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/CodeGen/CallingConvLower.h"
27 #include "llvm/CodeGen/IntrinsicLowering.h"
28 #include "llvm/CodeGen/MachineBasicBlock.h"
29 #include "llvm/CodeGen/MachineFrameInfo.h"
30 #include "llvm/CodeGen/MachineFunction.h"
31 #include "llvm/CodeGen/MachineInstrBuilder.h"
32 #include "llvm/CodeGen/MachineJumpTableInfo.h"
33 #include "llvm/CodeGen/MachineModuleInfo.h"
34 #include "llvm/CodeGen/MachineRegisterInfo.h"
35 #include "llvm/CodeGen/SelectionDAG.h"
36 #include "llvm/IR/CallingConv.h"
37 #include "llvm/IR/Constants.h"
38 #include "llvm/IR/Function.h"
39 #include "llvm/IR/GlobalValue.h"
40 #include "llvm/IR/IRBuilder.h"
41 #include "llvm/IR/Instruction.h"
42 #include "llvm/IR/Instructions.h"
43 #include "llvm/IR/Intrinsics.h"
44 #include "llvm/IR/Type.h"
45 #include "llvm/MC/MCSectionMachO.h"
46 #include "llvm/Support/CommandLine.h"
47 #include "llvm/Support/Debug.h"
48 #include "llvm/Support/ErrorHandling.h"
49 #include "llvm/Support/MathExtras.h"
50 #include "llvm/Target/TargetOptions.h"
54 #define DEBUG_TYPE "arm-isel"
56 STATISTIC(NumTailCalls, "Number of tail calls");
57 STATISTIC(NumMovwMovt, "Number of GAs materialized with movw + movt");
58 STATISTIC(NumLoopByVals, "Number of loops generated for byval arguments");
61 EnableARMLongCalls("arm-long-calls", cl::Hidden,
62 cl::desc("Generate calls via indirect call instructions"),
66 ARMInterworking("arm-interworking", cl::Hidden,
67 cl::desc("Enable / disable ARM interworking (for debugging only)"),
71 class ARMCCState : public CCState {
73 ARMCCState(CallingConv::ID CC, bool isVarArg, MachineFunction &MF,
74 SmallVectorImpl<CCValAssign> &locs, LLVMContext &C,
76 : CCState(CC, isVarArg, MF, locs, C) {
77 assert(((PC == Call) || (PC == Prologue)) &&
78 "ARMCCState users must specify whether their context is call"
79 "or prologue generation.");
85 // The APCS parameter registers.
86 static const MCPhysReg GPRArgRegs[] = {
87 ARM::R0, ARM::R1, ARM::R2, ARM::R3
90 void ARMTargetLowering::addTypeForNEON(MVT VT, MVT PromotedLdStVT,
91 MVT PromotedBitwiseVT) {
92 if (VT != PromotedLdStVT) {
93 setOperationAction(ISD::LOAD, VT, Promote);
94 AddPromotedToType (ISD::LOAD, VT, PromotedLdStVT);
96 setOperationAction(ISD::STORE, VT, Promote);
97 AddPromotedToType (ISD::STORE, VT, PromotedLdStVT);
100 MVT ElemTy = VT.getVectorElementType();
101 if (ElemTy != MVT::i64 && ElemTy != MVT::f64)
102 setOperationAction(ISD::SETCC, VT, Custom);
103 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
104 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
105 if (ElemTy == MVT::i32) {
106 setOperationAction(ISD::SINT_TO_FP, VT, Custom);
107 setOperationAction(ISD::UINT_TO_FP, VT, Custom);
108 setOperationAction(ISD::FP_TO_SINT, VT, Custom);
109 setOperationAction(ISD::FP_TO_UINT, VT, Custom);
111 setOperationAction(ISD::SINT_TO_FP, VT, Expand);
112 setOperationAction(ISD::UINT_TO_FP, VT, Expand);
113 setOperationAction(ISD::FP_TO_SINT, VT, Expand);
114 setOperationAction(ISD::FP_TO_UINT, VT, Expand);
116 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
117 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
118 setOperationAction(ISD::CONCAT_VECTORS, VT, Legal);
119 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Legal);
120 setOperationAction(ISD::SELECT, VT, Expand);
121 setOperationAction(ISD::SELECT_CC, VT, Expand);
122 setOperationAction(ISD::VSELECT, VT, Expand);
123 setOperationAction(ISD::SIGN_EXTEND_INREG, VT, Expand);
124 if (VT.isInteger()) {
125 setOperationAction(ISD::SHL, VT, Custom);
126 setOperationAction(ISD::SRA, VT, Custom);
127 setOperationAction(ISD::SRL, VT, Custom);
130 // Promote all bit-wise operations.
131 if (VT.isInteger() && VT != PromotedBitwiseVT) {
132 setOperationAction(ISD::AND, VT, Promote);
133 AddPromotedToType (ISD::AND, VT, PromotedBitwiseVT);
134 setOperationAction(ISD::OR, VT, Promote);
135 AddPromotedToType (ISD::OR, VT, PromotedBitwiseVT);
136 setOperationAction(ISD::XOR, VT, Promote);
137 AddPromotedToType (ISD::XOR, VT, PromotedBitwiseVT);
140 // Neon does not support vector divide/remainder operations.
141 setOperationAction(ISD::SDIV, VT, Expand);
142 setOperationAction(ISD::UDIV, VT, Expand);
143 setOperationAction(ISD::FDIV, VT, Expand);
144 setOperationAction(ISD::SREM, VT, Expand);
145 setOperationAction(ISD::UREM, VT, Expand);
146 setOperationAction(ISD::FREM, VT, Expand);
149 void ARMTargetLowering::addDRTypeForNEON(MVT VT) {
150 addRegisterClass(VT, &ARM::DPRRegClass);
151 addTypeForNEON(VT, MVT::f64, MVT::v2i32);
154 void ARMTargetLowering::addQRTypeForNEON(MVT VT) {
155 addRegisterClass(VT, &ARM::DPairRegClass);
156 addTypeForNEON(VT, MVT::v2f64, MVT::v4i32);
159 ARMTargetLowering::ARMTargetLowering(const TargetMachine &TM,
160 const ARMSubtarget &STI)
161 : TargetLowering(TM), Subtarget(&STI) {
162 RegInfo = Subtarget->getRegisterInfo();
163 Itins = Subtarget->getInstrItineraryData();
165 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
167 if (Subtarget->isTargetMachO()) {
168 // Uses VFP for Thumb libfuncs if available.
169 if (Subtarget->isThumb() && Subtarget->hasVFP2() &&
170 Subtarget->hasARMOps() && !TM.Options.UseSoftFloat) {
171 // Single-precision floating-point arithmetic.
172 setLibcallName(RTLIB::ADD_F32, "__addsf3vfp");
173 setLibcallName(RTLIB::SUB_F32, "__subsf3vfp");
174 setLibcallName(RTLIB::MUL_F32, "__mulsf3vfp");
175 setLibcallName(RTLIB::DIV_F32, "__divsf3vfp");
177 // Double-precision floating-point arithmetic.
178 setLibcallName(RTLIB::ADD_F64, "__adddf3vfp");
179 setLibcallName(RTLIB::SUB_F64, "__subdf3vfp");
180 setLibcallName(RTLIB::MUL_F64, "__muldf3vfp");
181 setLibcallName(RTLIB::DIV_F64, "__divdf3vfp");
183 // Single-precision comparisons.
184 setLibcallName(RTLIB::OEQ_F32, "__eqsf2vfp");
185 setLibcallName(RTLIB::UNE_F32, "__nesf2vfp");
186 setLibcallName(RTLIB::OLT_F32, "__ltsf2vfp");
187 setLibcallName(RTLIB::OLE_F32, "__lesf2vfp");
188 setLibcallName(RTLIB::OGE_F32, "__gesf2vfp");
189 setLibcallName(RTLIB::OGT_F32, "__gtsf2vfp");
190 setLibcallName(RTLIB::UO_F32, "__unordsf2vfp");
191 setLibcallName(RTLIB::O_F32, "__unordsf2vfp");
193 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
194 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETNE);
195 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
196 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
197 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
198 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
199 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
200 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
202 // Double-precision comparisons.
203 setLibcallName(RTLIB::OEQ_F64, "__eqdf2vfp");
204 setLibcallName(RTLIB::UNE_F64, "__nedf2vfp");
205 setLibcallName(RTLIB::OLT_F64, "__ltdf2vfp");
206 setLibcallName(RTLIB::OLE_F64, "__ledf2vfp");
207 setLibcallName(RTLIB::OGE_F64, "__gedf2vfp");
208 setLibcallName(RTLIB::OGT_F64, "__gtdf2vfp");
209 setLibcallName(RTLIB::UO_F64, "__unorddf2vfp");
210 setLibcallName(RTLIB::O_F64, "__unorddf2vfp");
212 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
213 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETNE);
214 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
215 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
216 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
217 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
218 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
219 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
221 // Floating-point to integer conversions.
222 // i64 conversions are done via library routines even when generating VFP
223 // instructions, so use the same ones.
224 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__fixdfsivfp");
225 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__fixunsdfsivfp");
226 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__fixsfsivfp");
227 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__fixunssfsivfp");
229 // Conversions between floating types.
230 setLibcallName(RTLIB::FPROUND_F64_F32, "__truncdfsf2vfp");
231 setLibcallName(RTLIB::FPEXT_F32_F64, "__extendsfdf2vfp");
233 // Integer to floating-point conversions.
234 // i64 conversions are done via library routines even when generating VFP
235 // instructions, so use the same ones.
236 // FIXME: There appears to be some naming inconsistency in ARM libgcc:
237 // e.g., __floatunsidf vs. __floatunssidfvfp.
238 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__floatsidfvfp");
239 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__floatunssidfvfp");
240 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__floatsisfvfp");
241 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__floatunssisfvfp");
245 // These libcalls are not available in 32-bit.
246 setLibcallName(RTLIB::SHL_I128, nullptr);
247 setLibcallName(RTLIB::SRL_I128, nullptr);
248 setLibcallName(RTLIB::SRA_I128, nullptr);
250 if (Subtarget->isAAPCS_ABI() && !Subtarget->isTargetMachO() &&
251 !Subtarget->isTargetWindows()) {
252 static const struct {
253 const RTLIB::Libcall Op;
254 const char * const Name;
255 const CallingConv::ID CC;
256 const ISD::CondCode Cond;
258 // Double-precision floating-point arithmetic helper functions
259 // RTABI chapter 4.1.2, Table 2
260 { RTLIB::ADD_F64, "__aeabi_dadd", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
261 { RTLIB::DIV_F64, "__aeabi_ddiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
262 { RTLIB::MUL_F64, "__aeabi_dmul", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
263 { RTLIB::SUB_F64, "__aeabi_dsub", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
265 // Double-precision floating-point comparison helper functions
266 // RTABI chapter 4.1.2, Table 3
267 { RTLIB::OEQ_F64, "__aeabi_dcmpeq", CallingConv::ARM_AAPCS, ISD::SETNE },
268 { RTLIB::UNE_F64, "__aeabi_dcmpeq", CallingConv::ARM_AAPCS, ISD::SETEQ },
269 { RTLIB::OLT_F64, "__aeabi_dcmplt", CallingConv::ARM_AAPCS, ISD::SETNE },
270 { RTLIB::OLE_F64, "__aeabi_dcmple", CallingConv::ARM_AAPCS, ISD::SETNE },
271 { RTLIB::OGE_F64, "__aeabi_dcmpge", CallingConv::ARM_AAPCS, ISD::SETNE },
272 { RTLIB::OGT_F64, "__aeabi_dcmpgt", CallingConv::ARM_AAPCS, ISD::SETNE },
273 { RTLIB::UO_F64, "__aeabi_dcmpun", CallingConv::ARM_AAPCS, ISD::SETNE },
274 { RTLIB::O_F64, "__aeabi_dcmpun", CallingConv::ARM_AAPCS, ISD::SETEQ },
276 // Single-precision floating-point arithmetic helper functions
277 // RTABI chapter 4.1.2, Table 4
278 { RTLIB::ADD_F32, "__aeabi_fadd", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
279 { RTLIB::DIV_F32, "__aeabi_fdiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
280 { RTLIB::MUL_F32, "__aeabi_fmul", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
281 { RTLIB::SUB_F32, "__aeabi_fsub", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
283 // Single-precision floating-point comparison helper functions
284 // RTABI chapter 4.1.2, Table 5
285 { RTLIB::OEQ_F32, "__aeabi_fcmpeq", CallingConv::ARM_AAPCS, ISD::SETNE },
286 { RTLIB::UNE_F32, "__aeabi_fcmpeq", CallingConv::ARM_AAPCS, ISD::SETEQ },
287 { RTLIB::OLT_F32, "__aeabi_fcmplt", CallingConv::ARM_AAPCS, ISD::SETNE },
288 { RTLIB::OLE_F32, "__aeabi_fcmple", CallingConv::ARM_AAPCS, ISD::SETNE },
289 { RTLIB::OGE_F32, "__aeabi_fcmpge", CallingConv::ARM_AAPCS, ISD::SETNE },
290 { RTLIB::OGT_F32, "__aeabi_fcmpgt", CallingConv::ARM_AAPCS, ISD::SETNE },
291 { RTLIB::UO_F32, "__aeabi_fcmpun", CallingConv::ARM_AAPCS, ISD::SETNE },
292 { RTLIB::O_F32, "__aeabi_fcmpun", CallingConv::ARM_AAPCS, ISD::SETEQ },
294 // Floating-point to integer conversions.
295 // RTABI chapter 4.1.2, Table 6
296 { RTLIB::FPTOSINT_F64_I32, "__aeabi_d2iz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
297 { RTLIB::FPTOUINT_F64_I32, "__aeabi_d2uiz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
298 { RTLIB::FPTOSINT_F64_I64, "__aeabi_d2lz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
299 { RTLIB::FPTOUINT_F64_I64, "__aeabi_d2ulz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
300 { RTLIB::FPTOSINT_F32_I32, "__aeabi_f2iz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
301 { RTLIB::FPTOUINT_F32_I32, "__aeabi_f2uiz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
302 { RTLIB::FPTOSINT_F32_I64, "__aeabi_f2lz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
303 { RTLIB::FPTOUINT_F32_I64, "__aeabi_f2ulz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
305 // Conversions between floating types.
306 // RTABI chapter 4.1.2, Table 7
307 { RTLIB::FPROUND_F64_F32, "__aeabi_d2f", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
308 { RTLIB::FPROUND_F64_F16, "__aeabi_d2h", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
309 { RTLIB::FPEXT_F32_F64, "__aeabi_f2d", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
311 // Integer to floating-point conversions.
312 // RTABI chapter 4.1.2, Table 8
313 { RTLIB::SINTTOFP_I32_F64, "__aeabi_i2d", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
314 { RTLIB::UINTTOFP_I32_F64, "__aeabi_ui2d", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
315 { RTLIB::SINTTOFP_I64_F64, "__aeabi_l2d", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
316 { RTLIB::UINTTOFP_I64_F64, "__aeabi_ul2d", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
317 { RTLIB::SINTTOFP_I32_F32, "__aeabi_i2f", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
318 { RTLIB::UINTTOFP_I32_F32, "__aeabi_ui2f", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
319 { RTLIB::SINTTOFP_I64_F32, "__aeabi_l2f", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
320 { RTLIB::UINTTOFP_I64_F32, "__aeabi_ul2f", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
322 // Long long helper functions
323 // RTABI chapter 4.2, Table 9
324 { RTLIB::MUL_I64, "__aeabi_lmul", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
325 { RTLIB::SHL_I64, "__aeabi_llsl", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
326 { RTLIB::SRL_I64, "__aeabi_llsr", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
327 { RTLIB::SRA_I64, "__aeabi_lasr", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
329 // Integer division functions
330 // RTABI chapter 4.3.1
331 { RTLIB::SDIV_I8, "__aeabi_idiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
332 { RTLIB::SDIV_I16, "__aeabi_idiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
333 { RTLIB::SDIV_I32, "__aeabi_idiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
334 { RTLIB::SDIV_I64, "__aeabi_ldivmod", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
335 { RTLIB::UDIV_I8, "__aeabi_uidiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
336 { RTLIB::UDIV_I16, "__aeabi_uidiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
337 { RTLIB::UDIV_I32, "__aeabi_uidiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
338 { RTLIB::UDIV_I64, "__aeabi_uldivmod", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
341 // RTABI chapter 4.3.4
342 { RTLIB::MEMCPY, "__aeabi_memcpy", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
343 { RTLIB::MEMMOVE, "__aeabi_memmove", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
344 { RTLIB::MEMSET, "__aeabi_memset", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
347 for (const auto &LC : LibraryCalls) {
348 setLibcallName(LC.Op, LC.Name);
349 setLibcallCallingConv(LC.Op, LC.CC);
350 if (LC.Cond != ISD::SETCC_INVALID)
351 setCmpLibcallCC(LC.Op, LC.Cond);
355 if (Subtarget->isTargetWindows()) {
356 static const struct {
357 const RTLIB::Libcall Op;
358 const char * const Name;
359 const CallingConv::ID CC;
361 { RTLIB::FPTOSINT_F32_I64, "__stoi64", CallingConv::ARM_AAPCS_VFP },
362 { RTLIB::FPTOSINT_F64_I64, "__dtoi64", CallingConv::ARM_AAPCS_VFP },
363 { RTLIB::FPTOUINT_F32_I64, "__stou64", CallingConv::ARM_AAPCS_VFP },
364 { RTLIB::FPTOUINT_F64_I64, "__dtou64", CallingConv::ARM_AAPCS_VFP },
365 { RTLIB::SINTTOFP_I64_F32, "__i64tos", CallingConv::ARM_AAPCS_VFP },
366 { RTLIB::SINTTOFP_I64_F64, "__i64tod", CallingConv::ARM_AAPCS_VFP },
367 { RTLIB::UINTTOFP_I64_F32, "__u64tos", CallingConv::ARM_AAPCS_VFP },
368 { RTLIB::UINTTOFP_I64_F64, "__u64tod", CallingConv::ARM_AAPCS_VFP },
371 for (const auto &LC : LibraryCalls) {
372 setLibcallName(LC.Op, LC.Name);
373 setLibcallCallingConv(LC.Op, LC.CC);
377 // Use divmod compiler-rt calls for iOS 5.0 and later.
378 if (Subtarget->getTargetTriple().isiOS() &&
379 !Subtarget->getTargetTriple().isOSVersionLT(5, 0)) {
380 setLibcallName(RTLIB::SDIVREM_I32, "__divmodsi4");
381 setLibcallName(RTLIB::UDIVREM_I32, "__udivmodsi4");
384 // The half <-> float conversion functions are always soft-float, but are
385 // needed for some targets which use a hard-float calling convention by
387 if (Subtarget->isAAPCS_ABI()) {
388 setLibcallCallingConv(RTLIB::FPROUND_F32_F16, CallingConv::ARM_AAPCS);
389 setLibcallCallingConv(RTLIB::FPROUND_F64_F16, CallingConv::ARM_AAPCS);
390 setLibcallCallingConv(RTLIB::FPEXT_F16_F32, CallingConv::ARM_AAPCS);
392 setLibcallCallingConv(RTLIB::FPROUND_F32_F16, CallingConv::ARM_APCS);
393 setLibcallCallingConv(RTLIB::FPROUND_F64_F16, CallingConv::ARM_APCS);
394 setLibcallCallingConv(RTLIB::FPEXT_F16_F32, CallingConv::ARM_APCS);
397 if (Subtarget->isThumb1Only())
398 addRegisterClass(MVT::i32, &ARM::tGPRRegClass);
400 addRegisterClass(MVT::i32, &ARM::GPRRegClass);
401 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
402 !Subtarget->isThumb1Only()) {
403 addRegisterClass(MVT::f32, &ARM::SPRRegClass);
404 addRegisterClass(MVT::f64, &ARM::DPRRegClass);
407 for (MVT VT : MVT::vector_valuetypes()) {
408 for (MVT InnerVT : MVT::vector_valuetypes()) {
409 setTruncStoreAction(VT, InnerVT, Expand);
410 setLoadExtAction(ISD::SEXTLOAD, VT, InnerVT, Expand);
411 setLoadExtAction(ISD::ZEXTLOAD, VT, InnerVT, Expand);
412 setLoadExtAction(ISD::EXTLOAD, VT, InnerVT, Expand);
415 setOperationAction(ISD::MULHS, VT, Expand);
416 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
417 setOperationAction(ISD::MULHU, VT, Expand);
418 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
420 setOperationAction(ISD::BSWAP, VT, Expand);
423 setOperationAction(ISD::ConstantFP, MVT::f32, Custom);
424 setOperationAction(ISD::ConstantFP, MVT::f64, Custom);
426 if (Subtarget->hasNEON()) {
427 addDRTypeForNEON(MVT::v2f32);
428 addDRTypeForNEON(MVT::v8i8);
429 addDRTypeForNEON(MVT::v4i16);
430 addDRTypeForNEON(MVT::v2i32);
431 addDRTypeForNEON(MVT::v1i64);
433 addQRTypeForNEON(MVT::v4f32);
434 addQRTypeForNEON(MVT::v2f64);
435 addQRTypeForNEON(MVT::v16i8);
436 addQRTypeForNEON(MVT::v8i16);
437 addQRTypeForNEON(MVT::v4i32);
438 addQRTypeForNEON(MVT::v2i64);
440 // v2f64 is legal so that QR subregs can be extracted as f64 elements, but
441 // neither Neon nor VFP support any arithmetic operations on it.
442 // The same with v4f32. But keep in mind that vadd, vsub, vmul are natively
443 // supported for v4f32.
444 setOperationAction(ISD::FADD, MVT::v2f64, Expand);
445 setOperationAction(ISD::FSUB, MVT::v2f64, Expand);
446 setOperationAction(ISD::FMUL, MVT::v2f64, Expand);
447 // FIXME: Code duplication: FDIV and FREM are expanded always, see
448 // ARMTargetLowering::addTypeForNEON method for details.
449 setOperationAction(ISD::FDIV, MVT::v2f64, Expand);
450 setOperationAction(ISD::FREM, MVT::v2f64, Expand);
451 // FIXME: Create unittest.
452 // In another words, find a way when "copysign" appears in DAG with vector
454 setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand);
455 // FIXME: Code duplication: SETCC has custom operation action, see
456 // ARMTargetLowering::addTypeForNEON method for details.
457 setOperationAction(ISD::SETCC, MVT::v2f64, Expand);
458 // FIXME: Create unittest for FNEG and for FABS.
459 setOperationAction(ISD::FNEG, MVT::v2f64, Expand);
460 setOperationAction(ISD::FABS, MVT::v2f64, Expand);
461 setOperationAction(ISD::FSQRT, MVT::v2f64, Expand);
462 setOperationAction(ISD::FSIN, MVT::v2f64, Expand);
463 setOperationAction(ISD::FCOS, MVT::v2f64, Expand);
464 setOperationAction(ISD::FPOWI, MVT::v2f64, Expand);
465 setOperationAction(ISD::FPOW, MVT::v2f64, Expand);
466 setOperationAction(ISD::FLOG, MVT::v2f64, Expand);
467 setOperationAction(ISD::FLOG2, MVT::v2f64, Expand);
468 setOperationAction(ISD::FLOG10, MVT::v2f64, Expand);
469 setOperationAction(ISD::FEXP, MVT::v2f64, Expand);
470 setOperationAction(ISD::FEXP2, MVT::v2f64, Expand);
471 // FIXME: Create unittest for FCEIL, FTRUNC, FRINT, FNEARBYINT, FFLOOR.
472 setOperationAction(ISD::FCEIL, MVT::v2f64, Expand);
473 setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand);
474 setOperationAction(ISD::FRINT, MVT::v2f64, Expand);
475 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand);
476 setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand);
477 setOperationAction(ISD::FMA, MVT::v2f64, Expand);
479 setOperationAction(ISD::FSQRT, MVT::v4f32, Expand);
480 setOperationAction(ISD::FSIN, MVT::v4f32, Expand);
481 setOperationAction(ISD::FCOS, MVT::v4f32, Expand);
482 setOperationAction(ISD::FPOWI, MVT::v4f32, Expand);
483 setOperationAction(ISD::FPOW, MVT::v4f32, Expand);
484 setOperationAction(ISD::FLOG, MVT::v4f32, Expand);
485 setOperationAction(ISD::FLOG2, MVT::v4f32, Expand);
486 setOperationAction(ISD::FLOG10, MVT::v4f32, Expand);
487 setOperationAction(ISD::FEXP, MVT::v4f32, Expand);
488 setOperationAction(ISD::FEXP2, MVT::v4f32, Expand);
489 setOperationAction(ISD::FCEIL, MVT::v4f32, Expand);
490 setOperationAction(ISD::FTRUNC, MVT::v4f32, Expand);
491 setOperationAction(ISD::FRINT, MVT::v4f32, Expand);
492 setOperationAction(ISD::FNEARBYINT, MVT::v4f32, Expand);
493 setOperationAction(ISD::FFLOOR, MVT::v4f32, Expand);
495 // Mark v2f32 intrinsics.
496 setOperationAction(ISD::FSQRT, MVT::v2f32, Expand);
497 setOperationAction(ISD::FSIN, MVT::v2f32, Expand);
498 setOperationAction(ISD::FCOS, MVT::v2f32, Expand);
499 setOperationAction(ISD::FPOWI, MVT::v2f32, Expand);
500 setOperationAction(ISD::FPOW, MVT::v2f32, Expand);
501 setOperationAction(ISD::FLOG, MVT::v2f32, Expand);
502 setOperationAction(ISD::FLOG2, MVT::v2f32, Expand);
503 setOperationAction(ISD::FLOG10, MVT::v2f32, Expand);
504 setOperationAction(ISD::FEXP, MVT::v2f32, Expand);
505 setOperationAction(ISD::FEXP2, MVT::v2f32, Expand);
506 setOperationAction(ISD::FCEIL, MVT::v2f32, Expand);
507 setOperationAction(ISD::FTRUNC, MVT::v2f32, Expand);
508 setOperationAction(ISD::FRINT, MVT::v2f32, Expand);
509 setOperationAction(ISD::FNEARBYINT, MVT::v2f32, Expand);
510 setOperationAction(ISD::FFLOOR, MVT::v2f32, Expand);
512 // Neon does not support some operations on v1i64 and v2i64 types.
513 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
514 // Custom handling for some quad-vector types to detect VMULL.
515 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
516 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
517 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
518 // Custom handling for some vector types to avoid expensive expansions
519 setOperationAction(ISD::SDIV, MVT::v4i16, Custom);
520 setOperationAction(ISD::SDIV, MVT::v8i8, Custom);
521 setOperationAction(ISD::UDIV, MVT::v4i16, Custom);
522 setOperationAction(ISD::UDIV, MVT::v8i8, Custom);
523 setOperationAction(ISD::SETCC, MVT::v1i64, Expand);
524 setOperationAction(ISD::SETCC, MVT::v2i64, Expand);
525 // Neon does not have single instruction SINT_TO_FP and UINT_TO_FP with
526 // a destination type that is wider than the source, and nor does
527 // it have a FP_TO_[SU]INT instruction with a narrower destination than
529 setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Custom);
530 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
531 setOperationAction(ISD::FP_TO_UINT, MVT::v4i16, Custom);
532 setOperationAction(ISD::FP_TO_SINT, MVT::v4i16, Custom);
534 setOperationAction(ISD::FP_ROUND, MVT::v2f32, Expand);
535 setOperationAction(ISD::FP_EXTEND, MVT::v2f64, Expand);
537 // NEON does not have single instruction CTPOP for vectors with element
538 // types wider than 8-bits. However, custom lowering can leverage the
539 // v8i8/v16i8 vcnt instruction.
540 setOperationAction(ISD::CTPOP, MVT::v2i32, Custom);
541 setOperationAction(ISD::CTPOP, MVT::v4i32, Custom);
542 setOperationAction(ISD::CTPOP, MVT::v4i16, Custom);
543 setOperationAction(ISD::CTPOP, MVT::v8i16, Custom);
545 // NEON only has FMA instructions as of VFP4.
546 if (!Subtarget->hasVFP4()) {
547 setOperationAction(ISD::FMA, MVT::v2f32, Expand);
548 setOperationAction(ISD::FMA, MVT::v4f32, Expand);
551 setTargetDAGCombine(ISD::INTRINSIC_VOID);
552 setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN);
553 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
554 setTargetDAGCombine(ISD::SHL);
555 setTargetDAGCombine(ISD::SRL);
556 setTargetDAGCombine(ISD::SRA);
557 setTargetDAGCombine(ISD::SIGN_EXTEND);
558 setTargetDAGCombine(ISD::ZERO_EXTEND);
559 setTargetDAGCombine(ISD::ANY_EXTEND);
560 setTargetDAGCombine(ISD::SELECT_CC);
561 setTargetDAGCombine(ISD::BUILD_VECTOR);
562 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
563 setTargetDAGCombine(ISD::INSERT_VECTOR_ELT);
564 setTargetDAGCombine(ISD::STORE);
565 setTargetDAGCombine(ISD::FP_TO_SINT);
566 setTargetDAGCombine(ISD::FP_TO_UINT);
567 setTargetDAGCombine(ISD::FDIV);
568 setTargetDAGCombine(ISD::LOAD);
570 // It is legal to extload from v4i8 to v4i16 or v4i32.
571 MVT Tys[6] = {MVT::v8i8, MVT::v4i8, MVT::v2i8,
572 MVT::v4i16, MVT::v2i16,
574 for (unsigned i = 0; i < 6; ++i) {
575 for (MVT VT : MVT::integer_vector_valuetypes()) {
576 setLoadExtAction(ISD::EXTLOAD, VT, Tys[i], Legal);
577 setLoadExtAction(ISD::ZEXTLOAD, VT, Tys[i], Legal);
578 setLoadExtAction(ISD::SEXTLOAD, VT, Tys[i], Legal);
583 // ARM and Thumb2 support UMLAL/SMLAL.
584 if (!Subtarget->isThumb1Only())
585 setTargetDAGCombine(ISD::ADDC);
587 if (Subtarget->isFPOnlySP()) {
588 // When targetting a floating-point unit with only single-precision
589 // operations, f64 is legal for the few double-precision instructions which
590 // are present However, no double-precision operations other than moves,
591 // loads and stores are provided by the hardware.
592 setOperationAction(ISD::FADD, MVT::f64, Expand);
593 setOperationAction(ISD::FSUB, MVT::f64, Expand);
594 setOperationAction(ISD::FMUL, MVT::f64, Expand);
595 setOperationAction(ISD::FMA, MVT::f64, Expand);
596 setOperationAction(ISD::FDIV, MVT::f64, Expand);
597 setOperationAction(ISD::FREM, MVT::f64, Expand);
598 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
599 setOperationAction(ISD::FGETSIGN, MVT::f64, Expand);
600 setOperationAction(ISD::FNEG, MVT::f64, Expand);
601 setOperationAction(ISD::FABS, MVT::f64, Expand);
602 setOperationAction(ISD::FSQRT, MVT::f64, Expand);
603 setOperationAction(ISD::FSIN, MVT::f64, Expand);
604 setOperationAction(ISD::FCOS, MVT::f64, Expand);
605 setOperationAction(ISD::FPOWI, MVT::f64, Expand);
606 setOperationAction(ISD::FPOW, MVT::f64, Expand);
607 setOperationAction(ISD::FLOG, MVT::f64, Expand);
608 setOperationAction(ISD::FLOG2, MVT::f64, Expand);
609 setOperationAction(ISD::FLOG10, MVT::f64, Expand);
610 setOperationAction(ISD::FEXP, MVT::f64, Expand);
611 setOperationAction(ISD::FEXP2, MVT::f64, Expand);
612 setOperationAction(ISD::FCEIL, MVT::f64, Expand);
613 setOperationAction(ISD::FTRUNC, MVT::f64, Expand);
614 setOperationAction(ISD::FRINT, MVT::f64, Expand);
615 setOperationAction(ISD::FNEARBYINT, MVT::f64, Expand);
616 setOperationAction(ISD::FFLOOR, MVT::f64, Expand);
617 setOperationAction(ISD::FP_ROUND, MVT::f32, Custom);
618 setOperationAction(ISD::FP_EXTEND, MVT::f64, Custom);
621 computeRegisterProperties();
623 // ARM does not have floating-point extending loads.
624 for (MVT VT : MVT::fp_valuetypes()) {
625 setLoadExtAction(ISD::EXTLOAD, VT, MVT::f32, Expand);
626 setLoadExtAction(ISD::EXTLOAD, VT, MVT::f16, Expand);
629 // ... or truncating stores
630 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
631 setTruncStoreAction(MVT::f32, MVT::f16, Expand);
632 setTruncStoreAction(MVT::f64, MVT::f16, Expand);
634 // ARM does not have i1 sign extending load.
635 for (MVT VT : MVT::integer_valuetypes())
636 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Promote);
638 // ARM supports all 4 flavors of integer indexed load / store.
639 if (!Subtarget->isThumb1Only()) {
640 for (unsigned im = (unsigned)ISD::PRE_INC;
641 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
642 setIndexedLoadAction(im, MVT::i1, Legal);
643 setIndexedLoadAction(im, MVT::i8, Legal);
644 setIndexedLoadAction(im, MVT::i16, Legal);
645 setIndexedLoadAction(im, MVT::i32, Legal);
646 setIndexedStoreAction(im, MVT::i1, Legal);
647 setIndexedStoreAction(im, MVT::i8, Legal);
648 setIndexedStoreAction(im, MVT::i16, Legal);
649 setIndexedStoreAction(im, MVT::i32, Legal);
653 setOperationAction(ISD::SADDO, MVT::i32, Custom);
654 setOperationAction(ISD::UADDO, MVT::i32, Custom);
655 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
656 setOperationAction(ISD::USUBO, MVT::i32, Custom);
658 // i64 operation support.
659 setOperationAction(ISD::MUL, MVT::i64, Expand);
660 setOperationAction(ISD::MULHU, MVT::i32, Expand);
661 if (Subtarget->isThumb1Only()) {
662 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
663 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
665 if (Subtarget->isThumb1Only() || !Subtarget->hasV6Ops()
666 || (Subtarget->isThumb2() && !Subtarget->hasThumb2DSP()))
667 setOperationAction(ISD::MULHS, MVT::i32, Expand);
669 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
670 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
671 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
672 setOperationAction(ISD::SRL, MVT::i64, Custom);
673 setOperationAction(ISD::SRA, MVT::i64, Custom);
675 if (!Subtarget->isThumb1Only()) {
676 // FIXME: We should do this for Thumb1 as well.
677 setOperationAction(ISD::ADDC, MVT::i32, Custom);
678 setOperationAction(ISD::ADDE, MVT::i32, Custom);
679 setOperationAction(ISD::SUBC, MVT::i32, Custom);
680 setOperationAction(ISD::SUBE, MVT::i32, Custom);
683 // ARM does not have ROTL.
684 setOperationAction(ISD::ROTL, MVT::i32, Expand);
685 setOperationAction(ISD::CTTZ, MVT::i32, Custom);
686 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
687 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
688 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
690 // These just redirect to CTTZ and CTLZ on ARM.
691 setOperationAction(ISD::CTTZ_ZERO_UNDEF , MVT::i32 , Expand);
692 setOperationAction(ISD::CTLZ_ZERO_UNDEF , MVT::i32 , Expand);
694 setOperationAction(ISD::READCYCLECOUNTER, MVT::i64, Custom);
696 // Only ARMv6 has BSWAP.
697 if (!Subtarget->hasV6Ops())
698 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
700 if (!(Subtarget->hasDivide() && Subtarget->isThumb2()) &&
701 !(Subtarget->hasDivideInARMMode() && !Subtarget->isThumb())) {
702 // These are expanded into libcalls if the cpu doesn't have HW divider.
703 setOperationAction(ISD::SDIV, MVT::i32, Expand);
704 setOperationAction(ISD::UDIV, MVT::i32, Expand);
707 // FIXME: Also set divmod for SREM on EABI
708 setOperationAction(ISD::SREM, MVT::i32, Expand);
709 setOperationAction(ISD::UREM, MVT::i32, Expand);
710 // Register based DivRem for AEABI (RTABI 4.2)
711 if (Subtarget->isTargetAEABI()) {
712 setLibcallName(RTLIB::SDIVREM_I8, "__aeabi_idivmod");
713 setLibcallName(RTLIB::SDIVREM_I16, "__aeabi_idivmod");
714 setLibcallName(RTLIB::SDIVREM_I32, "__aeabi_idivmod");
715 setLibcallName(RTLIB::SDIVREM_I64, "__aeabi_ldivmod");
716 setLibcallName(RTLIB::UDIVREM_I8, "__aeabi_uidivmod");
717 setLibcallName(RTLIB::UDIVREM_I16, "__aeabi_uidivmod");
718 setLibcallName(RTLIB::UDIVREM_I32, "__aeabi_uidivmod");
719 setLibcallName(RTLIB::UDIVREM_I64, "__aeabi_uldivmod");
721 setLibcallCallingConv(RTLIB::SDIVREM_I8, CallingConv::ARM_AAPCS);
722 setLibcallCallingConv(RTLIB::SDIVREM_I16, CallingConv::ARM_AAPCS);
723 setLibcallCallingConv(RTLIB::SDIVREM_I32, CallingConv::ARM_AAPCS);
724 setLibcallCallingConv(RTLIB::SDIVREM_I64, CallingConv::ARM_AAPCS);
725 setLibcallCallingConv(RTLIB::UDIVREM_I8, CallingConv::ARM_AAPCS);
726 setLibcallCallingConv(RTLIB::UDIVREM_I16, CallingConv::ARM_AAPCS);
727 setLibcallCallingConv(RTLIB::UDIVREM_I32, CallingConv::ARM_AAPCS);
728 setLibcallCallingConv(RTLIB::UDIVREM_I64, CallingConv::ARM_AAPCS);
730 setOperationAction(ISD::SDIVREM, MVT::i32, Custom);
731 setOperationAction(ISD::UDIVREM, MVT::i32, Custom);
733 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
734 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
737 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
738 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
739 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
740 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
741 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
743 setOperationAction(ISD::TRAP, MVT::Other, Legal);
745 // Use the default implementation.
746 setOperationAction(ISD::VASTART, MVT::Other, Custom);
747 setOperationAction(ISD::VAARG, MVT::Other, Expand);
748 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
749 setOperationAction(ISD::VAEND, MVT::Other, Expand);
750 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
751 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
753 if (!Subtarget->isTargetMachO()) {
754 // Non-MachO platforms may return values in these registers via the
755 // personality function.
756 setExceptionPointerRegister(ARM::R0);
757 setExceptionSelectorRegister(ARM::R1);
760 if (Subtarget->getTargetTriple().isWindowsItaniumEnvironment())
761 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
763 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
765 // ARMv6 Thumb1 (except for CPUs that support dmb / dsb) and earlier use
766 // the default expansion. If we are targeting a single threaded system,
767 // then set them all for expand so we can lower them later into their
769 if (TM.Options.ThreadModel == ThreadModel::Single)
770 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Expand);
771 else if (Subtarget->hasAnyDataBarrier() && !Subtarget->isThumb1Only()) {
772 // ATOMIC_FENCE needs custom lowering; the others should have been expanded
773 // to ldrex/strex loops already.
774 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Custom);
776 // On v8, we have particularly efficient implementations of atomic fences
777 // if they can be combined with nearby atomic loads and stores.
778 if (!Subtarget->hasV8Ops()) {
779 // Automatically insert fences (dmb ish) around ATOMIC_SWAP etc.
780 setInsertFencesForAtomic(true);
783 // If there's anything we can use as a barrier, go through custom lowering
785 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other,
786 Subtarget->hasAnyDataBarrier() ? Custom : Expand);
788 // Set them all for expansion, which will force libcalls.
789 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Expand);
790 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Expand);
791 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Expand);
792 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Expand);
793 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Expand);
794 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Expand);
795 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Expand);
796 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Expand);
797 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i32, Expand);
798 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i32, Expand);
799 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i32, Expand);
800 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i32, Expand);
801 // Mark ATOMIC_LOAD and ATOMIC_STORE custom so we can handle the
802 // Unordered/Monotonic case.
803 setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Custom);
804 setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Custom);
807 setOperationAction(ISD::PREFETCH, MVT::Other, Custom);
809 // Requires SXTB/SXTH, available on v6 and up in both ARM and Thumb modes.
810 if (!Subtarget->hasV6Ops()) {
811 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
812 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
814 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
816 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
817 !Subtarget->isThumb1Only()) {
818 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
819 // iff target supports vfp2.
820 setOperationAction(ISD::BITCAST, MVT::i64, Custom);
821 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
824 // We want to custom lower some of our intrinsics.
825 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
826 if (Subtarget->isTargetDarwin()) {
827 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
828 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
829 setLibcallName(RTLIB::UNWIND_RESUME, "_Unwind_SjLj_Resume");
832 setOperationAction(ISD::SETCC, MVT::i32, Expand);
833 setOperationAction(ISD::SETCC, MVT::f32, Expand);
834 setOperationAction(ISD::SETCC, MVT::f64, Expand);
835 setOperationAction(ISD::SELECT, MVT::i32, Custom);
836 setOperationAction(ISD::SELECT, MVT::f32, Custom);
837 setOperationAction(ISD::SELECT, MVT::f64, Custom);
838 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
839 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
840 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
842 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
843 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
844 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
845 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
846 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
848 // We don't support sin/cos/fmod/copysign/pow
849 setOperationAction(ISD::FSIN, MVT::f64, Expand);
850 setOperationAction(ISD::FSIN, MVT::f32, Expand);
851 setOperationAction(ISD::FCOS, MVT::f32, Expand);
852 setOperationAction(ISD::FCOS, MVT::f64, Expand);
853 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
854 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
855 setOperationAction(ISD::FREM, MVT::f64, Expand);
856 setOperationAction(ISD::FREM, MVT::f32, Expand);
857 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
858 !Subtarget->isThumb1Only()) {
859 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
860 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
862 setOperationAction(ISD::FPOW, MVT::f64, Expand);
863 setOperationAction(ISD::FPOW, MVT::f32, Expand);
865 if (!Subtarget->hasVFP4()) {
866 setOperationAction(ISD::FMA, MVT::f64, Expand);
867 setOperationAction(ISD::FMA, MVT::f32, Expand);
870 // Various VFP goodness
871 if (!TM.Options.UseSoftFloat && !Subtarget->isThumb1Only()) {
872 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
873 if (Subtarget->hasVFP2()) {
874 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
875 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
876 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
877 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
880 // FP-ARMv8 adds f64 <-> f16 conversion. Before that it should be expanded.
881 if (!Subtarget->hasFPARMv8() || Subtarget->isFPOnlySP()) {
882 setOperationAction(ISD::FP16_TO_FP, MVT::f64, Expand);
883 setOperationAction(ISD::FP_TO_FP16, MVT::f64, Expand);
886 // fp16 is a special v7 extension that adds f16 <-> f32 conversions.
887 if (!Subtarget->hasFP16()) {
888 setOperationAction(ISD::FP16_TO_FP, MVT::f32, Expand);
889 setOperationAction(ISD::FP_TO_FP16, MVT::f32, Expand);
893 // Combine sin / cos into one node or libcall if possible.
894 if (Subtarget->hasSinCos()) {
895 setLibcallName(RTLIB::SINCOS_F32, "sincosf");
896 setLibcallName(RTLIB::SINCOS_F64, "sincos");
897 if (Subtarget->getTargetTriple().isiOS()) {
898 // For iOS, we don't want to the normal expansion of a libcall to
899 // sincos. We want to issue a libcall to __sincos_stret.
900 setOperationAction(ISD::FSINCOS, MVT::f64, Custom);
901 setOperationAction(ISD::FSINCOS, MVT::f32, Custom);
905 // FP-ARMv8 implements a lot of rounding-like FP operations.
906 if (Subtarget->hasFPARMv8()) {
907 setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
908 setOperationAction(ISD::FCEIL, MVT::f32, Legal);
909 setOperationAction(ISD::FROUND, MVT::f32, Legal);
910 setOperationAction(ISD::FTRUNC, MVT::f32, Legal);
911 setOperationAction(ISD::FNEARBYINT, MVT::f32, Legal);
912 setOperationAction(ISD::FRINT, MVT::f32, Legal);
913 if (!Subtarget->isFPOnlySP()) {
914 setOperationAction(ISD::FFLOOR, MVT::f64, Legal);
915 setOperationAction(ISD::FCEIL, MVT::f64, Legal);
916 setOperationAction(ISD::FROUND, MVT::f64, Legal);
917 setOperationAction(ISD::FTRUNC, MVT::f64, Legal);
918 setOperationAction(ISD::FNEARBYINT, MVT::f64, Legal);
919 setOperationAction(ISD::FRINT, MVT::f64, Legal);
922 // We have target-specific dag combine patterns for the following nodes:
923 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
924 setTargetDAGCombine(ISD::ADD);
925 setTargetDAGCombine(ISD::SUB);
926 setTargetDAGCombine(ISD::MUL);
927 setTargetDAGCombine(ISD::AND);
928 setTargetDAGCombine(ISD::OR);
929 setTargetDAGCombine(ISD::XOR);
931 if (Subtarget->hasV6Ops())
932 setTargetDAGCombine(ISD::SRL);
934 setStackPointerRegisterToSaveRestore(ARM::SP);
936 if (TM.Options.UseSoftFloat || Subtarget->isThumb1Only() ||
937 !Subtarget->hasVFP2())
938 setSchedulingPreference(Sched::RegPressure);
940 setSchedulingPreference(Sched::Hybrid);
942 //// temporary - rewrite interface to use type
943 MaxStoresPerMemset = 8;
944 MaxStoresPerMemsetOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
945 MaxStoresPerMemcpy = 4; // For @llvm.memcpy -> sequence of stores
946 MaxStoresPerMemcpyOptSize = Subtarget->isTargetDarwin() ? 4 : 2;
947 MaxStoresPerMemmove = 4; // For @llvm.memmove -> sequence of stores
948 MaxStoresPerMemmoveOptSize = Subtarget->isTargetDarwin() ? 4 : 2;
950 // On ARM arguments smaller than 4 bytes are extended, so all arguments
951 // are at least 4 bytes aligned.
952 setMinStackArgumentAlignment(4);
954 // Prefer likely predicted branches to selects on out-of-order cores.
955 PredictableSelectIsExpensive = Subtarget->isLikeA9();
957 setMinFunctionAlignment(Subtarget->isThumb() ? 1 : 2);
960 // FIXME: It might make sense to define the representative register class as the
961 // nearest super-register that has a non-null superset. For example, DPR_VFP2 is
962 // a super-register of SPR, and DPR is a superset if DPR_VFP2. Consequently,
963 // SPR's representative would be DPR_VFP2. This should work well if register
964 // pressure tracking were modified such that a register use would increment the
965 // pressure of the register class's representative and all of it's super
966 // classes' representatives transitively. We have not implemented this because
967 // of the difficulty prior to coalescing of modeling operand register classes
968 // due to the common occurrence of cross class copies and subregister insertions
970 std::pair<const TargetRegisterClass*, uint8_t>
971 ARMTargetLowering::findRepresentativeClass(MVT VT) const{
972 const TargetRegisterClass *RRC = nullptr;
974 switch (VT.SimpleTy) {
976 return TargetLowering::findRepresentativeClass(VT);
977 // Use DPR as representative register class for all floating point
978 // and vector types. Since there are 32 SPR registers and 32 DPR registers so
979 // the cost is 1 for both f32 and f64.
980 case MVT::f32: case MVT::f64: case MVT::v8i8: case MVT::v4i16:
981 case MVT::v2i32: case MVT::v1i64: case MVT::v2f32:
982 RRC = &ARM::DPRRegClass;
983 // When NEON is used for SP, only half of the register file is available
984 // because operations that define both SP and DP results will be constrained
985 // to the VFP2 class (D0-D15). We currently model this constraint prior to
986 // coalescing by double-counting the SP regs. See the FIXME above.
