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 setOperationAction(ISD::FMINNUM, MVT::v2f32, Legal);
938 setOperationAction(ISD::FMAXNUM, MVT::v2f32, Legal);
939 setOperationAction(ISD::FMINNUM, MVT::v4f32, Legal);
940 setOperationAction(ISD::FMAXNUM, MVT::v4f32, Legal);
942 if (!Subtarget->isFPOnlySP()) {
943 setOperationAction(ISD::FFLOOR, MVT::f64, Legal);
944 setOperationAction(ISD::FCEIL, MVT::f64, Legal);
945 setOperationAction(ISD::FROUND, MVT::f64, Legal);
946 setOperationAction(ISD::FTRUNC, MVT::f64, Legal);
947 setOperationAction(ISD::FNEARBYINT, MVT::f64, Legal);
948 setOperationAction(ISD::FRINT, MVT::f64, Legal);
949 setOperationAction(ISD::FMINNUM, MVT::f64, Legal);
950 setOperationAction(ISD::FMAXNUM, MVT::f64, Legal);
954 if (Subtarget->hasVFP3()) {
955 setOperationAction(ISD::FMINNAN, MVT::f32, Legal);
956 setOperationAction(ISD::FMAXNAN, MVT::f32, Legal);
957 setOperationAction(ISD::FMINNAN, MVT::f64, Legal);
958 setOperationAction(ISD::FMAXNAN, MVT::f64, Legal);
960 if (Subtarget->hasNEON()) {
961 setOperationAction(ISD::FMINNAN, MVT::v2f32, Legal);
962 setOperationAction(ISD::FMAXNAN, MVT::v2f32, Legal);
963 setOperationAction(ISD::FMINNAN, MVT::v4f32, Legal);
964 setOperationAction(ISD::FMAXNAN, MVT::v4f32, Legal);
967 // We have target-specific dag combine patterns for the following nodes:
968 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
969 setTargetDAGCombine(ISD::ADD);
970 setTargetDAGCombine(ISD::SUB);
971 setTargetDAGCombine(ISD::MUL);
972 setTargetDAGCombine(ISD::AND);
973 setTargetDAGCombine(ISD::OR);
974 setTargetDAGCombine(ISD::XOR);
976 if (Subtarget->hasV6Ops())
977 setTargetDAGCombine(ISD::SRL);
979 setStackPointerRegisterToSaveRestore(ARM::SP);
981 if (Subtarget->useSoftFloat() || Subtarget->isThumb1Only() ||
982 !Subtarget->hasVFP2())
983 setSchedulingPreference(Sched::RegPressure);
985 setSchedulingPreference(Sched::Hybrid);
987 //// temporary - rewrite interface to use type
988 MaxStoresPerMemset = 8;
989 MaxStoresPerMemsetOptSize = 4;
990 MaxStoresPerMemcpy = 4; // For @llvm.memcpy -> sequence of stores
991 MaxStoresPerMemcpyOptSize = 2;
992 MaxStoresPerMemmove = 4; // For @llvm.memmove -> sequence of stores
993 MaxStoresPerMemmoveOptSize = 2;
995 // On ARM arguments smaller than 4 bytes are extended, so all arguments
996 // are at least 4 bytes aligned.
997 setMinStackArgumentAlignment(4);
999 // Prefer likely predicted branches to selects on out-of-order cores.
1000 PredictableSelectIsExpensive = Subtarget->isLikeA9();
1002 setMinFunctionAlignment(Subtarget->isThumb() ? 1 : 2);
1005 bool ARMTargetLowering::useSoftFloat() const {
1006 return Subtarget->useSoftFloat();
1009 // FIXME: It might make sense to define the representative register class as the
1010 // nearest super-register that has a non-null superset. For example, DPR_VFP2 is
1011 // a super-register of SPR, and DPR is a superset if DPR_VFP2. Consequently,
1012 // SPR's representative would be DPR_VFP2. This should work well if register
1013 // pressure tracking were modified such that a register use would increment the
1014 // pressure of the register class's representative and all of it's super
1015 // classes' representatives transitively. We have not implemented this because
1016 // of the difficulty prior to coalescing of modeling operand register classes
1017 // due to the common occurrence of cross class copies and subregister insertions
1019 std::pair<const TargetRegisterClass *, uint8_t>
1020 ARMTargetLowering::findRepresentativeClass(const TargetRegisterInfo *TRI,
1022 const TargetRegisterClass *RRC = nullptr;
1024 switch (VT.SimpleTy) {
1026 return TargetLowering::findRepresentativeClass(TRI, VT);
1027 // Use DPR as representative register class for all floating point
1028 // and vector types. Since there are 32 SPR registers and 32 DPR registers so
1029 // the cost is 1 for both f32 and f64.
1030 case MVT::f32: case MVT::f64: case MVT::v8i8: case MVT::v4i16:
1031 case MVT::v2i32: case MVT::v1i64: case MVT::v2f32:
1032 RRC = &ARM::DPRRegClass;
1033 // When NEON is used for SP, only half of the register file is available
1034 // because operations that define both SP and DP results will be constrained
1035 // to the VFP2 class (D0-D15). We currently model this constraint prior to
1036 // coalescing by double-counting the SP regs. See the FIXME above.
1037 if (Subtarget->useNEONForSinglePrecisionFP())
1040 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
1041 case MVT::v4f32: case MVT::v2f64:
1042 RRC = &ARM::DPRRegClass;
1046 RRC = &ARM::DPRRegClass;
1050 RRC = &ARM::DPRRegClass;
1054 return std::make_pair(RRC, Cost);
1057 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
1058 switch ((ARMISD::NodeType)Opcode) {
1059 case ARMISD::FIRST_NUMBER: break;
1060 case ARMISD::Wrapper: return "ARMISD::Wrapper";
1061 case ARMISD::WrapperPIC: return "ARMISD::WrapperPIC";
1062 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
1063 case ARMISD::COPY_STRUCT_BYVAL: return "ARMISD::COPY_STRUCT_BYVAL";
1064 case ARMISD::CALL: return "ARMISD::CALL";
1065 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
1066 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
1067 case ARMISD::tCALL: return "ARMISD::tCALL";
1068 case ARMISD::BRCOND: return "ARMISD::BRCOND";
1069 case ARMISD::BR_JT: return "ARMISD::BR_JT";
1070 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
1071 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
1072 case ARMISD::INTRET_FLAG: return "ARMISD::INTRET_FLAG";
1073 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
1074 case ARMISD::CMP: return "ARMISD::CMP";
1075 case ARMISD::CMN: return "ARMISD::CMN";
1076 case ARMISD::CMPZ: return "ARMISD::CMPZ";
1077 case ARMISD::CMPFP: return "ARMISD::CMPFP";
1078 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
1079 case ARMISD::BCC_i64: return "ARMISD::BCC_i64";
1080 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
1082 case ARMISD::CMOV: return "ARMISD::CMOV";
1084 case ARMISD::RBIT: return "ARMISD::RBIT";
1086 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
1087 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
1088 case ARMISD::RRX: return "ARMISD::RRX";
1090 case ARMISD::ADDC: return "ARMISD::ADDC";
1091 case ARMISD::ADDE: return "ARMISD::ADDE";
1092 case ARMISD::SUBC: return "ARMISD::SUBC";
1093 case ARMISD::SUBE: return "ARMISD::SUBE";
1095 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
1096 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
1098 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
1099 case ARMISD::EH_SJLJ_LONGJMP: return "ARMISD::EH_SJLJ_LONGJMP";
1100 case ARMISD::EH_SJLJ_SETUP_DISPATCH: return "ARMISD::EH_SJLJ_SETUP_DISPATCH";
1102 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN";
1104 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
1106 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
1108 case ARMISD::MEMBARRIER_MCR: return "ARMISD::MEMBARRIER_MCR";
1110 case ARMISD::PRELOAD: return "ARMISD::PRELOAD";
1112 case ARMISD::WIN__CHKSTK: return "ARMISD:::WIN__CHKSTK";
1114 case ARMISD::VCEQ: return "ARMISD::VCEQ";
1115 case ARMISD::VCEQZ: return "ARMISD::VCEQZ";
1116 case ARMISD::VCGE: return "ARMISD::VCGE";
1117 case ARMISD::VCGEZ: return "ARMISD::VCGEZ";
1118 case ARMISD::VCLEZ: return "ARMISD::VCLEZ";
1119 case ARMISD::VCGEU: return "ARMISD::VCGEU";
1120 case ARMISD::VCGT: return "ARMISD::VCGT";
1121 case ARMISD::VCGTZ: return "ARMISD::VCGTZ";
1122 case ARMISD::VCLTZ: return "ARMISD::VCLTZ";
1123 case ARMISD::VCGTU: return "ARMISD::VCGTU";
1124 case ARMISD::VTST: return "ARMISD::VTST";
1126 case ARMISD::VSHL: return "ARMISD::VSHL";
1127 case ARMISD::VSHRs: return "ARMISD::VSHRs";
1128 case ARMISD::VSHRu: return "ARMISD::VSHRu";
1129 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
1130 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
1131 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
1132 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
1133 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
1134 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
1135 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
1136 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
1137 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
1138 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
1139 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
1140 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
1141 case ARMISD::VSLI: return "ARMISD::VSLI";
1142 case ARMISD::VSRI: return "ARMISD::VSRI";
1143 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
1144 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
1145 case ARMISD::VMOVIMM: return "ARMISD::VMOVIMM";
1146 case ARMISD::VMVNIMM: return "ARMISD::VMVNIMM";
1147 case ARMISD::VMOVFPIMM: return "ARMISD::VMOVFPIMM";
1148 case ARMISD::VDUP: return "ARMISD::VDUP";
1149 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
1150 case ARMISD::VEXT: return "ARMISD::VEXT";
1151 case ARMISD::VREV64: return "ARMISD::VREV64";
1152 case ARMISD::VREV32: return "ARMISD::VREV32";
1153 case ARMISD::VREV16: return "ARMISD::VREV16";
1154 case ARMISD::VZIP: return "ARMISD::VZIP";
1155 case ARMISD::VUZP: return "ARMISD::VUZP";
1156 case ARMISD::VTRN: return "ARMISD::VTRN";
1157 case ARMISD::VTBL1: return "ARMISD::VTBL1";
1158 case ARMISD::VTBL2: return "ARMISD::VTBL2";
1159 case ARMISD::VMULLs: return "ARMISD::VMULLs";
1160 case ARMISD::VMULLu: return "ARMISD::VMULLu";
1161 case ARMISD::UMLAL: return "ARMISD::UMLAL";
1162 case ARMISD::SMLAL: return "ARMISD::SMLAL";
1163 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
1164 case ARMISD::BFI: return "ARMISD::BFI";
1165 case ARMISD::VORRIMM: return "ARMISD::VORRIMM";
1166 case ARMISD::VBICIMM: return "ARMISD::VBICIMM";
1167 case ARMISD::VBSL: return "ARMISD::VBSL";
1168 case ARMISD::VLD2DUP: return "ARMISD::VLD2DUP";
1169 case ARMISD::VLD3DUP: return "ARMISD::VLD3DUP";
1170 case ARMISD::VLD4DUP: return "ARMISD::VLD4DUP";
1171 case ARMISD::VLD1_UPD: return "ARMISD::VLD1_UPD";
1172 case ARMISD::VLD2_UPD: return "ARMISD::VLD2_UPD";
1173 case ARMISD::VLD3_UPD: return "ARMISD::VLD3_UPD";
1174 case ARMISD::VLD4_UPD: return "ARMISD::VLD4_UPD";
1175 case ARMISD::VLD2LN_UPD: return "ARMISD::VLD2LN_UPD";
1176 case ARMISD::VLD3LN_UPD: return "ARMISD::VLD3LN_UPD";
1177 case ARMISD::VLD4LN_UPD: return "ARMISD::VLD4LN_UPD";
1178 case ARMISD::VLD2DUP_UPD: return "ARMISD::VLD2DUP_UPD";
1179 case ARMISD::VLD3DUP_UPD: return "ARMISD::VLD3DUP_UPD";
1180 case ARMISD::VLD4DUP_UPD: return "ARMISD::VLD4DUP_UPD";
1181 case ARMISD::VST1_UPD: return "ARMISD::VST1_UPD";
1182 case ARMISD::VST2_UPD: return "ARMISD::VST2_UPD";
1183 case ARMISD::VST3_UPD: return "ARMISD::VST3_UPD";
1184 case ARMISD::VST4_UPD: return "ARMISD::VST4_UPD";
1185 case ARMISD::VST2LN_UPD: return "ARMISD::VST2LN_UPD";
1186 case ARMISD::VST3LN_UPD: return "ARMISD::VST3LN_UPD";
1187 case ARMISD::VST4LN_UPD: return "ARMISD::VST4LN_UPD";
1192 EVT ARMTargetLowering::getSetCCResultType(const DataLayout &DL, LLVMContext &,
1195 return getPointerTy(DL);
1196 return VT.changeVectorElementTypeToInteger();
1199 /// getRegClassFor - Return the register class that should be used for the
1200 /// specified value type.
1201 const TargetRegisterClass *ARMTargetLowering::getRegClassFor(MVT VT) const {
1202 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
1203 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
1204 // load / store 4 to 8 consecutive D registers.
1205 if (Subtarget->hasNEON()) {
1206 if (VT == MVT::v4i64)
1207 return &ARM::QQPRRegClass;
1208 if (VT == MVT::v8i64)
1209 return &ARM::QQQQPRRegClass;
1211 return TargetLowering::getRegClassFor(VT);
1214 // memcpy, and other memory intrinsics, typically tries to use LDM/STM if the
1215 // source/dest is aligned and the copy size is large enough. We therefore want
1216 // to align such objects passed to memory intrinsics.
1217 bool ARMTargetLowering::shouldAlignPointerArgs(CallInst *CI, unsigned &MinSize,
1218 unsigned &PrefAlign) const {
1219 if (!isa<MemIntrinsic>(CI))
1222 // On ARM11 onwards (excluding M class) 8-byte aligned LDM is typically 1
1223 // cycle faster than 4-byte aligned LDM.
1224 PrefAlign = (Subtarget->hasV6Ops() && !Subtarget->isMClass() ? 8 : 4);
1228 // Create a fast isel object.
1230 ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
1231 const TargetLibraryInfo *libInfo) const {
1232 return ARM::createFastISel(funcInfo, libInfo);
1235 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
1236 unsigned NumVals = N->getNumValues();
1238 return Sched::RegPressure;
1240 for (unsigned i = 0; i != NumVals; ++i) {
1241 EVT VT = N->getValueType(i);
1242 if (VT == MVT::Glue || VT == MVT::Other)
1244 if (VT.isFloatingPoint() || VT.isVector())
1248 if (!N->isMachineOpcode())
1249 return Sched::RegPressure;
1251 // Load are scheduled for latency even if there instruction itinerary
1252 // is not available.
1253 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
1254 const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
1256 if (MCID.getNumDefs() == 0)
1257 return Sched::RegPressure;
1258 if (!Itins->isEmpty() &&
1259 Itins->getOperandCycle(MCID.getSchedClass(), 0) > 2)
1262 return Sched::RegPressure;
1265 //===----------------------------------------------------------------------===//
1267 //===----------------------------------------------------------------------===//
1269 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
1270 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
1272 default: llvm_unreachable("Unknown condition code!");
1273 case ISD::SETNE: return ARMCC::NE;
1274 case ISD::SETEQ: return ARMCC::EQ;
1275 case ISD::SETGT: return ARMCC::GT;
1276 case ISD::SETGE: return ARMCC::GE;
1277 case ISD::SETLT: return ARMCC::LT;
1278 case ISD::SETLE: return ARMCC::LE;
1279 case ISD::SETUGT: return ARMCC::HI;
1280 case ISD::SETUGE: return ARMCC::HS;
1281 case ISD::SETULT: return ARMCC::LO;
1282 case ISD::SETULE: return ARMCC::LS;
1286 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
1287 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
1288 ARMCC::CondCodes &CondCode2) {
1289 CondCode2 = ARMCC::AL;
1291 default: llvm_unreachable("Unknown FP condition!");
1293 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
1295 case ISD::SETOGT: CondCode = ARMCC::GT; break;
1297 case ISD::SETOGE: CondCode = ARMCC::GE; break;
1298 case ISD::SETOLT: CondCode = ARMCC::MI; break;
1299 case ISD::SETOLE: CondCode = ARMCC::LS; break;
1300 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
1301 case ISD::SETO: CondCode = ARMCC::VC; break;
1302 case ISD::SETUO: CondCode = ARMCC::VS; break;
1303 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
1304 case ISD::SETUGT: CondCode = ARMCC::HI; break;
1305 case ISD::SETUGE: CondCode = ARMCC::PL; break;
1307 case ISD::SETULT: CondCode = ARMCC::LT; break;
1309 case ISD::SETULE: CondCode = ARMCC::LE; break;
1311 case ISD::SETUNE: CondCode = ARMCC::NE; break;
1315 //===----------------------------------------------------------------------===//
1316 // Calling Convention Implementation
1317 //===----------------------------------------------------------------------===//
1319 #include "ARMGenCallingConv.inc"
1321 /// getEffectiveCallingConv - Get the effective calling convention, taking into
1322 /// account presence of floating point hardware and calling convention
1323 /// limitations, such as support for variadic functions.
1325 ARMTargetLowering::getEffectiveCallingConv(CallingConv::ID CC,
1326 bool isVarArg) const {
1329 llvm_unreachable("Unsupported calling convention");
1330 case CallingConv::ARM_AAPCS:
1331 case CallingConv::ARM_APCS:
1332 case CallingConv::GHC:
1334 case CallingConv::ARM_AAPCS_VFP:
1335 return isVarArg ? CallingConv::ARM_AAPCS : CallingConv::ARM_AAPCS_VFP;
1336 case CallingConv::C:
1337 if (!Subtarget->isAAPCS_ABI())
1338 return CallingConv::ARM_APCS;
1339 else if (Subtarget->hasVFP2() && !Subtarget->isThumb1Only() &&
1340 getTargetMachine().Options.FloatABIType == FloatABI::Hard &&
1342 return CallingConv::ARM_AAPCS_VFP;
1344 return CallingConv::ARM_AAPCS;
1345 case CallingConv::Fast:
1346 if (!Subtarget->isAAPCS_ABI()) {
1347 if (Subtarget->hasVFP2() && !Subtarget->isThumb1Only() && !isVarArg)
1348 return CallingConv::Fast;
1349 return CallingConv::ARM_APCS;
1350 } else if (Subtarget->hasVFP2() && !Subtarget->isThumb1Only() && !isVarArg)
1351 return CallingConv::ARM_AAPCS_VFP;
1353 return CallingConv::ARM_AAPCS;
1357 /// CCAssignFnForNode - Selects the correct CCAssignFn for the given
1358 /// CallingConvention.
1359 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
1361 bool isVarArg) const {
1362 switch (getEffectiveCallingConv(CC, isVarArg)) {
1364 llvm_unreachable("Unsupported calling convention");
1365 case CallingConv::ARM_APCS:
1366 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1367 case CallingConv::ARM_AAPCS:
1368 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1369 case CallingConv::ARM_AAPCS_VFP:
1370 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1371 case CallingConv::Fast:
1372 return (Return ? RetFastCC_ARM_APCS : FastCC_ARM_APCS);
1373 case CallingConv::GHC:
1374 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS_GHC);
1378 /// LowerCallResult - Lower the result values of a call into the
1379 /// appropriate copies out of appropriate physical registers.
1381 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1382 CallingConv::ID CallConv, bool isVarArg,
1383 const SmallVectorImpl<ISD::InputArg> &Ins,
1384 SDLoc dl, SelectionDAG &DAG,
1385 SmallVectorImpl<SDValue> &InVals,
1386 bool isThisReturn, SDValue ThisVal) const {
1388 // Assign locations to each value returned by this call.
1389 SmallVector<CCValAssign, 16> RVLocs;
1390 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
1391 *DAG.getContext(), Call);
1392 CCInfo.AnalyzeCallResult(Ins,
1393 CCAssignFnForNode(CallConv, /* Return*/ true,
1396 // Copy all of the result registers out of their specified physreg.
1397 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1398 CCValAssign VA = RVLocs[i];
1400 // Pass 'this' value directly from the argument to return value, to avoid
1401 // reg unit interference
1402 if (i == 0 && isThisReturn) {
1403 assert(!VA.needsCustom() && VA.getLocVT() == MVT::i32 &&
1404 "unexpected return calling convention register assignment");
1405 InVals.push_back(ThisVal);
1410 if (VA.needsCustom()) {
1411 // Handle f64 or half of a v2f64.
1412 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1414 Chain = Lo.getValue(1);
1415 InFlag = Lo.getValue(2);
1416 VA = RVLocs[++i]; // skip ahead to next loc
1417 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1419 Chain = Hi.getValue(1);
1420 InFlag = Hi.getValue(2);
1421 if (!Subtarget->isLittle())
1423 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1425 if (VA.getLocVT() == MVT::v2f64) {
1426 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1427 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1428 DAG.getConstant(0, dl, MVT::i32));
1430 VA = RVLocs[++i]; // skip ahead to next loc
1431 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1432 Chain = Lo.getValue(1);
1433 InFlag = Lo.getValue(2);
1434 VA = RVLocs[++i]; // skip ahead to next loc
1435 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1436 Chain = Hi.getValue(1);
1437 InFlag = Hi.getValue(2);
1438 if (!Subtarget->isLittle())
1440 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1441 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1442 DAG.getConstant(1, dl, MVT::i32));
1445 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
1447 Chain = Val.getValue(1);
1448 InFlag = Val.getValue(2);
1451 switch (VA.getLocInfo()) {
1452 default: llvm_unreachable("Unknown loc info!");
1453 case CCValAssign::Full: break;
1454 case CCValAssign::BCvt:
1455 Val = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), Val);
1459 InVals.push_back(Val);
1465 /// LowerMemOpCallTo - Store the argument to the stack.
1467 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
1468 SDValue StackPtr, SDValue Arg,
1469 SDLoc dl, SelectionDAG &DAG,
1470 const CCValAssign &VA,
1471 ISD::ArgFlagsTy Flags) const {
1472 unsigned LocMemOffset = VA.getLocMemOffset();
1473 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset, dl);
1474 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(DAG.getDataLayout()),
1476 return DAG.getStore(Chain, dl, Arg, PtrOff,
1477 MachinePointerInfo::getStack(LocMemOffset),
1481 void ARMTargetLowering::PassF64ArgInRegs(SDLoc dl, SelectionDAG &DAG,
1482 SDValue Chain, SDValue &Arg,
1483 RegsToPassVector &RegsToPass,
1484 CCValAssign &VA, CCValAssign &NextVA,
1486 SmallVectorImpl<SDValue> &MemOpChains,
1487 ISD::ArgFlagsTy Flags) const {
1489 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1490 DAG.getVTList(MVT::i32, MVT::i32), Arg);
1491 unsigned id = Subtarget->isLittle() ? 0 : 1;
1492 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd.getValue(id)));
1494 if (NextVA.isRegLoc())
1495 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1-id)));
1497 assert(NextVA.isMemLoc());
1498 if (!StackPtr.getNode())
1499 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP,
1500 getPointerTy(DAG.getDataLayout()));
1502 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1-id),
1508 /// LowerCall - Lowering a call into a callseq_start <-
1509 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
1512 ARMTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
1513 SmallVectorImpl<SDValue> &InVals) const {
1514 SelectionDAG &DAG = CLI.DAG;
1516 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
1517 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
1518 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
1519 SDValue Chain = CLI.Chain;
1520 SDValue Callee = CLI.Callee;
1521 bool &isTailCall = CLI.IsTailCall;
1522 CallingConv::ID CallConv = CLI.CallConv;
1523 bool doesNotRet = CLI.DoesNotReturn;
1524 bool isVarArg = CLI.IsVarArg;
1526 MachineFunction &MF = DAG.getMachineFunction();
1527 bool isStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1528 bool isThisReturn = false;
1529 bool isSibCall = false;
1530 auto Attr = MF.getFunction()->getFnAttribute("disable-tail-calls");
1532 // Disable tail calls if they're not supported.
1533 if (!Subtarget->supportsTailCall() || Attr.getValueAsString() == "true")
1537 // Check if it's really possible to do a tail call.
1538 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1539 isVarArg, isStructRet, MF.getFunction()->hasStructRetAttr(),
1540 Outs, OutVals, Ins, DAG);
1541 if (!isTailCall && CLI.CS && CLI.CS->isMustTailCall())
1542 report_fatal_error("failed to perform tail call elimination on a call "
1543 "site marked musttail");
1544 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1545 // detected sibcalls.
1552 // Analyze operands of the call, assigning locations to each operand.
1553 SmallVector<CCValAssign, 16> ArgLocs;
1554 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs,
1555 *DAG.getContext(), Call);
1556 CCInfo.AnalyzeCallOperands(Outs,
1557 CCAssignFnForNode(CallConv, /* Return*/ false,
1560 // Get a count of how many bytes are to be pushed on the stack.
1561 unsigned NumBytes = CCInfo.getNextStackOffset();
1563 // For tail calls, memory operands are available in our caller's stack.
1567 // Adjust the stack pointer for the new arguments...
1568 // These operations are automatically eliminated by the prolog/epilog pass
1570 Chain = DAG.getCALLSEQ_START(Chain,
1571 DAG.getIntPtrConstant(NumBytes, dl, true), dl);
1574 DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy(DAG.getDataLayout()));
1576 RegsToPassVector RegsToPass;
1577 SmallVector<SDValue, 8> MemOpChains;
1579 // Walk the register/memloc assignments, inserting copies/loads. In the case
1580 // of tail call optimization, arguments are handled later.
1581 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1583 ++i, ++realArgIdx) {
1584 CCValAssign &VA = ArgLocs[i];
1585 SDValue Arg = OutVals[realArgIdx];
1586 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1587 bool isByVal = Flags.isByVal();
1589 // Promote the value if needed.
1590 switch (VA.getLocInfo()) {
1591 default: llvm_unreachable("Unknown loc info!");
1592 case CCValAssign::Full: break;
1593 case CCValAssign::SExt:
1594 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1596 case CCValAssign::ZExt:
1597 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1599 case CCValAssign::AExt:
1600 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1602 case CCValAssign::BCvt:
1603 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1607 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1608 if (VA.needsCustom()) {
1609 if (VA.getLocVT() == MVT::v2f64) {
1610 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1611 DAG.getConstant(0, dl, MVT::i32));
1612 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1613 DAG.getConstant(1, dl, MVT::i32));
1615 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1616 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1618 VA = ArgLocs[++i]; // skip ahead to next loc
1619 if (VA.isRegLoc()) {
1620 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1621 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1623 assert(VA.isMemLoc());
1625 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1626 dl, DAG, VA, Flags));
1629 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1630 StackPtr, MemOpChains, Flags);
1632 } else if (VA.isRegLoc()) {
1633 if (realArgIdx == 0 && Flags.isReturned() && Outs[0].VT == MVT::i32) {
1634 assert(VA.getLocVT() == MVT::i32 &&
1635 "unexpected calling convention register assignment");
1636 assert(!Ins.empty() && Ins[0].VT == MVT::i32 &&
1637 "unexpected use of 'returned'");
1638 isThisReturn = true;
1640 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1641 } else if (isByVal) {
1642 assert(VA.isMemLoc());
1643 unsigned offset = 0;
1645 // True if this byval aggregate will be split between registers
1647 unsigned ByValArgsCount = CCInfo.getInRegsParamsCount();
1648 unsigned CurByValIdx = CCInfo.getInRegsParamsProcessed();
1650 if (CurByValIdx < ByValArgsCount) {
1652 unsigned RegBegin, RegEnd;
1653 CCInfo.getInRegsParamInfo(CurByValIdx, RegBegin, RegEnd);
1656 DAG.getTargetLoweringInfo().getPointerTy(DAG.getDataLayout());
1658 for (i = 0, j = RegBegin; j < RegEnd; i++, j++) {
1659 SDValue Const = DAG.getConstant(4*i, dl, MVT::i32);
1660 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
1661 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
1662 MachinePointerInfo(),
1663 false, false, false,
1664 DAG.InferPtrAlignment(AddArg));
1665 MemOpChains.push_back(Load.getValue(1));
1666 RegsToPass.push_back(std::make_pair(j, Load));
1669 // If parameter size outsides register area, "offset" value
1670 // helps us to calculate stack slot for remained part properly.
1671 offset = RegEnd - RegBegin;
1673 CCInfo.nextInRegsParam();
1676 if (Flags.getByValSize() > 4*offset) {
1677 auto PtrVT = getPointerTy(DAG.getDataLayout());
1678 unsigned LocMemOffset = VA.getLocMemOffset();
1679 SDValue StkPtrOff = DAG.getIntPtrConstant(LocMemOffset, dl);
1680 SDValue Dst = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, StkPtrOff);
1681 SDValue SrcOffset = DAG.getIntPtrConstant(4*offset, dl);
1682 SDValue Src = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, SrcOffset);
1683 SDValue SizeNode = DAG.getConstant(Flags.getByValSize() - 4*offset, dl,
1685 SDValue AlignNode = DAG.getConstant(Flags.getByValAlign(), dl,
1688 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
1689 SDValue Ops[] = { Chain, Dst, Src, SizeNode, AlignNode};
1690 MemOpChains.push_back(DAG.getNode(ARMISD::COPY_STRUCT_BYVAL, dl, VTs,
1693 } else if (!isSibCall) {
1694 assert(VA.isMemLoc());
1696 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1697 dl, DAG, VA, Flags));
1701 if (!MemOpChains.empty())
1702 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains);
1704 // Build a sequence of copy-to-reg nodes chained together with token chain
1705 // and flag operands which copy the outgoing args into the appropriate regs.
1707 // Tail call byval lowering might overwrite argument registers so in case of
1708 // tail call optimization the copies to registers are lowered later.
1710 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1711 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1712 RegsToPass[i].second, InFlag);
1713 InFlag = Chain.getValue(1);
1716 // For tail calls lower the arguments to the 'real' stack slot.
1718 // Force all the incoming stack arguments to be loaded from the stack
1719 // before any new outgoing arguments are stored to the stack, because the
1720 // outgoing stack slots may alias the incoming argument stack slots, and
1721 // the alias isn't otherwise explicit. This is slightly more conservative
1722 // than necessary, because it means that each store effectively depends
1723 // on every argument instead of just those arguments it would clobber.
1725 // Do not flag preceding copytoreg stuff together with the following stuff.
1727 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1728 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1729 RegsToPass[i].second, InFlag);
1730 InFlag = Chain.getValue(1);
1735 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1736 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1737 // node so that legalize doesn't hack it.
1738 bool isDirect = false;
1739 bool isARMFunc = false;
1740 bool isLocalARMFunc = false;
1741 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1742 auto PtrVt = getPointerTy(DAG.getDataLayout());
1744 if (Subtarget->genLongCalls()) {
1745 assert((Subtarget->isTargetWindows() ||
1746 getTargetMachine().getRelocationModel() == Reloc::Static) &&
1747 "long-calls with non-static relocation model!");
1748 // Handle a global address or an external symbol. If it's not one of
1749 // those, the target's already in a register, so we don't need to do
1751 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1752 const GlobalValue *GV = G->getGlobal();
1753 // Create a constant pool entry for the callee address
1754 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1755 ARMConstantPoolValue *CPV =
1756 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 0);
1758 // Get the address of the callee into a register
1759 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVt, 4);
1760 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1761 Callee = DAG.getLoad(PtrVt, dl, DAG.getEntryNode(), CPAddr,
1762 MachinePointerInfo::getConstantPool(), false, false,
1764 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1765 const char *Sym = S->getSymbol();
1767 // Create a constant pool entry for the callee address
1768 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1769 ARMConstantPoolValue *CPV =
1770 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1771 ARMPCLabelIndex, 0);
1772 // Get the address of the callee into a register
1773 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVt, 4);
1774 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1775 Callee = DAG.getLoad(PtrVt, dl, DAG.getEntryNode(), CPAddr,
1776 MachinePointerInfo::getConstantPool(), false, false,
1779 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1780 const GlobalValue *GV = G->getGlobal();
1782 bool isDef = GV->isStrongDefinitionForLinker();
1783 bool isStub = (!isDef && Subtarget->isTargetMachO()) &&
1784 getTargetMachine().getRelocationModel() != Reloc::Static;
1785 isARMFunc = !Subtarget->isThumb() || (isStub && !Subtarget->isMClass());
1786 // ARM call to a local ARM function is predicable.
1787 isLocalARMFunc = !Subtarget->isThumb() && (isDef || !ARMInterworking);
1788 // tBX takes a register source operand.
1789 if (isStub && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1790 assert(Subtarget->isTargetMachO() && "WrapperPIC use on non-MachO?");
1791 Callee = DAG.getNode(
1792 ARMISD::WrapperPIC, dl, PtrVt,
1793 DAG.getTargetGlobalAddress(GV, dl, PtrVt, 0, ARMII::MO_NONLAZY));
1794 Callee = DAG.getLoad(PtrVt, dl, DAG.getEntryNode(), Callee,
1795 MachinePointerInfo::getGOT(), false, false, true, 0);
1796 } else if (Subtarget->isTargetCOFF()) {
1797 assert(Subtarget->isTargetWindows() &&
1798 "Windows is the only supported COFF target");
1799 unsigned TargetFlags = GV->hasDLLImportStorageClass()
1800 ? ARMII::MO_DLLIMPORT
1801 : ARMII::MO_NO_FLAG;
1803 DAG.getTargetGlobalAddress(GV, dl, PtrVt, /*Offset=*/0, TargetFlags);
1804 if (GV->hasDLLImportStorageClass())
1806 DAG.getLoad(PtrVt, dl, DAG.getEntryNode(),
1807 DAG.getNode(ARMISD::Wrapper, dl, PtrVt, Callee),
1808 MachinePointerInfo::getGOT(), false, false, false, 0);
1810 // On ELF targets for PIC code, direct calls should go through the PLT
1811 unsigned OpFlags = 0;
1812 if (Subtarget->isTargetELF() &&
1813 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1814 OpFlags = ARMII::MO_PLT;
1815 Callee = DAG.getTargetGlobalAddress(GV, dl, PtrVt, 0, OpFlags);
1817 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1819 bool isStub = Subtarget->isTargetMachO() &&
1820 getTargetMachine().getRelocationModel() != Reloc::Static;
1821 isARMFunc = !Subtarget->isThumb() || (isStub && !Subtarget->isMClass());
1822 // tBX takes a register source operand.
1823 const char *Sym = S->getSymbol();
1824 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1825 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1826 ARMConstantPoolValue *CPV =
1827 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1828 ARMPCLabelIndex, 4);
1829 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVt, 4);
1830 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1831 Callee = DAG.getLoad(PtrVt, dl, DAG.getEntryNode(), CPAddr,
1832 MachinePointerInfo::getConstantPool(), false, false,
1834 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, dl, MVT::i32);
1835 Callee = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVt, Callee, PICLabel);
1837 unsigned OpFlags = 0;
1838 // On ELF targets for PIC code, direct calls should go through the PLT
1839 if (Subtarget->isTargetELF() &&
1840 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1841 OpFlags = ARMII::MO_PLT;
1842 Callee = DAG.getTargetExternalSymbol(Sym, PtrVt, OpFlags);
1846 // FIXME: handle tail calls differently.
1848 if (Subtarget->isThumb()) {
1849 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1850 CallOpc = ARMISD::CALL_NOLINK;
1852 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1854 if (!isDirect && !Subtarget->hasV5TOps())
1855 CallOpc = ARMISD::CALL_NOLINK;
1856 else if (doesNotRet && isDirect && Subtarget->hasRAS() &&
1857 // Emit regular call when code size is the priority
1858 !MF.getFunction()->optForMinSize())
1859 // "mov lr, pc; b _foo" to avoid confusing the RSP
1860 CallOpc = ARMISD::CALL_NOLINK;
1862 CallOpc = isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL;
1865 std::vector<SDValue> Ops;
1866 Ops.push_back(Chain);
1867 Ops.push_back(Callee);
1869 // Add argument registers to the end of the list so that they are known live
1871 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1872 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1873 RegsToPass[i].second.getValueType()));
1875 // Add a register mask operand representing the call-preserved registers.
1877 const uint32_t *Mask;
1878 const ARMBaseRegisterInfo *ARI = Subtarget->getRegisterInfo();
1880 // For 'this' returns, use the R0-preserving mask if applicable
1881 Mask = ARI->getThisReturnPreservedMask(MF, CallConv);
1883 // Set isThisReturn to false if the calling convention is not one that
1884 // allows 'returned' to be modeled in this way, so LowerCallResult does
1885 // not try to pass 'this' straight through
1886 isThisReturn = false;
1887 Mask = ARI->getCallPreservedMask(MF, CallConv);
1890 Mask = ARI->getCallPreservedMask(MF, CallConv);
1892 assert(Mask && "Missing call preserved mask for calling convention");
1893 Ops.push_back(DAG.getRegisterMask(Mask));
1896 if (InFlag.getNode())
1897 Ops.push_back(InFlag);
1899 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
1901 MF.getFrameInfo()->setHasTailCall();
1902 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, Ops);
1905 // Returns a chain and a flag for retval copy to use.
1906 Chain = DAG.getNode(CallOpc, dl, NodeTys, Ops);
1907 InFlag = Chain.getValue(1);
1909 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, dl, true),
1910 DAG.getIntPtrConstant(0, dl, true), InFlag, dl);
1912 InFlag = Chain.getValue(1);
1914 // Handle result values, copying them out of physregs into vregs that we
1916 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins, dl, DAG,
1917 InVals, isThisReturn,
1918 isThisReturn ? OutVals[0] : SDValue());
1921 /// HandleByVal - Every parameter *after* a byval parameter is passed
1922 /// on the stack. Remember the next parameter register to allocate,
1923 /// and then confiscate the rest of the parameter registers to insure
1925 void ARMTargetLowering::HandleByVal(CCState *State, unsigned &Size,
1926 unsigned Align) const {
1927 assert((State->getCallOrPrologue() == Prologue ||
1928 State->getCallOrPrologue() == Call) &&
1929 "unhandled ParmContext");
1931 // Byval (as with any stack) slots are always at least 4 byte aligned.
1932 Align = std::max(Align, 4U);
1934 unsigned Reg = State->AllocateReg(GPRArgRegs);
1938 unsigned AlignInRegs = Align / 4;
1939 unsigned Waste = (ARM::R4 - Reg) % AlignInRegs;
1940 for (unsigned i = 0; i < Waste; ++i)
1941 Reg = State->AllocateReg(GPRArgRegs);
1946 unsigned Excess = 4 * (ARM::R4 - Reg);
1948 // Special case when NSAA != SP and parameter size greater than size of
1949 // all remained GPR regs. In that case we can't split parameter, we must
1950 // send it to stack. We also must set NCRN to R4, so waste all
1951 // remained registers.
1952 const unsigned NSAAOffset = State->getNextStackOffset();
1953 if (NSAAOffset != 0 && Size > Excess) {
1954 while (State->AllocateReg(GPRArgRegs))
1959 // First register for byval parameter is the first register that wasn't
1960 // allocated before this method call, so it would be "reg".
1961 // If parameter is small enough to be saved in range [reg, r4), then
1962 // the end (first after last) register would be reg + param-size-in-regs,
1963 // else parameter would be splitted between registers and stack,
1964 // end register would be r4 in this case.
1965 unsigned ByValRegBegin = Reg;
1966 unsigned ByValRegEnd = std::min<unsigned>(Reg + Size / 4, ARM::R4);
1967 State->addInRegsParamInfo(ByValRegBegin, ByValRegEnd);
1968 // Note, first register is allocated in the beginning of function already,
1969 // allocate remained amount of registers we need.
1970 for (unsigned i = Reg + 1; i != ByValRegEnd; ++i)
1971 State->AllocateReg(GPRArgRegs);
1972 // A byval parameter that is split between registers and memory needs its
1973 // size truncated here.
1974 // In the case where the entire structure fits in registers, we set the
1975 // size in memory to zero.
1976 Size = std::max<int>(Size - Excess, 0);
1979 /// MatchingStackOffset - Return true if the given stack call argument is
1980 /// already available in the same position (relatively) of the caller's
1981 /// incoming argument stack.
1983 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1984 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1985 const TargetInstrInfo *TII) {
1986 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1988 if (Arg.getOpcode() == ISD::CopyFromReg) {
1989 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1990 if (!TargetRegisterInfo::isVirtualRegister(VR))
1992 MachineInstr *Def = MRI->getVRegDef(VR);
1995 if (!Flags.isByVal()) {
1996 if (!TII->isLoadFromStackSlot(Def, FI))
2001 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
2002 if (Flags.isByVal())
2003 // ByVal argument is passed in as a pointer but it's now being
2004 // dereferenced. e.g.
2005 // define @foo(%struct.X* %A) {
2006 // tail call @bar(%struct.X* byval %A)
2009 SDValue Ptr = Ld->getBasePtr();
2010 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
2013 FI = FINode->getIndex();
2017 assert(FI != INT_MAX);
2018 if (!MFI->isFixedObjectIndex(FI))
2020 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
2023 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2024 /// for tail call optimization. Targets which want to do tail call
2025 /// optimization should implement this function.
2027 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2028 CallingConv::ID CalleeCC,
2030 bool isCalleeStructRet,
2031 bool isCallerStructRet,
2032 const SmallVectorImpl<ISD::OutputArg> &Outs,
2033 const SmallVectorImpl<SDValue> &OutVals,
2034 const SmallVectorImpl<ISD::InputArg> &Ins,
2035 SelectionDAG& DAG) const {
2036 const Function *CallerF = DAG.getMachineFunction().getFunction();
2037 CallingConv::ID CallerCC = CallerF->getCallingConv();
2038 bool CCMatch = CallerCC == CalleeCC;
2040 // Look for obvious safe cases to perform tail call optimization that do not
2041 // require ABI changes. This is what gcc calls sibcall.
2043 // Do not sibcall optimize vararg calls unless the call site is not passing
2045 if (isVarArg && !Outs.empty())
2048 // Exception-handling functions need a special set of instructions to indicate
2049 // a return to the hardware. Tail-calling another function would probably
2051 if (CallerF->hasFnAttribute("interrupt"))
2054 // Also avoid sibcall optimization if either caller or callee uses struct
2055 // return semantics.
2056 if (isCalleeStructRet || isCallerStructRet)
2059 // FIXME: Completely disable sibcall for Thumb1 since ThumbRegisterInfo::
2060 // emitEpilogue is not ready for them. Thumb tail calls also use t2B, as
2061 // the Thumb1 16-bit unconditional branch doesn't have sufficient relocation
2062 // support in the assembler and linker to be used. This would need to be
2063 // fixed to fully support tail calls in Thumb1.
2065 // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take
2066 // LR. This means if we need to reload LR, it takes an extra instructions,
2067 // which outweighs the value of the tail call; but here we don't know yet
2068 // whether LR is going to be used. Probably the right approach is to
2069 // generate the tail call here and turn it back into CALL/RET in
2070 // emitEpilogue if LR is used.
2072 // Thumb1 PIC calls to external symbols use BX, so they can be tail calls,
2073 // but we need to make sure there are enough registers; the only valid
2074 // registers are the 4 used for parameters. We don't currently do this
2076 if (Subtarget->isThumb1Only())
2079 // Externally-defined functions with weak linkage should not be
2080 // tail-called on ARM when the OS does not support dynamic
2081 // pre-emption of symbols, as the AAELF spec requires normal calls
2082 // to undefined weak functions to be replaced with a NOP or jump to the
2083 // next instruction. The behaviour of branch instructions in this
2084 // situation (as used for tail calls) is implementation-defined, so we
2085 // cannot rely on the linker replacing the tail call with a return.
2086 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2087 const GlobalValue *GV = G->getGlobal();
2088 const Triple &TT = getTargetMachine().getTargetTriple();
2089 if (GV->hasExternalWeakLinkage() &&
2090 (!TT.isOSWindows() || TT.isOSBinFormatELF() || TT.isOSBinFormatMachO()))
2094 // If the calling conventions do not match, then we'd better make sure the
2095 // results are returned in the same way as what the caller expects.
2097 SmallVector<CCValAssign, 16> RVLocs1;
2098 ARMCCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(), RVLocs1,
2099 *DAG.getContext(), Call);
2100 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
2102 SmallVector<CCValAssign, 16> RVLocs2;
2103 ARMCCState CCInfo2(CallerCC, false, DAG.getMachineFunction(), RVLocs2,
2104 *DAG.getContext(), Call);
2105 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
2107 if (RVLocs1.size() != RVLocs2.size())
2109 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
2110 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
2112 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
2114 if (RVLocs1[i].isRegLoc()) {
2115 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
2118 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
2124 // If Caller's vararg or byval argument has been split between registers and
2125 // stack, do not perform tail call, since part of the argument is in caller's
2127 const ARMFunctionInfo *AFI_Caller = DAG.getMachineFunction().
2128 getInfo<ARMFunctionInfo>();
2129 if (AFI_Caller->getArgRegsSaveSize())
2132 // If the callee takes no arguments then go on to check the results of the
2134 if (!Outs.empty()) {
2135 // Check if stack adjustment is needed. For now, do not do this if any
2136 // argument is passed on the stack.
2137 SmallVector<CCValAssign, 16> ArgLocs;
2138 ARMCCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
2139 *DAG.getContext(), Call);
2140 CCInfo.AnalyzeCallOperands(Outs,
2141 CCAssignFnForNode(CalleeCC, false, isVarArg));
2142 if (CCInfo.getNextStackOffset()) {
2143 MachineFunction &MF = DAG.getMachineFunction();
2145 // Check if the arguments are already laid out in the right way as
2146 // the caller's fixed stack objects.
2147 MachineFrameInfo *MFI = MF.getFrameInfo();
2148 const MachineRegisterInfo *MRI = &MF.getRegInfo();
2149 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
2150 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
2152 ++i, ++realArgIdx) {
2153 CCValAssign &VA = ArgLocs[i];
2154 EVT RegVT = VA.getLocVT();
2155 SDValue Arg = OutVals[realArgIdx];
2156 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
2157 if (VA.getLocInfo() == CCValAssign::Indirect)
2159 if (VA.needsCustom()) {
2160 // f64 and vector types are split into multiple registers or
2161 // register/stack-slot combinations. The types will not match
2162 // the registers; give up on memory f64 refs until we figure
2163 // out what to do about this.
2166 if (!ArgLocs[++i].isRegLoc())
2168 if (RegVT == MVT::v2f64) {
2169 if (!ArgLocs[++i].isRegLoc())
2171 if (!ArgLocs[++i].isRegLoc())
2174 } else if (!VA.isRegLoc()) {
2175 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
2187 ARMTargetLowering::CanLowerReturn(CallingConv::ID CallConv,
2188 MachineFunction &MF, bool isVarArg,
2189 const SmallVectorImpl<ISD::OutputArg> &Outs,
2190 LLVMContext &Context) const {
2191 SmallVector<CCValAssign, 16> RVLocs;
2192 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, Context);
2193 return CCInfo.CheckReturn(Outs, CCAssignFnForNode(CallConv, /*Return=*/true,
2197 static SDValue LowerInterruptReturn(SmallVectorImpl<SDValue> &RetOps,
2198 SDLoc DL, SelectionDAG &DAG) {
2199 const MachineFunction &MF = DAG.getMachineFunction();
2200 const Function *F = MF.getFunction();
2202 StringRef IntKind = F->getFnAttribute("interrupt").getValueAsString();
2204 // See ARM ARM v7 B1.8.3. On exception entry LR is set to a possibly offset
2205 // version of the "preferred return address". These offsets affect the return
2206 // instruction if this is a return from PL1 without hypervisor extensions.
2207 // IRQ/FIQ: +4 "subs pc, lr, #4"
2208 // SWI: 0 "subs pc, lr, #0"
2209 // ABORT: +4 "subs pc, lr, #4"
2210 // UNDEF: +4/+2 "subs pc, lr, #0"
2211 // UNDEF varies depending on where the exception came from ARM or Thumb
2212 // mode. Alongside GCC, we throw our hands up in disgust and pretend it's 0.
2215 if (IntKind == "" || IntKind == "IRQ" || IntKind == "FIQ" ||
2218 else if (IntKind == "SWI" || IntKind == "UNDEF")
2221 report_fatal_error("Unsupported interrupt attribute. If present, value "
2222 "must be one of: IRQ, FIQ, SWI, ABORT or UNDEF");
2224 RetOps.insert(RetOps.begin() + 1,
2225 DAG.getConstant(LROffset, DL, MVT::i32, false));
2227 return DAG.getNode(ARMISD::INTRET_FLAG, DL, MVT::Other, RetOps);
2231 ARMTargetLowering::LowerReturn(SDValue Chain,
2232 CallingConv::ID CallConv, bool isVarArg,
2233 const SmallVectorImpl<ISD::OutputArg> &Outs,
2234 const SmallVectorImpl<SDValue> &OutVals,
2235 SDLoc dl, SelectionDAG &DAG) const {
2237 // CCValAssign - represent the assignment of the return value to a location.
2238 SmallVector<CCValAssign, 16> RVLocs;
2240 // CCState - Info about the registers and stack slots.
2241 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
2242 *DAG.getContext(), Call);
2244 // Analyze outgoing return values.
2245 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
2249 SmallVector<SDValue, 4> RetOps;
2250 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
2251 bool isLittleEndian = Subtarget->isLittle();
2253 MachineFunction &MF = DAG.getMachineFunction();
2254 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2255 AFI->setReturnRegsCount(RVLocs.size());
2257 // Copy the result values into the output registers.
