1 //===-- ARMISelLowering.cpp - ARM DAG Lowering Implementation -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the interfaces that ARM uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
15 #include "ARMISelLowering.h"
16 #include "ARMCallingConv.h"
17 #include "ARMConstantPoolValue.h"
18 #include "ARMMachineFunctionInfo.h"
19 #include "ARMPerfectShuffle.h"
20 #include "ARMSubtarget.h"
21 #include "ARMTargetMachine.h"
22 #include "ARMTargetObjectFile.h"
23 #include "MCTargetDesc/ARMAddressingModes.h"
24 #include "llvm/ADT/Statistic.h"
25 #include "llvm/ADT/StringExtras.h"
26 #include "llvm/ADT/StringSwitch.h"
27 #include "llvm/CodeGen/CallingConvLower.h"
28 #include "llvm/CodeGen/IntrinsicLowering.h"
29 #include "llvm/CodeGen/MachineBasicBlock.h"
30 #include "llvm/CodeGen/MachineFrameInfo.h"
31 #include "llvm/CodeGen/MachineFunction.h"
32 #include "llvm/CodeGen/MachineInstrBuilder.h"
33 #include "llvm/CodeGen/MachineJumpTableInfo.h"
34 #include "llvm/CodeGen/MachineModuleInfo.h"
35 #include "llvm/CodeGen/MachineRegisterInfo.h"
36 #include "llvm/CodeGen/SelectionDAG.h"
37 #include "llvm/IR/CallingConv.h"
38 #include "llvm/IR/Constants.h"
39 #include "llvm/IR/Function.h"
40 #include "llvm/IR/GlobalValue.h"
41 #include "llvm/IR/IRBuilder.h"
42 #include "llvm/IR/Instruction.h"
43 #include "llvm/IR/Instructions.h"
44 #include "llvm/IR/IntrinsicInst.h"
45 #include "llvm/IR/Intrinsics.h"
46 #include "llvm/IR/Type.h"
47 #include "llvm/MC/MCSectionMachO.h"
48 #include "llvm/Support/CommandLine.h"
49 #include "llvm/Support/Debug.h"
50 #include "llvm/Support/ErrorHandling.h"
51 #include "llvm/Support/MathExtras.h"
52 #include "llvm/Support/raw_ostream.h"
53 #include "llvm/Target/TargetOptions.h"
57 #define DEBUG_TYPE "arm-isel"
59 STATISTIC(NumTailCalls, "Number of tail calls");
60 STATISTIC(NumMovwMovt, "Number of GAs materialized with movw + movt");
61 STATISTIC(NumLoopByVals, "Number of loops generated for byval arguments");
64 EnableARMLongCalls("arm-long-calls", cl::Hidden,
65 cl::desc("Generate calls via indirect call instructions"),
69 ARMInterworking("arm-interworking", cl::Hidden,
70 cl::desc("Enable / disable ARM interworking (for debugging only)"),
74 class ARMCCState : public CCState {
76 ARMCCState(CallingConv::ID CC, bool isVarArg, MachineFunction &MF,
77 SmallVectorImpl<CCValAssign> &locs, LLVMContext &C,
79 : CCState(CC, isVarArg, MF, locs, C) {
80 assert(((PC == Call) || (PC == Prologue)) &&
81 "ARMCCState users must specify whether their context is call"
82 "or prologue generation.");
88 // The APCS parameter registers.
89 static const MCPhysReg GPRArgRegs[] = {
90 ARM::R0, ARM::R1, ARM::R2, ARM::R3
93 void ARMTargetLowering::addTypeForNEON(MVT VT, MVT PromotedLdStVT,
94 MVT PromotedBitwiseVT) {
95 if (VT != PromotedLdStVT) {
96 setOperationAction(ISD::LOAD, VT, Promote);
97 AddPromotedToType (ISD::LOAD, VT, PromotedLdStVT);
99 setOperationAction(ISD::STORE, VT, Promote);
100 AddPromotedToType (ISD::STORE, VT, PromotedLdStVT);
103 MVT ElemTy = VT.getVectorElementType();
104 if (ElemTy != MVT::i64 && ElemTy != MVT::f64)
105 setOperationAction(ISD::SETCC, VT, Custom);
106 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
107 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
108 if (ElemTy == MVT::i32) {
109 setOperationAction(ISD::SINT_TO_FP, VT, Custom);
110 setOperationAction(ISD::UINT_TO_FP, VT, Custom);
111 setOperationAction(ISD::FP_TO_SINT, VT, Custom);
112 setOperationAction(ISD::FP_TO_UINT, VT, Custom);
114 setOperationAction(ISD::SINT_TO_FP, VT, Expand);
115 setOperationAction(ISD::UINT_TO_FP, VT, Expand);
116 setOperationAction(ISD::FP_TO_SINT, VT, Expand);
117 setOperationAction(ISD::FP_TO_UINT, VT, Expand);
119 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
120 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
121 setOperationAction(ISD::CONCAT_VECTORS, VT, Legal);
122 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Legal);
123 setOperationAction(ISD::SELECT, VT, Expand);
124 setOperationAction(ISD::SELECT_CC, VT, Expand);
125 setOperationAction(ISD::VSELECT, VT, Expand);
126 setOperationAction(ISD::SIGN_EXTEND_INREG, VT, Expand);
127 if (VT.isInteger()) {
128 setOperationAction(ISD::SHL, VT, Custom);
129 setOperationAction(ISD::SRA, VT, Custom);
130 setOperationAction(ISD::SRL, VT, Custom);
133 // Promote all bit-wise operations.
134 if (VT.isInteger() && VT != PromotedBitwiseVT) {
135 setOperationAction(ISD::AND, VT, Promote);
136 AddPromotedToType (ISD::AND, VT, PromotedBitwiseVT);
137 setOperationAction(ISD::OR, VT, Promote);
138 AddPromotedToType (ISD::OR, VT, PromotedBitwiseVT);
139 setOperationAction(ISD::XOR, VT, Promote);
140 AddPromotedToType (ISD::XOR, VT, PromotedBitwiseVT);
143 // Neon does not support vector divide/remainder operations.
144 setOperationAction(ISD::SDIV, VT, Expand);
145 setOperationAction(ISD::UDIV, VT, Expand);
146 setOperationAction(ISD::FDIV, VT, Expand);
147 setOperationAction(ISD::SREM, VT, Expand);
148 setOperationAction(ISD::UREM, VT, Expand);
149 setOperationAction(ISD::FREM, VT, Expand);
152 void ARMTargetLowering::addDRTypeForNEON(MVT VT) {
153 addRegisterClass(VT, &ARM::DPRRegClass);
154 addTypeForNEON(VT, MVT::f64, MVT::v2i32);
157 void ARMTargetLowering::addQRTypeForNEON(MVT VT) {
158 addRegisterClass(VT, &ARM::DPairRegClass);
159 addTypeForNEON(VT, MVT::v2f64, MVT::v4i32);
162 ARMTargetLowering::ARMTargetLowering(const TargetMachine &TM,
163 const ARMSubtarget &STI)
164 : TargetLowering(TM), Subtarget(&STI) {
165 RegInfo = Subtarget->getRegisterInfo();
166 Itins = Subtarget->getInstrItineraryData();
168 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
170 if (Subtarget->isTargetMachO()) {
171 // Uses VFP for Thumb libfuncs if available.
172 if (Subtarget->isThumb() && Subtarget->hasVFP2() &&
173 Subtarget->hasARMOps() && !Subtarget->useSoftFloat()) {
174 // Single-precision floating-point arithmetic.
175 setLibcallName(RTLIB::ADD_F32, "__addsf3vfp");
176 setLibcallName(RTLIB::SUB_F32, "__subsf3vfp");
177 setLibcallName(RTLIB::MUL_F32, "__mulsf3vfp");
178 setLibcallName(RTLIB::DIV_F32, "__divsf3vfp");
180 // Double-precision floating-point arithmetic.
181 setLibcallName(RTLIB::ADD_F64, "__adddf3vfp");
182 setLibcallName(RTLIB::SUB_F64, "__subdf3vfp");
183 setLibcallName(RTLIB::MUL_F64, "__muldf3vfp");
184 setLibcallName(RTLIB::DIV_F64, "__divdf3vfp");
186 // Single-precision comparisons.
187 setLibcallName(RTLIB::OEQ_F32, "__eqsf2vfp");
188 setLibcallName(RTLIB::UNE_F32, "__nesf2vfp");
189 setLibcallName(RTLIB::OLT_F32, "__ltsf2vfp");
190 setLibcallName(RTLIB::OLE_F32, "__lesf2vfp");
191 setLibcallName(RTLIB::OGE_F32, "__gesf2vfp");
192 setLibcallName(RTLIB::OGT_F32, "__gtsf2vfp");
193 setLibcallName(RTLIB::UO_F32, "__unordsf2vfp");
194 setLibcallName(RTLIB::O_F32, "__unordsf2vfp");
196 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
197 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETNE);
198 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
199 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
200 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
201 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
202 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
203 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
205 // Double-precision comparisons.
206 setLibcallName(RTLIB::OEQ_F64, "__eqdf2vfp");
207 setLibcallName(RTLIB::UNE_F64, "__nedf2vfp");
208 setLibcallName(RTLIB::OLT_F64, "__ltdf2vfp");
209 setLibcallName(RTLIB::OLE_F64, "__ledf2vfp");
210 setLibcallName(RTLIB::OGE_F64, "__gedf2vfp");
211 setLibcallName(RTLIB::OGT_F64, "__gtdf2vfp");
212 setLibcallName(RTLIB::UO_F64, "__unorddf2vfp");
213 setLibcallName(RTLIB::O_F64, "__unorddf2vfp");
215 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
216 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETNE);
217 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
218 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
219 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
220 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
221 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
222 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
224 // Floating-point to integer conversions.
225 // i64 conversions are done via library routines even when generating VFP
226 // instructions, so use the same ones.
227 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__fixdfsivfp");
228 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__fixunsdfsivfp");
229 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__fixsfsivfp");
230 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__fixunssfsivfp");
232 // Conversions between floating types.
233 setLibcallName(RTLIB::FPROUND_F64_F32, "__truncdfsf2vfp");
234 setLibcallName(RTLIB::FPEXT_F32_F64, "__extendsfdf2vfp");
236 // Integer to floating-point conversions.
237 // i64 conversions are done via library routines even when generating VFP
238 // instructions, so use the same ones.
239 // FIXME: There appears to be some naming inconsistency in ARM libgcc:
240 // e.g., __floatunsidf vs. __floatunssidfvfp.
241 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__floatsidfvfp");
242 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__floatunssidfvfp");
243 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__floatsisfvfp");
244 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__floatunssisfvfp");
248 // These libcalls are not available in 32-bit.
249 setLibcallName(RTLIB::SHL_I128, nullptr);
250 setLibcallName(RTLIB::SRL_I128, nullptr);
251 setLibcallName(RTLIB::SRA_I128, nullptr);
253 if (Subtarget->isAAPCS_ABI() && !Subtarget->isTargetMachO() &&
254 !Subtarget->isTargetWindows()) {
255 static const struct {
256 const RTLIB::Libcall Op;
257 const char * const Name;
258 const CallingConv::ID CC;
259 const ISD::CondCode Cond;
261 // Double-precision floating-point arithmetic helper functions
262 // RTABI chapter 4.1.2, Table 2
263 { RTLIB::ADD_F64, "__aeabi_dadd", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
264 { RTLIB::DIV_F64, "__aeabi_ddiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
265 { RTLIB::MUL_F64, "__aeabi_dmul", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
266 { RTLIB::SUB_F64, "__aeabi_dsub", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
268 // Double-precision floating-point comparison helper functions
269 // RTABI chapter 4.1.2, Table 3
270 { RTLIB::OEQ_F64, "__aeabi_dcmpeq", CallingConv::ARM_AAPCS, ISD::SETNE },
271 { RTLIB::UNE_F64, "__aeabi_dcmpeq", CallingConv::ARM_AAPCS, ISD::SETEQ },
272 { RTLIB::OLT_F64, "__aeabi_dcmplt", CallingConv::ARM_AAPCS, ISD::SETNE },
273 { RTLIB::OLE_F64, "__aeabi_dcmple", CallingConv::ARM_AAPCS, ISD::SETNE },
274 { RTLIB::OGE_F64, "__aeabi_dcmpge", CallingConv::ARM_AAPCS, ISD::SETNE },
275 { RTLIB::OGT_F64, "__aeabi_dcmpgt", CallingConv::ARM_AAPCS, ISD::SETNE },
276 { RTLIB::UO_F64, "__aeabi_dcmpun", CallingConv::ARM_AAPCS, ISD::SETNE },
277 { RTLIB::O_F64, "__aeabi_dcmpun", CallingConv::ARM_AAPCS, ISD::SETEQ },
279 // Single-precision floating-point arithmetic helper functions
280 // RTABI chapter 4.1.2, Table 4
281 { RTLIB::ADD_F32, "__aeabi_fadd", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
282 { RTLIB::DIV_F32, "__aeabi_fdiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
283 { RTLIB::MUL_F32, "__aeabi_fmul", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
284 { RTLIB::SUB_F32, "__aeabi_fsub", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
286 // Single-precision floating-point comparison helper functions
287 // RTABI chapter 4.1.2, Table 5
288 { RTLIB::OEQ_F32, "__aeabi_fcmpeq", CallingConv::ARM_AAPCS, ISD::SETNE },
289 { RTLIB::UNE_F32, "__aeabi_fcmpeq", CallingConv::ARM_AAPCS, ISD::SETEQ },
290 { RTLIB::OLT_F32, "__aeabi_fcmplt", CallingConv::ARM_AAPCS, ISD::SETNE },
291 { RTLIB::OLE_F32, "__aeabi_fcmple", CallingConv::ARM_AAPCS, ISD::SETNE },
292 { RTLIB::OGE_F32, "__aeabi_fcmpge", CallingConv::ARM_AAPCS, ISD::SETNE },
293 { RTLIB::OGT_F32, "__aeabi_fcmpgt", CallingConv::ARM_AAPCS, ISD::SETNE },
294 { RTLIB::UO_F32, "__aeabi_fcmpun", CallingConv::ARM_AAPCS, ISD::SETNE },
295 { RTLIB::O_F32, "__aeabi_fcmpun", CallingConv::ARM_AAPCS, ISD::SETEQ },
297 // Floating-point to integer conversions.
298 // RTABI chapter 4.1.2, Table 6
299 { RTLIB::FPTOSINT_F64_I32, "__aeabi_d2iz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
300 { RTLIB::FPTOUINT_F64_I32, "__aeabi_d2uiz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
301 { RTLIB::FPTOSINT_F64_I64, "__aeabi_d2lz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
302 { RTLIB::FPTOUINT_F64_I64, "__aeabi_d2ulz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
303 { RTLIB::FPTOSINT_F32_I32, "__aeabi_f2iz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
304 { RTLIB::FPTOUINT_F32_I32, "__aeabi_f2uiz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
305 { RTLIB::FPTOSINT_F32_I64, "__aeabi_f2lz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
306 { RTLIB::FPTOUINT_F32_I64, "__aeabi_f2ulz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
308 // Conversions between floating types.
309 // RTABI chapter 4.1.2, Table 7
310 { RTLIB::FPROUND_F64_F32, "__aeabi_d2f", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
311 { RTLIB::FPROUND_F64_F16, "__aeabi_d2h", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
312 { RTLIB::FPEXT_F32_F64, "__aeabi_f2d", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
314 // Integer to floating-point conversions.
315 // RTABI chapter 4.1.2, Table 8
316 { RTLIB::SINTTOFP_I32_F64, "__aeabi_i2d", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
317 { RTLIB::UINTTOFP_I32_F64, "__aeabi_ui2d", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
318 { RTLIB::SINTTOFP_I64_F64, "__aeabi_l2d", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
319 { RTLIB::UINTTOFP_I64_F64, "__aeabi_ul2d", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
320 { RTLIB::SINTTOFP_I32_F32, "__aeabi_i2f", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
321 { RTLIB::UINTTOFP_I32_F32, "__aeabi_ui2f", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
322 { RTLIB::SINTTOFP_I64_F32, "__aeabi_l2f", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
323 { RTLIB::UINTTOFP_I64_F32, "__aeabi_ul2f", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
325 // Long long helper functions
326 // RTABI chapter 4.2, Table 9
327 { RTLIB::MUL_I64, "__aeabi_lmul", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
328 { RTLIB::SHL_I64, "__aeabi_llsl", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
329 { RTLIB::SRL_I64, "__aeabi_llsr", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
330 { RTLIB::SRA_I64, "__aeabi_lasr", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
332 // Integer division functions
333 // RTABI chapter 4.3.1
334 { RTLIB::SDIV_I8, "__aeabi_idiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
335 { RTLIB::SDIV_I16, "__aeabi_idiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
336 { RTLIB::SDIV_I32, "__aeabi_idiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
337 { RTLIB::SDIV_I64, "__aeabi_ldivmod", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
338 { RTLIB::UDIV_I8, "__aeabi_uidiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
339 { RTLIB::UDIV_I16, "__aeabi_uidiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
340 { RTLIB::UDIV_I32, "__aeabi_uidiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
341 { RTLIB::UDIV_I64, "__aeabi_uldivmod", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
344 // RTABI chapter 4.3.4
345 { RTLIB::MEMCPY, "__aeabi_memcpy", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
346 { RTLIB::MEMMOVE, "__aeabi_memmove", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
347 { RTLIB::MEMSET, "__aeabi_memset", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
350 for (const auto &LC : LibraryCalls) {
351 setLibcallName(LC.Op, LC.Name);
352 setLibcallCallingConv(LC.Op, LC.CC);
353 if (LC.Cond != ISD::SETCC_INVALID)
354 setCmpLibcallCC(LC.Op, LC.Cond);
358 if (Subtarget->isTargetWindows()) {
359 static const struct {
360 const RTLIB::Libcall Op;
361 const char * const Name;
362 const CallingConv::ID CC;
364 { RTLIB::FPTOSINT_F32_I64, "__stoi64", CallingConv::ARM_AAPCS_VFP },
365 { RTLIB::FPTOSINT_F64_I64, "__dtoi64", CallingConv::ARM_AAPCS_VFP },
366 { RTLIB::FPTOUINT_F32_I64, "__stou64", CallingConv::ARM_AAPCS_VFP },
367 { RTLIB::FPTOUINT_F64_I64, "__dtou64", CallingConv::ARM_AAPCS_VFP },
368 { RTLIB::SINTTOFP_I64_F32, "__i64tos", CallingConv::ARM_AAPCS_VFP },
369 { RTLIB::SINTTOFP_I64_F64, "__i64tod", CallingConv::ARM_AAPCS_VFP },
370 { RTLIB::UINTTOFP_I64_F32, "__u64tos", CallingConv::ARM_AAPCS_VFP },
371 { RTLIB::UINTTOFP_I64_F64, "__u64tod", CallingConv::ARM_AAPCS_VFP },
374 for (const auto &LC : LibraryCalls) {
375 setLibcallName(LC.Op, LC.Name);
376 setLibcallCallingConv(LC.Op, LC.CC);
380 // Use divmod compiler-rt calls for iOS 5.0 and later.
381 if (Subtarget->getTargetTriple().isiOS() &&
382 !Subtarget->getTargetTriple().isOSVersionLT(5, 0)) {
383 setLibcallName(RTLIB::SDIVREM_I32, "__divmodsi4");
384 setLibcallName(RTLIB::UDIVREM_I32, "__udivmodsi4");
387 // The half <-> float conversion functions are always soft-float, but are
388 // needed for some targets which use a hard-float calling convention by
390 if (Subtarget->isAAPCS_ABI()) {
391 setLibcallCallingConv(RTLIB::FPROUND_F32_F16, CallingConv::ARM_AAPCS);
392 setLibcallCallingConv(RTLIB::FPROUND_F64_F16, CallingConv::ARM_AAPCS);
393 setLibcallCallingConv(RTLIB::FPEXT_F16_F32, CallingConv::ARM_AAPCS);
395 setLibcallCallingConv(RTLIB::FPROUND_F32_F16, CallingConv::ARM_APCS);
396 setLibcallCallingConv(RTLIB::FPROUND_F64_F16, CallingConv::ARM_APCS);
397 setLibcallCallingConv(RTLIB::FPEXT_F16_F32, CallingConv::ARM_APCS);
400 if (Subtarget->isThumb1Only())
401 addRegisterClass(MVT::i32, &ARM::tGPRRegClass);
403 addRegisterClass(MVT::i32, &ARM::GPRRegClass);
404 if (!Subtarget->useSoftFloat() && Subtarget->hasVFP2() &&
405 !Subtarget->isThumb1Only()) {
406 addRegisterClass(MVT::f32, &ARM::SPRRegClass);
407 addRegisterClass(MVT::f64, &ARM::DPRRegClass);
410 for (MVT VT : MVT::vector_valuetypes()) {
411 for (MVT InnerVT : MVT::vector_valuetypes()) {
412 setTruncStoreAction(VT, InnerVT, Expand);
413 setLoadExtAction(ISD::SEXTLOAD, VT, InnerVT, Expand);
414 setLoadExtAction(ISD::ZEXTLOAD, VT, InnerVT, Expand);
415 setLoadExtAction(ISD::EXTLOAD, VT, InnerVT, Expand);
418 setOperationAction(ISD::MULHS, VT, Expand);
419 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
420 setOperationAction(ISD::MULHU, VT, Expand);
421 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
423 setOperationAction(ISD::BSWAP, VT, Expand);
426 setOperationAction(ISD::ConstantFP, MVT::f32, Custom);
427 setOperationAction(ISD::ConstantFP, MVT::f64, Custom);
429 if (Subtarget->hasNEON()) {
430 addDRTypeForNEON(MVT::v2f32);
431 addDRTypeForNEON(MVT::v8i8);
432 addDRTypeForNEON(MVT::v4i16);
433 addDRTypeForNEON(MVT::v2i32);
434 addDRTypeForNEON(MVT::v1i64);
436 addQRTypeForNEON(MVT::v4f32);
437 addQRTypeForNEON(MVT::v2f64);
438 addQRTypeForNEON(MVT::v16i8);
439 addQRTypeForNEON(MVT::v8i16);
440 addQRTypeForNEON(MVT::v4i32);
441 addQRTypeForNEON(MVT::v2i64);
443 // v2f64 is legal so that QR subregs can be extracted as f64 elements, but
444 // neither Neon nor VFP support any arithmetic operations on it.
445 // The same with v4f32. But keep in mind that vadd, vsub, vmul are natively
446 // supported for v4f32.
447 setOperationAction(ISD::FADD, MVT::v2f64, Expand);
448 setOperationAction(ISD::FSUB, MVT::v2f64, Expand);
449 setOperationAction(ISD::FMUL, MVT::v2f64, Expand);
450 // FIXME: Code duplication: FDIV and FREM are expanded always, see
451 // ARMTargetLowering::addTypeForNEON method for details.
452 setOperationAction(ISD::FDIV, MVT::v2f64, Expand);
453 setOperationAction(ISD::FREM, MVT::v2f64, Expand);
454 // FIXME: Create unittest.
455 // In another words, find a way when "copysign" appears in DAG with vector
457 setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand);
458 // FIXME: Code duplication: SETCC has custom operation action, see
459 // ARMTargetLowering::addTypeForNEON method for details.
460 setOperationAction(ISD::SETCC, MVT::v2f64, Expand);
461 // FIXME: Create unittest for FNEG and for FABS.
462 setOperationAction(ISD::FNEG, MVT::v2f64, Expand);
463 setOperationAction(ISD::FABS, MVT::v2f64, Expand);
464 setOperationAction(ISD::FSQRT, MVT::v2f64, Expand);
465 setOperationAction(ISD::FSIN, MVT::v2f64, Expand);
466 setOperationAction(ISD::FCOS, MVT::v2f64, Expand);
467 setOperationAction(ISD::FPOWI, MVT::v2f64, Expand);
468 setOperationAction(ISD::FPOW, MVT::v2f64, Expand);
469 setOperationAction(ISD::FLOG, MVT::v2f64, Expand);
470 setOperationAction(ISD::FLOG2, MVT::v2f64, Expand);
471 setOperationAction(ISD::FLOG10, MVT::v2f64, Expand);
472 setOperationAction(ISD::FEXP, MVT::v2f64, Expand);
473 setOperationAction(ISD::FEXP2, MVT::v2f64, Expand);
474 // FIXME: Create unittest for FCEIL, FTRUNC, FRINT, FNEARBYINT, FFLOOR.
475 setOperationAction(ISD::FCEIL, MVT::v2f64, Expand);
476 setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand);
477 setOperationAction(ISD::FRINT, MVT::v2f64, Expand);
478 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand);
479 setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand);
480 setOperationAction(ISD::FMA, MVT::v2f64, Expand);
482 setOperationAction(ISD::FSQRT, MVT::v4f32, Expand);
483 setOperationAction(ISD::FSIN, MVT::v4f32, Expand);
484 setOperationAction(ISD::FCOS, MVT::v4f32, Expand);
485 setOperationAction(ISD::FPOWI, MVT::v4f32, Expand);
486 setOperationAction(ISD::FPOW, MVT::v4f32, Expand);
487 setOperationAction(ISD::FLOG, MVT::v4f32, Expand);
488 setOperationAction(ISD::FLOG2, MVT::v4f32, Expand);
489 setOperationAction(ISD::FLOG10, MVT::v4f32, Expand);
490 setOperationAction(ISD::FEXP, MVT::v4f32, Expand);
491 setOperationAction(ISD::FEXP2, MVT::v4f32, Expand);
492 setOperationAction(ISD::FCEIL, MVT::v4f32, Expand);
493 setOperationAction(ISD::FTRUNC, MVT::v4f32, Expand);
494 setOperationAction(ISD::FRINT, MVT::v4f32, Expand);
495 setOperationAction(ISD::FNEARBYINT, MVT::v4f32, Expand);
496 setOperationAction(ISD::FFLOOR, MVT::v4f32, Expand);
498 // Mark v2f32 intrinsics.
499 setOperationAction(ISD::FSQRT, MVT::v2f32, Expand);
500 setOperationAction(ISD::FSIN, MVT::v2f32, Expand);
501 setOperationAction(ISD::FCOS, MVT::v2f32, Expand);
502 setOperationAction(ISD::FPOWI, MVT::v2f32, Expand);
503 setOperationAction(ISD::FPOW, MVT::v2f32, Expand);
504 setOperationAction(ISD::FLOG, MVT::v2f32, Expand);
505 setOperationAction(ISD::FLOG2, MVT::v2f32, Expand);
506 setOperationAction(ISD::FLOG10, MVT::v2f32, Expand);
507 setOperationAction(ISD::FEXP, MVT::v2f32, Expand);
508 setOperationAction(ISD::FEXP2, MVT::v2f32, Expand);
509 setOperationAction(ISD::FCEIL, MVT::v2f32, Expand);
510 setOperationAction(ISD::FTRUNC, MVT::v2f32, Expand);
511 setOperationAction(ISD::FRINT, MVT::v2f32, Expand);
512 setOperationAction(ISD::FNEARBYINT, MVT::v2f32, Expand);
513 setOperationAction(ISD::FFLOOR, MVT::v2f32, Expand);
515 // Neon does not support some operations on v1i64 and v2i64 types.
516 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
517 // Custom handling for some quad-vector types to detect VMULL.
518 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
519 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
520 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
521 // Custom handling for some vector types to avoid expensive expansions
522 setOperationAction(ISD::SDIV, MVT::v4i16, Custom);
523 setOperationAction(ISD::SDIV, MVT::v8i8, Custom);
524 setOperationAction(ISD::UDIV, MVT::v4i16, Custom);
525 setOperationAction(ISD::UDIV, MVT::v8i8, Custom);
526 setOperationAction(ISD::SETCC, MVT::v1i64, Expand);
527 setOperationAction(ISD::SETCC, MVT::v2i64, Expand);
528 // Neon does not have single instruction SINT_TO_FP and UINT_TO_FP with
529 // a destination type that is wider than the source, and nor does
530 // it have a FP_TO_[SU]INT instruction with a narrower destination than
532 setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Custom);
533 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
534 setOperationAction(ISD::FP_TO_UINT, MVT::v4i16, Custom);
535 setOperationAction(ISD::FP_TO_SINT, MVT::v4i16, Custom);
537 setOperationAction(ISD::FP_ROUND, MVT::v2f32, Expand);
538 setOperationAction(ISD::FP_EXTEND, MVT::v2f64, Expand);
540 // NEON does not have single instruction CTPOP for vectors with element
541 // types wider than 8-bits. However, custom lowering can leverage the
542 // v8i8/v16i8 vcnt instruction.
543 setOperationAction(ISD::CTPOP, MVT::v2i32, Custom);
544 setOperationAction(ISD::CTPOP, MVT::v4i32, Custom);
545 setOperationAction(ISD::CTPOP, MVT::v4i16, Custom);
546 setOperationAction(ISD::CTPOP, MVT::v8i16, Custom);
548 // NEON only has FMA instructions as of VFP4.
549 if (!Subtarget->hasVFP4()) {
550 setOperationAction(ISD::FMA, MVT::v2f32, Expand);
551 setOperationAction(ISD::FMA, MVT::v4f32, Expand);
554 setTargetDAGCombine(ISD::INTRINSIC_VOID);
555 setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN);
556 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
557 setTargetDAGCombine(ISD::SHL);
558 setTargetDAGCombine(ISD::SRL);
559 setTargetDAGCombine(ISD::SRA);
560 setTargetDAGCombine(ISD::SIGN_EXTEND);
561 setTargetDAGCombine(ISD::ZERO_EXTEND);
562 setTargetDAGCombine(ISD::ANY_EXTEND);
563 setTargetDAGCombine(ISD::SELECT_CC);
564 setTargetDAGCombine(ISD::BUILD_VECTOR);
565 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
566 setTargetDAGCombine(ISD::INSERT_VECTOR_ELT);
567 setTargetDAGCombine(ISD::STORE);
568 setTargetDAGCombine(ISD::FP_TO_SINT);
569 setTargetDAGCombine(ISD::FP_TO_UINT);
570 setTargetDAGCombine(ISD::FDIV);
571 setTargetDAGCombine(ISD::LOAD);
573 // It is legal to extload from v4i8 to v4i16 or v4i32.
574 for (MVT Ty : {MVT::v8i8, MVT::v4i8, MVT::v2i8, MVT::v4i16, MVT::v2i16,
576 for (MVT VT : MVT::integer_vector_valuetypes()) {
577 setLoadExtAction(ISD::EXTLOAD, VT, Ty, Legal);
578 setLoadExtAction(ISD::ZEXTLOAD, VT, Ty, Legal);
579 setLoadExtAction(ISD::SEXTLOAD, VT, Ty, Legal);
584 // ARM and Thumb2 support UMLAL/SMLAL.
585 if (!Subtarget->isThumb1Only())
586 setTargetDAGCombine(ISD::ADDC);
588 if (Subtarget->isFPOnlySP()) {
589 // When targetting a floating-point unit with only single-precision
590 // operations, f64 is legal for the few double-precision instructions which
591 // are present However, no double-precision operations other than moves,
592 // loads and stores are provided by the hardware.
593 setOperationAction(ISD::FADD, MVT::f64, Expand);
594 setOperationAction(ISD::FSUB, MVT::f64, Expand);
595 setOperationAction(ISD::FMUL, MVT::f64, Expand);
596 setOperationAction(ISD::FMA, MVT::f64, Expand);
597 setOperationAction(ISD::FDIV, MVT::f64, Expand);
598 setOperationAction(ISD::FREM, MVT::f64, Expand);
599 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
600 setOperationAction(ISD::FGETSIGN, MVT::f64, Expand);
601 setOperationAction(ISD::FNEG, MVT::f64, Expand);
602 setOperationAction(ISD::FABS, MVT::f64, Expand);
603 setOperationAction(ISD::FSQRT, MVT::f64, Expand);
604 setOperationAction(ISD::FSIN, MVT::f64, Expand);
605 setOperationAction(ISD::FCOS, MVT::f64, Expand);
606 setOperationAction(ISD::FPOWI, MVT::f64, Expand);
607 setOperationAction(ISD::FPOW, MVT::f64, Expand);
608 setOperationAction(ISD::FLOG, MVT::f64, Expand);
609 setOperationAction(ISD::FLOG2, MVT::f64, Expand);
610 setOperationAction(ISD::FLOG10, MVT::f64, Expand);
611 setOperationAction(ISD::FEXP, MVT::f64, Expand);
612 setOperationAction(ISD::FEXP2, MVT::f64, Expand);
613 setOperationAction(ISD::FCEIL, MVT::f64, Expand);
614 setOperationAction(ISD::FTRUNC, MVT::f64, Expand);
615 setOperationAction(ISD::FRINT, MVT::f64, Expand);
616 setOperationAction(ISD::FNEARBYINT, MVT::f64, Expand);
617 setOperationAction(ISD::FFLOOR, MVT::f64, Expand);
618 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
619 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
620 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
621 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
622 setOperationAction(ISD::FP_TO_SINT, MVT::f64, Custom);
623 setOperationAction(ISD::FP_TO_UINT, MVT::f64, Custom);
624 setOperationAction(ISD::FP_ROUND, MVT::f32, Custom);
625 setOperationAction(ISD::FP_EXTEND, MVT::f64, Custom);
628 computeRegisterProperties(Subtarget->getRegisterInfo());
630 // ARM does not have floating-point extending loads.
631 for (MVT VT : MVT::fp_valuetypes()) {
632 setLoadExtAction(ISD::EXTLOAD, VT, MVT::f32, Expand);
633 setLoadExtAction(ISD::EXTLOAD, VT, MVT::f16, Expand);
636 // ... or truncating stores
637 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
638 setTruncStoreAction(MVT::f32, MVT::f16, Expand);
639 setTruncStoreAction(MVT::f64, MVT::f16, Expand);
641 // ARM does not have i1 sign extending load.
642 for (MVT VT : MVT::integer_valuetypes())
643 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Promote);
645 // ARM supports all 4 flavors of integer indexed load / store.
646 if (!Subtarget->isThumb1Only()) {
647 for (unsigned im = (unsigned)ISD::PRE_INC;
648 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
649 setIndexedLoadAction(im, MVT::i1, Legal);
650 setIndexedLoadAction(im, MVT::i8, Legal);
651 setIndexedLoadAction(im, MVT::i16, Legal);
652 setIndexedLoadAction(im, MVT::i32, Legal);
653 setIndexedStoreAction(im, MVT::i1, Legal);
654 setIndexedStoreAction(im, MVT::i8, Legal);
655 setIndexedStoreAction(im, MVT::i16, Legal);
656 setIndexedStoreAction(im, MVT::i32, Legal);
660 setOperationAction(ISD::SADDO, MVT::i32, Custom);
661 setOperationAction(ISD::UADDO, MVT::i32, Custom);
662 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
663 setOperationAction(ISD::USUBO, MVT::i32, Custom);
665 // i64 operation support.
666 setOperationAction(ISD::MUL, MVT::i64, Expand);
667 setOperationAction(ISD::MULHU, MVT::i32, Expand);
668 if (Subtarget->isThumb1Only()) {
669 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
670 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
672 if (Subtarget->isThumb1Only() || !Subtarget->hasV6Ops()
673 || (Subtarget->isThumb2() && !Subtarget->hasThumb2DSP()))
674 setOperationAction(ISD::MULHS, MVT::i32, Expand);
676 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
677 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
678 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
679 setOperationAction(ISD::SRL, MVT::i64, Custom);
680 setOperationAction(ISD::SRA, MVT::i64, Custom);
682 if (!Subtarget->isThumb1Only()) {
683 // FIXME: We should do this for Thumb1 as well.
684 setOperationAction(ISD::ADDC, MVT::i32, Custom);
685 setOperationAction(ISD::ADDE, MVT::i32, Custom);
686 setOperationAction(ISD::SUBC, MVT::i32, Custom);
687 setOperationAction(ISD::SUBE, MVT::i32, Custom);
690 // ARM does not have ROTL.
691 setOperationAction(ISD::ROTL, MVT::i32, Expand);
692 setOperationAction(ISD::CTTZ, MVT::i32, Custom);
693 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
694 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
695 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
697 // These just redirect to CTTZ and CTLZ on ARM.
698 setOperationAction(ISD::CTTZ_ZERO_UNDEF , MVT::i32 , Expand);
699 setOperationAction(ISD::CTLZ_ZERO_UNDEF , MVT::i32 , Expand);
701 setOperationAction(ISD::READCYCLECOUNTER, MVT::i64, Custom);
703 // Only ARMv6 has BSWAP.
704 if (!Subtarget->hasV6Ops())
705 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
707 if (!(Subtarget->hasDivide() && Subtarget->isThumb2()) &&
708 !(Subtarget->hasDivideInARMMode() && !Subtarget->isThumb())) {
709 // These are expanded into libcalls if the cpu doesn't have HW divider.
710 setOperationAction(ISD::SDIV, MVT::i32, Expand);
711 setOperationAction(ISD::UDIV, MVT::i32, Expand);
714 // FIXME: Also set divmod for SREM on EABI
715 setOperationAction(ISD::SREM, MVT::i32, Expand);
716 setOperationAction(ISD::UREM, MVT::i32, Expand);
717 // Register based DivRem for AEABI (RTABI 4.2)
718 if (Subtarget->isTargetAEABI()) {
719 setLibcallName(RTLIB::SDIVREM_I8, "__aeabi_idivmod");
720 setLibcallName(RTLIB::SDIVREM_I16, "__aeabi_idivmod");
721 setLibcallName(RTLIB::SDIVREM_I32, "__aeabi_idivmod");
722 setLibcallName(RTLIB::SDIVREM_I64, "__aeabi_ldivmod");
723 setLibcallName(RTLIB::UDIVREM_I8, "__aeabi_uidivmod");
724 setLibcallName(RTLIB::UDIVREM_I16, "__aeabi_uidivmod");
725 setLibcallName(RTLIB::UDIVREM_I32, "__aeabi_uidivmod");
726 setLibcallName(RTLIB::UDIVREM_I64, "__aeabi_uldivmod");
728 setLibcallCallingConv(RTLIB::SDIVREM_I8, CallingConv::ARM_AAPCS);
729 setLibcallCallingConv(RTLIB::SDIVREM_I16, CallingConv::ARM_AAPCS);
730 setLibcallCallingConv(RTLIB::SDIVREM_I32, CallingConv::ARM_AAPCS);
731 setLibcallCallingConv(RTLIB::SDIVREM_I64, CallingConv::ARM_AAPCS);
732 setLibcallCallingConv(RTLIB::UDIVREM_I8, CallingConv::ARM_AAPCS);
733 setLibcallCallingConv(RTLIB::UDIVREM_I16, CallingConv::ARM_AAPCS);
734 setLibcallCallingConv(RTLIB::UDIVREM_I32, CallingConv::ARM_AAPCS);
735 setLibcallCallingConv(RTLIB::UDIVREM_I64, CallingConv::ARM_AAPCS);
737 setOperationAction(ISD::SDIVREM, MVT::i32, Custom);
738 setOperationAction(ISD::UDIVREM, MVT::i32, Custom);
740 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
741 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
744 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
745 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
746 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
747 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
748 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
750 setOperationAction(ISD::TRAP, MVT::Other, Legal);
752 // Use the default implementation.
753 setOperationAction(ISD::VASTART, MVT::Other, Custom);
754 setOperationAction(ISD::VAARG, MVT::Other, Expand);
755 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
756 setOperationAction(ISD::VAEND, MVT::Other, Expand);
757 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
758 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
760 if (!Subtarget->isTargetMachO()) {
761 // Non-MachO platforms may return values in these registers via the
762 // personality function.
763 setExceptionPointerRegister(ARM::R0);
764 setExceptionSelectorRegister(ARM::R1);
767 if (Subtarget->getTargetTriple().isWindowsItaniumEnvironment())
768 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
770 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
772 // ARMv6 Thumb1 (except for CPUs that support dmb / dsb) and earlier use
773 // the default expansion. If we are targeting a single threaded system,
774 // then set them all for expand so we can lower them later into their
776 if (TM.Options.ThreadModel == ThreadModel::Single)
777 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Expand);
778 else if (Subtarget->hasAnyDataBarrier() && !Subtarget->isThumb1Only()) {
779 // ATOMIC_FENCE needs custom lowering; the others should have been expanded
780 // to ldrex/strex loops already.
781 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Custom);
783 // On v8, we have particularly efficient implementations of atomic fences
784 // if they can be combined with nearby atomic loads and stores.
785 if (!Subtarget->hasV8Ops()) {
786 // Automatically insert fences (dmb ish) around ATOMIC_SWAP etc.
787 setInsertFencesForAtomic(true);
790 // If there's anything we can use as a barrier, go through custom lowering
792 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other,
793 Subtarget->hasAnyDataBarrier() ? Custom : Expand);
795 // Set them all for expansion, which will force libcalls.
796 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Expand);
797 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Expand);
798 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Expand);
799 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Expand);
800 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Expand);
801 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Expand);
802 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Expand);
803 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Expand);
804 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i32, Expand);
805 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i32, Expand);
806 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i32, Expand);
807 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i32, Expand);
808 // Mark ATOMIC_LOAD and ATOMIC_STORE custom so we can handle the
809 // Unordered/Monotonic case.
810 setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Custom);
811 setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Custom);
814 setOperationAction(ISD::PREFETCH, MVT::Other, Custom);
816 // Requires SXTB/SXTH, available on v6 and up in both ARM and Thumb modes.
817 if (!Subtarget->hasV6Ops()) {
818 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
819 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
821 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
823 if (!Subtarget->useSoftFloat() && Subtarget->hasVFP2() &&
824 !Subtarget->isThumb1Only()) {
825 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
826 // iff target supports vfp2.
827 setOperationAction(ISD::BITCAST, MVT::i64, Custom);
828 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
831 // We want to custom lower some of our intrinsics.
832 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
833 if (Subtarget->isTargetDarwin()) {
834 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
835 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
836 setLibcallName(RTLIB::UNWIND_RESUME, "_Unwind_SjLj_Resume");
839 setOperationAction(ISD::SETCC, MVT::i32, Expand);
840 setOperationAction(ISD::SETCC, MVT::f32, Expand);
841 setOperationAction(ISD::SETCC, MVT::f64, Expand);
842 setOperationAction(ISD::SELECT, MVT::i32, Custom);
843 setOperationAction(ISD::SELECT, MVT::f32, Custom);
844 setOperationAction(ISD::SELECT, MVT::f64, Custom);
845 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
846 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
847 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
849 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
850 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
851 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
852 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
853 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
855 // We don't support sin/cos/fmod/copysign/pow
856 setOperationAction(ISD::FSIN, MVT::f64, Expand);
857 setOperationAction(ISD::FSIN, MVT::f32, Expand);
858 setOperationAction(ISD::FCOS, MVT::f32, Expand);
859 setOperationAction(ISD::FCOS, MVT::f64, Expand);
860 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
861 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
862 setOperationAction(ISD::FREM, MVT::f64, Expand);
863 setOperationAction(ISD::FREM, MVT::f32, Expand);
864 if (!Subtarget->useSoftFloat() && Subtarget->hasVFP2() &&
865 !Subtarget->isThumb1Only()) {
866 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
867 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
869 setOperationAction(ISD::FPOW, MVT::f64, Expand);
870 setOperationAction(ISD::FPOW, MVT::f32, Expand);
872 if (!Subtarget->hasVFP4()) {
873 setOperationAction(ISD::FMA, MVT::f64, Expand);
874 setOperationAction(ISD::FMA, MVT::f32, Expand);
877 // Various VFP goodness
878 if (!Subtarget->useSoftFloat() && !Subtarget->isThumb1Only()) {
879 // FP-ARMv8 adds f64 <-> f16 conversion. Before that it should be expanded.
880 if (!Subtarget->hasFPARMv8() || Subtarget->isFPOnlySP()) {
881 setOperationAction(ISD::FP16_TO_FP, MVT::f64, Expand);
882 setOperationAction(ISD::FP_TO_FP16, MVT::f64, Expand);
885 // fp16 is a special v7 extension that adds f16 <-> f32 conversions.
886 if (!Subtarget->hasFP16()) {
887 setOperationAction(ISD::FP16_TO_FP, MVT::f32, Expand);
888 setOperationAction(ISD::FP_TO_FP16, MVT::f32, Expand);
892 // Combine sin / cos into one node or libcall if possible.
893 if (Subtarget->hasSinCos()) {
894 setLibcallName(RTLIB::SINCOS_F32, "sincosf");
895 setLibcallName(RTLIB::SINCOS_F64, "sincos");
896 if (Subtarget->getTargetTriple().isiOS()) {
897 // For iOS, we don't want to the normal expansion of a libcall to
898 // sincos. We want to issue a libcall to __sincos_stret.
899 setOperationAction(ISD::FSINCOS, MVT::f64, Custom);
900 setOperationAction(ISD::FSINCOS, MVT::f32, Custom);
904 // FP-ARMv8 implements a lot of rounding-like FP operations.
905 if (Subtarget->hasFPARMv8()) {
906 setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
907 setOperationAction(ISD::FCEIL, MVT::f32, Legal);
908 setOperationAction(ISD::FROUND, MVT::f32, Legal);
909 setOperationAction(ISD::FTRUNC, MVT::f32, Legal);
910 setOperationAction(ISD::FNEARBYINT, MVT::f32, Legal);
911 setOperationAction(ISD::FRINT, MVT::f32, Legal);
912 if (!Subtarget->isFPOnlySP()) {
913 setOperationAction(ISD::FFLOOR, MVT::f64, Legal);
914 setOperationAction(ISD::FCEIL, MVT::f64, Legal);
915 setOperationAction(ISD::FROUND, MVT::f64, Legal);
916 setOperationAction(ISD::FTRUNC, MVT::f64, Legal);
917 setOperationAction(ISD::FNEARBYINT, MVT::f64, Legal);
918 setOperationAction(ISD::FRINT, MVT::f64, Legal);
921 // We have target-specific dag combine patterns for the following nodes:
922 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
923 setTargetDAGCombine(ISD::ADD);
924 setTargetDAGCombine(ISD::SUB);
925 setTargetDAGCombine(ISD::MUL);
926 setTargetDAGCombine(ISD::AND);
927 setTargetDAGCombine(ISD::OR);
928 setTargetDAGCombine(ISD::XOR);
930 if (Subtarget->hasV6Ops())
931 setTargetDAGCombine(ISD::SRL);
933 setStackPointerRegisterToSaveRestore(ARM::SP);
935 if (Subtarget->useSoftFloat() || Subtarget->isThumb1Only() ||
936 !Subtarget->hasVFP2())
937 setSchedulingPreference(Sched::RegPressure);
939 setSchedulingPreference(Sched::Hybrid);
941 //// temporary - rewrite interface to use type
942 MaxStoresPerMemset = 8;
943 MaxStoresPerMemsetOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
944 MaxStoresPerMemcpy = 4; // For @llvm.memcpy -> sequence of stores
945 MaxStoresPerMemcpyOptSize = Subtarget->isTargetDarwin() ? 4 : 2;
946 MaxStoresPerMemmove = 4; // For @llvm.memmove -> sequence of stores
947 MaxStoresPerMemmoveOptSize = Subtarget->isTargetDarwin() ? 4 : 2;
949 // On ARM arguments smaller than 4 bytes are extended, so all arguments
950 // are at least 4 bytes aligned.
951 setMinStackArgumentAlignment(4);
953 // Prefer likely predicted branches to selects on out-of-order cores.
954 PredictableSelectIsExpensive = Subtarget->isLikeA9();
956 setMinFunctionAlignment(Subtarget->isThumb() ? 1 : 2);
959 bool ARMTargetLowering::useSoftFloat() const {
960 return Subtarget->useSoftFloat();
963 // FIXME: It might make sense to define the representative register class as the
964 // nearest super-register that has a non-null superset. For example, DPR_VFP2 is
965 // a super-register of SPR, and DPR is a superset if DPR_VFP2. Consequently,
966 // SPR's representative would be DPR_VFP2. This should work well if register
967 // pressure tracking were modified such that a register use would increment the
968 // pressure of the register class's representative and all of it's super
969 // classes' representatives transitively. We have not implemented this because
970 // of the difficulty prior to coalescing of modeling operand register classes
971 // due to the common occurrence of cross class copies and subregister insertions
973 std::pair<const TargetRegisterClass *, uint8_t>
974 ARMTargetLowering::findRepresentativeClass(const TargetRegisterInfo *TRI,
976 const TargetRegisterClass *RRC = nullptr;
978 switch (VT.SimpleTy) {
980 return TargetLowering::findRepresentativeClass(TRI, VT);
981 // Use DPR as representative register class for all floating point
982 // and vector types. Since there are 32 SPR registers and 32 DPR registers so
983 // the cost is 1 for both f32 and f64.
984 case MVT::f32: case MVT::f64: case MVT::v8i8: case MVT::v4i16:
985 case MVT::v2i32: case MVT::v1i64: case MVT::v2f32:
986 RRC = &ARM::DPRRegClass;
987 // When NEON is used for SP, only half of the register file is available
988 // because operations that define both SP and DP results will be constrained
989 // to the VFP2 class (D0-D15). We currently model this constraint prior to
990 // coalescing by double-counting the SP regs. See the FIXME above.