987 if (Subtarget->useNEONForSinglePrecisionFP())
990 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
991 case MVT::v4f32: case MVT::v2f64:
992 RRC = &ARM::DPRRegClass;
996 RRC = &ARM::DPRRegClass;
1000 RRC = &ARM::DPRRegClass;
1004 return std::make_pair(RRC, Cost);
1007 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
1009 default: return nullptr;
1010 case ARMISD::Wrapper: return "ARMISD::Wrapper";
1011 case ARMISD::WrapperPIC: return "ARMISD::WrapperPIC";
1012 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
1013 case ARMISD::CALL: return "ARMISD::CALL";
1014 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
1015 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
1016 case ARMISD::tCALL: return "ARMISD::tCALL";
1017 case ARMISD::BRCOND: return "ARMISD::BRCOND";
1018 case ARMISD::BR_JT: return "ARMISD::BR_JT";
1019 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
1020 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
1021 case ARMISD::INTRET_FLAG: return "ARMISD::INTRET_FLAG";
1022 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
1023 case ARMISD::CMP: return "ARMISD::CMP";
1024 case ARMISD::CMN: return "ARMISD::CMN";
1025 case ARMISD::CMPZ: return "ARMISD::CMPZ";
1026 case ARMISD::CMPFP: return "ARMISD::CMPFP";
1027 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
1028 case ARMISD::BCC_i64: return "ARMISD::BCC_i64";
1029 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
1031 case ARMISD::CMOV: return "ARMISD::CMOV";
1033 case ARMISD::RBIT: return "ARMISD::RBIT";
1035 case ARMISD::FTOSI: return "ARMISD::FTOSI";
1036 case ARMISD::FTOUI: return "ARMISD::FTOUI";
1037 case ARMISD::SITOF: return "ARMISD::SITOF";
1038 case ARMISD::UITOF: return "ARMISD::UITOF";
1040 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
1041 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
1042 case ARMISD::RRX: return "ARMISD::RRX";
1044 case ARMISD::ADDC: return "ARMISD::ADDC";
1045 case ARMISD::ADDE: return "ARMISD::ADDE";
1046 case ARMISD::SUBC: return "ARMISD::SUBC";
1047 case ARMISD::SUBE: return "ARMISD::SUBE";
1049 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
1050 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
1052 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
1053 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
1055 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN";
1057 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
1059 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
1061 case ARMISD::MEMBARRIER_MCR: return "ARMISD::MEMBARRIER_MCR";
1063 case ARMISD::PRELOAD: return "ARMISD::PRELOAD";
1065 case ARMISD::WIN__CHKSTK: return "ARMISD:::WIN__CHKSTK";
1067 case ARMISD::VCEQ: return "ARMISD::VCEQ";
1068 case ARMISD::VCEQZ: return "ARMISD::VCEQZ";
1069 case ARMISD::VCGE: return "ARMISD::VCGE";
1070 case ARMISD::VCGEZ: return "ARMISD::VCGEZ";
1071 case ARMISD::VCLEZ: return "ARMISD::VCLEZ";
1072 case ARMISD::VCGEU: return "ARMISD::VCGEU";
1073 case ARMISD::VCGT: return "ARMISD::VCGT";
1074 case ARMISD::VCGTZ: return "ARMISD::VCGTZ";
1075 case ARMISD::VCLTZ: return "ARMISD::VCLTZ";
1076 case ARMISD::VCGTU: return "ARMISD::VCGTU";
1077 case ARMISD::VTST: return "ARMISD::VTST";
1079 case ARMISD::VSHL: return "ARMISD::VSHL";
1080 case ARMISD::VSHRs: return "ARMISD::VSHRs";
1081 case ARMISD::VSHRu: return "ARMISD::VSHRu";
1082 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
1083 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
1084 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
1085 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
1086 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
1087 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
1088 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
1089 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
1090 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
1091 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
1092 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
1093 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
1094 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
1095 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
1096 case ARMISD::VMOVIMM: return "ARMISD::VMOVIMM";
1097 case ARMISD::VMVNIMM: return "ARMISD::VMVNIMM";
1098 case ARMISD::VMOVFPIMM: return "ARMISD::VMOVFPIMM";
1099 case ARMISD::VDUP: return "ARMISD::VDUP";
1100 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
1101 case ARMISD::VEXT: return "ARMISD::VEXT";
1102 case ARMISD::VREV64: return "ARMISD::VREV64";
1103 case ARMISD::VREV32: return "ARMISD::VREV32";
1104 case ARMISD::VREV16: return "ARMISD::VREV16";
1105 case ARMISD::VZIP: return "ARMISD::VZIP";
1106 case ARMISD::VUZP: return "ARMISD::VUZP";
1107 case ARMISD::VTRN: return "ARMISD::VTRN";
1108 case ARMISD::VTBL1: return "ARMISD::VTBL1";
1109 case ARMISD::VTBL2: return "ARMISD::VTBL2";
1110 case ARMISD::VMULLs: return "ARMISD::VMULLs";
1111 case ARMISD::VMULLu: return "ARMISD::VMULLu";
1112 case ARMISD::UMLAL: return "ARMISD::UMLAL";
1113 case ARMISD::SMLAL: return "ARMISD::SMLAL";
1114 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
1115 case ARMISD::FMAX: return "ARMISD::FMAX";
1116 case ARMISD::FMIN: return "ARMISD::FMIN";
1117 case ARMISD::VMAXNM: return "ARMISD::VMAX";
1118 case ARMISD::VMINNM: return "ARMISD::VMIN";
1119 case ARMISD::BFI: return "ARMISD::BFI";
1120 case ARMISD::VORRIMM: return "ARMISD::VORRIMM";
1121 case ARMISD::VBICIMM: return "ARMISD::VBICIMM";
1122 case ARMISD::VBSL: return "ARMISD::VBSL";
1123 case ARMISD::VLD2DUP: return "ARMISD::VLD2DUP";
1124 case ARMISD::VLD3DUP: return "ARMISD::VLD3DUP";
1125 case ARMISD::VLD4DUP: return "ARMISD::VLD4DUP";
1126 case ARMISD::VLD1_UPD: return "ARMISD::VLD1_UPD";
1127 case ARMISD::VLD2_UPD: return "ARMISD::VLD2_UPD";
1128 case ARMISD::VLD3_UPD: return "ARMISD::VLD3_UPD";
1129 case ARMISD::VLD4_UPD: return "ARMISD::VLD4_UPD";
1130 case ARMISD::VLD2LN_UPD: return "ARMISD::VLD2LN_UPD";
1131 case ARMISD::VLD3LN_UPD: return "ARMISD::VLD3LN_UPD";
1132 case ARMISD::VLD4LN_UPD: return "ARMISD::VLD4LN_UPD";
1133 case ARMISD::VLD2DUP_UPD: return "ARMISD::VLD2DUP_UPD";
1134 case ARMISD::VLD3DUP_UPD: return "ARMISD::VLD3DUP_UPD";
1135 case ARMISD::VLD4DUP_UPD: return "ARMISD::VLD4DUP_UPD";
1136 case ARMISD::VST1_UPD: return "ARMISD::VST1_UPD";
1137 case ARMISD::VST2_UPD: return "ARMISD::VST2_UPD";
1138 case ARMISD::VST3_UPD: return "ARMISD::VST3_UPD";
1139 case ARMISD::VST4_UPD: return "ARMISD::VST4_UPD";
1140 case ARMISD::VST2LN_UPD: return "ARMISD::VST2LN_UPD";
1141 case ARMISD::VST3LN_UPD: return "ARMISD::VST3LN_UPD";
1142 case ARMISD::VST4LN_UPD: return "ARMISD::VST4LN_UPD";
1146 EVT ARMTargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
1147 if (!VT.isVector()) return getPointerTy();
1148 return VT.changeVectorElementTypeToInteger();
1151 /// getRegClassFor - Return the register class that should be used for the
1152 /// specified value type.
1153 const TargetRegisterClass *ARMTargetLowering::getRegClassFor(MVT VT) const {
1154 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
1155 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
1156 // load / store 4 to 8 consecutive D registers.
1157 if (Subtarget->hasNEON()) {
1158 if (VT == MVT::v4i64)
1159 return &ARM::QQPRRegClass;
1160 if (VT == MVT::v8i64)
1161 return &ARM::QQQQPRRegClass;
1163 return TargetLowering::getRegClassFor(VT);
1166 // Create a fast isel object.
1168 ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
1169 const TargetLibraryInfo *libInfo) const {
1170 return ARM::createFastISel(funcInfo, libInfo);
1173 /// getMaximalGlobalOffset - Returns the maximal possible offset which can
1174 /// be used for loads / stores from the global.
1175 unsigned ARMTargetLowering::getMaximalGlobalOffset() const {
1176 return (Subtarget->isThumb1Only() ? 127 : 4095);
1179 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
1180 unsigned NumVals = N->getNumValues();
1182 return Sched::RegPressure;
1184 for (unsigned i = 0; i != NumVals; ++i) {
1185 EVT VT = N->getValueType(i);
1186 if (VT == MVT::Glue || VT == MVT::Other)
1188 if (VT.isFloatingPoint() || VT.isVector())
1192 if (!N->isMachineOpcode())
1193 return Sched::RegPressure;
1195 // Load are scheduled for latency even if there instruction itinerary
1196 // is not available.
1197 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
1198 const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
1200 if (MCID.getNumDefs() == 0)
1201 return Sched::RegPressure;
1202 if (!Itins->isEmpty() &&
1203 Itins->getOperandCycle(MCID.getSchedClass(), 0) > 2)
1206 return Sched::RegPressure;
1209 //===----------------------------------------------------------------------===//
1211 //===----------------------------------------------------------------------===//
1213 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
1214 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
1216 default: llvm_unreachable("Unknown condition code!");
1217 case ISD::SETNE: return ARMCC::NE;
1218 case ISD::SETEQ: return ARMCC::EQ;
1219 case ISD::SETGT: return ARMCC::GT;
1220 case ISD::SETGE: return ARMCC::GE;
1221 case ISD::SETLT: return ARMCC::LT;
1222 case ISD::SETLE: return ARMCC::LE;
1223 case ISD::SETUGT: return ARMCC::HI;
1224 case ISD::SETUGE: return ARMCC::HS;
1225 case ISD::SETULT: return ARMCC::LO;
1226 case ISD::SETULE: return ARMCC::LS;
1230 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
1231 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
1232 ARMCC::CondCodes &CondCode2) {
1233 CondCode2 = ARMCC::AL;
1235 default: llvm_unreachable("Unknown FP condition!");
1237 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
1239 case ISD::SETOGT: CondCode = ARMCC::GT; break;
1241 case ISD::SETOGE: CondCode = ARMCC::GE; break;
1242 case ISD::SETOLT: CondCode = ARMCC::MI; break;
1243 case ISD::SETOLE: CondCode = ARMCC::LS; break;
1244 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
1245 case ISD::SETO: CondCode = ARMCC::VC; break;
1246 case ISD::SETUO: CondCode = ARMCC::VS; break;
1247 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
1248 case ISD::SETUGT: CondCode = ARMCC::HI; break;
1249 case ISD::SETUGE: CondCode = ARMCC::PL; break;
1251 case ISD::SETULT: CondCode = ARMCC::LT; break;
1253 case ISD::SETULE: CondCode = ARMCC::LE; break;
1255 case ISD::SETUNE: CondCode = ARMCC::NE; break;
1259 //===----------------------------------------------------------------------===//
1260 // Calling Convention Implementation
1261 //===----------------------------------------------------------------------===//
1263 #include "ARMGenCallingConv.inc"
1265 /// getEffectiveCallingConv - Get the effective calling convention, taking into
1266 /// account presence of floating point hardware and calling convention
1267 /// limitations, such as support for variadic functions.
1269 ARMTargetLowering::getEffectiveCallingConv(CallingConv::ID CC,
1270 bool isVarArg) const {
1273 llvm_unreachable("Unsupported calling convention");
1274 case CallingConv::ARM_AAPCS:
1275 case CallingConv::ARM_APCS:
1276 case CallingConv::GHC:
1278 case CallingConv::ARM_AAPCS_VFP:
1279 return isVarArg ? CallingConv::ARM_AAPCS : CallingConv::ARM_AAPCS_VFP;
1280 case CallingConv::C:
1281 if (!Subtarget->isAAPCS_ABI())
1282 return CallingConv::ARM_APCS;
1283 else if (Subtarget->hasVFP2() && !Subtarget->isThumb1Only() &&
1284 getTargetMachine().Options.FloatABIType == FloatABI::Hard &&
1286 return CallingConv::ARM_AAPCS_VFP;
1288 return CallingConv::ARM_AAPCS;
1289 case CallingConv::Fast:
1290 if (!Subtarget->isAAPCS_ABI()) {
1291 if (Subtarget->hasVFP2() && !Subtarget->isThumb1Only() && !isVarArg)
1292 return CallingConv::Fast;
1293 return CallingConv::ARM_APCS;
1294 } else if (Subtarget->hasVFP2() && !Subtarget->isThumb1Only() && !isVarArg)
1295 return CallingConv::ARM_AAPCS_VFP;
1297 return CallingConv::ARM_AAPCS;
1301 /// CCAssignFnForNode - Selects the correct CCAssignFn for the given
1302 /// CallingConvention.
1303 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
1305 bool isVarArg) const {
1306 switch (getEffectiveCallingConv(CC, isVarArg)) {
1308 llvm_unreachable("Unsupported calling convention");
1309 case CallingConv::ARM_APCS:
1310 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1311 case CallingConv::ARM_AAPCS:
1312 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1313 case CallingConv::ARM_AAPCS_VFP:
1314 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1315 case CallingConv::Fast:
1316 return (Return ? RetFastCC_ARM_APCS : FastCC_ARM_APCS);
1317 case CallingConv::GHC:
1318 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS_GHC);
1322 /// LowerCallResult - Lower the result values of a call into the
1323 /// appropriate copies out of appropriate physical registers.
1325 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1326 CallingConv::ID CallConv, bool isVarArg,
1327 const SmallVectorImpl<ISD::InputArg> &Ins,
1328 SDLoc dl, SelectionDAG &DAG,
1329 SmallVectorImpl<SDValue> &InVals,
1330 bool isThisReturn, SDValue ThisVal) const {
1332 // Assign locations to each value returned by this call.
1333 SmallVector<CCValAssign, 16> RVLocs;
1334 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
1335 *DAG.getContext(), Call);
1336 CCInfo.AnalyzeCallResult(Ins,
1337 CCAssignFnForNode(CallConv, /* Return*/ true,
1340 // Copy all of the result registers out of their specified physreg.
1341 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1342 CCValAssign VA = RVLocs[i];
1344 // Pass 'this' value directly from the argument to return value, to avoid
1345 // reg unit interference
1346 if (i == 0 && isThisReturn) {
1347 assert(!VA.needsCustom() && VA.getLocVT() == MVT::i32 &&
1348 "unexpected return calling convention register assignment");
1349 InVals.push_back(ThisVal);
1354 if (VA.needsCustom()) {
1355 // Handle f64 or half of a v2f64.
1356 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1358 Chain = Lo.getValue(1);
1359 InFlag = Lo.getValue(2);
1360 VA = RVLocs[++i]; // skip ahead to next loc
1361 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1363 Chain = Hi.getValue(1);
1364 InFlag = Hi.getValue(2);
1365 if (!Subtarget->isLittle())
1367 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1369 if (VA.getLocVT() == MVT::v2f64) {
1370 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1371 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1372 DAG.getConstant(0, MVT::i32));
1374 VA = RVLocs[++i]; // skip ahead to next loc
1375 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1376 Chain = Lo.getValue(1);
1377 InFlag = Lo.getValue(2);
1378 VA = RVLocs[++i]; // skip ahead to next loc
1379 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1380 Chain = Hi.getValue(1);
1381 InFlag = Hi.getValue(2);
1382 if (!Subtarget->isLittle())
1384 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1385 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1386 DAG.getConstant(1, MVT::i32));
1389 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
1391 Chain = Val.getValue(1);
1392 InFlag = Val.getValue(2);
1395 switch (VA.getLocInfo()) {
1396 default: llvm_unreachable("Unknown loc info!");
1397 case CCValAssign::Full: break;
1398 case CCValAssign::BCvt:
1399 Val = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), Val);
1403 InVals.push_back(Val);
1409 /// LowerMemOpCallTo - Store the argument to the stack.
1411 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
1412 SDValue StackPtr, SDValue Arg,
1413 SDLoc dl, SelectionDAG &DAG,
1414 const CCValAssign &VA,
1415 ISD::ArgFlagsTy Flags) const {
1416 unsigned LocMemOffset = VA.getLocMemOffset();
1417 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1418 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1419 return DAG.getStore(Chain, dl, Arg, PtrOff,
1420 MachinePointerInfo::getStack(LocMemOffset),
1424 void ARMTargetLowering::PassF64ArgInRegs(SDLoc dl, SelectionDAG &DAG,
1425 SDValue Chain, SDValue &Arg,
1426 RegsToPassVector &RegsToPass,
1427 CCValAssign &VA, CCValAssign &NextVA,
1429 SmallVectorImpl<SDValue> &MemOpChains,
1430 ISD::ArgFlagsTy Flags) const {
1432 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1433 DAG.getVTList(MVT::i32, MVT::i32), Arg);
1434 unsigned id = Subtarget->isLittle() ? 0 : 1;
1435 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd.getValue(id)));
1437 if (NextVA.isRegLoc())
1438 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1-id)));
1440 assert(NextVA.isMemLoc());
1441 if (!StackPtr.getNode())
1442 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1444 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1-id),
1450 /// LowerCall - Lowering a call into a callseq_start <-
1451 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
1454 ARMTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
1455 SmallVectorImpl<SDValue> &InVals) const {
1456 SelectionDAG &DAG = CLI.DAG;
1458 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
1459 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
1460 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
1461 SDValue Chain = CLI.Chain;
1462 SDValue Callee = CLI.Callee;
1463 bool &isTailCall = CLI.IsTailCall;
1464 CallingConv::ID CallConv = CLI.CallConv;
1465 bool doesNotRet = CLI.DoesNotReturn;
1466 bool isVarArg = CLI.IsVarArg;
1468 MachineFunction &MF = DAG.getMachineFunction();
1469 bool isStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1470 bool isThisReturn = false;
1471 bool isSibCall = false;
1473 // Disable tail calls if they're not supported.
1474 if (!Subtarget->supportsTailCall() || MF.getTarget().Options.DisableTailCalls)
1478 // Check if it's really possible to do a tail call.
1479 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1480 isVarArg, isStructRet, MF.getFunction()->hasStructRetAttr(),
1481 Outs, OutVals, Ins, DAG);
1482 if (!isTailCall && CLI.CS && CLI.CS->isMustTailCall())
1483 report_fatal_error("failed to perform tail call elimination on a call "
1484 "site marked musttail");
1485 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1486 // detected sibcalls.
1493 // Analyze operands of the call, assigning locations to each operand.
1494 SmallVector<CCValAssign, 16> ArgLocs;
1495 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs,
1496 *DAG.getContext(), Call);
1497 CCInfo.AnalyzeCallOperands(Outs,
1498 CCAssignFnForNode(CallConv, /* Return*/ false,
1501 // Get a count of how many bytes are to be pushed on the stack.
1502 unsigned NumBytes = CCInfo.getNextStackOffset();
1504 // For tail calls, memory operands are available in our caller's stack.
1508 // Adjust the stack pointer for the new arguments...
1509 // These operations are automatically eliminated by the prolog/epilog pass
1511 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true),
1514 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1516 RegsToPassVector RegsToPass;
1517 SmallVector<SDValue, 8> MemOpChains;
1519 // Walk the register/memloc assignments, inserting copies/loads. In the case
1520 // of tail call optimization, arguments are handled later.
1521 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1523 ++i, ++realArgIdx) {
1524 CCValAssign &VA = ArgLocs[i];
1525 SDValue Arg = OutVals[realArgIdx];
1526 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1527 bool isByVal = Flags.isByVal();
1529 // Promote the value if needed.
1530 switch (VA.getLocInfo()) {
1531 default: llvm_unreachable("Unknown loc info!");
1532 case CCValAssign::Full: break;
1533 case CCValAssign::SExt:
1534 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1536 case CCValAssign::ZExt:
1537 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1539 case CCValAssign::AExt:
1540 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1542 case CCValAssign::BCvt:
1543 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1547 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1548 if (VA.needsCustom()) {
1549 if (VA.getLocVT() == MVT::v2f64) {
1550 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1551 DAG.getConstant(0, MVT::i32));
1552 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1553 DAG.getConstant(1, MVT::i32));
1555 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1556 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1558 VA = ArgLocs[++i]; // skip ahead to next loc
1559 if (VA.isRegLoc()) {
1560 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1561 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1563 assert(VA.isMemLoc());
1565 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1566 dl, DAG, VA, Flags));
1569 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1570 StackPtr, MemOpChains, Flags);
1572 } else if (VA.isRegLoc()) {
1573 if (realArgIdx == 0 && Flags.isReturned() && Outs[0].VT == MVT::i32) {
1574 assert(VA.getLocVT() == MVT::i32 &&
1575 "unexpected calling convention register assignment");
1576 assert(!Ins.empty() && Ins[0].VT == MVT::i32 &&
1577 "unexpected use of 'returned'");
1578 isThisReturn = true;
1580 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1581 } else if (isByVal) {
1582 assert(VA.isMemLoc());
1583 unsigned offset = 0;
1585 // True if this byval aggregate will be split between registers
1587 unsigned ByValArgsCount = CCInfo.getInRegsParamsCount();
1588 unsigned CurByValIdx = CCInfo.getInRegsParamsProcessed();
1590 if (CurByValIdx < ByValArgsCount) {
1592 unsigned RegBegin, RegEnd;
1593 CCInfo.getInRegsParamInfo(CurByValIdx, RegBegin, RegEnd);
1595 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1597 for (i = 0, j = RegBegin; j < RegEnd; i++, j++) {
1598 SDValue Const = DAG.getConstant(4*i, MVT::i32);
1599 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
1600 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
1601 MachinePointerInfo(),
1602 false, false, false,
1603 DAG.InferPtrAlignment(AddArg));
1604 MemOpChains.push_back(Load.getValue(1));
1605 RegsToPass.push_back(std::make_pair(j, Load));
1608 // If parameter size outsides register area, "offset" value
1609 // helps us to calculate stack slot for remained part properly.
1610 offset = RegEnd - RegBegin;
1612 CCInfo.nextInRegsParam();
1615 if (Flags.getByValSize() > 4*offset) {
1616 unsigned LocMemOffset = VA.getLocMemOffset();
1617 SDValue StkPtrOff = DAG.getIntPtrConstant(LocMemOffset);
1618 SDValue Dst = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr,
1620 SDValue SrcOffset = DAG.getIntPtrConstant(4*offset);
1621 SDValue Src = DAG.getNode(ISD::ADD, dl, getPointerTy(), Arg, SrcOffset);
1622 SDValue SizeNode = DAG.getConstant(Flags.getByValSize() - 4*offset,
1624 SDValue AlignNode = DAG.getConstant(Flags.getByValAlign(), MVT::i32);
1626 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
1627 SDValue Ops[] = { Chain, Dst, Src, SizeNode, AlignNode};
1628 MemOpChains.push_back(DAG.getNode(ARMISD::COPY_STRUCT_BYVAL, dl, VTs,
1631 } else if (!isSibCall) {
1632 assert(VA.isMemLoc());
1634 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1635 dl, DAG, VA, Flags));
1639 if (!MemOpChains.empty())
1640 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains);
1642 // Build a sequence of copy-to-reg nodes chained together with token chain
1643 // and flag operands which copy the outgoing args into the appropriate regs.
1645 // Tail call byval lowering might overwrite argument registers so in case of
1646 // tail call optimization the copies to registers are lowered later.
1648 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1649 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1650 RegsToPass[i].second, InFlag);
1651 InFlag = Chain.getValue(1);
1654 // For tail calls lower the arguments to the 'real' stack slot.
1656 // Force all the incoming stack arguments to be loaded from the stack
1657 // before any new outgoing arguments are stored to the stack, because the
1658 // outgoing stack slots may alias the incoming argument stack slots, and
1659 // the alias isn't otherwise explicit. This is slightly more conservative
1660 // than necessary, because it means that each store effectively depends
1661 // on every argument instead of just those arguments it would clobber.
1663 // Do not flag preceding copytoreg stuff together with the following stuff.
1665 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1666 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1667 RegsToPass[i].second, InFlag);
1668 InFlag = Chain.getValue(1);
1673 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1674 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1675 // node so that legalize doesn't hack it.
1676 bool isDirect = false;
1677 bool isARMFunc = false;
1678 bool isLocalARMFunc = false;
1679 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1681 if (EnableARMLongCalls) {
1682 assert((Subtarget->isTargetWindows() ||
1683 getTargetMachine().getRelocationModel() == Reloc::Static) &&
1684 "long-calls with non-static relocation model!");
1685 // Handle a global address or an external symbol. If it's not one of
1686 // those, the target's already in a register, so we don't need to do
1688 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1689 const GlobalValue *GV = G->getGlobal();
1690 // Create a constant pool entry for the callee address
1691 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1692 ARMConstantPoolValue *CPV =
1693 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 0);
1695 // Get the address of the callee into a register
1696 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1697 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1698 Callee = DAG.getLoad(getPointerTy(), dl,
1699 DAG.getEntryNode(), CPAddr,
1700 MachinePointerInfo::getConstantPool(),
1701 false, false, false, 0);
1702 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1703 const char *Sym = S->getSymbol();
1705 // Create a constant pool entry for the callee address
1706 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1707 ARMConstantPoolValue *CPV =
1708 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1709 ARMPCLabelIndex, 0);
1710 // Get the address of the callee into a register
1711 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1712 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1713 Callee = DAG.getLoad(getPointerTy(), dl,
1714 DAG.getEntryNode(), CPAddr,
1715 MachinePointerInfo::getConstantPool(),
1716 false, false, false, 0);
1718 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1719 const GlobalValue *GV = G->getGlobal();
1721 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1722 bool isStub = (isExt && Subtarget->isTargetMachO()) &&
1723 getTargetMachine().getRelocationModel() != Reloc::Static;
1724 isARMFunc = !Subtarget->isThumb() || (isStub && !Subtarget->isMClass());
1725 // ARM call to a local ARM function is predicable.
1726 isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking);
1727 // tBX takes a register source operand.
1728 if (isStub && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1729 assert(Subtarget->isTargetMachO() && "WrapperPIC use on non-MachO?");
1730 Callee = DAG.getNode(ARMISD::WrapperPIC, dl, getPointerTy(),
1731 DAG.getTargetGlobalAddress(GV, dl, getPointerTy(),
1732 0, ARMII::MO_NONLAZY));
1733 Callee = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Callee,
1734 MachinePointerInfo::getGOT(), false, false, true, 0);
1735 } else if (Subtarget->isTargetCOFF()) {
1736 assert(Subtarget->isTargetWindows() &&
1737 "Windows is the only supported COFF target");
1738 unsigned TargetFlags = GV->hasDLLImportStorageClass()
1739 ? ARMII::MO_DLLIMPORT
1740 : ARMII::MO_NO_FLAG;
1741 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), /*Offset=*/0,
1743 if (GV->hasDLLImportStorageClass())
1744 Callee = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
1745 DAG.getNode(ARMISD::Wrapper, dl, getPointerTy(),
1746 Callee), MachinePointerInfo::getGOT(),
1747 false, false, false, 0);
1749 // On ELF targets for PIC code, direct calls should go through the PLT
1750 unsigned OpFlags = 0;
1751 if (Subtarget->isTargetELF() &&
1752 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1753 OpFlags = ARMII::MO_PLT;
1754 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
1756 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1758 bool isStub = Subtarget->isTargetMachO() &&
1759 getTargetMachine().getRelocationModel() != Reloc::Static;
1760 isARMFunc = !Subtarget->isThumb() || (isStub && !Subtarget->isMClass());
1761 // tBX takes a register source operand.
1762 const char *Sym = S->getSymbol();
1763 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1764 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1765 ARMConstantPoolValue *CPV =
1766 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1767 ARMPCLabelIndex, 4);
1768 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1769 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1770 Callee = DAG.getLoad(getPointerTy(), dl,
1771 DAG.getEntryNode(), CPAddr,
1772 MachinePointerInfo::getConstantPool(),
1773 false, false, false, 0);
1774 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1775 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1776 getPointerTy(), Callee, PICLabel);
1778 unsigned OpFlags = 0;
1779 // On ELF targets for PIC code, direct calls should go through the PLT
1780 if (Subtarget->isTargetELF() &&
1781 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1782 OpFlags = ARMII::MO_PLT;
1783 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlags);
1787 // FIXME: handle tail calls differently.
1789 bool HasMinSizeAttr = MF.getFunction()->getAttributes().hasAttribute(
1790 AttributeSet::FunctionIndex, Attribute::MinSize);
1791 if (Subtarget->isThumb()) {
1792 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1793 CallOpc = ARMISD::CALL_NOLINK;
1795 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1797 if (!isDirect && !Subtarget->hasV5TOps())
1798 CallOpc = ARMISD::CALL_NOLINK;
1799 else if (doesNotRet && isDirect && Subtarget->hasRAS() &&
1800 // Emit regular call when code size is the priority
1802 // "mov lr, pc; b _foo" to avoid confusing the RSP
1803 CallOpc = ARMISD::CALL_NOLINK;
1805 CallOpc = isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL;
1808 std::vector<SDValue> Ops;
1809 Ops.push_back(Chain);
1810 Ops.push_back(Callee);
1812 // Add argument registers to the end of the list so that they are known live
1814 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1815 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1816 RegsToPass[i].second.getValueType()));
1818 // Add a register mask operand representing the call-preserved registers.
1820 const uint32_t *Mask;
1821 const ARMBaseRegisterInfo *ARI = Subtarget->getRegisterInfo();
1823 // For 'this' returns, use the R0-preserving mask if applicable
1824 Mask = ARI->getThisReturnPreservedMask(CallConv);
1826 // Set isThisReturn to false if the calling convention is not one that
1827 // allows 'returned' to be modeled in this way, so LowerCallResult does
1828 // not try to pass 'this' straight through
1829 isThisReturn = false;
1830 Mask = ARI->getCallPreservedMask(CallConv);
1833 Mask = ARI->getCallPreservedMask(CallConv);
1835 assert(Mask && "Missing call preserved mask for calling convention");
1836 Ops.push_back(DAG.getRegisterMask(Mask));
1839 if (InFlag.getNode())
1840 Ops.push_back(InFlag);
1842 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
1844 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, Ops);
1846 // Returns a chain and a flag for retval copy to use.
1847 Chain = DAG.getNode(CallOpc, dl, NodeTys, Ops);
1848 InFlag = Chain.getValue(1);
1850 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1851 DAG.getIntPtrConstant(0, true), InFlag, dl);
1853 InFlag = Chain.getValue(1);
1855 // Handle result values, copying them out of physregs into vregs that we
1857 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins, dl, DAG,
1858 InVals, isThisReturn,
1859 isThisReturn ? OutVals[0] : SDValue());
1862 /// HandleByVal - Every parameter *after* a byval parameter is passed
1863 /// on the stack. Remember the next parameter register to allocate,
1864 /// and then confiscate the rest of the parameter registers to insure
1867 ARMTargetLowering::HandleByVal(
1868 CCState *State, unsigned &size, unsigned Align) const {
1869 unsigned reg = State->AllocateReg(GPRArgRegs, 4);
1870 assert((State->getCallOrPrologue() == Prologue ||
1871 State->getCallOrPrologue() == Call) &&
1872 "unhandled ParmContext");
1874 if ((ARM::R0 <= reg) && (reg <= ARM::R3)) {
1875 if (Subtarget->isAAPCS_ABI() && Align > 4) {
1876 unsigned AlignInRegs = Align / 4;
1877 unsigned Waste = (ARM::R4 - reg) % AlignInRegs;
1878 for (unsigned i = 0; i < Waste; ++i)
1879 reg = State->AllocateReg(GPRArgRegs, 4);
1882 unsigned excess = 4 * (ARM::R4 - reg);
1884 // Special case when NSAA != SP and parameter size greater than size of
1885 // all remained GPR regs. In that case we can't split parameter, we must
1886 // send it to stack. We also must set NCRN to R4, so waste all
1887 // remained registers.
1888 const unsigned NSAAOffset = State->getNextStackOffset();
1889 if (Subtarget->isAAPCS_ABI() && NSAAOffset != 0 && size > excess) {
1890 while (State->AllocateReg(GPRArgRegs, 4))
1895 // First register for byval parameter is the first register that wasn't
1896 // allocated before this method call, so it would be "reg".
1897 // If parameter is small enough to be saved in range [reg, r4), then
1898 // the end (first after last) register would be reg + param-size-in-regs,
1899 // else parameter would be splitted between registers and stack,
1900 // end register would be r4 in this case.
1901 unsigned ByValRegBegin = reg;
1902 unsigned ByValRegEnd = (size < excess) ? reg + size/4 : (unsigned)ARM::R4;
1903 State->addInRegsParamInfo(ByValRegBegin, ByValRegEnd);
1904 // Note, first register is allocated in the beginning of function already,
1905 // allocate remained amount of registers we need.
1906 for (unsigned i = reg+1; i != ByValRegEnd; ++i)
1907 State->AllocateReg(GPRArgRegs, 4);
1908 // A byval parameter that is split between registers and memory needs its
1909 // size truncated here.
1910 // In the case where the entire structure fits in registers, we set the
1911 // size in memory to zero.
1920 /// MatchingStackOffset - Return true if the given stack call argument is
1921 /// already available in the same position (relatively) of the caller's
1922 /// incoming argument stack.
1924 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1925 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1926 const TargetInstrInfo *TII) {
1927 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1929 if (Arg.getOpcode() == ISD::CopyFromReg) {
1930 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1931 if (!TargetRegisterInfo::isVirtualRegister(VR))
1933 MachineInstr *Def = MRI->getVRegDef(VR);
1936 if (!Flags.isByVal()) {
1937 if (!TII->isLoadFromStackSlot(Def, FI))
1942 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
1943 if (Flags.isByVal())
1944 // ByVal argument is passed in as a pointer but it's now being
1945 // dereferenced. e.g.
1946 // define @foo(%struct.X* %A) {
1947 // tail call @bar(%struct.X* byval %A)
1950 SDValue Ptr = Ld->getBasePtr();
1951 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
1954 FI = FINode->getIndex();
1958 assert(FI != INT_MAX);
1959 if (!MFI->isFixedObjectIndex(FI))
1961 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
1964 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
1965 /// for tail call optimization. Targets which want to do tail call
1966 /// optimization should implement this function.
1968 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1969 CallingConv::ID CalleeCC,
1971 bool isCalleeStructRet,
1972 bool isCallerStructRet,
1973 const SmallVectorImpl<ISD::OutputArg> &Outs,
1974 const SmallVectorImpl<SDValue> &OutVals,
1975 const SmallVectorImpl<ISD::InputArg> &Ins,
1976 SelectionDAG& DAG) const {
1977 const Function *CallerF = DAG.getMachineFunction().getFunction();
1978 CallingConv::ID CallerCC = CallerF->getCallingConv();
1979 bool CCMatch = CallerCC == CalleeCC;
1981 // Look for obvious safe cases to perform tail call optimization that do not
1982 // require ABI changes. This is what gcc calls sibcall.
1984 // Do not sibcall optimize vararg calls unless the call site is not passing
1986 if (isVarArg && !Outs.empty())
1989 // Exception-handling functions need a special set of instructions to indicate
1990 // a return to the hardware. Tail-calling another function would probably
1992 if (CallerF->hasFnAttribute("interrupt"))
1995 // Also avoid sibcall optimization if either caller or callee uses struct
1996 // return semantics.
1997 if (isCalleeStructRet || isCallerStructRet)
2000 // FIXME: Completely disable sibcall for Thumb1 since Thumb1RegisterInfo::
2001 // emitEpilogue is not ready for them. Thumb tail calls also use t2B, as
2002 // the Thumb1 16-bit unconditional branch doesn't have sufficient relocation
2003 // support in the assembler and linker to be used. This would need to be
2004 // fixed to fully support tail calls in Thumb1.
2006 // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take
2007 // LR. This means if we need to reload LR, it takes an extra instructions,
2008 // which outweighs the value of the tail call; but here we don't know yet
2009 // whether LR is going to be used. Probably the right approach is to
2010 // generate the tail call here and turn it back into CALL/RET in
2011 // emitEpilogue if LR is used.
2013 // Thumb1 PIC calls to external symbols use BX, so they can be tail calls,
2014 // but we need to make sure there are enough registers; the only valid
2015 // registers are the 4 used for parameters. We don't currently do this
2017 if (Subtarget->isThumb1Only())
2020 // Externally-defined functions with weak linkage should not be
2021 // tail-called on ARM when the OS does not support dynamic
2022 // pre-emption of symbols, as the AAELF spec requires normal calls
2023 // to undefined weak functions to be replaced with a NOP or jump to the
2024 // next instruction. The behaviour of branch instructions in this
2025 // situation (as used for tail calls) is implementation-defined, so we
2026 // cannot rely on the linker replacing the tail call with a return.
2027 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2028 const GlobalValue *GV = G->getGlobal();
2029 const Triple TT(getTargetMachine().getTargetTriple());
2030 if (GV->hasExternalWeakLinkage() &&
2031 (!TT.isOSWindows() || TT.isOSBinFormatELF() || TT.isOSBinFormatMachO()))
2035 // If the calling conventions do not match, then we'd better make sure the
2036 // results are returned in the same way as what the caller expects.
2038 SmallVector<CCValAssign, 16> RVLocs1;
2039 ARMCCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(), RVLocs1,
2040 *DAG.getContext(), Call);
2041 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
2043 SmallVector<CCValAssign, 16> RVLocs2;
2044 ARMCCState CCInfo2(CallerCC, false, DAG.getMachineFunction(), RVLocs2,
2045 *DAG.getContext(), Call);
2046 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
2048 if (RVLocs1.size() != RVLocs2.size())
2050 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
2051 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
2053 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
2055 if (RVLocs1[i].isRegLoc()) {
2056 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
2059 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
2065 // If Caller's vararg or byval argument has been split between registers and
2066 // stack, do not perform tail call, since part of the argument is in caller's
2068 const ARMFunctionInfo *AFI_Caller = DAG.getMachineFunction().
2069 getInfo<ARMFunctionInfo>();
2070 if (AFI_Caller->getArgRegsSaveSize())
2073 // If the callee takes no arguments then go on to check the results of the
2075 if (!Outs.empty()) {
2076 // Check if stack adjustment is needed. For now, do not do this if any
2077 // argument is passed on the stack.
2078 SmallVector<CCValAssign, 16> ArgLocs;
2079 ARMCCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
2080 *DAG.getContext(), Call);
2081 CCInfo.AnalyzeCallOperands(Outs,
2082 CCAssignFnForNode(CalleeCC, false, isVarArg));
2083 if (CCInfo.getNextStackOffset()) {
2084 MachineFunction &MF = DAG.getMachineFunction();
2086 // Check if the arguments are already laid out in the right way as
2087 // the caller's fixed stack objects.
2088 MachineFrameInfo *MFI = MF.getFrameInfo();
2089 const MachineRegisterInfo *MRI = &MF.getRegInfo();
2090 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
2091 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
2093 ++i, ++realArgIdx) {
2094 CCValAssign &VA = ArgLocs[i];
2095 EVT RegVT = VA.getLocVT();
2096 SDValue Arg = OutVals[realArgIdx];
2097 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
2098 if (VA.getLocInfo() == CCValAssign::Indirect)
2100 if (VA.needsCustom()) {
2101 // f64 and vector types are split into multiple registers or
2102 // register/stack-slot combinations. The types will not match
2103 // the registers; give up on memory f64 refs until we figure
2104 // out what to do about this.
2107 if (!ArgLocs[++i].isRegLoc())
2109 if (RegVT == MVT::v2f64) {
2110 if (!ArgLocs[++i].isRegLoc())
2112 if (!ArgLocs[++i].isRegLoc())
2115 } else if (!VA.isRegLoc()) {
2116 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
2128 ARMTargetLowering::CanLowerReturn(CallingConv::ID CallConv,
2129 MachineFunction &MF, bool isVarArg,
2130 const SmallVectorImpl<ISD::OutputArg> &Outs,
2131 LLVMContext &Context) const {
2132 SmallVector<CCValAssign, 16> RVLocs;
2133 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, Context);
2134 return CCInfo.CheckReturn(Outs, CCAssignFnForNode(CallConv, /*Return=*/true,
2138 static SDValue LowerInterruptReturn(SmallVectorImpl<SDValue> &RetOps,
2139 SDLoc DL, SelectionDAG &DAG) {
2140 const MachineFunction &MF = DAG.getMachineFunction();
2141 const Function *F = MF.getFunction();
2143 StringRef IntKind = F->getFnAttribute("interrupt").getValueAsString();
2145 // See ARM ARM v7 B1.8.3. On exception entry LR is set to a possibly offset
2146 // version of the "preferred return address". These offsets affect the return
2147 // instruction if this is a return from PL1 without hypervisor extensions.
2148 // IRQ/FIQ: +4 "subs pc, lr, #4"
2149 // SWI: 0 "subs pc, lr, #0"
2150 // ABORT: +4 "subs pc, lr, #4"
2151 // UNDEF: +4/+2 "subs pc, lr, #0"
2152 // UNDEF varies depending on where the exception came from ARM or Thumb
2153 // mode. Alongside GCC, we throw our hands up in disgust and pretend it's 0.
2156 if (IntKind == "" || IntKind == "IRQ" || IntKind == "FIQ" ||
2159 else if (IntKind == "SWI" || IntKind == "UNDEF")
2162 report_fatal_error("Unsupported interrupt attribute. If present, value "
2163 "must be one of: IRQ, FIQ, SWI, ABORT or UNDEF");
2165 RetOps.insert(RetOps.begin() + 1, DAG.getConstant(LROffset, MVT::i32, false));
2167 return DAG.getNode(ARMISD::INTRET_FLAG, DL, MVT::Other, RetOps);
2171 ARMTargetLowering::LowerReturn(SDValue Chain,
2172 CallingConv::ID CallConv, bool isVarArg,
2173 const SmallVectorImpl<ISD::OutputArg> &Outs,
2174 const SmallVectorImpl<SDValue> &OutVals,
2175 SDLoc dl, SelectionDAG &DAG) const {
2177 // CCValAssign - represent the assignment of the return value to a location.
2178 SmallVector<CCValAssign, 16> RVLocs;
2180 // CCState - Info about the registers and stack slots.
2181 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
2182 *DAG.getContext(), Call);
2184 // Analyze outgoing return values.
2185 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
2189 SmallVector<SDValue, 4> RetOps;
2190 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
2191 bool isLittleEndian = Subtarget->isLittle();
2193 MachineFunction &MF = DAG.getMachineFunction();
2194 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2195 AFI->setReturnRegsCount(RVLocs.size());
2197 // Copy the result values into the output registers.
2198 for (unsigned i = 0, realRVLocIdx = 0;
2200 ++i, ++realRVLocIdx) {
2201 CCValAssign &VA = RVLocs[i];
2202 assert(VA.isRegLoc() && "Can only return in registers!");
2204 SDValue Arg = OutVals[realRVLocIdx];
2206 switch (VA.getLocInfo()) {
2207 default: llvm_unreachable("Unknown loc info!");
2208 case CCValAssign::Full: break;
2209 case CCValAssign::BCvt:
2210 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
2214 if (VA.needsCustom()) {
2215 if (VA.getLocVT() == MVT::v2f64) {
2216 // Extract the first half and return it in two registers.
2217 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
2218 DAG.getConstant(0, MVT::i32));
2219 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
2220 DAG.getVTList(MVT::i32, MVT::i32), Half);
2222 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2223 HalfGPRs.getValue(isLittleEndian ? 0 : 1),
2225 Flag = Chain.getValue(1);
2226 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2227 VA = RVLocs[++i]; // skip ahead to next loc
2228 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2229 HalfGPRs.getValue(isLittleEndian ? 1 : 0),
2231 Flag = Chain.getValue(1);
2232 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2233 VA = RVLocs[++i]; // skip ahead to next loc
2235 // Extract the 2nd half and fall through to handle it as an f64 value.
2236 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
2237 DAG.getConstant(1, MVT::i32));
2239 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
2241 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
2242 DAG.getVTList(MVT::i32, MVT::i32), Arg);
2243 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2244 fmrrd.getValue(isLittleEndian ? 0 : 1),
2246 Flag = Chain.getValue(1);
2247 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2248 VA = RVLocs[++i]; // skip ahead to next loc
2249 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2250 fmrrd.getValue(isLittleEndian ? 1 : 0),
2253 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
2255 // Guarantee that all emitted copies are
2256 // stuck together, avoiding something bad.
2257 Flag = Chain.getValue(1);
2258 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2261 // Update chain and glue.
2264 RetOps.push_back(Flag);
2266 // CPUs which aren't M-class use a special sequence to return from
2267 // exceptions (roughly, any instruction setting pc and cpsr simultaneously,
2268 // though we use "subs pc, lr, #N").
2270 // M-class CPUs actually use a normal return sequence with a special
2271 // (hardware-provided) value in LR, so the normal code path works.
2272 if (DAG.getMachineFunction().getFunction()->hasFnAttribute("interrupt") &&
2273 !Subtarget->isMClass()) {
2274 if (Subtarget->isThumb1Only())
2275 report_fatal_error("interrupt attribute is not supported in Thumb1");
2276 return LowerInterruptReturn(RetOps, dl, DAG);
2279 return DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, RetOps);
2282 bool ARMTargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
2283 if (N->getNumValues() != 1)
2285 if (!N->hasNUsesOfValue(1, 0))
2288 SDValue TCChain = Chain;
2289 SDNode *Copy = *N->use_begin();
2290 if (Copy->getOpcode() == ISD::CopyToReg) {
2291 // If the copy has a glue operand, we conservatively assume it isn't safe to
2292 // perform a tail call.
2293 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
2295 TCChain = Copy->getOperand(0);
2296 } else if (Copy->getOpcode() == ARMISD::VMOVRRD) {
2297 SDNode *VMov = Copy;
2298 // f64 returned in a pair of GPRs.
2299 SmallPtrSet<SDNode*, 2> Copies;
2300 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
2302 if (UI->getOpcode() != ISD::CopyToReg)
2306 if (Copies.size() > 2)
2309 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
2311 SDValue UseChain = UI->getOperand(0);
2312 if (Copies.count(UseChain.getNode()))
2316 // We are at the top of this chain.
2317 // If the copy has a glue operand, we conservatively assume it
2318 // isn't safe to perform a tail call.
2319 if (UI->getOperand(UI->getNumOperands()-1).getValueType() == MVT::Glue)
2325 } else if (Copy->getOpcode() == ISD::BITCAST) {
2326 // f32 returned in a single GPR.
2327 if (!Copy->hasOneUse())
2329 Copy = *Copy->use_begin();
2330 if (Copy->getOpcode() != ISD::CopyToReg || !Copy->hasNUsesOfValue(1, 0))
2332 // If the copy has a glue operand, we conservatively assume it isn't safe to
2333 // perform a tail call.
2334 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
2336 TCChain = Copy->getOperand(0);
2341 bool HasRet = false;
2342 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
2344 if (UI->getOpcode() != ARMISD::RET_FLAG &&
2345 UI->getOpcode() != ARMISD::INTRET_FLAG)
2357 bool ARMTargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
2358 if (!Subtarget->supportsTailCall())
2361 if (!CI->isTailCall() || getTargetMachine().Options.DisableTailCalls)
2364 return !Subtarget->isThumb1Only();
2367 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
2368 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
2369 // one of the above mentioned nodes. It has to be wrapped because otherwise
2370 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
2371 // be used to form addressing mode. These wrapped nodes will be selected
2373 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
2374 EVT PtrVT = Op.getValueType();
2375 // FIXME there is no actual debug info here
2377 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
2379 if (CP->isMachineConstantPoolEntry())
2380 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
2381 CP->getAlignment());
2383 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
2384 CP->getAlignment());
2385 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
2388 unsigned ARMTargetLowering::getJumpTableEncoding() const {
2389 return MachineJumpTableInfo::EK_Inline;
2392 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
2393 SelectionDAG &DAG) const {
2394 MachineFunction &MF = DAG.getMachineFunction();
2395 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2396 unsigned ARMPCLabelIndex = 0;
2398 EVT PtrVT = getPointerTy();
2399 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
2400 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2402 if (RelocM == Reloc::Static) {
2403 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
2405 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2406 ARMPCLabelIndex = AFI->createPICLabelUId();
2407 ARMConstantPoolValue *CPV =
2408 ARMConstantPoolConstant::Create(BA, ARMPCLabelIndex,
2409 ARMCP::CPBlockAddress, PCAdj);
2410 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2412 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
2413 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
2414 MachinePointerInfo::getConstantPool(),
2415 false, false, false, 0);
2416 if (RelocM == Reloc::Static)
2418 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2419 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
2422 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
2424 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
2425 SelectionDAG &DAG) const {
2427 EVT PtrVT = getPointerTy();
2428 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2429 MachineFunction &MF = DAG.getMachineFunction();
2430 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2431 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2432 ARMConstantPoolValue *CPV =
2433 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2434 ARMCP::CPValue, PCAdj, ARMCP::TLSGD, true);
2435 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2436 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
2437 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
2438 MachinePointerInfo::getConstantPool(),
2439 false, false, false, 0);
2440 SDValue Chain = Argument.getValue(1);
2442 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2443 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
2445 // call __tls_get_addr.
2448 Entry.Node = Argument;
2449 Entry.Ty = (Type *) Type::getInt32Ty(*DAG.getContext());
2450 Args.push_back(Entry);
2452 // FIXME: is there useful debug info available here?
2453 TargetLowering::CallLoweringInfo CLI(DAG);
2454 CLI.setDebugLoc(dl).setChain(Chain)
2455 .setCallee(CallingConv::C, Type::getInt32Ty(*DAG.getContext()),
2456 DAG.getExternalSymbol("__tls_get_addr", PtrVT), std::move(Args),
2459 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
2460 return CallResult.first;
2463 // Lower ISD::GlobalTLSAddress using the "initial exec" or
2464 // "local exec" model.
2466 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
2468 TLSModel::Model model) const {
2469 const GlobalValue *GV = GA->getGlobal();
2472 SDValue Chain = DAG.getEntryNode();
2473 EVT PtrVT = getPointerTy();
2474 // Get the Thread Pointer
2475 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2477 if (model == TLSModel::InitialExec) {
2478 MachineFunction &MF = DAG.getMachineFunction();
2479 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2480 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2481 // Initial exec model.
2482 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2483 ARMConstantPoolValue *CPV =
2484 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2485 ARMCP::CPValue, PCAdj, ARMCP::GOTTPOFF,
2487 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2488 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2489 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2490 MachinePointerInfo::getConstantPool(),
2491 false, false, false, 0);
2492 Chain = Offset.getValue(1);
2494 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2495 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
2497 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2498 MachinePointerInfo::getConstantPool(),
2499 false, false, false, 0);
2502 assert(model == TLSModel::LocalExec);
2503 ARMConstantPoolValue *CPV =
2504 ARMConstantPoolConstant::Create(GV, ARMCP::TPOFF);
2505 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2506 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2507 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2508 MachinePointerInfo::getConstantPool(),
2509 false, false, false, 0);
2512 // The address of the thread local variable is the add of the thread
2513 // pointer with the offset of the variable.
2514 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
2518 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
2519 // TODO: implement the "local dynamic" model
2520 assert(Subtarget->isTargetELF() &&
2521 "TLS not implemented for non-ELF targets");
2522 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
2524 TLSModel::Model model = getTargetMachine().getTLSModel(GA->getGlobal());
2527 case TLSModel::GeneralDynamic:
2528 case TLSModel::LocalDynamic:
2529 return LowerToTLSGeneralDynamicModel(GA, DAG);
2530 case TLSModel::InitialExec:
2531 case TLSModel::LocalExec:
2532 return LowerToTLSExecModels(GA, DAG, model);
2534 llvm_unreachable("bogus TLS model");
2537 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
2538 SelectionDAG &DAG) const {
2539 EVT PtrVT = getPointerTy();
2541 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2542 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2543 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
2544 ARMConstantPoolValue *CPV =
2545 ARMConstantPoolConstant::Create(GV,
2546 UseGOTOFF ? ARMCP::GOTOFF : ARMCP::GOT);
2547 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2548 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2549 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
2551 MachinePointerInfo::getConstantPool(),
2552 false, false, false, 0);
2553 SDValue Chain = Result.getValue(1);
2554 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
2555 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
2557 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
2558 MachinePointerInfo::getGOT(),
2559 false, false, false, 0);
2563 // If we have T2 ops, we can materialize the address directly via movt/movw
2564 // pair. This is always cheaper.
2565 if (Subtarget->useMovt(DAG.getMachineFunction())) {
2567 // FIXME: Once remat is capable of dealing with instructions with register
2568 // operands, expand this into two nodes.
2569 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2570 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2572 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2573 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2574 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2575 MachinePointerInfo::getConstantPool(),
2576 false, false, false, 0);
2580 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
2581 SelectionDAG &DAG) const {
2582 EVT PtrVT = getPointerTy();
2584 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2585 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2587 if (Subtarget->useMovt(DAG.getMachineFunction()))
2590 // FIXME: Once remat is capable of dealing with instructions with register
2591 // operands, expand this into multiple nodes
2593 RelocM == Reloc::PIC_ ? ARMISD::WrapperPIC : ARMISD::Wrapper;
2595 SDValue G = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, ARMII::MO_NONLAZY);
2596 SDValue Result = DAG.getNode(Wrapper, dl, PtrVT, G);
2598 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2599 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
2600 MachinePointerInfo::getGOT(), false, false, false, 0);
2604 SDValue ARMTargetLowering::LowerGlobalAddressWindows(SDValue Op,
2605 SelectionDAG &DAG) const {
2606 assert(Subtarget->isTargetWindows() && "non-Windows COFF is not supported");
2607 assert(Subtarget->useMovt(DAG.getMachineFunction()) &&
2608 "Windows on ARM expects to use movw/movt");
2610 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2611 const ARMII::TOF TargetFlags =
2612 (GV->hasDLLImportStorageClass() ? ARMII::MO_DLLIMPORT : ARMII::MO_NO_FLAG);
2613 EVT PtrVT = getPointerTy();
2619 // FIXME: Once remat is capable of dealing with instructions with register
2620 // operands, expand this into two nodes.