2258 for (unsigned i = 0, realRVLocIdx = 0;
2260 ++i, ++realRVLocIdx) {
2261 CCValAssign &VA = RVLocs[i];
2262 assert(VA.isRegLoc() && "Can only return in registers!");
2264 SDValue Arg = OutVals[realRVLocIdx];
2266 switch (VA.getLocInfo()) {
2267 default: llvm_unreachable("Unknown loc info!");
2268 case CCValAssign::Full: break;
2269 case CCValAssign::BCvt:
2270 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
2274 if (VA.needsCustom()) {
2275 if (VA.getLocVT() == MVT::v2f64) {
2276 // Extract the first half and return it in two registers.
2277 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
2278 DAG.getConstant(0, dl, MVT::i32));
2279 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
2280 DAG.getVTList(MVT::i32, MVT::i32), Half);
2282 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2283 HalfGPRs.getValue(isLittleEndian ? 0 : 1),
2285 Flag = Chain.getValue(1);
2286 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2287 VA = RVLocs[++i]; // skip ahead to next loc
2288 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2289 HalfGPRs.getValue(isLittleEndian ? 1 : 0),
2291 Flag = Chain.getValue(1);
2292 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2293 VA = RVLocs[++i]; // skip ahead to next loc
2295 // Extract the 2nd half and fall through to handle it as an f64 value.
2296 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
2297 DAG.getConstant(1, dl, MVT::i32));
2299 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
2301 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
2302 DAG.getVTList(MVT::i32, MVT::i32), Arg);
2303 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2304 fmrrd.getValue(isLittleEndian ? 0 : 1),
2306 Flag = Chain.getValue(1);
2307 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2308 VA = RVLocs[++i]; // skip ahead to next loc
2309 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2310 fmrrd.getValue(isLittleEndian ? 1 : 0),
2313 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
2315 // Guarantee that all emitted copies are
2316 // stuck together, avoiding something bad.
2317 Flag = Chain.getValue(1);
2318 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2321 // Update chain and glue.
2324 RetOps.push_back(Flag);
2326 // CPUs which aren't M-class use a special sequence to return from
2327 // exceptions (roughly, any instruction setting pc and cpsr simultaneously,
2328 // though we use "subs pc, lr, #N").
2330 // M-class CPUs actually use a normal return sequence with a special
2331 // (hardware-provided) value in LR, so the normal code path works.
2332 if (DAG.getMachineFunction().getFunction()->hasFnAttribute("interrupt") &&
2333 !Subtarget->isMClass()) {
2334 if (Subtarget->isThumb1Only())
2335 report_fatal_error("interrupt attribute is not supported in Thumb1");
2336 return LowerInterruptReturn(RetOps, dl, DAG);
2339 return DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, RetOps);
2342 bool ARMTargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
2343 if (N->getNumValues() != 1)
2345 if (!N->hasNUsesOfValue(1, 0))
2348 SDValue TCChain = Chain;
2349 SDNode *Copy = *N->use_begin();
2350 if (Copy->getOpcode() == ISD::CopyToReg) {
2351 // If the copy has a glue operand, we conservatively assume it isn't safe to
2352 // perform a tail call.
2353 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
2355 TCChain = Copy->getOperand(0);
2356 } else if (Copy->getOpcode() == ARMISD::VMOVRRD) {
2357 SDNode *VMov = Copy;
2358 // f64 returned in a pair of GPRs.
2359 SmallPtrSet<SDNode*, 2> Copies;
2360 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
2362 if (UI->getOpcode() != ISD::CopyToReg)
2366 if (Copies.size() > 2)
2369 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
2371 SDValue UseChain = UI->getOperand(0);
2372 if (Copies.count(UseChain.getNode()))
2376 // We are at the top of this chain.
2377 // If the copy has a glue operand, we conservatively assume it
2378 // isn't safe to perform a tail call.
2379 if (UI->getOperand(UI->getNumOperands()-1).getValueType() == MVT::Glue)
2385 } else if (Copy->getOpcode() == ISD::BITCAST) {
2386 // f32 returned in a single GPR.
2387 if (!Copy->hasOneUse())
2389 Copy = *Copy->use_begin();
2390 if (Copy->getOpcode() != ISD::CopyToReg || !Copy->hasNUsesOfValue(1, 0))
2392 // If the copy has a glue operand, we conservatively assume it isn't safe to
2393 // perform a tail call.
2394 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
2396 TCChain = Copy->getOperand(0);
2401 bool HasRet = false;
2402 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
2404 if (UI->getOpcode() != ARMISD::RET_FLAG &&
2405 UI->getOpcode() != ARMISD::INTRET_FLAG)
2417 bool ARMTargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
2418 if (!Subtarget->supportsTailCall())
2422 CI->getParent()->getParent()->getFnAttribute("disable-tail-calls");
2423 if (!CI->isTailCall() || Attr.getValueAsString() == "true")
2426 return !Subtarget->isThumb1Only();
2429 // Trying to write a 64 bit value so need to split into two 32 bit values first,
2430 // and pass the lower and high parts through.
2431 static SDValue LowerWRITE_REGISTER(SDValue Op, SelectionDAG &DAG) {
2433 SDValue WriteValue = Op->getOperand(2);
2435 // This function is only supposed to be called for i64 type argument.
2436 assert(WriteValue.getValueType() == MVT::i64
2437 && "LowerWRITE_REGISTER called for non-i64 type argument.");
2439 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, WriteValue,
2440 DAG.getConstant(0, DL, MVT::i32));
2441 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, WriteValue,
2442 DAG.getConstant(1, DL, MVT::i32));
2443 SDValue Ops[] = { Op->getOperand(0), Op->getOperand(1), Lo, Hi };
2444 return DAG.getNode(ISD::WRITE_REGISTER, DL, MVT::Other, Ops);
2447 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
2448 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
2449 // one of the above mentioned nodes. It has to be wrapped because otherwise
2450 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
2451 // be used to form addressing mode. These wrapped nodes will be selected
2453 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
2454 EVT PtrVT = Op.getValueType();
2455 // FIXME there is no actual debug info here
2457 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
2459 if (CP->isMachineConstantPoolEntry())
2460 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
2461 CP->getAlignment());
2463 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
2464 CP->getAlignment());
2465 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
2468 unsigned ARMTargetLowering::getJumpTableEncoding() const {
2469 return MachineJumpTableInfo::EK_Inline;
2472 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
2473 SelectionDAG &DAG) const {
2474 MachineFunction &MF = DAG.getMachineFunction();
2475 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2476 unsigned ARMPCLabelIndex = 0;
2478 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2479 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
2480 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2482 if (RelocM == Reloc::Static) {
2483 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
2485 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2486 ARMPCLabelIndex = AFI->createPICLabelUId();
2487 ARMConstantPoolValue *CPV =
2488 ARMConstantPoolConstant::Create(BA, ARMPCLabelIndex,
2489 ARMCP::CPBlockAddress, PCAdj);
2490 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2492 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
2493 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
2494 MachinePointerInfo::getConstantPool(),
2495 false, false, false, 0);
2496 if (RelocM == Reloc::Static)
2498 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, DL, MVT::i32);
2499 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
2502 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
2504 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
2505 SelectionDAG &DAG) const {
2507 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2508 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2509 MachineFunction &MF = DAG.getMachineFunction();
2510 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2511 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2512 ARMConstantPoolValue *CPV =
2513 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2514 ARMCP::CPValue, PCAdj, ARMCP::TLSGD, true);
2515 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2516 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
2517 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
2518 MachinePointerInfo::getConstantPool(),
2519 false, false, false, 0);
2520 SDValue Chain = Argument.getValue(1);
2522 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, dl, MVT::i32);
2523 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
2525 // call __tls_get_addr.
2528 Entry.Node = Argument;
2529 Entry.Ty = (Type *) Type::getInt32Ty(*DAG.getContext());
2530 Args.push_back(Entry);
2532 // FIXME: is there useful debug info available here?
2533 TargetLowering::CallLoweringInfo CLI(DAG);
2534 CLI.setDebugLoc(dl).setChain(Chain)
2535 .setCallee(CallingConv::C, Type::getInt32Ty(*DAG.getContext()),
2536 DAG.getExternalSymbol("__tls_get_addr", PtrVT), std::move(Args),
2539 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
2540 return CallResult.first;
2543 // Lower ISD::GlobalTLSAddress using the "initial exec" or
2544 // "local exec" model.
2546 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
2548 TLSModel::Model model) const {
2549 const GlobalValue *GV = GA->getGlobal();
2552 SDValue Chain = DAG.getEntryNode();
2553 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2554 // Get the Thread Pointer
2555 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2557 if (model == TLSModel::InitialExec) {
2558 MachineFunction &MF = DAG.getMachineFunction();
2559 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2560 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2561 // Initial exec model.
2562 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2563 ARMConstantPoolValue *CPV =
2564 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2565 ARMCP::CPValue, PCAdj, ARMCP::GOTTPOFF,
2567 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2568 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2569 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2570 MachinePointerInfo::getConstantPool(),
2571 false, false, false, 0);
2572 Chain = Offset.getValue(1);
2574 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, dl, MVT::i32);
2575 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
2577 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2578 MachinePointerInfo::getConstantPool(),
2579 false, false, false, 0);
2582 assert(model == TLSModel::LocalExec);
2583 ARMConstantPoolValue *CPV =
2584 ARMConstantPoolConstant::Create(GV, ARMCP::TPOFF);
2585 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2586 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2587 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2588 MachinePointerInfo::getConstantPool(),
2589 false, false, false, 0);
2592 // The address of the thread local variable is the add of the thread
2593 // pointer with the offset of the variable.
2594 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
2598 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
2599 // TODO: implement the "local dynamic" model
2600 assert(Subtarget->isTargetELF() &&
2601 "TLS not implemented for non-ELF targets");
2602 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
2603 if (DAG.getTarget().Options.EmulatedTLS)
2604 return LowerToTLSEmulatedModel(GA, DAG);
2606 TLSModel::Model model = getTargetMachine().getTLSModel(GA->getGlobal());
2609 case TLSModel::GeneralDynamic:
2610 case TLSModel::LocalDynamic:
2611 return LowerToTLSGeneralDynamicModel(GA, DAG);
2612 case TLSModel::InitialExec:
2613 case TLSModel::LocalExec:
2614 return LowerToTLSExecModels(GA, DAG, model);
2616 llvm_unreachable("bogus TLS model");
2619 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
2620 SelectionDAG &DAG) const {
2621 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2623 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2624 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2625 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
2626 ARMConstantPoolValue *CPV =
2627 ARMConstantPoolConstant::Create(GV,
2628 UseGOTOFF ? ARMCP::GOTOFF : ARMCP::GOT);
2629 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2630 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2631 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
2633 MachinePointerInfo::getConstantPool(),
2634 false, false, false, 0);
2635 SDValue Chain = Result.getValue(1);
2636 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
2637 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
2639 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
2640 MachinePointerInfo::getGOT(),
2641 false, false, false, 0);
2645 // If we have T2 ops, we can materialize the address directly via movt/movw
2646 // pair. This is always cheaper.
2647 if (Subtarget->useMovt(DAG.getMachineFunction())) {
2649 // FIXME: Once remat is capable of dealing with instructions with register
2650 // operands, expand this into two nodes.
2651 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2652 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2654 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2655 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2656 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2657 MachinePointerInfo::getConstantPool(),
2658 false, false, false, 0);
2662 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
2663 SelectionDAG &DAG) const {
2664 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2666 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2667 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2669 if (Subtarget->useMovt(DAG.getMachineFunction()))
2672 // FIXME: Once remat is capable of dealing with instructions with register
2673 // operands, expand this into multiple nodes
2675 RelocM == Reloc::PIC_ ? ARMISD::WrapperPIC : ARMISD::Wrapper;
2677 SDValue G = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, ARMII::MO_NONLAZY);
2678 SDValue Result = DAG.getNode(Wrapper, dl, PtrVT, G);
2680 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2681 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
2682 MachinePointerInfo::getGOT(), false, false, false, 0);
2686 SDValue ARMTargetLowering::LowerGlobalAddressWindows(SDValue Op,
2687 SelectionDAG &DAG) const {
2688 assert(Subtarget->isTargetWindows() && "non-Windows COFF is not supported");
2689 assert(Subtarget->useMovt(DAG.getMachineFunction()) &&
2690 "Windows on ARM expects to use movw/movt");
2692 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2693 const ARMII::TOF TargetFlags =
2694 (GV->hasDLLImportStorageClass() ? ARMII::MO_DLLIMPORT : ARMII::MO_NO_FLAG);
2695 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2701 // FIXME: Once remat is capable of dealing with instructions with register
2702 // operands, expand this into two nodes.
2703 Result = DAG.getNode(ARMISD::Wrapper, DL, PtrVT,
2704 DAG.getTargetGlobalAddress(GV, DL, PtrVT, /*Offset=*/0,
2706 if (GV->hasDLLImportStorageClass())
2707 Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), Result,
2708 MachinePointerInfo::getGOT(), false, false, false, 0);
2712 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
2713 SelectionDAG &DAG) const {
2714 assert(Subtarget->isTargetELF() &&
2715 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
2716 MachineFunction &MF = DAG.getMachineFunction();
2717 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2718 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2719 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2721 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2722 ARMConstantPoolValue *CPV =
2723 ARMConstantPoolSymbol::Create(*DAG.getContext(), "_GLOBAL_OFFSET_TABLE_",
2724 ARMPCLabelIndex, PCAdj);
2725 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2726 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2727 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2728 MachinePointerInfo::getConstantPool(),
2729 false, false, false, 0);
2730 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, dl, MVT::i32);
2731 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2735 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
2737 SDValue Val = DAG.getConstant(0, dl, MVT::i32);
2738 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl,
2739 DAG.getVTList(MVT::i32, MVT::Other), Op.getOperand(0),
2740 Op.getOperand(1), Val);
2744 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
2746 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
2747 Op.getOperand(1), DAG.getConstant(0, dl, MVT::i32));
2750 SDValue ARMTargetLowering::LowerEH_SJLJ_SETUP_DISPATCH(SDValue Op,
2751 SelectionDAG &DAG) const {
2753 return DAG.getNode(ARMISD::EH_SJLJ_SETUP_DISPATCH, dl, MVT::Other,
2758 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
2759 const ARMSubtarget *Subtarget) const {
2760 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2763 default: return SDValue(); // Don't custom lower most intrinsics.
2764 case Intrinsic::arm_rbit: {
2765 assert(Op.getOperand(1).getValueType() == MVT::i32 &&
2766 "RBIT intrinsic must have i32 type!");
2767 return DAG.getNode(ARMISD::RBIT, dl, MVT::i32, Op.getOperand(1));
2769 case Intrinsic::arm_thread_pointer: {
2770 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2771 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2773 case Intrinsic::eh_sjlj_lsda: {
2774 MachineFunction &MF = DAG.getMachineFunction();
2775 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2776 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2777 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2778 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2780 unsigned PCAdj = (RelocM != Reloc::PIC_)
2781 ? 0 : (Subtarget->isThumb() ? 4 : 8);
2782 ARMConstantPoolValue *CPV =
2783 ARMConstantPoolConstant::Create(MF.getFunction(), ARMPCLabelIndex,
2784 ARMCP::CPLSDA, PCAdj);
2785 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2786 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2788 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2789 MachinePointerInfo::getConstantPool(),
2790 false, false, false, 0);
2792 if (RelocM == Reloc::PIC_) {
2793 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, dl, MVT::i32);
2794 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2798 case Intrinsic::arm_neon_vmulls:
2799 case Intrinsic::arm_neon_vmullu: {
2800 unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vmulls)
2801 ? ARMISD::VMULLs : ARMISD::VMULLu;
2802 return DAG.getNode(NewOpc, SDLoc(Op), Op.getValueType(),
2803 Op.getOperand(1), Op.getOperand(2));
2805 case Intrinsic::arm_neon_vminnm:
2806 case Intrinsic::arm_neon_vmaxnm: {
2807 unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vminnm)
2808 ? ISD::FMINNUM : ISD::FMAXNUM;
2809 return DAG.getNode(NewOpc, SDLoc(Op), Op.getValueType(),
2810 Op.getOperand(1), Op.getOperand(2));
2812 case Intrinsic::arm_neon_vmins:
2813 case Intrinsic::arm_neon_vmaxs: {
2814 // v{min,max}s is overloaded between signed integers and floats.
2815 if (!Op.getValueType().isFloatingPoint())
2817 unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vmins)
2818 ? ISD::FMINNAN : ISD::FMAXNAN;
2819 return DAG.getNode(NewOpc, SDLoc(Op), Op.getValueType(),
2820 Op.getOperand(1), Op.getOperand(2));
2825 static SDValue LowerATOMIC_FENCE(SDValue Op, SelectionDAG &DAG,
2826 const ARMSubtarget *Subtarget) {
2827 // FIXME: handle "fence singlethread" more efficiently.
2829 if (!Subtarget->hasDataBarrier()) {
2830 // Some ARMv6 cpus can support data barriers with an mcr instruction.
2831 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2833 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2834 "Unexpected ISD::ATOMIC_FENCE encountered. Should be libcall!");
2835 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2836 DAG.getConstant(0, dl, MVT::i32));
2839 ConstantSDNode *OrdN = cast<ConstantSDNode>(Op.getOperand(1));
2840 AtomicOrdering Ord = static_cast<AtomicOrdering>(OrdN->getZExtValue());
2841 ARM_MB::MemBOpt Domain = ARM_MB::ISH;
2842 if (Subtarget->isMClass()) {
2843 // Only a full system barrier exists in the M-class architectures.
2844 Domain = ARM_MB::SY;
2845 } else if (Subtarget->isSwift() && Ord == Release) {
2846 // Swift happens to implement ISHST barriers in a way that's compatible with
2847 // Release semantics but weaker than ISH so we'd be fools not to use
2848 // it. Beware: other processors probably don't!
2849 Domain = ARM_MB::ISHST;
2852 return DAG.getNode(ISD::INTRINSIC_VOID, dl, MVT::Other, Op.getOperand(0),
2853 DAG.getConstant(Intrinsic::arm_dmb, dl, MVT::i32),
2854 DAG.getConstant(Domain, dl, MVT::i32));
2857 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG,
2858 const ARMSubtarget *Subtarget) {
2859 // ARM pre v5TE and Thumb1 does not have preload instructions.
2860 if (!(Subtarget->isThumb2() ||
2861 (!Subtarget->isThumb1Only() && Subtarget->hasV5TEOps())))
2862 // Just preserve the chain.
2863 return Op.getOperand(0);
2866 unsigned isRead = ~cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() & 1;
2868 (!Subtarget->hasV7Ops() || !Subtarget->hasMPExtension()))
2869 // ARMv7 with MP extension has PLDW.
2870 return Op.getOperand(0);
2872 unsigned isData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
2873 if (Subtarget->isThumb()) {
2875 isRead = ~isRead & 1;
2876 isData = ~isData & 1;
2879 return DAG.getNode(ARMISD::PRELOAD, dl, MVT::Other, Op.getOperand(0),
2880 Op.getOperand(1), DAG.getConstant(isRead, dl, MVT::i32),
2881 DAG.getConstant(isData, dl, MVT::i32));
2884 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
2885 MachineFunction &MF = DAG.getMachineFunction();
2886 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
2888 // vastart just stores the address of the VarArgsFrameIndex slot into the
2889 // memory location argument.
2891 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy(DAG.getDataLayout());
2892 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2893 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
2894 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
2895 MachinePointerInfo(SV), false, false, 0);
2899 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
2900 SDValue &Root, SelectionDAG &DAG,
2902 MachineFunction &MF = DAG.getMachineFunction();
2903 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2905 const TargetRegisterClass *RC;
2906 if (AFI->isThumb1OnlyFunction())
2907 RC = &ARM::tGPRRegClass;
2909 RC = &ARM::GPRRegClass;
2911 // Transform the arguments stored in physical registers into virtual ones.
2912 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2913 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2916 if (NextVA.isMemLoc()) {
2917 MachineFrameInfo *MFI = MF.getFrameInfo();
2918 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true);
2920 // Create load node to retrieve arguments from the stack.
2921 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
2922 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
2923 MachinePointerInfo::getFixedStack(FI),
2924 false, false, false, 0);
2926 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
2927 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2929 if (!Subtarget->isLittle())
2930 std::swap (ArgValue, ArgValue2);
2931 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
2934 // The remaining GPRs hold either the beginning of variable-argument
2935 // data, or the beginning of an aggregate passed by value (usually
2936 // byval). Either way, we allocate stack slots adjacent to the data
2937 // provided by our caller, and store the unallocated registers there.
2938 // If this is a variadic function, the va_list pointer will begin with
2939 // these values; otherwise, this reassembles a (byval) structure that
2940 // was split between registers and memory.
2941 // Return: The frame index registers were stored into.
2943 ARMTargetLowering::StoreByValRegs(CCState &CCInfo, SelectionDAG &DAG,
2944 SDLoc dl, SDValue &Chain,
2945 const Value *OrigArg,
2946 unsigned InRegsParamRecordIdx,
2948 unsigned ArgSize) const {
2949 // Currently, two use-cases possible:
2950 // Case #1. Non-var-args function, and we meet first byval parameter.
2951 // Setup first unallocated register as first byval register;
2952 // eat all remained registers
2953 // (these two actions are performed by HandleByVal method).
2954 // Then, here, we initialize stack frame with
2955 // "store-reg" instructions.
2956 // Case #2. Var-args function, that doesn't contain byval parameters.
2957 // The same: eat all remained unallocated registers,
2958 // initialize stack frame.
2960 MachineFunction &MF = DAG.getMachineFunction();
2961 MachineFrameInfo *MFI = MF.getFrameInfo();
2962 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2963 unsigned RBegin, REnd;
2964 if (InRegsParamRecordIdx < CCInfo.getInRegsParamsCount()) {
2965 CCInfo.getInRegsParamInfo(InRegsParamRecordIdx, RBegin, REnd);
2967 unsigned RBeginIdx = CCInfo.getFirstUnallocated(GPRArgRegs);
2968 RBegin = RBeginIdx == 4 ? (unsigned)ARM::R4 : GPRArgRegs[RBeginIdx];
2973 ArgOffset = -4 * (ARM::R4 - RBegin);
2975 auto PtrVT = getPointerTy(DAG.getDataLayout());
2976 int FrameIndex = MFI->CreateFixedObject(ArgSize, ArgOffset, false);
2977 SDValue FIN = DAG.getFrameIndex(FrameIndex, PtrVT);
2979 SmallVector<SDValue, 4> MemOps;
2980 const TargetRegisterClass *RC =
2981 AFI->isThumb1OnlyFunction() ? &ARM::tGPRRegClass : &ARM::GPRRegClass;
2983 for (unsigned Reg = RBegin, i = 0; Reg < REnd; ++Reg, ++i) {
2984 unsigned VReg = MF.addLiveIn(Reg, RC);
2985 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2987 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2988 MachinePointerInfo(OrigArg, 4 * i), false, false, 0);
2989 MemOps.push_back(Store);
2990 FIN = DAG.getNode(ISD::ADD, dl, PtrVT, FIN, DAG.getConstant(4, dl, PtrVT));
2993 if (!MemOps.empty())
2994 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOps);
2998 // Setup stack frame, the va_list pointer will start from.
3000 ARMTargetLowering::VarArgStyleRegisters(CCState &CCInfo, SelectionDAG &DAG,
3001 SDLoc dl, SDValue &Chain,
3003 unsigned TotalArgRegsSaveSize,
3004 bool ForceMutable) const {
3005 MachineFunction &MF = DAG.getMachineFunction();
3006 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
3008 // Try to store any remaining integer argument regs
3009 // to their spots on the stack so that they may be loaded by deferencing
3010 // the result of va_next.
3011 // If there is no regs to be stored, just point address after last
3012 // argument passed via stack.
3013 int FrameIndex = StoreByValRegs(CCInfo, DAG, dl, Chain, nullptr,
3014 CCInfo.getInRegsParamsCount(),
3015 CCInfo.getNextStackOffset(), 4);
3016 AFI->setVarArgsFrameIndex(FrameIndex);
3020 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
3021 CallingConv::ID CallConv, bool isVarArg,
3022 const SmallVectorImpl<ISD::InputArg>
3024 SDLoc dl, SelectionDAG &DAG,
3025 SmallVectorImpl<SDValue> &InVals)
3027 MachineFunction &MF = DAG.getMachineFunction();
3028 MachineFrameInfo *MFI = MF.getFrameInfo();
3030 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
3032 // Assign locations to all of the incoming arguments.
3033 SmallVector<CCValAssign, 16> ArgLocs;
3034 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs,
3035 *DAG.getContext(), Prologue);
3036 CCInfo.AnalyzeFormalArguments(Ins,
3037 CCAssignFnForNode(CallConv, /* Return*/ false,
3040 SmallVector<SDValue, 16> ArgValues;
3042 Function::const_arg_iterator CurOrigArg = MF.getFunction()->arg_begin();
3043 unsigned CurArgIdx = 0;
3045 // Initially ArgRegsSaveSize is zero.
3046 // Then we increase this value each time we meet byval parameter.
3047 // We also increase this value in case of varargs function.
3048 AFI->setArgRegsSaveSize(0);
3050 // Calculate the amount of stack space that we need to allocate to store
3051 // byval and variadic arguments that are passed in registers.
3052 // We need to know this before we allocate the first byval or variadic
3053 // argument, as they will be allocated a stack slot below the CFA (Canonical
3054 // Frame Address, the stack pointer at entry to the function).
3055 unsigned ArgRegBegin = ARM::R4;
3056 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3057 if (CCInfo.getInRegsParamsProcessed() >= CCInfo.getInRegsParamsCount())
3060 CCValAssign &VA = ArgLocs[i];
3061 unsigned Index = VA.getValNo();
3062 ISD::ArgFlagsTy Flags = Ins[Index].Flags;
3063 if (!Flags.isByVal())
3066 assert(VA.isMemLoc() && "unexpected byval pointer in reg");
3067 unsigned RBegin, REnd;
3068 CCInfo.getInRegsParamInfo(CCInfo.getInRegsParamsProcessed(), RBegin, REnd);
3069 ArgRegBegin = std::min(ArgRegBegin, RBegin);
3071 CCInfo.nextInRegsParam();
3073 CCInfo.rewindByValRegsInfo();
3075 int lastInsIndex = -1;
3076 if (isVarArg && MFI->hasVAStart()) {
3077 unsigned RegIdx = CCInfo.getFirstUnallocated(GPRArgRegs);
3078 if (RegIdx != array_lengthof(GPRArgRegs))
3079 ArgRegBegin = std::min(ArgRegBegin, (unsigned)GPRArgRegs[RegIdx]);
3082 unsigned TotalArgRegsSaveSize = 4 * (ARM::R4 - ArgRegBegin);
3083 AFI->setArgRegsSaveSize(TotalArgRegsSaveSize);
3084 auto PtrVT = getPointerTy(DAG.getDataLayout());
3086 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3087 CCValAssign &VA = ArgLocs[i];
3088 if (Ins[VA.getValNo()].isOrigArg()) {
3089 std::advance(CurOrigArg,
3090 Ins[VA.getValNo()].getOrigArgIndex() - CurArgIdx);
3091 CurArgIdx = Ins[VA.getValNo()].getOrigArgIndex();
3093 // Arguments stored in registers.
3094 if (VA.isRegLoc()) {
3095 EVT RegVT = VA.getLocVT();
3097 if (VA.needsCustom()) {
3098 // f64 and vector types are split up into multiple registers or
3099 // combinations of registers and stack slots.
3100 if (VA.getLocVT() == MVT::v2f64) {
3101 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
3103 VA = ArgLocs[++i]; // skip ahead to next loc
3105 if (VA.isMemLoc()) {
3106 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true);
3107 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
3108 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
3109 MachinePointerInfo::getFixedStack(FI),
3110 false, false, false, 0);
3112 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
3115 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
3116 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
3117 ArgValue, ArgValue1,
3118 DAG.getIntPtrConstant(0, dl));
3119 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
3120 ArgValue, ArgValue2,
3121 DAG.getIntPtrConstant(1, dl));
3123 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
3126 const TargetRegisterClass *RC;
3128 if (RegVT == MVT::f32)
3129 RC = &ARM::SPRRegClass;
3130 else if (RegVT == MVT::f64)
3131 RC = &ARM::DPRRegClass;
3132 else if (RegVT == MVT::v2f64)
3133 RC = &ARM::QPRRegClass;
3134 else if (RegVT == MVT::i32)
3135 RC = AFI->isThumb1OnlyFunction() ? &ARM::tGPRRegClass
3136 : &ARM::GPRRegClass;
3138 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
3140 // Transform the arguments in physical registers into virtual ones.
3141 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
3142 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
3145 // If this is an 8 or 16-bit value, it is really passed promoted
3146 // to 32 bits. Insert an assert[sz]ext to capture this, then
3147 // truncate to the right size.
3148 switch (VA.getLocInfo()) {
3149 default: llvm_unreachable("Unknown loc info!");
3150 case CCValAssign::Full: break;
3151 case CCValAssign::BCvt:
3152 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
3154 case CCValAssign::SExt:
3155 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
3156 DAG.getValueType(VA.getValVT()));
3157 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
3159 case CCValAssign::ZExt:
3160 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
3161 DAG.getValueType(VA.getValVT()));
3162 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
3166 InVals.push_back(ArgValue);
3168 } else { // VA.isRegLoc()
3171 assert(VA.isMemLoc());
3172 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
3174 int index = VA.getValNo();
3176 // Some Ins[] entries become multiple ArgLoc[] entries.
3177 // Process them only once.
3178 if (index != lastInsIndex)
3180 ISD::ArgFlagsTy Flags = Ins[index].Flags;
3181 // FIXME: For now, all byval parameter objects are marked mutable.
3182 // This can be changed with more analysis.
3183 // In case of tail call optimization mark all arguments mutable.
3184 // Since they could be overwritten by lowering of arguments in case of
3186 if (Flags.isByVal()) {
3187 assert(Ins[index].isOrigArg() &&
3188 "Byval arguments cannot be implicit");
3189 unsigned CurByValIndex = CCInfo.getInRegsParamsProcessed();
3191 int FrameIndex = StoreByValRegs(CCInfo, DAG, dl, Chain, CurOrigArg,
3192 CurByValIndex, VA.getLocMemOffset(),
3193 Flags.getByValSize());
3194 InVals.push_back(DAG.getFrameIndex(FrameIndex, PtrVT));
3195 CCInfo.nextInRegsParam();
3197 unsigned FIOffset = VA.getLocMemOffset();
3198 int FI = MFI->CreateFixedObject(VA.getLocVT().getSizeInBits()/8,
3201 // Create load nodes to retrieve arguments from the stack.
3202 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
3203 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
3204 MachinePointerInfo::getFixedStack(FI),
3205 false, false, false, 0));
3207 lastInsIndex = index;
3213 if (isVarArg && MFI->hasVAStart())
3214 VarArgStyleRegisters(CCInfo, DAG, dl, Chain,
3215 CCInfo.getNextStackOffset(),
3216 TotalArgRegsSaveSize);
3218 AFI->setArgumentStackSize(CCInfo.getNextStackOffset());
3223 /// isFloatingPointZero - Return true if this is +0.0.
3224 static bool isFloatingPointZero(SDValue Op) {
3225 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
3226 return CFP->getValueAPF().isPosZero();
3227 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
3228 // Maybe this has already been legalized into the constant pool?
3229 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
3230 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
3231 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
3232 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
3233 return CFP->getValueAPF().isPosZero();
3235 } else if (Op->getOpcode() == ISD::BITCAST &&
3236 Op->getValueType(0) == MVT::f64) {
3237 // Handle (ISD::BITCAST (ARMISD::VMOVIMM (ISD::TargetConstant 0)) MVT::f64)
3238 // created by LowerConstantFP().
3239 SDValue BitcastOp = Op->getOperand(0);
3240 if (BitcastOp->getOpcode() == ARMISD::VMOVIMM) {
3241 SDValue MoveOp = BitcastOp->getOperand(0);
3242 if (MoveOp->getOpcode() == ISD::TargetConstant &&
3243 cast<ConstantSDNode>(MoveOp)->getZExtValue() == 0) {
3251 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
3252 /// the given operands.
3254 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
3255 SDValue &ARMcc, SelectionDAG &DAG,
3257 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
3258 unsigned C = RHSC->getZExtValue();
3259 if (!isLegalICmpImmediate(C)) {
3260 // Constant does not fit, try adjusting it by one?
3265 if (C != 0x80000000 && isLegalICmpImmediate(C-1)) {
3266 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
3267 RHS = DAG.getConstant(C - 1, dl, MVT::i32);
3272 if (C != 0 && isLegalICmpImmediate(C-1)) {
3273 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
3274 RHS = DAG.getConstant(C - 1, dl, MVT::i32);
3279 if (C != 0x7fffffff && isLegalICmpImmediate(C+1)) {
3280 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
3281 RHS = DAG.getConstant(C + 1, dl, MVT::i32);
3286 if (C != 0xffffffff && isLegalICmpImmediate(C+1)) {
3287 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
3288 RHS = DAG.getConstant(C + 1, dl, MVT::i32);
3295 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3296 ARMISD::NodeType CompareType;
3299 CompareType = ARMISD::CMP;
3304 CompareType = ARMISD::CMPZ;
3307 ARMcc = DAG.getConstant(CondCode, dl, MVT::i32);
3308 return DAG.getNode(CompareType, dl, MVT::Glue, LHS, RHS);
3311 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
3313 ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
3315 assert(!Subtarget->isFPOnlySP() || RHS.getValueType() != MVT::f64);
3317 if (!isFloatingPointZero(RHS))
3318 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Glue, LHS, RHS);
3320 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Glue, LHS);
3321 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Glue, Cmp);
3324 /// duplicateCmp - Glue values can have only one use, so this function
3325 /// duplicates a comparison node.
3327 ARMTargetLowering::duplicateCmp(SDValue Cmp, SelectionDAG &DAG) const {
3328 unsigned Opc = Cmp.getOpcode();
3330 if (Opc == ARMISD::CMP || Opc == ARMISD::CMPZ)
3331 return DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
3333 assert(Opc == ARMISD::FMSTAT && "unexpected comparison operation");
3334 Cmp = Cmp.getOperand(0);
3335 Opc = Cmp.getOpcode();
3336 if (Opc == ARMISD::CMPFP)
3337 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
3339 assert(Opc == ARMISD::CMPFPw0 && "unexpected operand of FMSTAT");
3340 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0));
3342 return DAG.getNode(ARMISD::FMSTAT, DL, MVT::Glue, Cmp);
3345 std::pair<SDValue, SDValue>
3346 ARMTargetLowering::getARMXALUOOp(SDValue Op, SelectionDAG &DAG,
3347 SDValue &ARMcc) const {
3348 assert(Op.getValueType() == MVT::i32 && "Unsupported value type");
3350 SDValue Value, OverflowCmp;
3351 SDValue LHS = Op.getOperand(0);
3352 SDValue RHS = Op.getOperand(1);
3355 // FIXME: We are currently always generating CMPs because we don't support
3356 // generating CMN through the backend. This is not as good as the natural
3357 // CMP case because it causes a register dependency and cannot be folded
3360 switch (Op.getOpcode()) {
3362 llvm_unreachable("Unknown overflow instruction!");
3364 ARMcc = DAG.getConstant(ARMCC::VC, dl, MVT::i32);
3365 Value = DAG.getNode(ISD::ADD, dl, Op.getValueType(), LHS, RHS);
3366 OverflowCmp = DAG.getNode(ARMISD::CMP, dl, MVT::Glue, Value, LHS);
3369 ARMcc = DAG.getConstant(ARMCC::HS, dl, MVT::i32);
3370 Value = DAG.getNode(ISD::ADD, dl, Op.getValueType(), LHS, RHS);
3371 OverflowCmp = DAG.getNode(ARMISD::CMP, dl, MVT::Glue, Value, LHS);
3374 ARMcc = DAG.getConstant(ARMCC::VC, dl, MVT::i32);
3375 Value = DAG.getNode(ISD::SUB, dl, Op.getValueType(), LHS, RHS);
3376 OverflowCmp = DAG.getNode(ARMISD::CMP, dl, MVT::Glue, LHS, RHS);
3379 ARMcc = DAG.getConstant(ARMCC::HS, dl, MVT::i32);
3380 Value = DAG.getNode(ISD::SUB, dl, Op.getValueType(), LHS, RHS);
3381 OverflowCmp = DAG.getNode(ARMISD::CMP, dl, MVT::Glue, LHS, RHS);
3385 return std::make_pair(Value, OverflowCmp);
3390 ARMTargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) const {
3391 // Let legalize expand this if it isn't a legal type yet.
3392 if (!DAG.getTargetLoweringInfo().isTypeLegal(Op.getValueType()))
3395 SDValue Value, OverflowCmp;
3397 std::tie(Value, OverflowCmp) = getARMXALUOOp(Op, DAG, ARMcc);
3398 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3400 // We use 0 and 1 as false and true values.
3401 SDValue TVal = DAG.getConstant(1, dl, MVT::i32);
3402 SDValue FVal = DAG.getConstant(0, dl, MVT::i32);
3403 EVT VT = Op.getValueType();
3405 SDValue Overflow = DAG.getNode(ARMISD::CMOV, dl, VT, TVal, FVal,
3406 ARMcc, CCR, OverflowCmp);
3408 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
3409 return DAG.getNode(ISD::MERGE_VALUES, dl, VTs, Value, Overflow);
3413 SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
3414 SDValue Cond = Op.getOperand(0);
3415 SDValue SelectTrue = Op.getOperand(1);
3416 SDValue SelectFalse = Op.getOperand(2);
3418 unsigned Opc = Cond.getOpcode();
3420 if (Cond.getResNo() == 1 &&
3421 (Opc == ISD::SADDO || Opc == ISD::UADDO || Opc == ISD::SSUBO ||
3422 Opc == ISD::USUBO)) {
3423 if (!DAG.getTargetLoweringInfo().isTypeLegal(Cond->getValueType(0)))
3426 SDValue Value, OverflowCmp;
3428 std::tie(Value, OverflowCmp) = getARMXALUOOp(Cond, DAG, ARMcc);
3429 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3430 EVT VT = Op.getValueType();
3432 return getCMOV(dl, VT, SelectTrue, SelectFalse, ARMcc, CCR,
3438 // (select (cmov 1, 0, cond), t, f) -> (cmov t, f, cond)
3439 // (select (cmov 0, 1, cond), t, f) -> (cmov f, t, cond)
3441 if (Cond.getOpcode() == ARMISD::CMOV && Cond.hasOneUse()) {
3442 const ConstantSDNode *CMOVTrue =
3443 dyn_cast<ConstantSDNode>(Cond.getOperand(0));
3444 const ConstantSDNode *CMOVFalse =
3445 dyn_cast<ConstantSDNode>(Cond.getOperand(1));
3447 if (CMOVTrue && CMOVFalse) {
3448 unsigned CMOVTrueVal = CMOVTrue->getZExtValue();
3449 unsigned CMOVFalseVal = CMOVFalse->getZExtValue();
3453 if (CMOVTrueVal == 1 && CMOVFalseVal == 0) {
3455 False = SelectFalse;
3456 } else if (CMOVTrueVal == 0 && CMOVFalseVal == 1) {
3461 if (True.getNode() && False.getNode()) {
3462 EVT VT = Op.getValueType();
3463 SDValue ARMcc = Cond.getOperand(2);
3464 SDValue CCR = Cond.getOperand(3);
3465 SDValue Cmp = duplicateCmp(Cond.getOperand(4), DAG);
3466 assert(True.getValueType() == VT);
3467 return getCMOV(dl, VT, True, False, ARMcc, CCR, Cmp, DAG);
3472 // ARM's BooleanContents value is UndefinedBooleanContent. Mask out the
3473 // undefined bits before doing a full-word comparison with zero.
3474 Cond = DAG.getNode(ISD::AND, dl, Cond.getValueType(), Cond,
3475 DAG.getConstant(1, dl, Cond.getValueType()));
3477 return DAG.getSelectCC(dl, Cond,
3478 DAG.getConstant(0, dl, Cond.getValueType()),
3479 SelectTrue, SelectFalse, ISD::SETNE);
3482 static void checkVSELConstraints(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
3483 bool &swpCmpOps, bool &swpVselOps) {
3484 // Start by selecting the GE condition code for opcodes that return true for
3486 if (CC == ISD::SETUGE || CC == ISD::SETOGE || CC == ISD::SETOLE ||
3488 CondCode = ARMCC::GE;
3490 // and GT for opcodes that return false for 'equality'.
3491 else if (CC == ISD::SETUGT || CC == ISD::SETOGT || CC == ISD::SETOLT ||
3493 CondCode = ARMCC::GT;
3495 // Since we are constrained to GE/GT, if the opcode contains 'less', we need
3496 // to swap the compare operands.
3497 if (CC == ISD::SETOLE || CC == ISD::SETULE || CC == ISD::SETOLT ||
3501 // Both GT and GE are ordered comparisons, and return false for 'unordered'.
3502 // If we have an unordered opcode, we need to swap the operands to the VSEL
3503 // instruction (effectively negating the condition).
3505 // This also has the effect of swapping which one of 'less' or 'greater'
3506 // returns true, so we also swap the compare operands. It also switches
3507 // whether we return true for 'equality', so we compensate by picking the
3508 // opposite condition code to our original choice.
3509 if (CC == ISD::SETULE || CC == ISD::SETULT || CC == ISD::SETUGE ||
3510 CC == ISD::SETUGT) {
3511 swpCmpOps = !swpCmpOps;
3512 swpVselOps = !swpVselOps;
3513 CondCode = CondCode == ARMCC::GT ? ARMCC::GE : ARMCC::GT;
3516 // 'ordered' is 'anything but unordered', so use the VS condition code and
3517 // swap the VSEL operands.
3518 if (CC == ISD::SETO) {
3519 CondCode = ARMCC::VS;
3523 // 'unordered or not equal' is 'anything but equal', so use the EQ condition
3524 // code and swap the VSEL operands.
3525 if (CC == ISD::SETUNE) {
3526 CondCode = ARMCC::EQ;
3531 SDValue ARMTargetLowering::getCMOV(SDLoc dl, EVT VT, SDValue FalseVal,
3532 SDValue TrueVal, SDValue ARMcc, SDValue CCR,
3533 SDValue Cmp, SelectionDAG &DAG) const {
3534 if (Subtarget->isFPOnlySP() && VT == MVT::f64) {
3535 FalseVal = DAG.getNode(ARMISD::VMOVRRD, dl,
3536 DAG.getVTList(MVT::i32, MVT::i32), FalseVal);
3537 TrueVal = DAG.getNode(ARMISD::VMOVRRD, dl,
3538 DAG.getVTList(MVT::i32, MVT::i32), TrueVal);
3540 SDValue TrueLow = TrueVal.getValue(0);
3541 SDValue TrueHigh = TrueVal.getValue(1);
3542 SDValue FalseLow = FalseVal.getValue(0);
3543 SDValue FalseHigh = FalseVal.getValue(1);
3545 SDValue Low = DAG.getNode(ARMISD::CMOV, dl, MVT::i32, FalseLow, TrueLow,
3547 SDValue High = DAG.getNode(ARMISD::CMOV, dl, MVT::i32, FalseHigh, TrueHigh,
3548 ARMcc, CCR, duplicateCmp(Cmp, DAG));
3550 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Low, High);
3552 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,
3557 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
3558 EVT VT = Op.getValueType();
3559 SDValue LHS = Op.getOperand(0);
3560 SDValue RHS = Op.getOperand(1);
3561 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
3562 SDValue TrueVal = Op.getOperand(2);
3563 SDValue FalseVal = Op.getOperand(3);
3566 if (Subtarget->isFPOnlySP() && LHS.getValueType() == MVT::f64) {
3567 DAG.getTargetLoweringInfo().softenSetCCOperands(DAG, MVT::f64, LHS, RHS, CC,
3570 // If softenSetCCOperands only returned one value, we should compare it to
3572 if (!RHS.getNode()) {
3573 RHS = DAG.getConstant(0, dl, LHS.getValueType());
3578 if (LHS.getValueType() == MVT::i32) {
3579 // Try to generate VSEL on ARMv8.
3580 // The VSEL instruction can't use all the usual ARM condition
3581 // codes: it only has two bits to select the condition code, so it's
3582 // constrained to use only GE, GT, VS and EQ.
3584 // To implement all the various ISD::SETXXX opcodes, we sometimes need to
3585 // swap the operands of the previous compare instruction (effectively
3586 // inverting the compare condition, swapping 'less' and 'greater') and
3587 // sometimes need to swap the operands to the VSEL (which inverts the
3588 // condition in the sense of firing whenever the previous condition didn't)
3589 if (Subtarget->hasFPARMv8() && (TrueVal.getValueType() == MVT::f32 ||
3590 TrueVal.getValueType() == MVT::f64)) {
3591 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3592 if (CondCode == ARMCC::LT || CondCode == ARMCC::LE ||
3593 CondCode == ARMCC::VC || CondCode == ARMCC::NE) {
3594 CC = ISD::getSetCCInverse(CC, true);
3595 std::swap(TrueVal, FalseVal);
3600 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3601 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3602 return getCMOV(dl, VT, FalseVal, TrueVal, ARMcc, CCR, Cmp, DAG);
3605 ARMCC::CondCodes CondCode, CondCode2;
3606 FPCCToARMCC(CC, CondCode, CondCode2);
3608 // Try to generate VMAXNM/VMINNM on ARMv8.
3609 if (Subtarget->hasFPARMv8() && (TrueVal.getValueType() == MVT::f32 ||
3610 TrueVal.getValueType() == MVT::f64)) {
3611 // We can use VMAXNM/VMINNM for a compare followed by a select with the
3612 // same operands, as follows:
3613 // c = fcmp [?gt, ?ge, ?lt, ?le] a, b
3615 // In NoNaNsFPMath the CC will have been changed from, e.g., 'ogt' to 'gt'.
3616 bool swapSides = false;
3617 if (!getTargetMachine().Options.NoNaNsFPMath) {
3618 // transformability may depend on which way around we compare
3626 // the non-NaN should be RHS
3627 swapSides = DAG.isKnownNeverNaN(LHS) && !DAG.isKnownNeverNaN(RHS);
3633 // the non-NaN should be LHS
3634 swapSides = DAG.isKnownNeverNaN(RHS) && !DAG.isKnownNeverNaN(LHS);
3638 swapSides = swapSides || (LHS == FalseVal && RHS == TrueVal);
3640 CC = ISD::getSetCCSwappedOperands(CC);
3641 std::swap(LHS, RHS);
3643 if (LHS == TrueVal && RHS == FalseVal) {
3644 bool canTransform = true;
3645 // FIXME: FastMathFlags::noSignedZeros() doesn't appear reachable from here
3646 if (!getTargetMachine().Options.UnsafeFPMath &&
3647 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
3648 const ConstantFPSDNode *Zero;
3655 // RHS must not be -0
3656 canTransform = (Zero = dyn_cast<ConstantFPSDNode>(RHS)) &&
3657 !Zero->isNegative();
3662 // LHS must not be -0
3663 canTransform = (Zero = dyn_cast<ConstantFPSDNode>(LHS)) &&
3664 !Zero->isNegative();
3669 // RHS must not be +0
3670 canTransform = (Zero = dyn_cast<ConstantFPSDNode>(RHS)) &&
3676 // LHS must not be +0
3677 canTransform = (Zero = dyn_cast<ConstantFPSDNode>(LHS)) &&
3683 // Note: If one of the elements in a pair is a number and the other
3684 // element is NaN, the corresponding result element is the number.
3685 // This is consistent with the IEEE 754-2008 standard.
3686 // Therefore, a > b ? a : b <=> vmax(a,b), if b is constant and a is NaN
3692 if (!DAG.isKnownNeverNaN(RHS))
3694 return DAG.getNode(ISD::FMAXNUM, dl, VT, LHS, RHS);
3697 if (!DAG.isKnownNeverNaN(LHS))
3701 return DAG.getNode(ISD::FMAXNUM, dl, VT, LHS, RHS);
3704 if (!DAG.isKnownNeverNaN(RHS))
3706 return DAG.getNode(ISD::FMINNUM, dl, VT, LHS, RHS);
3709 if (!DAG.isKnownNeverNaN(LHS))
3713 return DAG.getNode(ISD::FMINNUM, dl, VT, LHS, RHS);
3718 bool swpCmpOps = false;
3719 bool swpVselOps = false;
3720 checkVSELConstraints(CC, CondCode, swpCmpOps, swpVselOps);
3722 if (CondCode == ARMCC::GT || CondCode == ARMCC::GE ||
3723 CondCode == ARMCC::VS || CondCode == ARMCC::EQ) {
3725 std::swap(LHS, RHS);
3727 std::swap(TrueVal, FalseVal);
3731 SDValue ARMcc = DAG.getConstant(CondCode, dl, MVT::i32);
3732 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
3733 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3734 SDValue Result = getCMOV(dl, VT, FalseVal, TrueVal, ARMcc, CCR, Cmp, DAG);
3735 if (CondCode2 != ARMCC::AL) {
3736 SDValue ARMcc2 = DAG.getConstant(CondCode2, dl, MVT::i32);
3737 // FIXME: Needs another CMP because flag can have but one use.