991 if (Subtarget->useNEONForSinglePrecisionFP())
994 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
995 case MVT::v4f32: case MVT::v2f64:
996 RRC = &ARM::DPRRegClass;
1000 RRC = &ARM::DPRRegClass;
1004 RRC = &ARM::DPRRegClass;
1008 return std::make_pair(RRC, Cost);
1011 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
1012 switch ((ARMISD::NodeType)Opcode) {
1013 case ARMISD::FIRST_NUMBER: break;
1014 case ARMISD::Wrapper: return "ARMISD::Wrapper";
1015 case ARMISD::WrapperPIC: return "ARMISD::WrapperPIC";
1016 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
1017 case ARMISD::COPY_STRUCT_BYVAL: return "ARMISD::COPY_STRUCT_BYVAL";
1018 case ARMISD::CALL: return "ARMISD::CALL";
1019 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
1020 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
1021 case ARMISD::tCALL: return "ARMISD::tCALL";
1022 case ARMISD::BRCOND: return "ARMISD::BRCOND";
1023 case ARMISD::BR_JT: return "ARMISD::BR_JT";
1024 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
1025 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
1026 case ARMISD::INTRET_FLAG: return "ARMISD::INTRET_FLAG";
1027 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
1028 case ARMISD::CMP: return "ARMISD::CMP";
1029 case ARMISD::CMN: return "ARMISD::CMN";
1030 case ARMISD::CMPZ: return "ARMISD::CMPZ";
1031 case ARMISD::CMPFP: return "ARMISD::CMPFP";
1032 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
1033 case ARMISD::BCC_i64: return "ARMISD::BCC_i64";
1034 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
1036 case ARMISD::CMOV: return "ARMISD::CMOV";
1038 case ARMISD::RBIT: return "ARMISD::RBIT";
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::VSLI: return "ARMISD::VSLI";
1095 case ARMISD::VSRI: return "ARMISD::VSRI";
1096 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
1097 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
1098 case ARMISD::VMOVIMM: return "ARMISD::VMOVIMM";
1099 case ARMISD::VMVNIMM: return "ARMISD::VMVNIMM";
1100 case ARMISD::VMOVFPIMM: return "ARMISD::VMOVFPIMM";
1101 case ARMISD::VDUP: return "ARMISD::VDUP";
1102 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
1103 case ARMISD::VEXT: return "ARMISD::VEXT";
1104 case ARMISD::VREV64: return "ARMISD::VREV64";
1105 case ARMISD::VREV32: return "ARMISD::VREV32";
1106 case ARMISD::VREV16: return "ARMISD::VREV16";
1107 case ARMISD::VZIP: return "ARMISD::VZIP";
1108 case ARMISD::VUZP: return "ARMISD::VUZP";
1109 case ARMISD::VTRN: return "ARMISD::VTRN";
1110 case ARMISD::VTBL1: return "ARMISD::VTBL1";
1111 case ARMISD::VTBL2: return "ARMISD::VTBL2";
1112 case ARMISD::VMULLs: return "ARMISD::VMULLs";
1113 case ARMISD::VMULLu: return "ARMISD::VMULLu";
1114 case ARMISD::UMLAL: return "ARMISD::UMLAL";
1115 case ARMISD::SMLAL: return "ARMISD::SMLAL";
1116 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
1117 case ARMISD::FMAX: return "ARMISD::FMAX";
1118 case ARMISD::FMIN: return "ARMISD::FMIN";
1119 case ARMISD::VMAXNM: return "ARMISD::VMAX";
1120 case ARMISD::VMINNM: return "ARMISD::VMIN";
1121 case ARMISD::BFI: return "ARMISD::BFI";
1122 case ARMISD::VORRIMM: return "ARMISD::VORRIMM";
1123 case ARMISD::VBICIMM: return "ARMISD::VBICIMM";
1124 case ARMISD::VBSL: return "ARMISD::VBSL";
1125 case ARMISD::VLD2DUP: return "ARMISD::VLD2DUP";
1126 case ARMISD::VLD3DUP: return "ARMISD::VLD3DUP";
1127 case ARMISD::VLD4DUP: return "ARMISD::VLD4DUP";
1128 case ARMISD::VLD1_UPD: return "ARMISD::VLD1_UPD";
1129 case ARMISD::VLD2_UPD: return "ARMISD::VLD2_UPD";
1130 case ARMISD::VLD3_UPD: return "ARMISD::VLD3_UPD";
1131 case ARMISD::VLD4_UPD: return "ARMISD::VLD4_UPD";
1132 case ARMISD::VLD2LN_UPD: return "ARMISD::VLD2LN_UPD";
1133 case ARMISD::VLD3LN_UPD: return "ARMISD::VLD3LN_UPD";
1134 case ARMISD::VLD4LN_UPD: return "ARMISD::VLD4LN_UPD";
1135 case ARMISD::VLD2DUP_UPD: return "ARMISD::VLD2DUP_UPD";
1136 case ARMISD::VLD3DUP_UPD: return "ARMISD::VLD3DUP_UPD";
1137 case ARMISD::VLD4DUP_UPD: return "ARMISD::VLD4DUP_UPD";
1138 case ARMISD::VST1_UPD: return "ARMISD::VST1_UPD";
1139 case ARMISD::VST2_UPD: return "ARMISD::VST2_UPD";
1140 case ARMISD::VST3_UPD: return "ARMISD::VST3_UPD";
1141 case ARMISD::VST4_UPD: return "ARMISD::VST4_UPD";
1142 case ARMISD::VST2LN_UPD: return "ARMISD::VST2LN_UPD";
1143 case ARMISD::VST3LN_UPD: return "ARMISD::VST3LN_UPD";
1144 case ARMISD::VST4LN_UPD: return "ARMISD::VST4LN_UPD";
1149 EVT ARMTargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
1150 if (!VT.isVector()) return getPointerTy();
1151 return VT.changeVectorElementTypeToInteger();
1154 /// getRegClassFor - Return the register class that should be used for the
1155 /// specified value type.
1156 const TargetRegisterClass *ARMTargetLowering::getRegClassFor(MVT VT) const {
1157 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
1158 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
1159 // load / store 4 to 8 consecutive D registers.
1160 if (Subtarget->hasNEON()) {
1161 if (VT == MVT::v4i64)
1162 return &ARM::QQPRRegClass;
1163 if (VT == MVT::v8i64)
1164 return &ARM::QQQQPRRegClass;
1166 return TargetLowering::getRegClassFor(VT);
1169 // memcpy, and other memory intrinsics, typically tries to use LDM/STM if the
1170 // source/dest is aligned and the copy size is large enough. We therefore want
1171 // to align such objects passed to memory intrinsics.
1172 bool ARMTargetLowering::shouldAlignPointerArgs(CallInst *CI, unsigned &MinSize,
1173 unsigned &PrefAlign) const {
1174 if (!isa<MemIntrinsic>(CI))
1177 // On ARM11 onwards (excluding M class) 8-byte aligned LDM is typically 1
1178 // cycle faster than 4-byte aligned LDM.
1179 PrefAlign = (Subtarget->hasV6Ops() && !Subtarget->isMClass() ? 8 : 4);
1183 // Create a fast isel object.
1185 ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
1186 const TargetLibraryInfo *libInfo) const {
1187 return ARM::createFastISel(funcInfo, libInfo);
1190 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
1191 unsigned NumVals = N->getNumValues();
1193 return Sched::RegPressure;
1195 for (unsigned i = 0; i != NumVals; ++i) {
1196 EVT VT = N->getValueType(i);
1197 if (VT == MVT::Glue || VT == MVT::Other)
1199 if (VT.isFloatingPoint() || VT.isVector())
1203 if (!N->isMachineOpcode())
1204 return Sched::RegPressure;
1206 // Load are scheduled for latency even if there instruction itinerary
1207 // is not available.
1208 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
1209 const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
1211 if (MCID.getNumDefs() == 0)
1212 return Sched::RegPressure;
1213 if (!Itins->isEmpty() &&
1214 Itins->getOperandCycle(MCID.getSchedClass(), 0) > 2)
1217 return Sched::RegPressure;
1220 //===----------------------------------------------------------------------===//
1222 //===----------------------------------------------------------------------===//
1224 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
1225 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
1227 default: llvm_unreachable("Unknown condition code!");
1228 case ISD::SETNE: return ARMCC::NE;
1229 case ISD::SETEQ: return ARMCC::EQ;
1230 case ISD::SETGT: return ARMCC::GT;
1231 case ISD::SETGE: return ARMCC::GE;
1232 case ISD::SETLT: return ARMCC::LT;
1233 case ISD::SETLE: return ARMCC::LE;
1234 case ISD::SETUGT: return ARMCC::HI;
1235 case ISD::SETUGE: return ARMCC::HS;
1236 case ISD::SETULT: return ARMCC::LO;
1237 case ISD::SETULE: return ARMCC::LS;
1241 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
1242 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
1243 ARMCC::CondCodes &CondCode2) {
1244 CondCode2 = ARMCC::AL;
1246 default: llvm_unreachable("Unknown FP condition!");
1248 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
1250 case ISD::SETOGT: CondCode = ARMCC::GT; break;
1252 case ISD::SETOGE: CondCode = ARMCC::GE; break;
1253 case ISD::SETOLT: CondCode = ARMCC::MI; break;
1254 case ISD::SETOLE: CondCode = ARMCC::LS; break;
1255 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
1256 case ISD::SETO: CondCode = ARMCC::VC; break;
1257 case ISD::SETUO: CondCode = ARMCC::VS; break;
1258 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
1259 case ISD::SETUGT: CondCode = ARMCC::HI; break;
1260 case ISD::SETUGE: CondCode = ARMCC::PL; break;
1262 case ISD::SETULT: CondCode = ARMCC::LT; break;
1264 case ISD::SETULE: CondCode = ARMCC::LE; break;
1266 case ISD::SETUNE: CondCode = ARMCC::NE; break;
1270 //===----------------------------------------------------------------------===//
1271 // Calling Convention Implementation
1272 //===----------------------------------------------------------------------===//
1274 #include "ARMGenCallingConv.inc"
1276 /// getEffectiveCallingConv - Get the effective calling convention, taking into
1277 /// account presence of floating point hardware and calling convention
1278 /// limitations, such as support for variadic functions.
1280 ARMTargetLowering::getEffectiveCallingConv(CallingConv::ID CC,
1281 bool isVarArg) const {
1284 llvm_unreachable("Unsupported calling convention");
1285 case CallingConv::ARM_AAPCS:
1286 case CallingConv::ARM_APCS:
1287 case CallingConv::GHC:
1289 case CallingConv::ARM_AAPCS_VFP:
1290 return isVarArg ? CallingConv::ARM_AAPCS : CallingConv::ARM_AAPCS_VFP;
1291 case CallingConv::C:
1292 if (!Subtarget->isAAPCS_ABI())
1293 return CallingConv::ARM_APCS;
1294 else if (Subtarget->hasVFP2() && !Subtarget->isThumb1Only() &&
1295 getTargetMachine().Options.FloatABIType == FloatABI::Hard &&
1297 return CallingConv::ARM_AAPCS_VFP;
1299 return CallingConv::ARM_AAPCS;
1300 case CallingConv::Fast:
1301 if (!Subtarget->isAAPCS_ABI()) {
1302 if (Subtarget->hasVFP2() && !Subtarget->isThumb1Only() && !isVarArg)
1303 return CallingConv::Fast;
1304 return CallingConv::ARM_APCS;
1305 } else if (Subtarget->hasVFP2() && !Subtarget->isThumb1Only() && !isVarArg)
1306 return CallingConv::ARM_AAPCS_VFP;
1308 return CallingConv::ARM_AAPCS;
1312 /// CCAssignFnForNode - Selects the correct CCAssignFn for the given
1313 /// CallingConvention.
1314 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
1316 bool isVarArg) const {
1317 switch (getEffectiveCallingConv(CC, isVarArg)) {
1319 llvm_unreachable("Unsupported calling convention");
1320 case CallingConv::ARM_APCS:
1321 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1322 case CallingConv::ARM_AAPCS:
1323 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1324 case CallingConv::ARM_AAPCS_VFP:
1325 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1326 case CallingConv::Fast:
1327 return (Return ? RetFastCC_ARM_APCS : FastCC_ARM_APCS);
1328 case CallingConv::GHC:
1329 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS_GHC);
1333 /// LowerCallResult - Lower the result values of a call into the
1334 /// appropriate copies out of appropriate physical registers.
1336 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1337 CallingConv::ID CallConv, bool isVarArg,
1338 const SmallVectorImpl<ISD::InputArg> &Ins,
1339 SDLoc dl, SelectionDAG &DAG,
1340 SmallVectorImpl<SDValue> &InVals,
1341 bool isThisReturn, SDValue ThisVal) const {
1343 // Assign locations to each value returned by this call.
1344 SmallVector<CCValAssign, 16> RVLocs;
1345 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
1346 *DAG.getContext(), Call);
1347 CCInfo.AnalyzeCallResult(Ins,
1348 CCAssignFnForNode(CallConv, /* Return*/ true,
1351 // Copy all of the result registers out of their specified physreg.
1352 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1353 CCValAssign VA = RVLocs[i];
1355 // Pass 'this' value directly from the argument to return value, to avoid
1356 // reg unit interference
1357 if (i == 0 && isThisReturn) {
1358 assert(!VA.needsCustom() && VA.getLocVT() == MVT::i32 &&
1359 "unexpected return calling convention register assignment");
1360 InVals.push_back(ThisVal);
1365 if (VA.needsCustom()) {
1366 // Handle f64 or half of a v2f64.
1367 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1369 Chain = Lo.getValue(1);
1370 InFlag = Lo.getValue(2);
1371 VA = RVLocs[++i]; // skip ahead to next loc
1372 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1374 Chain = Hi.getValue(1);
1375 InFlag = Hi.getValue(2);
1376 if (!Subtarget->isLittle())
1378 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1380 if (VA.getLocVT() == MVT::v2f64) {
1381 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1382 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1383 DAG.getConstant(0, dl, MVT::i32));
1385 VA = RVLocs[++i]; // skip ahead to next loc
1386 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1387 Chain = Lo.getValue(1);
1388 InFlag = Lo.getValue(2);
1389 VA = RVLocs[++i]; // skip ahead to next loc
1390 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1391 Chain = Hi.getValue(1);
1392 InFlag = Hi.getValue(2);
1393 if (!Subtarget->isLittle())
1395 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1396 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1397 DAG.getConstant(1, dl, MVT::i32));
1400 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
1402 Chain = Val.getValue(1);
1403 InFlag = Val.getValue(2);
1406 switch (VA.getLocInfo()) {
1407 default: llvm_unreachable("Unknown loc info!");
1408 case CCValAssign::Full: break;
1409 case CCValAssign::BCvt:
1410 Val = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), Val);
1414 InVals.push_back(Val);
1420 /// LowerMemOpCallTo - Store the argument to the stack.
1422 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
1423 SDValue StackPtr, SDValue Arg,
1424 SDLoc dl, SelectionDAG &DAG,
1425 const CCValAssign &VA,
1426 ISD::ArgFlagsTy Flags) const {
1427 unsigned LocMemOffset = VA.getLocMemOffset();
1428 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset, dl);
1429 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1430 return DAG.getStore(Chain, dl, Arg, PtrOff,
1431 MachinePointerInfo::getStack(LocMemOffset),
1435 void ARMTargetLowering::PassF64ArgInRegs(SDLoc dl, SelectionDAG &DAG,
1436 SDValue Chain, SDValue &Arg,
1437 RegsToPassVector &RegsToPass,
1438 CCValAssign &VA, CCValAssign &NextVA,
1440 SmallVectorImpl<SDValue> &MemOpChains,
1441 ISD::ArgFlagsTy Flags) const {
1443 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1444 DAG.getVTList(MVT::i32, MVT::i32), Arg);
1445 unsigned id = Subtarget->isLittle() ? 0 : 1;
1446 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd.getValue(id)));
1448 if (NextVA.isRegLoc())
1449 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1-id)));
1451 assert(NextVA.isMemLoc());
1452 if (!StackPtr.getNode())
1453 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1455 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1-id),
1461 /// LowerCall - Lowering a call into a callseq_start <-
1462 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
1465 ARMTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
1466 SmallVectorImpl<SDValue> &InVals) const {
1467 SelectionDAG &DAG = CLI.DAG;
1469 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
1470 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
1471 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
1472 SDValue Chain = CLI.Chain;
1473 SDValue Callee = CLI.Callee;
1474 bool &isTailCall = CLI.IsTailCall;
1475 CallingConv::ID CallConv = CLI.CallConv;
1476 bool doesNotRet = CLI.DoesNotReturn;
1477 bool isVarArg = CLI.IsVarArg;
1479 MachineFunction &MF = DAG.getMachineFunction();
1480 bool isStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1481 bool isThisReturn = false;
1482 bool isSibCall = false;
1484 // Disable tail calls if they're not supported.
1485 if (!Subtarget->supportsTailCall() || MF.getTarget().Options.DisableTailCalls)
1489 // Check if it's really possible to do a tail call.
1490 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1491 isVarArg, isStructRet, MF.getFunction()->hasStructRetAttr(),
1492 Outs, OutVals, Ins, DAG);
1493 if (!isTailCall && CLI.CS && CLI.CS->isMustTailCall())
1494 report_fatal_error("failed to perform tail call elimination on a call "
1495 "site marked musttail");
1496 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1497 // detected sibcalls.
1504 // Analyze operands of the call, assigning locations to each operand.
1505 SmallVector<CCValAssign, 16> ArgLocs;
1506 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs,
1507 *DAG.getContext(), Call);
1508 CCInfo.AnalyzeCallOperands(Outs,
1509 CCAssignFnForNode(CallConv, /* Return*/ false,
1512 // Get a count of how many bytes are to be pushed on the stack.
1513 unsigned NumBytes = CCInfo.getNextStackOffset();
1515 // For tail calls, memory operands are available in our caller's stack.
1519 // Adjust the stack pointer for the new arguments...
1520 // These operations are automatically eliminated by the prolog/epilog pass
1522 Chain = DAG.getCALLSEQ_START(Chain,
1523 DAG.getIntPtrConstant(NumBytes, dl, true), dl);
1525 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1527 RegsToPassVector RegsToPass;
1528 SmallVector<SDValue, 8> MemOpChains;
1530 // Walk the register/memloc assignments, inserting copies/loads. In the case
1531 // of tail call optimization, arguments are handled later.
1532 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1534 ++i, ++realArgIdx) {
1535 CCValAssign &VA = ArgLocs[i];
1536 SDValue Arg = OutVals[realArgIdx];
1537 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1538 bool isByVal = Flags.isByVal();
1540 // Promote the value if needed.
1541 switch (VA.getLocInfo()) {
1542 default: llvm_unreachable("Unknown loc info!");
1543 case CCValAssign::Full: break;
1544 case CCValAssign::SExt:
1545 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1547 case CCValAssign::ZExt:
1548 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1550 case CCValAssign::AExt:
1551 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1553 case CCValAssign::BCvt:
1554 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1558 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1559 if (VA.needsCustom()) {
1560 if (VA.getLocVT() == MVT::v2f64) {
1561 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1562 DAG.getConstant(0, dl, MVT::i32));
1563 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1564 DAG.getConstant(1, dl, MVT::i32));
1566 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1567 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1569 VA = ArgLocs[++i]; // skip ahead to next loc
1570 if (VA.isRegLoc()) {
1571 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1572 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1574 assert(VA.isMemLoc());
1576 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1577 dl, DAG, VA, Flags));
1580 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1581 StackPtr, MemOpChains, Flags);
1583 } else if (VA.isRegLoc()) {
1584 if (realArgIdx == 0 && Flags.isReturned() && Outs[0].VT == MVT::i32) {
1585 assert(VA.getLocVT() == MVT::i32 &&
1586 "unexpected calling convention register assignment");
1587 assert(!Ins.empty() && Ins[0].VT == MVT::i32 &&
1588 "unexpected use of 'returned'");
1589 isThisReturn = true;
1591 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1592 } else if (isByVal) {
1593 assert(VA.isMemLoc());
1594 unsigned offset = 0;
1596 // True if this byval aggregate will be split between registers
1598 unsigned ByValArgsCount = CCInfo.getInRegsParamsCount();
1599 unsigned CurByValIdx = CCInfo.getInRegsParamsProcessed();
1601 if (CurByValIdx < ByValArgsCount) {
1603 unsigned RegBegin, RegEnd;
1604 CCInfo.getInRegsParamInfo(CurByValIdx, RegBegin, RegEnd);
1606 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1608 for (i = 0, j = RegBegin; j < RegEnd; i++, j++) {
1609 SDValue Const = DAG.getConstant(4*i, dl, MVT::i32);
1610 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
1611 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
1612 MachinePointerInfo(),
1613 false, false, false,
1614 DAG.InferPtrAlignment(AddArg));
1615 MemOpChains.push_back(Load.getValue(1));
1616 RegsToPass.push_back(std::make_pair(j, Load));
1619 // If parameter size outsides register area, "offset" value
1620 // helps us to calculate stack slot for remained part properly.
1621 offset = RegEnd - RegBegin;
1623 CCInfo.nextInRegsParam();
1626 if (Flags.getByValSize() > 4*offset) {
1627 unsigned LocMemOffset = VA.getLocMemOffset();
1628 SDValue StkPtrOff = DAG.getIntPtrConstant(LocMemOffset, dl);
1629 SDValue Dst = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr,
1631 SDValue SrcOffset = DAG.getIntPtrConstant(4*offset, dl);
1632 SDValue Src = DAG.getNode(ISD::ADD, dl, getPointerTy(), Arg, SrcOffset);
1633 SDValue SizeNode = DAG.getConstant(Flags.getByValSize() - 4*offset, dl,
1635 SDValue AlignNode = DAG.getConstant(Flags.getByValAlign(), dl,
1638 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
1639 SDValue Ops[] = { Chain, Dst, Src, SizeNode, AlignNode};
1640 MemOpChains.push_back(DAG.getNode(ARMISD::COPY_STRUCT_BYVAL, dl, VTs,
1643 } else if (!isSibCall) {
1644 assert(VA.isMemLoc());
1646 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1647 dl, DAG, VA, Flags));
1651 if (!MemOpChains.empty())
1652 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains);
1654 // Build a sequence of copy-to-reg nodes chained together with token chain
1655 // and flag operands which copy the outgoing args into the appropriate regs.
1657 // Tail call byval lowering might overwrite argument registers so in case of
1658 // tail call optimization the copies to registers are lowered later.
1660 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1661 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1662 RegsToPass[i].second, InFlag);
1663 InFlag = Chain.getValue(1);
1666 // For tail calls lower the arguments to the 'real' stack slot.
1668 // Force all the incoming stack arguments to be loaded from the stack
1669 // before any new outgoing arguments are stored to the stack, because the
1670 // outgoing stack slots may alias the incoming argument stack slots, and
1671 // the alias isn't otherwise explicit. This is slightly more conservative
1672 // than necessary, because it means that each store effectively depends
1673 // on every argument instead of just those arguments it would clobber.
1675 // Do not flag preceding copytoreg stuff together with the following stuff.
1677 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1678 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1679 RegsToPass[i].second, InFlag);
1680 InFlag = Chain.getValue(1);
1685 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1686 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1687 // node so that legalize doesn't hack it.
1688 bool isDirect = false;
1689 bool isARMFunc = false;
1690 bool isLocalARMFunc = false;
1691 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1693 if (EnableARMLongCalls) {
1694 assert((Subtarget->isTargetWindows() ||
1695 getTargetMachine().getRelocationModel() == Reloc::Static) &&
1696 "long-calls with non-static relocation model!");
1697 // Handle a global address or an external symbol. If it's not one of
1698 // those, the target's already in a register, so we don't need to do
1700 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1701 const GlobalValue *GV = G->getGlobal();
1702 // Create a constant pool entry for the callee address
1703 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1704 ARMConstantPoolValue *CPV =
1705 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 0);
1707 // Get the address of the callee into a register
1708 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1709 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1710 Callee = DAG.getLoad(getPointerTy(), dl,
1711 DAG.getEntryNode(), CPAddr,
1712 MachinePointerInfo::getConstantPool(),
1713 false, false, false, 0);
1714 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1715 const char *Sym = S->getSymbol();
1717 // Create a constant pool entry for the callee address
1718 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1719 ARMConstantPoolValue *CPV =
1720 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1721 ARMPCLabelIndex, 0);
1722 // Get the address of the callee into a register
1723 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1724 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1725 Callee = DAG.getLoad(getPointerTy(), dl,
1726 DAG.getEntryNode(), CPAddr,
1727 MachinePointerInfo::getConstantPool(),
1728 false, false, false, 0);
1730 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1731 const GlobalValue *GV = G->getGlobal();
1733 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1734 bool isStub = (isExt && Subtarget->isTargetMachO()) &&
1735 getTargetMachine().getRelocationModel() != Reloc::Static;
1736 isARMFunc = !Subtarget->isThumb() || (isStub && !Subtarget->isMClass());
1737 // ARM call to a local ARM function is predicable.
1738 isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking);
1739 // tBX takes a register source operand.
1740 if (isStub && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1741 assert(Subtarget->isTargetMachO() && "WrapperPIC use on non-MachO?");
1742 Callee = DAG.getNode(ARMISD::WrapperPIC, dl, getPointerTy(),
1743 DAG.getTargetGlobalAddress(GV, dl, getPointerTy(),
1744 0, ARMII::MO_NONLAZY));
1745 Callee = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Callee,
1746 MachinePointerInfo::getGOT(), false, false, true, 0);
1747 } else if (Subtarget->isTargetCOFF()) {
1748 assert(Subtarget->isTargetWindows() &&
1749 "Windows is the only supported COFF target");
1750 unsigned TargetFlags = GV->hasDLLImportStorageClass()
1751 ? ARMII::MO_DLLIMPORT
1752 : ARMII::MO_NO_FLAG;
1753 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), /*Offset=*/0,
1755 if (GV->hasDLLImportStorageClass())
1756 Callee = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
1757 DAG.getNode(ARMISD::Wrapper, dl, getPointerTy(),
1758 Callee), MachinePointerInfo::getGOT(),
1759 false, false, false, 0);
1761 // On ELF targets for PIC code, direct calls should go through the PLT
1762 unsigned OpFlags = 0;
1763 if (Subtarget->isTargetELF() &&
1764 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1765 OpFlags = ARMII::MO_PLT;
1766 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
1768 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1770 bool isStub = Subtarget->isTargetMachO() &&
1771 getTargetMachine().getRelocationModel() != Reloc::Static;
1772 isARMFunc = !Subtarget->isThumb() || (isStub && !Subtarget->isMClass());
1773 // tBX takes a register source operand.
1774 const char *Sym = S->getSymbol();
1775 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1776 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1777 ARMConstantPoolValue *CPV =
1778 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1779 ARMPCLabelIndex, 4);
1780 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1781 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1782 Callee = DAG.getLoad(getPointerTy(), dl,
1783 DAG.getEntryNode(), CPAddr,
1784 MachinePointerInfo::getConstantPool(),
1785 false, false, false, 0);
1786 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, dl, MVT::i32);
1787 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1788 getPointerTy(), Callee, PICLabel);
1790 unsigned OpFlags = 0;
1791 // On ELF targets for PIC code, direct calls should go through the PLT
1792 if (Subtarget->isTargetELF() &&
1793 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1794 OpFlags = ARMII::MO_PLT;
1795 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlags);
1799 // FIXME: handle tail calls differently.
1801 bool HasMinSizeAttr = MF.getFunction()->hasFnAttribute(Attribute::MinSize);
1802 if (Subtarget->isThumb()) {
1803 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1804 CallOpc = ARMISD::CALL_NOLINK;
1806 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1808 if (!isDirect && !Subtarget->hasV5TOps())
1809 CallOpc = ARMISD::CALL_NOLINK;
1810 else if (doesNotRet && isDirect && Subtarget->hasRAS() &&
1811 // Emit regular call when code size is the priority
1813 // "mov lr, pc; b _foo" to avoid confusing the RSP
1814 CallOpc = ARMISD::CALL_NOLINK;
1816 CallOpc = isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL;
1819 std::vector<SDValue> Ops;
1820 Ops.push_back(Chain);
1821 Ops.push_back(Callee);
1823 // Add argument registers to the end of the list so that they are known live
1825 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1826 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1827 RegsToPass[i].second.getValueType()));
1829 // Add a register mask operand representing the call-preserved registers.
1831 const uint32_t *Mask;
1832 const ARMBaseRegisterInfo *ARI = Subtarget->getRegisterInfo();
1834 // For 'this' returns, use the R0-preserving mask if applicable
1835 Mask = ARI->getThisReturnPreservedMask(MF, CallConv);
1837 // Set isThisReturn to false if the calling convention is not one that
1838 // allows 'returned' to be modeled in this way, so LowerCallResult does
1839 // not try to pass 'this' straight through
1840 isThisReturn = false;
1841 Mask = ARI->getCallPreservedMask(MF, CallConv);
1844 Mask = ARI->getCallPreservedMask(MF, CallConv);
1846 assert(Mask && "Missing call preserved mask for calling convention");
1847 Ops.push_back(DAG.getRegisterMask(Mask));
1850 if (InFlag.getNode())
1851 Ops.push_back(InFlag);
1853 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
1855 MF.getFrameInfo()->setHasTailCall();
1856 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, Ops);
1859 // Returns a chain and a flag for retval copy to use.
1860 Chain = DAG.getNode(CallOpc, dl, NodeTys, Ops);
1861 InFlag = Chain.getValue(1);
1863 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, dl, true),
1864 DAG.getIntPtrConstant(0, dl, true), InFlag, dl);
1866 InFlag = Chain.getValue(1);
1868 // Handle result values, copying them out of physregs into vregs that we
1870 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins, dl, DAG,
1871 InVals, isThisReturn,
1872 isThisReturn ? OutVals[0] : SDValue());
1875 /// HandleByVal - Every parameter *after* a byval parameter is passed
1876 /// on the stack. Remember the next parameter register to allocate,
1877 /// and then confiscate the rest of the parameter registers to insure
1879 void ARMTargetLowering::HandleByVal(CCState *State, unsigned &Size,
1880 unsigned Align) const {
1881 assert((State->getCallOrPrologue() == Prologue ||
1882 State->getCallOrPrologue() == Call) &&
1883 "unhandled ParmContext");
1885 // Byval (as with any stack) slots are always at least 4 byte aligned.
1886 Align = std::max(Align, 4U);
1888 unsigned Reg = State->AllocateReg(GPRArgRegs);
1892 unsigned AlignInRegs = Align / 4;
1893 unsigned Waste = (ARM::R4 - Reg) % AlignInRegs;
1894 for (unsigned i = 0; i < Waste; ++i)
1895 Reg = State->AllocateReg(GPRArgRegs);
1900 unsigned Excess = 4 * (ARM::R4 - Reg);
1902 // Special case when NSAA != SP and parameter size greater than size of
1903 // all remained GPR regs. In that case we can't split parameter, we must
1904 // send it to stack. We also must set NCRN to R4, so waste all
1905 // remained registers.
1906 const unsigned NSAAOffset = State->getNextStackOffset();
1907 if (NSAAOffset != 0 && Size > Excess) {
1908 while (State->AllocateReg(GPRArgRegs))
1913 // First register for byval parameter is the first register that wasn't
1914 // allocated before this method call, so it would be "reg".
1915 // If parameter is small enough to be saved in range [reg, r4), then
1916 // the end (first after last) register would be reg + param-size-in-regs,
1917 // else parameter would be splitted between registers and stack,
1918 // end register would be r4 in this case.
1919 unsigned ByValRegBegin = Reg;
1920 unsigned ByValRegEnd = std::min<unsigned>(Reg + Size / 4, ARM::R4);
1921 State->addInRegsParamInfo(ByValRegBegin, ByValRegEnd);
1922 // Note, first register is allocated in the beginning of function already,
1923 // allocate remained amount of registers we need.
1924 for (unsigned i = Reg + 1; i != ByValRegEnd; ++i)
1925 State->AllocateReg(GPRArgRegs);
1926 // A byval parameter that is split between registers and memory needs its
1927 // size truncated here.
1928 // In the case where the entire structure fits in registers, we set the
1929 // size in memory to zero.
1930 Size = std::max<int>(Size - Excess, 0);
1934 /// MatchingStackOffset - Return true if the given stack call argument is
1935 /// already available in the same position (relatively) of the caller's
1936 /// incoming argument stack.
1938 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1939 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1940 const TargetInstrInfo *TII) {
1941 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1943 if (Arg.getOpcode() == ISD::CopyFromReg) {
1944 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1945 if (!TargetRegisterInfo::isVirtualRegister(VR))
1947 MachineInstr *Def = MRI->getVRegDef(VR);
1950 if (!Flags.isByVal()) {
1951 if (!TII->isLoadFromStackSlot(Def, FI))
1956 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
1957 if (Flags.isByVal())
1958 // ByVal argument is passed in as a pointer but it's now being
1959 // dereferenced. e.g.
1960 // define @foo(%struct.X* %A) {
1961 // tail call @bar(%struct.X* byval %A)
1964 SDValue Ptr = Ld->getBasePtr();
1965 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
1968 FI = FINode->getIndex();
1972 assert(FI != INT_MAX);
1973 if (!MFI->isFixedObjectIndex(FI))
1975 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
1978 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
1979 /// for tail call optimization. Targets which want to do tail call
1980 /// optimization should implement this function.
1982 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1983 CallingConv::ID CalleeCC,
1985 bool isCalleeStructRet,
1986 bool isCallerStructRet,
1987 const SmallVectorImpl<ISD::OutputArg> &Outs,
1988 const SmallVectorImpl<SDValue> &OutVals,
1989 const SmallVectorImpl<ISD::InputArg> &Ins,
1990 SelectionDAG& DAG) const {
1991 const Function *CallerF = DAG.getMachineFunction().getFunction();
1992 CallingConv::ID CallerCC = CallerF->getCallingConv();
1993 bool CCMatch = CallerCC == CalleeCC;
1995 // Look for obvious safe cases to perform tail call optimization that do not
1996 // require ABI changes. This is what gcc calls sibcall.
1998 // Do not sibcall optimize vararg calls unless the call site is not passing
2000 if (isVarArg && !Outs.empty())
2003 // Exception-handling functions need a special set of instructions to indicate
2004 // a return to the hardware. Tail-calling another function would probably
2006 if (CallerF->hasFnAttribute("interrupt"))
2009 // Also avoid sibcall optimization if either caller or callee uses struct
2010 // return semantics.
2011 if (isCalleeStructRet || isCallerStructRet)
2014 // FIXME: Completely disable sibcall for Thumb1 since ThumbRegisterInfo::
2015 // emitEpilogue is not ready for them. Thumb tail calls also use t2B, as
2016 // the Thumb1 16-bit unconditional branch doesn't have sufficient relocation
2017 // support in the assembler and linker to be used. This would need to be
2018 // fixed to fully support tail calls in Thumb1.
2020 // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take
2021 // LR. This means if we need to reload LR, it takes an extra instructions,
2022 // which outweighs the value of the tail call; but here we don't know yet
2023 // whether LR is going to be used. Probably the right approach is to
2024 // generate the tail call here and turn it back into CALL/RET in
2025 // emitEpilogue if LR is used.
2027 // Thumb1 PIC calls to external symbols use BX, so they can be tail calls,
2028 // but we need to make sure there are enough registers; the only valid
2029 // registers are the 4 used for parameters. We don't currently do this
2031 if (Subtarget->isThumb1Only())
2034 // Externally-defined functions with weak linkage should not be
2035 // tail-called on ARM when the OS does not support dynamic
2036 // pre-emption of symbols, as the AAELF spec requires normal calls
2037 // to undefined weak functions to be replaced with a NOP or jump to the
2038 // next instruction. The behaviour of branch instructions in this
2039 // situation (as used for tail calls) is implementation-defined, so we
2040 // cannot rely on the linker replacing the tail call with a return.
2041 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2042 const GlobalValue *GV = G->getGlobal();
2043 const Triple TT(getTargetMachine().getTargetTriple());
2044 if (GV->hasExternalWeakLinkage() &&
2045 (!TT.isOSWindows() || TT.isOSBinFormatELF() || TT.isOSBinFormatMachO()))
2049 // If the calling conventions do not match, then we'd better make sure the
2050 // results are returned in the same way as what the caller expects.
2052 SmallVector<CCValAssign, 16> RVLocs1;
2053 ARMCCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(), RVLocs1,
2054 *DAG.getContext(), Call);
2055 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
2057 SmallVector<CCValAssign, 16> RVLocs2;
2058 ARMCCState CCInfo2(CallerCC, false, DAG.getMachineFunction(), RVLocs2,
2059 *DAG.getContext(), Call);
2060 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
2062 if (RVLocs1.size() != RVLocs2.size())
2064 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
2065 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
2067 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
2069 if (RVLocs1[i].isRegLoc()) {
2070 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
2073 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
2079 // If Caller's vararg or byval argument has been split between registers and
2080 // stack, do not perform tail call, since part of the argument is in caller's
2082 const ARMFunctionInfo *AFI_Caller = DAG.getMachineFunction().
2083 getInfo<ARMFunctionInfo>();
2084 if (AFI_Caller->getArgRegsSaveSize())
2087 // If the callee takes no arguments then go on to check the results of the
2089 if (!Outs.empty()) {
2090 // Check if stack adjustment is needed. For now, do not do this if any
2091 // argument is passed on the stack.
2092 SmallVector<CCValAssign, 16> ArgLocs;
2093 ARMCCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
2094 *DAG.getContext(), Call);
2095 CCInfo.AnalyzeCallOperands(Outs,
2096 CCAssignFnForNode(CalleeCC, false, isVarArg));
2097 if (CCInfo.getNextStackOffset()) {
2098 MachineFunction &MF = DAG.getMachineFunction();
2100 // Check if the arguments are already laid out in the right way as
2101 // the caller's fixed stack objects.
2102 MachineFrameInfo *MFI = MF.getFrameInfo();
2103 const MachineRegisterInfo *MRI = &MF.getRegInfo();
2104 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
2105 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
2107 ++i, ++realArgIdx) {
2108 CCValAssign &VA = ArgLocs[i];
2109 EVT RegVT = VA.getLocVT();
2110 SDValue Arg = OutVals[realArgIdx];
2111 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
2112 if (VA.getLocInfo() == CCValAssign::Indirect)
2114 if (VA.needsCustom()) {
2115 // f64 and vector types are split into multiple registers or
2116 // register/stack-slot combinations. The types will not match
2117 // the registers; give up on memory f64 refs until we figure
2118 // out what to do about this.
2121 if (!ArgLocs[++i].isRegLoc())
2123 if (RegVT == MVT::v2f64) {
2124 if (!ArgLocs[++i].isRegLoc())
2126 if (!ArgLocs[++i].isRegLoc())
2129 } else if (!VA.isRegLoc()) {
2130 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
2142 ARMTargetLowering::CanLowerReturn(CallingConv::ID CallConv,
2143 MachineFunction &MF, bool isVarArg,
2144 const SmallVectorImpl<ISD::OutputArg> &Outs,
2145 LLVMContext &Context) const {
2146 SmallVector<CCValAssign, 16> RVLocs;
2147 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, Context);
2148 return CCInfo.CheckReturn(Outs, CCAssignFnForNode(CallConv, /*Return=*/true,
2152 static SDValue LowerInterruptReturn(SmallVectorImpl<SDValue> &RetOps,
2153 SDLoc DL, SelectionDAG &DAG) {
2154 const MachineFunction &MF = DAG.getMachineFunction();
2155 const Function *F = MF.getFunction();
2157 StringRef IntKind = F->getFnAttribute("interrupt").getValueAsString();
2159 // See ARM ARM v7 B1.8.3. On exception entry LR is set to a possibly offset
2160 // version of the "preferred return address". These offsets affect the return
2161 // instruction if this is a return from PL1 without hypervisor extensions.
2162 // IRQ/FIQ: +4 "subs pc, lr, #4"
2163 // SWI: 0 "subs pc, lr, #0"
2164 // ABORT: +4 "subs pc, lr, #4"
2165 // UNDEF: +4/+2 "subs pc, lr, #0"
2166 // UNDEF varies depending on where the exception came from ARM or Thumb
2167 // mode. Alongside GCC, we throw our hands up in disgust and pretend it's 0.
2170 if (IntKind == "" || IntKind == "IRQ" || IntKind == "FIQ" ||
2173 else if (IntKind == "SWI" || IntKind == "UNDEF")
2176 report_fatal_error("Unsupported interrupt attribute. If present, value "
2177 "must be one of: IRQ, FIQ, SWI, ABORT or UNDEF");
2179 RetOps.insert(RetOps.begin() + 1,
2180 DAG.getConstant(LROffset, DL, MVT::i32, false));
2182 return DAG.getNode(ARMISD::INTRET_FLAG, DL, MVT::Other, RetOps);
2186 ARMTargetLowering::LowerReturn(SDValue Chain,
2187 CallingConv::ID CallConv, bool isVarArg,
2188 const SmallVectorImpl<ISD::OutputArg> &Outs,
2189 const SmallVectorImpl<SDValue> &OutVals,
2190 SDLoc dl, SelectionDAG &DAG) const {
2192 // CCValAssign - represent the assignment of the return value to a location.
2193 SmallVector<CCValAssign, 16> RVLocs;
2195 // CCState - Info about the registers and stack slots.
2196 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
2197 *DAG.getContext(), Call);
2199 // Analyze outgoing return values.
2200 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
2204 SmallVector<SDValue, 4> RetOps;
2205 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
2206 bool isLittleEndian = Subtarget->isLittle();
2208 MachineFunction &MF = DAG.getMachineFunction();
2209 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2210 AFI->setReturnRegsCount(RVLocs.size());
2212 // Copy the result values into the output registers.
2213 for (unsigned i = 0, realRVLocIdx = 0;
2215 ++i, ++realRVLocIdx) {
2216 CCValAssign &VA = RVLocs[i];
2217 assert(VA.isRegLoc() && "Can only return in registers!");
2219 SDValue Arg = OutVals[realRVLocIdx];
2221 switch (VA.getLocInfo()) {
2222 default: llvm_unreachable("Unknown loc info!");
2223 case CCValAssign::Full: break;
2224 case CCValAssign::BCvt:
2225 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
2229 if (VA.needsCustom()) {
2230 if (VA.getLocVT() == MVT::v2f64) {
2231 // Extract the first half and return it in two registers.
2232 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
2233 DAG.getConstant(0, dl, MVT::i32));
2234 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
2235 DAG.getVTList(MVT::i32, MVT::i32), Half);
2237 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2238 HalfGPRs.getValue(isLittleEndian ? 0 : 1),
2240 Flag = Chain.getValue(1);
2241 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2242 VA = RVLocs[++i]; // skip ahead to next loc
2243 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2244 HalfGPRs.getValue(isLittleEndian ? 1 : 0),
2246 Flag = Chain.getValue(1);
2247 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2248 VA = RVLocs[++i]; // skip ahead to next loc
2250 // Extract the 2nd half and fall through to handle it as an f64 value.
2251 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
2252 DAG.getConstant(1, dl, MVT::i32));
2254 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
2256 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
2257 DAG.getVTList(MVT::i32, MVT::i32), Arg);
2258 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2259 fmrrd.getValue(isLittleEndian ? 0 : 1),
2261 Flag = Chain.getValue(1);
2262 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2263 VA = RVLocs[++i]; // skip ahead to next loc
2264 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2265 fmrrd.getValue(isLittleEndian ? 1 : 0),
2268 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
2270 // Guarantee that all emitted copies are
2271 // stuck together, avoiding something bad.
2272 Flag = Chain.getValue(1);
2273 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2276 // Update chain and glue.
2279 RetOps.push_back(Flag);
2281 // CPUs which aren't M-class use a special sequence to return from
2282 // exceptions (roughly, any instruction setting pc and cpsr simultaneously,
2283 // though we use "subs pc, lr, #N").
2285 // M-class CPUs actually use a normal return sequence with a special
2286 // (hardware-provided) value in LR, so the normal code path works.
2287 if (DAG.getMachineFunction().getFunction()->hasFnAttribute("interrupt") &&
2288 !Subtarget->isMClass()) {
2289 if (Subtarget->isThumb1Only())
2290 report_fatal_error("interrupt attribute is not supported in Thumb1");
2291 return LowerInterruptReturn(RetOps, dl, DAG);
2294 return DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, RetOps);
2297 bool ARMTargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
2298 if (N->getNumValues() != 1)
2300 if (!N->hasNUsesOfValue(1, 0))
2303 SDValue TCChain = Chain;
2304 SDNode *Copy = *N->use_begin();
2305 if (Copy->getOpcode() == ISD::CopyToReg) {
2306 // If the copy has a glue operand, we conservatively assume it isn't safe to
2307 // perform a tail call.
2308 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
2310 TCChain = Copy->getOperand(0);
2311 } else if (Copy->getOpcode() == ARMISD::VMOVRRD) {
2312 SDNode *VMov = Copy;
2313 // f64 returned in a pair of GPRs.
2314 SmallPtrSet<SDNode*, 2> Copies;
2315 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
2317 if (UI->getOpcode() != ISD::CopyToReg)
2321 if (Copies.size() > 2)
2324 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
2326 SDValue UseChain = UI->getOperand(0);
2327 if (Copies.count(UseChain.getNode()))
2331 // We are at the top of this chain.
2332 // If the copy has a glue operand, we conservatively assume it
2333 // isn't safe to perform a tail call.
2334 if (UI->getOperand(UI->getNumOperands()-1).getValueType() == MVT::Glue)
2340 } else if (Copy->getOpcode() == ISD::BITCAST) {
2341 // f32 returned in a single GPR.
2342 if (!Copy->hasOneUse())
2344 Copy = *Copy->use_begin();
2345 if (Copy->getOpcode() != ISD::CopyToReg || !Copy->hasNUsesOfValue(1, 0))
2347 // If the copy has a glue operand, we conservatively assume it isn't safe to
2348 // perform a tail call.
2349 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
2351 TCChain = Copy->getOperand(0);
2356 bool HasRet = false;
2357 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
2359 if (UI->getOpcode() != ARMISD::RET_FLAG &&
2360 UI->getOpcode() != ARMISD::INTRET_FLAG)
2372 bool ARMTargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
2373 if (!Subtarget->supportsTailCall())
2376 if (!CI->isTailCall() || getTargetMachine().Options.DisableTailCalls)
2379 return !Subtarget->isThumb1Only();
2382 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
2383 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
2384 // one of the above mentioned nodes. It has to be wrapped because otherwise
2385 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
2386 // be used to form addressing mode. These wrapped nodes will be selected
2388 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
2389 EVT PtrVT = Op.getValueType();
2390 // FIXME there is no actual debug info here
2392 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
2394 if (CP->isMachineConstantPoolEntry())
2395 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
2396 CP->getAlignment());
2398 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
2399 CP->getAlignment());
2400 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
2403 unsigned ARMTargetLowering::getJumpTableEncoding() const {
2404 return MachineJumpTableInfo::EK_Inline;
2407 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
2408 SelectionDAG &DAG) const {
2409 MachineFunction &MF = DAG.getMachineFunction();
2410 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2411 unsigned ARMPCLabelIndex = 0;
2413 EVT PtrVT = getPointerTy();
2414 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
2415 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2417 if (RelocM == Reloc::Static) {
2418 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
2420 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2421 ARMPCLabelIndex = AFI->createPICLabelUId();
2422 ARMConstantPoolValue *CPV =
2423 ARMConstantPoolConstant::Create(BA, ARMPCLabelIndex,
2424 ARMCP::CPBlockAddress, PCAdj);
2425 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2427 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
2428 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
2429 MachinePointerInfo::getConstantPool(),
2430 false, false, false, 0);
2431 if (RelocM == Reloc::Static)
2433 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, DL, MVT::i32);
2434 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
2437 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
2439 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
2440 SelectionDAG &DAG) const {
2442 EVT PtrVT = getPointerTy();
2443 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2444 MachineFunction &MF = DAG.getMachineFunction();
2445 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2446 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2447 ARMConstantPoolValue *CPV =
2448 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2449 ARMCP::CPValue, PCAdj, ARMCP::TLSGD, true);
2450 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2451 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
2452 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
2453 MachinePointerInfo::getConstantPool(),
2454 false, false, false, 0);
2455 SDValue Chain = Argument.getValue(1);
2457 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, dl, MVT::i32);
2458 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
2460 // call __tls_get_addr.
2463 Entry.Node = Argument;
2464 Entry.Ty = (Type *) Type::getInt32Ty(*DAG.getContext());
2465 Args.push_back(Entry);
2467 // FIXME: is there useful debug info available here?
2468 TargetLowering::CallLoweringInfo CLI(DAG);
2469 CLI.setDebugLoc(dl).setChain(Chain)
2470 .setCallee(CallingConv::C, Type::getInt32Ty(*DAG.getContext()),
2471 DAG.getExternalSymbol("__tls_get_addr", PtrVT), std::move(Args),
2474 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
2475 return CallResult.first;
2478 // Lower ISD::GlobalTLSAddress using the "initial exec" or
2479 // "local exec" model.
2481 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
2483 TLSModel::Model model) const {
2484 const GlobalValue *GV = GA->getGlobal();
2487 SDValue Chain = DAG.getEntryNode();
2488 EVT PtrVT = getPointerTy();
2489 // Get the Thread Pointer
2490 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2492 if (model == TLSModel::InitialExec) {
2493 MachineFunction &MF = DAG.getMachineFunction();
2494 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2495 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2496 // Initial exec model.
2497 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2498 ARMConstantPoolValue *CPV =
2499 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2500 ARMCP::CPValue, PCAdj, ARMCP::GOTTPOFF,
2502 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2503 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2504 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2505 MachinePointerInfo::getConstantPool(),
2506 false, false, false, 0);
2507 Chain = Offset.getValue(1);
2509 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, dl, MVT::i32);
2510 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
2512 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2513 MachinePointerInfo::getConstantPool(),
2514 false, false, false, 0);
2517 assert(model == TLSModel::LocalExec);
2518 ARMConstantPoolValue *CPV =
2519 ARMConstantPoolConstant::Create(GV, ARMCP::TPOFF);
2520 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2521 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2522 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2523 MachinePointerInfo::getConstantPool(),
2524 false, false, false, 0);
2527 // The address of the thread local variable is the add of the thread
2528 // pointer with the offset of the variable.
2529 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
2533 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
2534 // TODO: implement the "local dynamic" model
2535 assert(Subtarget->isTargetELF() &&
2536 "TLS not implemented for non-ELF targets");
2537 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
2539 TLSModel::Model model = getTargetMachine().getTLSModel(GA->getGlobal());
2542 case TLSModel::GeneralDynamic:
2543 case TLSModel::LocalDynamic:
2544 return LowerToTLSGeneralDynamicModel(GA, DAG);
2545 case TLSModel::InitialExec:
2546 case TLSModel::LocalExec:
2547 return LowerToTLSExecModels(GA, DAG, model);
2549 llvm_unreachable("bogus TLS model");
2552 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
2553 SelectionDAG &DAG) const {
2554 EVT PtrVT = getPointerTy();
2556 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2557 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2558 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
2559 ARMConstantPoolValue *CPV =
2560 ARMConstantPoolConstant::Create(GV,
2561 UseGOTOFF ? ARMCP::GOTOFF : ARMCP::GOT);
2562 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2563 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2564 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
2566 MachinePointerInfo::getConstantPool(),
2567 false, false, false, 0);
2568 SDValue Chain = Result.getValue(1);
2569 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
2570 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
2572 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
2573 MachinePointerInfo::getGOT(),
2574 false, false, false, 0);
2578 // If we have T2 ops, we can materialize the address directly via movt/movw
2579 // pair. This is always cheaper.
2580 if (Subtarget->useMovt(DAG.getMachineFunction())) {
2582 // FIXME: Once remat is capable of dealing with instructions with register
2583 // operands, expand this into two nodes.
2584 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2585 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2587 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2588 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2589 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2590 MachinePointerInfo::getConstantPool(),
2591 false, false, false, 0);
2595 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
2596 SelectionDAG &DAG) const {
2597 EVT PtrVT = getPointerTy();
2599 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2600 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2602 if (Subtarget->useMovt(DAG.getMachineFunction()))
2605 // FIXME: Once remat is capable of dealing with instructions with register
2606 // operands, expand this into multiple nodes
2608 RelocM == Reloc::PIC_ ? ARMISD::WrapperPIC : ARMISD::Wrapper;
2610 SDValue G = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, ARMII::MO_NONLAZY);
2611 SDValue Result = DAG.getNode(Wrapper, dl, PtrVT, G);
2613 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2614 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
2615 MachinePointerInfo::getGOT(), false, false, false, 0);
2619 SDValue ARMTargetLowering::LowerGlobalAddressWindows(SDValue Op,
2620 SelectionDAG &DAG) const {
2621 assert(Subtarget->isTargetWindows() && "non-Windows COFF is not supported");
2622 assert(Subtarget->useMovt(DAG.getMachineFunction()) &&
2623 "Windows on ARM expects to use movw/movt");
2625 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2626 const ARMII::TOF TargetFlags =
2627 (GV->hasDLLImportStorageClass() ? ARMII::MO_DLLIMPORT : ARMII::MO_NO_FLAG);
2628 EVT PtrVT = getPointerTy();
2634 // FIXME: Once remat is capable of dealing with instructions with register
2635 // operands, expand this into two nodes.