2621 Result = DAG.getNode(ARMISD::Wrapper, DL, PtrVT,
2622 DAG.getTargetGlobalAddress(GV, DL, PtrVT, /*Offset=*/0,
2624 if (GV->hasDLLImportStorageClass())
2625 Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), Result,
2626 MachinePointerInfo::getGOT(), false, false, false, 0);
2630 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
2631 SelectionDAG &DAG) const {
2632 assert(Subtarget->isTargetELF() &&
2633 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
2634 MachineFunction &MF = DAG.getMachineFunction();
2635 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2636 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2637 EVT PtrVT = getPointerTy();
2639 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2640 ARMConstantPoolValue *CPV =
2641 ARMConstantPoolSymbol::Create(*DAG.getContext(), "_GLOBAL_OFFSET_TABLE_",
2642 ARMPCLabelIndex, PCAdj);
2643 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2644 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2645 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2646 MachinePointerInfo::getConstantPool(),
2647 false, false, false, 0);
2648 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2649 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2653 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
2655 SDValue Val = DAG.getConstant(0, MVT::i32);
2656 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl,
2657 DAG.getVTList(MVT::i32, MVT::Other), Op.getOperand(0),
2658 Op.getOperand(1), Val);
2662 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
2664 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
2665 Op.getOperand(1), DAG.getConstant(0, MVT::i32));
2669 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
2670 const ARMSubtarget *Subtarget) const {
2671 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2674 default: return SDValue(); // Don't custom lower most intrinsics.
2675 case Intrinsic::arm_rbit: {
2676 assert(Op.getOperand(1).getValueType() == MVT::i32 &&
2677 "RBIT intrinsic must have i32 type!");
2678 return DAG.getNode(ARMISD::RBIT, dl, MVT::i32, Op.getOperand(1));
2680 case Intrinsic::arm_thread_pointer: {
2681 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2682 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2684 case Intrinsic::eh_sjlj_lsda: {
2685 MachineFunction &MF = DAG.getMachineFunction();
2686 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2687 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2688 EVT PtrVT = getPointerTy();
2689 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2691 unsigned PCAdj = (RelocM != Reloc::PIC_)
2692 ? 0 : (Subtarget->isThumb() ? 4 : 8);
2693 ARMConstantPoolValue *CPV =
2694 ARMConstantPoolConstant::Create(MF.getFunction(), ARMPCLabelIndex,
2695 ARMCP::CPLSDA, PCAdj);
2696 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2697 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2699 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2700 MachinePointerInfo::getConstantPool(),
2701 false, false, false, 0);
2703 if (RelocM == Reloc::PIC_) {
2704 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2705 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2709 case Intrinsic::arm_neon_vmulls:
2710 case Intrinsic::arm_neon_vmullu: {
2711 unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vmulls)
2712 ? ARMISD::VMULLs : ARMISD::VMULLu;
2713 return DAG.getNode(NewOpc, SDLoc(Op), Op.getValueType(),
2714 Op.getOperand(1), Op.getOperand(2));
2719 static SDValue LowerATOMIC_FENCE(SDValue Op, SelectionDAG &DAG,
2720 const ARMSubtarget *Subtarget) {
2721 // FIXME: handle "fence singlethread" more efficiently.
2723 if (!Subtarget->hasDataBarrier()) {
2724 // Some ARMv6 cpus can support data barriers with an mcr instruction.
2725 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2727 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2728 "Unexpected ISD::ATOMIC_FENCE encountered. Should be libcall!");
2729 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2730 DAG.getConstant(0, MVT::i32));
2733 ConstantSDNode *OrdN = cast<ConstantSDNode>(Op.getOperand(1));
2734 AtomicOrdering Ord = static_cast<AtomicOrdering>(OrdN->getZExtValue());
2735 ARM_MB::MemBOpt Domain = ARM_MB::ISH;
2736 if (Subtarget->isMClass()) {
2737 // Only a full system barrier exists in the M-class architectures.
2738 Domain = ARM_MB::SY;
2739 } else if (Subtarget->isSwift() && Ord == Release) {
2740 // Swift happens to implement ISHST barriers in a way that's compatible with
2741 // Release semantics but weaker than ISH so we'd be fools not to use
2742 // it. Beware: other processors probably don't!
2743 Domain = ARM_MB::ISHST;
2746 return DAG.getNode(ISD::INTRINSIC_VOID, dl, MVT::Other, Op.getOperand(0),
2747 DAG.getConstant(Intrinsic::arm_dmb, MVT::i32),
2748 DAG.getConstant(Domain, MVT::i32));
2751 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG,
2752 const ARMSubtarget *Subtarget) {
2753 // ARM pre v5TE and Thumb1 does not have preload instructions.
2754 if (!(Subtarget->isThumb2() ||
2755 (!Subtarget->isThumb1Only() && Subtarget->hasV5TEOps())))
2756 // Just preserve the chain.
2757 return Op.getOperand(0);
2760 unsigned isRead = ~cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() & 1;
2762 (!Subtarget->hasV7Ops() || !Subtarget->hasMPExtension()))
2763 // ARMv7 with MP extension has PLDW.
2764 return Op.getOperand(0);
2766 unsigned isData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
2767 if (Subtarget->isThumb()) {
2769 isRead = ~isRead & 1;
2770 isData = ~isData & 1;
2773 return DAG.getNode(ARMISD::PRELOAD, dl, MVT::Other, Op.getOperand(0),
2774 Op.getOperand(1), DAG.getConstant(isRead, MVT::i32),
2775 DAG.getConstant(isData, MVT::i32));
2778 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
2779 MachineFunction &MF = DAG.getMachineFunction();
2780 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
2782 // vastart just stores the address of the VarArgsFrameIndex slot into the
2783 // memory location argument.
2785 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2786 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2787 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
2788 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
2789 MachinePointerInfo(SV), false, false, 0);
2793 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
2794 SDValue &Root, SelectionDAG &DAG,
2796 MachineFunction &MF = DAG.getMachineFunction();
2797 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2799 const TargetRegisterClass *RC;
2800 if (AFI->isThumb1OnlyFunction())
2801 RC = &ARM::tGPRRegClass;
2803 RC = &ARM::GPRRegClass;
2805 // Transform the arguments stored in physical registers into virtual ones.
2806 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2807 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2810 if (NextVA.isMemLoc()) {
2811 MachineFrameInfo *MFI = MF.getFrameInfo();
2812 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true);
2814 // Create load node to retrieve arguments from the stack.
2815 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2816 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
2817 MachinePointerInfo::getFixedStack(FI),
2818 false, false, false, 0);
2820 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
2821 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2823 if (!Subtarget->isLittle())
2824 std::swap (ArgValue, ArgValue2);
2825 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
2829 ARMTargetLowering::computeRegArea(CCState &CCInfo, MachineFunction &MF,
2830 unsigned InRegsParamRecordIdx,
2832 unsigned &ArgRegsSize,
2833 unsigned &ArgRegsSaveSize)
2836 if (InRegsParamRecordIdx < CCInfo.getInRegsParamsCount()) {
2837 unsigned RBegin, REnd;
2838 CCInfo.getInRegsParamInfo(InRegsParamRecordIdx, RBegin, REnd);
2839 NumGPRs = REnd - RBegin;
2841 unsigned int firstUnalloced;
2842 firstUnalloced = CCInfo.getFirstUnallocated(GPRArgRegs,
2843 sizeof(GPRArgRegs) /
2844 sizeof(GPRArgRegs[0]));
2845 NumGPRs = (firstUnalloced <= 3) ? (4 - firstUnalloced) : 0;
2848 unsigned Align = Subtarget->getFrameLowering()->getStackAlignment();
2849 ArgRegsSize = NumGPRs * 4;
2851 // If parameter is split between stack and GPRs...
2852 if (NumGPRs && Align > 4 &&
2853 (ArgRegsSize < ArgSize ||
2854 InRegsParamRecordIdx >= CCInfo.getInRegsParamsCount())) {
2855 // Add padding for part of param recovered from GPRs. For example,
2856 // if Align == 8, its last byte must be at address K*8 - 1.
2857 // We need to do it, since remained (stack) part of parameter has
2858 // stack alignment, and we need to "attach" "GPRs head" without gaps
2861 // |---- 8 bytes block ----| |---- 8 bytes block ----| |---- 8 bytes...
2862 // [ [padding] [GPRs head] ] [ Tail passed via stack ....
2864 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2866 OffsetToAlignment(ArgRegsSize + AFI->getArgRegsSaveSize(), Align);
2867 ArgRegsSaveSize = ArgRegsSize + Padding;
2869 // We don't need to extend regs save size for byval parameters if they
2870 // are passed via GPRs only.
2871 ArgRegsSaveSize = ArgRegsSize;
2874 // The remaining GPRs hold either the beginning of variable-argument
2875 // data, or the beginning of an aggregate passed by value (usually
2876 // byval). Either way, we allocate stack slots adjacent to the data
2877 // provided by our caller, and store the unallocated registers there.
2878 // If this is a variadic function, the va_list pointer will begin with
2879 // these values; otherwise, this reassembles a (byval) structure that
2880 // was split between registers and memory.
2881 // Return: The frame index registers were stored into.
2883 ARMTargetLowering::StoreByValRegs(CCState &CCInfo, SelectionDAG &DAG,
2884 SDLoc dl, SDValue &Chain,
2885 const Value *OrigArg,
2886 unsigned InRegsParamRecordIdx,
2887 unsigned OffsetFromOrigArg,
2891 unsigned ByValStoreOffset,
2892 unsigned TotalArgRegsSaveSize) const {
2894 // Currently, two use-cases possible:
2895 // Case #1. Non-var-args function, and we meet first byval parameter.
2896 // Setup first unallocated register as first byval register;
2897 // eat all remained registers
2898 // (these two actions are performed by HandleByVal method).
2899 // Then, here, we initialize stack frame with
2900 // "store-reg" instructions.
2901 // Case #2. Var-args function, that doesn't contain byval parameters.
2902 // The same: eat all remained unallocated registers,
2903 // initialize stack frame.
2905 MachineFunction &MF = DAG.getMachineFunction();
2906 MachineFrameInfo *MFI = MF.getFrameInfo();
2907 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2908 unsigned firstRegToSaveIndex, lastRegToSaveIndex;
2909 unsigned RBegin, REnd;
2910 if (InRegsParamRecordIdx < CCInfo.getInRegsParamsCount()) {
2911 CCInfo.getInRegsParamInfo(InRegsParamRecordIdx, RBegin, REnd);
2912 firstRegToSaveIndex = RBegin - ARM::R0;
2913 lastRegToSaveIndex = REnd - ARM::R0;
2915 firstRegToSaveIndex = CCInfo.getFirstUnallocated
2916 (GPRArgRegs, array_lengthof(GPRArgRegs));
2917 lastRegToSaveIndex = 4;
2920 unsigned ArgRegsSize, ArgRegsSaveSize;
2921 computeRegArea(CCInfo, MF, InRegsParamRecordIdx, ArgSize,
2922 ArgRegsSize, ArgRegsSaveSize);
2924 // Store any by-val regs to their spots on the stack so that they may be
2925 // loaded by deferencing the result of formal parameter pointer or va_next.
2926 // Note: once stack area for byval/varargs registers
2927 // was initialized, it can't be initialized again.
2928 if (ArgRegsSaveSize) {
2929 unsigned Padding = ArgRegsSaveSize - ArgRegsSize;
2932 assert(AFI->getStoredByValParamsPadding() == 0 &&
2933 "The only parameter may be padded.");
2934 AFI->setStoredByValParamsPadding(Padding);
2937 int FrameIndex = MFI->CreateFixedObject(ArgRegsSaveSize,
2940 (int64_t)TotalArgRegsSaveSize,
2942 SDValue FIN = DAG.getFrameIndex(FrameIndex, getPointerTy());
2944 MFI->CreateFixedObject(Padding,
2945 ArgOffset + ByValStoreOffset -
2946 (int64_t)ArgRegsSaveSize,
2950 SmallVector<SDValue, 4> MemOps;
2951 for (unsigned i = 0; firstRegToSaveIndex < lastRegToSaveIndex;
2952 ++firstRegToSaveIndex, ++i) {
2953 const TargetRegisterClass *RC;
2954 if (AFI->isThumb1OnlyFunction())
2955 RC = &ARM::tGPRRegClass;
2957 RC = &ARM::GPRRegClass;
2959 unsigned VReg = MF.addLiveIn(GPRArgRegs[firstRegToSaveIndex], RC);
2960 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2962 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2963 MachinePointerInfo(OrigArg, OffsetFromOrigArg + 4*i),
2965 MemOps.push_back(Store);
2966 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
2967 DAG.getConstant(4, getPointerTy()));
2970 AFI->setArgRegsSaveSize(ArgRegsSaveSize + AFI->getArgRegsSaveSize());
2972 if (!MemOps.empty())
2973 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOps);
2977 // We cannot allocate a zero-byte object for the first variadic argument,
2978 // so just make up a size.
2981 // This will point to the next argument passed via stack.
2982 return MFI->CreateFixedObject(
2983 ArgSize, ArgOffset, !ForceMutable);
2987 // Setup stack frame, the va_list pointer will start from.
2989 ARMTargetLowering::VarArgStyleRegisters(CCState &CCInfo, SelectionDAG &DAG,
2990 SDLoc dl, SDValue &Chain,
2992 unsigned TotalArgRegsSaveSize,
2993 bool ForceMutable) const {
2994 MachineFunction &MF = DAG.getMachineFunction();
2995 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2997 // Try to store any remaining integer argument regs
2998 // to their spots on the stack so that they may be loaded by deferencing
2999 // the result of va_next.
3000 // If there is no regs to be stored, just point address after last
3001 // argument passed via stack.
3003 StoreByValRegs(CCInfo, DAG, dl, Chain, nullptr,
3004 CCInfo.getInRegsParamsCount(), 0, ArgOffset, 0, ForceMutable,
3005 0, TotalArgRegsSaveSize);
3007 AFI->setVarArgsFrameIndex(FrameIndex);
3011 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
3012 CallingConv::ID CallConv, bool isVarArg,
3013 const SmallVectorImpl<ISD::InputArg>
3015 SDLoc dl, SelectionDAG &DAG,
3016 SmallVectorImpl<SDValue> &InVals)
3018 MachineFunction &MF = DAG.getMachineFunction();
3019 MachineFrameInfo *MFI = MF.getFrameInfo();
3021 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
3023 // Assign locations to all of the incoming arguments.
3024 SmallVector<CCValAssign, 16> ArgLocs;
3025 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs,
3026 *DAG.getContext(), Prologue);
3027 CCInfo.AnalyzeFormalArguments(Ins,
3028 CCAssignFnForNode(CallConv, /* Return*/ false,
3031 SmallVector<SDValue, 16> ArgValues;
3032 int lastInsIndex = -1;
3034 Function::const_arg_iterator CurOrigArg = MF.getFunction()->arg_begin();
3035 unsigned CurArgIdx = 0;
3037 // Initially ArgRegsSaveSize is zero.
3038 // Then we increase this value each time we meet byval parameter.
3039 // We also increase this value in case of varargs function.
3040 AFI->setArgRegsSaveSize(0);
3042 unsigned ByValStoreOffset = 0;
3043 unsigned TotalArgRegsSaveSize = 0;
3044 unsigned ArgRegsSaveSizeMaxAlign = 4;
3046 // Calculate the amount of stack space that we need to allocate to store
3047 // byval and variadic arguments that are passed in registers.
3048 // We need to know this before we allocate the first byval or variadic
3049 // argument, as they will be allocated a stack slot below the CFA (Canonical
3050 // Frame Address, the stack pointer at entry to the function).
3051 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3052 CCValAssign &VA = ArgLocs[i];
3053 if (VA.isMemLoc()) {
3054 int index = VA.getValNo();
3055 if (index != lastInsIndex) {
3056 ISD::ArgFlagsTy Flags = Ins[index].Flags;
3057 if (Flags.isByVal()) {
3058 unsigned ExtraArgRegsSize;
3059 unsigned ExtraArgRegsSaveSize;
3060 computeRegArea(CCInfo, MF, CCInfo.getInRegsParamsProcessed(),
3061 Flags.getByValSize(),
3062 ExtraArgRegsSize, ExtraArgRegsSaveSize);
3064 TotalArgRegsSaveSize += ExtraArgRegsSaveSize;
3065 if (Flags.getByValAlign() > ArgRegsSaveSizeMaxAlign)
3066 ArgRegsSaveSizeMaxAlign = Flags.getByValAlign();
3067 CCInfo.nextInRegsParam();
3069 lastInsIndex = index;
3073 CCInfo.rewindByValRegsInfo();
3075 if (isVarArg && MFI->hasVAStart()) {
3076 unsigned ExtraArgRegsSize;
3077 unsigned ExtraArgRegsSaveSize;
3078 computeRegArea(CCInfo, MF, CCInfo.getInRegsParamsCount(), 0,
3079 ExtraArgRegsSize, ExtraArgRegsSaveSize);
3080 TotalArgRegsSaveSize += ExtraArgRegsSaveSize;
3082 // If the arg regs save area contains N-byte aligned values, the
3083 // bottom of it must be at least N-byte aligned.
3084 TotalArgRegsSaveSize = RoundUpToAlignment(TotalArgRegsSaveSize, ArgRegsSaveSizeMaxAlign);
3085 TotalArgRegsSaveSize = std::min(TotalArgRegsSaveSize, 16U);
3087 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3088 CCValAssign &VA = ArgLocs[i];
3089 std::advance(CurOrigArg, Ins[VA.getValNo()].OrigArgIndex - CurArgIdx);
3090 CurArgIdx = Ins[VA.getValNo()].OrigArgIndex;
3091 // Arguments stored in registers.
3092 if (VA.isRegLoc()) {
3093 EVT RegVT = VA.getLocVT();
3095 if (VA.needsCustom()) {
3096 // f64 and vector types are split up into multiple registers or
3097 // combinations of registers and stack slots.
3098 if (VA.getLocVT() == MVT::v2f64) {
3099 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
3101 VA = ArgLocs[++i]; // skip ahead to next loc
3103 if (VA.isMemLoc()) {
3104 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true);
3105 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
3106 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
3107 MachinePointerInfo::getFixedStack(FI),
3108 false, false, false, 0);
3110 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
3113 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
3114 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
3115 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
3116 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
3117 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
3119 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
3122 const TargetRegisterClass *RC;
3124 if (RegVT == MVT::f32)
3125 RC = &ARM::SPRRegClass;
3126 else if (RegVT == MVT::f64)
3127 RC = &ARM::DPRRegClass;
3128 else if (RegVT == MVT::v2f64)
3129 RC = &ARM::QPRRegClass;
3130 else if (RegVT == MVT::i32)
3131 RC = AFI->isThumb1OnlyFunction() ? &ARM::tGPRRegClass
3132 : &ARM::GPRRegClass;
3134 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
3136 // Transform the arguments in physical registers into virtual ones.
3137 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
3138 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
3141 // If this is an 8 or 16-bit value, it is really passed promoted
3142 // to 32 bits. Insert an assert[sz]ext to capture this, then
3143 // truncate to the right size.
3144 switch (VA.getLocInfo()) {
3145 default: llvm_unreachable("Unknown loc info!");
3146 case CCValAssign::Full: break;
3147 case CCValAssign::BCvt:
3148 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
3150 case CCValAssign::SExt:
3151 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
3152 DAG.getValueType(VA.getValVT()));
3153 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
3155 case CCValAssign::ZExt:
3156 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
3157 DAG.getValueType(VA.getValVT()));
3158 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
3162 InVals.push_back(ArgValue);
3164 } else { // VA.isRegLoc()
3167 assert(VA.isMemLoc());
3168 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
3170 int index = ArgLocs[i].getValNo();
3172 // Some Ins[] entries become multiple ArgLoc[] entries.
3173 // Process them only once.
3174 if (index != lastInsIndex)
3176 ISD::ArgFlagsTy Flags = Ins[index].Flags;
3177 // FIXME: For now, all byval parameter objects are marked mutable.
3178 // This can be changed with more analysis.
3179 // In case of tail call optimization mark all arguments mutable.
3180 // Since they could be overwritten by lowering of arguments in case of
3182 if (Flags.isByVal()) {
3183 unsigned CurByValIndex = CCInfo.getInRegsParamsProcessed();
3185 ByValStoreOffset = RoundUpToAlignment(ByValStoreOffset, Flags.getByValAlign());
3186 int FrameIndex = StoreByValRegs(
3187 CCInfo, DAG, dl, Chain, CurOrigArg,
3189 Ins[VA.getValNo()].PartOffset,
3190 VA.getLocMemOffset(),
3191 Flags.getByValSize(),
3192 true /*force mutable frames*/,
3194 TotalArgRegsSaveSize);
3195 ByValStoreOffset += Flags.getByValSize();
3196 ByValStoreOffset = std::min(ByValStoreOffset, 16U);
3197 InVals.push_back(DAG.getFrameIndex(FrameIndex, getPointerTy()));
3198 CCInfo.nextInRegsParam();
3200 unsigned FIOffset = VA.getLocMemOffset();
3201 int FI = MFI->CreateFixedObject(VA.getLocVT().getSizeInBits()/8,
3204 // Create load nodes to retrieve arguments from the stack.
3205 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
3206 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
3207 MachinePointerInfo::getFixedStack(FI),
3208 false, false, false, 0));
3210 lastInsIndex = index;
3216 if (isVarArg && MFI->hasVAStart())
3217 VarArgStyleRegisters(CCInfo, DAG, dl, Chain,
3218 CCInfo.getNextStackOffset(),
3219 TotalArgRegsSaveSize);
3221 AFI->setArgumentStackSize(CCInfo.getNextStackOffset());
3226 /// isFloatingPointZero - Return true if this is +0.0.
3227 static bool isFloatingPointZero(SDValue Op) {
3228 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
3229 return CFP->getValueAPF().isPosZero();
3230 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
3231 // Maybe this has already been legalized into the constant pool?
3232 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
3233 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
3234 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
3235 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
3236 return CFP->getValueAPF().isPosZero();
3238 } else if (Op->getOpcode() == ISD::BITCAST &&
3239 Op->getValueType(0) == MVT::f64) {
3240 // Handle (ISD::BITCAST (ARMISD::VMOVIMM (ISD::TargetConstant 0)) MVT::f64)
3241 // created by LowerConstantFP().
3242 SDValue BitcastOp = Op->getOperand(0);
3243 if (BitcastOp->getOpcode() == ARMISD::VMOVIMM) {
3244 SDValue MoveOp = BitcastOp->getOperand(0);
3245 if (MoveOp->getOpcode() == ISD::TargetConstant &&
3246 cast<ConstantSDNode>(MoveOp)->getZExtValue() == 0) {
3254 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
3255 /// the given operands.
3257 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
3258 SDValue &ARMcc, SelectionDAG &DAG,
3260 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
3261 unsigned C = RHSC->getZExtValue();
3262 if (!isLegalICmpImmediate(C)) {
3263 // Constant does not fit, try adjusting it by one?
3268 if (C != 0x80000000 && isLegalICmpImmediate(C-1)) {
3269 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
3270 RHS = DAG.getConstant(C-1, MVT::i32);
3275 if (C != 0 && isLegalICmpImmediate(C-1)) {
3276 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
3277 RHS = DAG.getConstant(C-1, MVT::i32);
3282 if (C != 0x7fffffff && isLegalICmpImmediate(C+1)) {
3283 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
3284 RHS = DAG.getConstant(C+1, MVT::i32);
3289 if (C != 0xffffffff && isLegalICmpImmediate(C+1)) {
3290 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
3291 RHS = DAG.getConstant(C+1, MVT::i32);
3298 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3299 ARMISD::NodeType CompareType;
3302 CompareType = ARMISD::CMP;
3307 CompareType = ARMISD::CMPZ;
3310 ARMcc = DAG.getConstant(CondCode, MVT::i32);
3311 return DAG.getNode(CompareType, dl, MVT::Glue, LHS, RHS);
3314 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
3316 ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
3318 assert(!Subtarget->isFPOnlySP() || RHS.getValueType() != MVT::f64);
3320 if (!isFloatingPointZero(RHS))
3321 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Glue, LHS, RHS);
3323 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Glue, LHS);
3324 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Glue, Cmp);
3327 /// duplicateCmp - Glue values can have only one use, so this function
3328 /// duplicates a comparison node.
3330 ARMTargetLowering::duplicateCmp(SDValue Cmp, SelectionDAG &DAG) const {
3331 unsigned Opc = Cmp.getOpcode();
3333 if (Opc == ARMISD::CMP || Opc == ARMISD::CMPZ)
3334 return DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
3336 assert(Opc == ARMISD::FMSTAT && "unexpected comparison operation");
3337 Cmp = Cmp.getOperand(0);
3338 Opc = Cmp.getOpcode();
3339 if (Opc == ARMISD::CMPFP)
3340 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
3342 assert(Opc == ARMISD::CMPFPw0 && "unexpected operand of FMSTAT");
3343 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0));
3345 return DAG.getNode(ARMISD::FMSTAT, DL, MVT::Glue, Cmp);
3348 std::pair<SDValue, SDValue>
3349 ARMTargetLowering::getARMXALUOOp(SDValue Op, SelectionDAG &DAG,
3350 SDValue &ARMcc) const {
3351 assert(Op.getValueType() == MVT::i32 && "Unsupported value type");
3353 SDValue Value, OverflowCmp;
3354 SDValue LHS = Op.getOperand(0);
3355 SDValue RHS = Op.getOperand(1);
3358 // FIXME: We are currently always generating CMPs because we don't support
3359 // generating CMN through the backend. This is not as good as the natural
3360 // CMP case because it causes a register dependency and cannot be folded
3363 switch (Op.getOpcode()) {
3365 llvm_unreachable("Unknown overflow instruction!");
3367 ARMcc = DAG.getConstant(ARMCC::VC, MVT::i32);
3368 Value = DAG.getNode(ISD::ADD, SDLoc(Op), Op.getValueType(), LHS, RHS);
3369 OverflowCmp = DAG.getNode(ARMISD::CMP, SDLoc(Op), MVT::Glue, Value, LHS);
3372 ARMcc = DAG.getConstant(ARMCC::HS, MVT::i32);
3373 Value = DAG.getNode(ISD::ADD, SDLoc(Op), Op.getValueType(), LHS, RHS);
3374 OverflowCmp = DAG.getNode(ARMISD::CMP, SDLoc(Op), MVT::Glue, Value, LHS);
3377 ARMcc = DAG.getConstant(ARMCC::VC, MVT::i32);
3378 Value = DAG.getNode(ISD::SUB, SDLoc(Op), Op.getValueType(), LHS, RHS);
3379 OverflowCmp = DAG.getNode(ARMISD::CMP, SDLoc(Op), MVT::Glue, LHS, RHS);
3382 ARMcc = DAG.getConstant(ARMCC::HS, MVT::i32);
3383 Value = DAG.getNode(ISD::SUB, SDLoc(Op), Op.getValueType(), LHS, RHS);
3384 OverflowCmp = DAG.getNode(ARMISD::CMP, SDLoc(Op), MVT::Glue, LHS, RHS);
3388 return std::make_pair(Value, OverflowCmp);
3393 ARMTargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) const {
3394 // Let legalize expand this if it isn't a legal type yet.
3395 if (!DAG.getTargetLoweringInfo().isTypeLegal(Op.getValueType()))
3398 SDValue Value, OverflowCmp;
3400 std::tie(Value, OverflowCmp) = getARMXALUOOp(Op, DAG, ARMcc);
3401 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3402 // We use 0 and 1 as false and true values.
3403 SDValue TVal = DAG.getConstant(1, MVT::i32);
3404 SDValue FVal = DAG.getConstant(0, MVT::i32);
3405 EVT VT = Op.getValueType();
3407 SDValue Overflow = DAG.getNode(ARMISD::CMOV, SDLoc(Op), VT, TVal, FVal,
3408 ARMcc, CCR, OverflowCmp);
3410 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
3411 return DAG.getNode(ISD::MERGE_VALUES, SDLoc(Op), VTs, Value, Overflow);
3415 SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
3416 SDValue Cond = Op.getOperand(0);
3417 SDValue SelectTrue = Op.getOperand(1);
3418 SDValue SelectFalse = Op.getOperand(2);
3420 unsigned Opc = Cond.getOpcode();
3422 if (Cond.getResNo() == 1 &&
3423 (Opc == ISD::SADDO || Opc == ISD::UADDO || Opc == ISD::SSUBO ||
3424 Opc == ISD::USUBO)) {
3425 if (!DAG.getTargetLoweringInfo().isTypeLegal(Cond->getValueType(0)))
3428 SDValue Value, OverflowCmp;
3430 std::tie(Value, OverflowCmp) = getARMXALUOOp(Cond, DAG, ARMcc);
3431 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3432 EVT VT = Op.getValueType();
3434 return getCMOV(SDLoc(Op), VT, SelectTrue, SelectFalse, ARMcc, CCR,
3440 // (select (cmov 1, 0, cond), t, f) -> (cmov t, f, cond)
3441 // (select (cmov 0, 1, cond), t, f) -> (cmov f, t, cond)
3443 if (Cond.getOpcode() == ARMISD::CMOV && Cond.hasOneUse()) {
3444 const ConstantSDNode *CMOVTrue =
3445 dyn_cast<ConstantSDNode>(Cond.getOperand(0));
3446 const ConstantSDNode *CMOVFalse =
3447 dyn_cast<ConstantSDNode>(Cond.getOperand(1));
3449 if (CMOVTrue && CMOVFalse) {
3450 unsigned CMOVTrueVal = CMOVTrue->getZExtValue();
3451 unsigned CMOVFalseVal = CMOVFalse->getZExtValue();
3455 if (CMOVTrueVal == 1 && CMOVFalseVal == 0) {
3457 False = SelectFalse;
3458 } else if (CMOVTrueVal == 0 && CMOVFalseVal == 1) {
3463 if (True.getNode() && False.getNode()) {
3464 EVT VT = Op.getValueType();
3465 SDValue ARMcc = Cond.getOperand(2);
3466 SDValue CCR = Cond.getOperand(3);
3467 SDValue Cmp = duplicateCmp(Cond.getOperand(4), DAG);
3468 assert(True.getValueType() == VT);
3469 return getCMOV(dl, VT, True, False, ARMcc, CCR, Cmp, DAG);
3474 // ARM's BooleanContents value is UndefinedBooleanContent. Mask out the
3475 // undefined bits before doing a full-word comparison with zero.
3476 Cond = DAG.getNode(ISD::AND, dl, Cond.getValueType(), Cond,
3477 DAG.getConstant(1, Cond.getValueType()));
3479 return DAG.getSelectCC(dl, Cond,
3480 DAG.getConstant(0, Cond.getValueType()),
3481 SelectTrue, SelectFalse, ISD::SETNE);
3484 static ISD::CondCode getInverseCCForVSEL(ISD::CondCode CC) {
3485 if (CC == ISD::SETNE)
3487 return ISD::getSetCCInverse(CC, true);
3490 static void checkVSELConstraints(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
3491 bool &swpCmpOps, bool &swpVselOps) {
3492 // Start by selecting the GE condition code for opcodes that return true for
3494 if (CC == ISD::SETUGE || CC == ISD::SETOGE || CC == ISD::SETOLE ||
3496 CondCode = ARMCC::GE;
3498 // and GT for opcodes that return false for 'equality'.
3499 else if (CC == ISD::SETUGT || CC == ISD::SETOGT || CC == ISD::SETOLT ||
3501 CondCode = ARMCC::GT;
3503 // Since we are constrained to GE/GT, if the opcode contains 'less', we need
3504 // to swap the compare operands.
3505 if (CC == ISD::SETOLE || CC == ISD::SETULE || CC == ISD::SETOLT ||
3509 // Both GT and GE are ordered comparisons, and return false for 'unordered'.
3510 // If we have an unordered opcode, we need to swap the operands to the VSEL
3511 // instruction (effectively negating the condition).
3513 // This also has the effect of swapping which one of 'less' or 'greater'
3514 // returns true, so we also swap the compare operands. It also switches
3515 // whether we return true for 'equality', so we compensate by picking the
3516 // opposite condition code to our original choice.
3517 if (CC == ISD::SETULE || CC == ISD::SETULT || CC == ISD::SETUGE ||
3518 CC == ISD::SETUGT) {
3519 swpCmpOps = !swpCmpOps;
3520 swpVselOps = !swpVselOps;
3521 CondCode = CondCode == ARMCC::GT ? ARMCC::GE : ARMCC::GT;
3524 // 'ordered' is 'anything but unordered', so use the VS condition code and
3525 // swap the VSEL operands.
3526 if (CC == ISD::SETO) {
3527 CondCode = ARMCC::VS;
3531 // 'unordered or not equal' is 'anything but equal', so use the EQ condition
3532 // code and swap the VSEL operands.
3533 if (CC == ISD::SETUNE) {
3534 CondCode = ARMCC::EQ;
3539 SDValue ARMTargetLowering::getCMOV(SDLoc dl, EVT VT, SDValue FalseVal,
3540 SDValue TrueVal, SDValue ARMcc, SDValue CCR,
3541 SDValue Cmp, SelectionDAG &DAG) const {
3542 if (Subtarget->isFPOnlySP() && VT == MVT::f64) {
3543 FalseVal = DAG.getNode(ARMISD::VMOVRRD, dl,
3544 DAG.getVTList(MVT::i32, MVT::i32), FalseVal);
3545 TrueVal = DAG.getNode(ARMISD::VMOVRRD, dl,
3546 DAG.getVTList(MVT::i32, MVT::i32), TrueVal);
3548 SDValue TrueLow = TrueVal.getValue(0);
3549 SDValue TrueHigh = TrueVal.getValue(1);
3550 SDValue FalseLow = FalseVal.getValue(0);
3551 SDValue FalseHigh = FalseVal.getValue(1);
3553 SDValue Low = DAG.getNode(ARMISD::CMOV, dl, MVT::i32, FalseLow, TrueLow,
3555 SDValue High = DAG.getNode(ARMISD::CMOV, dl, MVT::i32, FalseHigh, TrueHigh,
3556 ARMcc, CCR, duplicateCmp(Cmp, DAG));
3558 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Low, High);
3560 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,
3565 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
3566 EVT VT = Op.getValueType();
3567 SDValue LHS = Op.getOperand(0);
3568 SDValue RHS = Op.getOperand(1);
3569 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
3570 SDValue TrueVal = Op.getOperand(2);
3571 SDValue FalseVal = Op.getOperand(3);
3574 if (Subtarget->isFPOnlySP() && LHS.getValueType() == MVT::f64) {
3575 DAG.getTargetLoweringInfo().softenSetCCOperands(DAG, MVT::f64, LHS, RHS, CC,
3578 // If softenSetCCOperands only returned one value, we should compare it to
3580 if (!RHS.getNode()) {
3581 RHS = DAG.getConstant(0, LHS.getValueType());
3586 if (LHS.getValueType() == MVT::i32) {
3587 // Try to generate VSEL on ARMv8.
3588 // The VSEL instruction can't use all the usual ARM condition
3589 // codes: it only has two bits to select the condition code, so it's
3590 // constrained to use only GE, GT, VS and EQ.
3592 // To implement all the various ISD::SETXXX opcodes, we sometimes need to
3593 // swap the operands of the previous compare instruction (effectively
3594 // inverting the compare condition, swapping 'less' and 'greater') and
3595 // sometimes need to swap the operands to the VSEL (which inverts the
3596 // condition in the sense of firing whenever the previous condition didn't)
3597 if (Subtarget->hasFPARMv8() && (TrueVal.getValueType() == MVT::f32 ||
3598 TrueVal.getValueType() == MVT::f64)) {
3599 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3600 if (CondCode == ARMCC::LT || CondCode == ARMCC::LE ||
3601 CondCode == ARMCC::VC || CondCode == ARMCC::NE) {
3602 CC = getInverseCCForVSEL(CC);
3603 std::swap(TrueVal, FalseVal);
3608 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3609 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3610 return getCMOV(dl, VT, FalseVal, TrueVal, ARMcc, CCR, Cmp, DAG);
3613 ARMCC::CondCodes CondCode, CondCode2;
3614 FPCCToARMCC(CC, CondCode, CondCode2);
3616 // Try to generate VSEL on ARMv8.
3617 if (Subtarget->hasFPARMv8() && (TrueVal.getValueType() == MVT::f32 ||
3618 TrueVal.getValueType() == MVT::f64)) {
3619 // We can select VMAXNM/VMINNM from a compare followed by a select with the
3620 // same operands, as follows:
3621 // c = fcmp [ogt, olt, ugt, ult] a, b
3623 // We only do this in unsafe-fp-math, because signed zeros and NaNs are
3624 // handled differently than the original code sequence.
3625 if (getTargetMachine().Options.UnsafeFPMath) {
3626 if (LHS == TrueVal && RHS == FalseVal) {
3627 if (CC == ISD::SETOGT || CC == ISD::SETUGT)
3628 return DAG.getNode(ARMISD::VMAXNM, dl, VT, TrueVal, FalseVal);
3629 if (CC == ISD::SETOLT || CC == ISD::SETULT)
3630 return DAG.getNode(ARMISD::VMINNM, dl, VT, TrueVal, FalseVal);
3631 } else if (LHS == FalseVal && RHS == TrueVal) {
3632 if (CC == ISD::SETOLT || CC == ISD::SETULT)
3633 return DAG.getNode(ARMISD::VMAXNM, dl, VT, TrueVal, FalseVal);
3634 if (CC == ISD::SETOGT || CC == ISD::SETUGT)
3635 return DAG.getNode(ARMISD::VMINNM, dl, VT, TrueVal, FalseVal);
3639 bool swpCmpOps = false;
3640 bool swpVselOps = false;
3641 checkVSELConstraints(CC, CondCode, swpCmpOps, swpVselOps);
3643 if (CondCode == ARMCC::GT || CondCode == ARMCC::GE ||
3644 CondCode == ARMCC::VS || CondCode == ARMCC::EQ) {
3646 std::swap(LHS, RHS);
3648 std::swap(TrueVal, FalseVal);
3652 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
3653 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
3654 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3655 SDValue Result = getCMOV(dl, VT, FalseVal, TrueVal, ARMcc, CCR, Cmp, DAG);
3656 if (CondCode2 != ARMCC::AL) {
3657 SDValue ARMcc2 = DAG.getConstant(CondCode2, MVT::i32);
3658 // FIXME: Needs another CMP because flag can have but one use.
3659 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
3660 Result = getCMOV(dl, VT, Result, TrueVal, ARMcc2, CCR, Cmp2, DAG);
3665 /// canChangeToInt - Given the fp compare operand, return true if it is suitable
3666 /// to morph to an integer compare sequence.
3667 static bool canChangeToInt(SDValue Op, bool &SeenZero,
3668 const ARMSubtarget *Subtarget) {
3669 SDNode *N = Op.getNode();
3670 if (!N->hasOneUse())
3671 // Otherwise it requires moving the value from fp to integer registers.
3673 if (!N->getNumValues())
3675 EVT VT = Op.getValueType();
3676 if (VT != MVT::f32 && !Subtarget->isFPBrccSlow())
3677 // f32 case is generally profitable. f64 case only makes sense when vcmpe +
3678 // vmrs are very slow, e.g. cortex-a8.
3681 if (isFloatingPointZero(Op)) {
3685 return ISD::isNormalLoad(N);
3688 static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) {
3689 if (isFloatingPointZero(Op))
3690 return DAG.getConstant(0, MVT::i32);
3692 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op))
3693 return DAG.getLoad(MVT::i32, SDLoc(Op),
3694 Ld->getChain(), Ld->getBasePtr(), Ld->getPointerInfo(),
3695 Ld->isVolatile(), Ld->isNonTemporal(),
3696 Ld->isInvariant(), Ld->getAlignment());
3698 llvm_unreachable("Unknown VFP cmp argument!");
3701 static void expandf64Toi32(SDValue Op, SelectionDAG &DAG,
3702 SDValue &RetVal1, SDValue &RetVal2) {
3703 if (isFloatingPointZero(Op)) {
3704 RetVal1 = DAG.getConstant(0, MVT::i32);
3705 RetVal2 = DAG.getConstant(0, MVT::i32);
3709 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) {
3710 SDValue Ptr = Ld->getBasePtr();
3711 RetVal1 = DAG.getLoad(MVT::i32, SDLoc(Op),
3712 Ld->getChain(), Ptr,
3713 Ld->getPointerInfo(),
3714 Ld->isVolatile(), Ld->isNonTemporal(),
3715 Ld->isInvariant(), Ld->getAlignment());
3717 EVT PtrType = Ptr.getValueType();
3718 unsigned NewAlign = MinAlign(Ld->getAlignment(), 4);
3719 SDValue NewPtr = DAG.getNode(ISD::ADD, SDLoc(Op),
3720 PtrType, Ptr, DAG.getConstant(4, PtrType));
3721 RetVal2 = DAG.getLoad(MVT::i32, SDLoc(Op),
3722 Ld->getChain(), NewPtr,
3723 Ld->getPointerInfo().getWithOffset(4),
3724 Ld->isVolatile(), Ld->isNonTemporal(),
3725 Ld->isInvariant(), NewAlign);
3729 llvm_unreachable("Unknown VFP cmp argument!");
3732 /// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some
3733 /// f32 and even f64 comparisons to integer ones.
3735 ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const {
3736 SDValue Chain = Op.getOperand(0);
3737 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3738 SDValue LHS = Op.getOperand(2);
3739 SDValue RHS = Op.getOperand(3);
3740 SDValue Dest = Op.getOperand(4);
3743 bool LHSSeenZero = false;
3744 bool LHSOk = canChangeToInt(LHS, LHSSeenZero, Subtarget);
3745 bool RHSSeenZero = false;
3746 bool RHSOk = canChangeToInt(RHS, RHSSeenZero, Subtarget);
3747 if (LHSOk && RHSOk && (LHSSeenZero || RHSSeenZero)) {
3748 // If unsafe fp math optimization is enabled and there are no other uses of
3749 // the CMP operands, and the condition code is EQ or NE, we can optimize it
3750 // to an integer comparison.
3751 if (CC == ISD::SETOEQ)
3753 else if (CC == ISD::SETUNE)
3756 SDValue Mask = DAG.getConstant(0x7fffffff, MVT::i32);
3758 if (LHS.getValueType() == MVT::f32) {
3759 LHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3760 bitcastf32Toi32(LHS, DAG), Mask);
3761 RHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3762 bitcastf32Toi32(RHS, DAG), Mask);
3763 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3764 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3765 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3766 Chain, Dest, ARMcc, CCR, Cmp);
3771 expandf64Toi32(LHS, DAG, LHS1, LHS2);
3772 expandf64Toi32(RHS, DAG, RHS1, RHS2);
3773 LHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, LHS2, Mask);
3774 RHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, RHS2, Mask);
3775 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3776 ARMcc = DAG.getConstant(CondCode, MVT::i32);
3777 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3778 SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest };
3779 return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops);
3785 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
3786 SDValue Chain = Op.getOperand(0);
3787 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3788 SDValue LHS = Op.getOperand(2);
3789 SDValue RHS = Op.getOperand(3);
3790 SDValue Dest = Op.getOperand(4);
3793 if (Subtarget->isFPOnlySP() && LHS.getValueType() == MVT::f64) {
3794 DAG.getTargetLoweringInfo().softenSetCCOperands(DAG, MVT::f64, LHS, RHS, CC,
3797 // If softenSetCCOperands only returned one value, we should compare it to
3799 if (!RHS.getNode()) {
3800 RHS = DAG.getConstant(0, LHS.getValueType());
3805 if (LHS.getValueType() == MVT::i32) {
3807 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3808 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3809 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3810 Chain, Dest, ARMcc, CCR, Cmp);
3813 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3815 if (getTargetMachine().Options.UnsafeFPMath &&
3816 (CC == ISD::SETEQ || CC == ISD::SETOEQ ||
3817 CC == ISD::SETNE || CC == ISD::SETUNE)) {
3818 SDValue Result = OptimizeVFPBrcond(Op, DAG);
3819 if (Result.getNode())
3823 ARMCC::CondCodes CondCode, CondCode2;
3824 FPCCToARMCC(CC, CondCode, CondCode2);
3826 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
3827 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
3828 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3829 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3830 SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp };
3831 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops);
3832 if (CondCode2 != ARMCC::AL) {
3833 ARMcc = DAG.getConstant(CondCode2, MVT::i32);
3834 SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) };
3835 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops);
3840 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
3841 SDValue Chain = Op.getOperand(0);
3842 SDValue Table = Op.getOperand(1);
3843 SDValue Index = Op.getOperand(2);
3846 EVT PTy = getPointerTy();
3847 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
3848 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
3849 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
3850 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
3851 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
3852 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
3853 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
3854 if (Subtarget->isThumb2()) {
3855 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
3856 // which does another jump to the destination. This also makes it easier
3857 // to translate it to TBB / TBH later.
3858 // FIXME: This might not work if the function is extremely large.
3859 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
3860 Addr, Op.getOperand(2), JTI, UId);
3862 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
3863 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
3864 MachinePointerInfo::getJumpTable(),
3865 false, false, false, 0);
3866 Chain = Addr.getValue(1);
3867 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
3868 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3870 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
3871 MachinePointerInfo::getJumpTable(),
3872 false, false, false, 0);
3873 Chain = Addr.getValue(1);
3874 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3878 static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3879 EVT VT = Op.getValueType();
3882 if (Op.getValueType().getVectorElementType() == MVT::i32) {
3883 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::f32)
3885 return DAG.UnrollVectorOp(Op.getNode());
3888 assert(Op.getOperand(0).getValueType() == MVT::v4f32 &&
3889 "Invalid type for custom lowering!");
3890 if (VT != MVT::v4i16)
3891 return DAG.UnrollVectorOp(Op.getNode());
3893 Op = DAG.getNode(Op.getOpcode(), dl, MVT::v4i32, Op.getOperand(0));
3894 return DAG.getNode(ISD::TRUNCATE, dl, VT, Op);
3897 SDValue ARMTargetLowering::LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) const {
3898 EVT VT = Op.getValueType();
3900 return LowerVectorFP_TO_INT(Op, DAG);
3902 if (Subtarget->isFPOnlySP() && Op.getOperand(0).getValueType() == MVT::f64) {
3904 if (Op.getOpcode() == ISD::FP_TO_SINT)
3905 LC = RTLIB::getFPTOSINT(Op.getOperand(0).getValueType(),
3908 LC = RTLIB::getFPTOUINT(Op.getOperand(0).getValueType(),
3910 return makeLibCall(DAG, LC, Op.getValueType(), &Op.getOperand(0), 1,
3911 /*isSigned*/ false, SDLoc(Op)).first;
3917 switch (Op.getOpcode()) {
3918 default: llvm_unreachable("Invalid opcode!");
3919 case ISD::FP_TO_SINT:
3920 Opc = ARMISD::FTOSI;
3922 case ISD::FP_TO_UINT:
3923 Opc = ARMISD::FTOUI;
3926 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
3927 return DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
3930 static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3931 EVT VT = Op.getValueType();
3934 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::i32) {
3935 if (VT.getVectorElementType() == MVT::f32)
3937 return DAG.UnrollVectorOp(Op.getNode());
3940 assert(Op.getOperand(0).getValueType() == MVT::v4i16 &&
3941 "Invalid type for custom lowering!");
3942 if (VT != MVT::v4f32)
3943 return DAG.UnrollVectorOp(Op.getNode());
3947 switch (Op.getOpcode()) {
3948 default: llvm_unreachable("Invalid opcode!");
3949 case ISD::SINT_TO_FP:
3950 CastOpc = ISD::SIGN_EXTEND;
3951 Opc = ISD::SINT_TO_FP;
3953 case ISD::UINT_TO_FP:
3954 CastOpc = ISD::ZERO_EXTEND;
3955 Opc = ISD::UINT_TO_FP;
3959 Op = DAG.getNode(CastOpc, dl, MVT::v4i32, Op.getOperand(0));
3960 return DAG.getNode(Opc, dl, VT, Op);
3963 SDValue ARMTargetLowering::LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) const {
3964 EVT VT = Op.getValueType();
3966 return LowerVectorINT_TO_FP(Op, DAG);
3968 if (Subtarget->isFPOnlySP() && Op.getValueType() == MVT::f64) {
3970 if (Op.getOpcode() == ISD::SINT_TO_FP)
3971 LC = RTLIB::getSINTTOFP(Op.getOperand(0).getValueType(),
3974 LC = RTLIB::getUINTTOFP(Op.getOperand(0).getValueType(),
3976 return makeLibCall(DAG, LC, Op.getValueType(), &Op.getOperand(0), 1,
3977 /*isSigned*/ false, SDLoc(Op)).first;
3983 switch (Op.getOpcode()) {
3984 default: llvm_unreachable("Invalid opcode!");
3985 case ISD::SINT_TO_FP:
3986 Opc = ARMISD::SITOF;
3988 case ISD::UINT_TO_FP:
3989 Opc = ARMISD::UITOF;
3993 Op = DAG.getNode(ISD::BITCAST, dl, MVT::f32, Op.getOperand(0));
3994 return DAG.getNode(Opc, dl, VT, Op);
3997 SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
3998 // Implement fcopysign with a fabs and a conditional fneg.
3999 SDValue Tmp0 = Op.getOperand(0);
4000 SDValue Tmp1 = Op.getOperand(1);
4002 EVT VT = Op.getValueType();
4003 EVT SrcVT = Tmp1.getValueType();
4004 bool InGPR = Tmp0.getOpcode() == ISD::BITCAST ||
4005 Tmp0.getOpcode() == ARMISD::VMOVDRR;
4006 bool UseNEON = !InGPR && Subtarget->hasNEON();
4009 // Use VBSL to copy the sign bit.