3738 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
3739 Result = getCMOV(dl, VT, Result, TrueVal, ARMcc2, CCR, Cmp2, DAG);
3744 /// canChangeToInt - Given the fp compare operand, return true if it is suitable
3745 /// to morph to an integer compare sequence.
3746 static bool canChangeToInt(SDValue Op, bool &SeenZero,
3747 const ARMSubtarget *Subtarget) {
3748 SDNode *N = Op.getNode();
3749 if (!N->hasOneUse())
3750 // Otherwise it requires moving the value from fp to integer registers.
3752 if (!N->getNumValues())
3754 EVT VT = Op.getValueType();
3755 if (VT != MVT::f32 && !Subtarget->isFPBrccSlow())
3756 // f32 case is generally profitable. f64 case only makes sense when vcmpe +
3757 // vmrs are very slow, e.g. cortex-a8.
3760 if (isFloatingPointZero(Op)) {
3764 return ISD::isNormalLoad(N);
3767 static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) {
3768 if (isFloatingPointZero(Op))
3769 return DAG.getConstant(0, SDLoc(Op), MVT::i32);
3771 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op))
3772 return DAG.getLoad(MVT::i32, SDLoc(Op),
3773 Ld->getChain(), Ld->getBasePtr(), Ld->getPointerInfo(),
3774 Ld->isVolatile(), Ld->isNonTemporal(),
3775 Ld->isInvariant(), Ld->getAlignment());
3777 llvm_unreachable("Unknown VFP cmp argument!");
3780 static void expandf64Toi32(SDValue Op, SelectionDAG &DAG,
3781 SDValue &RetVal1, SDValue &RetVal2) {
3784 if (isFloatingPointZero(Op)) {
3785 RetVal1 = DAG.getConstant(0, dl, MVT::i32);
3786 RetVal2 = DAG.getConstant(0, dl, MVT::i32);
3790 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) {
3791 SDValue Ptr = Ld->getBasePtr();
3792 RetVal1 = DAG.getLoad(MVT::i32, dl,
3793 Ld->getChain(), Ptr,
3794 Ld->getPointerInfo(),
3795 Ld->isVolatile(), Ld->isNonTemporal(),
3796 Ld->isInvariant(), Ld->getAlignment());
3798 EVT PtrType = Ptr.getValueType();
3799 unsigned NewAlign = MinAlign(Ld->getAlignment(), 4);
3800 SDValue NewPtr = DAG.getNode(ISD::ADD, dl,
3801 PtrType, Ptr, DAG.getConstant(4, dl, PtrType));
3802 RetVal2 = DAG.getLoad(MVT::i32, dl,
3803 Ld->getChain(), NewPtr,
3804 Ld->getPointerInfo().getWithOffset(4),
3805 Ld->isVolatile(), Ld->isNonTemporal(),
3806 Ld->isInvariant(), NewAlign);
3810 llvm_unreachable("Unknown VFP cmp argument!");
3813 /// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some
3814 /// f32 and even f64 comparisons to integer ones.
3816 ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const {
3817 SDValue Chain = Op.getOperand(0);
3818 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3819 SDValue LHS = Op.getOperand(2);
3820 SDValue RHS = Op.getOperand(3);
3821 SDValue Dest = Op.getOperand(4);
3824 bool LHSSeenZero = false;
3825 bool LHSOk = canChangeToInt(LHS, LHSSeenZero, Subtarget);
3826 bool RHSSeenZero = false;
3827 bool RHSOk = canChangeToInt(RHS, RHSSeenZero, Subtarget);
3828 if (LHSOk && RHSOk && (LHSSeenZero || RHSSeenZero)) {
3829 // If unsafe fp math optimization is enabled and there are no other uses of
3830 // the CMP operands, and the condition code is EQ or NE, we can optimize it
3831 // to an integer comparison.
3832 if (CC == ISD::SETOEQ)
3834 else if (CC == ISD::SETUNE)
3837 SDValue Mask = DAG.getConstant(0x7fffffff, dl, MVT::i32);
3839 if (LHS.getValueType() == MVT::f32) {
3840 LHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3841 bitcastf32Toi32(LHS, DAG), Mask);
3842 RHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3843 bitcastf32Toi32(RHS, DAG), Mask);
3844 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3845 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3846 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3847 Chain, Dest, ARMcc, CCR, Cmp);
3852 expandf64Toi32(LHS, DAG, LHS1, LHS2);
3853 expandf64Toi32(RHS, DAG, RHS1, RHS2);
3854 LHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, LHS2, Mask);
3855 RHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, RHS2, Mask);
3856 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3857 ARMcc = DAG.getConstant(CondCode, dl, MVT::i32);
3858 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3859 SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest };
3860 return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops);
3866 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
3867 SDValue Chain = Op.getOperand(0);
3868 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3869 SDValue LHS = Op.getOperand(2);
3870 SDValue RHS = Op.getOperand(3);
3871 SDValue Dest = Op.getOperand(4);
3874 if (Subtarget->isFPOnlySP() && LHS.getValueType() == MVT::f64) {
3875 DAG.getTargetLoweringInfo().softenSetCCOperands(DAG, MVT::f64, LHS, RHS, CC,
3878 // If softenSetCCOperands only returned one value, we should compare it to
3880 if (!RHS.getNode()) {
3881 RHS = DAG.getConstant(0, dl, LHS.getValueType());
3886 if (LHS.getValueType() == MVT::i32) {
3888 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3889 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3890 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3891 Chain, Dest, ARMcc, CCR, Cmp);
3894 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3896 if (getTargetMachine().Options.UnsafeFPMath &&
3897 (CC == ISD::SETEQ || CC == ISD::SETOEQ ||
3898 CC == ISD::SETNE || CC == ISD::SETUNE)) {
3899 SDValue Result = OptimizeVFPBrcond(Op, DAG);
3900 if (Result.getNode())
3904 ARMCC::CondCodes CondCode, CondCode2;
3905 FPCCToARMCC(CC, CondCode, CondCode2);
3907 SDValue ARMcc = DAG.getConstant(CondCode, dl, MVT::i32);
3908 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
3909 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3910 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3911 SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp };
3912 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops);
3913 if (CondCode2 != ARMCC::AL) {
3914 ARMcc = DAG.getConstant(CondCode2, dl, MVT::i32);
3915 SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) };
3916 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops);
3921 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
3922 SDValue Chain = Op.getOperand(0);
3923 SDValue Table = Op.getOperand(1);
3924 SDValue Index = Op.getOperand(2);
3927 EVT PTy = getPointerTy(DAG.getDataLayout());
3928 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
3929 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
3930 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI);
3931 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, dl, PTy));
3932 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
3933 if (Subtarget->isThumb2()) {
3934 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
3935 // which does another jump to the destination. This also makes it easier
3936 // to translate it to TBB / TBH later.
3937 // FIXME: This might not work if the function is extremely large.
3938 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
3939 Addr, Op.getOperand(2), JTI);
3941 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
3942 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
3943 MachinePointerInfo::getJumpTable(),
3944 false, false, false, 0);
3945 Chain = Addr.getValue(1);
3946 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
3947 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI);
3949 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
3950 MachinePointerInfo::getJumpTable(),
3951 false, false, false, 0);
3952 Chain = Addr.getValue(1);
3953 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI);
3957 static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3958 EVT VT = Op.getValueType();
3961 if (Op.getValueType().getVectorElementType() == MVT::i32) {
3962 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::f32)
3964 return DAG.UnrollVectorOp(Op.getNode());
3967 assert(Op.getOperand(0).getValueType() == MVT::v4f32 &&
3968 "Invalid type for custom lowering!");
3969 if (VT != MVT::v4i16)
3970 return DAG.UnrollVectorOp(Op.getNode());
3972 Op = DAG.getNode(Op.getOpcode(), dl, MVT::v4i32, Op.getOperand(0));
3973 return DAG.getNode(ISD::TRUNCATE, dl, VT, Op);
3976 SDValue ARMTargetLowering::LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) const {
3977 EVT VT = Op.getValueType();
3979 return LowerVectorFP_TO_INT(Op, DAG);
3980 if (Subtarget->isFPOnlySP() && Op.getOperand(0).getValueType() == MVT::f64) {
3982 if (Op.getOpcode() == ISD::FP_TO_SINT)
3983 LC = RTLIB::getFPTOSINT(Op.getOperand(0).getValueType(),
3986 LC = RTLIB::getFPTOUINT(Op.getOperand(0).getValueType(),
3988 return makeLibCall(DAG, LC, Op.getValueType(), &Op.getOperand(0), 1,
3989 /*isSigned*/ false, SDLoc(Op)).first;
3995 static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3996 EVT VT = Op.getValueType();
3999 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::i32) {
4000 if (VT.getVectorElementType() == MVT::f32)
4002 return DAG.UnrollVectorOp(Op.getNode());
4005 assert(Op.getOperand(0).getValueType() == MVT::v4i16 &&
4006 "Invalid type for custom lowering!");
4007 if (VT != MVT::v4f32)
4008 return DAG.UnrollVectorOp(Op.getNode());
4012 switch (Op.getOpcode()) {
4013 default: llvm_unreachable("Invalid opcode!");
4014 case ISD::SINT_TO_FP:
4015 CastOpc = ISD::SIGN_EXTEND;
4016 Opc = ISD::SINT_TO_FP;
4018 case ISD::UINT_TO_FP:
4019 CastOpc = ISD::ZERO_EXTEND;
4020 Opc = ISD::UINT_TO_FP;
4024 Op = DAG.getNode(CastOpc, dl, MVT::v4i32, Op.getOperand(0));
4025 return DAG.getNode(Opc, dl, VT, Op);
4028 SDValue ARMTargetLowering::LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) const {
4029 EVT VT = Op.getValueType();
4031 return LowerVectorINT_TO_FP(Op, DAG);
4032 if (Subtarget->isFPOnlySP() && Op.getValueType() == MVT::f64) {
4034 if (Op.getOpcode() == ISD::SINT_TO_FP)
4035 LC = RTLIB::getSINTTOFP(Op.getOperand(0).getValueType(),
4038 LC = RTLIB::getUINTTOFP(Op.getOperand(0).getValueType(),
4040 return makeLibCall(DAG, LC, Op.getValueType(), &Op.getOperand(0), 1,
4041 /*isSigned*/ false, SDLoc(Op)).first;
4047 SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
4048 // Implement fcopysign with a fabs and a conditional fneg.
4049 SDValue Tmp0 = Op.getOperand(0);
4050 SDValue Tmp1 = Op.getOperand(1);
4052 EVT VT = Op.getValueType();
4053 EVT SrcVT = Tmp1.getValueType();
4054 bool InGPR = Tmp0.getOpcode() == ISD::BITCAST ||
4055 Tmp0.getOpcode() == ARMISD::VMOVDRR;
4056 bool UseNEON = !InGPR && Subtarget->hasNEON();
4059 // Use VBSL to copy the sign bit.
4060 unsigned EncodedVal = ARM_AM::createNEONModImm(0x6, 0x80);
4061 SDValue Mask = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v2i32,
4062 DAG.getTargetConstant(EncodedVal, dl, MVT::i32));
4063 EVT OpVT = (VT == MVT::f32) ? MVT::v2i32 : MVT::v1i64;
4065 Mask = DAG.getNode(ARMISD::VSHL, dl, OpVT,
4066 DAG.getNode(ISD::BITCAST, dl, OpVT, Mask),
4067 DAG.getConstant(32, dl, MVT::i32));
4068 else /*if (VT == MVT::f32)*/
4069 Tmp0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp0);
4070 if (SrcVT == MVT::f32) {
4071 Tmp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp1);
4073 Tmp1 = DAG.getNode(ARMISD::VSHL, dl, OpVT,
4074 DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1),
4075 DAG.getConstant(32, dl, MVT::i32));
4076 } else if (VT == MVT::f32)
4077 Tmp1 = DAG.getNode(ARMISD::VSHRu, dl, MVT::v1i64,
4078 DAG.getNode(ISD::BITCAST, dl, MVT::v1i64, Tmp1),
4079 DAG.getConstant(32, dl, MVT::i32));
4080 Tmp0 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp0);
4081 Tmp1 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1);
4083 SDValue AllOnes = DAG.getTargetConstant(ARM_AM::createNEONModImm(0xe, 0xff),
4085 AllOnes = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v8i8, AllOnes);
4086 SDValue MaskNot = DAG.getNode(ISD::XOR, dl, OpVT, Mask,
4087 DAG.getNode(ISD::BITCAST, dl, OpVT, AllOnes));
4089 SDValue Res = DAG.getNode(ISD::OR, dl, OpVT,
4090 DAG.getNode(ISD::AND, dl, OpVT, Tmp1, Mask),
4091 DAG.getNode(ISD::AND, dl, OpVT, Tmp0, MaskNot));
4092 if (VT == MVT::f32) {
4093 Res = DAG.getNode(ISD::BITCAST, dl, MVT::v2f32, Res);
4094 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, Res,
4095 DAG.getConstant(0, dl, MVT::i32));
4097 Res = DAG.getNode(ISD::BITCAST, dl, MVT::f64, Res);
4103 // Bitcast operand 1 to i32.
4104 if (SrcVT == MVT::f64)
4105 Tmp1 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
4107 Tmp1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp1);
4109 // Or in the signbit with integer operations.
4110 SDValue Mask1 = DAG.getConstant(0x80000000, dl, MVT::i32);
4111 SDValue Mask2 = DAG.getConstant(0x7fffffff, dl, MVT::i32);
4112 Tmp1 = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp1, Mask1);
4113 if (VT == MVT::f32) {
4114 Tmp0 = DAG.getNode(ISD::AND, dl, MVT::i32,
4115 DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp0), Mask2);
4116 return DAG.getNode(ISD::BITCAST, dl, MVT::f32,
4117 DAG.getNode(ISD::OR, dl, MVT::i32, Tmp0, Tmp1));
4120 // f64: Or the high part with signbit and then combine two parts.
4121 Tmp0 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
4123 SDValue Lo = Tmp0.getValue(0);
4124 SDValue Hi = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp0.getValue(1), Mask2);
4125 Hi = DAG.getNode(ISD::OR, dl, MVT::i32, Hi, Tmp1);
4126 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
4129 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
4130 MachineFunction &MF = DAG.getMachineFunction();
4131 MachineFrameInfo *MFI = MF.getFrameInfo();
4132 MFI->setReturnAddressIsTaken(true);
4134 if (verifyReturnAddressArgumentIsConstant(Op, DAG))
4137 EVT VT = Op.getValueType();
4139 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
4141 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
4142 SDValue Offset = DAG.getConstant(4, dl, MVT::i32);
4143 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
4144 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
4145 MachinePointerInfo(), false, false, false, 0);
4148 // Return LR, which contains the return address. Mark it an implicit live-in.
4149 unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32));
4150 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
4153 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
4154 const ARMBaseRegisterInfo &ARI =
4155 *static_cast<const ARMBaseRegisterInfo*>(RegInfo);
4156 MachineFunction &MF = DAG.getMachineFunction();
4157 MachineFrameInfo *MFI = MF.getFrameInfo();
4158 MFI->setFrameAddressIsTaken(true);
4160 EVT VT = Op.getValueType();
4161 SDLoc dl(Op); // FIXME probably not meaningful
4162 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
4163 unsigned FrameReg = ARI.getFrameRegister(MF);
4164 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
4166 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
4167 MachinePointerInfo(),
4168 false, false, false, 0);
4172 // FIXME? Maybe this could be a TableGen attribute on some registers and
4173 // this table could be generated automatically from RegInfo.
4174 unsigned ARMTargetLowering::getRegisterByName(const char* RegName, EVT VT,
4175 SelectionDAG &DAG) const {
4176 unsigned Reg = StringSwitch<unsigned>(RegName)
4177 .Case("sp", ARM::SP)
4181 report_fatal_error(Twine("Invalid register name \""
4182 + StringRef(RegName) + "\"."));
4185 // Result is 64 bit value so split into two 32 bit values and return as a
4187 static void ExpandREAD_REGISTER(SDNode *N, SmallVectorImpl<SDValue> &Results,
4188 SelectionDAG &DAG) {
4191 // This function is only supposed to be called for i64 type destination.
4192 assert(N->getValueType(0) == MVT::i64
4193 && "ExpandREAD_REGISTER called for non-i64 type result.");
4195 SDValue Read = DAG.getNode(ISD::READ_REGISTER, DL,
4196 DAG.getVTList(MVT::i32, MVT::i32, MVT::Other),
4200 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Read.getValue(0),
4202 Results.push_back(Read.getOperand(0));
4205 /// ExpandBITCAST - If the target supports VFP, this function is called to
4206 /// expand a bit convert where either the source or destination type is i64 to
4207 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
4208 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
4209 /// vectors), since the legalizer won't know what to do with that.
4210 static SDValue ExpandBITCAST(SDNode *N, SelectionDAG &DAG) {
4211 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4213 SDValue Op = N->getOperand(0);
4215 // This function is only supposed to be called for i64 types, either as the
4216 // source or destination of the bit convert.
4217 EVT SrcVT = Op.getValueType();
4218 EVT DstVT = N->getValueType(0);
4219 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
4220 "ExpandBITCAST called for non-i64 type");
4222 // Turn i64->f64 into VMOVDRR.
4223 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
4224 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
4225 DAG.getConstant(0, dl, MVT::i32));
4226 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
4227 DAG.getConstant(1, dl, MVT::i32));
4228 return DAG.getNode(ISD::BITCAST, dl, DstVT,
4229 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
4232 // Turn f64->i64 into VMOVRRD.
4233 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
4235 if (DAG.getDataLayout().isBigEndian() && SrcVT.isVector() &&
4236 SrcVT.getVectorNumElements() > 1)
4237 Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
4238 DAG.getVTList(MVT::i32, MVT::i32),
4239 DAG.getNode(ARMISD::VREV64, dl, SrcVT, Op));
4241 Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
4242 DAG.getVTList(MVT::i32, MVT::i32), Op);
4243 // Merge the pieces into a single i64 value.
4244 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
4250 /// getZeroVector - Returns a vector of specified type with all zero elements.
4251 /// Zero vectors are used to represent vector negation and in those cases
4252 /// will be implemented with the NEON VNEG instruction. However, VNEG does
4253 /// not support i64 elements, so sometimes the zero vectors will need to be
4254 /// explicitly constructed. Regardless, use a canonical VMOV to create the
4256 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, SDLoc dl) {
4257 assert(VT.isVector() && "Expected a vector type");
4258 // The canonical modified immediate encoding of a zero vector is....0!
4259 SDValue EncodedVal = DAG.getTargetConstant(0, dl, MVT::i32);
4260 EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
4261 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal);
4262 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4265 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
4266 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
4267 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
4268 SelectionDAG &DAG) const {
4269 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4270 EVT VT = Op.getValueType();
4271 unsigned VTBits = VT.getSizeInBits();
4273 SDValue ShOpLo = Op.getOperand(0);
4274 SDValue ShOpHi = Op.getOperand(1);
4275 SDValue ShAmt = Op.getOperand(2);
4277 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
4279 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
4281 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
4282 DAG.getConstant(VTBits, dl, MVT::i32), ShAmt);
4283 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
4284 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
4285 DAG.getConstant(VTBits, dl, MVT::i32));
4286 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
4287 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
4288 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
4290 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
4291 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, dl, MVT::i32),
4292 ISD::SETGE, ARMcc, DAG, dl);
4293 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
4294 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc,
4297 SDValue Ops[2] = { Lo, Hi };
4298 return DAG.getMergeValues(Ops, dl);
4301 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
4302 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
4303 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
4304 SelectionDAG &DAG) const {
4305 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4306 EVT VT = Op.getValueType();
4307 unsigned VTBits = VT.getSizeInBits();
4309 SDValue ShOpLo = Op.getOperand(0);
4310 SDValue ShOpHi = Op.getOperand(1);
4311 SDValue ShAmt = Op.getOperand(2);
4314 assert(Op.getOpcode() == ISD::SHL_PARTS);
4315 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
4316 DAG.getConstant(VTBits, dl, MVT::i32), ShAmt);
4317 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
4318 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
4319 DAG.getConstant(VTBits, dl, MVT::i32));
4320 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
4321 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
4323 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
4324 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
4325 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, dl, MVT::i32),
4326 ISD::SETGE, ARMcc, DAG, dl);
4327 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
4328 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc,
4331 SDValue Ops[2] = { Lo, Hi };
4332 return DAG.getMergeValues(Ops, dl);
4335 SDValue ARMTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
4336 SelectionDAG &DAG) const {
4337 // The rounding mode is in bits 23:22 of the FPSCR.
4338 // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0
4339 // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3)
4340 // so that the shift + and get folded into a bitfield extract.
4342 SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::i32,
4343 DAG.getConstant(Intrinsic::arm_get_fpscr, dl,
4345 SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPSCR,
4346 DAG.getConstant(1U << 22, dl, MVT::i32));
4347 SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds,
4348 DAG.getConstant(22, dl, MVT::i32));
4349 return DAG.getNode(ISD::AND, dl, MVT::i32, RMODE,
4350 DAG.getConstant(3, dl, MVT::i32));
4353 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
4354 const ARMSubtarget *ST) {
4356 EVT VT = N->getValueType(0);
4357 if (VT.isVector()) {
4358 assert(ST->hasNEON());
4360 // Compute the least significant set bit: LSB = X & -X
4361 SDValue X = N->getOperand(0);
4362 SDValue NX = DAG.getNode(ISD::SUB, dl, VT, getZeroVector(VT, DAG, dl), X);
4363 SDValue LSB = DAG.getNode(ISD::AND, dl, VT, X, NX);
4365 EVT ElemTy = VT.getVectorElementType();
4367 if (ElemTy == MVT::i8) {
4368 // Compute with: cttz(x) = ctpop(lsb - 1)
4369 SDValue One = DAG.getNode(ARMISD::VMOVIMM, dl, VT,
4370 DAG.getTargetConstant(1, dl, ElemTy));
4371 SDValue Bits = DAG.getNode(ISD::SUB, dl, VT, LSB, One);
4372 return DAG.getNode(ISD::CTPOP, dl, VT, Bits);
4375 if ((ElemTy == MVT::i16 || ElemTy == MVT::i32) &&
4376 (N->getOpcode() == ISD::CTTZ_ZERO_UNDEF)) {
4377 // Compute with: cttz(x) = (width - 1) - ctlz(lsb), if x != 0
4378 unsigned NumBits = ElemTy.getSizeInBits();
4379 SDValue WidthMinus1 =
4380 DAG.getNode(ARMISD::VMOVIMM, dl, VT,
4381 DAG.getTargetConstant(NumBits - 1, dl, ElemTy));
4382 SDValue CTLZ = DAG.getNode(ISD::CTLZ, dl, VT, LSB);
4383 return DAG.getNode(ISD::SUB, dl, VT, WidthMinus1, CTLZ);
4386 // Compute with: cttz(x) = ctpop(lsb - 1)
4388 // Since we can only compute the number of bits in a byte with vcnt.8, we
4389 // have to gather the result with pairwise addition (vpaddl) for i16, i32,
4394 if (ElemTy == MVT::i64) {
4395 // Load constant 0xffff'ffff'ffff'ffff to register.
4396 SDValue FF = DAG.getNode(ARMISD::VMOVIMM, dl, VT,
4397 DAG.getTargetConstant(0x1eff, dl, MVT::i32));
4398 Bits = DAG.getNode(ISD::ADD, dl, VT, LSB, FF);
4400 SDValue One = DAG.getNode(ARMISD::VMOVIMM, dl, VT,
4401 DAG.getTargetConstant(1, dl, ElemTy));
4402 Bits = DAG.getNode(ISD::SUB, dl, VT, LSB, One);
4405 // Count #bits with vcnt.8.
4406 EVT VT8Bit = VT.is64BitVector() ? MVT::v8i8 : MVT::v16i8;
4407 SDValue BitsVT8 = DAG.getNode(ISD::BITCAST, dl, VT8Bit, Bits);
4408 SDValue Cnt8 = DAG.getNode(ISD::CTPOP, dl, VT8Bit, BitsVT8);
4410 // Gather the #bits with vpaddl (pairwise add.)
4411 EVT VT16Bit = VT.is64BitVector() ? MVT::v4i16 : MVT::v8i16;
4412 SDValue Cnt16 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT16Bit,
4413 DAG.getTargetConstant(Intrinsic::arm_neon_vpaddlu, dl, MVT::i32),
4415 if (ElemTy == MVT::i16)
4418 EVT VT32Bit = VT.is64BitVector() ? MVT::v2i32 : MVT::v4i32;
4419 SDValue Cnt32 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT32Bit,
4420 DAG.getTargetConstant(Intrinsic::arm_neon_vpaddlu, dl, MVT::i32),
4422 if (ElemTy == MVT::i32)
4425 assert(ElemTy == MVT::i64);
4426 SDValue Cnt64 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
4427 DAG.getTargetConstant(Intrinsic::arm_neon_vpaddlu, dl, MVT::i32),
4432 if (!ST->hasV6T2Ops())
4435 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
4436 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
4439 /// getCTPOP16BitCounts - Returns a v8i8/v16i8 vector containing the bit-count
4440 /// for each 16-bit element from operand, repeated. The basic idea is to
4441 /// leverage vcnt to get the 8-bit counts, gather and add the results.
4443 /// Trace for v4i16:
4444 /// input = [v0 v1 v2 v3 ] (vi 16-bit element)
4445 /// cast: N0 = [w0 w1 w2 w3 w4 w5 w6 w7] (v0 = [w0 w1], wi 8-bit element)
4446 /// vcnt: N1 = [b0 b1 b2 b3 b4 b5 b6 b7] (bi = bit-count of 8-bit element wi)
4447 /// vrev: N2 = [b1 b0 b3 b2 b5 b4 b7 b6]
4448 /// [b0 b1 b2 b3 b4 b5 b6 b7]
4449 /// +[b1 b0 b3 b2 b5 b4 b7 b6]
4450 /// N3=N1+N2 = [k0 k0 k1 k1 k2 k2 k3 k3] (k0 = b0+b1 = bit-count of 16-bit v0,
4451 /// vuzp: = [k0 k1 k2 k3 k0 k1 k2 k3] each ki is 8-bits)
4452 static SDValue getCTPOP16BitCounts(SDNode *N, SelectionDAG &DAG) {
4453 EVT VT = N->getValueType(0);
4456 EVT VT8Bit = VT.is64BitVector() ? MVT::v8i8 : MVT::v16i8;
4457 SDValue N0 = DAG.getNode(ISD::BITCAST, DL, VT8Bit, N->getOperand(0));
4458 SDValue N1 = DAG.getNode(ISD::CTPOP, DL, VT8Bit, N0);
4459 SDValue N2 = DAG.getNode(ARMISD::VREV16, DL, VT8Bit, N1);
4460 SDValue N3 = DAG.getNode(ISD::ADD, DL, VT8Bit, N1, N2);
4461 return DAG.getNode(ARMISD::VUZP, DL, VT8Bit, N3, N3);
4464 /// lowerCTPOP16BitElements - Returns a v4i16/v8i16 vector containing the
4465 /// bit-count for each 16-bit element from the operand. We need slightly
4466 /// different sequencing for v4i16 and v8i16 to stay within NEON's available
4467 /// 64/128-bit registers.
4469 /// Trace for v4i16:
4470 /// input = [v0 v1 v2 v3 ] (vi 16-bit element)
4471 /// v8i8: BitCounts = [k0 k1 k2 k3 k0 k1 k2 k3 ] (ki is the bit-count of vi)
4472 /// v8i16:Extended = [k0 k1 k2 k3 k0 k1 k2 k3 ]
4473 /// v4i16:Extracted = [k0 k1 k2 k3 ]
4474 static SDValue lowerCTPOP16BitElements(SDNode *N, SelectionDAG &DAG) {
4475 EVT VT = N->getValueType(0);
4478 SDValue BitCounts = getCTPOP16BitCounts(N, DAG);
4479 if (VT.is64BitVector()) {
4480 SDValue Extended = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v8i16, BitCounts);
4481 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i16, Extended,
4482 DAG.getIntPtrConstant(0, DL));
4484 SDValue Extracted = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v8i8,
4485 BitCounts, DAG.getIntPtrConstant(0, DL));
4486 return DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v8i16, Extracted);
4490 /// lowerCTPOP32BitElements - Returns a v2i32/v4i32 vector containing the
4491 /// bit-count for each 32-bit element from the operand. The idea here is
4492 /// to split the vector into 16-bit elements, leverage the 16-bit count
4493 /// routine, and then combine the results.
4495 /// Trace for v2i32 (v4i32 similar with Extracted/Extended exchanged):
4496 /// input = [v0 v1 ] (vi: 32-bit elements)
4497 /// Bitcast = [w0 w1 w2 w3 ] (wi: 16-bit elements, v0 = [w0 w1])
4498 /// Counts16 = [k0 k1 k2 k3 ] (ki: 16-bit elements, bit-count of wi)
4499 /// vrev: N0 = [k1 k0 k3 k2 ]
4501 /// N1 =+[k1 k0 k3 k2 ]
4503 /// N2 =+[k1 k3 k0 k2 ]
4505 /// Extended =+[k1 k3 k0 k2 ]
4507 /// Extracted=+[k1 k3 ]
4509 static SDValue lowerCTPOP32BitElements(SDNode *N, SelectionDAG &DAG) {
4510 EVT VT = N->getValueType(0);
4513 EVT VT16Bit = VT.is64BitVector() ? MVT::v4i16 : MVT::v8i16;
4515 SDValue Bitcast = DAG.getNode(ISD::BITCAST, DL, VT16Bit, N->getOperand(0));
4516 SDValue Counts16 = lowerCTPOP16BitElements(Bitcast.getNode(), DAG);
4517 SDValue N0 = DAG.getNode(ARMISD::VREV32, DL, VT16Bit, Counts16);
4518 SDValue N1 = DAG.getNode(ISD::ADD, DL, VT16Bit, Counts16, N0);
4519 SDValue N2 = DAG.getNode(ARMISD::VUZP, DL, VT16Bit, N1, N1);
4521 if (VT.is64BitVector()) {
4522 SDValue Extended = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v4i32, N2);
4523 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i32, Extended,
4524 DAG.getIntPtrConstant(0, DL));
4526 SDValue Extracted = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i16, N2,
4527 DAG.getIntPtrConstant(0, DL));
4528 return DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v4i32, Extracted);
4532 static SDValue LowerCTPOP(SDNode *N, SelectionDAG &DAG,
4533 const ARMSubtarget *ST) {
4534 EVT VT = N->getValueType(0);
4536 assert(ST->hasNEON() && "Custom ctpop lowering requires NEON.");
4537 assert((VT == MVT::v2i32 || VT == MVT::v4i32 ||
4538 VT == MVT::v4i16 || VT == MVT::v8i16) &&
4539 "Unexpected type for custom ctpop lowering");
4541 if (VT.getVectorElementType() == MVT::i32)
4542 return lowerCTPOP32BitElements(N, DAG);
4544 return lowerCTPOP16BitElements(N, DAG);
4547 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
4548 const ARMSubtarget *ST) {
4549 EVT VT = N->getValueType(0);
4555 // Lower vector shifts on NEON to use VSHL.
4556 assert(ST->hasNEON() && "unexpected vector shift");
4558 // Left shifts translate directly to the vshiftu intrinsic.
4559 if (N->getOpcode() == ISD::SHL)
4560 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
4561 DAG.getConstant(Intrinsic::arm_neon_vshiftu, dl,
4563 N->getOperand(0), N->getOperand(1));
4565 assert((N->getOpcode() == ISD::SRA ||
4566 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
4568 // NEON uses the same intrinsics for both left and right shifts. For
4569 // right shifts, the shift amounts are negative, so negate the vector of
4571 EVT ShiftVT = N->getOperand(1).getValueType();
4572 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
4573 getZeroVector(ShiftVT, DAG, dl),
4575 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
4576 Intrinsic::arm_neon_vshifts :
4577 Intrinsic::arm_neon_vshiftu);
4578 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
4579 DAG.getConstant(vshiftInt, dl, MVT::i32),
4580 N->getOperand(0), NegatedCount);
4583 static SDValue Expand64BitShift(SDNode *N, SelectionDAG &DAG,
4584 const ARMSubtarget *ST) {
4585 EVT VT = N->getValueType(0);
4588 // We can get here for a node like i32 = ISD::SHL i32, i64
4592 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
4593 "Unknown shift to lower!");
4595 // We only lower SRA, SRL of 1 here, all others use generic lowering.
4596 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
4597 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
4600 // If we are in thumb mode, we don't have RRX.
4601 if (ST->isThumb1Only()) return SDValue();
4603 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
4604 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
4605 DAG.getConstant(0, dl, MVT::i32));
4606 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
4607 DAG.getConstant(1, dl, MVT::i32));
4609 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
4610 // captures the result into a carry flag.
4611 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
4612 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Glue), Hi);
4614 // The low part is an ARMISD::RRX operand, which shifts the carry in.
4615 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
4617 // Merge the pieces into a single i64 value.
4618 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
4621 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
4622 SDValue TmpOp0, TmpOp1;
4623 bool Invert = false;
4627 SDValue Op0 = Op.getOperand(0);
4628 SDValue Op1 = Op.getOperand(1);
4629 SDValue CC = Op.getOperand(2);
4630 EVT CmpVT = Op0.getValueType().changeVectorElementTypeToInteger();
4631 EVT VT = Op.getValueType();
4632 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
4635 if (Op1.getValueType().isFloatingPoint()) {
4636 switch (SetCCOpcode) {
4637 default: llvm_unreachable("Illegal FP comparison");
4639 case ISD::SETNE: Invert = true; // Fallthrough
4641 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
4643 case ISD::SETLT: Swap = true; // Fallthrough
4645 case ISD::SETGT: Opc = ARMISD::VCGT; break;
4647 case ISD::SETLE: Swap = true; // Fallthrough
4649 case ISD::SETGE: Opc = ARMISD::VCGE; break;
4650 case ISD::SETUGE: Swap = true; // Fallthrough
4651 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
4652 case ISD::SETUGT: Swap = true; // Fallthrough
4653 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
4654 case ISD::SETUEQ: Invert = true; // Fallthrough
4656 // Expand this to (OLT | OGT).
4660 Op0 = DAG.getNode(ARMISD::VCGT, dl, CmpVT, TmpOp1, TmpOp0);
4661 Op1 = DAG.getNode(ARMISD::VCGT, dl, CmpVT, TmpOp0, TmpOp1);
4663 case ISD::SETUO: Invert = true; // Fallthrough
4665 // Expand this to (OLT | OGE).
4669 Op0 = DAG.getNode(ARMISD::VCGT, dl, CmpVT, TmpOp1, TmpOp0);
4670 Op1 = DAG.getNode(ARMISD::VCGE, dl, CmpVT, TmpOp0, TmpOp1);
4674 // Integer comparisons.
4675 switch (SetCCOpcode) {
4676 default: llvm_unreachable("Illegal integer comparison");
4677 case ISD::SETNE: Invert = true;
4678 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
4679 case ISD::SETLT: Swap = true;
4680 case ISD::SETGT: Opc = ARMISD::VCGT; break;
4681 case ISD::SETLE: Swap = true;
4682 case ISD::SETGE: Opc = ARMISD::VCGE; break;
4683 case ISD::SETULT: Swap = true;
4684 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
4685 case ISD::SETULE: Swap = true;
4686 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
4689 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
4690 if (Opc == ARMISD::VCEQ) {
4693 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
4695 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
4698 // Ignore bitconvert.
4699 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BITCAST)
4700 AndOp = AndOp.getOperand(0);
4702 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
4704 Op0 = DAG.getNode(ISD::BITCAST, dl, CmpVT, AndOp.getOperand(0));
4705 Op1 = DAG.getNode(ISD::BITCAST, dl, CmpVT, AndOp.getOperand(1));
4712 std::swap(Op0, Op1);
4714 // If one of the operands is a constant vector zero, attempt to fold the
4715 // comparison to a specialized compare-against-zero form.
4717 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
4719 else if (ISD::isBuildVectorAllZeros(Op0.getNode())) {
4720 if (Opc == ARMISD::VCGE)
4721 Opc = ARMISD::VCLEZ;
4722 else if (Opc == ARMISD::VCGT)
4723 Opc = ARMISD::VCLTZ;
4728 if (SingleOp.getNode()) {
4731 Result = DAG.getNode(ARMISD::VCEQZ, dl, CmpVT, SingleOp); break;
4733 Result = DAG.getNode(ARMISD::VCGEZ, dl, CmpVT, SingleOp); break;
4735 Result = DAG.getNode(ARMISD::VCLEZ, dl, CmpVT, SingleOp); break;
4737 Result = DAG.getNode(ARMISD::VCGTZ, dl, CmpVT, SingleOp); break;
4739 Result = DAG.getNode(ARMISD::VCLTZ, dl, CmpVT, SingleOp); break;
4741 Result = DAG.getNode(Opc, dl, CmpVT, Op0, Op1);
4744 Result = DAG.getNode(Opc, dl, CmpVT, Op0, Op1);
4747 Result = DAG.getSExtOrTrunc(Result, dl, VT);
4750 Result = DAG.getNOT(dl, Result, VT);
4755 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
4756 /// valid vector constant for a NEON instruction with a "modified immediate"
4757 /// operand (e.g., VMOV). If so, return the encoded value.
4758 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
4759 unsigned SplatBitSize, SelectionDAG &DAG,
4760 SDLoc dl, EVT &VT, bool is128Bits,
4761 NEONModImmType type) {
4762 unsigned OpCmode, Imm;
4764 // SplatBitSize is set to the smallest size that splats the vector, so a
4765 // zero vector will always have SplatBitSize == 8. However, NEON modified
4766 // immediate instructions others than VMOV do not support the 8-bit encoding
4767 // of a zero vector, and the default encoding of zero is supposed to be the
4772 switch (SplatBitSize) {
4774 if (type != VMOVModImm)
4776 // Any 1-byte value is OK. Op=0, Cmode=1110.
4777 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
4780 VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
4784 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
4785 VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
4786 if ((SplatBits & ~0xff) == 0) {
4787 // Value = 0x00nn: Op=x, Cmode=100x.
4792 if ((SplatBits & ~0xff00) == 0) {
4793 // Value = 0xnn00: Op=x, Cmode=101x.
4795 Imm = SplatBits >> 8;
4801 // NEON's 32-bit VMOV supports splat values where:
4802 // * only one byte is nonzero, or
4803 // * the least significant byte is 0xff and the second byte is nonzero, or
4804 // * the least significant 2 bytes are 0xff and the third is nonzero.
4805 VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
4806 if ((SplatBits & ~0xff) == 0) {
4807 // Value = 0x000000nn: Op=x, Cmode=000x.
4812 if ((SplatBits & ~0xff00) == 0) {
4813 // Value = 0x0000nn00: Op=x, Cmode=001x.
4815 Imm = SplatBits >> 8;
4818 if ((SplatBits & ~0xff0000) == 0) {
4819 // Value = 0x00nn0000: Op=x, Cmode=010x.
4821 Imm = SplatBits >> 16;
4824 if ((SplatBits & ~0xff000000) == 0) {
4825 // Value = 0xnn000000: Op=x, Cmode=011x.
4827 Imm = SplatBits >> 24;
4831 // cmode == 0b1100 and cmode == 0b1101 are not supported for VORR or VBIC
4832 if (type == OtherModImm) return SDValue();
4834 if ((SplatBits & ~0xffff) == 0 &&
4835 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
4836 // Value = 0x0000nnff: Op=x, Cmode=1100.
4838 Imm = SplatBits >> 8;
4842 if ((SplatBits & ~0xffffff) == 0 &&
4843 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
4844 // Value = 0x00nnffff: Op=x, Cmode=1101.
4846 Imm = SplatBits >> 16;
4850 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
4851 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
4852 // VMOV.I32. A (very) minor optimization would be to replicate the value
4853 // and fall through here to test for a valid 64-bit splat. But, then the
4854 // caller would also need to check and handle the change in size.
4858 if (type != VMOVModImm)
4860 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
4861 uint64_t BitMask = 0xff;
4863 unsigned ImmMask = 1;
4865 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
4866 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
4869 } else if ((SplatBits & BitMask) != 0) {
4876 if (DAG.getDataLayout().isBigEndian())
4877 // swap higher and lower 32 bit word
4878 Imm = ((Imm & 0xf) << 4) | ((Imm & 0xf0) >> 4);
4880 // Op=1, Cmode=1110.
4882 VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
4887 llvm_unreachable("unexpected size for isNEONModifiedImm");
4890 unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm);
4891 return DAG.getTargetConstant(EncodedVal, dl, MVT::i32);
4894 SDValue ARMTargetLowering::LowerConstantFP(SDValue Op, SelectionDAG &DAG,
4895 const ARMSubtarget *ST) const {
4899 bool IsDouble = Op.getValueType() == MVT::f64;
4900 ConstantFPSDNode *CFP = cast<ConstantFPSDNode>(Op);
4902 // Use the default (constant pool) lowering for double constants when we have
4904 if (IsDouble && Subtarget->isFPOnlySP())
4907 // Try splatting with a VMOV.f32...
4908 APFloat FPVal = CFP->getValueAPF();
4909 int ImmVal = IsDouble ? ARM_AM::getFP64Imm(FPVal) : ARM_AM::getFP32Imm(FPVal);
4912 if (IsDouble || !ST->useNEONForSinglePrecisionFP()) {
4913 // We have code in place to select a valid ConstantFP already, no need to
4918 // It's a float and we are trying to use NEON operations where
4919 // possible. Lower it to a splat followed by an extract.
4921 SDValue NewVal = DAG.getTargetConstant(ImmVal, DL, MVT::i32);
4922 SDValue VecConstant = DAG.getNode(ARMISD::VMOVFPIMM, DL, MVT::v2f32,
4924 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecConstant,
4925 DAG.getConstant(0, DL, MVT::i32));
4928 // The rest of our options are NEON only, make sure that's allowed before
4930 if (!ST->hasNEON() || (!IsDouble && !ST->useNEONForSinglePrecisionFP()))
4934 uint64_t iVal = FPVal.bitcastToAPInt().getZExtValue();
4936 // It wouldn't really be worth bothering for doubles except for one very
4937 // important value, which does happen to match: 0.0. So make sure we don't do
4939 if (IsDouble && (iVal & 0xffffffff) != (iVal >> 32))
4942 // Try a VMOV.i32 (FIXME: i8, i16, or i64 could work too).
4943 SDValue NewVal = isNEONModifiedImm(iVal & 0xffffffffU, 0, 32, DAG, SDLoc(Op),
4944 VMovVT, false, VMOVModImm);
4945 if (NewVal != SDValue()) {
4947 SDValue VecConstant = DAG.getNode(ARMISD::VMOVIMM, DL, VMovVT,
4950 return DAG.getNode(ISD::BITCAST, DL, MVT::f64, VecConstant);
4952 // It's a float: cast and extract a vector element.
4953 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
4955 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
4956 DAG.getConstant(0, DL, MVT::i32));
4959 // Finally, try a VMVN.i32
4960 NewVal = isNEONModifiedImm(~iVal & 0xffffffffU, 0, 32, DAG, SDLoc(Op), VMovVT,
4962 if (NewVal != SDValue()) {
4964 SDValue VecConstant = DAG.getNode(ARMISD::VMVNIMM, DL, VMovVT, NewVal);
4967 return DAG.getNode(ISD::BITCAST, DL, MVT::f64, VecConstant);
4969 // It's a float: cast and extract a vector element.
4970 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
4972 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
4973 DAG.getConstant(0, DL, MVT::i32));
4979 // check if an VEXT instruction can handle the shuffle mask when the
4980 // vector sources of the shuffle are the same.
4981 static bool isSingletonVEXTMask(ArrayRef<int> M, EVT VT, unsigned &Imm) {
4982 unsigned NumElts = VT.getVectorNumElements();
4984 // Assume that the first shuffle index is not UNDEF. Fail if it is.
4990 // If this is a VEXT shuffle, the immediate value is the index of the first
4991 // element. The other shuffle indices must be the successive elements after
4993 unsigned ExpectedElt = Imm;
4994 for (unsigned i = 1; i < NumElts; ++i) {
4995 // Increment the expected index. If it wraps around, just follow it
4996 // back to index zero and keep going.
4998 if (ExpectedElt == NumElts)
5001 if (M[i] < 0) continue; // ignore UNDEF indices
5002 if (ExpectedElt != static_cast<unsigned>(M[i]))
5010 static bool isVEXTMask(ArrayRef<int> M, EVT VT,
5011 bool &ReverseVEXT, unsigned &Imm) {
5012 unsigned NumElts = VT.getVectorNumElements();
5013 ReverseVEXT = false;
5015 // Assume that the first shuffle index is not UNDEF. Fail if it is.
5021 // If this is a VEXT shuffle, the immediate value is the index of the first
5022 // element. The other shuffle indices must be the successive elements after
5024 unsigned ExpectedElt = Imm;
5025 for (unsigned i = 1; i < NumElts; ++i) {
5026 // Increment the expected index. If it wraps around, it may still be
5027 // a VEXT but the source vectors must be swapped.
5029 if (ExpectedElt == NumElts * 2) {
5034 if (M[i] < 0) continue; // ignore UNDEF indices
5035 if (ExpectedElt != static_cast<unsigned>(M[i]))
5039 // Adjust the index value if the source operands will be swapped.
5046 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
5047 /// instruction with the specified blocksize. (The order of the elements
5048 /// within each block of the vector is reversed.)
5049 static bool isVREVMask(ArrayRef<int> M, EVT VT, unsigned BlockSize) {
5050 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
5051 "Only possible block sizes for VREV are: 16, 32, 64");
5053 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
5057 unsigned NumElts = VT.getVectorNumElements();
5058 unsigned BlockElts = M[0] + 1;
5059 // If the first shuffle index is UNDEF, be optimistic.
5061 BlockElts = BlockSize / EltSz;
5063 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
5066 for (unsigned i = 0; i < NumElts; ++i) {
5067 if (M[i] < 0) continue; // ignore UNDEF indices
5068 if ((unsigned) M[i] != (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
5075 static bool isVTBLMask(ArrayRef<int> M, EVT VT) {
5076 // We can handle <8 x i8> vector shuffles. If the index in the mask is out of
5077 // range, then 0 is placed into the resulting vector. So pretty much any mask
5078 // of 8 elements can work here.
5079 return VT == MVT::v8i8 && M.size() == 8;
5082 // Checks whether the shuffle mask represents a vector transpose (VTRN) by
5083 // checking that pairs of elements in the shuffle mask represent the same index
5084 // in each vector, incrementing the expected index by 2 at each step.
5085 // e.g. For v1,v2 of type v4i32 a valid shuffle mask is: [0, 4, 2, 6]
5086 // v1={a,b,c,d} => x=shufflevector v1, v2 shufflemask => x={a,e,c,g}
5088 // WhichResult gives the offset for each element in the mask based on which
5089 // of the two results it belongs to.
5091 // The transpose can be represented either as:
5092 // result1 = shufflevector v1, v2, result1_shuffle_mask
5093 // result2 = shufflevector v1, v2, result2_shuffle_mask
5094 // where v1/v2 and the shuffle masks have the same number of elements
5095 // (here WhichResult (see below) indicates which result is being checked)
5098 // results = shufflevector v1, v2, shuffle_mask
5099 // where both results are returned in one vector and the shuffle mask has twice
5100 // as many elements as v1/v2 (here WhichResult will always be 0 if true) here we
5101 // want to check the low half and high half of the shuffle mask as if it were
5103 static bool isVTRNMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
5104 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
5108 unsigned NumElts = VT.getVectorNumElements();
5109 if (M.size() != NumElts && M.size() != NumElts*2)
5112 // If the mask is twice as long as the result then we need to check the upper
5113 // and lower parts of the mask
5114 for (unsigned i = 0; i < M.size(); i += NumElts) {
5115 WhichResult = M[i] == 0 ? 0 : 1;
5116 for (unsigned j = 0; j < NumElts; j += 2) {
5117 if ((M[i+j] >= 0 && (unsigned) M[i+j] != j + WhichResult) ||
5118 (M[i+j+1] >= 0 && (unsigned) M[i+j+1] != j + NumElts + WhichResult))
5123 if (M.size() == NumElts*2)
5129 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
5130 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
5131 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
5132 static bool isVTRN_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
5133 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
5137 unsigned NumElts = VT.getVectorNumElements();
5138 if (M.size() != NumElts && M.size() != NumElts*2)
5141 for (unsigned i = 0; i < M.size(); i += NumElts) {
5142 WhichResult = M[i] == 0 ? 0 : 1;
5143 for (unsigned j = 0; j < NumElts; j += 2) {
5144 if ((M[i+j] >= 0 && (unsigned) M[i+j] != j + WhichResult) ||
5145 (M[i+j+1] >= 0 && (unsigned) M[i+j+1] != j + WhichResult))
5150 if (M.size() == NumElts*2)
5156 // Checks whether the shuffle mask represents a vector unzip (VUZP) by checking
5157 // that the mask elements are either all even and in steps of size 2 or all odd
5158 // and in steps of size 2.
5159 // e.g. For v1,v2 of type v4i32 a valid shuffle mask is: [0, 2, 4, 6]
5160 // v1={a,b,c,d} => x=shufflevector v1, v2 shufflemask => x={a,c,e,g}
5162 // Requires similar checks to that of isVTRNMask with
5163 // respect the how results are returned.