2636 Result = DAG.getNode(ARMISD::Wrapper, DL, PtrVT,
2637 DAG.getTargetGlobalAddress(GV, DL, PtrVT, /*Offset=*/0,
2639 if (GV->hasDLLImportStorageClass())
2640 Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), Result,
2641 MachinePointerInfo::getGOT(), false, false, false, 0);
2645 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
2646 SelectionDAG &DAG) const {
2647 assert(Subtarget->isTargetELF() &&
2648 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
2649 MachineFunction &MF = DAG.getMachineFunction();
2650 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2651 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2652 EVT PtrVT = getPointerTy();
2654 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2655 ARMConstantPoolValue *CPV =
2656 ARMConstantPoolSymbol::Create(*DAG.getContext(), "_GLOBAL_OFFSET_TABLE_",
2657 ARMPCLabelIndex, PCAdj);
2658 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2659 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2660 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2661 MachinePointerInfo::getConstantPool(),
2662 false, false, false, 0);
2663 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, dl, MVT::i32);
2664 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2668 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
2670 SDValue Val = DAG.getConstant(0, dl, MVT::i32);
2671 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl,
2672 DAG.getVTList(MVT::i32, MVT::Other), Op.getOperand(0),
2673 Op.getOperand(1), Val);
2677 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
2679 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
2680 Op.getOperand(1), DAG.getConstant(0, dl, MVT::i32));
2684 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
2685 const ARMSubtarget *Subtarget) const {
2686 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2689 default: return SDValue(); // Don't custom lower most intrinsics.
2690 case Intrinsic::arm_rbit: {
2691 assert(Op.getOperand(1).getValueType() == MVT::i32 &&
2692 "RBIT intrinsic must have i32 type!");
2693 return DAG.getNode(ARMISD::RBIT, dl, MVT::i32, Op.getOperand(1));
2695 case Intrinsic::arm_thread_pointer: {
2696 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2697 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2699 case Intrinsic::eh_sjlj_lsda: {
2700 MachineFunction &MF = DAG.getMachineFunction();
2701 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2702 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2703 EVT PtrVT = getPointerTy();
2704 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2706 unsigned PCAdj = (RelocM != Reloc::PIC_)
2707 ? 0 : (Subtarget->isThumb() ? 4 : 8);
2708 ARMConstantPoolValue *CPV =
2709 ARMConstantPoolConstant::Create(MF.getFunction(), ARMPCLabelIndex,
2710 ARMCP::CPLSDA, PCAdj);
2711 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2712 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2714 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2715 MachinePointerInfo::getConstantPool(),
2716 false, false, false, 0);
2718 if (RelocM == Reloc::PIC_) {
2719 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, dl, MVT::i32);
2720 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2724 case Intrinsic::arm_neon_vmulls:
2725 case Intrinsic::arm_neon_vmullu: {
2726 unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vmulls)
2727 ? ARMISD::VMULLs : ARMISD::VMULLu;
2728 return DAG.getNode(NewOpc, SDLoc(Op), Op.getValueType(),
2729 Op.getOperand(1), Op.getOperand(2));
2734 static SDValue LowerATOMIC_FENCE(SDValue Op, SelectionDAG &DAG,
2735 const ARMSubtarget *Subtarget) {
2736 // FIXME: handle "fence singlethread" more efficiently.
2738 if (!Subtarget->hasDataBarrier()) {
2739 // Some ARMv6 cpus can support data barriers with an mcr instruction.
2740 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2742 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2743 "Unexpected ISD::ATOMIC_FENCE encountered. Should be libcall!");
2744 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2745 DAG.getConstant(0, dl, MVT::i32));
2748 ConstantSDNode *OrdN = cast<ConstantSDNode>(Op.getOperand(1));
2749 AtomicOrdering Ord = static_cast<AtomicOrdering>(OrdN->getZExtValue());
2750 ARM_MB::MemBOpt Domain = ARM_MB::ISH;
2751 if (Subtarget->isMClass()) {
2752 // Only a full system barrier exists in the M-class architectures.
2753 Domain = ARM_MB::SY;
2754 } else if (Subtarget->isSwift() && Ord == Release) {
2755 // Swift happens to implement ISHST barriers in a way that's compatible with
2756 // Release semantics but weaker than ISH so we'd be fools not to use
2757 // it. Beware: other processors probably don't!
2758 Domain = ARM_MB::ISHST;
2761 return DAG.getNode(ISD::INTRINSIC_VOID, dl, MVT::Other, Op.getOperand(0),
2762 DAG.getConstant(Intrinsic::arm_dmb, dl, MVT::i32),
2763 DAG.getConstant(Domain, dl, MVT::i32));
2766 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG,
2767 const ARMSubtarget *Subtarget) {
2768 // ARM pre v5TE and Thumb1 does not have preload instructions.
2769 if (!(Subtarget->isThumb2() ||
2770 (!Subtarget->isThumb1Only() && Subtarget->hasV5TEOps())))
2771 // Just preserve the chain.
2772 return Op.getOperand(0);
2775 unsigned isRead = ~cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() & 1;
2777 (!Subtarget->hasV7Ops() || !Subtarget->hasMPExtension()))
2778 // ARMv7 with MP extension has PLDW.
2779 return Op.getOperand(0);
2781 unsigned isData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
2782 if (Subtarget->isThumb()) {
2784 isRead = ~isRead & 1;
2785 isData = ~isData & 1;
2788 return DAG.getNode(ARMISD::PRELOAD, dl, MVT::Other, Op.getOperand(0),
2789 Op.getOperand(1), DAG.getConstant(isRead, dl, MVT::i32),
2790 DAG.getConstant(isData, dl, MVT::i32));
2793 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
2794 MachineFunction &MF = DAG.getMachineFunction();
2795 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
2797 // vastart just stores the address of the VarArgsFrameIndex slot into the
2798 // memory location argument.
2800 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2801 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2802 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
2803 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
2804 MachinePointerInfo(SV), false, false, 0);
2808 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
2809 SDValue &Root, SelectionDAG &DAG,
2811 MachineFunction &MF = DAG.getMachineFunction();
2812 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2814 const TargetRegisterClass *RC;
2815 if (AFI->isThumb1OnlyFunction())
2816 RC = &ARM::tGPRRegClass;
2818 RC = &ARM::GPRRegClass;
2820 // Transform the arguments stored in physical registers into virtual ones.
2821 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2822 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2825 if (NextVA.isMemLoc()) {
2826 MachineFrameInfo *MFI = MF.getFrameInfo();
2827 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true);
2829 // Create load node to retrieve arguments from the stack.
2830 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2831 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
2832 MachinePointerInfo::getFixedStack(FI),
2833 false, false, false, 0);
2835 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
2836 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2838 if (!Subtarget->isLittle())
2839 std::swap (ArgValue, ArgValue2);
2840 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
2843 // The remaining GPRs hold either the beginning of variable-argument
2844 // data, or the beginning of an aggregate passed by value (usually
2845 // byval). Either way, we allocate stack slots adjacent to the data
2846 // provided by our caller, and store the unallocated registers there.
2847 // If this is a variadic function, the va_list pointer will begin with
2848 // these values; otherwise, this reassembles a (byval) structure that
2849 // was split between registers and memory.
2850 // Return: The frame index registers were stored into.
2852 ARMTargetLowering::StoreByValRegs(CCState &CCInfo, SelectionDAG &DAG,
2853 SDLoc dl, SDValue &Chain,
2854 const Value *OrigArg,
2855 unsigned InRegsParamRecordIdx,
2857 unsigned ArgSize) const {
2858 // Currently, two use-cases possible:
2859 // Case #1. Non-var-args function, and we meet first byval parameter.
2860 // Setup first unallocated register as first byval register;
2861 // eat all remained registers
2862 // (these two actions are performed by HandleByVal method).
2863 // Then, here, we initialize stack frame with
2864 // "store-reg" instructions.
2865 // Case #2. Var-args function, that doesn't contain byval parameters.
2866 // The same: eat all remained unallocated registers,
2867 // initialize stack frame.
2869 MachineFunction &MF = DAG.getMachineFunction();
2870 MachineFrameInfo *MFI = MF.getFrameInfo();
2871 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2872 unsigned RBegin, REnd;
2873 if (InRegsParamRecordIdx < CCInfo.getInRegsParamsCount()) {
2874 CCInfo.getInRegsParamInfo(InRegsParamRecordIdx, RBegin, REnd);
2876 unsigned RBeginIdx = CCInfo.getFirstUnallocated(GPRArgRegs);
2877 RBegin = RBeginIdx == 4 ? (unsigned)ARM::R4 : GPRArgRegs[RBeginIdx];
2882 ArgOffset = -4 * (ARM::R4 - RBegin);
2884 int FrameIndex = MFI->CreateFixedObject(ArgSize, ArgOffset, false);
2885 SDValue FIN = DAG.getFrameIndex(FrameIndex, getPointerTy());
2887 SmallVector<SDValue, 4> MemOps;
2888 const TargetRegisterClass *RC =
2889 AFI->isThumb1OnlyFunction() ? &ARM::tGPRRegClass : &ARM::GPRRegClass;
2891 for (unsigned Reg = RBegin, i = 0; Reg < REnd; ++Reg, ++i) {
2892 unsigned VReg = MF.addLiveIn(Reg, RC);
2893 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2895 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2896 MachinePointerInfo(OrigArg, 4 * i), false, false, 0);
2897 MemOps.push_back(Store);
2898 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
2899 DAG.getConstant(4, dl, getPointerTy()));
2902 if (!MemOps.empty())
2903 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOps);
2907 // Setup stack frame, the va_list pointer will start from.
2909 ARMTargetLowering::VarArgStyleRegisters(CCState &CCInfo, SelectionDAG &DAG,
2910 SDLoc dl, SDValue &Chain,
2912 unsigned TotalArgRegsSaveSize,
2913 bool ForceMutable) const {
2914 MachineFunction &MF = DAG.getMachineFunction();
2915 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2917 // Try to store any remaining integer argument regs
2918 // to their spots on the stack so that they may be loaded by deferencing
2919 // the result of va_next.
2920 // If there is no regs to be stored, just point address after last
2921 // argument passed via stack.
2922 int FrameIndex = StoreByValRegs(CCInfo, DAG, dl, Chain, nullptr,
2923 CCInfo.getInRegsParamsCount(),
2924 CCInfo.getNextStackOffset(), 4);
2925 AFI->setVarArgsFrameIndex(FrameIndex);
2929 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
2930 CallingConv::ID CallConv, bool isVarArg,
2931 const SmallVectorImpl<ISD::InputArg>
2933 SDLoc dl, SelectionDAG &DAG,
2934 SmallVectorImpl<SDValue> &InVals)
2936 MachineFunction &MF = DAG.getMachineFunction();
2937 MachineFrameInfo *MFI = MF.getFrameInfo();
2939 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2941 // Assign locations to all of the incoming arguments.
2942 SmallVector<CCValAssign, 16> ArgLocs;
2943 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs,
2944 *DAG.getContext(), Prologue);
2945 CCInfo.AnalyzeFormalArguments(Ins,
2946 CCAssignFnForNode(CallConv, /* Return*/ false,
2949 SmallVector<SDValue, 16> ArgValues;
2951 Function::const_arg_iterator CurOrigArg = MF.getFunction()->arg_begin();
2952 unsigned CurArgIdx = 0;
2954 // Initially ArgRegsSaveSize is zero.
2955 // Then we increase this value each time we meet byval parameter.
2956 // We also increase this value in case of varargs function.
2957 AFI->setArgRegsSaveSize(0);
2959 // Calculate the amount of stack space that we need to allocate to store
2960 // byval and variadic arguments that are passed in registers.
2961 // We need to know this before we allocate the first byval or variadic
2962 // argument, as they will be allocated a stack slot below the CFA (Canonical
2963 // Frame Address, the stack pointer at entry to the function).
2964 unsigned ArgRegBegin = ARM::R4;
2965 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2966 if (CCInfo.getInRegsParamsProcessed() >= CCInfo.getInRegsParamsCount())
2969 CCValAssign &VA = ArgLocs[i];
2970 unsigned Index = VA.getValNo();
2971 ISD::ArgFlagsTy Flags = Ins[Index].Flags;
2972 if (!Flags.isByVal())
2975 assert(VA.isMemLoc() && "unexpected byval pointer in reg");
2976 unsigned RBegin, REnd;
2977 CCInfo.getInRegsParamInfo(CCInfo.getInRegsParamsProcessed(), RBegin, REnd);
2978 ArgRegBegin = std::min(ArgRegBegin, RBegin);
2980 CCInfo.nextInRegsParam();
2982 CCInfo.rewindByValRegsInfo();
2984 int lastInsIndex = -1;
2985 if (isVarArg && MFI->hasVAStart()) {
2986 unsigned RegIdx = CCInfo.getFirstUnallocated(GPRArgRegs);
2987 if (RegIdx != array_lengthof(GPRArgRegs))
2988 ArgRegBegin = std::min(ArgRegBegin, (unsigned)GPRArgRegs[RegIdx]);
2991 unsigned TotalArgRegsSaveSize = 4 * (ARM::R4 - ArgRegBegin);
2992 AFI->setArgRegsSaveSize(TotalArgRegsSaveSize);
2994 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2995 CCValAssign &VA = ArgLocs[i];
2996 if (Ins[VA.getValNo()].isOrigArg()) {
2997 std::advance(CurOrigArg,
2998 Ins[VA.getValNo()].getOrigArgIndex() - CurArgIdx);
2999 CurArgIdx = Ins[VA.getValNo()].getOrigArgIndex();
3001 // Arguments stored in registers.
3002 if (VA.isRegLoc()) {
3003 EVT RegVT = VA.getLocVT();
3005 if (VA.needsCustom()) {
3006 // f64 and vector types are split up into multiple registers or
3007 // combinations of registers and stack slots.
3008 if (VA.getLocVT() == MVT::v2f64) {
3009 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
3011 VA = ArgLocs[++i]; // skip ahead to next loc
3013 if (VA.isMemLoc()) {
3014 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true);
3015 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
3016 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
3017 MachinePointerInfo::getFixedStack(FI),
3018 false, false, false, 0);
3020 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
3023 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
3024 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
3025 ArgValue, ArgValue1,
3026 DAG.getIntPtrConstant(0, dl));
3027 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
3028 ArgValue, ArgValue2,
3029 DAG.getIntPtrConstant(1, dl));
3031 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
3034 const TargetRegisterClass *RC;
3036 if (RegVT == MVT::f32)
3037 RC = &ARM::SPRRegClass;
3038 else if (RegVT == MVT::f64)
3039 RC = &ARM::DPRRegClass;
3040 else if (RegVT == MVT::v2f64)
3041 RC = &ARM::QPRRegClass;
3042 else if (RegVT == MVT::i32)
3043 RC = AFI->isThumb1OnlyFunction() ? &ARM::tGPRRegClass
3044 : &ARM::GPRRegClass;
3046 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
3048 // Transform the arguments in physical registers into virtual ones.
3049 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
3050 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
3053 // If this is an 8 or 16-bit value, it is really passed promoted
3054 // to 32 bits. Insert an assert[sz]ext to capture this, then
3055 // truncate to the right size.
3056 switch (VA.getLocInfo()) {
3057 default: llvm_unreachable("Unknown loc info!");
3058 case CCValAssign::Full: break;
3059 case CCValAssign::BCvt:
3060 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
3062 case CCValAssign::SExt:
3063 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
3064 DAG.getValueType(VA.getValVT()));
3065 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
3067 case CCValAssign::ZExt:
3068 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
3069 DAG.getValueType(VA.getValVT()));
3070 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
3074 InVals.push_back(ArgValue);
3076 } else { // VA.isRegLoc()
3079 assert(VA.isMemLoc());
3080 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
3082 int index = VA.getValNo();
3084 // Some Ins[] entries become multiple ArgLoc[] entries.
3085 // Process them only once.
3086 if (index != lastInsIndex)
3088 ISD::ArgFlagsTy Flags = Ins[index].Flags;
3089 // FIXME: For now, all byval parameter objects are marked mutable.
3090 // This can be changed with more analysis.
3091 // In case of tail call optimization mark all arguments mutable.
3092 // Since they could be overwritten by lowering of arguments in case of
3094 if (Flags.isByVal()) {
3095 assert(Ins[index].isOrigArg() &&
3096 "Byval arguments cannot be implicit");
3097 unsigned CurByValIndex = CCInfo.getInRegsParamsProcessed();
3099 int FrameIndex = StoreByValRegs(CCInfo, DAG, dl, Chain, CurOrigArg,
3100 CurByValIndex, VA.getLocMemOffset(),
3101 Flags.getByValSize());
3102 InVals.push_back(DAG.getFrameIndex(FrameIndex, getPointerTy()));
3103 CCInfo.nextInRegsParam();
3105 unsigned FIOffset = VA.getLocMemOffset();
3106 int FI = MFI->CreateFixedObject(VA.getLocVT().getSizeInBits()/8,
3109 // Create load nodes to retrieve arguments from the stack.
3110 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
3111 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
3112 MachinePointerInfo::getFixedStack(FI),
3113 false, false, false, 0));
3115 lastInsIndex = index;
3121 if (isVarArg && MFI->hasVAStart())
3122 VarArgStyleRegisters(CCInfo, DAG, dl, Chain,
3123 CCInfo.getNextStackOffset(),
3124 TotalArgRegsSaveSize);
3126 AFI->setArgumentStackSize(CCInfo.getNextStackOffset());
3131 /// isFloatingPointZero - Return true if this is +0.0.
3132 static bool isFloatingPointZero(SDValue Op) {
3133 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
3134 return CFP->getValueAPF().isPosZero();
3135 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
3136 // Maybe this has already been legalized into the constant pool?
3137 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
3138 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
3139 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
3140 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
3141 return CFP->getValueAPF().isPosZero();
3143 } else if (Op->getOpcode() == ISD::BITCAST &&
3144 Op->getValueType(0) == MVT::f64) {
3145 // Handle (ISD::BITCAST (ARMISD::VMOVIMM (ISD::TargetConstant 0)) MVT::f64)
3146 // created by LowerConstantFP().
3147 SDValue BitcastOp = Op->getOperand(0);
3148 if (BitcastOp->getOpcode() == ARMISD::VMOVIMM) {
3149 SDValue MoveOp = BitcastOp->getOperand(0);
3150 if (MoveOp->getOpcode() == ISD::TargetConstant &&
3151 cast<ConstantSDNode>(MoveOp)->getZExtValue() == 0) {
3159 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
3160 /// the given operands.
3162 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
3163 SDValue &ARMcc, SelectionDAG &DAG,
3165 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
3166 unsigned C = RHSC->getZExtValue();
3167 if (!isLegalICmpImmediate(C)) {
3168 // Constant does not fit, try adjusting it by one?
3173 if (C != 0x80000000 && isLegalICmpImmediate(C-1)) {
3174 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
3175 RHS = DAG.getConstant(C - 1, dl, MVT::i32);
3180 if (C != 0 && isLegalICmpImmediate(C-1)) {
3181 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
3182 RHS = DAG.getConstant(C - 1, dl, MVT::i32);
3187 if (C != 0x7fffffff && isLegalICmpImmediate(C+1)) {
3188 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
3189 RHS = DAG.getConstant(C + 1, dl, MVT::i32);
3194 if (C != 0xffffffff && isLegalICmpImmediate(C+1)) {
3195 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
3196 RHS = DAG.getConstant(C + 1, dl, MVT::i32);
3203 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3204 ARMISD::NodeType CompareType;
3207 CompareType = ARMISD::CMP;
3212 CompareType = ARMISD::CMPZ;
3215 ARMcc = DAG.getConstant(CondCode, dl, MVT::i32);
3216 return DAG.getNode(CompareType, dl, MVT::Glue, LHS, RHS);
3219 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
3221 ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
3223 assert(!Subtarget->isFPOnlySP() || RHS.getValueType() != MVT::f64);
3225 if (!isFloatingPointZero(RHS))
3226 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Glue, LHS, RHS);
3228 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Glue, LHS);
3229 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Glue, Cmp);
3232 /// duplicateCmp - Glue values can have only one use, so this function
3233 /// duplicates a comparison node.
3235 ARMTargetLowering::duplicateCmp(SDValue Cmp, SelectionDAG &DAG) const {
3236 unsigned Opc = Cmp.getOpcode();
3238 if (Opc == ARMISD::CMP || Opc == ARMISD::CMPZ)
3239 return DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
3241 assert(Opc == ARMISD::FMSTAT && "unexpected comparison operation");
3242 Cmp = Cmp.getOperand(0);
3243 Opc = Cmp.getOpcode();
3244 if (Opc == ARMISD::CMPFP)
3245 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
3247 assert(Opc == ARMISD::CMPFPw0 && "unexpected operand of FMSTAT");
3248 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0));
3250 return DAG.getNode(ARMISD::FMSTAT, DL, MVT::Glue, Cmp);
3253 std::pair<SDValue, SDValue>
3254 ARMTargetLowering::getARMXALUOOp(SDValue Op, SelectionDAG &DAG,
3255 SDValue &ARMcc) const {
3256 assert(Op.getValueType() == MVT::i32 && "Unsupported value type");
3258 SDValue Value, OverflowCmp;
3259 SDValue LHS = Op.getOperand(0);
3260 SDValue RHS = Op.getOperand(1);
3263 // FIXME: We are currently always generating CMPs because we don't support
3264 // generating CMN through the backend. This is not as good as the natural
3265 // CMP case because it causes a register dependency and cannot be folded
3268 switch (Op.getOpcode()) {
3270 llvm_unreachable("Unknown overflow instruction!");
3272 ARMcc = DAG.getConstant(ARMCC::VC, dl, MVT::i32);
3273 Value = DAG.getNode(ISD::ADD, dl, Op.getValueType(), LHS, RHS);
3274 OverflowCmp = DAG.getNode(ARMISD::CMP, dl, MVT::Glue, Value, LHS);
3277 ARMcc = DAG.getConstant(ARMCC::HS, dl, MVT::i32);
3278 Value = DAG.getNode(ISD::ADD, dl, Op.getValueType(), LHS, RHS);
3279 OverflowCmp = DAG.getNode(ARMISD::CMP, dl, MVT::Glue, Value, LHS);
3282 ARMcc = DAG.getConstant(ARMCC::VC, dl, MVT::i32);
3283 Value = DAG.getNode(ISD::SUB, dl, Op.getValueType(), LHS, RHS);
3284 OverflowCmp = DAG.getNode(ARMISD::CMP, dl, MVT::Glue, LHS, RHS);
3287 ARMcc = DAG.getConstant(ARMCC::HS, dl, MVT::i32);
3288 Value = DAG.getNode(ISD::SUB, dl, Op.getValueType(), LHS, RHS);
3289 OverflowCmp = DAG.getNode(ARMISD::CMP, dl, MVT::Glue, LHS, RHS);
3293 return std::make_pair(Value, OverflowCmp);
3298 ARMTargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) const {
3299 // Let legalize expand this if it isn't a legal type yet.
3300 if (!DAG.getTargetLoweringInfo().isTypeLegal(Op.getValueType()))
3303 SDValue Value, OverflowCmp;
3305 std::tie(Value, OverflowCmp) = getARMXALUOOp(Op, DAG, ARMcc);
3306 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3308 // We use 0 and 1 as false and true values.
3309 SDValue TVal = DAG.getConstant(1, dl, MVT::i32);
3310 SDValue FVal = DAG.getConstant(0, dl, MVT::i32);
3311 EVT VT = Op.getValueType();
3313 SDValue Overflow = DAG.getNode(ARMISD::CMOV, dl, VT, TVal, FVal,
3314 ARMcc, CCR, OverflowCmp);
3316 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
3317 return DAG.getNode(ISD::MERGE_VALUES, dl, VTs, Value, Overflow);
3321 SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
3322 SDValue Cond = Op.getOperand(0);
3323 SDValue SelectTrue = Op.getOperand(1);
3324 SDValue SelectFalse = Op.getOperand(2);
3326 unsigned Opc = Cond.getOpcode();
3328 if (Cond.getResNo() == 1 &&
3329 (Opc == ISD::SADDO || Opc == ISD::UADDO || Opc == ISD::SSUBO ||
3330 Opc == ISD::USUBO)) {
3331 if (!DAG.getTargetLoweringInfo().isTypeLegal(Cond->getValueType(0)))
3334 SDValue Value, OverflowCmp;
3336 std::tie(Value, OverflowCmp) = getARMXALUOOp(Cond, DAG, ARMcc);
3337 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3338 EVT VT = Op.getValueType();
3340 return getCMOV(dl, VT, SelectTrue, SelectFalse, ARMcc, CCR,
3346 // (select (cmov 1, 0, cond), t, f) -> (cmov t, f, cond)
3347 // (select (cmov 0, 1, cond), t, f) -> (cmov f, t, cond)
3349 if (Cond.getOpcode() == ARMISD::CMOV && Cond.hasOneUse()) {
3350 const ConstantSDNode *CMOVTrue =
3351 dyn_cast<ConstantSDNode>(Cond.getOperand(0));
3352 const ConstantSDNode *CMOVFalse =
3353 dyn_cast<ConstantSDNode>(Cond.getOperand(1));
3355 if (CMOVTrue && CMOVFalse) {
3356 unsigned CMOVTrueVal = CMOVTrue->getZExtValue();
3357 unsigned CMOVFalseVal = CMOVFalse->getZExtValue();
3361 if (CMOVTrueVal == 1 && CMOVFalseVal == 0) {
3363 False = SelectFalse;
3364 } else if (CMOVTrueVal == 0 && CMOVFalseVal == 1) {
3369 if (True.getNode() && False.getNode()) {
3370 EVT VT = Op.getValueType();
3371 SDValue ARMcc = Cond.getOperand(2);
3372 SDValue CCR = Cond.getOperand(3);
3373 SDValue Cmp = duplicateCmp(Cond.getOperand(4), DAG);
3374 assert(True.getValueType() == VT);
3375 return getCMOV(dl, VT, True, False, ARMcc, CCR, Cmp, DAG);
3380 // ARM's BooleanContents value is UndefinedBooleanContent. Mask out the
3381 // undefined bits before doing a full-word comparison with zero.
3382 Cond = DAG.getNode(ISD::AND, dl, Cond.getValueType(), Cond,
3383 DAG.getConstant(1, dl, Cond.getValueType()));
3385 return DAG.getSelectCC(dl, Cond,
3386 DAG.getConstant(0, dl, Cond.getValueType()),
3387 SelectTrue, SelectFalse, ISD::SETNE);
3390 static void checkVSELConstraints(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
3391 bool &swpCmpOps, bool &swpVselOps) {
3392 // Start by selecting the GE condition code for opcodes that return true for
3394 if (CC == ISD::SETUGE || CC == ISD::SETOGE || CC == ISD::SETOLE ||
3396 CondCode = ARMCC::GE;
3398 // and GT for opcodes that return false for 'equality'.
3399 else if (CC == ISD::SETUGT || CC == ISD::SETOGT || CC == ISD::SETOLT ||
3401 CondCode = ARMCC::GT;
3403 // Since we are constrained to GE/GT, if the opcode contains 'less', we need
3404 // to swap the compare operands.
3405 if (CC == ISD::SETOLE || CC == ISD::SETULE || CC == ISD::SETOLT ||
3409 // Both GT and GE are ordered comparisons, and return false for 'unordered'.
3410 // If we have an unordered opcode, we need to swap the operands to the VSEL
3411 // instruction (effectively negating the condition).
3413 // This also has the effect of swapping which one of 'less' or 'greater'
3414 // returns true, so we also swap the compare operands. It also switches
3415 // whether we return true for 'equality', so we compensate by picking the
3416 // opposite condition code to our original choice.
3417 if (CC == ISD::SETULE || CC == ISD::SETULT || CC == ISD::SETUGE ||
3418 CC == ISD::SETUGT) {
3419 swpCmpOps = !swpCmpOps;
3420 swpVselOps = !swpVselOps;
3421 CondCode = CondCode == ARMCC::GT ? ARMCC::GE : ARMCC::GT;
3424 // 'ordered' is 'anything but unordered', so use the VS condition code and
3425 // swap the VSEL operands.
3426 if (CC == ISD::SETO) {
3427 CondCode = ARMCC::VS;
3431 // 'unordered or not equal' is 'anything but equal', so use the EQ condition
3432 // code and swap the VSEL operands.
3433 if (CC == ISD::SETUNE) {
3434 CondCode = ARMCC::EQ;
3439 SDValue ARMTargetLowering::getCMOV(SDLoc dl, EVT VT, SDValue FalseVal,
3440 SDValue TrueVal, SDValue ARMcc, SDValue CCR,
3441 SDValue Cmp, SelectionDAG &DAG) const {
3442 if (Subtarget->isFPOnlySP() && VT == MVT::f64) {
3443 FalseVal = DAG.getNode(ARMISD::VMOVRRD, dl,
3444 DAG.getVTList(MVT::i32, MVT::i32), FalseVal);
3445 TrueVal = DAG.getNode(ARMISD::VMOVRRD, dl,
3446 DAG.getVTList(MVT::i32, MVT::i32), TrueVal);
3448 SDValue TrueLow = TrueVal.getValue(0);
3449 SDValue TrueHigh = TrueVal.getValue(1);
3450 SDValue FalseLow = FalseVal.getValue(0);
3451 SDValue FalseHigh = FalseVal.getValue(1);
3453 SDValue Low = DAG.getNode(ARMISD::CMOV, dl, MVT::i32, FalseLow, TrueLow,
3455 SDValue High = DAG.getNode(ARMISD::CMOV, dl, MVT::i32, FalseHigh, TrueHigh,
3456 ARMcc, CCR, duplicateCmp(Cmp, DAG));
3458 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Low, High);
3460 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,
3465 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
3466 EVT VT = Op.getValueType();
3467 SDValue LHS = Op.getOperand(0);
3468 SDValue RHS = Op.getOperand(1);
3469 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
3470 SDValue TrueVal = Op.getOperand(2);
3471 SDValue FalseVal = Op.getOperand(3);
3474 if (Subtarget->isFPOnlySP() && LHS.getValueType() == MVT::f64) {
3475 DAG.getTargetLoweringInfo().softenSetCCOperands(DAG, MVT::f64, LHS, RHS, CC,
3478 // If softenSetCCOperands only returned one value, we should compare it to
3480 if (!RHS.getNode()) {
3481 RHS = DAG.getConstant(0, dl, LHS.getValueType());
3486 if (LHS.getValueType() == MVT::i32) {
3487 // Try to generate VSEL on ARMv8.
3488 // The VSEL instruction can't use all the usual ARM condition
3489 // codes: it only has two bits to select the condition code, so it's
3490 // constrained to use only GE, GT, VS and EQ.
3492 // To implement all the various ISD::SETXXX opcodes, we sometimes need to
3493 // swap the operands of the previous compare instruction (effectively
3494 // inverting the compare condition, swapping 'less' and 'greater') and
3495 // sometimes need to swap the operands to the VSEL (which inverts the
3496 // condition in the sense of firing whenever the previous condition didn't)
3497 if (Subtarget->hasFPARMv8() && (TrueVal.getValueType() == MVT::f32 ||
3498 TrueVal.getValueType() == MVT::f64)) {
3499 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3500 if (CondCode == ARMCC::LT || CondCode == ARMCC::LE ||
3501 CondCode == ARMCC::VC || CondCode == ARMCC::NE) {
3502 CC = ISD::getSetCCInverse(CC, true);
3503 std::swap(TrueVal, FalseVal);
3508 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3509 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3510 return getCMOV(dl, VT, FalseVal, TrueVal, ARMcc, CCR, Cmp, DAG);
3513 ARMCC::CondCodes CondCode, CondCode2;
3514 FPCCToARMCC(CC, CondCode, CondCode2);
3516 // Try to generate VMAXNM/VMINNM on ARMv8.
3517 if (Subtarget->hasFPARMv8() && (TrueVal.getValueType() == MVT::f32 ||
3518 TrueVal.getValueType() == MVT::f64)) {
3519 // We can use VMAXNM/VMINNM for a compare followed by a select with the
3520 // same operands, as follows:
3521 // c = fcmp [?gt, ?ge, ?lt, ?le] a, b
3523 // In NoNaNsFPMath the CC will have been changed from, e.g., 'ogt' to 'gt'.
3524 // FIXME: There is similar code that allows some extensions in
3525 // AArch64TargetLowering::LowerSELECT_CC that should be shared with this
3527 bool swapSides = false;
3528 if (!getTargetMachine().Options.NoNaNsFPMath) {
3529 // transformability may depend on which way around we compare
3537 // the non-NaN should be RHS
3538 swapSides = DAG.isKnownNeverNaN(LHS) && !DAG.isKnownNeverNaN(RHS);
3544 // the non-NaN should be LHS
3545 swapSides = DAG.isKnownNeverNaN(RHS) && !DAG.isKnownNeverNaN(LHS);
3549 swapSides = swapSides || (LHS == FalseVal && RHS == TrueVal);
3551 CC = ISD::getSetCCSwappedOperands(CC);
3552 std::swap(LHS, RHS);
3554 if (LHS == TrueVal && RHS == FalseVal) {
3555 bool canTransform = true;
3556 // FIXME: FastMathFlags::noSignedZeros() doesn't appear reachable from here
3557 if (!getTargetMachine().Options.UnsafeFPMath &&
3558 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
3559 const ConstantFPSDNode *Zero;
3566 // RHS must not be -0
3567 canTransform = (Zero = dyn_cast<ConstantFPSDNode>(RHS)) &&
3568 !Zero->isNegative();
3573 // LHS must not be -0
3574 canTransform = (Zero = dyn_cast<ConstantFPSDNode>(LHS)) &&
3575 !Zero->isNegative();
3580 // RHS must not be +0
3581 canTransform = (Zero = dyn_cast<ConstantFPSDNode>(RHS)) &&
3587 // LHS must not be +0
3588 canTransform = (Zero = dyn_cast<ConstantFPSDNode>(LHS)) &&
3594 // Note: If one of the elements in a pair is a number and the other
3595 // element is NaN, the corresponding result element is the number.
3596 // This is consistent with the IEEE 754-2008 standard.
3597 // Therefore, a > b ? a : b <=> vmax(a,b), if b is constant and a is NaN
3603 if (!DAG.isKnownNeverNaN(RHS))
3605 return DAG.getNode(ARMISD::VMAXNM, dl, VT, LHS, RHS);
3608 if (!DAG.isKnownNeverNaN(LHS))
3612 return DAG.getNode(ARMISD::VMAXNM, dl, VT, LHS, RHS);
3615 if (!DAG.isKnownNeverNaN(RHS))
3617 return DAG.getNode(ARMISD::VMINNM, dl, VT, LHS, RHS);
3620 if (!DAG.isKnownNeverNaN(LHS))
3624 return DAG.getNode(ARMISD::VMINNM, dl, VT, LHS, RHS);
3629 bool swpCmpOps = false;
3630 bool swpVselOps = false;
3631 checkVSELConstraints(CC, CondCode, swpCmpOps, swpVselOps);
3633 if (CondCode == ARMCC::GT || CondCode == ARMCC::GE ||
3634 CondCode == ARMCC::VS || CondCode == ARMCC::EQ) {
3636 std::swap(LHS, RHS);
3638 std::swap(TrueVal, FalseVal);
3642 SDValue ARMcc = DAG.getConstant(CondCode, dl, MVT::i32);
3643 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
3644 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3645 SDValue Result = getCMOV(dl, VT, FalseVal, TrueVal, ARMcc, CCR, Cmp, DAG);
3646 if (CondCode2 != ARMCC::AL) {
3647 SDValue ARMcc2 = DAG.getConstant(CondCode2, dl, MVT::i32);
3648 // FIXME: Needs another CMP because flag can have but one use.
3649 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
3650 Result = getCMOV(dl, VT, Result, TrueVal, ARMcc2, CCR, Cmp2, DAG);
3655 /// canChangeToInt - Given the fp compare operand, return true if it is suitable
3656 /// to morph to an integer compare sequence.
3657 static bool canChangeToInt(SDValue Op, bool &SeenZero,
3658 const ARMSubtarget *Subtarget) {
3659 SDNode *N = Op.getNode();
3660 if (!N->hasOneUse())
3661 // Otherwise it requires moving the value from fp to integer registers.
3663 if (!N->getNumValues())
3665 EVT VT = Op.getValueType();
3666 if (VT != MVT::f32 && !Subtarget->isFPBrccSlow())
3667 // f32 case is generally profitable. f64 case only makes sense when vcmpe +
3668 // vmrs are very slow, e.g. cortex-a8.
3671 if (isFloatingPointZero(Op)) {
3675 return ISD::isNormalLoad(N);
3678 static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) {
3679 if (isFloatingPointZero(Op))
3680 return DAG.getConstant(0, SDLoc(Op), MVT::i32);
3682 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op))
3683 return DAG.getLoad(MVT::i32, SDLoc(Op),
3684 Ld->getChain(), Ld->getBasePtr(), Ld->getPointerInfo(),
3685 Ld->isVolatile(), Ld->isNonTemporal(),
3686 Ld->isInvariant(), Ld->getAlignment());
3688 llvm_unreachable("Unknown VFP cmp argument!");
3691 static void expandf64Toi32(SDValue Op, SelectionDAG &DAG,
3692 SDValue &RetVal1, SDValue &RetVal2) {
3695 if (isFloatingPointZero(Op)) {
3696 RetVal1 = DAG.getConstant(0, dl, MVT::i32);
3697 RetVal2 = DAG.getConstant(0, dl, MVT::i32);
3701 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) {
3702 SDValue Ptr = Ld->getBasePtr();
3703 RetVal1 = DAG.getLoad(MVT::i32, dl,
3704 Ld->getChain(), Ptr,
3705 Ld->getPointerInfo(),
3706 Ld->isVolatile(), Ld->isNonTemporal(),
3707 Ld->isInvariant(), Ld->getAlignment());
3709 EVT PtrType = Ptr.getValueType();
3710 unsigned NewAlign = MinAlign(Ld->getAlignment(), 4);
3711 SDValue NewPtr = DAG.getNode(ISD::ADD, dl,
3712 PtrType, Ptr, DAG.getConstant(4, dl, PtrType));
3713 RetVal2 = DAG.getLoad(MVT::i32, dl,
3714 Ld->getChain(), NewPtr,
3715 Ld->getPointerInfo().getWithOffset(4),
3716 Ld->isVolatile(), Ld->isNonTemporal(),
3717 Ld->isInvariant(), NewAlign);
3721 llvm_unreachable("Unknown VFP cmp argument!");
3724 /// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some
3725 /// f32 and even f64 comparisons to integer ones.
3727 ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const {
3728 SDValue Chain = Op.getOperand(0);
3729 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3730 SDValue LHS = Op.getOperand(2);
3731 SDValue RHS = Op.getOperand(3);
3732 SDValue Dest = Op.getOperand(4);
3735 bool LHSSeenZero = false;
3736 bool LHSOk = canChangeToInt(LHS, LHSSeenZero, Subtarget);
3737 bool RHSSeenZero = false;
3738 bool RHSOk = canChangeToInt(RHS, RHSSeenZero, Subtarget);
3739 if (LHSOk && RHSOk && (LHSSeenZero || RHSSeenZero)) {
3740 // If unsafe fp math optimization is enabled and there are no other uses of
3741 // the CMP operands, and the condition code is EQ or NE, we can optimize it
3742 // to an integer comparison.
3743 if (CC == ISD::SETOEQ)
3745 else if (CC == ISD::SETUNE)
3748 SDValue Mask = DAG.getConstant(0x7fffffff, dl, MVT::i32);
3750 if (LHS.getValueType() == MVT::f32) {
3751 LHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3752 bitcastf32Toi32(LHS, DAG), Mask);
3753 RHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3754 bitcastf32Toi32(RHS, DAG), Mask);
3755 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3756 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3757 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3758 Chain, Dest, ARMcc, CCR, Cmp);
3763 expandf64Toi32(LHS, DAG, LHS1, LHS2);
3764 expandf64Toi32(RHS, DAG, RHS1, RHS2);
3765 LHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, LHS2, Mask);
3766 RHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, RHS2, Mask);
3767 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3768 ARMcc = DAG.getConstant(CondCode, dl, MVT::i32);
3769 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3770 SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest };
3771 return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops);
3777 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
3778 SDValue Chain = Op.getOperand(0);
3779 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3780 SDValue LHS = Op.getOperand(2);
3781 SDValue RHS = Op.getOperand(3);
3782 SDValue Dest = Op.getOperand(4);
3785 if (Subtarget->isFPOnlySP() && LHS.getValueType() == MVT::f64) {
3786 DAG.getTargetLoweringInfo().softenSetCCOperands(DAG, MVT::f64, LHS, RHS, CC,
3789 // If softenSetCCOperands only returned one value, we should compare it to
3791 if (!RHS.getNode()) {
3792 RHS = DAG.getConstant(0, dl, LHS.getValueType());
3797 if (LHS.getValueType() == MVT::i32) {
3799 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3800 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3801 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3802 Chain, Dest, ARMcc, CCR, Cmp);
3805 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3807 if (getTargetMachine().Options.UnsafeFPMath &&
3808 (CC == ISD::SETEQ || CC == ISD::SETOEQ ||
3809 CC == ISD::SETNE || CC == ISD::SETUNE)) {
3810 SDValue Result = OptimizeVFPBrcond(Op, DAG);
3811 if (Result.getNode())
3815 ARMCC::CondCodes CondCode, CondCode2;
3816 FPCCToARMCC(CC, CondCode, CondCode2);
3818 SDValue ARMcc = DAG.getConstant(CondCode, dl, MVT::i32);
3819 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
3820 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3821 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3822 SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp };
3823 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops);
3824 if (CondCode2 != ARMCC::AL) {
3825 ARMcc = DAG.getConstant(CondCode2, dl, MVT::i32);
3826 SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) };
3827 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops);
3832 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
3833 SDValue Chain = Op.getOperand(0);
3834 SDValue Table = Op.getOperand(1);
3835 SDValue Index = Op.getOperand(2);
3838 EVT PTy = getPointerTy();
3839 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
3840 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
3841 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI);
3842 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, dl, PTy));
3843 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
3844 if (Subtarget->isThumb2()) {
3845 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
3846 // which does another jump to the destination. This also makes it easier
3847 // to translate it to TBB / TBH later.
3848 // FIXME: This might not work if the function is extremely large.
3849 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
3850 Addr, Op.getOperand(2), JTI);
3852 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
3853 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
3854 MachinePointerInfo::getJumpTable(),
3855 false, false, false, 0);
3856 Chain = Addr.getValue(1);
3857 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
3858 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI);
3860 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
3861 MachinePointerInfo::getJumpTable(),
3862 false, false, false, 0);
3863 Chain = Addr.getValue(1);
3864 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI);
3868 static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3869 EVT VT = Op.getValueType();
3872 if (Op.getValueType().getVectorElementType() == MVT::i32) {
3873 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::f32)
3875 return DAG.UnrollVectorOp(Op.getNode());
3878 assert(Op.getOperand(0).getValueType() == MVT::v4f32 &&
3879 "Invalid type for custom lowering!");
3880 if (VT != MVT::v4i16)
3881 return DAG.UnrollVectorOp(Op.getNode());
3883 Op = DAG.getNode(Op.getOpcode(), dl, MVT::v4i32, Op.getOperand(0));
3884 return DAG.getNode(ISD::TRUNCATE, dl, VT, Op);
3887 SDValue ARMTargetLowering::LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) const {
3888 EVT VT = Op.getValueType();
3890 return LowerVectorFP_TO_INT(Op, DAG);
3891 if (Subtarget->isFPOnlySP() && Op.getOperand(0).getValueType() == MVT::f64) {
3893 if (Op.getOpcode() == ISD::FP_TO_SINT)
3894 LC = RTLIB::getFPTOSINT(Op.getOperand(0).getValueType(),
3897 LC = RTLIB::getFPTOUINT(Op.getOperand(0).getValueType(),
3899 return makeLibCall(DAG, LC, Op.getValueType(), &Op.getOperand(0), 1,
3900 /*isSigned*/ false, SDLoc(Op)).first;
3906 static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3907 EVT VT = Op.getValueType();
3910 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::i32) {
3911 if (VT.getVectorElementType() == MVT::f32)
3913 return DAG.UnrollVectorOp(Op.getNode());
3916 assert(Op.getOperand(0).getValueType() == MVT::v4i16 &&
3917 "Invalid type for custom lowering!");
3918 if (VT != MVT::v4f32)
3919 return DAG.UnrollVectorOp(Op.getNode());
3923 switch (Op.getOpcode()) {
3924 default: llvm_unreachable("Invalid opcode!");
3925 case ISD::SINT_TO_FP:
3926 CastOpc = ISD::SIGN_EXTEND;
3927 Opc = ISD::SINT_TO_FP;
3929 case ISD::UINT_TO_FP:
3930 CastOpc = ISD::ZERO_EXTEND;
3931 Opc = ISD::UINT_TO_FP;
3935 Op = DAG.getNode(CastOpc, dl, MVT::v4i32, Op.getOperand(0));
3936 return DAG.getNode(Opc, dl, VT, Op);
3939 SDValue ARMTargetLowering::LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) const {
3940 EVT VT = Op.getValueType();
3942 return LowerVectorINT_TO_FP(Op, DAG);
3943 if (Subtarget->isFPOnlySP() && Op.getValueType() == MVT::f64) {
3945 if (Op.getOpcode() == ISD::SINT_TO_FP)
3946 LC = RTLIB::getSINTTOFP(Op.getOperand(0).getValueType(),
3949 LC = RTLIB::getUINTTOFP(Op.getOperand(0).getValueType(),
3951 return makeLibCall(DAG, LC, Op.getValueType(), &Op.getOperand(0), 1,
3952 /*isSigned*/ false, SDLoc(Op)).first;
3958 SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
3959 // Implement fcopysign with a fabs and a conditional fneg.
3960 SDValue Tmp0 = Op.getOperand(0);
3961 SDValue Tmp1 = Op.getOperand(1);
3963 EVT VT = Op.getValueType();
3964 EVT SrcVT = Tmp1.getValueType();
3965 bool InGPR = Tmp0.getOpcode() == ISD::BITCAST ||
3966 Tmp0.getOpcode() == ARMISD::VMOVDRR;
3967 bool UseNEON = !InGPR && Subtarget->hasNEON();
3970 // Use VBSL to copy the sign bit.
3971 unsigned EncodedVal = ARM_AM::createNEONModImm(0x6, 0x80);
3972 SDValue Mask = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v2i32,
3973 DAG.getTargetConstant(EncodedVal, dl, MVT::i32));
3974 EVT OpVT = (VT == MVT::f32) ? MVT::v2i32 : MVT::v1i64;
3976 Mask = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3977 DAG.getNode(ISD::BITCAST, dl, OpVT, Mask),
3978 DAG.getConstant(32, dl, MVT::i32));
3979 else /*if (VT == MVT::f32)*/
3980 Tmp0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp0);
3981 if (SrcVT == MVT::f32) {
3982 Tmp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp1);
3984 Tmp1 = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3985 DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1),
3986 DAG.getConstant(32, dl, MVT::i32));
3987 } else if (VT == MVT::f32)
3988 Tmp1 = DAG.getNode(ARMISD::VSHRu, dl, MVT::v1i64,
3989 DAG.getNode(ISD::BITCAST, dl, MVT::v1i64, Tmp1),
3990 DAG.getConstant(32, dl, MVT::i32));
3991 Tmp0 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp0);
3992 Tmp1 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1);
3994 SDValue AllOnes = DAG.getTargetConstant(ARM_AM::createNEONModImm(0xe, 0xff),
3996 AllOnes = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v8i8, AllOnes);
3997 SDValue MaskNot = DAG.getNode(ISD::XOR, dl, OpVT, Mask,
3998 DAG.getNode(ISD::BITCAST, dl, OpVT, AllOnes));
4000 SDValue Res = DAG.getNode(ISD::OR, dl, OpVT,
4001 DAG.getNode(ISD::AND, dl, OpVT, Tmp1, Mask),
4002 DAG.getNode(ISD::AND, dl, OpVT, Tmp0, MaskNot));
4003 if (VT == MVT::f32) {
4004 Res = DAG.getNode(ISD::BITCAST, dl, MVT::v2f32, Res);
4005 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, Res,
4006 DAG.getConstant(0, dl, MVT::i32));
4008 Res = DAG.getNode(ISD::BITCAST, dl, MVT::f64, Res);
4014 // Bitcast operand 1 to i32.
4015 if (SrcVT == MVT::f64)
4016 Tmp1 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
4018 Tmp1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp1);
4020 // Or in the signbit with integer operations.
4021 SDValue Mask1 = DAG.getConstant(0x80000000, dl, MVT::i32);
4022 SDValue Mask2 = DAG.getConstant(0x7fffffff, dl, MVT::i32);
4023 Tmp1 = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp1, Mask1);
4024 if (VT == MVT::f32) {
4025 Tmp0 = DAG.getNode(ISD::AND, dl, MVT::i32,
4026 DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp0), Mask2);
4027 return DAG.getNode(ISD::BITCAST, dl, MVT::f32,
4028 DAG.getNode(ISD::OR, dl, MVT::i32, Tmp0, Tmp1));
4031 // f64: Or the high part with signbit and then combine two parts.
4032 Tmp0 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
4034 SDValue Lo = Tmp0.getValue(0);
4035 SDValue Hi = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp0.getValue(1), Mask2);
4036 Hi = DAG.getNode(ISD::OR, dl, MVT::i32, Hi, Tmp1);
4037 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
4040 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
4041 MachineFunction &MF = DAG.getMachineFunction();
4042 MachineFrameInfo *MFI = MF.getFrameInfo();
4043 MFI->setReturnAddressIsTaken(true);
4045 if (verifyReturnAddressArgumentIsConstant(Op, DAG))
4048 EVT VT = Op.getValueType();
4050 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
4052 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
4053 SDValue Offset = DAG.getConstant(4, dl, MVT::i32);
4054 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
4055 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
4056 MachinePointerInfo(), false, false, false, 0);
4059 // Return LR, which contains the return address. Mark it an implicit live-in.
4060 unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32));
4061 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
4064 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
4065 const ARMBaseRegisterInfo &ARI =
4066 *static_cast<const ARMBaseRegisterInfo*>(RegInfo);
4067 MachineFunction &MF = DAG.getMachineFunction();
4068 MachineFrameInfo *MFI = MF.getFrameInfo();
4069 MFI->setFrameAddressIsTaken(true);
4071 EVT VT = Op.getValueType();
4072 SDLoc dl(Op); // FIXME probably not meaningful
4073 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
4074 unsigned FrameReg = ARI.getFrameRegister(MF);
4075 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
4077 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
4078 MachinePointerInfo(),
4079 false, false, false, 0);
4083 // FIXME? Maybe this could be a TableGen attribute on some registers and
4084 // this table could be generated automatically from RegInfo.