4010 unsigned EncodedVal = ARM_AM::createNEONModImm(0x6, 0x80);
4011 SDValue Mask = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v2i32,
4012 DAG.getTargetConstant(EncodedVal, MVT::i32));
4013 EVT OpVT = (VT == MVT::f32) ? MVT::v2i32 : MVT::v1i64;
4015 Mask = DAG.getNode(ARMISD::VSHL, dl, OpVT,
4016 DAG.getNode(ISD::BITCAST, dl, OpVT, Mask),
4017 DAG.getConstant(32, MVT::i32));
4018 else /*if (VT == MVT::f32)*/
4019 Tmp0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp0);
4020 if (SrcVT == MVT::f32) {
4021 Tmp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp1);
4023 Tmp1 = DAG.getNode(ARMISD::VSHL, dl, OpVT,
4024 DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1),
4025 DAG.getConstant(32, MVT::i32));
4026 } else if (VT == MVT::f32)
4027 Tmp1 = DAG.getNode(ARMISD::VSHRu, dl, MVT::v1i64,
4028 DAG.getNode(ISD::BITCAST, dl, MVT::v1i64, Tmp1),
4029 DAG.getConstant(32, MVT::i32));
4030 Tmp0 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp0);
4031 Tmp1 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1);
4033 SDValue AllOnes = DAG.getTargetConstant(ARM_AM::createNEONModImm(0xe, 0xff),
4035 AllOnes = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v8i8, AllOnes);
4036 SDValue MaskNot = DAG.getNode(ISD::XOR, dl, OpVT, Mask,
4037 DAG.getNode(ISD::BITCAST, dl, OpVT, AllOnes));
4039 SDValue Res = DAG.getNode(ISD::OR, dl, OpVT,
4040 DAG.getNode(ISD::AND, dl, OpVT, Tmp1, Mask),
4041 DAG.getNode(ISD::AND, dl, OpVT, Tmp0, MaskNot));
4042 if (VT == MVT::f32) {
4043 Res = DAG.getNode(ISD::BITCAST, dl, MVT::v2f32, Res);
4044 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, Res,
4045 DAG.getConstant(0, MVT::i32));
4047 Res = DAG.getNode(ISD::BITCAST, dl, MVT::f64, Res);
4053 // Bitcast operand 1 to i32.
4054 if (SrcVT == MVT::f64)
4055 Tmp1 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
4057 Tmp1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp1);
4059 // Or in the signbit with integer operations.
4060 SDValue Mask1 = DAG.getConstant(0x80000000, MVT::i32);
4061 SDValue Mask2 = DAG.getConstant(0x7fffffff, MVT::i32);
4062 Tmp1 = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp1, Mask1);
4063 if (VT == MVT::f32) {
4064 Tmp0 = DAG.getNode(ISD::AND, dl, MVT::i32,
4065 DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp0), Mask2);
4066 return DAG.getNode(ISD::BITCAST, dl, MVT::f32,
4067 DAG.getNode(ISD::OR, dl, MVT::i32, Tmp0, Tmp1));
4070 // f64: Or the high part with signbit and then combine two parts.
4071 Tmp0 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
4073 SDValue Lo = Tmp0.getValue(0);
4074 SDValue Hi = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp0.getValue(1), Mask2);
4075 Hi = DAG.getNode(ISD::OR, dl, MVT::i32, Hi, Tmp1);
4076 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
4079 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
4080 MachineFunction &MF = DAG.getMachineFunction();
4081 MachineFrameInfo *MFI = MF.getFrameInfo();
4082 MFI->setReturnAddressIsTaken(true);
4084 if (verifyReturnAddressArgumentIsConstant(Op, DAG))
4087 EVT VT = Op.getValueType();
4089 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
4091 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
4092 SDValue Offset = DAG.getConstant(4, MVT::i32);
4093 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
4094 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
4095 MachinePointerInfo(), false, false, false, 0);
4098 // Return LR, which contains the return address. Mark it an implicit live-in.
4099 unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32));
4100 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
4103 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
4104 const ARMBaseRegisterInfo &ARI =
4105 *static_cast<const ARMBaseRegisterInfo*>(RegInfo);
4106 MachineFunction &MF = DAG.getMachineFunction();
4107 MachineFrameInfo *MFI = MF.getFrameInfo();
4108 MFI->setFrameAddressIsTaken(true);
4110 EVT VT = Op.getValueType();
4111 SDLoc dl(Op); // FIXME probably not meaningful
4112 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
4113 unsigned FrameReg = ARI.getFrameRegister(MF);
4114 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
4116 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
4117 MachinePointerInfo(),
4118 false, false, false, 0);
4122 // FIXME? Maybe this could be a TableGen attribute on some registers and
4123 // this table could be generated automatically from RegInfo.
4124 unsigned ARMTargetLowering::getRegisterByName(const char* RegName,
4126 unsigned Reg = StringSwitch<unsigned>(RegName)
4127 .Case("sp", ARM::SP)
4131 report_fatal_error("Invalid register name global variable");
4134 /// ExpandBITCAST - If the target supports VFP, this function is called to
4135 /// expand a bit convert where either the source or destination type is i64 to
4136 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
4137 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
4138 /// vectors), since the legalizer won't know what to do with that.
4139 static SDValue ExpandBITCAST(SDNode *N, SelectionDAG &DAG) {
4140 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4142 SDValue Op = N->getOperand(0);
4144 // This function is only supposed to be called for i64 types, either as the
4145 // source or destination of the bit convert.
4146 EVT SrcVT = Op.getValueType();
4147 EVT DstVT = N->getValueType(0);
4148 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
4149 "ExpandBITCAST called for non-i64 type");
4151 // Turn i64->f64 into VMOVDRR.
4152 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
4153 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
4154 DAG.getConstant(0, MVT::i32));
4155 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
4156 DAG.getConstant(1, MVT::i32));
4157 return DAG.getNode(ISD::BITCAST, dl, DstVT,
4158 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
4161 // Turn f64->i64 into VMOVRRD.
4162 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
4164 if (TLI.isBigEndian() && SrcVT.isVector() &&
4165 SrcVT.getVectorNumElements() > 1)
4166 Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
4167 DAG.getVTList(MVT::i32, MVT::i32),
4168 DAG.getNode(ARMISD::VREV64, dl, SrcVT, Op));
4170 Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
4171 DAG.getVTList(MVT::i32, MVT::i32), Op);
4172 // Merge the pieces into a single i64 value.
4173 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
4179 /// getZeroVector - Returns a vector of specified type with all zero elements.
4180 /// Zero vectors are used to represent vector negation and in those cases
4181 /// will be implemented with the NEON VNEG instruction. However, VNEG does
4182 /// not support i64 elements, so sometimes the zero vectors will need to be
4183 /// explicitly constructed. Regardless, use a canonical VMOV to create the
4185 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, SDLoc dl) {
4186 assert(VT.isVector() && "Expected a vector type");
4187 // The canonical modified immediate encoding of a zero vector is....0!
4188 SDValue EncodedVal = DAG.getTargetConstant(0, MVT::i32);
4189 EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
4190 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal);
4191 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4194 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
4195 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
4196 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
4197 SelectionDAG &DAG) const {
4198 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4199 EVT VT = Op.getValueType();
4200 unsigned VTBits = VT.getSizeInBits();
4202 SDValue ShOpLo = Op.getOperand(0);
4203 SDValue ShOpHi = Op.getOperand(1);
4204 SDValue ShAmt = Op.getOperand(2);
4206 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
4208 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
4210 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
4211 DAG.getConstant(VTBits, MVT::i32), ShAmt);
4212 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
4213 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
4214 DAG.getConstant(VTBits, MVT::i32));
4215 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
4216 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
4217 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
4219 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
4220 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
4222 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
4223 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc,
4226 SDValue Ops[2] = { Lo, Hi };
4227 return DAG.getMergeValues(Ops, dl);
4230 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
4231 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
4232 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
4233 SelectionDAG &DAG) const {
4234 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4235 EVT VT = Op.getValueType();
4236 unsigned VTBits = VT.getSizeInBits();
4238 SDValue ShOpLo = Op.getOperand(0);
4239 SDValue ShOpHi = Op.getOperand(1);
4240 SDValue ShAmt = Op.getOperand(2);
4243 assert(Op.getOpcode() == ISD::SHL_PARTS);
4244 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
4245 DAG.getConstant(VTBits, MVT::i32), ShAmt);
4246 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
4247 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
4248 DAG.getConstant(VTBits, MVT::i32));
4249 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
4250 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
4252 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
4253 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
4254 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
4256 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
4257 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc,
4260 SDValue Ops[2] = { Lo, Hi };
4261 return DAG.getMergeValues(Ops, dl);
4264 SDValue ARMTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
4265 SelectionDAG &DAG) const {
4266 // The rounding mode is in bits 23:22 of the FPSCR.
4267 // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0
4268 // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3)
4269 // so that the shift + and get folded into a bitfield extract.
4271 SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::i32,
4272 DAG.getConstant(Intrinsic::arm_get_fpscr,
4274 SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPSCR,
4275 DAG.getConstant(1U << 22, MVT::i32));
4276 SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds,
4277 DAG.getConstant(22, MVT::i32));
4278 return DAG.getNode(ISD::AND, dl, MVT::i32, RMODE,
4279 DAG.getConstant(3, MVT::i32));
4282 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
4283 const ARMSubtarget *ST) {
4284 EVT VT = N->getValueType(0);
4287 if (!ST->hasV6T2Ops())
4290 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
4291 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
4294 /// getCTPOP16BitCounts - Returns a v8i8/v16i8 vector containing the bit-count
4295 /// for each 16-bit element from operand, repeated. The basic idea is to
4296 /// leverage vcnt to get the 8-bit counts, gather and add the results.
4298 /// Trace for v4i16:
4299 /// input = [v0 v1 v2 v3 ] (vi 16-bit element)
4300 /// cast: N0 = [w0 w1 w2 w3 w4 w5 w6 w7] (v0 = [w0 w1], wi 8-bit element)
4301 /// vcnt: N1 = [b0 b1 b2 b3 b4 b5 b6 b7] (bi = bit-count of 8-bit element wi)
4302 /// vrev: N2 = [b1 b0 b3 b2 b5 b4 b7 b6]
4303 /// [b0 b1 b2 b3 b4 b5 b6 b7]
4304 /// +[b1 b0 b3 b2 b5 b4 b7 b6]
4305 /// N3=N1+N2 = [k0 k0 k1 k1 k2 k2 k3 k3] (k0 = b0+b1 = bit-count of 16-bit v0,
4306 /// vuzp: = [k0 k1 k2 k3 k0 k1 k2 k3] each ki is 8-bits)
4307 static SDValue getCTPOP16BitCounts(SDNode *N, SelectionDAG &DAG) {
4308 EVT VT = N->getValueType(0);
4311 EVT VT8Bit = VT.is64BitVector() ? MVT::v8i8 : MVT::v16i8;
4312 SDValue N0 = DAG.getNode(ISD::BITCAST, DL, VT8Bit, N->getOperand(0));
4313 SDValue N1 = DAG.getNode(ISD::CTPOP, DL, VT8Bit, N0);
4314 SDValue N2 = DAG.getNode(ARMISD::VREV16, DL, VT8Bit, N1);
4315 SDValue N3 = DAG.getNode(ISD::ADD, DL, VT8Bit, N1, N2);
4316 return DAG.getNode(ARMISD::VUZP, DL, VT8Bit, N3, N3);
4319 /// lowerCTPOP16BitElements - Returns a v4i16/v8i16 vector containing the
4320 /// bit-count for each 16-bit element from the operand. We need slightly
4321 /// different sequencing for v4i16 and v8i16 to stay within NEON's available
4322 /// 64/128-bit registers.
4324 /// Trace for v4i16:
4325 /// input = [v0 v1 v2 v3 ] (vi 16-bit element)
4326 /// v8i8: BitCounts = [k0 k1 k2 k3 k0 k1 k2 k3 ] (ki is the bit-count of vi)
4327 /// v8i16:Extended = [k0 k1 k2 k3 k0 k1 k2 k3 ]
4328 /// v4i16:Extracted = [k0 k1 k2 k3 ]
4329 static SDValue lowerCTPOP16BitElements(SDNode *N, SelectionDAG &DAG) {
4330 EVT VT = N->getValueType(0);
4333 SDValue BitCounts = getCTPOP16BitCounts(N, DAG);
4334 if (VT.is64BitVector()) {
4335 SDValue Extended = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v8i16, BitCounts);
4336 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i16, Extended,
4337 DAG.getIntPtrConstant(0));
4339 SDValue Extracted = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v8i8,
4340 BitCounts, DAG.getIntPtrConstant(0));
4341 return DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v8i16, Extracted);
4345 /// lowerCTPOP32BitElements - Returns a v2i32/v4i32 vector containing the
4346 /// bit-count for each 32-bit element from the operand. The idea here is
4347 /// to split the vector into 16-bit elements, leverage the 16-bit count
4348 /// routine, and then combine the results.
4350 /// Trace for v2i32 (v4i32 similar with Extracted/Extended exchanged):
4351 /// input = [v0 v1 ] (vi: 32-bit elements)
4352 /// Bitcast = [w0 w1 w2 w3 ] (wi: 16-bit elements, v0 = [w0 w1])
4353 /// Counts16 = [k0 k1 k2 k3 ] (ki: 16-bit elements, bit-count of wi)
4354 /// vrev: N0 = [k1 k0 k3 k2 ]
4356 /// N1 =+[k1 k0 k3 k2 ]
4358 /// N2 =+[k1 k3 k0 k2 ]
4360 /// Extended =+[k1 k3 k0 k2 ]
4362 /// Extracted=+[k1 k3 ]
4364 static SDValue lowerCTPOP32BitElements(SDNode *N, SelectionDAG &DAG) {
4365 EVT VT = N->getValueType(0);
4368 EVT VT16Bit = VT.is64BitVector() ? MVT::v4i16 : MVT::v8i16;
4370 SDValue Bitcast = DAG.getNode(ISD::BITCAST, DL, VT16Bit, N->getOperand(0));
4371 SDValue Counts16 = lowerCTPOP16BitElements(Bitcast.getNode(), DAG);
4372 SDValue N0 = DAG.getNode(ARMISD::VREV32, DL, VT16Bit, Counts16);
4373 SDValue N1 = DAG.getNode(ISD::ADD, DL, VT16Bit, Counts16, N0);
4374 SDValue N2 = DAG.getNode(ARMISD::VUZP, DL, VT16Bit, N1, N1);
4376 if (VT.is64BitVector()) {
4377 SDValue Extended = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v4i32, N2);
4378 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i32, Extended,
4379 DAG.getIntPtrConstant(0));
4381 SDValue Extracted = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i16, N2,
4382 DAG.getIntPtrConstant(0));
4383 return DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v4i32, Extracted);
4387 static SDValue LowerCTPOP(SDNode *N, SelectionDAG &DAG,
4388 const ARMSubtarget *ST) {
4389 EVT VT = N->getValueType(0);
4391 assert(ST->hasNEON() && "Custom ctpop lowering requires NEON.");
4392 assert((VT == MVT::v2i32 || VT == MVT::v4i32 ||
4393 VT == MVT::v4i16 || VT == MVT::v8i16) &&
4394 "Unexpected type for custom ctpop lowering");
4396 if (VT.getVectorElementType() == MVT::i32)
4397 return lowerCTPOP32BitElements(N, DAG);
4399 return lowerCTPOP16BitElements(N, DAG);
4402 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
4403 const ARMSubtarget *ST) {
4404 EVT VT = N->getValueType(0);
4410 // Lower vector shifts on NEON to use VSHL.
4411 assert(ST->hasNEON() && "unexpected vector shift");
4413 // Left shifts translate directly to the vshiftu intrinsic.
4414 if (N->getOpcode() == ISD::SHL)
4415 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
4416 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
4417 N->getOperand(0), N->getOperand(1));
4419 assert((N->getOpcode() == ISD::SRA ||
4420 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
4422 // NEON uses the same intrinsics for both left and right shifts. For
4423 // right shifts, the shift amounts are negative, so negate the vector of
4425 EVT ShiftVT = N->getOperand(1).getValueType();
4426 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
4427 getZeroVector(ShiftVT, DAG, dl),
4429 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
4430 Intrinsic::arm_neon_vshifts :
4431 Intrinsic::arm_neon_vshiftu);
4432 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
4433 DAG.getConstant(vshiftInt, MVT::i32),
4434 N->getOperand(0), NegatedCount);
4437 static SDValue Expand64BitShift(SDNode *N, SelectionDAG &DAG,
4438 const ARMSubtarget *ST) {
4439 EVT VT = N->getValueType(0);
4442 // We can get here for a node like i32 = ISD::SHL i32, i64
4446 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
4447 "Unknown shift to lower!");
4449 // We only lower SRA, SRL of 1 here, all others use generic lowering.
4450 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
4451 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
4454 // If we are in thumb mode, we don't have RRX.
4455 if (ST->isThumb1Only()) return SDValue();
4457 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
4458 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
4459 DAG.getConstant(0, MVT::i32));
4460 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
4461 DAG.getConstant(1, MVT::i32));
4463 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
4464 // captures the result into a carry flag.
4465 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
4466 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Glue), Hi);
4468 // The low part is an ARMISD::RRX operand, which shifts the carry in.
4469 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
4471 // Merge the pieces into a single i64 value.
4472 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
4475 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
4476 SDValue TmpOp0, TmpOp1;
4477 bool Invert = false;
4481 SDValue Op0 = Op.getOperand(0);
4482 SDValue Op1 = Op.getOperand(1);
4483 SDValue CC = Op.getOperand(2);
4484 EVT VT = Op.getValueType();
4485 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
4488 if (Op1.getValueType().isFloatingPoint()) {
4489 switch (SetCCOpcode) {
4490 default: llvm_unreachable("Illegal FP comparison");
4492 case ISD::SETNE: Invert = true; // Fallthrough
4494 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
4496 case ISD::SETLT: Swap = true; // Fallthrough
4498 case ISD::SETGT: Opc = ARMISD::VCGT; break;
4500 case ISD::SETLE: Swap = true; // Fallthrough
4502 case ISD::SETGE: Opc = ARMISD::VCGE; break;
4503 case ISD::SETUGE: Swap = true; // Fallthrough
4504 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
4505 case ISD::SETUGT: Swap = true; // Fallthrough
4506 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
4507 case ISD::SETUEQ: Invert = true; // Fallthrough
4509 // Expand this to (OLT | OGT).
4513 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
4514 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
4516 case ISD::SETUO: Invert = true; // Fallthrough
4518 // Expand this to (OLT | OGE).
4522 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
4523 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
4527 // Integer comparisons.
4528 switch (SetCCOpcode) {
4529 default: llvm_unreachable("Illegal integer comparison");
4530 case ISD::SETNE: Invert = true;
4531 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
4532 case ISD::SETLT: Swap = true;
4533 case ISD::SETGT: Opc = ARMISD::VCGT; break;
4534 case ISD::SETLE: Swap = true;
4535 case ISD::SETGE: Opc = ARMISD::VCGE; break;
4536 case ISD::SETULT: Swap = true;
4537 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
4538 case ISD::SETULE: Swap = true;
4539 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
4542 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
4543 if (Opc == ARMISD::VCEQ) {
4546 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
4548 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
4551 // Ignore bitconvert.
4552 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BITCAST)
4553 AndOp = AndOp.getOperand(0);
4555 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
4557 Op0 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(0));
4558 Op1 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(1));
4565 std::swap(Op0, Op1);
4567 // If one of the operands is a constant vector zero, attempt to fold the
4568 // comparison to a specialized compare-against-zero form.
4570 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
4572 else if (ISD::isBuildVectorAllZeros(Op0.getNode())) {
4573 if (Opc == ARMISD::VCGE)
4574 Opc = ARMISD::VCLEZ;
4575 else if (Opc == ARMISD::VCGT)
4576 Opc = ARMISD::VCLTZ;
4581 if (SingleOp.getNode()) {
4584 Result = DAG.getNode(ARMISD::VCEQZ, dl, VT, SingleOp); break;
4586 Result = DAG.getNode(ARMISD::VCGEZ, dl, VT, SingleOp); break;
4588 Result = DAG.getNode(ARMISD::VCLEZ, dl, VT, SingleOp); break;
4590 Result = DAG.getNode(ARMISD::VCGTZ, dl, VT, SingleOp); break;
4592 Result = DAG.getNode(ARMISD::VCLTZ, dl, VT, SingleOp); break;
4594 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
4597 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
4601 Result = DAG.getNOT(dl, Result, VT);
4606 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
4607 /// valid vector constant for a NEON instruction with a "modified immediate"
4608 /// operand (e.g., VMOV). If so, return the encoded value.
4609 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
4610 unsigned SplatBitSize, SelectionDAG &DAG,
4611 EVT &VT, bool is128Bits, NEONModImmType type) {
4612 unsigned OpCmode, Imm;
4614 // SplatBitSize is set to the smallest size that splats the vector, so a
4615 // zero vector will always have SplatBitSize == 8. However, NEON modified
4616 // immediate instructions others than VMOV do not support the 8-bit encoding
4617 // of a zero vector, and the default encoding of zero is supposed to be the
4622 switch (SplatBitSize) {
4624 if (type != VMOVModImm)
4626 // Any 1-byte value is OK. Op=0, Cmode=1110.
4627 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
4630 VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
4634 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
4635 VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
4636 if ((SplatBits & ~0xff) == 0) {
4637 // Value = 0x00nn: Op=x, Cmode=100x.
4642 if ((SplatBits & ~0xff00) == 0) {
4643 // Value = 0xnn00: Op=x, Cmode=101x.
4645 Imm = SplatBits >> 8;
4651 // NEON's 32-bit VMOV supports splat values where:
4652 // * only one byte is nonzero, or
4653 // * the least significant byte is 0xff and the second byte is nonzero, or
4654 // * the least significant 2 bytes are 0xff and the third is nonzero.
4655 VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
4656 if ((SplatBits & ~0xff) == 0) {
4657 // Value = 0x000000nn: Op=x, Cmode=000x.
4662 if ((SplatBits & ~0xff00) == 0) {
4663 // Value = 0x0000nn00: Op=x, Cmode=001x.
4665 Imm = SplatBits >> 8;
4668 if ((SplatBits & ~0xff0000) == 0) {
4669 // Value = 0x00nn0000: Op=x, Cmode=010x.
4671 Imm = SplatBits >> 16;
4674 if ((SplatBits & ~0xff000000) == 0) {
4675 // Value = 0xnn000000: Op=x, Cmode=011x.
4677 Imm = SplatBits >> 24;
4681 // cmode == 0b1100 and cmode == 0b1101 are not supported for VORR or VBIC
4682 if (type == OtherModImm) return SDValue();
4684 if ((SplatBits & ~0xffff) == 0 &&
4685 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
4686 // Value = 0x0000nnff: Op=x, Cmode=1100.
4688 Imm = SplatBits >> 8;
4692 if ((SplatBits & ~0xffffff) == 0 &&
4693 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
4694 // Value = 0x00nnffff: Op=x, Cmode=1101.
4696 Imm = SplatBits >> 16;
4700 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
4701 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
4702 // VMOV.I32. A (very) minor optimization would be to replicate the value
4703 // and fall through here to test for a valid 64-bit splat. But, then the
4704 // caller would also need to check and handle the change in size.
4708 if (type != VMOVModImm)
4710 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
4711 uint64_t BitMask = 0xff;
4713 unsigned ImmMask = 1;
4715 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
4716 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
4719 } else if ((SplatBits & BitMask) != 0) {
4726 if (DAG.getTargetLoweringInfo().isBigEndian())
4727 // swap higher and lower 32 bit word
4728 Imm = ((Imm & 0xf) << 4) | ((Imm & 0xf0) >> 4);
4730 // Op=1, Cmode=1110.
4732 VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
4737 llvm_unreachable("unexpected size for isNEONModifiedImm");
4740 unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm);
4741 return DAG.getTargetConstant(EncodedVal, MVT::i32);
4744 SDValue ARMTargetLowering::LowerConstantFP(SDValue Op, SelectionDAG &DAG,
4745 const ARMSubtarget *ST) const {
4749 bool IsDouble = Op.getValueType() == MVT::f64;
4750 ConstantFPSDNode *CFP = cast<ConstantFPSDNode>(Op);
4752 // Use the default (constant pool) lowering for double constants when we have
4754 if (IsDouble && Subtarget->isFPOnlySP())
4757 // Try splatting with a VMOV.f32...
4758 APFloat FPVal = CFP->getValueAPF();
4759 int ImmVal = IsDouble ? ARM_AM::getFP64Imm(FPVal) : ARM_AM::getFP32Imm(FPVal);
4762 if (IsDouble || !ST->useNEONForSinglePrecisionFP()) {
4763 // We have code in place to select a valid ConstantFP already, no need to
4768 // It's a float and we are trying to use NEON operations where
4769 // possible. Lower it to a splat followed by an extract.
4771 SDValue NewVal = DAG.getTargetConstant(ImmVal, MVT::i32);
4772 SDValue VecConstant = DAG.getNode(ARMISD::VMOVFPIMM, DL, MVT::v2f32,
4774 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecConstant,
4775 DAG.getConstant(0, MVT::i32));
4778 // The rest of our options are NEON only, make sure that's allowed before
4780 if (!ST->hasNEON() || (!IsDouble && !ST->useNEONForSinglePrecisionFP()))
4784 uint64_t iVal = FPVal.bitcastToAPInt().getZExtValue();
4786 // It wouldn't really be worth bothering for doubles except for one very
4787 // important value, which does happen to match: 0.0. So make sure we don't do
4789 if (IsDouble && (iVal & 0xffffffff) != (iVal >> 32))
4792 // Try a VMOV.i32 (FIXME: i8, i16, or i64 could work too).
4793 SDValue NewVal = isNEONModifiedImm(iVal & 0xffffffffU, 0, 32, DAG, VMovVT,
4795 if (NewVal != SDValue()) {
4797 SDValue VecConstant = DAG.getNode(ARMISD::VMOVIMM, DL, VMovVT,
4800 return DAG.getNode(ISD::BITCAST, DL, MVT::f64, VecConstant);
4802 // It's a float: cast and extract a vector element.
4803 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
4805 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
4806 DAG.getConstant(0, MVT::i32));
4809 // Finally, try a VMVN.i32
4810 NewVal = isNEONModifiedImm(~iVal & 0xffffffffU, 0, 32, DAG, VMovVT,
4812 if (NewVal != SDValue()) {
4814 SDValue VecConstant = DAG.getNode(ARMISD::VMVNIMM, DL, VMovVT, NewVal);
4817 return DAG.getNode(ISD::BITCAST, DL, MVT::f64, VecConstant);
4819 // It's a float: cast and extract a vector element.
4820 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
4822 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
4823 DAG.getConstant(0, MVT::i32));
4829 // check if an VEXT instruction can handle the shuffle mask when the
4830 // vector sources of the shuffle are the same.
4831 static bool isSingletonVEXTMask(ArrayRef<int> M, EVT VT, unsigned &Imm) {
4832 unsigned NumElts = VT.getVectorNumElements();
4834 // Assume that the first shuffle index is not UNDEF. Fail if it is.
4840 // If this is a VEXT shuffle, the immediate value is the index of the first
4841 // element. The other shuffle indices must be the successive elements after
4843 unsigned ExpectedElt = Imm;
4844 for (unsigned i = 1; i < NumElts; ++i) {
4845 // Increment the expected index. If it wraps around, just follow it
4846 // back to index zero and keep going.
4848 if (ExpectedElt == NumElts)
4851 if (M[i] < 0) continue; // ignore UNDEF indices
4852 if (ExpectedElt != static_cast<unsigned>(M[i]))
4860 static bool isVEXTMask(ArrayRef<int> M, EVT VT,
4861 bool &ReverseVEXT, unsigned &Imm) {
4862 unsigned NumElts = VT.getVectorNumElements();
4863 ReverseVEXT = false;
4865 // Assume that the first shuffle index is not UNDEF. Fail if it is.
4871 // If this is a VEXT shuffle, the immediate value is the index of the first
4872 // element. The other shuffle indices must be the successive elements after
4874 unsigned ExpectedElt = Imm;
4875 for (unsigned i = 1; i < NumElts; ++i) {
4876 // Increment the expected index. If it wraps around, it may still be
4877 // a VEXT but the source vectors must be swapped.
4879 if (ExpectedElt == NumElts * 2) {
4884 if (M[i] < 0) continue; // ignore UNDEF indices
4885 if (ExpectedElt != static_cast<unsigned>(M[i]))
4889 // Adjust the index value if the source operands will be swapped.
4896 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
4897 /// instruction with the specified blocksize. (The order of the elements
4898 /// within each block of the vector is reversed.)
4899 static bool isVREVMask(ArrayRef<int> M, EVT VT, unsigned BlockSize) {
4900 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
4901 "Only possible block sizes for VREV are: 16, 32, 64");
4903 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4907 unsigned NumElts = VT.getVectorNumElements();
4908 unsigned BlockElts = M[0] + 1;
4909 // If the first shuffle index is UNDEF, be optimistic.
4911 BlockElts = BlockSize / EltSz;
4913 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
4916 for (unsigned i = 0; i < NumElts; ++i) {
4917 if (M[i] < 0) continue; // ignore UNDEF indices
4918 if ((unsigned) M[i] != (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
4925 static bool isVTBLMask(ArrayRef<int> M, EVT VT) {
4926 // We can handle <8 x i8> vector shuffles. If the index in the mask is out of
4927 // range, then 0 is placed into the resulting vector. So pretty much any mask
4928 // of 8 elements can work here.
4929 return VT == MVT::v8i8 && M.size() == 8;
4932 static bool isVTRNMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4933 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4937 unsigned NumElts = VT.getVectorNumElements();
4938 WhichResult = (M[0] == 0 ? 0 : 1);
4939 for (unsigned i = 0; i < NumElts; i += 2) {
4940 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
4941 (M[i+1] >= 0 && (unsigned) M[i+1] != i + NumElts + WhichResult))
4947 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
4948 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4949 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
4950 static bool isVTRN_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4951 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4955 unsigned NumElts = VT.getVectorNumElements();
4956 WhichResult = (M[0] == 0 ? 0 : 1);
4957 for (unsigned i = 0; i < NumElts; i += 2) {
4958 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
4959 (M[i+1] >= 0 && (unsigned) M[i+1] != i + WhichResult))
4965 static bool isVUZPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4966 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4970 unsigned NumElts = VT.getVectorNumElements();
4971 WhichResult = (M[0] == 0 ? 0 : 1);
4972 for (unsigned i = 0; i != NumElts; ++i) {
4973 if (M[i] < 0) continue; // ignore UNDEF indices
4974 if ((unsigned) M[i] != 2 * i + WhichResult)
4978 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4979 if (VT.is64BitVector() && EltSz == 32)
4985 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
4986 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4987 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
4988 static bool isVUZP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4989 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4993 unsigned Half = VT.getVectorNumElements() / 2;
4994 WhichResult = (M[0] == 0 ? 0 : 1);
4995 for (unsigned j = 0; j != 2; ++j) {
4996 unsigned Idx = WhichResult;
4997 for (unsigned i = 0; i != Half; ++i) {
4998 int MIdx = M[i + j * Half];
4999 if (MIdx >= 0 && (unsigned) MIdx != Idx)
5005 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
5006 if (VT.is64BitVector() && EltSz == 32)
5012 static bool isVZIPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
5013 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
5017 unsigned NumElts = VT.getVectorNumElements();
5018 WhichResult = (M[0] == 0 ? 0 : 1);
5019 unsigned Idx = WhichResult * NumElts / 2;
5020 for (unsigned i = 0; i != NumElts; i += 2) {
5021 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
5022 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx + NumElts))
5027 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
5028 if (VT.is64BitVector() && EltSz == 32)
5034 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
5035 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
5036 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
5037 static bool isVZIP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
5038 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
5042 unsigned NumElts = VT.getVectorNumElements();
5043 WhichResult = (M[0] == 0 ? 0 : 1);
5044 unsigned Idx = WhichResult * NumElts / 2;
5045 for (unsigned i = 0; i != NumElts; i += 2) {
5046 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
5047 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx))
5052 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
5053 if (VT.is64BitVector() && EltSz == 32)
5059 /// \return true if this is a reverse operation on an vector.
5060 static bool isReverseMask(ArrayRef<int> M, EVT VT) {
5061 unsigned NumElts = VT.getVectorNumElements();
5062 // Make sure the mask has the right size.
5063 if (NumElts != M.size())
5066 // Look for <15, ..., 3, -1, 1, 0>.
5067 for (unsigned i = 0; i != NumElts; ++i)
5068 if (M[i] >= 0 && M[i] != (int) (NumElts - 1 - i))
5074 // If N is an integer constant that can be moved into a register in one
5075 // instruction, return an SDValue of such a constant (will become a MOV
5076 // instruction). Otherwise return null.
5077 static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG,
5078 const ARMSubtarget *ST, SDLoc dl) {
5080 if (!isa<ConstantSDNode>(N))
5082 Val = cast<ConstantSDNode>(N)->getZExtValue();
5084 if (ST->isThumb1Only()) {
5085 if (Val <= 255 || ~Val <= 255)
5086 return DAG.getConstant(Val, MVT::i32);
5088 if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1)
5089 return DAG.getConstant(Val, MVT::i32);
5094 // If this is a case we can't handle, return null and let the default
5095 // expansion code take care of it.
5096 SDValue ARMTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
5097 const ARMSubtarget *ST) const {
5098 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
5100 EVT VT = Op.getValueType();
5102 APInt SplatBits, SplatUndef;
5103 unsigned SplatBitSize;
5105 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
5106 if (SplatBitSize <= 64) {
5107 // Check if an immediate VMOV works.
5109 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
5110 SplatUndef.getZExtValue(), SplatBitSize,
5111 DAG, VmovVT, VT.is128BitVector(),
5113 if (Val.getNode()) {
5114 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
5115 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
5118 // Try an immediate VMVN.
5119 uint64_t NegatedImm = (~SplatBits).getZExtValue();
5120 Val = isNEONModifiedImm(NegatedImm,
5121 SplatUndef.getZExtValue(), SplatBitSize,
5122 DAG, VmovVT, VT.is128BitVector(),
5124 if (Val.getNode()) {
5125 SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
5126 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
5129 // Use vmov.f32 to materialize other v2f32 and v4f32 splats.
5130 if ((VT == MVT::v2f32 || VT == MVT::v4f32) && SplatBitSize == 32) {
5131 int ImmVal = ARM_AM::getFP32Imm(SplatBits);
5133 SDValue Val = DAG.getTargetConstant(ImmVal, MVT::i32);
5134 return DAG.getNode(ARMISD::VMOVFPIMM, dl, VT, Val);
5140 // Scan through the operands to see if only one value is used.
5142 // As an optimisation, even if more than one value is used it may be more
5143 // profitable to splat with one value then change some lanes.
5145 // Heuristically we decide to do this if the vector has a "dominant" value,
5146 // defined as splatted to more than half of the lanes.
5147 unsigned NumElts = VT.getVectorNumElements();
5148 bool isOnlyLowElement = true;
5149 bool usesOnlyOneValue = true;
5150 bool hasDominantValue = false;
5151 bool isConstant = true;
5153 // Map of the number of times a particular SDValue appears in the
5155 DenseMap<SDValue, unsigned> ValueCounts;
5157 for (unsigned i = 0; i < NumElts; ++i) {
5158 SDValue V = Op.getOperand(i);
5159 if (V.getOpcode() == ISD::UNDEF)
5162 isOnlyLowElement = false;
5163 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
5166 ValueCounts.insert(std::make_pair(V, 0));
5167 unsigned &Count = ValueCounts[V];
5169 // Is this value dominant? (takes up more than half of the lanes)
5170 if (++Count > (NumElts / 2)) {
5171 hasDominantValue = true;
5175 if (ValueCounts.size() != 1)
5176 usesOnlyOneValue = false;
5177 if (!Value.getNode() && ValueCounts.size() > 0)
5178 Value = ValueCounts.begin()->first;
5180 if (ValueCounts.size() == 0)
5181 return DAG.getUNDEF(VT);
5183 // Loads are better lowered with insert_vector_elt/ARMISD::BUILD_VECTOR.
5184 // Keep going if we are hitting this case.
5185 if (isOnlyLowElement && !ISD::isNormalLoad(Value.getNode()))
5186 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
5188 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5190 // Use VDUP for non-constant splats. For f32 constant splats, reduce to
5191 // i32 and try again.
5192 if (hasDominantValue && EltSize <= 32) {
5196 // If we are VDUPing a value that comes directly from a vector, that will
5197 // cause an unnecessary move to and from a GPR, where instead we could
5198 // just use VDUPLANE. We can only do this if the lane being extracted
5199 // is at a constant index, as the VDUP from lane instructions only have
5200 // constant-index forms.
5201 if (Value->getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5202 isa<ConstantSDNode>(Value->getOperand(1))) {
5203 // We need to create a new undef vector to use for the VDUPLANE if the
5204 // size of the vector from which we get the value is different than the
5205 // size of the vector that we need to create. We will insert the element
5206 // such that the register coalescer will remove unnecessary copies.
5207 if (VT != Value->getOperand(0).getValueType()) {
5208 ConstantSDNode *constIndex;
5209 constIndex = dyn_cast<ConstantSDNode>(Value->getOperand(1));
5210 assert(constIndex && "The index is not a constant!");
5211 unsigned index = constIndex->getAPIntValue().getLimitedValue() %
5212 VT.getVectorNumElements();
5213 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
5214 DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DAG.getUNDEF(VT),
5215 Value, DAG.getConstant(index, MVT::i32)),
5216 DAG.getConstant(index, MVT::i32));
5218 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
5219 Value->getOperand(0), Value->getOperand(1));
5221 N = DAG.getNode(ARMISD::VDUP, dl, VT, Value);
5223 if (!usesOnlyOneValue) {
5224 // The dominant value was splatted as 'N', but we now have to insert
5225 // all differing elements.
5226 for (unsigned I = 0; I < NumElts; ++I) {
5227 if (Op.getOperand(I) == Value)
5229 SmallVector<SDValue, 3> Ops;
5231 Ops.push_back(Op.getOperand(I));
5232 Ops.push_back(DAG.getConstant(I, MVT::i32));
5233 N = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Ops);
5238 if (VT.getVectorElementType().isFloatingPoint()) {
5239 SmallVector<SDValue, 8> Ops;
5240 for (unsigned i = 0; i < NumElts; ++i)
5241 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, MVT::i32,
5243 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
5244 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, Ops);
5245 Val = LowerBUILD_VECTOR(Val, DAG, ST);
5247 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
5249 if (usesOnlyOneValue) {
5250 SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
5251 if (isConstant && Val.getNode())
5252 return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
5256 // If all elements are constants and the case above didn't get hit, fall back
5257 // to the default expansion, which will generate a load from the constant
5262 // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
5264 SDValue shuffle = ReconstructShuffle(Op, DAG);
5265 if (shuffle != SDValue())
5269 // Vectors with 32- or 64-bit elements can be built by directly assigning
5270 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
5271 // will be legalized.
5272 if (EltSize >= 32) {
5273 // Do the expansion with floating-point types, since that is what the VFP
5274 // registers are defined to use, and since i64 is not legal.
5275 EVT EltVT = EVT::getFloatingPointVT(EltSize);
5276 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
5277 SmallVector<SDValue, 8> Ops;
5278 for (unsigned i = 0; i < NumElts; ++i)
5279 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, EltVT, Op.getOperand(i)));
5280 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, Ops);
5281 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
5284 // If all else fails, just use a sequence of INSERT_VECTOR_ELT when we
5285 // know the default expansion would otherwise fall back on something even
5286 // worse. For a vector with one or two non-undef values, that's
5287 // scalar_to_vector for the elements followed by a shuffle (provided the
5288 // shuffle is valid for the target) and materialization element by element
5289 // on the stack followed by a load for everything else.
5290 if (!isConstant && !usesOnlyOneValue) {
5291 SDValue Vec = DAG.getUNDEF(VT);
5292 for (unsigned i = 0 ; i < NumElts; ++i) {
5293 SDValue V = Op.getOperand(i);
5294 if (V.getOpcode() == ISD::UNDEF)
5296 SDValue LaneIdx = DAG.getConstant(i, MVT::i32);
5297 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Vec, V, LaneIdx);
5305 // Gather data to see if the operation can be modelled as a
5306 // shuffle in combination with VEXTs.
5307 SDValue ARMTargetLowering::ReconstructShuffle(SDValue Op,
5308 SelectionDAG &DAG) const {
5310 EVT VT = Op.getValueType();
5311 unsigned NumElts = VT.getVectorNumElements();
5313 SmallVector<SDValue, 2> SourceVecs;
5314 SmallVector<unsigned, 2> MinElts;
5315 SmallVector<unsigned, 2> MaxElts;
5317 for (unsigned i = 0; i < NumElts; ++i) {
5318 SDValue V = Op.getOperand(i);
5319 if (V.getOpcode() == ISD::UNDEF)
5321 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
5322 // A shuffle can only come from building a vector from various
5323 // elements of other vectors.
5325 } else if (V.getOperand(0).getValueType().getVectorElementType() !=
5326 VT.getVectorElementType()) {
5327 // This code doesn't know how to handle shuffles where the vector
5328 // element types do not match (this happens because type legalization
5329 // promotes the return type of EXTRACT_VECTOR_ELT).
5330 // FIXME: It might be appropriate to extend this code to handle
5331 // mismatched types.
5335 // Record this extraction against the appropriate vector if possible...
5336 SDValue SourceVec = V.getOperand(0);
5337 // If the element number isn't a constant, we can't effectively
5338 // analyze what's going on.
5339 if (!isa<ConstantSDNode>(V.getOperand(1)))
5341 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
5342 bool FoundSource = false;
5343 for (unsigned j = 0; j < SourceVecs.size(); ++j) {
5344 if (SourceVecs[j] == SourceVec) {
5345 if (MinElts[j] > EltNo)
5347 if (MaxElts[j] < EltNo)
5354 // Or record a new source if not...
5356 SourceVecs.push_back(SourceVec);
5357 MinElts.push_back(EltNo);
5358 MaxElts.push_back(EltNo);
5362 // Currently only do something sane when at most two source vectors
5364 if (SourceVecs.size() > 2)
5367 SDValue ShuffleSrcs[2] = {DAG.getUNDEF(VT), DAG.getUNDEF(VT) };
5368 int VEXTOffsets[2] = {0, 0};
5370 // This loop extracts the usage patterns of the source vectors
5371 // and prepares appropriate SDValues for a shuffle if possible.
5372 for (unsigned i = 0; i < SourceVecs.size(); ++i) {
5373 if (SourceVecs[i].getValueType() == VT) {
5374 // No VEXT necessary
5375 ShuffleSrcs[i] = SourceVecs[i];
5378 } else if (SourceVecs[i].getValueType().getVectorNumElements() < NumElts) {
5379 // It probably isn't worth padding out a smaller vector just to
5380 // break it down again in a shuffle.
5384 // Since only 64-bit and 128-bit vectors are legal on ARM and
5385 // we've eliminated the other cases...
5386 assert(SourceVecs[i].getValueType().getVectorNumElements() == 2*NumElts &&
5387 "unexpected vector sizes in ReconstructShuffle");
5389 if (MaxElts[i] - MinElts[i] >= NumElts) {
5390 // Span too large for a VEXT to cope
5394 if (MinElts[i] >= NumElts) {
5395 // The extraction can just take the second half
5396 VEXTOffsets[i] = NumElts;
5397 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5399 DAG.getIntPtrConstant(NumElts));
5400 } else if (MaxElts[i] < NumElts) {
5401 // The extraction can just take the first half
5403 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5405 DAG.getIntPtrConstant(0));
5407 // An actual VEXT is needed
5408 VEXTOffsets[i] = MinElts[i];
5409 SDValue VEXTSrc1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5411 DAG.getIntPtrConstant(0));
5412 SDValue VEXTSrc2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5414 DAG.getIntPtrConstant(NumElts));
5415 ShuffleSrcs[i] = DAG.getNode(ARMISD::VEXT, dl, VT, VEXTSrc1, VEXTSrc2,
5416 DAG.getConstant(VEXTOffsets[i], MVT::i32));
5420 SmallVector<int, 8> Mask;
5422 for (unsigned i = 0; i < NumElts; ++i) {
5423 SDValue Entry = Op.getOperand(i);
5424 if (Entry.getOpcode() == ISD::UNDEF) {
5429 SDValue ExtractVec = Entry.getOperand(0);
5430 int ExtractElt = cast<ConstantSDNode>(Op.getOperand(i)
5431 .getOperand(1))->getSExtValue();
5432 if (ExtractVec == SourceVecs[0]) {
5433 Mask.push_back(ExtractElt - VEXTOffsets[0]);
5435 Mask.push_back(ExtractElt + NumElts - VEXTOffsets[1]);
5439 // Final check before we try to produce nonsense...
5440 if (isShuffleMaskLegal(Mask, VT))
5441 return DAG.getVectorShuffle(VT, dl, ShuffleSrcs[0], ShuffleSrcs[1],
5447 /// isShuffleMaskLegal - Targets can use this to indicate that they only
5448 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
5449 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
5450 /// are assumed to be legal.
5452 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
5454 if (VT.getVectorNumElements() == 4 &&
5455 (VT.is128BitVector() || VT.is64BitVector())) {
5456 unsigned PFIndexes[4];
5457 for (unsigned i = 0; i != 4; ++i) {
5461 PFIndexes[i] = M[i];
5464 // Compute the index in the perfect shuffle table.
5465 unsigned PFTableIndex =
5466 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
5467 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
5468 unsigned Cost = (PFEntry >> 30);
5475 unsigned Imm, WhichResult;
5477 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5478 return (EltSize >= 32 ||
5479 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
5480 isVREVMask(M, VT, 64) ||
5481 isVREVMask(M, VT, 32) ||
5482 isVREVMask(M, VT, 16) ||
5483 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
5484 isVTBLMask(M, VT) ||
5485 isVTRNMask(M, VT, WhichResult) ||
5486 isVUZPMask(M, VT, WhichResult) ||
5487 isVZIPMask(M, VT, WhichResult) ||
5488 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
5489 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
5490 isVZIP_v_undef_Mask(M, VT, WhichResult) ||
5491 ((VT == MVT::v8i16 || VT == MVT::v16i8) && isReverseMask(M, VT)));
5494 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
5495 /// the specified operations to build the shuffle.
5496 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
5497 SDValue RHS, SelectionDAG &DAG,
5499 unsigned OpNum = (PFEntry >> 26) & 0x0F;
5500 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
5501 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
5504 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
5513 OP_VUZPL, // VUZP, left result
5514 OP_VUZPR, // VUZP, right result
5515 OP_VZIPL, // VZIP, left result
5516 OP_VZIPR, // VZIP, right result
5517 OP_VTRNL, // VTRN, left result
5518 OP_VTRNR // VTRN, right result
5521 if (OpNum == OP_COPY) {
5522 if (LHSID == (1*9+2)*9+3) return LHS;
5523 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
5527 SDValue OpLHS, OpRHS;
5528 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
5529 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
5530 EVT VT = OpLHS.getValueType();
5533 default: llvm_unreachable("Unknown shuffle opcode!");
5535 // VREV divides the vector in half and swaps within the half.
5536 if (VT.getVectorElementType() == MVT::i32 ||
5537 VT.getVectorElementType() == MVT::f32)
5538 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
5539 // vrev <4 x i16> -> VREV32
5540 if (VT.getVectorElementType() == MVT::i16)
5541 return DAG.getNode(ARMISD::VREV32, dl, VT, OpLHS);
5542 // vrev <4 x i8> -> VREV16
5543 assert(VT.getVectorElementType() == MVT::i8);
5544 return DAG.getNode(ARMISD::VREV16, dl, VT, OpLHS);
5549 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
5550 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
5554 return DAG.getNode(ARMISD::VEXT, dl, VT,
5556 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
5559 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
5560 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
5563 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
5564 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
5567 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
5568 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
5572 static SDValue LowerVECTOR_SHUFFLEv8i8(SDValue Op,
5573 ArrayRef<int> ShuffleMask,
5574 SelectionDAG &DAG) {
5575 // Check to see if we can use the VTBL instruction.
5576 SDValue V1 = Op.getOperand(0);
5577 SDValue V2 = Op.getOperand(1);
5580 SmallVector<SDValue, 8> VTBLMask;
5581 for (ArrayRef<int>::iterator
5582 I = ShuffleMask.begin(), E = ShuffleMask.end(); I != E; ++I)
5583 VTBLMask.push_back(DAG.getConstant(*I, MVT::i32));
5585 if (V2.getNode()->getOpcode() == ISD::UNDEF)
5586 return DAG.getNode(ARMISD::VTBL1, DL, MVT::v8i8, V1,
5587 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8, VTBLMask));
5589 return DAG.getNode(ARMISD::VTBL2, DL, MVT::v8i8, V1, V2,
5590 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8, VTBLMask));
5593 static SDValue LowerReverse_VECTOR_SHUFFLEv16i8_v8i16(SDValue Op,
5594 SelectionDAG &DAG) {
5596 SDValue OpLHS = Op.getOperand(0);
5597 EVT VT = OpLHS.getValueType();
5599 assert((VT == MVT::v8i16 || VT == MVT::v16i8) &&
5600 "Expect an v8i16/v16i8 type");
5601 OpLHS = DAG.getNode(ARMISD::VREV64, DL, VT, OpLHS);
5602 // For a v16i8 type: After the VREV, we have got <8, ...15, 8, ..., 0>. Now,
5603 // extract the first 8 bytes into the top double word and the last 8 bytes
5604 // into the bottom double word. The v8i16 case is similar.
5605 unsigned ExtractNum = (VT == MVT::v16i8) ? 8 : 4;
5606 return DAG.getNode(ARMISD::VEXT, DL, VT, OpLHS, OpLHS,
5607 DAG.getConstant(ExtractNum, MVT::i32));
5610 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
5611 SDValue V1 = Op.getOperand(0);
5612 SDValue V2 = Op.getOperand(1);
5614 EVT VT = Op.getValueType();
5615 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
5617 // Convert shuffles that are directly supported on NEON to target-specific
5618 // DAG nodes, instead of keeping them as shuffles and matching them again
5619 // during code selection. This is more efficient and avoids the possibility
5620 // of inconsistencies between legalization and selection.
5621 // FIXME: floating-point vectors should be canonicalized to integer vectors
5622 // of the same time so that they get CSEd properly.
5623 ArrayRef<int> ShuffleMask = SVN->getMask();
5625 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5626 if (EltSize <= 32) {
5627 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
5628 int Lane = SVN->getSplatIndex();
5629 // If this is undef splat, generate it via "just" vdup, if possible.
5630 if (Lane == -1) Lane = 0;
5632 // Test if V1 is a SCALAR_TO_VECTOR.
5633 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
5634 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
5636 // Test if V1 is a BUILD_VECTOR which is equivalent to a SCALAR_TO_VECTOR
5637 // (and probably will turn into a SCALAR_TO_VECTOR once legalization
5639 if (Lane == 0 && V1.getOpcode() == ISD::BUILD_VECTOR &&
5640 !isa<ConstantSDNode>(V1.getOperand(0))) {
5641 bool IsScalarToVector = true;
5642 for (unsigned i = 1, e = V1.getNumOperands(); i != e; ++i)
5643 if (V1.getOperand(i).getOpcode() != ISD::UNDEF) {
5644 IsScalarToVector = false;
5647 if (IsScalarToVector)
5648 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
5650 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
5651 DAG.getConstant(Lane, MVT::i32));
5656 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
5659 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
5660 DAG.getConstant(Imm, MVT::i32));
5663 if (isVREVMask(ShuffleMask, VT, 64))
5664 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
5665 if (isVREVMask(ShuffleMask, VT, 32))
5666 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
5667 if (isVREVMask(ShuffleMask, VT, 16))
5668 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
5670 if (V2->getOpcode() == ISD::UNDEF &&
5671 isSingletonVEXTMask(ShuffleMask, VT, Imm)) {
5672 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V1,
5673 DAG.getConstant(Imm, MVT::i32));
5676 // Check for Neon shuffles that modify both input vectors in place.