5164 static bool isVUZPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
5165 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
5169 unsigned NumElts = VT.getVectorNumElements();
5170 if (M.size() != NumElts && M.size() != NumElts*2)
5173 for (unsigned i = 0; i < M.size(); i += NumElts) {
5174 WhichResult = M[i] == 0 ? 0 : 1;
5175 for (unsigned j = 0; j < NumElts; ++j) {
5176 if (M[i+j] >= 0 && (unsigned) M[i+j] != 2 * j + WhichResult)
5181 if (M.size() == NumElts*2)
5184 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
5185 if (VT.is64BitVector() && EltSz == 32)
5191 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
5192 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
5193 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
5194 static bool isVUZP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
5195 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
5199 unsigned NumElts = VT.getVectorNumElements();
5200 if (M.size() != NumElts && M.size() != NumElts*2)
5203 unsigned Half = NumElts / 2;
5204 for (unsigned i = 0; i < M.size(); i += NumElts) {
5205 WhichResult = M[i] == 0 ? 0 : 1;
5206 for (unsigned j = 0; j < NumElts; j += Half) {
5207 unsigned Idx = WhichResult;
5208 for (unsigned k = 0; k < Half; ++k) {
5209 int MIdx = M[i + j + k];
5210 if (MIdx >= 0 && (unsigned) MIdx != Idx)
5217 if (M.size() == NumElts*2)
5220 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
5221 if (VT.is64BitVector() && EltSz == 32)
5227 // Checks whether the shuffle mask represents a vector zip (VZIP) by checking
5228 // that pairs of elements of the shufflemask represent the same index in each
5229 // vector incrementing sequentially through the vectors.
5230 // e.g. For v1,v2 of type v4i32 a valid shuffle mask is: [0, 4, 1, 5]
5231 // v1={a,b,c,d} => x=shufflevector v1, v2 shufflemask => x={a,e,b,f}
5233 // Requires similar checks to that of isVTRNMask with respect the how results
5235 static bool isVZIPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
5236 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
5240 unsigned NumElts = VT.getVectorNumElements();
5241 if (M.size() != NumElts && M.size() != NumElts*2)
5244 for (unsigned i = 0; i < M.size(); i += NumElts) {
5245 WhichResult = M[i] == 0 ? 0 : 1;
5246 unsigned Idx = WhichResult * NumElts / 2;
5247 for (unsigned j = 0; j < NumElts; j += 2) {
5248 if ((M[i+j] >= 0 && (unsigned) M[i+j] != Idx) ||
5249 (M[i+j+1] >= 0 && (unsigned) M[i+j+1] != Idx + NumElts))
5255 if (M.size() == NumElts*2)
5258 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
5259 if (VT.is64BitVector() && EltSz == 32)
5265 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
5266 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
5267 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
5268 static bool isVZIP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
5269 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
5273 unsigned NumElts = VT.getVectorNumElements();
5274 if (M.size() != NumElts && M.size() != NumElts*2)
5277 for (unsigned i = 0; i < M.size(); i += NumElts) {
5278 WhichResult = M[i] == 0 ? 0 : 1;
5279 unsigned Idx = WhichResult * NumElts / 2;
5280 for (unsigned j = 0; j < NumElts; j += 2) {
5281 if ((M[i+j] >= 0 && (unsigned) M[i+j] != Idx) ||
5282 (M[i+j+1] >= 0 && (unsigned) M[i+j+1] != Idx))
5288 if (M.size() == NumElts*2)
5291 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
5292 if (VT.is64BitVector() && EltSz == 32)
5298 /// Check if \p ShuffleMask is a NEON two-result shuffle (VZIP, VUZP, VTRN),
5299 /// and return the corresponding ARMISD opcode if it is, or 0 if it isn't.
5300 static unsigned isNEONTwoResultShuffleMask(ArrayRef<int> ShuffleMask, EVT VT,
5301 unsigned &WhichResult,
5304 if (isVTRNMask(ShuffleMask, VT, WhichResult))
5305 return ARMISD::VTRN;
5306 if (isVUZPMask(ShuffleMask, VT, WhichResult))
5307 return ARMISD::VUZP;
5308 if (isVZIPMask(ShuffleMask, VT, WhichResult))
5309 return ARMISD::VZIP;
5312 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
5313 return ARMISD::VTRN;
5314 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
5315 return ARMISD::VUZP;
5316 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
5317 return ARMISD::VZIP;
5322 /// \return true if this is a reverse operation on an vector.
5323 static bool isReverseMask(ArrayRef<int> M, EVT VT) {
5324 unsigned NumElts = VT.getVectorNumElements();
5325 // Make sure the mask has the right size.
5326 if (NumElts != M.size())
5329 // Look for <15, ..., 3, -1, 1, 0>.
5330 for (unsigned i = 0; i != NumElts; ++i)
5331 if (M[i] >= 0 && M[i] != (int) (NumElts - 1 - i))
5337 // If N is an integer constant that can be moved into a register in one
5338 // instruction, return an SDValue of such a constant (will become a MOV
5339 // instruction). Otherwise return null.
5340 static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG,
5341 const ARMSubtarget *ST, SDLoc dl) {
5343 if (!isa<ConstantSDNode>(N))
5345 Val = cast<ConstantSDNode>(N)->getZExtValue();
5347 if (ST->isThumb1Only()) {
5348 if (Val <= 255 || ~Val <= 255)
5349 return DAG.getConstant(Val, dl, MVT::i32);
5351 if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1)
5352 return DAG.getConstant(Val, dl, MVT::i32);
5357 // If this is a case we can't handle, return null and let the default
5358 // expansion code take care of it.
5359 SDValue ARMTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
5360 const ARMSubtarget *ST) const {
5361 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
5363 EVT VT = Op.getValueType();
5365 APInt SplatBits, SplatUndef;
5366 unsigned SplatBitSize;
5368 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
5369 if (SplatBitSize <= 64) {
5370 // Check if an immediate VMOV works.
5372 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
5373 SplatUndef.getZExtValue(), SplatBitSize,
5374 DAG, dl, VmovVT, VT.is128BitVector(),
5376 if (Val.getNode()) {
5377 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
5378 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
5381 // Try an immediate VMVN.
5382 uint64_t NegatedImm = (~SplatBits).getZExtValue();
5383 Val = isNEONModifiedImm(NegatedImm,
5384 SplatUndef.getZExtValue(), SplatBitSize,
5385 DAG, dl, VmovVT, VT.is128BitVector(),
5387 if (Val.getNode()) {
5388 SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
5389 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
5392 // Use vmov.f32 to materialize other v2f32 and v4f32 splats.
5393 if ((VT == MVT::v2f32 || VT == MVT::v4f32) && SplatBitSize == 32) {
5394 int ImmVal = ARM_AM::getFP32Imm(SplatBits);
5396 SDValue Val = DAG.getTargetConstant(ImmVal, dl, MVT::i32);
5397 return DAG.getNode(ARMISD::VMOVFPIMM, dl, VT, Val);
5403 // Scan through the operands to see if only one value is used.
5405 // As an optimisation, even if more than one value is used it may be more
5406 // profitable to splat with one value then change some lanes.
5408 // Heuristically we decide to do this if the vector has a "dominant" value,
5409 // defined as splatted to more than half of the lanes.
5410 unsigned NumElts = VT.getVectorNumElements();
5411 bool isOnlyLowElement = true;
5412 bool usesOnlyOneValue = true;
5413 bool hasDominantValue = false;
5414 bool isConstant = true;
5416 // Map of the number of times a particular SDValue appears in the
5418 DenseMap<SDValue, unsigned> ValueCounts;
5420 for (unsigned i = 0; i < NumElts; ++i) {
5421 SDValue V = Op.getOperand(i);
5422 if (V.getOpcode() == ISD::UNDEF)
5425 isOnlyLowElement = false;
5426 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
5429 ValueCounts.insert(std::make_pair(V, 0));
5430 unsigned &Count = ValueCounts[V];
5432 // Is this value dominant? (takes up more than half of the lanes)
5433 if (++Count > (NumElts / 2)) {
5434 hasDominantValue = true;
5438 if (ValueCounts.size() != 1)
5439 usesOnlyOneValue = false;
5440 if (!Value.getNode() && ValueCounts.size() > 0)
5441 Value = ValueCounts.begin()->first;
5443 if (ValueCounts.size() == 0)
5444 return DAG.getUNDEF(VT);
5446 // Loads are better lowered with insert_vector_elt/ARMISD::BUILD_VECTOR.
5447 // Keep going if we are hitting this case.
5448 if (isOnlyLowElement && !ISD::isNormalLoad(Value.getNode()))
5449 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
5451 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5453 // Use VDUP for non-constant splats. For f32 constant splats, reduce to
5454 // i32 and try again.
5455 if (hasDominantValue && EltSize <= 32) {
5459 // If we are VDUPing a value that comes directly from a vector, that will
5460 // cause an unnecessary move to and from a GPR, where instead we could
5461 // just use VDUPLANE. We can only do this if the lane being extracted
5462 // is at a constant index, as the VDUP from lane instructions only have
5463 // constant-index forms.
5464 if (Value->getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5465 isa<ConstantSDNode>(Value->getOperand(1))) {
5466 // We need to create a new undef vector to use for the VDUPLANE if the
5467 // size of the vector from which we get the value is different than the
5468 // size of the vector that we need to create. We will insert the element
5469 // such that the register coalescer will remove unnecessary copies.
5470 if (VT != Value->getOperand(0).getValueType()) {
5471 ConstantSDNode *constIndex;
5472 constIndex = dyn_cast<ConstantSDNode>(Value->getOperand(1));
5473 assert(constIndex && "The index is not a constant!");
5474 unsigned index = constIndex->getAPIntValue().getLimitedValue() %
5475 VT.getVectorNumElements();
5476 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
5477 DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DAG.getUNDEF(VT),
5478 Value, DAG.getConstant(index, dl, MVT::i32)),
5479 DAG.getConstant(index, dl, MVT::i32));
5481 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
5482 Value->getOperand(0), Value->getOperand(1));
5484 N = DAG.getNode(ARMISD::VDUP, dl, VT, Value);
5486 if (!usesOnlyOneValue) {
5487 // The dominant value was splatted as 'N', but we now have to insert
5488 // all differing elements.
5489 for (unsigned I = 0; I < NumElts; ++I) {
5490 if (Op.getOperand(I) == Value)
5492 SmallVector<SDValue, 3> Ops;
5494 Ops.push_back(Op.getOperand(I));
5495 Ops.push_back(DAG.getConstant(I, dl, MVT::i32));
5496 N = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Ops);
5501 if (VT.getVectorElementType().isFloatingPoint()) {
5502 SmallVector<SDValue, 8> Ops;
5503 for (unsigned i = 0; i < NumElts; ++i)
5504 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, MVT::i32,
5506 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
5507 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, Ops);
5508 Val = LowerBUILD_VECTOR(Val, DAG, ST);
5510 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
5512 if (usesOnlyOneValue) {
5513 SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
5514 if (isConstant && Val.getNode())
5515 return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
5519 // If all elements are constants and the case above didn't get hit, fall back
5520 // to the default expansion, which will generate a load from the constant
5525 // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
5527 SDValue shuffle = ReconstructShuffle(Op, DAG);
5528 if (shuffle != SDValue())
5532 // Vectors with 32- or 64-bit elements can be built by directly assigning
5533 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
5534 // will be legalized.
5535 if (EltSize >= 32) {
5536 // Do the expansion with floating-point types, since that is what the VFP
5537 // registers are defined to use, and since i64 is not legal.
5538 EVT EltVT = EVT::getFloatingPointVT(EltSize);
5539 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
5540 SmallVector<SDValue, 8> Ops;
5541 for (unsigned i = 0; i < NumElts; ++i)
5542 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, EltVT, Op.getOperand(i)));
5543 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, Ops);
5544 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
5547 // If all else fails, just use a sequence of INSERT_VECTOR_ELT when we
5548 // know the default expansion would otherwise fall back on something even
5549 // worse. For a vector with one or two non-undef values, that's
5550 // scalar_to_vector for the elements followed by a shuffle (provided the
5551 // shuffle is valid for the target) and materialization element by element
5552 // on the stack followed by a load for everything else.
5553 if (!isConstant && !usesOnlyOneValue) {
5554 SDValue Vec = DAG.getUNDEF(VT);
5555 for (unsigned i = 0 ; i < NumElts; ++i) {
5556 SDValue V = Op.getOperand(i);
5557 if (V.getOpcode() == ISD::UNDEF)
5559 SDValue LaneIdx = DAG.getConstant(i, dl, MVT::i32);
5560 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Vec, V, LaneIdx);
5568 /// getExtFactor - Determine the adjustment factor for the position when
5569 /// generating an "extract from vector registers" instruction.
5570 static unsigned getExtFactor(SDValue &V) {
5571 EVT EltType = V.getValueType().getVectorElementType();
5572 return EltType.getSizeInBits() / 8;
5575 // Gather data to see if the operation can be modelled as a
5576 // shuffle in combination with VEXTs.
5577 SDValue ARMTargetLowering::ReconstructShuffle(SDValue Op,
5578 SelectionDAG &DAG) const {
5579 assert(Op.getOpcode() == ISD::BUILD_VECTOR && "Unknown opcode!");
5581 EVT VT = Op.getValueType();
5582 unsigned NumElts = VT.getVectorNumElements();
5584 struct ShuffleSourceInfo {
5589 // We may insert some combination of BITCASTs and VEXT nodes to force Vec to
5590 // be compatible with the shuffle we intend to construct. As a result
5591 // ShuffleVec will be some sliding window into the original Vec.
5594 // Code should guarantee that element i in Vec starts at element "WindowBase
5595 // + i * WindowScale in ShuffleVec".
5599 bool operator ==(SDValue OtherVec) { return Vec == OtherVec; }
5600 ShuffleSourceInfo(SDValue Vec)
5601 : Vec(Vec), MinElt(UINT_MAX), MaxElt(0), ShuffleVec(Vec), WindowBase(0),
5605 // First gather all vectors used as an immediate source for this BUILD_VECTOR
5607 SmallVector<ShuffleSourceInfo, 2> Sources;
5608 for (unsigned i = 0; i < NumElts; ++i) {
5609 SDValue V = Op.getOperand(i);
5610 if (V.getOpcode() == ISD::UNDEF)
5612 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
5613 // A shuffle can only come from building a vector from various
5614 // elements of other vectors.
5618 // Add this element source to the list if it's not already there.
5619 SDValue SourceVec = V.getOperand(0);
5620 auto Source = std::find(Sources.begin(), Sources.end(), SourceVec);
5621 if (Source == Sources.end())
5622 Source = Sources.insert(Sources.end(), ShuffleSourceInfo(SourceVec));
5624 // Update the minimum and maximum lane number seen.
5625 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
5626 Source->MinElt = std::min(Source->MinElt, EltNo);
5627 Source->MaxElt = std::max(Source->MaxElt, EltNo);
5630 // Currently only do something sane when at most two source vectors
5632 if (Sources.size() > 2)
5635 // Find out the smallest element size among result and two sources, and use
5636 // it as element size to build the shuffle_vector.
5637 EVT SmallestEltTy = VT.getVectorElementType();
5638 for (auto &Source : Sources) {
5639 EVT SrcEltTy = Source.Vec.getValueType().getVectorElementType();
5640 if (SrcEltTy.bitsLT(SmallestEltTy))
5641 SmallestEltTy = SrcEltTy;
5643 unsigned ResMultiplier =
5644 VT.getVectorElementType().getSizeInBits() / SmallestEltTy.getSizeInBits();
5645 NumElts = VT.getSizeInBits() / SmallestEltTy.getSizeInBits();
5646 EVT ShuffleVT = EVT::getVectorVT(*DAG.getContext(), SmallestEltTy, NumElts);
5648 // If the source vector is too wide or too narrow, we may nevertheless be able
5649 // to construct a compatible shuffle either by concatenating it with UNDEF or
5650 // extracting a suitable range of elements.
5651 for (auto &Src : Sources) {
5652 EVT SrcVT = Src.ShuffleVec.getValueType();
5654 if (SrcVT.getSizeInBits() == VT.getSizeInBits())
5657 // This stage of the search produces a source with the same element type as
5658 // the original, but with a total width matching the BUILD_VECTOR output.
5659 EVT EltVT = SrcVT.getVectorElementType();
5660 unsigned NumSrcElts = VT.getSizeInBits() / EltVT.getSizeInBits();
5661 EVT DestVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumSrcElts);
5663 if (SrcVT.getSizeInBits() < VT.getSizeInBits()) {
5664 if (2 * SrcVT.getSizeInBits() != VT.getSizeInBits())
5666 // We can pad out the smaller vector for free, so if it's part of a
5669 DAG.getNode(ISD::CONCAT_VECTORS, dl, DestVT, Src.ShuffleVec,
5670 DAG.getUNDEF(Src.ShuffleVec.getValueType()));
5674 if (SrcVT.getSizeInBits() != 2 * VT.getSizeInBits())
5677 if (Src.MaxElt - Src.MinElt >= NumSrcElts) {
5678 // Span too large for a VEXT to cope
5682 if (Src.MinElt >= NumSrcElts) {
5683 // The extraction can just take the second half
5685 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec,
5686 DAG.getConstant(NumSrcElts, dl, MVT::i32));
5687 Src.WindowBase = -NumSrcElts;
5688 } else if (Src.MaxElt < NumSrcElts) {
5689 // The extraction can just take the first half
5691 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec,
5692 DAG.getConstant(0, dl, MVT::i32));
5694 // An actual VEXT is needed
5696 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec,
5697 DAG.getConstant(0, dl, MVT::i32));
5699 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec,
5700 DAG.getConstant(NumSrcElts, dl, MVT::i32));
5701 unsigned Imm = Src.MinElt * getExtFactor(VEXTSrc1);
5703 Src.ShuffleVec = DAG.getNode(ARMISD::VEXT, dl, DestVT, VEXTSrc1,
5705 DAG.getConstant(Imm, dl, MVT::i32));
5706 Src.WindowBase = -Src.MinElt;
5710 // Another possible incompatibility occurs from the vector element types. We
5711 // can fix this by bitcasting the source vectors to the same type we intend
5713 for (auto &Src : Sources) {
5714 EVT SrcEltTy = Src.ShuffleVec.getValueType().getVectorElementType();
5715 if (SrcEltTy == SmallestEltTy)
5717 assert(ShuffleVT.getVectorElementType() == SmallestEltTy);
5718 Src.ShuffleVec = DAG.getNode(ISD::BITCAST, dl, ShuffleVT, Src.ShuffleVec);
5719 Src.WindowScale = SrcEltTy.getSizeInBits() / SmallestEltTy.getSizeInBits();
5720 Src.WindowBase *= Src.WindowScale;
5723 // Final sanity check before we try to actually produce a shuffle.
5725 for (auto Src : Sources)
5726 assert(Src.ShuffleVec.getValueType() == ShuffleVT);
5729 // The stars all align, our next step is to produce the mask for the shuffle.
5730 SmallVector<int, 8> Mask(ShuffleVT.getVectorNumElements(), -1);
5731 int BitsPerShuffleLane = ShuffleVT.getVectorElementType().getSizeInBits();
5732 for (unsigned i = 0; i < VT.getVectorNumElements(); ++i) {
5733 SDValue Entry = Op.getOperand(i);
5734 if (Entry.getOpcode() == ISD::UNDEF)
5737 auto Src = std::find(Sources.begin(), Sources.end(), Entry.getOperand(0));
5738 int EltNo = cast<ConstantSDNode>(Entry.getOperand(1))->getSExtValue();
5740 // EXTRACT_VECTOR_ELT performs an implicit any_ext; BUILD_VECTOR an implicit
5741 // trunc. So only std::min(SrcBits, DestBits) actually get defined in this
5743 EVT OrigEltTy = Entry.getOperand(0).getValueType().getVectorElementType();
5744 int BitsDefined = std::min(OrigEltTy.getSizeInBits(),
5745 VT.getVectorElementType().getSizeInBits());
5746 int LanesDefined = BitsDefined / BitsPerShuffleLane;
5748 // This source is expected to fill ResMultiplier lanes of the final shuffle,
5749 // starting at the appropriate offset.
5750 int *LaneMask = &Mask[i * ResMultiplier];
5752 int ExtractBase = EltNo * Src->WindowScale + Src->WindowBase;
5753 ExtractBase += NumElts * (Src - Sources.begin());
5754 for (int j = 0; j < LanesDefined; ++j)
5755 LaneMask[j] = ExtractBase + j;
5758 // Final check before we try to produce nonsense...
5759 if (!isShuffleMaskLegal(Mask, ShuffleVT))
5762 // We can't handle more than two sources. This should have already
5763 // been checked before this point.
5764 assert(Sources.size() <= 2 && "Too many sources!");
5766 SDValue ShuffleOps[] = { DAG.getUNDEF(ShuffleVT), DAG.getUNDEF(ShuffleVT) };
5767 for (unsigned i = 0; i < Sources.size(); ++i)
5768 ShuffleOps[i] = Sources[i].ShuffleVec;
5770 SDValue Shuffle = DAG.getVectorShuffle(ShuffleVT, dl, ShuffleOps[0],
5771 ShuffleOps[1], &Mask[0]);
5772 return DAG.getNode(ISD::BITCAST, dl, VT, Shuffle);
5775 /// isShuffleMaskLegal - Targets can use this to indicate that they only
5776 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
5777 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
5778 /// are assumed to be legal.
5780 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
5782 if (VT.getVectorNumElements() == 4 &&
5783 (VT.is128BitVector() || VT.is64BitVector())) {
5784 unsigned PFIndexes[4];
5785 for (unsigned i = 0; i != 4; ++i) {
5789 PFIndexes[i] = M[i];
5792 // Compute the index in the perfect shuffle table.
5793 unsigned PFTableIndex =
5794 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
5795 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
5796 unsigned Cost = (PFEntry >> 30);
5802 bool ReverseVEXT, isV_UNDEF;
5803 unsigned Imm, WhichResult;
5805 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5806 return (EltSize >= 32 ||
5807 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
5808 isVREVMask(M, VT, 64) ||
5809 isVREVMask(M, VT, 32) ||
5810 isVREVMask(M, VT, 16) ||
5811 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
5812 isVTBLMask(M, VT) ||
5813 isNEONTwoResultShuffleMask(M, VT, WhichResult, isV_UNDEF) ||
5814 ((VT == MVT::v8i16 || VT == MVT::v16i8) && isReverseMask(M, VT)));
5817 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
5818 /// the specified operations to build the shuffle.
5819 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
5820 SDValue RHS, SelectionDAG &DAG,
5822 unsigned OpNum = (PFEntry >> 26) & 0x0F;
5823 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
5824 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
5827 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
5836 OP_VUZPL, // VUZP, left result
5837 OP_VUZPR, // VUZP, right result
5838 OP_VZIPL, // VZIP, left result
5839 OP_VZIPR, // VZIP, right result
5840 OP_VTRNL, // VTRN, left result
5841 OP_VTRNR // VTRN, right result
5844 if (OpNum == OP_COPY) {
5845 if (LHSID == (1*9+2)*9+3) return LHS;
5846 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
5850 SDValue OpLHS, OpRHS;
5851 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
5852 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
5853 EVT VT = OpLHS.getValueType();
5856 default: llvm_unreachable("Unknown shuffle opcode!");
5858 // VREV divides the vector in half and swaps within the half.
5859 if (VT.getVectorElementType() == MVT::i32 ||
5860 VT.getVectorElementType() == MVT::f32)
5861 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
5862 // vrev <4 x i16> -> VREV32
5863 if (VT.getVectorElementType() == MVT::i16)
5864 return DAG.getNode(ARMISD::VREV32, dl, VT, OpLHS);
5865 // vrev <4 x i8> -> VREV16
5866 assert(VT.getVectorElementType() == MVT::i8);
5867 return DAG.getNode(ARMISD::VREV16, dl, VT, OpLHS);
5872 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
5873 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, dl, MVT::i32));
5877 return DAG.getNode(ARMISD::VEXT, dl, VT,
5879 DAG.getConstant(OpNum - OP_VEXT1 + 1, dl, MVT::i32));
5882 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
5883 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
5886 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
5887 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
5890 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
5891 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
5895 static SDValue LowerVECTOR_SHUFFLEv8i8(SDValue Op,
5896 ArrayRef<int> ShuffleMask,
5897 SelectionDAG &DAG) {
5898 // Check to see if we can use the VTBL instruction.
5899 SDValue V1 = Op.getOperand(0);
5900 SDValue V2 = Op.getOperand(1);
5903 SmallVector<SDValue, 8> VTBLMask;
5904 for (ArrayRef<int>::iterator
5905 I = ShuffleMask.begin(), E = ShuffleMask.end(); I != E; ++I)
5906 VTBLMask.push_back(DAG.getConstant(*I, DL, MVT::i32));
5908 if (V2.getNode()->getOpcode() == ISD::UNDEF)
5909 return DAG.getNode(ARMISD::VTBL1, DL, MVT::v8i8, V1,
5910 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8, VTBLMask));
5912 return DAG.getNode(ARMISD::VTBL2, DL, MVT::v8i8, V1, V2,
5913 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8, VTBLMask));
5916 static SDValue LowerReverse_VECTOR_SHUFFLEv16i8_v8i16(SDValue Op,
5917 SelectionDAG &DAG) {
5919 SDValue OpLHS = Op.getOperand(0);
5920 EVT VT = OpLHS.getValueType();
5922 assert((VT == MVT::v8i16 || VT == MVT::v16i8) &&
5923 "Expect an v8i16/v16i8 type");
5924 OpLHS = DAG.getNode(ARMISD::VREV64, DL, VT, OpLHS);
5925 // For a v16i8 type: After the VREV, we have got <8, ...15, 8, ..., 0>. Now,
5926 // extract the first 8 bytes into the top double word and the last 8 bytes
5927 // into the bottom double word. The v8i16 case is similar.
5928 unsigned ExtractNum = (VT == MVT::v16i8) ? 8 : 4;
5929 return DAG.getNode(ARMISD::VEXT, DL, VT, OpLHS, OpLHS,
5930 DAG.getConstant(ExtractNum, DL, MVT::i32));
5933 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
5934 SDValue V1 = Op.getOperand(0);
5935 SDValue V2 = Op.getOperand(1);
5937 EVT VT = Op.getValueType();
5938 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
5940 // Convert shuffles that are directly supported on NEON to target-specific
5941 // DAG nodes, instead of keeping them as shuffles and matching them again
5942 // during code selection. This is more efficient and avoids the possibility
5943 // of inconsistencies between legalization and selection.
5944 // FIXME: floating-point vectors should be canonicalized to integer vectors
5945 // of the same time so that they get CSEd properly.
5946 ArrayRef<int> ShuffleMask = SVN->getMask();
5948 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5949 if (EltSize <= 32) {
5950 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
5951 int Lane = SVN->getSplatIndex();
5952 // If this is undef splat, generate it via "just" vdup, if possible.
5953 if (Lane == -1) Lane = 0;
5955 // Test if V1 is a SCALAR_TO_VECTOR.
5956 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
5957 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
5959 // Test if V1 is a BUILD_VECTOR which is equivalent to a SCALAR_TO_VECTOR
5960 // (and probably will turn into a SCALAR_TO_VECTOR once legalization
5962 if (Lane == 0 && V1.getOpcode() == ISD::BUILD_VECTOR &&
5963 !isa<ConstantSDNode>(V1.getOperand(0))) {
5964 bool IsScalarToVector = true;
5965 for (unsigned i = 1, e = V1.getNumOperands(); i != e; ++i)
5966 if (V1.getOperand(i).getOpcode() != ISD::UNDEF) {
5967 IsScalarToVector = false;
5970 if (IsScalarToVector)
5971 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
5973 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
5974 DAG.getConstant(Lane, dl, MVT::i32));
5979 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
5982 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
5983 DAG.getConstant(Imm, dl, MVT::i32));
5986 if (isVREVMask(ShuffleMask, VT, 64))
5987 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
5988 if (isVREVMask(ShuffleMask, VT, 32))
5989 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
5990 if (isVREVMask(ShuffleMask, VT, 16))
5991 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
5993 if (V2->getOpcode() == ISD::UNDEF &&
5994 isSingletonVEXTMask(ShuffleMask, VT, Imm)) {
5995 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V1,
5996 DAG.getConstant(Imm, dl, MVT::i32));
5999 // Check for Neon shuffles that modify both input vectors in place.
6000 // If both results are used, i.e., if there are two shuffles with the same
6001 // source operands and with masks corresponding to both results of one of
6002 // these operations, DAG memoization will ensure that a single node is
6003 // used for both shuffles.
6004 unsigned WhichResult;
6006 if (unsigned ShuffleOpc = isNEONTwoResultShuffleMask(
6007 ShuffleMask, VT, WhichResult, isV_UNDEF)) {
6010 return DAG.getNode(ShuffleOpc, dl, DAG.getVTList(VT, VT), V1, V2)
6011 .getValue(WhichResult);
6014 // Also check for these shuffles through CONCAT_VECTORS: we canonicalize
6015 // shuffles that produce a result larger than their operands with:
6016 // shuffle(concat(v1, undef), concat(v2, undef))
6018 // shuffle(concat(v1, v2), undef)
6019 // because we can access quad vectors (see PerformVECTOR_SHUFFLECombine).
6021 // This is useful in the general case, but there are special cases where
6022 // native shuffles produce larger results: the two-result ops.
6024 // Look through the concat when lowering them:
6025 // shuffle(concat(v1, v2), undef)
6027 // concat(VZIP(v1, v2):0, :1)
6029 if (V1->getOpcode() == ISD::CONCAT_VECTORS &&
6030 V2->getOpcode() == ISD::UNDEF) {
6031 SDValue SubV1 = V1->getOperand(0);
6032 SDValue SubV2 = V1->getOperand(1);
6033 EVT SubVT = SubV1.getValueType();
6035 // We expect these to have been canonicalized to -1.
6036 assert(std::all_of(ShuffleMask.begin(), ShuffleMask.end(), [&](int i) {
6037 return i < (int)VT.getVectorNumElements();
6038 }) && "Unexpected shuffle index into UNDEF operand!");
6040 if (unsigned ShuffleOpc = isNEONTwoResultShuffleMask(
6041 ShuffleMask, SubVT, WhichResult, isV_UNDEF)) {
6044 assert((WhichResult == 0) &&
6045 "In-place shuffle of concat can only have one result!");
6046 SDValue Res = DAG.getNode(ShuffleOpc, dl, DAG.getVTList(SubVT, SubVT),
6048 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, Res.getValue(0),
6054 // If the shuffle is not directly supported and it has 4 elements, use
6055 // the PerfectShuffle-generated table to synthesize it from other shuffles.
6056 unsigned NumElts = VT.getVectorNumElements();
6058 unsigned PFIndexes[4];
6059 for (unsigned i = 0; i != 4; ++i) {
6060 if (ShuffleMask[i] < 0)
6063 PFIndexes[i] = ShuffleMask[i];
6066 // Compute the index in the perfect shuffle table.
6067 unsigned PFTableIndex =
6068 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
6069 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
6070 unsigned Cost = (PFEntry >> 30);
6073 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
6076 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
6077 if (EltSize >= 32) {
6078 // Do the expansion with floating-point types, since that is what the VFP
6079 // registers are defined to use, and since i64 is not legal.
6080 EVT EltVT = EVT::getFloatingPointVT(EltSize);
6081 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
6082 V1 = DAG.getNode(ISD::BITCAST, dl, VecVT, V1);
6083 V2 = DAG.getNode(ISD::BITCAST, dl, VecVT, V2);
6084 SmallVector<SDValue, 8> Ops;
6085 for (unsigned i = 0; i < NumElts; ++i) {
6086 if (ShuffleMask[i] < 0)
6087 Ops.push_back(DAG.getUNDEF(EltVT));
6089 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
6090 ShuffleMask[i] < (int)NumElts ? V1 : V2,
6091 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
6094 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, Ops);
6095 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
6098 if ((VT == MVT::v8i16 || VT == MVT::v16i8) && isReverseMask(ShuffleMask, VT))
6099 return LowerReverse_VECTOR_SHUFFLEv16i8_v8i16(Op, DAG);
6101 if (VT == MVT::v8i8) {
6102 SDValue NewOp = LowerVECTOR_SHUFFLEv8i8(Op, ShuffleMask, DAG);
6103 if (NewOp.getNode())
6110 static SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
6111 // INSERT_VECTOR_ELT is legal only for immediate indexes.
6112 SDValue Lane = Op.getOperand(2);
6113 if (!isa<ConstantSDNode>(Lane))
6119 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
6120 // EXTRACT_VECTOR_ELT is legal only for immediate indexes.
6121 SDValue Lane = Op.getOperand(1);
6122 if (!isa<ConstantSDNode>(Lane))
6125 SDValue Vec = Op.getOperand(0);
6126 if (Op.getValueType() == MVT::i32 &&
6127 Vec.getValueType().getVectorElementType().getSizeInBits() < 32) {
6129 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
6135 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
6136 // The only time a CONCAT_VECTORS operation can have legal types is when
6137 // two 64-bit vectors are concatenated to a 128-bit vector.
6138 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
6139 "unexpected CONCAT_VECTORS");
6141 SDValue Val = DAG.getUNDEF(MVT::v2f64);
6142 SDValue Op0 = Op.getOperand(0);
6143 SDValue Op1 = Op.getOperand(1);
6144 if (Op0.getOpcode() != ISD::UNDEF)
6145 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
6146 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op0),
6147 DAG.getIntPtrConstant(0, dl));
6148 if (Op1.getOpcode() != ISD::UNDEF)
6149 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
6150 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op1),
6151 DAG.getIntPtrConstant(1, dl));
6152 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Val);
6155 /// isExtendedBUILD_VECTOR - Check if N is a constant BUILD_VECTOR where each
6156 /// element has been zero/sign-extended, depending on the isSigned parameter,
6157 /// from an integer type half its size.
6158 static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG,
6160 // A v2i64 BUILD_VECTOR will have been legalized to a BITCAST from v4i32.
6161 EVT VT = N->getValueType(0);
6162 if (VT == MVT::v2i64 && N->getOpcode() == ISD::BITCAST) {
6163 SDNode *BVN = N->getOperand(0).getNode();
6164 if (BVN->getValueType(0) != MVT::v4i32 ||
6165 BVN->getOpcode() != ISD::BUILD_VECTOR)
6167 unsigned LoElt = DAG.getDataLayout().isBigEndian() ? 1 : 0;
6168 unsigned HiElt = 1 - LoElt;
6169 ConstantSDNode *Lo0 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt));
6170 ConstantSDNode *Hi0 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt));
6171 ConstantSDNode *Lo1 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt+2));
6172 ConstantSDNode *Hi1 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt+2));
6173 if (!Lo0 || !Hi0 || !Lo1 || !Hi1)
6176 if (Hi0->getSExtValue() == Lo0->getSExtValue() >> 32 &&
6177 Hi1->getSExtValue() == Lo1->getSExtValue() >> 32)
6180 if (Hi0->isNullValue() && Hi1->isNullValue())
6186 if (N->getOpcode() != ISD::BUILD_VECTOR)
6189 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
6190 SDNode *Elt = N->getOperand(i).getNode();
6191 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
6192 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
6193 unsigned HalfSize = EltSize / 2;
6195 if (!isIntN(HalfSize, C->getSExtValue()))
6198 if (!isUIntN(HalfSize, C->getZExtValue()))
6209 /// isSignExtended - Check if a node is a vector value that is sign-extended
6210 /// or a constant BUILD_VECTOR with sign-extended elements.
6211 static bool isSignExtended(SDNode *N, SelectionDAG &DAG) {
6212 if (N->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N))
6214 if (isExtendedBUILD_VECTOR(N, DAG, true))
6219 /// isZeroExtended - Check if a node is a vector value that is zero-extended
6220 /// or a constant BUILD_VECTOR with zero-extended elements.
6221 static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) {
6222 if (N->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N))
6224 if (isExtendedBUILD_VECTOR(N, DAG, false))
6229 static EVT getExtensionTo64Bits(const EVT &OrigVT) {
6230 if (OrigVT.getSizeInBits() >= 64)
6233 assert(OrigVT.isSimple() && "Expecting a simple value type");
6235 MVT::SimpleValueType OrigSimpleTy = OrigVT.getSimpleVT().SimpleTy;
6236 switch (OrigSimpleTy) {
6237 default: llvm_unreachable("Unexpected Vector Type");
6246 /// AddRequiredExtensionForVMULL - Add a sign/zero extension to extend the total
6247 /// value size to 64 bits. We need a 64-bit D register as an operand to VMULL.
6248 /// We insert the required extension here to get the vector to fill a D register.
6249 static SDValue AddRequiredExtensionForVMULL(SDValue N, SelectionDAG &DAG,
6252 unsigned ExtOpcode) {
6253 // The vector originally had a size of OrigTy. It was then extended to ExtTy.
6254 // We expect the ExtTy to be 128-bits total. If the OrigTy is less than
6255 // 64-bits we need to insert a new extension so that it will be 64-bits.
6256 assert(ExtTy.is128BitVector() && "Unexpected extension size");
6257 if (OrigTy.getSizeInBits() >= 64)
6260 // Must extend size to at least 64 bits to be used as an operand for VMULL.
6261 EVT NewVT = getExtensionTo64Bits(OrigTy);
6263 return DAG.getNode(ExtOpcode, SDLoc(N), NewVT, N);
6266 /// SkipLoadExtensionForVMULL - return a load of the original vector size that
6267 /// does not do any sign/zero extension. If the original vector is less
6268 /// than 64 bits, an appropriate extension will be added after the load to
6269 /// reach a total size of 64 bits. We have to add the extension separately
6270 /// because ARM does not have a sign/zero extending load for vectors.
6271 static SDValue SkipLoadExtensionForVMULL(LoadSDNode *LD, SelectionDAG& DAG) {
6272 EVT ExtendedTy = getExtensionTo64Bits(LD->getMemoryVT());
6274 // The load already has the right type.
6275 if (ExtendedTy == LD->getMemoryVT())
6276 return DAG.getLoad(LD->getMemoryVT(), SDLoc(LD), LD->getChain(),
6277 LD->getBasePtr(), LD->getPointerInfo(), LD->isVolatile(),
6278 LD->isNonTemporal(), LD->isInvariant(),
6279 LD->getAlignment());
6281 // We need to create a zextload/sextload. We cannot just create a load
6282 // followed by a zext/zext node because LowerMUL is also run during normal
6283 // operation legalization where we can't create illegal types.
6284 return DAG.getExtLoad(LD->getExtensionType(), SDLoc(LD), ExtendedTy,
6285 LD->getChain(), LD->getBasePtr(), LD->getPointerInfo(),
6286 LD->getMemoryVT(), LD->isVolatile(), LD->isInvariant(),
6287 LD->isNonTemporal(), LD->getAlignment());
6290 /// SkipExtensionForVMULL - For a node that is a SIGN_EXTEND, ZERO_EXTEND,
6291 /// extending load, or BUILD_VECTOR with extended elements, return the
6292 /// unextended value. The unextended vector should be 64 bits so that it can
6293 /// be used as an operand to a VMULL instruction. If the original vector size
6294 /// before extension is less than 64 bits we add a an extension to resize
6295 /// the vector to 64 bits.
6296 static SDValue SkipExtensionForVMULL(SDNode *N, SelectionDAG &DAG) {
6297 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
6298 return AddRequiredExtensionForVMULL(N->getOperand(0), DAG,
6299 N->getOperand(0)->getValueType(0),
6303 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
6304 return SkipLoadExtensionForVMULL(LD, DAG);
6306 // Otherwise, the value must be a BUILD_VECTOR. For v2i64, it will
6307 // have been legalized as a BITCAST from v4i32.
6308 if (N->getOpcode() == ISD::BITCAST) {
6309 SDNode *BVN = N->getOperand(0).getNode();
6310 assert(BVN->getOpcode() == ISD::BUILD_VECTOR &&
6311 BVN->getValueType(0) == MVT::v4i32 && "expected v4i32 BUILD_VECTOR");
6312 unsigned LowElt = DAG.getDataLayout().isBigEndian() ? 1 : 0;
6313 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), MVT::v2i32,
6314 BVN->getOperand(LowElt), BVN->getOperand(LowElt+2));
6316 // Construct a new BUILD_VECTOR with elements truncated to half the size.
6317 assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR");
6318 EVT VT = N->getValueType(0);
6319 unsigned EltSize = VT.getVectorElementType().getSizeInBits() / 2;
6320 unsigned NumElts = VT.getVectorNumElements();
6321 MVT TruncVT = MVT::getIntegerVT(EltSize);
6322 SmallVector<SDValue, 8> Ops;
6324 for (unsigned i = 0; i != NumElts; ++i) {
6325 ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i));
6326 const APInt &CInt = C->getAPIntValue();
6327 // Element types smaller than 32 bits are not legal, so use i32 elements.
6328 // The values are implicitly truncated so sext vs. zext doesn't matter.
6329 Ops.push_back(DAG.getConstant(CInt.zextOrTrunc(32), dl, MVT::i32));
6331 return DAG.getNode(ISD::BUILD_VECTOR, dl,
6332 MVT::getVectorVT(TruncVT, NumElts), Ops);
6335 static bool isAddSubSExt(SDNode *N, SelectionDAG &DAG) {
6336 unsigned Opcode = N->getOpcode();
6337 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
6338 SDNode *N0 = N->getOperand(0).getNode();
6339 SDNode *N1 = N->getOperand(1).getNode();
6340 return N0->hasOneUse() && N1->hasOneUse() &&
6341 isSignExtended(N0, DAG) && isSignExtended(N1, DAG);
6346 static bool isAddSubZExt(SDNode *N, SelectionDAG &DAG) {
6347 unsigned Opcode = N->getOpcode();
6348 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
6349 SDNode *N0 = N->getOperand(0).getNode();
6350 SDNode *N1 = N->getOperand(1).getNode();
6351 return N0->hasOneUse() && N1->hasOneUse() &&
6352 isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG);
6357 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
6358 // Multiplications are only custom-lowered for 128-bit vectors so that
6359 // VMULL can be detected. Otherwise v2i64 multiplications are not legal.
6360 EVT VT = Op.getValueType();
6361 assert(VT.is128BitVector() && VT.isInteger() &&
6362 "unexpected type for custom-lowering ISD::MUL");
6363 SDNode *N0 = Op.getOperand(0).getNode();
6364 SDNode *N1 = Op.getOperand(1).getNode();
6365 unsigned NewOpc = 0;
6367 bool isN0SExt = isSignExtended(N0, DAG);
6368 bool isN1SExt = isSignExtended(N1, DAG);
6369 if (isN0SExt && isN1SExt)
6370 NewOpc = ARMISD::VMULLs;
6372 bool isN0ZExt = isZeroExtended(N0, DAG);
6373 bool isN1ZExt = isZeroExtended(N1, DAG);
6374 if (isN0ZExt && isN1ZExt)
6375 NewOpc = ARMISD::VMULLu;
6376 else if (isN1SExt || isN1ZExt) {
6377 // Look for (s/zext A + s/zext B) * (s/zext C). We want to turn these
6378 // into (s/zext A * s/zext C) + (s/zext B * s/zext C)
6379 if (isN1SExt && isAddSubSExt(N0, DAG)) {
6380 NewOpc = ARMISD::VMULLs;
6382 } else if (isN1ZExt && isAddSubZExt(N0, DAG)) {
6383 NewOpc = ARMISD::VMULLu;
6385 } else if (isN0ZExt && isAddSubZExt(N1, DAG)) {
6387 NewOpc = ARMISD::VMULLu;
6393 if (VT == MVT::v2i64)
6394 // Fall through to expand this. It is not legal.
6397 // Other vector multiplications are legal.
6402 // Legalize to a VMULL instruction.
6405 SDValue Op1 = SkipExtensionForVMULL(N1, DAG);
6407 Op0 = SkipExtensionForVMULL(N0, DAG);
6408 assert(Op0.getValueType().is64BitVector() &&
6409 Op1.getValueType().is64BitVector() &&
6410 "unexpected types for extended operands to VMULL");
6411 return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
6414 // Optimizing (zext A + zext B) * C, to (VMULL A, C) + (VMULL B, C) during
6415 // isel lowering to take advantage of no-stall back to back vmul + vmla.
6422 SDValue N00 = SkipExtensionForVMULL(N0->getOperand(0).getNode(), DAG);
6423 SDValue N01 = SkipExtensionForVMULL(N0->getOperand(1).getNode(), DAG);
6424 EVT Op1VT = Op1.getValueType();
6425 return DAG.getNode(N0->getOpcode(), DL, VT,
6426 DAG.getNode(NewOpc, DL, VT,
6427 DAG.getNode(ISD::BITCAST, DL, Op1VT, N00), Op1),
6428 DAG.getNode(NewOpc, DL, VT,
6429 DAG.getNode(ISD::BITCAST, DL, Op1VT, N01), Op1));
6433 LowerSDIV_v4i8(SDValue X, SDValue Y, SDLoc dl, SelectionDAG &DAG) {
6435 // float4 xf = vcvt_f32_s32(vmovl_s16(a.lo));
6436 // float4 yf = vcvt_f32_s32(vmovl_s16(b.lo));
6437 X = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, X);
6438 Y = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, Y);
6439 X = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, X);
6440 Y = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, Y);
6441 // Get reciprocal estimate.
6442 // float4 recip = vrecpeq_f32(yf);
6443 Y = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6444 DAG.getConstant(Intrinsic::arm_neon_vrecpe, dl, MVT::i32),
6446 // Because char has a smaller range than uchar, we can actually get away
6447 // without any newton steps. This requires that we use a weird bias
6448 // of 0xb000, however (again, this has been exhaustively tested).
6449 // float4 result = as_float4(as_int4(xf*recip) + 0xb000);
6450 X = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, X, Y);
6451 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, X);
6452 Y = DAG.getConstant(0xb000, dl, MVT::i32);
6453 Y = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Y, Y, Y, Y);
6454 X = DAG.getNode(ISD::ADD, dl, MVT::v4i32, X, Y);
6455 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, X);
6456 // Convert back to short.
6457 X = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, X);
6458 X = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, X);
6463 LowerSDIV_v4i16(SDValue N0, SDValue N1, SDLoc dl, SelectionDAG &DAG) {
6465 // Convert to float.
6466 // float4 yf = vcvt_f32_s32(vmovl_s16(y));
6467 // float4 xf = vcvt_f32_s32(vmovl_s16(x));
6468 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N0);
6469 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N1);
6470 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
6471 N1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
6473 // Use reciprocal estimate and one refinement step.
6474 // float4 recip = vrecpeq_f32(yf);
6475 // recip *= vrecpsq_f32(yf, recip);
6476 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6477 DAG.getConstant(Intrinsic::arm_neon_vrecpe, dl, MVT::i32),
6479 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6480 DAG.getConstant(Intrinsic::arm_neon_vrecps, dl, MVT::i32),
6482 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
6483 // Because short has a smaller range than ushort, we can actually get away
6484 // with only a single newton step. This requires that we use a weird bias
6485 // of 89, however (again, this has been exhaustively tested).
6486 // float4 result = as_float4(as_int4(xf*recip) + 0x89);
6487 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
6488 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
6489 N1 = DAG.getConstant(0x89, dl, MVT::i32);
6490 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
6491 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
6492 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
6493 // Convert back to integer and return.
6494 // return vmovn_s32(vcvt_s32_f32(result));
6495 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
6496 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
6500 static SDValue LowerSDIV(SDValue Op, SelectionDAG &DAG) {
6501 EVT VT = Op.getValueType();
6502 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
6503 "unexpected type for custom-lowering ISD::SDIV");
6506 SDValue N0 = Op.getOperand(0);
6507 SDValue N1 = Op.getOperand(1);
6510 if (VT == MVT::v8i8) {
6511 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N0);
6512 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N1);
6514 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
6515 DAG.getIntPtrConstant(4, dl));
6516 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
6517 DAG.getIntPtrConstant(4, dl));
6518 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
6519 DAG.getIntPtrConstant(0, dl));
6520 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
6521 DAG.getIntPtrConstant(0, dl));
6523 N0 = LowerSDIV_v4i8(N0, N1, dl, DAG); // v4i16
6524 N2 = LowerSDIV_v4i8(N2, N3, dl, DAG); // v4i16
6526 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
6527 N0 = LowerCONCAT_VECTORS(N0, DAG);
6529 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v8i8, N0);
6532 return LowerSDIV_v4i16(N0, N1, dl, DAG);
6535 static SDValue LowerUDIV(SDValue Op, SelectionDAG &DAG) {
6536 EVT VT = Op.getValueType();
6537 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
6538 "unexpected type for custom-lowering ISD::UDIV");
6541 SDValue N0 = Op.getOperand(0);
6542 SDValue N1 = Op.getOperand(1);
6545 if (VT == MVT::v8i8) {
6546 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N0);
6547 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N1);
6549 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
6550 DAG.getIntPtrConstant(4, dl));
6551 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
6552 DAG.getIntPtrConstant(4, dl));
6553 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
6554 DAG.getIntPtrConstant(0, dl));
6555 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
6556 DAG.getIntPtrConstant(0, dl));
6558 N0 = LowerSDIV_v4i16(N0, N1, dl, DAG); // v4i16
6559 N2 = LowerSDIV_v4i16(N2, N3, dl, DAG); // v4i16
6561 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
6562 N0 = LowerCONCAT_VECTORS(N0, DAG);
6564 N0 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v8i8,
6565 DAG.getConstant(Intrinsic::arm_neon_vqmovnsu, dl,
6571 // v4i16 sdiv ... Convert to float.
6572 // float4 yf = vcvt_f32_s32(vmovl_u16(y));
6573 // float4 xf = vcvt_f32_s32(vmovl_u16(x));
6574 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N0);
6575 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N1);
6576 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
6577 SDValue BN1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
6579 // Use reciprocal estimate and two refinement steps.
6580 // float4 recip = vrecpeq_f32(yf);
6581 // recip *= vrecpsq_f32(yf, recip);
6582 // recip *= vrecpsq_f32(yf, recip);
6583 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6584 DAG.getConstant(Intrinsic::arm_neon_vrecpe, dl, MVT::i32),
6586 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6587 DAG.getConstant(Intrinsic::arm_neon_vrecps, dl, MVT::i32),
6589 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
6590 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6591 DAG.getConstant(Intrinsic::arm_neon_vrecps, dl, MVT::i32),
6593 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
6594 // Simply multiplying by the reciprocal estimate can leave us a few ulps
6595 // too low, so we add 2 ulps (exhaustive testing shows that this is enough,
6596 // and that it will never cause us to return an answer too large).
6597 // float4 result = as_float4(as_int4(xf*recip) + 2);
6598 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
6599 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
6600 N1 = DAG.getConstant(2, dl, MVT::i32);
6601 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
6602 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
6603 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
6604 // Convert back to integer and return.
6605 // return vmovn_u32(vcvt_s32_f32(result));
6606 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
6607 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
6611 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
6612 EVT VT = Op.getNode()->getValueType(0);
6613 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
6616 bool ExtraOp = false;
6617 switch (Op.getOpcode()) {
6618 default: llvm_unreachable("Invalid code");
6619 case ISD::ADDC: Opc = ARMISD::ADDC; break;
6620 case ISD::ADDE: Opc = ARMISD::ADDE; ExtraOp = true; break;
6621 case ISD::SUBC: Opc = ARMISD::SUBC; break;
6622 case ISD::SUBE: Opc = ARMISD::SUBE; ExtraOp = true; break;
6626 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
6628 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
6629 Op.getOperand(1), Op.getOperand(2));
6632 SDValue ARMTargetLowering::LowerFSINCOS(SDValue Op, SelectionDAG &DAG) const {
6633 assert(Subtarget->isTargetDarwin());
6635 // For iOS, we want to call an alternative entry point: __sincos_stret,
6636 // return values are passed via sret.
6638 SDValue Arg = Op.getOperand(0);
6639 EVT ArgVT = Arg.getValueType();
6640 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
6641 auto PtrVT = getPointerTy(DAG.getDataLayout());
6643 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
6645 // Pair of floats / doubles used to pass the result.
6646 StructType *RetTy = StructType::get(ArgTy, ArgTy, nullptr);
6648 // Create stack object for sret.
6649 auto &DL = DAG.getDataLayout();
6650 const uint64_t ByteSize = DL.getTypeAllocSize(RetTy);
6651 const unsigned StackAlign = DL.getPrefTypeAlignment(RetTy);
6652 int FrameIdx = FrameInfo->CreateStackObject(ByteSize, StackAlign, false);
6653 SDValue SRet = DAG.getFrameIndex(FrameIdx, getPointerTy(DL));
6659 Entry.Ty = RetTy->getPointerTo();
6660 Entry.isSExt = false;
6661 Entry.isZExt = false;
6662 Entry.isSRet = true;
6663 Args.push_back(Entry);
6667 Entry.isSExt = false;
6668 Entry.isZExt = false;
6669 Args.push_back(Entry);
6671 const char *LibcallName = (ArgVT == MVT::f64)
6672 ? "__sincos_stret" : "__sincosf_stret";
6673 SDValue Callee = DAG.getExternalSymbol(LibcallName, getPointerTy(DL));
6675 TargetLowering::CallLoweringInfo CLI(DAG);
6676 CLI.setDebugLoc(dl).setChain(DAG.getEntryNode())
6677 .setCallee(CallingConv::C, Type::getVoidTy(*DAG.getContext()), Callee,
6679 .setDiscardResult();
6681 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
6683 SDValue LoadSin = DAG.getLoad(ArgVT, dl, CallResult.second, SRet,
6684 MachinePointerInfo(), false, false, false, 0);
6686 // Address of cos field.
6687 SDValue Add = DAG.getNode(ISD::ADD, dl, PtrVT, SRet,
6688 DAG.getIntPtrConstant(ArgVT.getStoreSize(), dl));
6689 SDValue LoadCos = DAG.getLoad(ArgVT, dl, LoadSin.getValue(1), Add,
6690 MachinePointerInfo(), false, false, false, 0);
6692 SDVTList Tys = DAG.getVTList(ArgVT, ArgVT);
6693 return DAG.getNode(ISD::MERGE_VALUES, dl, Tys,
6694 LoadSin.getValue(0), LoadCos.getValue(0));
6697 static SDValue LowerAtomicLoadStore(SDValue Op, SelectionDAG &DAG) {
6698 // Monotonic load/store is legal for all targets
6699 if (cast<AtomicSDNode>(Op)->getOrdering() <= Monotonic)
6702 // Acquire/Release load/store is not legal for targets without a
6703 // dmb or equivalent available.
6707 static void ReplaceREADCYCLECOUNTER(SDNode *N,
6708 SmallVectorImpl<SDValue> &Results,
6710 const ARMSubtarget *Subtarget) {
6712 SDValue Cycles32, OutChain;
6714 if (Subtarget->hasPerfMon()) {
6715 // Under Power Management extensions, the cycle-count is:
6716 // mrc p15, #0, <Rt>, c9, c13, #0
6717 SDValue Ops[] = { N->getOperand(0), // Chain
6718 DAG.getConstant(Intrinsic::arm_mrc, DL, MVT::i32),
6719 DAG.getConstant(15, DL, MVT::i32),
6720 DAG.getConstant(0, DL, MVT::i32),
6721 DAG.getConstant(9, DL, MVT::i32),
6722 DAG.getConstant(13, DL, MVT::i32),
6723 DAG.getConstant(0, DL, MVT::i32)
6726 Cycles32 = DAG.getNode(ISD::INTRINSIC_W_CHAIN, DL,
6727 DAG.getVTList(MVT::i32, MVT::Other), Ops);
6728 OutChain = Cycles32.getValue(1);
6730 // Intrinsic is defined to return 0 on unsupported platforms. Technically
6731 // there are older ARM CPUs that have implementation-specific ways of
6732 // obtaining this information (FIXME!).
6733 Cycles32 = DAG.getConstant(0, DL, MVT::i32);
6734 OutChain = DAG.getEntryNode();
6738 SDValue Cycles64 = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64,
6739 Cycles32, DAG.getConstant(0, DL, MVT::i32));
6740 Results.push_back(Cycles64);
6741 Results.push_back(OutChain);
6744 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
6745 switch (Op.getOpcode()) {
6746 default: llvm_unreachable("Don't know how to custom lower this!");
6747 case ISD::WRITE_REGISTER: return LowerWRITE_REGISTER(Op, DAG);
6748 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
6749 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
6750 case ISD::GlobalAddress:
6751 switch (Subtarget->getTargetTriple().getObjectFormat()) {
6752 default: llvm_unreachable("unknown object format");
6754 return LowerGlobalAddressWindows(Op, DAG);
6756 return LowerGlobalAddressELF(Op, DAG);
6758 return LowerGlobalAddressDarwin(Op, DAG);
6760 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
6761 case ISD::SELECT: return LowerSELECT(Op, DAG);
6762 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
6763 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
6764 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
6765 case ISD::VASTART: return LowerVASTART(Op, DAG);
6766 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, DAG, Subtarget);
6767 case ISD::PREFETCH: return LowerPREFETCH(Op, DAG, Subtarget);
6768 case ISD::SINT_TO_FP:
6769 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
6770 case ISD::FP_TO_SINT:
6771 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
6772 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
6773 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
6774 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
6775 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
6776 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
6777 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
6778 case ISD::EH_SJLJ_SETUP_DISPATCH: return LowerEH_SJLJ_SETUP_DISPATCH(Op, DAG);
6779 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
6781 case ISD::BITCAST: return ExpandBITCAST(Op.getNode(), DAG);
6784 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
6785 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
6786 case ISD::SRL_PARTS:
6787 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
6789 case ISD::CTTZ_ZERO_UNDEF: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
6790 case ISD::CTPOP: return LowerCTPOP(Op.getNode(), DAG, Subtarget);
6791 case ISD::SETCC: return LowerVSETCC(Op, DAG);
6792 case ISD::ConstantFP: return LowerConstantFP(Op, DAG, Subtarget);
6793 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG, Subtarget);
6794 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
6795 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
6796 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
6797 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
6798 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
6799 case ISD::MUL: return LowerMUL(Op, DAG);
6800 case ISD::SDIV: return LowerSDIV(Op, DAG);
6801 case ISD::UDIV: return LowerUDIV(Op, DAG);
6805 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
6810 return LowerXALUO(Op, DAG);
6811 case ISD::ATOMIC_LOAD:
6812 case ISD::ATOMIC_STORE: return LowerAtomicLoadStore(Op, DAG);
6813 case ISD::FSINCOS: return LowerFSINCOS(Op, DAG);
6815 case ISD::UDIVREM: return LowerDivRem(Op, DAG);
6816 case ISD::DYNAMIC_STACKALLOC:
6817 if (Subtarget->getTargetTriple().isWindowsItaniumEnvironment())
6818 return LowerDYNAMIC_STACKALLOC(Op, DAG);
6819 llvm_unreachable("Don't know how to custom lower this!");
6820 case ISD::FP_ROUND: return LowerFP_ROUND(Op, DAG);
6821 case ISD::FP_EXTEND: return LowerFP_EXTEND(Op, DAG);
6825 /// ReplaceNodeResults - Replace the results of node with an illegal result
6826 /// type with new values built out of custom code.
6827 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
6828 SmallVectorImpl<SDValue>&Results,
6829 SelectionDAG &DAG) const {
6831 switch (N->getOpcode()) {
6833 llvm_unreachable("Don't know how to custom expand this!");
6834 case ISD::READ_REGISTER:
6835 ExpandREAD_REGISTER(N, Results, DAG);
6838 Res = ExpandBITCAST(N, DAG);
6842 Res = Expand64BitShift(N, DAG, Subtarget);
6844 case ISD::READCYCLECOUNTER:
6845 ReplaceREADCYCLECOUNTER(N, Results, DAG, Subtarget);
6849 Results.push_back(Res);
6852 //===----------------------------------------------------------------------===//
6853 // ARM Scheduler Hooks
6854 //===----------------------------------------------------------------------===//
6856 /// SetupEntryBlockForSjLj - Insert code into the entry block that creates and
6857 /// registers the function context.
6858 void ARMTargetLowering::
6859 SetupEntryBlockForSjLj(MachineInstr *MI, MachineBasicBlock *MBB,
6860 MachineBasicBlock *DispatchBB, int FI) const {
6861 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
6862 DebugLoc dl = MI->getDebugLoc();
6863 MachineFunction *MF = MBB->getParent();
6864 MachineRegisterInfo *MRI = &MF->getRegInfo();
6865 MachineConstantPool *MCP = MF->getConstantPool();
6866 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
6867 const Function *F = MF->getFunction();
6869 bool isThumb = Subtarget->isThumb();
6870 bool isThumb2 = Subtarget->isThumb2();
6872 unsigned PCLabelId = AFI->createPICLabelUId();
6873 unsigned PCAdj = (isThumb || isThumb2) ? 4 : 8;
6874 ARMConstantPoolValue *CPV =
6875 ARMConstantPoolMBB::Create(F->getContext(), DispatchBB, PCLabelId, PCAdj);
6876 unsigned CPI = MCP->getConstantPoolIndex(CPV, 4);
6878 const TargetRegisterClass *TRC = isThumb ? &ARM::tGPRRegClass
6879 : &ARM::GPRRegClass;
6881 // Grab constant pool and fixed stack memory operands.
6882 MachineMemOperand *CPMMO =
6883 MF->getMachineMemOperand(MachinePointerInfo::getConstantPool(),
6884 MachineMemOperand::MOLoad, 4, 4);
6886 MachineMemOperand *FIMMOSt =
6887 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
6888 MachineMemOperand::MOStore, 4, 4);
6890 // Load the address of the dispatch MBB into the jump buffer.
6892 // Incoming value: jbuf
6893 // ldr.n r5, LCPI1_1
6896 // str r5, [$jbuf, #+4] ; &jbuf[1]
6897 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6898 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2LDRpci), NewVReg1)
6899 .addConstantPoolIndex(CPI)
6900 .addMemOperand(CPMMO));
6901 // Set the low bit because of thumb mode.
6902 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6904 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2ORRri), NewVReg2)
6905 .addReg(NewVReg1, RegState::Kill)
6907 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6908 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg3)
6909 .addReg(NewVReg2, RegState::Kill)
6911 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2STRi12))
6912 .addReg(NewVReg3, RegState::Kill)
6914 .addImm(36) // &jbuf[1] :: pc
6915 .addMemOperand(FIMMOSt));
6916 } else if (isThumb) {
6917 // Incoming value: jbuf
6918 // ldr.n r1, LCPI1_4
6922 // add r2, $jbuf, #+4 ; &jbuf[1]
6924 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6925 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tLDRpci), NewVReg1)
6926 .addConstantPoolIndex(CPI)
6927 .addMemOperand(CPMMO));
6928 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6929 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg2)
6930 .addReg(NewVReg1, RegState::Kill)
6932 // Set the low bit because of thumb mode.
6933 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6934 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tMOVi8), NewVReg3)
6935 .addReg(ARM::CPSR, RegState::Define)
6937 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6938 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tORR), NewVReg4)
6939 .addReg(ARM::CPSR, RegState::Define)
6940 .addReg(NewVReg2, RegState::Kill)
6941 .addReg(NewVReg3, RegState::Kill));
6942 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6943 BuildMI(*MBB, MI, dl, TII->get(ARM::tADDframe), NewVReg5)
6945 .addImm(36); // &jbuf[1] :: pc
6946 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tSTRi))
6947 .addReg(NewVReg4, RegState::Kill)
6948 .addReg(NewVReg5, RegState::Kill)
6950 .addMemOperand(FIMMOSt));
6952 // Incoming value: jbuf
6955 // str r1, [$jbuf, #+4] ; &jbuf[1]
6956 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6957 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::LDRi12), NewVReg1)
6958 .addConstantPoolIndex(CPI)
6960 .addMemOperand(CPMMO));
6961 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6962 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::PICADD), NewVReg2)
6963 .addReg(NewVReg1, RegState::Kill)
6964 .addImm(PCLabelId));
6965 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::STRi12))
6966 .addReg(NewVReg2, RegState::Kill)
6968 .addImm(36) // &jbuf[1] :: pc
6969 .addMemOperand(FIMMOSt));
6973 void ARMTargetLowering::EmitSjLjDispatchBlock(MachineInstr *MI,
6974 MachineBasicBlock *MBB) const {
6975 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
6976 DebugLoc dl = MI->getDebugLoc();
6977 MachineFunction *MF = MBB->getParent();
6978 MachineRegisterInfo *MRI = &MF->getRegInfo();
6979 MachineFrameInfo *MFI = MF->getFrameInfo();
6980 int FI = MFI->getFunctionContextIndex();
6982 const TargetRegisterClass *TRC = Subtarget->isThumb() ? &ARM::tGPRRegClass
6983 : &ARM::GPRnopcRegClass;
6985 // Get a mapping of the call site numbers to all of the landing pads they're
6987 DenseMap<unsigned, SmallVector<MachineBasicBlock*, 2> > CallSiteNumToLPad;
6988 unsigned MaxCSNum = 0;
6989 MachineModuleInfo &MMI = MF->getMMI();
6990 for (MachineFunction::iterator BB = MF->begin(), E = MF->end(); BB != E;
6992 if (!BB->isLandingPad()) continue;
6994 // FIXME: We should assert that the EH_LABEL is the first MI in the landing
6996 for (MachineBasicBlock::iterator
6997 II = BB->begin(), IE = BB->end(); II != IE; ++II) {
6998 if (!II->isEHLabel()) continue;
7000 MCSymbol *Sym = II->getOperand(0).getMCSymbol();
7001 if (!MMI.hasCallSiteLandingPad(Sym)) continue;
7003 SmallVectorImpl<unsigned> &CallSiteIdxs = MMI.getCallSiteLandingPad(Sym);
7004 for (SmallVectorImpl<unsigned>::iterator
7005 CSI = CallSiteIdxs.begin(), CSE = CallSiteIdxs.end();
7006 CSI != CSE; ++CSI) {
7007 CallSiteNumToLPad[*CSI].push_back(BB);
7008 MaxCSNum = std::max(MaxCSNum, *CSI);
7014 // Get an ordered list of the machine basic blocks for the jump table.
7015 std::vector<MachineBasicBlock*> LPadList;
7016 SmallPtrSet<MachineBasicBlock*, 64> InvokeBBs;
7017 LPadList.reserve(CallSiteNumToLPad.size());
7018 for (unsigned I = 1; I <= MaxCSNum; ++I) {
7019 SmallVectorImpl<MachineBasicBlock*> &MBBList = CallSiteNumToLPad[I];
7020 for (SmallVectorImpl<MachineBasicBlock*>::iterator
7021 II = MBBList.begin(), IE = MBBList.end(); II != IE; ++II) {
7022 LPadList.push_back(*II);
7023 InvokeBBs.insert((*II)->pred_begin(), (*II)->pred_end());
7027 assert(!LPadList.empty() &&
7028 "No landing pad destinations for the dispatch jump table!");
7030 // Create the jump table and associated information.
7031 MachineJumpTableInfo *JTI =
7032 MF->getOrCreateJumpTableInfo(MachineJumpTableInfo::EK_Inline);
7033 unsigned MJTI = JTI->createJumpTableIndex(LPadList);
7034 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
7036 // Create the MBBs for the dispatch code.
7038 // Shove the dispatch's address into the return slot in the function context.
7039 MachineBasicBlock *DispatchBB = MF->CreateMachineBasicBlock();
7040 DispatchBB->setIsLandingPad();
7042 MachineBasicBlock *TrapBB = MF->CreateMachineBasicBlock();
7043 unsigned trap_opcode;
7044 if (Subtarget->isThumb())
7045 trap_opcode = ARM::tTRAP;
7047 trap_opcode = Subtarget->useNaClTrap() ? ARM::TRAPNaCl : ARM::TRAP;
7049 BuildMI(TrapBB, dl, TII->get(trap_opcode));
7050 DispatchBB->addSuccessor(TrapBB);
7052 MachineBasicBlock *DispContBB = MF->CreateMachineBasicBlock();
7053 DispatchBB->addSuccessor(DispContBB);
7056 MF->insert(MF->end(), DispatchBB);
7057 MF->insert(MF->end(), DispContBB);
7058 MF->insert(MF->end(), TrapBB);
7060 // Insert code into the entry block that creates and registers the function
7062 SetupEntryBlockForSjLj(MI, MBB, DispatchBB, FI);
7064 MachineMemOperand *FIMMOLd =
7065 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
7066 MachineMemOperand::MOLoad |
7067 MachineMemOperand::MOVolatile, 4, 4);
7069 MachineInstrBuilder MIB;
7070 MIB = BuildMI(DispatchBB, dl, TII->get(ARM::Int_eh_sjlj_dispatchsetup));
7072 const ARMBaseInstrInfo *AII = static_cast<const ARMBaseInstrInfo*>(TII);
7073 const ARMBaseRegisterInfo &RI = AII->getRegisterInfo();
7075 // Add a register mask with no preserved registers. This results in all
7076 // registers being marked as clobbered.
7077 MIB.addRegMask(RI.getNoPreservedMask());
7079 unsigned NumLPads = LPadList.size();
7080 if (Subtarget->isThumb2()) {
7081 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
7082 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2LDRi12), NewVReg1)
7085 .addMemOperand(FIMMOLd));
7087 if (NumLPads < 256) {
7088 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPri))
7090 .addImm(LPadList.size()));
7092 unsigned VReg1 = MRI->createVirtualRegister(TRC);
7093 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVi16), VReg1)
7094 .addImm(NumLPads & 0xFFFF));
7096 unsigned VReg2 = VReg1;
7097 if ((NumLPads & 0xFFFF0000) != 0) {
7098 VReg2 = MRI->createVirtualRegister(TRC);
7099 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVTi16), VReg2)
7101 .addImm(NumLPads >> 16));
7104 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPrr))
7109 BuildMI(DispatchBB, dl, TII->get(ARM::t2Bcc))
7114 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
7115 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::t2LEApcrelJT),NewVReg3)
7116 .addJumpTableIndex(MJTI));
7118 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
7121 BuildMI(DispContBB, dl, TII->get(ARM::t2ADDrs), NewVReg4)
7122 .addReg(NewVReg3, RegState::Kill)
7124 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
7126 BuildMI(DispContBB, dl, TII->get(ARM::t2BR_JT))
7127 .addReg(NewVReg4, RegState::Kill)
7129 .addJumpTableIndex(MJTI);
7130 } else if (Subtarget->isThumb()) {
7131 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
7132 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRspi), NewVReg1)
7135 .addMemOperand(FIMMOLd));
7137 if (NumLPads < 256) {
7138 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPi8))
7142 MachineConstantPool *ConstantPool = MF->getConstantPool();
7143 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
7144 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
7146 // MachineConstantPool wants an explicit alignment.
7147 unsigned Align = MF->getDataLayout().getPrefTypeAlignment(Int32Ty);
7149 Align = MF->getDataLayout().getTypeAllocSize(C->getType());
7150 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
7152 unsigned VReg1 = MRI->createVirtualRegister(TRC);
7153 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRpci))
7154 .addReg(VReg1, RegState::Define)
7155 .addConstantPoolIndex(Idx));
7156 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPr))
7161 BuildMI(DispatchBB, dl, TII->get(ARM::tBcc))
7166 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
7167 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLSLri), NewVReg2)
7168 .addReg(ARM::CPSR, RegState::Define)
7172 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
7173 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLEApcrelJT), NewVReg3)
7174 .addJumpTableIndex(MJTI));
7176 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
7177 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg4)
7178 .addReg(ARM::CPSR, RegState::Define)
7179 .addReg(NewVReg2, RegState::Kill)
7182 MachineMemOperand *JTMMOLd =
7183 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
7184 MachineMemOperand::MOLoad, 4, 4);
7186 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
7187 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLDRi), NewVReg5)
7188 .addReg(NewVReg4, RegState::Kill)
7190 .addMemOperand(JTMMOLd));
7192 unsigned NewVReg6 = NewVReg5;
7193 if (RelocM == Reloc::PIC_) {
7194 NewVReg6 = MRI->createVirtualRegister(TRC);
7195 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg6)
7196 .addReg(ARM::CPSR, RegState::Define)
7197 .addReg(NewVReg5, RegState::Kill)
7201 BuildMI(DispContBB, dl, TII->get(ARM::tBR_JTr))
7202 .addReg(NewVReg6, RegState::Kill)
7203 .addJumpTableIndex(MJTI);
7205 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
7206 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRi12), NewVReg1)
7209 .addMemOperand(FIMMOLd));
7211 if (NumLPads < 256) {
7212 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPri))
7215 } else if (Subtarget->hasV6T2Ops() && isUInt<16>(NumLPads)) {
7216 unsigned VReg1 = MRI->createVirtualRegister(TRC);
7217 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVi16), VReg1)
7218 .addImm(NumLPads & 0xFFFF));
7220 unsigned VReg2 = VReg1;
7221 if ((NumLPads & 0xFFFF0000) != 0) {
7222 VReg2 = MRI->createVirtualRegister(TRC);
7223 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVTi16), VReg2)
7225 .addImm(NumLPads >> 16));
7228 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
7232 MachineConstantPool *ConstantPool = MF->getConstantPool();
7233 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
7234 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
7236 // MachineConstantPool wants an explicit alignment.
7237 unsigned Align = MF->getDataLayout().getPrefTypeAlignment(Int32Ty);
7239 Align = MF->getDataLayout().getTypeAllocSize(C->getType());
7240 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
7242 unsigned VReg1 = MRI->createVirtualRegister(TRC);
7243 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRcp))
7244 .addReg(VReg1, RegState::Define)
7245 .addConstantPoolIndex(Idx)
7247 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
7249 .addReg(VReg1, RegState::Kill));
7252 BuildMI(DispatchBB, dl, TII->get(ARM::Bcc))
7257 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
7259 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::MOVsi), NewVReg3)
7261 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
7262 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
7263 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::LEApcrelJT), NewVReg4)
7264 .addJumpTableIndex(MJTI));
7266 MachineMemOperand *JTMMOLd =
7267 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
7268 MachineMemOperand::MOLoad, 4, 4);
7269 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
7271 BuildMI(DispContBB, dl, TII->get(ARM::LDRrs), NewVReg5)
7272 .addReg(NewVReg3, RegState::Kill)
7275 .addMemOperand(JTMMOLd));
7277 if (RelocM == Reloc::PIC_) {
7278 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTadd))
7279 .addReg(NewVReg5, RegState::Kill)
7281 .addJumpTableIndex(MJTI);
7283 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTr))
7284 .addReg(NewVReg5, RegState::Kill)
7285 .addJumpTableIndex(MJTI);
7289 // Add the jump table entries as successors to the MBB.
7290 SmallPtrSet<MachineBasicBlock*, 8> SeenMBBs;
7291 for (std::vector<MachineBasicBlock*>::iterator
7292 I = LPadList.begin(), E = LPadList.end(); I != E; ++I) {
7293 MachineBasicBlock *CurMBB = *I;
7294 if (SeenMBBs.insert(CurMBB).second)
7295 DispContBB->addSuccessor(CurMBB);
7298 // N.B. the order the invoke BBs are processed in doesn't matter here.
7299 const MCPhysReg *SavedRegs = RI.getCalleeSavedRegs(MF);
7300 SmallVector<MachineBasicBlock*, 64> MBBLPads;
7301 for (MachineBasicBlock *BB : InvokeBBs) {
7303 // Remove the landing pad successor from the invoke block and replace it
7304 // with the new dispatch block.
7305 SmallVector<MachineBasicBlock*, 4> Successors(BB->succ_begin(),
7307 while (!Successors.empty()) {
7308 MachineBasicBlock *SMBB = Successors.pop_back_val();
7309 if (SMBB->isLandingPad()) {
7310 BB->removeSuccessor(SMBB);
7311 MBBLPads.push_back(SMBB);
7315 BB->addSuccessor(DispatchBB);
7317 // Find the invoke call and mark all of the callee-saved registers as
7318 // 'implicit defined' so that they're spilled. This prevents code from
7319 // moving instructions to before the EH block, where they will never be
7321 for (MachineBasicBlock::reverse_iterator
7322 II = BB->rbegin(), IE = BB->rend(); II != IE; ++II) {
7323 if (!II->isCall()) continue;
7325 DenseMap<unsigned, bool> DefRegs;
7326 for (MachineInstr::mop_iterator
7327 OI = II->operands_begin(), OE = II->operands_end();
7329 if (!OI->isReg()) continue;
7330 DefRegs[OI->getReg()] = true;
7333 MachineInstrBuilder MIB(*MF, &*II);
7335 for (unsigned i = 0; SavedRegs[i] != 0; ++i) {
7336 unsigned Reg = SavedRegs[i];
7337 if (Subtarget->isThumb2() &&
7338 !ARM::tGPRRegClass.contains(Reg) &&
7339 !ARM::hGPRRegClass.contains(Reg))
7341 if (Subtarget->isThumb1Only() && !ARM::tGPRRegClass.contains(Reg))
7343 if (!Subtarget->isThumb() && !ARM::GPRRegClass.contains(Reg))
7346 MIB.addReg(Reg, RegState::ImplicitDefine | RegState::Dead);
7353 // Mark all former landing pads as non-landing pads. The dispatch is the only
7355 for (SmallVectorImpl<MachineBasicBlock*>::iterator
7356 I = MBBLPads.begin(), E = MBBLPads.end(); I != E; ++I)
7357 (*I)->setIsLandingPad(false);
7359 // The instruction is gone now.
7360 MI->eraseFromParent();
7364 MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) {
7365 for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
7366 E = MBB->succ_end(); I != E; ++I)
7369 llvm_unreachable("Expecting a BB with two successors!");
7372 /// Return the load opcode for a given load size. If load size >= 8,
7373 /// neon opcode will be returned.
7374 static unsigned getLdOpcode(unsigned LdSize, bool IsThumb1, bool IsThumb2) {
7376 return LdSize == 16 ? ARM::VLD1q32wb_fixed
7377 : LdSize == 8 ? ARM::VLD1d32wb_fixed : 0;
7379 return LdSize == 4 ? ARM::tLDRi
7380 : LdSize == 2 ? ARM::tLDRHi
7381 : LdSize == 1 ? ARM::tLDRBi : 0;
7383 return LdSize == 4 ? ARM::t2LDR_POST
7384 : LdSize == 2 ? ARM::t2LDRH_POST
7385 : LdSize == 1 ? ARM::t2LDRB_POST : 0;
7386 return LdSize == 4 ? ARM::LDR_POST_IMM
7387 : LdSize == 2 ? ARM::LDRH_POST
7388 : LdSize == 1 ? ARM::LDRB_POST_IMM : 0;
7391 /// Return the store opcode for a given store size. If store size >= 8,
7392 /// neon opcode will be returned.
7393 static unsigned getStOpcode(unsigned StSize, bool IsThumb1, bool IsThumb2) {
7395 return StSize == 16 ? ARM::VST1q32wb_fixed
7396 : StSize == 8 ? ARM::VST1d32wb_fixed : 0;
7398 return StSize == 4 ? ARM::tSTRi
7399 : StSize == 2 ? ARM::tSTRHi
7400 : StSize == 1 ? ARM::tSTRBi : 0;
7402 return StSize == 4 ? ARM::t2STR_POST
7403 : StSize == 2 ? ARM::t2STRH_POST
7404 : StSize == 1 ? ARM::t2STRB_POST : 0;
7405 return StSize == 4 ? ARM::STR_POST_IMM
7406 : StSize == 2 ? ARM::STRH_POST
7407 : StSize == 1 ? ARM::STRB_POST_IMM : 0;
7410 /// Emit a post-increment load operation with given size. The instructions
7411 /// will be added to BB at Pos.
7412 static void emitPostLd(MachineBasicBlock *BB, MachineInstr *Pos,
7413 const TargetInstrInfo *TII, DebugLoc dl,
7414 unsigned LdSize, unsigned Data, unsigned AddrIn,
7415 unsigned AddrOut, bool IsThumb1, bool IsThumb2) {
7416 unsigned LdOpc = getLdOpcode(LdSize, IsThumb1, IsThumb2);
7417 assert(LdOpc != 0 && "Should have a load opcode");
7419 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
7420 .addReg(AddrOut, RegState::Define).addReg(AddrIn)
7422 } else if (IsThumb1) {
7423 // load + update AddrIn
7424 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
7425 .addReg(AddrIn).addImm(0));
7426 MachineInstrBuilder MIB =
7427 BuildMI(*BB, Pos, dl, TII->get(ARM::tADDi8), AddrOut);
7428 MIB = AddDefaultT1CC(MIB);
7429 MIB.addReg(AddrIn).addImm(LdSize);
7430 AddDefaultPred(MIB);
7431 } else if (IsThumb2) {
7432 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
7433 .addReg(AddrOut, RegState::Define).addReg(AddrIn)
7436 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
7437 .addReg(AddrOut, RegState::Define).addReg(AddrIn)
7438 .addReg(0).addImm(LdSize));
7442 /// Emit a post-increment store operation with given size. The instructions
7443 /// will be added to BB at Pos.
7444 static void emitPostSt(MachineBasicBlock *BB, MachineInstr *Pos,
7445 const TargetInstrInfo *TII, DebugLoc dl,
7446 unsigned StSize, unsigned Data, unsigned AddrIn,
7447 unsigned AddrOut, bool IsThumb1, bool IsThumb2) {
7448 unsigned StOpc = getStOpcode(StSize, IsThumb1, IsThumb2);
7449 assert(StOpc != 0 && "Should have a store opcode");
7451 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc), AddrOut)
7452 .addReg(AddrIn).addImm(0).addReg(Data));
7453 } else if (IsThumb1) {
7454 // store + update AddrIn
7455 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc)).addReg(Data)
7456 .addReg(AddrIn).addImm(0));
7457 MachineInstrBuilder MIB =
7458 BuildMI(*BB, Pos, dl, TII->get(ARM::tADDi8), AddrOut);
7459 MIB = AddDefaultT1CC(MIB);
7460 MIB.addReg(AddrIn).addImm(StSize);
7461 AddDefaultPred(MIB);
7462 } else if (IsThumb2) {
7463 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc), AddrOut)
7464 .addReg(Data).addReg(AddrIn).addImm(StSize));
7466 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc), AddrOut)
7467 .addReg(Data).addReg(AddrIn).addReg(0)
7473 ARMTargetLowering::EmitStructByval(MachineInstr *MI,
7474 MachineBasicBlock *BB) const {
7475 // This pseudo instruction has 3 operands: dst, src, size
7476 // We expand it to a loop if size > Subtarget->getMaxInlineSizeThreshold().
7477 // Otherwise, we will generate unrolled scalar copies.
7478 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
7479 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7480 MachineFunction::iterator It = BB;
7483 unsigned dest = MI->getOperand(0).getReg();
7484 unsigned src = MI->getOperand(1).getReg();
7485 unsigned SizeVal = MI->getOperand(2).getImm();
7486 unsigned Align = MI->getOperand(3).getImm();
7487 DebugLoc dl = MI->getDebugLoc();
7489 MachineFunction *MF = BB->getParent();
7490 MachineRegisterInfo &MRI = MF->getRegInfo();
7491 unsigned UnitSize = 0;
7492 const TargetRegisterClass *TRC = nullptr;
7493 const TargetRegisterClass *VecTRC = nullptr;
7495 bool IsThumb1 = Subtarget->isThumb1Only();
7496 bool IsThumb2 = Subtarget->isThumb2();
7500 } else if (Align & 2) {
7503 // Check whether we can use NEON instructions.
7504 if (!MF->getFunction()->hasFnAttribute(Attribute::NoImplicitFloat) &&
7505 Subtarget->hasNEON()) {
7506 if ((Align % 16 == 0) && SizeVal >= 16)
7508 else if ((Align % 8 == 0) && SizeVal >= 8)
7511 // Can't use NEON instructions.
7516 // Select the correct opcode and register class for unit size load/store
7517 bool IsNeon = UnitSize >= 8;
7518 TRC = (IsThumb1 || IsThumb2) ? &ARM::tGPRRegClass : &ARM::GPRRegClass;
7520 VecTRC = UnitSize == 16 ? &ARM::DPairRegClass
7521 : UnitSize == 8 ? &ARM::DPRRegClass
7524 unsigned BytesLeft = SizeVal % UnitSize;
7525 unsigned LoopSize = SizeVal - BytesLeft;
7527 if (SizeVal <= Subtarget->getMaxInlineSizeThreshold()) {
7528 // Use LDR and STR to copy.
7529 // [scratch, srcOut] = LDR_POST(srcIn, UnitSize)
7530 // [destOut] = STR_POST(scratch, destIn, UnitSize)
7531 unsigned srcIn = src;
7532 unsigned destIn = dest;
7533 for (unsigned i = 0; i < LoopSize; i+=UnitSize) {
7534 unsigned srcOut = MRI.createVirtualRegister(TRC);
7535 unsigned destOut = MRI.createVirtualRegister(TRC);
7536 unsigned scratch = MRI.createVirtualRegister(IsNeon ? VecTRC : TRC);
7537 emitPostLd(BB, MI, TII, dl, UnitSize, scratch, srcIn, srcOut,
7538 IsThumb1, IsThumb2);
7539 emitPostSt(BB, MI, TII, dl, UnitSize, scratch, destIn, destOut,
7540 IsThumb1, IsThumb2);
7545 // Handle the leftover bytes with LDRB and STRB.
7546 // [scratch, srcOut] = LDRB_POST(srcIn, 1)
7547 // [destOut] = STRB_POST(scratch, destIn, 1)
7548 for (unsigned i = 0; i < BytesLeft; i++) {
7549 unsigned srcOut = MRI.createVirtualRegister(TRC);
7550 unsigned destOut = MRI.createVirtualRegister(TRC);
7551 unsigned scratch = MRI.createVirtualRegister(TRC);
7552 emitPostLd(BB, MI, TII, dl, 1, scratch, srcIn, srcOut,
7553 IsThumb1, IsThumb2);
7554 emitPostSt(BB, MI, TII, dl, 1, scratch, destIn, destOut,
7555 IsThumb1, IsThumb2);
7559 MI->eraseFromParent(); // The instruction is gone now.
7563 // Expand the pseudo op to a loop.
7566 // movw varEnd, # --> with thumb2
7568 // ldrcp varEnd, idx --> without thumb2
7569 // fallthrough --> loopMBB
7571 // PHI varPhi, varEnd, varLoop
7572 // PHI srcPhi, src, srcLoop
7573 // PHI destPhi, dst, destLoop
7574 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
7575 // [destLoop] = STR_POST(scratch, destPhi, UnitSize)
7576 // subs varLoop, varPhi, #UnitSize
7578 // fallthrough --> exitMBB
7580 // epilogue to handle left-over bytes
7581 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
7582 // [destOut] = STRB_POST(scratch, destLoop, 1)
7583 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
7584 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
7585 MF->insert(It, loopMBB);
7586 MF->insert(It, exitMBB);
7588 // Transfer the remainder of BB and its successor edges to exitMBB.
7589 exitMBB->splice(exitMBB->begin(), BB,
7590 std::next(MachineBasicBlock::iterator(MI)), BB->end());
7591 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
7593 // Load an immediate to varEnd.
7594 unsigned varEnd = MRI.createVirtualRegister(TRC);
7595 if (Subtarget->useMovt(*MF)) {
7596 unsigned Vtmp = varEnd;
7597 if ((LoopSize & 0xFFFF0000) != 0)
7598 Vtmp = MRI.createVirtualRegister(TRC);
7599 AddDefaultPred(BuildMI(BB, dl,
7600 TII->get(IsThumb2 ? ARM::t2MOVi16 : ARM::MOVi16),
7601 Vtmp).addImm(LoopSize & 0xFFFF));
7603 if ((LoopSize & 0xFFFF0000) != 0)
7604 AddDefaultPred(BuildMI(BB, dl,
7605 TII->get(IsThumb2 ? ARM::t2MOVTi16 : ARM::MOVTi16),
7608 .addImm(LoopSize >> 16));
7610 MachineConstantPool *ConstantPool = MF->getConstantPool();
7611 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
7612 const Constant *C = ConstantInt::get(Int32Ty, LoopSize);
7614 // MachineConstantPool wants an explicit alignment.
7615 unsigned Align = MF->getDataLayout().getPrefTypeAlignment(Int32Ty);
7617 Align = MF->getDataLayout().getTypeAllocSize(C->getType());
7618 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
7621 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(ARM::tLDRpci)).addReg(
7622 varEnd, RegState::Define).addConstantPoolIndex(Idx));
7624 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(ARM::LDRcp)).addReg(
7625 varEnd, RegState::Define).addConstantPoolIndex(Idx).addImm(0));
7627 BB->addSuccessor(loopMBB);
7629 // Generate the loop body:
7630 // varPhi = PHI(varLoop, varEnd)
7631 // srcPhi = PHI(srcLoop, src)
7632 // destPhi = PHI(destLoop, dst)
7633 MachineBasicBlock *entryBB = BB;
7635 unsigned varLoop = MRI.createVirtualRegister(TRC);
7636 unsigned varPhi = MRI.createVirtualRegister(TRC);
7637 unsigned srcLoop = MRI.createVirtualRegister(TRC);
7638 unsigned srcPhi = MRI.createVirtualRegister(TRC);
7639 unsigned destLoop = MRI.createVirtualRegister(TRC);
7640 unsigned destPhi = MRI.createVirtualRegister(TRC);
7642 BuildMI(*BB, BB->begin(), dl, TII->get(ARM::PHI), varPhi)
7643 .addReg(varLoop).addMBB(loopMBB)
7644 .addReg(varEnd).addMBB(entryBB);
7645 BuildMI(BB, dl, TII->get(ARM::PHI), srcPhi)
7646 .addReg(srcLoop).addMBB(loopMBB)
7647 .addReg(src).addMBB(entryBB);
7648 BuildMI(BB, dl, TII->get(ARM::PHI), destPhi)
7649 .addReg(destLoop).addMBB(loopMBB)
7650 .addReg(dest).addMBB(entryBB);
7652 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
7653 // [destLoop] = STR_POST(scratch, destPhi, UnitSiz)
7654 unsigned scratch = MRI.createVirtualRegister(IsNeon ? VecTRC : TRC);
7655 emitPostLd(BB, BB->end(), TII, dl, UnitSize, scratch, srcPhi, srcLoop,
7656 IsThumb1, IsThumb2);
7657 emitPostSt(BB, BB->end(), TII, dl, UnitSize, scratch, destPhi, destLoop,
7658 IsThumb1, IsThumb2);
7660 // Decrement loop variable by UnitSize.
7662 MachineInstrBuilder MIB =
7663 BuildMI(*BB, BB->end(), dl, TII->get(ARM::tSUBi8), varLoop);
7664 MIB = AddDefaultT1CC(MIB);
7665 MIB.addReg(varPhi).addImm(UnitSize);
7666 AddDefaultPred(MIB);
7668 MachineInstrBuilder MIB =
7669 BuildMI(*BB, BB->end(), dl,
7670 TII->get(IsThumb2 ? ARM::t2SUBri : ARM::SUBri), varLoop);
7671 AddDefaultCC(AddDefaultPred(MIB.addReg(varPhi).addImm(UnitSize)));
7672 MIB->getOperand(5).setReg(ARM::CPSR);
7673 MIB->getOperand(5).setIsDef(true);
7675 BuildMI(*BB, BB->end(), dl,
7676 TII->get(IsThumb1 ? ARM::tBcc : IsThumb2 ? ARM::t2Bcc : ARM::Bcc))
7677 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
7679 // loopMBB can loop back to loopMBB or fall through to exitMBB.
7680 BB->addSuccessor(loopMBB);
7681 BB->addSuccessor(exitMBB);
7683 // Add epilogue to handle BytesLeft.
7685 MachineInstr *StartOfExit = exitMBB->begin();
7687 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
7688 // [destOut] = STRB_POST(scratch, destLoop, 1)
7689 unsigned srcIn = srcLoop;
7690 unsigned destIn = destLoop;
7691 for (unsigned i = 0; i < BytesLeft; i++) {
7692 unsigned srcOut = MRI.createVirtualRegister(TRC);
7693 unsigned destOut = MRI.createVirtualRegister(TRC);
7694 unsigned scratch = MRI.createVirtualRegister(TRC);
7695 emitPostLd(BB, StartOfExit, TII, dl, 1, scratch, srcIn, srcOut,
7696 IsThumb1, IsThumb2);
7697 emitPostSt(BB, StartOfExit, TII, dl, 1, scratch, destIn, destOut,
7698 IsThumb1, IsThumb2);
7703 MI->eraseFromParent(); // The instruction is gone now.
7708 ARMTargetLowering::EmitLowered__chkstk(MachineInstr *MI,
7709 MachineBasicBlock *MBB) const {
7710 const TargetMachine &TM = getTargetMachine();
7711 const TargetInstrInfo &TII = *Subtarget->getInstrInfo();
7712 DebugLoc DL = MI->getDebugLoc();
7714 assert(Subtarget->isTargetWindows() &&
7715 "__chkstk is only supported on Windows");
7716 assert(Subtarget->isThumb2() && "Windows on ARM requires Thumb-2 mode");
7718 // __chkstk takes the number of words to allocate on the stack in R4, and
7719 // returns the stack adjustment in number of bytes in R4. This will not
7720 // clober any other registers (other than the obvious lr).
7722 // Although, technically, IP should be considered a register which may be
7723 // clobbered, the call itself will not touch it. Windows on ARM is a pure
7724 // thumb-2 environment, so there is no interworking required. As a result, we
7725 // do not expect a veneer to be emitted by the linker, clobbering IP.
7727 // Each module receives its own copy of __chkstk, so no import thunk is
7728 // required, again, ensuring that IP is not clobbered.
7730 // Finally, although some linkers may theoretically provide a trampoline for
7731 // out of range calls (which is quite common due to a 32M range limitation of
7732 // branches for Thumb), we can generate the long-call version via
7733 // -mcmodel=large, alleviating the need for the trampoline which may clobber
7736 switch (TM.getCodeModel()) {
7737 case CodeModel::Small:
7738 case CodeModel::Medium:
7739 case CodeModel::Default:
7740 case CodeModel::Kernel:
7741 BuildMI(*MBB, MI, DL, TII.get(ARM::tBL))
7742 .addImm((unsigned)ARMCC::AL).addReg(0)
7743 .addExternalSymbol("__chkstk")
7744 .addReg(ARM::R4, RegState::Implicit | RegState::Kill)
7745 .addReg(ARM::R4, RegState::Implicit | RegState::Define)
7746 .addReg(ARM::R12, RegState::Implicit | RegState::Define | RegState::Dead);
7748 case CodeModel::Large:
7749 case CodeModel::JITDefault: {
7750 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
7751 unsigned Reg = MRI.createVirtualRegister(&ARM::rGPRRegClass);
7753 BuildMI(*MBB, MI, DL, TII.get(ARM::t2MOVi32imm), Reg)
7754 .addExternalSymbol("__chkstk");
7755 BuildMI(*MBB, MI, DL, TII.get(ARM::tBLXr))
7756 .addImm((unsigned)ARMCC::AL).addReg(0)
7757 .addReg(Reg, RegState::Kill)
7758 .addReg(ARM::R4, RegState::Implicit | RegState::Kill)
7759 .addReg(ARM::R4, RegState::Implicit | RegState::Define)
7760 .addReg(ARM::R12, RegState::Implicit | RegState::Define | RegState::Dead);
7765 AddDefaultCC(AddDefaultPred(BuildMI(*MBB, MI, DL, TII.get(ARM::t2SUBrr),
7767 .addReg(ARM::SP).addReg(ARM::R4)));
7769 MI->eraseFromParent();
7774 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
7775 MachineBasicBlock *BB) const {
7776 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
7777 DebugLoc dl = MI->getDebugLoc();
7778 bool isThumb2 = Subtarget->isThumb2();
7779 switch (MI->getOpcode()) {
7782 llvm_unreachable("Unexpected instr type to insert");
7784 // The Thumb2 pre-indexed stores have the same MI operands, they just
7785 // define them differently in the .td files from the isel patterns, so
7786 // they need pseudos.
7787 case ARM::t2STR_preidx:
7788 MI->setDesc(TII->get(ARM::t2STR_PRE));
7790 case ARM::t2STRB_preidx:
7791 MI->setDesc(TII->get(ARM::t2STRB_PRE));
7793 case ARM::t2STRH_preidx:
7794 MI->setDesc(TII->get(ARM::t2STRH_PRE));
7797 case ARM::STRi_preidx:
7798 case ARM::STRBi_preidx: {
7799 unsigned NewOpc = MI->getOpcode() == ARM::STRi_preidx ?
7800 ARM::STR_PRE_IMM : ARM::STRB_PRE_IMM;
7801 // Decode the offset.
7802 unsigned Offset = MI->getOperand(4).getImm();
7803 bool isSub = ARM_AM::getAM2Op(Offset) == ARM_AM::sub;
7804 Offset = ARM_AM::getAM2Offset(Offset);
7808 MachineMemOperand *MMO = *MI->memoperands_begin();
7809 BuildMI(*BB, MI, dl, TII->get(NewOpc))
7810 .addOperand(MI->getOperand(0)) // Rn_wb
7811 .addOperand(MI->getOperand(1)) // Rt
7812 .addOperand(MI->getOperand(2)) // Rn
7813 .addImm(Offset) // offset (skip GPR==zero_reg)
7814 .addOperand(MI->getOperand(5)) // pred
7815 .addOperand(MI->getOperand(6))
7816 .addMemOperand(MMO);
7817 MI->eraseFromParent();
7820 case ARM::STRr_preidx:
7821 case ARM::STRBr_preidx:
7822 case ARM::STRH_preidx: {
7824 switch (MI->getOpcode()) {
7825 default: llvm_unreachable("unexpected opcode!");
7826 case ARM::STRr_preidx: NewOpc = ARM::STR_PRE_REG; break;
7827 case ARM::STRBr_preidx: NewOpc = ARM::STRB_PRE_REG; break;
7828 case ARM::STRH_preidx: NewOpc = ARM::STRH_PRE; break;
7830 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(NewOpc));
7831 for (unsigned i = 0; i < MI->getNumOperands(); ++i)
7832 MIB.addOperand(MI->getOperand(i));
7833 MI->eraseFromParent();
7837 case ARM::tMOVCCr_pseudo: {
7838 // To "insert" a SELECT_CC instruction, we actually have to insert the
7839 // diamond control-flow pattern. The incoming instruction knows the
7840 // destination vreg to set, the condition code register to branch on, the
7841 // true/false values to select between, and a branch opcode to use.
7842 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7843 MachineFunction::iterator It = BB;
7849 // cmpTY ccX, r1, r2
7851 // fallthrough --> copy0MBB
7852 MachineBasicBlock *thisMBB = BB;
7853 MachineFunction *F = BB->getParent();
7854 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
7855 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
7856 F->insert(It, copy0MBB);
7857 F->insert(It, sinkMBB);
7859 // Transfer the remainder of BB and its successor edges to sinkMBB.