4085 unsigned ARMTargetLowering::getRegisterByName(const char* RegName,
4087 unsigned Reg = StringSwitch<unsigned>(RegName)
4088 .Case("sp", ARM::SP)
4092 report_fatal_error("Invalid register name global variable");
4095 /// ExpandBITCAST - If the target supports VFP, this function is called to
4096 /// expand a bit convert where either the source or destination type is i64 to
4097 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
4098 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
4099 /// vectors), since the legalizer won't know what to do with that.
4100 static SDValue ExpandBITCAST(SDNode *N, SelectionDAG &DAG) {
4101 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4103 SDValue Op = N->getOperand(0);
4105 // This function is only supposed to be called for i64 types, either as the
4106 // source or destination of the bit convert.
4107 EVT SrcVT = Op.getValueType();
4108 EVT DstVT = N->getValueType(0);
4109 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
4110 "ExpandBITCAST called for non-i64 type");
4112 // Turn i64->f64 into VMOVDRR.
4113 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
4114 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
4115 DAG.getConstant(0, dl, MVT::i32));
4116 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
4117 DAG.getConstant(1, dl, MVT::i32));
4118 return DAG.getNode(ISD::BITCAST, dl, DstVT,
4119 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
4122 // Turn f64->i64 into VMOVRRD.
4123 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
4125 if (TLI.isBigEndian() && SrcVT.isVector() &&
4126 SrcVT.getVectorNumElements() > 1)
4127 Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
4128 DAG.getVTList(MVT::i32, MVT::i32),
4129 DAG.getNode(ARMISD::VREV64, dl, SrcVT, Op));
4131 Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
4132 DAG.getVTList(MVT::i32, MVT::i32), Op);
4133 // Merge the pieces into a single i64 value.
4134 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
4140 /// getZeroVector - Returns a vector of specified type with all zero elements.
4141 /// Zero vectors are used to represent vector negation and in those cases
4142 /// will be implemented with the NEON VNEG instruction. However, VNEG does
4143 /// not support i64 elements, so sometimes the zero vectors will need to be
4144 /// explicitly constructed. Regardless, use a canonical VMOV to create the
4146 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, SDLoc dl) {
4147 assert(VT.isVector() && "Expected a vector type");
4148 // The canonical modified immediate encoding of a zero vector is....0!
4149 SDValue EncodedVal = DAG.getTargetConstant(0, dl, MVT::i32);
4150 EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
4151 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal);
4152 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4155 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
4156 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
4157 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
4158 SelectionDAG &DAG) const {
4159 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4160 EVT VT = Op.getValueType();
4161 unsigned VTBits = VT.getSizeInBits();
4163 SDValue ShOpLo = Op.getOperand(0);
4164 SDValue ShOpHi = Op.getOperand(1);
4165 SDValue ShAmt = Op.getOperand(2);
4167 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
4169 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
4171 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
4172 DAG.getConstant(VTBits, dl, MVT::i32), ShAmt);
4173 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
4174 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
4175 DAG.getConstant(VTBits, dl, MVT::i32));
4176 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
4177 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
4178 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
4180 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
4181 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, dl, MVT::i32),
4182 ISD::SETGE, ARMcc, DAG, dl);
4183 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
4184 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc,
4187 SDValue Ops[2] = { Lo, Hi };
4188 return DAG.getMergeValues(Ops, dl);
4191 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
4192 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
4193 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
4194 SelectionDAG &DAG) const {
4195 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4196 EVT VT = Op.getValueType();
4197 unsigned VTBits = VT.getSizeInBits();
4199 SDValue ShOpLo = Op.getOperand(0);
4200 SDValue ShOpHi = Op.getOperand(1);
4201 SDValue ShAmt = Op.getOperand(2);
4204 assert(Op.getOpcode() == ISD::SHL_PARTS);
4205 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
4206 DAG.getConstant(VTBits, dl, MVT::i32), ShAmt);
4207 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
4208 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
4209 DAG.getConstant(VTBits, dl, MVT::i32));
4210 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
4211 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
4213 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
4214 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
4215 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, dl, MVT::i32),
4216 ISD::SETGE, ARMcc, DAG, dl);
4217 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
4218 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc,
4221 SDValue Ops[2] = { Lo, Hi };
4222 return DAG.getMergeValues(Ops, dl);
4225 SDValue ARMTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
4226 SelectionDAG &DAG) const {
4227 // The rounding mode is in bits 23:22 of the FPSCR.
4228 // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0
4229 // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3)
4230 // so that the shift + and get folded into a bitfield extract.
4232 SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::i32,
4233 DAG.getConstant(Intrinsic::arm_get_fpscr, dl,
4235 SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPSCR,
4236 DAG.getConstant(1U << 22, dl, MVT::i32));
4237 SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds,
4238 DAG.getConstant(22, dl, MVT::i32));
4239 return DAG.getNode(ISD::AND, dl, MVT::i32, RMODE,
4240 DAG.getConstant(3, dl, MVT::i32));
4243 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
4244 const ARMSubtarget *ST) {
4245 EVT VT = N->getValueType(0);
4248 if (!ST->hasV6T2Ops())
4251 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
4252 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
4255 /// getCTPOP16BitCounts - Returns a v8i8/v16i8 vector containing the bit-count
4256 /// for each 16-bit element from operand, repeated. The basic idea is to
4257 /// leverage vcnt to get the 8-bit counts, gather and add the results.
4259 /// Trace for v4i16:
4260 /// input = [v0 v1 v2 v3 ] (vi 16-bit element)
4261 /// cast: N0 = [w0 w1 w2 w3 w4 w5 w6 w7] (v0 = [w0 w1], wi 8-bit element)
4262 /// vcnt: N1 = [b0 b1 b2 b3 b4 b5 b6 b7] (bi = bit-count of 8-bit element wi)
4263 /// vrev: N2 = [b1 b0 b3 b2 b5 b4 b7 b6]
4264 /// [b0 b1 b2 b3 b4 b5 b6 b7]
4265 /// +[b1 b0 b3 b2 b5 b4 b7 b6]
4266 /// N3=N1+N2 = [k0 k0 k1 k1 k2 k2 k3 k3] (k0 = b0+b1 = bit-count of 16-bit v0,
4267 /// vuzp: = [k0 k1 k2 k3 k0 k1 k2 k3] each ki is 8-bits)
4268 static SDValue getCTPOP16BitCounts(SDNode *N, SelectionDAG &DAG) {
4269 EVT VT = N->getValueType(0);
4272 EVT VT8Bit = VT.is64BitVector() ? MVT::v8i8 : MVT::v16i8;
4273 SDValue N0 = DAG.getNode(ISD::BITCAST, DL, VT8Bit, N->getOperand(0));
4274 SDValue N1 = DAG.getNode(ISD::CTPOP, DL, VT8Bit, N0);
4275 SDValue N2 = DAG.getNode(ARMISD::VREV16, DL, VT8Bit, N1);
4276 SDValue N3 = DAG.getNode(ISD::ADD, DL, VT8Bit, N1, N2);
4277 return DAG.getNode(ARMISD::VUZP, DL, VT8Bit, N3, N3);
4280 /// lowerCTPOP16BitElements - Returns a v4i16/v8i16 vector containing the
4281 /// bit-count for each 16-bit element from the operand. We need slightly
4282 /// different sequencing for v4i16 and v8i16 to stay within NEON's available
4283 /// 64/128-bit registers.
4285 /// Trace for v4i16:
4286 /// input = [v0 v1 v2 v3 ] (vi 16-bit element)
4287 /// v8i8: BitCounts = [k0 k1 k2 k3 k0 k1 k2 k3 ] (ki is the bit-count of vi)
4288 /// v8i16:Extended = [k0 k1 k2 k3 k0 k1 k2 k3 ]
4289 /// v4i16:Extracted = [k0 k1 k2 k3 ]
4290 static SDValue lowerCTPOP16BitElements(SDNode *N, SelectionDAG &DAG) {
4291 EVT VT = N->getValueType(0);
4294 SDValue BitCounts = getCTPOP16BitCounts(N, DAG);
4295 if (VT.is64BitVector()) {
4296 SDValue Extended = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v8i16, BitCounts);
4297 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i16, Extended,
4298 DAG.getIntPtrConstant(0, DL));
4300 SDValue Extracted = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v8i8,
4301 BitCounts, DAG.getIntPtrConstant(0, DL));
4302 return DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v8i16, Extracted);
4306 /// lowerCTPOP32BitElements - Returns a v2i32/v4i32 vector containing the
4307 /// bit-count for each 32-bit element from the operand. The idea here is
4308 /// to split the vector into 16-bit elements, leverage the 16-bit count
4309 /// routine, and then combine the results.
4311 /// Trace for v2i32 (v4i32 similar with Extracted/Extended exchanged):
4312 /// input = [v0 v1 ] (vi: 32-bit elements)
4313 /// Bitcast = [w0 w1 w2 w3 ] (wi: 16-bit elements, v0 = [w0 w1])
4314 /// Counts16 = [k0 k1 k2 k3 ] (ki: 16-bit elements, bit-count of wi)
4315 /// vrev: N0 = [k1 k0 k3 k2 ]
4317 /// N1 =+[k1 k0 k3 k2 ]
4319 /// N2 =+[k1 k3 k0 k2 ]
4321 /// Extended =+[k1 k3 k0 k2 ]
4323 /// Extracted=+[k1 k3 ]
4325 static SDValue lowerCTPOP32BitElements(SDNode *N, SelectionDAG &DAG) {
4326 EVT VT = N->getValueType(0);
4329 EVT VT16Bit = VT.is64BitVector() ? MVT::v4i16 : MVT::v8i16;
4331 SDValue Bitcast = DAG.getNode(ISD::BITCAST, DL, VT16Bit, N->getOperand(0));
4332 SDValue Counts16 = lowerCTPOP16BitElements(Bitcast.getNode(), DAG);
4333 SDValue N0 = DAG.getNode(ARMISD::VREV32, DL, VT16Bit, Counts16);
4334 SDValue N1 = DAG.getNode(ISD::ADD, DL, VT16Bit, Counts16, N0);
4335 SDValue N2 = DAG.getNode(ARMISD::VUZP, DL, VT16Bit, N1, N1);
4337 if (VT.is64BitVector()) {
4338 SDValue Extended = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v4i32, N2);
4339 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i32, Extended,
4340 DAG.getIntPtrConstant(0, DL));
4342 SDValue Extracted = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i16, N2,
4343 DAG.getIntPtrConstant(0, DL));
4344 return DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v4i32, Extracted);
4348 static SDValue LowerCTPOP(SDNode *N, SelectionDAG &DAG,
4349 const ARMSubtarget *ST) {
4350 EVT VT = N->getValueType(0);
4352 assert(ST->hasNEON() && "Custom ctpop lowering requires NEON.");
4353 assert((VT == MVT::v2i32 || VT == MVT::v4i32 ||
4354 VT == MVT::v4i16 || VT == MVT::v8i16) &&
4355 "Unexpected type for custom ctpop lowering");
4357 if (VT.getVectorElementType() == MVT::i32)
4358 return lowerCTPOP32BitElements(N, DAG);
4360 return lowerCTPOP16BitElements(N, DAG);
4363 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
4364 const ARMSubtarget *ST) {
4365 EVT VT = N->getValueType(0);
4371 // Lower vector shifts on NEON to use VSHL.
4372 assert(ST->hasNEON() && "unexpected vector shift");
4374 // Left shifts translate directly to the vshiftu intrinsic.
4375 if (N->getOpcode() == ISD::SHL)
4376 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
4377 DAG.getConstant(Intrinsic::arm_neon_vshiftu, dl,
4379 N->getOperand(0), N->getOperand(1));
4381 assert((N->getOpcode() == ISD::SRA ||
4382 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
4384 // NEON uses the same intrinsics for both left and right shifts. For
4385 // right shifts, the shift amounts are negative, so negate the vector of
4387 EVT ShiftVT = N->getOperand(1).getValueType();
4388 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
4389 getZeroVector(ShiftVT, DAG, dl),
4391 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
4392 Intrinsic::arm_neon_vshifts :
4393 Intrinsic::arm_neon_vshiftu);
4394 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
4395 DAG.getConstant(vshiftInt, dl, MVT::i32),
4396 N->getOperand(0), NegatedCount);
4399 static SDValue Expand64BitShift(SDNode *N, SelectionDAG &DAG,
4400 const ARMSubtarget *ST) {
4401 EVT VT = N->getValueType(0);
4404 // We can get here for a node like i32 = ISD::SHL i32, i64
4408 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
4409 "Unknown shift to lower!");
4411 // We only lower SRA, SRL of 1 here, all others use generic lowering.
4412 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
4413 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
4416 // If we are in thumb mode, we don't have RRX.
4417 if (ST->isThumb1Only()) return SDValue();
4419 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
4420 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
4421 DAG.getConstant(0, dl, MVT::i32));
4422 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
4423 DAG.getConstant(1, dl, MVT::i32));
4425 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
4426 // captures the result into a carry flag.
4427 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
4428 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Glue), Hi);
4430 // The low part is an ARMISD::RRX operand, which shifts the carry in.
4431 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
4433 // Merge the pieces into a single i64 value.
4434 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
4437 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
4438 SDValue TmpOp0, TmpOp1;
4439 bool Invert = false;
4443 SDValue Op0 = Op.getOperand(0);
4444 SDValue Op1 = Op.getOperand(1);
4445 SDValue CC = Op.getOperand(2);
4446 EVT CmpVT = Op0.getValueType().changeVectorElementTypeToInteger();
4447 EVT VT = Op.getValueType();
4448 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
4451 if (Op1.getValueType().isFloatingPoint()) {
4452 switch (SetCCOpcode) {
4453 default: llvm_unreachable("Illegal FP comparison");
4455 case ISD::SETNE: Invert = true; // Fallthrough
4457 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
4459 case ISD::SETLT: Swap = true; // Fallthrough
4461 case ISD::SETGT: Opc = ARMISD::VCGT; break;
4463 case ISD::SETLE: Swap = true; // Fallthrough
4465 case ISD::SETGE: Opc = ARMISD::VCGE; break;
4466 case ISD::SETUGE: Swap = true; // Fallthrough
4467 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
4468 case ISD::SETUGT: Swap = true; // Fallthrough
4469 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
4470 case ISD::SETUEQ: Invert = true; // Fallthrough
4472 // Expand this to (OLT | OGT).
4476 Op0 = DAG.getNode(ARMISD::VCGT, dl, CmpVT, TmpOp1, TmpOp0);
4477 Op1 = DAG.getNode(ARMISD::VCGT, dl, CmpVT, TmpOp0, TmpOp1);
4479 case ISD::SETUO: Invert = true; // Fallthrough
4481 // Expand this to (OLT | OGE).
4485 Op0 = DAG.getNode(ARMISD::VCGT, dl, CmpVT, TmpOp1, TmpOp0);
4486 Op1 = DAG.getNode(ARMISD::VCGE, dl, CmpVT, TmpOp0, TmpOp1);
4490 // Integer comparisons.
4491 switch (SetCCOpcode) {
4492 default: llvm_unreachable("Illegal integer comparison");
4493 case ISD::SETNE: Invert = true;
4494 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
4495 case ISD::SETLT: Swap = true;
4496 case ISD::SETGT: Opc = ARMISD::VCGT; break;
4497 case ISD::SETLE: Swap = true;
4498 case ISD::SETGE: Opc = ARMISD::VCGE; break;
4499 case ISD::SETULT: Swap = true;
4500 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
4501 case ISD::SETULE: Swap = true;
4502 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
4505 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
4506 if (Opc == ARMISD::VCEQ) {
4509 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
4511 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
4514 // Ignore bitconvert.
4515 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BITCAST)
4516 AndOp = AndOp.getOperand(0);
4518 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
4520 Op0 = DAG.getNode(ISD::BITCAST, dl, CmpVT, AndOp.getOperand(0));
4521 Op1 = DAG.getNode(ISD::BITCAST, dl, CmpVT, AndOp.getOperand(1));
4528 std::swap(Op0, Op1);
4530 // If one of the operands is a constant vector zero, attempt to fold the
4531 // comparison to a specialized compare-against-zero form.
4533 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
4535 else if (ISD::isBuildVectorAllZeros(Op0.getNode())) {
4536 if (Opc == ARMISD::VCGE)
4537 Opc = ARMISD::VCLEZ;
4538 else if (Opc == ARMISD::VCGT)
4539 Opc = ARMISD::VCLTZ;
4544 if (SingleOp.getNode()) {
4547 Result = DAG.getNode(ARMISD::VCEQZ, dl, CmpVT, SingleOp); break;
4549 Result = DAG.getNode(ARMISD::VCGEZ, dl, CmpVT, SingleOp); break;
4551 Result = DAG.getNode(ARMISD::VCLEZ, dl, CmpVT, SingleOp); break;
4553 Result = DAG.getNode(ARMISD::VCGTZ, dl, CmpVT, SingleOp); break;
4555 Result = DAG.getNode(ARMISD::VCLTZ, dl, CmpVT, SingleOp); break;
4557 Result = DAG.getNode(Opc, dl, CmpVT, Op0, Op1);
4560 Result = DAG.getNode(Opc, dl, CmpVT, Op0, Op1);
4563 Result = DAG.getSExtOrTrunc(Result, dl, VT);
4566 Result = DAG.getNOT(dl, Result, VT);
4571 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
4572 /// valid vector constant for a NEON instruction with a "modified immediate"
4573 /// operand (e.g., VMOV). If so, return the encoded value.
4574 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
4575 unsigned SplatBitSize, SelectionDAG &DAG,
4576 SDLoc dl, EVT &VT, bool is128Bits,
4577 NEONModImmType type) {
4578 unsigned OpCmode, Imm;
4580 // SplatBitSize is set to the smallest size that splats the vector, so a
4581 // zero vector will always have SplatBitSize == 8. However, NEON modified
4582 // immediate instructions others than VMOV do not support the 8-bit encoding
4583 // of a zero vector, and the default encoding of zero is supposed to be the
4588 switch (SplatBitSize) {
4590 if (type != VMOVModImm)
4592 // Any 1-byte value is OK. Op=0, Cmode=1110.
4593 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
4596 VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
4600 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
4601 VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
4602 if ((SplatBits & ~0xff) == 0) {
4603 // Value = 0x00nn: Op=x, Cmode=100x.
4608 if ((SplatBits & ~0xff00) == 0) {
4609 // Value = 0xnn00: Op=x, Cmode=101x.
4611 Imm = SplatBits >> 8;
4617 // NEON's 32-bit VMOV supports splat values where:
4618 // * only one byte is nonzero, or
4619 // * the least significant byte is 0xff and the second byte is nonzero, or
4620 // * the least significant 2 bytes are 0xff and the third is nonzero.
4621 VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
4622 if ((SplatBits & ~0xff) == 0) {
4623 // Value = 0x000000nn: Op=x, Cmode=000x.
4628 if ((SplatBits & ~0xff00) == 0) {
4629 // Value = 0x0000nn00: Op=x, Cmode=001x.
4631 Imm = SplatBits >> 8;
4634 if ((SplatBits & ~0xff0000) == 0) {
4635 // Value = 0x00nn0000: Op=x, Cmode=010x.
4637 Imm = SplatBits >> 16;
4640 if ((SplatBits & ~0xff000000) == 0) {
4641 // Value = 0xnn000000: Op=x, Cmode=011x.
4643 Imm = SplatBits >> 24;
4647 // cmode == 0b1100 and cmode == 0b1101 are not supported for VORR or VBIC
4648 if (type == OtherModImm) return SDValue();
4650 if ((SplatBits & ~0xffff) == 0 &&
4651 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
4652 // Value = 0x0000nnff: Op=x, Cmode=1100.
4654 Imm = SplatBits >> 8;
4658 if ((SplatBits & ~0xffffff) == 0 &&
4659 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
4660 // Value = 0x00nnffff: Op=x, Cmode=1101.
4662 Imm = SplatBits >> 16;
4666 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
4667 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
4668 // VMOV.I32. A (very) minor optimization would be to replicate the value
4669 // and fall through here to test for a valid 64-bit splat. But, then the
4670 // caller would also need to check and handle the change in size.
4674 if (type != VMOVModImm)
4676 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
4677 uint64_t BitMask = 0xff;
4679 unsigned ImmMask = 1;
4681 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
4682 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
4685 } else if ((SplatBits & BitMask) != 0) {
4692 if (DAG.getTargetLoweringInfo().isBigEndian())
4693 // swap higher and lower 32 bit word
4694 Imm = ((Imm & 0xf) << 4) | ((Imm & 0xf0) >> 4);
4696 // Op=1, Cmode=1110.
4698 VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
4703 llvm_unreachable("unexpected size for isNEONModifiedImm");
4706 unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm);
4707 return DAG.getTargetConstant(EncodedVal, dl, MVT::i32);
4710 SDValue ARMTargetLowering::LowerConstantFP(SDValue Op, SelectionDAG &DAG,
4711 const ARMSubtarget *ST) const {
4715 bool IsDouble = Op.getValueType() == MVT::f64;
4716 ConstantFPSDNode *CFP = cast<ConstantFPSDNode>(Op);
4718 // Use the default (constant pool) lowering for double constants when we have
4720 if (IsDouble && Subtarget->isFPOnlySP())
4723 // Try splatting with a VMOV.f32...
4724 APFloat FPVal = CFP->getValueAPF();
4725 int ImmVal = IsDouble ? ARM_AM::getFP64Imm(FPVal) : ARM_AM::getFP32Imm(FPVal);
4728 if (IsDouble || !ST->useNEONForSinglePrecisionFP()) {
4729 // We have code in place to select a valid ConstantFP already, no need to
4734 // It's a float and we are trying to use NEON operations where
4735 // possible. Lower it to a splat followed by an extract.
4737 SDValue NewVal = DAG.getTargetConstant(ImmVal, DL, MVT::i32);
4738 SDValue VecConstant = DAG.getNode(ARMISD::VMOVFPIMM, DL, MVT::v2f32,
4740 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecConstant,
4741 DAG.getConstant(0, DL, MVT::i32));
4744 // The rest of our options are NEON only, make sure that's allowed before
4746 if (!ST->hasNEON() || (!IsDouble && !ST->useNEONForSinglePrecisionFP()))
4750 uint64_t iVal = FPVal.bitcastToAPInt().getZExtValue();
4752 // It wouldn't really be worth bothering for doubles except for one very
4753 // important value, which does happen to match: 0.0. So make sure we don't do
4755 if (IsDouble && (iVal & 0xffffffff) != (iVal >> 32))
4758 // Try a VMOV.i32 (FIXME: i8, i16, or i64 could work too).
4759 SDValue NewVal = isNEONModifiedImm(iVal & 0xffffffffU, 0, 32, DAG, SDLoc(Op),
4760 VMovVT, false, VMOVModImm);
4761 if (NewVal != SDValue()) {
4763 SDValue VecConstant = DAG.getNode(ARMISD::VMOVIMM, DL, VMovVT,
4766 return DAG.getNode(ISD::BITCAST, DL, MVT::f64, VecConstant);
4768 // It's a float: cast and extract a vector element.
4769 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
4771 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
4772 DAG.getConstant(0, DL, MVT::i32));
4775 // Finally, try a VMVN.i32
4776 NewVal = isNEONModifiedImm(~iVal & 0xffffffffU, 0, 32, DAG, SDLoc(Op), VMovVT,
4778 if (NewVal != SDValue()) {
4780 SDValue VecConstant = DAG.getNode(ARMISD::VMVNIMM, DL, VMovVT, NewVal);
4783 return DAG.getNode(ISD::BITCAST, DL, MVT::f64, VecConstant);
4785 // It's a float: cast and extract a vector element.
4786 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
4788 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
4789 DAG.getConstant(0, DL, MVT::i32));
4795 // check if an VEXT instruction can handle the shuffle mask when the
4796 // vector sources of the shuffle are the same.
4797 static bool isSingletonVEXTMask(ArrayRef<int> M, EVT VT, unsigned &Imm) {
4798 unsigned NumElts = VT.getVectorNumElements();
4800 // Assume that the first shuffle index is not UNDEF. Fail if it is.
4806 // If this is a VEXT shuffle, the immediate value is the index of the first
4807 // element. The other shuffle indices must be the successive elements after
4809 unsigned ExpectedElt = Imm;
4810 for (unsigned i = 1; i < NumElts; ++i) {
4811 // Increment the expected index. If it wraps around, just follow it
4812 // back to index zero and keep going.
4814 if (ExpectedElt == NumElts)
4817 if (M[i] < 0) continue; // ignore UNDEF indices
4818 if (ExpectedElt != static_cast<unsigned>(M[i]))
4826 static bool isVEXTMask(ArrayRef<int> M, EVT VT,
4827 bool &ReverseVEXT, unsigned &Imm) {
4828 unsigned NumElts = VT.getVectorNumElements();
4829 ReverseVEXT = false;
4831 // Assume that the first shuffle index is not UNDEF. Fail if it is.
4837 // If this is a VEXT shuffle, the immediate value is the index of the first
4838 // element. The other shuffle indices must be the successive elements after
4840 unsigned ExpectedElt = Imm;
4841 for (unsigned i = 1; i < NumElts; ++i) {
4842 // Increment the expected index. If it wraps around, it may still be
4843 // a VEXT but the source vectors must be swapped.
4845 if (ExpectedElt == NumElts * 2) {
4850 if (M[i] < 0) continue; // ignore UNDEF indices
4851 if (ExpectedElt != static_cast<unsigned>(M[i]))
4855 // Adjust the index value if the source operands will be swapped.
4862 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
4863 /// instruction with the specified blocksize. (The order of the elements
4864 /// within each block of the vector is reversed.)
4865 static bool isVREVMask(ArrayRef<int> M, EVT VT, unsigned BlockSize) {
4866 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
4867 "Only possible block sizes for VREV are: 16, 32, 64");
4869 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4873 unsigned NumElts = VT.getVectorNumElements();
4874 unsigned BlockElts = M[0] + 1;
4875 // If the first shuffle index is UNDEF, be optimistic.
4877 BlockElts = BlockSize / EltSz;
4879 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
4882 for (unsigned i = 0; i < NumElts; ++i) {
4883 if (M[i] < 0) continue; // ignore UNDEF indices
4884 if ((unsigned) M[i] != (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
4891 static bool isVTBLMask(ArrayRef<int> M, EVT VT) {
4892 // We can handle <8 x i8> vector shuffles. If the index in the mask is out of
4893 // range, then 0 is placed into the resulting vector. So pretty much any mask
4894 // of 8 elements can work here.
4895 return VT == MVT::v8i8 && M.size() == 8;
4898 static bool isVTRNMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4899 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4903 unsigned NumElts = VT.getVectorNumElements();
4904 WhichResult = (M[0] == 0 ? 0 : 1);
4905 for (unsigned i = 0; i < NumElts; i += 2) {
4906 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
4907 (M[i+1] >= 0 && (unsigned) M[i+1] != i + NumElts + WhichResult))
4913 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
4914 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4915 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
4916 static bool isVTRN_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4917 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4921 unsigned NumElts = VT.getVectorNumElements();
4922 WhichResult = (M[0] == 0 ? 0 : 1);
4923 for (unsigned i = 0; i < NumElts; i += 2) {
4924 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
4925 (M[i+1] >= 0 && (unsigned) M[i+1] != i + WhichResult))
4931 static bool isVUZPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4932 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4936 unsigned NumElts = VT.getVectorNumElements();
4937 WhichResult = (M[0] == 0 ? 0 : 1);
4938 for (unsigned i = 0; i != NumElts; ++i) {
4939 if (M[i] < 0) continue; // ignore UNDEF indices
4940 if ((unsigned) M[i] != 2 * i + WhichResult)
4944 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4945 if (VT.is64BitVector() && EltSz == 32)
4951 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
4952 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4953 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
4954 static bool isVUZP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4955 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4959 unsigned Half = VT.getVectorNumElements() / 2;
4960 WhichResult = (M[0] == 0 ? 0 : 1);
4961 for (unsigned j = 0; j != 2; ++j) {
4962 unsigned Idx = WhichResult;
4963 for (unsigned i = 0; i != Half; ++i) {
4964 int MIdx = M[i + j * Half];
4965 if (MIdx >= 0 && (unsigned) MIdx != Idx)
4971 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4972 if (VT.is64BitVector() && EltSz == 32)
4978 static bool isVZIPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4979 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4983 unsigned NumElts = VT.getVectorNumElements();
4984 WhichResult = (M[0] == 0 ? 0 : 1);
4985 unsigned Idx = WhichResult * NumElts / 2;
4986 for (unsigned i = 0; i != NumElts; i += 2) {
4987 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
4988 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx + NumElts))
4993 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4994 if (VT.is64BitVector() && EltSz == 32)
5000 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
5001 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
5002 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
5003 static bool isVZIP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
5004 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
5008 unsigned NumElts = VT.getVectorNumElements();
5009 WhichResult = (M[0] == 0 ? 0 : 1);
5010 unsigned Idx = WhichResult * NumElts / 2;
5011 for (unsigned i = 0; i != NumElts; i += 2) {
5012 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
5013 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx))
5018 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
5019 if (VT.is64BitVector() && EltSz == 32)
5025 /// \return true if this is a reverse operation on an vector.
5026 static bool isReverseMask(ArrayRef<int> M, EVT VT) {
5027 unsigned NumElts = VT.getVectorNumElements();
5028 // Make sure the mask has the right size.
5029 if (NumElts != M.size())
5032 // Look for <15, ..., 3, -1, 1, 0>.
5033 for (unsigned i = 0; i != NumElts; ++i)
5034 if (M[i] >= 0 && M[i] != (int) (NumElts - 1 - i))
5040 // If N is an integer constant that can be moved into a register in one
5041 // instruction, return an SDValue of such a constant (will become a MOV
5042 // instruction). Otherwise return null.
5043 static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG,
5044 const ARMSubtarget *ST, SDLoc dl) {
5046 if (!isa<ConstantSDNode>(N))
5048 Val = cast<ConstantSDNode>(N)->getZExtValue();
5050 if (ST->isThumb1Only()) {
5051 if (Val <= 255 || ~Val <= 255)
5052 return DAG.getConstant(Val, dl, MVT::i32);
5054 if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1)
5055 return DAG.getConstant(Val, dl, MVT::i32);
5060 // If this is a case we can't handle, return null and let the default
5061 // expansion code take care of it.
5062 SDValue ARMTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
5063 const ARMSubtarget *ST) const {
5064 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
5066 EVT VT = Op.getValueType();
5068 APInt SplatBits, SplatUndef;
5069 unsigned SplatBitSize;
5071 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
5072 if (SplatBitSize <= 64) {
5073 // Check if an immediate VMOV works.
5075 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
5076 SplatUndef.getZExtValue(), SplatBitSize,
5077 DAG, dl, VmovVT, VT.is128BitVector(),
5079 if (Val.getNode()) {
5080 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
5081 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
5084 // Try an immediate VMVN.
5085 uint64_t NegatedImm = (~SplatBits).getZExtValue();
5086 Val = isNEONModifiedImm(NegatedImm,
5087 SplatUndef.getZExtValue(), SplatBitSize,
5088 DAG, dl, VmovVT, VT.is128BitVector(),
5090 if (Val.getNode()) {
5091 SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
5092 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
5095 // Use vmov.f32 to materialize other v2f32 and v4f32 splats.
5096 if ((VT == MVT::v2f32 || VT == MVT::v4f32) && SplatBitSize == 32) {
5097 int ImmVal = ARM_AM::getFP32Imm(SplatBits);
5099 SDValue Val = DAG.getTargetConstant(ImmVal, dl, MVT::i32);
5100 return DAG.getNode(ARMISD::VMOVFPIMM, dl, VT, Val);
5106 // Scan through the operands to see if only one value is used.
5108 // As an optimisation, even if more than one value is used it may be more
5109 // profitable to splat with one value then change some lanes.
5111 // Heuristically we decide to do this if the vector has a "dominant" value,
5112 // defined as splatted to more than half of the lanes.
5113 unsigned NumElts = VT.getVectorNumElements();
5114 bool isOnlyLowElement = true;
5115 bool usesOnlyOneValue = true;
5116 bool hasDominantValue = false;
5117 bool isConstant = true;
5119 // Map of the number of times a particular SDValue appears in the
5121 DenseMap<SDValue, unsigned> ValueCounts;
5123 for (unsigned i = 0; i < NumElts; ++i) {
5124 SDValue V = Op.getOperand(i);
5125 if (V.getOpcode() == ISD::UNDEF)
5128 isOnlyLowElement = false;
5129 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
5132 ValueCounts.insert(std::make_pair(V, 0));
5133 unsigned &Count = ValueCounts[V];
5135 // Is this value dominant? (takes up more than half of the lanes)
5136 if (++Count > (NumElts / 2)) {
5137 hasDominantValue = true;
5141 if (ValueCounts.size() != 1)
5142 usesOnlyOneValue = false;
5143 if (!Value.getNode() && ValueCounts.size() > 0)
5144 Value = ValueCounts.begin()->first;
5146 if (ValueCounts.size() == 0)
5147 return DAG.getUNDEF(VT);
5149 // Loads are better lowered with insert_vector_elt/ARMISD::BUILD_VECTOR.
5150 // Keep going if we are hitting this case.
5151 if (isOnlyLowElement && !ISD::isNormalLoad(Value.getNode()))
5152 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
5154 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5156 // Use VDUP for non-constant splats. For f32 constant splats, reduce to
5157 // i32 and try again.
5158 if (hasDominantValue && EltSize <= 32) {
5162 // If we are VDUPing a value that comes directly from a vector, that will
5163 // cause an unnecessary move to and from a GPR, where instead we could
5164 // just use VDUPLANE. We can only do this if the lane being extracted
5165 // is at a constant index, as the VDUP from lane instructions only have
5166 // constant-index forms.
5167 if (Value->getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5168 isa<ConstantSDNode>(Value->getOperand(1))) {
5169 // We need to create a new undef vector to use for the VDUPLANE if the
5170 // size of the vector from which we get the value is different than the
5171 // size of the vector that we need to create. We will insert the element
5172 // such that the register coalescer will remove unnecessary copies.
5173 if (VT != Value->getOperand(0).getValueType()) {
5174 ConstantSDNode *constIndex;
5175 constIndex = dyn_cast<ConstantSDNode>(Value->getOperand(1));
5176 assert(constIndex && "The index is not a constant!");
5177 unsigned index = constIndex->getAPIntValue().getLimitedValue() %
5178 VT.getVectorNumElements();
5179 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
5180 DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DAG.getUNDEF(VT),
5181 Value, DAG.getConstant(index, dl, MVT::i32)),
5182 DAG.getConstant(index, dl, MVT::i32));
5184 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
5185 Value->getOperand(0), Value->getOperand(1));
5187 N = DAG.getNode(ARMISD::VDUP, dl, VT, Value);
5189 if (!usesOnlyOneValue) {
5190 // The dominant value was splatted as 'N', but we now have to insert
5191 // all differing elements.
5192 for (unsigned I = 0; I < NumElts; ++I) {
5193 if (Op.getOperand(I) == Value)
5195 SmallVector<SDValue, 3> Ops;
5197 Ops.push_back(Op.getOperand(I));
5198 Ops.push_back(DAG.getConstant(I, dl, MVT::i32));
5199 N = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Ops);
5204 if (VT.getVectorElementType().isFloatingPoint()) {
5205 SmallVector<SDValue, 8> Ops;
5206 for (unsigned i = 0; i < NumElts; ++i)
5207 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, MVT::i32,
5209 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
5210 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, Ops);
5211 Val = LowerBUILD_VECTOR(Val, DAG, ST);
5213 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
5215 if (usesOnlyOneValue) {
5216 SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
5217 if (isConstant && Val.getNode())
5218 return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
5222 // If all elements are constants and the case above didn't get hit, fall back
5223 // to the default expansion, which will generate a load from the constant
5228 // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
5230 SDValue shuffle = ReconstructShuffle(Op, DAG);
5231 if (shuffle != SDValue())
5235 // Vectors with 32- or 64-bit elements can be built by directly assigning
5236 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
5237 // will be legalized.
5238 if (EltSize >= 32) {
5239 // Do the expansion with floating-point types, since that is what the VFP
5240 // registers are defined to use, and since i64 is not legal.
5241 EVT EltVT = EVT::getFloatingPointVT(EltSize);
5242 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
5243 SmallVector<SDValue, 8> Ops;
5244 for (unsigned i = 0; i < NumElts; ++i)
5245 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, EltVT, Op.getOperand(i)));
5246 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, Ops);
5247 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
5250 // If all else fails, just use a sequence of INSERT_VECTOR_ELT when we
5251 // know the default expansion would otherwise fall back on something even
5252 // worse. For a vector with one or two non-undef values, that's
5253 // scalar_to_vector for the elements followed by a shuffle (provided the
5254 // shuffle is valid for the target) and materialization element by element
5255 // on the stack followed by a load for everything else.
5256 if (!isConstant && !usesOnlyOneValue) {
5257 SDValue Vec = DAG.getUNDEF(VT);
5258 for (unsigned i = 0 ; i < NumElts; ++i) {
5259 SDValue V = Op.getOperand(i);
5260 if (V.getOpcode() == ISD::UNDEF)
5262 SDValue LaneIdx = DAG.getConstant(i, dl, MVT::i32);
5263 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Vec, V, LaneIdx);
5271 // Gather data to see if the operation can be modelled as a
5272 // shuffle in combination with VEXTs.
5273 SDValue ARMTargetLowering::ReconstructShuffle(SDValue Op,
5274 SelectionDAG &DAG) const {
5276 EVT VT = Op.getValueType();
5277 unsigned NumElts = VT.getVectorNumElements();
5279 SmallVector<SDValue, 2> SourceVecs;
5280 SmallVector<unsigned, 2> MinElts;
5281 SmallVector<unsigned, 2> MaxElts;
5283 for (unsigned i = 0; i < NumElts; ++i) {
5284 SDValue V = Op.getOperand(i);
5285 if (V.getOpcode() == ISD::UNDEF)
5287 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
5288 // A shuffle can only come from building a vector from various
5289 // elements of other vectors.
5291 } else if (V.getOperand(0).getValueType().getVectorElementType() !=
5292 VT.getVectorElementType()) {
5293 // This code doesn't know how to handle shuffles where the vector
5294 // element types do not match (this happens because type legalization
5295 // promotes the return type of EXTRACT_VECTOR_ELT).
5296 // FIXME: It might be appropriate to extend this code to handle
5297 // mismatched types.
5301 // Record this extraction against the appropriate vector if possible...
5302 SDValue SourceVec = V.getOperand(0);
5303 // If the element number isn't a constant, we can't effectively
5304 // analyze what's going on.
5305 if (!isa<ConstantSDNode>(V.getOperand(1)))
5307 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
5308 bool FoundSource = false;
5309 for (unsigned j = 0; j < SourceVecs.size(); ++j) {
5310 if (SourceVecs[j] == SourceVec) {
5311 if (MinElts[j] > EltNo)
5313 if (MaxElts[j] < EltNo)
5320 // Or record a new source if not...
5322 SourceVecs.push_back(SourceVec);
5323 MinElts.push_back(EltNo);
5324 MaxElts.push_back(EltNo);
5328 // Currently only do something sane when at most two source vectors
5330 if (SourceVecs.size() > 2)
5333 SDValue ShuffleSrcs[2] = {DAG.getUNDEF(VT), DAG.getUNDEF(VT) };
5334 int VEXTOffsets[2] = {0, 0};
5336 // This loop extracts the usage patterns of the source vectors
5337 // and prepares appropriate SDValues for a shuffle if possible.
5338 for (unsigned i = 0; i < SourceVecs.size(); ++i) {
5339 if (SourceVecs[i].getValueType() == VT) {
5340 // No VEXT necessary
5341 ShuffleSrcs[i] = SourceVecs[i];
5344 } else if (SourceVecs[i].getValueType().getVectorNumElements() < NumElts) {
5345 // It probably isn't worth padding out a smaller vector just to
5346 // break it down again in a shuffle.
5350 // Since only 64-bit and 128-bit vectors are legal on ARM and
5351 // we've eliminated the other cases...
5352 assert(SourceVecs[i].getValueType().getVectorNumElements() == 2*NumElts &&
5353 "unexpected vector sizes in ReconstructShuffle");
5355 if (MaxElts[i] - MinElts[i] >= NumElts) {
5356 // Span too large for a VEXT to cope
5360 if (MinElts[i] >= NumElts) {
5361 // The extraction can just take the second half
5362 VEXTOffsets[i] = NumElts;
5363 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5365 DAG.getIntPtrConstant(NumElts, dl));
5366 } else if (MaxElts[i] < NumElts) {
5367 // The extraction can just take the first half
5369 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5371 DAG.getIntPtrConstant(0, dl));
5373 // An actual VEXT is needed
5374 VEXTOffsets[i] = MinElts[i];
5375 SDValue VEXTSrc1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5377 DAG.getIntPtrConstant(0, dl));
5378 SDValue VEXTSrc2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5380 DAG.getIntPtrConstant(NumElts, dl));
5381 ShuffleSrcs[i] = DAG.getNode(ARMISD::VEXT, dl, VT, VEXTSrc1, VEXTSrc2,
5382 DAG.getConstant(VEXTOffsets[i], dl,
5387 SmallVector<int, 8> Mask;
5389 for (unsigned i = 0; i < NumElts; ++i) {
5390 SDValue Entry = Op.getOperand(i);
5391 if (Entry.getOpcode() == ISD::UNDEF) {
5396 SDValue ExtractVec = Entry.getOperand(0);
5397 int ExtractElt = cast<ConstantSDNode>(Op.getOperand(i)
5398 .getOperand(1))->getSExtValue();
5399 if (ExtractVec == SourceVecs[0]) {
5400 Mask.push_back(ExtractElt - VEXTOffsets[0]);
5402 Mask.push_back(ExtractElt + NumElts - VEXTOffsets[1]);
5406 // Final check before we try to produce nonsense...
5407 if (isShuffleMaskLegal(Mask, VT))
5408 return DAG.getVectorShuffle(VT, dl, ShuffleSrcs[0], ShuffleSrcs[1],
5414 /// isShuffleMaskLegal - Targets can use this to indicate that they only
5415 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
5416 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
5417 /// are assumed to be legal.
5419 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
5421 if (VT.getVectorNumElements() == 4 &&
5422 (VT.is128BitVector() || VT.is64BitVector())) {
5423 unsigned PFIndexes[4];
5424 for (unsigned i = 0; i != 4; ++i) {
5428 PFIndexes[i] = M[i];
5431 // Compute the index in the perfect shuffle table.
5432 unsigned PFTableIndex =
5433 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
5434 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
5435 unsigned Cost = (PFEntry >> 30);
5442 unsigned Imm, WhichResult;
5444 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5445 return (EltSize >= 32 ||
5446 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
5447 isVREVMask(M, VT, 64) ||
5448 isVREVMask(M, VT, 32) ||
5449 isVREVMask(M, VT, 16) ||
5450 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
5451 isVTBLMask(M, VT) ||
5452 isVTRNMask(M, VT, WhichResult) ||
5453 isVUZPMask(M, VT, WhichResult) ||
5454 isVZIPMask(M, VT, WhichResult) ||
5455 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
5456 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
5457 isVZIP_v_undef_Mask(M, VT, WhichResult) ||
5458 ((VT == MVT::v8i16 || VT == MVT::v16i8) && isReverseMask(M, VT)));
5461 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
5462 /// the specified operations to build the shuffle.
5463 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
5464 SDValue RHS, SelectionDAG &DAG,
5466 unsigned OpNum = (PFEntry >> 26) & 0x0F;
5467 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
5468 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
5471 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
5480 OP_VUZPL, // VUZP, left result
5481 OP_VUZPR, // VUZP, right result
5482 OP_VZIPL, // VZIP, left result
5483 OP_VZIPR, // VZIP, right result
5484 OP_VTRNL, // VTRN, left result
5485 OP_VTRNR // VTRN, right result
5488 if (OpNum == OP_COPY) {
5489 if (LHSID == (1*9+2)*9+3) return LHS;
5490 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
5494 SDValue OpLHS, OpRHS;
5495 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
5496 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
5497 EVT VT = OpLHS.getValueType();
5500 default: llvm_unreachable("Unknown shuffle opcode!");
5502 // VREV divides the vector in half and swaps within the half.
5503 if (VT.getVectorElementType() == MVT::i32 ||
5504 VT.getVectorElementType() == MVT::f32)
5505 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
5506 // vrev <4 x i16> -> VREV32
5507 if (VT.getVectorElementType() == MVT::i16)
5508 return DAG.getNode(ARMISD::VREV32, dl, VT, OpLHS);
5509 // vrev <4 x i8> -> VREV16
5510 assert(VT.getVectorElementType() == MVT::i8);
5511 return DAG.getNode(ARMISD::VREV16, dl, VT, OpLHS);
5516 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
5517 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, dl, MVT::i32));
5521 return DAG.getNode(ARMISD::VEXT, dl, VT,
5523 DAG.getConstant(OpNum - OP_VEXT1 + 1, dl, MVT::i32));
5526 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
5527 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
5530 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
5531 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
5534 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
5535 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
5539 static SDValue LowerVECTOR_SHUFFLEv8i8(SDValue Op,
5540 ArrayRef<int> ShuffleMask,
5541 SelectionDAG &DAG) {
5542 // Check to see if we can use the VTBL instruction.
5543 SDValue V1 = Op.getOperand(0);
5544 SDValue V2 = Op.getOperand(1);
5547 SmallVector<SDValue, 8> VTBLMask;
5548 for (ArrayRef<int>::iterator
5549 I = ShuffleMask.begin(), E = ShuffleMask.end(); I != E; ++I)
5550 VTBLMask.push_back(DAG.getConstant(*I, DL, MVT::i32));
5552 if (V2.getNode()->getOpcode() == ISD::UNDEF)
5553 return DAG.getNode(ARMISD::VTBL1, DL, MVT::v8i8, V1,
5554 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8, VTBLMask));
5556 return DAG.getNode(ARMISD::VTBL2, DL, MVT::v8i8, V1, V2,
5557 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8, VTBLMask));
5560 static SDValue LowerReverse_VECTOR_SHUFFLEv16i8_v8i16(SDValue Op,
5561 SelectionDAG &DAG) {
5563 SDValue OpLHS = Op.getOperand(0);
5564 EVT VT = OpLHS.getValueType();
5566 assert((VT == MVT::v8i16 || VT == MVT::v16i8) &&
5567 "Expect an v8i16/v16i8 type");
5568 OpLHS = DAG.getNode(ARMISD::VREV64, DL, VT, OpLHS);
5569 // For a v16i8 type: After the VREV, we have got <8, ...15, 8, ..., 0>. Now,
5570 // extract the first 8 bytes into the top double word and the last 8 bytes
5571 // into the bottom double word. The v8i16 case is similar.
5572 unsigned ExtractNum = (VT == MVT::v16i8) ? 8 : 4;
5573 return DAG.getNode(ARMISD::VEXT, DL, VT, OpLHS, OpLHS,
5574 DAG.getConstant(ExtractNum, DL, MVT::i32));
5577 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
5578 SDValue V1 = Op.getOperand(0);
5579 SDValue V2 = Op.getOperand(1);
5581 EVT VT = Op.getValueType();
5582 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
5584 // Convert shuffles that are directly supported on NEON to target-specific
5585 // DAG nodes, instead of keeping them as shuffles and matching them again
5586 // during code selection. This is more efficient and avoids the possibility
5587 // of inconsistencies between legalization and selection.
5588 // FIXME: floating-point vectors should be canonicalized to integer vectors
5589 // of the same time so that they get CSEd properly.
5590 ArrayRef<int> ShuffleMask = SVN->getMask();
5592 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5593 if (EltSize <= 32) {
5594 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
5595 int Lane = SVN->getSplatIndex();
5596 // If this is undef splat, generate it via "just" vdup, if possible.
5597 if (Lane == -1) Lane = 0;
5599 // Test if V1 is a SCALAR_TO_VECTOR.
5600 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
5601 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
5603 // Test if V1 is a BUILD_VECTOR which is equivalent to a SCALAR_TO_VECTOR
5604 // (and probably will turn into a SCALAR_TO_VECTOR once legalization
5606 if (Lane == 0 && V1.getOpcode() == ISD::BUILD_VECTOR &&
5607 !isa<ConstantSDNode>(V1.getOperand(0))) {
5608 bool IsScalarToVector = true;
5609 for (unsigned i = 1, e = V1.getNumOperands(); i != e; ++i)
5610 if (V1.getOperand(i).getOpcode() != ISD::UNDEF) {
5611 IsScalarToVector = false;
5614 if (IsScalarToVector)
5615 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
5617 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
5618 DAG.getConstant(Lane, dl, MVT::i32));
5623 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
5626 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
5627 DAG.getConstant(Imm, dl, MVT::i32));
5630 if (isVREVMask(ShuffleMask, VT, 64))
5631 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
5632 if (isVREVMask(ShuffleMask, VT, 32))
5633 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
5634 if (isVREVMask(ShuffleMask, VT, 16))
5635 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
5637 if (V2->getOpcode() == ISD::UNDEF &&
5638 isSingletonVEXTMask(ShuffleMask, VT, Imm)) {
5639 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V1,
5640 DAG.getConstant(Imm, dl, MVT::i32));
5643 // Check for Neon shuffles that modify both input vectors in place.
5644 // If both results are used, i.e., if there are two shuffles with the same
5645 // source operands and with masks corresponding to both results of one of
5646 // these operations, DAG memoization will ensure that a single node is
5647 // used for both shuffles.
5648 unsigned WhichResult;
5649 if (isVTRNMask(ShuffleMask, VT, WhichResult))
5650 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
5651 V1, V2).getValue(WhichResult);
5652 if (isVUZPMask(ShuffleMask, VT, WhichResult))
5653 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
5654 V1, V2).getValue(WhichResult);
5655 if (isVZIPMask(ShuffleMask, VT, WhichResult))
5656 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
5657 V1, V2).getValue(WhichResult);
5659 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
5660 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
5661 V1, V1).getValue(WhichResult);
5662 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
5663 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
5664 V1, V1).getValue(WhichResult);
5665 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
5666 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
5667 V1, V1).getValue(WhichResult);
5670 // If the shuffle is not directly supported and it has 4 elements, use
5671 // the PerfectShuffle-generated table to synthesize it from other shuffles.
5672 unsigned NumElts = VT.getVectorNumElements();
5674 unsigned PFIndexes[4];
5675 for (unsigned i = 0; i != 4; ++i) {
5676 if (ShuffleMask[i] < 0)
5679 PFIndexes[i] = ShuffleMask[i];
5682 // Compute the index in the perfect shuffle table.
5683 unsigned PFTableIndex =
5684 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
5685 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
5686 unsigned Cost = (PFEntry >> 30);
5689 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
5692 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
5693 if (EltSize >= 32) {
5694 // Do the expansion with floating-point types, since that is what the VFP
5695 // registers are defined to use, and since i64 is not legal.