5677 // If both results are used, i.e., if there are two shuffles with the same
5678 // source operands and with masks corresponding to both results of one of
5679 // these operations, DAG memoization will ensure that a single node is
5680 // used for both shuffles.
5681 unsigned WhichResult;
5682 if (isVTRNMask(ShuffleMask, VT, WhichResult))
5683 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
5684 V1, V2).getValue(WhichResult);
5685 if (isVUZPMask(ShuffleMask, VT, WhichResult))
5686 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
5687 V1, V2).getValue(WhichResult);
5688 if (isVZIPMask(ShuffleMask, VT, WhichResult))
5689 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
5690 V1, V2).getValue(WhichResult);
5692 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
5693 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
5694 V1, V1).getValue(WhichResult);
5695 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
5696 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
5697 V1, V1).getValue(WhichResult);
5698 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
5699 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
5700 V1, V1).getValue(WhichResult);
5703 // If the shuffle is not directly supported and it has 4 elements, use
5704 // the PerfectShuffle-generated table to synthesize it from other shuffles.
5705 unsigned NumElts = VT.getVectorNumElements();
5707 unsigned PFIndexes[4];
5708 for (unsigned i = 0; i != 4; ++i) {
5709 if (ShuffleMask[i] < 0)
5712 PFIndexes[i] = ShuffleMask[i];
5715 // Compute the index in the perfect shuffle table.
5716 unsigned PFTableIndex =
5717 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
5718 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
5719 unsigned Cost = (PFEntry >> 30);
5722 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
5725 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
5726 if (EltSize >= 32) {
5727 // Do the expansion with floating-point types, since that is what the VFP
5728 // registers are defined to use, and since i64 is not legal.
5729 EVT EltVT = EVT::getFloatingPointVT(EltSize);
5730 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
5731 V1 = DAG.getNode(ISD::BITCAST, dl, VecVT, V1);
5732 V2 = DAG.getNode(ISD::BITCAST, dl, VecVT, V2);
5733 SmallVector<SDValue, 8> Ops;
5734 for (unsigned i = 0; i < NumElts; ++i) {
5735 if (ShuffleMask[i] < 0)
5736 Ops.push_back(DAG.getUNDEF(EltVT));
5738 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
5739 ShuffleMask[i] < (int)NumElts ? V1 : V2,
5740 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
5743 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, Ops);
5744 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
5747 if ((VT == MVT::v8i16 || VT == MVT::v16i8) && isReverseMask(ShuffleMask, VT))
5748 return LowerReverse_VECTOR_SHUFFLEv16i8_v8i16(Op, DAG);
5750 if (VT == MVT::v8i8) {
5751 SDValue NewOp = LowerVECTOR_SHUFFLEv8i8(Op, ShuffleMask, DAG);
5752 if (NewOp.getNode())
5759 static SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
5760 // INSERT_VECTOR_ELT is legal only for immediate indexes.
5761 SDValue Lane = Op.getOperand(2);
5762 if (!isa<ConstantSDNode>(Lane))
5768 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
5769 // EXTRACT_VECTOR_ELT is legal only for immediate indexes.
5770 SDValue Lane = Op.getOperand(1);
5771 if (!isa<ConstantSDNode>(Lane))
5774 SDValue Vec = Op.getOperand(0);
5775 if (Op.getValueType() == MVT::i32 &&
5776 Vec.getValueType().getVectorElementType().getSizeInBits() < 32) {
5778 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
5784 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
5785 // The only time a CONCAT_VECTORS operation can have legal types is when
5786 // two 64-bit vectors are concatenated to a 128-bit vector.
5787 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
5788 "unexpected CONCAT_VECTORS");
5790 SDValue Val = DAG.getUNDEF(MVT::v2f64);
5791 SDValue Op0 = Op.getOperand(0);
5792 SDValue Op1 = Op.getOperand(1);
5793 if (Op0.getOpcode() != ISD::UNDEF)
5794 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
5795 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op0),
5796 DAG.getIntPtrConstant(0));
5797 if (Op1.getOpcode() != ISD::UNDEF)
5798 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
5799 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op1),
5800 DAG.getIntPtrConstant(1));
5801 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Val);
5804 /// isExtendedBUILD_VECTOR - Check if N is a constant BUILD_VECTOR where each
5805 /// element has been zero/sign-extended, depending on the isSigned parameter,
5806 /// from an integer type half its size.
5807 static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG,
5809 // A v2i64 BUILD_VECTOR will have been legalized to a BITCAST from v4i32.
5810 EVT VT = N->getValueType(0);
5811 if (VT == MVT::v2i64 && N->getOpcode() == ISD::BITCAST) {
5812 SDNode *BVN = N->getOperand(0).getNode();
5813 if (BVN->getValueType(0) != MVT::v4i32 ||
5814 BVN->getOpcode() != ISD::BUILD_VECTOR)
5816 unsigned LoElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
5817 unsigned HiElt = 1 - LoElt;
5818 ConstantSDNode *Lo0 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt));
5819 ConstantSDNode *Hi0 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt));
5820 ConstantSDNode *Lo1 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt+2));
5821 ConstantSDNode *Hi1 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt+2));
5822 if (!Lo0 || !Hi0 || !Lo1 || !Hi1)
5825 if (Hi0->getSExtValue() == Lo0->getSExtValue() >> 32 &&
5826 Hi1->getSExtValue() == Lo1->getSExtValue() >> 32)
5829 if (Hi0->isNullValue() && Hi1->isNullValue())
5835 if (N->getOpcode() != ISD::BUILD_VECTOR)
5838 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
5839 SDNode *Elt = N->getOperand(i).getNode();
5840 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
5841 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5842 unsigned HalfSize = EltSize / 2;
5844 if (!isIntN(HalfSize, C->getSExtValue()))
5847 if (!isUIntN(HalfSize, C->getZExtValue()))
5858 /// isSignExtended - Check if a node is a vector value that is sign-extended
5859 /// or a constant BUILD_VECTOR with sign-extended elements.
5860 static bool isSignExtended(SDNode *N, SelectionDAG &DAG) {
5861 if (N->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N))
5863 if (isExtendedBUILD_VECTOR(N, DAG, true))
5868 /// isZeroExtended - Check if a node is a vector value that is zero-extended
5869 /// or a constant BUILD_VECTOR with zero-extended elements.
5870 static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) {
5871 if (N->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N))
5873 if (isExtendedBUILD_VECTOR(N, DAG, false))
5878 static EVT getExtensionTo64Bits(const EVT &OrigVT) {
5879 if (OrigVT.getSizeInBits() >= 64)
5882 assert(OrigVT.isSimple() && "Expecting a simple value type");
5884 MVT::SimpleValueType OrigSimpleTy = OrigVT.getSimpleVT().SimpleTy;
5885 switch (OrigSimpleTy) {
5886 default: llvm_unreachable("Unexpected Vector Type");
5895 /// AddRequiredExtensionForVMULL - Add a sign/zero extension to extend the total
5896 /// value size to 64 bits. We need a 64-bit D register as an operand to VMULL.
5897 /// We insert the required extension here to get the vector to fill a D register.
5898 static SDValue AddRequiredExtensionForVMULL(SDValue N, SelectionDAG &DAG,
5901 unsigned ExtOpcode) {
5902 // The vector originally had a size of OrigTy. It was then extended to ExtTy.
5903 // We expect the ExtTy to be 128-bits total. If the OrigTy is less than
5904 // 64-bits we need to insert a new extension so that it will be 64-bits.
5905 assert(ExtTy.is128BitVector() && "Unexpected extension size");
5906 if (OrigTy.getSizeInBits() >= 64)
5909 // Must extend size to at least 64 bits to be used as an operand for VMULL.
5910 EVT NewVT = getExtensionTo64Bits(OrigTy);
5912 return DAG.getNode(ExtOpcode, SDLoc(N), NewVT, N);
5915 /// SkipLoadExtensionForVMULL - return a load of the original vector size that
5916 /// does not do any sign/zero extension. If the original vector is less
5917 /// than 64 bits, an appropriate extension will be added after the load to
5918 /// reach a total size of 64 bits. We have to add the extension separately
5919 /// because ARM does not have a sign/zero extending load for vectors.
5920 static SDValue SkipLoadExtensionForVMULL(LoadSDNode *LD, SelectionDAG& DAG) {
5921 EVT ExtendedTy = getExtensionTo64Bits(LD->getMemoryVT());
5923 // The load already has the right type.
5924 if (ExtendedTy == LD->getMemoryVT())
5925 return DAG.getLoad(LD->getMemoryVT(), SDLoc(LD), LD->getChain(),
5926 LD->getBasePtr(), LD->getPointerInfo(), LD->isVolatile(),
5927 LD->isNonTemporal(), LD->isInvariant(),
5928 LD->getAlignment());
5930 // We need to create a zextload/sextload. We cannot just create a load
5931 // followed by a zext/zext node because LowerMUL is also run during normal
5932 // operation legalization where we can't create illegal types.
5933 return DAG.getExtLoad(LD->getExtensionType(), SDLoc(LD), ExtendedTy,
5934 LD->getChain(), LD->getBasePtr(), LD->getPointerInfo(),
5935 LD->getMemoryVT(), LD->isVolatile(), LD->isInvariant(),
5936 LD->isNonTemporal(), LD->getAlignment());
5939 /// SkipExtensionForVMULL - For a node that is a SIGN_EXTEND, ZERO_EXTEND,
5940 /// extending load, or BUILD_VECTOR with extended elements, return the
5941 /// unextended value. The unextended vector should be 64 bits so that it can
5942 /// be used as an operand to a VMULL instruction. If the original vector size
5943 /// before extension is less than 64 bits we add a an extension to resize
5944 /// the vector to 64 bits.
5945 static SDValue SkipExtensionForVMULL(SDNode *N, SelectionDAG &DAG) {
5946 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
5947 return AddRequiredExtensionForVMULL(N->getOperand(0), DAG,
5948 N->getOperand(0)->getValueType(0),
5952 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
5953 return SkipLoadExtensionForVMULL(LD, DAG);
5955 // Otherwise, the value must be a BUILD_VECTOR. For v2i64, it will
5956 // have been legalized as a BITCAST from v4i32.
5957 if (N->getOpcode() == ISD::BITCAST) {
5958 SDNode *BVN = N->getOperand(0).getNode();
5959 assert(BVN->getOpcode() == ISD::BUILD_VECTOR &&
5960 BVN->getValueType(0) == MVT::v4i32 && "expected v4i32 BUILD_VECTOR");
5961 unsigned LowElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
5962 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), MVT::v2i32,
5963 BVN->getOperand(LowElt), BVN->getOperand(LowElt+2));
5965 // Construct a new BUILD_VECTOR with elements truncated to half the size.
5966 assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR");
5967 EVT VT = N->getValueType(0);
5968 unsigned EltSize = VT.getVectorElementType().getSizeInBits() / 2;
5969 unsigned NumElts = VT.getVectorNumElements();
5970 MVT TruncVT = MVT::getIntegerVT(EltSize);
5971 SmallVector<SDValue, 8> Ops;
5972 for (unsigned i = 0; i != NumElts; ++i) {
5973 ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i));
5974 const APInt &CInt = C->getAPIntValue();
5975 // Element types smaller than 32 bits are not legal, so use i32 elements.
5976 // The values are implicitly truncated so sext vs. zext doesn't matter.
5977 Ops.push_back(DAG.getConstant(CInt.zextOrTrunc(32), MVT::i32));
5979 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N),
5980 MVT::getVectorVT(TruncVT, NumElts), Ops);
5983 static bool isAddSubSExt(SDNode *N, SelectionDAG &DAG) {
5984 unsigned Opcode = N->getOpcode();
5985 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
5986 SDNode *N0 = N->getOperand(0).getNode();
5987 SDNode *N1 = N->getOperand(1).getNode();
5988 return N0->hasOneUse() && N1->hasOneUse() &&
5989 isSignExtended(N0, DAG) && isSignExtended(N1, DAG);
5994 static bool isAddSubZExt(SDNode *N, SelectionDAG &DAG) {
5995 unsigned Opcode = N->getOpcode();
5996 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
5997 SDNode *N0 = N->getOperand(0).getNode();
5998 SDNode *N1 = N->getOperand(1).getNode();
5999 return N0->hasOneUse() && N1->hasOneUse() &&
6000 isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG);
6005 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
6006 // Multiplications are only custom-lowered for 128-bit vectors so that
6007 // VMULL can be detected. Otherwise v2i64 multiplications are not legal.
6008 EVT VT = Op.getValueType();
6009 assert(VT.is128BitVector() && VT.isInteger() &&
6010 "unexpected type for custom-lowering ISD::MUL");
6011 SDNode *N0 = Op.getOperand(0).getNode();
6012 SDNode *N1 = Op.getOperand(1).getNode();
6013 unsigned NewOpc = 0;
6015 bool isN0SExt = isSignExtended(N0, DAG);
6016 bool isN1SExt = isSignExtended(N1, DAG);
6017 if (isN0SExt && isN1SExt)
6018 NewOpc = ARMISD::VMULLs;
6020 bool isN0ZExt = isZeroExtended(N0, DAG);
6021 bool isN1ZExt = isZeroExtended(N1, DAG);
6022 if (isN0ZExt && isN1ZExt)
6023 NewOpc = ARMISD::VMULLu;
6024 else if (isN1SExt || isN1ZExt) {
6025 // Look for (s/zext A + s/zext B) * (s/zext C). We want to turn these
6026 // into (s/zext A * s/zext C) + (s/zext B * s/zext C)
6027 if (isN1SExt && isAddSubSExt(N0, DAG)) {
6028 NewOpc = ARMISD::VMULLs;
6030 } else if (isN1ZExt && isAddSubZExt(N0, DAG)) {
6031 NewOpc = ARMISD::VMULLu;
6033 } else if (isN0ZExt && isAddSubZExt(N1, DAG)) {
6035 NewOpc = ARMISD::VMULLu;
6041 if (VT == MVT::v2i64)
6042 // Fall through to expand this. It is not legal.
6045 // Other vector multiplications are legal.
6050 // Legalize to a VMULL instruction.
6053 SDValue Op1 = SkipExtensionForVMULL(N1, DAG);
6055 Op0 = SkipExtensionForVMULL(N0, DAG);
6056 assert(Op0.getValueType().is64BitVector() &&
6057 Op1.getValueType().is64BitVector() &&
6058 "unexpected types for extended operands to VMULL");
6059 return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
6062 // Optimizing (zext A + zext B) * C, to (VMULL A, C) + (VMULL B, C) during
6063 // isel lowering to take advantage of no-stall back to back vmul + vmla.
6070 SDValue N00 = SkipExtensionForVMULL(N0->getOperand(0).getNode(), DAG);
6071 SDValue N01 = SkipExtensionForVMULL(N0->getOperand(1).getNode(), DAG);
6072 EVT Op1VT = Op1.getValueType();
6073 return DAG.getNode(N0->getOpcode(), DL, VT,
6074 DAG.getNode(NewOpc, DL, VT,
6075 DAG.getNode(ISD::BITCAST, DL, Op1VT, N00), Op1),
6076 DAG.getNode(NewOpc, DL, VT,
6077 DAG.getNode(ISD::BITCAST, DL, Op1VT, N01), Op1));
6081 LowerSDIV_v4i8(SDValue X, SDValue Y, SDLoc dl, SelectionDAG &DAG) {
6083 // float4 xf = vcvt_f32_s32(vmovl_s16(a.lo));
6084 // float4 yf = vcvt_f32_s32(vmovl_s16(b.lo));
6085 X = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, X);
6086 Y = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, Y);
6087 X = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, X);
6088 Y = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, Y);
6089 // Get reciprocal estimate.
6090 // float4 recip = vrecpeq_f32(yf);
6091 Y = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6092 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), Y);
6093 // Because char has a smaller range than uchar, we can actually get away
6094 // without any newton steps. This requires that we use a weird bias
6095 // of 0xb000, however (again, this has been exhaustively tested).
6096 // float4 result = as_float4(as_int4(xf*recip) + 0xb000);
6097 X = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, X, Y);
6098 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, X);
6099 Y = DAG.getConstant(0xb000, MVT::i32);
6100 Y = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Y, Y, Y, Y);
6101 X = DAG.getNode(ISD::ADD, dl, MVT::v4i32, X, Y);
6102 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, X);
6103 // Convert back to short.
6104 X = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, X);
6105 X = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, X);
6110 LowerSDIV_v4i16(SDValue N0, SDValue N1, SDLoc dl, SelectionDAG &DAG) {
6112 // Convert to float.
6113 // float4 yf = vcvt_f32_s32(vmovl_s16(y));
6114 // float4 xf = vcvt_f32_s32(vmovl_s16(x));
6115 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N0);
6116 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N1);
6117 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
6118 N1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
6120 // Use reciprocal estimate and one refinement step.
6121 // float4 recip = vrecpeq_f32(yf);
6122 // recip *= vrecpsq_f32(yf, recip);
6123 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6124 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), N1);
6125 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6126 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
6128 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
6129 // Because short has a smaller range than ushort, we can actually get away
6130 // with only a single newton step. This requires that we use a weird bias
6131 // of 89, however (again, this has been exhaustively tested).
6132 // float4 result = as_float4(as_int4(xf*recip) + 0x89);
6133 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
6134 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
6135 N1 = DAG.getConstant(0x89, MVT::i32);
6136 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
6137 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
6138 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
6139 // Convert back to integer and return.
6140 // return vmovn_s32(vcvt_s32_f32(result));
6141 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
6142 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
6146 static SDValue LowerSDIV(SDValue Op, SelectionDAG &DAG) {
6147 EVT VT = Op.getValueType();
6148 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
6149 "unexpected type for custom-lowering ISD::SDIV");
6152 SDValue N0 = Op.getOperand(0);
6153 SDValue N1 = Op.getOperand(1);
6156 if (VT == MVT::v8i8) {
6157 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N0);
6158 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N1);
6160 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
6161 DAG.getIntPtrConstant(4));
6162 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
6163 DAG.getIntPtrConstant(4));
6164 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
6165 DAG.getIntPtrConstant(0));
6166 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
6167 DAG.getIntPtrConstant(0));
6169 N0 = LowerSDIV_v4i8(N0, N1, dl, DAG); // v4i16
6170 N2 = LowerSDIV_v4i8(N2, N3, dl, DAG); // v4i16
6172 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
6173 N0 = LowerCONCAT_VECTORS(N0, DAG);
6175 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v8i8, N0);
6178 return LowerSDIV_v4i16(N0, N1, dl, DAG);
6181 static SDValue LowerUDIV(SDValue Op, SelectionDAG &DAG) {
6182 EVT VT = Op.getValueType();
6183 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
6184 "unexpected type for custom-lowering ISD::UDIV");
6187 SDValue N0 = Op.getOperand(0);
6188 SDValue N1 = Op.getOperand(1);
6191 if (VT == MVT::v8i8) {
6192 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N0);
6193 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N1);
6195 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
6196 DAG.getIntPtrConstant(4));
6197 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
6198 DAG.getIntPtrConstant(4));
6199 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
6200 DAG.getIntPtrConstant(0));
6201 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
6202 DAG.getIntPtrConstant(0));
6204 N0 = LowerSDIV_v4i16(N0, N1, dl, DAG); // v4i16
6205 N2 = LowerSDIV_v4i16(N2, N3, dl, DAG); // v4i16
6207 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
6208 N0 = LowerCONCAT_VECTORS(N0, DAG);
6210 N0 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v8i8,
6211 DAG.getConstant(Intrinsic::arm_neon_vqmovnsu, MVT::i32),
6216 // v4i16 sdiv ... Convert to float.
6217 // float4 yf = vcvt_f32_s32(vmovl_u16(y));
6218 // float4 xf = vcvt_f32_s32(vmovl_u16(x));
6219 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N0);
6220 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N1);
6221 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
6222 SDValue BN1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
6224 // Use reciprocal estimate and two refinement steps.
6225 // float4 recip = vrecpeq_f32(yf);
6226 // recip *= vrecpsq_f32(yf, recip);
6227 // recip *= vrecpsq_f32(yf, recip);
6228 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6229 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), BN1);
6230 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6231 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
6233 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
6234 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6235 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
6237 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
6238 // Simply multiplying by the reciprocal estimate can leave us a few ulps
6239 // too low, so we add 2 ulps (exhaustive testing shows that this is enough,
6240 // and that it will never cause us to return an answer too large).
6241 // float4 result = as_float4(as_int4(xf*recip) + 2);
6242 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
6243 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
6244 N1 = DAG.getConstant(2, MVT::i32);
6245 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
6246 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
6247 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
6248 // Convert back to integer and return.
6249 // return vmovn_u32(vcvt_s32_f32(result));
6250 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
6251 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
6255 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
6256 EVT VT = Op.getNode()->getValueType(0);
6257 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
6260 bool ExtraOp = false;
6261 switch (Op.getOpcode()) {
6262 default: llvm_unreachable("Invalid code");
6263 case ISD::ADDC: Opc = ARMISD::ADDC; break;
6264 case ISD::ADDE: Opc = ARMISD::ADDE; ExtraOp = true; break;
6265 case ISD::SUBC: Opc = ARMISD::SUBC; break;
6266 case ISD::SUBE: Opc = ARMISD::SUBE; ExtraOp = true; break;
6270 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
6272 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
6273 Op.getOperand(1), Op.getOperand(2));
6276 SDValue ARMTargetLowering::LowerFSINCOS(SDValue Op, SelectionDAG &DAG) const {
6277 assert(Subtarget->isTargetDarwin());
6279 // For iOS, we want to call an alternative entry point: __sincos_stret,
6280 // return values are passed via sret.
6282 SDValue Arg = Op.getOperand(0);
6283 EVT ArgVT = Arg.getValueType();
6284 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
6286 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
6287 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6289 // Pair of floats / doubles used to pass the result.
6290 StructType *RetTy = StructType::get(ArgTy, ArgTy, nullptr);
6292 // Create stack object for sret.
6293 const uint64_t ByteSize = TLI.getDataLayout()->getTypeAllocSize(RetTy);
6294 const unsigned StackAlign = TLI.getDataLayout()->getPrefTypeAlignment(RetTy);
6295 int FrameIdx = FrameInfo->CreateStackObject(ByteSize, StackAlign, false);
6296 SDValue SRet = DAG.getFrameIndex(FrameIdx, TLI.getPointerTy());
6302 Entry.Ty = RetTy->getPointerTo();
6303 Entry.isSExt = false;
6304 Entry.isZExt = false;
6305 Entry.isSRet = true;
6306 Args.push_back(Entry);
6310 Entry.isSExt = false;
6311 Entry.isZExt = false;
6312 Args.push_back(Entry);
6314 const char *LibcallName = (ArgVT == MVT::f64)
6315 ? "__sincos_stret" : "__sincosf_stret";
6316 SDValue Callee = DAG.getExternalSymbol(LibcallName, getPointerTy());
6318 TargetLowering::CallLoweringInfo CLI(DAG);
6319 CLI.setDebugLoc(dl).setChain(DAG.getEntryNode())
6320 .setCallee(CallingConv::C, Type::getVoidTy(*DAG.getContext()), Callee,
6322 .setDiscardResult();
6324 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
6326 SDValue LoadSin = DAG.getLoad(ArgVT, dl, CallResult.second, SRet,
6327 MachinePointerInfo(), false, false, false, 0);
6329 // Address of cos field.
6330 SDValue Add = DAG.getNode(ISD::ADD, dl, getPointerTy(), SRet,
6331 DAG.getIntPtrConstant(ArgVT.getStoreSize()));
6332 SDValue LoadCos = DAG.getLoad(ArgVT, dl, LoadSin.getValue(1), Add,
6333 MachinePointerInfo(), false, false, false, 0);
6335 SDVTList Tys = DAG.getVTList(ArgVT, ArgVT);
6336 return DAG.getNode(ISD::MERGE_VALUES, dl, Tys,
6337 LoadSin.getValue(0), LoadCos.getValue(0));
6340 static SDValue LowerAtomicLoadStore(SDValue Op, SelectionDAG &DAG) {
6341 // Monotonic load/store is legal for all targets
6342 if (cast<AtomicSDNode>(Op)->getOrdering() <= Monotonic)
6345 // Acquire/Release load/store is not legal for targets without a
6346 // dmb or equivalent available.
6350 static void ReplaceREADCYCLECOUNTER(SDNode *N,
6351 SmallVectorImpl<SDValue> &Results,
6353 const ARMSubtarget *Subtarget) {
6355 SDValue Cycles32, OutChain;
6357 if (Subtarget->hasPerfMon()) {
6358 // Under Power Management extensions, the cycle-count is:
6359 // mrc p15, #0, <Rt>, c9, c13, #0
6360 SDValue Ops[] = { N->getOperand(0), // Chain
6361 DAG.getConstant(Intrinsic::arm_mrc, MVT::i32),
6362 DAG.getConstant(15, MVT::i32),
6363 DAG.getConstant(0, MVT::i32),
6364 DAG.getConstant(9, MVT::i32),
6365 DAG.getConstant(13, MVT::i32),
6366 DAG.getConstant(0, MVT::i32)
6369 Cycles32 = DAG.getNode(ISD::INTRINSIC_W_CHAIN, DL,
6370 DAG.getVTList(MVT::i32, MVT::Other), Ops);
6371 OutChain = Cycles32.getValue(1);
6373 // Intrinsic is defined to return 0 on unsupported platforms. Technically
6374 // there are older ARM CPUs that have implementation-specific ways of
6375 // obtaining this information (FIXME!).
6376 Cycles32 = DAG.getConstant(0, MVT::i32);
6377 OutChain = DAG.getEntryNode();
6381 SDValue Cycles64 = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64,
6382 Cycles32, DAG.getConstant(0, MVT::i32));
6383 Results.push_back(Cycles64);
6384 Results.push_back(OutChain);
6387 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
6388 switch (Op.getOpcode()) {
6389 default: llvm_unreachable("Don't know how to custom lower this!");
6390 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
6391 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
6392 case ISD::GlobalAddress:
6393 switch (Subtarget->getTargetTriple().getObjectFormat()) {
6394 default: llvm_unreachable("unknown object format");
6396 return LowerGlobalAddressWindows(Op, DAG);
6398 return LowerGlobalAddressELF(Op, DAG);
6400 return LowerGlobalAddressDarwin(Op, DAG);
6402 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
6403 case ISD::SELECT: return LowerSELECT(Op, DAG);
6404 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
6405 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
6406 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
6407 case ISD::VASTART: return LowerVASTART(Op, DAG);
6408 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, DAG, Subtarget);
6409 case ISD::PREFETCH: return LowerPREFETCH(Op, DAG, Subtarget);
6410 case ISD::SINT_TO_FP:
6411 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
6412 case ISD::FP_TO_SINT:
6413 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
6414 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
6415 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
6416 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
6417 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
6418 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
6419 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
6420 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
6422 case ISD::BITCAST: return ExpandBITCAST(Op.getNode(), DAG);
6425 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
6426 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
6427 case ISD::SRL_PARTS:
6428 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
6429 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
6430 case ISD::CTPOP: return LowerCTPOP(Op.getNode(), DAG, Subtarget);
6431 case ISD::SETCC: return LowerVSETCC(Op, DAG);
6432 case ISD::ConstantFP: return LowerConstantFP(Op, DAG, Subtarget);
6433 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG, Subtarget);
6434 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
6435 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
6436 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
6437 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
6438 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
6439 case ISD::MUL: return LowerMUL(Op, DAG);
6440 case ISD::SDIV: return LowerSDIV(Op, DAG);
6441 case ISD::UDIV: return LowerUDIV(Op, DAG);
6445 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
6450 return LowerXALUO(Op, DAG);
6451 case ISD::ATOMIC_LOAD:
6452 case ISD::ATOMIC_STORE: return LowerAtomicLoadStore(Op, DAG);
6453 case ISD::FSINCOS: return LowerFSINCOS(Op, DAG);
6455 case ISD::UDIVREM: return LowerDivRem(Op, DAG);
6456 case ISD::DYNAMIC_STACKALLOC:
6457 if (Subtarget->getTargetTriple().isWindowsItaniumEnvironment())
6458 return LowerDYNAMIC_STACKALLOC(Op, DAG);
6459 llvm_unreachable("Don't know how to custom lower this!");
6460 case ISD::FP_ROUND: return LowerFP_ROUND(Op, DAG);
6461 case ISD::FP_EXTEND: return LowerFP_EXTEND(Op, DAG);
6465 /// ReplaceNodeResults - Replace the results of node with an illegal result
6466 /// type with new values built out of custom code.
6467 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
6468 SmallVectorImpl<SDValue>&Results,
6469 SelectionDAG &DAG) const {
6471 switch (N->getOpcode()) {
6473 llvm_unreachable("Don't know how to custom expand this!");
6475 Res = ExpandBITCAST(N, DAG);
6479 Res = Expand64BitShift(N, DAG, Subtarget);
6481 case ISD::READCYCLECOUNTER:
6482 ReplaceREADCYCLECOUNTER(N, Results, DAG, Subtarget);
6486 Results.push_back(Res);
6489 //===----------------------------------------------------------------------===//
6490 // ARM Scheduler Hooks
6491 //===----------------------------------------------------------------------===//
6493 /// SetupEntryBlockForSjLj - Insert code into the entry block that creates and
6494 /// registers the function context.
6495 void ARMTargetLowering::
6496 SetupEntryBlockForSjLj(MachineInstr *MI, MachineBasicBlock *MBB,
6497 MachineBasicBlock *DispatchBB, int FI) const {
6498 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
6499 DebugLoc dl = MI->getDebugLoc();
6500 MachineFunction *MF = MBB->getParent();
6501 MachineRegisterInfo *MRI = &MF->getRegInfo();
6502 MachineConstantPool *MCP = MF->getConstantPool();
6503 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
6504 const Function *F = MF->getFunction();
6506 bool isThumb = Subtarget->isThumb();
6507 bool isThumb2 = Subtarget->isThumb2();
6509 unsigned PCLabelId = AFI->createPICLabelUId();
6510 unsigned PCAdj = (isThumb || isThumb2) ? 4 : 8;
6511 ARMConstantPoolValue *CPV =
6512 ARMConstantPoolMBB::Create(F->getContext(), DispatchBB, PCLabelId, PCAdj);
6513 unsigned CPI = MCP->getConstantPoolIndex(CPV, 4);
6515 const TargetRegisterClass *TRC = isThumb ? &ARM::tGPRRegClass
6516 : &ARM::GPRRegClass;
6518 // Grab constant pool and fixed stack memory operands.
6519 MachineMemOperand *CPMMO =
6520 MF->getMachineMemOperand(MachinePointerInfo::getConstantPool(),
6521 MachineMemOperand::MOLoad, 4, 4);
6523 MachineMemOperand *FIMMOSt =
6524 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
6525 MachineMemOperand::MOStore, 4, 4);
6527 // Load the address of the dispatch MBB into the jump buffer.
6529 // Incoming value: jbuf
6530 // ldr.n r5, LCPI1_1
6533 // str r5, [$jbuf, #+4] ; &jbuf[1]
6534 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6535 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2LDRpci), NewVReg1)
6536 .addConstantPoolIndex(CPI)
6537 .addMemOperand(CPMMO));
6538 // Set the low bit because of thumb mode.
6539 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6541 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2ORRri), NewVReg2)
6542 .addReg(NewVReg1, RegState::Kill)
6544 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6545 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg3)
6546 .addReg(NewVReg2, RegState::Kill)
6548 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2STRi12))
6549 .addReg(NewVReg3, RegState::Kill)
6551 .addImm(36) // &jbuf[1] :: pc
6552 .addMemOperand(FIMMOSt));
6553 } else if (isThumb) {
6554 // Incoming value: jbuf
6555 // ldr.n r1, LCPI1_4
6559 // add r2, $jbuf, #+4 ; &jbuf[1]
6561 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6562 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tLDRpci), NewVReg1)
6563 .addConstantPoolIndex(CPI)
6564 .addMemOperand(CPMMO));
6565 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6566 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg2)
6567 .addReg(NewVReg1, RegState::Kill)
6569 // Set the low bit because of thumb mode.
6570 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6571 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tMOVi8), NewVReg3)
6572 .addReg(ARM::CPSR, RegState::Define)
6574 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6575 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tORR), NewVReg4)
6576 .addReg(ARM::CPSR, RegState::Define)
6577 .addReg(NewVReg2, RegState::Kill)
6578 .addReg(NewVReg3, RegState::Kill));
6579 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6580 BuildMI(*MBB, MI, dl, TII->get(ARM::tADDframe), NewVReg5)
6582 .addImm(36); // &jbuf[1] :: pc
6583 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tSTRi))
6584 .addReg(NewVReg4, RegState::Kill)
6585 .addReg(NewVReg5, RegState::Kill)
6587 .addMemOperand(FIMMOSt));
6589 // Incoming value: jbuf
6592 // str r1, [$jbuf, #+4] ; &jbuf[1]
6593 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6594 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::LDRi12), NewVReg1)
6595 .addConstantPoolIndex(CPI)
6597 .addMemOperand(CPMMO));
6598 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6599 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::PICADD), NewVReg2)
6600 .addReg(NewVReg1, RegState::Kill)
6601 .addImm(PCLabelId));
6602 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::STRi12))
6603 .addReg(NewVReg2, RegState::Kill)
6605 .addImm(36) // &jbuf[1] :: pc
6606 .addMemOperand(FIMMOSt));
6610 MachineBasicBlock *ARMTargetLowering::
6611 EmitSjLjDispatchBlock(MachineInstr *MI, MachineBasicBlock *MBB) const {
6612 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
6613 DebugLoc dl = MI->getDebugLoc();
6614 MachineFunction *MF = MBB->getParent();
6615 MachineRegisterInfo *MRI = &MF->getRegInfo();
6616 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
6617 MachineFrameInfo *MFI = MF->getFrameInfo();
6618 int FI = MFI->getFunctionContextIndex();
6620 const TargetRegisterClass *TRC = Subtarget->isThumb() ? &ARM::tGPRRegClass
6621 : &ARM::GPRnopcRegClass;
6623 // Get a mapping of the call site numbers to all of the landing pads they're
6625 DenseMap<unsigned, SmallVector<MachineBasicBlock*, 2> > CallSiteNumToLPad;
6626 unsigned MaxCSNum = 0;
6627 MachineModuleInfo &MMI = MF->getMMI();
6628 for (MachineFunction::iterator BB = MF->begin(), E = MF->end(); BB != E;
6630 if (!BB->isLandingPad()) continue;
6632 // FIXME: We should assert that the EH_LABEL is the first MI in the landing
6634 for (MachineBasicBlock::iterator
6635 II = BB->begin(), IE = BB->end(); II != IE; ++II) {
6636 if (!II->isEHLabel()) continue;
6638 MCSymbol *Sym = II->getOperand(0).getMCSymbol();
6639 if (!MMI.hasCallSiteLandingPad(Sym)) continue;
6641 SmallVectorImpl<unsigned> &CallSiteIdxs = MMI.getCallSiteLandingPad(Sym);
6642 for (SmallVectorImpl<unsigned>::iterator
6643 CSI = CallSiteIdxs.begin(), CSE = CallSiteIdxs.end();
6644 CSI != CSE; ++CSI) {
6645 CallSiteNumToLPad[*CSI].push_back(BB);
6646 MaxCSNum = std::max(MaxCSNum, *CSI);
6652 // Get an ordered list of the machine basic blocks for the jump table.
6653 std::vector<MachineBasicBlock*> LPadList;
6654 SmallPtrSet<MachineBasicBlock*, 64> InvokeBBs;
6655 LPadList.reserve(CallSiteNumToLPad.size());
6656 for (unsigned I = 1; I <= MaxCSNum; ++I) {
6657 SmallVectorImpl<MachineBasicBlock*> &MBBList = CallSiteNumToLPad[I];
6658 for (SmallVectorImpl<MachineBasicBlock*>::iterator
6659 II = MBBList.begin(), IE = MBBList.end(); II != IE; ++II) {
6660 LPadList.push_back(*II);
6661 InvokeBBs.insert((*II)->pred_begin(), (*II)->pred_end());
6665 assert(!LPadList.empty() &&
6666 "No landing pad destinations for the dispatch jump table!");
6668 // Create the jump table and associated information.
6669 MachineJumpTableInfo *JTI =
6670 MF->getOrCreateJumpTableInfo(MachineJumpTableInfo::EK_Inline);
6671 unsigned MJTI = JTI->createJumpTableIndex(LPadList);
6672 unsigned UId = AFI->createJumpTableUId();
6673 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
6675 // Create the MBBs for the dispatch code.
6677 // Shove the dispatch's address into the return slot in the function context.
6678 MachineBasicBlock *DispatchBB = MF->CreateMachineBasicBlock();
6679 DispatchBB->setIsLandingPad();
6681 MachineBasicBlock *TrapBB = MF->CreateMachineBasicBlock();
6682 unsigned trap_opcode;
6683 if (Subtarget->isThumb())
6684 trap_opcode = ARM::tTRAP;
6686 trap_opcode = Subtarget->useNaClTrap() ? ARM::TRAPNaCl : ARM::TRAP;
6688 BuildMI(TrapBB, dl, TII->get(trap_opcode));
6689 DispatchBB->addSuccessor(TrapBB);
6691 MachineBasicBlock *DispContBB = MF->CreateMachineBasicBlock();
6692 DispatchBB->addSuccessor(DispContBB);
6695 MF->insert(MF->end(), DispatchBB);
6696 MF->insert(MF->end(), DispContBB);
6697 MF->insert(MF->end(), TrapBB);
6699 // Insert code into the entry block that creates and registers the function
6701 SetupEntryBlockForSjLj(MI, MBB, DispatchBB, FI);
6703 MachineMemOperand *FIMMOLd =
6704 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
6705 MachineMemOperand::MOLoad |
6706 MachineMemOperand::MOVolatile, 4, 4);
6708 MachineInstrBuilder MIB;
6709 MIB = BuildMI(DispatchBB, dl, TII->get(ARM::Int_eh_sjlj_dispatchsetup));
6711 const ARMBaseInstrInfo *AII = static_cast<const ARMBaseInstrInfo*>(TII);
6712 const ARMBaseRegisterInfo &RI = AII->getRegisterInfo();
6714 // Add a register mask with no preserved registers. This results in all
6715 // registers being marked as clobbered.
6716 MIB.addRegMask(RI.getNoPreservedMask());
6718 unsigned NumLPads = LPadList.size();
6719 if (Subtarget->isThumb2()) {
6720 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6721 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2LDRi12), NewVReg1)
6724 .addMemOperand(FIMMOLd));
6726 if (NumLPads < 256) {
6727 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPri))
6729 .addImm(LPadList.size()));
6731 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6732 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVi16), VReg1)
6733 .addImm(NumLPads & 0xFFFF));
6735 unsigned VReg2 = VReg1;
6736 if ((NumLPads & 0xFFFF0000) != 0) {
6737 VReg2 = MRI->createVirtualRegister(TRC);
6738 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVTi16), VReg2)
6740 .addImm(NumLPads >> 16));
6743 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPrr))
6748 BuildMI(DispatchBB, dl, TII->get(ARM::t2Bcc))
6753 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6754 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::t2LEApcrelJT),NewVReg3)
6755 .addJumpTableIndex(MJTI)
6758 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6761 BuildMI(DispContBB, dl, TII->get(ARM::t2ADDrs), NewVReg4)
6762 .addReg(NewVReg3, RegState::Kill)
6764 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
6766 BuildMI(DispContBB, dl, TII->get(ARM::t2BR_JT))
6767 .addReg(NewVReg4, RegState::Kill)
6769 .addJumpTableIndex(MJTI)
6771 } else if (Subtarget->isThumb()) {
6772 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6773 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRspi), NewVReg1)
6776 .addMemOperand(FIMMOLd));
6778 if (NumLPads < 256) {
6779 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPi8))
6783 MachineConstantPool *ConstantPool = MF->getConstantPool();
6784 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
6785 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
6787 // MachineConstantPool wants an explicit alignment.
6788 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
6790 Align = getDataLayout()->getTypeAllocSize(C->getType());
6791 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6793 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6794 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRpci))
6795 .addReg(VReg1, RegState::Define)
6796 .addConstantPoolIndex(Idx));
6797 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPr))
6802 BuildMI(DispatchBB, dl, TII->get(ARM::tBcc))
6807 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6808 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLSLri), NewVReg2)
6809 .addReg(ARM::CPSR, RegState::Define)
6813 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6814 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLEApcrelJT), NewVReg3)
6815 .addJumpTableIndex(MJTI)
6818 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6819 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg4)
6820 .addReg(ARM::CPSR, RegState::Define)
6821 .addReg(NewVReg2, RegState::Kill)
6824 MachineMemOperand *JTMMOLd =
6825 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
6826 MachineMemOperand::MOLoad, 4, 4);
6828 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6829 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLDRi), NewVReg5)
6830 .addReg(NewVReg4, RegState::Kill)
6832 .addMemOperand(JTMMOLd));
6834 unsigned NewVReg6 = NewVReg5;
6835 if (RelocM == Reloc::PIC_) {
6836 NewVReg6 = MRI->createVirtualRegister(TRC);
6837 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg6)
6838 .addReg(ARM::CPSR, RegState::Define)
6839 .addReg(NewVReg5, RegState::Kill)
6843 BuildMI(DispContBB, dl, TII->get(ARM::tBR_JTr))
6844 .addReg(NewVReg6, RegState::Kill)
6845 .addJumpTableIndex(MJTI)
6848 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6849 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRi12), NewVReg1)
6852 .addMemOperand(FIMMOLd));
6854 if (NumLPads < 256) {
6855 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPri))
6858 } else if (Subtarget->hasV6T2Ops() && isUInt<16>(NumLPads)) {
6859 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6860 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVi16), VReg1)
6861 .addImm(NumLPads & 0xFFFF));
6863 unsigned VReg2 = VReg1;
6864 if ((NumLPads & 0xFFFF0000) != 0) {
6865 VReg2 = MRI->createVirtualRegister(TRC);
6866 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVTi16), VReg2)
6868 .addImm(NumLPads >> 16));
6871 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
6875 MachineConstantPool *ConstantPool = MF->getConstantPool();
6876 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
6877 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
6879 // MachineConstantPool wants an explicit alignment.
6880 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
6882 Align = getDataLayout()->getTypeAllocSize(C->getType());
6883 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6885 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6886 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRcp))
6887 .addReg(VReg1, RegState::Define)
6888 .addConstantPoolIndex(Idx)
6890 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
6892 .addReg(VReg1, RegState::Kill));
6895 BuildMI(DispatchBB, dl, TII->get(ARM::Bcc))
6900 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6902 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::MOVsi), NewVReg3)
6904 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
6905 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6906 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::LEApcrelJT), NewVReg4)
6907 .addJumpTableIndex(MJTI)
6910 MachineMemOperand *JTMMOLd =
6911 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
6912 MachineMemOperand::MOLoad, 4, 4);
6913 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6915 BuildMI(DispContBB, dl, TII->get(ARM::LDRrs), NewVReg5)
6916 .addReg(NewVReg3, RegState::Kill)
6919 .addMemOperand(JTMMOLd));
6921 if (RelocM == Reloc::PIC_) {
6922 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTadd))
6923 .addReg(NewVReg5, RegState::Kill)
6925 .addJumpTableIndex(MJTI)
6928 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTr))
6929 .addReg(NewVReg5, RegState::Kill)
6930 .addJumpTableIndex(MJTI)
6935 // Add the jump table entries as successors to the MBB.
6936 SmallPtrSet<MachineBasicBlock*, 8> SeenMBBs;
6937 for (std::vector<MachineBasicBlock*>::iterator
6938 I = LPadList.begin(), E = LPadList.end(); I != E; ++I) {
6939 MachineBasicBlock *CurMBB = *I;
6940 if (SeenMBBs.insert(CurMBB).second)
6941 DispContBB->addSuccessor(CurMBB);
6944 // N.B. the order the invoke BBs are processed in doesn't matter here.
6945 const MCPhysReg *SavedRegs = RI.getCalleeSavedRegs(MF);
6946 SmallVector<MachineBasicBlock*, 64> MBBLPads;
6947 for (MachineBasicBlock *BB : InvokeBBs) {
6949 // Remove the landing pad successor from the invoke block and replace it
6950 // with the new dispatch block.
6951 SmallVector<MachineBasicBlock*, 4> Successors(BB->succ_begin(),
6953 while (!Successors.empty()) {
6954 MachineBasicBlock *SMBB = Successors.pop_back_val();
6955 if (SMBB->isLandingPad()) {
6956 BB->removeSuccessor(SMBB);
6957 MBBLPads.push_back(SMBB);
6961 BB->addSuccessor(DispatchBB);
6963 // Find the invoke call and mark all of the callee-saved registers as
6964 // 'implicit defined' so that they're spilled. This prevents code from
6965 // moving instructions to before the EH block, where they will never be
6967 for (MachineBasicBlock::reverse_iterator
6968 II = BB->rbegin(), IE = BB->rend(); II != IE; ++II) {
6969 if (!II->isCall()) continue;
6971 DenseMap<unsigned, bool> DefRegs;
6972 for (MachineInstr::mop_iterator
6973 OI = II->operands_begin(), OE = II->operands_end();
6975 if (!OI->isReg()) continue;
6976 DefRegs[OI->getReg()] = true;
6979 MachineInstrBuilder MIB(*MF, &*II);
6981 for (unsigned i = 0; SavedRegs[i] != 0; ++i) {
6982 unsigned Reg = SavedRegs[i];
6983 if (Subtarget->isThumb2() &&
6984 !ARM::tGPRRegClass.contains(Reg) &&
6985 !ARM::hGPRRegClass.contains(Reg))
6987 if (Subtarget->isThumb1Only() && !ARM::tGPRRegClass.contains(Reg))
6989 if (!Subtarget->isThumb() && !ARM::GPRRegClass.contains(Reg))
6992 MIB.addReg(Reg, RegState::ImplicitDefine | RegState::Dead);
6999 // Mark all former landing pads as non-landing pads. The dispatch is the only
7001 for (SmallVectorImpl<MachineBasicBlock*>::iterator
7002 I = MBBLPads.begin(), E = MBBLPads.end(); I != E; ++I)
7003 (*I)->setIsLandingPad(false);
7005 // The instruction is gone now.
7006 MI->eraseFromParent();
7012 MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) {
7013 for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
7014 E = MBB->succ_end(); I != E; ++I)
7017 llvm_unreachable("Expecting a BB with two successors!");
7020 /// Return the load opcode for a given load size. If load size >= 8,
7021 /// neon opcode will be returned.
7022 static unsigned getLdOpcode(unsigned LdSize, bool IsThumb1, bool IsThumb2) {
7024 return LdSize == 16 ? ARM::VLD1q32wb_fixed
7025 : LdSize == 8 ? ARM::VLD1d32wb_fixed : 0;
7027 return LdSize == 4 ? ARM::tLDRi
7028 : LdSize == 2 ? ARM::tLDRHi
7029 : LdSize == 1 ? ARM::tLDRBi : 0;
7031 return LdSize == 4 ? ARM::t2LDR_POST
7032 : LdSize == 2 ? ARM::t2LDRH_POST
7033 : LdSize == 1 ? ARM::t2LDRB_POST : 0;
7034 return LdSize == 4 ? ARM::LDR_POST_IMM
7035 : LdSize == 2 ? ARM::LDRH_POST
7036 : LdSize == 1 ? ARM::LDRB_POST_IMM : 0;
7039 /// Return the store opcode for a given store size. If store size >= 8,
7040 /// neon opcode will be returned.
7041 static unsigned getStOpcode(unsigned StSize, bool IsThumb1, bool IsThumb2) {
7043 return StSize == 16 ? ARM::VST1q32wb_fixed
7044 : StSize == 8 ? ARM::VST1d32wb_fixed : 0;
7046 return StSize == 4 ? ARM::tSTRi
7047 : StSize == 2 ? ARM::tSTRHi
7048 : StSize == 1 ? ARM::tSTRBi : 0;
7050 return StSize == 4 ? ARM::t2STR_POST
7051 : StSize == 2 ? ARM::t2STRH_POST
7052 : StSize == 1 ? ARM::t2STRB_POST : 0;
7053 return StSize == 4 ? ARM::STR_POST_IMM
7054 : StSize == 2 ? ARM::STRH_POST
7055 : StSize == 1 ? ARM::STRB_POST_IMM : 0;
7058 /// Emit a post-increment load operation with given size. The instructions
7059 /// will be added to BB at Pos.
7060 static void emitPostLd(MachineBasicBlock *BB, MachineInstr *Pos,
7061 const TargetInstrInfo *TII, DebugLoc dl,
7062 unsigned LdSize, unsigned Data, unsigned AddrIn,
7063 unsigned AddrOut, bool IsThumb1, bool IsThumb2) {
7064 unsigned LdOpc = getLdOpcode(LdSize, IsThumb1, IsThumb2);
7065 assert(LdOpc != 0 && "Should have a load opcode");
7067 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
7068 .addReg(AddrOut, RegState::Define).addReg(AddrIn)
7070 } else if (IsThumb1) {
7071 // load + update AddrIn
7072 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
7073 .addReg(AddrIn).addImm(0));
7074 MachineInstrBuilder MIB =
7075 BuildMI(*BB, Pos, dl, TII->get(ARM::tADDi8), AddrOut);
7076 MIB = AddDefaultT1CC(MIB);
7077 MIB.addReg(AddrIn).addImm(LdSize);
7078 AddDefaultPred(MIB);
7079 } else if (IsThumb2) {
7080 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
7081 .addReg(AddrOut, RegState::Define).addReg(AddrIn)
7084 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
7085 .addReg(AddrOut, RegState::Define).addReg(AddrIn)
7086 .addReg(0).addImm(LdSize));
7090 /// Emit a post-increment store operation with given size. The instructions
7091 /// will be added to BB at Pos.
7092 static void emitPostSt(MachineBasicBlock *BB, MachineInstr *Pos,
7093 const TargetInstrInfo *TII, DebugLoc dl,
7094 unsigned StSize, unsigned Data, unsigned AddrIn,
7095 unsigned AddrOut, bool IsThumb1, bool IsThumb2) {
7096 unsigned StOpc = getStOpcode(StSize, IsThumb1, IsThumb2);
7097 assert(StOpc != 0 && "Should have a store opcode");
7099 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc), AddrOut)
7100 .addReg(AddrIn).addImm(0).addReg(Data));
7101 } else if (IsThumb1) {
7102 // store + update AddrIn
7103 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc)).addReg(Data)
7104 .addReg(AddrIn).addImm(0));
7105 MachineInstrBuilder MIB =
7106 BuildMI(*BB, Pos, dl, TII->get(ARM::tADDi8), AddrOut);
7107 MIB = AddDefaultT1CC(MIB);
7108 MIB.addReg(AddrIn).addImm(StSize);
7109 AddDefaultPred(MIB);
7110 } else if (IsThumb2) {
7111 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc), AddrOut)
7112 .addReg(Data).addReg(AddrIn).addImm(StSize));
7114 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc), AddrOut)
7115 .addReg(Data).addReg(AddrIn).addReg(0)
7121 ARMTargetLowering::EmitStructByval(MachineInstr *MI,
7122 MachineBasicBlock *BB) const {
7123 // This pseudo instruction has 3 operands: dst, src, size
7124 // We expand it to a loop if size > Subtarget->getMaxInlineSizeThreshold().
7125 // Otherwise, we will generate unrolled scalar copies.
7126 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
7127 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7128 MachineFunction::iterator It = BB;
7131 unsigned dest = MI->getOperand(0).getReg();
7132 unsigned src = MI->getOperand(1).getReg();
7133 unsigned SizeVal = MI->getOperand(2).getImm();
7134 unsigned Align = MI->getOperand(3).getImm();
7135 DebugLoc dl = MI->getDebugLoc();
7137 MachineFunction *MF = BB->getParent();
7138 MachineRegisterInfo &MRI = MF->getRegInfo();
7139 unsigned UnitSize = 0;
7140 const TargetRegisterClass *TRC = nullptr;
7141 const TargetRegisterClass *VecTRC = nullptr;
7143 bool IsThumb1 = Subtarget->isThumb1Only();
7144 bool IsThumb2 = Subtarget->isThumb2();
7148 } else if (Align & 2) {
7151 // Check whether we can use NEON instructions.
7152 if (!MF->getFunction()->getAttributes().
7153 hasAttribute(AttributeSet::FunctionIndex,
7154 Attribute::NoImplicitFloat) &&
7155 Subtarget->hasNEON()) {
7156 if ((Align % 16 == 0) && SizeVal >= 16)
7158 else if ((Align % 8 == 0) && SizeVal >= 8)
7161 // Can't use NEON instructions.
7166 // Select the correct opcode and register class for unit size load/store
7167 bool IsNeon = UnitSize >= 8;
7168 TRC = (IsThumb1 || IsThumb2) ? &ARM::tGPRRegClass : &ARM::GPRRegClass;
7170 VecTRC = UnitSize == 16 ? &ARM::DPairRegClass
7171 : UnitSize == 8 ? &ARM::DPRRegClass
7174 unsigned BytesLeft = SizeVal % UnitSize;
7175 unsigned LoopSize = SizeVal - BytesLeft;
7177 if (SizeVal <= Subtarget->getMaxInlineSizeThreshold()) {
7178 // Use LDR and STR to copy.
7179 // [scratch, srcOut] = LDR_POST(srcIn, UnitSize)
7180 // [destOut] = STR_POST(scratch, destIn, UnitSize)
7181 unsigned srcIn = src;
7182 unsigned destIn = dest;
7183 for (unsigned i = 0; i < LoopSize; i+=UnitSize) {
7184 unsigned srcOut = MRI.createVirtualRegister(TRC);
7185 unsigned destOut = MRI.createVirtualRegister(TRC);
7186 unsigned scratch = MRI.createVirtualRegister(IsNeon ? VecTRC : TRC);
7187 emitPostLd(BB, MI, TII, dl, UnitSize, scratch, srcIn, srcOut,
7188 IsThumb1, IsThumb2);
7189 emitPostSt(BB, MI, TII, dl, UnitSize, scratch, destIn, destOut,
7190 IsThumb1, IsThumb2);
7195 // Handle the leftover bytes with LDRB and STRB.
7196 // [scratch, srcOut] = LDRB_POST(srcIn, 1)
7197 // [destOut] = STRB_POST(scratch, destIn, 1)
7198 for (unsigned i = 0; i < BytesLeft; i++) {
7199 unsigned srcOut = MRI.createVirtualRegister(TRC);
7200 unsigned destOut = MRI.createVirtualRegister(TRC);
7201 unsigned scratch = MRI.createVirtualRegister(TRC);
7202 emitPostLd(BB, MI, TII, dl, 1, scratch, srcIn, srcOut,
7203 IsThumb1, IsThumb2);
7204 emitPostSt(BB, MI, TII, dl, 1, scratch, destIn, destOut,
7205 IsThumb1, IsThumb2);
7209 MI->eraseFromParent(); // The instruction is gone now.
7213 // Expand the pseudo op to a loop.
7216 // movw varEnd, # --> with thumb2
7218 // ldrcp varEnd, idx --> without thumb2
7219 // fallthrough --> loopMBB
7221 // PHI varPhi, varEnd, varLoop
7222 // PHI srcPhi, src, srcLoop
7223 // PHI destPhi, dst, destLoop
7224 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
7225 // [destLoop] = STR_POST(scratch, destPhi, UnitSize)
7226 // subs varLoop, varPhi, #UnitSize
7228 // fallthrough --> exitMBB
7230 // epilogue to handle left-over bytes
7231 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
7232 // [destOut] = STRB_POST(scratch, destLoop, 1)
7233 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
7234 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
7235 MF->insert(It, loopMBB);
7236 MF->insert(It, exitMBB);
7238 // Transfer the remainder of BB and its successor edges to exitMBB.
7239 exitMBB->splice(exitMBB->begin(), BB,
7240 std::next(MachineBasicBlock::iterator(MI)), BB->end());
7241 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
7243 // Load an immediate to varEnd.
7244 unsigned varEnd = MRI.createVirtualRegister(TRC);
7246 unsigned Vtmp = varEnd;
7247 if ((LoopSize & 0xFFFF0000) != 0)
7248 Vtmp = MRI.createVirtualRegister(TRC);
7249 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2MOVi16), Vtmp)
7250 .addImm(LoopSize & 0xFFFF));
7252 if ((LoopSize & 0xFFFF0000) != 0)
7253 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2MOVTi16), varEnd)
7254 .addReg(Vtmp).addImm(LoopSize >> 16));
7256 MachineConstantPool *ConstantPool = MF->getConstantPool();
7257 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
7258 const Constant *C = ConstantInt::get(Int32Ty, LoopSize);
7260 // MachineConstantPool wants an explicit alignment.
7261 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
7263 Align = getDataLayout()->getTypeAllocSize(C->getType());
7264 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
7267 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(ARM::tLDRpci)).addReg(
7268 varEnd, RegState::Define).addConstantPoolIndex(Idx));
7270 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(ARM::LDRcp)).addReg(
7271 varEnd, RegState::Define).addConstantPoolIndex(Idx).addImm(0));
7273 BB->addSuccessor(loopMBB);
7275 // Generate the loop body:
7276 // varPhi = PHI(varLoop, varEnd)
7277 // srcPhi = PHI(srcLoop, src)
7278 // destPhi = PHI(destLoop, dst)
7279 MachineBasicBlock *entryBB = BB;
7281 unsigned varLoop = MRI.createVirtualRegister(TRC);
7282 unsigned varPhi = MRI.createVirtualRegister(TRC);
7283 unsigned srcLoop = MRI.createVirtualRegister(TRC);
7284 unsigned srcPhi = MRI.createVirtualRegister(TRC);
7285 unsigned destLoop = MRI.createVirtualRegister(TRC);
7286 unsigned destPhi = MRI.createVirtualRegister(TRC);
7288 BuildMI(*BB, BB->begin(), dl, TII->get(ARM::PHI), varPhi)
7289 .addReg(varLoop).addMBB(loopMBB)
7290 .addReg(varEnd).addMBB(entryBB);
7291 BuildMI(BB, dl, TII->get(ARM::PHI), srcPhi)
7292 .addReg(srcLoop).addMBB(loopMBB)
7293 .addReg(src).addMBB(entryBB);
7294 BuildMI(BB, dl, TII->get(ARM::PHI), destPhi)
7295 .addReg(destLoop).addMBB(loopMBB)
7296 .addReg(dest).addMBB(entryBB);
7298 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
7299 // [destLoop] = STR_POST(scratch, destPhi, UnitSiz)
7300 unsigned scratch = MRI.createVirtualRegister(IsNeon ? VecTRC : TRC);
7301 emitPostLd(BB, BB->end(), TII, dl, UnitSize, scratch, srcPhi, srcLoop,
7302 IsThumb1, IsThumb2);
7303 emitPostSt(BB, BB->end(), TII, dl, UnitSize, scratch, destPhi, destLoop,
7304 IsThumb1, IsThumb2);
7306 // Decrement loop variable by UnitSize.
7308 MachineInstrBuilder MIB =
7309 BuildMI(*BB, BB->end(), dl, TII->get(ARM::tSUBi8), varLoop);
7310 MIB = AddDefaultT1CC(MIB);
7311 MIB.addReg(varPhi).addImm(UnitSize);
7312 AddDefaultPred(MIB);
7314 MachineInstrBuilder MIB =
7315 BuildMI(*BB, BB->end(), dl,
7316 TII->get(IsThumb2 ? ARM::t2SUBri : ARM::SUBri), varLoop);
7317 AddDefaultCC(AddDefaultPred(MIB.addReg(varPhi).addImm(UnitSize)));
7318 MIB->getOperand(5).setReg(ARM::CPSR);
7319 MIB->getOperand(5).setIsDef(true);
7321 BuildMI(*BB, BB->end(), dl,
7322 TII->get(IsThumb1 ? ARM::tBcc : IsThumb2 ? ARM::t2Bcc : ARM::Bcc))
7323 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
7325 // loopMBB can loop back to loopMBB or fall through to exitMBB.
7326 BB->addSuccessor(loopMBB);
7327 BB->addSuccessor(exitMBB);
7329 // Add epilogue to handle BytesLeft.
7331 MachineInstr *StartOfExit = exitMBB->begin();
7333 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
7334 // [destOut] = STRB_POST(scratch, destLoop, 1)
7335 unsigned srcIn = srcLoop;
7336 unsigned destIn = destLoop;
7337 for (unsigned i = 0; i < BytesLeft; i++) {
7338 unsigned srcOut = MRI.createVirtualRegister(TRC);
7339 unsigned destOut = MRI.createVirtualRegister(TRC);
7340 unsigned scratch = MRI.createVirtualRegister(TRC);
7341 emitPostLd(BB, StartOfExit, TII, dl, 1, scratch, srcIn, srcOut,
7342 IsThumb1, IsThumb2);
7343 emitPostSt(BB, StartOfExit, TII, dl, 1, scratch, destIn, destOut,
7344 IsThumb1, IsThumb2);
7349 MI->eraseFromParent(); // The instruction is gone now.
7354 ARMTargetLowering::EmitLowered__chkstk(MachineInstr *MI,
7355 MachineBasicBlock *MBB) const {
7356 const TargetMachine &TM = getTargetMachine();
7357 const TargetInstrInfo &TII = *Subtarget->getInstrInfo();
7358 DebugLoc DL = MI->getDebugLoc();
7360 assert(Subtarget->isTargetWindows() &&
7361 "__chkstk is only supported on Windows");
7362 assert(Subtarget->isThumb2() && "Windows on ARM requires Thumb-2 mode");
7364 // __chkstk takes the number of words to allocate on the stack in R4, and
7365 // returns the stack adjustment in number of bytes in R4. This will not
7366 // clober any other registers (other than the obvious lr).
7368 // Although, technically, IP should be considered a register which may be
7369 // clobbered, the call itself will not touch it. Windows on ARM is a pure
7370 // thumb-2 environment, so there is no interworking required. As a result, we
7371 // do not expect a veneer to be emitted by the linker, clobbering IP.
7373 // Each module receives its own copy of __chkstk, so no import thunk is
7374 // required, again, ensuring that IP is not clobbered.
7376 // Finally, although some linkers may theoretically provide a trampoline for
7377 // out of range calls (which is quite common due to a 32M range limitation of
7378 // branches for Thumb), we can generate the long-call version via
7379 // -mcmodel=large, alleviating the need for the trampoline which may clobber
7382 switch (TM.getCodeModel()) {
7383 case CodeModel::Small:
7384 case CodeModel::Medium:
7385 case CodeModel::Default:
7386 case CodeModel::Kernel:
7387 BuildMI(*MBB, MI, DL, TII.get(ARM::tBL))
7388 .addImm((unsigned)ARMCC::AL).addReg(0)
7389 .addExternalSymbol("__chkstk")
7390 .addReg(ARM::R4, RegState::Implicit | RegState::Kill)
7391 .addReg(ARM::R4, RegState::Implicit | RegState::Define)
7392 .addReg(ARM::R12, RegState::Implicit | RegState::Define | RegState::Dead);
7394 case CodeModel::Large:
7395 case CodeModel::JITDefault: {
7396 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
7397 unsigned Reg = MRI.createVirtualRegister(&ARM::rGPRRegClass);
7399 BuildMI(*MBB, MI, DL, TII.get(ARM::t2MOVi32imm), Reg)
7400 .addExternalSymbol("__chkstk");
7401 BuildMI(*MBB, MI, DL, TII.get(ARM::tBLXr))
7402 .addImm((unsigned)ARMCC::AL).addReg(0)
7403 .addReg(Reg, RegState::Kill)
7404 .addReg(ARM::R4, RegState::Implicit | RegState::Kill)
7405 .addReg(ARM::R4, RegState::Implicit | RegState::Define)
7406 .addReg(ARM::R12, RegState::Implicit | RegState::Define | RegState::Dead);
7411 AddDefaultCC(AddDefaultPred(BuildMI(*MBB, MI, DL, TII.get(ARM::t2SUBrr),
7413 .addReg(ARM::SP).addReg(ARM::R4)));
7415 MI->eraseFromParent();
7420 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
7421 MachineBasicBlock *BB) const {
7422 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
7423 DebugLoc dl = MI->getDebugLoc();
7424 bool isThumb2 = Subtarget->isThumb2();
7425 switch (MI->getOpcode()) {
7428 llvm_unreachable("Unexpected instr type to insert");
7430 // The Thumb2 pre-indexed stores have the same MI operands, they just
7431 // define them differently in the .td files from the isel patterns, so
7432 // they need pseudos.
7433 case ARM::t2STR_preidx:
7434 MI->setDesc(TII->get(ARM::t2STR_PRE));
7436 case ARM::t2STRB_preidx:
7437 MI->setDesc(TII->get(ARM::t2STRB_PRE));
7439 case ARM::t2STRH_preidx:
7440 MI->setDesc(TII->get(ARM::t2STRH_PRE));
7443 case ARM::STRi_preidx:
7444 case ARM::STRBi_preidx: {
7445 unsigned NewOpc = MI->getOpcode() == ARM::STRi_preidx ?
7446 ARM::STR_PRE_IMM : ARM::STRB_PRE_IMM;
7447 // Decode the offset.
7448 unsigned Offset = MI->getOperand(4).getImm();
7449 bool isSub = ARM_AM::getAM2Op(Offset) == ARM_AM::sub;
7450 Offset = ARM_AM::getAM2Offset(Offset);
7454 MachineMemOperand *MMO = *MI->memoperands_begin();
7455 BuildMI(*BB, MI, dl, TII->get(NewOpc))
7456 .addOperand(MI->getOperand(0)) // Rn_wb
7457 .addOperand(MI->getOperand(1)) // Rt
7458 .addOperand(MI->getOperand(2)) // Rn
7459 .addImm(Offset) // offset (skip GPR==zero_reg)
7460 .addOperand(MI->getOperand(5)) // pred
7461 .addOperand(MI->getOperand(6))
7462 .addMemOperand(MMO);
7463 MI->eraseFromParent();
7466 case ARM::STRr_preidx:
7467 case ARM::STRBr_preidx:
7468 case ARM::STRH_preidx: {
7470 switch (MI->getOpcode()) {
7471 default: llvm_unreachable("unexpected opcode!");
7472 case ARM::STRr_preidx: NewOpc = ARM::STR_PRE_REG; break;
7473 case ARM::STRBr_preidx: NewOpc = ARM::STRB_PRE_REG; break;
7474 case ARM::STRH_preidx: NewOpc = ARM::STRH_PRE; break;
7476 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(NewOpc));
7477 for (unsigned i = 0; i < MI->getNumOperands(); ++i)
7478 MIB.addOperand(MI->getOperand(i));
7479 MI->eraseFromParent();
7483 case ARM::tMOVCCr_pseudo: {
7484 // To "insert" a SELECT_CC instruction, we actually have to insert the
7485 // diamond control-flow pattern. The incoming instruction knows the
7486 // destination vreg to set, the condition code register to branch on, the
7487 // true/false values to select between, and a branch opcode to use.
7488 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7489 MachineFunction::iterator It = BB;
7495 // cmpTY ccX, r1, r2
7497 // fallthrough --> copy0MBB
7498 MachineBasicBlock *thisMBB = BB;
7499 MachineFunction *F = BB->getParent();
7500 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
7501 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
7502 F->insert(It, copy0MBB);
7503 F->insert(It, sinkMBB);
7505 // Transfer the remainder of BB and its successor edges to sinkMBB.
7506 sinkMBB->splice(sinkMBB->begin(), BB,
7507 std::next(MachineBasicBlock::iterator(MI)), BB->end());
7508 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
7510 BB->addSuccessor(copy0MBB);
7511 BB->addSuccessor(sinkMBB);
7513 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
7514 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
7517 // %FalseValue = ...
7518 // # fallthrough to sinkMBB
7521 // Update machine-CFG edges
7522 BB->addSuccessor(sinkMBB);
7525 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
7528 BuildMI(*BB, BB->begin(), dl,
7529 TII->get(ARM::PHI), MI->getOperand(0).getReg())
7530 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
7531 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
7533 MI->eraseFromParent(); // The pseudo instruction is gone now.
7538 case ARM::BCCZi64: {
7539 // If there is an unconditional branch to the other successor, remove it.
7540 BB->erase(std::next(MachineBasicBlock::iterator(MI)), BB->end());
7542 // Compare both parts that make up the double comparison separately for
7544 bool RHSisZero = MI->getOpcode() == ARM::BCCZi64;
7546 unsigned LHS1 = MI->getOperand(1).getReg();
7547 unsigned LHS2 = MI->getOperand(2).getReg();
7549 AddDefaultPred(BuildMI(BB, dl,
7550 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7551 .addReg(LHS1).addImm(0));
7552 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7553 .addReg(LHS2).addImm(0)
7554 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
7556 unsigned RHS1 = MI->getOperand(3).getReg();
7557 unsigned RHS2 = MI->getOperand(4).getReg();
7558 AddDefaultPred(BuildMI(BB, dl,
7559 TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
7560 .addReg(LHS1).addReg(RHS1));
7561 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
7562 .addReg(LHS2).addReg(RHS2)
7563 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
7566 MachineBasicBlock *destMBB = MI->getOperand(RHSisZero ? 3 : 5).getMBB();
7567 MachineBasicBlock *exitMBB = OtherSucc(BB, destMBB);
7568 if (MI->getOperand(0).getImm() == ARMCC::NE)
7569 std::swap(destMBB, exitMBB);
7571 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
7572 .addMBB(destMBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
7574 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2B)).addMBB(exitMBB));
7576 BuildMI(BB, dl, TII->get(ARM::B)) .addMBB(exitMBB);
7578 MI->eraseFromParent(); // The pseudo instruction is gone now.
7582 case ARM::Int_eh_sjlj_setjmp:
7583 case ARM::Int_eh_sjlj_setjmp_nofp:
7584 case ARM::tInt_eh_sjlj_setjmp:
7585 case ARM::t2Int_eh_sjlj_setjmp:
7586 case ARM::t2Int_eh_sjlj_setjmp_nofp:
7587 EmitSjLjDispatchBlock(MI, BB);
7592 // To insert an ABS instruction, we have to insert the
7593 // diamond control-flow pattern. The incoming instruction knows the
7594 // source vreg to test against 0, the destination vreg to set,
7595 // the condition code register to branch on, the
7596 // true/false values to select between, and a branch opcode to use.
7601 // BCC (branch to SinkBB if V0 >= 0)
7602 // RSBBB: V3 = RSBri V2, 0 (compute ABS if V2 < 0)
7603 // SinkBB: V1 = PHI(V2, V3)
7604 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7605 MachineFunction::iterator BBI = BB;
7607 MachineFunction *Fn = BB->getParent();
7608 MachineBasicBlock *RSBBB = Fn->CreateMachineBasicBlock(LLVM_BB);
7609 MachineBasicBlock *SinkBB = Fn->CreateMachineBasicBlock(LLVM_BB);
7610 Fn->insert(BBI, RSBBB);
7611 Fn->insert(BBI, SinkBB);
7613 unsigned int ABSSrcReg = MI->getOperand(1).getReg();
7614 unsigned int ABSDstReg = MI->getOperand(0).getReg();
7615 bool isThumb2 = Subtarget->isThumb2();
7616 MachineRegisterInfo &MRI = Fn->getRegInfo();
7617 // In Thumb mode S must not be specified if source register is the SP or
7618 // PC and if destination register is the SP, so restrict register class
7619 unsigned NewRsbDstReg =
7620 MRI.createVirtualRegister(isThumb2 ? &ARM::rGPRRegClass : &ARM::GPRRegClass);
7622 // Transfer the remainder of BB and its successor edges to sinkMBB.
7623 SinkBB->splice(SinkBB->begin(), BB,
7624 std::next(MachineBasicBlock::iterator(MI)), BB->end());
7625 SinkBB->transferSuccessorsAndUpdatePHIs(BB);
7627 BB->addSuccessor(RSBBB);
7628 BB->addSuccessor(SinkBB);
7630 // fall through to SinkMBB
7631 RSBBB->addSuccessor(SinkBB);
7633 // insert a cmp at the end of BB
7634 AddDefaultPred(BuildMI(BB, dl,
7635 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7636 .addReg(ABSSrcReg).addImm(0));
7638 // insert a bcc with opposite CC to ARMCC::MI at the end of BB
7640 TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)).addMBB(SinkBB)
7641 .addImm(ARMCC::getOppositeCondition(ARMCC::MI)).addReg(ARM::CPSR);
7643 // insert rsbri in RSBBB
7644 // Note: BCC and rsbri will be converted into predicated rsbmi
7645 // by if-conversion pass
7646 BuildMI(*RSBBB, RSBBB->begin(), dl,
7647 TII->get(isThumb2 ? ARM::t2RSBri : ARM::RSBri), NewRsbDstReg)
7648 .addReg(ABSSrcReg, RegState::Kill)
7649 .addImm(0).addImm((unsigned)ARMCC::AL).addReg(0).addReg(0);
7651 // insert PHI in SinkBB,
7652 // reuse ABSDstReg to not change uses of ABS instruction
7653 BuildMI(*SinkBB, SinkBB->begin(), dl,
7654 TII->get(ARM::PHI), ABSDstReg)
7655 .addReg(NewRsbDstReg).addMBB(RSBBB)
7656 .addReg(ABSSrcReg).addMBB(BB);
7658 // remove ABS instruction
7659 MI->eraseFromParent();
7661 // return last added BB
7664 case ARM::COPY_STRUCT_BYVAL_I32:
7666 return EmitStructByval(MI, BB);
7667 case ARM::WIN__CHKSTK:
7668 return EmitLowered__chkstk(MI, BB);
7672 void ARMTargetLowering::AdjustInstrPostInstrSelection(MachineInstr *MI,
7673 SDNode *Node) const {
7674 const MCInstrDesc *MCID = &MI->getDesc();
7675 // Adjust potentially 's' setting instructions after isel, i.e. ADC, SBC, RSB,
7676 // RSC. Coming out of isel, they have an implicit CPSR def, but the optional
7677 // operand is still set to noreg. If needed, set the optional operand's
7678 // register to CPSR, and remove the redundant implicit def.
7680 // e.g. ADCS (..., CPSR<imp-def>) -> ADC (... opt:CPSR<def>).
7682 // Rename pseudo opcodes.
7683 unsigned NewOpc = convertAddSubFlagsOpcode(MI->getOpcode());
7685 const ARMBaseInstrInfo *TII = Subtarget->getInstrInfo();
7686 MCID = &TII->get(NewOpc);
7688 assert(MCID->getNumOperands() == MI->getDesc().getNumOperands() + 1 &&
7689 "converted opcode should be the same except for cc_out");
7693 // Add the optional cc_out operand
7694 MI->addOperand(MachineOperand::CreateReg(0, /*isDef=*/true));
7696 unsigned ccOutIdx = MCID->getNumOperands() - 1;
7698 // Any ARM instruction that sets the 's' bit should specify an optional
7699 // "cc_out" operand in the last operand position.
7700 if (!MI->hasOptionalDef() || !MCID->OpInfo[ccOutIdx].isOptionalDef()) {
7701 assert(!NewOpc && "Optional cc_out operand required");
7704 // Look for an implicit def of CPSR added by MachineInstr ctor. Remove it
7705 // since we already have an optional CPSR def.
7706 bool definesCPSR = false;
7707 bool deadCPSR = false;
7708 for (unsigned i = MCID->getNumOperands(), e = MI->getNumOperands();
7710 const MachineOperand &MO = MI->getOperand(i);
7711 if (MO.isReg() && MO.isDef() && MO.getReg() == ARM::CPSR) {
7715 MI->RemoveOperand(i);
7720 assert(!NewOpc && "Optional cc_out operand required");
7723 assert(deadCPSR == !Node->hasAnyUseOfValue(1) && "inconsistent dead flag");
7725 assert(!MI->getOperand(ccOutIdx).getReg() &&
7726 "expect uninitialized optional cc_out operand");
7730 // If this instruction was defined with an optional CPSR def and its dag node
7731 // had a live implicit CPSR def, then activate the optional CPSR def.
7732 MachineOperand &MO = MI->getOperand(ccOutIdx);
7733 MO.setReg(ARM::CPSR);
7737 //===----------------------------------------------------------------------===//
7738 // ARM Optimization Hooks
7739 //===----------------------------------------------------------------------===//
7741 // Helper function that checks if N is a null or all ones constant.
7742 static inline bool isZeroOrAllOnes(SDValue N, bool AllOnes) {
7743 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N);
7746 return AllOnes ? C->isAllOnesValue() : C->isNullValue();
7749 // Return true if N is conditionally 0 or all ones.
7750 // Detects these expressions where cc is an i1 value:
7752 // (select cc 0, y) [AllOnes=0]
7753 // (select cc y, 0) [AllOnes=0]
7754 // (zext cc) [AllOnes=0]
7755 // (sext cc) [AllOnes=0/1]
7756 // (select cc -1, y) [AllOnes=1]
7757 // (select cc y, -1) [AllOnes=1]
7759 // Invert is set when N is the null/all ones constant when CC is false.
7760 // OtherOp is set to the alternative value of N.
7761 static bool isConditionalZeroOrAllOnes(SDNode *N, bool AllOnes,
7762 SDValue &CC, bool &Invert,
7764 SelectionDAG &DAG) {
7765 switch (N->getOpcode()) {
7766 default: return false;
7768 CC = N->getOperand(0);
7769 SDValue N1 = N->getOperand(1);
7770 SDValue N2 = N->getOperand(2);
7771 if (isZeroOrAllOnes(N1, AllOnes)) {
7776 if (isZeroOrAllOnes(N2, AllOnes)) {
7783 case ISD::ZERO_EXTEND:
7784 // (zext cc) can never be the all ones value.
7788 case ISD::SIGN_EXTEND: {
7789 EVT VT = N->getValueType(0);
7790 CC = N->getOperand(0);
7791 if (CC.getValueType() != MVT::i1)
7795 // When looking for an AllOnes constant, N is an sext, and the 'other'
7797 OtherOp = DAG.getConstant(0, VT);
7798 else if (N->getOpcode() == ISD::ZERO_EXTEND)
7799 // When looking for a 0 constant, N can be zext or sext.
7800 OtherOp = DAG.getConstant(1, VT);
7802 OtherOp = DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), VT);
7808 // Combine a constant select operand into its use:
7810 // (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
7811 // (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
7812 // (and (select cc, -1, c), x) -> (select cc, x, (and, x, c)) [AllOnes=1]
7813 // (or (select cc, 0, c), x) -> (select cc, x, (or, x, c))
7814 // (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c))
7816 // The transform is rejected if the select doesn't have a constant operand that
7817 // is null, or all ones when AllOnes is set.
7819 // Also recognize sext/zext from i1:
7821 // (add (zext cc), x) -> (select cc (add x, 1), x)
7822 // (add (sext cc), x) -> (select cc (add x, -1), x)
7824 // These transformations eventually create predicated instructions.
7826 // @param N The node to transform.
7827 // @param Slct The N operand that is a select.
7828 // @param OtherOp The other N operand (x above).
7829 // @param DCI Context.
7830 // @param AllOnes Require the select constant to be all ones instead of null.
7831 // @returns The new node, or SDValue() on failure.
7833 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
7834 TargetLowering::DAGCombinerInfo &DCI,
7835 bool AllOnes = false) {
7836 SelectionDAG &DAG = DCI.DAG;
7837 EVT VT = N->getValueType(0);
7838 SDValue NonConstantVal;
7841 if (!isConditionalZeroOrAllOnes(Slct.getNode(), AllOnes, CCOp, SwapSelectOps,
7842 NonConstantVal, DAG))
7845 // Slct is now know to be the desired identity constant when CC is true.
7846 SDValue TrueVal = OtherOp;
7847 SDValue FalseVal = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
7848 OtherOp, NonConstantVal);
7849 // Unless SwapSelectOps says CC should be false.
7851 std::swap(TrueVal, FalseVal);
7853 return DAG.getNode(ISD::SELECT, SDLoc(N), VT,
7854 CCOp, TrueVal, FalseVal);
7857 // Attempt combineSelectAndUse on each operand of a commutative operator N.
7859 SDValue combineSelectAndUseCommutative(SDNode *N, bool AllOnes,
7860 TargetLowering::DAGCombinerInfo &DCI) {
7861 SDValue N0 = N->getOperand(0);
7862 SDValue N1 = N->getOperand(1);
7863 if (N0.getNode()->hasOneUse()) {
7864 SDValue Result = combineSelectAndUse(N, N0, N1, DCI, AllOnes);
7865 if (Result.getNode())
7868 if (N1.getNode()->hasOneUse()) {
7869 SDValue Result = combineSelectAndUse(N, N1, N0, DCI, AllOnes);
7870 if (Result.getNode())
7876 // AddCombineToVPADDL- For pair-wise add on neon, use the vpaddl instruction
7877 // (only after legalization).
7878 static SDValue AddCombineToVPADDL(SDNode *N, SDValue N0, SDValue N1,
7879 TargetLowering::DAGCombinerInfo &DCI,
7880 const ARMSubtarget *Subtarget) {
7882 // Only perform optimization if after legalize, and if NEON is available. We
7883 // also expected both operands to be BUILD_VECTORs.
7884 if (DCI.isBeforeLegalize() || !Subtarget->hasNEON()
7885 || N0.getOpcode() != ISD::BUILD_VECTOR
7886 || N1.getOpcode() != ISD::BUILD_VECTOR)
7889 // Check output type since VPADDL operand elements can only be 8, 16, or 32.
7890 EVT VT = N->getValueType(0);
7891 if (!VT.isInteger() || VT.getVectorElementType() == MVT::i64)
7894 // Check that the vector operands are of the right form.
7895 // N0 and N1 are BUILD_VECTOR nodes with N number of EXTRACT_VECTOR
7896 // operands, where N is the size of the formed vector.
7897 // Each EXTRACT_VECTOR should have the same input vector and odd or even
7898 // index such that we have a pair wise add pattern.
7900 // Grab the vector that all EXTRACT_VECTOR nodes should be referencing.
7901 if (N0->getOperand(0)->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
7903 SDValue Vec = N0->getOperand(0)->getOperand(0);
7904 SDNode *V = Vec.getNode();
7905 unsigned nextIndex = 0;
7907 // For each operands to the ADD which are BUILD_VECTORs,
7908 // check to see if each of their operands are an EXTRACT_VECTOR with
7909 // the same vector and appropriate index.
7910 for (unsigned i = 0, e = N0->getNumOperands(); i != e; ++i) {
7911 if (N0->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT
7912 && N1->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
7914 SDValue ExtVec0 = N0->getOperand(i);
7915 SDValue ExtVec1 = N1->getOperand(i);
7917 // First operand is the vector, verify its the same.
7918 if (V != ExtVec0->getOperand(0).getNode() ||
7919 V != ExtVec1->getOperand(0).getNode())
7922 // Second is the constant, verify its correct.
7923 ConstantSDNode *C0 = dyn_cast<ConstantSDNode>(ExtVec0->getOperand(1));
7924 ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(ExtVec1->getOperand(1));
7926 // For the constant, we want to see all the even or all the odd.
7927 if (!C0 || !C1 || C0->getZExtValue() != nextIndex
7928 || C1->getZExtValue() != nextIndex+1)
7937 // Create VPADDL node.
7938 SelectionDAG &DAG = DCI.DAG;
7939 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7941 // Build operand list.
7942 SmallVector<SDValue, 8> Ops;
7943 Ops.push_back(DAG.getConstant(Intrinsic::arm_neon_vpaddls,
7944 TLI.getPointerTy()));
7946 // Input is the vector.
7949 // Get widened type and narrowed type.
7951 unsigned numElem = VT.getVectorNumElements();
7953 EVT inputLaneType = Vec.getValueType().getVectorElementType();
7954 switch (inputLaneType.getSimpleVT().SimpleTy) {
7955 case MVT::i8: widenType = MVT::getVectorVT(MVT::i16, numElem); break;
7956 case MVT::i16: widenType = MVT::getVectorVT(MVT::i32, numElem); break;
7957 case MVT::i32: widenType = MVT::getVectorVT(MVT::i64, numElem); break;
7959 llvm_unreachable("Invalid vector element type for padd optimization.");
7962 SDValue tmp = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N), widenType, Ops);
7963 unsigned ExtOp = VT.bitsGT(tmp.getValueType()) ? ISD::ANY_EXTEND : ISD::TRUNCATE;
7964 return DAG.getNode(ExtOp, SDLoc(N), VT, tmp);
7967 static SDValue findMUL_LOHI(SDValue V) {
7968 if (V->getOpcode() == ISD::UMUL_LOHI ||
7969 V->getOpcode() == ISD::SMUL_LOHI)
7974 static SDValue AddCombineTo64bitMLAL(SDNode *AddcNode,
7975 TargetLowering::DAGCombinerInfo &DCI,
7976 const ARMSubtarget *Subtarget) {
7978 if (Subtarget->isThumb1Only()) return SDValue();
7980 // Only perform the checks after legalize when the pattern is available.
7981 if (DCI.isBeforeLegalize()) return SDValue();
7983 // Look for multiply add opportunities.
7984 // The pattern is a ISD::UMUL_LOHI followed by two add nodes, where
7985 // each add nodes consumes a value from ISD::UMUL_LOHI and there is
7986 // a glue link from the first add to the second add.
7987 // If we find this pattern, we can replace the U/SMUL_LOHI, ADDC, and ADDE by
7988 // a S/UMLAL instruction.
7991 // \ / \ [no multiline comment]
7997 assert(AddcNode->getOpcode() == ISD::ADDC && "Expect an ADDC");
7998 SDValue AddcOp0 = AddcNode->getOperand(0);
7999 SDValue AddcOp1 = AddcNode->getOperand(1);
8001 // Check if the two operands are from the same mul_lohi node.
8002 if (AddcOp0.getNode() == AddcOp1.getNode())
8005 assert(AddcNode->getNumValues() == 2 &&
8006 AddcNode->getValueType(0) == MVT::i32 &&
8007 "Expect ADDC with two result values. First: i32");
8009 // Check that we have a glued ADDC node.
8010 if (AddcNode->getValueType(1) != MVT::Glue)
8013 // Check that the ADDC adds the low result of the S/UMUL_LOHI.
8014 if (AddcOp0->getOpcode() != ISD::UMUL_LOHI &&
8015 AddcOp0->getOpcode() != ISD::SMUL_LOHI &&
8016 AddcOp1->getOpcode() != ISD::UMUL_LOHI &&
8017 AddcOp1->getOpcode() != ISD::SMUL_LOHI)
8020 // Look for the glued ADDE.
8021 SDNode* AddeNode = AddcNode->getGluedUser();
8025 // Make sure it is really an ADDE.
8026 if (AddeNode->getOpcode() != ISD::ADDE)
8029 assert(AddeNode->getNumOperands() == 3 &&
8030 AddeNode->getOperand(2).getValueType() == MVT::Glue &&
8031 "ADDE node has the wrong inputs");
8033 // Check for the triangle shape.
8034 SDValue AddeOp0 = AddeNode->getOperand(0);
8035 SDValue AddeOp1 = AddeNode->getOperand(1);
8037 // Make sure that the ADDE operands are not coming from the same node.
8038 if (AddeOp0.getNode() == AddeOp1.getNode())
8041 // Find the MUL_LOHI node walking up ADDE's operands.
8042 bool IsLeftOperandMUL = false;
8043 SDValue MULOp = findMUL_LOHI(AddeOp0);
8044 if (MULOp == SDValue())
8045 MULOp = findMUL_LOHI(AddeOp1);
8047 IsLeftOperandMUL = true;
8048 if (MULOp == SDValue())
8051 // Figure out the right opcode.
8052 unsigned Opc = MULOp->getOpcode();
8053 unsigned FinalOpc = (Opc == ISD::SMUL_LOHI) ? ARMISD::SMLAL : ARMISD::UMLAL;
8055 // Figure out the high and low input values to the MLAL node.
8056 SDValue* HiAdd = nullptr;
8057 SDValue* LoMul = nullptr;
8058 SDValue* LowAdd = nullptr;
8060 // Ensure that ADDE is from high result of ISD::SMUL_LOHI.
8061 if ((AddeOp0 != MULOp.getValue(1)) && (AddeOp1 != MULOp.getValue(1)))
8064 if (IsLeftOperandMUL)
8070 // Ensure that LoMul and LowAdd are taken from correct ISD::SMUL_LOHI node
8071 // whose low result is fed to the ADDC we are checking.
8073 if (AddcOp0 == MULOp.getValue(0)) {
8077 if (AddcOp1 == MULOp.getValue(0)) {
8085 // Create the merged node.
8086 SelectionDAG &DAG = DCI.DAG;
8088 // Build operand list.
8089 SmallVector<SDValue, 8> Ops;
8090 Ops.push_back(LoMul->getOperand(0));
8091 Ops.push_back(LoMul->getOperand(1));
8092 Ops.push_back(*LowAdd);
8093 Ops.push_back(*HiAdd);
8095 SDValue MLALNode = DAG.getNode(FinalOpc, SDLoc(AddcNode),
8096 DAG.getVTList(MVT::i32, MVT::i32), Ops);
8098 // Replace the ADDs' nodes uses by the MLA node's values.
8099 SDValue HiMLALResult(MLALNode.getNode(), 1);
8100 DAG.ReplaceAllUsesOfValueWith(SDValue(AddeNode, 0), HiMLALResult);
8102 SDValue LoMLALResult(MLALNode.getNode(), 0);
8103 DAG.ReplaceAllUsesOfValueWith(SDValue(AddcNode, 0), LoMLALResult);
8105 // Return original node to notify the driver to stop replacing.
8106 SDValue resNode(AddcNode, 0);
8110 /// PerformADDCCombine - Target-specific dag combine transform from
8111 /// ISD::ADDC, ISD::ADDE, and ISD::MUL_LOHI to MLAL.
8112 static SDValue PerformADDCCombine(SDNode *N,
8113 TargetLowering::DAGCombinerInfo &DCI,
8114 const ARMSubtarget *Subtarget) {
8116 return AddCombineTo64bitMLAL(N, DCI, Subtarget);
8120 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
8121 /// operands N0 and N1. This is a helper for PerformADDCombine that is
8122 /// called with the default operands, and if that fails, with commuted
8124 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
8125 TargetLowering::DAGCombinerInfo &DCI,
8126 const ARMSubtarget *Subtarget){
8128 // Attempt to create vpaddl for this add.
8129 SDValue Result = AddCombineToVPADDL(N, N0, N1, DCI, Subtarget);
8130 if (Result.getNode())
8133 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
8134 if (N0.getNode()->hasOneUse()) {
8135 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
8136 if (Result.getNode()) return Result;
8141 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
8143 static SDValue PerformADDCombine(SDNode *N,
8144 TargetLowering::DAGCombinerInfo &DCI,
8145 const ARMSubtarget *Subtarget) {
8146 SDValue N0 = N->getOperand(0);
8147 SDValue N1 = N->getOperand(1);
8149 // First try with the default operand order.
8150 SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI, Subtarget);
8151 if (Result.getNode())
8154 // If that didn't work, try again with the operands commuted.
8155 return PerformADDCombineWithOperands(N, N1, N0, DCI, Subtarget);
8158 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
8160 static SDValue PerformSUBCombine(SDNode *N,
8161 TargetLowering::DAGCombinerInfo &DCI) {
8162 SDValue N0 = N->getOperand(0);
8163 SDValue N1 = N->getOperand(1);
8165 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
8166 if (N1.getNode()->hasOneUse()) {
8167 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
8168 if (Result.getNode()) return Result;
8174 /// PerformVMULCombine
8175 /// Distribute (A + B) * C to (A * C) + (B * C) to take advantage of the
8176 /// special multiplier accumulator forwarding.
8182 // However, for (A + B) * (A + B),
8189 static SDValue PerformVMULCombine(SDNode *N,
8190 TargetLowering::DAGCombinerInfo &DCI,
8191 const ARMSubtarget *Subtarget) {
8192 if (!Subtarget->hasVMLxForwarding())
8195 SelectionDAG &DAG = DCI.DAG;
8196 SDValue N0 = N->getOperand(0);
8197 SDValue N1 = N->getOperand(1);
8198 unsigned Opcode = N0.getOpcode();
8199 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
8200 Opcode != ISD::FADD && Opcode != ISD::FSUB) {
8201 Opcode = N1.getOpcode();
8202 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
8203 Opcode != ISD::FADD && Opcode != ISD::FSUB)
8211 EVT VT = N->getValueType(0);
8213 SDValue N00 = N0->getOperand(0);
8214 SDValue N01 = N0->getOperand(1);
8215 return DAG.getNode(Opcode, DL, VT,
8216 DAG.getNode(ISD::MUL, DL, VT, N00, N1),
8217 DAG.getNode(ISD::MUL, DL, VT, N01, N1));
8220 static SDValue PerformMULCombine(SDNode *N,
8221 TargetLowering::DAGCombinerInfo &DCI,
8222 const ARMSubtarget *Subtarget) {
8223 SelectionDAG &DAG = DCI.DAG;
8225 if (Subtarget->isThumb1Only())
8228 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
8231 EVT VT = N->getValueType(0);
8232 if (VT.is64BitVector() || VT.is128BitVector())
8233 return PerformVMULCombine(N, DCI, Subtarget);
8237 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
8241 int64_t MulAmt = C->getSExtValue();
8242 unsigned ShiftAmt = countTrailingZeros<uint64_t>(MulAmt);
8244 ShiftAmt = ShiftAmt & (32 - 1);
8245 SDValue V = N->getOperand(0);
8249 MulAmt >>= ShiftAmt;
8252 if (isPowerOf2_32(MulAmt - 1)) {
8253 // (mul x, 2^N + 1) => (add (shl x, N), x)
8254 Res = DAG.getNode(ISD::ADD, DL, VT,
8256 DAG.getNode(ISD::SHL, DL, VT,
8258 DAG.getConstant(Log2_32(MulAmt - 1),
8260 } else if (isPowerOf2_32(MulAmt + 1)) {
8261 // (mul x, 2^N - 1) => (sub (shl x, N), x)
8262 Res = DAG.getNode(ISD::SUB, DL, VT,
8263 DAG.getNode(ISD::SHL, DL, VT,
8265 DAG.getConstant(Log2_32(MulAmt + 1),
8271 uint64_t MulAmtAbs = -MulAmt;
8272 if (isPowerOf2_32(MulAmtAbs + 1)) {
8273 // (mul x, -(2^N - 1)) => (sub x, (shl x, N))
8274 Res = DAG.getNode(ISD::SUB, DL, VT,
8276 DAG.getNode(ISD::SHL, DL, VT,
8278 DAG.getConstant(Log2_32(MulAmtAbs + 1),
8280 } else if (isPowerOf2_32(MulAmtAbs - 1)) {
8281 // (mul x, -(2^N + 1)) => - (add (shl x, N), x)
8282 Res = DAG.getNode(ISD::ADD, DL, VT,
8284 DAG.getNode(ISD::SHL, DL, VT,
8286 DAG.getConstant(Log2_32(MulAmtAbs-1),
8288 Res = DAG.getNode(ISD::SUB, DL, VT,
8289 DAG.getConstant(0, MVT::i32),Res);
8296 Res = DAG.getNode(ISD::SHL, DL, VT,
8297 Res, DAG.getConstant(ShiftAmt, MVT::i32));
8299 // Do not add new nodes to DAG combiner worklist.
8300 DCI.CombineTo(N, Res, false);
8304 static SDValue PerformANDCombine(SDNode *N,
8305 TargetLowering::DAGCombinerInfo &DCI,
8306 const ARMSubtarget *Subtarget) {
8308 // Attempt to use immediate-form VBIC
8309 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
8311 EVT VT = N->getValueType(0);
8312 SelectionDAG &DAG = DCI.DAG;
8314 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8317 APInt SplatBits, SplatUndef;
8318 unsigned SplatBitSize;
8321 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
8322 if (SplatBitSize <= 64) {
8324 SDValue Val = isNEONModifiedImm((~SplatBits).getZExtValue(),
8325 SplatUndef.getZExtValue(), SplatBitSize,
8326 DAG, VbicVT, VT.is128BitVector(),
8328 if (Val.getNode()) {
8330 DAG.getNode(ISD::BITCAST, dl, VbicVT, N->getOperand(0));
8331 SDValue Vbic = DAG.getNode(ARMISD::VBICIMM, dl, VbicVT, Input, Val);
8332 return DAG.getNode(ISD::BITCAST, dl, VT, Vbic);
8337 if (!Subtarget->isThumb1Only()) {
8338 // fold (and (select cc, -1, c), x) -> (select cc, x, (and, x, c))
8339 SDValue Result = combineSelectAndUseCommutative(N, true, DCI);
8340 if (Result.getNode())
8347 /// PerformORCombine - Target-specific dag combine xforms for ISD::OR
8348 static SDValue PerformORCombine(SDNode *N,
8349 TargetLowering::DAGCombinerInfo &DCI,
8350 const ARMSubtarget *Subtarget) {
8351 // Attempt to use immediate-form VORR
8352 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
8354 EVT VT = N->getValueType(0);
8355 SelectionDAG &DAG = DCI.DAG;
8357 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8360 APInt SplatBits, SplatUndef;
8361 unsigned SplatBitSize;
8363 if (BVN && Subtarget->hasNEON() &&
8364 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
8365 if (SplatBitSize <= 64) {
8367 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
8368 SplatUndef.getZExtValue(), SplatBitSize,
8369 DAG, VorrVT, VT.is128BitVector(),
8371 if (Val.getNode()) {
8373 DAG.getNode(ISD::BITCAST, dl, VorrVT, N->getOperand(0));
8374 SDValue Vorr = DAG.getNode(ARMISD::VORRIMM, dl, VorrVT, Input, Val);
8375 return DAG.getNode(ISD::BITCAST, dl, VT, Vorr);
8380 if (!Subtarget->isThumb1Only()) {
8381 // fold (or (select cc, 0, c), x) -> (select cc, x, (or, x, c))
8382 SDValue Result = combineSelectAndUseCommutative(N, false, DCI);
8383 if (Result.getNode())
8387 // The code below optimizes (or (and X, Y), Z).
8388 // The AND operand needs to have a single user to make these optimizations
8390 SDValue N0 = N->getOperand(0);
8391 if (N0.getOpcode() != ISD::AND || !N0.hasOneUse())
8393 SDValue N1 = N->getOperand(1);
8395 // (or (and B, A), (and C, ~A)) => (VBSL A, B, C) when A is a constant.
8396 if (Subtarget->hasNEON() && N1.getOpcode() == ISD::AND && VT.isVector() &&
8397 DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
8399 unsigned SplatBitSize;
8402 APInt SplatBits0, SplatBits1;
8403 BuildVectorSDNode *BVN0 = dyn_cast<BuildVectorSDNode>(N0->getOperand(1));
8404 BuildVectorSDNode *BVN1 = dyn_cast<BuildVectorSDNode>(N1->getOperand(1));
8405 // Ensure that the second operand of both ands are constants
8406 if (BVN0 && BVN0->isConstantSplat(SplatBits0, SplatUndef, SplatBitSize,
8407 HasAnyUndefs) && !HasAnyUndefs) {
8408 if (BVN1 && BVN1->isConstantSplat(SplatBits1, SplatUndef, SplatBitSize,
8409 HasAnyUndefs) && !HasAnyUndefs) {
8410 // Ensure that the bit width of the constants are the same and that
8411 // the splat arguments are logical inverses as per the pattern we
8412 // are trying to simplify.
8413 if (SplatBits0.getBitWidth() == SplatBits1.getBitWidth() &&
8414 SplatBits0 == ~SplatBits1) {
8415 // Canonicalize the vector type to make instruction selection
8417 EVT CanonicalVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
8418 SDValue Result = DAG.getNode(ARMISD::VBSL, dl, CanonicalVT,
8422 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
8428 // Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when
8431 // BFI is only available on V6T2+
8432 if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops())
8436 // 1) or (and A, mask), val => ARMbfi A, val, mask
8437 // iff (val & mask) == val
8439 // 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
8440 // 2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2)
8441 // && mask == ~mask2
8442 // 2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2)
8443 // && ~mask == mask2
8444 // (i.e., copy a bitfield value into another bitfield of the same width)
8449 SDValue N00 = N0.getOperand(0);
8451 // The value and the mask need to be constants so we can verify this is
8452 // actually a bitfield set. If the mask is 0xffff, we can do better
8453 // via a movt instruction, so don't use BFI in that case.
8454 SDValue MaskOp = N0.getOperand(1);
8455 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(MaskOp);
8458 unsigned Mask = MaskC->getZExtValue();
8462 // Case (1): or (and A, mask), val => ARMbfi A, val, mask
8463 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
8465 unsigned Val = N1C->getZExtValue();
8466 if ((Val & ~Mask) != Val)
8469 if (ARM::isBitFieldInvertedMask(Mask)) {
8470 Val >>= countTrailingZeros(~Mask);
8472 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00,
8473 DAG.getConstant(Val, MVT::i32),
8474 DAG.getConstant(Mask, MVT::i32));
8476 // Do not add new nodes to DAG combiner worklist.
8477 DCI.CombineTo(N, Res, false);
8480 } else if (N1.getOpcode() == ISD::AND) {
8481 // case (2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
8482 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
8485 unsigned Mask2 = N11C->getZExtValue();
8487 // Mask and ~Mask2 (or reverse) must be equivalent for the BFI pattern
8489 if (ARM::isBitFieldInvertedMask(Mask) &&
8491 // The pack halfword instruction works better for masks that fit it,
8492 // so use that when it's available.
8493 if (Subtarget->hasT2ExtractPack() &&
8494 (Mask == 0xffff || Mask == 0xffff0000))
8497 unsigned amt = countTrailingZeros(Mask2);
8498 Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0),
8499 DAG.getConstant(amt, MVT::i32));
8500 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00, Res,
8501 DAG.getConstant(Mask, MVT::i32));
8502 // Do not add new nodes to DAG combiner worklist.
8503 DCI.CombineTo(N, Res, false);
8505 } else if (ARM::isBitFieldInvertedMask(~Mask) &&
8507 // The pack halfword instruction works better for masks that fit it,
8508 // so use that when it's available.
8509 if (Subtarget->hasT2ExtractPack() &&
8510 (Mask2 == 0xffff || Mask2 == 0xffff0000))
8513 unsigned lsb = countTrailingZeros(Mask);
8514 Res = DAG.getNode(ISD::SRL, DL, VT, N00,
8515 DAG.getConstant(lsb, MVT::i32));
8516 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res,
8517 DAG.getConstant(Mask2, MVT::i32));
8518 // Do not add new nodes to DAG combiner worklist.
8519 DCI.CombineTo(N, Res, false);
8524 if (DAG.MaskedValueIsZero(N1, MaskC->getAPIntValue()) &&
8525 N00.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N00.getOperand(1)) &&
8526 ARM::isBitFieldInvertedMask(~Mask)) {
8527 // Case (3): or (and (shl A, #shamt), mask), B => ARMbfi B, A, ~mask
8528 // where lsb(mask) == #shamt and masked bits of B are known zero.
8529 SDValue ShAmt = N00.getOperand(1);
8530 unsigned ShAmtC = cast<ConstantSDNode>(ShAmt)->getZExtValue();
8531 unsigned LSB = countTrailingZeros(Mask);
8535 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1, N00.getOperand(0),
8536 DAG.getConstant(~Mask, MVT::i32));
8538 // Do not add new nodes to DAG combiner worklist.
8539 DCI.CombineTo(N, Res, false);
8545 static SDValue PerformXORCombine(SDNode *N,
8546 TargetLowering::DAGCombinerInfo &DCI,
8547 const ARMSubtarget *Subtarget) {
8548 EVT VT = N->getValueType(0);
8549 SelectionDAG &DAG = DCI.DAG;
8551 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8554 if (!Subtarget->isThumb1Only()) {
8555 // fold (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c))
8556 SDValue Result = combineSelectAndUseCommutative(N, false, DCI);
8557 if (Result.getNode())
8564 /// PerformBFICombine - (bfi A, (and B, Mask1), Mask2) -> (bfi A, B, Mask2) iff
8565 /// the bits being cleared by the AND are not demanded by the BFI.
8566 static SDValue PerformBFICombine(SDNode *N,
8567 TargetLowering::DAGCombinerInfo &DCI) {
8568 SDValue N1 = N->getOperand(1);
8569 if (N1.getOpcode() == ISD::AND) {
8570 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
8573 unsigned InvMask = cast<ConstantSDNode>(N->getOperand(2))->getZExtValue();
8574 unsigned LSB = countTrailingZeros(~InvMask);
8575 unsigned Width = (32 - countLeadingZeros(~InvMask)) - LSB;
8577 static_cast<unsigned>(std::numeric_limits<unsigned>::digits) &&
8578 "undefined behavior");
8579 unsigned Mask = (1u << Width) - 1;
8580 unsigned Mask2 = N11C->getZExtValue();
8581 if ((Mask & (~Mask2)) == 0)
8582 return DCI.DAG.getNode(ARMISD::BFI, SDLoc(N), N->getValueType(0),
8583 N->getOperand(0), N1.getOperand(0),
8589 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
8590 /// ARMISD::VMOVRRD.
8591 static SDValue PerformVMOVRRDCombine(SDNode *N,
8592 TargetLowering::DAGCombinerInfo &DCI,
8593 const ARMSubtarget *Subtarget) {
8594 // vmovrrd(vmovdrr x, y) -> x,y
8595 SDValue InDouble = N->getOperand(0);
8596 if (InDouble.getOpcode() == ARMISD::VMOVDRR && !Subtarget->isFPOnlySP())
8597 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
8599 // vmovrrd(load f64) -> (load i32), (load i32)
8600 SDNode *InNode = InDouble.getNode();
8601 if (ISD::isNormalLoad(InNode) && InNode->hasOneUse() &&
8602 InNode->getValueType(0) == MVT::f64 &&
8603 InNode->getOperand(1).getOpcode() == ISD::FrameIndex &&
8604 !cast<LoadSDNode>(InNode)->isVolatile()) {
8605 // TODO: Should this be done for non-FrameIndex operands?
8606 LoadSDNode *LD = cast<LoadSDNode>(InNode);
8608 SelectionDAG &DAG = DCI.DAG;
8610 SDValue BasePtr = LD->getBasePtr();
8611 SDValue NewLD1 = DAG.getLoad(MVT::i32, DL, LD->getChain(), BasePtr,
8612 LD->getPointerInfo(), LD->isVolatile(),
8613 LD->isNonTemporal(), LD->isInvariant(),
8614 LD->getAlignment());
8616 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
8617 DAG.getConstant(4, MVT::i32));
8618 SDValue NewLD2 = DAG.getLoad(MVT::i32, DL, NewLD1.getValue(1), OffsetPtr,
8619 LD->getPointerInfo(), LD->isVolatile(),
8620 LD->isNonTemporal(), LD->isInvariant(),
8621 std::min(4U, LD->getAlignment() / 2));
8623 DAG.ReplaceAllUsesOfValueWith(SDValue(LD, 1), NewLD2.getValue(1));
8624 if (DCI.DAG.getTargetLoweringInfo().isBigEndian())
8625 std::swap (NewLD1, NewLD2);
8626 SDValue Result = DCI.CombineTo(N, NewLD1, NewLD2);
8633 /// PerformVMOVDRRCombine - Target-specific dag combine xforms for
8634 /// ARMISD::VMOVDRR. This is also used for BUILD_VECTORs with 2 operands.
8635 static SDValue PerformVMOVDRRCombine(SDNode *N, SelectionDAG &DAG) {
8636 // N=vmovrrd(X); vmovdrr(N:0, N:1) -> bit_convert(X)
8637 SDValue Op0 = N->getOperand(0);
8638 SDValue Op1 = N->getOperand(1);
8639 if (Op0.getOpcode() == ISD::BITCAST)
8640 Op0 = Op0.getOperand(0);
8641 if (Op1.getOpcode() == ISD::BITCAST)
8642 Op1 = Op1.getOperand(0);
8643 if (Op0.getOpcode() == ARMISD::VMOVRRD &&
8644 Op0.getNode() == Op1.getNode() &&
8645 Op0.getResNo() == 0 && Op1.getResNo() == 1)
8646 return DAG.getNode(ISD::BITCAST, SDLoc(N),
8647 N->getValueType(0), Op0.getOperand(0));
8651 /// hasNormalLoadOperand - Check if any of the operands of a BUILD_VECTOR node
8652 /// are normal, non-volatile loads. If so, it is profitable to bitcast an
8653 /// i64 vector to have f64 elements, since the value can then be loaded
8654 /// directly into a VFP register.
8655 static bool hasNormalLoadOperand(SDNode *N) {
8656 unsigned NumElts = N->getValueType(0).getVectorNumElements();
8657 for (unsigned i = 0; i < NumElts; ++i) {
8658 SDNode *Elt = N->getOperand(i).getNode();
8659 if (ISD::isNormalLoad(Elt) && !cast<LoadSDNode>(Elt)->isVolatile())
8665 /// PerformBUILD_VECTORCombine - Target-specific dag combine xforms for
8666 /// ISD::BUILD_VECTOR.
8667 static SDValue PerformBUILD_VECTORCombine(SDNode *N,
8668 TargetLowering::DAGCombinerInfo &DCI,
8669 const ARMSubtarget *Subtarget) {
8670 // build_vector(N=ARMISD::VMOVRRD(X), N:1) -> bit_convert(X):
8671 // VMOVRRD is introduced when legalizing i64 types. It forces the i64 value
8672 // into a pair of GPRs, which is fine when the value is used as a scalar,
8673 // but if the i64 value is converted to a vector, we need to undo the VMOVRRD.
8674 SelectionDAG &DAG = DCI.DAG;
8675 if (N->getNumOperands() == 2) {
8676 SDValue RV = PerformVMOVDRRCombine(N, DAG);
8681 // Load i64 elements as f64 values so that type legalization does not split
8682 // them up into i32 values.
8683 EVT VT = N->getValueType(0);
8684 if (VT.getVectorElementType() != MVT::i64 || !hasNormalLoadOperand(N))
8687 SmallVector<SDValue, 8> Ops;
8688 unsigned NumElts = VT.getVectorNumElements();
8689 for (unsigned i = 0; i < NumElts; ++i) {
8690 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(i));
8692 // Make the DAGCombiner fold the bitcast.
8693 DCI.AddToWorklist(V.getNode());
8695 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, NumElts);
8696 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, FloatVT, Ops);
8697 return DAG.getNode(ISD::BITCAST, dl, VT, BV);
8700 /// \brief Target-specific dag combine xforms for ARMISD::BUILD_VECTOR.
8702 PerformARMBUILD_VECTORCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
8703 // ARMISD::BUILD_VECTOR is introduced when legalizing ISD::BUILD_VECTOR.
8704 // At that time, we may have inserted bitcasts from integer to float.
8705 // If these bitcasts have survived DAGCombine, change the lowering of this
8706 // BUILD_VECTOR in something more vector friendly, i.e., that does not
8707 // force to use floating point types.
8709 // Make sure we can change the type of the vector.
8710 // This is possible iff:
8711 // 1. The vector is only used in a bitcast to a integer type. I.e.,
8712 // 1.1. Vector is used only once.
8713 // 1.2. Use is a bit convert to an integer type.
8714 // 2. The size of its operands are 32-bits (64-bits are not legal).
8715 EVT VT = N->getValueType(0);
8716 EVT EltVT = VT.getVectorElementType();
8718 // Check 1.1. and 2.
8719 if (EltVT.getSizeInBits() != 32 || !N->hasOneUse())
8722 // By construction, the input type must be float.
8723 assert(EltVT == MVT::f32 && "Unexpected type!");
8726 SDNode *Use = *N->use_begin();
8727 if (Use->getOpcode() != ISD::BITCAST ||
8728 Use->getValueType(0).isFloatingPoint())
8731 // Check profitability.
8732 // Model is, if more than half of the relevant operands are bitcast from
8733 // i32, turn the build_vector into a sequence of insert_vector_elt.
8734 // Relevant operands are everything that is not statically
8735 // (i.e., at compile time) bitcasted.
8736 unsigned NumOfBitCastedElts = 0;
8737 unsigned NumElts = VT.getVectorNumElements();
8738 unsigned NumOfRelevantElts = NumElts;
8739 for (unsigned Idx = 0; Idx < NumElts; ++Idx) {
8740 SDValue Elt = N->getOperand(Idx);
8741 if (Elt->getOpcode() == ISD::BITCAST) {
8742 // Assume only bit cast to i32 will go away.
8743 if (Elt->getOperand(0).getValueType() == MVT::i32)
8744 ++NumOfBitCastedElts;
8745 } else if (Elt.getOpcode() == ISD::UNDEF || isa<ConstantSDNode>(Elt))
8746 // Constants are statically casted, thus do not count them as
8747 // relevant operands.
8748 --NumOfRelevantElts;
8751 // Check if more than half of the elements require a non-free bitcast.
8752 if (NumOfBitCastedElts <= NumOfRelevantElts / 2)
8755 SelectionDAG &DAG = DCI.DAG;
8756 // Create the new vector type.
8757 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
8758 // Check if the type is legal.
8759 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8760 if (!TLI.isTypeLegal(VecVT))
8764 // ARMISD::BUILD_VECTOR E1, E2, ..., EN.
8765 // => BITCAST INSERT_VECTOR_ELT
8766 // (INSERT_VECTOR_ELT (...), (BITCAST EN-1), N-1),
8768 SDValue Vec = DAG.getUNDEF(VecVT);
8770 for (unsigned Idx = 0 ; Idx < NumElts; ++Idx) {
8771 SDValue V = N->getOperand(Idx);
8772 if (V.getOpcode() == ISD::UNDEF)
8774 if (V.getOpcode() == ISD::BITCAST &&
8775 V->getOperand(0).getValueType() == MVT::i32)
8776 // Fold obvious case.
8777 V = V.getOperand(0);
8779 V = DAG.getNode(ISD::BITCAST, SDLoc(V), MVT::i32, V);
8780 // Make the DAGCombiner fold the bitcasts.
8781 DCI.AddToWorklist(V.getNode());
8783 SDValue LaneIdx = DAG.getConstant(Idx, MVT::i32);
8784 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VecVT, Vec, V, LaneIdx);
8786 Vec = DAG.getNode(ISD::BITCAST, dl, VT, Vec);
8787 // Make the DAGCombiner fold the bitcasts.
8788 DCI.AddToWorklist(Vec.getNode());
8792 /// PerformInsertEltCombine - Target-specific dag combine xforms for
8793 /// ISD::INSERT_VECTOR_ELT.
8794 static SDValue PerformInsertEltCombine(SDNode *N,
8795 TargetLowering::DAGCombinerInfo &DCI) {
8796 // Bitcast an i64 load inserted into a vector to f64.
8797 // Otherwise, the i64 value will be legalized to a pair of i32 values.
8798 EVT VT = N->getValueType(0);
8799 SDNode *Elt = N->getOperand(1).getNode();
8800 if (VT.getVectorElementType() != MVT::i64 ||
8801 !ISD::isNormalLoad(Elt) || cast<LoadSDNode>(Elt)->isVolatile())
8804 SelectionDAG &DAG = DCI.DAG;
8806 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
8807 VT.getVectorNumElements());
8808 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, N->getOperand(0));
8809 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(1));
8810 // Make the DAGCombiner fold the bitcasts.
8811 DCI.AddToWorklist(Vec.getNode());
8812 DCI.AddToWorklist(V.getNode());
8813 SDValue InsElt = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, FloatVT,
8814 Vec, V, N->getOperand(2));
8815 return DAG.getNode(ISD::BITCAST, dl, VT, InsElt);
8818 /// PerformVECTOR_SHUFFLECombine - Target-specific dag combine xforms for
8819 /// ISD::VECTOR_SHUFFLE.
8820 static SDValue PerformVECTOR_SHUFFLECombine(SDNode *N, SelectionDAG &DAG) {
8821 // The LLVM shufflevector instruction does not require the shuffle mask
8822 // length to match the operand vector length, but ISD::VECTOR_SHUFFLE does
8823 // have that requirement. When translating to ISD::VECTOR_SHUFFLE, if the
8824 // operands do not match the mask length, they are extended by concatenating
8825 // them with undef vectors. That is probably the right thing for other
8826 // targets, but for NEON it is better to concatenate two double-register
8827 // size vector operands into a single quad-register size vector. Do that
8828 // transformation here:
8829 // shuffle(concat(v1, undef), concat(v2, undef)) ->
8830 // shuffle(concat(v1, v2), undef)
8831 SDValue Op0 = N->getOperand(0);
8832 SDValue Op1 = N->getOperand(1);
8833 if (Op0.getOpcode() != ISD::CONCAT_VECTORS ||
8834 Op1.getOpcode() != ISD::CONCAT_VECTORS ||
8835 Op0.getNumOperands() != 2 ||
8836 Op1.getNumOperands() != 2)
8838 SDValue Concat0Op1 = Op0.getOperand(1);
8839 SDValue Concat1Op1 = Op1.getOperand(1);
8840 if (Concat0Op1.getOpcode() != ISD::UNDEF ||
8841 Concat1Op1.getOpcode() != ISD::UNDEF)
8843 // Skip the transformation if any of the types are illegal.
8844 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8845 EVT VT = N->getValueType(0);
8846 if (!TLI.isTypeLegal(VT) ||
8847 !TLI.isTypeLegal(Concat0Op1.getValueType()) ||
8848 !TLI.isTypeLegal(Concat1Op1.getValueType()))
8851 SDValue NewConcat = DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT,
8852 Op0.getOperand(0), Op1.getOperand(0));
8853 // Translate the shuffle mask.
8854 SmallVector<int, 16> NewMask;
8855 unsigned NumElts = VT.getVectorNumElements();
8856 unsigned HalfElts = NumElts/2;
8857 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
8858 for (unsigned n = 0; n < NumElts; ++n) {
8859 int MaskElt = SVN->getMaskElt(n);
8861 if (MaskElt < (int)HalfElts)
8863 else if (MaskElt >= (int)NumElts && MaskElt < (int)(NumElts + HalfElts))
8864 NewElt = HalfElts + MaskElt - NumElts;
8865 NewMask.push_back(NewElt);
8867 return DAG.getVectorShuffle(VT, SDLoc(N), NewConcat,
8868 DAG.getUNDEF(VT), NewMask.data());
8871 /// CombineBaseUpdate - Target-specific DAG combine function for VLDDUP,
8872 /// NEON load/store intrinsics, and generic vector load/stores, to merge
8873 /// base address updates.
8874 /// For generic load/stores, the memory type is assumed to be a vector.
8875 /// The caller is assumed to have checked legality.
8876 static SDValue CombineBaseUpdate(SDNode *N,
8877 TargetLowering::DAGCombinerInfo &DCI) {
8878 SelectionDAG &DAG = DCI.DAG;
8879 bool isIntrinsic = (N->getOpcode() == ISD::INTRINSIC_VOID ||
8880 N->getOpcode() == ISD::INTRINSIC_W_CHAIN);
8881 bool isStore = N->getOpcode() == ISD::STORE;
8882 unsigned AddrOpIdx = ((isIntrinsic || isStore) ? 2 : 1);
8883 SDValue Addr = N->getOperand(AddrOpIdx);
8885 // Search for a use of the address operand that is an increment.
8886 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(),
8887 UE = Addr.getNode()->use_end(); UI != UE; ++UI) {
8889 if (User->getOpcode() != ISD::ADD ||
8890 UI.getUse().getResNo() != Addr.getResNo())
8893 // Check that the add is independent of the load/store. Otherwise, folding
8894 // it would create a cycle.
8895 if (User->isPredecessorOf(N) || N->isPredecessorOf(User))
8898 // Find the new opcode for the updating load/store.
8900 bool isLaneOp = false;
8901 unsigned NewOpc = 0;
8902 unsigned NumVecs = 0;
8904 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
8906 default: llvm_unreachable("unexpected intrinsic for Neon base update");
8907 case Intrinsic::arm_neon_vld1: NewOpc = ARMISD::VLD1_UPD;
8909 case Intrinsic::arm_neon_vld2: NewOpc = ARMISD::VLD2_UPD;
8911 case Intrinsic::arm_neon_vld3: NewOpc = ARMISD::VLD3_UPD;
8913 case Intrinsic::arm_neon_vld4: NewOpc = ARMISD::VLD4_UPD;
8915 case Intrinsic::arm_neon_vld2lane: NewOpc = ARMISD::VLD2LN_UPD;
8916 NumVecs = 2; isLaneOp = true; break;
8917 case Intrinsic::arm_neon_vld3lane: NewOpc = ARMISD::VLD3LN_UPD;
8918 NumVecs = 3; isLaneOp = true; break;
8919 case Intrinsic::arm_neon_vld4lane: NewOpc = ARMISD::VLD4LN_UPD;
8920 NumVecs = 4; isLaneOp = true; break;
8921 case Intrinsic::arm_neon_vst1: NewOpc = ARMISD::VST1_UPD;
8922 NumVecs = 1; isLoad = false; break;
8923 case Intrinsic::arm_neon_vst2: NewOpc = ARMISD::VST2_UPD;
8924 NumVecs = 2; isLoad = false; break;
8925 case Intrinsic::arm_neon_vst3: NewOpc = ARMISD::VST3_UPD;
8926 NumVecs = 3; isLoad = false; break;
8927 case Intrinsic::arm_neon_vst4: NewOpc = ARMISD::VST4_UPD;
8928 NumVecs = 4; isLoad = false; break;
8929 case Intrinsic::arm_neon_vst2lane: NewOpc = ARMISD::VST2LN_UPD;
8930 NumVecs = 2; isLoad = false; isLaneOp = true; break;
8931 case Intrinsic::arm_neon_vst3lane: NewOpc = ARMISD::VST3LN_UPD;
8932 NumVecs = 3; isLoad = false; isLaneOp = true; break;
8933 case Intrinsic::arm_neon_vst4lane: NewOpc = ARMISD::VST4LN_UPD;
8934 NumVecs = 4; isLoad = false; isLaneOp = true; break;
8938 switch (N->getOpcode()) {
8939 default: llvm_unreachable("unexpected opcode for Neon base update");
8940 case ARMISD::VLD2DUP: NewOpc = ARMISD::VLD2DUP_UPD; NumVecs = 2; break;
8941 case ARMISD::VLD3DUP: NewOpc = ARMISD::VLD3DUP_UPD; NumVecs = 3; break;
8942 case ARMISD::VLD4DUP: NewOpc = ARMISD::VLD4DUP_UPD; NumVecs = 4; break;
8943 case ISD::LOAD: NewOpc = ARMISD::VLD1_UPD;
8944 NumVecs = 1; isLaneOp = false; break;
8945 case ISD::STORE: NewOpc = ARMISD::VST1_UPD;
8946 NumVecs = 1; isLoad = false; isLaneOp = false; break;
8950 // Find the size of memory referenced by the load/store.
8953 VecTy = N->getValueType(0);
8954 else if (isIntrinsic)
8955 VecTy = N->getOperand(AddrOpIdx+1).getValueType();
8957 VecTy = N->getOperand(1).getValueType();
8959 unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8;
8961 NumBytes /= VecTy.getVectorNumElements();
8963 // If the increment is a constant, it must match the memory ref size.
8964 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
8965 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
8966 uint64_t IncVal = CInc->getZExtValue();
8967 if (IncVal != NumBytes)
8969 } else if (NumBytes >= 3 * 16) {
8970 // VLD3/4 and VST3/4 for 128-bit vectors are implemented with two
8971 // separate instructions that make it harder to use a non-constant update.
8975 EVT AlignedVecTy = VecTy;
8977 // If this is a less-than-standard-aligned load/store, change the type to
8978 // match the standard alignment.
8979 // The alignment is overlooked when selecting _UPD variants; and it's
8980 // easier to introduce bitcasts here than fix that.
8981 // There are 3 ways to get to this base-update combine:
8982 // - intrinsics: they are assumed to be properly aligned (to the standard
8983 // alignment of the memory type), so we don't need to do anything.
8984 // - ARMISD::VLDx nodes: they are only generated from the aforementioned
8985 // intrinsics, so, likewise, there's nothing to do.
8986 // - generic load/store instructions: the alignment is specified as an
8987 // explicit operand, rather than implicitly as the standard alignment
8988 // of the memory type (like the intrisics). We need to change the
8989 // memory type to match the explicit alignment. That way, we don't
8990 // generate non-standard-aligned ARMISD::VLDx nodes.
8991 if (LSBaseSDNode *LSN = dyn_cast<LSBaseSDNode>(N)) {
8992 unsigned Alignment = LSN->getAlignment();
8995 if (Alignment < VecTy.getScalarSizeInBits() / 8) {
8996 MVT EltTy = MVT::getIntegerVT(Alignment * 8);
8997 assert(NumVecs == 1 && "Unexpected multi-element generic load/store.");
8998 assert(!isLaneOp && "Unexpected generic load/store lane.");
8999 unsigned NumElts = NumBytes / (EltTy.getSizeInBits() / 8);
9000 AlignedVecTy = MVT::getVectorVT(EltTy, NumElts);
9004 // Create the new updating load/store node.
9005 // First, create an SDVTList for the new updating node's results.
9007 unsigned NumResultVecs = (isLoad ? NumVecs : 0);
9009 for (n = 0; n < NumResultVecs; ++n)
9010 Tys[n] = AlignedVecTy;
9011 Tys[n++] = MVT::i32;
9012 Tys[n] = MVT::Other;
9013 SDVTList SDTys = DAG.getVTList(makeArrayRef(Tys, NumResultVecs+2));
9015 // Then, gather the new node's operands.
9016 SmallVector<SDValue, 8> Ops;
9017 Ops.push_back(N->getOperand(0)); // incoming chain
9018 Ops.push_back(N->getOperand(AddrOpIdx));
9020 if (StoreSDNode *StN = dyn_cast<StoreSDNode>(N)) {
9021 // Try to match the intrinsic's signature
9022 Ops.push_back(StN->getValue());
9023 Ops.push_back(DAG.getConstant(StN->getAlignment(), MVT::i32));
9025 for (unsigned i = AddrOpIdx + 1; i < N->getNumOperands(); ++i)
9026 Ops.push_back(N->getOperand(i));
9029 // If this is a non-standard-aligned store, the penultimate operand is the
9030 // stored value. Bitcast it to the aligned type.
9031 if (AlignedVecTy != VecTy && N->getOpcode() == ISD::STORE) {
9032 SDValue &StVal = Ops[Ops.size()-2];
9033 StVal = DAG.getNode(ISD::BITCAST, SDLoc(N), AlignedVecTy, StVal);
9036 MemSDNode *MemInt = cast<MemSDNode>(N);
9037 SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, SDLoc(N), SDTys,
9039 MemInt->getMemOperand());
9042 std::vector<SDValue> NewResults;
9043 for (unsigned i = 0; i < NumResultVecs; ++i) {
9044 NewResults.push_back(SDValue(UpdN.getNode(), i));
9047 // If this is an non-standard-aligned load, the first result is the loaded
9048 // value. Bitcast it to the expected result type.
9049 if (AlignedVecTy != VecTy && N->getOpcode() == ISD::LOAD) {
9050 SDValue &LdVal = NewResults[0];
9051 LdVal = DAG.getNode(ISD::BITCAST, SDLoc(N), VecTy, LdVal);
9054 NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs+1)); // chain
9055 DCI.CombineTo(N, NewResults);
9056 DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs));
9063 static SDValue PerformVLDCombine(SDNode *N,
9064 TargetLowering::DAGCombinerInfo &DCI) {
9065 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
9068 return CombineBaseUpdate(N, DCI);
9071 /// CombineVLDDUP - For a VDUPLANE node N, check if its source operand is a
9072 /// vldN-lane (N > 1) intrinsic, and if all the other uses of that intrinsic
9073 /// are also VDUPLANEs. If so, combine them to a vldN-dup operation and
9075 static bool CombineVLDDUP(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
9076 SelectionDAG &DAG = DCI.DAG;
9077 EVT VT = N->getValueType(0);
9078 // vldN-dup instructions only support 64-bit vectors for N > 1.
9079 if (!VT.is64BitVector())
9082 // Check if the VDUPLANE operand is a vldN-dup intrinsic.
9083 SDNode *VLD = N->getOperand(0).getNode();
9084 if (VLD->getOpcode() != ISD::INTRINSIC_W_CHAIN)
9086 unsigned NumVecs = 0;
9087 unsigned NewOpc = 0;
9088 unsigned IntNo = cast<ConstantSDNode>(VLD->getOperand(1))->getZExtValue();
9089 if (IntNo == Intrinsic::arm_neon_vld2lane) {
9091 NewOpc = ARMISD::VLD2DUP;
9092 } else if (IntNo == Intrinsic::arm_neon_vld3lane) {
9094 NewOpc = ARMISD::VLD3DUP;
9095 } else if (IntNo == Intrinsic::arm_neon_vld4lane) {
9097 NewOpc = ARMISD::VLD4DUP;
9102 // First check that all the vldN-lane uses are VDUPLANEs and that the lane
9103 // numbers match the load.
9104 unsigned VLDLaneNo =
9105 cast<ConstantSDNode>(VLD->getOperand(NumVecs+3))->getZExtValue();
9106 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
9108 // Ignore uses of the chain result.
9109 if (UI.getUse().getResNo() == NumVecs)
9112 if (User->getOpcode() != ARMISD::VDUPLANE ||
9113 VLDLaneNo != cast<ConstantSDNode>(User->getOperand(1))->getZExtValue())
9117 // Create the vldN-dup node.
9120 for (n = 0; n < NumVecs; ++n)
9122 Tys[n] = MVT::Other;
9123 SDVTList SDTys = DAG.getVTList(makeArrayRef(Tys, NumVecs+1));
9124 SDValue Ops[] = { VLD->getOperand(0), VLD->getOperand(2) };
9125 MemIntrinsicSDNode *VLDMemInt = cast<MemIntrinsicSDNode>(VLD);
9126 SDValue VLDDup = DAG.getMemIntrinsicNode(NewOpc, SDLoc(VLD), SDTys,
9127 Ops, VLDMemInt->getMemoryVT(),
9128 VLDMemInt->getMemOperand());
9131 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
9133 unsigned ResNo = UI.getUse().getResNo();
9134 // Ignore uses of the chain result.
9135 if (ResNo == NumVecs)
9138 DCI.CombineTo(User, SDValue(VLDDup.getNode(), ResNo));
9141 // Now the vldN-lane intrinsic is dead except for its chain result.
9142 // Update uses of the chain.
9143 std::vector<SDValue> VLDDupResults;
9144 for (unsigned n = 0; n < NumVecs; ++n)
9145 VLDDupResults.push_back(SDValue(VLDDup.getNode(), n));
9146 VLDDupResults.push_back(SDValue(VLDDup.getNode(), NumVecs));
9147 DCI.CombineTo(VLD, VLDDupResults);
9152 /// PerformVDUPLANECombine - Target-specific dag combine xforms for
9153 /// ARMISD::VDUPLANE.
9154 static SDValue PerformVDUPLANECombine(SDNode *N,
9155 TargetLowering::DAGCombinerInfo &DCI) {
9156 SDValue Op = N->getOperand(0);
9158 // If the source is a vldN-lane (N > 1) intrinsic, and all the other uses
9159 // of that intrinsic are also VDUPLANEs, combine them to a vldN-dup operation.
9160 if (CombineVLDDUP(N, DCI))
9161 return SDValue(N, 0);
9163 // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is
9164 // redundant. Ignore bit_converts for now; element sizes are checked below.
9165 while (Op.getOpcode() == ISD::BITCAST)
9166 Op = Op.getOperand(0);
9167 if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM)
9170 // Make sure the VMOV element size is not bigger than the VDUPLANE elements.
9171 unsigned EltSize = Op.getValueType().getVectorElementType().getSizeInBits();
9172 // The canonical VMOV for a zero vector uses a 32-bit element size.
9173 unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
9175 if (ARM_AM::decodeNEONModImm(Imm, EltBits) == 0)
9177 EVT VT = N->getValueType(0);
9178 if (EltSize > VT.getVectorElementType().getSizeInBits())
9181 return DCI.DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Op);
9184 static SDValue PerformLOADCombine(SDNode *N,
9185 TargetLowering::DAGCombinerInfo &DCI) {
9186 EVT VT = N->getValueType(0);
9188 // If this is a legal vector load, try to combine it into a VLD1_UPD.
9189 if (ISD::isNormalLoad(N) && VT.isVector() &&
9190 DCI.DAG.getTargetLoweringInfo().isTypeLegal(VT))
9191 return CombineBaseUpdate(N, DCI);
9196 /// PerformSTORECombine - Target-specific dag combine xforms for
9198 static SDValue PerformSTORECombine(SDNode *N,
9199 TargetLowering::DAGCombinerInfo &DCI) {
9200 StoreSDNode *St = cast<StoreSDNode>(N);
9201 if (St->isVolatile())
9204 // Optimize trunc store (of multiple scalars) to shuffle and store. First,
9205 // pack all of the elements in one place. Next, store to memory in fewer
9207 SDValue StVal = St->getValue();
9208 EVT VT = StVal.getValueType();
9209 if (St->isTruncatingStore() && VT.isVector()) {
9210 SelectionDAG &DAG = DCI.DAG;
9211 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9212 EVT StVT = St->getMemoryVT();
9213 unsigned NumElems = VT.getVectorNumElements();
9214 assert(StVT != VT && "Cannot truncate to the same type");
9215 unsigned FromEltSz = VT.getVectorElementType().getSizeInBits();
9216 unsigned ToEltSz = StVT.getVectorElementType().getSizeInBits();
9218 // From, To sizes and ElemCount must be pow of two
9219 if (!isPowerOf2_32(NumElems * FromEltSz * ToEltSz)) return SDValue();
9221 // We are going to use the original vector elt for storing.
9222 // Accumulated smaller vector elements must be a multiple of the store size.
9223 if (0 != (NumElems * FromEltSz) % ToEltSz) return SDValue();
9225 unsigned SizeRatio = FromEltSz / ToEltSz;
9226 assert(SizeRatio * NumElems * ToEltSz == VT.getSizeInBits());
9228 // Create a type on which we perform the shuffle.
9229 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(), StVT.getScalarType(),
9230 NumElems*SizeRatio);
9231 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
9234 SDValue WideVec = DAG.getNode(ISD::BITCAST, DL, WideVecVT, StVal);
9235 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
9236 for (unsigned i = 0; i < NumElems; ++i)
9237 ShuffleVec[i] = TLI.isBigEndian() ? (i+1) * SizeRatio - 1 : i * SizeRatio;
9239 // Can't shuffle using an illegal type.
9240 if (!TLI.isTypeLegal(WideVecVT)) return SDValue();
9242 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, DL, WideVec,
9243 DAG.getUNDEF(WideVec.getValueType()),
9245 // At this point all of the data is stored at the bottom of the
9246 // register. We now need to save it to mem.
9248 // Find the largest store unit
9249 MVT StoreType = MVT::i8;
9250 for (MVT Tp : MVT::integer_valuetypes()) {
9251 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToEltSz)
9254 // Didn't find a legal store type.
9255 if (!TLI.isTypeLegal(StoreType))
9258 // Bitcast the original vector into a vector of store-size units
9259 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
9260 StoreType, VT.getSizeInBits()/EVT(StoreType).getSizeInBits());
9261 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
9262 SDValue ShuffWide = DAG.getNode(ISD::BITCAST, DL, StoreVecVT, Shuff);
9263 SmallVector<SDValue, 8> Chains;
9264 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits()/8,
9265 TLI.getPointerTy());
9266 SDValue BasePtr = St->getBasePtr();
9268 // Perform one or more big stores into memory.
9269 unsigned E = (ToEltSz*NumElems)/StoreType.getSizeInBits();
9270 for (unsigned I = 0; I < E; I++) {
9271 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
9272 StoreType, ShuffWide,
9273 DAG.getIntPtrConstant(I));
9274 SDValue Ch = DAG.getStore(St->getChain(), DL, SubVec, BasePtr,
9275 St->getPointerInfo(), St->isVolatile(),
9276 St->isNonTemporal(), St->getAlignment());
9277 BasePtr = DAG.getNode(ISD::ADD, DL, BasePtr.getValueType(), BasePtr,
9279 Chains.push_back(Ch);
9281 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
9284 if (!ISD::isNormalStore(St))
9287 // Split a store of a VMOVDRR into two integer stores to avoid mixing NEON and
9288 // ARM stores of arguments in the same cache line.
9289 if (StVal.getNode()->getOpcode() == ARMISD::VMOVDRR &&
9290 StVal.getNode()->hasOneUse()) {
9291 SelectionDAG &DAG = DCI.DAG;
9292 bool isBigEndian = DAG.getTargetLoweringInfo().isBigEndian();
9294 SDValue BasePtr = St->getBasePtr();
9295 SDValue NewST1 = DAG.getStore(St->getChain(), DL,
9296 StVal.getNode()->getOperand(isBigEndian ? 1 : 0 ),
9297 BasePtr, St->getPointerInfo(), St->isVolatile(),
9298 St->isNonTemporal(), St->getAlignment());
9300 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
9301 DAG.getConstant(4, MVT::i32));
9302 return DAG.getStore(NewST1.getValue(0), DL,
9303 StVal.getNode()->getOperand(isBigEndian ? 0 : 1),
9304 OffsetPtr, St->getPointerInfo(), St->isVolatile(),
9305 St->isNonTemporal(),
9306 std::min(4U, St->getAlignment() / 2));
9309 if (StVal.getValueType() == MVT::i64 &&
9310 StVal.getNode()->getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
9312 // Bitcast an i64 store extracted from a vector to f64.
9313 // Otherwise, the i64 value will be legalized to a pair of i32 values.
9314 SelectionDAG &DAG = DCI.DAG;
9316 SDValue IntVec = StVal.getOperand(0);
9317 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
9318 IntVec.getValueType().getVectorNumElements());
9319 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, IntVec);
9320 SDValue ExtElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
9321 Vec, StVal.getOperand(1));
9323 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ExtElt);
9324 // Make the DAGCombiner fold the bitcasts.
9325 DCI.AddToWorklist(Vec.getNode());
9326 DCI.AddToWorklist(ExtElt.getNode());
9327 DCI.AddToWorklist(V.getNode());
9328 return DAG.getStore(St->getChain(), dl, V, St->getBasePtr(),
9329 St->getPointerInfo(), St->isVolatile(),
9330 St->isNonTemporal(), St->getAlignment(),
9334 // If this is a legal vector store, try to combine it into a VST1_UPD.
9335 if (ISD::isNormalStore(N) && VT.isVector() &&
9336 DCI.DAG.getTargetLoweringInfo().isTypeLegal(VT))
9337 return CombineBaseUpdate(N, DCI);
9342 // isConstVecPow2 - Return true if each vector element is a power of 2, all
9343 // elements are the same constant, C, and Log2(C) ranges from 1 to 32.
9344 static bool isConstVecPow2(SDValue ConstVec, bool isSigned, uint64_t &C)
9348 for (unsigned I = 0, E = ConstVec.getValueType().getVectorNumElements();
9350 ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(ConstVec.getOperand(I));
9355 APFloat APF = C->getValueAPF();
9356 if (APF.convertToInteger(&cN, 64, isSigned, APFloat::rmTowardZero, &isExact)
9357 != APFloat::opOK || !isExact)
9360 c0 = (I == 0) ? cN : c0;
9361 if (!isPowerOf2_64(cN) || c0 != cN || Log2_64(c0) < 1 || Log2_64(c0) > 32)
9368 /// PerformVCVTCombine - VCVT (floating-point to fixed-point, Advanced SIMD)
9369 /// can replace combinations of VMUL and VCVT (floating-point to integer)
9370 /// when the VMUL has a constant operand that is a power of 2.
9372 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
9373 /// vmul.f32 d16, d17, d16
9374 /// vcvt.s32.f32 d16, d16
9376 /// vcvt.s32.f32 d16, d16, #3
9377 static SDValue PerformVCVTCombine(SDNode *N,
9378 TargetLowering::DAGCombinerInfo &DCI,
9379 const ARMSubtarget *Subtarget) {
9380 SelectionDAG &DAG = DCI.DAG;
9381 SDValue Op = N->getOperand(0);
9383 if (!Subtarget->hasNEON() || !Op.getValueType().isVector() ||
9384 Op.getOpcode() != ISD::FMUL)
9388 SDValue N0 = Op->getOperand(0);
9389 SDValue ConstVec = Op->getOperand(1);
9390 bool isSigned = N->getOpcode() == ISD::FP_TO_SINT;
9392 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
9393 !isConstVecPow2(ConstVec, isSigned, C))
9396 MVT FloatTy = Op.getSimpleValueType().getVectorElementType();
9397 MVT IntTy = N->getSimpleValueType(0).getVectorElementType();
9398 unsigned NumLanes = Op.getValueType().getVectorNumElements();
9399 if (FloatTy.getSizeInBits() != 32 || IntTy.getSizeInBits() > 32 ||
9401 // These instructions only exist converting from f32 to i32. We can handle
9402 // smaller integers by generating an extra truncate, but larger ones would
9403 // be lossy. We also can't handle more then 4 lanes, since these intructions
9404 // only support v2i32/v4i32 types.
9408 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfp2fxs :
9409 Intrinsic::arm_neon_vcvtfp2fxu;
9410 SDValue FixConv = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N),
9411 NumLanes == 2 ? MVT::v2i32 : MVT::v4i32,
9412 DAG.getConstant(IntrinsicOpcode, MVT::i32), N0,
9413 DAG.getConstant(Log2_64(C), MVT::i32));
9415 if (IntTy.getSizeInBits() < FloatTy.getSizeInBits())
9416 FixConv = DAG.getNode(ISD::TRUNCATE, SDLoc(N), N->getValueType(0), FixConv);
9421 /// PerformVDIVCombine - VCVT (fixed-point to floating-point, Advanced SIMD)
9422 /// can replace combinations of VCVT (integer to floating-point) and VDIV
9423 /// when the VDIV has a constant operand that is a power of 2.
9425 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
9426 /// vcvt.f32.s32 d16, d16
9427 /// vdiv.f32 d16, d17, d16
9429 /// vcvt.f32.s32 d16, d16, #3
9430 static SDValue PerformVDIVCombine(SDNode *N,
9431 TargetLowering::DAGCombinerInfo &DCI,
9432 const ARMSubtarget *Subtarget) {
9433 SelectionDAG &DAG = DCI.DAG;
9434 SDValue Op = N->getOperand(0);
9435 unsigned OpOpcode = Op.getNode()->getOpcode();
9437 if (!Subtarget->hasNEON() || !N->getValueType(0).isVector() ||
9438 (OpOpcode != ISD::SINT_TO_FP && OpOpcode != ISD::UINT_TO_FP))
9442 SDValue ConstVec = N->getOperand(1);
9443 bool isSigned = OpOpcode == ISD::SINT_TO_FP;
9445 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
9446 !isConstVecPow2(ConstVec, isSigned, C))
9449 MVT FloatTy = N->getSimpleValueType(0).getVectorElementType();
9450 MVT IntTy = Op.getOperand(0).getSimpleValueType().getVectorElementType();
9451 if (FloatTy.getSizeInBits() != 32 || IntTy.getSizeInBits() > 32) {
9452 // These instructions only exist converting from i32 to f32. We can handle
9453 // smaller integers by generating an extra extend, but larger ones would
9458 SDValue ConvInput = Op.getOperand(0);
9459 unsigned NumLanes = Op.getValueType().getVectorNumElements();
9460 if (IntTy.getSizeInBits() < FloatTy.getSizeInBits())
9461 ConvInput = DAG.getNode(isSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND,
9462 SDLoc(N), NumLanes == 2 ? MVT::v2i32 : MVT::v4i32,
9465 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfxs2fp :
9466 Intrinsic::arm_neon_vcvtfxu2fp;
9467 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N),
9469 DAG.getConstant(IntrinsicOpcode, MVT::i32),
9470 ConvInput, DAG.getConstant(Log2_64(C), MVT::i32));
9473 /// Getvshiftimm - Check if this is a valid build_vector for the immediate
9474 /// operand of a vector shift operation, where all the elements of the
9475 /// build_vector must have the same constant integer value.
9476 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
9477 // Ignore bit_converts.
9478 while (Op.getOpcode() == ISD::BITCAST)
9479 Op = Op.getOperand(0);
9480 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
9481 APInt SplatBits, SplatUndef;
9482 unsigned SplatBitSize;
9484 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
9485 HasAnyUndefs, ElementBits) ||
9486 SplatBitSize > ElementBits)
9488 Cnt = SplatBits.getSExtValue();
9492 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
9493 /// operand of a vector shift left operation. That value must be in the range:
9494 /// 0 <= Value < ElementBits for a left shift; or
9495 /// 0 <= Value <= ElementBits for a long left shift.
9496 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
9497 assert(VT.isVector() && "vector shift count is not a vector type");
9498 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
9499 if (! getVShiftImm(Op, ElementBits, Cnt))
9501 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
9504 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
9505 /// operand of a vector shift right operation. For a shift opcode, the value
9506 /// is positive, but for an intrinsic the value count must be negative. The
9507 /// absolute value must be in the range:
9508 /// 1 <= |Value| <= ElementBits for a right shift; or
9509 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
9510 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
9512 assert(VT.isVector() && "vector shift count is not a vector type");
9513 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
9514 if (! getVShiftImm(Op, ElementBits, Cnt))
9518 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
9521 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
9522 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
9523 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
9526 // Don't do anything for most intrinsics.
9529 // Vector shifts: check for immediate versions and lower them.
9530 // Note: This is done during DAG combining instead of DAG legalizing because
9531 // the build_vectors for 64-bit vector element shift counts are generally
9532 // not legal, and it is hard to see their values after they get legalized to
9533 // loads from a constant pool.
9534 case Intrinsic::arm_neon_vshifts:
9535 case Intrinsic::arm_neon_vshiftu:
9536 case Intrinsic::arm_neon_vrshifts:
9537 case Intrinsic::arm_neon_vrshiftu:
9538 case Intrinsic::arm_neon_vrshiftn:
9539 case Intrinsic::arm_neon_vqshifts:
9540 case Intrinsic::arm_neon_vqshiftu:
9541 case Intrinsic::arm_neon_vqshiftsu:
9542 case Intrinsic::arm_neon_vqshiftns:
9543 case Intrinsic::arm_neon_vqshiftnu:
9544 case Intrinsic::arm_neon_vqshiftnsu:
9545 case Intrinsic::arm_neon_vqrshiftns:
9546 case Intrinsic::arm_neon_vqrshiftnu:
9547 case Intrinsic::arm_neon_vqrshiftnsu: {
9548 EVT VT = N->getOperand(1).getValueType();
9550 unsigned VShiftOpc = 0;
9553 case Intrinsic::arm_neon_vshifts:
9554 case Intrinsic::arm_neon_vshiftu:
9555 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
9556 VShiftOpc = ARMISD::VSHL;
9559 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
9560 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
9561 ARMISD::VSHRs : ARMISD::VSHRu);
9566 case Intrinsic::arm_neon_vrshifts:
9567 case Intrinsic::arm_neon_vrshiftu:
9568 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
9572 case Intrinsic::arm_neon_vqshifts:
9573 case Intrinsic::arm_neon_vqshiftu:
9574 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
9578 case Intrinsic::arm_neon_vqshiftsu:
9579 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
9581 llvm_unreachable("invalid shift count for vqshlu intrinsic");
9583 case Intrinsic::arm_neon_vrshiftn:
9584 case Intrinsic::arm_neon_vqshiftns:
9585 case Intrinsic::arm_neon_vqshiftnu:
9586 case Intrinsic::arm_neon_vqshiftnsu:
9587 case Intrinsic::arm_neon_vqrshiftns:
9588 case Intrinsic::arm_neon_vqrshiftnu:
9589 case Intrinsic::arm_neon_vqrshiftnsu:
9590 // Narrowing shifts require an immediate right shift.
9591 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
9593 llvm_unreachable("invalid shift count for narrowing vector shift "
9597 llvm_unreachable("unhandled vector shift");
9601 case Intrinsic::arm_neon_vshifts:
9602 case Intrinsic::arm_neon_vshiftu:
9603 // Opcode already set above.
9605 case Intrinsic::arm_neon_vrshifts:
9606 VShiftOpc = ARMISD::VRSHRs; break;
9607 case Intrinsic::arm_neon_vrshiftu:
9608 VShiftOpc = ARMISD::VRSHRu; break;
9609 case Intrinsic::arm_neon_vrshiftn:
9610 VShiftOpc = ARMISD::VRSHRN; break;
9611 case Intrinsic::arm_neon_vqshifts:
9612 VShiftOpc = ARMISD::VQSHLs; break;
9613 case Intrinsic::arm_neon_vqshiftu:
9614 VShiftOpc = ARMISD::VQSHLu; break;
9615 case Intrinsic::arm_neon_vqshiftsu:
9616 VShiftOpc = ARMISD::VQSHLsu; break;
9617 case Intrinsic::arm_neon_vqshiftns:
9618 VShiftOpc = ARMISD::VQSHRNs; break;
9619 case Intrinsic::arm_neon_vqshiftnu:
9620 VShiftOpc = ARMISD::VQSHRNu; break;
9621 case Intrinsic::arm_neon_vqshiftnsu:
9622 VShiftOpc = ARMISD::VQSHRNsu; break;
9623 case Intrinsic::arm_neon_vqrshiftns:
9624 VShiftOpc = ARMISD::VQRSHRNs; break;
9625 case Intrinsic::arm_neon_vqrshiftnu:
9626 VShiftOpc = ARMISD::VQRSHRNu; break;
9627 case Intrinsic::arm_neon_vqrshiftnsu:
9628 VShiftOpc = ARMISD::VQRSHRNsu; break;
9631 return DAG.getNode(VShiftOpc, SDLoc(N), N->getValueType(0),
9632 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
9635 case Intrinsic::arm_neon_vshiftins: {
9636 EVT VT = N->getOperand(1).getValueType();
9638 unsigned VShiftOpc = 0;
9640 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
9641 VShiftOpc = ARMISD::VSLI;
9642 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
9643 VShiftOpc = ARMISD::VSRI;
9645 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
9648 return DAG.getNode(VShiftOpc, SDLoc(N), N->getValueType(0),
9649 N->getOperand(1), N->getOperand(2),
9650 DAG.getConstant(Cnt, MVT::i32));
9653 case Intrinsic::arm_neon_vqrshifts:
9654 case Intrinsic::arm_neon_vqrshiftu:
9655 // No immediate versions of these to check for.
9662 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
9663 /// lowers them. As with the vector shift intrinsics, this is done during DAG
9664 /// combining instead of DAG legalizing because the build_vectors for 64-bit
9665 /// vector element shift counts are generally not legal, and it is hard to see
9666 /// their values after they get legalized to loads from a constant pool.
9667 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
9668 const ARMSubtarget *ST) {
9669 EVT VT = N->getValueType(0);
9670 if (N->getOpcode() == ISD::SRL && VT == MVT::i32 && ST->hasV6Ops()) {
9671 // Canonicalize (srl (bswap x), 16) to (rotr (bswap x), 16) if the high
9672 // 16-bits of x is zero. This optimizes rev + lsr 16 to rev16.
9673 SDValue N1 = N->getOperand(1);
9674 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) {
9675 SDValue N0 = N->getOperand(0);
9676 if (C->getZExtValue() == 16 && N0.getOpcode() == ISD::BSWAP &&
9677 DAG.MaskedValueIsZero(N0.getOperand(0),
9678 APInt::getHighBitsSet(32, 16)))
9679 return DAG.getNode(ISD::ROTR, SDLoc(N), VT, N0, N1);
9683 // Nothing to be done for scalar shifts.
9684 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9685 if (!VT.isVector() || !TLI.isTypeLegal(VT))
9688 assert(ST->hasNEON() && "unexpected vector shift");
9691 switch (N->getOpcode()) {
9692 default: llvm_unreachable("unexpected shift opcode");
9695 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
9696 return DAG.getNode(ARMISD::VSHL, SDLoc(N), VT, N->getOperand(0),
9697 DAG.getConstant(Cnt, MVT::i32));
9702 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
9703 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
9704 ARMISD::VSHRs : ARMISD::VSHRu);
9705 return DAG.getNode(VShiftOpc, SDLoc(N), VT, N->getOperand(0),
9706 DAG.getConstant(Cnt, MVT::i32));
9712 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
9713 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
9714 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
9715 const ARMSubtarget *ST) {
9716 SDValue N0 = N->getOperand(0);
9718 // Check for sign- and zero-extensions of vector extract operations of 8-
9719 // and 16-bit vector elements. NEON supports these directly. They are
9720 // handled during DAG combining because type legalization will promote them
9721 // to 32-bit types and it is messy to recognize the operations after that.
9722 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
9723 SDValue Vec = N0.getOperand(0);
9724 SDValue Lane = N0.getOperand(1);
9725 EVT VT = N->getValueType(0);
9726 EVT EltVT = N0.getValueType();
9727 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9729 if (VT == MVT::i32 &&
9730 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
9731 TLI.isTypeLegal(Vec.getValueType()) &&
9732 isa<ConstantSDNode>(Lane)) {
9735 switch (N->getOpcode()) {
9736 default: llvm_unreachable("unexpected opcode");
9737 case ISD::SIGN_EXTEND:
9738 Opc = ARMISD::VGETLANEs;
9740 case ISD::ZERO_EXTEND:
9741 case ISD::ANY_EXTEND:
9742 Opc = ARMISD::VGETLANEu;
9745 return DAG.getNode(Opc, SDLoc(N), VT, Vec, Lane);
9752 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
9753 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
9754 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
9755 const ARMSubtarget *ST) {
9756 // If the target supports NEON, try to use vmax/vmin instructions for f32
9757 // selects like "x < y ? x : y". Unless the NoNaNsFPMath option is set,
9758 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
9759 // a NaN; only do the transformation when it matches that behavior.
9761 // For now only do this when using NEON for FP operations; if using VFP, it
9762 // is not obvious that the benefit outweighs the cost of switching to the
9764 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
9765 N->getValueType(0) != MVT::f32)
9768 SDValue CondLHS = N->getOperand(0);
9769 SDValue CondRHS = N->getOperand(1);
9770 SDValue LHS = N->getOperand(2);
9771 SDValue RHS = N->getOperand(3);
9772 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
9774 unsigned Opcode = 0;
9776 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
9777 IsReversed = false; // x CC y ? x : y
9778 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
9779 IsReversed = true ; // x CC y ? y : x
9793 // If LHS is NaN, an ordered comparison will be false and the result will
9794 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
9795 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
9796 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
9797 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
9799 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
9800 // will return -0, so vmin can only be used for unsafe math or if one of
9801 // the operands is known to be nonzero.
9802 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
9803 !DAG.getTarget().Options.UnsafeFPMath &&
9804 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9806 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
9815 // If LHS is NaN, an ordered comparison will be false and the result will
9816 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
9817 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
9818 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
9819 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
9821 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
9822 // will return +0, so vmax can only be used for unsafe math or if one of
9823 // the operands is known to be nonzero.
9824 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
9825 !DAG.getTarget().Options.UnsafeFPMath &&
9826 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9828 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
9834 return DAG.getNode(Opcode, SDLoc(N), N->getValueType(0), LHS, RHS);
9837 /// PerformCMOVCombine - Target-specific DAG combining for ARMISD::CMOV.
9839 ARMTargetLowering::PerformCMOVCombine(SDNode *N, SelectionDAG &DAG) const {
9840 SDValue Cmp = N->getOperand(4);
9841 if (Cmp.getOpcode() != ARMISD::CMPZ)
9842 // Only looking at EQ and NE cases.
9845 EVT VT = N->getValueType(0);
9847 SDValue LHS = Cmp.getOperand(0);
9848 SDValue RHS = Cmp.getOperand(1);
9849 SDValue FalseVal = N->getOperand(0);
9850 SDValue TrueVal = N->getOperand(1);
9851 SDValue ARMcc = N->getOperand(2);
9852 ARMCC::CondCodes CC =
9853 (ARMCC::CondCodes)cast<ConstantSDNode>(ARMcc)->getZExtValue();
9871 /// FIXME: Turn this into a target neutral optimization?
9873 if (CC == ARMCC::NE && FalseVal == RHS && FalseVal != LHS) {
9874 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, TrueVal, ARMcc,
9875 N->getOperand(3), Cmp);
9876 } else if (CC == ARMCC::EQ && TrueVal == RHS) {
9878 SDValue NewCmp = getARMCmp(LHS, RHS, ISD::SETNE, ARMcc, DAG, dl);
9879 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, FalseVal, ARMcc,
9880 N->getOperand(3), NewCmp);
9883 if (Res.getNode()) {
9884 APInt KnownZero, KnownOne;
9885 DAG.computeKnownBits(SDValue(N,0), KnownZero, KnownOne);
9886 // Capture demanded bits information that would be otherwise lost.
9887 if (KnownZero == 0xfffffffe)
9888 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
9889 DAG.getValueType(MVT::i1));
9890 else if (KnownZero == 0xffffff00)
9891 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
9892 DAG.getValueType(MVT::i8));
9893 else if (KnownZero == 0xffff0000)
9894 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
9895 DAG.getValueType(MVT::i16));
9901 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
9902 DAGCombinerInfo &DCI) const {
9903 switch (N->getOpcode()) {
9905 case ISD::ADDC: return PerformADDCCombine(N, DCI, Subtarget);
9906 case ISD::ADD: return PerformADDCombine(N, DCI, Subtarget);
9907 case ISD::SUB: return PerformSUBCombine(N, DCI);
9908 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
9909 case ISD::OR: return PerformORCombine(N, DCI, Subtarget);
9910 case ISD::XOR: return PerformXORCombine(N, DCI, Subtarget);
9911 case ISD::AND: return PerformANDCombine(N, DCI, Subtarget);
9912 case ARMISD::BFI: return PerformBFICombine(N, DCI);
9913 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI, Subtarget);
9914 case ARMISD::VMOVDRR: return PerformVMOVDRRCombine(N, DCI.DAG);
9915 case ISD::STORE: return PerformSTORECombine(N, DCI);
9916 case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DCI, Subtarget);
9917 case ISD::INSERT_VECTOR_ELT: return PerformInsertEltCombine(N, DCI);
9918 case ISD::VECTOR_SHUFFLE: return PerformVECTOR_SHUFFLECombine(N, DCI.DAG);
9919 case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI);
9920 case ISD::FP_TO_SINT:
9921 case ISD::FP_TO_UINT: return PerformVCVTCombine(N, DCI, Subtarget);
9922 case ISD::FDIV: return PerformVDIVCombine(N, DCI, Subtarget);
9923 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
9926 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
9927 case ISD::SIGN_EXTEND:
9928 case ISD::ZERO_EXTEND:
9929 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
9930 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
9931 case ARMISD::CMOV: return PerformCMOVCombine(N, DCI.DAG);
9932 case ISD::LOAD: return PerformLOADCombine(N, DCI);
9933 case ARMISD::VLD2DUP:
9934 case ARMISD::VLD3DUP:
9935 case ARMISD::VLD4DUP:
9936 return PerformVLDCombine(N, DCI);
9937 case ARMISD::BUILD_VECTOR:
9938 return PerformARMBUILD_VECTORCombine(N, DCI);
9939 case ISD::INTRINSIC_VOID:
9940 case ISD::INTRINSIC_W_CHAIN:
9941 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
9942 case Intrinsic::arm_neon_vld1:
9943 case Intrinsic::arm_neon_vld2:
9944 case Intrinsic::arm_neon_vld3:
9945 case Intrinsic::arm_neon_vld4:
9946 case Intrinsic::arm_neon_vld2lane:
9947 case Intrinsic::arm_neon_vld3lane:
9948 case Intrinsic::arm_neon_vld4lane:
9949 case Intrinsic::arm_neon_vst1:
9950 case Intrinsic::arm_neon_vst2:
9951 case Intrinsic::arm_neon_vst3:
9952 case Intrinsic::arm_neon_vst4:
9953 case Intrinsic::arm_neon_vst2lane:
9954 case Intrinsic::arm_neon_vst3lane:
9955 case Intrinsic::arm_neon_vst4lane:
9956 return PerformVLDCombine(N, DCI);
9964 bool ARMTargetLowering::isDesirableToTransformToIntegerOp(unsigned Opc,
9966 return (VT == MVT::f32) && (Opc == ISD::LOAD || Opc == ISD::STORE);
9969 bool ARMTargetLowering::allowsMisalignedMemoryAccesses(EVT VT,
9973 // The AllowsUnaliged flag models the SCTLR.A setting in ARM cpus
9974 bool AllowsUnaligned = Subtarget->allowsUnalignedMem();
9976 switch (VT.getSimpleVT().SimpleTy) {
9982 // Unaligned access can use (for example) LRDB, LRDH, LDR
9983 if (AllowsUnaligned) {
9985 *Fast = Subtarget->hasV7Ops();
9992 // For any little-endian targets with neon, we can support unaligned ld/st
9993 // of D and Q (e.g. {D0,D1}) registers by using vld1.i8/vst1.i8.
9994 // A big-endian target may also explicitly support unaligned accesses
9995 if (Subtarget->hasNEON() && (AllowsUnaligned || isLittleEndian())) {
10005 static bool memOpAlign(unsigned DstAlign, unsigned SrcAlign,
10006 unsigned AlignCheck) {
10007 return ((SrcAlign == 0 || SrcAlign % AlignCheck == 0) &&
10008 (DstAlign == 0 || DstAlign % AlignCheck == 0));
10011 EVT ARMTargetLowering::getOptimalMemOpType(uint64_t Size,
10012 unsigned DstAlign, unsigned SrcAlign,
10013 bool IsMemset, bool ZeroMemset,
10015 MachineFunction &MF) const {
10016 const Function *F = MF.getFunction();
10018 // See if we can use NEON instructions for this...
10019 if ((!IsMemset || ZeroMemset) &&
10020 Subtarget->hasNEON() &&
10021 !F->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
10022 Attribute::NoImplicitFloat)) {
10025 (memOpAlign(SrcAlign, DstAlign, 16) ||
10026 (allowsMisalignedMemoryAccesses(MVT::v2f64, 0, 1, &Fast) && Fast))) {
10028 } else if (Size >= 8 &&
10029 (memOpAlign(SrcAlign, DstAlign, 8) ||
10030 (allowsMisalignedMemoryAccesses(MVT::f64, 0, 1, &Fast) &&
10036 // Lowering to i32/i16 if the size permits.
10039 else if (Size >= 2)
10042 // Let the target-independent logic figure it out.
10046 bool ARMTargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
10047 if (Val.getOpcode() != ISD::LOAD)
10050 EVT VT1 = Val.getValueType();
10051 if (!VT1.isSimple() || !VT1.isInteger() ||
10052 !VT2.isSimple() || !VT2.isInteger())
10055 switch (VT1.getSimpleVT().SimpleTy) {
10060 // 8-bit and 16-bit loads implicitly zero-extend to 32-bits.
10067 bool ARMTargetLowering::allowTruncateForTailCall(Type *Ty1, Type *Ty2) const {
10068 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
10071 if (!isTypeLegal(EVT::getEVT(Ty1)))
10074 assert(Ty1->getPrimitiveSizeInBits() <= 64 && "i128 is probably not a noop");
10076 // Assuming the caller doesn't have a zeroext or signext return parameter,
10077 // truncation all the way down to i1 is valid.
10082 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
10086 unsigned Scale = 1;
10087 switch (VT.getSimpleVT().SimpleTy) {
10088 default: return false;
10103 if ((V & (Scale - 1)) != 0)
10106 return V == (V & ((1LL << 5) - 1));
10109 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
10110 const ARMSubtarget *Subtarget) {
10111 bool isNeg = false;
10117 switch (VT.getSimpleVT().SimpleTy) {
10118 default: return false;
10123 // + imm12 or - imm8
10125 return V == (V & ((1LL << 8) - 1));
10126 return V == (V & ((1LL << 12) - 1));
10129 // Same as ARM mode. FIXME: NEON?
10130 if (!Subtarget->hasVFP2())
10135 return V == (V & ((1LL << 8) - 1));
10139 /// isLegalAddressImmediate - Return true if the integer value can be used
10140 /// as the offset of the target addressing mode for load / store of the
10142 static bool isLegalAddressImmediate(int64_t V, EVT VT,
10143 const ARMSubtarget *Subtarget) {
10147 if (!VT.isSimple())
10150 if (Subtarget->isThumb1Only())
10151 return isLegalT1AddressImmediate(V, VT);
10152 else if (Subtarget->isThumb2())
10153 return isLegalT2AddressImmediate(V, VT, Subtarget);
10158 switch (VT.getSimpleVT().SimpleTy) {
10159 default: return false;
10164 return V == (V & ((1LL << 12) - 1));
10167 return V == (V & ((1LL << 8) - 1));
10170 if (!Subtarget->hasVFP2()) // FIXME: NEON?
10175 return V == (V & ((1LL << 8) - 1));
10179 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
10181 int Scale = AM.Scale;
10185 switch (VT.getSimpleVT().SimpleTy) {
10186 default: return false;
10194 Scale = Scale & ~1;
10195 return Scale == 2 || Scale == 4 || Scale == 8;
10198 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
10202 // Note, we allow "void" uses (basically, uses that aren't loads or
10203 // stores), because arm allows folding a scale into many arithmetic
10204 // operations. This should be made more precise and revisited later.
10206 // Allow r << imm, but the imm has to be a multiple of two.
10207 if (Scale & 1) return false;
10208 return isPowerOf2_32(Scale);
10212 /// isLegalAddressingMode - Return true if the addressing mode represented
10213 /// by AM is legal for this target, for a load/store of the specified type.
10214 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
10216 EVT VT = getValueType(Ty, true);
10217 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
10220 // Can never fold addr of global into load/store.
10224 switch (AM.Scale) {
10225 case 0: // no scale reg, must be "r+i" or "r", or "i".
10228 if (Subtarget->isThumb1Only())
10232 // ARM doesn't support any R+R*scale+imm addr modes.
10236 if (!VT.isSimple())
10239 if (Subtarget->isThumb2())
10240 return isLegalT2ScaledAddressingMode(AM, VT);
10242 int Scale = AM.Scale;
10243 switch (VT.getSimpleVT().SimpleTy) {
10244 default: return false;
10248 if (Scale < 0) Scale = -Scale;
10252 return isPowerOf2_32(Scale & ~1);
10256 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
10261 // Note, we allow "void" uses (basically, uses that aren't loads or
10262 // stores), because arm allows folding a scale into many arithmetic
10263 // operations. This should be made more precise and revisited later.
10265 // Allow r << imm, but the imm has to be a multiple of two.
10266 if (Scale & 1) return false;
10267 return isPowerOf2_32(Scale);
10273 /// isLegalICmpImmediate - Return true if the specified immediate is legal
10274 /// icmp immediate, that is the target has icmp instructions which can compare
10275 /// a register against the immediate without having to materialize the
10276 /// immediate into a register.
10277 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
10278 // Thumb2 and ARM modes can use cmn for negative immediates.
10279 if (!Subtarget->isThumb())
10280 return ARM_AM::getSOImmVal(llvm::abs64(Imm)) != -1;
10281 if (Subtarget->isThumb2())
10282 return ARM_AM::getT2SOImmVal(llvm::abs64(Imm)) != -1;
10283 // Thumb1 doesn't have cmn, and only 8-bit immediates.
10284 return Imm >= 0 && Imm <= 255;
10287 /// isLegalAddImmediate - Return true if the specified immediate is a legal add
10288 /// *or sub* immediate, that is the target has add or sub instructions which can
10289 /// add a register with the immediate without having to materialize the
10290 /// immediate into a register.
10291 bool ARMTargetLowering::isLegalAddImmediate(int64_t Imm) const {
10292 // Same encoding for add/sub, just flip the sign.
10293 int64_t AbsImm = llvm::abs64(Imm);
10294 if (!Subtarget->isThumb())
10295 return ARM_AM::getSOImmVal(AbsImm) != -1;
10296 if (Subtarget->isThumb2())
10297 return ARM_AM::getT2SOImmVal(AbsImm) != -1;
10298 // Thumb1 only has 8-bit unsigned immediate.
10299 return AbsImm >= 0 && AbsImm <= 255;
10302 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
10303 bool isSEXTLoad, SDValue &Base,
10304 SDValue &Offset, bool &isInc,
10305 SelectionDAG &DAG) {
10306 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
10309 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
10310 // AddressingMode 3
10311 Base = Ptr->getOperand(0);
10312 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
10313 int RHSC = (int)RHS->getZExtValue();
10314 if (RHSC < 0 && RHSC > -256) {
10315 assert(Ptr->getOpcode() == ISD::ADD);
10317 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
10321 isInc = (Ptr->getOpcode() == ISD::ADD);
10322 Offset = Ptr->getOperand(1);
10324 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
10325 // AddressingMode 2
10326 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
10327 int RHSC = (int)RHS->getZExtValue();
10328 if (RHSC < 0 && RHSC > -0x1000) {
10329 assert(Ptr->getOpcode() == ISD::ADD);
10331 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
10332 Base = Ptr->getOperand(0);
10337 if (Ptr->getOpcode() == ISD::ADD) {
10339 ARM_AM::ShiftOpc ShOpcVal=
10340 ARM_AM::getShiftOpcForNode(Ptr->getOperand(0).getOpcode());
10341 if (ShOpcVal != ARM_AM::no_shift) {
10342 Base = Ptr->getOperand(1);
10343 Offset = Ptr->getOperand(0);
10345 Base = Ptr->getOperand(0);
10346 Offset = Ptr->getOperand(1);
10351 isInc = (Ptr->getOpcode() == ISD::ADD);
10352 Base = Ptr->getOperand(0);
10353 Offset = Ptr->getOperand(1);
10357 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
10361 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
10362 bool isSEXTLoad, SDValue &Base,
10363 SDValue &Offset, bool &isInc,
10364 SelectionDAG &DAG) {
10365 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
10368 Base = Ptr->getOperand(0);
10369 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
10370 int RHSC = (int)RHS->getZExtValue();
10371 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
10372 assert(Ptr->getOpcode() == ISD::ADD);
10374 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
10376 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
10377 isInc = Ptr->getOpcode() == ISD::ADD;
10378 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
10386 /// getPreIndexedAddressParts - returns true by value, base pointer and
10387 /// offset pointer and addressing mode by reference if the node's address
10388 /// can be legally represented as pre-indexed load / store address.
10390 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
10392 ISD::MemIndexedMode &AM,
10393 SelectionDAG &DAG) const {
10394 if (Subtarget->isThumb1Only())
10399 bool isSEXTLoad = false;
10400 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
10401 Ptr = LD->getBasePtr();
10402 VT = LD->getMemoryVT();
10403 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
10404 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
10405 Ptr = ST->getBasePtr();
10406 VT = ST->getMemoryVT();
10411 bool isLegal = false;
10412 if (Subtarget->isThumb2())
10413 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
10414 Offset, isInc, DAG);
10416 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
10417 Offset, isInc, DAG);
10421 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
10425 /// getPostIndexedAddressParts - returns true by value, base pointer and
10426 /// offset pointer and addressing mode by reference if this node can be
10427 /// combined with a load / store to form a post-indexed load / store.
10428 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
10431 ISD::MemIndexedMode &AM,
10432 SelectionDAG &DAG) const {
10433 if (Subtarget->isThumb1Only())
10438 bool isSEXTLoad = false;
10439 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
10440 VT = LD->getMemoryVT();
10441 Ptr = LD->getBasePtr();
10442 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
10443 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
10444 VT = ST->getMemoryVT();
10445 Ptr = ST->getBasePtr();
10450 bool isLegal = false;
10451 if (Subtarget->isThumb2())
10452 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
10455 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
10461 // Swap base ptr and offset to catch more post-index load / store when
10462 // it's legal. In Thumb2 mode, offset must be an immediate.
10463 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
10464 !Subtarget->isThumb2())
10465 std::swap(Base, Offset);
10467 // Post-indexed load / store update the base pointer.
10472 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
10476 void ARMTargetLowering::computeKnownBitsForTargetNode(const SDValue Op,
10479 const SelectionDAG &DAG,
10480 unsigned Depth) const {
10481 unsigned BitWidth = KnownOne.getBitWidth();
10482 KnownZero = KnownOne = APInt(BitWidth, 0);
10483 switch (Op.getOpcode()) {
10489 // These nodes' second result is a boolean
10490 if (Op.getResNo() == 0)
10492 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - 1);
10494 case ARMISD::CMOV: {
10495 // Bits are known zero/one if known on the LHS and RHS.
10496 DAG.computeKnownBits(Op.getOperand(0), KnownZero, KnownOne, Depth+1);
10497 if (KnownZero == 0 && KnownOne == 0) return;
10499 APInt KnownZeroRHS, KnownOneRHS;
10500 DAG.computeKnownBits(Op.getOperand(1), KnownZeroRHS, KnownOneRHS, Depth+1);
10501 KnownZero &= KnownZeroRHS;
10502 KnownOne &= KnownOneRHS;
10505 case ISD::INTRINSIC_W_CHAIN: {
10506 ConstantSDNode *CN = cast<ConstantSDNode>(Op->getOperand(1));
10507 Intrinsic::ID IntID = static_cast<Intrinsic::ID>(CN->getZExtValue());
10510 case Intrinsic::arm_ldaex:
10511 case Intrinsic::arm_ldrex: {
10512 EVT VT = cast<MemIntrinsicSDNode>(Op)->getMemoryVT();
10513 unsigned MemBits = VT.getScalarType().getSizeInBits();
10514 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - MemBits);
10522 //===----------------------------------------------------------------------===//
10523 // ARM Inline Assembly Support
10524 //===----------------------------------------------------------------------===//
10526 bool ARMTargetLowering::ExpandInlineAsm(CallInst *CI) const {
10527 // Looking for "rev" which is V6+.
10528 if (!Subtarget->hasV6Ops())
10531 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
10532 std::string AsmStr = IA->getAsmString();
10533 SmallVector<StringRef, 4> AsmPieces;
10534 SplitString(AsmStr, AsmPieces, ";\n");
10536 switch (AsmPieces.size()) {
10537 default: return false;
10539 AsmStr = AsmPieces[0];
10541 SplitString(AsmStr, AsmPieces, " \t,");
10544 if (AsmPieces.size() == 3 &&
10545 AsmPieces[0] == "rev" && AsmPieces[1] == "$0" && AsmPieces[2] == "$1" &&
10546 IA->getConstraintString().compare(0, 4, "=l,l") == 0) {
10547 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
10548 if (Ty && Ty->getBitWidth() == 32)
10549 return IntrinsicLowering::LowerToByteSwap(CI);
10557 /// getConstraintType - Given a constraint letter, return the type of
10558 /// constraint it is for this target.
10559 ARMTargetLowering::ConstraintType
10560 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
10561 if (Constraint.size() == 1) {
10562 switch (Constraint[0]) {
10564 case 'l': return C_RegisterClass;
10565 case 'w': return C_RegisterClass;
10566 case 'h': return C_RegisterClass;
10567 case 'x': return C_RegisterClass;
10568 case 't': return C_RegisterClass;
10569 case 'j': return C_Other; // Constant for movw.
10570 // An address with a single base register. Due to the way we
10571 // currently handle addresses it is the same as an 'r' memory constraint.
10572 case 'Q': return C_Memory;
10574 } else if (Constraint.size() == 2) {
10575 switch (Constraint[0]) {
10577 // All 'U+' constraints are addresses.
10578 case 'U': return C_Memory;
10581 return TargetLowering::getConstraintType(Constraint);
10584 /// Examine constraint type and operand type and determine a weight value.
10585 /// This object must already have been set up with the operand type
10586 /// and the current alternative constraint selected.
10587 TargetLowering::ConstraintWeight
10588 ARMTargetLowering::getSingleConstraintMatchWeight(
10589 AsmOperandInfo &info, const char *constraint) const {
10590 ConstraintWeight weight = CW_Invalid;
10591 Value *CallOperandVal = info.CallOperandVal;
10592 // If we don't have a value, we can't do a match,
10593 // but allow it at the lowest weight.
10594 if (!CallOperandVal)
10596 Type *type = CallOperandVal->getType();
10597 // Look at the constraint type.
10598 switch (*constraint) {
10600 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
10603 if (type->isIntegerTy()) {
10604 if (Subtarget->isThumb())
10605 weight = CW_SpecificReg;
10607 weight = CW_Register;
10611 if (type->isFloatingPointTy())
10612 weight = CW_Register;
10618 typedef std::pair<unsigned, const TargetRegisterClass*> RCPair;
10620 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
10622 if (Constraint.size() == 1) {
10623 // GCC ARM Constraint Letters
10624 switch (Constraint[0]) {
10625 case 'l': // Low regs or general regs.
10626 if (Subtarget->isThumb())
10627 return RCPair(0U, &ARM::tGPRRegClass);
10628 return RCPair(0U, &ARM::GPRRegClass);
10629 case 'h': // High regs or no regs.
10630 if (Subtarget->isThumb())
10631 return RCPair(0U, &ARM::hGPRRegClass);
10634 if (Subtarget->isThumb1Only())
10635 return RCPair(0U, &ARM::tGPRRegClass);
10636 return RCPair(0U, &ARM::GPRRegClass);
10638 if (VT == MVT::Other)
10640 if (VT == MVT::f32)
10641 return RCPair(0U, &ARM::SPRRegClass);
10642 if (VT.getSizeInBits() == 64)
10643 return RCPair(0U, &ARM::DPRRegClass);
10644 if (VT.getSizeInBits() == 128)
10645 return RCPair(0U, &ARM::QPRRegClass);
10648 if (VT == MVT::Other)
10650 if (VT == MVT::f32)
10651 return RCPair(0U, &ARM::SPR_8RegClass);
10652 if (VT.getSizeInBits() == 64)
10653 return RCPair(0U, &ARM::DPR_8RegClass);
10654 if (VT.getSizeInBits() == 128)
10655 return RCPair(0U, &ARM::QPR_8RegClass);
10658 if (VT == MVT::f32)
10659 return RCPair(0U, &ARM::SPRRegClass);
10663 if (StringRef("{cc}").equals_lower(Constraint))
10664 return std::make_pair(unsigned(ARM::CPSR), &ARM::CCRRegClass);
10666 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
10669 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
10670 /// vector. If it is invalid, don't add anything to Ops.
10671 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
10672 std::string &Constraint,
10673 std::vector<SDValue>&Ops,
10674 SelectionDAG &DAG) const {
10677 // Currently only support length 1 constraints.
10678 if (Constraint.length() != 1) return;
10680 char ConstraintLetter = Constraint[0];
10681 switch (ConstraintLetter) {
10684 case 'I': case 'J': case 'K': case 'L':
10685 case 'M': case 'N': case 'O':
10686 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
10690 int64_t CVal64 = C->getSExtValue();
10691 int CVal = (int) CVal64;
10692 // None of these constraints allow values larger than 32 bits. Check
10693 // that the value fits in an int.
10694 if (CVal != CVal64)
10697 switch (ConstraintLetter) {
10699 // Constant suitable for movw, must be between 0 and
10701 if (Subtarget->hasV6T2Ops())
10702 if (CVal >= 0 && CVal <= 65535)
10706 if (Subtarget->isThumb1Only()) {
10707 // This must be a constant between 0 and 255, for ADD
10709 if (CVal >= 0 && CVal <= 255)
10711 } else if (Subtarget->isThumb2()) {
10712 // A constant that can be used as an immediate value in a
10713 // data-processing instruction.
10714 if (ARM_AM::getT2SOImmVal(CVal) != -1)
10717 // A constant that can be used as an immediate value in a
10718 // data-processing instruction.
10719 if (ARM_AM::getSOImmVal(CVal) != -1)
10725 if (Subtarget->isThumb()) { // FIXME thumb2
10726 // This must be a constant between -255 and -1, for negated ADD
10727 // immediates. This can be used in GCC with an "n" modifier that
10728 // prints the negated value, for use with SUB instructions. It is
10729 // not useful otherwise but is implemented for compatibility.
10730 if (CVal >= -255 && CVal <= -1)
10733 // This must be a constant between -4095 and 4095. It is not clear
10734 // what this constraint is intended for. Implemented for
10735 // compatibility with GCC.
10736 if (CVal >= -4095 && CVal <= 4095)
10742 if (Subtarget->isThumb1Only()) {
10743 // A 32-bit value where only one byte has a nonzero value. Exclude
10744 // zero to match GCC. This constraint is used by GCC internally for
10745 // constants that can be loaded with a move/shift combination.
10746 // It is not useful otherwise but is implemented for compatibility.
10747 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
10749 } else if (Subtarget->isThumb2()) {
10750 // A constant whose bitwise inverse can be used as an immediate
10751 // value in a data-processing instruction. This can be used in GCC
10752 // with a "B" modifier that prints the inverted value, for use with
10753 // BIC and MVN instructions. It is not useful otherwise but is
10754 // implemented for compatibility.
10755 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
10758 // A constant whose bitwise inverse can be used as an immediate
10759 // value in a data-processing instruction. This can be used in GCC
10760 // with a "B" modifier that prints the inverted value, for use with
10761 // BIC and MVN instructions. It is not useful otherwise but is
10762 // implemented for compatibility.
10763 if (ARM_AM::getSOImmVal(~CVal) != -1)
10769 if (Subtarget->isThumb1Only()) {
10770 // This must be a constant between -7 and 7,
10771 // for 3-operand ADD/SUB immediate instructions.
10772 if (CVal >= -7 && CVal < 7)
10774 } else if (Subtarget->isThumb2()) {
10775 // A constant whose negation can be used as an immediate value in a
10776 // data-processing instruction. This can be used in GCC with an "n"
10777 // modifier that prints the negated value, for use with SUB
10778 // instructions. It is not useful otherwise but is implemented for
10780 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
10783 // A constant whose negation can be used as an immediate value in a
10784 // data-processing instruction. This can be used in GCC with an "n"
10785 // modifier that prints the negated value, for use with SUB
10786 // instructions. It is not useful otherwise but is implemented for
10788 if (ARM_AM::getSOImmVal(-CVal) != -1)
10794 if (Subtarget->isThumb()) { // FIXME thumb2
10795 // This must be a multiple of 4 between 0 and 1020, for
10796 // ADD sp + immediate.
10797 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
10800 // A power of two or a constant between 0 and 32. This is used in
10801 // GCC for the shift amount on shifted register operands, but it is
10802 // useful in general for any shift amounts.
10803 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
10809 if (Subtarget->isThumb()) { // FIXME thumb2
10810 // This must be a constant between 0 and 31, for shift amounts.
10811 if (CVal >= 0 && CVal <= 31)
10817 if (Subtarget->isThumb()) { // FIXME thumb2
10818 // This must be a multiple of 4 between -508 and 508, for
10819 // ADD/SUB sp = sp + immediate.
10820 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
10825 Result = DAG.getTargetConstant(CVal, Op.getValueType());
10829 if (Result.getNode()) {
10830 Ops.push_back(Result);
10833 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
10836 SDValue ARMTargetLowering::LowerDivRem(SDValue Op, SelectionDAG &DAG) const {
10837 assert(Subtarget->isTargetAEABI() && "Register-based DivRem lowering only");
10838 unsigned Opcode = Op->getOpcode();
10839 assert((Opcode == ISD::SDIVREM || Opcode == ISD::UDIVREM) &&
10840 "Invalid opcode for Div/Rem lowering");
10841 bool isSigned = (Opcode == ISD::SDIVREM);
10842 EVT VT = Op->getValueType(0);
10843 Type *Ty = VT.getTypeForEVT(*DAG.getContext());
10846 switch (VT.getSimpleVT().SimpleTy) {
10847 default: llvm_unreachable("Unexpected request for libcall!");
10848 case MVT::i8: LC = isSigned ? RTLIB::SDIVREM_I8 : RTLIB::UDIVREM_I8; break;
10849 case MVT::i16: LC = isSigned ? RTLIB::SDIVREM_I16 : RTLIB::UDIVREM_I16; break;
10850 case MVT::i32: LC = isSigned ? RTLIB::SDIVREM_I32 : RTLIB::UDIVREM_I32; break;
10851 case MVT::i64: LC = isSigned ? RTLIB::SDIVREM_I64 : RTLIB::UDIVREM_I64; break;
10854 SDValue InChain = DAG.getEntryNode();
10856 TargetLowering::ArgListTy Args;
10857 TargetLowering::ArgListEntry Entry;
10858 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i) {
10859 EVT ArgVT = Op->getOperand(i).getValueType();
10860 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
10861 Entry.Node = Op->getOperand(i);
10863 Entry.isSExt = isSigned;
10864 Entry.isZExt = !isSigned;
10865 Args.push_back(Entry);
10868 SDValue Callee = DAG.getExternalSymbol(getLibcallName(LC),
10871 Type *RetTy = (Type*)StructType::get(Ty, Ty, nullptr);
10874 TargetLowering::CallLoweringInfo CLI(DAG);
10875 CLI.setDebugLoc(dl).setChain(InChain)
10876 .setCallee(getLibcallCallingConv(LC), RetTy, Callee, std::move(Args), 0)
10877 .setInRegister().setSExtResult(isSigned).setZExtResult(!isSigned);
10879 std::pair<SDValue, SDValue> CallInfo = LowerCallTo(CLI);
10880 return CallInfo.first;
10884 ARMTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const {
10885 assert(Subtarget->isTargetWindows() && "unsupported target platform");
10889 SDValue Chain = Op.getOperand(0);
10890 SDValue Size = Op.getOperand(1);
10892 SDValue Words = DAG.getNode(ISD::SRL, DL, MVT::i32, Size,
10893 DAG.getConstant(2, MVT::i32));
10896 Chain = DAG.getCopyToReg(Chain, DL, ARM::R4, Words, Flag);
10897 Flag = Chain.getValue(1);
10899 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
10900 Chain = DAG.getNode(ARMISD::WIN__CHKSTK, DL, NodeTys, Chain, Flag);
10902 SDValue NewSP = DAG.getCopyFromReg(Chain, DL, ARM::SP, MVT::i32);
10903 Chain = NewSP.getValue(1);
10905 SDValue Ops[2] = { NewSP, Chain };
10906 return DAG.getMergeValues(Ops, DL);
10909 SDValue ARMTargetLowering::LowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) const {
10910 assert(Op.getValueType() == MVT::f64 && Subtarget->isFPOnlySP() &&
10911 "Unexpected type for custom-lowering FP_EXTEND");
10914 LC = RTLIB::getFPEXT(Op.getOperand(0).getValueType(), Op.getValueType());
10916 SDValue SrcVal = Op.getOperand(0);
10917 return makeLibCall(DAG, LC, Op.getValueType(), &SrcVal, 1,
10918 /*isSigned*/ false, SDLoc(Op)).first;
10921 SDValue ARMTargetLowering::LowerFP_ROUND(SDValue Op, SelectionDAG &DAG) const {
10922 assert(Op.getOperand(0).getValueType() == MVT::f64 &&
10923 Subtarget->isFPOnlySP() &&
10924 "Unexpected type for custom-lowering FP_ROUND");
10927 LC = RTLIB::getFPROUND(Op.getOperand(0).getValueType(), Op.getValueType());
10929 SDValue SrcVal = Op.getOperand(0);
10930 return makeLibCall(DAG, LC, Op.getValueType(), &SrcVal, 1,
10931 /*isSigned*/ false, SDLoc(Op)).first;
10935 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
10936 // The ARM target isn't yet aware of offsets.
10940 bool ARM::isBitFieldInvertedMask(unsigned v) {
10941 if (v == 0xffffffff)
10944 // there can be 1's on either or both "outsides", all the "inside"
10945 // bits must be 0's
10946 unsigned TO = CountTrailingOnes_32(v);
10947 unsigned LO = CountLeadingOnes_32(v);
10948 v = (v >> TO) << TO;
10949 v = (v << LO) >> LO;
10953 /// isFPImmLegal - Returns true if the target can instruction select the
10954 /// specified FP immediate natively. If false, the legalizer will
10955 /// materialize the FP immediate as a load from a constant pool.
10956 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
10957 if (!Subtarget->hasVFP3())
10959 if (VT == MVT::f32)
10960 return ARM_AM::getFP32Imm(Imm) != -1;
10961 if (VT == MVT::f64 && !Subtarget->isFPOnlySP())
10962 return ARM_AM::getFP64Imm(Imm) != -1;
10966 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
10967 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
10968 /// specified in the intrinsic calls.
10969 bool ARMTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
10971 unsigned Intrinsic) const {
10972 switch (Intrinsic) {
10973 case Intrinsic::arm_neon_vld1:
10974 case Intrinsic::arm_neon_vld2:
10975 case Intrinsic::arm_neon_vld3:
10976 case Intrinsic::arm_neon_vld4:
10977 case Intrinsic::arm_neon_vld2lane:
10978 case Intrinsic::arm_neon_vld3lane:
10979 case Intrinsic::arm_neon_vld4lane: {
10980 Info.opc = ISD::INTRINSIC_W_CHAIN;
10981 // Conservatively set memVT to the entire set of vectors loaded.
10982 uint64_t NumElts = getDataLayout()->getTypeAllocSize(I.getType()) / 8;
10983 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
10984 Info.ptrVal = I.getArgOperand(0);
10986 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
10987 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
10988 Info.vol = false; // volatile loads with NEON intrinsics not supported
10989 Info.readMem = true;
10990 Info.writeMem = false;
10993 case Intrinsic::arm_neon_vst1:
10994 case Intrinsic::arm_neon_vst2:
10995 case Intrinsic::arm_neon_vst3:
10996 case Intrinsic::arm_neon_vst4:
10997 case Intrinsic::arm_neon_vst2lane:
10998 case Intrinsic::arm_neon_vst3lane:
10999 case Intrinsic::arm_neon_vst4lane: {
11000 Info.opc = ISD::INTRINSIC_VOID;
11001 // Conservatively set memVT to the entire set of vectors stored.
11002 unsigned NumElts = 0;
11003 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
11004 Type *ArgTy = I.getArgOperand(ArgI)->getType();
11005 if (!ArgTy->isVectorTy())
11007 NumElts += getDataLayout()->getTypeAllocSize(ArgTy) / 8;
11009 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
11010 Info.ptrVal = I.getArgOperand(0);
11012 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
11013 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
11014 Info.vol = false; // volatile stores with NEON intrinsics not supported
11015 Info.readMem = false;
11016 Info.writeMem = true;
11019 case Intrinsic::arm_ldaex:
11020 case Intrinsic::arm_ldrex: {
11021 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(0)->getType());
11022 Info.opc = ISD::INTRINSIC_W_CHAIN;
11023 Info.memVT = MVT::getVT(PtrTy->getElementType());
11024 Info.ptrVal = I.getArgOperand(0);
11026 Info.align = getDataLayout()->getABITypeAlignment(PtrTy->getElementType());
11028 Info.readMem = true;
11029 Info.writeMem = false;
11032 case Intrinsic::arm_stlex:
11033 case Intrinsic::arm_strex: {
11034 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(1)->getType());
11035 Info.opc = ISD::INTRINSIC_W_CHAIN;
11036 Info.memVT = MVT::getVT(PtrTy->getElementType());
11037 Info.ptrVal = I.getArgOperand(1);
11039 Info.align = getDataLayout()->getABITypeAlignment(PtrTy->getElementType());
11041 Info.readMem = false;
11042 Info.writeMem = true;
11045 case Intrinsic::arm_stlexd:
11046 case Intrinsic::arm_strexd: {
11047 Info.opc = ISD::INTRINSIC_W_CHAIN;
11048 Info.memVT = MVT::i64;
11049 Info.ptrVal = I.getArgOperand(2);
11053 Info.readMem = false;
11054 Info.writeMem = true;
11057 case Intrinsic::arm_ldaexd:
11058 case Intrinsic::arm_ldrexd: {
11059 Info.opc = ISD::INTRINSIC_W_CHAIN;
11060 Info.memVT = MVT::i64;
11061 Info.ptrVal = I.getArgOperand(0);
11065 Info.readMem = true;
11066 Info.writeMem = false;
11076 /// \brief Returns true if it is beneficial to convert a load of a constant
11077 /// to just the constant itself.
11078 bool ARMTargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
11080 assert(Ty->isIntegerTy());
11082 unsigned Bits = Ty->getPrimitiveSizeInBits();
11083 if (Bits == 0 || Bits > 32)
11088 bool ARMTargetLowering::hasLoadLinkedStoreConditional() const { return true; }
11090 Instruction* ARMTargetLowering::makeDMB(IRBuilder<> &Builder,
11091 ARM_MB::MemBOpt Domain) const {
11092 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
11094 // First, if the target has no DMB, see what fallback we can use.
11095 if (!Subtarget->hasDataBarrier()) {
11096 // Some ARMv6 cpus can support data barriers with an mcr instruction.
11097 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
11099 if (Subtarget->hasV6Ops() && !Subtarget->isThumb()) {
11100 Function *MCR = llvm::Intrinsic::getDeclaration(M, Intrinsic::arm_mcr);
11101 Value* args[6] = {Builder.getInt32(15), Builder.getInt32(0),
11102 Builder.getInt32(0), Builder.getInt32(7),
11103 Builder.getInt32(10), Builder.getInt32(5)};
11104 return Builder.CreateCall(MCR, args);
11106 // Instead of using barriers, atomic accesses on these subtargets use
11108 llvm_unreachable("makeDMB on a target so old that it has no barriers");
11111 Function *DMB = llvm::Intrinsic::getDeclaration(M, Intrinsic::arm_dmb);
11112 // Only a full system barrier exists in the M-class architectures.
11113 Domain = Subtarget->isMClass() ? ARM_MB::SY : Domain;
11114 Constant *CDomain = Builder.getInt32(Domain);
11115 return Builder.CreateCall(DMB, CDomain);
11119 // Based on http://www.cl.cam.ac.uk/~pes20/cpp/cpp0xmappings.html
11120 Instruction* ARMTargetLowering::emitLeadingFence(IRBuilder<> &Builder,
11121 AtomicOrdering Ord, bool IsStore,
11122 bool IsLoad) const {
11123 if (!getInsertFencesForAtomic())
11129 llvm_unreachable("Invalid fence: unordered/non-atomic");
11132 return nullptr; // Nothing to do
11133 case SequentiallyConsistent:
11135 return nullptr; // Nothing to do
11138 case AcquireRelease:
11139 if (Subtarget->isSwift())
11140 return makeDMB(Builder, ARM_MB::ISHST);
11141 // FIXME: add a comment with a link to documentation justifying this.
11143 return makeDMB(Builder, ARM_MB::ISH);
11145 llvm_unreachable("Unknown fence ordering in emitLeadingFence");
11148 Instruction* ARMTargetLowering::emitTrailingFence(IRBuilder<> &Builder,
11149 AtomicOrdering Ord, bool IsStore,
11150 bool IsLoad) const {
11151 if (!getInsertFencesForAtomic())
11157 llvm_unreachable("Invalid fence: unordered/not-atomic");
11160 return nullptr; // Nothing to do
11162 case AcquireRelease:
11163 case SequentiallyConsistent:
11164 return makeDMB(Builder, ARM_MB::ISH);
11166 llvm_unreachable("Unknown fence ordering in emitTrailingFence");
11169 // Loads and stores less than 64-bits are already atomic; ones above that
11170 // are doomed anyway, so defer to the default libcall and blame the OS when
11171 // things go wrong. Cortex M doesn't have ldrexd/strexd though, so don't emit
11172 // anything for those.
11173 bool ARMTargetLowering::shouldExpandAtomicStoreInIR(StoreInst *SI) const {
11174 unsigned Size = SI->getValueOperand()->getType()->getPrimitiveSizeInBits();
11175 return (Size == 64) && !Subtarget->isMClass();
11178 // Loads and stores less than 64-bits are already atomic; ones above that
11179 // are doomed anyway, so defer to the default libcall and blame the OS when
11180 // things go wrong. Cortex M doesn't have ldrexd/strexd though, so don't emit
11181 // anything for those.
11182 // FIXME: ldrd and strd are atomic if the CPU has LPAE (e.g. A15 has that
11183 // guarantee, see DDI0406C ARM architecture reference manual,
11184 // sections A8.8.72-74 LDRD)
11185 bool ARMTargetLowering::shouldExpandAtomicLoadInIR(LoadInst *LI) const {
11186 unsigned Size = LI->getType()->getPrimitiveSizeInBits();
11187 return (Size == 64) && !Subtarget->isMClass();
11190 // For the real atomic operations, we have ldrex/strex up to 32 bits,
11191 // and up to 64 bits on the non-M profiles
11192 bool ARMTargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
11193 unsigned Size = AI->getType()->getPrimitiveSizeInBits();
11194 return Size <= (Subtarget->isMClass() ? 32U : 64U);
11197 // This has so far only been implemented for MachO.
11198 bool ARMTargetLowering::useLoadStackGuardNode() const {
11199 return Subtarget->isTargetMachO();
11202 bool ARMTargetLowering::canCombineStoreAndExtract(Type *VectorTy, Value *Idx,
11203 unsigned &Cost) const {
11204 // If we do not have NEON, vector types are not natively supported.
11205 if (!Subtarget->hasNEON())
11208 // Floating point values and vector values map to the same register file.
11209 // Therefore, althought we could do a store extract of a vector type, this is
11210 // better to leave at float as we have more freedom in the addressing mode for
11212 if (VectorTy->isFPOrFPVectorTy())
11215 // If the index is unknown at compile time, this is very expensive to lower
11216 // and it is not possible to combine the store with the extract.
11217 if (!isa<ConstantInt>(Idx))
11220 assert(VectorTy->isVectorTy() && "VectorTy is not a vector type");
11221 unsigned BitWidth = cast<VectorType>(VectorTy)->getBitWidth();
11222 // We can do a store + vector extract on any vector that fits perfectly in a D
11224 if (BitWidth == 64 || BitWidth == 128) {
11231 Value *ARMTargetLowering::emitLoadLinked(IRBuilder<> &Builder, Value *Addr,
11232 AtomicOrdering Ord) const {
11233 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
11234 Type *ValTy = cast<PointerType>(Addr->getType())->getElementType();
11235 bool IsAcquire = isAtLeastAcquire(Ord);
11237 // Since i64 isn't legal and intrinsics don't get type-lowered, the ldrexd
11238 // intrinsic must return {i32, i32} and we have to recombine them into a
11239 // single i64 here.
11240 if (ValTy->getPrimitiveSizeInBits() == 64) {
11241 Intrinsic::ID Int =
11242 IsAcquire ? Intrinsic::arm_ldaexd : Intrinsic::arm_ldrexd;
11243 Function *Ldrex = llvm::Intrinsic::getDeclaration(M, Int);
11245 Addr = Builder.CreateBitCast(Addr, Type::getInt8PtrTy(M->getContext()));
11246 Value *LoHi = Builder.CreateCall(Ldrex, Addr, "lohi");
11248 Value *Lo = Builder.CreateExtractValue(LoHi, 0, "lo");
11249 Value *Hi = Builder.CreateExtractValue(LoHi, 1, "hi");
11250 if (!Subtarget->isLittle())
11251 std::swap (Lo, Hi);
11252 Lo = Builder.CreateZExt(Lo, ValTy, "lo64");
11253 Hi = Builder.CreateZExt(Hi, ValTy, "hi64");
11254 return Builder.CreateOr(
11255 Lo, Builder.CreateShl(Hi, ConstantInt::get(ValTy, 32)), "val64");
11258 Type *Tys[] = { Addr->getType() };
11259 Intrinsic::ID Int = IsAcquire ? Intrinsic::arm_ldaex : Intrinsic::arm_ldrex;
11260 Function *Ldrex = llvm::Intrinsic::getDeclaration(M, Int, Tys);
11262 return Builder.CreateTruncOrBitCast(
11263 Builder.CreateCall(Ldrex, Addr),
11264 cast<PointerType>(Addr->getType())->getElementType());
11267 Value *ARMTargetLowering::emitStoreConditional(IRBuilder<> &Builder, Value *Val,
11269 AtomicOrdering Ord) const {
11270 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
11271 bool IsRelease = isAtLeastRelease(Ord);
11273 // Since the intrinsics must have legal type, the i64 intrinsics take two
11274 // parameters: "i32, i32". We must marshal Val into the appropriate form
11275 // before the call.
11276 if (Val->getType()->getPrimitiveSizeInBits() == 64) {
11277 Intrinsic::ID Int =
11278 IsRelease ? Intrinsic::arm_stlexd : Intrinsic::arm_strexd;
11279 Function *Strex = Intrinsic::getDeclaration(M, Int);
11280 Type *Int32Ty = Type::getInt32Ty(M->getContext());
11282 Value *Lo = Builder.CreateTrunc(Val, Int32Ty, "lo");
11283 Value *Hi = Builder.CreateTrunc(Builder.CreateLShr(Val, 32), Int32Ty, "hi");
11284 if (!Subtarget->isLittle())
11285 std::swap (Lo, Hi);
11286 Addr = Builder.CreateBitCast(Addr, Type::getInt8PtrTy(M->getContext()));
11287 return Builder.CreateCall3(Strex, Lo, Hi, Addr);
11290 Intrinsic::ID Int = IsRelease ? Intrinsic::arm_stlex : Intrinsic::arm_strex;
11291 Type *Tys[] = { Addr->getType() };
11292 Function *Strex = Intrinsic::getDeclaration(M, Int, Tys);
11294 return Builder.CreateCall2(
11295 Strex, Builder.CreateZExtOrBitCast(
11296 Val, Strex->getFunctionType()->getParamType(0)),
11308 static bool isHomogeneousAggregate(Type *Ty, HABaseType &Base,
11309 uint64_t &Members) {
11310 if (const StructType *ST = dyn_cast<StructType>(Ty)) {
11311 for (unsigned i = 0; i < ST->getNumElements(); ++i) {
11312 uint64_t SubMembers = 0;
11313 if (!isHomogeneousAggregate(ST->getElementType(i), Base, SubMembers))
11315 Members += SubMembers;
11317 } else if (const ArrayType *AT = dyn_cast<ArrayType>(Ty)) {
11318 uint64_t SubMembers = 0;
11319 if (!isHomogeneousAggregate(AT->getElementType(), Base, SubMembers))
11321 Members += SubMembers * AT->getNumElements();
11322 } else if (Ty->isFloatTy()) {
11323 if (Base != HA_UNKNOWN && Base != HA_FLOAT)
11327 } else if (Ty->isDoubleTy()) {
11328 if (Base != HA_UNKNOWN && Base != HA_DOUBLE)
11332 } else if (const VectorType *VT = dyn_cast<VectorType>(Ty)) {
11339 return VT->getBitWidth() == 64;
11341 return VT->getBitWidth() == 128;
11343 switch (VT->getBitWidth()) {
11356 return (Members > 0 && Members <= 4);
11359 /// \brief Return true if a type is an AAPCS-VFP homogeneous aggregate.
11360 bool ARMTargetLowering::functionArgumentNeedsConsecutiveRegisters(
11361 Type *Ty, CallingConv::ID CallConv, bool isVarArg) const {
11362 if (getEffectiveCallingConv(CallConv, isVarArg) !=
11363 CallingConv::ARM_AAPCS_VFP)
11366 HABaseType Base = HA_UNKNOWN;
11367 uint64_t Members = 0;
11368 bool result = isHomogeneousAggregate(Ty, Base, Members);
11369 DEBUG(dbgs() << "isHA: " << result << " "; Ty->dump());