7860 sinkMBB->splice(sinkMBB->begin(), BB,
7861 std::next(MachineBasicBlock::iterator(MI)), BB->end());
7862 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
7864 BB->addSuccessor(copy0MBB);
7865 BB->addSuccessor(sinkMBB);
7867 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
7868 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
7871 // %FalseValue = ...
7872 // # fallthrough to sinkMBB
7875 // Update machine-CFG edges
7876 BB->addSuccessor(sinkMBB);
7879 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
7882 BuildMI(*BB, BB->begin(), dl,
7883 TII->get(ARM::PHI), MI->getOperand(0).getReg())
7884 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
7885 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
7887 MI->eraseFromParent(); // The pseudo instruction is gone now.
7892 case ARM::BCCZi64: {
7893 // If there is an unconditional branch to the other successor, remove it.
7894 BB->erase(std::next(MachineBasicBlock::iterator(MI)), BB->end());
7896 // Compare both parts that make up the double comparison separately for
7898 bool RHSisZero = MI->getOpcode() == ARM::BCCZi64;
7900 unsigned LHS1 = MI->getOperand(1).getReg();
7901 unsigned LHS2 = MI->getOperand(2).getReg();
7903 AddDefaultPred(BuildMI(BB, dl,
7904 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7905 .addReg(LHS1).addImm(0));
7906 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7907 .addReg(LHS2).addImm(0)
7908 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
7910 unsigned RHS1 = MI->getOperand(3).getReg();
7911 unsigned RHS2 = MI->getOperand(4).getReg();
7912 AddDefaultPred(BuildMI(BB, dl,
7913 TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
7914 .addReg(LHS1).addReg(RHS1));
7915 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
7916 .addReg(LHS2).addReg(RHS2)
7917 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
7920 MachineBasicBlock *destMBB = MI->getOperand(RHSisZero ? 3 : 5).getMBB();
7921 MachineBasicBlock *exitMBB = OtherSucc(BB, destMBB);
7922 if (MI->getOperand(0).getImm() == ARMCC::NE)
7923 std::swap(destMBB, exitMBB);
7925 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
7926 .addMBB(destMBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
7928 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2B)).addMBB(exitMBB));
7930 BuildMI(BB, dl, TII->get(ARM::B)) .addMBB(exitMBB);
7932 MI->eraseFromParent(); // The pseudo instruction is gone now.
7936 case ARM::Int_eh_sjlj_setjmp:
7937 case ARM::Int_eh_sjlj_setjmp_nofp:
7938 case ARM::tInt_eh_sjlj_setjmp:
7939 case ARM::t2Int_eh_sjlj_setjmp:
7940 case ARM::t2Int_eh_sjlj_setjmp_nofp:
7943 case ARM::Int_eh_sjlj_setup_dispatch:
7944 EmitSjLjDispatchBlock(MI, BB);
7949 // To insert an ABS instruction, we have to insert the
7950 // diamond control-flow pattern. The incoming instruction knows the
7951 // source vreg to test against 0, the destination vreg to set,
7952 // the condition code register to branch on, the
7953 // true/false values to select between, and a branch opcode to use.
7958 // BCC (branch to SinkBB if V0 >= 0)
7959 // RSBBB: V3 = RSBri V2, 0 (compute ABS if V2 < 0)
7960 // SinkBB: V1 = PHI(V2, V3)
7961 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7962 MachineFunction::iterator BBI = BB;
7964 MachineFunction *Fn = BB->getParent();
7965 MachineBasicBlock *RSBBB = Fn->CreateMachineBasicBlock(LLVM_BB);
7966 MachineBasicBlock *SinkBB = Fn->CreateMachineBasicBlock(LLVM_BB);
7967 Fn->insert(BBI, RSBBB);
7968 Fn->insert(BBI, SinkBB);
7970 unsigned int ABSSrcReg = MI->getOperand(1).getReg();
7971 unsigned int ABSDstReg = MI->getOperand(0).getReg();
7972 bool ABSSrcKIll = MI->getOperand(1).isKill();
7973 bool isThumb2 = Subtarget->isThumb2();
7974 MachineRegisterInfo &MRI = Fn->getRegInfo();
7975 // In Thumb mode S must not be specified if source register is the SP or
7976 // PC and if destination register is the SP, so restrict register class
7977 unsigned NewRsbDstReg =
7978 MRI.createVirtualRegister(isThumb2 ? &ARM::rGPRRegClass : &ARM::GPRRegClass);
7980 // Transfer the remainder of BB and its successor edges to sinkMBB.
7981 SinkBB->splice(SinkBB->begin(), BB,
7982 std::next(MachineBasicBlock::iterator(MI)), BB->end());
7983 SinkBB->transferSuccessorsAndUpdatePHIs(BB);
7985 BB->addSuccessor(RSBBB);
7986 BB->addSuccessor(SinkBB);
7988 // fall through to SinkMBB
7989 RSBBB->addSuccessor(SinkBB);
7991 // insert a cmp at the end of BB
7992 AddDefaultPred(BuildMI(BB, dl,
7993 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7994 .addReg(ABSSrcReg).addImm(0));
7996 // insert a bcc with opposite CC to ARMCC::MI at the end of BB
7998 TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)).addMBB(SinkBB)
7999 .addImm(ARMCC::getOppositeCondition(ARMCC::MI)).addReg(ARM::CPSR);
8001 // insert rsbri in RSBBB
8002 // Note: BCC and rsbri will be converted into predicated rsbmi
8003 // by if-conversion pass
8004 BuildMI(*RSBBB, RSBBB->begin(), dl,
8005 TII->get(isThumb2 ? ARM::t2RSBri : ARM::RSBri), NewRsbDstReg)
8006 .addReg(ABSSrcReg, ABSSrcKIll ? RegState::Kill : 0)
8007 .addImm(0).addImm((unsigned)ARMCC::AL).addReg(0).addReg(0);
8009 // insert PHI in SinkBB,
8010 // reuse ABSDstReg to not change uses of ABS instruction
8011 BuildMI(*SinkBB, SinkBB->begin(), dl,
8012 TII->get(ARM::PHI), ABSDstReg)
8013 .addReg(NewRsbDstReg).addMBB(RSBBB)
8014 .addReg(ABSSrcReg).addMBB(BB);
8016 // remove ABS instruction
8017 MI->eraseFromParent();
8019 // return last added BB
8022 case ARM::COPY_STRUCT_BYVAL_I32:
8024 return EmitStructByval(MI, BB);
8025 case ARM::WIN__CHKSTK:
8026 return EmitLowered__chkstk(MI, BB);
8030 void ARMTargetLowering::AdjustInstrPostInstrSelection(MachineInstr *MI,
8031 SDNode *Node) const {
8032 const MCInstrDesc *MCID = &MI->getDesc();
8033 // Adjust potentially 's' setting instructions after isel, i.e. ADC, SBC, RSB,
8034 // RSC. Coming out of isel, they have an implicit CPSR def, but the optional
8035 // operand is still set to noreg. If needed, set the optional operand's
8036 // register to CPSR, and remove the redundant implicit def.
8038 // e.g. ADCS (..., CPSR<imp-def>) -> ADC (... opt:CPSR<def>).
8040 // Rename pseudo opcodes.
8041 unsigned NewOpc = convertAddSubFlagsOpcode(MI->getOpcode());
8043 const ARMBaseInstrInfo *TII = Subtarget->getInstrInfo();
8044 MCID = &TII->get(NewOpc);
8046 assert(MCID->getNumOperands() == MI->getDesc().getNumOperands() + 1 &&
8047 "converted opcode should be the same except for cc_out");
8051 // Add the optional cc_out operand
8052 MI->addOperand(MachineOperand::CreateReg(0, /*isDef=*/true));
8054 unsigned ccOutIdx = MCID->getNumOperands() - 1;
8056 // Any ARM instruction that sets the 's' bit should specify an optional
8057 // "cc_out" operand in the last operand position.
8058 if (!MI->hasOptionalDef() || !MCID->OpInfo[ccOutIdx].isOptionalDef()) {
8059 assert(!NewOpc && "Optional cc_out operand required");
8062 // Look for an implicit def of CPSR added by MachineInstr ctor. Remove it
8063 // since we already have an optional CPSR def.
8064 bool definesCPSR = false;
8065 bool deadCPSR = false;
8066 for (unsigned i = MCID->getNumOperands(), e = MI->getNumOperands();
8068 const MachineOperand &MO = MI->getOperand(i);
8069 if (MO.isReg() && MO.isDef() && MO.getReg() == ARM::CPSR) {
8073 MI->RemoveOperand(i);
8078 assert(!NewOpc && "Optional cc_out operand required");
8081 assert(deadCPSR == !Node->hasAnyUseOfValue(1) && "inconsistent dead flag");
8083 assert(!MI->getOperand(ccOutIdx).getReg() &&
8084 "expect uninitialized optional cc_out operand");
8088 // If this instruction was defined with an optional CPSR def and its dag node
8089 // had a live implicit CPSR def, then activate the optional CPSR def.
8090 MachineOperand &MO = MI->getOperand(ccOutIdx);
8091 MO.setReg(ARM::CPSR);
8095 //===----------------------------------------------------------------------===//
8096 // ARM Optimization Hooks
8097 //===----------------------------------------------------------------------===//
8099 // Helper function that checks if N is a null or all ones constant.
8100 static inline bool isZeroOrAllOnes(SDValue N, bool AllOnes) {
8101 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N);
8104 return AllOnes ? C->isAllOnesValue() : C->isNullValue();
8107 // Return true if N is conditionally 0 or all ones.
8108 // Detects these expressions where cc is an i1 value:
8110 // (select cc 0, y) [AllOnes=0]
8111 // (select cc y, 0) [AllOnes=0]
8112 // (zext cc) [AllOnes=0]
8113 // (sext cc) [AllOnes=0/1]
8114 // (select cc -1, y) [AllOnes=1]
8115 // (select cc y, -1) [AllOnes=1]
8117 // Invert is set when N is the null/all ones constant when CC is false.
8118 // OtherOp is set to the alternative value of N.
8119 static bool isConditionalZeroOrAllOnes(SDNode *N, bool AllOnes,
8120 SDValue &CC, bool &Invert,
8122 SelectionDAG &DAG) {
8123 switch (N->getOpcode()) {
8124 default: return false;
8126 CC = N->getOperand(0);
8127 SDValue N1 = N->getOperand(1);
8128 SDValue N2 = N->getOperand(2);
8129 if (isZeroOrAllOnes(N1, AllOnes)) {
8134 if (isZeroOrAllOnes(N2, AllOnes)) {
8141 case ISD::ZERO_EXTEND:
8142 // (zext cc) can never be the all ones value.
8146 case ISD::SIGN_EXTEND: {
8148 EVT VT = N->getValueType(0);
8149 CC = N->getOperand(0);
8150 if (CC.getValueType() != MVT::i1)
8154 // When looking for an AllOnes constant, N is an sext, and the 'other'
8156 OtherOp = DAG.getConstant(0, dl, VT);
8157 else if (N->getOpcode() == ISD::ZERO_EXTEND)
8158 // When looking for a 0 constant, N can be zext or sext.
8159 OtherOp = DAG.getConstant(1, dl, VT);
8161 OtherOp = DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), dl,
8168 // Combine a constant select operand into its use:
8170 // (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
8171 // (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
8172 // (and (select cc, -1, c), x) -> (select cc, x, (and, x, c)) [AllOnes=1]
8173 // (or (select cc, 0, c), x) -> (select cc, x, (or, x, c))
8174 // (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c))
8176 // The transform is rejected if the select doesn't have a constant operand that
8177 // is null, or all ones when AllOnes is set.
8179 // Also recognize sext/zext from i1:
8181 // (add (zext cc), x) -> (select cc (add x, 1), x)
8182 // (add (sext cc), x) -> (select cc (add x, -1), x)
8184 // These transformations eventually create predicated instructions.
8186 // @param N The node to transform.
8187 // @param Slct The N operand that is a select.
8188 // @param OtherOp The other N operand (x above).
8189 // @param DCI Context.
8190 // @param AllOnes Require the select constant to be all ones instead of null.
8191 // @returns The new node, or SDValue() on failure.
8193 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
8194 TargetLowering::DAGCombinerInfo &DCI,
8195 bool AllOnes = false) {
8196 SelectionDAG &DAG = DCI.DAG;
8197 EVT VT = N->getValueType(0);
8198 SDValue NonConstantVal;
8201 if (!isConditionalZeroOrAllOnes(Slct.getNode(), AllOnes, CCOp, SwapSelectOps,
8202 NonConstantVal, DAG))
8205 // Slct is now know to be the desired identity constant when CC is true.
8206 SDValue TrueVal = OtherOp;
8207 SDValue FalseVal = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
8208 OtherOp, NonConstantVal);
8209 // Unless SwapSelectOps says CC should be false.
8211 std::swap(TrueVal, FalseVal);
8213 return DAG.getNode(ISD::SELECT, SDLoc(N), VT,
8214 CCOp, TrueVal, FalseVal);
8217 // Attempt combineSelectAndUse on each operand of a commutative operator N.
8219 SDValue combineSelectAndUseCommutative(SDNode *N, bool AllOnes,
8220 TargetLowering::DAGCombinerInfo &DCI) {
8221 SDValue N0 = N->getOperand(0);
8222 SDValue N1 = N->getOperand(1);
8223 if (N0.getNode()->hasOneUse()) {
8224 SDValue Result = combineSelectAndUse(N, N0, N1, DCI, AllOnes);
8225 if (Result.getNode())
8228 if (N1.getNode()->hasOneUse()) {
8229 SDValue Result = combineSelectAndUse(N, N1, N0, DCI, AllOnes);
8230 if (Result.getNode())
8236 // AddCombineToVPADDL- For pair-wise add on neon, use the vpaddl instruction
8237 // (only after legalization).
8238 static SDValue AddCombineToVPADDL(SDNode *N, SDValue N0, SDValue N1,
8239 TargetLowering::DAGCombinerInfo &DCI,
8240 const ARMSubtarget *Subtarget) {
8242 // Only perform optimization if after legalize, and if NEON is available. We
8243 // also expected both operands to be BUILD_VECTORs.
8244 if (DCI.isBeforeLegalize() || !Subtarget->hasNEON()
8245 || N0.getOpcode() != ISD::BUILD_VECTOR
8246 || N1.getOpcode() != ISD::BUILD_VECTOR)
8249 // Check output type since VPADDL operand elements can only be 8, 16, or 32.
8250 EVT VT = N->getValueType(0);
8251 if (!VT.isInteger() || VT.getVectorElementType() == MVT::i64)
8254 // Check that the vector operands are of the right form.
8255 // N0 and N1 are BUILD_VECTOR nodes with N number of EXTRACT_VECTOR
8256 // operands, where N is the size of the formed vector.
8257 // Each EXTRACT_VECTOR should have the same input vector and odd or even
8258 // index such that we have a pair wise add pattern.
8260 // Grab the vector that all EXTRACT_VECTOR nodes should be referencing.
8261 if (N0->getOperand(0)->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
8263 SDValue Vec = N0->getOperand(0)->getOperand(0);
8264 SDNode *V = Vec.getNode();
8265 unsigned nextIndex = 0;
8267 // For each operands to the ADD which are BUILD_VECTORs,
8268 // check to see if each of their operands are an EXTRACT_VECTOR with
8269 // the same vector and appropriate index.
8270 for (unsigned i = 0, e = N0->getNumOperands(); i != e; ++i) {
8271 if (N0->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT
8272 && N1->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
8274 SDValue ExtVec0 = N0->getOperand(i);
8275 SDValue ExtVec1 = N1->getOperand(i);
8277 // First operand is the vector, verify its the same.
8278 if (V != ExtVec0->getOperand(0).getNode() ||
8279 V != ExtVec1->getOperand(0).getNode())
8282 // Second is the constant, verify its correct.
8283 ConstantSDNode *C0 = dyn_cast<ConstantSDNode>(ExtVec0->getOperand(1));
8284 ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(ExtVec1->getOperand(1));
8286 // For the constant, we want to see all the even or all the odd.
8287 if (!C0 || !C1 || C0->getZExtValue() != nextIndex
8288 || C1->getZExtValue() != nextIndex+1)
8297 // Create VPADDL node.
8298 SelectionDAG &DAG = DCI.DAG;
8299 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8303 // Build operand list.
8304 SmallVector<SDValue, 8> Ops;
8305 Ops.push_back(DAG.getConstant(Intrinsic::arm_neon_vpaddls, dl,
8306 TLI.getPointerTy(DAG.getDataLayout())));
8308 // Input is the vector.
8311 // Get widened type and narrowed type.
8313 unsigned numElem = VT.getVectorNumElements();
8315 EVT inputLaneType = Vec.getValueType().getVectorElementType();
8316 switch (inputLaneType.getSimpleVT().SimpleTy) {
8317 case MVT::i8: widenType = MVT::getVectorVT(MVT::i16, numElem); break;
8318 case MVT::i16: widenType = MVT::getVectorVT(MVT::i32, numElem); break;
8319 case MVT::i32: widenType = MVT::getVectorVT(MVT::i64, numElem); break;
8321 llvm_unreachable("Invalid vector element type for padd optimization.");
8324 SDValue tmp = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, widenType, Ops);
8325 unsigned ExtOp = VT.bitsGT(tmp.getValueType()) ? ISD::ANY_EXTEND : ISD::TRUNCATE;
8326 return DAG.getNode(ExtOp, dl, VT, tmp);
8329 static SDValue findMUL_LOHI(SDValue V) {
8330 if (V->getOpcode() == ISD::UMUL_LOHI ||
8331 V->getOpcode() == ISD::SMUL_LOHI)
8336 static SDValue AddCombineTo64bitMLAL(SDNode *AddcNode,
8337 TargetLowering::DAGCombinerInfo &DCI,
8338 const ARMSubtarget *Subtarget) {
8340 if (Subtarget->isThumb1Only()) return SDValue();
8342 // Only perform the checks after legalize when the pattern is available.
8343 if (DCI.isBeforeLegalize()) return SDValue();
8345 // Look for multiply add opportunities.
8346 // The pattern is a ISD::UMUL_LOHI followed by two add nodes, where
8347 // each add nodes consumes a value from ISD::UMUL_LOHI and there is
8348 // a glue link from the first add to the second add.
8349 // If we find this pattern, we can replace the U/SMUL_LOHI, ADDC, and ADDE by
8350 // a S/UMLAL instruction.
8353 // / \ [no multiline comment]
8359 assert(AddcNode->getOpcode() == ISD::ADDC && "Expect an ADDC");
8360 SDValue AddcOp0 = AddcNode->getOperand(0);
8361 SDValue AddcOp1 = AddcNode->getOperand(1);
8363 // Check if the two operands are from the same mul_lohi node.
8364 if (AddcOp0.getNode() == AddcOp1.getNode())
8367 assert(AddcNode->getNumValues() == 2 &&
8368 AddcNode->getValueType(0) == MVT::i32 &&
8369 "Expect ADDC with two result values. First: i32");
8371 // Check that we have a glued ADDC node.
8372 if (AddcNode->getValueType(1) != MVT::Glue)
8375 // Check that the ADDC adds the low result of the S/UMUL_LOHI.
8376 if (AddcOp0->getOpcode() != ISD::UMUL_LOHI &&
8377 AddcOp0->getOpcode() != ISD::SMUL_LOHI &&
8378 AddcOp1->getOpcode() != ISD::UMUL_LOHI &&
8379 AddcOp1->getOpcode() != ISD::SMUL_LOHI)
8382 // Look for the glued ADDE.
8383 SDNode* AddeNode = AddcNode->getGluedUser();
8387 // Make sure it is really an ADDE.
8388 if (AddeNode->getOpcode() != ISD::ADDE)
8391 assert(AddeNode->getNumOperands() == 3 &&
8392 AddeNode->getOperand(2).getValueType() == MVT::Glue &&
8393 "ADDE node has the wrong inputs");
8395 // Check for the triangle shape.
8396 SDValue AddeOp0 = AddeNode->getOperand(0);
8397 SDValue AddeOp1 = AddeNode->getOperand(1);
8399 // Make sure that the ADDE operands are not coming from the same node.
8400 if (AddeOp0.getNode() == AddeOp1.getNode())
8403 // Find the MUL_LOHI node walking up ADDE's operands.
8404 bool IsLeftOperandMUL = false;
8405 SDValue MULOp = findMUL_LOHI(AddeOp0);
8406 if (MULOp == SDValue())
8407 MULOp = findMUL_LOHI(AddeOp1);
8409 IsLeftOperandMUL = true;
8410 if (MULOp == SDValue())
8413 // Figure out the right opcode.
8414 unsigned Opc = MULOp->getOpcode();
8415 unsigned FinalOpc = (Opc == ISD::SMUL_LOHI) ? ARMISD::SMLAL : ARMISD::UMLAL;
8417 // Figure out the high and low input values to the MLAL node.
8418 SDValue* HiAdd = nullptr;
8419 SDValue* LoMul = nullptr;
8420 SDValue* LowAdd = nullptr;
8422 // Ensure that ADDE is from high result of ISD::SMUL_LOHI.
8423 if ((AddeOp0 != MULOp.getValue(1)) && (AddeOp1 != MULOp.getValue(1)))
8426 if (IsLeftOperandMUL)
8432 // Ensure that LoMul and LowAdd are taken from correct ISD::SMUL_LOHI node
8433 // whose low result is fed to the ADDC we are checking.
8435 if (AddcOp0 == MULOp.getValue(0)) {
8439 if (AddcOp1 == MULOp.getValue(0)) {
8447 // Create the merged node.
8448 SelectionDAG &DAG = DCI.DAG;
8450 // Build operand list.
8451 SmallVector<SDValue, 8> Ops;
8452 Ops.push_back(LoMul->getOperand(0));
8453 Ops.push_back(LoMul->getOperand(1));
8454 Ops.push_back(*LowAdd);
8455 Ops.push_back(*HiAdd);
8457 SDValue MLALNode = DAG.getNode(FinalOpc, SDLoc(AddcNode),
8458 DAG.getVTList(MVT::i32, MVT::i32), Ops);
8460 // Replace the ADDs' nodes uses by the MLA node's values.
8461 SDValue HiMLALResult(MLALNode.getNode(), 1);
8462 DAG.ReplaceAllUsesOfValueWith(SDValue(AddeNode, 0), HiMLALResult);
8464 SDValue LoMLALResult(MLALNode.getNode(), 0);
8465 DAG.ReplaceAllUsesOfValueWith(SDValue(AddcNode, 0), LoMLALResult);
8467 // Return original node to notify the driver to stop replacing.
8468 SDValue resNode(AddcNode, 0);
8472 /// PerformADDCCombine - Target-specific dag combine transform from
8473 /// ISD::ADDC, ISD::ADDE, and ISD::MUL_LOHI to MLAL.
8474 static SDValue PerformADDCCombine(SDNode *N,
8475 TargetLowering::DAGCombinerInfo &DCI,
8476 const ARMSubtarget *Subtarget) {
8478 return AddCombineTo64bitMLAL(N, DCI, Subtarget);
8482 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
8483 /// operands N0 and N1. This is a helper for PerformADDCombine that is
8484 /// called with the default operands, and if that fails, with commuted
8486 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
8487 TargetLowering::DAGCombinerInfo &DCI,
8488 const ARMSubtarget *Subtarget){
8490 // Attempt to create vpaddl for this add.
8491 SDValue Result = AddCombineToVPADDL(N, N0, N1, DCI, Subtarget);
8492 if (Result.getNode())
8495 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
8496 if (N0.getNode()->hasOneUse()) {
8497 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
8498 if (Result.getNode()) return Result;
8503 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
8505 static SDValue PerformADDCombine(SDNode *N,
8506 TargetLowering::DAGCombinerInfo &DCI,
8507 const ARMSubtarget *Subtarget) {
8508 SDValue N0 = N->getOperand(0);
8509 SDValue N1 = N->getOperand(1);
8511 // First try with the default operand order.
8512 SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI, Subtarget);
8513 if (Result.getNode())
8516 // If that didn't work, try again with the operands commuted.
8517 return PerformADDCombineWithOperands(N, N1, N0, DCI, Subtarget);
8520 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
8522 static SDValue PerformSUBCombine(SDNode *N,
8523 TargetLowering::DAGCombinerInfo &DCI) {
8524 SDValue N0 = N->getOperand(0);
8525 SDValue N1 = N->getOperand(1);
8527 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
8528 if (N1.getNode()->hasOneUse()) {
8529 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
8530 if (Result.getNode()) return Result;
8536 /// PerformVMULCombine
8537 /// Distribute (A + B) * C to (A * C) + (B * C) to take advantage of the
8538 /// special multiplier accumulator forwarding.
8544 // However, for (A + B) * (A + B),
8551 static SDValue PerformVMULCombine(SDNode *N,
8552 TargetLowering::DAGCombinerInfo &DCI,
8553 const ARMSubtarget *Subtarget) {
8554 if (!Subtarget->hasVMLxForwarding())
8557 SelectionDAG &DAG = DCI.DAG;
8558 SDValue N0 = N->getOperand(0);
8559 SDValue N1 = N->getOperand(1);
8560 unsigned Opcode = N0.getOpcode();
8561 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
8562 Opcode != ISD::FADD && Opcode != ISD::FSUB) {
8563 Opcode = N1.getOpcode();
8564 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
8565 Opcode != ISD::FADD && Opcode != ISD::FSUB)
8573 EVT VT = N->getValueType(0);
8575 SDValue N00 = N0->getOperand(0);
8576 SDValue N01 = N0->getOperand(1);
8577 return DAG.getNode(Opcode, DL, VT,
8578 DAG.getNode(ISD::MUL, DL, VT, N00, N1),
8579 DAG.getNode(ISD::MUL, DL, VT, N01, N1));
8582 static SDValue PerformMULCombine(SDNode *N,
8583 TargetLowering::DAGCombinerInfo &DCI,
8584 const ARMSubtarget *Subtarget) {
8585 SelectionDAG &DAG = DCI.DAG;
8587 if (Subtarget->isThumb1Only())
8590 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
8593 EVT VT = N->getValueType(0);
8594 if (VT.is64BitVector() || VT.is128BitVector())
8595 return PerformVMULCombine(N, DCI, Subtarget);
8599 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
8603 int64_t MulAmt = C->getSExtValue();
8604 unsigned ShiftAmt = countTrailingZeros<uint64_t>(MulAmt);
8606 ShiftAmt = ShiftAmt & (32 - 1);
8607 SDValue V = N->getOperand(0);
8611 MulAmt >>= ShiftAmt;
8614 if (isPowerOf2_32(MulAmt - 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(MulAmt - 1), DL,
8622 } else if (isPowerOf2_32(MulAmt + 1)) {
8623 // (mul x, 2^N - 1) => (sub (shl x, N), x)
8624 Res = DAG.getNode(ISD::SUB, DL, VT,
8625 DAG.getNode(ISD::SHL, DL, VT,
8627 DAG.getConstant(Log2_32(MulAmt + 1), DL,
8633 uint64_t MulAmtAbs = -MulAmt;
8634 if (isPowerOf2_32(MulAmtAbs + 1)) {
8635 // (mul x, -(2^N - 1)) => (sub x, (shl x, N))
8636 Res = DAG.getNode(ISD::SUB, DL, VT,
8638 DAG.getNode(ISD::SHL, DL, VT,
8640 DAG.getConstant(Log2_32(MulAmtAbs + 1), DL,
8642 } else if (isPowerOf2_32(MulAmtAbs - 1)) {
8643 // (mul x, -(2^N + 1)) => - (add (shl x, N), x)
8644 Res = DAG.getNode(ISD::ADD, DL, VT,
8646 DAG.getNode(ISD::SHL, DL, VT,
8648 DAG.getConstant(Log2_32(MulAmtAbs - 1), DL,
8650 Res = DAG.getNode(ISD::SUB, DL, VT,
8651 DAG.getConstant(0, DL, MVT::i32), Res);
8658 Res = DAG.getNode(ISD::SHL, DL, VT,
8659 Res, DAG.getConstant(ShiftAmt, DL, MVT::i32));
8661 // Do not add new nodes to DAG combiner worklist.
8662 DCI.CombineTo(N, Res, false);
8666 static SDValue PerformANDCombine(SDNode *N,
8667 TargetLowering::DAGCombinerInfo &DCI,
8668 const ARMSubtarget *Subtarget) {
8670 // Attempt to use immediate-form VBIC
8671 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
8673 EVT VT = N->getValueType(0);
8674 SelectionDAG &DAG = DCI.DAG;
8676 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8679 APInt SplatBits, SplatUndef;
8680 unsigned SplatBitSize;
8683 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
8684 if (SplatBitSize <= 64) {
8686 SDValue Val = isNEONModifiedImm((~SplatBits).getZExtValue(),
8687 SplatUndef.getZExtValue(), SplatBitSize,
8688 DAG, dl, VbicVT, VT.is128BitVector(),
8690 if (Val.getNode()) {
8692 DAG.getNode(ISD::BITCAST, dl, VbicVT, N->getOperand(0));
8693 SDValue Vbic = DAG.getNode(ARMISD::VBICIMM, dl, VbicVT, Input, Val);
8694 return DAG.getNode(ISD::BITCAST, dl, VT, Vbic);
8699 if (!Subtarget->isThumb1Only()) {
8700 // fold (and (select cc, -1, c), x) -> (select cc, x, (and, x, c))
8701 SDValue Result = combineSelectAndUseCommutative(N, true, DCI);
8702 if (Result.getNode())
8709 /// PerformORCombine - Target-specific dag combine xforms for ISD::OR
8710 static SDValue PerformORCombine(SDNode *N,
8711 TargetLowering::DAGCombinerInfo &DCI,
8712 const ARMSubtarget *Subtarget) {
8713 // Attempt to use immediate-form VORR
8714 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
8716 EVT VT = N->getValueType(0);
8717 SelectionDAG &DAG = DCI.DAG;
8719 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8722 APInt SplatBits, SplatUndef;
8723 unsigned SplatBitSize;
8725 if (BVN && Subtarget->hasNEON() &&
8726 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
8727 if (SplatBitSize <= 64) {
8729 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
8730 SplatUndef.getZExtValue(), SplatBitSize,
8731 DAG, dl, VorrVT, VT.is128BitVector(),
8733 if (Val.getNode()) {
8735 DAG.getNode(ISD::BITCAST, dl, VorrVT, N->getOperand(0));
8736 SDValue Vorr = DAG.getNode(ARMISD::VORRIMM, dl, VorrVT, Input, Val);
8737 return DAG.getNode(ISD::BITCAST, dl, VT, Vorr);
8742 if (!Subtarget->isThumb1Only()) {
8743 // fold (or (select cc, 0, c), x) -> (select cc, x, (or, x, c))
8744 SDValue Result = combineSelectAndUseCommutative(N, false, DCI);
8745 if (Result.getNode())
8749 // The code below optimizes (or (and X, Y), Z).
8750 // The AND operand needs to have a single user to make these optimizations
8752 SDValue N0 = N->getOperand(0);
8753 if (N0.getOpcode() != ISD::AND || !N0.hasOneUse())
8755 SDValue N1 = N->getOperand(1);
8757 // (or (and B, A), (and C, ~A)) => (VBSL A, B, C) when A is a constant.
8758 if (Subtarget->hasNEON() && N1.getOpcode() == ISD::AND && VT.isVector() &&
8759 DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
8761 unsigned SplatBitSize;
8764 APInt SplatBits0, SplatBits1;
8765 BuildVectorSDNode *BVN0 = dyn_cast<BuildVectorSDNode>(N0->getOperand(1));
8766 BuildVectorSDNode *BVN1 = dyn_cast<BuildVectorSDNode>(N1->getOperand(1));
8767 // Ensure that the second operand of both ands are constants
8768 if (BVN0 && BVN0->isConstantSplat(SplatBits0, SplatUndef, SplatBitSize,
8769 HasAnyUndefs) && !HasAnyUndefs) {
8770 if (BVN1 && BVN1->isConstantSplat(SplatBits1, SplatUndef, SplatBitSize,
8771 HasAnyUndefs) && !HasAnyUndefs) {
8772 // Ensure that the bit width of the constants are the same and that
8773 // the splat arguments are logical inverses as per the pattern we
8774 // are trying to simplify.
8775 if (SplatBits0.getBitWidth() == SplatBits1.getBitWidth() &&
8776 SplatBits0 == ~SplatBits1) {
8777 // Canonicalize the vector type to make instruction selection
8779 EVT CanonicalVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
8780 SDValue Result = DAG.getNode(ARMISD::VBSL, dl, CanonicalVT,
8784 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
8790 // Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when
8793 // BFI is only available on V6T2+
8794 if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops())
8798 // 1) or (and A, mask), val => ARMbfi A, val, mask
8799 // iff (val & mask) == val
8801 // 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
8802 // 2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2)
8803 // && mask == ~mask2
8804 // 2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2)
8805 // && ~mask == mask2
8806 // (i.e., copy a bitfield value into another bitfield of the same width)
8811 SDValue N00 = N0.getOperand(0);
8813 // The value and the mask need to be constants so we can verify this is
8814 // actually a bitfield set. If the mask is 0xffff, we can do better
8815 // via a movt instruction, so don't use BFI in that case.
8816 SDValue MaskOp = N0.getOperand(1);
8817 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(MaskOp);
8820 unsigned Mask = MaskC->getZExtValue();
8824 // Case (1): or (and A, mask), val => ARMbfi A, val, mask
8825 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
8827 unsigned Val = N1C->getZExtValue();
8828 if ((Val & ~Mask) != Val)
8831 if (ARM::isBitFieldInvertedMask(Mask)) {
8832 Val >>= countTrailingZeros(~Mask);
8834 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00,
8835 DAG.getConstant(Val, DL, MVT::i32),
8836 DAG.getConstant(Mask, DL, MVT::i32));
8838 // Do not add new nodes to DAG combiner worklist.
8839 DCI.CombineTo(N, Res, false);
8842 } else if (N1.getOpcode() == ISD::AND) {
8843 // case (2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
8844 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
8847 unsigned Mask2 = N11C->getZExtValue();
8849 // Mask and ~Mask2 (or reverse) must be equivalent for the BFI pattern
8851 if (ARM::isBitFieldInvertedMask(Mask) &&
8853 // The pack halfword instruction works better for masks that fit it,
8854 // so use that when it's available.
8855 if (Subtarget->hasT2ExtractPack() &&
8856 (Mask == 0xffff || Mask == 0xffff0000))
8859 unsigned amt = countTrailingZeros(Mask2);
8860 Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0),
8861 DAG.getConstant(amt, DL, MVT::i32));
8862 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00, Res,
8863 DAG.getConstant(Mask, DL, MVT::i32));
8864 // Do not add new nodes to DAG combiner worklist.
8865 DCI.CombineTo(N, Res, false);
8867 } else if (ARM::isBitFieldInvertedMask(~Mask) &&
8869 // The pack halfword instruction works better for masks that fit it,
8870 // so use that when it's available.
8871 if (Subtarget->hasT2ExtractPack() &&
8872 (Mask2 == 0xffff || Mask2 == 0xffff0000))
8875 unsigned lsb = countTrailingZeros(Mask);
8876 Res = DAG.getNode(ISD::SRL, DL, VT, N00,
8877 DAG.getConstant(lsb, DL, MVT::i32));
8878 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res,
8879 DAG.getConstant(Mask2, DL, MVT::i32));
8880 // Do not add new nodes to DAG combiner worklist.
8881 DCI.CombineTo(N, Res, false);
8886 if (DAG.MaskedValueIsZero(N1, MaskC->getAPIntValue()) &&
8887 N00.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N00.getOperand(1)) &&
8888 ARM::isBitFieldInvertedMask(~Mask)) {
8889 // Case (3): or (and (shl A, #shamt), mask), B => ARMbfi B, A, ~mask
8890 // where lsb(mask) == #shamt and masked bits of B are known zero.
8891 SDValue ShAmt = N00.getOperand(1);
8892 unsigned ShAmtC = cast<ConstantSDNode>(ShAmt)->getZExtValue();
8893 unsigned LSB = countTrailingZeros(Mask);
8897 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1, N00.getOperand(0),
8898 DAG.getConstant(~Mask, DL, MVT::i32));
8900 // Do not add new nodes to DAG combiner worklist.
8901 DCI.CombineTo(N, Res, false);
8907 static SDValue PerformXORCombine(SDNode *N,
8908 TargetLowering::DAGCombinerInfo &DCI,
8909 const ARMSubtarget *Subtarget) {
8910 EVT VT = N->getValueType(0);
8911 SelectionDAG &DAG = DCI.DAG;
8913 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8916 if (!Subtarget->isThumb1Only()) {
8917 // fold (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c))
8918 SDValue Result = combineSelectAndUseCommutative(N, false, DCI);
8919 if (Result.getNode())
8926 /// PerformBFICombine - (bfi A, (and B, Mask1), Mask2) -> (bfi A, B, Mask2) iff
8927 /// the bits being cleared by the AND are not demanded by the BFI.
8928 static SDValue PerformBFICombine(SDNode *N,
8929 TargetLowering::DAGCombinerInfo &DCI) {
8930 SDValue N1 = N->getOperand(1);
8931 if (N1.getOpcode() == ISD::AND) {
8932 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
8935 unsigned InvMask = cast<ConstantSDNode>(N->getOperand(2))->getZExtValue();
8936 unsigned LSB = countTrailingZeros(~InvMask);
8937 unsigned Width = (32 - countLeadingZeros(~InvMask)) - LSB;
8939 static_cast<unsigned>(std::numeric_limits<unsigned>::digits) &&
8940 "undefined behavior");
8941 unsigned Mask = (1u << Width) - 1;
8942 unsigned Mask2 = N11C->getZExtValue();
8943 if ((Mask & (~Mask2)) == 0)
8944 return DCI.DAG.getNode(ARMISD::BFI, SDLoc(N), N->getValueType(0),
8945 N->getOperand(0), N1.getOperand(0),
8951 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
8952 /// ARMISD::VMOVRRD.
8953 static SDValue PerformVMOVRRDCombine(SDNode *N,
8954 TargetLowering::DAGCombinerInfo &DCI,
8955 const ARMSubtarget *Subtarget) {
8956 // vmovrrd(vmovdrr x, y) -> x,y
8957 SDValue InDouble = N->getOperand(0);
8958 if (InDouble.getOpcode() == ARMISD::VMOVDRR && !Subtarget->isFPOnlySP())
8959 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
8961 // vmovrrd(load f64) -> (load i32), (load i32)
8962 SDNode *InNode = InDouble.getNode();
8963 if (ISD::isNormalLoad(InNode) && InNode->hasOneUse() &&
8964 InNode->getValueType(0) == MVT::f64 &&
8965 InNode->getOperand(1).getOpcode() == ISD::FrameIndex &&
8966 !cast<LoadSDNode>(InNode)->isVolatile()) {
8967 // TODO: Should this be done for non-FrameIndex operands?
8968 LoadSDNode *LD = cast<LoadSDNode>(InNode);
8970 SelectionDAG &DAG = DCI.DAG;
8972 SDValue BasePtr = LD->getBasePtr();
8973 SDValue NewLD1 = DAG.getLoad(MVT::i32, DL, LD->getChain(), BasePtr,
8974 LD->getPointerInfo(), LD->isVolatile(),
8975 LD->isNonTemporal(), LD->isInvariant(),
8976 LD->getAlignment());
8978 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
8979 DAG.getConstant(4, DL, MVT::i32));
8980 SDValue NewLD2 = DAG.getLoad(MVT::i32, DL, NewLD1.getValue(1), OffsetPtr,
8981 LD->getPointerInfo(), LD->isVolatile(),
8982 LD->isNonTemporal(), LD->isInvariant(),
8983 std::min(4U, LD->getAlignment() / 2));
8985 DAG.ReplaceAllUsesOfValueWith(SDValue(LD, 1), NewLD2.getValue(1));
8986 if (DCI.DAG.getDataLayout().isBigEndian())
8987 std::swap (NewLD1, NewLD2);
8988 SDValue Result = DCI.CombineTo(N, NewLD1, NewLD2);
8995 /// PerformVMOVDRRCombine - Target-specific dag combine xforms for
8996 /// ARMISD::VMOVDRR. This is also used for BUILD_VECTORs with 2 operands.
8997 static SDValue PerformVMOVDRRCombine(SDNode *N, SelectionDAG &DAG) {
8998 // N=vmovrrd(X); vmovdrr(N:0, N:1) -> bit_convert(X)
8999 SDValue Op0 = N->getOperand(0);
9000 SDValue Op1 = N->getOperand(1);
9001 if (Op0.getOpcode() == ISD::BITCAST)
9002 Op0 = Op0.getOperand(0);
9003 if (Op1.getOpcode() == ISD::BITCAST)
9004 Op1 = Op1.getOperand(0);
9005 if (Op0.getOpcode() == ARMISD::VMOVRRD &&
9006 Op0.getNode() == Op1.getNode() &&
9007 Op0.getResNo() == 0 && Op1.getResNo() == 1)
9008 return DAG.getNode(ISD::BITCAST, SDLoc(N),
9009 N->getValueType(0), Op0.getOperand(0));
9013 /// hasNormalLoadOperand - Check if any of the operands of a BUILD_VECTOR node
9014 /// are normal, non-volatile loads. If so, it is profitable to bitcast an
9015 /// i64 vector to have f64 elements, since the value can then be loaded
9016 /// directly into a VFP register.
9017 static bool hasNormalLoadOperand(SDNode *N) {
9018 unsigned NumElts = N->getValueType(0).getVectorNumElements();
9019 for (unsigned i = 0; i < NumElts; ++i) {
9020 SDNode *Elt = N->getOperand(i).getNode();
9021 if (ISD::isNormalLoad(Elt) && !cast<LoadSDNode>(Elt)->isVolatile())
9027 /// PerformBUILD_VECTORCombine - Target-specific dag combine xforms for
9028 /// ISD::BUILD_VECTOR.
9029 static SDValue PerformBUILD_VECTORCombine(SDNode *N,
9030 TargetLowering::DAGCombinerInfo &DCI,
9031 const ARMSubtarget *Subtarget) {
9032 // build_vector(N=ARMISD::VMOVRRD(X), N:1) -> bit_convert(X):
9033 // VMOVRRD is introduced when legalizing i64 types. It forces the i64 value
9034 // into a pair of GPRs, which is fine when the value is used as a scalar,
9035 // but if the i64 value is converted to a vector, we need to undo the VMOVRRD.
9036 SelectionDAG &DAG = DCI.DAG;
9037 if (N->getNumOperands() == 2) {
9038 SDValue RV = PerformVMOVDRRCombine(N, DAG);
9043 // Load i64 elements as f64 values so that type legalization does not split
9044 // them up into i32 values.
9045 EVT VT = N->getValueType(0);
9046 if (VT.getVectorElementType() != MVT::i64 || !hasNormalLoadOperand(N))
9049 SmallVector<SDValue, 8> Ops;
9050 unsigned NumElts = VT.getVectorNumElements();
9051 for (unsigned i = 0; i < NumElts; ++i) {
9052 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(i));
9054 // Make the DAGCombiner fold the bitcast.
9055 DCI.AddToWorklist(V.getNode());
9057 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, NumElts);
9058 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, FloatVT, Ops);
9059 return DAG.getNode(ISD::BITCAST, dl, VT, BV);
9062 /// \brief Target-specific dag combine xforms for ARMISD::BUILD_VECTOR.
9064 PerformARMBUILD_VECTORCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
9065 // ARMISD::BUILD_VECTOR is introduced when legalizing ISD::BUILD_VECTOR.
9066 // At that time, we may have inserted bitcasts from integer to float.
9067 // If these bitcasts have survived DAGCombine, change the lowering of this
9068 // BUILD_VECTOR in something more vector friendly, i.e., that does not
9069 // force to use floating point types.
9071 // Make sure we can change the type of the vector.
9072 // This is possible iff:
9073 // 1. The vector is only used in a bitcast to a integer type. I.e.,
9074 // 1.1. Vector is used only once.
9075 // 1.2. Use is a bit convert to an integer type.
9076 // 2. The size of its operands are 32-bits (64-bits are not legal).
9077 EVT VT = N->getValueType(0);
9078 EVT EltVT = VT.getVectorElementType();
9080 // Check 1.1. and 2.
9081 if (EltVT.getSizeInBits() != 32 || !N->hasOneUse())
9084 // By construction, the input type must be float.
9085 assert(EltVT == MVT::f32 && "Unexpected type!");
9088 SDNode *Use = *N->use_begin();
9089 if (Use->getOpcode() != ISD::BITCAST ||
9090 Use->getValueType(0).isFloatingPoint())
9093 // Check profitability.
9094 // Model is, if more than half of the relevant operands are bitcast from
9095 // i32, turn the build_vector into a sequence of insert_vector_elt.
9096 // Relevant operands are everything that is not statically
9097 // (i.e., at compile time) bitcasted.
9098 unsigned NumOfBitCastedElts = 0;
9099 unsigned NumElts = VT.getVectorNumElements();
9100 unsigned NumOfRelevantElts = NumElts;
9101 for (unsigned Idx = 0; Idx < NumElts; ++Idx) {
9102 SDValue Elt = N->getOperand(Idx);
9103 if (Elt->getOpcode() == ISD::BITCAST) {
9104 // Assume only bit cast to i32 will go away.
9105 if (Elt->getOperand(0).getValueType() == MVT::i32)
9106 ++NumOfBitCastedElts;
9107 } else if (Elt.getOpcode() == ISD::UNDEF || isa<ConstantSDNode>(Elt))
9108 // Constants are statically casted, thus do not count them as
9109 // relevant operands.
9110 --NumOfRelevantElts;
9113 // Check if more than half of the elements require a non-free bitcast.
9114 if (NumOfBitCastedElts <= NumOfRelevantElts / 2)
9117 SelectionDAG &DAG = DCI.DAG;
9118 // Create the new vector type.
9119 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
9120 // Check if the type is legal.
9121 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9122 if (!TLI.isTypeLegal(VecVT))
9126 // ARMISD::BUILD_VECTOR E1, E2, ..., EN.
9127 // => BITCAST INSERT_VECTOR_ELT
9128 // (INSERT_VECTOR_ELT (...), (BITCAST EN-1), N-1),
9130 SDValue Vec = DAG.getUNDEF(VecVT);
9132 for (unsigned Idx = 0 ; Idx < NumElts; ++Idx) {
9133 SDValue V = N->getOperand(Idx);
9134 if (V.getOpcode() == ISD::UNDEF)
9136 if (V.getOpcode() == ISD::BITCAST &&
9137 V->getOperand(0).getValueType() == MVT::i32)
9138 // Fold obvious case.
9139 V = V.getOperand(0);
9141 V = DAG.getNode(ISD::BITCAST, SDLoc(V), MVT::i32, V);
9142 // Make the DAGCombiner fold the bitcasts.
9143 DCI.AddToWorklist(V.getNode());
9145 SDValue LaneIdx = DAG.getConstant(Idx, dl, MVT::i32);
9146 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VecVT, Vec, V, LaneIdx);
9148 Vec = DAG.getNode(ISD::BITCAST, dl, VT, Vec);
9149 // Make the DAGCombiner fold the bitcasts.
9150 DCI.AddToWorklist(Vec.getNode());
9154 /// PerformInsertEltCombine - Target-specific dag combine xforms for
9155 /// ISD::INSERT_VECTOR_ELT.
9156 static SDValue PerformInsertEltCombine(SDNode *N,
9157 TargetLowering::DAGCombinerInfo &DCI) {
9158 // Bitcast an i64 load inserted into a vector to f64.
9159 // Otherwise, the i64 value will be legalized to a pair of i32 values.
9160 EVT VT = N->getValueType(0);
9161 SDNode *Elt = N->getOperand(1).getNode();
9162 if (VT.getVectorElementType() != MVT::i64 ||
9163 !ISD::isNormalLoad(Elt) || cast<LoadSDNode>(Elt)->isVolatile())
9166 SelectionDAG &DAG = DCI.DAG;
9168 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
9169 VT.getVectorNumElements());
9170 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, N->getOperand(0));
9171 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(1));
9172 // Make the DAGCombiner fold the bitcasts.
9173 DCI.AddToWorklist(Vec.getNode());
9174 DCI.AddToWorklist(V.getNode());
9175 SDValue InsElt = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, FloatVT,
9176 Vec, V, N->getOperand(2));
9177 return DAG.getNode(ISD::BITCAST, dl, VT, InsElt);
9180 /// PerformVECTOR_SHUFFLECombine - Target-specific dag combine xforms for
9181 /// ISD::VECTOR_SHUFFLE.
9182 static SDValue PerformVECTOR_SHUFFLECombine(SDNode *N, SelectionDAG &DAG) {
9183 // The LLVM shufflevector instruction does not require the shuffle mask
9184 // length to match the operand vector length, but ISD::VECTOR_SHUFFLE does
9185 // have that requirement. When translating to ISD::VECTOR_SHUFFLE, if the
9186 // operands do not match the mask length, they are extended by concatenating
9187 // them with undef vectors. That is probably the right thing for other
9188 // targets, but for NEON it is better to concatenate two double-register
9189 // size vector operands into a single quad-register size vector. Do that
9190 // transformation here:
9191 // shuffle(concat(v1, undef), concat(v2, undef)) ->
9192 // shuffle(concat(v1, v2), undef)
9193 SDValue Op0 = N->getOperand(0);
9194 SDValue Op1 = N->getOperand(1);
9195 if (Op0.getOpcode() != ISD::CONCAT_VECTORS ||
9196 Op1.getOpcode() != ISD::CONCAT_VECTORS ||
9197 Op0.getNumOperands() != 2 ||
9198 Op1.getNumOperands() != 2)
9200 SDValue Concat0Op1 = Op0.getOperand(1);
9201 SDValue Concat1Op1 = Op1.getOperand(1);
9202 if (Concat0Op1.getOpcode() != ISD::UNDEF ||
9203 Concat1Op1.getOpcode() != ISD::UNDEF)
9205 // Skip the transformation if any of the types are illegal.