5696 EVT EltVT = EVT::getFloatingPointVT(EltSize);
5697 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
5698 V1 = DAG.getNode(ISD::BITCAST, dl, VecVT, V1);
5699 V2 = DAG.getNode(ISD::BITCAST, dl, VecVT, V2);
5700 SmallVector<SDValue, 8> Ops;
5701 for (unsigned i = 0; i < NumElts; ++i) {
5702 if (ShuffleMask[i] < 0)
5703 Ops.push_back(DAG.getUNDEF(EltVT));
5705 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
5706 ShuffleMask[i] < (int)NumElts ? V1 : V2,
5707 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
5710 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, Ops);
5711 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
5714 if ((VT == MVT::v8i16 || VT == MVT::v16i8) && isReverseMask(ShuffleMask, VT))
5715 return LowerReverse_VECTOR_SHUFFLEv16i8_v8i16(Op, DAG);
5717 if (VT == MVT::v8i8) {
5718 SDValue NewOp = LowerVECTOR_SHUFFLEv8i8(Op, ShuffleMask, DAG);
5719 if (NewOp.getNode())
5726 static SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
5727 // INSERT_VECTOR_ELT is legal only for immediate indexes.
5728 SDValue Lane = Op.getOperand(2);
5729 if (!isa<ConstantSDNode>(Lane))
5735 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
5736 // EXTRACT_VECTOR_ELT is legal only for immediate indexes.
5737 SDValue Lane = Op.getOperand(1);
5738 if (!isa<ConstantSDNode>(Lane))
5741 SDValue Vec = Op.getOperand(0);
5742 if (Op.getValueType() == MVT::i32 &&
5743 Vec.getValueType().getVectorElementType().getSizeInBits() < 32) {
5745 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
5751 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
5752 // The only time a CONCAT_VECTORS operation can have legal types is when
5753 // two 64-bit vectors are concatenated to a 128-bit vector.
5754 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
5755 "unexpected CONCAT_VECTORS");
5757 SDValue Val = DAG.getUNDEF(MVT::v2f64);
5758 SDValue Op0 = Op.getOperand(0);
5759 SDValue Op1 = Op.getOperand(1);
5760 if (Op0.getOpcode() != ISD::UNDEF)
5761 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
5762 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op0),
5763 DAG.getIntPtrConstant(0, dl));
5764 if (Op1.getOpcode() != ISD::UNDEF)
5765 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
5766 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op1),
5767 DAG.getIntPtrConstant(1, dl));
5768 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Val);
5771 /// isExtendedBUILD_VECTOR - Check if N is a constant BUILD_VECTOR where each
5772 /// element has been zero/sign-extended, depending on the isSigned parameter,
5773 /// from an integer type half its size.
5774 static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG,
5776 // A v2i64 BUILD_VECTOR will have been legalized to a BITCAST from v4i32.
5777 EVT VT = N->getValueType(0);
5778 if (VT == MVT::v2i64 && N->getOpcode() == ISD::BITCAST) {
5779 SDNode *BVN = N->getOperand(0).getNode();
5780 if (BVN->getValueType(0) != MVT::v4i32 ||
5781 BVN->getOpcode() != ISD::BUILD_VECTOR)
5783 unsigned LoElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
5784 unsigned HiElt = 1 - LoElt;
5785 ConstantSDNode *Lo0 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt));
5786 ConstantSDNode *Hi0 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt));
5787 ConstantSDNode *Lo1 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt+2));
5788 ConstantSDNode *Hi1 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt+2));
5789 if (!Lo0 || !Hi0 || !Lo1 || !Hi1)
5792 if (Hi0->getSExtValue() == Lo0->getSExtValue() >> 32 &&
5793 Hi1->getSExtValue() == Lo1->getSExtValue() >> 32)
5796 if (Hi0->isNullValue() && Hi1->isNullValue())
5802 if (N->getOpcode() != ISD::BUILD_VECTOR)
5805 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
5806 SDNode *Elt = N->getOperand(i).getNode();
5807 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
5808 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5809 unsigned HalfSize = EltSize / 2;
5811 if (!isIntN(HalfSize, C->getSExtValue()))
5814 if (!isUIntN(HalfSize, C->getZExtValue()))
5825 /// isSignExtended - Check if a node is a vector value that is sign-extended
5826 /// or a constant BUILD_VECTOR with sign-extended elements.
5827 static bool isSignExtended(SDNode *N, SelectionDAG &DAG) {
5828 if (N->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N))
5830 if (isExtendedBUILD_VECTOR(N, DAG, true))
5835 /// isZeroExtended - Check if a node is a vector value that is zero-extended
5836 /// or a constant BUILD_VECTOR with zero-extended elements.
5837 static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) {
5838 if (N->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N))
5840 if (isExtendedBUILD_VECTOR(N, DAG, false))
5845 static EVT getExtensionTo64Bits(const EVT &OrigVT) {
5846 if (OrigVT.getSizeInBits() >= 64)
5849 assert(OrigVT.isSimple() && "Expecting a simple value type");
5851 MVT::SimpleValueType OrigSimpleTy = OrigVT.getSimpleVT().SimpleTy;
5852 switch (OrigSimpleTy) {
5853 default: llvm_unreachable("Unexpected Vector Type");
5862 /// AddRequiredExtensionForVMULL - Add a sign/zero extension to extend the total
5863 /// value size to 64 bits. We need a 64-bit D register as an operand to VMULL.
5864 /// We insert the required extension here to get the vector to fill a D register.
5865 static SDValue AddRequiredExtensionForVMULL(SDValue N, SelectionDAG &DAG,
5868 unsigned ExtOpcode) {
5869 // The vector originally had a size of OrigTy. It was then extended to ExtTy.
5870 // We expect the ExtTy to be 128-bits total. If the OrigTy is less than
5871 // 64-bits we need to insert a new extension so that it will be 64-bits.
5872 assert(ExtTy.is128BitVector() && "Unexpected extension size");
5873 if (OrigTy.getSizeInBits() >= 64)
5876 // Must extend size to at least 64 bits to be used as an operand for VMULL.
5877 EVT NewVT = getExtensionTo64Bits(OrigTy);
5879 return DAG.getNode(ExtOpcode, SDLoc(N), NewVT, N);
5882 /// SkipLoadExtensionForVMULL - return a load of the original vector size that
5883 /// does not do any sign/zero extension. If the original vector is less
5884 /// than 64 bits, an appropriate extension will be added after the load to
5885 /// reach a total size of 64 bits. We have to add the extension separately
5886 /// because ARM does not have a sign/zero extending load for vectors.
5887 static SDValue SkipLoadExtensionForVMULL(LoadSDNode *LD, SelectionDAG& DAG) {
5888 EVT ExtendedTy = getExtensionTo64Bits(LD->getMemoryVT());
5890 // The load already has the right type.
5891 if (ExtendedTy == LD->getMemoryVT())
5892 return DAG.getLoad(LD->getMemoryVT(), SDLoc(LD), LD->getChain(),
5893 LD->getBasePtr(), LD->getPointerInfo(), LD->isVolatile(),
5894 LD->isNonTemporal(), LD->isInvariant(),
5895 LD->getAlignment());
5897 // We need to create a zextload/sextload. We cannot just create a load
5898 // followed by a zext/zext node because LowerMUL is also run during normal
5899 // operation legalization where we can't create illegal types.
5900 return DAG.getExtLoad(LD->getExtensionType(), SDLoc(LD), ExtendedTy,
5901 LD->getChain(), LD->getBasePtr(), LD->getPointerInfo(),
5902 LD->getMemoryVT(), LD->isVolatile(), LD->isInvariant(),
5903 LD->isNonTemporal(), LD->getAlignment());
5906 /// SkipExtensionForVMULL - For a node that is a SIGN_EXTEND, ZERO_EXTEND,
5907 /// extending load, or BUILD_VECTOR with extended elements, return the
5908 /// unextended value. The unextended vector should be 64 bits so that it can
5909 /// be used as an operand to a VMULL instruction. If the original vector size
5910 /// before extension is less than 64 bits we add a an extension to resize
5911 /// the vector to 64 bits.
5912 static SDValue SkipExtensionForVMULL(SDNode *N, SelectionDAG &DAG) {
5913 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
5914 return AddRequiredExtensionForVMULL(N->getOperand(0), DAG,
5915 N->getOperand(0)->getValueType(0),
5919 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
5920 return SkipLoadExtensionForVMULL(LD, DAG);
5922 // Otherwise, the value must be a BUILD_VECTOR. For v2i64, it will
5923 // have been legalized as a BITCAST from v4i32.
5924 if (N->getOpcode() == ISD::BITCAST) {
5925 SDNode *BVN = N->getOperand(0).getNode();
5926 assert(BVN->getOpcode() == ISD::BUILD_VECTOR &&
5927 BVN->getValueType(0) == MVT::v4i32 && "expected v4i32 BUILD_VECTOR");
5928 unsigned LowElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
5929 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), MVT::v2i32,
5930 BVN->getOperand(LowElt), BVN->getOperand(LowElt+2));
5932 // Construct a new BUILD_VECTOR with elements truncated to half the size.
5933 assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR");
5934 EVT VT = N->getValueType(0);
5935 unsigned EltSize = VT.getVectorElementType().getSizeInBits() / 2;
5936 unsigned NumElts = VT.getVectorNumElements();
5937 MVT TruncVT = MVT::getIntegerVT(EltSize);
5938 SmallVector<SDValue, 8> Ops;
5940 for (unsigned i = 0; i != NumElts; ++i) {
5941 ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i));
5942 const APInt &CInt = C->getAPIntValue();
5943 // Element types smaller than 32 bits are not legal, so use i32 elements.
5944 // The values are implicitly truncated so sext vs. zext doesn't matter.
5945 Ops.push_back(DAG.getConstant(CInt.zextOrTrunc(32), dl, MVT::i32));
5947 return DAG.getNode(ISD::BUILD_VECTOR, dl,
5948 MVT::getVectorVT(TruncVT, NumElts), Ops);
5951 static bool isAddSubSExt(SDNode *N, SelectionDAG &DAG) {
5952 unsigned Opcode = N->getOpcode();
5953 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
5954 SDNode *N0 = N->getOperand(0).getNode();
5955 SDNode *N1 = N->getOperand(1).getNode();
5956 return N0->hasOneUse() && N1->hasOneUse() &&
5957 isSignExtended(N0, DAG) && isSignExtended(N1, DAG);
5962 static bool isAddSubZExt(SDNode *N, SelectionDAG &DAG) {
5963 unsigned Opcode = N->getOpcode();
5964 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
5965 SDNode *N0 = N->getOperand(0).getNode();
5966 SDNode *N1 = N->getOperand(1).getNode();
5967 return N0->hasOneUse() && N1->hasOneUse() &&
5968 isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG);
5973 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
5974 // Multiplications are only custom-lowered for 128-bit vectors so that
5975 // VMULL can be detected. Otherwise v2i64 multiplications are not legal.
5976 EVT VT = Op.getValueType();
5977 assert(VT.is128BitVector() && VT.isInteger() &&
5978 "unexpected type for custom-lowering ISD::MUL");
5979 SDNode *N0 = Op.getOperand(0).getNode();
5980 SDNode *N1 = Op.getOperand(1).getNode();
5981 unsigned NewOpc = 0;
5983 bool isN0SExt = isSignExtended(N0, DAG);
5984 bool isN1SExt = isSignExtended(N1, DAG);
5985 if (isN0SExt && isN1SExt)
5986 NewOpc = ARMISD::VMULLs;
5988 bool isN0ZExt = isZeroExtended(N0, DAG);
5989 bool isN1ZExt = isZeroExtended(N1, DAG);
5990 if (isN0ZExt && isN1ZExt)
5991 NewOpc = ARMISD::VMULLu;
5992 else if (isN1SExt || isN1ZExt) {
5993 // Look for (s/zext A + s/zext B) * (s/zext C). We want to turn these
5994 // into (s/zext A * s/zext C) + (s/zext B * s/zext C)
5995 if (isN1SExt && isAddSubSExt(N0, DAG)) {
5996 NewOpc = ARMISD::VMULLs;
5998 } else if (isN1ZExt && isAddSubZExt(N0, DAG)) {
5999 NewOpc = ARMISD::VMULLu;
6001 } else if (isN0ZExt && isAddSubZExt(N1, DAG)) {
6003 NewOpc = ARMISD::VMULLu;
6009 if (VT == MVT::v2i64)
6010 // Fall through to expand this. It is not legal.
6013 // Other vector multiplications are legal.
6018 // Legalize to a VMULL instruction.
6021 SDValue Op1 = SkipExtensionForVMULL(N1, DAG);
6023 Op0 = SkipExtensionForVMULL(N0, DAG);
6024 assert(Op0.getValueType().is64BitVector() &&
6025 Op1.getValueType().is64BitVector() &&
6026 "unexpected types for extended operands to VMULL");
6027 return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
6030 // Optimizing (zext A + zext B) * C, to (VMULL A, C) + (VMULL B, C) during
6031 // isel lowering to take advantage of no-stall back to back vmul + vmla.
6038 SDValue N00 = SkipExtensionForVMULL(N0->getOperand(0).getNode(), DAG);
6039 SDValue N01 = SkipExtensionForVMULL(N0->getOperand(1).getNode(), DAG);
6040 EVT Op1VT = Op1.getValueType();
6041 return DAG.getNode(N0->getOpcode(), DL, VT,
6042 DAG.getNode(NewOpc, DL, VT,
6043 DAG.getNode(ISD::BITCAST, DL, Op1VT, N00), Op1),
6044 DAG.getNode(NewOpc, DL, VT,
6045 DAG.getNode(ISD::BITCAST, DL, Op1VT, N01), Op1));
6049 LowerSDIV_v4i8(SDValue X, SDValue Y, SDLoc dl, SelectionDAG &DAG) {
6051 // float4 xf = vcvt_f32_s32(vmovl_s16(a.lo));
6052 // float4 yf = vcvt_f32_s32(vmovl_s16(b.lo));
6053 X = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, X);
6054 Y = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, Y);
6055 X = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, X);
6056 Y = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, Y);
6057 // Get reciprocal estimate.
6058 // float4 recip = vrecpeq_f32(yf);
6059 Y = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6060 DAG.getConstant(Intrinsic::arm_neon_vrecpe, dl, MVT::i32),
6062 // Because char has a smaller range than uchar, we can actually get away
6063 // without any newton steps. This requires that we use a weird bias
6064 // of 0xb000, however (again, this has been exhaustively tested).
6065 // float4 result = as_float4(as_int4(xf*recip) + 0xb000);
6066 X = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, X, Y);
6067 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, X);
6068 Y = DAG.getConstant(0xb000, dl, MVT::i32);
6069 Y = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Y, Y, Y, Y);
6070 X = DAG.getNode(ISD::ADD, dl, MVT::v4i32, X, Y);
6071 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, X);
6072 // Convert back to short.
6073 X = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, X);
6074 X = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, X);
6079 LowerSDIV_v4i16(SDValue N0, SDValue N1, SDLoc dl, SelectionDAG &DAG) {
6081 // Convert to float.
6082 // float4 yf = vcvt_f32_s32(vmovl_s16(y));
6083 // float4 xf = vcvt_f32_s32(vmovl_s16(x));
6084 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N0);
6085 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N1);
6086 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
6087 N1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
6089 // Use reciprocal estimate and one refinement step.
6090 // float4 recip = vrecpeq_f32(yf);
6091 // recip *= vrecpsq_f32(yf, recip);
6092 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6093 DAG.getConstant(Intrinsic::arm_neon_vrecpe, dl, MVT::i32),
6095 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6096 DAG.getConstant(Intrinsic::arm_neon_vrecps, dl, MVT::i32),
6098 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
6099 // Because short has a smaller range than ushort, we can actually get away
6100 // with only a single newton step. This requires that we use a weird bias
6101 // of 89, however (again, this has been exhaustively tested).
6102 // float4 result = as_float4(as_int4(xf*recip) + 0x89);
6103 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
6104 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
6105 N1 = DAG.getConstant(0x89, dl, MVT::i32);
6106 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
6107 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
6108 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
6109 // Convert back to integer and return.
6110 // return vmovn_s32(vcvt_s32_f32(result));
6111 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
6112 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
6116 static SDValue LowerSDIV(SDValue Op, SelectionDAG &DAG) {
6117 EVT VT = Op.getValueType();
6118 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
6119 "unexpected type for custom-lowering ISD::SDIV");
6122 SDValue N0 = Op.getOperand(0);
6123 SDValue N1 = Op.getOperand(1);
6126 if (VT == MVT::v8i8) {
6127 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N0);
6128 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N1);
6130 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
6131 DAG.getIntPtrConstant(4, dl));
6132 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
6133 DAG.getIntPtrConstant(4, dl));
6134 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
6135 DAG.getIntPtrConstant(0, dl));
6136 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
6137 DAG.getIntPtrConstant(0, dl));
6139 N0 = LowerSDIV_v4i8(N0, N1, dl, DAG); // v4i16
6140 N2 = LowerSDIV_v4i8(N2, N3, dl, DAG); // v4i16
6142 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
6143 N0 = LowerCONCAT_VECTORS(N0, DAG);
6145 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v8i8, N0);
6148 return LowerSDIV_v4i16(N0, N1, dl, DAG);
6151 static SDValue LowerUDIV(SDValue Op, SelectionDAG &DAG) {
6152 EVT VT = Op.getValueType();
6153 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
6154 "unexpected type for custom-lowering ISD::UDIV");
6157 SDValue N0 = Op.getOperand(0);
6158 SDValue N1 = Op.getOperand(1);
6161 if (VT == MVT::v8i8) {
6162 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N0);
6163 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N1);
6165 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
6166 DAG.getIntPtrConstant(4, dl));
6167 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
6168 DAG.getIntPtrConstant(4, dl));
6169 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
6170 DAG.getIntPtrConstant(0, dl));
6171 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
6172 DAG.getIntPtrConstant(0, dl));
6174 N0 = LowerSDIV_v4i16(N0, N1, dl, DAG); // v4i16
6175 N2 = LowerSDIV_v4i16(N2, N3, dl, DAG); // v4i16
6177 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
6178 N0 = LowerCONCAT_VECTORS(N0, DAG);
6180 N0 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v8i8,
6181 DAG.getConstant(Intrinsic::arm_neon_vqmovnsu, dl,
6187 // v4i16 sdiv ... Convert to float.
6188 // float4 yf = vcvt_f32_s32(vmovl_u16(y));
6189 // float4 xf = vcvt_f32_s32(vmovl_u16(x));
6190 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N0);
6191 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N1);
6192 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
6193 SDValue BN1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
6195 // Use reciprocal estimate and two refinement steps.
6196 // float4 recip = vrecpeq_f32(yf);
6197 // recip *= vrecpsq_f32(yf, recip);
6198 // recip *= vrecpsq_f32(yf, recip);
6199 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6200 DAG.getConstant(Intrinsic::arm_neon_vrecpe, dl, MVT::i32),
6202 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6203 DAG.getConstant(Intrinsic::arm_neon_vrecps, dl, MVT::i32),
6205 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
6206 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6207 DAG.getConstant(Intrinsic::arm_neon_vrecps, dl, MVT::i32),
6209 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
6210 // Simply multiplying by the reciprocal estimate can leave us a few ulps
6211 // too low, so we add 2 ulps (exhaustive testing shows that this is enough,
6212 // and that it will never cause us to return an answer too large).
6213 // float4 result = as_float4(as_int4(xf*recip) + 2);
6214 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
6215 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
6216 N1 = DAG.getConstant(2, dl, MVT::i32);
6217 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
6218 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
6219 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
6220 // Convert back to integer and return.
6221 // return vmovn_u32(vcvt_s32_f32(result));
6222 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
6223 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
6227 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
6228 EVT VT = Op.getNode()->getValueType(0);
6229 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
6232 bool ExtraOp = false;
6233 switch (Op.getOpcode()) {
6234 default: llvm_unreachable("Invalid code");
6235 case ISD::ADDC: Opc = ARMISD::ADDC; break;
6236 case ISD::ADDE: Opc = ARMISD::ADDE; ExtraOp = true; break;
6237 case ISD::SUBC: Opc = ARMISD::SUBC; break;
6238 case ISD::SUBE: Opc = ARMISD::SUBE; ExtraOp = true; break;
6242 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
6244 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
6245 Op.getOperand(1), Op.getOperand(2));
6248 SDValue ARMTargetLowering::LowerFSINCOS(SDValue Op, SelectionDAG &DAG) const {
6249 assert(Subtarget->isTargetDarwin());
6251 // For iOS, we want to call an alternative entry point: __sincos_stret,
6252 // return values are passed via sret.
6254 SDValue Arg = Op.getOperand(0);
6255 EVT ArgVT = Arg.getValueType();
6256 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
6258 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
6259 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6261 // Pair of floats / doubles used to pass the result.
6262 StructType *RetTy = StructType::get(ArgTy, ArgTy, nullptr);
6264 // Create stack object for sret.
6265 const uint64_t ByteSize = TLI.getDataLayout()->getTypeAllocSize(RetTy);
6266 const unsigned StackAlign = TLI.getDataLayout()->getPrefTypeAlignment(RetTy);
6267 int FrameIdx = FrameInfo->CreateStackObject(ByteSize, StackAlign, false);
6268 SDValue SRet = DAG.getFrameIndex(FrameIdx, TLI.getPointerTy());
6274 Entry.Ty = RetTy->getPointerTo();
6275 Entry.isSExt = false;
6276 Entry.isZExt = false;
6277 Entry.isSRet = true;
6278 Args.push_back(Entry);
6282 Entry.isSExt = false;
6283 Entry.isZExt = false;
6284 Args.push_back(Entry);
6286 const char *LibcallName = (ArgVT == MVT::f64)
6287 ? "__sincos_stret" : "__sincosf_stret";
6288 SDValue Callee = DAG.getExternalSymbol(LibcallName, getPointerTy());
6290 TargetLowering::CallLoweringInfo CLI(DAG);
6291 CLI.setDebugLoc(dl).setChain(DAG.getEntryNode())
6292 .setCallee(CallingConv::C, Type::getVoidTy(*DAG.getContext()), Callee,
6294 .setDiscardResult();
6296 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
6298 SDValue LoadSin = DAG.getLoad(ArgVT, dl, CallResult.second, SRet,
6299 MachinePointerInfo(), false, false, false, 0);
6301 // Address of cos field.
6302 SDValue Add = DAG.getNode(ISD::ADD, dl, getPointerTy(), SRet,
6303 DAG.getIntPtrConstant(ArgVT.getStoreSize(), dl));
6304 SDValue LoadCos = DAG.getLoad(ArgVT, dl, LoadSin.getValue(1), Add,
6305 MachinePointerInfo(), false, false, false, 0);
6307 SDVTList Tys = DAG.getVTList(ArgVT, ArgVT);
6308 return DAG.getNode(ISD::MERGE_VALUES, dl, Tys,
6309 LoadSin.getValue(0), LoadCos.getValue(0));
6312 static SDValue LowerAtomicLoadStore(SDValue Op, SelectionDAG &DAG) {
6313 // Monotonic load/store is legal for all targets
6314 if (cast<AtomicSDNode>(Op)->getOrdering() <= Monotonic)
6317 // Acquire/Release load/store is not legal for targets without a
6318 // dmb or equivalent available.
6322 static void ReplaceREADCYCLECOUNTER(SDNode *N,
6323 SmallVectorImpl<SDValue> &Results,
6325 const ARMSubtarget *Subtarget) {
6327 SDValue Cycles32, OutChain;
6329 if (Subtarget->hasPerfMon()) {
6330 // Under Power Management extensions, the cycle-count is:
6331 // mrc p15, #0, <Rt>, c9, c13, #0
6332 SDValue Ops[] = { N->getOperand(0), // Chain
6333 DAG.getConstant(Intrinsic::arm_mrc, DL, MVT::i32),
6334 DAG.getConstant(15, DL, MVT::i32),
6335 DAG.getConstant(0, DL, MVT::i32),
6336 DAG.getConstant(9, DL, MVT::i32),
6337 DAG.getConstant(13, DL, MVT::i32),
6338 DAG.getConstant(0, DL, MVT::i32)
6341 Cycles32 = DAG.getNode(ISD::INTRINSIC_W_CHAIN, DL,
6342 DAG.getVTList(MVT::i32, MVT::Other), Ops);
6343 OutChain = Cycles32.getValue(1);
6345 // Intrinsic is defined to return 0 on unsupported platforms. Technically
6346 // there are older ARM CPUs that have implementation-specific ways of
6347 // obtaining this information (FIXME!).
6348 Cycles32 = DAG.getConstant(0, DL, MVT::i32);
6349 OutChain = DAG.getEntryNode();
6353 SDValue Cycles64 = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64,
6354 Cycles32, DAG.getConstant(0, DL, MVT::i32));
6355 Results.push_back(Cycles64);
6356 Results.push_back(OutChain);
6359 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
6360 switch (Op.getOpcode()) {
6361 default: llvm_unreachable("Don't know how to custom lower this!");
6362 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
6363 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
6364 case ISD::GlobalAddress:
6365 switch (Subtarget->getTargetTriple().getObjectFormat()) {
6366 default: llvm_unreachable("unknown object format");
6368 return LowerGlobalAddressWindows(Op, DAG);
6370 return LowerGlobalAddressELF(Op, DAG);
6372 return LowerGlobalAddressDarwin(Op, DAG);
6374 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
6375 case ISD::SELECT: return LowerSELECT(Op, DAG);
6376 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
6377 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
6378 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
6379 case ISD::VASTART: return LowerVASTART(Op, DAG);
6380 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, DAG, Subtarget);
6381 case ISD::PREFETCH: return LowerPREFETCH(Op, DAG, Subtarget);
6382 case ISD::SINT_TO_FP:
6383 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
6384 case ISD::FP_TO_SINT:
6385 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
6386 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
6387 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
6388 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
6389 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
6390 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
6391 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
6392 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
6394 case ISD::BITCAST: return ExpandBITCAST(Op.getNode(), DAG);
6397 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
6398 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
6399 case ISD::SRL_PARTS:
6400 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
6401 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
6402 case ISD::CTPOP: return LowerCTPOP(Op.getNode(), DAG, Subtarget);
6403 case ISD::SETCC: return LowerVSETCC(Op, DAG);
6404 case ISD::ConstantFP: return LowerConstantFP(Op, DAG, Subtarget);
6405 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG, Subtarget);
6406 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
6407 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
6408 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
6409 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
6410 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
6411 case ISD::MUL: return LowerMUL(Op, DAG);
6412 case ISD::SDIV: return LowerSDIV(Op, DAG);
6413 case ISD::UDIV: return LowerUDIV(Op, DAG);
6417 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
6422 return LowerXALUO(Op, DAG);
6423 case ISD::ATOMIC_LOAD:
6424 case ISD::ATOMIC_STORE: return LowerAtomicLoadStore(Op, DAG);
6425 case ISD::FSINCOS: return LowerFSINCOS(Op, DAG);
6427 case ISD::UDIVREM: return LowerDivRem(Op, DAG);
6428 case ISD::DYNAMIC_STACKALLOC:
6429 if (Subtarget->getTargetTriple().isWindowsItaniumEnvironment())
6430 return LowerDYNAMIC_STACKALLOC(Op, DAG);
6431 llvm_unreachable("Don't know how to custom lower this!");
6432 case ISD::FP_ROUND: return LowerFP_ROUND(Op, DAG);
6433 case ISD::FP_EXTEND: return LowerFP_EXTEND(Op, DAG);
6437 /// ReplaceNodeResults - Replace the results of node with an illegal result
6438 /// type with new values built out of custom code.
6439 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
6440 SmallVectorImpl<SDValue>&Results,
6441 SelectionDAG &DAG) const {
6443 switch (N->getOpcode()) {
6445 llvm_unreachable("Don't know how to custom expand this!");
6447 Res = ExpandBITCAST(N, DAG);
6451 Res = Expand64BitShift(N, DAG, Subtarget);
6453 case ISD::READCYCLECOUNTER:
6454 ReplaceREADCYCLECOUNTER(N, Results, DAG, Subtarget);
6458 Results.push_back(Res);
6461 //===----------------------------------------------------------------------===//
6462 // ARM Scheduler Hooks
6463 //===----------------------------------------------------------------------===//
6465 /// SetupEntryBlockForSjLj - Insert code into the entry block that creates and
6466 /// registers the function context.
6467 void ARMTargetLowering::
6468 SetupEntryBlockForSjLj(MachineInstr *MI, MachineBasicBlock *MBB,
6469 MachineBasicBlock *DispatchBB, int FI) const {
6470 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
6471 DebugLoc dl = MI->getDebugLoc();
6472 MachineFunction *MF = MBB->getParent();
6473 MachineRegisterInfo *MRI = &MF->getRegInfo();
6474 MachineConstantPool *MCP = MF->getConstantPool();
6475 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
6476 const Function *F = MF->getFunction();
6478 bool isThumb = Subtarget->isThumb();
6479 bool isThumb2 = Subtarget->isThumb2();
6481 unsigned PCLabelId = AFI->createPICLabelUId();
6482 unsigned PCAdj = (isThumb || isThumb2) ? 4 : 8;
6483 ARMConstantPoolValue *CPV =
6484 ARMConstantPoolMBB::Create(F->getContext(), DispatchBB, PCLabelId, PCAdj);
6485 unsigned CPI = MCP->getConstantPoolIndex(CPV, 4);
6487 const TargetRegisterClass *TRC = isThumb ? &ARM::tGPRRegClass
6488 : &ARM::GPRRegClass;
6490 // Grab constant pool and fixed stack memory operands.
6491 MachineMemOperand *CPMMO =
6492 MF->getMachineMemOperand(MachinePointerInfo::getConstantPool(),
6493 MachineMemOperand::MOLoad, 4, 4);
6495 MachineMemOperand *FIMMOSt =
6496 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
6497 MachineMemOperand::MOStore, 4, 4);
6499 // Load the address of the dispatch MBB into the jump buffer.
6501 // Incoming value: jbuf
6502 // ldr.n r5, LCPI1_1
6505 // str r5, [$jbuf, #+4] ; &jbuf[1]
6506 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6507 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2LDRpci), NewVReg1)
6508 .addConstantPoolIndex(CPI)
6509 .addMemOperand(CPMMO));
6510 // Set the low bit because of thumb mode.
6511 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6513 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2ORRri), NewVReg2)
6514 .addReg(NewVReg1, RegState::Kill)
6516 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6517 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg3)
6518 .addReg(NewVReg2, RegState::Kill)
6520 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2STRi12))
6521 .addReg(NewVReg3, RegState::Kill)
6523 .addImm(36) // &jbuf[1] :: pc
6524 .addMemOperand(FIMMOSt));
6525 } else if (isThumb) {
6526 // Incoming value: jbuf
6527 // ldr.n r1, LCPI1_4
6531 // add r2, $jbuf, #+4 ; &jbuf[1]
6533 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6534 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tLDRpci), NewVReg1)
6535 .addConstantPoolIndex(CPI)
6536 .addMemOperand(CPMMO));
6537 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6538 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg2)
6539 .addReg(NewVReg1, RegState::Kill)
6541 // Set the low bit because of thumb mode.
6542 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6543 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tMOVi8), NewVReg3)
6544 .addReg(ARM::CPSR, RegState::Define)
6546 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6547 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tORR), NewVReg4)
6548 .addReg(ARM::CPSR, RegState::Define)
6549 .addReg(NewVReg2, RegState::Kill)
6550 .addReg(NewVReg3, RegState::Kill));
6551 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6552 BuildMI(*MBB, MI, dl, TII->get(ARM::tADDframe), NewVReg5)
6554 .addImm(36); // &jbuf[1] :: pc
6555 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tSTRi))
6556 .addReg(NewVReg4, RegState::Kill)
6557 .addReg(NewVReg5, RegState::Kill)
6559 .addMemOperand(FIMMOSt));
6561 // Incoming value: jbuf
6564 // str r1, [$jbuf, #+4] ; &jbuf[1]
6565 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6566 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::LDRi12), NewVReg1)
6567 .addConstantPoolIndex(CPI)
6569 .addMemOperand(CPMMO));
6570 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6571 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::PICADD), NewVReg2)
6572 .addReg(NewVReg1, RegState::Kill)
6573 .addImm(PCLabelId));
6574 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::STRi12))
6575 .addReg(NewVReg2, RegState::Kill)
6577 .addImm(36) // &jbuf[1] :: pc
6578 .addMemOperand(FIMMOSt));
6582 void ARMTargetLowering::EmitSjLjDispatchBlock(MachineInstr *MI,
6583 MachineBasicBlock *MBB) const {
6584 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
6585 DebugLoc dl = MI->getDebugLoc();
6586 MachineFunction *MF = MBB->getParent();
6587 MachineRegisterInfo *MRI = &MF->getRegInfo();
6588 MachineFrameInfo *MFI = MF->getFrameInfo();
6589 int FI = MFI->getFunctionContextIndex();
6591 const TargetRegisterClass *TRC = Subtarget->isThumb() ? &ARM::tGPRRegClass
6592 : &ARM::GPRnopcRegClass;
6594 // Get a mapping of the call site numbers to all of the landing pads they're
6596 DenseMap<unsigned, SmallVector<MachineBasicBlock*, 2> > CallSiteNumToLPad;
6597 unsigned MaxCSNum = 0;
6598 MachineModuleInfo &MMI = MF->getMMI();
6599 for (MachineFunction::iterator BB = MF->begin(), E = MF->end(); BB != E;
6601 if (!BB->isLandingPad()) continue;
6603 // FIXME: We should assert that the EH_LABEL is the first MI in the landing
6605 for (MachineBasicBlock::iterator
6606 II = BB->begin(), IE = BB->end(); II != IE; ++II) {
6607 if (!II->isEHLabel()) continue;
6609 MCSymbol *Sym = II->getOperand(0).getMCSymbol();
6610 if (!MMI.hasCallSiteLandingPad(Sym)) continue;
6612 SmallVectorImpl<unsigned> &CallSiteIdxs = MMI.getCallSiteLandingPad(Sym);
6613 for (SmallVectorImpl<unsigned>::iterator
6614 CSI = CallSiteIdxs.begin(), CSE = CallSiteIdxs.end();
6615 CSI != CSE; ++CSI) {
6616 CallSiteNumToLPad[*CSI].push_back(BB);
6617 MaxCSNum = std::max(MaxCSNum, *CSI);
6623 // Get an ordered list of the machine basic blocks for the jump table.
6624 std::vector<MachineBasicBlock*> LPadList;
6625 SmallPtrSet<MachineBasicBlock*, 64> InvokeBBs;
6626 LPadList.reserve(CallSiteNumToLPad.size());
6627 for (unsigned I = 1; I <= MaxCSNum; ++I) {
6628 SmallVectorImpl<MachineBasicBlock*> &MBBList = CallSiteNumToLPad[I];
6629 for (SmallVectorImpl<MachineBasicBlock*>::iterator
6630 II = MBBList.begin(), IE = MBBList.end(); II != IE; ++II) {
6631 LPadList.push_back(*II);
6632 InvokeBBs.insert((*II)->pred_begin(), (*II)->pred_end());
6636 assert(!LPadList.empty() &&
6637 "No landing pad destinations for the dispatch jump table!");
6639 // Create the jump table and associated information.
6640 MachineJumpTableInfo *JTI =
6641 MF->getOrCreateJumpTableInfo(MachineJumpTableInfo::EK_Inline);
6642 unsigned MJTI = JTI->createJumpTableIndex(LPadList);
6643 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
6645 // Create the MBBs for the dispatch code.
6647 // Shove the dispatch's address into the return slot in the function context.
6648 MachineBasicBlock *DispatchBB = MF->CreateMachineBasicBlock();
6649 DispatchBB->setIsLandingPad();
6651 MachineBasicBlock *TrapBB = MF->CreateMachineBasicBlock();
6652 unsigned trap_opcode;
6653 if (Subtarget->isThumb())
6654 trap_opcode = ARM::tTRAP;
6656 trap_opcode = Subtarget->useNaClTrap() ? ARM::TRAPNaCl : ARM::TRAP;
6658 BuildMI(TrapBB, dl, TII->get(trap_opcode));
6659 DispatchBB->addSuccessor(TrapBB);
6661 MachineBasicBlock *DispContBB = MF->CreateMachineBasicBlock();
6662 DispatchBB->addSuccessor(DispContBB);
6665 MF->insert(MF->end(), DispatchBB);
6666 MF->insert(MF->end(), DispContBB);
6667 MF->insert(MF->end(), TrapBB);
6669 // Insert code into the entry block that creates and registers the function
6671 SetupEntryBlockForSjLj(MI, MBB, DispatchBB, FI);
6673 MachineMemOperand *FIMMOLd =
6674 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
6675 MachineMemOperand::MOLoad |
6676 MachineMemOperand::MOVolatile, 4, 4);
6678 MachineInstrBuilder MIB;
6679 MIB = BuildMI(DispatchBB, dl, TII->get(ARM::Int_eh_sjlj_dispatchsetup));
6681 const ARMBaseInstrInfo *AII = static_cast<const ARMBaseInstrInfo*>(TII);
6682 const ARMBaseRegisterInfo &RI = AII->getRegisterInfo();
6684 // Add a register mask with no preserved registers. This results in all
6685 // registers being marked as clobbered.
6686 MIB.addRegMask(RI.getNoPreservedMask());
6688 unsigned NumLPads = LPadList.size();
6689 if (Subtarget->isThumb2()) {
6690 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6691 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2LDRi12), NewVReg1)
6694 .addMemOperand(FIMMOLd));
6696 if (NumLPads < 256) {
6697 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPri))
6699 .addImm(LPadList.size()));
6701 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6702 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVi16), VReg1)
6703 .addImm(NumLPads & 0xFFFF));
6705 unsigned VReg2 = VReg1;
6706 if ((NumLPads & 0xFFFF0000) != 0) {
6707 VReg2 = MRI->createVirtualRegister(TRC);
6708 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVTi16), VReg2)
6710 .addImm(NumLPads >> 16));
6713 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPrr))
6718 BuildMI(DispatchBB, dl, TII->get(ARM::t2Bcc))
6723 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6724 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::t2LEApcrelJT),NewVReg3)
6725 .addJumpTableIndex(MJTI));
6727 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6730 BuildMI(DispContBB, dl, TII->get(ARM::t2ADDrs), NewVReg4)
6731 .addReg(NewVReg3, RegState::Kill)
6733 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
6735 BuildMI(DispContBB, dl, TII->get(ARM::t2BR_JT))
6736 .addReg(NewVReg4, RegState::Kill)
6738 .addJumpTableIndex(MJTI);
6739 } else if (Subtarget->isThumb()) {
6740 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6741 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRspi), NewVReg1)
6744 .addMemOperand(FIMMOLd));
6746 if (NumLPads < 256) {
6747 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPi8))
6751 MachineConstantPool *ConstantPool = MF->getConstantPool();
6752 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
6753 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
6755 // MachineConstantPool wants an explicit alignment.
6756 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
6758 Align = getDataLayout()->getTypeAllocSize(C->getType());
6759 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6761 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6762 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRpci))
6763 .addReg(VReg1, RegState::Define)
6764 .addConstantPoolIndex(Idx));
6765 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPr))
6770 BuildMI(DispatchBB, dl, TII->get(ARM::tBcc))
6775 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6776 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLSLri), NewVReg2)
6777 .addReg(ARM::CPSR, RegState::Define)
6781 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6782 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLEApcrelJT), NewVReg3)
6783 .addJumpTableIndex(MJTI));
6785 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6786 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg4)
6787 .addReg(ARM::CPSR, RegState::Define)
6788 .addReg(NewVReg2, RegState::Kill)
6791 MachineMemOperand *JTMMOLd =
6792 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
6793 MachineMemOperand::MOLoad, 4, 4);
6795 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6796 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLDRi), NewVReg5)
6797 .addReg(NewVReg4, RegState::Kill)
6799 .addMemOperand(JTMMOLd));
6801 unsigned NewVReg6 = NewVReg5;
6802 if (RelocM == Reloc::PIC_) {
6803 NewVReg6 = MRI->createVirtualRegister(TRC);
6804 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg6)
6805 .addReg(ARM::CPSR, RegState::Define)
6806 .addReg(NewVReg5, RegState::Kill)
6810 BuildMI(DispContBB, dl, TII->get(ARM::tBR_JTr))
6811 .addReg(NewVReg6, RegState::Kill)
6812 .addJumpTableIndex(MJTI);
6814 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6815 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRi12), NewVReg1)
6818 .addMemOperand(FIMMOLd));
6820 if (NumLPads < 256) {
6821 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPri))
6824 } else if (Subtarget->hasV6T2Ops() && isUInt<16>(NumLPads)) {
6825 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6826 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVi16), VReg1)
6827 .addImm(NumLPads & 0xFFFF));
6829 unsigned VReg2 = VReg1;
6830 if ((NumLPads & 0xFFFF0000) != 0) {
6831 VReg2 = MRI->createVirtualRegister(TRC);
6832 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVTi16), VReg2)
6834 .addImm(NumLPads >> 16));
6837 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
6841 MachineConstantPool *ConstantPool = MF->getConstantPool();
6842 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
6843 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
6845 // MachineConstantPool wants an explicit alignment.
6846 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
6848 Align = getDataLayout()->getTypeAllocSize(C->getType());
6849 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6851 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6852 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRcp))
6853 .addReg(VReg1, RegState::Define)
6854 .addConstantPoolIndex(Idx)
6856 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
6858 .addReg(VReg1, RegState::Kill));
6861 BuildMI(DispatchBB, dl, TII->get(ARM::Bcc))
6866 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6868 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::MOVsi), NewVReg3)
6870 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
6871 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6872 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::LEApcrelJT), NewVReg4)
6873 .addJumpTableIndex(MJTI));
6875 MachineMemOperand *JTMMOLd =
6876 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
6877 MachineMemOperand::MOLoad, 4, 4);
6878 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6880 BuildMI(DispContBB, dl, TII->get(ARM::LDRrs), NewVReg5)
6881 .addReg(NewVReg3, RegState::Kill)
6884 .addMemOperand(JTMMOLd));
6886 if (RelocM == Reloc::PIC_) {
6887 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTadd))
6888 .addReg(NewVReg5, RegState::Kill)
6890 .addJumpTableIndex(MJTI);
6892 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTr))
6893 .addReg(NewVReg5, RegState::Kill)
6894 .addJumpTableIndex(MJTI);
6898 // Add the jump table entries as successors to the MBB.
6899 SmallPtrSet<MachineBasicBlock*, 8> SeenMBBs;
6900 for (std::vector<MachineBasicBlock*>::iterator
6901 I = LPadList.begin(), E = LPadList.end(); I != E; ++I) {
6902 MachineBasicBlock *CurMBB = *I;
6903 if (SeenMBBs.insert(CurMBB).second)
6904 DispContBB->addSuccessor(CurMBB);
6907 // N.B. the order the invoke BBs are processed in doesn't matter here.
6908 const MCPhysReg *SavedRegs = RI.getCalleeSavedRegs(MF);
6909 SmallVector<MachineBasicBlock*, 64> MBBLPads;
6910 for (MachineBasicBlock *BB : InvokeBBs) {
6912 // Remove the landing pad successor from the invoke block and replace it
6913 // with the new dispatch block.
6914 SmallVector<MachineBasicBlock*, 4> Successors(BB->succ_begin(),
6916 while (!Successors.empty()) {
6917 MachineBasicBlock *SMBB = Successors.pop_back_val();
6918 if (SMBB->isLandingPad()) {
6919 BB->removeSuccessor(SMBB);
6920 MBBLPads.push_back(SMBB);
6924 BB->addSuccessor(DispatchBB);
6926 // Find the invoke call and mark all of the callee-saved registers as
6927 // 'implicit defined' so that they're spilled. This prevents code from
6928 // moving instructions to before the EH block, where they will never be
6930 for (MachineBasicBlock::reverse_iterator
6931 II = BB->rbegin(), IE = BB->rend(); II != IE; ++II) {
6932 if (!II->isCall()) continue;
6934 DenseMap<unsigned, bool> DefRegs;
6935 for (MachineInstr::mop_iterator
6936 OI = II->operands_begin(), OE = II->operands_end();
6938 if (!OI->isReg()) continue;
6939 DefRegs[OI->getReg()] = true;
6942 MachineInstrBuilder MIB(*MF, &*II);
6944 for (unsigned i = 0; SavedRegs[i] != 0; ++i) {
6945 unsigned Reg = SavedRegs[i];
6946 if (Subtarget->isThumb2() &&
6947 !ARM::tGPRRegClass.contains(Reg) &&
6948 !ARM::hGPRRegClass.contains(Reg))
6950 if (Subtarget->isThumb1Only() && !ARM::tGPRRegClass.contains(Reg))
6952 if (!Subtarget->isThumb() && !ARM::GPRRegClass.contains(Reg))
6955 MIB.addReg(Reg, RegState::ImplicitDefine | RegState::Dead);
6962 // Mark all former landing pads as non-landing pads. The dispatch is the only
6964 for (SmallVectorImpl<MachineBasicBlock*>::iterator
6965 I = MBBLPads.begin(), E = MBBLPads.end(); I != E; ++I)
6966 (*I)->setIsLandingPad(false);
6968 // The instruction is gone now.
6969 MI->eraseFromParent();
6973 MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) {
6974 for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
6975 E = MBB->succ_end(); I != E; ++I)
6978 llvm_unreachable("Expecting a BB with two successors!");
6981 /// Return the load opcode for a given load size. If load size >= 8,
6982 /// neon opcode will be returned.
6983 static unsigned getLdOpcode(unsigned LdSize, bool IsThumb1, bool IsThumb2) {
6985 return LdSize == 16 ? ARM::VLD1q32wb_fixed
6986 : LdSize == 8 ? ARM::VLD1d32wb_fixed : 0;
6988 return LdSize == 4 ? ARM::tLDRi
6989 : LdSize == 2 ? ARM::tLDRHi
6990 : LdSize == 1 ? ARM::tLDRBi : 0;
6992 return LdSize == 4 ? ARM::t2LDR_POST
6993 : LdSize == 2 ? ARM::t2LDRH_POST
6994 : LdSize == 1 ? ARM::t2LDRB_POST : 0;
6995 return LdSize == 4 ? ARM::LDR_POST_IMM
6996 : LdSize == 2 ? ARM::LDRH_POST
6997 : LdSize == 1 ? ARM::LDRB_POST_IMM : 0;
7000 /// Return the store opcode for a given store size. If store size >= 8,
7001 /// neon opcode will be returned.
7002 static unsigned getStOpcode(unsigned StSize, bool IsThumb1, bool IsThumb2) {
7004 return StSize == 16 ? ARM::VST1q32wb_fixed
7005 : StSize == 8 ? ARM::VST1d32wb_fixed : 0;
7007 return StSize == 4 ? ARM::tSTRi
7008 : StSize == 2 ? ARM::tSTRHi
7009 : StSize == 1 ? ARM::tSTRBi : 0;
7011 return StSize == 4 ? ARM::t2STR_POST
7012 : StSize == 2 ? ARM::t2STRH_POST
7013 : StSize == 1 ? ARM::t2STRB_POST : 0;
7014 return StSize == 4 ? ARM::STR_POST_IMM
7015 : StSize == 2 ? ARM::STRH_POST
7016 : StSize == 1 ? ARM::STRB_POST_IMM : 0;
7019 /// Emit a post-increment load operation with given size. The instructions
7020 /// will be added to BB at Pos.
7021 static void emitPostLd(MachineBasicBlock *BB, MachineInstr *Pos,
7022 const TargetInstrInfo *TII, DebugLoc dl,
7023 unsigned LdSize, unsigned Data, unsigned AddrIn,
7024 unsigned AddrOut, bool IsThumb1, bool IsThumb2) {
7025 unsigned LdOpc = getLdOpcode(LdSize, IsThumb1, IsThumb2);
7026 assert(LdOpc != 0 && "Should have a load opcode");
7028 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
7029 .addReg(AddrOut, RegState::Define).addReg(AddrIn)
7031 } else if (IsThumb1) {
7032 // load + update AddrIn
7033 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
7034 .addReg(AddrIn).addImm(0));
7035 MachineInstrBuilder MIB =
7036 BuildMI(*BB, Pos, dl, TII->get(ARM::tADDi8), AddrOut);
7037 MIB = AddDefaultT1CC(MIB);
7038 MIB.addReg(AddrIn).addImm(LdSize);
7039 AddDefaultPred(MIB);
7040 } else if (IsThumb2) {
7041 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
7042 .addReg(AddrOut, RegState::Define).addReg(AddrIn)
7045 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
7046 .addReg(AddrOut, RegState::Define).addReg(AddrIn)
7047 .addReg(0).addImm(LdSize));
7051 /// Emit a post-increment store operation with given size. The instructions
7052 /// will be added to BB at Pos.
7053 static void emitPostSt(MachineBasicBlock *BB, MachineInstr *Pos,
7054 const TargetInstrInfo *TII, DebugLoc dl,
7055 unsigned StSize, unsigned Data, unsigned AddrIn,
7056 unsigned AddrOut, bool IsThumb1, bool IsThumb2) {
7057 unsigned StOpc = getStOpcode(StSize, IsThumb1, IsThumb2);
7058 assert(StOpc != 0 && "Should have a store opcode");
7060 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc), AddrOut)
7061 .addReg(AddrIn).addImm(0).addReg(Data));
7062 } else if (IsThumb1) {
7063 // store + update AddrIn
7064 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc)).addReg(Data)
7065 .addReg(AddrIn).addImm(0));
7066 MachineInstrBuilder MIB =
7067 BuildMI(*BB, Pos, dl, TII->get(ARM::tADDi8), AddrOut);
7068 MIB = AddDefaultT1CC(MIB);
7069 MIB.addReg(AddrIn).addImm(StSize);
7070 AddDefaultPred(MIB);
7071 } else if (IsThumb2) {
7072 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc), AddrOut)
7073 .addReg(Data).addReg(AddrIn).addImm(StSize));
7075 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc), AddrOut)
7076 .addReg(Data).addReg(AddrIn).addReg(0)
7082 ARMTargetLowering::EmitStructByval(MachineInstr *MI,
7083 MachineBasicBlock *BB) const {
7084 // This pseudo instruction has 3 operands: dst, src, size
7085 // We expand it to a loop if size > Subtarget->getMaxInlineSizeThreshold().
7086 // Otherwise, we will generate unrolled scalar copies.
7087 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
7088 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7089 MachineFunction::iterator It = BB;
7092 unsigned dest = MI->getOperand(0).getReg();
7093 unsigned src = MI->getOperand(1).getReg();
7094 unsigned SizeVal = MI->getOperand(2).getImm();
7095 unsigned Align = MI->getOperand(3).getImm();
7096 DebugLoc dl = MI->getDebugLoc();
7098 MachineFunction *MF = BB->getParent();
7099 MachineRegisterInfo &MRI = MF->getRegInfo();
7100 unsigned UnitSize = 0;
7101 const TargetRegisterClass *TRC = nullptr;
7102 const TargetRegisterClass *VecTRC = nullptr;
7104 bool IsThumb1 = Subtarget->isThumb1Only();
7105 bool IsThumb2 = Subtarget->isThumb2();
7109 } else if (Align & 2) {
7112 // Check whether we can use NEON instructions.
7113 if (!MF->getFunction()->hasFnAttribute(Attribute::NoImplicitFloat) &&
7114 Subtarget->hasNEON()) {
7115 if ((Align % 16 == 0) && SizeVal >= 16)
7117 else if ((Align % 8 == 0) && SizeVal >= 8)
7120 // Can't use NEON instructions.
7125 // Select the correct opcode and register class for unit size load/store
7126 bool IsNeon = UnitSize >= 8;
7127 TRC = (IsThumb1 || IsThumb2) ? &ARM::tGPRRegClass : &ARM::GPRRegClass;
7129 VecTRC = UnitSize == 16 ? &ARM::DPairRegClass
7130 : UnitSize == 8 ? &ARM::DPRRegClass
7133 unsigned BytesLeft = SizeVal % UnitSize;
7134 unsigned LoopSize = SizeVal - BytesLeft;
7136 if (SizeVal <= Subtarget->getMaxInlineSizeThreshold()) {
7137 // Use LDR and STR to copy.
7138 // [scratch, srcOut] = LDR_POST(srcIn, UnitSize)
7139 // [destOut] = STR_POST(scratch, destIn, UnitSize)
7140 unsigned srcIn = src;
7141 unsigned destIn = dest;
7142 for (unsigned i = 0; i < LoopSize; i+=UnitSize) {
7143 unsigned srcOut = MRI.createVirtualRegister(TRC);
7144 unsigned destOut = MRI.createVirtualRegister(TRC);
7145 unsigned scratch = MRI.createVirtualRegister(IsNeon ? VecTRC : TRC);
7146 emitPostLd(BB, MI, TII, dl, UnitSize, scratch, srcIn, srcOut,
7147 IsThumb1, IsThumb2);
7148 emitPostSt(BB, MI, TII, dl, UnitSize, scratch, destIn, destOut,
7149 IsThumb1, IsThumb2);
7154 // Handle the leftover bytes with LDRB and STRB.
7155 // [scratch, srcOut] = LDRB_POST(srcIn, 1)
7156 // [destOut] = STRB_POST(scratch, destIn, 1)
7157 for (unsigned i = 0; i < BytesLeft; i++) {
7158 unsigned srcOut = MRI.createVirtualRegister(TRC);
7159 unsigned destOut = MRI.createVirtualRegister(TRC);
7160 unsigned scratch = MRI.createVirtualRegister(TRC);
7161 emitPostLd(BB, MI, TII, dl, 1, scratch, srcIn, srcOut,
7162 IsThumb1, IsThumb2);
7163 emitPostSt(BB, MI, TII, dl, 1, scratch, destIn, destOut,
7164 IsThumb1, IsThumb2);
7168 MI->eraseFromParent(); // The instruction is gone now.
7172 // Expand the pseudo op to a loop.
7175 // movw varEnd, # --> with thumb2
7177 // ldrcp varEnd, idx --> without thumb2
7178 // fallthrough --> loopMBB
7180 // PHI varPhi, varEnd, varLoop
7181 // PHI srcPhi, src, srcLoop
7182 // PHI destPhi, dst, destLoop
7183 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
7184 // [destLoop] = STR_POST(scratch, destPhi, UnitSize)
7185 // subs varLoop, varPhi, #UnitSize
7187 // fallthrough --> exitMBB
7189 // epilogue to handle left-over bytes
7190 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
7191 // [destOut] = STRB_POST(scratch, destLoop, 1)
7192 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
7193 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
7194 MF->insert(It, loopMBB);
7195 MF->insert(It, exitMBB);
7197 // Transfer the remainder of BB and its successor edges to exitMBB.
7198 exitMBB->splice(exitMBB->begin(), BB,
7199 std::next(MachineBasicBlock::iterator(MI)), BB->end());
7200 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
7202 // Load an immediate to varEnd.
7203 unsigned varEnd = MRI.createVirtualRegister(TRC);
7204 if (Subtarget->useMovt(*MF)) {
7205 unsigned Vtmp = varEnd;
7206 if ((LoopSize & 0xFFFF0000) != 0)
7207 Vtmp = MRI.createVirtualRegister(TRC);
7208 AddDefaultPred(BuildMI(BB, dl,
7209 TII->get(IsThumb2 ? ARM::t2MOVi16 : ARM::MOVi16),
7210 Vtmp).addImm(LoopSize & 0xFFFF));
7212 if ((LoopSize & 0xFFFF0000) != 0)
7213 AddDefaultPred(BuildMI(BB, dl,
7214 TII->get(IsThumb2 ? ARM::t2MOVTi16 : ARM::MOVTi16),
7217 .addImm(LoopSize >> 16));
7219 MachineConstantPool *ConstantPool = MF->getConstantPool();
7220 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
7221 const Constant *C = ConstantInt::get(Int32Ty, LoopSize);
7223 // MachineConstantPool wants an explicit alignment.
7224 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
7226 Align = getDataLayout()->getTypeAllocSize(C->getType());
7227 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
7230 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(ARM::tLDRpci)).addReg(
7231 varEnd, RegState::Define).addConstantPoolIndex(Idx));
7233 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(ARM::LDRcp)).addReg(
7234 varEnd, RegState::Define).addConstantPoolIndex(Idx).addImm(0));
7236 BB->addSuccessor(loopMBB);
7238 // Generate the loop body:
7239 // varPhi = PHI(varLoop, varEnd)
7240 // srcPhi = PHI(srcLoop, src)
7241 // destPhi = PHI(destLoop, dst)
7242 MachineBasicBlock *entryBB = BB;
7244 unsigned varLoop = MRI.createVirtualRegister(TRC);
7245 unsigned varPhi = MRI.createVirtualRegister(TRC);
7246 unsigned srcLoop = MRI.createVirtualRegister(TRC);
7247 unsigned srcPhi = MRI.createVirtualRegister(TRC);
7248 unsigned destLoop = MRI.createVirtualRegister(TRC);
7249 unsigned destPhi = MRI.createVirtualRegister(TRC);
7251 BuildMI(*BB, BB->begin(), dl, TII->get(ARM::PHI), varPhi)
7252 .addReg(varLoop).addMBB(loopMBB)
7253 .addReg(varEnd).addMBB(entryBB);
7254 BuildMI(BB, dl, TII->get(ARM::PHI), srcPhi)
7255 .addReg(srcLoop).addMBB(loopMBB)
7256 .addReg(src).addMBB(entryBB);
7257 BuildMI(BB, dl, TII->get(ARM::PHI), destPhi)
7258 .addReg(destLoop).addMBB(loopMBB)
7259 .addReg(dest).addMBB(entryBB);
7261 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
7262 // [destLoop] = STR_POST(scratch, destPhi, UnitSiz)
7263 unsigned scratch = MRI.createVirtualRegister(IsNeon ? VecTRC : TRC);
7264 emitPostLd(BB, BB->end(), TII, dl, UnitSize, scratch, srcPhi, srcLoop,
7265 IsThumb1, IsThumb2);
7266 emitPostSt(BB, BB->end(), TII, dl, UnitSize, scratch, destPhi, destLoop,
7267 IsThumb1, IsThumb2);
7269 // Decrement loop variable by UnitSize.
7271 MachineInstrBuilder MIB =
7272 BuildMI(*BB, BB->end(), dl, TII->get(ARM::tSUBi8), varLoop);
7273 MIB = AddDefaultT1CC(MIB);
7274 MIB.addReg(varPhi).addImm(UnitSize);
7275 AddDefaultPred(MIB);
7277 MachineInstrBuilder MIB =
7278 BuildMI(*BB, BB->end(), dl,
7279 TII->get(IsThumb2 ? ARM::t2SUBri : ARM::SUBri), varLoop);
7280 AddDefaultCC(AddDefaultPred(MIB.addReg(varPhi).addImm(UnitSize)));
7281 MIB->getOperand(5).setReg(ARM::CPSR);
7282 MIB->getOperand(5).setIsDef(true);
7284 BuildMI(*BB, BB->end(), dl,
7285 TII->get(IsThumb1 ? ARM::tBcc : IsThumb2 ? ARM::t2Bcc : ARM::Bcc))
7286 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
7288 // loopMBB can loop back to loopMBB or fall through to exitMBB.
7289 BB->addSuccessor(loopMBB);
7290 BB->addSuccessor(exitMBB);
7292 // Add epilogue to handle BytesLeft.
7294 MachineInstr *StartOfExit = exitMBB->begin();
7296 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
7297 // [destOut] = STRB_POST(scratch, destLoop, 1)
7298 unsigned srcIn = srcLoop;
7299 unsigned destIn = destLoop;
7300 for (unsigned i = 0; i < BytesLeft; i++) {
7301 unsigned srcOut = MRI.createVirtualRegister(TRC);
7302 unsigned destOut = MRI.createVirtualRegister(TRC);
7303 unsigned scratch = MRI.createVirtualRegister(TRC);
7304 emitPostLd(BB, StartOfExit, TII, dl, 1, scratch, srcIn, srcOut,
7305 IsThumb1, IsThumb2);
7306 emitPostSt(BB, StartOfExit, TII, dl, 1, scratch, destIn, destOut,
7307 IsThumb1, IsThumb2);
7312 MI->eraseFromParent(); // The instruction is gone now.
7317 ARMTargetLowering::EmitLowered__chkstk(MachineInstr *MI,
7318 MachineBasicBlock *MBB) const {
7319 const TargetMachine &TM = getTargetMachine();
7320 const TargetInstrInfo &TII = *Subtarget->getInstrInfo();
7321 DebugLoc DL = MI->getDebugLoc();
7323 assert(Subtarget->isTargetWindows() &&
7324 "__chkstk is only supported on Windows");
7325 assert(Subtarget->isThumb2() && "Windows on ARM requires Thumb-2 mode");
7327 // __chkstk takes the number of words to allocate on the stack in R4, and
7328 // returns the stack adjustment in number of bytes in R4. This will not
7329 // clober any other registers (other than the obvious lr).
7331 // Although, technically, IP should be considered a register which may be
7332 // clobbered, the call itself will not touch it. Windows on ARM is a pure
7333 // thumb-2 environment, so there is no interworking required. As a result, we
7334 // do not expect a veneer to be emitted by the linker, clobbering IP.
7336 // Each module receives its own copy of __chkstk, so no import thunk is
7337 // required, again, ensuring that IP is not clobbered.
7339 // Finally, although some linkers may theoretically provide a trampoline for
7340 // out of range calls (which is quite common due to a 32M range limitation of
7341 // branches for Thumb), we can generate the long-call version via
7342 // -mcmodel=large, alleviating the need for the trampoline which may clobber
7345 switch (TM.getCodeModel()) {
7346 case CodeModel::Small:
7347 case CodeModel::Medium:
7348 case CodeModel::Default:
7349 case CodeModel::Kernel:
7350 BuildMI(*MBB, MI, DL, TII.get(ARM::tBL))
7351 .addImm((unsigned)ARMCC::AL).addReg(0)
7352 .addExternalSymbol("__chkstk")
7353 .addReg(ARM::R4, RegState::Implicit | RegState::Kill)
7354 .addReg(ARM::R4, RegState::Implicit | RegState::Define)
7355 .addReg(ARM::R12, RegState::Implicit | RegState::Define | RegState::Dead);
7357 case CodeModel::Large:
7358 case CodeModel::JITDefault: {
7359 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
7360 unsigned Reg = MRI.createVirtualRegister(&ARM::rGPRRegClass);
7362 BuildMI(*MBB, MI, DL, TII.get(ARM::t2MOVi32imm), Reg)
7363 .addExternalSymbol("__chkstk");
7364 BuildMI(*MBB, MI, DL, TII.get(ARM::tBLXr))
7365 .addImm((unsigned)ARMCC::AL).addReg(0)
7366 .addReg(Reg, RegState::Kill)
7367 .addReg(ARM::R4, RegState::Implicit | RegState::Kill)
7368 .addReg(ARM::R4, RegState::Implicit | RegState::Define)
7369 .addReg(ARM::R12, RegState::Implicit | RegState::Define | RegState::Dead);
7374 AddDefaultCC(AddDefaultPred(BuildMI(*MBB, MI, DL, TII.get(ARM::t2SUBrr),
7376 .addReg(ARM::SP).addReg(ARM::R4)));
7378 MI->eraseFromParent();
7383 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
7384 MachineBasicBlock *BB) const {
7385 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
7386 DebugLoc dl = MI->getDebugLoc();
7387 bool isThumb2 = Subtarget->isThumb2();
7388 switch (MI->getOpcode()) {
7391 llvm_unreachable("Unexpected instr type to insert");
7393 // The Thumb2 pre-indexed stores have the same MI operands, they just
7394 // define them differently in the .td files from the isel patterns, so
7395 // they need pseudos.
7396 case ARM::t2STR_preidx:
7397 MI->setDesc(TII->get(ARM::t2STR_PRE));
7399 case ARM::t2STRB_preidx:
7400 MI->setDesc(TII->get(ARM::t2STRB_PRE));
7402 case ARM::t2STRH_preidx:
7403 MI->setDesc(TII->get(ARM::t2STRH_PRE));
7406 case ARM::STRi_preidx:
7407 case ARM::STRBi_preidx: {
7408 unsigned NewOpc = MI->getOpcode() == ARM::STRi_preidx ?
7409 ARM::STR_PRE_IMM : ARM::STRB_PRE_IMM;
7410 // Decode the offset.
7411 unsigned Offset = MI->getOperand(4).getImm();
7412 bool isSub = ARM_AM::getAM2Op(Offset) == ARM_AM::sub;
7413 Offset = ARM_AM::getAM2Offset(Offset);
7417 MachineMemOperand *MMO = *MI->memoperands_begin();
7418 BuildMI(*BB, MI, dl, TII->get(NewOpc))
7419 .addOperand(MI->getOperand(0)) // Rn_wb
7420 .addOperand(MI->getOperand(1)) // Rt
7421 .addOperand(MI->getOperand(2)) // Rn
7422 .addImm(Offset) // offset (skip GPR==zero_reg)
7423 .addOperand(MI->getOperand(5)) // pred
7424 .addOperand(MI->getOperand(6))
7425 .addMemOperand(MMO);
7426 MI->eraseFromParent();
7429 case ARM::STRr_preidx:
7430 case ARM::STRBr_preidx:
7431 case ARM::STRH_preidx: {
7433 switch (MI->getOpcode()) {
7434 default: llvm_unreachable("unexpected opcode!");
7435 case ARM::STRr_preidx: NewOpc = ARM::STR_PRE_REG; break;
7436 case ARM::STRBr_preidx: NewOpc = ARM::STRB_PRE_REG; break;
7437 case ARM::STRH_preidx: NewOpc = ARM::STRH_PRE; break;
7439 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(NewOpc));
7440 for (unsigned i = 0; i < MI->getNumOperands(); ++i)
7441 MIB.addOperand(MI->getOperand(i));
7442 MI->eraseFromParent();
7446 case ARM::tMOVCCr_pseudo: {
7447 // To "insert" a SELECT_CC instruction, we actually have to insert the
7448 // diamond control-flow pattern. The incoming instruction knows the
7449 // destination vreg to set, the condition code register to branch on, the
7450 // true/false values to select between, and a branch opcode to use.
7451 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7452 MachineFunction::iterator It = BB;
7458 // cmpTY ccX, r1, r2
7460 // fallthrough --> copy0MBB
7461 MachineBasicBlock *thisMBB = BB;
7462 MachineFunction *F = BB->getParent();
7463 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
7464 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
7465 F->insert(It, copy0MBB);
7466 F->insert(It, sinkMBB);
7468 // Transfer the remainder of BB and its successor edges to sinkMBB.
7469 sinkMBB->splice(sinkMBB->begin(), BB,
7470 std::next(MachineBasicBlock::iterator(MI)), BB->end());
7471 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
7473 BB->addSuccessor(copy0MBB);
7474 BB->addSuccessor(sinkMBB);
7476 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
7477 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
7480 // %FalseValue = ...
7481 // # fallthrough to sinkMBB
7484 // Update machine-CFG edges
7485 BB->addSuccessor(sinkMBB);
7488 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
7491 BuildMI(*BB, BB->begin(), dl,
7492 TII->get(ARM::PHI), MI->getOperand(0).getReg())
7493 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
7494 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
7496 MI->eraseFromParent(); // The pseudo instruction is gone now.
7501 case ARM::BCCZi64: {
7502 // If there is an unconditional branch to the other successor, remove it.
7503 BB->erase(std::next(MachineBasicBlock::iterator(MI)), BB->end());
7505 // Compare both parts that make up the double comparison separately for
7507 bool RHSisZero = MI->getOpcode() == ARM::BCCZi64;
7509 unsigned LHS1 = MI->getOperand(1).getReg();
7510 unsigned LHS2 = MI->getOperand(2).getReg();
7512 AddDefaultPred(BuildMI(BB, dl,
7513 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7514 .addReg(LHS1).addImm(0));
7515 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7516 .addReg(LHS2).addImm(0)
7517 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
7519 unsigned RHS1 = MI->getOperand(3).getReg();
7520 unsigned RHS2 = MI->getOperand(4).getReg();
7521 AddDefaultPred(BuildMI(BB, dl,
7522 TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
7523 .addReg(LHS1).addReg(RHS1));
7524 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
7525 .addReg(LHS2).addReg(RHS2)
7526 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
7529 MachineBasicBlock *destMBB = MI->getOperand(RHSisZero ? 3 : 5).getMBB();
7530 MachineBasicBlock *exitMBB = OtherSucc(BB, destMBB);
7531 if (MI->getOperand(0).getImm() == ARMCC::NE)
7532 std::swap(destMBB, exitMBB);
7534 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
7535 .addMBB(destMBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
7537 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2B)).addMBB(exitMBB));
7539 BuildMI(BB, dl, TII->get(ARM::B)) .addMBB(exitMBB);
7541 MI->eraseFromParent(); // The pseudo instruction is gone now.
7545 case ARM::Int_eh_sjlj_setjmp:
7546 case ARM::Int_eh_sjlj_setjmp_nofp:
7547 case ARM::tInt_eh_sjlj_setjmp:
7548 case ARM::t2Int_eh_sjlj_setjmp:
7549 case ARM::t2Int_eh_sjlj_setjmp_nofp:
7550 EmitSjLjDispatchBlock(MI, BB);
7555 // To insert an ABS instruction, we have to insert the
7556 // diamond control-flow pattern. The incoming instruction knows the
7557 // source vreg to test against 0, the destination vreg to set,
7558 // the condition code register to branch on, the
7559 // true/false values to select between, and a branch opcode to use.
7564 // BCC (branch to SinkBB if V0 >= 0)
7565 // RSBBB: V3 = RSBri V2, 0 (compute ABS if V2 < 0)
7566 // SinkBB: V1 = PHI(V2, V3)
7567 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7568 MachineFunction::iterator BBI = BB;
7570 MachineFunction *Fn = BB->getParent();
7571 MachineBasicBlock *RSBBB = Fn->CreateMachineBasicBlock(LLVM_BB);
7572 MachineBasicBlock *SinkBB = Fn->CreateMachineBasicBlock(LLVM_BB);
7573 Fn->insert(BBI, RSBBB);
7574 Fn->insert(BBI, SinkBB);
7576 unsigned int ABSSrcReg = MI->getOperand(1).getReg();
7577 unsigned int ABSDstReg = MI->getOperand(0).getReg();
7578 bool ABSSrcKIll = MI->getOperand(1).isKill();
7579 bool isThumb2 = Subtarget->isThumb2();
7580 MachineRegisterInfo &MRI = Fn->getRegInfo();
7581 // In Thumb mode S must not be specified if source register is the SP or
7582 // PC and if destination register is the SP, so restrict register class
7583 unsigned NewRsbDstReg =
7584 MRI.createVirtualRegister(isThumb2 ? &ARM::rGPRRegClass : &ARM::GPRRegClass);
7586 // Transfer the remainder of BB and its successor edges to sinkMBB.
7587 SinkBB->splice(SinkBB->begin(), BB,
7588 std::next(MachineBasicBlock::iterator(MI)), BB->end());
7589 SinkBB->transferSuccessorsAndUpdatePHIs(BB);
7591 BB->addSuccessor(RSBBB);
7592 BB->addSuccessor(SinkBB);
7594 // fall through to SinkMBB
7595 RSBBB->addSuccessor(SinkBB);
7597 // insert a cmp at the end of BB
7598 AddDefaultPred(BuildMI(BB, dl,
7599 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7600 .addReg(ABSSrcReg).addImm(0));
7602 // insert a bcc with opposite CC to ARMCC::MI at the end of BB
7604 TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)).addMBB(SinkBB)
7605 .addImm(ARMCC::getOppositeCondition(ARMCC::MI)).addReg(ARM::CPSR);
7607 // insert rsbri in RSBBB
7608 // Note: BCC and rsbri will be converted into predicated rsbmi
7609 // by if-conversion pass
7610 BuildMI(*RSBBB, RSBBB->begin(), dl,
7611 TII->get(isThumb2 ? ARM::t2RSBri : ARM::RSBri), NewRsbDstReg)
7612 .addReg(ABSSrcReg, ABSSrcKIll ? RegState::Kill : 0)
7613 .addImm(0).addImm((unsigned)ARMCC::AL).addReg(0).addReg(0);
7615 // insert PHI in SinkBB,
7616 // reuse ABSDstReg to not change uses of ABS instruction
7617 BuildMI(*SinkBB, SinkBB->begin(), dl,
7618 TII->get(ARM::PHI), ABSDstReg)
7619 .addReg(NewRsbDstReg).addMBB(RSBBB)
7620 .addReg(ABSSrcReg).addMBB(BB);
7622 // remove ABS instruction
7623 MI->eraseFromParent();
7625 // return last added BB
7628 case ARM::COPY_STRUCT_BYVAL_I32:
7630 return EmitStructByval(MI, BB);
7631 case ARM::WIN__CHKSTK:
7632 return EmitLowered__chkstk(MI, BB);
7636 void ARMTargetLowering::AdjustInstrPostInstrSelection(MachineInstr *MI,
7637 SDNode *Node) const {
7638 const MCInstrDesc *MCID = &MI->getDesc();
7639 // Adjust potentially 's' setting instructions after isel, i.e. ADC, SBC, RSB,
7640 // RSC. Coming out of isel, they have an implicit CPSR def, but the optional
7641 // operand is still set to noreg. If needed, set the optional operand's
7642 // register to CPSR, and remove the redundant implicit def.
7644 // e.g. ADCS (..., CPSR<imp-def>) -> ADC (... opt:CPSR<def>).
7646 // Rename pseudo opcodes.
7647 unsigned NewOpc = convertAddSubFlagsOpcode(MI->getOpcode());
7649 const ARMBaseInstrInfo *TII = Subtarget->getInstrInfo();
7650 MCID = &TII->get(NewOpc);
7652 assert(MCID->getNumOperands() == MI->getDesc().getNumOperands() + 1 &&
7653 "converted opcode should be the same except for cc_out");
7657 // Add the optional cc_out operand
7658 MI->addOperand(MachineOperand::CreateReg(0, /*isDef=*/true));
7660 unsigned ccOutIdx = MCID->getNumOperands() - 1;
7662 // Any ARM instruction that sets the 's' bit should specify an optional
7663 // "cc_out" operand in the last operand position.
7664 if (!MI->hasOptionalDef() || !MCID->OpInfo[ccOutIdx].isOptionalDef()) {
7665 assert(!NewOpc && "Optional cc_out operand required");
7668 // Look for an implicit def of CPSR added by MachineInstr ctor. Remove it
7669 // since we already have an optional CPSR def.
7670 bool definesCPSR = false;
7671 bool deadCPSR = false;
7672 for (unsigned i = MCID->getNumOperands(), e = MI->getNumOperands();
7674 const MachineOperand &MO = MI->getOperand(i);
7675 if (MO.isReg() && MO.isDef() && MO.getReg() == ARM::CPSR) {
7679 MI->RemoveOperand(i);
7684 assert(!NewOpc && "Optional cc_out operand required");
7687 assert(deadCPSR == !Node->hasAnyUseOfValue(1) && "inconsistent dead flag");
7689 assert(!MI->getOperand(ccOutIdx).getReg() &&
7690 "expect uninitialized optional cc_out operand");
7694 // If this instruction was defined with an optional CPSR def and its dag node
7695 // had a live implicit CPSR def, then activate the optional CPSR def.
7696 MachineOperand &MO = MI->getOperand(ccOutIdx);
7697 MO.setReg(ARM::CPSR);
7701 //===----------------------------------------------------------------------===//
7702 // ARM Optimization Hooks
7703 //===----------------------------------------------------------------------===//
7705 // Helper function that checks if N is a null or all ones constant.
7706 static inline bool isZeroOrAllOnes(SDValue N, bool AllOnes) {
7707 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N);
7710 return AllOnes ? C->isAllOnesValue() : C->isNullValue();
7713 // Return true if N is conditionally 0 or all ones.
7714 // Detects these expressions where cc is an i1 value:
7716 // (select cc 0, y) [AllOnes=0]
7717 // (select cc y, 0) [AllOnes=0]
7718 // (zext cc) [AllOnes=0]
7719 // (sext cc) [AllOnes=0/1]
7720 // (select cc -1, y) [AllOnes=1]
7721 // (select cc y, -1) [AllOnes=1]
7723 // Invert is set when N is the null/all ones constant when CC is false.
7724 // OtherOp is set to the alternative value of N.
7725 static bool isConditionalZeroOrAllOnes(SDNode *N, bool AllOnes,
7726 SDValue &CC, bool &Invert,
7728 SelectionDAG &DAG) {
7729 switch (N->getOpcode()) {
7730 default: return false;
7732 CC = N->getOperand(0);
7733 SDValue N1 = N->getOperand(1);
7734 SDValue N2 = N->getOperand(2);
7735 if (isZeroOrAllOnes(N1, AllOnes)) {
7740 if (isZeroOrAllOnes(N2, AllOnes)) {
7747 case ISD::ZERO_EXTEND:
7748 // (zext cc) can never be the all ones value.
7752 case ISD::SIGN_EXTEND: {
7754 EVT VT = N->getValueType(0);
7755 CC = N->getOperand(0);
7756 if (CC.getValueType() != MVT::i1)
7760 // When looking for an AllOnes constant, N is an sext, and the 'other'
7762 OtherOp = DAG.getConstant(0, dl, VT);
7763 else if (N->getOpcode() == ISD::ZERO_EXTEND)
7764 // When looking for a 0 constant, N can be zext or sext.
7765 OtherOp = DAG.getConstant(1, dl, VT);
7767 OtherOp = DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), dl,
7774 // Combine a constant select operand into its use:
7776 // (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
7777 // (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
7778 // (and (select cc, -1, c), x) -> (select cc, x, (and, x, c)) [AllOnes=1]
7779 // (or (select cc, 0, c), x) -> (select cc, x, (or, x, c))
7780 // (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c))
7782 // The transform is rejected if the select doesn't have a constant operand that
7783 // is null, or all ones when AllOnes is set.
7785 // Also recognize sext/zext from i1:
7787 // (add (zext cc), x) -> (select cc (add x, 1), x)
7788 // (add (sext cc), x) -> (select cc (add x, -1), x)
7790 // These transformations eventually create predicated instructions.
7792 // @param N The node to transform.
7793 // @param Slct The N operand that is a select.
7794 // @param OtherOp The other N operand (x above).
7795 // @param DCI Context.
7796 // @param AllOnes Require the select constant to be all ones instead of null.
7797 // @returns The new node, or SDValue() on failure.
7799 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
7800 TargetLowering::DAGCombinerInfo &DCI,
7801 bool AllOnes = false) {
7802 SelectionDAG &DAG = DCI.DAG;
7803 EVT VT = N->getValueType(0);
7804 SDValue NonConstantVal;
7807 if (!isConditionalZeroOrAllOnes(Slct.getNode(), AllOnes, CCOp, SwapSelectOps,
7808 NonConstantVal, DAG))
7811 // Slct is now know to be the desired identity constant when CC is true.
7812 SDValue TrueVal = OtherOp;
7813 SDValue FalseVal = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
7814 OtherOp, NonConstantVal);
7815 // Unless SwapSelectOps says CC should be false.
7817 std::swap(TrueVal, FalseVal);
7819 return DAG.getNode(ISD::SELECT, SDLoc(N), VT,
7820 CCOp, TrueVal, FalseVal);
7823 // Attempt combineSelectAndUse on each operand of a commutative operator N.
7825 SDValue combineSelectAndUseCommutative(SDNode *N, bool AllOnes,
7826 TargetLowering::DAGCombinerInfo &DCI) {
7827 SDValue N0 = N->getOperand(0);
7828 SDValue N1 = N->getOperand(1);
7829 if (N0.getNode()->hasOneUse()) {
7830 SDValue Result = combineSelectAndUse(N, N0, N1, DCI, AllOnes);
7831 if (Result.getNode())
7834 if (N1.getNode()->hasOneUse()) {
7835 SDValue Result = combineSelectAndUse(N, N1, N0, DCI, AllOnes);
7836 if (Result.getNode())
7842 // AddCombineToVPADDL- For pair-wise add on neon, use the vpaddl instruction
7843 // (only after legalization).
7844 static SDValue AddCombineToVPADDL(SDNode *N, SDValue N0, SDValue N1,
7845 TargetLowering::DAGCombinerInfo &DCI,
7846 const ARMSubtarget *Subtarget) {
7848 // Only perform optimization if after legalize, and if NEON is available. We
7849 // also expected both operands to be BUILD_VECTORs.
7850 if (DCI.isBeforeLegalize() || !Subtarget->hasNEON()
7851 || N0.getOpcode() != ISD::BUILD_VECTOR
7852 || N1.getOpcode() != ISD::BUILD_VECTOR)
7855 // Check output type since VPADDL operand elements can only be 8, 16, or 32.
7856 EVT VT = N->getValueType(0);
7857 if (!VT.isInteger() || VT.getVectorElementType() == MVT::i64)
7860 // Check that the vector operands are of the right form.
7861 // N0 and N1 are BUILD_VECTOR nodes with N number of EXTRACT_VECTOR
7862 // operands, where N is the size of the formed vector.
7863 // Each EXTRACT_VECTOR should have the same input vector and odd or even
7864 // index such that we have a pair wise add pattern.
7866 // Grab the vector that all EXTRACT_VECTOR nodes should be referencing.
7867 if (N0->getOperand(0)->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
7869 SDValue Vec = N0->getOperand(0)->getOperand(0);
7870 SDNode *V = Vec.getNode();
7871 unsigned nextIndex = 0;
7873 // For each operands to the ADD which are BUILD_VECTORs,
7874 // check to see if each of their operands are an EXTRACT_VECTOR with
7875 // the same vector and appropriate index.
7876 for (unsigned i = 0, e = N0->getNumOperands(); i != e; ++i) {
7877 if (N0->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT
7878 && N1->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
7880 SDValue ExtVec0 = N0->getOperand(i);
7881 SDValue ExtVec1 = N1->getOperand(i);
7883 // First operand is the vector, verify its the same.
7884 if (V != ExtVec0->getOperand(0).getNode() ||
7885 V != ExtVec1->getOperand(0).getNode())
7888 // Second is the constant, verify its correct.
7889 ConstantSDNode *C0 = dyn_cast<ConstantSDNode>(ExtVec0->getOperand(1));
7890 ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(ExtVec1->getOperand(1));
7892 // For the constant, we want to see all the even or all the odd.
7893 if (!C0 || !C1 || C0->getZExtValue() != nextIndex
7894 || C1->getZExtValue() != nextIndex+1)
7903 // Create VPADDL node.
7904 SelectionDAG &DAG = DCI.DAG;
7905 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7909 // Build operand list.
7910 SmallVector<SDValue, 8> Ops;
7911 Ops.push_back(DAG.getConstant(Intrinsic::arm_neon_vpaddls, dl,
7912 TLI.getPointerTy()));
7914 // Input is the vector.
7917 // Get widened type and narrowed type.
7919 unsigned numElem = VT.getVectorNumElements();
7921 EVT inputLaneType = Vec.getValueType().getVectorElementType();
7922 switch (inputLaneType.getSimpleVT().SimpleTy) {
7923 case MVT::i8: widenType = MVT::getVectorVT(MVT::i16, numElem); break;
7924 case MVT::i16: widenType = MVT::getVectorVT(MVT::i32, numElem); break;
7925 case MVT::i32: widenType = MVT::getVectorVT(MVT::i64, numElem); break;
7927 llvm_unreachable("Invalid vector element type for padd optimization.");
7930 SDValue tmp = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, widenType, Ops);
7931 unsigned ExtOp = VT.bitsGT(tmp.getValueType()) ? ISD::ANY_EXTEND : ISD::TRUNCATE;
7932 return DAG.getNode(ExtOp, dl, VT, tmp);
7935 static SDValue findMUL_LOHI(SDValue V) {
7936 if (V->getOpcode() == ISD::UMUL_LOHI ||
7937 V->getOpcode() == ISD::SMUL_LOHI)
7942 static SDValue AddCombineTo64bitMLAL(SDNode *AddcNode,
7943 TargetLowering::DAGCombinerInfo &DCI,
7944 const ARMSubtarget *Subtarget) {
7946 if (Subtarget->isThumb1Only()) return SDValue();
7948 // Only perform the checks after legalize when the pattern is available.
7949 if (DCI.isBeforeLegalize()) return SDValue();
7951 // Look for multiply add opportunities.
7952 // The pattern is a ISD::UMUL_LOHI followed by two add nodes, where
7953 // each add nodes consumes a value from ISD::UMUL_LOHI and there is
7954 // a glue link from the first add to the second add.
7955 // If we find this pattern, we can replace the U/SMUL_LOHI, ADDC, and ADDE by
7956 // a S/UMLAL instruction.
7959 // \ / \ [no multiline comment]
7965 assert(AddcNode->getOpcode() == ISD::ADDC && "Expect an ADDC");
7966 SDValue AddcOp0 = AddcNode->getOperand(0);
7967 SDValue AddcOp1 = AddcNode->getOperand(1);
7969 // Check if the two operands are from the same mul_lohi node.
7970 if (AddcOp0.getNode() == AddcOp1.getNode())
7973 assert(AddcNode->getNumValues() == 2 &&
7974 AddcNode->getValueType(0) == MVT::i32 &&
7975 "Expect ADDC with two result values. First: i32");
7977 // Check that we have a glued ADDC node.
7978 if (AddcNode->getValueType(1) != MVT::Glue)
7981 // Check that the ADDC adds the low result of the S/UMUL_LOHI.
7982 if (AddcOp0->getOpcode() != ISD::UMUL_LOHI &&
7983 AddcOp0->getOpcode() != ISD::SMUL_LOHI &&
7984 AddcOp1->getOpcode() != ISD::UMUL_LOHI &&
7985 AddcOp1->getOpcode() != ISD::SMUL_LOHI)
7988 // Look for the glued ADDE.
7989 SDNode* AddeNode = AddcNode->getGluedUser();
7993 // Make sure it is really an ADDE.
7994 if (AddeNode->getOpcode() != ISD::ADDE)
7997 assert(AddeNode->getNumOperands() == 3 &&
7998 AddeNode->getOperand(2).getValueType() == MVT::Glue &&
7999 "ADDE node has the wrong inputs");
8001 // Check for the triangle shape.
8002 SDValue AddeOp0 = AddeNode->getOperand(0);
8003 SDValue AddeOp1 = AddeNode->getOperand(1);
8005 // Make sure that the ADDE operands are not coming from the same node.
8006 if (AddeOp0.getNode() == AddeOp1.getNode())
8009 // Find the MUL_LOHI node walking up ADDE's operands.
8010 bool IsLeftOperandMUL = false;
8011 SDValue MULOp = findMUL_LOHI(AddeOp0);
8012 if (MULOp == SDValue())
8013 MULOp = findMUL_LOHI(AddeOp1);
8015 IsLeftOperandMUL = true;
8016 if (MULOp == SDValue())
8019 // Figure out the right opcode.
8020 unsigned Opc = MULOp->getOpcode();
8021 unsigned FinalOpc = (Opc == ISD::SMUL_LOHI) ? ARMISD::SMLAL : ARMISD::UMLAL;
8023 // Figure out the high and low input values to the MLAL node.
8024 SDValue* HiAdd = nullptr;
8025 SDValue* LoMul = nullptr;
8026 SDValue* LowAdd = nullptr;
8028 // Ensure that ADDE is from high result of ISD::SMUL_LOHI.
8029 if ((AddeOp0 != MULOp.getValue(1)) && (AddeOp1 != MULOp.getValue(1)))
8032 if (IsLeftOperandMUL)
8038 // Ensure that LoMul and LowAdd are taken from correct ISD::SMUL_LOHI node
8039 // whose low result is fed to the ADDC we are checking.
8041 if (AddcOp0 == MULOp.getValue(0)) {
8045 if (AddcOp1 == MULOp.getValue(0)) {
8053 // Create the merged node.
8054 SelectionDAG &DAG = DCI.DAG;
8056 // Build operand list.
8057 SmallVector<SDValue, 8> Ops;
8058 Ops.push_back(LoMul->getOperand(0));
8059 Ops.push_back(LoMul->getOperand(1));
8060 Ops.push_back(*LowAdd);
8061 Ops.push_back(*HiAdd);
8063 SDValue MLALNode = DAG.getNode(FinalOpc, SDLoc(AddcNode),
8064 DAG.getVTList(MVT::i32, MVT::i32), Ops);
8066 // Replace the ADDs' nodes uses by the MLA node's values.
8067 SDValue HiMLALResult(MLALNode.getNode(), 1);
8068 DAG.ReplaceAllUsesOfValueWith(SDValue(AddeNode, 0), HiMLALResult);
8070 SDValue LoMLALResult(MLALNode.getNode(), 0);
8071 DAG.ReplaceAllUsesOfValueWith(SDValue(AddcNode, 0), LoMLALResult);
8073 // Return original node to notify the driver to stop replacing.
8074 SDValue resNode(AddcNode, 0);
8078 /// PerformADDCCombine - Target-specific dag combine transform from
8079 /// ISD::ADDC, ISD::ADDE, and ISD::MUL_LOHI to MLAL.
8080 static SDValue PerformADDCCombine(SDNode *N,
8081 TargetLowering::DAGCombinerInfo &DCI,
8082 const ARMSubtarget *Subtarget) {
8084 return AddCombineTo64bitMLAL(N, DCI, Subtarget);
8088 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
8089 /// operands N0 and N1. This is a helper for PerformADDCombine that is
8090 /// called with the default operands, and if that fails, with commuted
8092 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
8093 TargetLowering::DAGCombinerInfo &DCI,
8094 const ARMSubtarget *Subtarget){
8096 // Attempt to create vpaddl for this add.
8097 SDValue Result = AddCombineToVPADDL(N, N0, N1, DCI, Subtarget);
8098 if (Result.getNode())
8101 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
8102 if (N0.getNode()->hasOneUse()) {
8103 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
8104 if (Result.getNode()) return Result;
8109 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
8111 static SDValue PerformADDCombine(SDNode *N,
8112 TargetLowering::DAGCombinerInfo &DCI,
8113 const ARMSubtarget *Subtarget) {
8114 SDValue N0 = N->getOperand(0);
8115 SDValue N1 = N->getOperand(1);
8117 // First try with the default operand order.
8118 SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI, Subtarget);
8119 if (Result.getNode())
8122 // If that didn't work, try again with the operands commuted.
8123 return PerformADDCombineWithOperands(N, N1, N0, DCI, Subtarget);
8126 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
8128 static SDValue PerformSUBCombine(SDNode *N,
8129 TargetLowering::DAGCombinerInfo &DCI) {
8130 SDValue N0 = N->getOperand(0);
8131 SDValue N1 = N->getOperand(1);
8133 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
8134 if (N1.getNode()->hasOneUse()) {
8135 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
8136 if (Result.getNode()) return Result;
8142 /// PerformVMULCombine
8143 /// Distribute (A + B) * C to (A * C) + (B * C) to take advantage of the
8144 /// special multiplier accumulator forwarding.
8150 // However, for (A + B) * (A + B),
8157 static SDValue PerformVMULCombine(SDNode *N,
8158 TargetLowering::DAGCombinerInfo &DCI,
8159 const ARMSubtarget *Subtarget) {
8160 if (!Subtarget->hasVMLxForwarding())
8163 SelectionDAG &DAG = DCI.DAG;
8164 SDValue N0 = N->getOperand(0);
8165 SDValue N1 = N->getOperand(1);
8166 unsigned Opcode = N0.getOpcode();
8167 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
8168 Opcode != ISD::FADD && Opcode != ISD::FSUB) {
8169 Opcode = N1.getOpcode();
8170 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
8171 Opcode != ISD::FADD && Opcode != ISD::FSUB)
8179 EVT VT = N->getValueType(0);
8181 SDValue N00 = N0->getOperand(0);
8182 SDValue N01 = N0->getOperand(1);
8183 return DAG.getNode(Opcode, DL, VT,
8184 DAG.getNode(ISD::MUL, DL, VT, N00, N1),
8185 DAG.getNode(ISD::MUL, DL, VT, N01, N1));
8188 static SDValue PerformMULCombine(SDNode *N,
8189 TargetLowering::DAGCombinerInfo &DCI,
8190 const ARMSubtarget *Subtarget) {
8191 SelectionDAG &DAG = DCI.DAG;
8193 if (Subtarget->isThumb1Only())
8196 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
8199 EVT VT = N->getValueType(0);
8200 if (VT.is64BitVector() || VT.is128BitVector())
8201 return PerformVMULCombine(N, DCI, Subtarget);
8205 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
8209 int64_t MulAmt = C->getSExtValue();
8210 unsigned ShiftAmt = countTrailingZeros<uint64_t>(MulAmt);
8212 ShiftAmt = ShiftAmt & (32 - 1);
8213 SDValue V = N->getOperand(0);
8217 MulAmt >>= ShiftAmt;
8220 if (isPowerOf2_32(MulAmt - 1)) {
8221 // (mul x, 2^N + 1) => (add (shl x, N), x)
8222 Res = DAG.getNode(ISD::ADD, DL, VT,
8224 DAG.getNode(ISD::SHL, DL, VT,
8226 DAG.getConstant(Log2_32(MulAmt - 1), DL,
8228 } else if (isPowerOf2_32(MulAmt + 1)) {
8229 // (mul x, 2^N - 1) => (sub (shl x, N), x)
8230 Res = DAG.getNode(ISD::SUB, DL, VT,
8231 DAG.getNode(ISD::SHL, DL, VT,
8233 DAG.getConstant(Log2_32(MulAmt + 1), DL,
8239 uint64_t MulAmtAbs = -MulAmt;
8240 if (isPowerOf2_32(MulAmtAbs + 1)) {
8241 // (mul x, -(2^N - 1)) => (sub x, (shl x, N))
8242 Res = DAG.getNode(ISD::SUB, DL, VT,
8244 DAG.getNode(ISD::SHL, DL, VT,
8246 DAG.getConstant(Log2_32(MulAmtAbs + 1), DL,
8248 } else if (isPowerOf2_32(MulAmtAbs - 1)) {
8249 // (mul x, -(2^N + 1)) => - (add (shl x, N), x)
8250 Res = DAG.getNode(ISD::ADD, DL, VT,
8252 DAG.getNode(ISD::SHL, DL, VT,
8254 DAG.getConstant(Log2_32(MulAmtAbs - 1), DL,
8256 Res = DAG.getNode(ISD::SUB, DL, VT,
8257 DAG.getConstant(0, DL, MVT::i32), Res);
8264 Res = DAG.getNode(ISD::SHL, DL, VT,
8265 Res, DAG.getConstant(ShiftAmt, DL, MVT::i32));
8267 // Do not add new nodes to DAG combiner worklist.
8268 DCI.CombineTo(N, Res, false);
8272 static SDValue PerformANDCombine(SDNode *N,
8273 TargetLowering::DAGCombinerInfo &DCI,
8274 const ARMSubtarget *Subtarget) {
8276 // Attempt to use immediate-form VBIC
8277 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
8279 EVT VT = N->getValueType(0);
8280 SelectionDAG &DAG = DCI.DAG;
8282 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8285 APInt SplatBits, SplatUndef;
8286 unsigned SplatBitSize;
8289 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
8290 if (SplatBitSize <= 64) {
8292 SDValue Val = isNEONModifiedImm((~SplatBits).getZExtValue(),
8293 SplatUndef.getZExtValue(), SplatBitSize,
8294 DAG, dl, VbicVT, VT.is128BitVector(),
8296 if (Val.getNode()) {
8298 DAG.getNode(ISD::BITCAST, dl, VbicVT, N->getOperand(0));
8299 SDValue Vbic = DAG.getNode(ARMISD::VBICIMM, dl, VbicVT, Input, Val);
8300 return DAG.getNode(ISD::BITCAST, dl, VT, Vbic);
8305 if (!Subtarget->isThumb1Only()) {
8306 // fold (and (select cc, -1, c), x) -> (select cc, x, (and, x, c))
8307 SDValue Result = combineSelectAndUseCommutative(N, true, DCI);
8308 if (Result.getNode())
8315 /// PerformORCombine - Target-specific dag combine xforms for ISD::OR
8316 static SDValue PerformORCombine(SDNode *N,
8317 TargetLowering::DAGCombinerInfo &DCI,
8318 const ARMSubtarget *Subtarget) {
8319 // Attempt to use immediate-form VORR
8320 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
8322 EVT VT = N->getValueType(0);
8323 SelectionDAG &DAG = DCI.DAG;
8325 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8328 APInt SplatBits, SplatUndef;
8329 unsigned SplatBitSize;
8331 if (BVN && Subtarget->hasNEON() &&
8332 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
8333 if (SplatBitSize <= 64) {
8335 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
8336 SplatUndef.getZExtValue(), SplatBitSize,
8337 DAG, dl, VorrVT, VT.is128BitVector(),
8339 if (Val.getNode()) {
8341 DAG.getNode(ISD::BITCAST, dl, VorrVT, N->getOperand(0));
8342 SDValue Vorr = DAG.getNode(ARMISD::VORRIMM, dl, VorrVT, Input, Val);
8343 return DAG.getNode(ISD::BITCAST, dl, VT, Vorr);
8348 if (!Subtarget->isThumb1Only()) {
8349 // fold (or (select cc, 0, c), x) -> (select cc, x, (or, x, c))
8350 SDValue Result = combineSelectAndUseCommutative(N, false, DCI);
8351 if (Result.getNode())
8355 // The code below optimizes (or (and X, Y), Z).
8356 // The AND operand needs to have a single user to make these optimizations
8358 SDValue N0 = N->getOperand(0);
8359 if (N0.getOpcode() != ISD::AND || !N0.hasOneUse())
8361 SDValue N1 = N->getOperand(1);
8363 // (or (and B, A), (and C, ~A)) => (VBSL A, B, C) when A is a constant.
8364 if (Subtarget->hasNEON() && N1.getOpcode() == ISD::AND && VT.isVector() &&
8365 DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
8367 unsigned SplatBitSize;
8370 APInt SplatBits0, SplatBits1;
8371 BuildVectorSDNode *BVN0 = dyn_cast<BuildVectorSDNode>(N0->getOperand(1));
8372 BuildVectorSDNode *BVN1 = dyn_cast<BuildVectorSDNode>(N1->getOperand(1));
8373 // Ensure that the second operand of both ands are constants
8374 if (BVN0 && BVN0->isConstantSplat(SplatBits0, SplatUndef, SplatBitSize,
8375 HasAnyUndefs) && !HasAnyUndefs) {
8376 if (BVN1 && BVN1->isConstantSplat(SplatBits1, SplatUndef, SplatBitSize,
8377 HasAnyUndefs) && !HasAnyUndefs) {
8378 // Ensure that the bit width of the constants are the same and that
8379 // the splat arguments are logical inverses as per the pattern we
8380 // are trying to simplify.
8381 if (SplatBits0.getBitWidth() == SplatBits1.getBitWidth() &&
8382 SplatBits0 == ~SplatBits1) {
8383 // Canonicalize the vector type to make instruction selection
8385 EVT CanonicalVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
8386 SDValue Result = DAG.getNode(ARMISD::VBSL, dl, CanonicalVT,
8390 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
8396 // Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when
8399 // BFI is only available on V6T2+
8400 if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops())
8404 // 1) or (and A, mask), val => ARMbfi A, val, mask
8405 // iff (val & mask) == val
8407 // 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
8408 // 2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2)
8409 // && mask == ~mask2
8410 // 2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2)
8411 // && ~mask == mask2
8412 // (i.e., copy a bitfield value into another bitfield of the same width)
8417 SDValue N00 = N0.getOperand(0);
8419 // The value and the mask need to be constants so we can verify this is
8420 // actually a bitfield set. If the mask is 0xffff, we can do better
8421 // via a movt instruction, so don't use BFI in that case.
8422 SDValue MaskOp = N0.getOperand(1);
8423 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(MaskOp);
8426 unsigned Mask = MaskC->getZExtValue();
8430 // Case (1): or (and A, mask), val => ARMbfi A, val, mask
8431 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
8433 unsigned Val = N1C->getZExtValue();
8434 if ((Val & ~Mask) != Val)
8437 if (ARM::isBitFieldInvertedMask(Mask)) {
8438 Val >>= countTrailingZeros(~Mask);
8440 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00,
8441 DAG.getConstant(Val, DL, MVT::i32),
8442 DAG.getConstant(Mask, DL, MVT::i32));
8444 // Do not add new nodes to DAG combiner worklist.
8445 DCI.CombineTo(N, Res, false);
8448 } else if (N1.getOpcode() == ISD::AND) {
8449 // case (2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
8450 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
8453 unsigned Mask2 = N11C->getZExtValue();
8455 // Mask and ~Mask2 (or reverse) must be equivalent for the BFI pattern
8457 if (ARM::isBitFieldInvertedMask(Mask) &&
8459 // The pack halfword instruction works better for masks that fit it,
8460 // so use that when it's available.
8461 if (Subtarget->hasT2ExtractPack() &&
8462 (Mask == 0xffff || Mask == 0xffff0000))
8465 unsigned amt = countTrailingZeros(Mask2);
8466 Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0),
8467 DAG.getConstant(amt, DL, MVT::i32));
8468 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00, Res,
8469 DAG.getConstant(Mask, DL, MVT::i32));
8470 // Do not add new nodes to DAG combiner worklist.
8471 DCI.CombineTo(N, Res, false);
8473 } else if (ARM::isBitFieldInvertedMask(~Mask) &&
8475 // The pack halfword instruction works better for masks that fit it,
8476 // so use that when it's available.
8477 if (Subtarget->hasT2ExtractPack() &&
8478 (Mask2 == 0xffff || Mask2 == 0xffff0000))
8481 unsigned lsb = countTrailingZeros(Mask);
8482 Res = DAG.getNode(ISD::SRL, DL, VT, N00,
8483 DAG.getConstant(lsb, DL, MVT::i32));
8484 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res,
8485 DAG.getConstant(Mask2, DL, MVT::i32));
8486 // Do not add new nodes to DAG combiner worklist.
8487 DCI.CombineTo(N, Res, false);
8492 if (DAG.MaskedValueIsZero(N1, MaskC->getAPIntValue()) &&
8493 N00.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N00.getOperand(1)) &&
8494 ARM::isBitFieldInvertedMask(~Mask)) {
8495 // Case (3): or (and (shl A, #shamt), mask), B => ARMbfi B, A, ~mask
8496 // where lsb(mask) == #shamt and masked bits of B are known zero.
8497 SDValue ShAmt = N00.getOperand(1);
8498 unsigned ShAmtC = cast<ConstantSDNode>(ShAmt)->getZExtValue();
8499 unsigned LSB = countTrailingZeros(Mask);
8503 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1, N00.getOperand(0),
8504 DAG.getConstant(~Mask, DL, MVT::i32));
8506 // Do not add new nodes to DAG combiner worklist.
8507 DCI.CombineTo(N, Res, false);
8513 static SDValue PerformXORCombine(SDNode *N,
8514 TargetLowering::DAGCombinerInfo &DCI,
8515 const ARMSubtarget *Subtarget) {
8516 EVT VT = N->getValueType(0);
8517 SelectionDAG &DAG = DCI.DAG;
8519 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8522 if (!Subtarget->isThumb1Only()) {
8523 // fold (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c))
8524 SDValue Result = combineSelectAndUseCommutative(N, false, DCI);
8525 if (Result.getNode())
8532 /// PerformBFICombine - (bfi A, (and B, Mask1), Mask2) -> (bfi A, B, Mask2) iff
8533 /// the bits being cleared by the AND are not demanded by the BFI.
8534 static SDValue PerformBFICombine(SDNode *N,
8535 TargetLowering::DAGCombinerInfo &DCI) {
8536 SDValue N1 = N->getOperand(1);
8537 if (N1.getOpcode() == ISD::AND) {
8538 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
8541 unsigned InvMask = cast<ConstantSDNode>(N->getOperand(2))->getZExtValue();
8542 unsigned LSB = countTrailingZeros(~InvMask);
8543 unsigned Width = (32 - countLeadingZeros(~InvMask)) - LSB;
8545 static_cast<unsigned>(std::numeric_limits<unsigned>::digits) &&
8546 "undefined behavior");
8547 unsigned Mask = (1u << Width) - 1;
8548 unsigned Mask2 = N11C->getZExtValue();
8549 if ((Mask & (~Mask2)) == 0)
8550 return DCI.DAG.getNode(ARMISD::BFI, SDLoc(N), N->getValueType(0),
8551 N->getOperand(0), N1.getOperand(0),
8557 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
8558 /// ARMISD::VMOVRRD.
8559 static SDValue PerformVMOVRRDCombine(SDNode *N,
8560 TargetLowering::DAGCombinerInfo &DCI,
8561 const ARMSubtarget *Subtarget) {
8562 // vmovrrd(vmovdrr x, y) -> x,y
8563 SDValue InDouble = N->getOperand(0);
8564 if (InDouble.getOpcode() == ARMISD::VMOVDRR && !Subtarget->isFPOnlySP())
8565 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
8567 // vmovrrd(load f64) -> (load i32), (load i32)
8568 SDNode *InNode = InDouble.getNode();
8569 if (ISD::isNormalLoad(InNode) && InNode->hasOneUse() &&
8570 InNode->getValueType(0) == MVT::f64 &&
8571 InNode->getOperand(1).getOpcode() == ISD::FrameIndex &&
8572 !cast<LoadSDNode>(InNode)->isVolatile()) {
8573 // TODO: Should this be done for non-FrameIndex operands?
8574 LoadSDNode *LD = cast<LoadSDNode>(InNode);
8576 SelectionDAG &DAG = DCI.DAG;
8578 SDValue BasePtr = LD->getBasePtr();
8579 SDValue NewLD1 = DAG.getLoad(MVT::i32, DL, LD->getChain(), BasePtr,
8580 LD->getPointerInfo(), LD->isVolatile(),
8581 LD->isNonTemporal(), LD->isInvariant(),
8582 LD->getAlignment());
8584 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
8585 DAG.getConstant(4, DL, MVT::i32));
8586 SDValue NewLD2 = DAG.getLoad(MVT::i32, DL, NewLD1.getValue(1), OffsetPtr,
8587 LD->getPointerInfo(), LD->isVolatile(),
8588 LD->isNonTemporal(), LD->isInvariant(),
8589 std::min(4U, LD->getAlignment() / 2));
8591 DAG.ReplaceAllUsesOfValueWith(SDValue(LD, 1), NewLD2.getValue(1));
8592 if (DCI.DAG.getTargetLoweringInfo().isBigEndian())
8593 std::swap (NewLD1, NewLD2);
8594 SDValue Result = DCI.CombineTo(N, NewLD1, NewLD2);
8601 /// PerformVMOVDRRCombine - Target-specific dag combine xforms for
8602 /// ARMISD::VMOVDRR. This is also used for BUILD_VECTORs with 2 operands.
8603 static SDValue PerformVMOVDRRCombine(SDNode *N, SelectionDAG &DAG) {
8604 // N=vmovrrd(X); vmovdrr(N:0, N:1) -> bit_convert(X)
8605 SDValue Op0 = N->getOperand(0);
8606 SDValue Op1 = N->getOperand(1);
8607 if (Op0.getOpcode() == ISD::BITCAST)
8608 Op0 = Op0.getOperand(0);
8609 if (Op1.getOpcode() == ISD::BITCAST)
8610 Op1 = Op1.getOperand(0);
8611 if (Op0.getOpcode() == ARMISD::VMOVRRD &&
8612 Op0.getNode() == Op1.getNode() &&
8613 Op0.getResNo() == 0 && Op1.getResNo() == 1)
8614 return DAG.getNode(ISD::BITCAST, SDLoc(N),
8615 N->getValueType(0), Op0.getOperand(0));
8619 /// hasNormalLoadOperand - Check if any of the operands of a BUILD_VECTOR node
8620 /// are normal, non-volatile loads. If so, it is profitable to bitcast an
8621 /// i64 vector to have f64 elements, since the value can then be loaded
8622 /// directly into a VFP register.
8623 static bool hasNormalLoadOperand(SDNode *N) {
8624 unsigned NumElts = N->getValueType(0).getVectorNumElements();
8625 for (unsigned i = 0; i < NumElts; ++i) {
8626 SDNode *Elt = N->getOperand(i).getNode();
8627 if (ISD::isNormalLoad(Elt) && !cast<LoadSDNode>(Elt)->isVolatile())
8633 /// PerformBUILD_VECTORCombine - Target-specific dag combine xforms for
8634 /// ISD::BUILD_VECTOR.
8635 static SDValue PerformBUILD_VECTORCombine(SDNode *N,
8636 TargetLowering::DAGCombinerInfo &DCI,
8637 const ARMSubtarget *Subtarget) {
8638 // build_vector(N=ARMISD::VMOVRRD(X), N:1) -> bit_convert(X):
8639 // VMOVRRD is introduced when legalizing i64 types. It forces the i64 value
8640 // into a pair of GPRs, which is fine when the value is used as a scalar,
8641 // but if the i64 value is converted to a vector, we need to undo the VMOVRRD.
8642 SelectionDAG &DAG = DCI.DAG;
8643 if (N->getNumOperands() == 2) {
8644 SDValue RV = PerformVMOVDRRCombine(N, DAG);
8649 // Load i64 elements as f64 values so that type legalization does not split
8650 // them up into i32 values.
8651 EVT VT = N->getValueType(0);
8652 if (VT.getVectorElementType() != MVT::i64 || !hasNormalLoadOperand(N))
8655 SmallVector<SDValue, 8> Ops;
8656 unsigned NumElts = VT.getVectorNumElements();
8657 for (unsigned i = 0; i < NumElts; ++i) {
8658 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(i));
8660 // Make the DAGCombiner fold the bitcast.
8661 DCI.AddToWorklist(V.getNode());
8663 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, NumElts);
8664 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, FloatVT, Ops);
8665 return DAG.getNode(ISD::BITCAST, dl, VT, BV);
8668 /// \brief Target-specific dag combine xforms for ARMISD::BUILD_VECTOR.
8670 PerformARMBUILD_VECTORCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
8671 // ARMISD::BUILD_VECTOR is introduced when legalizing ISD::BUILD_VECTOR.
8672 // At that time, we may have inserted bitcasts from integer to float.
8673 // If these bitcasts have survived DAGCombine, change the lowering of this
8674 // BUILD_VECTOR in something more vector friendly, i.e., that does not
8675 // force to use floating point types.
8677 // Make sure we can change the type of the vector.
8678 // This is possible iff:
8679 // 1. The vector is only used in a bitcast to a integer type. I.e.,
8680 // 1.1. Vector is used only once.
8681 // 1.2. Use is a bit convert to an integer type.
8682 // 2. The size of its operands are 32-bits (64-bits are not legal).
8683 EVT VT = N->getValueType(0);
8684 EVT EltVT = VT.getVectorElementType();
8686 // Check 1.1. and 2.
8687 if (EltVT.getSizeInBits() != 32 || !N->hasOneUse())
8690 // By construction, the input type must be float.
8691 assert(EltVT == MVT::f32 && "Unexpected type!");
8694 SDNode *Use = *N->use_begin();
8695 if (Use->getOpcode() != ISD::BITCAST ||
8696 Use->getValueType(0).isFloatingPoint())
8699 // Check profitability.
8700 // Model is, if more than half of the relevant operands are bitcast from
8701 // i32, turn the build_vector into a sequence of insert_vector_elt.
8702 // Relevant operands are everything that is not statically
8703 // (i.e., at compile time) bitcasted.
8704 unsigned NumOfBitCastedElts = 0;
8705 unsigned NumElts = VT.getVectorNumElements();
8706 unsigned NumOfRelevantElts = NumElts;
8707 for (unsigned Idx = 0; Idx < NumElts; ++Idx) {
8708 SDValue Elt = N->getOperand(Idx);
8709 if (Elt->getOpcode() == ISD::BITCAST) {
8710 // Assume only bit cast to i32 will go away.
8711 if (Elt->getOperand(0).getValueType() == MVT::i32)
8712 ++NumOfBitCastedElts;
8713 } else if (Elt.getOpcode() == ISD::UNDEF || isa<ConstantSDNode>(Elt))
8714 // Constants are statically casted, thus do not count them as
8715 // relevant operands.
8716 --NumOfRelevantElts;
8719 // Check if more than half of the elements require a non-free bitcast.
8720 if (NumOfBitCastedElts <= NumOfRelevantElts / 2)
8723 SelectionDAG &DAG = DCI.DAG;
8724 // Create the new vector type.
8725 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
8726 // Check if the type is legal.
8727 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8728 if (!TLI.isTypeLegal(VecVT))
8732 // ARMISD::BUILD_VECTOR E1, E2, ..., EN.
8733 // => BITCAST INSERT_VECTOR_ELT
8734 // (INSERT_VECTOR_ELT (...), (BITCAST EN-1), N-1),
8736 SDValue Vec = DAG.getUNDEF(VecVT);
8738 for (unsigned Idx = 0 ; Idx < NumElts; ++Idx) {
8739 SDValue V = N->getOperand(Idx);
8740 if (V.getOpcode() == ISD::UNDEF)
8742 if (V.getOpcode() == ISD::BITCAST &&
8743 V->getOperand(0).getValueType() == MVT::i32)
8744 // Fold obvious case.
8745 V = V.getOperand(0);
8747 V = DAG.getNode(ISD::BITCAST, SDLoc(V), MVT::i32, V);
8748 // Make the DAGCombiner fold the bitcasts.
8749 DCI.AddToWorklist(V.getNode());
8751 SDValue LaneIdx = DAG.getConstant(Idx, dl, MVT::i32);
8752 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VecVT, Vec, V, LaneIdx);
8754 Vec = DAG.getNode(ISD::BITCAST, dl, VT, Vec);
8755 // Make the DAGCombiner fold the bitcasts.
8756 DCI.AddToWorklist(Vec.getNode());
8760 /// PerformInsertEltCombine - Target-specific dag combine xforms for
8761 /// ISD::INSERT_VECTOR_ELT.
8762 static SDValue PerformInsertEltCombine(SDNode *N,
8763 TargetLowering::DAGCombinerInfo &DCI) {
8764 // Bitcast an i64 load inserted into a vector to f64.
8765 // Otherwise, the i64 value will be legalized to a pair of i32 values.
8766 EVT VT = N->getValueType(0);
8767 SDNode *Elt = N->getOperand(1).getNode();
8768 if (VT.getVectorElementType() != MVT::i64 ||
8769 !ISD::isNormalLoad(Elt) || cast<LoadSDNode>(Elt)->isVolatile())
8772 SelectionDAG &DAG = DCI.DAG;
8774 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
8775 VT.getVectorNumElements());
8776 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, N->getOperand(0));
8777 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(1));
8778 // Make the DAGCombiner fold the bitcasts.
8779 DCI.AddToWorklist(Vec.getNode());
8780 DCI.AddToWorklist(V.getNode());
8781 SDValue InsElt = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, FloatVT,
8782 Vec, V, N->getOperand(2));
8783 return DAG.getNode(ISD::BITCAST, dl, VT, InsElt);
8786 /// PerformVECTOR_SHUFFLECombine - Target-specific dag combine xforms for
8787 /// ISD::VECTOR_SHUFFLE.
8788 static SDValue PerformVECTOR_SHUFFLECombine(SDNode *N, SelectionDAG &DAG) {
8789 // The LLVM shufflevector instruction does not require the shuffle mask
8790 // length to match the operand vector length, but ISD::VECTOR_SHUFFLE does
8791 // have that requirement. When translating to ISD::VECTOR_SHUFFLE, if the
8792 // operands do not match the mask length, they are extended by concatenating
8793 // them with undef vectors. That is probably the right thing for other
8794 // targets, but for NEON it is better to concatenate two double-register
8795 // size vector operands into a single quad-register size vector. Do that
8796 // transformation here:
8797 // shuffle(concat(v1, undef), concat(v2, undef)) ->
8798 // shuffle(concat(v1, v2), undef)
8799 SDValue Op0 = N->getOperand(0);
8800 SDValue Op1 = N->getOperand(1);
8801 if (Op0.getOpcode() != ISD::CONCAT_VECTORS ||
8802 Op1.getOpcode() != ISD::CONCAT_VECTORS ||
8803 Op0.getNumOperands() != 2 ||
8804 Op1.getNumOperands() != 2)
8806 SDValue Concat0Op1 = Op0.getOperand(1);
8807 SDValue Concat1Op1 = Op1.getOperand(1);
8808 if (Concat0Op1.getOpcode() != ISD::UNDEF ||
8809 Concat1Op1.getOpcode() != ISD::UNDEF)
8811 // Skip the transformation if any of the types are illegal.
8812 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8813 EVT VT = N->getValueType(0);
8814 if (!TLI.isTypeLegal(VT) ||
8815 !TLI.isTypeLegal(Concat0Op1.getValueType()) ||
8816 !TLI.isTypeLegal(Concat1Op1.getValueType()))
8819 SDValue NewConcat = DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT,
8820 Op0.getOperand(0), Op1.getOperand(0));
8821 // Translate the shuffle mask.
8822 SmallVector<int, 16> NewMask;
8823 unsigned NumElts = VT.getVectorNumElements();
8824 unsigned HalfElts = NumElts/2;
8825 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
8826 for (unsigned n = 0; n < NumElts; ++n) {
8827 int MaskElt = SVN->getMaskElt(n);
8829 if (MaskElt < (int)HalfElts)
8831 else if (MaskElt >= (int)NumElts && MaskElt < (int)(NumElts + HalfElts))
8832 NewElt = HalfElts + MaskElt - NumElts;
8833 NewMask.push_back(NewElt);
8835 return DAG.getVectorShuffle(VT, SDLoc(N), NewConcat,
8836 DAG.getUNDEF(VT), NewMask.data());
8839 /// CombineBaseUpdate - Target-specific DAG combine function for VLDDUP,
8840 /// NEON load/store intrinsics, and generic vector load/stores, to merge
8841 /// base address updates.
8842 /// For generic load/stores, the memory type is assumed to be a vector.
8843 /// The caller is assumed to have checked legality.
8844 static SDValue CombineBaseUpdate(SDNode *N,
8845 TargetLowering::DAGCombinerInfo &DCI) {
8846 SelectionDAG &DAG = DCI.DAG;
8847 const bool isIntrinsic = (N->getOpcode() == ISD::INTRINSIC_VOID ||
8848 N->getOpcode() == ISD::INTRINSIC_W_CHAIN);
8849 const bool isStore = N->getOpcode() == ISD::STORE;
8850 const unsigned AddrOpIdx = ((isIntrinsic || isStore) ? 2 : 1);
8851 SDValue Addr = N->getOperand(AddrOpIdx);
8852 MemSDNode *MemN = cast<MemSDNode>(N);
8855 // Search for a use of the address operand that is an increment.
8856 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(),
8857 UE = Addr.getNode()->use_end(); UI != UE; ++UI) {
8859 if (User->getOpcode() != ISD::ADD ||
8860 UI.getUse().getResNo() != Addr.getResNo())
8863 // Check that the add is independent of the load/store. Otherwise, folding
8864 // it would create a cycle.
8865 if (User->isPredecessorOf(N) || N->isPredecessorOf(User))
8868 // Find the new opcode for the updating load/store.
8869 bool isLoadOp = true;
8870 bool isLaneOp = false;
8871 unsigned NewOpc = 0;
8872 unsigned NumVecs = 0;
8874 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
8876 default: llvm_unreachable("unexpected intrinsic for Neon base update");
8877 case Intrinsic::arm_neon_vld1: NewOpc = ARMISD::VLD1_UPD;
8879 case Intrinsic::arm_neon_vld2: NewOpc = ARMISD::VLD2_UPD;
8881 case Intrinsic::arm_neon_vld3: NewOpc = ARMISD::VLD3_UPD;
8883 case Intrinsic::arm_neon_vld4: NewOpc = ARMISD::VLD4_UPD;
8885 case Intrinsic::arm_neon_vld2lane: NewOpc = ARMISD::VLD2LN_UPD;
8886 NumVecs = 2; isLaneOp = true; break;
8887 case Intrinsic::arm_neon_vld3lane: NewOpc = ARMISD::VLD3LN_UPD;
8888 NumVecs = 3; isLaneOp = true; break;
8889 case Intrinsic::arm_neon_vld4lane: NewOpc = ARMISD::VLD4LN_UPD;
8890 NumVecs = 4; isLaneOp = true; break;
8891 case Intrinsic::arm_neon_vst1: NewOpc = ARMISD::VST1_UPD;
8892 NumVecs = 1; isLoadOp = false; break;
8893 case Intrinsic::arm_neon_vst2: NewOpc = ARMISD::VST2_UPD;
8894 NumVecs = 2; isLoadOp = false; break;
8895 case Intrinsic::arm_neon_vst3: NewOpc = ARMISD::VST3_UPD;
8896 NumVecs = 3; isLoadOp = false; break;
8897 case Intrinsic::arm_neon_vst4: NewOpc = ARMISD::VST4_UPD;
8898 NumVecs = 4; isLoadOp = false; break;
8899 case Intrinsic::arm_neon_vst2lane: NewOpc = ARMISD::VST2LN_UPD;
8900 NumVecs = 2; isLoadOp = false; isLaneOp = true; break;
8901 case Intrinsic::arm_neon_vst3lane: NewOpc = ARMISD::VST3LN_UPD;
8902 NumVecs = 3; isLoadOp = false; isLaneOp = true; break;
8903 case Intrinsic::arm_neon_vst4lane: NewOpc = ARMISD::VST4LN_UPD;
8904 NumVecs = 4; isLoadOp = false; isLaneOp = true; break;
8908 switch (N->getOpcode()) {
8909 default: llvm_unreachable("unexpected opcode for Neon base update");
8910 case ARMISD::VLD2DUP: NewOpc = ARMISD::VLD2DUP_UPD; NumVecs = 2; break;
8911 case ARMISD::VLD3DUP: NewOpc = ARMISD::VLD3DUP_UPD; NumVecs = 3; break;
8912 case ARMISD::VLD4DUP: NewOpc = ARMISD::VLD4DUP_UPD; NumVecs = 4; break;
8913 case ISD::LOAD: NewOpc = ARMISD::VLD1_UPD;
8914 NumVecs = 1; isLaneOp = false; break;
8915 case ISD::STORE: NewOpc = ARMISD::VST1_UPD;
8916 NumVecs = 1; isLaneOp = false; isLoadOp = false; break;
8920 // Find the size of memory referenced by the load/store.
8923 VecTy = N->getValueType(0);
8924 } else if (isIntrinsic) {
8925 VecTy = N->getOperand(AddrOpIdx+1).getValueType();
8927 assert(isStore && "Node has to be a load, a store, or an intrinsic!");
8928 VecTy = N->getOperand(1).getValueType();
8931 unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8;
8933 NumBytes /= VecTy.getVectorNumElements();
8935 // If the increment is a constant, it must match the memory ref size.
8936 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
8937 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
8938 uint64_t IncVal = CInc->getZExtValue();
8939 if (IncVal != NumBytes)
8941 } else if (NumBytes >= 3 * 16) {
8942 // VLD3/4 and VST3/4 for 128-bit vectors are implemented with two
8943 // separate instructions that make it harder to use a non-constant update.
8947 // OK, we found an ADD we can fold into the base update.
8948 // Now, create a _UPD node, taking care of not breaking alignment.
8950 EVT AlignedVecTy = VecTy;
8951 unsigned Alignment = MemN->getAlignment();
8953 // If this is a less-than-standard-aligned load/store, change the type to
8954 // match the standard alignment.
8955 // The alignment is overlooked when selecting _UPD variants; and it's
8956 // easier to introduce bitcasts here than fix that.
8957 // There are 3 ways to get to this base-update combine:
8958 // - intrinsics: they are assumed to be properly aligned (to the standard
8959 // alignment of the memory type), so we don't need to do anything.
8960 // - ARMISD::VLDx nodes: they are only generated from the aforementioned
8961 // intrinsics, so, likewise, there's nothing to do.
8962 // - generic load/store instructions: the alignment is specified as an
8963 // explicit operand, rather than implicitly as the standard alignment
8964 // of the memory type (like the intrisics). We need to change the
8965 // memory type to match the explicit alignment. That way, we don't
8966 // generate non-standard-aligned ARMISD::VLDx nodes.
8967 if (isa<LSBaseSDNode>(N)) {
8970 if (Alignment < VecTy.getScalarSizeInBits() / 8) {
8971 MVT EltTy = MVT::getIntegerVT(Alignment * 8);
8972 assert(NumVecs == 1 && "Unexpected multi-element generic load/store.");
8973 assert(!isLaneOp && "Unexpected generic load/store lane.");
8974 unsigned NumElts = NumBytes / (EltTy.getSizeInBits() / 8);
8975 AlignedVecTy = MVT::getVectorVT(EltTy, NumElts);
8977 // Don't set an explicit alignment on regular load/stores that we want
8978 // to transform to VLD/VST 1_UPD nodes.
8979 // This matches the behavior of regular load/stores, which only get an
8980 // explicit alignment if the MMO alignment is larger than the standard
8981 // alignment of the memory type.
8982 // Intrinsics, however, always get an explicit alignment, set to the
8983 // alignment of the MMO.
8987 // Create the new updating load/store node.
8988 // First, create an SDVTList for the new updating node's results.
8990 unsigned NumResultVecs = (isLoadOp ? NumVecs : 0);
8992 for (n = 0; n < NumResultVecs; ++n)
8993 Tys[n] = AlignedVecTy;
8994 Tys[n++] = MVT::i32;
8995 Tys[n] = MVT::Other;
8996 SDVTList SDTys = DAG.getVTList(makeArrayRef(Tys, NumResultVecs+2));
8998 // Then, gather the new node's operands.
8999 SmallVector<SDValue, 8> Ops;
9000 Ops.push_back(N->getOperand(0)); // incoming chain
9001 Ops.push_back(N->getOperand(AddrOpIdx));
9004 if (StoreSDNode *StN = dyn_cast<StoreSDNode>(N)) {
9005 // Try to match the intrinsic's signature
9006 Ops.push_back(StN->getValue());
9008 // Loads (and of course intrinsics) match the intrinsics' signature,
9009 // so just add all but the alignment operand.
9010 for (unsigned i = AddrOpIdx + 1; i < N->getNumOperands() - 1; ++i)
9011 Ops.push_back(N->getOperand(i));
9014 // For all node types, the alignment operand is always the last one.
9015 Ops.push_back(DAG.getConstant(Alignment, dl, MVT::i32));
9017 // If this is a non-standard-aligned STORE, the penultimate operand is the
9018 // stored value. Bitcast it to the aligned type.
9019 if (AlignedVecTy != VecTy && N->getOpcode() == ISD::STORE) {
9020 SDValue &StVal = Ops[Ops.size()-2];
9021 StVal = DAG.getNode(ISD::BITCAST, dl, AlignedVecTy, StVal);
9024 SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, dl, SDTys,
9026 MemN->getMemOperand());
9029 SmallVector<SDValue, 5> NewResults;
9030 for (unsigned i = 0; i < NumResultVecs; ++i)
9031 NewResults.push_back(SDValue(UpdN.getNode(), i));
9033 // If this is an non-standard-aligned LOAD, the first result is the loaded
9034 // value. Bitcast it to the expected result type.
9035 if (AlignedVecTy != VecTy && N->getOpcode() == ISD::LOAD) {
9036 SDValue &LdVal = NewResults[0];
9037 LdVal = DAG.getNode(ISD::BITCAST, dl, VecTy, LdVal);
9040 NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs+1)); // chain
9041 DCI.CombineTo(N, NewResults);
9042 DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs));
9049 static SDValue PerformVLDCombine(SDNode *N,
9050 TargetLowering::DAGCombinerInfo &DCI) {
9051 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
9054 return CombineBaseUpdate(N, DCI);
9057 /// CombineVLDDUP - For a VDUPLANE node N, check if its source operand is a
9058 /// vldN-lane (N > 1) intrinsic, and if all the other uses of that intrinsic
9059 /// are also VDUPLANEs. If so, combine them to a vldN-dup operation and
9061 static bool CombineVLDDUP(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
9062 SelectionDAG &DAG = DCI.DAG;
9063 EVT VT = N->getValueType(0);
9064 // vldN-dup instructions only support 64-bit vectors for N > 1.
9065 if (!VT.is64BitVector())
9068 // Check if the VDUPLANE operand is a vldN-dup intrinsic.
9069 SDNode *VLD = N->getOperand(0).getNode();
9070 if (VLD->getOpcode() != ISD::INTRINSIC_W_CHAIN)
9072 unsigned NumVecs = 0;
9073 unsigned NewOpc = 0;
9074 unsigned IntNo = cast<ConstantSDNode>(VLD->getOperand(1))->getZExtValue();
9075 if (IntNo == Intrinsic::arm_neon_vld2lane) {
9077 NewOpc = ARMISD::VLD2DUP;
9078 } else if (IntNo == Intrinsic::arm_neon_vld3lane) {
9080 NewOpc = ARMISD::VLD3DUP;
9081 } else if (IntNo == Intrinsic::arm_neon_vld4lane) {
9083 NewOpc = ARMISD::VLD4DUP;
9088 // First check that all the vldN-lane uses are VDUPLANEs and that the lane
9089 // numbers match the load.
9090 unsigned VLDLaneNo =
9091 cast<ConstantSDNode>(VLD->getOperand(NumVecs+3))->getZExtValue();
9092 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
9094 // Ignore uses of the chain result.
9095 if (UI.getUse().getResNo() == NumVecs)
9098 if (User->getOpcode() != ARMISD::VDUPLANE ||
9099 VLDLaneNo != cast<ConstantSDNode>(User->getOperand(1))->getZExtValue())
9103 // Create the vldN-dup node.
9106 for (n = 0; n < NumVecs; ++n)
9108 Tys[n] = MVT::Other;
9109 SDVTList SDTys = DAG.getVTList(makeArrayRef(Tys, NumVecs+1));
9110 SDValue Ops[] = { VLD->getOperand(0), VLD->getOperand(2) };
9111 MemIntrinsicSDNode *VLDMemInt = cast<MemIntrinsicSDNode>(VLD);
9112 SDValue VLDDup = DAG.getMemIntrinsicNode(NewOpc, SDLoc(VLD), SDTys,
9113 Ops, VLDMemInt->getMemoryVT(),
9114 VLDMemInt->getMemOperand());
9117 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
9119 unsigned ResNo = UI.getUse().getResNo();
9120 // Ignore uses of the chain result.
9121 if (ResNo == NumVecs)
9124 DCI.CombineTo(User, SDValue(VLDDup.getNode(), ResNo));
9127 // Now the vldN-lane intrinsic is dead except for its chain result.
9128 // Update uses of the chain.
9129 std::vector<SDValue> VLDDupResults;
9130 for (unsigned n = 0; n < NumVecs; ++n)
9131 VLDDupResults.push_back(SDValue(VLDDup.getNode(), n));
9132 VLDDupResults.push_back(SDValue(VLDDup.getNode(), NumVecs));
9133 DCI.CombineTo(VLD, VLDDupResults);
9138 /// PerformVDUPLANECombine - Target-specific dag combine xforms for
9139 /// ARMISD::VDUPLANE.
9140 static SDValue PerformVDUPLANECombine(SDNode *N,
9141 TargetLowering::DAGCombinerInfo &DCI) {
9142 SDValue Op = N->getOperand(0);
9144 // If the source is a vldN-lane (N > 1) intrinsic, and all the other uses
9145 // of that intrinsic are also VDUPLANEs, combine them to a vldN-dup operation.
9146 if (CombineVLDDUP(N, DCI))
9147 return SDValue(N, 0);
9149 // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is
9150 // redundant. Ignore bit_converts for now; element sizes are checked below.
9151 while (Op.getOpcode() == ISD::BITCAST)
9152 Op = Op.getOperand(0);
9153 if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM)
9156 // Make sure the VMOV element size is not bigger than the VDUPLANE elements.
9157 unsigned EltSize = Op.getValueType().getVectorElementType().getSizeInBits();
9158 // The canonical VMOV for a zero vector uses a 32-bit element size.
9159 unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
9161 if (ARM_AM::decodeNEONModImm(Imm, EltBits) == 0)
9163 EVT VT = N->getValueType(0);
9164 if (EltSize > VT.getVectorElementType().getSizeInBits())
9167 return DCI.DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Op);
9170 static SDValue PerformLOADCombine(SDNode *N,
9171 TargetLowering::DAGCombinerInfo &DCI) {
9172 EVT VT = N->getValueType(0);
9174 // If this is a legal vector load, try to combine it into a VLD1_UPD.
9175 if (ISD::isNormalLoad(N) && VT.isVector() &&
9176 DCI.DAG.getTargetLoweringInfo().isTypeLegal(VT))
9177 return CombineBaseUpdate(N, DCI);
9182 /// PerformSTORECombine - Target-specific dag combine xforms for
9184 static SDValue PerformSTORECombine(SDNode *N,
9185 TargetLowering::DAGCombinerInfo &DCI) {
9186 StoreSDNode *St = cast<StoreSDNode>(N);
9187 if (St->isVolatile())
9190 // Optimize trunc store (of multiple scalars) to shuffle and store. First,
9191 // pack all of the elements in one place. Next, store to memory in fewer
9193 SDValue StVal = St->getValue();
9194 EVT VT = StVal.getValueType();
9195 if (St->isTruncatingStore() && VT.isVector()) {
9196 SelectionDAG &DAG = DCI.DAG;
9197 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9198 EVT StVT = St->getMemoryVT();
9199 unsigned NumElems = VT.getVectorNumElements();
9200 assert(StVT != VT && "Cannot truncate to the same type");
9201 unsigned FromEltSz = VT.getVectorElementType().getSizeInBits();
9202 unsigned ToEltSz = StVT.getVectorElementType().getSizeInBits();
9204 // From, To sizes and ElemCount must be pow of two
9205 if (!isPowerOf2_32(NumElems * FromEltSz * ToEltSz)) return SDValue();
9207 // We are going to use the original vector elt for storing.
9208 // Accumulated smaller vector elements must be a multiple of the store size.
9209 if (0 != (NumElems * FromEltSz) % ToEltSz) return SDValue();
9211 unsigned SizeRatio = FromEltSz / ToEltSz;
9212 assert(SizeRatio * NumElems * ToEltSz == VT.getSizeInBits());
9214 // Create a type on which we perform the shuffle.
9215 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(), StVT.getScalarType(),
9216 NumElems*SizeRatio);
9217 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
9220 SDValue WideVec = DAG.getNode(ISD::BITCAST, DL, WideVecVT, StVal);
9221 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
9222 for (unsigned i = 0; i < NumElems; ++i)
9223 ShuffleVec[i] = TLI.isBigEndian() ? (i+1) * SizeRatio - 1 : i * SizeRatio;
9225 // Can't shuffle using an illegal type.
9226 if (!TLI.isTypeLegal(WideVecVT)) return SDValue();
9228 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, DL, WideVec,
9229 DAG.getUNDEF(WideVec.getValueType()),
9231 // At this point all of the data is stored at the bottom of the
9232 // register. We now need to save it to mem.
9234 // Find the largest store unit
9235 MVT StoreType = MVT::i8;
9236 for (MVT Tp : MVT::integer_valuetypes()) {
9237 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToEltSz)
9240 // Didn't find a legal store type.
9241 if (!TLI.isTypeLegal(StoreType))
9244 // Bitcast the original vector into a vector of store-size units
9245 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
9246 StoreType, VT.getSizeInBits()/EVT(StoreType).getSizeInBits());
9247 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
9248 SDValue ShuffWide = DAG.getNode(ISD::BITCAST, DL, StoreVecVT, Shuff);
9249 SmallVector<SDValue, 8> Chains;
9250 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits()/8, DL,
9251 TLI.getPointerTy());
9252 SDValue BasePtr = St->getBasePtr();
9254 // Perform one or more big stores into memory.
9255 unsigned E = (ToEltSz*NumElems)/StoreType.getSizeInBits();
9256 for (unsigned I = 0; I < E; I++) {
9257 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
9258 StoreType, ShuffWide,
9259 DAG.getIntPtrConstant(I, DL));
9260 SDValue Ch = DAG.getStore(St->getChain(), DL, SubVec, BasePtr,
9261 St->getPointerInfo(), St->isVolatile(),
9262 St->isNonTemporal(), St->getAlignment());
9263 BasePtr = DAG.getNode(ISD::ADD, DL, BasePtr.getValueType(), BasePtr,
9265 Chains.push_back(Ch);
9267 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
9270 if (!ISD::isNormalStore(St))
9273 // Split a store of a VMOVDRR into two integer stores to avoid mixing NEON and
9274 // ARM stores of arguments in the same cache line.
9275 if (StVal.getNode()->getOpcode() == ARMISD::VMOVDRR &&
9276 StVal.getNode()->hasOneUse()) {
9277 SelectionDAG &DAG = DCI.DAG;
9278 bool isBigEndian = DAG.getTargetLoweringInfo().isBigEndian();
9280 SDValue BasePtr = St->getBasePtr();
9281 SDValue NewST1 = DAG.getStore(St->getChain(), DL,
9282 StVal.getNode()->getOperand(isBigEndian ? 1 : 0 ),
9283 BasePtr, St->getPointerInfo(), St->isVolatile(),
9284 St->isNonTemporal(), St->getAlignment());
9286 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
9287 DAG.getConstant(4, DL, MVT::i32));
9288 return DAG.getStore(NewST1.getValue(0), DL,
9289 StVal.getNode()->getOperand(isBigEndian ? 0 : 1),
9290 OffsetPtr, St->getPointerInfo(), St->isVolatile(),
9291 St->isNonTemporal(),
9292 std::min(4U, St->getAlignment() / 2));
9295 if (StVal.getValueType() == MVT::i64 &&
9296 StVal.getNode()->getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
9298 // Bitcast an i64 store extracted from a vector to f64.
9299 // Otherwise, the i64 value will be legalized to a pair of i32 values.
9300 SelectionDAG &DAG = DCI.DAG;
9302 SDValue IntVec = StVal.getOperand(0);
9303 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
9304 IntVec.getValueType().getVectorNumElements());
9305 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, IntVec);
9306 SDValue ExtElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
9307 Vec, StVal.getOperand(1));
9309 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ExtElt);
9310 // Make the DAGCombiner fold the bitcasts.
9311 DCI.AddToWorklist(Vec.getNode());
9312 DCI.AddToWorklist(ExtElt.getNode());
9313 DCI.AddToWorklist(V.getNode());
9314 return DAG.getStore(St->getChain(), dl, V, St->getBasePtr(),
9315 St->getPointerInfo(), St->isVolatile(),
9316 St->isNonTemporal(), St->getAlignment(),
9320 // If this is a legal vector store, try to combine it into a VST1_UPD.
9321 if (ISD::isNormalStore(N) && VT.isVector() &&
9322 DCI.DAG.getTargetLoweringInfo().isTypeLegal(VT))
9323 return CombineBaseUpdate(N, DCI);
9328 // isConstVecPow2 - Return true if each vector element is a power of 2, all
9329 // elements are the same constant, C, and Log2(C) ranges from 1 to 32.
9330 static bool isConstVecPow2(SDValue ConstVec, bool isSigned, uint64_t &C)
9334 for (unsigned I = 0, E = ConstVec.getValueType().getVectorNumElements();
9336 ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(ConstVec.getOperand(I));
9341 APFloat APF = C->getValueAPF();
9342 if (APF.convertToInteger(&cN, 64, isSigned, APFloat::rmTowardZero, &isExact)
9343 != APFloat::opOK || !isExact)
9346 c0 = (I == 0) ? cN : c0;
9347 if (!isPowerOf2_64(cN) || c0 != cN || Log2_64(c0) < 1 || Log2_64(c0) > 32)
9354 /// PerformVCVTCombine - VCVT (floating-point to fixed-point, Advanced SIMD)
9355 /// can replace combinations of VMUL and VCVT (floating-point to integer)
9356 /// when the VMUL has a constant operand that is a power of 2.
9358 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
9359 /// vmul.f32 d16, d17, d16
9360 /// vcvt.s32.f32 d16, d16
9362 /// vcvt.s32.f32 d16, d16, #3
9363 static SDValue PerformVCVTCombine(SDNode *N,
9364 TargetLowering::DAGCombinerInfo &DCI,
9365 const ARMSubtarget *Subtarget) {
9366 SelectionDAG &DAG = DCI.DAG;
9367 SDValue Op = N->getOperand(0);
9369 if (!Subtarget->hasNEON() || !Op.getValueType().isVector() ||
9370 Op.getOpcode() != ISD::FMUL)
9374 SDValue N0 = Op->getOperand(0);
9375 SDValue ConstVec = Op->getOperand(1);
9376 bool isSigned = N->getOpcode() == ISD::FP_TO_SINT;
9378 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
9379 !isConstVecPow2(ConstVec, isSigned, C))
9382 MVT FloatTy = Op.getSimpleValueType().getVectorElementType();
9383 MVT IntTy = N->getSimpleValueType(0).getVectorElementType();
9384 unsigned NumLanes = Op.getValueType().getVectorNumElements();
9385 if (FloatTy.getSizeInBits() != 32 || IntTy.getSizeInBits() > 32 ||
9387 // These instructions only exist converting from f32 to i32. We can handle
9388 // smaller integers by generating an extra truncate, but larger ones would
9389 // be lossy. We also can't handle more then 4 lanes, since these intructions
9390 // only support v2i32/v4i32 types.
9395 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfp2fxs :
9396 Intrinsic::arm_neon_vcvtfp2fxu;
9397 SDValue FixConv = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl,
9398 NumLanes == 2 ? MVT::v2i32 : MVT::v4i32,
9399 DAG.getConstant(IntrinsicOpcode, dl, MVT::i32),
9401 DAG.getConstant(Log2_64(C), dl, MVT::i32));
9403 if (IntTy.getSizeInBits() < FloatTy.getSizeInBits())
9404 FixConv = DAG.getNode(ISD::TRUNCATE, dl, N->getValueType(0), FixConv);
9409 /// PerformVDIVCombine - VCVT (fixed-point to floating-point, Advanced SIMD)
9410 /// can replace combinations of VCVT (integer to floating-point) and VDIV
9411 /// when the VDIV has a constant operand that is a power of 2.
9413 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
9414 /// vcvt.f32.s32 d16, d16
9415 /// vdiv.f32 d16, d17, d16
9417 /// vcvt.f32.s32 d16, d16, #3
9418 static SDValue PerformVDIVCombine(SDNode *N,
9419 TargetLowering::DAGCombinerInfo &DCI,
9420 const ARMSubtarget *Subtarget) {
9421 SelectionDAG &DAG = DCI.DAG;
9422 SDValue Op = N->getOperand(0);
9423 unsigned OpOpcode = Op.getNode()->getOpcode();
9425 if (!Subtarget->hasNEON() || !N->getValueType(0).isVector() ||
9426 (OpOpcode != ISD::SINT_TO_FP && OpOpcode != ISD::UINT_TO_FP))
9430 SDValue ConstVec = N->getOperand(1);
9431 bool isSigned = OpOpcode == ISD::SINT_TO_FP;
9433 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
9434 !isConstVecPow2(ConstVec, isSigned, C))
9437 MVT FloatTy = N->getSimpleValueType(0).getVectorElementType();
9438 MVT IntTy = Op.getOperand(0).getSimpleValueType().getVectorElementType();
9439 if (FloatTy.getSizeInBits() != 32 || IntTy.getSizeInBits() > 32) {
9440 // These instructions only exist converting from i32 to f32. We can handle
9441 // smaller integers by generating an extra extend, but larger ones would
9447 SDValue ConvInput = Op.getOperand(0);
9448 unsigned NumLanes = Op.getValueType().getVectorNumElements();
9449 if (IntTy.getSizeInBits() < FloatTy.getSizeInBits())
9450 ConvInput = DAG.getNode(isSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND,
9451 dl, NumLanes == 2 ? MVT::v2i32 : MVT::v4i32,
9454 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfxs2fp :
9455 Intrinsic::arm_neon_vcvtfxu2fp;
9456 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl,
9458 DAG.getConstant(IntrinsicOpcode, dl, MVT::i32),
9459 ConvInput, DAG.getConstant(Log2_64(C), dl, MVT::i32));
9462 /// Getvshiftimm - Check if this is a valid build_vector for the immediate
9463 /// operand of a vector shift operation, where all the elements of the
9464 /// build_vector must have the same constant integer value.
9465 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
9466 // Ignore bit_converts.
9467 while (Op.getOpcode() == ISD::BITCAST)
9468 Op = Op.getOperand(0);
9469 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
9470 APInt SplatBits, SplatUndef;
9471 unsigned SplatBitSize;
9473 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
9474 HasAnyUndefs, ElementBits) ||
9475 SplatBitSize > ElementBits)
9477 Cnt = SplatBits.getSExtValue();
9481 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
9482 /// operand of a vector shift left operation. That value must be in the range:
9483 /// 0 <= Value < ElementBits for a left shift; or
9484 /// 0 <= Value <= ElementBits for a long left shift.
9485 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
9486 assert(VT.isVector() && "vector shift count is not a vector type");
9487 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
9488 if (! getVShiftImm(Op, ElementBits, Cnt))
9490 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
9493 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
9494 /// operand of a vector shift right operation. For a shift opcode, the value
9495 /// is positive, but for an intrinsic the value count must be negative. The
9496 /// absolute value must be in the range:
9497 /// 1 <= |Value| <= ElementBits for a right shift; or
9498 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
9499 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
9501 assert(VT.isVector() && "vector shift count is not a vector type");
9502 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
9503 if (! getVShiftImm(Op, ElementBits, Cnt))
9507 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
9510 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
9511 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
9512 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
9515 // Don't do anything for most intrinsics.
9518 // Vector shifts: check for immediate versions and lower them.
9519 // Note: This is done during DAG combining instead of DAG legalizing because
9520 // the build_vectors for 64-bit vector element shift counts are generally
9521 // not legal, and it is hard to see their values after they get legalized to
9522 // loads from a constant pool.
9523 case Intrinsic::arm_neon_vshifts:
9524 case Intrinsic::arm_neon_vshiftu:
9525 case Intrinsic::arm_neon_vrshifts:
9526 case Intrinsic::arm_neon_vrshiftu:
9527 case Intrinsic::arm_neon_vrshiftn:
9528 case Intrinsic::arm_neon_vqshifts:
9529 case Intrinsic::arm_neon_vqshiftu:
9530 case Intrinsic::arm_neon_vqshiftsu:
9531 case Intrinsic::arm_neon_vqshiftns:
9532 case Intrinsic::arm_neon_vqshiftnu:
9533 case Intrinsic::arm_neon_vqshiftnsu:
9534 case Intrinsic::arm_neon_vqrshiftns:
9535 case Intrinsic::arm_neon_vqrshiftnu:
9536 case Intrinsic::arm_neon_vqrshiftnsu: {
9537 EVT VT = N->getOperand(1).getValueType();
9539 unsigned VShiftOpc = 0;
9542 case Intrinsic::arm_neon_vshifts:
9543 case Intrinsic::arm_neon_vshiftu:
9544 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
9545 VShiftOpc = ARMISD::VSHL;
9548 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
9549 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
9550 ARMISD::VSHRs : ARMISD::VSHRu);
9555 case Intrinsic::arm_neon_vrshifts:
9556 case Intrinsic::arm_neon_vrshiftu:
9557 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
9561 case Intrinsic::arm_neon_vqshifts:
9562 case Intrinsic::arm_neon_vqshiftu:
9563 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
9567 case Intrinsic::arm_neon_vqshiftsu:
9568 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
9570 llvm_unreachable("invalid shift count for vqshlu intrinsic");
9572 case Intrinsic::arm_neon_vrshiftn:
9573 case Intrinsic::arm_neon_vqshiftns:
9574 case Intrinsic::arm_neon_vqshiftnu:
9575 case Intrinsic::arm_neon_vqshiftnsu:
9576 case Intrinsic::arm_neon_vqrshiftns:
9577 case Intrinsic::arm_neon_vqrshiftnu:
9578 case Intrinsic::arm_neon_vqrshiftnsu:
9579 // Narrowing shifts require an immediate right shift.
9580 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
9582 llvm_unreachable("invalid shift count for narrowing vector shift "
9586 llvm_unreachable("unhandled vector shift");
9590 case Intrinsic::arm_neon_vshifts:
9591 case Intrinsic::arm_neon_vshiftu:
9592 // Opcode already set above.
9594 case Intrinsic::arm_neon_vrshifts:
9595 VShiftOpc = ARMISD::VRSHRs; break;
9596 case Intrinsic::arm_neon_vrshiftu:
9597 VShiftOpc = ARMISD::VRSHRu; break;
9598 case Intrinsic::arm_neon_vrshiftn:
9599 VShiftOpc = ARMISD::VRSHRN; break;
9600 case Intrinsic::arm_neon_vqshifts:
9601 VShiftOpc = ARMISD::VQSHLs; break;
9602 case Intrinsic::arm_neon_vqshiftu:
9603 VShiftOpc = ARMISD::VQSHLu; break;
9604 case Intrinsic::arm_neon_vqshiftsu:
9605 VShiftOpc = ARMISD::VQSHLsu; break;
9606 case Intrinsic::arm_neon_vqshiftns:
9607 VShiftOpc = ARMISD::VQSHRNs; break;
9608 case Intrinsic::arm_neon_vqshiftnu:
9609 VShiftOpc = ARMISD::VQSHRNu; break;
9610 case Intrinsic::arm_neon_vqshiftnsu:
9611 VShiftOpc = ARMISD::VQSHRNsu; break;
9612 case Intrinsic::arm_neon_vqrshiftns:
9613 VShiftOpc = ARMISD::VQRSHRNs; break;
9614 case Intrinsic::arm_neon_vqrshiftnu:
9615 VShiftOpc = ARMISD::VQRSHRNu; break;
9616 case Intrinsic::arm_neon_vqrshiftnsu:
9617 VShiftOpc = ARMISD::VQRSHRNsu; break;
9621 return DAG.getNode(VShiftOpc, dl, N->getValueType(0),
9622 N->getOperand(1), DAG.getConstant(Cnt, dl, MVT::i32));
9625 case Intrinsic::arm_neon_vshiftins: {
9626 EVT VT = N->getOperand(1).getValueType();
9628 unsigned VShiftOpc = 0;
9630 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
9631 VShiftOpc = ARMISD::VSLI;
9632 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
9633 VShiftOpc = ARMISD::VSRI;
9635 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
9639 return DAG.getNode(VShiftOpc, dl, N->getValueType(0),
9640 N->getOperand(1), N->getOperand(2),
9641 DAG.getConstant(Cnt, dl, MVT::i32));
9644 case Intrinsic::arm_neon_vqrshifts:
9645 case Intrinsic::arm_neon_vqrshiftu:
9646 // No immediate versions of these to check for.
9653 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
9654 /// lowers them. As with the vector shift intrinsics, this is done during DAG
9655 /// combining instead of DAG legalizing because the build_vectors for 64-bit
9656 /// vector element shift counts are generally not legal, and it is hard to see
9657 /// their values after they get legalized to loads from a constant pool.
9658 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
9659 const ARMSubtarget *ST) {
9660 EVT VT = N->getValueType(0);
9661 if (N->getOpcode() == ISD::SRL && VT == MVT::i32 && ST->hasV6Ops()) {
9662 // Canonicalize (srl (bswap x), 16) to (rotr (bswap x), 16) if the high
9663 // 16-bits of x is zero. This optimizes rev + lsr 16 to rev16.
9664 SDValue N1 = N->getOperand(1);
9665 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) {
9666 SDValue N0 = N->getOperand(0);
9667 if (C->getZExtValue() == 16 && N0.getOpcode() == ISD::BSWAP &&
9668 DAG.MaskedValueIsZero(N0.getOperand(0),
9669 APInt::getHighBitsSet(32, 16)))
9670 return DAG.getNode(ISD::ROTR, SDLoc(N), VT, N0, N1);
9674 // Nothing to be done for scalar shifts.
9675 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9676 if (!VT.isVector() || !TLI.isTypeLegal(VT))
9679 assert(ST->hasNEON() && "unexpected vector shift");
9682 switch (N->getOpcode()) {
9683 default: llvm_unreachable("unexpected shift opcode");
9686 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt)) {
9688 return DAG.getNode(ARMISD::VSHL, dl, VT, N->getOperand(0),
9689 DAG.getConstant(Cnt, dl, MVT::i32));
9695 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
9696 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
9697 ARMISD::VSHRs : ARMISD::VSHRu);
9699 return DAG.getNode(VShiftOpc, dl, VT, N->getOperand(0),
9700 DAG.getConstant(Cnt, dl, MVT::i32));
9706 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
9707 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
9708 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
9709 const ARMSubtarget *ST) {
9710 SDValue N0 = N->getOperand(0);
9712 // Check for sign- and zero-extensions of vector extract operations of 8-
9713 // and 16-bit vector elements. NEON supports these directly. They are
9714 // handled during DAG combining because type legalization will promote them
9715 // to 32-bit types and it is messy to recognize the operations after that.
9716 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
9717 SDValue Vec = N0.getOperand(0);
9718 SDValue Lane = N0.getOperand(1);
9719 EVT VT = N->getValueType(0);
9720 EVT EltVT = N0.getValueType();
9721 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9723 if (VT == MVT::i32 &&
9724 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
9725 TLI.isTypeLegal(Vec.getValueType()) &&
9726 isa<ConstantSDNode>(Lane)) {
9729 switch (N->getOpcode()) {
9730 default: llvm_unreachable("unexpected opcode");
9731 case ISD::SIGN_EXTEND:
9732 Opc = ARMISD::VGETLANEs;
9734 case ISD::ZERO_EXTEND:
9735 case ISD::ANY_EXTEND:
9736 Opc = ARMISD::VGETLANEu;
9739 return DAG.getNode(Opc, SDLoc(N), VT, Vec, Lane);
9746 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
9747 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
9748 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
9749 const ARMSubtarget *ST) {
9750 // If the target supports NEON, try to use vmax/vmin instructions for f32
9751 // selects like "x < y ? x : y". Unless the NoNaNsFPMath option is set,
9752 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
9753 // a NaN; only do the transformation when it matches that behavior.
9755 // For now only do this when using NEON for FP operations; if using VFP, it
9756 // is not obvious that the benefit outweighs the cost of switching to the
9758 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
9759 N->getValueType(0) != MVT::f32)
9762 SDValue CondLHS = N->getOperand(0);
9763 SDValue CondRHS = N->getOperand(1);
9764 SDValue LHS = N->getOperand(2);
9765 SDValue RHS = N->getOperand(3);
9766 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
9768 unsigned Opcode = 0;
9770 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
9771 IsReversed = false; // x CC y ? x : y
9772 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
9773 IsReversed = true ; // x CC y ? y : x
9787 // If LHS is NaN, an ordered comparison will be false and the result will
9788 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
9789 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
9790 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
9791 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
9793 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
9794 // will return -0, so vmin can only be used for unsafe math or if one of
9795 // the operands is known to be nonzero.
9796 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
9797 !DAG.getTarget().Options.UnsafeFPMath &&
9798 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9800 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
9809 // If LHS is NaN, an ordered comparison will be false and the result will
9810 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
9811 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
9812 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
9813 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
9815 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
9816 // will return +0, so vmax can only be used for unsafe math or if one of
9817 // the operands is known to be nonzero.
9818 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
9819 !DAG.getTarget().Options.UnsafeFPMath &&
9820 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9822 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
9828 return DAG.getNode(Opcode, SDLoc(N), N->getValueType(0), LHS, RHS);
9831 /// PerformCMOVCombine - Target-specific DAG combining for ARMISD::CMOV.
9833 ARMTargetLowering::PerformCMOVCombine(SDNode *N, SelectionDAG &DAG) const {
9834 SDValue Cmp = N->getOperand(4);
9835 if (Cmp.getOpcode() != ARMISD::CMPZ)
9836 // Only looking at EQ and NE cases.
9839 EVT VT = N->getValueType(0);
9841 SDValue LHS = Cmp.getOperand(0);
9842 SDValue RHS = Cmp.getOperand(1);
9843 SDValue FalseVal = N->getOperand(0);
9844 SDValue TrueVal = N->getOperand(1);
9845 SDValue ARMcc = N->getOperand(2);
9846 ARMCC::CondCodes CC =
9847 (ARMCC::CondCodes)cast<ConstantSDNode>(ARMcc)->getZExtValue();
9865 /// FIXME: Turn this into a target neutral optimization?
9867 if (CC == ARMCC::NE && FalseVal == RHS && FalseVal != LHS) {
9868 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, TrueVal, ARMcc,
9869 N->getOperand(3), Cmp);
9870 } else if (CC == ARMCC::EQ && TrueVal == RHS) {
9872 SDValue NewCmp = getARMCmp(LHS, RHS, ISD::SETNE, ARMcc, DAG, dl);
9873 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, FalseVal, ARMcc,
9874 N->getOperand(3), NewCmp);
9877 if (Res.getNode()) {
9878 APInt KnownZero, KnownOne;
9879 DAG.computeKnownBits(SDValue(N,0), KnownZero, KnownOne);
9880 // Capture demanded bits information that would be otherwise lost.
9881 if (KnownZero == 0xfffffffe)
9882 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
9883 DAG.getValueType(MVT::i1));
9884 else if (KnownZero == 0xffffff00)
9885 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
9886 DAG.getValueType(MVT::i8));
9887 else if (KnownZero == 0xffff0000)
9888 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
9889 DAG.getValueType(MVT::i16));
9895 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
9896 DAGCombinerInfo &DCI) const {
9897 switch (N->getOpcode()) {
9899 case ISD::ADDC: return PerformADDCCombine(N, DCI, Subtarget);
9900 case ISD::ADD: return PerformADDCombine(N, DCI, Subtarget);
9901 case ISD::SUB: return PerformSUBCombine(N, DCI);
9902 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
9903 case ISD::OR: return PerformORCombine(N, DCI, Subtarget);
9904 case ISD::XOR: return PerformXORCombine(N, DCI, Subtarget);
9905 case ISD::AND: return PerformANDCombine(N, DCI, Subtarget);
9906 case ARMISD::BFI: return PerformBFICombine(N, DCI);
9907 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI, Subtarget);
9908 case ARMISD::VMOVDRR: return PerformVMOVDRRCombine(N, DCI.DAG);
9909 case ISD::STORE: return PerformSTORECombine(N, DCI);
9910 case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DCI, Subtarget);
9911 case ISD::INSERT_VECTOR_ELT: return PerformInsertEltCombine(N, DCI);
9912 case ISD::VECTOR_SHUFFLE: return PerformVECTOR_SHUFFLECombine(N, DCI.DAG);
9913 case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI);
9914 case ISD::FP_TO_SINT:
9915 case ISD::FP_TO_UINT: return PerformVCVTCombine(N, DCI, Subtarget);
9916 case ISD::FDIV: return PerformVDIVCombine(N, DCI, Subtarget);
9917 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
9920 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
9921 case ISD::SIGN_EXTEND:
9922 case ISD::ZERO_EXTEND:
9923 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
9924 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
9925 case ARMISD::CMOV: return PerformCMOVCombine(N, DCI.DAG);
9926 case ISD::LOAD: return PerformLOADCombine(N, DCI);
9927 case ARMISD::VLD2DUP:
9928 case ARMISD::VLD3DUP:
9929 case ARMISD::VLD4DUP:
9930 return PerformVLDCombine(N, DCI);
9931 case ARMISD::BUILD_VECTOR:
9932 return PerformARMBUILD_VECTORCombine(N, DCI);
9933 case ISD::INTRINSIC_VOID:
9934 case ISD::INTRINSIC_W_CHAIN:
9935 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
9936 case Intrinsic::arm_neon_vld1:
9937 case Intrinsic::arm_neon_vld2:
9938 case Intrinsic::arm_neon_vld3:
9939 case Intrinsic::arm_neon_vld4:
9940 case Intrinsic::arm_neon_vld2lane:
9941 case Intrinsic::arm_neon_vld3lane:
9942 case Intrinsic::arm_neon_vld4lane:
9943 case Intrinsic::arm_neon_vst1:
9944 case Intrinsic::arm_neon_vst2:
9945 case Intrinsic::arm_neon_vst3:
9946 case Intrinsic::arm_neon_vst4:
9947 case Intrinsic::arm_neon_vst2lane:
9948 case Intrinsic::arm_neon_vst3lane:
9949 case Intrinsic::arm_neon_vst4lane:
9950 return PerformVLDCombine(N, DCI);
9958 bool ARMTargetLowering::isDesirableToTransformToIntegerOp(unsigned Opc,
9960 return (VT == MVT::f32) && (Opc == ISD::LOAD || Opc == ISD::STORE);
9963 bool ARMTargetLowering::allowsMisalignedMemoryAccesses(EVT VT,
9967 // The AllowsUnaliged flag models the SCTLR.A setting in ARM cpus
9968 bool AllowsUnaligned = Subtarget->allowsUnalignedMem();
9970 switch (VT.getSimpleVT().SimpleTy) {
9976 // Unaligned access can use (for example) LRDB, LRDH, LDR
9977 if (AllowsUnaligned) {
9979 *Fast = Subtarget->hasV7Ops();
9986 // For any little-endian targets with neon, we can support unaligned ld/st
9987 // of D and Q (e.g. {D0,D1}) registers by using vld1.i8/vst1.i8.
9988 // A big-endian target may also explicitly support unaligned accesses
9989 if (Subtarget->hasNEON() && (AllowsUnaligned || isLittleEndian())) {
9999 static bool memOpAlign(unsigned DstAlign, unsigned SrcAlign,
10000 unsigned AlignCheck) {
10001 return ((SrcAlign == 0 || SrcAlign % AlignCheck == 0) &&
10002 (DstAlign == 0 || DstAlign % AlignCheck == 0));
10005 EVT ARMTargetLowering::getOptimalMemOpType(uint64_t Size,
10006 unsigned DstAlign, unsigned SrcAlign,
10007 bool IsMemset, bool ZeroMemset,
10009 MachineFunction &MF) const {
10010 const Function *F = MF.getFunction();
10012 // See if we can use NEON instructions for this...
10013 if ((!IsMemset || ZeroMemset) && Subtarget->hasNEON() &&
10014 !F->hasFnAttribute(Attribute::NoImplicitFloat)) {
10017 (memOpAlign(SrcAlign, DstAlign, 16) ||
10018 (allowsMisalignedMemoryAccesses(MVT::v2f64, 0, 1, &Fast) && Fast))) {
10020 } else if (Size >= 8 &&
10021 (memOpAlign(SrcAlign, DstAlign, 8) ||
10022 (allowsMisalignedMemoryAccesses(MVT::f64, 0, 1, &Fast) &&
10028 // Lowering to i32/i16 if the size permits.
10031 else if (Size >= 2)
10034 // Let the target-independent logic figure it out.
10038 bool ARMTargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
10039 if (Val.getOpcode() != ISD::LOAD)
10042 EVT VT1 = Val.getValueType();
10043 if (!VT1.isSimple() || !VT1.isInteger() ||
10044 !VT2.isSimple() || !VT2.isInteger())
10047 switch (VT1.getSimpleVT().SimpleTy) {
10052 // 8-bit and 16-bit loads implicitly zero-extend to 32-bits.
10059 bool ARMTargetLowering::isVectorLoadExtDesirable(SDValue ExtVal) const {
10060 EVT VT = ExtVal.getValueType();
10062 if (!isTypeLegal(VT))
10065 // Don't create a loadext if we can fold the extension into a wide/long
10067 // If there's more than one user instruction, the loadext is desirable no
10068 // matter what. There can be two uses by the same instruction.
10069 if (ExtVal->use_empty() ||
10070 !ExtVal->use_begin()->isOnlyUserOf(ExtVal.getNode()))
10073 SDNode *U = *ExtVal->use_begin();
10074 if ((U->getOpcode() == ISD::ADD || U->getOpcode() == ISD::SUB ||
10075 U->getOpcode() == ISD::SHL || U->getOpcode() == ARMISD::VSHL))
10081 bool ARMTargetLowering::allowTruncateForTailCall(Type *Ty1, Type *Ty2) const {
10082 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
10085 if (!isTypeLegal(EVT::getEVT(Ty1)))
10088 assert(Ty1->getPrimitiveSizeInBits() <= 64 && "i128 is probably not a noop");
10090 // Assuming the caller doesn't have a zeroext or signext return parameter,
10091 // truncation all the way down to i1 is valid.
10096 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
10100 unsigned Scale = 1;
10101 switch (VT.getSimpleVT().SimpleTy) {
10102 default: return false;
10117 if ((V & (Scale - 1)) != 0)
10120 return V == (V & ((1LL << 5) - 1));
10123 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
10124 const ARMSubtarget *Subtarget) {
10125 bool isNeg = false;
10131 switch (VT.getSimpleVT().SimpleTy) {
10132 default: return false;
10137 // + imm12 or - imm8
10139 return V == (V & ((1LL << 8) - 1));
10140 return V == (V & ((1LL << 12) - 1));
10143 // Same as ARM mode. FIXME: NEON?
10144 if (!Subtarget->hasVFP2())
10149 return V == (V & ((1LL << 8) - 1));
10153 /// isLegalAddressImmediate - Return true if the integer value can be used
10154 /// as the offset of the target addressing mode for load / store of the
10156 static bool isLegalAddressImmediate(int64_t V, EVT VT,
10157 const ARMSubtarget *Subtarget) {
10161 if (!VT.isSimple())
10164 if (Subtarget->isThumb1Only())
10165 return isLegalT1AddressImmediate(V, VT);
10166 else if (Subtarget->isThumb2())
10167 return isLegalT2AddressImmediate(V, VT, Subtarget);
10172 switch (VT.getSimpleVT().SimpleTy) {
10173 default: return false;
10178 return V == (V & ((1LL << 12) - 1));
10181 return V == (V & ((1LL << 8) - 1));
10184 if (!Subtarget->hasVFP2()) // FIXME: NEON?
10189 return V == (V & ((1LL << 8) - 1));
10193 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
10195 int Scale = AM.Scale;
10199 switch (VT.getSimpleVT().SimpleTy) {
10200 default: return false;
10208 Scale = Scale & ~1;
10209 return Scale == 2 || Scale == 4 || Scale == 8;
10212 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
10216 // Note, we allow "void" uses (basically, uses that aren't loads or
10217 // stores), because arm allows folding a scale into many arithmetic
10218 // operations. This should be made more precise and revisited later.
10220 // Allow r << imm, but the imm has to be a multiple of two.
10221 if (Scale & 1) return false;
10222 return isPowerOf2_32(Scale);
10226 /// isLegalAddressingMode - Return true if the addressing mode represented
10227 /// by AM is legal for this target, for a load/store of the specified type.
10228 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
10230 EVT VT = getValueType(Ty, true);
10231 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
10234 // Can never fold addr of global into load/store.
10238 switch (AM.Scale) {
10239 case 0: // no scale reg, must be "r+i" or "r", or "i".
10242 if (Subtarget->isThumb1Only())
10246 // ARM doesn't support any R+R*scale+imm addr modes.
10250 if (!VT.isSimple())
10253 if (Subtarget->isThumb2())
10254 return isLegalT2ScaledAddressingMode(AM, VT);
10256 int Scale = AM.Scale;
10257 switch (VT.getSimpleVT().SimpleTy) {
10258 default: return false;
10262 if (Scale < 0) Scale = -Scale;
10266 return isPowerOf2_32(Scale & ~1);
10270 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
10275 // Note, we allow "void" uses (basically, uses that aren't loads or
10276 // stores), because arm allows folding a scale into many arithmetic
10277 // operations. This should be made more precise and revisited later.
10279 // Allow r << imm, but the imm has to be a multiple of two.
10280 if (Scale & 1) return false;
10281 return isPowerOf2_32(Scale);
10287 /// isLegalICmpImmediate - Return true if the specified immediate is legal
10288 /// icmp immediate, that is the target has icmp instructions which can compare
10289 /// a register against the immediate without having to materialize the
10290 /// immediate into a register.
10291 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
10292 // Thumb2 and ARM modes can use cmn for negative immediates.
10293 if (!Subtarget->isThumb())
10294 return ARM_AM::getSOImmVal(std::abs(Imm)) != -1;
10295 if (Subtarget->isThumb2())
10296 return ARM_AM::getT2SOImmVal(std::abs(Imm)) != -1;
10297 // Thumb1 doesn't have cmn, and only 8-bit immediates.
10298 return Imm >= 0 && Imm <= 255;
10301 /// isLegalAddImmediate - Return true if the specified immediate is a legal add
10302 /// *or sub* immediate, that is the target has add or sub instructions which can
10303 /// add a register with the immediate without having to materialize the
10304 /// immediate into a register.
10305 bool ARMTargetLowering::isLegalAddImmediate(int64_t Imm) const {
10306 // Same encoding for add/sub, just flip the sign.
10307 int64_t AbsImm = std::abs(Imm);
10308 if (!Subtarget->isThumb())
10309 return ARM_AM::getSOImmVal(AbsImm) != -1;
10310 if (Subtarget->isThumb2())
10311 return ARM_AM::getT2SOImmVal(AbsImm) != -1;
10312 // Thumb1 only has 8-bit unsigned immediate.
10313 return AbsImm >= 0 && AbsImm <= 255;
10316 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
10317 bool isSEXTLoad, SDValue &Base,
10318 SDValue &Offset, bool &isInc,
10319 SelectionDAG &DAG) {
10320 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
10323 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
10324 // AddressingMode 3
10325 Base = Ptr->getOperand(0);
10326 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
10327 int RHSC = (int)RHS->getZExtValue();
10328 if (RHSC < 0 && RHSC > -256) {
10329 assert(Ptr->getOpcode() == ISD::ADD);
10331 Offset = DAG.getConstant(-RHSC, SDLoc(Ptr), RHS->getValueType(0));
10335 isInc = (Ptr->getOpcode() == ISD::ADD);
10336 Offset = Ptr->getOperand(1);
10338 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
10339 // AddressingMode 2
10340 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
10341 int RHSC = (int)RHS->getZExtValue();
10342 if (RHSC < 0 && RHSC > -0x1000) {
10343 assert(Ptr->getOpcode() == ISD::ADD);
10345 Offset = DAG.getConstant(-RHSC, SDLoc(Ptr), RHS->getValueType(0));
10346 Base = Ptr->getOperand(0);
10351 if (Ptr->getOpcode() == ISD::ADD) {
10353 ARM_AM::ShiftOpc ShOpcVal=
10354 ARM_AM::getShiftOpcForNode(Ptr->getOperand(0).getOpcode());
10355 if (ShOpcVal != ARM_AM::no_shift) {
10356 Base = Ptr->getOperand(1);
10357 Offset = Ptr->getOperand(0);
10359 Base = Ptr->getOperand(0);
10360 Offset = Ptr->getOperand(1);
10365 isInc = (Ptr->getOpcode() == ISD::ADD);
10366 Base = Ptr->getOperand(0);
10367 Offset = Ptr->getOperand(1);
10371 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
10375 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
10376 bool isSEXTLoad, SDValue &Base,
10377 SDValue &Offset, bool &isInc,
10378 SelectionDAG &DAG) {
10379 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
10382 Base = Ptr->getOperand(0);
10383 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
10384 int RHSC = (int)RHS->getZExtValue();
10385 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
10386 assert(Ptr->getOpcode() == ISD::ADD);
10388 Offset = DAG.getConstant(-RHSC, SDLoc(Ptr), RHS->getValueType(0));
10390 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
10391 isInc = Ptr->getOpcode() == ISD::ADD;
10392 Offset = DAG.getConstant(RHSC, SDLoc(Ptr), RHS->getValueType(0));
10400 /// getPreIndexedAddressParts - returns true by value, base pointer and
10401 /// offset pointer and addressing mode by reference if the node's address
10402 /// can be legally represented as pre-indexed load / store address.
10404 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
10406 ISD::MemIndexedMode &AM,
10407 SelectionDAG &DAG) const {
10408 if (Subtarget->isThumb1Only())
10413 bool isSEXTLoad = false;
10414 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
10415 Ptr = LD->getBasePtr();
10416 VT = LD->getMemoryVT();
10417 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
10418 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
10419 Ptr = ST->getBasePtr();
10420 VT = ST->getMemoryVT();
10425 bool isLegal = false;
10426 if (Subtarget->isThumb2())
10427 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
10428 Offset, isInc, DAG);
10430 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
10431 Offset, isInc, DAG);
10435 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
10439 /// getPostIndexedAddressParts - returns true by value, base pointer and
10440 /// offset pointer and addressing mode by reference if this node can be
10441 /// combined with a load / store to form a post-indexed load / store.
10442 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
10445 ISD::MemIndexedMode &AM,
10446 SelectionDAG &DAG) const {
10447 if (Subtarget->isThumb1Only())
10452 bool isSEXTLoad = false;
10453 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
10454 VT = LD->getMemoryVT();
10455 Ptr = LD->getBasePtr();
10456 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
10457 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
10458 VT = ST->getMemoryVT();
10459 Ptr = ST->getBasePtr();
10464 bool isLegal = false;
10465 if (Subtarget->isThumb2())
10466 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
10469 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
10475 // Swap base ptr and offset to catch more post-index load / store when
10476 // it's legal. In Thumb2 mode, offset must be an immediate.
10477 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
10478 !Subtarget->isThumb2())
10479 std::swap(Base, Offset);
10481 // Post-indexed load / store update the base pointer.
10486 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
10490 void ARMTargetLowering::computeKnownBitsForTargetNode(const SDValue Op,
10493 const SelectionDAG &DAG,
10494 unsigned Depth) const {
10495 unsigned BitWidth = KnownOne.getBitWidth();
10496 KnownZero = KnownOne = APInt(BitWidth, 0);
10497 switch (Op.getOpcode()) {
10503 // These nodes' second result is a boolean
10504 if (Op.getResNo() == 0)
10506 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - 1);
10508 case ARMISD::CMOV: {
10509 // Bits are known zero/one if known on the LHS and RHS.
10510 DAG.computeKnownBits(Op.getOperand(0), KnownZero, KnownOne, Depth+1);
10511 if (KnownZero == 0 && KnownOne == 0) return;
10513 APInt KnownZeroRHS, KnownOneRHS;
10514 DAG.computeKnownBits(Op.getOperand(1), KnownZeroRHS, KnownOneRHS, Depth+1);
10515 KnownZero &= KnownZeroRHS;
10516 KnownOne &= KnownOneRHS;
10519 case ISD::INTRINSIC_W_CHAIN: {
10520 ConstantSDNode *CN = cast<ConstantSDNode>(Op->getOperand(1));
10521 Intrinsic::ID IntID = static_cast<Intrinsic::ID>(CN->getZExtValue());
10524 case Intrinsic::arm_ldaex:
10525 case Intrinsic::arm_ldrex: {
10526 EVT VT = cast<MemIntrinsicSDNode>(Op)->getMemoryVT();
10527 unsigned MemBits = VT.getScalarType().getSizeInBits();
10528 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - MemBits);
10536 //===----------------------------------------------------------------------===//
10537 // ARM Inline Assembly Support
10538 //===----------------------------------------------------------------------===//
10540 bool ARMTargetLowering::ExpandInlineAsm(CallInst *CI) const {
10541 // Looking for "rev" which is V6+.
10542 if (!Subtarget->hasV6Ops())
10545 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
10546 std::string AsmStr = IA->getAsmString();
10547 SmallVector<StringRef, 4> AsmPieces;
10548 SplitString(AsmStr, AsmPieces, ";\n");
10550 switch (AsmPieces.size()) {
10551 default: return false;
10553 AsmStr = AsmPieces[0];
10555 SplitString(AsmStr, AsmPieces, " \t,");
10558 if (AsmPieces.size() == 3 &&
10559 AsmPieces[0] == "rev" && AsmPieces[1] == "$0" && AsmPieces[2] == "$1" &&
10560 IA->getConstraintString().compare(0, 4, "=l,l") == 0) {
10561 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
10562 if (Ty && Ty->getBitWidth() == 32)
10563 return IntrinsicLowering::LowerToByteSwap(CI);
10571 /// getConstraintType - Given a constraint letter, return the type of
10572 /// constraint it is for this target.
10573 ARMTargetLowering::ConstraintType
10574 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
10575 if (Constraint.size() == 1) {
10576 switch (Constraint[0]) {
10578 case 'l': return C_RegisterClass;
10579 case 'w': return C_RegisterClass;
10580 case 'h': return C_RegisterClass;
10581 case 'x': return C_RegisterClass;
10582 case 't': return C_RegisterClass;
10583 case 'j': return C_Other; // Constant for movw.
10584 // An address with a single base register. Due to the way we
10585 // currently handle addresses it is the same as an 'r' memory constraint.
10586 case 'Q': return C_Memory;
10588 } else if (Constraint.size() == 2) {
10589 switch (Constraint[0]) {
10591 // All 'U+' constraints are addresses.
10592 case 'U': return C_Memory;
10595 return TargetLowering::getConstraintType(Constraint);
10598 /// Examine constraint type and operand type and determine a weight value.
10599 /// This object must already have been set up with the operand type
10600 /// and the current alternative constraint selected.
10601 TargetLowering::ConstraintWeight
10602 ARMTargetLowering::getSingleConstraintMatchWeight(
10603 AsmOperandInfo &info, const char *constraint) const {
10604 ConstraintWeight weight = CW_Invalid;
10605 Value *CallOperandVal = info.CallOperandVal;
10606 // If we don't have a value, we can't do a match,
10607 // but allow it at the lowest weight.
10608 if (!CallOperandVal)
10610 Type *type = CallOperandVal->getType();
10611 // Look at the constraint type.
10612 switch (*constraint) {
10614 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
10617 if (type->isIntegerTy()) {
10618 if (Subtarget->isThumb())
10619 weight = CW_SpecificReg;
10621 weight = CW_Register;
10625 if (type->isFloatingPointTy())
10626 weight = CW_Register;
10632 typedef std::pair<unsigned, const TargetRegisterClass*> RCPair;
10634 ARMTargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
10635 const std::string &Constraint,
10637 if (Constraint.size() == 1) {
10638 // GCC ARM Constraint Letters
10639 switch (Constraint[0]) {
10640 case 'l': // Low regs or general regs.
10641 if (Subtarget->isThumb())
10642 return RCPair(0U, &ARM::tGPRRegClass);
10643 return RCPair(0U, &ARM::GPRRegClass);
10644 case 'h': // High regs or no regs.
10645 if (Subtarget->isThumb())
10646 return RCPair(0U, &ARM::hGPRRegClass);
10649 if (Subtarget->isThumb1Only())
10650 return RCPair(0U, &ARM::tGPRRegClass);
10651 return RCPair(0U, &ARM::GPRRegClass);
10653 if (VT == MVT::Other)
10655 if (VT == MVT::f32)
10656 return RCPair(0U, &ARM::SPRRegClass);
10657 if (VT.getSizeInBits() == 64)
10658 return RCPair(0U, &ARM::DPRRegClass);
10659 if (VT.getSizeInBits() == 128)
10660 return RCPair(0U, &ARM::QPRRegClass);
10663 if (VT == MVT::Other)
10665 if (VT == MVT::f32)
10666 return RCPair(0U, &ARM::SPR_8RegClass);
10667 if (VT.getSizeInBits() == 64)
10668 return RCPair(0U, &ARM::DPR_8RegClass);
10669 if (VT.getSizeInBits() == 128)
10670 return RCPair(0U, &ARM::QPR_8RegClass);
10673 if (VT == MVT::f32)
10674 return RCPair(0U, &ARM::SPRRegClass);
10678 if (StringRef("{cc}").equals_lower(Constraint))
10679 return std::make_pair(unsigned(ARM::CPSR), &ARM::CCRRegClass);
10681 return TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
10684 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
10685 /// vector. If it is invalid, don't add anything to Ops.
10686 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
10687 std::string &Constraint,
10688 std::vector<SDValue>&Ops,
10689 SelectionDAG &DAG) const {
10692 // Currently only support length 1 constraints.
10693 if (Constraint.length() != 1) return;
10695 char ConstraintLetter = Constraint[0];
10696 switch (ConstraintLetter) {
10699 case 'I': case 'J': case 'K': case 'L':
10700 case 'M': case 'N': case 'O':
10701 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
10705 int64_t CVal64 = C->getSExtValue();
10706 int CVal = (int) CVal64;
10707 // None of these constraints allow values larger than 32 bits. Check
10708 // that the value fits in an int.
10709 if (CVal != CVal64)
10712 switch (ConstraintLetter) {
10714 // Constant suitable for movw, must be between 0 and
10716 if (Subtarget->hasV6T2Ops())
10717 if (CVal >= 0 && CVal <= 65535)
10721 if (Subtarget->isThumb1Only()) {
10722 // This must be a constant between 0 and 255, for ADD
10724 if (CVal >= 0 && CVal <= 255)
10726 } else if (Subtarget->isThumb2()) {
10727 // A constant that can be used as an immediate value in a
10728 // data-processing instruction.
10729 if (ARM_AM::getT2SOImmVal(CVal) != -1)
10732 // A constant that can be used as an immediate value in a
10733 // data-processing instruction.
10734 if (ARM_AM::getSOImmVal(CVal) != -1)
10740 if (Subtarget->isThumb()) { // FIXME thumb2
10741 // This must be a constant between -255 and -1, for negated ADD
10742 // immediates. This can be used in GCC with an "n" modifier that
10743 // prints the negated value, for use with SUB instructions. It is
10744 // not useful otherwise but is implemented for compatibility.
10745 if (CVal >= -255 && CVal <= -1)
10748 // This must be a constant between -4095 and 4095. It is not clear
10749 // what this constraint is intended for. Implemented for
10750 // compatibility with GCC.
10751 if (CVal >= -4095 && CVal <= 4095)
10757 if (Subtarget->isThumb1Only()) {
10758 // A 32-bit value where only one byte has a nonzero value. Exclude
10759 // zero to match GCC. This constraint is used by GCC internally for
10760 // constants that can be loaded with a move/shift combination.
10761 // It is not useful otherwise but is implemented for compatibility.
10762 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
10764 } else if (Subtarget->isThumb2()) {
10765 // A constant whose bitwise inverse can be used as an immediate
10766 // value in a data-processing instruction. This can be used in GCC
10767 // with a "B" modifier that prints the inverted value, for use with
10768 // BIC and MVN instructions. It is not useful otherwise but is
10769 // implemented for compatibility.
10770 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
10773 // A constant whose bitwise inverse can be used as an immediate
10774 // value in a data-processing instruction. This can be used in GCC
10775 // with a "B" modifier that prints the inverted value, for use with
10776 // BIC and MVN instructions. It is not useful otherwise but is
10777 // implemented for compatibility.
10778 if (ARM_AM::getSOImmVal(~CVal) != -1)
10784 if (Subtarget->isThumb1Only()) {
10785 // This must be a constant between -7 and 7,
10786 // for 3-operand ADD/SUB immediate instructions.
10787 if (CVal >= -7 && CVal < 7)
10789 } else if (Subtarget->isThumb2()) {
10790 // A constant whose negation can be used as an immediate value in a
10791 // data-processing instruction. This can be used in GCC with an "n"
10792 // modifier that prints the negated value, for use with SUB
10793 // instructions. It is not useful otherwise but is implemented for
10795 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
10798 // A constant whose negation can be used as an immediate value in a
10799 // data-processing instruction. This can be used in GCC with an "n"
10800 // modifier that prints the negated value, for use with SUB
10801 // instructions. It is not useful otherwise but is implemented for
10803 if (ARM_AM::getSOImmVal(-CVal) != -1)
10809 if (Subtarget->isThumb()) { // FIXME thumb2
10810 // This must be a multiple of 4 between 0 and 1020, for
10811 // ADD sp + immediate.
10812 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
10815 // A power of two or a constant between 0 and 32. This is used in
10816 // GCC for the shift amount on shifted register operands, but it is
10817 // useful in general for any shift amounts.
10818 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
10824 if (Subtarget->isThumb()) { // FIXME thumb2
10825 // This must be a constant between 0 and 31, for shift amounts.
10826 if (CVal >= 0 && CVal <= 31)
10832 if (Subtarget->isThumb()) { // FIXME thumb2
10833 // This must be a multiple of 4 between -508 and 508, for
10834 // ADD/SUB sp = sp + immediate.
10835 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
10840 Result = DAG.getTargetConstant(CVal, SDLoc(Op), Op.getValueType());
10844 if (Result.getNode()) {
10845 Ops.push_back(Result);
10848 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
10851 SDValue ARMTargetLowering::LowerDivRem(SDValue Op, SelectionDAG &DAG) const {
10852 assert(Subtarget->isTargetAEABI() && "Register-based DivRem lowering only");
10853 unsigned Opcode = Op->getOpcode();
10854 assert((Opcode == ISD::SDIVREM || Opcode == ISD::UDIVREM) &&
10855 "Invalid opcode for Div/Rem lowering");
10856 bool isSigned = (Opcode == ISD::SDIVREM);
10857 EVT VT = Op->getValueType(0);
10858 Type *Ty = VT.getTypeForEVT(*DAG.getContext());
10861 switch (VT.getSimpleVT().SimpleTy) {
10862 default: llvm_unreachable("Unexpected request for libcall!");
10863 case MVT::i8: LC = isSigned ? RTLIB::SDIVREM_I8 : RTLIB::UDIVREM_I8; break;
10864 case MVT::i16: LC = isSigned ? RTLIB::SDIVREM_I16 : RTLIB::UDIVREM_I16; break;
10865 case MVT::i32: LC = isSigned ? RTLIB::SDIVREM_I32 : RTLIB::UDIVREM_I32; break;
10866 case MVT::i64: LC = isSigned ? RTLIB::SDIVREM_I64 : RTLIB::UDIVREM_I64; break;
10869 SDValue InChain = DAG.getEntryNode();
10871 TargetLowering::ArgListTy Args;
10872 TargetLowering::ArgListEntry Entry;
10873 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i) {
10874 EVT ArgVT = Op->getOperand(i).getValueType();
10875 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
10876 Entry.Node = Op->getOperand(i);
10878 Entry.isSExt = isSigned;
10879 Entry.isZExt = !isSigned;
10880 Args.push_back(Entry);
10883 SDValue Callee = DAG.getExternalSymbol(getLibcallName(LC),
10886 Type *RetTy = (Type*)StructType::get(Ty, Ty, nullptr);
10889 TargetLowering::CallLoweringInfo CLI(DAG);
10890 CLI.setDebugLoc(dl).setChain(InChain)
10891 .setCallee(getLibcallCallingConv(LC), RetTy, Callee, std::move(Args), 0)
10892 .setInRegister().setSExtResult(isSigned).setZExtResult(!isSigned);
10894 std::pair<SDValue, SDValue> CallInfo = LowerCallTo(CLI);
10895 return CallInfo.first;
10899 ARMTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const {
10900 assert(Subtarget->isTargetWindows() && "unsupported target platform");
10904 SDValue Chain = Op.getOperand(0);
10905 SDValue Size = Op.getOperand(1);
10907 SDValue Words = DAG.getNode(ISD::SRL, DL, MVT::i32, Size,
10908 DAG.getConstant(2, DL, MVT::i32));
10911 Chain = DAG.getCopyToReg(Chain, DL, ARM::R4, Words, Flag);
10912 Flag = Chain.getValue(1);
10914 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
10915 Chain = DAG.getNode(ARMISD::WIN__CHKSTK, DL, NodeTys, Chain, Flag);
10917 SDValue NewSP = DAG.getCopyFromReg(Chain, DL, ARM::SP, MVT::i32);
10918 Chain = NewSP.getValue(1);
10920 SDValue Ops[2] = { NewSP, Chain };
10921 return DAG.getMergeValues(Ops, DL);
10924 SDValue ARMTargetLowering::LowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) const {
10925 assert(Op.getValueType() == MVT::f64 && Subtarget->isFPOnlySP() &&
10926 "Unexpected type for custom-lowering FP_EXTEND");
10929 LC = RTLIB::getFPEXT(Op.getOperand(0).getValueType(), Op.getValueType());
10931 SDValue SrcVal = Op.getOperand(0);
10932 return makeLibCall(DAG, LC, Op.getValueType(), &SrcVal, 1,
10933 /*isSigned*/ false, SDLoc(Op)).first;
10936 SDValue ARMTargetLowering::LowerFP_ROUND(SDValue Op, SelectionDAG &DAG) const {
10937 assert(Op.getOperand(0).getValueType() == MVT::f64 &&
10938 Subtarget->isFPOnlySP() &&
10939 "Unexpected type for custom-lowering FP_ROUND");
10942 LC = RTLIB::getFPROUND(Op.getOperand(0).getValueType(), Op.getValueType());
10944 SDValue SrcVal = Op.getOperand(0);
10945 return makeLibCall(DAG, LC, Op.getValueType(), &SrcVal, 1,
10946 /*isSigned*/ false, SDLoc(Op)).first;
10950 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
10951 // The ARM target isn't yet aware of offsets.
10955 bool ARM::isBitFieldInvertedMask(unsigned v) {
10956 if (v == 0xffffffff)
10959 // there can be 1's on either or both "outsides", all the "inside"
10960 // bits must be 0's
10961 return isShiftedMask_32(~v);
10964 /// isFPImmLegal - Returns true if the target can instruction select the
10965 /// specified FP immediate natively. If false, the legalizer will
10966 /// materialize the FP immediate as a load from a constant pool.
10967 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
10968 if (!Subtarget->hasVFP3())
10970 if (VT == MVT::f32)
10971 return ARM_AM::getFP32Imm(Imm) != -1;
10972 if (VT == MVT::f64 && !Subtarget->isFPOnlySP())
10973 return ARM_AM::getFP64Imm(Imm) != -1;
10977 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
10978 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
10979 /// specified in the intrinsic calls.
10980 bool ARMTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
10982 unsigned Intrinsic) const {
10983 switch (Intrinsic) {
10984 case Intrinsic::arm_neon_vld1:
10985 case Intrinsic::arm_neon_vld2:
10986 case Intrinsic::arm_neon_vld3:
10987 case Intrinsic::arm_neon_vld4:
10988 case Intrinsic::arm_neon_vld2lane:
10989 case Intrinsic::arm_neon_vld3lane:
10990 case Intrinsic::arm_neon_vld4lane: {
10991 Info.opc = ISD::INTRINSIC_W_CHAIN;
10992 // Conservatively set memVT to the entire set of vectors loaded.
10993 uint64_t NumElts = getDataLayout()->getTypeAllocSize(I.getType()) / 8;
10994 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
10995 Info.ptrVal = I.getArgOperand(0);
10997 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
10998 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
10999 Info.vol = false; // volatile loads with NEON intrinsics not supported
11000 Info.readMem = true;
11001 Info.writeMem = false;
11004 case Intrinsic::arm_neon_vst1:
11005 case Intrinsic::arm_neon_vst2:
11006 case Intrinsic::arm_neon_vst3:
11007 case Intrinsic::arm_neon_vst4:
11008 case Intrinsic::arm_neon_vst2lane:
11009 case Intrinsic::arm_neon_vst3lane:
11010 case Intrinsic::arm_neon_vst4lane: {
11011 Info.opc = ISD::INTRINSIC_VOID;
11012 // Conservatively set memVT to the entire set of vectors stored.
11013 unsigned NumElts = 0;
11014 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
11015 Type *ArgTy = I.getArgOperand(ArgI)->getType();
11016 if (!ArgTy->isVectorTy())
11018 NumElts += getDataLayout()->getTypeAllocSize(ArgTy) / 8;
11020 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
11021 Info.ptrVal = I.getArgOperand(0);
11023 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
11024 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
11025 Info.vol = false; // volatile stores with NEON intrinsics not supported
11026 Info.readMem = false;
11027 Info.writeMem = true;
11030 case Intrinsic::arm_ldaex:
11031 case Intrinsic::arm_ldrex: {
11032 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(0)->getType());
11033 Info.opc = ISD::INTRINSIC_W_CHAIN;
11034 Info.memVT = MVT::getVT(PtrTy->getElementType());
11035 Info.ptrVal = I.getArgOperand(0);
11037 Info.align = getDataLayout()->getABITypeAlignment(PtrTy->getElementType());
11039 Info.readMem = true;
11040 Info.writeMem = false;
11043 case Intrinsic::arm_stlex:
11044 case Intrinsic::arm_strex: {
11045 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(1)->getType());
11046 Info.opc = ISD::INTRINSIC_W_CHAIN;
11047 Info.memVT = MVT::getVT(PtrTy->getElementType());
11048 Info.ptrVal = I.getArgOperand(1);
11050 Info.align = getDataLayout()->getABITypeAlignment(PtrTy->getElementType());
11052 Info.readMem = false;
11053 Info.writeMem = true;
11056 case Intrinsic::arm_stlexd:
11057 case Intrinsic::arm_strexd: {
11058 Info.opc = ISD::INTRINSIC_W_CHAIN;
11059 Info.memVT = MVT::i64;
11060 Info.ptrVal = I.getArgOperand(2);
11064 Info.readMem = false;
11065 Info.writeMem = true;
11068 case Intrinsic::arm_ldaexd:
11069 case Intrinsic::arm_ldrexd: {
11070 Info.opc = ISD::INTRINSIC_W_CHAIN;
11071 Info.memVT = MVT::i64;
11072 Info.ptrVal = I.getArgOperand(0);
11076 Info.readMem = true;
11077 Info.writeMem = false;
11087 /// \brief Returns true if it is beneficial to convert a load of a constant
11088 /// to just the constant itself.
11089 bool ARMTargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
11091 assert(Ty->isIntegerTy());
11093 unsigned Bits = Ty->getPrimitiveSizeInBits();
11094 if (Bits == 0 || Bits > 32)
11099 bool ARMTargetLowering::hasLoadLinkedStoreConditional() const { return true; }
11101 Instruction* ARMTargetLowering::makeDMB(IRBuilder<> &Builder,
11102 ARM_MB::MemBOpt Domain) const {
11103 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
11105 // First, if the target has no DMB, see what fallback we can use.
11106 if (!Subtarget->hasDataBarrier()) {
11107 // Some ARMv6 cpus can support data barriers with an mcr instruction.
11108 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
11110 if (Subtarget->hasV6Ops() && !Subtarget->isThumb()) {
11111 Function *MCR = llvm::Intrinsic::getDeclaration(M, Intrinsic::arm_mcr);
11112 Value* args[6] = {Builder.getInt32(15), Builder.getInt32(0),
11113 Builder.getInt32(0), Builder.getInt32(7),
11114 Builder.getInt32(10), Builder.getInt32(5)};
11115 return Builder.CreateCall(MCR, args);
11117 // Instead of using barriers, atomic accesses on these subtargets use
11119 llvm_unreachable("makeDMB on a target so old that it has no barriers");
11122 Function *DMB = llvm::Intrinsic::getDeclaration(M, Intrinsic::arm_dmb);
11123 // Only a full system barrier exists in the M-class architectures.
11124 Domain = Subtarget->isMClass() ? ARM_MB::SY : Domain;
11125 Constant *CDomain = Builder.getInt32(Domain);
11126 return Builder.CreateCall(DMB, CDomain);
11130 // Based on http://www.cl.cam.ac.uk/~pes20/cpp/cpp0xmappings.html
11131 Instruction* ARMTargetLowering::emitLeadingFence(IRBuilder<> &Builder,
11132 AtomicOrdering Ord, bool IsStore,
11133 bool IsLoad) const {
11134 if (!getInsertFencesForAtomic())
11140 llvm_unreachable("Invalid fence: unordered/non-atomic");
11143 return nullptr; // Nothing to do
11144 case SequentiallyConsistent:
11146 return nullptr; // Nothing to do
11149 case AcquireRelease:
11150 if (Subtarget->isSwift())
11151 return makeDMB(Builder, ARM_MB::ISHST);
11152 // FIXME: add a comment with a link to documentation justifying this.
11154 return makeDMB(Builder, ARM_MB::ISH);
11156 llvm_unreachable("Unknown fence ordering in emitLeadingFence");
11159 Instruction* ARMTargetLowering::emitTrailingFence(IRBuilder<> &Builder,
11160 AtomicOrdering Ord, bool IsStore,
11161 bool IsLoad) const {
11162 if (!getInsertFencesForAtomic())
11168 llvm_unreachable("Invalid fence: unordered/not-atomic");
11171 return nullptr; // Nothing to do
11173 case AcquireRelease:
11174 case SequentiallyConsistent:
11175 return makeDMB(Builder, ARM_MB::ISH);
11177 llvm_unreachable("Unknown fence ordering in emitTrailingFence");
11180 // Loads and stores less than 64-bits are already atomic; ones above that
11181 // are doomed anyway, so defer to the default libcall and blame the OS when
11182 // things go wrong. Cortex M doesn't have ldrexd/strexd though, so don't emit
11183 // anything for those.
11184 bool ARMTargetLowering::shouldExpandAtomicStoreInIR(StoreInst *SI) const {
11185 unsigned Size = SI->getValueOperand()->getType()->getPrimitiveSizeInBits();
11186 return (Size == 64) && !Subtarget->isMClass();
11189 // Loads and stores less than 64-bits are already atomic; ones above that
11190 // are doomed anyway, so defer to the default libcall and blame the OS when
11191 // things go wrong. Cortex M doesn't have ldrexd/strexd though, so don't emit
11192 // anything for those.
11193 // FIXME: ldrd and strd are atomic if the CPU has LPAE (e.g. A15 has that
11194 // guarantee, see DDI0406C ARM architecture reference manual,
11195 // sections A8.8.72-74 LDRD)
11196 bool ARMTargetLowering::shouldExpandAtomicLoadInIR(LoadInst *LI) const {
11197 unsigned Size = LI->getType()->getPrimitiveSizeInBits();
11198 return (Size == 64) && !Subtarget->isMClass();
11201 // For the real atomic operations, we have ldrex/strex up to 32 bits,
11202 // and up to 64 bits on the non-M profiles
11203 TargetLoweringBase::AtomicRMWExpansionKind
11204 ARMTargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
11205 unsigned Size = AI->getType()->getPrimitiveSizeInBits();
11206 return (Size <= (Subtarget->isMClass() ? 32U : 64U))
11207 ? AtomicRMWExpansionKind::LLSC
11208 : AtomicRMWExpansionKind::None;
11211 // This has so far only been implemented for MachO.
11212 bool ARMTargetLowering::useLoadStackGuardNode() const {
11213 return Subtarget->isTargetMachO();
11216 bool ARMTargetLowering::canCombineStoreAndExtract(Type *VectorTy, Value *Idx,
11217 unsigned &Cost) const {
11218 // If we do not have NEON, vector types are not natively supported.
11219 if (!Subtarget->hasNEON())
11222 // Floating point values and vector values map to the same register file.
11223 // Therefore, althought we could do a store extract of a vector type, this is
11224 // better to leave at float as we have more freedom in the addressing mode for
11226 if (VectorTy->isFPOrFPVectorTy())
11229 // If the index is unknown at compile time, this is very expensive to lower
11230 // and it is not possible to combine the store with the extract.
11231 if (!isa<ConstantInt>(Idx))
11234 assert(VectorTy->isVectorTy() && "VectorTy is not a vector type");
11235 unsigned BitWidth = cast<VectorType>(VectorTy)->getBitWidth();
11236 // We can do a store + vector extract on any vector that fits perfectly in a D
11238 if (BitWidth == 64 || BitWidth == 128) {
11245 Value *ARMTargetLowering::emitLoadLinked(IRBuilder<> &Builder, Value *Addr,
11246 AtomicOrdering Ord) const {
11247 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
11248 Type *ValTy = cast<PointerType>(Addr->getType())->getElementType();
11249 bool IsAcquire = isAtLeastAcquire(Ord);
11251 // Since i64 isn't legal and intrinsics don't get type-lowered, the ldrexd
11252 // intrinsic must return {i32, i32} and we have to recombine them into a
11253 // single i64 here.
11254 if (ValTy->getPrimitiveSizeInBits() == 64) {
11255 Intrinsic::ID Int =
11256 IsAcquire ? Intrinsic::arm_ldaexd : Intrinsic::arm_ldrexd;
11257 Function *Ldrex = llvm::Intrinsic::getDeclaration(M, Int);
11259 Addr = Builder.CreateBitCast(Addr, Type::getInt8PtrTy(M->getContext()));
11260 Value *LoHi = Builder.CreateCall(Ldrex, Addr, "lohi");
11262 Value *Lo = Builder.CreateExtractValue(LoHi, 0, "lo");
11263 Value *Hi = Builder.CreateExtractValue(LoHi, 1, "hi");
11264 if (!Subtarget->isLittle())
11265 std::swap (Lo, Hi);
11266 Lo = Builder.CreateZExt(Lo, ValTy, "lo64");
11267 Hi = Builder.CreateZExt(Hi, ValTy, "hi64");
11268 return Builder.CreateOr(
11269 Lo, Builder.CreateShl(Hi, ConstantInt::get(ValTy, 32)), "val64");
11272 Type *Tys[] = { Addr->getType() };
11273 Intrinsic::ID Int = IsAcquire ? Intrinsic::arm_ldaex : Intrinsic::arm_ldrex;
11274 Function *Ldrex = llvm::Intrinsic::getDeclaration(M, Int, Tys);
11276 return Builder.CreateTruncOrBitCast(
11277 Builder.CreateCall(Ldrex, Addr),
11278 cast<PointerType>(Addr->getType())->getElementType());
11281 Value *ARMTargetLowering::emitStoreConditional(IRBuilder<> &Builder, Value *Val,
11283 AtomicOrdering Ord) const {
11284 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
11285 bool IsRelease = isAtLeastRelease(Ord);
11287 // Since the intrinsics must have legal type, the i64 intrinsics take two
11288 // parameters: "i32, i32". We must marshal Val into the appropriate form
11289 // before the call.
11290 if (Val->getType()->getPrimitiveSizeInBits() == 64) {
11291 Intrinsic::ID Int =
11292 IsRelease ? Intrinsic::arm_stlexd : Intrinsic::arm_strexd;
11293 Function *Strex = Intrinsic::getDeclaration(M, Int);
11294 Type *Int32Ty = Type::getInt32Ty(M->getContext());
11296 Value *Lo = Builder.CreateTrunc(Val, Int32Ty, "lo");
11297 Value *Hi = Builder.CreateTrunc(Builder.CreateLShr(Val, 32), Int32Ty, "hi");
11298 if (!Subtarget->isLittle())
11299 std::swap (Lo, Hi);
11300 Addr = Builder.CreateBitCast(Addr, Type::getInt8PtrTy(M->getContext()));
11301 return Builder.CreateCall3(Strex, Lo, Hi, Addr);
11304 Intrinsic::ID Int = IsRelease ? Intrinsic::arm_stlex : Intrinsic::arm_strex;
11305 Type *Tys[] = { Addr->getType() };
11306 Function *Strex = Intrinsic::getDeclaration(M, Int, Tys);
11308 return Builder.CreateCall2(
11309 Strex, Builder.CreateZExtOrBitCast(
11310 Val, Strex->getFunctionType()->getParamType(0)),
11322 static bool isHomogeneousAggregate(Type *Ty, HABaseType &Base,
11323 uint64_t &Members) {
11324 if (const StructType *ST = dyn_cast<StructType>(Ty)) {
11325 for (unsigned i = 0; i < ST->getNumElements(); ++i) {
11326 uint64_t SubMembers = 0;
11327 if (!isHomogeneousAggregate(ST->getElementType(i), Base, SubMembers))
11329 Members += SubMembers;
11331 } else if (const ArrayType *AT = dyn_cast<ArrayType>(Ty)) {
11332 uint64_t SubMembers = 0;
11333 if (!isHomogeneousAggregate(AT->getElementType(), Base, SubMembers))
11335 Members += SubMembers * AT->getNumElements();
11336 } else if (Ty->isFloatTy()) {
11337 if (Base != HA_UNKNOWN && