTER(Op, DAG);
6720 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
6721 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
6722 case ISD::GlobalAddress:
6723 switch (Subtarget->getTargetTriple().getObjectFormat()) {
6724 default: llvm_unreachable("unknown object format");
6726 return LowerGlobalAddressWindows(Op, DAG);
6728 return LowerGlobalAddressELF(Op, DAG);
6730 return LowerGlobalAddressDarwin(Op, DAG);
6732 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
6733 case ISD::SELECT: return LowerSELECT(Op, DAG);
6734 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
6735 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
6736 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
6737 case ISD::VASTART: return LowerVASTART(Op, DAG);
6738 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, DAG, Subtarget);
6739 case ISD::PREFETCH: return LowerPREFETCH(Op, DAG, Subtarget);
6740 case ISD::SINT_TO_FP:
6741 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
6742 case ISD::FP_TO_SINT:
6743 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
6744 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
6745 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
6746 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
6747 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
6748 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
6749 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
6750 case ISD::EH_SJLJ_SETUP_DISPATCH: return LowerEH_SJLJ_SETUP_DISPATCH(Op, DAG);
6751 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
6753 case ISD::BITCAST: return ExpandBITCAST(Op.getNode(), DAG);
6756 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
6757 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
6758 case ISD::SRL_PARTS:
6759 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
6761 case ISD::CTTZ_ZERO_UNDEF: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
6762 case ISD::CTPOP: return LowerCTPOP(Op.getNode(), DAG, Subtarget);
6763 case ISD::SETCC: return LowerVSETCC(Op, DAG);
6764 case ISD::ConstantFP: return LowerConstantFP(Op, DAG, Subtarget);
6765 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG, Subtarget);
6766 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
6767 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
6768 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
6769 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
6770 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
6771 case ISD::MUL: return LowerMUL(Op, DAG);
6772 case ISD::SDIV: return LowerSDIV(Op, DAG);
6773 case ISD::UDIV: return LowerUDIV(Op, DAG);
6777 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
6782 return LowerXALUO(Op, DAG);
6783 case ISD::ATOMIC_LOAD:
6784 case ISD::ATOMIC_STORE: return LowerAtomicLoadStore(Op, DAG);
6785 case ISD::FSINCOS: return LowerFSINCOS(Op, DAG);
6787 case ISD::UDIVREM: return LowerDivRem(Op, DAG);
6788 case ISD::DYNAMIC_STACKALLOC:
6789 if (Subtarget->getTargetTriple().isWindowsItaniumEnvironment())
6790 return LowerDYNAMIC_STACKALLOC(Op, DAG);
6791 llvm_unreachable("Don't know how to custom lower this!");
6792 case ISD::FP_ROUND: return LowerFP_ROUND(Op, DAG);
6793 case ISD::FP_EXTEND: return LowerFP_EXTEND(Op, DAG);
6797 /// ReplaceNodeResults - Replace the results of node with an illegal result
6798 /// type with new values built out of custom code.
6799 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
6800 SmallVectorImpl<SDValue>&Results,
6801 SelectionDAG &DAG) const {
6803 switch (N->getOpcode()) {
6805 llvm_unreachable("Don't know how to custom expand this!");
6806 case ISD::READ_REGISTER:
6807 ExpandREAD_REGISTER(N, Results, DAG);
6810 Res = ExpandBITCAST(N, DAG);
6814 Res = Expand64BitShift(N, DAG, Subtarget);
6816 case ISD::READCYCLECOUNTER:
6817 ReplaceREADCYCLECOUNTER(N, Results, DAG, Subtarget);
6821 Results.push_back(Res);
6824 //===----------------------------------------------------------------------===//
6825 // ARM Scheduler Hooks
6826 //===----------------------------------------------------------------------===//
6828 /// SetupEntryBlockForSjLj - Insert code into the entry block that creates and
6829 /// registers the function context.
6830 void ARMTargetLowering::
6831 SetupEntryBlockForSjLj(MachineInstr *MI, MachineBasicBlock *MBB,
6832 MachineBasicBlock *DispatchBB, int FI) const {
6833 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
6834 DebugLoc dl = MI->getDebugLoc();
6835 MachineFunction *MF = MBB->getParent();
6836 MachineRegisterInfo *MRI = &MF->getRegInfo();
6837 MachineConstantPool *MCP = MF->getConstantPool();
6838 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
6839 const Function *F = MF->getFunction();
6841 bool isThumb = Subtarget->isThumb();
6842 bool isThumb2 = Subtarget->isThumb2();
6844 unsigned PCLabelId = AFI->createPICLabelUId();
6845 unsigned PCAdj = (isThumb || isThumb2) ? 4 : 8;
6846 ARMConstantPoolValue *CPV =
6847 ARMConstantPoolMBB::Create(F->getContext(), DispatchBB, PCLabelId, PCAdj);
6848 unsigned CPI = MCP->getConstantPoolIndex(CPV, 4);
6850 const TargetRegisterClass *TRC = isThumb ? &ARM::tGPRRegClass
6851 : &ARM::GPRRegClass;
6853 // Grab constant pool and fixed stack memory operands.
6854 MachineMemOperand *CPMMO =
6855 MF->getMachineMemOperand(MachinePointerInfo::getConstantPool(),
6856 MachineMemOperand::MOLoad, 4, 4);
6858 MachineMemOperand *FIMMOSt =
6859 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
6860 MachineMemOperand::MOStore, 4, 4);
6862 // Load the address of the dispatch MBB into the jump buffer.
6864 // Incoming value: jbuf
6865 // ldr.n r5, LCPI1_1
6868 // str r5, [$jbuf, #+4] ; &jbuf[1]
6869 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6870 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2LDRpci), NewVReg1)
6871 .addConstantPoolIndex(CPI)
6872 .addMemOperand(CPMMO));
6873 // Set the low bit because of thumb mode.
6874 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6876 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2ORRri), NewVReg2)
6877 .addReg(NewVReg1, RegState::Kill)
6879 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6880 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg3)
6881 .addReg(NewVReg2, RegState::Kill)
6883 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2STRi12))
6884 .addReg(NewVReg3, RegState::Kill)
6886 .addImm(36) // &jbuf[1] :: pc
6887 .addMemOperand(FIMMOSt));
6888 } else if (isThumb) {
6889 // Incoming value: jbuf
6890 // ldr.n r1, LCPI1_4
6894 // add r2, $jbuf, #+4 ; &jbuf[1]
6896 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6897 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tLDRpci), NewVReg1)
6898 .addConstantPoolIndex(CPI)
6899 .addMemOperand(CPMMO));
6900 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6901 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg2)
6902 .addReg(NewVReg1, RegState::Kill)
6904 // Set the low bit because of thumb mode.
6905 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6906 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tMOVi8), NewVReg3)
6907 .addReg(ARM::CPSR, RegState::Define)
6909 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6910 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tORR), NewVReg4)
6911 .addReg(ARM::CPSR, RegState::Define)
6912 .addReg(NewVReg2, RegState::Kill)
6913 .addReg(NewVReg3, RegState::Kill));
6914 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6915 BuildMI(*MBB, MI, dl, TII->get(ARM::tADDframe), NewVReg5)
6917 .addImm(36); // &jbuf[1] :: pc
6918 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tSTRi))
6919 .addReg(NewVReg4, RegState::Kill)
6920 .addReg(NewVReg5, RegState::Kill)
6922 .addMemOperand(FIMMOSt));
6924 // Incoming value: jbuf
6927 // str r1, [$jbuf, #+4] ; &jbuf[1]
6928 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6929 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::LDRi12), NewVReg1)
6930 .addConstantPoolIndex(CPI)
6932 .addMemOperand(CPMMO));
6933 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6934 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::PICADD), NewVReg2)
6935 .addReg(NewVReg1, RegState::Kill)
6936 .addImm(PCLabelId));
6937 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::STRi12))
6938 .addReg(NewVReg2, RegState::Kill)
6940 .addImm(36) // &jbuf[1] :: pc
6941 .addMemOperand(FIMMOSt));
6945 void ARMTargetLowering::EmitSjLjDispatchBlock(MachineInstr *MI,
6946 MachineBasicBlock *MBB) const {
6947 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
6948 DebugLoc dl = MI->getDebugLoc();
6949 MachineFunction *MF = MBB->getParent();
6950 MachineRegisterInfo *MRI = &MF->getRegInfo();
6951 MachineFrameInfo *MFI = MF->getFrameInfo();
6952 int FI = MFI->getFunctionContextIndex();
6954 const TargetRegisterClass *TRC = Subtarget->isThumb() ? &ARM::tGPRRegClass
6955 : &ARM::GPRnopcRegClass;
6957 // Get a mapping of the call site numbers to all of the landing pads they're
6959 DenseMap<unsigned, SmallVector<MachineBasicBlock*, 2> > CallSiteNumToLPad;
6960 unsigned MaxCSNum = 0;
6961 MachineModuleInfo &MMI = MF->getMMI();
6962 for (MachineFunction::iterator BB = MF->begin(), E = MF->end(); BB != E;
6964 if (!BB->isLandingPad()) continue;
6966 // FIXME: We should assert that the EH_LABEL is the first MI in the landing
6968 for (MachineBasicBlock::iterator
6969 II = BB->begin(), IE = BB->end(); II != IE; ++II) {
6970 if (!II->isEHLabel()) continue;
6972 MCSymbol *Sym = II->getOperand(0).getMCSymbol();
6973 if (!MMI.hasCallSiteLandingPad(Sym)) continue;
6975 SmallVectorImpl<unsigned> &CallSiteIdxs = MMI.getCallSiteLandingPad(Sym);
6976 for (SmallVectorImpl<unsigned>::iterator
6977 CSI = CallSiteIdxs.begin(), CSE = CallSiteIdxs.end();
6978 CSI != CSE; ++CSI) {
6979 CallSiteNumToLPad[*CSI].push_back(BB);
6980 MaxCSNum = std::max(MaxCSNum, *CSI);
6986 // Get an ordered list of the machine basic blocks for the jump table.
6987 std::vector<MachineBasicBlock*> LPadList;
6988 SmallPtrSet<MachineBasicBlock*, 64> InvokeBBs;
6989 LPadList.reserve(CallSiteNumToLPad.size());
6990 for (unsigned I = 1; I <= MaxCSNum; ++I) {
6991 SmallVectorImpl<MachineBasicBlock*> &MBBList = CallSiteNumToLPad[I];
6992 for (SmallVectorImpl<MachineBasicBlock*>::iterator
6993 II = MBBList.begin(), IE = MBBList.end(); II != IE; ++II) {
6994 LPadList.push_back(*II);
6995 InvokeBBs.insert((*II)->pred_begin(), (*II)->pred_end());
6999 assert(!LPadList.empty() &&
7000 "No landing pad destinations for the dispatch jump table!");
7002 // Create the jump table and associated information.
7003 MachineJumpTableInfo *JTI =
7004 MF->getOrCreateJumpTableInfo(MachineJumpTableInfo::EK_Inline);
7005 unsigned MJTI = JTI->createJumpTableIndex(LPadList);
7006 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
7008 // Create the MBBs for the dispatch code.
7010 // Shove the dispatch's address into the return slot in the function context.
7011 MachineBasicBlock *DispatchBB = MF->CreateMachineBasicBlock();
7012 DispatchBB->setIsLandingPad();
7014 MachineBasicBlock *TrapBB = MF->CreateMachineBasicBlock();
7015 unsigned trap_opcode;
7016 if (Subtarget->isThumb())
7017 trap_opcode = ARM::tTRAP;
7019 trap_opcode = Subtarget->useNaClTrap() ? ARM::TRAPNaCl : ARM::TRAP;
7021 BuildMI(TrapBB, dl, TII->get(trap_opcode));
7022 DispatchBB->addSuccessor(TrapBB);
7024 MachineBasicBlock *DispContBB = MF->CreateMachineBasicBlock();
7025 DispatchBB->addSuccessor(DispContBB);
7028 MF->insert(MF->end(), DispatchBB);
7029 MF->insert(MF->end(), DispContBB);
7030 MF->insert(MF->end(), TrapBB);
7032 // Insert code into the entry block that creates and registers the function
7034 SetupEntryBlockForSjLj(MI, MBB, DispatchBB, FI);
7036 MachineMemOperand *FIMMOLd =
7037 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
7038 MachineMemOperand::MOLoad |
7039 MachineMemOperand::MOVolatile, 4, 4);
7041 MachineInstrBuilder MIB;
7042 MIB = BuildMI(DispatchBB, dl, TII->get(ARM::Int_eh_sjlj_dispatchsetup));
7044 const ARMBaseInstrInfo *AII = static_cast<const ARMBaseInstrInfo*>(TII);
7045 const ARMBaseRegisterInfo &RI = AII->getRegisterInfo();
7047 // Add a register mask with no preserved registers. This results in all
7048 // registers being marked as clobbered.
7049 MIB.addRegMask(RI.getNoPreservedMask());
7051 unsigned NumLPads = LPadList.size();
7052 if (Subtarget->isThumb2()) {
7053 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
7054 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2LDRi12), NewVReg1)
7057 .addMemOperand(FIMMOLd));
7059 if (NumLPads < 256) {
7060 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPri))
7062 .addImm(LPadList.size()));
7064 unsigned VReg1 = MRI->createVirtualRegister(TRC);
7065 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVi16), VReg1)
7066 .addImm(NumLPads & 0xFFFF));
7068 unsigned VReg2 = VReg1;
7069 if ((NumLPads & 0xFFFF0000) != 0) {
7070 VReg2 = MRI->createVirtualRegister(TRC);
7071 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVTi16), VReg2)
7073 .addImm(NumLPads >> 16));
7076 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPrr))
7081 BuildMI(DispatchBB, dl, TII->get(ARM::t2Bcc))
7086 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
7087 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::t2LEApcrelJT),NewVReg3)
7088 .addJumpTableIndex(MJTI));
7090 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
7093 BuildMI(DispContBB, dl, TII->get(ARM::t2ADDrs), NewVReg4)
7094 .addReg(NewVReg3, RegState::Kill)
7096 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
7098 BuildMI(DispContBB, dl, TII->get(ARM::t2BR_JT))
7099 .addReg(NewVReg4, RegState::Kill)
7101 .addJumpTableIndex(MJTI);
7102 } else if (Subtarget->isThumb()) {
7103 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
7104 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRspi), NewVReg1)
7107 .addMemOperand(FIMMOLd));
7109 if (NumLPads < 256) {
7110 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPi8))
7114 MachineConstantPool *ConstantPool = MF->getConstantPool();
7115 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
7116 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
7118 // MachineConstantPool wants an explicit alignment.
7119 unsigned Align = MF->getDataLayout().getPrefTypeAlignment(Int32Ty);
7121 Align = MF->getDataLayout().getTypeAllocSize(C->getType());
7122 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
7124 unsigned VReg1 = MRI->createVirtualRegister(TRC);
7125 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRpci))
7126 .addReg(VReg1, RegState::Define)
7127 .addConstantPoolIndex(Idx));
7128 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPr))
7133 BuildMI(DispatchBB, dl, TII->get(ARM::tBcc))
7138 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
7139 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLSLri), NewVReg2)
7140 .addReg(ARM::CPSR, RegState::Define)
7144 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
7145 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLEApcrelJT), NewVReg3)
7146 .addJumpTableIndex(MJTI));
7148 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
7149 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg4)
7150 .addReg(ARM::CPSR, RegState::Define)
7151 .addReg(NewVReg2, RegState::Kill)
7154 MachineMemOperand *JTMMOLd =
7155 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
7156 MachineMemOperand::MOLoad, 4, 4);
7158 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
7159 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLDRi), NewVReg5)
7160 .addReg(NewVReg4, RegState::Kill)
7162 .addMemOperand(JTMMOLd));
7164 unsigned NewVReg6 = NewVReg5;
7165 if (RelocM == Reloc::PIC_) {
7166 NewVReg6 = MRI->createVirtualRegister(TRC);
7167 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg6)
7168 .addReg(ARM::CPSR, RegState::Define)
7169 .addReg(NewVReg5, RegState::Kill)
7173 BuildMI(DispContBB, dl, TII->get(ARM::tBR_JTr))
7174 .addReg(NewVReg6, RegState::Kill)
7175 .addJumpTableIndex(MJTI);
7177 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
7178 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRi12), NewVReg1)
7181 .addMemOperand(FIMMOLd));
7183 if (NumLPads < 256) {
7184 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPri))
7187 } else if (Subtarget->hasV6T2Ops() && isUInt<16>(NumLPads)) {
7188 unsigned VReg1 = MRI->createVirtualRegister(TRC);
7189 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVi16), VReg1)
7190 .addImm(NumLPads & 0xFFFF));
7192 unsigned VReg2 = VReg1;
7193 if ((NumLPads & 0xFFFF0000) != 0) {
7194 VReg2 = MRI->createVirtualRegister(TRC);
7195 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVTi16), VReg2)
7197 .addImm(NumLPads >> 16));
7200 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
7204 MachineConstantPool *ConstantPool = MF->getConstantPool();
7205 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
7206 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
7208 // MachineConstantPool wants an explicit alignment.
7209 unsigned Align = MF->getDataLayout().getPrefTypeAlignment(Int32Ty);
7211 Align = MF->getDataLayout().getTypeAllocSize(C->getType());
7212 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
7214 unsigned VReg1 = MRI->createVirtualRegister(TRC);
7215 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRcp))
7216 .addReg(VReg1, RegState::Define)
7217 .addConstantPoolIndex(Idx)
7219 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
7221 .addReg(VReg1, RegState::Kill));
7224 BuildMI(DispatchBB, dl, TII->get(ARM::Bcc))
7229 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
7231 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::MOVsi), NewVReg3)
7233 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
7234 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
7235 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::LEApcrelJT), NewVReg4)
7236 .addJumpTableIndex(MJTI));
7238 MachineMemOperand *JTMMOLd =
7239 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
7240 MachineMemOperand::MOLoad, 4, 4);
7241 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
7243 BuildMI(DispContBB, dl, TII->get(ARM::LDRrs), NewVReg5)
7244 .addReg(NewVReg3, RegState::Kill)
7247 .addMemOperand(JTMMOLd));
7249 if (RelocM == Reloc::PIC_) {
7250 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTadd))
7251 .addReg(NewVReg5, RegState::Kill)
7253 .addJumpTableIndex(MJTI);
7255 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTr))
7256 .addReg(NewVReg5, RegState::Kill)
7257 .addJumpTableIndex(MJTI);
7261 // Add the jump table entries as successors to the MBB.
7262 SmallPtrSet<MachineBasicBlock*, 8> SeenMBBs;
7263 for (std::vector<MachineBasicBlock*>::iterator
7264 I = LPadList.begin(), E = LPadList.end(); I != E; ++I) {
7265 MachineBasicBlock *CurMBB = *I;
7266 if (SeenMBBs.insert(CurMBB).second)
7267 DispContBB->addSuccessor(CurMBB);
7270 // N.B. the order the invoke BBs are processed in doesn't matter here.
7271 const MCPhysReg *SavedRegs = RI.getCalleeSavedRegs(MF);
7272 SmallVector<MachineBasicBlock*, 64> MBBLPads;
7273 for (MachineBasicBlock *BB : InvokeBBs) {
7275 // Remove the landing pad successor from the invoke block and replace it
7276 // with the new dispatch block.
7277 SmallVector<MachineBasicBlock*, 4> Successors(BB->succ_begin(),
7279 while (!Successors.empty()) {
7280 MachineBasicBlock *SMBB = Successors.pop_back_val();
7281 if (SMBB->isLandingPad()) {
7282 BB->removeSuccessor(SMBB);
7283 MBBLPads.push_back(SMBB);
7287 BB->addSuccessor(DispatchBB);
7289 // Find the invoke call and mark all of the callee-saved registers as
7290 // 'implicit defined' so that they're spilled. This prevents code from
7291 // moving instructions to before the EH block, where they will never be
7293 for (MachineBasicBlock::reverse_iterator
7294 II = BB->rbegin(), IE = BB->rend(); II != IE; ++II) {
7295 if (!II->isCall()) continue;
7297 DenseMap<unsigned, bool> DefRegs;
7298 for (MachineInstr::mop_iterator
7299 OI = II->operands_begin(), OE = II->operands_end();
7301 if (!OI->isReg()) continue;
7302 DefRegs[OI->getReg()] = true;
7305 MachineInstrBuilder MIB(*MF, &*II);
7307 for (unsigned i = 0; SavedRegs[i] != 0; ++i) {
7308 unsigned Reg = SavedRegs[i];
7309 if (Subtarget->isThumb2() &&
7310 !ARM::tGPRRegClass.contains(Reg) &&
7311 !ARM::hGPRRegClass.contains(Reg))
7313 if (Subtarget->isThumb1Only() && !ARM::tGPRRegClass.contains(Reg))
7315 if (!Subtarget->isThumb() && !ARM::GPRRegClass.contains(Reg))
7318 MIB.addReg(Reg, RegState::ImplicitDefine | RegState::Dead);
7325 // Mark all former landing pads as non-landing pads. The dispatch is the only
7327 for (SmallVectorImpl<MachineBasicBlock*>::iterator
7328 I = MBBLPads.begin(), E = MBBLPads.end(); I != E; ++I)
7329 (*I)->setIsLandingPad(false);
7331 // The instruction is gone now.
7332 MI->eraseFromParent();
7336 MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) {
7337 for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
7338 E = MBB->succ_end(); I != E; ++I)
7341 llvm_unreachable("Expecting a BB with two successors!");
7344 /// Return the load opcode for a given load size. If load size >= 8,
7345 /// neon opcode will be returned.
7346 static unsigned getLdOpcode(unsigned LdSize, bool IsThumb1, bool IsThumb2) {
7348 return LdSize == 16 ? ARM::VLD1q32wb_fixed
7349 : LdSize == 8 ? ARM::VLD1d32wb_fixed : 0;
7351 return LdSize == 4 ? ARM::tLDRi
7352 : LdSize == 2 ? ARM::tLDRHi
7353 : LdSize == 1 ? ARM::tLDRBi : 0;
7355 return LdSize == 4 ? ARM::t2LDR_POST
7356 : LdSize == 2 ? ARM::t2LDRH_POST
7357 : LdSize == 1 ? ARM::t2LDRB_POST : 0;
7358 return LdSize == 4 ? ARM::LDR_POST_IMM
7359 : LdSize == 2 ? ARM::LDRH_POST
7360 : LdSize == 1 ? ARM::LDRB_POST_IMM : 0;
7363 /// Return the store opcode for a given store size. If store size >= 8,
7364 /// neon opcode will be returned.
7365 static unsigned getStOpcode(unsigned StSize, bool IsThumb1, bool IsThumb2) {
7367 return StSize == 16 ? ARM::VST1q32wb_fixed
7368 : StSize == 8 ? ARM::VST1d32wb_fixed : 0;
7370 return StSize == 4 ? ARM::tSTRi
7371 : StSize == 2 ? ARM::tSTRHi
7372 : StSize == 1 ? ARM::tSTRBi : 0;
7374 return StSize == 4 ? ARM::t2STR_POST
7375 : StSize == 2 ? ARM::t2STRH_POST
7376 : StSize == 1 ? ARM::t2STRB_POST : 0;
7377 return StSize == 4 ? ARM::STR_POST_IMM
7378 : StSize == 2 ? ARM::STRH_POST
7379 : StSize == 1 ? ARM::STRB_POST_IMM : 0;
7382 /// Emit a post-increment load operation with given size. The instructions
7383 /// will be added to BB at Pos.
7384 static void emitPostLd(MachineBasicBlock *BB, MachineInstr *Pos,
7385 const TargetInstrInfo *TII, DebugLoc dl,
7386 unsigned LdSize, unsigned Data, unsigned AddrIn,
7387 unsigned AddrOut, bool IsThumb1, bool IsThumb2) {
7388 unsigned LdOpc = getLdOpcode(LdSize, IsThumb1, IsThumb2);
7389 assert(LdOpc != 0 && "Should have a load opcode");
7391 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
7392 .addReg(AddrOut, RegState::Define).addReg(AddrIn)
7394 } else if (IsThumb1) {
7395 // load + update AddrIn
7396 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
7397 .addReg(AddrIn).addImm(0));
7398 MachineInstrBuilder MIB =
7399 BuildMI(*BB, Pos, dl, TII->get(ARM::tADDi8), AddrOut);
7400 MIB = AddDefaultT1CC(MIB);
7401 MIB.addReg(AddrIn).addImm(LdSize);
7402 AddDefaultPred(MIB);
7403 } else if (IsThumb2) {
7404 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
7405 .addReg(AddrOut, RegState::Define).addReg(AddrIn)
7408 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
7409 .addReg(AddrOut, RegState::Define).addReg(AddrIn)
7410 .addReg(0).addImm(LdSize));
7414 /// Emit a post-increment store operation with given size. The instructions
7415 /// will be added to BB at Pos.
7416 static void emitPostSt(MachineBasicBlock *BB, MachineInstr *Pos,
7417 const TargetInstrInfo *TII, DebugLoc dl,
7418 unsigned StSize, unsigned Data, unsigned AddrIn,
7419 unsigned AddrOut, bool IsThumb1, bool IsThumb2) {
7420 unsigned StOpc = getStOpcode(StSize, IsThumb1, IsThumb2);
7421 assert(StOpc != 0 && "Should have a store opcode");
7423 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc), AddrOut)
7424 .addReg(AddrIn).addImm(0).addReg(Data));
7425 } else if (IsThumb1) {
7426 // store + update AddrIn
7427 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc)).addReg(Data)
7428 .addReg(AddrIn).addImm(0));
7429 MachineInstrBuilder MIB =
7430 BuildMI(*BB, Pos, dl, TII->get(ARM::tADDi8), AddrOut);
7431 MIB = AddDefaultT1CC(MIB);
7432 MIB.addReg(AddrIn).addImm(StSize);
7433 AddDefaultPred(MIB);
7434 } else if (IsThumb2) {
7435 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc), AddrOut)
7436 .addReg(Data).addReg(AddrIn).addImm(StSize));
7438 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc), AddrOut)
7439 .addReg(Data).addReg(AddrIn).addReg(0)
7445 ARMTargetLowering::EmitStructByval(MachineInstr *MI,
7446 MachineBasicBlock *BB) const {
7447 // This pseudo instruction has 3 operands: dst, src, size
7448 // We expand it to a loop if size > Subtarget->getMaxInlineSizeThreshold().
7449 // Otherwise, we will generate unrolled scalar copies.
7450 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
7451 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7452 MachineFunction::iterator It = BB;
7455 unsigned dest = MI->getOperand(0).getReg();
7456 unsigned src = MI->getOperand(1).getReg();
7457 unsigned SizeVal = MI->getOperand(2).getImm();
7458 unsigned Align = MI->getOperand(3).getImm();
7459 DebugLoc dl = MI->getDebugLoc();
7461 MachineFunction *MF = BB->getParent();
7462 MachineRegisterInfo &MRI = MF->getRegInfo();
7463 unsigned UnitSize = 0;
7464 const TargetRegisterClass *TRC = nullptr;
7465 const TargetRegisterClass *VecTRC = nullptr;
7467 bool IsThumb1 = Subtarget->isThumb1Only();
7468 bool IsThumb2 = Subtarget->isThumb2();
7472 } else if (Align & 2) {
7475 // Check whether we can use NEON instructions.
7476 if (!MF->getFunction()->hasFnAttribute(Attribute::NoImplicitFloat) &&
7477 Subtarget->hasNEON()) {
7478 if ((Align % 16 == 0) && SizeVal >= 16)
7480 else if ((Align % 8 == 0) && SizeVal >= 8)
7483 // Can't use NEON instructions.
7488 // Select the correct opcode and register class for unit size load/store
7489 bool IsNeon = UnitSize >= 8;
7490 TRC = (IsThumb1 || IsThumb2) ? &ARM::tGPRRegClass : &ARM::GPRRegClass;
7492 VecTRC = UnitSize == 16 ? &ARM::DPairRegClass
7493 : UnitSize == 8 ? &ARM::DPRRegClass
7496 unsigned BytesLeft = SizeVal % UnitSize;
7497 unsigned LoopSize = SizeVal - BytesLeft;
7499 if (SizeVal <= Subtarget->getMaxInlineSizeThreshold()) {
7500 // Use LDR and STR to copy.
7501 // [scratch, srcOut] = LDR_POST(srcIn, UnitSize)
7502 // [destOut] = STR_POST(scratch, destIn, UnitSize)
7503 unsigned srcIn = src;
7504 unsigned destIn = dest;
7505 for (unsigned i = 0; i < LoopSize; i+=UnitSize) {
7506 unsigned srcOut = MRI.createVirtualRegister(TRC);
7507 unsigned destOut = MRI.createVirtualRegister(TRC);
7508 unsigned scratch = MRI.createVirtualRegister(IsNeon ? VecTRC : TRC);
7509 emitPostLd(BB, MI, TII, dl, UnitSize, scratch, srcIn, srcOut,
7510 IsThumb1, IsThumb2);
7511 emitPostSt(BB, MI, TII, dl, UnitSize, scratch, destIn, destOut,
7512 IsThumb1, IsThumb2);
7517 // Handle the leftover bytes with LDRB and STRB.
7518 // [scratch, srcOut] = LDRB_POST(srcIn, 1)
7519 // [destOut] = STRB_POST(scratch, destIn, 1)
7520 for (unsigned i = 0; i < BytesLeft; i++) {
7521 unsigned srcOut = MRI.createVirtualRegister(TRC);
7522 unsigned destOut = MRI.createVirtualRegister(TRC);
7523 unsigned scratch = MRI.createVirtualRegister(TRC);
7524 emitPostLd(BB, MI, TII, dl, 1, scratch, srcIn, srcOut,
7525 IsThumb1, IsThumb2);
7526 emitPostSt(BB, MI, TII, dl, 1, scratch, destIn, destOut,
7527 IsThumb1, IsThumb2);
7531 MI->eraseFromParent(); // The instruction is gone now.
7535 // Expand the pseudo op to a loop.
7538 // movw varEnd, # --> with thumb2
7540 // ldrcp varEnd, idx --> without thumb2
7541 // fallthrough --> loopMBB
7543 // PHI varPhi, varEnd, varLoop
7544 // PHI srcPhi, src, srcLoop
7545 // PHI destPhi, dst, destLoop
7546 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
7547 // [destLoop] = STR_POST(scratch, destPhi, UnitSize)
7548 // subs varLoop, varPhi, #UnitSize
7550 // fallthrough --> exitMBB
7552 // epilogue to handle left-over bytes
7553 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
7554 // [destOut] = STRB_POST(scratch, destLoop, 1)
7555 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
7556 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
7557 MF->insert(It, loopMBB);
7558 MF->insert(It, exitMBB);
7560 // Transfer the remainder of BB and its successor edges to exitMBB.
7561 exitMBB->splice(exitMBB->begin(), BB,
7562 std::next(MachineBasicBlock::iterator(MI)), BB->end());
7563 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
7565 // Load an immediate to varEnd.
7566 unsigned varEnd = MRI.createVirtualRegister(TRC);
7567 if (Subtarget->useMovt(*MF)) {
7568 unsigned Vtmp = varEnd;
7569 if ((LoopSize & 0xFFFF0000) != 0)
7570 Vtmp = MRI.createVirtualRegister(TRC);
7571 AddDefaultPred(BuildMI(BB, dl,
7572 TII->get(IsThumb2 ? ARM::t2MOVi16 : ARM::MOVi16),
7573 Vtmp).addImm(LoopSize & 0xFFFF));
7575 if ((LoopSize & 0xFFFF0000) != 0)
7576 AddDefaultPred(BuildMI(BB, dl,
7577 TII->get(IsThumb2 ? ARM::t2MOVTi16 : ARM::MOVTi16),
7580 .addImm(LoopSize >> 16));
7582 MachineConstantPool *ConstantPool = MF->getConstantPool();
7583 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
7584 const Constant *C = ConstantInt::get(Int32Ty, LoopSize);
7586 // MachineConstantPool wants an explicit alignment.
7587 unsigned Align = MF->getDataLayout().getPrefTypeAlignment(Int32Ty);
7589 Align = MF->getDataLayout().getTypeAllocSize(C->getType());
7590 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
7593 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(ARM::tLDRpci)).addReg(
7594 varEnd, RegState::Define).addConstantPoolIndex(Idx));
7596 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(ARM::LDRcp)).addReg(
7597 varEnd, RegState::Define).addConstantPoolIndex(Idx).addImm(0));
7599 BB->addSuccessor(loopMBB);
7601 // Generate the loop body:
7602 // varPhi = PHI(varLoop, varEnd)
7603 // srcPhi = PHI(srcLoop, src)
7604 // destPhi = PHI(destLoop, dst)
7605 MachineBasicBlock *entryBB = BB;
7607 unsigned varLoop = MRI.createVirtualRegister(TRC);
7608 unsigned varPhi = MRI.createVirtualRegister(TRC);
7609 unsigned srcLoop = MRI.createVirtualRegister(TRC);
7610 unsigned srcPhi = MRI.createVirtualRegister(TRC);
7611 unsigned destLoop = MRI.createVirtualRegister(TRC);
7612 unsigned destPhi = MRI.createVirtualRegister(TRC);
7614 BuildMI(*BB, BB->begin(), dl, TII->get(ARM::PHI), varPhi)
7615 .addReg(varLoop).addMBB(loopMBB)
7616 .addReg(varEnd).addMBB(entryBB);
7617 BuildMI(BB, dl, TII->get(ARM::PHI), srcPhi)
7618 .addReg(srcLoop).addMBB(loopMBB)
7619 .addReg(src).addMBB(entryBB);
7620 BuildMI(BB, dl, TII->get(ARM::PHI), destPhi)
7621 .addReg(destLoop).addMBB(loopMBB)
7622 .addReg(dest).addMBB(entryBB);
7624 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
7625 // [destLoop] = STR_POST(scratch, destPhi, UnitSiz)
7626 unsigned scratch = MRI.createVirtualRegister(IsNeon ? VecTRC : TRC);
7627 emitPostLd(BB, BB->end(), TII, dl, UnitSize, scratch, srcPhi, srcLoop,
7628 IsThumb1, IsThumb2);
7629 emitPostSt(BB, BB->end(), TII, dl, UnitSize, scratch, destPhi, destLoop,
7630 IsThumb1, IsThumb2);
7632 // Decrement loop variable by UnitSize.
7634 MachineInstrBuilder MIB =
7635 BuildMI(*BB, BB->end(), dl, TII->get(ARM::tSUBi8), varLoop);
7636 MIB = AddDefaultT1CC(MIB);
7637 MIB.addReg(varPhi).addImm(UnitSize);
7638 AddDefaultPred(MIB);
7640 MachineInstrBuilder MIB =
7641 BuildMI(*BB, BB->end(), dl,
7642 TII->get(IsThumb2 ? ARM::t2SUBri : ARM::SUBri), varLoop);
7643 AddDefaultCC(AddDefaultPred(MIB.addReg(varPhi).addImm(UnitSize)));
7644 MIB->getOperand(5).setReg(ARM::CPSR);
7645 MIB->getOperand(5).setIsDef(true);
7647 BuildMI(*BB, BB->end(), dl,
7648 TII->get(IsThumb1 ? ARM::tBcc : IsThumb2 ? ARM::t2Bcc : ARM::Bcc))
7649 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
7651 // loopMBB can loop back to loopMBB or fall through to exitMBB.
7652 BB->addSuccessor(loopMBB);
7653 BB->addSuccessor(exitMBB);
7655 // Add epilogue to handle BytesLeft.
7657 MachineInstr *StartOfExit = exitMBB->begin();
7659 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
7660 // [destOut] = STRB_POST(scratch, destLoop, 1)
7661 unsigned srcIn = srcLoop;
7662 unsigned destIn = destLoop;
7663 for (unsigned i = 0; i < BytesLeft; i++) {
7664 unsigned srcOut = MRI.createVirtualRegister(TRC);
7665 unsigned destOut = MRI.createVirtualRegister(TRC);
7666 unsigned scratch = MRI.createVirtualRegister(TRC);
7667 emitPostLd(BB, StartOfExit, TII, dl, 1, scratch, srcIn, srcOut,
7668 IsThumb1, IsThumb2);
7669 emitPostSt(BB, StartOfExit, TII, dl, 1, scratch, destIn, destOut,
7670 IsThumb1, IsThumb2);
7675 MI->eraseFromParent(); // The instruction is gone now.
7680 ARMTargetLowering::EmitLowered__chkstk(MachineInstr *MI,
7681 MachineBasicBlock *MBB) const {
7682 const TargetMachine &TM = getTargetMachine();
7683 const TargetInstrInfo &TII = *Subtarget->getInstrInfo();
7684 DebugLoc DL = MI->getDebugLoc();
7686 assert(Subtarget->isTargetWindows() &&
7687 "__chkstk is only supported on Windows");
7688 assert(Subtarget->isThumb2() && "Windows on ARM requires Thumb-2 mode");
7690 // __chkstk takes the number of words to allocate on the stack in R4, and
7691 // returns the stack adjustment in number of bytes in R4. This will not
7692 // clober any other registers (other than the obvious lr).
7694 // Although, technically, IP should be considered a register which may be
7695 // clobbered, the call itself will not touch it. Windows on ARM is a pure
7696 // thumb-2 environment, so there is no interworking required. As a result, we
7697 // do not expect a veneer to be emitted by the linker, clobbering IP.
7699 // Each module receives its own copy of __chkstk, so no import thunk is
7700 // required, again, ensuring that IP is not clobbered.
7702 // Finally, although some linkers may theoretically provide a trampoline for
7703 // out of range calls (which is quite common due to a 32M range limitation of
7704 // branches for Thumb), we can generate the long-call version via
7705 // -mcmodel=large, alleviating the need for the trampoline which may clobber
7708 switch (TM.getCodeModel()) {
7709 case CodeModel::Small:
7710 case CodeModel::Medium:
7711 case CodeModel::Default:
7712 case CodeModel::Kernel:
7713 BuildMI(*MBB, MI, DL, TII.get(ARM::tBL))
7714 .addImm((unsigned)ARMCC::AL).addReg(0)
7715 .addExternalSymbol("__chkstk")
7716 .addReg(ARM::R4, RegState::Implicit | RegState::Kill)
7717 .addReg(ARM::R4, RegState::Implicit | RegState::Define)
7718 .addReg(ARM::R12, RegState::Implicit | RegState::Define | RegState::Dead);
7720 case CodeModel::Large:
7721 case CodeModel::JITDefault: {
7722 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
7723 unsigned Reg = MRI.createVirtualRegister(&ARM::rGPRRegClass);
7725 BuildMI(*MBB, MI, DL, TII.get(ARM::t2MOVi32imm), Reg)
7726 .addExternalSymbol("__chkstk");
7727 BuildMI(*MBB, MI, DL, TII.get(ARM::tBLXr))
7728 .addImm((unsigned)ARMCC::AL).addReg(0)
7729 .addReg(Reg, RegState::Kill)
7730 .addReg(ARM::R4, RegState::Implicit | RegState::Kill)
7731 .addReg(ARM::R4, RegState::Implicit | RegState::Define)
7732 .addReg(ARM::R12, RegState::Implicit | RegState::Define | RegState::Dead);
7737 AddDefaultCC(AddDefaultPred(BuildMI(*MBB, MI, DL, TII.get(ARM::t2SUBrr),
7739 .addReg(ARM::SP).addReg(ARM::R4)));
7741 MI->eraseFromParent();
7746 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
7747 MachineBasicBlock *BB) const {
7748 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
7749 DebugLoc dl = MI->getDebugLoc();
7750 bool isThumb2 = Subtarget->isThumb2();
7751 switch (MI->getOpcode()) {
7754 llvm_unreachable("Unexpected instr type to insert");
7756 // The Thumb2 pre-indexed stores have the same MI operands, they just
7757 // define them differently in the .td files from the isel patterns, so
7758 // they need pseudos.
7759 case ARM::t2STR_preidx:
7760 MI->setDesc(TII->get(ARM::t2STR_PRE));
7762 case ARM::t2STRB_preidx:
7763 MI->setDesc(TII->get(ARM::t2STRB_PRE));
7765 case ARM::t2STRH_preidx:
7766 MI->setDesc(TII->get(ARM::t2STRH_PRE));
7769 case ARM::STRi_preidx:
7770 case ARM::STRBi_preidx: {
7771 unsigned NewOpc = MI->getOpcode() == ARM::STRi_preidx ?
7772 ARM::STR_PRE_IMM : ARM::STRB_PRE_IMM;
7773 // Decode the offset.
7774 unsigned Offset = MI->getOperand(4).getImm();
7775 bool isSub = ARM_AM::getAM2Op(Offset) == ARM_AM::sub;
7776 Offset = ARM_AM::getAM2Offset(Offset);
7780 MachineMemOperand *MMO = *MI->memoperands_begin();
7781 BuildMI(*BB, MI, dl, TII->get(NewOpc))
7782 .addOperand(MI->getOperand(0)) // Rn_wb
7783 .addOperand(MI->getOperand(1)) // Rt
7784 .addOperand(MI->getOperand(2)) // Rn
7785 .addImm(Offset) // offset (skip GPR==zero_reg)
7786 .addOperand(MI->getOperand(5)) // pred
7787 .addOperand(MI->getOperand(6))
7788 .addMemOperand(MMO);
7789 MI->eraseFromParent();
7792 case ARM::STRr_preidx:
7793 case ARM::STRBr_preidx:
7794 case ARM::STRH_preidx: {
7796 switch (MI->getOpcode()) {
7797 default: llvm_unreachable("unexpected opcode!");
7798 case ARM::STRr_preidx: NewOpc = ARM::STR_PRE_REG; break;
7799 case ARM::STRBr_preidx: NewOpc = ARM::STRB_PRE_REG; break;
7800 case ARM::STRH_preidx: NewOpc = ARM::STRH_PRE; break;
7802 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(NewOpc));
7803 for (unsigned i = 0; i < MI->getNumOperands(); ++i)
7804 MIB.addOperand(MI->getOperand(i));
7805 MI->eraseFromParent();
7809 case ARM::tMOVCCr_pseudo: {
7810 // To "insert" a SELECT_CC instruction, we actually have to insert the
7811 // diamond control-flow pattern. The incoming instruction knows the
7812 // destination vreg to set, the condition code register to branch on, the
7813 // true/false values to select between, and a branch opcode to use.
7814 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7815 MachineFunction::iterator It = BB;
7821 // cmpTY ccX, r1, r2
7823 // fallthrough --> copy0MBB
7824 MachineBasicBlock *thisMBB = BB;
7825 MachineFunction *F = BB->getParent();
7826 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
7827 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
7828 F->insert(It, copy0MBB);
7829 F->insert(It, sinkMBB);
7831 // Transfer the remainder of BB and its successor edges to sinkMBB.
7832 sinkMBB->splice(sinkMBB->begin(), BB,
7833 std::next(MachineBasicBlock::iterator(MI)), BB->end());
7834 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
7836 BB->addSuccessor(copy0MBB);
7837 BB->addSuccessor(sinkMBB);
7839 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
7840 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
7843 // %FalseValue = ...
7844 // # fallthrough to sinkMBB
7847 // Update machine-CFG edges
7848 BB->addSuccessor(sinkMBB);
7851 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
7854 BuildMI(*BB, BB->begin(), dl,
7855 TII->get(ARM::PHI), MI->getOperand(0).getReg())
7856 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
7857 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
7859 MI->eraseFromParent(); // The pseudo instruction is gone now.
7864 case ARM::BCCZi64: {
7865 // If there is an unconditional branch to the other successor, remove it.
7866 BB->erase(std::next(MachineBasicBlock::iterator(MI)), BB->end());
7868 // Compare both parts that make up the double comparison separately for
7870 bool RHSisZero = MI->getOpcode() == ARM::BCCZi64;
7872 unsigned LHS1 = MI->getOperand(1).getReg();
7873 unsigned LHS2 = MI->getOperand(2).getReg();
7875 AddDefaultPred(BuildMI(BB, dl,
7876 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7877 .addReg(LHS1).addImm(0));
7878 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7879 .addReg(LHS2).addImm(0)
7880 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
7882 unsigned RHS1 = MI->getOperand(3).getReg();
7883 unsigned RHS2 = MI->getOperand(4).getReg();
7884 AddDefaultPred(BuildMI(BB, dl,
7885 TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
7886 .addReg(LHS1).addReg(RHS1));
7887 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
7888 .addReg(LHS2).addReg(RHS2)
7889 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
7892 MachineBasicBlock *destMBB = MI->getOperand(RHSisZero ? 3 : 5).getMBB();
7893 MachineBasicBlock *exitMBB = OtherSucc(BB, destMBB);
7894 if (MI->getOperand(0).getImm() == ARMCC::NE)
7895 std::swap(destMBB, exitMBB);
7897 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
7898 .addMBB(destMBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
7900 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2B)).addMBB(exitMBB));
7902 BuildMI(BB, dl, TII->get(ARM::B)) .addMBB(exitMBB);
7904 MI->eraseFromParent(); // The pseudo instruction is gone now.
7908 case ARM::Int_eh_sjlj_setjmp:
7909 case ARM::Int_eh_sjlj_setjmp_nofp:
7910 case ARM::tInt_eh_sjlj_setjmp:
7911 case ARM::t2Int_eh_sjlj_setjmp:
7912 case ARM::t2Int_eh_sjlj_setjmp_nofp:
7915 case ARM::Int_eh_sjlj_setup_dispatch:
7916 EmitSjLjDispatchBlock(MI, BB);
7921 // To insert an ABS instruction, we have to insert the
7922 // diamond control-flow pattern. The incoming instruction knows the
7923 // source vreg to test against 0, the destination vreg to set,
7924 // the condition code register to branch on, the
7925 // true/false values to select between, and a branch opcode to use.
7930 // BCC (branch to SinkBB if V0 >= 0)
7931 // RSBBB: V3 = RSBri V2, 0 (compute ABS if V2 < 0)
7932 // SinkBB: V1 = PHI(V2, V3)
7933 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7934 MachineFunction::iterator BBI = BB;
7936 MachineFunction *Fn = BB->getParent();
7937 MachineBasicBlock *RSBBB = Fn->CreateMachineBasicBlock(LLVM_BB);
7938 MachineBasicBlock *SinkBB = Fn->CreateMachineBasicBlock(LLVM_BB);
7939 Fn->insert(BBI, RSBBB);
7940 Fn->insert(BBI, SinkBB);
7942 unsigned int ABSSrcReg = MI->getOperand(1).getReg();
7943 unsigned int ABSDstReg = MI->getOperand(0).getReg();
7944 bool ABSSrcKIll = MI->getOperand(1).isKill();
7945 bool isThumb2 = Subtarget->isThumb2();
7946 MachineRegisterInfo &MRI = Fn->getRegInfo();
7947 // In Thumb mode S must not be specified if source register is the SP or
7948 // PC and if destination register is the SP, so restrict register class
7949 unsigned NewRsbDstReg =
7950 MRI.createVirtualRegister(isThumb2 ? &ARM::rGPRRegClass : &ARM::GPRRegClass);
7952 // Transfer the remainder of BB and its successor edges to sinkMBB.
7953 SinkBB->splice(SinkBB->begin(), BB,
7954 std::next(MachineBasicBlock::iterator(MI)), BB->end());
7955 SinkBB->transferSuccessorsAndUpdatePHIs(BB);
7957 BB->addSuccessor(RSBBB);
7958 BB->addSuccessor(SinkBB);
7960 // fall through to SinkMBB
7961 RSBBB->addSuccessor(SinkBB);
7963 // insert a cmp at the end of BB
7964 AddDefaultPred(BuildMI(BB, dl,
7965 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7966 .addReg(ABSSrcReg).addImm(0));
7968 // insert a bcc with opposite CC to ARMCC::MI at the end of BB
7970 TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)).addMBB(SinkBB)
7971 .addImm(ARMCC::getOppositeCondition(ARMCC::MI)).addReg(ARM::CPSR);
7973 // insert rsbri in RSBBB
7974 // Note: BCC and rsbri will be converted into predicated rsbmi
7975 // by if-conversion pass
7976 BuildMI(*RSBBB, RSBBB->begin(), dl,
7977 TII->get(isThumb2 ? ARM::t2RSBri : ARM::RSBri), NewRsbDstReg)
7978 .addReg(ABSSrcReg, ABSSrcKIll ? RegState::Kill : 0)
7979 .addImm(0).addImm((unsigned)ARMCC::AL).addReg(0).addReg(0);
7981 // insert PHI in SinkBB,
7982 // reuse ABSDstReg to not change uses of ABS instruction
7983 BuildMI(*SinkBB, SinkBB->begin(), dl,
7984 TII->get(ARM::PHI), ABSDstReg)
7985 .addReg(NewRsbDstReg).addMBB(RSBBB)
7986 .addReg(ABSSrcReg).addMBB(BB);
7988 // remove ABS instruction
7989 MI->eraseFromParent();
7991 // return last added BB
7994 case ARM::COPY_STRUCT_BYVAL_I32:
7996 return EmitStructByval(MI, BB);
7997 case ARM::WIN__CHKSTK:
7998 return EmitLowered__chkstk(MI, BB);
8002 void ARMTargetLowering::AdjustInstrPostInstrSelection(MachineInstr *MI,
8003 SDNode *Node) const {
8004 const MCInstrDesc *MCID = &MI->getDesc();
8005 // Adjust potentially 's' setting instructions after isel, i.e. ADC, SBC, RSB,
8006 // RSC. Coming out of isel, they have an implicit CPSR def, but the optional
8007 // operand is still set to noreg. If needed, set the optional operand's
8008 // register to CPSR, and remove the redundant implicit def.
8010 // e.g. ADCS (..., CPSR<imp-def>) -> ADC (... opt:CPSR<def>).
8012 // Rename pseudo opcodes.
8013 unsigned NewOpc = convertAddSubFlagsOpcode(MI->getOpcode());
8015 const ARMBaseInstrInfo *TII = Subtarget->getInstrInfo();
8016 MCID = &TII->get(NewOpc);
8018 assert(MCID->getNumOperands() == MI->getDesc().getNumOperands() + 1 &&
8019 "converted opcode should be the same except for cc_out");
8023 // Add the optional cc_out operand
8024 MI->addOperand(MachineOperand::CreateReg(0, /*isDef=*/true));
8026 unsigned ccOutIdx = MCID->getNumOperands() - 1;
8028 // Any ARM instruction that sets the 's' bit should specify an optional
8029 // "cc_out" operand in the last operand position.
8030 if (!MI->hasOptionalDef() || !MCID->OpInfo[ccOutIdx].isOptionalDef()) {
8031 assert(!NewOpc && "Optional cc_out operand required");
8034 // Look for an implicit def of CPSR added by MachineInstr ctor. Remove it
8035 // since we already have an optional CPSR def.
8036 bool definesCPSR = false;
8037 bool deadCPSR = false;
8038 for (unsigned i = MCID->getNumOperands(), e = MI->getNumOperands();
8040 const MachineOperand &MO = MI->getOperand(i);
8041 if (MO.isReg() && MO.isDef() && MO.getReg() == ARM::CPSR) {
8045 MI->RemoveOperand(i);
8050 assert(!NewOpc && "Optional cc_out operand required");
8053 assert(deadCPSR == !Node->hasAnyUseOfValue(1) && "inconsistent dead flag");
8055 assert(!MI->getOperand(ccOutIdx).getReg() &&
8056 "expect uninitialized optional cc_out operand");
8060 // If this instruction was defined with an optional CPSR def and its dag node
8061 // had a live implicit CPSR def, then activate the optional CPSR def.
8062 MachineOperand &MO = MI->getOperand(ccOutIdx);
8063 MO.setReg(ARM::CPSR);
8067 //===----------------------------------------------------------------------===//
8068 // ARM Optimization Hooks
8069 //===----------------------------------------------------------------------===//
8071 // Helper function that checks if N is a null or all ones constant.
8072 static inline bool isZeroOrAllOnes(SDValue N, bool AllOnes) {
8073 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N);
8076 return AllOnes ? C->isAllOnesValue() : C->isNullValue();
8079 // Return true if N is conditionally 0 or all ones.
8080 // Detects these expressions where cc is an i1 value:
8082 // (select cc 0, y) [AllOnes=0]
8083 // (select cc y, 0) [AllOnes=0]
8084 // (zext cc) [AllOnes=0]
8085 // (sext cc) [AllOnes=0/1]
8086 // (select cc -1, y) [AllOnes=1]
8087 // (select cc y, -1) [AllOnes=1]
8089 // Invert is set when N is the null/all ones constant when CC is false.
8090 // OtherOp is set to the alternative value of N.
8091 static bool isConditionalZeroOrAllOnes(SDNode *N, bool AllOnes,
8092 SDValue &CC, bool &Invert,
8094 SelectionDAG &DAG) {
8095 switch (N->getOpcode()) {
8096 default: return false;
8098 CC = N->getOperand(0);
8099 SDValue N1 = N->getOperand(1);
8100 SDValue N2 = N->getOperand(2);
8101 if (isZeroOrAllOnes(N1, AllOnes)) {
8106 if (isZeroOrAllOnes(N2, AllOnes)) {
8113 case ISD::ZERO_EXTEND:
8114 // (zext cc) can never be the all ones value.
8118 case ISD::SIGN_EXTEND: {
8120 EVT VT = N->getValueType(0);
8121 CC = N->getOperand(0);
8122 if (CC.getValueType() != MVT::i1)
8126 // When looking for an AllOnes constant, N is an sext, and the 'other'
8128 OtherOp = DAG.getConstant(0, dl, VT);
8129 else if (N->getOpcode() == ISD::ZERO_EXTEND)
8130 // When looking for a 0 constant, N can be zext or sext.
8131 OtherOp = DAG.getConstant(1, dl, VT);
8133 OtherOp = DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), dl,
8140 // Combine a constant select operand into its use:
8142 // (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
8143 // (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
8144 // (and (select cc, -1, c), x) -> (select cc, x, (and, x, c)) [AllOnes=1]
8145 // (or (select cc, 0, c), x) -> (select cc, x, (or, x, c))
8146 // (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c))
8148 // The transform is rejected if the select doesn't have a constant operand that
8149 // is null, or all ones when AllOnes is set.
8151 // Also recognize sext/zext from i1:
8153 // (add (zext cc), x) -> (select cc (add x, 1), x)
8154 // (add (sext cc), x) -> (select cc (add x, -1), x)
8156 // These transformations eventually create predicated instructions.
8158 // @param N The node to transform.
8159 // @param Slct The N operand that is a select.
8160 // @param OtherOp The other N operand (x above).
8161 // @param DCI Context.
8162 // @param AllOnes Require the select constant to be all ones instead of null.
8163 // @returns The new node, or SDValue() on failure.
8165 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
8166 TargetLowering::DAGCombinerInfo &DCI,
8167 bool AllOnes = false) {
8168 SelectionDAG &DAG = DCI.DAG;
8169 EVT VT = N->getValueType(0);
8170 SDValue NonConstantVal;
8173 if (!isConditionalZeroOrAllOnes(Slct.getNode(), AllOnes, CCOp, SwapSelectOps,
8174 NonConstantVal, DAG))
8177 // Slct is now know to be the desired identity constant when CC is true.
8178 SDValue TrueVal = OtherOp;
8179 SDValue FalseVal = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
8180 OtherOp, NonConstantVal);
8181 // Unless SwapSelectOps says CC should be false.
8183 std::swap(TrueVal, FalseVal);
8185 return DAG.getNode(ISD::SELECT, SDLoc(N), VT,
8186 CCOp, TrueVal, FalseVal);
8189 // Attempt combineSelectAndUse on each operand of a commutative operator N.
8191 SDValue combineSelectAndUseCommutative(SDNode *N, bool AllOnes,
8192 TargetLowering::DAGCombinerInfo &DCI) {
8193 SDValue N0 = N->getOperand(0);
8194 SDValue N1 = N->getOperand(1);
8195 if (N0.getNode()->hasOneUse()) {
8196 SDValue Result = combineSelectAndUse(N, N0, N1, DCI, AllOnes);
8197 if (Result.getNode())
8200 if (N1.getNode()->hasOneUse()) {
8201 SDValue Result = combineSelectAndUse(N, N1, N0, DCI, AllOnes);
8202 if (Result.getNode())
8208 // AddCombineToVPADDL- For pair-wise add on neon, use the vpaddl instruction
8209 // (only after legalization).
8210 static SDValue AddCombineToVPADDL(SDNode *N, SDValue N0, SDValue N1,
8211 TargetLowering::DAGCombinerInfo &DCI,
8212 const ARMSubtarget *Subtarget) {
8214 // Only perform optimization if after legalize, and if NEON is available. We
8215 // also expected both operands to be BUILD_VECTORs.
8216 if (DCI.isBeforeLegalize() || !Subtarget->hasNEON()
8217 || N0.getOpcode() != ISD::BUILD_VECTOR
8218 || N1.getOpcode() != ISD::BUILD_VECTOR)
8221 // Check output type since VPADDL operand elements can only be 8, 16, or 32.
8222 EVT VT = N->getValueType(0);
8223 if (!VT.isInteger() || VT.getVectorElementType() == MVT::i64)
8226 // Check that the vector operands are of the right form.
8227 // N0 and N1 are BUILD_VECTOR nodes with N number of EXTRACT_VECTOR
8228 // operands, where N is the size of the formed vector.
8229 // Each EXTRACT_VECTOR should have the same input vector and odd or even
8230 // index such that we have a pair wise add pattern.
8232 // Grab the vector that all EXTRACT_VECTOR nodes should be referencing.
8233 if (N0->getOperand(0)->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
8235 SDValue Vec = N0->getOperand(0)->getOperand(0);
8236 SDNode *V = Vec.getNode();
8237 unsigned nextIndex = 0;
8239 // For each operands to the ADD which are BUILD_VECTORs,
8240 // check to see if each of their operands are an EXTRACT_VECTOR with
8241 // the same vector and appropriate index.
8242 for (unsigned i = 0, e = N0->getNumOperands(); i != e; ++i) {
8243 if (N0->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT
8244 && N1->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
8246 SDValue ExtVec0 = N0->getOperand(i);
8247 SDValue ExtVec1 = N1->getOperand(i);
8249 // First operand is the vector, verify its the same.
8250 if (V != ExtVec0->getOperand(0).getNode() ||
8251 V != ExtVec1->getOperand(0).getNode())
8254 // Second is the constant, verify its correct.
8255 ConstantSDNode *C0 = dyn_cast<ConstantSDNode>(ExtVec0->getOperand(1));
8256 ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(ExtVec1->getOperand(1));
8258 // For the constant, we want to see all the even or all the odd.
8259 if (!C0 || !C1 || C0->getZExtValue() != nextIndex
8260 || C1->getZExtValue() != nextIndex+1)
8269 // Create VPADDL node.
8270 SelectionDAG &DAG = DCI.DAG;
8271 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8275 // Build operand list.
8276 SmallVector<SDValue, 8> Ops;
8277 Ops.push_back(DAG.getConstant(Intrinsic::arm_neon_vpaddls, dl,
8278 TLI.getPointerTy(DAG.getDataLayout())));
8280 // Input is the vector.
8283 // Get widened type and narrowed type.
8285 unsigned numElem = VT.getVectorNumElements();
8287 EVT inputLaneType = Vec.getValueType().getVectorElementType();
8288 switch (inputLaneType.getSimpleVT().SimpleTy) {
8289 case MVT::i8: widenType = MVT::getVectorVT(MVT::i16, numElem); break;
8290 case MVT::i16: widenType = MVT::getVectorVT(MVT::i32, numElem); break;
8291 case MVT::i32: widenType = MVT::getVectorVT(MVT::i64, numElem); break;
8293 llvm_unreachable("Invalid vector element type for padd optimization.");
8296 SDValue tmp = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, widenType, Ops);
8297 unsigned ExtOp = VT.bitsGT(tmp.getValueType()) ? ISD::ANY_EXTEND : ISD::TRUNCATE;
8298 return DAG.getNode(ExtOp, dl, VT, tmp);
8301 static SDValue findMUL_LOHI(SDValue V) {
8302 if (V->getOpcode() == ISD::UMUL_LOHI ||
8303 V->getOpcode() == ISD::SMUL_LOHI)
8308 static SDValue AddCombineTo64bitMLAL(SDNode *AddcNode,
8309 TargetLowering::DAGCombinerInfo &DCI,
8310 const ARMSubtarget *Subtarget) {
8312 if (Subtarget->isThumb1Only()) return SDValue();
8314 // Only perform the checks after legalize when the pattern is available.
8315 if (DCI.isBeforeLegalize()) return SDValue();
8317 // Look for multiply add opportunities.
8318 // The pattern is a ISD::UMUL_LOHI followed by two add nodes, where
8319 // each add nodes consumes a value from ISD::UMUL_LOHI and there is
8320 // a glue link from the first add to the second add.
8321 // If we find this pattern, we can replace the U/SMUL_LOHI, ADDC, and ADDE by
8322 // a S/UMLAL instruction.
8325 // / \ [no multiline comment]
8331 assert(AddcNode->getOpcode() == ISD::ADDC && "Expect an ADDC");
8332 SDValue AddcOp0 = AddcNode->getOperand(0);
8333 SDValue AddcOp1 = AddcNode->getOperand(1);
8335 // Check if the two operands are from the same mul_lohi node.
8336 if (AddcOp0.getNode() == AddcOp1.getNode())
8339 assert(AddcNode->getNumValues() == 2 &&
8340 AddcNode->getValueType(0) == MVT::i32 &&
8341 "Expect ADDC with two result values. First: i32");
8343 // Check that we have a glued ADDC node.
8344 if (AddcNode->getValueType(1) != MVT::Glue)
8347 // Check that the ADDC adds the low result of the S/UMUL_LOHI.
8348 if (AddcOp0->getOpcode() != ISD::UMUL_LOHI &&
8349 AddcOp0->getOpcode() != ISD::SMUL_LOHI &&
8350 AddcOp1->getOpcode() != ISD::UMUL_LOHI &&
8351 AddcOp1->getOpcode() != ISD::SMUL_LOHI)
8354 // Look for the glued ADDE.
8355 SDNode* AddeNode = AddcNode->getGluedUser();
8359 // Make sure it is really an ADDE.
8360 if (AddeNode->getOpcode() != ISD::ADDE)
8363 assert(AddeNode->getNumOperands() == 3 &&
8364 AddeNode->getOperand(2).getValueType() == MVT::Glue &&
8365 "ADDE node has the wrong inputs");
8367 // Check for the triangle shape.
8368 SDValue AddeOp0 = AddeNode->getOperand(0);
8369 SDValue AddeOp1 = AddeNode->getOperand(1);
8371 // Make sure that the ADDE operands are not coming from the same node.
8372 if (AddeOp0.getNode() == AddeOp1.getNode())
8375 // Find the MUL_LOHI node walking up ADDE's operands.
8376 bool IsLeftOperandMUL = false;
8377 SDValue MULOp = findMUL_LOHI(AddeOp0);
8378 if (MULOp == SDValue())
8379 MULOp = findMUL_LOHI(AddeOp1);
8381 IsLeftOperandMUL = true;
8382 if (MULOp == SDValue())
8385 // Figure out the right opcode.
8386 unsigned Opc = MULOp->getOpcode();
8387 unsigned FinalOpc = (Opc == ISD::SMUL_LOHI) ? ARMISD::SMLAL : ARMISD::UMLAL;
8389 // Figure out the high and low input values to the MLAL node.
8390 SDValue* HiAdd = nullptr;
8391 SDValue* LoMul = nullptr;
8392 SDValue* LowAdd = nullptr;
8394 // Ensure that ADDE is from high result of ISD::SMUL_LOHI.
8395 if ((AddeOp0 != MULOp.getValue(1)) && (AddeOp1 != MULOp.getValue(1)))
8398 if (IsLeftOperandMUL)
8404 // Ensure that LoMul and LowAdd are taken from correct ISD::SMUL_LOHI node
8405 // whose low result is fed to the ADDC we are checking.
8407 if (AddcOp0 == MULOp.getValue(0)) {
8411 if (AddcOp1 == MULOp.getValue(0)) {
8419 // Create the merged node.
8420 SelectionDAG &DAG = DCI.DAG;
8422 // Build operand list.
8423 SmallVector<SDValue, 8> Ops;
8424 Ops.push_back(LoMul->getOperand(0));
8425 Ops.push_back(LoMul->getOperand(1));
8426 Ops.push_back(*LowAdd);
8427 Ops.push_back(*HiAdd);
8429 SDValue MLALNode = DAG.getNode(FinalOpc, SDLoc(AddcNode),
8430 DAG.getVTList(MVT::i32, MVT::i32), Ops);
8432 // Replace the ADDs' nodes uses by the MLA node's values.
8433 SDValue HiMLALResult(MLALNode.getNode(), 1);
8434 DAG.ReplaceAllUsesOfValueWith(SDValue(AddeNode, 0), HiMLALResult);
8436 SDValue LoMLALResult(MLALNode.getNode(), 0);
8437 DAG.ReplaceAllUsesOfValueWith(SDValue(AddcNode, 0), LoMLALResult);
8439 // Return original node to notify the driver to stop replacing.
8440 SDValue resNode(AddcNode, 0);
8444 /// PerformADDCCombine - Target-specific dag combine transform from
8445 /// ISD::ADDC, ISD::ADDE, and ISD::MUL_LOHI to MLAL.
8446 static SDValue PerformADDCCombine(SDNode *N,
8447 TargetLowering::DAGCombinerInfo &DCI,
8448 const ARMSubtarget *Subtarget) {
8450 return AddCombineTo64bitMLAL(N, DCI, Subtarget);
8454 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
8455 /// operands N0 and N1. This is a helper for PerformADDCombine that is
8456 /// called with the default operands, and if that fails, with commuted
8458 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
8459 TargetLowering::DAGCombinerInfo &DCI,
8460 const ARMSubtarget *Subtarget){
8462 // Attempt to create vpaddl for this add.
8463 SDValue Result = AddCombineToVPADDL(N, N0, N1, DCI, Subtarget);
8464 if (Result.getNode())
8467 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
8468 if (N0.getNode()->hasOneUse()) {
8469 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
8470 if (Result.getNode()) return Result;
8475 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
8477 static SDValue PerformADDCombine(SDNode *N,
8478 TargetLowering::DAGCombinerInfo &DCI,
8479 const ARMSubtarget *Subtarget) {
8480 SDValue N0 = N->getOperand(0);
8481 SDValue N1 = N->getOperand(1);
8483 // First try with the default operand order.
8484 SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI, Subtarget);
8485 if (Result.getNode())
8488 // If that didn't work, try again with the operands commuted.
8489 return PerformADDCombineWithOperands(N, N1, N0, DCI, Subtarget);
8492 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
8494 static SDValue PerformSUBCombine(SDNode *N,
8495 TargetLowering::DAGCombinerInfo &DCI) {
8496 SDValue N0 = N->getOperand(0);
8497 SDValue N1 = N->getOperand(1);
8499 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
8500 if (N1.getNode()->hasOneUse()) {
8501 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
8502 if (Result.getNode()) return Result;
8508 /// PerformVMULCombine
8509 /// Distribute (A + B) * C to (A * C) + (B * C) to take advantage of the
8510 /// special multiplier accumulator forwarding.
8516 // However, for (A + B) * (A + B),
8523 static SDValue PerformVMULCombine(SDNode *N,
8524 TargetLowering::DAGCombinerInfo &DCI,
8525 const ARMSubtarget *Subtarget) {
8526 if (!Subtarget->hasVMLxForwarding())
8529 SelectionDAG &DAG = DCI.DAG;
8530 SDValue N0 = N->getOperand(0);
8531 SDValue N1 = N->getOperand(1);
8532 unsigned Opcode = N0.getOpcode();
8533 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
8534 Opcode != ISD::FADD && Opcode != ISD::FSUB) {
8535 Opcode = N1.getOpcode();
8536 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
8537 Opcode != ISD::FADD && Opcode != ISD::FSUB)
8545 EVT VT = N->getValueType(0);
8547 SDValue N00 = N0->getOperand(0);
8548 SDValue N01 = N0->getOperand(1);
8549 return DAG.getNode(Opcode, DL, VT,
8550 DAG.getNode(ISD::MUL, DL, VT, N00, N1),
8551 DAG.getNode(ISD::MUL, DL, VT, N01, N1));
8554 static SDValue PerformMULCombine(SDNode *N,
8555 TargetLowering::DAGCombinerInfo &DCI,
8556 const ARMSubtarget *Subtarget) {
8557 SelectionDAG &DAG = DCI.DAG;
8559 if (Subtarget->isThumb1Only())
8562 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
8565 EVT VT = N->getValueType(0);
8566 if (VT.is64BitVector() || VT.is128BitVector())
8567 return PerformVMULCombine(N, DCI, Subtarget);
8571 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
8575 int64_t MulAmt = C->getSExtValue();
8576 unsigned ShiftAmt = countTrailingZeros<uint64_t>(MulAmt);
8578 ShiftAmt = ShiftAmt & (32 - 1);
8579 SDValue V = N->getOperand(0);
8583 MulAmt >>= ShiftAmt;
8586 if (isPowerOf2_32(MulAmt - 1)) {
8587 // (mul x, 2^N + 1) => (add (shl x, N), x)
8588 Res = DAG.getNode(ISD::ADD, DL, VT,
8590 DAG.getNode(ISD::SHL, DL, VT,
8592 DAG.getConstant(Log2_32(MulAmt - 1), DL,
8594 } else if (isPowerOf2_32(MulAmt + 1)) {
8595 // (mul x, 2^N - 1) => (sub (shl x, N), x)
8596 Res = DAG.getNode(ISD::SUB, DL, VT,
8597 DAG.getNode(ISD::SHL, DL, VT,
8599 DAG.getConstant(Log2_32(MulAmt + 1), DL,
8605 uint64_t MulAmtAbs = -MulAmt;
8606 if (isPowerOf2_32(MulAmtAbs + 1)) {
8607 // (mul x, -(2^N - 1)) => (sub x, (shl x, N))
8608 Res = DAG.getNode(ISD::SUB, DL, VT,
8610 DAG.getNode(ISD::SHL, DL, VT,
8612 DAG.getConstant(Log2_32(MulAmtAbs + 1), DL,
8614 } else if (isPowerOf2_32(MulAmtAbs - 1)) {
8615 // (mul x, -(2^N + 1)) => - (add (shl x, N), x)
8616 Res = DAG.getNode(ISD::ADD, DL, VT,
8618 DAG.getNode(ISD::SHL, DL, VT,
8620 DAG.getConstant(Log2_32(MulAmtAbs - 1), DL,
8622 Res = DAG.getNode(ISD::SUB, DL, VT,
8623 DAG.getConstant(0, DL, MVT::i32), Res);
8630 Res = DAG.getNode(ISD::SHL, DL, VT,
8631 Res, DAG.getConstant(ShiftAmt, DL, MVT::i32));
8633 // Do not add new nodes to DAG combiner worklist.
8634 DCI.CombineTo(N, Res, false);
8638 static SDValue PerformANDCombine(SDNode *N,
8639 TargetLowering::DAGCombinerInfo &DCI,
8640 const ARMSubtarget *Subtarget) {
8642 // Attempt to use immediate-form VBIC
8643 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
8645 EVT VT = N->getValueType(0);
8646 SelectionDAG &DAG = DCI.DAG;
8648 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8651 APInt SplatBits, SplatUndef;
8652 unsigned SplatBitSize;
8655 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
8656 if (SplatBitSize <= 64) {
8658 SDValue Val = isNEONModifiedImm((~SplatBits).getZExtValue(),
8659 SplatUndef.getZExtValue(), SplatBitSize,
8660 DAG, dl, VbicVT, VT.is128BitVector(),
8662 if (Val.getNode()) {
8664 DAG.getNode(ISD::BITCAST, dl, VbicVT, N->getOperand(0));
8665 SDValue Vbic = DAG.getNode(ARMISD::VBICIMM, dl, VbicVT, Input, Val);
8666 return DAG.getNode(ISD::BITCAST, dl, VT, Vbic);
8671 if (!Subtarget->isThumb1Only()) {
8672 // fold (and (select cc, -1, c), x) -> (select cc, x, (and, x, c))
8673 SDValue Result = combineSelectAndUseCommutative(N, true, DCI);
8674 if (Result.getNode())
8681 /// PerformORCombine - Target-specific dag combine xforms for ISD::OR
8682 static SDValue PerformORCombine(SDNode *N,
8683 TargetLowering::DAGCombinerInfo &DCI,
8684 const ARMSubtarget *Subtarget) {
8685 // Attempt to use immediate-form VORR
8686 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
8688 EVT VT = N->getValueType(0);
8689 SelectionDAG &DAG = DCI.DAG;
8691 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8694 APInt SplatBits, SplatUndef;
8695 unsigned SplatBitSize;
8697 if (BVN && Subtarget->hasNEON() &&
8698 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
8699 if (SplatBitSize <= 64) {
8701 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
8702 SplatUndef.getZExtValue(), SplatBitSize,
8703 DAG, dl, VorrVT, VT.is128BitVector(),
8705 if (Val.getNode()) {
8707 DAG.getNode(ISD::BITCAST, dl, VorrVT, N->getOperand(0));
8708 SDValue Vorr = DAG.getNode(ARMISD::VORRIMM, dl, VorrVT, Input, Val);
8709 return DAG.getNode(ISD::BITCAST, dl, VT, Vorr);
8714 if (!Subtarget->isThumb1Only()) {
8715 // fold (or (select cc, 0, c), x) -> (select cc, x, (or, x, c))
8716 SDValue Result = combineSelectAndUseCommutative(N, false, DCI);
8717 if (Result.getNode())
8721 // The code below optimizes (or (and X, Y), Z).
8722 // The AND operand needs to have a single user to make these optimizations
8724 SDValue N0 = N->getOperand(0);
8725 if (N0.getOpcode() != ISD::AND || !N0.hasOneUse())
8727 SDValue N1 = N->getOperand(1);
8729 // (or (and B, A), (and C, ~A)) => (VBSL A, B, C) when A is a constant.
8730 if (Subtarget->hasNEON() && N1.getOpcode() == ISD::AND && VT.isVector() &&
8731 DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
8733 unsigned SplatBitSize;
8736 APInt SplatBits0, SplatBits1;
8737 BuildVectorSDNode *BVN0 = dyn_cast<BuildVectorSDNode>(N0->getOperand(1));
8738 BuildVectorSDNode *BVN1 = dyn_cast<BuildVectorSDNode>(N1->getOperand(1));
8739 // Ensure that the second operand of both ands are constants
8740 if (BVN0 && BVN0->isConstantSplat(SplatBits0, SplatUndef, SplatBitSize,
8741 HasAnyUndefs) && !HasAnyUndefs) {
8742 if (BVN1 && BVN1->isConstantSplat(SplatBits1, SplatUndef, SplatBitSize,
8743 HasAnyUndefs) && !HasAnyUndefs) {
8744 // Ensure that the bit width of the constants are the same and that
8745 // the splat arguments are logical inverses as per the pattern we
8746 // are trying to simplify.
8747 if (SplatBits0.getBitWidth() == SplatBits1.getBitWidth() &&
8748 SplatBits0 == ~SplatBits1) {
8749 // Canonicalize the vector type to make instruction selection
8751 EVT CanonicalVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
8752 SDValue Result = DAG.getNode(ARMISD::VBSL, dl, CanonicalVT,
8756 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
8762 // Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when
8765 // BFI is only available on V6T2+
8766 if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops())
8770 // 1) or (and A, mask), val => ARMbfi A, val, mask
8771 // iff (val & mask) == val
8773 // 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
8774 // 2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2)
8775 // && mask == ~mask2
8776 // 2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2)
8777 // && ~mask == mask2
8778 // (i.e., copy a bitfield value into another bitfield of the same width)
8783 SDValue N00 = N0.getOperand(0);
8785 // The value and the mask need to be constants so we can verify this is
8786 // actually a bitfield set. If the mask is 0xffff, we can do better
8787 // via a movt instruction, so don't use BFI in that case.
8788 SDValue MaskOp = N0.getOperand(1);
8789 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(MaskOp);
8792 unsigned Mask = MaskC->getZExtValue();
8796 // Case (1): or (and A, mask), val => ARMbfi A, val, mask
8797 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
8799 unsigned Val = N1C->getZExtValue();
8800 if ((Val & ~Mask) != Val)
8803 if (ARM::isBitFieldInvertedMask(Mask)) {
8804 Val >>= countTrailingZeros(~Mask);
8806 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00,
8807 DAG.getConstant(Val, DL, MVT::i32),
8808 DAG.getConstant(Mask, DL, MVT::i32));
8810 // Do not add new nodes to DAG combiner worklist.
8811 DCI.CombineTo(N, Res, false);
8814 } else if (N1.getOpcode() == ISD::AND) {
8815 // case (2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
8816 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
8819 unsigned Mask2 = N11C->getZExtValue();
8821 // Mask and ~Mask2 (or reverse) must be equivalent for the BFI pattern
8823 if (ARM::isBitFieldInvertedMask(Mask) &&
8825 // The pack halfword instruction works better for masks that fit it,
8826 // so use that when it's available.
8827 if (Subtarget->hasT2ExtractPack() &&
8828 (Mask == 0xffff || Mask == 0xffff0000))
8831 unsigned amt = countTrailingZeros(Mask2);
8832 Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0),
8833 DAG.getConstant(amt, DL, MVT::i32));
8834 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00, Res,
8835 DAG.getConstant(Mask, DL, MVT::i32));
8836 // Do not add new nodes to DAG combiner worklist.
8837 DCI.CombineTo(N, Res, false);
8839 } else if (ARM::isBitFieldInvertedMask(~Mask) &&
8841 // The pack halfword instruction works better for masks that fit it,
8842 // so use that when it's available.
8843 if (Subtarget->hasT2ExtractPack() &&
8844 (Mask2 == 0xffff || Mask2 == 0xffff0000))
8847 unsigned lsb = countTrailingZeros(Mask);
8848 Res = DAG.getNode(ISD::SRL, DL, VT, N00,
8849 DAG.getConstant(lsb, DL, MVT::i32));
8850 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res,
8851 DAG.getConstant(Mask2, DL, MVT::i32));
8852 // Do not add new nodes to DAG combiner worklist.
8853 DCI.CombineTo(N, Res, false);
8858 if (DAG.MaskedValueIsZero(N1, MaskC->getAPIntValue()) &&
8859 N00.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N00.getOperand(1)) &&
8860 ARM::isBitFieldInvertedMask(~Mask)) {
8861 // Case (3): or (and (shl A, #shamt), mask), B => ARMbfi B, A, ~mask
8862 // where lsb(mask) == #shamt and masked bits of B are known zero.
8863 SDValue ShAmt = N00.getOperand(1);
8864 unsigned ShAmtC = cast<ConstantSDNode>(ShAmt)->getZExtValue();
8865 unsigned LSB = countTrailingZeros(Mask);
8869 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1, N00.getOperand(0),
8870 DAG.getConstant(~Mask, DL, MVT::i32));
8872 // Do not add new nodes to DAG combiner worklist.
8873 DCI.CombineTo(N, Res, false);
8879 static SDValue PerformXORCombine(SDNode *N,
8880 TargetLowering::DAGCombinerInfo &DCI,
8881 const ARMSubtarget *Subtarget) {
8882 EVT VT = N->getValueType(0);
8883 SelectionDAG &DAG = DCI.DAG;
8885 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8888 if (!Subtarget->isThumb1Only()) {
8889 // fold (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c))
8890 SDValue Result = combineSelectAndUseCommutative(N, false, DCI);
8891 if (Result.getNode())
8898 /// PerformBFICombine - (bfi A, (and B, Mask1), Mask2) -> (bfi A, B, Mask2) iff
8899 /// the bits being cleared by the AND are not demanded by the BFI.
8900 static SDValue PerformBFICombine(SDNode *N,
8901 TargetLowering::DAGCombinerInfo &DCI) {
8902 SDValue N1 = N->getOperand(1);
8903 if (N1.getOpcode() == ISD::AND) {
8904 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
8907 unsigned InvMask = cast<ConstantSDNode>(N->getOperand(2))->getZExtValue();
8908 unsigned LSB = countTrailingZeros(~InvMask);
8909 unsigned Width = (32 - countLeadingZeros(~InvMask)) - LSB;
8911 static_cast<unsigned>(std::numeric_limits<unsigned>::digits) &&
8912 "undefined behavior");
8913 unsigned Mask = (1u << Width) - 1;
8914 unsigned Mask2 = N11C->getZExtValue();
8915 if ((Mask & (~Mask2)) == 0)
8916 return DCI.DAG.getNode(ARMISD::BFI, SDLoc(N), N->getValueType(0),
8917 N->getOperand(0), N1.getOperand(0),
8923 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
8924 /// ARMISD::VMOVRRD.
8925 static SDValue PerformVMOVRRDCombine(SDNode *N,
8926 TargetLowering::DAGCombinerInfo &DCI,
8927 const ARMSubtarget *Subtarget) {
8928 // vmovrrd(vmovdrr x, y) -> x,y
8929 SDValue InDouble = N->getOperand(0);
8930 if (InDouble.getOpcode() == ARMISD::VMOVDRR && !Subtarget->isFPOnlySP())
8931 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
8933 // vmovrrd(load f64) -> (load i32), (load i32)
8934 SDNode *InNode = InDouble.getNode();
8935 if (ISD::isNormalLoad(InNode) && InNode->hasOneUse() &&
8936 InNode->getValueType(0) == MVT::f64 &&
8937 InNode->getOperand(1).getOpcode() == ISD::FrameIndex &&
8938 !cast<LoadSDNode>(InNode)->isVolatile()) {
8939 // TODO: Should this be done for non-FrameIndex operands?
8940 LoadSDNode *LD = cast<LoadSDNode>(InNode);
8942 SelectionDAG &DAG = DCI.DAG;
8944 SDValue BasePtr = LD->getBasePtr();
8945 SDValue NewLD1 = DAG.getLoad(MVT::i32, DL, LD->getChain(), BasePtr,
8946 LD->getPointerInfo(), LD->isVolatile(),
8947 LD->isNonTemporal(), LD->isInvariant(),
8948 LD->getAlignment());
8950 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
8951 DAG.getConstant(4, DL, MVT::i32));
8952 SDValue NewLD2 = DAG.getLoad(MVT::i32, DL, NewLD1.getValue(1), OffsetPtr,
8953 LD->getPointerInfo(), LD->isVolatile(),
8954 LD->isNonTemporal(), LD->isInvariant(),
8955 std::min(4U, LD->getAlignment() / 2));
8957 DAG.ReplaceAllUsesOfValueWith(SDValue(LD, 1), NewLD2.getValue(1));
8958 if (DCI.DAG.getDataLayout().isBigEndian())
8959 std::swap (NewLD1, NewLD2);
8960 SDValue Result = DCI.CombineTo(N, NewLD1, NewLD2);
8967 /// PerformVMOVDRRCombine - Target-specific dag combine xforms for
8968 /// ARMISD::VMOVDRR. This is also used for BUILD_VECTORs with 2 operands.
8969 static SDValue PerformVMOVDRRCombine(SDNode *N, SelectionDAG &DAG) {
8970 // N=vmovrrd(X); vmovdrr(N:0, N:1) -> bit_convert(X)
8971 SDValue Op0 = N->getOperand(0);
8972 SDValue Op1 = N->getOperand(1);
8973 if (Op0.getOpcode() == ISD::BITCAST)
8974 Op0 = Op0.getOperand(0);
8975 if (Op1.getOpcode() == ISD::BITCAST)
8976 Op1 = Op1.getOperand(0);
8977 if (Op0.getOpcode() == ARMISD::VMOVRRD &&
8978 Op0.getNode() == Op1.getNode() &&
8979 Op0.getResNo() == 0 && Op1.getResNo() == 1)
8980 return DAG.getNode(ISD::BITCAST, SDLoc(N),
8981 N->getValueType(0), Op0.getOperand(0));
8985 /// hasNormalLoadOperand - Check if any of the operands of a BUILD_VECTOR node
8986 /// are normal, non-volatile loads. If so, it is profitable to bitcast an
8987 /// i64 vector to have f64 elements, since the value can then be loaded
8988 /// directly into a VFP register.
8989 static bool hasNormalLoadOperand(SDNode *N) {
8990 unsigned NumElts = N->getValueType(0).getVectorNumElements();
8991 for (unsigned i = 0; i < NumElts; ++i) {
8992 SDNode *Elt = N->getOperand(i).getNode();
8993 if (ISD::isNormalLoad(Elt) && !cast<LoadSDNode>(Elt)->isVolatile())
8999 /// PerformBUILD_VECTORCombine - Target-specific dag combine xforms for
9000 /// ISD::BUILD_VECTOR.
9001 static SDValue PerformBUILD_VECTORCombine(SDNode *N,
9002 TargetLowering::DAGCombinerInfo &DCI,
9003 const ARMSubtarget *Subtarget) {
9004 // build_vector(N=ARMISD::VMOVRRD(X), N:1) -> bit_convert(X):
9005 // VMOVRRD is introduced when legalizing i64 types. It forces the i64 value
9006 // into a pair of GPRs, which is fine when the value is used as a scalar,
9007 // but if the i64 value is converted to a vector, we need to undo the VMOVRRD.
9008 SelectionDAG &DAG = DCI.DAG;
9009 if (N->getNumOperands() == 2) {
9010 SDValue RV = PerformVMOVDRRCombine(N, DAG);
9015 // Load i64 elements as f64 values so that type legalization does not split
9016 // them up into i32 values.
9017 EVT VT = N->getValueType(0);
9018 if (VT.getVectorElementType() != MVT::i64 || !hasNormalLoadOperand(N))
9021 SmallVector<SDValue, 8> Ops;
9022 unsigned NumElts = VT.getVectorNumElements();
9023 for (unsigned i = 0; i < NumElts; ++i) {
9024 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(i));
9026 // Make the DAGCombiner fold the bitcast.
9027 DCI.AddToWorklist(V.getNode());
9029 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, NumElts);
9030 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, FloatVT, Ops);
9031 return DAG.getNode(ISD::BITCAST, dl, VT, BV);
9034 /// \brief Target-specific dag combine xforms for ARMISD::BUILD_VECTOR.
9036 PerformARMBUILD_VECTORCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
9037 // ARMISD::BUILD_VECTOR is introduced when legalizing ISD::BUILD_VECTOR.
9038 // At that time, we may have inserted bitcasts from integer to float.
9039 // If these bitcasts have survived DAGCombine, change the lowering of this
9040 // BUILD_VECTOR in something more vector friendly, i.e., that does not
9041 // force to use floating point types.
9043 // Make sure we can change the type of the vector.
9044 // This is possible iff:
9045 // 1. The vector is only used in a bitcast to a integer type. I.e.,
9046 // 1.1. Vector is used only once.
9047 // 1.2. Use is a bit convert to an integer type.
9048 // 2. The size of its operands are 32-bits (64-bits are not legal).
9049 EVT VT = N->getValueType(0);
9050 EVT EltVT = VT.getVectorElementType();
9052 // Check 1.1. and 2.
9053 if (EltVT.getSizeInBits() != 32 || !N->hasOneUse())
9056 // By construction, the input type must be float.
9057 assert(EltVT == MVT::f32 && "Unexpected type!");
9060 SDNode *Use = *N->use_begin();
9061 if (Use->getOpcode() != ISD::BITCAST ||
9062 Use->getValueType(0).isFloatingPoint())
9065 // Check profitability.
9066 // Model is, if more than half of the relevant operands are bitcast from
9067 // i32, turn the build_vector into a sequence of insert_vector_elt.
9068 // Relevant operands are everything that is not statically
9069 // (i.e., at compile time) bitcasted.
9070 unsigned NumOfBitCastedElts = 0;
9071 unsigned NumElts = VT.getVectorNumElements();
9072 unsigned NumOfRelevantElts = NumElts;
9073 for (unsigned Idx = 0; Idx < NumElts; ++Idx) {
9074 SDValue Elt = N->getOperand(Idx);
9075 if (Elt->getOpcode() == ISD::BITCAST) {
9076 // Assume only bit cast to i32 will go away.
9077 if (Elt->getOperand(0).getValueType() == MVT::i32)
9078 ++NumOfBitCastedElts;
9079 } else if (Elt.getOpcode() == ISD::UNDEF || isa<ConstantSDNode>(Elt))
9080 // Constants are statically casted, thus do not count them as
9081 // relevant operands.
9082 --NumOfRelevantElts;
9085 // Check if more than half of the elements require a non-free bitcast.
9086 if (NumOfBitCastedElts <= NumOfRelevantElts / 2)
9089 SelectionDAG &DAG = DCI.DAG;
9090 // Create the new vector type.
9091 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
9092 // Check if the type is legal.
9093 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9094 if (!TLI.isTypeLegal(VecVT))
9098 // ARMISD::BUILD_VECTOR E1, E2, ..., EN.
9099 // => BITCAST INSERT_VECTOR_ELT
9100 // (INSERT_VECTOR_ELT (...), (BITCAST EN-1), N-1),
9102 SDValue Vec = DAG.getUNDEF(VecVT);
9104 for (unsigned Idx = 0 ; Idx < NumElts; ++Idx) {
9105 SDValue V = N->getOperand(Idx);
9106 if (V.getOpcode() == ISD::UNDEF)
9108 if (V.getOpcode() == ISD::BITCAST &&
9109 V->getOperand(0).getValueType() == MVT::i32)
9110 // Fold obvious case.
9111 V = V.getOperand(0);
9113 V = DAG.getNode(ISD::BITCAST, SDLoc(V), MVT::i32, V);
9114 // Make the DAGCombiner fold the bitcasts.
9115 DCI.AddToWorklist(V.getNode());
9117 SDValue LaneIdx = DAG.getConstant(Idx, dl, MVT::i32);
9118 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VecVT, Vec, V, LaneIdx);
9120 Vec = DAG.getNode(ISD::BITCAST, dl, VT, Vec);
9121 // Make the DAGCombiner fold the bitcasts.
9122 DCI.AddToWorklist(Vec.getNode());
9126 /// PerformInsertEltCombine - Target-specific dag combine xforms for
9127 /// ISD::INSERT_VECTOR_ELT.
9128 static SDValue PerformInsertEltCombine(SDNode *N,
9129 TargetLowering::DAGCombinerInfo &DCI) {
9130 // Bitcast an i64 load inserted into a vector to f64.
9131 // Otherwise, the i64 value will be legalized to a pair of i32 values.
9132 EVT VT = N->getValueType(0);
9133 SDNode *Elt = N->getOperand(1).getNode();
9134 if (VT.getVectorElementType() != MVT::i64 ||
9135 !ISD::isNormalLoad(Elt) || cast<LoadSDNode>(Elt)->isVolatile())
9138 SelectionDAG &DAG = DCI.DAG;
9140 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
9141 VT.getVectorNumElements());
9142 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, N->getOperand(0));
9143 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(1));
9144 // Make the DAGCombiner fold the bitcasts.
9145 DCI.AddToWorklist(Vec.getNode());
9146 DCI.AddToWorklist(V.getNode());
9147 SDValue InsElt = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, FloatVT,
9148 Vec, V, N->getOperand(2));
9149 return DAG.getNode(ISD::BITCAST, dl, VT, InsElt);
9152 /// PerformVECTOR_SHUFFLECombine - Target-specific dag combine xforms for
9153 /// ISD::VECTOR_SHUFFLE.
9154 static SDValue PerformVECTOR_SHUFFLECombine(SDNode *N, SelectionDAG &DAG) {
9155 // The LLVM shufflevector instruction does not require the shuffle mask
9156 // length to match the operand vector length, but ISD::VECTOR_SHUFFLE does
9157 // have that requirement. When translating to ISD::VECTOR_SHUFFLE, if the
9158 // operands do not match the mask length, they are extended by concatenating
9159 // them with undef vectors. That is probably the right thing for other
9160 // targets, but for NEON it is better to concatenate two double-register
9161 // size vector operands into a single quad-register size vector. Do that
9162 // transformation here:
9163 // shuffle(concat(v1, undef), concat(v2, undef)) ->
9164 // shuffle(concat(v1, v2), undef)
9165 SDValue Op0 = N->getOperand(0);
9166 SDValue Op1 = N->getOperand(1);
9167 if (Op0.getOpcode() != ISD::CONCAT_VECTORS ||
9168 Op1.getOpcode() != ISD::CONCAT_VECTORS ||
9169 Op0.getNumOperands() != 2 ||
9170 Op1.getNumOperands() != 2)
9172 SDValue Concat0Op1 = Op0.getOperand(1);
9173 SDValue Concat1Op1 = Op1.getOperand(1);
9174 if (Concat0Op1.getOpcode() != ISD::UNDEF ||
9175 Concat1Op1.getOpcode() != ISD::UNDEF)
9177 // Skip the transformation if any of the types are illegal.
9178 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9179 EVT VT = N->getValueType(0);
9180 if (!TLI.isTypeLegal(VT) ||
9181 !TLI.isTypeLegal(Concat0Op1.getValueType()) ||
9182 !TLI.isTypeLegal(Concat1Op1.getValueType()))
9185 SDValue NewConcat = DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT,
9186 Op0.getOperand(0), Op1.getOperand(0));
9187 // Translate the shuffle mask.
9188 SmallVector<int, 16> NewMask;
9189 unsigned NumElts = VT.getVectorNumElements();
9190 unsigned HalfElts = NumElts/2;
9191 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
9192 for (unsigned n = 0; n < NumElts; ++n) {
9193 int MaskElt = SVN->getMaskElt(n);
9195 if (MaskElt < (int)HalfElts)
9197 else if (MaskElt >= (int)NumElts && MaskElt < (int)(NumElts + HalfElts))
9198 NewElt = HalfElts + MaskElt - NumElts;
9199 NewMask.push_back(NewElt);
9201 return DAG.getVectorShuffle(VT, SDLoc(N), NewConcat,
9202 DAG.getUNDEF(VT), NewMask.data());
9205 /// CombineBaseUpdate - Target-specific DAG combine function for VLDDUP,
9206 /// NEON load/store intrinsics, and generic vector load/stores, to merge
9207 /// base address updates.
9208 /// For generic load/stores, the memory type is assumed to be a vector.
9209 /// The caller is assumed to have checked legality.
9210 static SDValue CombineBaseUpdate(SDNode *N,
9211 TargetLowering::DAGCombinerInfo &DCI) {
9212 SelectionDAG &DAG = DCI.DAG;
9213 const bool isIntrinsic = (N->getOpcode() == ISD::INTRINSIC_VOID ||
9214 N->getOpcode() == ISD::INTRINSIC_W_CHAIN);
9215 const bool isStore = N->getOpcode() == ISD::STORE;
9216 const unsigned AddrOpIdx = ((isIntrinsic || isStore) ? 2 : 1);
9217 SDValue Addr = N->getOperand(AddrOpIdx);
9218 MemSDNode *MemN = cast<MemSDNode>(N);
9221 // Search for a use of the address operand that is an increment.
9222 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(),
9223 UE = Addr.getNode()->use_end(); UI != UE; ++UI) {
9225 if (User->getOpcode() != ISD::ADD ||
9226 UI.getUse().getResNo() != Addr.getResNo())
9229 // Check that the add is independent of the load/store. Otherwise, folding
9230 // it would create a cycle.
9231 if (User->isPredecessorOf(N) || N->isPredecessorOf(User))
9234 // Find the new opcode for the updating load/store.
9235 bool isLoadOp = true;
9236 bool isLaneOp = false;
9237 unsigned NewOpc = 0;
9238 unsigned NumVecs = 0;
9240 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
9242 default: llvm_unreachable("unexpected intrinsic for Neon base update");
9243 case Intrinsic::arm_neon_vld1: NewOpc = ARMISD::VLD1_UPD;
9245 case Intrinsic::arm_neon_vld2: NewOpc = ARMISD::VLD2_UPD;
9247 case Intrinsic::arm_neon_vld3: NewOpc = ARMISD::VLD3_UPD;
9249 case Intrinsic::arm_neon_vld4: NewOpc = ARMISD::VLD4_UPD;
9251 case Intrinsic::arm_neon_vld2lane: NewOpc = ARMISD::VLD2LN_UPD;
9252 NumVecs = 2; isLaneOp = true; break;
9253 case Intrinsic::arm_neon_vld3lane: NewOpc = ARMISD::VLD3LN_UPD;
9254 NumVecs = 3; isLaneOp = true; break;
9255 case Intrinsic::arm_neon_vld4lane: NewOpc = ARMISD::VLD4LN_UPD;
9256 NumVecs = 4; isLaneOp = true; break;
9257 case Intrinsic::arm_neon_vst1: NewOpc = ARMISD::VST1_UPD;
9258 NumVecs = 1; isLoadOp = false; break;
9259 case Intrinsic::arm_neon_vst2: NewOpc = ARMISD::VST2_UPD;
9260 NumVecs = 2; isLoadOp = false; break;
9261 case Intrinsic::arm_neon_vst3: NewOpc = ARMISD::VST3_UPD;
9262 NumVecs = 3; isLoadOp = false; break;
9263 case Intrinsic::arm_neon_vst4: NewOpc = ARMISD::VST4_UPD;
9264 NumVecs = 4; isLoadOp = false; break;
9265 case Intrinsic::arm_neon_vst2lane: NewOpc = ARMISD::VST2LN_UPD;
9266 NumVecs = 2; isLoadOp = false; isLaneOp = true; break;
9267 case Intrinsic::arm_neon_vst3lane: NewOpc = ARMISD::VST3LN_UPD;
9268 NumVecs = 3; isLoadOp = false; isLaneOp = true; break;
9269 case Intrinsic::arm_neon_vst4lane: NewOpc = ARMISD::VST4LN_UPD;
9270 NumVecs = 4; isLoadOp = false; isLaneOp = true; break;
9274 switch (N->getOpcode()) {
9275 default: llvm_unreachable("unexpected opcode for Neon base update");
9276 case ARMISD::VLD2DUP: NewOpc = ARMISD::VLD2DUP_UPD; NumVecs = 2; break;
9277 case ARMISD::VLD3DUP: NewOpc = ARMISD::VLD3DUP_UPD; NumVecs = 3; break;
9278 case ARMISD::VLD4DUP: NewOpc = ARMISD::VLD4DUP_UPD; NumVecs = 4; break;
9279 case ISD::LOAD: NewOpc = ARMISD::VLD1_UPD;
9280 NumVecs = 1; isLaneOp = false; break;
9281 case ISD::STORE: NewOpc = ARMISD::VST1_UPD;
9282 NumVecs = 1; isLaneOp = false; isLoadOp = false; break;
9286 // Find the size of memory referenced by the load/store.
9289 VecTy = N->getValueType(0);
9290 } else if (isIntrinsic) {
9291 VecTy = N->getOperand(AddrOpIdx+1).getValueType();
9293 assert(isStore && "Node has to be a load, a store, or an intrinsic!");
9294 VecTy = N->getOperand(1).getValueType();
9297 unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8;
9299 NumBytes /= VecTy.getVectorNumElements();
9301 // If the increment is a constant, it must match the memory ref size.
9302 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
9303 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
9304 uint64_t IncVal = CInc->getZExtValue();
9305 if (IncVal != NumBytes)
9307 } else if (NumBytes >= 3 * 16) {
9308 // VLD3/4 and VST3/4 for 128-bit vectors are implemented with two
9309 // separate instructions that make it harder to use a non-constant update.
9313 // OK, we found an ADD we can fold into the base update.
9314 // Now, create a _UPD node, taking care of not breaking alignment.
9316 EVT AlignedVecTy = VecTy;
9317 unsigned Alignment = MemN->getAlignment();
9319 // If this is a less-than-standard-aligned load/store, change the type to
9320 // match the standard alignment.
9321 // The alignment is overlooked when selecting _UPD variants; and it's
9322 // easier to introduce bitcasts here than fix that.
9323 // There are 3 ways to get to this base-update combine:
9324 // - intrinsics: they are assumed to be properly aligned (to the standard
9325 // alignment of the memory type), so we don't need to do anything.
9326 // - ARMISD::VLDx nodes: they are only generated from the aforementioned
9327 // intrinsics, so, likewise, there's nothing to do.
9328 // - generic load/store instructions: the alignment is specified as an
9329 // explicit operand, rather than implicitly as the standard alignment
9330 // of the memory type (like the intrisics). We need to change the
9331 // memory type to match the explicit alignment. That way, we don't
9332 // generate non-standard-aligned ARMISD::VLDx nodes.
9333 if (isa<LSBaseSDNode>(N)) {
9336 if (Alignment < VecTy.getScalarSizeInBits() / 8) {
9337 MVT EltTy = MVT::getIntegerVT(Alignment * 8);
9338 assert(NumVecs == 1 && "Unexpected multi-element generic load/store.");
9339 assert(!isLaneOp && "Unexpected generic load/store lane.");
9340 unsigned NumElts = NumBytes / (EltTy.getSizeInBits() / 8);
9341 AlignedVecTy = MVT::getVectorVT(EltTy, NumElts);
9343 // Don't set an explicit alignment on regular load/stores that we want
9344 // to transform to VLD/VST 1_UPD nodes.
9345 // This matches the behavior of regular load/stores, which only get an
9346 // explicit alignment if the MMO alignment is larger than the standard
9347 // alignment of the memory type.
9348 // Intrinsics, however, always get an explicit alignment, set to the
9349 // alignment of the MMO.
9353 // Create the new updating load/store node.
9354 // First, create an SDVTList for the new updating node's results.
9356 unsigned NumResultVecs = (isLoadOp ? NumVecs : 0);
9358 for (n = 0; n < NumResultVecs; ++n)
9359 Tys[n] = AlignedVecTy;
9360 Tys[n++] = MVT::i32;
9361 Tys[n] = MVT::Other;
9362 SDVTList SDTys = DAG.getVTList(makeArrayRef(Tys, NumResultVecs+2));
9364 // Then, gather the new node's operands.
9365 SmallVector<SDValue, 8> Ops;
9366 Ops.push_back(N->getOperand(0)); // incoming chain
9367 Ops.push_back(N->getOperand(AddrOpIdx));
9370 if (StoreSDNode *StN = dyn_cast<StoreSDNode>(N)) {
9371 // Try to match the intrinsic's signature
9372 Ops.push_back(StN->getValue());
9374 // Loads (and of course intrinsics) match the intrinsics' signature,
9375 // so just add all but the alignment operand.
9376 for (unsigned i = AddrOpIdx + 1; i < N->getNumOperands() - 1; ++i)
9377 Ops.push_back(N->getOperand(i));
9380 // For all node types, the alignment operand is always the last one.
9381 Ops.push_back(DAG.getConstant(Alignment, dl, MVT::i32));
9383 // If this is a non-standard-aligned STORE, the penultimate operand is the
9384 // stored value. Bitcast it to the aligned type.
9385 if (AlignedVecTy != VecTy && N->getOpcode() == ISD::STORE) {
9386 SDValue &StVal = Ops[Ops.size()-2];
9387 StVal = DAG.getNode(ISD::BITCAST, dl, AlignedVecTy, StVal);
9390 SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, dl, SDTys,
9392 MemN->getMemOperand());
9395 SmallVector<SDValue, 5> NewResults;
9396 for (unsigned i = 0; i < NumResultVecs; ++i)
9397 NewResults.push_back(SDValue(UpdN.getNode(), i));
9399 // If this is an non-standard-aligned LOAD, the first result is the loaded
9400 // value. Bitcast it to the expected result type.
9401 if (AlignedVecTy != VecTy && N->getOpcode() == ISD::LOAD) {
9402 SDValue &LdVal = NewResults[0];
9403 LdVal = DAG.getNode(ISD::BITCAST, dl, VecTy, LdVal);
9406 NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs+1)); // chain
9407 DCI.CombineTo(N, NewResults);
9408 DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs));
9415 static SDValue PerformVLDCombine(SDNode *N,
9416 TargetLowering::DAGCombinerInfo &DCI) {
9417 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
9420 return CombineBaseUpdate(N, DCI);
9423 /// CombineVLDDUP - For a VDUPLANE node N, check if its source operand is a
9424 /// vldN-lane (N > 1) intrinsic, and if all the other uses of that intrinsic
9425 /// are also VDUPLANEs. If so, combine them to a vldN-dup operation and
9427 static bool CombineVLDDUP(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
9428 SelectionDAG &DAG = DCI.DAG;
9429 EVT VT = N->getValueType(0);
9430 // vldN-dup instructions only support 64-bit vectors for N > 1.
9431 if (!VT.is64BitVector())
9434 // Check if the VDUPLANE operand is a vldN-dup intrinsic.
9435 SDNode *VLD = N->getOperand(0).getNode();
9436 if (VLD->getOpcode() != ISD::INTRINSIC_W_CHAIN)
9438 unsigned NumVecs = 0;
9439 unsigned NewOpc = 0;
9440 unsigned IntNo = cast<ConstantSDNode>(VLD->getOperand(1))->getZExtValue();
9441 if (IntNo == Intrinsic::arm_neon_vld2lane) {
9443 NewOpc = ARMISD::VLD2DUP;
9444 } else if (IntNo == Intrinsic::arm_neon_vld3lane) {
9446 NewOpc = ARMISD::VLD3DUP;
9447 } else if (IntNo == Intrinsic::arm_neon_vld4lane) {
9449 NewOpc = ARMISD::VLD4DUP;
9454 // First check that all the vldN-lane uses are VDUPLANEs and that the lane
9455 // numbers match the load.
9456 unsigned VLDLaneNo =
9457 cast<ConstantSDNode>(VLD->getOperand(NumVecs+3))->getZExtValue();
9458 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
9460 // Ignore uses of the chain result.
9461 if (UI.getUse().getResNo() == NumVecs)
9464 if (User->getOpcode() != ARMISD::VDUPLANE ||
9465 VLDLaneNo != cast<ConstantSDNode>(User->getOperand(1))->getZExtValue())
9469 // Create the vldN-dup node.
9472 for (n = 0; n < NumVecs; ++n)
9474 Tys[n] = MVT::Other;
9475 SDVTList SDTys = DAG.getVTList(makeArrayRef(Tys, NumVecs+1));
9476 SDValue Ops[] = { VLD->getOperand(0), VLD->getOperand(2) };
9477 MemIntrinsicSDNode *VLDMemInt = cast<MemIntrinsicSDNode>(VLD);
9478 SDValue VLDDup = DAG.getMemIntrinsicNode(NewOpc, SDLoc(VLD), SDTys,
9479 Ops, VLDMemInt->getMemoryVT(),
9480 VLDMemInt->getMemOperand());
9483 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
9485 unsigned ResNo = UI.getUse().getResNo();
9486 // Ignore uses of the chain result.
9487 if (ResNo == NumVecs)
9490 DCI.CombineTo(User, SDValue(VLDDup.getNode(), ResNo));
9493 // Now the vldN-lane intrinsic is dead except for its chain result.
9494 // Update uses of the chain.
9495 std::vector<SDValue> VLDDupResults;
9496 for (unsigned n = 0; n < NumVecs; ++n)
9497 VLDDupResults.push_back(SDValue(VLDDup.getNode(), n));
9498 VLDDupResults.push_back(SDValue(VLDDup.getNode(), NumVecs));
9499 DCI.CombineTo(VLD, VLDDupResults);
9504 /// PerformVDUPLANECombine - Target-specific dag combine xforms for
9505 /// ARMISD::VDUPLANE.
9506 static SDValue PerformVDUPLANECombine(SDNode *N,
9507 TargetLowering::DAGCombinerInfo &DCI) {
9508 SDValue Op = N->getOperand(0);
9510 // If the source is a vldN-lane (N > 1) intrinsic, and all the other uses
9511 // of that intrinsic are also VDUPLANEs, combine them to a vldN-dup operation.
9512 if (CombineVLDDUP(N, DCI))
9513 return SDValue(N, 0);
9515 // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is
9516 // redundant. Ignore bit_converts for now; element sizes are checked below.
9517 while (Op.getOpcode() == ISD::BITCAST)
9518 Op = Op.getOperand(0);
9519 if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM)
9522 // Make sure the VMOV element size is not bigger than the VDUPLANE elements.
9523 unsigned EltSize = Op.getValueType().getVectorElementType().getSizeInBits();
9524 // The canonical VMOV for a zero vector uses a 32-bit element size.
9525 unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
9527 if (ARM_AM::decodeNEONModImm(Imm, EltBits) == 0)
9529 EVT VT = N->getValueType(0);
9530 if (EltSize > VT.getVectorElementType().getSizeInBits())
9533 return DCI.DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Op);
9536 static SDValue PerformLOADCombine(SDNode *N,
9537 TargetLowering::DAGCombinerInfo &DCI) {
9538 EVT VT = N->getValueType(0);
9540 // If this is a legal vector load, try to combine it into a VLD1_UPD.
9541 if (ISD::isNormalLoad(N) && VT.isVector() &&
9542 DCI.DAG.getTargetLoweringInfo().isTypeLegal(VT))
9543 return CombineBaseUpdate(N, DCI);
9548 /// PerformSTORECombine - Target-specific dag combine xforms for
9550 static SDValue PerformSTORECombine(SDNode *N,
9551 TargetLowering::DAGCombinerInfo &DCI) {
9552 StoreSDNode *St = cast<StoreSDNode>(N);
9553 if (St->isVolatile())
9556 // Optimize trunc store (of multiple scalars) to shuffle and store. First,
9557 // pack all of the elements in one place. Next, store to memory in fewer
9559 SDValue StVal = St->getValue();
9560 EVT VT = StVal.getValueType();
9561 if (St->isTruncatingStore() && VT.isVector()) {
9562 SelectionDAG &DAG = DCI.DAG;
9563 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9564 EVT StVT = St->getMemoryVT();
9565 unsigned NumElems = VT.getVectorNumElements();
9566 assert(StVT != VT && "Cannot truncate to the same type");
9567 unsigned FromEltSz = VT.getVectorElementType().getSizeInBits();
9568 unsigned ToEltSz = StVT.getVectorElementType().getSizeInBits();
9570 // From, To sizes and ElemCount must be pow of two
9571 if (!isPowerOf2_32(NumElems * FromEltSz * ToEltSz)) return SDValue();
9573 // We are going to use the original vector elt for storing.
9574 // Accumulated smaller vector elements must be a multiple of the store size.
9575 if (0 != (NumElems * FromEltSz) % ToEltSz) return SDValue();
9577 unsigned SizeRatio = FromEltSz / ToEltSz;
9578 assert(SizeRatio * NumElems * ToEltSz == VT.getSizeInBits());
9580 // Create a type on which we perform the shuffle.
9581 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(), StVT.getScalarType(),
9582 NumElems*SizeRatio);
9583 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
9586 SDValue WideVec = DAG.getNode(ISD::BITCAST, DL, WideVecVT, StVal);
9587 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
9588 for (unsigned i = 0; i < NumElems; ++i)
9589 ShuffleVec[i] = DAG.getDataLayout().isBigEndian()
9590 ? (i + 1) * SizeRatio - 1
9593 // Can't shuffle using an illegal type.
9594 if (!TLI.isTypeLegal(WideVecVT)) return SDValue();
9596 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, DL, WideVec,
9597 DAG.getUNDEF(WideVec.getValueType()),
9599 // At this point all of the data is stored at the bottom of the
9600 // register. We now need to save it to mem.
9602 // Find the largest store unit
9603 MVT StoreType = MVT::i8;
9604 for (MVT Tp : MVT::integer_valuetypes()) {
9605 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToEltSz)
9608 // Didn't find a legal store type.
9609 if (!TLI.isTypeLegal(StoreType))
9612 // Bitcast the original vector into a vector of store-size units
9613 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
9614 StoreType, VT.getSizeInBits()/EVT(StoreType).getSizeInBits());
9615 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
9616 SDValue ShuffWide = DAG.getNode(ISD::BITCAST, DL, StoreVecVT, Shuff);
9617 SmallVector<SDValue, 8> Chains;
9618 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits() / 8, DL,
9619 TLI.getPointerTy(DAG.getDataLayout()));
9620 SDValue BasePtr = St->getBasePtr();
9622 // Perform one or more big stores into memory.
9623 unsigned E = (ToEltSz*NumElems)/StoreType.getSizeInBits();
9624 for (unsigned I = 0; I < E; I++) {
9625 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
9626 StoreType, ShuffWide,
9627 DAG.getIntPtrConstant(I, DL));
9628 SDValue Ch = DAG.getStore(St->getChain(), DL, SubVec, BasePtr,
9629 St->getPointerInfo(), St->isVolatile(),
9630 St->isNonTemporal(), St->getAlignment());
9631 BasePtr = DAG.getNode(ISD::ADD, DL, BasePtr.getValueType(), BasePtr,
9633 Chains.push_back(Ch);
9635 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
9638 if (!ISD::isNormalStore(St))
9641 // Split a store of a VMOVDRR into two integer stores to avoid mixing NEON and
9642 // ARM stores of arguments in the same cache line.
9643 if (StVal.getNode()->getOpcode() == ARMISD::VMOVDRR &&
9644 StVal.getNode()->hasOneUse()) {
9645 SelectionDAG &DAG = DCI.DAG;
9646 bool isBigEndian = DAG.getDataLayout().isBigEndian();
9648 SDValue BasePtr = St->getBasePtr();
9649 SDValue NewST1 = DAG.getStore(St->getChain(), DL,
9650 StVal.getNode()->getOperand(isBigEndian ? 1 : 0 ),
9651 BasePtr, St->getPointerInfo(), St->isVolatile(),
9652 St->isNonTemporal(), St->getAlignment());
9654 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
9655 DAG.getConstant(4, DL, MVT::i32));
9656 return DAG.getStore(NewST1.getValue(0), DL,
9657 StVal.getNode()->getOperand(isBigEndian ? 0 : 1),
9658 OffsetPtr, St->getPointerInfo(), St->isVolatile(),
9659 St->isNonTemporal(),
9660 std::min(4U, St->getAlignment() / 2));
9663 if (StVal.getValueType() == MVT::i64 &&
9664 StVal.getNode()->getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
9666 // Bitcast an i64 store extracted from a vector to f64.
9667 // Otherwise, the i64 value will be legalized to a pair of i32 values.
9668 SelectionDAG &DAG = DCI.DAG;
9670 SDValue IntVec = StVal.getOperand(0);
9671 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
9672 IntVec.getValueType().getVectorNumElements());
9673 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, IntVec);
9674 SDValue ExtElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
9675 Vec, StVal.getOperand(1));
9677 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ExtElt);
9678 // Make the DAGCombiner fold the bitcasts.
9679 DCI.AddToWorklist(Vec.getNode());
9680 DCI.AddToWorklist(ExtElt.getNode());
9681 DCI.AddToWorklist(V.getNode());
9682 return DAG.getStore(St->getChain(), dl, V, St->getBasePtr(),
9683 St->getPointerInfo(), St->isVolatile(),
9684 St->isNonTemporal(), St->getAlignment(),
9688 // If this is a legal vector store, try to combine it into a VST1_UPD.
9689 if (ISD::isNormalStore(N) && VT.isVector() &&
9690 DCI.DAG.getTargetLoweringInfo().isTypeLegal(VT))
9691 return CombineBaseUpdate(N, DCI);
9696 // isConstVecPow2 - Return true if each vector element is a power of 2, all
9697 // elements are the same constant, C, and Log2(C) ranges from 1 to 32.
9698 static bool isConstVecPow2(SDValue ConstVec, bool isSigned, uint64_t &C)
9702 for (unsigned I = 0, E = ConstVec.getValueType().getVectorNumElements();
9704 ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(ConstVec.getOperand(I));
9709 APFloat APF = C->getValueAPF();
9710 if (APF.convertToInteger(&cN, 64, isSigned, APFloat::rmTowardZero, &isExact)
9711 != APFloat::opOK || !isExact)
9714 c0 = (I == 0) ? cN : c0;
9715 if (!isPowerOf2_64(cN) || c0 != cN || Log2_64(c0) < 1 || Log2_64(c0) > 32)
9722 /// PerformVCVTCombine - VCVT (floating-point to fixed-point, Advanced SIMD)
9723 /// can replace combinations of VMUL and VCVT (floating-point to integer)
9724 /// when the VMUL has a constant operand that is a power of 2.
9726 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
9727 /// vmul.f32 d16, d17, d16
9728 /// vcvt.s32.f32 d16, d16
9730 /// vcvt.s32.f32 d16, d16, #3
9731 static SDValue PerformVCVTCombine(SDNode *N,
9732 TargetLowering::DAGCombinerInfo &DCI,
9733 const ARMSubtarget *Subtarget) {
9734 SelectionDAG &DAG = DCI.DAG;
9735 SDValue Op = N->getOperand(0);
9737 if (!Subtarget->hasNEON() || !Op.getValueType().isVector() ||
9738 Op.getOpcode() != ISD::FMUL)
9742 SDValue N0 = Op->getOperand(0);
9743 SDValue ConstVec = Op->getOperand(1);
9744 bool isSigned = N->getOpcode() == ISD::FP_TO_SINT;
9746 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
9747 !isConstVecPow2(ConstVec, isSigned, C))
9750 MVT FloatTy = Op.getSimpleValueType().getVectorElementType();
9751 MVT IntTy = N->getSimpleValueType(0).getVectorElementType();
9752 unsigned NumLanes = Op.getValueType().getVectorNumElements();
9753 if (FloatTy.getSizeInBits() != 32 || IntTy.getSizeInBits() > 32 ||
9755 // These instructions only exist converting from f32 to i32. We can handle
9756 // smaller integers by generating an extra truncate, but larger ones would
9757 // be lossy. We also can't handle more then 4 lanes, since these intructions
9758 // only support v2i32/v4i32 types.
9763 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfp2fxs :
9764 Intrinsic::arm_neon_vcvtfp2fxu;
9765 SDValue FixConv = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl,
9766 NumLanes == 2 ? MVT::v2i32 : MVT::v4i32,
9767 DAG.getConstant(IntrinsicOpcode, dl, MVT::i32),
9769 DAG.getConstant(Log2_64(C), dl, MVT::i32));
9771 if (IntTy.getSizeInBits() < FloatTy.getSizeInBits())
9772 FixConv = DAG.getNode(ISD::TRUNCATE, dl, N->getValueType(0), FixConv);
9777 /// PerformVDIVCombine - VCVT (fixed-point to floating-point, Advanced SIMD)
9778 /// can replace combinations of VCVT (integer to floating-point) and VDIV
9779 /// when the VDIV has a constant operand that is a power of 2.
9781 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
9782 /// vcvt.f32.s32 d16, d16
9783 /// vdiv.f32 d16, d17, d16
9785 /// vcvt.f32.s32 d16, d16, #3
9786 static SDValue PerformVDIVCombine(SDNode *N,
9787 TargetLowering::DAGCombinerInfo &DCI,
9788 const ARMSubtarget *Subtarget) {
9789 SelectionDAG &DAG = DCI.DAG;
9790 SDValue Op = N->getOperand(0);
9791 unsigned OpOpcode = Op.getNode()->getOpcode();
9793 if (!Subtarget->hasNEON() || !N->getValueType(0).isVector() ||
9794 (OpOpcode != ISD::SINT_TO_FP && OpOpcode != ISD::UINT_TO_FP))
9798 SDValue ConstVec = N->getOperand(1);
9799 bool isSigned = OpOpcode == ISD::SINT_TO_FP;
9801 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
9802 !isConstVecPow2(ConstVec, isSigned, C))
9805 MVT FloatTy = N->getSimpleValueType(0).getVectorElementType();
9806 MVT IntTy = Op.getOperand(0).getSimpleValueType().getVectorElementType();
9807 if (FloatTy.getSizeInBits() != 32 || IntTy.getSizeInBits() > 32) {
9808 // These instructions only exist converting from i32 to f32. We can handle
9809 // smaller integers by generating an extra extend, but larger ones would
9815 SDValue ConvInput = Op.getOperand(0);
9816 unsigned NumLanes = Op.getValueType().getVectorNumElements();
9817 if (IntTy.getSizeInBits() < FloatTy.getSizeInBits())
9818 ConvInput = DAG.getNode(isSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND,
9819 dl, NumLanes == 2 ? MVT::v2i32 : MVT::v4i32,
9822 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfxs2fp :
9823 Intrinsic::arm_neon_vcvtfxu2fp;
9824 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl,
9826 DAG.getConstant(IntrinsicOpcode, dl, MVT::i32),
9827 ConvInput, DAG.getConstant(Log2_64(C), dl, MVT::i32));
9830 /// Getvshiftimm - Check if this is a valid build_vector for the immediate
9831 /// operand of a vector shift operation, where all the elements of the
9832 /// build_vector must have the same constant integer value.
9833 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
9834 // Ignore bit_converts.
9835 while (Op.getOpcode() == ISD::BITCAST)
9836 Op = Op.getOperand(0);
9837 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
9838 APInt SplatBits, SplatUndef;
9839 unsigned SplatBitSize;
9841 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
9842 HasAnyUndefs, ElementBits) ||
9843 SplatBitSize > ElementBits)
9845 Cnt = SplatBits.getSExtValue();
9849 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
9850 /// operand of a vector shift left operation. That value must be in the range:
9851 /// 0 <= Value < ElementBits for a left shift; or
9852 /// 0 <= Value <= ElementBits for a long left shift.
9853 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
9854 assert(VT.isVector() && "vector shift count is not a vector type");
9855 int64_t ElementBits = VT.getVectorElementType().getSizeInBits();
9856 if (! getVShiftImm(Op, ElementBits, Cnt))
9858 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
9861 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
9862 /// operand of a vector shift right operation. For a shift opcode, the value
9863 /// is positive, but for an intrinsic the value count must be negative. The
9864 /// absolute value must be in the range:
9865 /// 1 <= |Value| <= ElementBits for a right shift; or
9866 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
9867 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
9869 assert(VT.isVector() && "vector shift count is not a vector type");
9870 int64_t ElementBits = VT.getVectorElementType().getSizeInBits();
9871 if (! getVShiftImm(Op, ElementBits, Cnt))
9874 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
9875 if (Cnt >= -(isNarrow ? ElementBits/2 : ElementBits) && Cnt <= -1) {
9882 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
9883 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
9884 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
9887 // Don't do anything for most intrinsics.
9890 case Intrinsic::arm_neon_vabds:
9891 if (!N->getValueType(0).isInteger())
9893 return DAG.getNode(ISD::SABSDIFF, SDLoc(N), N->getValueType(0),
9894 N->getOperand(1), N->getOperand(2));
9895 case Intrinsic::arm_neon_vabdu:
9896 return DAG.getNode(ISD::UABSDIFF, SDLoc(N), N->getValueType(0),
9897 N->getOperand(1), N->getOperand(2));
9899 // Vector shifts: check for immediate versions and lower them.
9900 // Note: This is done during DAG combining instead of DAG legalizing because
9901 // the build_vectors for 64-bit vector element shift counts are generally
9902 // not legal, and it is hard to see their values after they get legalized to
9903 // loads from a constant pool.
9904 case Intrinsic::arm_neon_vshifts:
9905 case Intrinsic::arm_neon_vshiftu:
9906 case Intrinsic::arm_neon_vrshifts:
9907 case Intrinsic::arm_neon_vrshiftu:
9908 case Intrinsic::arm_neon_vrshiftn:
9909 case Intrinsic::arm_neon_vqshifts:
9910 case Intrinsic::arm_neon_vqshiftu:
9911 case Intrinsic::arm_neon_vqshiftsu:
9912 case Intrinsic::arm_neon_vqshiftns:
9913 case Intrinsic::arm_neon_vqshiftnu:
9914 case Intrinsic::arm_neon_vqshiftnsu:
9915 case Intrinsic::arm_neon_vqrshiftns:
9916 case Intrinsic::arm_neon_vqrshiftnu:
9917 case Intrinsic::arm_neon_vqrshiftnsu: {
9918 EVT VT = N->getOperand(1).getValueType();
9920 unsigned VShiftOpc = 0;
9923 case Intrinsic::arm_neon_vshifts:
9924 case Intrinsic::arm_neon_vshiftu:
9925 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
9926 VShiftOpc = ARMISD::VSHL;
9929 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
9930 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
9931 ARMISD::VSHRs : ARMISD::VSHRu);
9936 case Intrinsic::arm_neon_vrshifts:
9937 case Intrinsic::arm_neon_vrshiftu:
9938 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
9942 case Intrinsic::arm_neon_vqshifts:
9943 case Intrinsic::arm_neon_vqshiftu:
9944 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
9948 case Intrinsic::arm_neon_vqshiftsu:
9949 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
9951 llvm_unreachable("invalid shift count for vqshlu intrinsic");
9953 case Intrinsic::arm_neon_vrshiftn:
9954 case Intrinsic::arm_neon_vqshiftns:
9955 case Intrinsic::arm_neon_vqshiftnu:
9956 case Intrinsic::arm_neon_vqshiftnsu:
9957 case Intrinsic::arm_neon_vqrshiftns:
9958 case Intrinsic::arm_neon_vqrshiftnu:
9959 case Intrinsic::arm_neon_vqrshiftnsu:
9960 // Narrowing shifts require an immediate right shift.
9961 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
9963 llvm_unreachable("invalid shift count for narrowing vector shift "
9967 llvm_unreachable("unhandled vector shift");
9971 case Intrinsic::arm_neon_vshifts:
9972 case Intrinsic::arm_neon_vshiftu:
9973 // Opcode already set above.
9975 case Intrinsic::arm_neon_vrshifts:
9976 VShiftOpc = ARMISD::VRSHRs; break;
9977 case Intrinsic::arm_neon_vrshiftu:
9978 VShiftOpc = ARMISD::VRSHRu; break;
9979 case Intrinsic::arm_neon_vrshiftn:
9980 VShiftOpc = ARMISD::VRSHRN; break;
9981 case Intrinsic::arm_neon_vqshifts:
9982 VShiftOpc = ARMISD::VQSHLs; break;
9983 case Intrinsic::arm_neon_vqshiftu:
9984 VShiftOpc = ARMISD::VQSHLu; break;
9985 case Intrinsic::arm_neon_vqshiftsu:
9986 VShiftOpc = ARMISD::VQSHLsu; break;
9987 case Intrinsic::arm_neon_vqshiftns:
9988 VShiftOpc = ARMISD::VQSHRNs; break;
9989 case Intrinsic::arm_neon_vqshiftnu:
9990 VShiftOpc = ARMISD::VQSHRNu; break;
9991 case Intrinsic::arm_neon_vqshiftnsu:
9992 VShiftOpc = ARMISD::VQSHRNsu; break;
9993 case Intrinsic::arm_neon_vqrshiftns:
9994 VShiftOpc = ARMISD::VQRSHRNs; break;
9995 case Intrinsic::arm_neon_vqrshiftnu:
9996 VShiftOpc = ARMISD::VQRSHRNu; break;
9997 case Intrinsic::arm_neon_vqrshiftnsu:
9998 VShiftOpc = ARMISD::VQRSHRNsu; break;
10002 return DAG.getNode(VShiftOpc, dl, N->getValueType(0),
10003 N->getOperand(1), DAG.getConstant(Cnt, dl, MVT::i32));
10006 case Intrinsic::arm_neon_vshiftins: {
10007 EVT VT = N->getOperand(1).getValueType();
10009 unsigned VShiftOpc = 0;
10011 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
10012 VShiftOpc = ARMISD::VSLI;
10013 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
10014 VShiftOpc = ARMISD::VSRI;
10016 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
10020 return DAG.getNode(VShiftOpc, dl, N->getValueType(0),
10021 N->getOperand(1), N->getOperand(2),
10022 DAG.getConstant(Cnt, dl, MVT::i32));
10025 case Intrinsic::arm_neon_vqrshifts:
10026 case Intrinsic::arm_neon_vqrshiftu:
10027 // No immediate versions of these to check for.
10034 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
10035 /// lowers them. As with the vector shift intrinsics, this is done during DAG
10036 /// combining instead of DAG legalizing because the build_vectors for 64-bit
10037 /// vector element shift counts are generally not legal, and it is hard to see
10038 /// their values after they get legalized to loads from a constant pool.
10039 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
10040 const ARMSubtarget *ST) {
10041 EVT VT = N->getValueType(0);
10042 if (N->getOpcode() == ISD::SRL && VT == MVT::i32 && ST->hasV6Ops()) {
10043 // Canonicalize (srl (bswap x), 16) to (rotr (bswap x), 16) if the high
10044 // 16-bits of x is zero. This optimizes rev + lsr 16 to rev16.
10045 SDValue N1 = N->getOperand(1);
10046 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) {
10047 SDValue N0 = N->getOperand(0);
10048 if (C->getZExtValue() == 16 && N0.getOpcode() == ISD::BSWAP &&
10049 DAG.MaskedValueIsZero(N0.getOperand(0),
10050 APInt::getHighBitsSet(32, 16)))
10051 return DAG.getNode(ISD::ROTR, SDLoc(N), VT, N0, N1);
10055 // Nothing to be done for scalar shifts.
10056 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
10057 if (!VT.isVector() || !TLI.isTypeLegal(VT))
10060 assert(ST->hasNEON() && "unexpected vector shift");
10063 switch (N->getOpcode()) {
10064 default: llvm_unreachable("unexpected shift opcode");
10067 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt)) {
10069 return DAG.getNode(ARMISD::VSHL, dl, VT, N->getOperand(0),
10070 DAG.getConstant(Cnt, dl, MVT::i32));
10076 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
10077 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
10078 ARMISD::VSHRs : ARMISD::VSHRu);
10080 return DAG.getNode(VShiftOpc, dl, VT, N->getOperand(0),
10081 DAG.getConstant(Cnt, dl, MVT::i32));
10087 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
10088 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
10089 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
10090 const ARMSubtarget *ST) {
10091 SDValue N0 = N->getOperand(0);
10093 // Check for sign- and zero-extensions of vector extract operations of 8-
10094 // and 16-bit vector elements. NEON supports these directly. They are
10095 // handled during DAG combining because type legalization will promote them
10096 // to 32-bit types and it is messy to recognize the operations after that.
10097 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
10098 SDValue Vec = N0.getOperand(0);
10099 SDValue Lane = N0.getOperand(1);
10100 EVT VT = N->getValueType(0);
10101 EVT EltVT = N0.getValueType();
10102 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
10104 if (VT == MVT::i32 &&
10105 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
10106 TLI.isTypeLegal(Vec.getValueType()) &&
10107 isa<ConstantSDNode>(Lane)) {
10110 switch (N->getOpcode()) {
10111 default: llvm_unreachable("unexpected opcode");
10112 case ISD::SIGN_EXTEND:
10113 Opc = ARMISD::VGETLANEs;
10115 case ISD::ZERO_EXTEND:
10116 case ISD::ANY_EXTEND:
10117 Opc = ARMISD::VGETLANEu;
10120 return DAG.getNode(Opc, SDLoc(N), VT, Vec, Lane);
10127 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
10128 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
10129 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
10130 const ARMSubtarget *ST) {
10131 // If the target supports NEON, try to use vmax/vmin instructions for f32
10132 // selects like "x < y ? x : y". Unless the NoNaNsFPMath option is set,
10133 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
10134 // a NaN; only do the transformation when it matches that behavior.
10136 // For now only do this when using NEON for FP operations; if using VFP, it
10137 // is not obvious that the benefit outweighs the cost of switching to the
10139 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
10140 N->getValueType(0) != MVT::f32)
10143 SDValue CondLHS = N->getOperand(0);
10144 SDValue CondRHS = N->getOperand(1);
10145 SDValue LHS = N->getOperand(2);
10146 SDValue RHS = N->getOperand(3);
10147 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
10149 unsigned Opcode = 0;
10151 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
10152 IsReversed = false; // x CC y ? x : y
10153 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
10154 IsReversed = true ; // x CC y ? y : x
10168 // If LHS is NaN, an ordered comparison will be false and the result will
10169 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
10170 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
10171 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
10172 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
10174 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
10175 // will return -0, so vmin can only be used for unsafe math or if one of
10176 // the operands is known to be nonzero.
10177 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
10178 !DAG.getTarget().Options.UnsafeFPMath &&
10179 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
10181 Opcode = IsReversed ? ISD::FMAXNAN : ISD::FMINNAN;
10190 // If LHS is NaN, an ordered comparison will be false and the result will
10191 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
10192 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
10193 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
10194 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
10196 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
10197 // will return +0, so vmax can only be used for unsafe math or if one of
10198 // the operands is known to be nonzero.
10199 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
10200 !DAG.getTarget().Options.UnsafeFPMath &&
10201 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
10203 Opcode = IsReversed ? ISD::FMINNAN : ISD::FMAXNAN;
10209 return DAG.getNode(Opcode, SDLoc(N), N->getValueType(0), LHS, RHS);
10212 /// PerformCMOVCombine - Target-specific DAG combining for ARMISD::CMOV.
10214 ARMTargetLowering::PerformCMOVCombine(SDNode *N, SelectionDAG &DAG) const {
10215 SDValue Cmp = N->getOperand(4);
10216 if (Cmp.getOpcode() != ARMISD::CMPZ)
10217 // Only looking at EQ and NE cases.
10220 EVT VT = N->getValueType(0);
10222 SDValue LHS = Cmp.getOperand(0);
10223 SDValue RHS = Cmp.getOperand(1);
10224 SDValue FalseVal = N->getOperand(0);
10225 SDValue TrueVal = N->getOperand(1);
10226 SDValue ARMcc = N->getOperand(2);
10227 ARMCC::CondCodes CC =
10228 (ARMCC::CondCodes)cast<ConstantSDNode>(ARMcc)->getZExtValue();
10246 /// FIXME: Turn this into a target neutral optimization?
10248 if (CC == ARMCC::NE && FalseVal == RHS && FalseVal != LHS) {
10249 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, TrueVal, ARMcc,
10250 N->getOperand(3), Cmp);
10251 } else if (CC == ARMCC::EQ && TrueVal == RHS) {
10253 SDValue NewCmp = getARMCmp(LHS, RHS, ISD::SETNE, ARMcc, DAG, dl);
10254 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, FalseVal, ARMcc,
10255 N->getOperand(3), NewCmp);
10258 if (Res.getNode()) {
10259 APInt KnownZero, KnownOne;
10260 DAG.computeKnownBits(SDValue(N,0), KnownZero, KnownOne);
10261 // Capture demanded bits information that would be otherwise lost.
10262 if (KnownZero == 0xfffffffe)
10263 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
10264 DAG.getValueType(MVT::i1));
10265 else if (KnownZero == 0xffffff00)
10266 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
10267 DAG.getValueType(MVT::i8));
10268 else if (KnownZero == 0xffff0000)
10269 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
10270 DAG.getValueType(MVT::i16));
10276 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
10277 DAGCombinerInfo &DCI) const {
10278 switch (N->getOpcode()) {
10280 case ISD::ADDC: return PerformADDCCombine(N, DCI, Subtarget);
10281 case ISD::ADD: return PerformADDCombine(N, DCI, Subtarget);
10282 case ISD::SUB: return PerformSUBCombine(N, DCI);
10283 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
10284 case ISD::OR: return PerformORCombine(N, DCI, Subtarget);
10285 case ISD::XOR: return PerformXORCombine(N, DCI, Subtarget);
10286 case ISD::AND: return PerformANDCombine(N, DCI, Subtarget);
10287 case ARMISD::BFI: return PerformBFICombine(N, DCI);
10288 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI, Subtarget);
10289 case ARMISD::VMOVDRR: return PerformVMOVDRRCombine(N, DCI.DAG);
10290 case ISD::STORE: return PerformSTORECombine(N, DCI);
10291 case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DCI, Subtarget);
10292 case ISD::INSERT_VECTOR_ELT: return PerformInsertEltCombine(N, DCI);
10293 case ISD::VECTOR_SHUFFLE: return PerformVECTOR_SHUFFLECombine(N, DCI.DAG);
10294 case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI);
10295 case ISD::FP_TO_SINT:
10296 case ISD::FP_TO_UINT: return PerformVCVTCombine(N, DCI, Subtarget);
10297 case ISD::FDIV: return PerformVDIVCombine(N, DCI, Subtarget);
10298 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
10301 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
10302 case ISD::SIGN_EXTEND:
10303 case ISD::ZERO_EXTEND:
10304 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
10305 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
10306 case ARMISD::CMOV: return PerformCMOVCombine(N, DCI.DAG);
10307 case ISD::LOAD: return PerformLOADCombine(N, DCI);
10308 case ARMISD::VLD2DUP:
10309 case ARMISD::VLD3DUP:
10310 case ARMISD::VLD4DUP:
10311 return PerformVLDCombine(N, DCI);
10312 case ARMISD::BUILD_VECTOR:
10313 return PerformARMBUILD_VECTORCombine(N, DCI);
10314 case ISD::INTRINSIC_VOID:
10315 case ISD::INTRINSIC_W_CHAIN:
10316 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
10317 case Intrinsic::arm_neon_vld1:
10318 case Intrinsic::arm_neon_vld2:
10319 case Intrinsic::arm_neon_vld3:
10320 case Intrinsic::arm_neon_vld4:
10321 case Intrinsic::arm_neon_vld2lane:
10322 case Intrinsic::arm_neon_vld3lane:
10323 case Intrinsic::arm_neon_vld4lane:
10324 case Intrinsic::arm_neon_vst1:
10325 case Intrinsic::arm_neon_vst2:
10326 case Intrinsic::arm_neon_vst3:
10327 case Intrinsic::arm_neon_vst4:
10328 case Intrinsic::arm_neon_vst2lane:
10329 case Intrinsic::arm_neon_vst3lane:
10330 case Intrinsic::arm_neon_vst4lane:
10331 return PerformVLDCombine(N, DCI);
10339 bool ARMTargetLowering::isDesirableToTransformToIntegerOp(unsigned Opc,
10341 return (VT == MVT::f32) && (Opc == ISD::LOAD || Opc == ISD::STORE);
10344 bool ARMTargetLowering::allowsMisalignedMemoryAccesses(EVT VT,
10347 bool *Fast) const {
10348 // The AllowsUnaliged flag models the SCTLR.A setting in ARM cpus
10349 bool AllowsUnaligned = Subtarget->allowsUnalignedMem();
10351 switch (VT.getSimpleVT().SimpleTy) {
10357 // Unaligned access can use (for example) LRDB, LRDH, LDR
10358 if (AllowsUnaligned) {
10360 *Fast = Subtarget->hasV7Ops();
10367 // For any little-endian targets with neon, we can support unaligned ld/st
10368 // of D and Q (e.g. {D0,D1}) registers by using vld1.i8/vst1.i8.
10369 // A big-endian target may also explicitly support unaligned accesses
10370 if (Subtarget->hasNEON() && (AllowsUnaligned || Subtarget->isLittle())) {
10380 static bool memOpAlign(unsigned DstAlign, unsigned SrcAlign,
10381 unsigned AlignCheck) {
10382 return ((SrcAlign == 0 || SrcAlign % AlignCheck == 0) &&
10383 (DstAlign == 0 || DstAlign % AlignCheck == 0));
10386 EVT ARMTargetLowering::getOptimalMemOpType(uint64_t Size,
10387 unsigned DstAlign, unsigned SrcAlign,
10388 bool IsMemset, bool ZeroMemset,
10390 MachineFunction &MF) const {
10391 const Function *F = MF.getFunction();
10393 // See if we can use NEON instructions for this...
10394 if ((!IsMemset || ZeroMemset) && Subtarget->hasNEON() &&
10395 !F->hasFnAttribute(Attribute::NoImplicitFloat)) {
10398 (memOpAlign(SrcAlign, DstAlign, 16) ||
10399 (allowsMisalignedMemoryAccesses(MVT::v2f64, 0, 1, &Fast) && Fast))) {
10401 } else if (Size >= 8 &&
10402 (memOpAlign(SrcAlign, DstAlign, 8) ||
10403 (allowsMisalignedMemoryAccesses(MVT::f64, 0, 1, &Fast) &&
10409 // Lowering to i32/i16 if the size permits.
10412 else if (Size >= 2)
10415 // Let the target-independent logic figure it out.
10419 bool ARMTargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
10420 if (Val.getOpcode() != ISD::LOAD)
10423 EVT VT1 = Val.getValueType();
10424 if (!VT1.isSimple() || !VT1.isInteger() ||
10425 !VT2.isSimple() || !VT2.isInteger())
10428 switch (VT1.getSimpleVT().SimpleTy) {
10433 // 8-bit and 16-bit loads implicitly zero-extend to 32-bits.
10440 bool ARMTargetLowering::isVectorLoadExtDesirable(SDValue ExtVal) const {
10441 EVT VT = ExtVal.getValueType();
10443 if (!isTypeLegal(VT))
10446 // Don't create a loadext if we can fold the extension into a wide/long
10448 // If there's more than one user instruction, the loadext is desirable no
10449 // matter what. There can be two uses by the same instruction.
10450 if (ExtVal->use_empty() ||
10451 !ExtVal->use_begin()->isOnlyUserOf(ExtVal.getNode()))
10454 SDNode *U = *ExtVal->use_begin();
10455 if ((U->getOpcode() == ISD::ADD || U->getOpcode() == ISD::SUB ||
10456 U->getOpcode() == ISD::SHL || U->getOpcode() == ARMISD::VSHL))
10462 bool ARMTargetLowering::allowTruncateForTailCall(Type *Ty1, Type *Ty2) const {
10463 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
10466 if (!isTypeLegal(EVT::getEVT(Ty1)))
10469 assert(Ty1->getPrimitiveSizeInBits() <= 64 && "i128 is probably not a noop");
10471 // Assuming the caller doesn't have a zeroext or signext return parameter,
10472 // truncation all the way down to i1 is valid.
10477 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
10481 unsigned Scale = 1;
10482 switch (VT.getSimpleVT().SimpleTy) {
10483 default: return false;
10498 if ((V & (Scale - 1)) != 0)
10501 return V == (V & ((1LL << 5) - 1));
10504 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
10505 const ARMSubtarget *Subtarget) {
10506 bool isNeg = false;
10512 switch (VT.getSimpleVT().SimpleTy) {
10513 default: return false;
10518 // + imm12 or - imm8
10520 return V == (V & ((1LL << 8) - 1));
10521 return V == (V & ((1LL << 12) - 1));
10524 // Same as ARM mode. FIXME: NEON?
10525 if (!Subtarget->hasVFP2())
10530 return V == (V & ((1LL << 8) - 1));
10534 /// isLegalAddressImmediate - Return true if the integer value can be used
10535 /// as the offset of the target addressing mode for load / store of the
10537 static bool isLegalAddressImmediate(int64_t V, EVT VT,
10538 const ARMSubtarget *Subtarget) {
10542 if (!VT.isSimple())
10545 if (Subtarget->isThumb1Only())
10546 return isLegalT1AddressImmediate(V, VT);
10547 else if (Subtarget->isThumb2())
10548 return isLegalT2AddressImmediate(V, VT, Subtarget);
10553 switch (VT.getSimpleVT().SimpleTy) {
10554 default: return false;
10559 return V == (V & ((1LL << 12) - 1));
10562 return V == (V & ((1LL << 8) - 1));
10565 if (!Subtarget->hasVFP2()) // FIXME: NEON?
10570 return V == (V & ((1LL << 8) - 1));
10574 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
10576 int Scale = AM.Scale;
10580 switch (VT.getSimpleVT().SimpleTy) {
10581 default: return false;
10589 Scale = Scale & ~1;
10590 return Scale == 2 || Scale == 4 || Scale == 8;
10593 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
10597 // Note, we allow "void" uses (basically, uses that aren't loads or
10598 // stores), because arm allows folding a scale into many arithmetic
10599 // operations. This should be made more precise and revisited later.
10601 // Allow r << imm, but the imm has to be a multiple of two.
10602 if (Scale & 1) return false;
10603 return isPowerOf2_32(Scale);
10607 /// isLegalAddressingMode - Return true if the addressing mode represented
10608 /// by AM is legal for this target, for a load/store of the specified type.
10609 bool ARMTargetLowering::isLegalAddressingMode(const DataLayout &DL,
10610 const AddrMode &AM, Type *Ty,
10611 unsigned AS) const {
10612 EVT VT = getValueType(DL, Ty, true);
10613 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
10616 // Can never fold addr of global into load/store.
10620 switch (AM.Scale) {
10621 case 0: // no scale reg, must be "r+i" or "r", or "i".
10624 if (Subtarget->isThumb1Only())
10628 // ARM doesn't support any R+R*scale+imm addr modes.
10632 if (!VT.isSimple())
10635 if (Subtarget->isThumb2())
10636 return isLegalT2ScaledAddressingMode(AM, VT);
10638 int Scale = AM.Scale;
10639 switch (VT.getSimpleVT().SimpleTy) {
10640 default: return false;
10644 if (Scale < 0) Scale = -Scale;
10648 return isPowerOf2_32(Scale & ~1);
10652 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
10657 // Note, we allow "void" uses (basically, uses that aren't loads or
10658 // stores), because arm allows folding a scale into many arithmetic
10659 // operations. This should be made more precise and revisited later.
10661 // Allow r << imm, but the imm has to be a multiple of two.
10662 if (Scale & 1) return false;
10663 return isPowerOf2_32(Scale);
10669 /// isLegalICmpImmediate - Return true if the specified immediate is legal
10670 /// icmp immediate, that is the target has icmp instructions which can compare
10671 /// a register against the immediate without having to materialize the
10672 /// immediate into a register.
10673 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
10674 // Thumb2 and ARM modes can use cmn for negative immediates.
10675 if (!Subtarget->isThumb())
10676 return ARM_AM::getSOImmVal(std::abs(Imm)) != -1;
10677 if (Subtarget->isThumb2())
10678 return ARM_AM::getT2SOImmVal(std::abs(Imm)) != -1;
10679 // Thumb1 doesn't have cmn, and only 8-bit immediates.
10680 return Imm >= 0 && Imm <= 255;
10683 /// isLegalAddImmediate - Return true if the specified immediate is a legal add
10684 /// *or sub* immediate, that is the target has add or sub instructions which can
10685 /// add a register with the immediate without having to materialize the
10686 /// immediate into a register.
10687 bool ARMTargetLowering::isLegalAddImmediate(int64_t Imm) const {
10688 // Same encoding for add/sub, just flip the sign.
10689 int64_t AbsImm = std::abs(Imm);
10690 if (!Subtarget->isThumb())
10691 return ARM_AM::getSOImmVal(AbsImm) != -1;
10692 if (Subtarget->isThumb2())
10693 return ARM_AM::getT2SOImmVal(AbsImm) != -1;
10694 // Thumb1 only has 8-bit unsigned immediate.
10695 return AbsImm >= 0 && AbsImm <= 255;
10698 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
10699 bool isSEXTLoad, SDValue &Base,
10700 SDValue &Offset, bool &isInc,
10701 SelectionDAG &DAG) {
10702 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
10705 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
10706 // AddressingMode 3
10707 Base = Ptr->getOperand(0);
10708 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
10709 int RHSC = (int)RHS->getZExtValue();
10710 if (RHSC < 0 && RHSC > -256) {
10711 assert(Ptr->getOpcode() == ISD::ADD);
10713 Offset = DAG.getConstant(-RHSC, SDLoc(Ptr), RHS->getValueType(0));
10717 isInc = (Ptr->getOpcode() == ISD::ADD);
10718 Offset = Ptr->getOperand(1);
10720 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
10721 // AddressingMode 2
10722 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
10723 int RHSC = (int)RHS->getZExtValue();
10724 if (RHSC < 0 && RHSC > -0x1000) {
10725 assert(Ptr->getOpcode() == ISD::ADD);
10727 Offset = DAG.getConstant(-RHSC, SDLoc(Ptr), RHS->getValueType(0));
10728 Base = Ptr->getOperand(0);
10733 if (Ptr->getOpcode() == ISD::ADD) {
10735 ARM_AM::ShiftOpc ShOpcVal=
10736 ARM_AM::getShiftOpcForNode(Ptr->getOperand(0).getOpcode());
10737 if (ShOpcVal != ARM_AM::no_shift) {
10738 Base = Ptr->getOperand(1);
10739 Offset = Ptr->getOperand(0);
10741 Base = Ptr->getOperand(0);
10742 Offset = Ptr->getOperand(1);
10747 isInc = (Ptr->getOpcode() == ISD::ADD);
10748 Base = Ptr->getOperand(0);
10749 Offset = Ptr->getOperand(1);
10753 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
10757 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
10758 bool isSEXTLoad, SDValue &Base,
10759 SDValue &Offset, bool &isInc,
10760 SelectionDAG &DAG) {
10761 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
10764 Base = Ptr->getOperand(0);
10765 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
10766 int RHSC = (int)RHS->getZExtValue();
10767 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
10768 assert(Ptr->getOpcode() == ISD::ADD);
10770 Offset = DAG.getConstant(-RHSC, SDLoc(Ptr), RHS->getValueType(0));
10772 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
10773 isInc = Ptr->getOpcode() == ISD::ADD;
10774 Offset = DAG.getConstant(RHSC, SDLoc(Ptr), RHS->getValueType(0));
10782 /// getPreIndexedAddressParts - returns true by value, base pointer and
10783 /// offset pointer and addressing mode by reference if the node's address
10784 /// can be legally represented as pre-indexed load / store address.
10786 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
10788 ISD::MemIndexedMode &AM,
10789 SelectionDAG &DAG) const {
10790 if (Subtarget->isThumb1Only())
10795 bool isSEXTLoad = false;
10796 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
10797 Ptr = LD->getBasePtr();
10798 VT = LD->getMemoryVT();
10799 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
10800 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
10801 Ptr = ST->getBasePtr();
10802 VT = ST->getMemoryVT();
10807 bool isLegal = false;
10808 if (Subtarget->isThumb2())
10809 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
10810 Offset, isInc, DAG);
10812 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
10813 Offset, isInc, DAG);
10817 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
10821 /// getPostIndexedAddressParts - returns true by value, base pointer and
10822 /// offset pointer and addressing mode by reference if this node can be
10823 /// combined with a load / store to form a post-indexed load / store.
10824 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
10827 ISD::MemIndexedMode &AM,
10828 SelectionDAG &DAG) const {
10829 if (Subtarget->isThumb1Only())
10834 bool isSEXTLoad = false;
10835 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
10836 VT = LD->getMemoryVT();
10837 Ptr = LD->getBasePtr();
10838 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
10839 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
10840 VT = ST->getMemoryVT();
10841 Ptr = ST->getBasePtr();
10846 bool isLegal = false;
10847 if (Subtarget->isThumb2())
10848 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
10851 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
10857 // Swap base ptr and offset to catch more post-index load / store when
10858 // it's legal. In Thumb2 mode, offset must be an immediate.
10859 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
10860 !Subtarget->isThumb2())
10861 std::swap(Base, Offset);
10863 // Post-indexed load / store update the base pointer.
10868 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
10872 void ARMTargetLowering::computeKnownBitsForTargetNode(const SDValue Op,
10875 const SelectionDAG &DAG,
10876 unsigned Depth) const {
10877 unsigned BitWidth = KnownOne.getBitWidth();
10878 KnownZero = KnownOne = APInt(BitWidth, 0);
10879 switch (Op.getOpcode()) {
10885 // These nodes' second result is a boolean
10886 if (Op.getResNo() == 0)
10888 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - 1);
10890 case ARMISD::CMOV: {
10891 // Bits are known zero/one if known on the LHS and RHS.
10892 DAG.computeKnownBits(Op.getOperand(0), KnownZero, KnownOne, Depth+1);
10893 if (KnownZero == 0 && KnownOne == 0) return;
10895 APInt KnownZeroRHS, KnownOneRHS;
10896 DAG.computeKnownBits(Op.getOperand(1), KnownZeroRHS, KnownOneRHS, Depth+1);
10897 KnownZero &= KnownZeroRHS;
10898 KnownOne &= KnownOneRHS;
10901 case ISD::INTRINSIC_W_CHAIN: {
10902 ConstantSDNode *CN = cast<ConstantSDNode>(Op->getOperand(1));
10903 Intrinsic::ID IntID = static_cast<Intrinsic::ID>(CN->getZExtValue());
10906 case Intrinsic::arm_ldaex:
10907 case Intrinsic::arm_ldrex: {
10908 EVT VT = cast<MemIntrinsicSDNode>(Op)->getMemoryVT();
10909 unsigned MemBits = VT.getScalarType().getSizeInBits();
10910 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - MemBits);
10918 //===----------------------------------------------------------------------===//
10919 // ARM Inline Assembly Support
10920 //===----------------------------------------------------------------------===//
10922 bool ARMTargetLowering::ExpandInlineAsm(CallInst *CI) const {
10923 // Looking for "rev" which is V6+.
10924 if (!Subtarget->hasV6Ops())
10927 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
10928 std::string AsmStr = IA->getAsmString();
10929 SmallVector<StringRef, 4> AsmPieces;
10930 SplitString(AsmStr, AsmPieces, ";\n");
10932 switch (AsmPieces.size()) {
10933 default: return false;
10935 AsmStr = AsmPieces[0];
10937 SplitString(AsmStr, AsmPieces, " \t,");
10940 if (AsmPieces.size() == 3 &&
10941 AsmPieces[0] == "rev" && AsmPieces[1] == "$0" && AsmPieces[2] == "$1" &&
10942 IA->getConstraintString().compare(0, 4, "=l,l") == 0) {
10943 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
10944 if (Ty && Ty->getBitWidth() == 32)
10945 return IntrinsicLowering::LowerToByteSwap(CI);
10953 /// getConstraintType - Given a constraint letter, return the type of
10954 /// constraint it is for this target.
10955 ARMTargetLowering::ConstraintType
10956 ARMTargetLowering::getConstraintType(StringRef Constraint) const {
10957 if (Constraint.size() == 1) {
10958 switch (Constraint[0]) {
10960 case 'l': return C_RegisterClass;
10961 case 'w': return C_RegisterClass;
10962 case 'h': return C_RegisterClass;
10963 case 'x': return C_RegisterClass;
10964 case 't': return C_RegisterClass;
10965 case 'j': return C_Other; // Constant for movw.
10966 // An address with a single base register. Due to the way we
10967 // currently handle addresses it is the same as an 'r' memory constraint.
10968 case 'Q': return C_Memory;
10970 } else if (Constraint.size() == 2) {
10971 switch (Constraint[0]) {
10973 // All 'U+' constraints are addresses.
10974 case 'U': return C_Memory;
10977 return TargetLowering::getConstraintType(Constraint);
10980 /// Examine constraint type and operand type and determine a weight value.
10981 /// This object must already have been set up with the operand type
10982 /// and the current alternative constraint selected.
10983 TargetLowering::ConstraintWeight
10984 ARMTargetLowering::getSingleConstraintMatchWeight(
10985 AsmOperandInfo &info, const char *constraint) const {
10986 ConstraintWeight weight = CW_Invalid;
10987 Value *CallOperandVal = info.CallOperandVal;
10988 // If we don't have a value, we can't do a match,
10989 // but allow it at the lowest weight.
10990 if (!CallOperandVal)
10992 Type *type = CallOperandVal->getType();
10993 // Look at the constraint type.
10994 switch (*constraint) {
10996 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
10999 if (type->isIntegerTy()) {
11000 if (Subtarget->isThumb())
11001 weight = CW_SpecificReg;
11003 weight = CW_Register;
11007 if (type->isFloatingPointTy())
11008 weight = CW_Register;
11014 typedef std::pair<unsigned, const TargetRegisterClass*> RCPair;
11015 RCPair ARMTargetLowering::getRegForInlineAsmConstraint(
11016 const TargetRegisterInfo *TRI, StringRef Constraint, MVT VT) const {
11017 if (Constraint.size() == 1) {
11018 // GCC ARM Constraint Letters
11019 switch (Constraint[0]) {
11020 case 'l': // Low regs or general regs.
11021 if (Subtarget->isThumb())
11022 return RCPair(0U, &ARM::tGPRRegClass);
11023 return RCPair(0U, &ARM::GPRRegClass);
11024 case 'h': // High regs or no regs.
11025 if (Subtarget->isThumb())
11026 return RCPair(0U, &ARM::hGPRRegClass);
11029 if (Subtarget->isThumb1Only())
11030 return RCPair(0U, &ARM::tGPRRegClass);
11031 return RCPair(0U, &ARM::GPRRegClass);
11033 if (VT == MVT::Other)
11035 if (VT == MVT::f32)
11036 return RCPair(0U, &ARM::SPRRegClass);
11037 if (VT.getSizeInBits() == 64)
11038 return RCPair(0U, &ARM::DPRRegClass);
11039 if (VT.getSizeInBits() == 128)
11040 return RCPair(0U, &ARM::QPRRegClass);
11043 if (VT == MVT::Other)
11045 if (VT == MVT::f32)
11046 return RCPair(0U, &ARM::SPR_8RegClass);
11047 if (VT.getSizeInBits() == 64)
11048 return RCPair(0U, &ARM::DPR_8RegClass);
11049 if (VT.getSizeInBits() == 128)
11050 return RCPair(0U, &ARM::QPR_8RegClass);
11053 if (VT == MVT::f32)
11054 return RCPair(0U, &ARM::SPRRegClass);
11058 if (StringRef("{cc}").equals_lower(Constraint))
11059 return std::make_pair(unsigned(ARM::CPSR), &ARM::CCRRegClass);
11061 return TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
11064 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
11065 /// vector. If it is invalid, don't add anything to Ops.
11066 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
11067 std::string &Constraint,
11068 std::vector<SDValue>&Ops,
11069 SelectionDAG &DAG) const {
11072 // Currently only support length 1 constraints.
11073 if (Constraint.length() != 1) return;
11075 char ConstraintLetter = Constraint[0];
11076 switch (ConstraintLetter) {
11079 case 'I': case 'J': case 'K': case 'L':
11080 case 'M': case 'N': case 'O':
11081 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
11085 int64_t CVal64 = C->getSExtValue();
11086 int CVal = (int) CVal64;
11087 // None of these constraints allow values larger than 32 bits. Check
11088 // that the value fits in an int.
11089 if (CVal != CVal64)
11092 switch (ConstraintLetter) {
11094 // Constant suitable for movw, must be between 0 and
11096 if (Subtarget->hasV6T2Ops())
11097 if (CVal >= 0 && CVal <= 65535)
11101 if (Subtarget->isThumb1Only()) {
11102 // This must be a constant between 0 and 255, for ADD
11104 if (CVal >= 0 && CVal <= 255)
11106 } else if (Subtarget->isThumb2()) {
11107 // A constant that can be used as an immediate value in a
11108 // data-processing instruction.
11109 if (ARM_AM::getT2SOImmVal(CVal) != -1)
11112 // A constant that can be used as an immediate value in a
11113 // data-processing instruction.
11114 if (ARM_AM::getSOImmVal(CVal) != -1)
11120 if (Subtarget->isThumb()) { // FIXME thumb2
11121 // This must be a constant between -255 and -1, for negated ADD
11122 // immediates. This can be used in GCC with an "n" modifier that
11123 // prints the negated value, for use with SUB instructions. It is
11124 // not useful otherwise but is implemented for compatibility.
11125 if (CVal >= -255 && CVal <= -1)
11128 // This must be a constant between -4095 and 4095. It is not clear
11129 // what this constraint is intended for. Implemented for
11130 // compatibility with GCC.
11131 if (CVal >= -4095 && CVal <= 4095)
11137 if (Subtarget->isThumb1Only()) {
11138 // A 32-bit value where only one byte has a nonzero value. Exclude
11139 // zero to match GCC. This constraint is used by GCC internally for
11140 // constants that can be loaded with a move/shift combination.
11141 // It is not useful otherwise but is implemented for compatibility.
11142 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
11144 } else if (Subtarget->isThumb2()) {
11145 // A constant whose bitwise inverse can be used as an immediate
11146 // value in a data-processing instruction. This can be used in GCC
11147 // with a "B" modifier that prints the inverted value, for use with
11148 // BIC and MVN instructions. It is not useful otherwise but is
11149 // implemented for compatibility.
11150 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
11153 // A constant whose bitwise inverse can be used as an immediate
11154 // value in a data-processing instruction. This can be used in GCC
11155 // with a "B" modifier that prints the inverted value, for use with
11156 // BIC and MVN instructions. It is not useful otherwise but is
11157 // implemented for compatibility.
11158 if (ARM_AM::getSOImmVal(~CVal) != -1)
11164 if (Subtarget->isThumb1Only()) {
11165 // This must be a constant between -7 and 7,
11166 // for 3-operand ADD/SUB immediate instructions.
11167 if (CVal >= -7 && CVal < 7)
11169 } else if (Subtarget->isThumb2()) {
11170 // A constant whose negation can be used as an immediate value in a
11171 // data-processing instruction. This can be used in GCC with an "n"
11172 // modifier that prints the negated value, for use with SUB
11173 // instructions. It is not useful otherwise but is implemented for
11175 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
11178 // A constant whose negation can be used as an immediate value in a
11179 // data-processing instruction. This can be used in GCC with an "n"
11180 // modifier that prints the negated value, for use with SUB
11181 // instructions. It is not useful otherwise but is implemented for
11183 if (ARM_AM::getSOImmVal(-CVal) != -1)
11189 if (Subtarget->isThumb()) { // FIXME thumb2
11190 // This must be a multiple of 4 between 0 and 1020, for
11191 // ADD sp + immediate.
11192 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
11195 // A power of two or a constant between 0 and 32. This is used in
11196 // GCC for the shift amount on shifted register operands, but it is
11197 // useful in general for any shift amounts.
11198 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
11204 if (Subtarget->isThumb()) { // FIXME thumb2
11205 // This must be a constant between 0 and 31, for shift amounts.
11206 if (CVal >= 0 && CVal <= 31)
11212 if (Subtarget->isThumb()) { // FIXME thumb2
11213 // This must be a multiple of 4 between -508 and 508, for
11214 // ADD/SUB sp = sp + immediate.
11215 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
11220 Result = DAG.getTargetConstant(CVal, SDLoc(Op), Op.getValueType());
11224 if (Result.getNode()) {
11225 Ops.push_back(Result);
11228 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
11231 SDValue ARMTargetLowering::LowerDivRem(SDValue Op, SelectionDAG &DAG) const {
11232 assert((Subtarget->isTargetAEABI() || Subtarget->isTargetAndroid()) &&
11233 "Register-based DivRem lowering only");
11234 unsigned Opcode = Op->getOpcode();
11235 assert((Opcode == ISD::SDIVREM || Opcode == ISD::UDIVREM) &&
11236 "Invalid opcode for Div/Rem lowering");
11237 bool isSigned = (Opcode == ISD::SDIVREM);
11238 EVT VT = Op->getValueType(0);
11239 Type *Ty = VT.getTypeForEVT(*DAG.getContext());
11242 switch (VT.getSimpleVT().SimpleTy) {
11243 default: llvm_unreachable("Unexpected request for libcall!");
11244 case MVT::i8: LC = isSigned ? RTLIB::SDIVREM_I8 : RTLIB::UDIVREM_I8; break;
11245 case MVT::i16: LC = isSigned ? RTLIB::SDIVREM_I16 : RTLIB::UDIVREM_I16; break;
11246 case MVT::i32: LC = isSigned ? RTLIB::SDIVREM_I32 : RTLIB::UDIVREM_I32; break;
11247 case MVT::i64: LC = isSigned ? RTLIB::SDIVREM_I64 : RTLIB::UDIVREM_I64; break;
11250 SDValue InChain = DAG.getEntryNode();
11252 TargetLowering::ArgListTy Args;
11253 TargetLowering::ArgListEntry Entry;
11254 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i) {
11255 EVT ArgVT = Op->getOperand(i).getValueType();
11256 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
11257 Entry.Node = Op->getOperand(i);
11259 Entry.isSExt = isSigned;
11260 Entry.isZExt = !isSigned;
11261 Args.push_back(Entry);
11264 SDValue Callee = DAG.getExternalSymbol(getLibcallName(LC),
11265 getPointerTy(DAG.getDataLayout()));
11267 Type *RetTy = (Type*)StructType::get(Ty, Ty, nullptr);
11270 TargetLowering::CallLoweringInfo CLI(DAG);
11271 CLI.setDebugLoc(dl).setChain(InChain)
11272 .setCallee(getLibcallCallingConv(LC), RetTy, Callee, std::move(Args), 0)
11273 .setInRegister().setSExtResult(isSigned).setZExtResult(!isSigned);
11275 std::pair<SDValue, SDValue> CallInfo = LowerCallTo(CLI);
11276 return CallInfo.first;
11280 ARMTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const {
11281 assert(Subtarget->isTargetWindows() && "unsupported target platform");
11285 SDValue Chain = Op.getOperand(0);
11286 SDValue Size = Op.getOperand(1);
11288 SDValue Words = DAG.getNode(ISD::SRL, DL, MVT::i32, Size,
11289 DAG.getConstant(2, DL, MVT::i32));
11292 Chain = DAG.getCopyToReg(Chain, DL, ARM::R4, Words, Flag);
11293 Flag = Chain.getValue(1);
11295 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
11296 Chain = DAG.getNode(ARMISD::WIN__CHKSTK, DL, NodeTys, Chain, Flag);
11298 SDValue NewSP = DAG.getCopyFromReg(Chain, DL, ARM::SP, MVT::i32);
11299 Chain = NewSP.getValue(1);
11301 SDValue Ops[2] = { NewSP, Chain };
11302 return DAG.getMergeValues(Ops, DL);
11305 SDValue ARMTargetLowering::LowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) const {
11306 assert(Op.getValueType() == MVT::f64 && Subtarget->isFPOnlySP() &&
11307 "Unexpected type for custom-lowering FP_EXTEND");
11310 LC = RTLIB::getFPEXT(Op.getOperand(0).getValueType(), Op.getValueType());
11312 SDValue SrcVal = Op.getOperand(0);
11313 return makeLibCall(DAG, LC, Op.getValueType(), &SrcVal, 1,
11314 /*isSigned*/ false, SDLoc(Op)).first;
11317 SDValue ARMTargetLowering::LowerFP_ROUND(SDValue Op, SelectionDAG &DAG) const {
11318 assert(Op.getOperand(0).getValueType() == MVT::f64 &&
11319 Subtarget->isFPOnlySP() &&
11320 "Unexpected type for custom-lowering FP_ROUND");
11323 LC = RTLIB::getFPROUND(Op.getOperand(0).getValueType(), Op.getValueType());
11325 SDValue SrcVal = Op.getOperand(0);
11326 return makeLibCall(DAG, LC, Op.getValueType(), &SrcVal, 1,
11327 /*isSigned*/ false, SDLoc(Op)).first;
11331 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
11332 // The ARM target isn't yet aware of offsets.
11336 bool ARM::isBitFieldInvertedMask(unsigned v) {
11337 if (v == 0xffffffff)
11340 // there can be 1's on either or both "outsides", all the "inside"
11341 // bits must be 0's
11342 return isShiftedMask_32(~v);
11345 /// isFPImmLegal - Returns true if the target can instruction select the
11346 /// specified FP immediate natively. If false, the legalizer will
11347 /// materialize the FP immediate as a load from a constant pool.
11348 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
11349 if (!Subtarget->hasVFP3())
11351 if (VT == MVT::f32)
11352 return ARM_AM::getFP32Imm(Imm) != -1;
11353 if (VT == MVT::f64 && !Subtarget->isFPOnlySP())
11354 return ARM_AM::getFP64Imm(Imm) != -1;
11358 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
11359 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
11360 /// specified in the intrinsic calls.
11361 bool ARMTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
11363 unsigned Intrinsic) const {
11364 switch (Intrinsic) {
11365 case Intrinsic::arm_neon_vld1:
11366 case Intrinsic::arm_neon_vld2:
11367 case Intrinsic::arm_neon_vld3:
11368 case Intrinsic::arm_neon_vld4:
11369 case Intrinsic::arm_neon_vld2lane:
11370 case Intrinsic::arm_neon_vld3lane:
11371 case Intrinsic::arm_neon_vld4lane: {
11372 Info.opc = ISD::INTRINSIC_W_CHAIN;
11373 // Conservatively set memVT to the entire set of vectors loaded.
11374 auto &DL = I.getCalledFunction()->getParent()->getDataLayout();
11375 uint64_t NumElts = DL.getTypeAllocSize(I.getType()) / 8;
11376 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
11377 Info.ptrVal = I.getArgOperand(0);
11379 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
11380 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
11381 Info.vol = false; // volatile loads with NEON intrinsics not supported
11382 Info.readMem = true;
11383 Info.writeMem = false;
11386 case Intrinsic::arm_neon_vst1:
11387 case Intrinsic::arm_neon_vst2:
11388 case Intrinsic::arm_neon_vst3:
11389 case Intrinsic::arm_neon_vst4:
11390 case Intrinsic::arm_neon_vst2lane:
11391 case Intrinsic::arm_neon_vst3lane:
11392 case Intrinsic::arm_neon_vst4lane: {
11393 Info.opc = ISD::INTRINSIC_VOID;
11394 // Conservatively set memVT to the entire set of vectors stored.
11395 auto &DL = I.getCalledFunction()->getParent()->getDataLayout();
11396 unsigned NumElts = 0;
11397 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
11398 Type *ArgTy = I.getArgOperand(ArgI)->getType();
11399 if (!ArgTy->isVectorTy())
11401 NumElts += DL.getTypeAllocSize(ArgTy) / 8;
11403 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
11404 Info.ptrVal = I.getArgOperand(0);
11406 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
11407 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
11408 Info.vol = false; // volatile stores with NEON intrinsics not supported
11409 Info.readMem = false;
11410 Info.writeMem = true;
11413 case Intrinsic::arm_ldaex:
11414 case Intrinsic::arm_ldrex: {
11415 auto &DL = I.getCalledFunction()->getParent()->getDataLayout();
11416 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(0)->getType());
11417 Info.opc = ISD::INTRINSIC_W_CHAIN;
11418 Info.memVT = MVT::getVT(PtrTy->getElementType());
11419 Info.ptrVal = I.getArgOperand(0);
11421 Info.align = DL.getABITypeAlignment(PtrTy->getElementType());
11423 Info.readMem = true;
11424 Info.writeMem = false;
11427 case Intrinsic::arm_stlex:
11428 case Intrinsic::arm_strex: {
11429 auto &DL = I.getCalledFunction()->getParent()->getDataLayout();
11430 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(1)->getType());
11431 Info.opc = ISD::INTRINSIC_W_CHAIN;
11432 Info.memVT = MVT::getVT(PtrTy->getElementType());
11433 Info.ptrVal = I.getArgOperand(1);
11435 Info.align = DL.getABITypeAlignment(PtrTy->getElementType());
11437 Info.readMem = false;
11438 Info.writeMem = true;
11441 case Intrinsic::arm_stlexd:
11442 case Intrinsic::arm_strexd: {
11443 Info.opc = ISD::INTRINSIC_W_CHAIN;
11444 Info.memVT = MVT::i64;
11445 Info.ptrVal = I.getArgOperand(2);
11449 Info.readMem = false;
11450 Info.writeMem = true;
11453 case Intrinsic::arm_ldaexd:
11454 case Intrinsic::arm_ldrexd: {
11455 Info.opc = ISD::INTRINSIC_W_CHAIN;
11456 Info.memVT = MVT::i64;
11457 Info.ptrVal = I.getArgOperand(0);
11461 Info.readMem = true;
11462 Info.writeMem = false;
11472 /// \brief Returns true if it is beneficial to convert a load of a constant
11473 /// to just the constant itself.
11474 bool ARMTargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
11476 assert(Ty->isIntegerTy());
11478 unsigned Bits = Ty->getPrimitiveSizeInBits();
11479 if (Bits == 0 || Bits > 32)
11484 bool ARMTargetLowering::hasLoadLinkedStoreConditional() const { return true; }
11486 Instruction* ARMTargetLowering::makeDMB(IRBuilder<> &Builder,
11487 ARM_MB::MemBOpt Domain) const {
11488 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
11490 // First, if the target has no DMB, see what fallback we can use.
11491 if (!Subtarget->hasDataBarrier()) {
11492 // Some ARMv6 cpus can support data barriers with an mcr instruction.
11493 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
11495 if (Subtarget->hasV6Ops() && !Subtarget->isThumb()) {
11496 Function *MCR = llvm::Intrinsic::getDeclaration(M, Intrinsic::arm_mcr);
11497 Value* args[6] = {Builder.getInt32(15), Builder.getInt32(0),
11498 Builder.getInt32(0), Builder.getInt32(7),
11499 Builder.getInt32(10), Builder.getInt32(5)};
11500 return Builder.CreateCall(MCR, args);
11502 // Instead of using barriers, atomic accesses on these subtargets use
11504 llvm_unreachable("makeDMB on a target so old that it has no barriers");
11507 Function *DMB = llvm::Intrinsic::getDeclaration(M, Intrinsic::arm_dmb);
11508 // Only a full system barrier exists in the M-class architectures.
11509 Domain = Subtarget->isMClass() ? ARM_MB::SY : Domain;
11510 Constant *CDomain = Builder.getInt32(Domain);
11511 return Builder.CreateCall(DMB, CDomain);
11515 // Based on http://www.cl.cam.ac.uk/~pes20/cpp/cpp0xmappings.html
11516 Instruction* ARMTargetLowering::emitLeadingFence(IRBuilder<> &Builder,
11517 AtomicOrdering Ord, bool IsStore,
11518 bool IsLoad) const {
11519 if (!getInsertFencesForAtomic())
11525 llvm_unreachable("Invalid fence: unordered/non-atomic");
11528 return nullptr; // Nothing to do
11529 case SequentiallyConsistent:
11531 return nullptr; // Nothing to do
11534 case AcquireRelease:
11535 if (Subtarget->isSwift())
11536 return makeDMB(Builder, ARM_MB::ISHST);
11537 // FIXME: add a comment with a link to documentation justifying this.
11539 return makeDMB(Builder, ARM_MB::ISH);
11541 llvm_unreachable("Unknown fence ordering in emitLeadingFence");
11544 Instruction* ARMTargetLowering::emitTrailingFence(IRBuilder<> &Builder,
11545 AtomicOrdering Ord, bool IsStore,
11546 bool IsLoad) const {
11547 if (!getInsertFencesForAtomic())
11553 llvm_unreachable("Invalid fence: unordered/not-atomic");
11556 return nullptr; // Nothing to do
11558 case AcquireRelease:
11559 case SequentiallyConsistent:
11560 return makeDMB(Builder, ARM_MB::ISH);
11562 llvm_unreachable("Unknown fence ordering in emitTrailingFence");
11565 // Loads and stores less than 64-bits are already atomic; ones above that
11566 // are doomed anyway, so defer to the default libcall and blame the OS when
11567 // things go wrong. Cortex M doesn't have ldrexd/strexd though, so don't emit
11568 // anything for those.
11569 bool ARMTargetLowering::shouldExpandAtomicStoreInIR(StoreInst *SI) const {
11570 unsigned Size = SI->getValueOperand()->getType()->getPrimitiveSizeInBits();
11571 return (Size == 64) && !Subtarget->isMClass();
11574 // Loads and stores less than 64-bits are already atomic; ones above that
11575 // are doomed anyway, so defer to the default libcall and blame the OS when
11576 // things go wrong. Cortex M doesn't have ldrexd/strexd though, so don't emit
11577 // anything for those.
11578 // FIXME: ldrd and strd are atomic if the CPU has LPAE (e.g. A15 has that
11579 // guarantee, see DDI0406C ARM architecture reference manual,
11580 // sections A8.8.72-74 LDRD)
11581 bool ARMTargetLowering::shouldExpandAtomicLoadInIR(LoadInst *LI) const {
11582 unsigned Size = LI->getType()->getPrimitiveSizeInBits();
11583 return (Size == 64) && !Subtarget->isMClass();
11586 // For the real atomic operations, we have ldrex/strex up to 32 bits,
11587 // and up to 64 bits on the non-M profiles
11588 TargetLoweringBase::AtomicRMWExpansionKind
11589 ARMTargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
11590 unsigned Size = AI->getType()->getPrimitiveSizeInBits();
11591 return (Size <= (Subtarget->isMClass() ? 32U : 64U))
11592 ? AtomicRMWExpansionKind::LLSC
11593 : AtomicRMWExpansionKind::None;
11596 // This has so far only been implemented for MachO.
11597 bool ARMTargetLowering::useLoadStackGuardNode() const {
11598 return Subtarget->isTargetMachO();
11601 bool ARMTargetLowering::canCombineStoreAndExtract(Type *VectorTy, Value *Idx,
11602 unsigned &Cost) const {
11603 // If we do not have NEON, vector types are not natively supported.
11604 if (!Subtarget->hasNEON())
11607 // Floating point values and vector values map to the same register file.
11608 // Therefore, although we could do a store extract of a vector type, this is
11609 // better to leave at float as we have more freedom in the addressing mode for
11611 if (VectorTy->isFPOrFPVectorTy())
11614 // If the index is unknown at compile time, this is very expensive to lower
11615 // and it is not possible to combine the store with the extract.
11616 if (!isa<ConstantInt>(Idx))
11619 assert(VectorTy->isVectorTy() && "VectorTy is not a vector type");
11620 unsigned BitWidth = cast<VectorType>(VectorTy)->getBitWidth();
11621 // We can do a store + vector extract on any vector that fits perfectly in a D
11623 if (BitWidth == 64 || BitWidth == 128) {
11630 Value *ARMTargetLowering::emitLoadLinked(IRBuilder<> &Builder, Value *Addr,
11631 AtomicOrdering Ord) const {
11632 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
11633 Type *ValTy = cast<PointerType>(Addr->getType())->getElementType();
11634 bool IsAcquire = isAtLeastAcquire(Ord);
11636 // Since i64 isn't legal and intrinsics don't get type-lowered, the ldrexd
11637 // intrinsic must return {i32, i32} and we have to recombine them into a
11638 // single i64 here.
11639 if (ValTy->getPrimitiveSizeInBits() == 64) {
11640 Intrinsic::ID Int =
11641 IsAcquire ? Intrinsic::arm_ldaexd : Intrinsic::arm_ldrexd;
11642 Function *Ldrex = llvm::Intrinsic::getDeclaration(M, Int);
11644 Addr = Builder.CreateBitCast(Addr, Type::getInt8PtrTy(M->getContext()));
11645 Value *LoHi = Builder.CreateCall(Ldrex, Addr, "lohi");
11647 Value *Lo = Builder.CreateExtractValue(LoHi, 0, "lo");
11648 Value *Hi = Builder.CreateExtractValue(LoHi, 1, "hi");
11649 if (!Subtarget->isLittle())
11650 std::swap (Lo, Hi);
11651 Lo = Builder.CreateZExt(Lo, ValTy, "lo64");
11652 Hi = Builder.CreateZExt(Hi, ValTy, "hi64");
11653 return Builder.CreateOr(
11654 Lo, Builder.CreateShl(Hi, ConstantInt::get(ValTy, 32)), "val64");
11657 Type *Tys[] = { Addr->getType() };
11658 Intrinsic::ID Int = IsAcquire ? Intrinsic::arm_ldaex : Intrinsic::arm_ldrex;
11659 Function *Ldrex = llvm::Intrinsic::getDeclaration(M, Int, Tys);
11661 return Builder.CreateTruncOrBitCast(
11662 Builder.CreateCall(Ldrex, Addr),
11663 cast<PointerType>(Addr->getType())->getElementType());
11666 Value *ARMTargetLowering::emitStoreConditional(IRBuilder<> &Builder, Value *Val,
11668 AtomicOrdering Ord) const {
11669 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
11670 bool IsRelease = isAtLeastRelease(Ord);
11672 // Since the intrinsics must have legal type, the i64 intrinsics take two
11673 // parameters: "i32, i32". We must marshal Val into the appropriate form
11674 // before the call.
11675 if (Val->getType()->getPrimitiveSizeInBits() == 64) {
11676 Intrinsic::ID Int =
11677 IsRelease ? Intrinsic::arm_stlexd : Intrinsic::arm_strexd;
11678 Function *Strex = Intrinsic::getDeclaration(M, Int);
11679 Type *Int32Ty = Type::getInt32Ty(M->getContext());
11681 Value *Lo = Builder.CreateTrunc(Val, Int32Ty, "lo");
11682 Value *Hi = Builder.CreateTrunc(Builder.CreateLShr(Val, 32), Int32Ty, "hi");
11683 if (!Subtarget->isLittle())
11684 std::swap (Lo, Hi);
11685 Addr = Builder.CreateBitCast(Addr, Type::getInt8PtrTy(M->getContext()));
11686 return Builder.CreateCall(Strex, {Lo, Hi, Addr});
11689 Intrinsic::ID Int = IsRelease ? Intrinsic::arm_stlex : Intrinsic::arm_strex;
11690 Type *Tys[] = { Addr->getType() };
11691 Function *Strex = Intrinsic::getDeclaration(M, Int, Tys);
11693 return Builder.CreateCall(
11694 Strex, {Builder.CreateZExtOrBitCast(
11695 Val, Strex->getFunctionType()->getParamType(0)),
11699 /// \brief Lower an interleaved load into a vldN intrinsic.
11701 /// E.g. Lower an interleaved load (Factor = 2):
11702 /// %wide.vec = load <8 x i32>, <8 x i32>* %ptr, align 4
11703 /// %v0 = shuffle %wide.vec, undef, <0, 2, 4, 6> ; Extract even elements
11704 /// %v1 = shuffle %wide.vec, undef, <1, 3, 5, 7> ; Extract odd elements
11707 /// %vld2 = { <4 x i32>, <4 x i32> } call llvm.arm.neon.vld2(%ptr, 4)
11708 /// %vec0 = extractelement { <4 x i32>, <4 x i32> } %vld2, i32 0
11709 /// %vec1 = extractelement { <4 x i32>, <4 x i32> } %vld2, i32 1
11710 bool ARMTargetLowering::lowerInterleavedLoad(
11711 LoadInst *LI, ArrayRef<ShuffleVectorInst *> Shuffles,
11712 ArrayRef<unsigned> Indices, unsigned Factor) const {
11713 assert(Factor >= 2 && Factor <= getMaxSupportedInterleaveFactor() &&
11714 "Invalid interleave factor");
11715 assert(!Shuffles.empty() && "Empty shufflevector input");
11716 assert(Shuffles.size() == Indices.size() &&
11717 "Unmatched number of shufflevectors and indices");
11719 VectorType *VecTy = Shuffles[0]->getType();
11720 Type *EltTy = VecTy->getVectorElementType();
11722 const DataLayout &DL = LI->getModule()->getDataLayout();
11723 unsigned VecSize = DL.getTypeAllocSizeInBits(VecTy);
11724 bool EltIs64Bits = DL.getTypeAllocSizeInBits(EltTy) == 64;
11726 // Skip illegal vector types and vector types of i64/f64 element (vldN doesn't
11727 // support i64/f64 element).
11728 if ((VecSize != 64 && VecSize != 128) || EltIs64Bits)
11731 // A pointer vector can not be the return type of the ldN intrinsics. Need to
11732 // load integer vectors first and then convert to pointer vectors.
11733 if (EltTy->isPointerTy())
11735 VectorType::get(DL.getIntPtrType(EltTy), VecTy->getVectorNumElements());
11737 static const Intrinsic::ID LoadInts[3] = {Intrinsic::arm_neon_vld2,
11738 Intrinsic::arm_neon_vld3,
11739 Intrinsic::arm_neon_vld4};
11741 Function *VldnFunc =
11742 Intrinsic::getDeclaration(LI->getModule(), LoadInts[Factor - 2], VecTy);
11744 IRBuilder<> Builder(LI);
11745 SmallVector<Value *, 2> Ops;
11747 Type *Int8Ptr = Builder.getInt8PtrTy(LI->getPointerAddressSpace());
11748 Ops.push_back(Builder.CreateBitCast(LI->getPointerOperand(), Int8Ptr));
11749 Ops.push_back(Builder.getInt32(LI->getAlignment()));
11751 CallInst *VldN = Builder.CreateCall(VldnFunc, Ops, "vldN");
11753 // Replace uses of each shufflevector with the corresponding vector loaded
11755 for (unsigned i = 0; i < Shuffles.size(); i++) {
11756 ShuffleVectorInst *SV = Shuffles[i];
11757 unsigned Index = Indices[i];
11759 Value *SubVec = Builder.CreateExtractValue(VldN, Index);
11761 // Convert the integer vector to pointer vector if the element is pointer.
11762 if (EltTy->isPointerTy())
11763 SubVec = Builder.CreateIntToPtr(SubVec, SV->getType());
11765 SV->replaceAllUsesWith(SubVec);
11771 /// \brief Get a mask consisting of sequential integers starting from \p Start.
11773 /// I.e. <Start, Start + 1, ..., Start + NumElts - 1>
11774 static Constant *getSequentialMask(IRBuilder<> &Builder, unsigned Start,
11775 unsigned NumElts) {
11776 SmallVector<Constant *, 16> Mask;
11777 for (unsigned i = 0; i < NumElts; i++)
11778 Mask.push_back(Builder.getInt32(Start + i));
11780 return ConstantVector::get(Mask);
11783 /// \brief Lower an interleaved store into a vstN intrinsic.
11785 /// E.g. Lower an interleaved store (Factor = 3):
11786 /// %i.vec = shuffle <8 x i32> %v0, <8 x i32> %v1,
11787 /// <0, 4, 8, 1, 5, 9, 2, 6, 10, 3, 7, 11>
11788 /// store <12 x i32> %i.vec, <12 x i32>* %ptr, align 4
11791 /// %sub.v0 = shuffle <8 x i32> %v0, <8 x i32> v1, <0, 1, 2, 3>
11792 /// %sub.v1 = shuffle <8 x i32> %v0, <8 x i32> v1, <4, 5, 6, 7>
11793 /// %sub.v2 = shuffle <8 x i32> %v0, <8 x i32> v1, <8, 9, 10, 11>
11794 /// call void llvm.arm.neon.vst3(%ptr, %sub.v0, %sub.v1, %sub.v2, 4)
11796 /// Note that the new shufflevectors will be removed and we'll only generate one
11797 /// vst3 instruction in CodeGen.
11798 bool ARMTargetLowering::lowerInterleavedStore(StoreInst *SI,
11799 ShuffleVectorInst *SVI,
11800 unsigned Factor) const {
11801 assert(Factor >= 2 && Factor <= getMaxSupportedInterleaveFactor() &&
11802 "Invalid interleave factor");
11804 VectorType *VecTy = SVI->getType();
11805 assert(VecTy->getVectorNumElements() % Factor == 0 &&
11806 "Invalid interleaved store");
11808 unsigned NumSubElts = VecTy->getVectorNumElements() / Factor;
11809 Type *EltTy = VecTy->getVectorElementType();
11810 VectorType *SubVecTy = VectorType::get(EltTy, NumSubElts);
11812 const DataLayout &DL = SI->getModule()->getDataLayout();
11813 unsigned SubVecSize = DL.getTypeAllocSizeInBits(SubVecTy);
11814 bool EltIs64Bits = DL.getTypeAllocSizeInBits(EltTy) == 64;
11816 // Skip illegal sub vector types and vector types of i64/f64 element (vstN
11817 // doesn't support i64/f64 element).
11818 if ((SubVecSize != 64 && SubVecSize != 128) || EltIs64Bits)
11821 Value *Op0 = SVI->getOperand(0);
11822 Value *Op1 = SVI->getOperand(1);
11823 IRBuilder<> Builder(SI);
11825 // StN intrinsics don't support pointer vectors as arguments. Convert pointer
11826 // vectors to integer vectors.
11827 if (EltTy->isPointerTy()) {
11828 Type *IntTy = DL.getIntPtrType(EltTy);
11830 // Convert to the corresponding integer vector.
11832 VectorType::get(IntTy, Op0->getType()->getVectorNumElements());
11833 Op0 = Builder.CreatePtrToInt(Op0, IntVecTy);
11834 Op1 = Builder.CreatePtrToInt(Op1, IntVecTy);
11836 SubVecTy = VectorType::get(IntTy, NumSubElts);
11839 static Intrinsic::ID StoreInts[3] = {Intrinsic::arm_neon_vst2,
11840 Intrinsic::arm_neon_vst3,
11841 Intrinsic::arm_neon_vst4};
11842 Function *VstNFunc = Intrinsic::getDeclaration(
11843 SI->getModule(), StoreInts[Factor - 2], SubVecTy);
11845 SmallVector<Value *, 6> Ops;
11847 Type *Int8Ptr = Builder.getInt8PtrTy(SI->getPointerAddressSpace());
11848 Ops.push_back(Builder.CreateBitCast(SI->getPointerOperand(), Int8Ptr));
11850 // Split the shufflevector operands into sub vectors for the new vstN call.
11851 for (unsigned i = 0; i < Factor; i++)
11852 Ops.push_back(Builder.CreateShuffleVector(
11853 Op0, Op1, getSequentialMask(Builder, NumSubElts * i, NumSubElts)));
11855 Ops.push_back(Builder.getInt32(SI->getAlignment()));
11856 Builder.CreateCall(VstNFunc, Ops);
11868 static bool isHomogeneousAggregate(Type *Ty, HABaseType &Base,
11869 uint64_t &Members) {
11870 if (auto *ST = dyn_cast<StructType>(Ty)) {
11871 for (unsigned i = 0; i < ST->getNumElements(); ++i) {
11872 uint64_t SubMembers = 0;
11873 if (!isHomogeneousAggregate(ST->getElementType(i), Base, SubMembers))
11875 Members += SubMembers;
11877 } else if (auto *AT = dyn_cast<ArrayType>(Ty)) {
11878 uint64_t SubMembers = 0;
11879 if (!isHomogeneousAggregate(AT->getElementType(), Base, SubMembers))
11881 Members += SubMembers * AT->getNumElements();
11882 } else if (Ty->isFloatTy()) {
11883 if (Base != HA_UNKNOWN && Base != HA_FLOAT)
11887 } else if (Ty->isDoubleTy()) {
11888 if (Base != HA_UNKNOWN && Base != HA_DOUBLE)
11892 } else if (auto *VT = dyn_cast<VectorType>(Ty)) {
11899 return VT->getBitWidth() == 64;
11901 return VT->getBitWidth() == 128;
11903 switch (VT->getBitWidth()) {
11916 return (Members > 0 && Members <= 4);
11919 /// \brief Return true if a type is an AAPCS-VFP homogeneous aggregate or one of
11920 /// [N x i32] or [N x i64]. This allows front-ends to skip emitting padding when
11921 /// passing according to AAPCS rules.
11922 bool ARMTargetLowering::functionArgumentNeedsConsecutiveRegisters(
11923 Type *Ty, CallingConv::ID CallConv, bool isVarArg) const {
11924 if (getEffectiveCallingConv(CallConv, isVarArg) !=
11925 CallingConv::ARM_AAPCS_VFP)
11928 HABaseType Base = HA_UNKNOWN;
11929 uint64_t Members = 0;
11930 bool IsHA = isHomogeneousAggregate(Ty, Base, Members);
11931 DEBUG(dbgs() << "isHA: " << IsHA << " "; Ty->dump());
11933 bool IsIntArray = Ty->isArrayTy() && Ty->getArrayElementType()->isIntegerTy();
11934 return IsHA || IsIntArray;