9206 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9207 EVT VT = N->getValueType(0);
9208 if (!TLI.isTypeLegal(VT) ||
9209 !TLI.isTypeLegal(Concat0Op1.getValueType()) ||
9210 !TLI.isTypeLegal(Concat1Op1.getValueType()))
9213 SDValue NewConcat = DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT,
9214 Op0.getOperand(0), Op1.getOperand(0));
9215 // Translate the shuffle mask.
9216 SmallVector<int, 16> NewMask;
9217 unsigned NumElts = VT.getVectorNumElements();
9218 unsigned HalfElts = NumElts/2;
9219 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
9220 for (unsigned n = 0; n < NumElts; ++n) {
9221 int MaskElt = SVN->getMaskElt(n);
9223 if (MaskElt < (int)HalfElts)
9225 else if (MaskElt >= (int)NumElts && MaskElt < (int)(NumElts + HalfElts))
9226 NewElt = HalfElts + MaskElt - NumElts;
9227 NewMask.push_back(NewElt);
9229 return DAG.getVectorShuffle(VT, SDLoc(N), NewConcat,
9230 DAG.getUNDEF(VT), NewMask.data());
9233 /// CombineBaseUpdate - Target-specific DAG combine function for VLDDUP,
9234 /// NEON load/store intrinsics, and generic vector load/stores, to merge
9235 /// base address updates.
9236 /// For generic load/stores, the memory type is assumed to be a vector.
9237 /// The caller is assumed to have checked legality.
9238 static SDValue CombineBaseUpdate(SDNode *N,
9239 TargetLowering::DAGCombinerInfo &DCI) {
9240 SelectionDAG &DAG = DCI.DAG;
9241 const bool isIntrinsic = (N->getOpcode() == ISD::INTRINSIC_VOID ||
9242 N->getOpcode() == ISD::INTRINSIC_W_CHAIN);
9243 const bool isStore = N->getOpcode() == ISD::STORE;
9244 const unsigned AddrOpIdx = ((isIntrinsic || isStore) ? 2 : 1);
9245 SDValue Addr = N->getOperand(AddrOpIdx);
9246 MemSDNode *MemN = cast<MemSDNode>(N);
9249 // Search for a use of the address operand that is an increment.
9250 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(),
9251 UE = Addr.getNode()->use_end(); UI != UE; ++UI) {
9253 if (User->getOpcode() != ISD::ADD ||
9254 UI.getUse().getResNo() != Addr.getResNo())
9257 // Check that the add is independent of the load/store. Otherwise, folding
9258 // it would create a cycle.
9259 if (User->isPredecessorOf(N) || N->isPredecessorOf(User))
9262 // Find the new opcode for the updating load/store.
9263 bool isLoadOp = true;
9264 bool isLaneOp = false;
9265 unsigned NewOpc = 0;
9266 unsigned NumVecs = 0;
9268 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
9270 default: llvm_unreachable("unexpected intrinsic for Neon base update");
9271 case Intrinsic::arm_neon_vld1: NewOpc = ARMISD::VLD1_UPD;
9273 case Intrinsic::arm_neon_vld2: NewOpc = ARMISD::VLD2_UPD;
9275 case Intrinsic::arm_neon_vld3: NewOpc = ARMISD::VLD3_UPD;
9277 case Intrinsic::arm_neon_vld4: NewOpc = ARMISD::VLD4_UPD;
9279 case Intrinsic::arm_neon_vld2lane: NewOpc = ARMISD::VLD2LN_UPD;
9280 NumVecs = 2; isLaneOp = true; break;
9281 case Intrinsic::arm_neon_vld3lane: NewOpc = ARMISD::VLD3LN_UPD;
9282 NumVecs = 3; isLaneOp = true; break;
9283 case Intrinsic::arm_neon_vld4lane: NewOpc = ARMISD::VLD4LN_UPD;
9284 NumVecs = 4; isLaneOp = true; break;
9285 case Intrinsic::arm_neon_vst1: NewOpc = ARMISD::VST1_UPD;
9286 NumVecs = 1; isLoadOp = false; break;
9287 case Intrinsic::arm_neon_vst2: NewOpc = ARMISD::VST2_UPD;
9288 NumVecs = 2; isLoadOp = false; break;
9289 case Intrinsic::arm_neon_vst3: NewOpc = ARMISD::VST3_UPD;
9290 NumVecs = 3; isLoadOp = false; break;
9291 case Intrinsic::arm_neon_vst4: NewOpc = ARMISD::VST4_UPD;
9292 NumVecs = 4; isLoadOp = false; break;
9293 case Intrinsic::arm_neon_vst2lane: NewOpc = ARMISD::VST2LN_UPD;
9294 NumVecs = 2; isLoadOp = false; isLaneOp = true; break;
9295 case Intrinsic::arm_neon_vst3lane: NewOpc = ARMISD::VST3LN_UPD;
9296 NumVecs = 3; isLoadOp = false; isLaneOp = true; break;
9297 case Intrinsic::arm_neon_vst4lane: NewOpc = ARMISD::VST4LN_UPD;
9298 NumVecs = 4; isLoadOp = false; isLaneOp = true; break;
9302 switch (N->getOpcode()) {
9303 default: llvm_unreachable("unexpected opcode for Neon base update");
9304 case ARMISD::VLD2DUP: NewOpc = ARMISD::VLD2DUP_UPD; NumVecs = 2; break;
9305 case ARMISD::VLD3DUP: NewOpc = ARMISD::VLD3DUP_UPD; NumVecs = 3; break;
9306 case ARMISD::VLD4DUP: NewOpc = ARMISD::VLD4DUP_UPD; NumVecs = 4; break;
9307 case ISD::LOAD: NewOpc = ARMISD::VLD1_UPD;
9308 NumVecs = 1; isLaneOp = false; break;
9309 case ISD::STORE: NewOpc = ARMISD::VST1_UPD;
9310 NumVecs = 1; isLaneOp = false; isLoadOp = false; break;
9314 // Find the size of memory referenced by the load/store.
9317 VecTy = N->getValueType(0);
9318 } else if (isIntrinsic) {
9319 VecTy = N->getOperand(AddrOpIdx+1).getValueType();
9321 assert(isStore && "Node has to be a load, a store, or an intrinsic!");
9322 VecTy = N->getOperand(1).getValueType();
9325 unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8;
9327 NumBytes /= VecTy.getVectorNumElements();
9329 // If the increment is a constant, it must match the memory ref size.
9330 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
9331 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
9332 uint64_t IncVal = CInc->getZExtValue();
9333 if (IncVal != NumBytes)
9335 } else if (NumBytes >= 3 * 16) {
9336 // VLD3/4 and VST3/4 for 128-bit vectors are implemented with two
9337 // separate instructions that make it harder to use a non-constant update.
9341 // OK, we found an ADD we can fold into the base update.
9342 // Now, create a _UPD node, taking care of not breaking alignment.
9344 EVT AlignedVecTy = VecTy;
9345 unsigned Alignment = MemN->getAlignment();
9347 // If this is a less-than-standard-aligned load/store, change the type to
9348 // match the standard alignment.
9349 // The alignment is overlooked when selecting _UPD variants; and it's
9350 // easier to introduce bitcasts here than fix that.
9351 // There are 3 ways to get to this base-update combine:
9352 // - intrinsics: they are assumed to be properly aligned (to the standard
9353 // alignment of the memory type), so we don't need to do anything.
9354 // - ARMISD::VLDx nodes: they are only generated from the aforementioned
9355 // intrinsics, so, likewise, there's nothing to do.
9356 // - generic load/store instructions: the alignment is specified as an
9357 // explicit operand, rather than implicitly as the standard alignment
9358 // of the memory type (like the intrisics). We need to change the
9359 // memory type to match the explicit alignment. That way, we don't
9360 // generate non-standard-aligned ARMISD::VLDx nodes.
9361 if (isa<LSBaseSDNode>(N)) {
9364 if (Alignment < VecTy.getScalarSizeInBits() / 8) {
9365 MVT EltTy = MVT::getIntegerVT(Alignment * 8);
9366 assert(NumVecs == 1 && "Unexpected multi-element generic load/store.");
9367 assert(!isLaneOp && "Unexpected generic load/store lane.");
9368 unsigned NumElts = NumBytes / (EltTy.getSizeInBits() / 8);
9369 AlignedVecTy = MVT::getVectorVT(EltTy, NumElts);
9371 // Don't set an explicit alignment on regular load/stores that we want
9372 // to transform to VLD/VST 1_UPD nodes.
9373 // This matches the behavior of regular load/stores, which only get an
9374 // explicit alignment if the MMO alignment is larger than the standard
9375 // alignment of the memory type.
9376 // Intrinsics, however, always get an explicit alignment, set to the
9377 // alignment of the MMO.
9381 // Create the new updating load/store node.
9382 // First, create an SDVTList for the new updating node's results.
9384 unsigned NumResultVecs = (isLoadOp ? NumVecs : 0);
9386 for (n = 0; n < NumResultVecs; ++n)
9387 Tys[n] = AlignedVecTy;
9388 Tys[n++] = MVT::i32;
9389 Tys[n] = MVT::Other;
9390 SDVTList SDTys = DAG.getVTList(makeArrayRef(Tys, NumResultVecs+2));
9392 // Then, gather the new node's operands.
9393 SmallVector<SDValue, 8> Ops;
9394 Ops.push_back(N->getOperand(0)); // incoming chain
9395 Ops.push_back(N->getOperand(AddrOpIdx));
9398 if (StoreSDNode *StN = dyn_cast<StoreSDNode>(N)) {
9399 // Try to match the intrinsic's signature
9400 Ops.push_back(StN->getValue());
9402 // Loads (and of course intrinsics) match the intrinsics' signature,
9403 // so just add all but the alignment operand.
9404 for (unsigned i = AddrOpIdx + 1; i < N->getNumOperands() - 1; ++i)
9405 Ops.push_back(N->getOperand(i));
9408 // For all node types, the alignment operand is always the last one.
9409 Ops.push_back(DAG.getConstant(Alignment, dl, MVT::i32));
9411 // If this is a non-standard-aligned STORE, the penultimate operand is the
9412 // stored value. Bitcast it to the aligned type.
9413 if (AlignedVecTy != VecTy && N->getOpcode() == ISD::STORE) {
9414 SDValue &StVal = Ops[Ops.size()-2];
9415 StVal = DAG.getNode(ISD::BITCAST, dl, AlignedVecTy, StVal);
9418 SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, dl, SDTys,
9420 MemN->getMemOperand());
9423 SmallVector<SDValue, 5> NewResults;
9424 for (unsigned i = 0; i < NumResultVecs; ++i)
9425 NewResults.push_back(SDValue(UpdN.getNode(), i));
9427 // If this is an non-standard-aligned LOAD, the first result is the loaded
9428 // value. Bitcast it to the expected result type.
9429 if (AlignedVecTy != VecTy && N->getOpcode() == ISD::LOAD) {
9430 SDValue &LdVal = NewResults[0];
9431 LdVal = DAG.getNode(ISD::BITCAST, dl, VecTy, LdVal);
9434 NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs+1)); // chain
9435 DCI.CombineTo(N, NewResults);
9436 DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs));
9443 static SDValue PerformVLDCombine(SDNode *N,
9444 TargetLowering::DAGCombinerInfo &DCI) {
9445 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
9448 return CombineBaseUpdate(N, DCI);
9451 /// CombineVLDDUP - For a VDUPLANE node N, check if its source operand is a
9452 /// vldN-lane (N > 1) intrinsic, and if all the other uses of that intrinsic
9453 /// are also VDUPLANEs. If so, combine them to a vldN-dup operation and
9455 static bool CombineVLDDUP(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
9456 SelectionDAG &DAG = DCI.DAG;
9457 EVT VT = N->getValueType(0);
9458 // vldN-dup instructions only support 64-bit vectors for N > 1.
9459 if (!VT.is64BitVector())
9462 // Check if the VDUPLANE operand is a vldN-dup intrinsic.
9463 SDNode *VLD = N->getOperand(0).getNode();
9464 if (VLD->getOpcode() != ISD::INTRINSIC_W_CHAIN)
9466 unsigned NumVecs = 0;
9467 unsigned NewOpc = 0;
9468 unsigned IntNo = cast<ConstantSDNode>(VLD->getOperand(1))->getZExtValue();
9469 if (IntNo == Intrinsic::arm_neon_vld2lane) {
9471 NewOpc = ARMISD::VLD2DUP;
9472 } else if (IntNo == Intrinsic::arm_neon_vld3lane) {
9474 NewOpc = ARMISD::VLD3DUP;
9475 } else if (IntNo == Intrinsic::arm_neon_vld4lane) {
9477 NewOpc = ARMISD::VLD4DUP;
9482 // First check that all the vldN-lane uses are VDUPLANEs and that the lane
9483 // numbers match the load.
9484 unsigned VLDLaneNo =
9485 cast<ConstantSDNode>(VLD->getOperand(NumVecs+3))->getZExtValue();
9486 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
9488 // Ignore uses of the chain result.
9489 if (UI.getUse().getResNo() == NumVecs)
9492 if (User->getOpcode() != ARMISD::VDUPLANE ||
9493 VLDLaneNo != cast<ConstantSDNode>(User->getOperand(1))->getZExtValue())
9497 // Create the vldN-dup node.
9500 for (n = 0; n < NumVecs; ++n)
9502 Tys[n] = MVT::Other;
9503 SDVTList SDTys = DAG.getVTList(makeArrayRef(Tys, NumVecs+1));
9504 SDValue Ops[] = { VLD->getOperand(0), VLD->getOperand(2) };
9505 MemIntrinsicSDNode *VLDMemInt = cast<MemIntrinsicSDNode>(VLD);
9506 SDValue VLDDup = DAG.getMemIntrinsicNode(NewOpc, SDLoc(VLD), SDTys,
9507 Ops, VLDMemInt->getMemoryVT(),
9508 VLDMemInt->getMemOperand());
9511 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
9513 unsigned ResNo = UI.getUse().getResNo();
9514 // Ignore uses of the chain result.
9515 if (ResNo == NumVecs)
9518 DCI.CombineTo(User, SDValue(VLDDup.getNode(), ResNo));
9521 // Now the vldN-lane intrinsic is dead except for its chain result.
9522 // Update uses of the chain.
9523 std::vector<SDValue> VLDDupResults;
9524 for (unsigned n = 0; n < NumVecs; ++n)
9525 VLDDupResults.push_back(SDValue(VLDDup.getNode(), n));
9526 VLDDupResults.push_back(SDValue(VLDDup.getNode(), NumVecs));
9527 DCI.CombineTo(VLD, VLDDupResults);
9532 /// PerformVDUPLANECombine - Target-specific dag combine xforms for
9533 /// ARMISD::VDUPLANE.
9534 static SDValue PerformVDUPLANECombine(SDNode *N,
9535 TargetLowering::DAGCombinerInfo &DCI) {
9536 SDValue Op = N->getOperand(0);
9538 // If the source is a vldN-lane (N > 1) intrinsic, and all the other uses
9539 // of that intrinsic are also VDUPLANEs, combine them to a vldN-dup operation.
9540 if (CombineVLDDUP(N, DCI))
9541 return SDValue(N, 0);
9543 // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is
9544 // redundant. Ignore bit_converts for now; element sizes are checked below.
9545 while (Op.getOpcode() == ISD::BITCAST)
9546 Op = Op.getOperand(0);
9547 if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM)
9550 // Make sure the VMOV element size is not bigger than the VDUPLANE elements.
9551 unsigned EltSize = Op.getValueType().getVectorElementType().getSizeInBits();
9552 // The canonical VMOV for a zero vector uses a 32-bit element size.
9553 unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
9555 if (ARM_AM::decodeNEONModImm(Imm, EltBits) == 0)
9557 EVT VT = N->getValueType(0);
9558 if (EltSize > VT.getVectorElementType().getSizeInBits())
9561 return DCI.DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Op);
9564 static SDValue PerformLOADCombine(SDNode *N,
9565 TargetLowering::DAGCombinerInfo &DCI) {
9566 EVT VT = N->getValueType(0);
9568 // If this is a legal vector load, try to combine it into a VLD1_UPD.
9569 if (ISD::isNormalLoad(N) && VT.isVector() &&
9570 DCI.DAG.getTargetLoweringInfo().isTypeLegal(VT))
9571 return CombineBaseUpdate(N, DCI);
9576 /// PerformSTORECombine - Target-specific dag combine xforms for
9578 static SDValue PerformSTORECombine(SDNode *N,
9579 TargetLowering::DAGCombinerInfo &DCI) {
9580 StoreSDNode *St = cast<StoreSDNode>(N);
9581 if (St->isVolatile())
9584 // Optimize trunc store (of multiple scalars) to shuffle and store. First,
9585 // pack all of the elements in one place. Next, store to memory in fewer
9587 SDValue StVal = St->getValue();
9588 EVT VT = StVal.getValueType();
9589 if (St->isTruncatingStore() && VT.isVector()) {
9590 SelectionDAG &DAG = DCI.DAG;
9591 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9592 EVT StVT = St->getMemoryVT();
9593 unsigned NumElems = VT.getVectorNumElements();
9594 assert(StVT != VT && "Cannot truncate to the same type");
9595 unsigned FromEltSz = VT.getVectorElementType().getSizeInBits();
9596 unsigned ToEltSz = StVT.getVectorElementType().getSizeInBits();
9598 // From, To sizes and ElemCount must be pow of two
9599 if (!isPowerOf2_32(NumElems * FromEltSz * ToEltSz)) return SDValue();
9601 // We are going to use the original vector elt for storing.
9602 // Accumulated smaller vector elements must be a multiple of the store size.
9603 if (0 != (NumElems * FromEltSz) % ToEltSz) return SDValue();
9605 unsigned SizeRatio = FromEltSz / ToEltSz;
9606 assert(SizeRatio * NumElems * ToEltSz == VT.getSizeInBits());
9608 // Create a type on which we perform the shuffle.
9609 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(), StVT.getScalarType(),
9610 NumElems*SizeRatio);
9611 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
9614 SDValue WideVec = DAG.getNode(ISD::BITCAST, DL, WideVecVT, StVal);
9615 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
9616 for (unsigned i = 0; i < NumElems; ++i)
9617 ShuffleVec[i] = DAG.getDataLayout().isBigEndian()
9618 ? (i + 1) * SizeRatio - 1
9621 // Can't shuffle using an illegal type.
9622 if (!TLI.isTypeLegal(WideVecVT)) return SDValue();
9624 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, DL, WideVec,
9625 DAG.getUNDEF(WideVec.getValueType()),
9627 // At this point all of the data is stored at the bottom of the
9628 // register. We now need to save it to mem.
9630 // Find the largest store unit
9631 MVT StoreType = MVT::i8;
9632 for (MVT Tp : MVT::integer_valuetypes()) {
9633 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToEltSz)
9636 // Didn't find a legal store type.
9637 if (!TLI.isTypeLegal(StoreType))
9640 // Bitcast the original vector into a vector of store-size units
9641 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
9642 StoreType, VT.getSizeInBits()/EVT(StoreType).getSizeInBits());
9643 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
9644 SDValue ShuffWide = DAG.getNode(ISD::BITCAST, DL, StoreVecVT, Shuff);
9645 SmallVector<SDValue, 8> Chains;
9646 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits() / 8, DL,
9647 TLI.getPointerTy(DAG.getDataLayout()));
9648 SDValue BasePtr = St->getBasePtr();
9650 // Perform one or more big stores into memory.
9651 unsigned E = (ToEltSz*NumElems)/StoreType.getSizeInBits();
9652 for (unsigned I = 0; I < E; I++) {
9653 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
9654 StoreType, ShuffWide,
9655 DAG.getIntPtrConstant(I, DL));
9656 SDValue Ch = DAG.getStore(St->getChain(), DL, SubVec, BasePtr,
9657 St->getPointerInfo(), St->isVolatile(),
9658 St->isNonTemporal(), St->getAlignment());
9659 BasePtr = DAG.getNode(ISD::ADD, DL, BasePtr.getValueType(), BasePtr,
9661 Chains.push_back(Ch);
9663 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
9666 if (!ISD::isNormalStore(St))
9669 // Split a store of a VMOVDRR into two integer stores to avoid mixing NEON and
9670 // ARM stores of arguments in the same cache line.
9671 if (StVal.getNode()->getOpcode() == ARMISD::VMOVDRR &&
9672 StVal.getNode()->hasOneUse()) {
9673 SelectionDAG &DAG = DCI.DAG;
9674 bool isBigEndian = DAG.getDataLayout().isBigEndian();
9676 SDValue BasePtr = St->getBasePtr();
9677 SDValue NewST1 = DAG.getStore(St->getChain(), DL,
9678 StVal.getNode()->getOperand(isBigEndian ? 1 : 0 ),
9679 BasePtr, St->getPointerInfo(), St->isVolatile(),
9680 St->isNonTemporal(), St->getAlignment());
9682 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
9683 DAG.getConstant(4, DL, MVT::i32));
9684 return DAG.getStore(NewST1.getValue(0), DL,
9685 StVal.getNode()->getOperand(isBigEndian ? 0 : 1),
9686 OffsetPtr, St->getPointerInfo(), St->isVolatile(),
9687 St->isNonTemporal(),
9688 std::min(4U, St->getAlignment() / 2));
9691 if (StVal.getValueType() == MVT::i64 &&
9692 StVal.getNode()->getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
9694 // Bitcast an i64 store extracted from a vector to f64.
9695 // Otherwise, the i64 value will be legalized to a pair of i32 values.
9696 SelectionDAG &DAG = DCI.DAG;
9698 SDValue IntVec = StVal.getOperand(0);
9699 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
9700 IntVec.getValueType().getVectorNumElements());
9701 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, IntVec);
9702 SDValue ExtElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
9703 Vec, StVal.getOperand(1));
9705 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ExtElt);
9706 // Make the DAGCombiner fold the bitcasts.
9707 DCI.AddToWorklist(Vec.getNode());
9708 DCI.AddToWorklist(ExtElt.getNode());
9709 DCI.AddToWorklist(V.getNode());
9710 return DAG.getStore(St->getChain(), dl, V, St->getBasePtr(),
9711 St->getPointerInfo(), St->isVolatile(),
9712 St->isNonTemporal(), St->getAlignment(),
9716 // If this is a legal vector store, try to combine it into a VST1_UPD.
9717 if (ISD::isNormalStore(N) && VT.isVector() &&
9718 DCI.DAG.getTargetLoweringInfo().isTypeLegal(VT))
9719 return CombineBaseUpdate(N, DCI);
9724 // isConstVecPow2 - Return true if each vector element is a power of 2, all
9725 // elements are the same constant, C, and Log2(C) ranges from 1 to 32.
9726 static bool isConstVecPow2(SDValue ConstVec, bool isSigned, uint64_t &C)
9730 for (unsigned I = 0, E = ConstVec.getValueType().getVectorNumElements();
9732 ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(ConstVec.getOperand(I));
9737 APFloat APF = C->getValueAPF();
9738 if (APF.convertToInteger(&cN, 64, isSigned, APFloat::rmTowardZero, &isExact)
9739 != APFloat::opOK || !isExact)
9742 c0 = (I == 0) ? cN : c0;
9743 if (!isPowerOf2_64(cN) || c0 != cN || Log2_64(c0) < 1 || Log2_64(c0) > 32)
9750 /// PerformVCVTCombine - VCVT (floating-point to fixed-point, Advanced SIMD)
9751 /// can replace combinations of VMUL and VCVT (floating-point to integer)
9752 /// when the VMUL has a constant operand that is a power of 2.
9754 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
9755 /// vmul.f32 d16, d17, d16
9756 /// vcvt.s32.f32 d16, d16
9758 /// vcvt.s32.f32 d16, d16, #3
9759 static SDValue PerformVCVTCombine(SDNode *N,
9760 TargetLowering::DAGCombinerInfo &DCI,
9761 const ARMSubtarget *Subtarget) {
9762 SelectionDAG &DAG = DCI.DAG;
9763 SDValue Op = N->getOperand(0);
9765 if (!Subtarget->hasNEON() || !Op.getValueType().isVector() ||
9766 Op.getOpcode() != ISD::FMUL)
9770 SDValue N0 = Op->getOperand(0);
9771 SDValue ConstVec = Op->getOperand(1);
9772 bool isSigned = N->getOpcode() == ISD::FP_TO_SINT;
9774 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
9775 !isConstVecPow2(ConstVec, isSigned, C))
9778 MVT FloatTy = Op.getSimpleValueType().getVectorElementType();
9779 MVT IntTy = N->getSimpleValueType(0).getVectorElementType();
9780 unsigned NumLanes = Op.getValueType().getVectorNumElements();
9781 if (FloatTy.getSizeInBits() != 32 || IntTy.getSizeInBits() > 32 ||
9783 // These instructions only exist converting from f32 to i32. We can handle
9784 // smaller integers by generating an extra truncate, but larger ones would
9785 // be lossy. We also can't handle more then 4 lanes, since these intructions
9786 // only support v2i32/v4i32 types.
9791 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfp2fxs :
9792 Intrinsic::arm_neon_vcvtfp2fxu;
9793 SDValue FixConv = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl,
9794 NumLanes == 2 ? MVT::v2i32 : MVT::v4i32,
9795 DAG.getConstant(IntrinsicOpcode, dl, MVT::i32),
9797 DAG.getConstant(Log2_64(C), dl, MVT::i32));
9799 if (IntTy.getSizeInBits() < FloatTy.getSizeInBits())
9800 FixConv = DAG.getNode(ISD::TRUNCATE, dl, N->getValueType(0), FixConv);
9805 /// PerformVDIVCombine - VCVT (fixed-point to floating-point, Advanced SIMD)
9806 /// can replace combinations of VCVT (integer to floating-point) and VDIV
9807 /// when the VDIV has a constant operand that is a power of 2.
9809 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
9810 /// vcvt.f32.s32 d16, d16
9811 /// vdiv.f32 d16, d17, d16
9813 /// vcvt.f32.s32 d16, d16, #3
9814 static SDValue PerformVDIVCombine(SDNode *N,
9815 TargetLowering::DAGCombinerInfo &DCI,
9816 const ARMSubtarget *Subtarget) {
9817 SelectionDAG &DAG = DCI.DAG;
9818 SDValue Op = N->getOperand(0);
9819 unsigned OpOpcode = Op.getNode()->getOpcode();
9821 if (!Subtarget->hasNEON() || !N->getValueType(0).isVector() ||
9822 (OpOpcode != ISD::SINT_TO_FP && OpOpcode != ISD::UINT_TO_FP))
9826 SDValue ConstVec = N->getOperand(1);
9827 bool isSigned = OpOpcode == ISD::SINT_TO_FP;
9829 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
9830 !isConstVecPow2(ConstVec, isSigned, C))
9833 MVT FloatTy = N->getSimpleValueType(0).getVectorElementType();
9834 MVT IntTy = Op.getOperand(0).getSimpleValueType().getVectorElementType();
9835 if (FloatTy.getSizeInBits() != 32 || IntTy.getSizeInBits() > 32) {
9836 // These instructions only exist converting from i32 to f32. We can handle
9837 // smaller integers by generating an extra extend, but larger ones would
9843 SDValue ConvInput = Op.getOperand(0);
9844 unsigned NumLanes = Op.getValueType().getVectorNumElements();
9845 if (IntTy.getSizeInBits() < FloatTy.getSizeInBits())
9846 ConvInput = DAG.getNode(isSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND,
9847 dl, NumLanes == 2 ? MVT::v2i32 : MVT::v4i32,
9850 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfxs2fp :
9851 Intrinsic::arm_neon_vcvtfxu2fp;
9852 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl,
9854 DAG.getConstant(IntrinsicOpcode, dl, MVT::i32),
9855 ConvInput, DAG.getConstant(Log2_64(C), dl, MVT::i32));
9858 /// Getvshiftimm - Check if this is a valid build_vector for the immediate
9859 /// operand of a vector shift operation, where all the elements of the
9860 /// build_vector must have the same constant integer value.
9861 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
9862 // Ignore bit_converts.
9863 while (Op.getOpcode() == ISD::BITCAST)
9864 Op = Op.getOperand(0);
9865 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
9866 APInt SplatBits, SplatUndef;
9867 unsigned SplatBitSize;
9869 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
9870 HasAnyUndefs, ElementBits) ||
9871 SplatBitSize > ElementBits)
9873 Cnt = SplatBits.getSExtValue();
9877 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
9878 /// operand of a vector shift left operation. That value must be in the range:
9879 /// 0 <= Value < ElementBits for a left shift; or
9880 /// 0 <= Value <= ElementBits for a long left shift.
9881 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
9882 assert(VT.isVector() && "vector shift count is not a vector type");
9883 int64_t ElementBits = VT.getVectorElementType().getSizeInBits();
9884 if (! getVShiftImm(Op, ElementBits, Cnt))
9886 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
9889 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
9890 /// operand of a vector shift right operation. For a shift opcode, the value
9891 /// is positive, but for an intrinsic the value count must be negative. The
9892 /// absolute value must be in the range:
9893 /// 1 <= |Value| <= ElementBits for a right shift; or
9894 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
9895 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
9897 assert(VT.isVector() && "vector shift count is not a vector type");
9898 int64_t ElementBits = VT.getVectorElementType().getSizeInBits();
9899 if (! getVShiftImm(Op, ElementBits, Cnt))
9902 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
9903 if (Cnt >= -(isNarrow ? ElementBits/2 : ElementBits) && Cnt <= -1) {
9910 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
9911 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
9912 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
9915 // Don't do anything for most intrinsics.
9918 case Intrinsic::arm_neon_vabds:
9919 if (!N->getValueType(0).isInteger())
9921 return DAG.getNode(ISD::SABSDIFF, SDLoc(N), N->getValueType(0),
9922 N->getOperand(1), N->getOperand(2));
9923 case Intrinsic::arm_neon_vabdu:
9924 return DAG.getNode(ISD::UABSDIFF, SDLoc(N), N->getValueType(0),
9925 N->getOperand(1), N->getOperand(2));
9927 // Vector shifts: check for immediate versions and lower them.
9928 // Note: This is done during DAG combining instead of DAG legalizing because
9929 // the build_vectors for 64-bit vector element shift counts are generally
9930 // not legal, and it is hard to see their values after they get legalized to
9931 // loads from a constant pool.
9932 case Intrinsic::arm_neon_vshifts:
9933 case Intrinsic::arm_neon_vshiftu:
9934 case Intrinsic::arm_neon_vrshifts:
9935 case Intrinsic::arm_neon_vrshiftu:
9936 case Intrinsic::arm_neon_vrshiftn:
9937 case Intrinsic::arm_neon_vqshifts:
9938 case Intrinsic::arm_neon_vqshiftu:
9939 case Intrinsic::arm_neon_vqshiftsu:
9940 case Intrinsic::arm_neon_vqshiftns:
9941 case Intrinsic::arm_neon_vqshiftnu:
9942 case Intrinsic::arm_neon_vqshiftnsu:
9943 case Intrinsic::arm_neon_vqrshiftns:
9944 case Intrinsic::arm_neon_vqrshiftnu:
9945 case Intrinsic::arm_neon_vqrshiftnsu: {
9946 EVT VT = N->getOperand(1).getValueType();
9948 unsigned VShiftOpc = 0;
9951 case Intrinsic::arm_neon_vshifts:
9952 case Intrinsic::arm_neon_vshiftu:
9953 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
9954 VShiftOpc = ARMISD::VSHL;
9957 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
9958 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
9959 ARMISD::VSHRs : ARMISD::VSHRu);
9964 case Intrinsic::arm_neon_vrshifts:
9965 case Intrinsic::arm_neon_vrshiftu:
9966 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
9970 case Intrinsic::arm_neon_vqshifts:
9971 case Intrinsic::arm_neon_vqshiftu:
9972 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
9976 case Intrinsic::arm_neon_vqshiftsu:
9977 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
9979 llvm_unreachable("invalid shift count for vqshlu intrinsic");
9981 case Intrinsic::arm_neon_vrshiftn:
9982 case Intrinsic::arm_neon_vqshiftns:
9983 case Intrinsic::arm_neon_vqshiftnu:
9984 case Intrinsic::arm_neon_vqshiftnsu:
9985 case Intrinsic::arm_neon_vqrshiftns:
9986 case Intrinsic::arm_neon_vqrshiftnu:
9987 case Intrinsic::arm_neon_vqrshiftnsu:
9988 // Narrowing shifts require an immediate right shift.
9989 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
9991 llvm_unreachable("invalid shift count for narrowing vector shift "
9995 llvm_unreachable("unhandled vector shift");
9999 case Intrinsic::arm_neon_vshifts:
10000 case Intrinsic::arm_neon_vshiftu:
10001 // Opcode already set above.
10003 case Intrinsic::arm_neon_vrshifts:
10004 VShiftOpc = ARMISD::VRSHRs; break;
10005 case Intrinsic::arm_neon_vrshiftu:
10006 VShiftOpc = ARMISD::VRSHRu; break;
10007 case Intrinsic::arm_neon_vrshiftn:
10008 VShiftOpc = ARMISD::VRSHRN; break;
10009 case Intrinsic::arm_neon_vqshifts:
10010 VShiftOpc = ARMISD::VQSHLs; break;
10011 case Intrinsic::arm_neon_vqshiftu:
10012 VShiftOpc = ARMISD::VQSHLu; break;
10013 case Intrinsic::arm_neon_vqshiftsu:
10014 VShiftOpc = ARMISD::VQSHLsu; break;
10015 case Intrinsic::arm_neon_vqshiftns:
10016 VShiftOpc = ARMISD::VQSHRNs; break;
10017 case Intrinsic::arm_neon_vqshiftnu:
10018 VShiftOpc = ARMISD::VQSHRNu; break;
10019 case Intrinsic::arm_neon_vqshiftnsu:
10020 VShiftOpc = ARMISD::VQSHRNsu; break;
10021 case Intrinsic::arm_neon_vqrshiftns:
10022 VShiftOpc = ARMISD::VQRSHRNs; break;
10023 case Intrinsic::arm_neon_vqrshiftnu:
10024 VShiftOpc = ARMISD::VQRSHRNu; break;
10025 case Intrinsic::arm_neon_vqrshiftnsu:
10026 VShiftOpc = ARMISD::VQRSHRNsu; break;
10030 return DAG.getNode(VShiftOpc, dl, N->getValueType(0),
10031 N->getOperand(1), DAG.getConstant(Cnt, dl, MVT::i32));
10034 case Intrinsic::arm_neon_vshiftins: {
10035 EVT VT = N->getOperand(1).getValueType();
10037 unsigned VShiftOpc = 0;
10039 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
10040 VShiftOpc = ARMISD::VSLI;
10041 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
10042 VShiftOpc = ARMISD::VSRI;
10044 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
10048 return DAG.getNode(VShiftOpc, dl, N->getValueType(0),
10049 N->getOperand(1), N->getOperand(2),
10050 DAG.getConstant(Cnt, dl, MVT::i32));
10053 case Intrinsic::arm_neon_vqrshifts:
10054 case Intrinsic::arm_neon_vqrshiftu:
10055 // No immediate versions of these to check for.
10062 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
10063 /// lowers them. As with the vector shift intrinsics, this is done during DAG
10064 /// combining instead of DAG legalizing because the build_vectors for 64-bit
10065 /// vector element shift counts are generally not legal, and it is hard to see
10066 /// their values after they get legalized to loads from a constant pool.
10067 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
10068 const ARMSubtarget *ST) {
10069 EVT VT = N->getValueType(0);
10070 if (N->getOpcode() == ISD::SRL && VT == MVT::i32 && ST->hasV6Ops()) {
10071 // Canonicalize (srl (bswap x), 16) to (rotr (bswap x), 16) if the high
10072 // 16-bits of x is zero. This optimizes rev + lsr 16 to rev16.
10073 SDValue N1 = N->getOperand(1);
10074 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) {
10075 SDValue N0 = N->getOperand(0);
10076 if (C->getZExtValue() == 16 && N0.getOpcode() == ISD::BSWAP &&
10077 DAG.MaskedValueIsZero(N0.getOperand(0),
10078 APInt::getHighBitsSet(32, 16)))
10079 return DAG.getNode(ISD::ROTR, SDLoc(N), VT, N0, N1);
10083 // Nothing to be done for scalar shifts.
10084 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
10085 if (!VT.isVector() || !TLI.isTypeLegal(VT))
10088 assert(ST->hasNEON() && "unexpected vector shift");
10091 switch (N->getOpcode()) {
10092 default: llvm_unreachable("unexpected shift opcode");
10095 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt)) {
10097 return DAG.getNode(ARMISD::VSHL, dl, VT, N->getOperand(0),
10098 DAG.getConstant(Cnt, dl, MVT::i32));
10104 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
10105 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
10106 ARMISD::VSHRs : ARMISD::VSHRu);
10108 return DAG.getNode(VShiftOpc, dl, VT, N->getOperand(0),
10109 DAG.getConstant(Cnt, dl, MVT::i32));
10115 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
10116 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
10117 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
10118 const ARMSubtarget *ST) {
10119 SDValue N0 = N->getOperand(0);
10121 // Check for sign- and zero-extensions of vector extract operations of 8-
10122 // and 16-bit vector elements. NEON supports these directly. They are
10123 // handled during DAG combining because type legalization will promote them
10124 // to 32-bit types and it is messy to recognize the operations after that.
10125 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
10126 SDValue Vec = N0.getOperand(0);
10127 SDValue Lane = N0.getOperand(1);
10128 EVT VT = N->getValueType(0);
10129 EVT EltVT = N0.getValueType();
10130 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
10132 if (VT == MVT::i32 &&
10133 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
10134 TLI.isTypeLegal(Vec.getValueType()) &&
10135 isa<ConstantSDNode>(Lane)) {
10138 switch (N->getOpcode()) {
10139 default: llvm_unreachable("unexpected opcode");
10140 case ISD::SIGN_EXTEND:
10141 Opc = ARMISD::VGETLANEs;
10143 case ISD::ZERO_EXTEND:
10144 case ISD::ANY_EXTEND:
10145 Opc = ARMISD::VGETLANEu;
10148 return DAG.getNode(Opc, SDLoc(N), VT, Vec, Lane);
10155 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
10156 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
10157 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
10158 const ARMSubtarget *ST) {
10159 // If the target supports NEON, try to use vmax/vmin instructions for f32
10160 // selects like "x < y ? x : y". Unless the NoNaNsFPMath option is set,
10161 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
10162 // a NaN; only do the transformation when it matches that behavior.
10164 // For now only do this when using NEON for FP operations; if using VFP, it
10165 // is not obvious that the benefit outweighs the cost of switching to the
10167 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
10168 N->getValueType(0) != MVT::f32)
10171 SDValue CondLHS = N->getOperand(0);
10172 SDValue CondRHS = N->getOperand(1);
10173 SDValue LHS = N->getOperand(2);
10174 SDValue RHS = N->getOperand(3);
10175 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
10177 unsigned Opcode = 0;
10179 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
10180 IsReversed = false; // x CC y ? x : y
10181 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
10182 IsReversed = true ; // x CC y ? y : x
10196 // If LHS is NaN, an ordered comparison will be false and the result will
10197 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
10198 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
10199 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
10200 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
10202 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
10203 // will return -0, so vmin can only be used for unsafe math or if one of
10204 // the operands is known to be nonzero.
10205 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
10206 !DAG.getTarget().Options.UnsafeFPMath &&
10207 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
10209 Opcode = IsReversed ? ISD::FMAXNAN : ISD::FMINNAN;
10218 // If LHS is NaN, an ordered comparison will be false and the result will
10219 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
10220 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
10221 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
10222 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
10224 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
10225 // will return +0, so vmax can only be used for unsafe math or if one of
10226 // the operands is known to be nonzero.
10227 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
10228 !DAG.getTarget().Options.UnsafeFPMath &&
10229 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
10231 Opcode = IsReversed ? ISD::FMINNAN : ISD::FMAXNAN;
10237 return DAG.getNode(Opcode, SDLoc(N), N->getValueType(0), LHS, RHS);
10240 /// PerformCMOVCombine - Target-specific DAG combining for ARMISD::CMOV.
10242 ARMTargetLowering::PerformCMOVCombine(SDNode *N, SelectionDAG &DAG) const {
10243 SDValue Cmp = N->getOperand(4);
10244 if (Cmp.getOpcode() != ARMISD::CMPZ)
10245 // Only looking at EQ and NE cases.
10248 EVT VT = N->getValueType(0);
10250 SDValue LHS = Cmp.getOperand(0);
10251 SDValue RHS = Cmp.getOperand(1);
10252 SDValue FalseVal = N->getOperand(0);
10253 SDValue TrueVal = N->getOperand(1);
10254 SDValue ARMcc = N->getOperand(2);
10255 ARMCC::CondCodes CC =
10256 (ARMCC::CondCodes)cast<ConstantSDNode>(ARMcc)->getZExtValue();
10274 /// FIXME: Turn this into a target neutral optimization?
10276 if (CC == ARMCC::NE && FalseVal == RHS && FalseVal != LHS) {
10277 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, TrueVal, ARMcc,
10278 N->getOperand(3), Cmp);
10279 } else if (CC == ARMCC::EQ && TrueVal == RHS) {
10281 SDValue NewCmp = getARMCmp(LHS, RHS, ISD::SETNE, ARMcc, DAG, dl);
10282 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, FalseVal, ARMcc,
10283 N->getOperand(3), NewCmp);
10286 if (Res.getNode()) {
10287 APInt KnownZero, KnownOne;
10288 DAG.computeKnownBits(SDValue(N,0), KnownZero, KnownOne);
10289 // Capture demanded bits information that would be otherwise lost.
10290 if (KnownZero == 0xfffffffe)
10291 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
10292 DAG.getValueType(MVT::i1));
10293 else if (KnownZero == 0xffffff00)
10294 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
10295 DAG.getValueType(MVT::i8));
10296 else if (KnownZero == 0xffff0000)
10297 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
10298 DAG.getValueType(MVT::i16));
10304 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
10305 DAGCombinerInfo &DCI) const {
10306 switch (N->getOpcode()) {
10308 case ISD::ADDC: return PerformADDCCombine(N, DCI, Subtarget);
10309 case ISD::ADD: return PerformADDCombine(N, DCI, Subtarget);
10310 case ISD::SUB: return PerformSUBCombine(N, DCI);
10311 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
10312 case ISD::OR: return PerformORCombine(N, DCI, Subtarget);
10313 case ISD::XOR: return PerformXORCombine(N, DCI, Subtarget);
10314 case ISD::AND: return PerformANDCombine(N, DCI, Subtarget);
10315 case ARMISD::BFI: return PerformBFICombine(N, DCI);
10316 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI, Subtarget);
10317 case ARMISD::VMOVDRR: return PerformVMOVDRRCombine(N, DCI.DAG);
10318 case ISD::STORE: return PerformSTORECombine(N, DCI);
10319 case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DCI, Subtarget);
10320 case ISD::INSERT_VECTOR_ELT: return PerformInsertEltCombine(N, DCI);
10321 case ISD::VECTOR_SHUFFLE: return PerformVECTOR_SHUFFLECombine(N, DCI.DAG);
10322 case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI);
10323 case ISD::FP_TO_SINT:
10324 case ISD::FP_TO_UINT: return PerformVCVTCombine(N, DCI, Subtarget);
10325 case ISD::FDIV: return PerformVDIVCombine(N, DCI, Subtarget);
10326 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
10329 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
10330 case ISD::SIGN_EXTEND:
10331 case ISD::ZERO_EXTEND:
10332 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
10333 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
10334 case ARMISD::CMOV: return PerformCMOVCombine(N, DCI.DAG);
10335 case ISD::LOAD: return PerformLOADCombine(N, DCI);
10336 case ARMISD::VLD2DUP:
10337 case ARMISD::VLD3DUP:
10338 case ARMISD::VLD4DUP:
10339 return PerformVLDCombine(N, DCI);
10340 case ARMISD::BUILD_VECTOR:
10341 return PerformARMBUILD_VECTORCombine(N, DCI);
10342 case ISD::INTRINSIC_VOID:
10343 case ISD::INTRINSIC_W_CHAIN:
10344 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
10345 case Intrinsic::arm_neon_vld1:
10346 case Intrinsic::arm_neon_vld2:
10347 case Intrinsic::arm_neon_vld3:
10348 case Intrinsic::arm_neon_vld4:
10349 case Intrinsic::arm_neon_vld2lane:
10350 case Intrinsic::arm_neon_vld3lane:
10351 case Intrinsic::arm_neon_vld4lane:
10352 case Intrinsic::arm_neon_vst1:
10353 case Intrinsic::arm_neon_vst2:
10354 case Intrinsic::arm_neon_vst3:
10355 case Intrinsic::arm_neon_vst4:
10356 case Intrinsic::arm_neon_vst2lane:
10357 case Intrinsic::arm_neon_vst3lane:
10358 case Intrinsic::arm_neon_vst4lane:
10359 return PerformVLDCombine(N, DCI);
10367 bool ARMTargetLowering::isDesirableToTransformToIntegerOp(unsigned Opc,
10369 return (VT == MVT::f32) && (Opc == ISD::LOAD || Opc == ISD::STORE);
10372 bool ARMTargetLowering::allowsMisalignedMemoryAccesses(EVT VT,
10375 bool *Fast) const {
10376 // The AllowsUnaliged flag models the SCTLR.A setting in ARM cpus
10377 bool AllowsUnaligned = Subtarget->allowsUnalignedMem();
10379 switch (VT.getSimpleVT().SimpleTy) {
10385 // Unaligned access can use (for example) LRDB, LRDH, LDR
10386 if (AllowsUnaligned) {
10388 *Fast = Subtarget->hasV7Ops();
10395 // For any little-endian targets with neon, we can support unaligned ld/st
10396 // of D and Q (e.g. {D0,D1}) registers by using vld1.i8/vst1.i8.
10397 // A big-endian target may also explicitly support unaligned accesses
10398 if (Subtarget->hasNEON() && (AllowsUnaligned || Subtarget->isLittle())) {
10408 static bool memOpAlign(unsigned DstAlign, unsigned SrcAlign,
10409 unsigned AlignCheck) {
10410 return ((SrcAlign == 0 || SrcAlign % AlignCheck == 0) &&
10411 (DstAlign == 0 || DstAlign % AlignCheck == 0));
10414 EVT ARMTargetLowering::getOptimalMemOpType(uint64_t Size,
10415 unsigned DstAlign, unsigned SrcAlign,
10416 bool IsMemset, bool ZeroMemset,
10418 MachineFunction &MF) const {
10419 const Function *F = MF.getFunction();
10421 // See if we can use NEON instructions for this...
10422 if ((!IsMemset || ZeroMemset) && Subtarget->hasNEON() &&
10423 !F->hasFnAttribute(Attribute::NoImplicitFloat)) {
10426 (memOpAlign(SrcAlign, DstAlign, 16) ||
10427 (allowsMisalignedMemoryAccesses(MVT::v2f64, 0, 1, &Fast) && Fast))) {
10429 } else if (Size >= 8 &&
10430 (memOpAlign(SrcAlign, DstAlign, 8) ||
10431 (allowsMisalignedMemoryAccesses(MVT::f64, 0, 1, &Fast) &&
10437 // Lowering to i32/i16 if the size permits.
10440 else if (Size >= 2)
10443 // Let the target-independent logic figure it out.
10447 bool ARMTargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
10448 if (Val.getOpcode() != ISD::LOAD)
10451 EVT VT1 = Val.getValueType();
10452 if (!VT1.isSimple() || !VT1.isInteger() ||
10453 !VT2.isSimple() || !VT2.isInteger())
10456 switch (VT1.getSimpleVT().SimpleTy) {
10461 // 8-bit and 16-bit loads implicitly zero-extend to 32-bits.
10468 bool ARMTargetLowering::isVectorLoadExtDesirable(SDValue ExtVal) const {
10469 EVT VT = ExtVal.getValueType();
10471 if (!isTypeLegal(VT))
10474 // Don't create a loadext if we can fold the extension into a wide/long
10476 // If there's more than one user instruction, the loadext is desirable no
10477 // matter what. There can be two uses by the same instruction.
10478 if (ExtVal->use_empty() ||
10479 !ExtVal->use_begin()->isOnlyUserOf(ExtVal.getNode()))
10482 SDNode *U = *ExtVal->use_begin();
10483 if ((U->getOpcode() == ISD::ADD || U->getOpcode() == ISD::SUB ||
10484 U->getOpcode() == ISD::SHL || U->getOpcode() == ARMISD::VSHL))
10490 bool ARMTargetLowering::allowTruncateForTailCall(Type *Ty1, Type *Ty2) const {
10491 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
10494 if (!isTypeLegal(EVT::getEVT(Ty1)))
10497 assert(Ty1->getPrimitiveSizeInBits() <= 64 && "i128 is probably not a noop");
10499 // Assuming the caller doesn't have a zeroext or signext return parameter,
10500 // truncation all the way down to i1 is valid.
10505 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
10509 unsigned Scale = 1;
10510 switch (VT.getSimpleVT().SimpleTy) {
10511 default: return false;
10526 if ((V & (Scale - 1)) != 0)
10529 return V == (V & ((1LL << 5) - 1));
10532 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
10533 const ARMSubtarget *Subtarget) {
10534 bool isNeg = false;
10540 switch (VT.getSimpleVT().SimpleTy) {
10541 default: return false;
10546 // + imm12 or - imm8
10548 return V == (V & ((1LL << 8) - 1));
10549 return V == (V & ((1LL << 12) - 1));
10552 // Same as ARM mode. FIXME: NEON?
10553 if (!Subtarget->hasVFP2())
10558 return V == (V & ((1LL << 8) - 1));
10562 /// isLegalAddressImmediate - Return true if the integer value can be used
10563 /// as the offset of the target addressing mode for load / store of the
10565 static bool isLegalAddressImmediate(int64_t V, EVT VT,
10566 const ARMSubtarget *Subtarget) {
10570 if (!VT.isSimple())
10573 if (Subtarget->isThumb1Only())
10574 return isLegalT1AddressImmediate(V, VT);
10575 else if (Subtarget->isThumb2())
10576 return isLegalT2AddressImmediate(V, VT, Subtarget);
10581 switch (VT.getSimpleVT().SimpleTy) {
10582 default: return false;
10587 return V == (V & ((1LL << 12) - 1));
10590 return V == (V & ((1LL << 8) - 1));
10593 if (!Subtarget->hasVFP2()) // FIXME: NEON?
10598 return V == (V & ((1LL << 8) - 1));
10602 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
10604 int Scale = AM.Scale;
10608 switch (VT.getSimpleVT().SimpleTy) {
10609 default: return false;
10617 Scale = Scale & ~1;
10618 return Scale == 2 || Scale == 4 || Scale == 8;
10621 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
10625 // Note, we allow "void" uses (basically, uses that aren't loads or
10626 // stores), because arm allows folding a scale into many arithmetic
10627 // operations. This should be made more precise and revisited later.
10629 // Allow r << imm, but the imm has to be a multiple of two.
10630 if (Scale & 1) return false;
10631 return isPowerOf2_32(Scale);
10635 /// isLegalAddressingMode - Return true if the addressing mode represented
10636 /// by AM is legal for this target, for a load/store of the specified type.
10637 bool ARMTargetLowering::isLegalAddressingMode(const DataLayout &DL,
10638 const AddrMode &AM, Type *Ty,
10639 unsigned AS) const {
10640 EVT VT = getValueType(DL, Ty, true);
10641 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
10644 // Can never fold addr of global into load/store.
10648 switch (AM.Scale) {
10649 case 0: // no scale reg, must be "r+i" or "r", or "i".
10652 if (Subtarget->isThumb1Only())
10656 // ARM doesn't support any R+R*scale+imm addr modes.
10660 if (!VT.isSimple())
10663 if (Subtarget->isThumb2())
10664 return isLegalT2ScaledAddressingMode(AM, VT);
10666 int Scale = AM.Scale;
10667 switch (VT.getSimpleVT().SimpleTy) {
10668 default: return false;
10672 if (Scale < 0) Scale = -Scale;
10676 return isPowerOf2_32(Scale & ~1);
10680 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
10685 // Note, we allow "void" uses (basically, uses that aren't loads or
10686 // stores), because arm allows folding a scale into many arithmetic
10687 // operations. This should be made more precise and revisited later.
10689 // Allow r << imm, but the imm has to be a multiple of two.
10690 if (Scale & 1) return false;
10691 return isPowerOf2_32(Scale);
10697 /// isLegalICmpImmediate - Return true if the specified immediate is legal
10698 /// icmp immediate, that is the target has icmp instructions which can compare
10699 /// a register against the immediate without having to materialize the
10700 /// immediate into a register.
10701 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
10702 // Thumb2 and ARM modes can use cmn for negative immediates.
10703 if (!Subtarget->isThumb())
10704 return ARM_AM::getSOImmVal(std::abs(Imm)) != -1;
10705 if (Subtarget->isThumb2())
10706 return ARM_AM::getT2SOImmVal(std::abs(Imm)) != -1;
10707 // Thumb1 doesn't have cmn, and only 8-bit immediates.
10708 return Imm >= 0 && Imm <= 255;
10711 /// isLegalAddImmediate - Return true if the specified immediate is a legal add
10712 /// *or sub* immediate, that is the target has add or sub instructions which can
10713 /// add a register with the immediate without having to materialize the
10714 /// immediate into a register.
10715 bool ARMTargetLowering::isLegalAddImmediate(int64_t Imm) const {
10716 // Same encoding for add/sub, just flip the sign.
10717 int64_t AbsImm = std::abs(Imm);
10718 if (!Subtarget->isThumb())
10719 return ARM_AM::getSOImmVal(AbsImm) != -1;
10720 if (Subtarget->isThumb2())
10721 return ARM_AM::getT2SOImmVal(AbsImm) != -1;
10722 // Thumb1 only has 8-bit unsigned immediate.
10723 return AbsImm >= 0 && AbsImm <= 255;
10726 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
10727 bool isSEXTLoad, SDValue &Base,
10728 SDValue &Offset, bool &isInc,
10729 SelectionDAG &DAG) {
10730 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
10733 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
10734 // AddressingMode 3
10735 Base = Ptr->getOperand(0);
10736 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
10737 int RHSC = (int)RHS->getZExtValue();
10738 if (RHSC < 0 && RHSC > -256) {
10739 assert(Ptr->getOpcode() == ISD::ADD);
10741 Offset = DAG.getConstant(-RHSC, SDLoc(Ptr), RHS->getValueType(0));
10745 isInc = (Ptr->getOpcode() == ISD::ADD);
10746 Offset = Ptr->getOperand(1);
10748 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
10749 // AddressingMode 2
10750 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
10751 int RHSC = (int)RHS->getZExtValue();
10752 if (RHSC < 0 && RHSC > -0x1000) {
10753 assert(Ptr->getOpcode() == ISD::ADD);
10755 Offset = DAG.getConstant(-RHSC, SDLoc(Ptr), RHS->getValueType(0));
10756 Base = Ptr->getOperand(0);
10761 if (Ptr->getOpcode() == ISD::ADD) {
10763 ARM_AM::ShiftOpc ShOpcVal=
10764 ARM_AM::getShiftOpcForNode(Ptr->getOperand(0).getOpcode());
10765 if (ShOpcVal != ARM_AM::no_shift) {
10766 Base = Ptr->getOperand(1);
10767 Offset = Ptr->getOperand(0);
10769 Base = Ptr->getOperand(0);
10770 Offset = Ptr->getOperand(1);
10775 isInc = (Ptr->getOpcode() == ISD::ADD);
10776 Base = Ptr->getOperand(0);
10777 Offset = Ptr->getOperand(1);
10781 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
10785 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
10786 bool isSEXTLoad, SDValue &Base,
10787 SDValue &Offset, bool &isInc,
10788 SelectionDAG &DAG) {
10789 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
10792 Base = Ptr->getOperand(0);
10793 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
10794 int RHSC = (int)RHS->getZExtValue();
10795 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
10796 assert(Ptr->getOpcode() == ISD::ADD);
10798 Offset = DAG.getConstant(-RHSC, SDLoc(Ptr), RHS->getValueType(0));
10800 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
10801 isInc = Ptr->getOpcode() == ISD::ADD;
10802 Offset = DAG.getConstant(RHSC, SDLoc(Ptr), RHS->getValueType(0));
10810 /// getPreIndexedAddressParts - returns true by value, base pointer and
10811 /// offset pointer and addressing mode by reference if the node's address
10812 /// can be legally represented as pre-indexed load / store address.
10814 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
10816 ISD::MemIndexedMode &AM,
10817 SelectionDAG &DAG) const {
10818 if (Subtarget->isThumb1Only())
10823 bool isSEXTLoad = false;
10824 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
10825 Ptr = LD->getBasePtr();
10826 VT = LD->getMemoryVT();
10827 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
10828 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
10829 Ptr = ST->getBasePtr();
10830 VT = ST->getMemoryVT();
10835 bool isLegal = false;
10836 if (Subtarget->isThumb2())
10837 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
10838 Offset, isInc, DAG);
10840 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
10841 Offset, isInc, DAG);
10845 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
10849 /// getPostIndexedAddressParts - returns true by value, base pointer and
10850 /// offset pointer and addressing mode by reference if this node can be
10851 /// combined with a load / store to form a post-indexed load / store.
10852 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
10855 ISD::MemIndexedMode &AM,
10856 SelectionDAG &DAG) const {
10857 if (Subtarget->isThumb1Only())
10862 bool isSEXTLoad = false;
10863 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
10864 VT = LD->getMemoryVT();
10865 Ptr = LD->getBasePtr();
10866 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
10867 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
10868 VT = ST->getMemoryVT();
10869 Ptr = ST->getBasePtr();
10874 bool isLegal = false;
10875 if (Subtarget->isThumb2())
10876 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
10879 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
10885 // Swap base ptr and offset to catch more post-index load / store when
10886 // it's legal. In Thumb2 mode, offset must be an immediate.
10887 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
10888 !Subtarget->isThumb2())
10889 std::swap(Base, Offset);
10891 // Post-indexed load / store update the base pointer.
10896 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
10900 void ARMTargetLowering::computeKnownBitsForTargetNode(const SDValue Op,
10903 const SelectionDAG &DAG,
10904 unsigned Depth) const {
10905 unsigned BitWidth = KnownOne.getBitWidth();
10906 KnownZero = KnownOne = APInt(BitWidth, 0);
10907 switch (Op.getOpcode()) {
10913 // These nodes' second result is a boolean
10914 if (Op.getResNo() == 0)
10916 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - 1);
10918 case ARMISD::CMOV: {
10919 // Bits are known zero/one if known on the LHS and RHS.
10920 DAG.computeKnownBits(Op.getOperand(0), KnownZero, KnownOne, Depth+1);
10921 if (KnownZero == 0 && KnownOne == 0) return;
10923 APInt KnownZeroRHS, KnownOneRHS;
10924 DAG.computeKnownBits(Op.getOperand(1), KnownZeroRHS, KnownOneRHS, Depth+1);
10925 KnownZero &= KnownZeroRHS;
10926 KnownOne &= KnownOneRHS;
10929 case ISD::INTRINSIC_W_CHAIN: {
10930 ConstantSDNode *CN = cast<ConstantSDNode>(Op->getOperand(1));
10931 Intrinsic::ID IntID = static_cast<Intrinsic::ID>(CN->getZExtValue());
10934 case Intrinsic::arm_ldaex:
10935 case Intrinsic::arm_ldrex: {
10936 EVT VT = cast<MemIntrinsicSDNode>(Op)->getMemoryVT();
10937 unsigned MemBits = VT.getScalarType().getSizeInBits();
10938 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - MemBits);
10946 //===----------------------------------------------------------------------===//
10947 // ARM Inline Assembly Support
10948 //===----------------------------------------------------------------------===//
10950 bool ARMTargetLowering::ExpandInlineAsm(CallInst *CI) const {
10951 // Looking for "rev" which is V6+.
10952 if (!Subtarget->hasV6Ops())
10955 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
10956 std::string AsmStr = IA->getAsmString();
10957 SmallVector<StringRef, 4> AsmPieces;
10958 SplitString(AsmStr, AsmPieces, ";\n");
10960 switch (AsmPieces.size()) {
10961 default: return false;
10963 AsmStr = AsmPieces[0];
10965 SplitString(AsmStr, AsmPieces, " \t,");
10968 if (AsmPieces.size() == 3 &&
10969 AsmPieces[0] == "rev" && AsmPieces[1] == "$0" && AsmPieces[2] == "$1" &&
10970 IA->getConstraintString().compare(0, 4, "=l,l") == 0) {
10971 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
10972 if (Ty && Ty->getBitWidth() == 32)
10973 return IntrinsicLowering::LowerToByteSwap(CI);
10981 /// getConstraintType - Given a constraint letter, return the type of
10982 /// constraint it is for this target.
10983 ARMTargetLowering::ConstraintType
10984 ARMTargetLowering::getConstraintType(StringRef Constraint) const {
10985 if (Constraint.size() == 1) {
10986 switch (Constraint[0]) {
10988 case 'l': return C_RegisterClass;
10989 case 'w': return C_RegisterClass;
10990 case 'h': return C_RegisterClass;
10991 case 'x': return C_RegisterClass;
10992 case 't': return C_RegisterClass;
10993 case 'j': return C_Other; // Constant for movw.
10994 // An address with a single base register. Due to the way we
10995 // currently handle addresses it is the same as an 'r' memory constraint.
10996 case 'Q': return C_Memory;
10998 } else if (Constraint.size() == 2) {
10999 switch (Constraint[0]) {
11001 // All 'U+' constraints are addresses.
11002 case 'U': return C_Memory;
11005 return TargetLowering::getConstraintType(Constraint);
11008 /// Examine constraint type and operand type and determine a weight value.
11009 /// This object must already have been set up with the operand type
11010 /// and the current alternative constraint selected.
11011 TargetLowering::ConstraintWeight
11012 ARMTargetLowering::getSingleConstraintMatchWeight(
11013 AsmOperandInfo &info, const char *constraint) const {
11014 ConstraintWeight weight = CW_Invalid;
11015 Value *CallOperandVal = info.CallOperandVal;
11016 // If we don't have a value, we can't do a match,
11017 // but allow it at the lowest weight.
11018 if (!CallOperandVal)
11020 Type *type = CallOperandVal->getType();
11021 // Look at the constraint type.
11022 switch (*constraint) {
11024 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
11027 if (type->isIntegerTy()) {
11028 if (Subtarget->isThumb())
11029 weight = CW_SpecificReg;
11031 weight = CW_Register;
11035 if (type->isFloatingPointTy())
11036 weight = CW_Register;
11042 typedef std::pair<unsigned, const TargetRegisterClass*> RCPair;
11043 RCPair ARMTargetLowering::getRegForInlineAsmConstraint(
11044 const TargetRegisterInfo *TRI, StringRef Constraint, MVT VT) const {
11045 if (Constraint.size() == 1) {
11046 // GCC ARM Constraint Letters
11047 switch (Constraint[0]) {
11048 case 'l': // Low regs or general regs.
11049 if (Subtarget->isThumb())
11050 return RCPair(0U, &ARM::tGPRRegClass);
11051 return RCPair(0U, &ARM::GPRRegClass);
11052 case 'h': // High regs or no regs.
11053 if (Subtarget->isThumb())
11054 return RCPair(0U, &ARM::hGPRRegClass);
11057 if (Subtarget->isThumb1Only())
11058 return RCPair(0U, &ARM::tGPRRegClass);
11059 return RCPair(0U, &ARM::GPRRegClass);
11061 if (VT == MVT::Other)
11063 if (VT == MVT::f32)
11064 return RCPair(0U, &ARM::SPRRegClass);
11065 if (VT.getSizeInBits() == 64)
11066 return RCPair(0U, &ARM::DPRRegClass);
11067 if (VT.getSizeInBits() == 128)
11068 return RCPair(0U, &ARM::QPRRegClass);
11071 if (VT == MVT::Other)
11073 if (VT == MVT::f32)
11074 return RCPair(0U, &ARM::SPR_8RegClass);
11075 if (VT.getSizeInBits() == 64)
11076 return RCPair(0U, &ARM::DPR_8RegClass);
11077 if (VT.getSizeInBits() == 128)
11078 return RCPair(0U, &ARM::QPR_8RegClass);
11081 if (VT == MVT::f32)
11082 return RCPair(0U, &ARM::SPRRegClass);
11086 if (StringRef("{cc}").equals_lower(Constraint))
11087 return std::make_pair(unsigned(ARM::CPSR), &ARM::CCRRegClass);
11089 return TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
11092 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
11093 /// vector. If it is invalid, don't add anything to Ops.
11094 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
11095 std::string &Constraint,
11096 std::vector<SDValue>&Ops,
11097 SelectionDAG &DAG) const {
11100 // Currently only support length 1 constraints.
11101 if (Constraint.length() != 1) return;
11103 char ConstraintLetter = Constraint[0];
11104 switch (ConstraintLetter) {
11107 case 'I': case 'J': case 'K': case 'L':
11108 case 'M': case 'N': case 'O':
11109 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
11113 int64_t CVal64 = C->getSExtValue();
11114 int CVal = (int) CVal64;
11115 // None of these constraints allow values larger than 32 bits. Check
11116 // that the value fits in an int.
11117 if (CVal != CVal64)
11120 switch (ConstraintLetter) {
11122 // Constant suitable for movw, must be between 0 and
11124 if (Subtarget->hasV6T2Ops())
11125 if (CVal >= 0 && CVal <= 65535)
11129 if (Subtarget->isThumb1Only()) {
11130 // This must be a constant between 0 and 255, for ADD
11132 if (CVal >= 0 && CVal <= 255)
11134 } else if (Subtarget->isThumb2()) {
11135 // A constant that can be used as an immediate value in a
11136 // data-processing instruction.
11137 if (ARM_AM::getT2SOImmVal(CVal) != -1)
11140 // A constant that can be used as an immediate value in a
11141 // data-processing instruction.
11142 if (ARM_AM::getSOImmVal(CVal) != -1)
11148 if (Subtarget->isThumb()) { // FIXME thumb2
11149 // This must be a constant between -255 and -1, for negated ADD
11150 // immediates. This can be used in GCC with an "n" modifier that
11151 // prints the negated value, for use with SUB instructions. It is
11152 // not useful otherwise but is implemented for compatibility.
11153 if (CVal >= -255 && CVal <= -1)
11156 // This must be a constant between -4095 and 4095. It is not clear
11157 // what this constraint is intended for. Implemented for
11158 // compatibility with GCC.
11159 if (CVal >= -4095 && CVal <= 4095)
11165 if (Subtarget->isThumb1Only()) {
11166 // A 32-bit value where only one byte has a nonzero value. Exclude
11167 // zero to match GCC. This constraint is used by GCC internally for
11168 // constants that can be loaded with a move/shift combination.
11169 // It is not useful otherwise but is implemented for compatibility.
11170 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
11172 } else if (Subtarget->isThumb2()) {
11173 // A constant whose bitwise inverse can be used as an immediate
11174 // value in a data-processing instruction. This can be used in GCC
11175 // with a "B" modifier that prints the inverted value, for use with
11176 // BIC and MVN instructions. It is not useful otherwise but is
11177 // implemented for compatibility.
11178 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
11181 // A constant whose bitwise inverse can be used as an immediate
11182 // value in a data-processing instruction. This can be used in GCC
11183 // with a "B" modifier that prints the inverted value, for use with
11184 // BIC and MVN instructions. It is not useful otherwise but is
11185 // implemented for compatibility.
11186 if (ARM_AM::getSOImmVal(~CVal) != -1)
11192 if (Subtarget->isThumb1Only()) {
11193 // This must be a constant between -7 and 7,
11194 // for 3-operand ADD/SUB immediate instructions.
11195 if (CVal >= -7 && CVal < 7)
11197 } else if (Subtarget->isThumb2()) {
11198 // A constant whose negation can be used as an immediate value in a
11199 // data-processing instruction. This can be used in GCC with an "n"
11200 // modifier that prints the negated value, for use with SUB
11201 // instructions. It is not useful otherwise but is implemented for
11203 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
11206 // A constant whose negation can be used as an immediate value in a
11207 // data-processing instruction. This can be used in GCC with an "n"
11208 // modifier that prints the negated value, for use with SUB
11209 // instructions. It is not useful otherwise but is implemented for
11211 if (ARM_AM::getSOImmVal(-CVal) != -1)
11217 if (Subtarget->isThumb()) { // FIXME thumb2
11218 // This must be a multiple of 4 between 0 and 1020, for
11219 // ADD sp + immediate.
11220 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
11223 // A power of two or a constant between 0 and 32. This is used in
11224 // GCC for the shift amount on shifted register operands, but it is
11225 // useful in general for any shift amounts.
11226 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
11232 if (Subtarget->isThumb()) { // FIXME thumb2
11233 // This must be a constant between 0 and 31, for shift amounts.
11234 if (CVal >= 0 && CVal <= 31)
11240 if (Subtarget->isThumb()) { // FIXME thumb2
11241 // This must be a multiple of 4 between -508 and 508, for
11242 // ADD/SUB sp = sp + immediate.
11243 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
11248 Result = DAG.getTargetConstant(CVal, SDLoc(Op), Op.getValueType());
11252 if (Result.getNode()) {
11253 Ops.push_back(Result);
11256 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
11259 SDValue ARMTargetLowering::LowerDivRem(SDValue Op, SelectionDAG &DAG) const {
11260 assert((Subtarget->isTargetAEABI() || Subtarget->isTargetAndroid()) &&
11261 "Register-based DivRem lowering only");
11262 unsigned Opcode = Op->getOpcode();
11263 assert((Opcode == ISD::SDIVREM || Opcode == ISD::UDIVREM) &&
11264 "Invalid opcode for Div/Rem lowering");
11265 bool isSigned = (Opcode == ISD::SDIVREM);
11266 EVT VT = Op->getValueType(0);
11267 Type *Ty = VT.getTypeForEVT(*DAG.getContext());
11270 switch (VT.getSimpleVT().SimpleTy) {
11271 default: llvm_unreachable("Unexpected request for libcall!");
11272 case MVT::i8: LC = isSigned ? RTLIB::SDIVREM_I8 : RTLIB::UDIVREM_I8; break;
11273 case MVT::i16: LC = isSigned ? RTLIB::SDIVREM_I16 : RTLIB::UDIVREM_I16; break;
11274 case MVT::i32: LC = isSigned ? RTLIB::SDIVREM_I32 : RTLIB::UDIVREM_I32; break;
11275 case MVT::i64: LC = isSigned ? RTLIB::SDIVREM_I64 : RTLIB::UDIVREM_I64; break;
11278 SDValue InChain = DAG.getEntryNode();
11280 TargetLowering::ArgListTy Args;
11281 TargetLowering::ArgListEntry Entry;
11282 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i) {
11283 EVT ArgVT = Op->getOperand(i).getValueType();
11284 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
11285 Entry.Node = Op->getOperand(i);
11287 Entry.isSExt = isSigned;
11288 Entry.isZExt = !isSigned;
11289 Args.push_back(Entry);
11292 SDValue Callee = DAG.getExternalSymbol(getLibcallName(LC),
11293 getPointerTy(DAG.getDataLayout()));
11295 Type *RetTy = (Type*)StructType::get(Ty, Ty, nullptr);
11298 TargetLowering::CallLoweringInfo CLI(DAG);
11299 CLI.setDebugLoc(dl).setChain(InChain)
11300 .setCallee(getLibcallCallingConv(LC), RetTy, Callee, std::move(Args), 0)
11301 .setInRegister().setSExtResult(isSigned).setZExtResult(!isSigned);
11303 std::pair<SDValue, SDValue> CallInfo = LowerCallTo(CLI);
11304 return CallInfo.first;
11308 ARMTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const {
11309 assert(Subtarget->isTargetWindows() && "unsupported target platform");
11313 SDValue Chain = Op.getOperand(0);
11314 SDValue Size = Op.getOperand(1);
11316 SDValue Words = DAG.getNode(ISD::SRL, DL, MVT::i32, Size,
11317 DAG.getConstant(2, DL, MVT::i32));
11320 Chain = DAG.getCopyToReg(Chain, DL, ARM::R4, Words, Flag);
11321 Flag = Chain.getValue(1);
11323 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
11324 Chain = DAG.getNode(ARMISD::WIN__CHKSTK, DL, NodeTys, Chain, Flag);
11326 SDValue NewSP = DAG.getCopyFromReg(Chain, DL, ARM::SP, MVT::i32);
11327 Chain = NewSP.getValue(1);
11329 SDValue Ops[2] = { NewSP, Chain };
11330 return DAG.getMergeValues(Ops, DL);
11333 SDValue ARMTargetLowering::LowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) const {
11334 assert(Op.getValueType() == MVT::f64 && Subtarget->isFPOnlySP() &&
11335 "Unexpected type for custom-lowering FP_EXTEND");
11338 LC = RTLIB::getFPEXT(Op.getOperand(0).getValueType(), Op.getValueType());
11340 SDValue SrcVal = Op.getOperand(0);
11341 return makeLibCall(DAG, LC, Op.getValueType(), &SrcVal, 1,
11342 /*isSigned*/ false, SDLoc(Op)).first;
11345 SDValue ARMTargetLowering::LowerFP_ROUND(SDValue Op, SelectionDAG &DAG) const {
11346 assert(Op.getOperand(0).getValueType() == MVT::f64 &&
11347 Subtarget->isFPOnlySP() &&
11348 "Unexpected type for custom-lowering FP_ROUND");
11351 LC = RTLIB::getFPROUND(Op.getOperand(0).getValueType(), Op.getValueType());
11353 SDValue SrcVal = Op.getOperand(0);
11354 return makeLibCall(DAG, LC, Op.getValueType(), &SrcVal, 1,
11355 /*isSigned*/ false, SDLoc(Op)).first;
11359 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
11360 // The ARM target isn't yet aware of offsets.
11364 bool ARM::isBitFieldInvertedMask(unsigned v) {
11365 if (v == 0xffffffff)
11368 // there can be 1's on either or both "outsides", all the "inside"
11369 // bits must be 0's
11370 return isShiftedMask_32(~v);
11373 /// isFPImmLegal - Returns true if the target can instruction select the
11374 /// specified FP immediate natively. If false, the legalizer will
11375 /// materialize the FP immediate as a load from a constant pool.
11376 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
11377 if (!Subtarget->hasVFP3())
11379 if (VT == MVT::f32)
11380 return ARM_AM::getFP32Imm(Imm) != -1;
11381 if (VT == MVT::f64 && !Subtarget->isFPOnlySP())
11382 return ARM_AM::getFP64Imm(Imm) != -1;
11386 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
11387 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
11388 /// specified in the intrinsic calls.
11389 bool ARMTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
11391 unsigned Intrinsic) const {
11392 switch (Intrinsic) {
11393 case Intrinsic::arm_neon_vld1:
11394 case Intrinsic::arm_neon_vld2:
11395 case Intrinsic::arm_neon_vld3:
11396 case Intrinsic::arm_neon_vld4:
11397 case Intrinsic::arm_neon_vld2lane:
11398 case Intrinsic::arm_neon_vld3lane:
11399 case Intrinsic::arm_neon_vld4lane: {
11400 Info.opc = ISD::INTRINSIC_W_CHAIN;
11401 // Conservatively set memVT to the entire set of vectors loaded.
11402 auto &DL = I.getCalledFunction()->getParent()->getDataLayout();
11403 uint64_t NumElts = DL.getTypeAllocSize(I.getType()) / 8;
11404 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
11405 Info.ptrVal = I.getArgOperand(0);
11407 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
11408 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
11409 Info.vol = false; // volatile loads with NEON intrinsics not supported
11410 Info.readMem = true;
11411 Info.writeMem = false;
11414 case Intrinsic::arm_neon_vst1:
11415 case Intrinsic::arm_neon_vst2:
11416 case Intrinsic::arm_neon_vst3:
11417 case Intrinsic::arm_neon_vst4:
11418 case Intrinsic::arm_neon_vst2lane:
11419 case Intrinsic::arm_neon_vst3lane:
11420 case Intrinsic::arm_neon_vst4lane: {
11421 Info.opc = ISD::INTRINSIC_VOID;
11422 // Conservatively set memVT to the entire set of vectors stored.
11423 auto &DL = I.getCalledFunction()->getParent()->getDataLayout();
11424 unsigned NumElts = 0;
11425 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
11426 Type *ArgTy = I.getArgOperand(ArgI)->getType();
11427 if (!ArgTy->isVectorTy())
11429 NumElts += DL.getTypeAllocSize(ArgTy) / 8;
11431 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
11432 Info.ptrVal = I.getArgOperand(0);
11434 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
11435 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
11436 Info.vol = false; // volatile stores with NEON intrinsics not supported
11437 Info.readMem = false;
11438 Info.writeMem = true;
11441 case Intrinsic::arm_ldaex:
11442 case Intrinsic::arm_ldrex: {
11443 auto &DL = I.getCalledFunction()->getParent()->getDataLayout();
11444 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(0)->getType());
11445 Info.opc = ISD::INTRINSIC_W_CHAIN;
11446 Info.memVT = MVT::getVT(PtrTy->getElementType());
11447 Info.ptrVal = I.getArgOperand(0);
11449 Info.align = DL.getABITypeAlignment(PtrTy->getElementType());
11451 Info.readMem = true;
11452 Info.writeMem = false;
11455 case Intrinsic::arm_stlex:
11456 case Intrinsic::arm_strex: {
11457 auto &DL = I.getCalledFunction()->getParent()->getDataLayout();
11458 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(1)->getType());
11459 Info.opc = ISD::INTRINSIC_W_CHAIN;
11460 Info.memVT = MVT::getVT(PtrTy->getElementType());
11461 Info.ptrVal = I.getArgOperand(1);
11463 Info.align = DL.getABITypeAlignment(PtrTy->getElementType());
11465 Info.readMem = false;
11466 Info.writeMem = true;
11469 case Intrinsic::arm_stlexd:
11470 case Intrinsic::arm_strexd: {
11471 Info.opc = ISD::INTRINSIC_W_CHAIN;
11472 Info.memVT = MVT::i64;
11473 Info.ptrVal = I.getArgOperand(2);
11477 Info.readMem = false;
11478 Info.writeMem = true;
11481 case Intrinsic::arm_ldaexd:
11482 case Intrinsic::arm_ldrexd: {
11483 Info.opc = ISD::INTRINSIC_W_CHAIN;
11484 Info.memVT = MVT::i64;
11485 Info.ptrVal = I.getArgOperand(0);
11489 Info.readMem = true;
11490 Info.writeMem = false;
11500 /// \brief Returns true if it is beneficial to convert a load of a constant
11501 /// to just the constant itself.
11502 bool ARMTargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
11504 assert(Ty->isIntegerTy());
11506 unsigned Bits = Ty->getPrimitiveSizeInBits();
11507 if (Bits == 0 || Bits > 32)
11512 bool ARMTargetLowering::hasLoadLinkedStoreConditional() const { return true; }
11514 Instruction* ARMTargetLowering::makeDMB(IRBuilder<> &Builder,
11515 ARM_MB::MemBOpt Domain) const {
11516 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
11518 // First, if the target has no DMB, see what fallback we can use.
11519 if (!Subtarget->hasDataBarrier()) {
11520 // Some ARMv6 cpus can support data barriers with an mcr instruction.
11521 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
11523 if (Subtarget->hasV6Ops() && !Subtarget->isThumb()) {
11524 Function *MCR = llvm::Intrinsic::getDeclaration(M, Intrinsic::arm_mcr);
11525 Value* args[6] = {Builder.getInt32(15), Builder.getInt32(0),
11526 Builder.getInt32(0), Builder.getInt32(7),
11527 Builder.getInt32(10), Builder.getInt32(5)};
11528 return Builder.CreateCall(MCR, args);
11530 // Instead of using barriers, atomic accesses on these subtargets use
11532 llvm_unreachable("makeDMB on a target so old that it has no barriers");
11535 Function *DMB = llvm::Intrinsic::getDeclaration(M, Intrinsic::arm_dmb);
11536 // Only a full system barrier exists in the M-class architectures.
11537 Domain = Subtarget->isMClass() ? ARM_MB::SY : Domain;
11538 Constant *CDomain = Builder.getInt32(Domain);
11539 return Builder.CreateCall(DMB, CDomain);
11543 // Based on http://www.cl.cam.ac.uk/~pes20/cpp/cpp0xmappings.html
11544 Instruction* ARMTargetLowering::emitLeadingFence(IRBuilder<> &Builder,
11545 AtomicOrdering Ord, bool IsStore,
11546 bool IsLoad) const {
11547 if (!getInsertFencesForAtomic())
11553 llvm_unreachable("Invalid fence: unordered/non-atomic");
11556 return nullptr; // Nothing to do
11557 case SequentiallyConsistent:
11559 return nullptr; // Nothing to do
11562 case AcquireRelease:
11563 if (Subtarget->isSwift())
11564 return makeDMB(Builder, ARM_MB::ISHST);
11565 // FIXME: add a comment with a link to documentation justifying this.
11567 return makeDMB(Builder, ARM_MB::ISH);
11569 llvm_unreachable("Unknown fence ordering in emitLeadingFence");
11572 Instruction* ARMTargetLowering::emitTrailingFence(IRBuilder<> &Builder,
11573 AtomicOrdering Ord, bool IsStore,
11574 bool IsLoad) const {
11575 if (!getInsertFencesForAtomic())
11581 llvm_unreachable("Invalid fence: unordered/not-atomic");
11584 return nullptr; // Nothing to do
11586 case AcquireRelease:
11587 case SequentiallyConsistent:
11588 return makeDMB(Builder, ARM_MB::ISH);
11590 llvm_unreachable("Unknown fence ordering in emitTrailingFence");
11593 // Loads and stores less than 64-bits are already atomic; ones above that
11594 // are doomed anyway, so defer to the default libcall and blame the OS when
11595 // things go wrong. Cortex M doesn't have ldrexd/strexd though, so don't emit
11596 // anything for those.
11597 bool ARMTargetLowering::shouldExpandAtomicStoreInIR(StoreInst *SI) const {
11598 unsigned Size = SI->getValueOperand()->getType()->getPrimitiveSizeInBits();
11599 return (Size == 64) && !Subtarget->isMClass();
11602 // Loads and stores less than 64-bits are already atomic; ones above that
11603 // are doomed anyway, so defer to the default libcall and blame the OS when
11604 // things go wrong. Cortex M doesn't have ldrexd/strexd though, so don't emit
11605 // anything for those.
11606 // FIXME: ldrd and strd are atomic if the CPU has LPAE (e.g. A15 has that
11607 // guarantee, see DDI0406C ARM architecture reference manual,
11608 // sections A8.8.72-74 LDRD)
11609 bool ARMTargetLowering::shouldExpandAtomicLoadInIR(LoadInst *LI) const {
11610 unsigned Size = LI->getType()->getPrimitiveSizeInBits();
11611 return (Size == 64) && !Subtarget->isMClass();
11614 // For the real atomic operations, we have ldrex/strex up to 32 bits,
11615 // and up to 64 bits on the non-M profiles
11616 TargetLoweringBase::AtomicRMWExpansionKind
11617 ARMTargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
11618 unsigned Size = AI->getType()->getPrimitiveSizeInBits();
11619 return (Size <= (Subtarget->isMClass() ? 32U : 64U))
11620 ? AtomicRMWExpansionKind::LLSC
11621 : AtomicRMWExpansionKind::None;
11624 // This has so far only been implemented for MachO.
11625 bool ARMTargetLowering::useLoadStackGuardNode() const {
11626 return Subtarget->isTargetMachO();
11629 bool ARMTargetLowering::canCombineStoreAndExtract(Type *VectorTy, Value *Idx,
11630 unsigned &Cost) const {
11631 // If we do not have NEON, vector types are not natively supported.
11632 if (!Subtarget->hasNEON())
11635 // Floating point values and vector values map to the same register file.
11636 // Therefore, although we could do a store extract of a vector type, this is
11637 // better to leave at float as we have more freedom in the addressing mode for
11639 if (VectorTy->isFPOrFPVectorTy())
11642 // If the index is unknown at compile time, this is very expensive to lower
11643 // and it is not possible to combine the store with the extract.
11644 if (!isa<ConstantInt>(Idx))
11647 assert(VectorTy->isVectorTy() && "VectorTy is not a vector type");
11648 unsigned BitWidth = cast<VectorType>(VectorTy)->getBitWidth();
11649 // We can do a store + vector extract on any vector that fits perfectly in a D
11651 if (BitWidth == 64 || BitWidth == 128) {
11658 Value *ARMTargetLowering::emitLoadLinked(IRBuilder<> &Builder, Value *Addr,
11659 AtomicOrdering Ord) const {
11660 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
11661 Type *ValTy = cast<PointerType>(Addr->getType())->getElementType();
11662 bool IsAcquire = isAtLeastAcquire(Ord);
11664 // Since i64 isn't legal and intrinsics don't get type-lowered, the ldrexd
11665 // intrinsic must return {i32, i32} and we have to recombine them into a
11666 // single i64 here.
11667 if (ValTy->getPrimitiveSizeInBits() == 64) {
11668 Intrinsic::ID Int =
11669 IsAcquire ? Intrinsic::arm_ldaexd : Intrinsic::arm_ldrexd;
11670 Function *Ldrex = llvm::Intrinsic::getDeclaration(M, Int);
11672 Addr = Builder.CreateBitCast(Addr, Type::getInt8PtrTy(M->getContext()));
11673 Value *LoHi = Builder.CreateCall(Ldrex, Addr, "lohi");
11675 Value *Lo = Builder.CreateExtractValue(LoHi, 0, "lo");
11676 Value *Hi = Builder.CreateExtractValue(LoHi, 1, "hi");
11677 if (!Subtarget->isLittle())
11678 std::swap (Lo, Hi);
11679 Lo = Builder.CreateZExt(Lo, ValTy, "lo64");
11680 Hi = Builder.CreateZExt(Hi, ValTy, "hi64");
11681 return Builder.CreateOr(
11682 Lo, Builder.CreateShl(Hi, ConstantInt::get(ValTy, 32)), "val64");
11685 Type *Tys[] = { Addr->getType() };
11686 Intrinsic::ID Int = IsAcquire ? Intrinsic::arm_ldaex : Intrinsic::arm_ldrex;
11687 Function *Ldrex = llvm::Intrinsic::getDeclaration(M, Int, Tys);
11689 return Builder.CreateTruncOrBitCast(
11690 Builder.CreateCall(Ldrex, Addr),
11691 cast<PointerType>(Addr->getType())->getElementType());
11694 Value *ARMTargetLowering::emitStoreConditional(IRBuilder<> &Builder, Value *Val,
11696 AtomicOrdering Ord) const {
11697 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
11698 bool IsRelease = isAtLeastRelease(Ord);
11700 // Since the intrinsics must have legal type, the i64 intrinsics take two
11701 // parameters: "i32, i32". We must marshal Val into the appropriate form
11702 // before the call.
11703 if (Val->getType()->getPrimitiveSizeInBits() == 64) {
11704 Intrinsic::ID Int =
11705 IsRelease ? Intrinsic::arm_stlexd : Intrinsic::arm_strexd;
11706 Function *Strex = Intrinsic::getDeclaration(M, Int);
11707 Type *Int32Ty = Type::getInt32Ty(M->getContext());
11709 Value *Lo = Builder.CreateTrunc(Val, Int32Ty, "lo");
11710 Value *Hi = Builder.CreateTrunc(Builder.CreateLShr(Val, 32), Int32Ty, "hi");
11711 if (!Subtarget->isLittle())
11712 std::swap (Lo, Hi);
11713 Addr = Builder.CreateBitCast(Addr, Type::getInt8PtrTy(M->getContext()));
11714 return Builder.CreateCall(Strex, {Lo, Hi, Addr});
11717 Intrinsic::ID Int = IsRelease ? Intrinsic::arm_stlex : Intrinsic::arm_strex;
11718 Type *Tys[] = { Addr->getType() };
11719 Function *Strex = Intrinsic::getDeclaration(M, Int, Tys);
11721 return Builder.CreateCall(
11722 Strex, {Builder.CreateZExtOrBitCast(
11723 Val, Strex->getFunctionType()->getParamType(0)),
11727 /// \brief Lower an interleaved load into a vldN intrinsic.
11729 /// E.g. Lower an interleaved load (Factor = 2):
11730 /// %wide.vec = load <8 x i32>, <8 x i32>* %ptr, align 4
11731 /// %v0 = shuffle %wide.vec, undef, <0, 2, 4, 6> ; Extract even elements
11732 /// %v1 = shuffle %wide.vec, undef, <1, 3, 5, 7> ; Extract odd elements
11735 /// %vld2 = { <4 x i32>, <4 x i32> } call llvm.arm.neon.vld2(%ptr, 4)
11736 /// %vec0 = extractelement { <4 x i32>, <4 x i32> } %vld2, i32 0
11737 /// %vec1 = extractelement { <4 x i32>, <4 x i32> } %vld2, i32 1
11738 bool ARMTargetLowering::lowerInterleavedLoad(
11739 LoadInst *LI, ArrayRef<ShuffleVectorInst *> Shuffles,
11740 ArrayRef<unsigned> Indices, unsigned Factor) const {
11741 assert(Factor >= 2 && Factor <= getMaxSupportedInterleaveFactor() &&
11742 "Invalid interleave factor");
11743 assert(!Shuffles.empty() && "Empty shufflevector input");
11744 assert(Shuffles.size() == Indices.size() &&
11745 "Unmatched number of shufflevectors and indices");
11747 VectorType *VecTy = Shuffles[0]->getType();
11748 Type *EltTy = VecTy->getVectorElementType();
11750 const DataLayout &DL = LI->getModule()->getDataLayout();
11751 unsigned VecSize = DL.getTypeAllocSizeInBits(VecTy);
11752 bool EltIs64Bits = DL.getTypeAllocSizeInBits(EltTy) == 64;
11754 // Skip illegal vector types and vector types of i64/f64 element (vldN doesn't
11755 // support i64/f64 element).
11756 if ((VecSize != 64 && VecSize != 128) || EltIs64Bits)
11759 // A pointer vector can not be the return type of the ldN intrinsics. Need to
11760 // load integer vectors first and then convert to pointer vectors.
11761 if (EltTy->isPointerTy())
11763 VectorType::get(DL.getIntPtrType(EltTy), VecTy->getVectorNumElements());
11765 static const Intrinsic::ID LoadInts[3] = {Intrinsic::arm_neon_vld2,
11766 Intrinsic::arm_neon_vld3,
11767 Intrinsic::arm_neon_vld4};
11769 Function *VldnFunc =
11770 Intrinsic::getDeclaration(LI->getModule(), LoadInts[Factor - 2], VecTy);
11772 IRBuilder<> Builder(LI);
11773 SmallVector<Value *, 2> Ops;
11775 Type *Int8Ptr = Builder.getInt8PtrTy(LI->getPointerAddressSpace());
11776 Ops.push_back(Builder.CreateBitCast(LI->getPointerOperand(), Int8Ptr));
11777 Ops.push_back(Builder.getInt32(LI->getAlignment()));
11779 CallInst *VldN = Builder.CreateCall(VldnFunc, Ops, "vldN");
11781 // Replace uses of each shufflevector with the corresponding vector loaded
11783 for (unsigned i = 0; i < Shuffles.size(); i++) {
11784 ShuffleVectorInst *SV = Shuffles[i];
11785 unsigned Index = Indices[i];
11787 Value *SubVec = Builder.CreateExtractValue(VldN, Index);
11789 // Convert the integer vector to pointer vector if the element is pointer.
11790 if (EltTy->isPointerTy())
11791 SubVec = Builder.CreateIntToPtr(SubVec, SV->getType());
11793 SV->replaceAllUsesWith(SubVec);
11799 /// \brief Get a mask consisting of sequential integers starting from \p Start.
11801 /// I.e. <Start, Start + 1, ..., Start + NumElts - 1>
11802 static Constant *getSequentialMask(IRBuilder<> &Builder, unsigned Start,
11803 unsigned NumElts) {
11804 SmallVector<Constant *, 16> Mask;
11805 for (unsigned i = 0; i < NumElts; i++)
11806 Mask.push_back(Builder.getInt32(Start + i));
11808 return ConstantVector::get(Mask);
11811 /// \brief Lower an interleaved store into a vstN intrinsic.
11813 /// E.g. Lower an interleaved store (Factor = 3):
11814 /// %i.vec = shuffle <8 x i32> %v0, <8 x i32> %v1,
11815 /// <0, 4, 8, 1, 5, 9, 2, 6, 10, 3, 7, 11>
11816 /// store <12 x i32> %i.vec, <12 x i32>* %ptr, align 4
11819 /// %sub.v0 = shuffle <8 x i32> %v0, <8 x i32> v1, <0, 1, 2, 3>
11820 /// %sub.v1 = shuffle <8 x i32> %v0, <8 x i32> v1, <4, 5, 6, 7>
11821 /// %sub.v2 = shuffle <8 x i32> %v0, <8 x i32> v1, <8, 9, 10, 11>
11822 /// call void llvm.arm.neon.vst3(%ptr, %sub.v0, %sub.v1, %sub.v2, 4)
11824 /// Note that the new shufflevectors will be removed and we'll only generate one
11825 /// vst3 instruction in CodeGen.
11826 bool ARMTargetLowering::lowerInterleavedStore(StoreInst *SI,
11827 ShuffleVectorInst *SVI,
11828 unsigned Factor) const {
11829 assert(Factor >= 2 && Factor <= getMaxSupportedInterleaveFactor() &&
11830 "Invalid interleave factor");
11832 VectorType *VecTy = SVI->getType();
11833 assert(VecTy->getVectorNumElements() % Factor == 0 &&
11834 "Invalid interleaved store");
11836 unsigned NumSubElts = VecTy->getVectorNumElements() / Factor;
11837 Type *EltTy = VecTy->getVectorElementType();
11838 VectorType *SubVecTy = VectorType::get(EltTy, NumSubElts);
11840 const DataLayout &DL = SI->getModule()->getDataLayout();
11841 unsigned SubVecSize = DL.getTypeAllocSizeInBits(SubVecTy);
11842 bool EltIs64Bits = DL.getTypeAllocSizeInBits(EltTy) == 64;
11844 // Skip illegal sub vector types and vector types of i64/f64 element (vstN
11845 // doesn't support i64/f64 element).
11846 if ((SubVecSize != 64 && SubVecSize != 128) || EltIs64Bits)
11849 Value *Op0 = SVI->getOperand(0);
11850 Value *Op1 = SVI->getOperand(1);
11851 IRBuilder<> Builder(SI);
11853 // StN intrinsics don't support pointer vectors as arguments. Convert pointer
11854 // vectors to integer vectors.
11855 if (EltTy->isPointerTy()) {
11856 Type *IntTy = DL.getIntPtrType(EltTy);
11858 // Convert to the corresponding integer vector.
11860 VectorType::get(IntTy, Op0->getType()->getVectorNumElements());
11861 Op0 = Builder.CreatePtrToInt(Op0, IntVecTy);
11862 Op1 = Builder.CreatePtrToInt(Op1, IntVecTy);
11864 SubVecTy = VectorType::get(IntTy, NumSubElts);
11867 static Intrinsic::ID StoreInts[3] = {Intrinsic::arm_neon_vst2,
11868 Intrinsic::arm_neon_vst3,
11869 Intrinsic::arm_neon_vst4};
11870 Function *VstNFunc = Intrinsic::getDeclaration(
11871 SI->getModule(), StoreInts[Factor - 2], SubVecTy);
11873 SmallVector<Value *, 6> Ops;
11875 Type *Int8Ptr = Builder.getInt8PtrTy(SI->getPointerAddressSpace());
11876 Ops.push_back(Builder.CreateBitCast(SI->getPointerOperand(), Int8Ptr));
11878 // Split the shufflevector operands into sub vectors for the new vstN call.
11879 for (unsigned i = 0; i < Factor; i++)
11880 Ops.push_back(Builder.CreateShuffleVector(
11881 Op0, Op1, getSequentialMask(Builder, NumSubElts * i, NumSubElts)));
11883 Ops.push_back(Builder.getInt32(SI->getAlignment()));
11884 Builder.CreateCall(VstNFunc, Ops);
11896 static bool isHomogeneousAggregate(Type *Ty, HABaseType &Base,
11897 uint64_t &Members) {
11898 if (auto *ST = dyn_cast<StructType>(Ty)) {
11899 for (unsigned i = 0; i < ST->getNumElements(); ++i) {
11900 uint64_t SubMembers = 0;
11901 if (!isHomogeneousAggregate(ST->getElementType(i), Base, SubMembers))
11903 Members += SubMembers;
11905 } else if (auto *AT = dyn_cast<ArrayType>(Ty)) {
11906 uint64_t SubMembers = 0;
11907 if (!isHomogeneousAggregate(AT->getElementType(), Base, SubMembers))
11909 Members += SubMembers * AT->getNumElements();
11910 } else if (Ty->isFloatTy()) {
11911 if (Base != HA_UNKNOWN && Base != HA_FLOAT)
11915 } else if (Ty->isDoubleTy()) {
11916 if (Base != HA_UNKNOWN && Base != HA_DOUBLE)
11920 } else if (auto *VT = dyn_cast<VectorType>(Ty)) {
11927 return VT->getBitWidth() == 64;
11929 return VT->getBitWidth() == 128;
11931 switch (VT->getBitWidth()) {
11944 return (Members > 0 && Members <= 4);
11947 /// \brief Return true if a type is an AAPCS-VFP homogeneous aggregate or one of
11948 /// [N x i32] or [N x i64]. This allows front-ends to skip emitting padding when
11949 /// passing according to AAPCS rules.
11950 bool ARMTargetLowering::functionArgumentNeedsConsecutiveRegisters(
11951 Type *Ty, CallingConv::ID CallConv, bool isVarArg) const {
11952 if (getEffectiveCallingConv(CallConv, isVarArg) !=
11953 CallingConv::ARM_AAPCS_VFP)
11956 HABaseType Base = HA_UNKNOWN;
11957 uint64_t Members = 0;
11958 bool IsHA = isHomogeneousAggregate(Ty, Base, Members);
11959 DEBUG(dbgs() << "isHA: " << IsHA << " "; Ty->dump());
11961 bool IsIntArray = Ty->isArrayTy() && Ty->getArrayElementType()->isIntegerTy();
11962 return IsHA || IsIntArray;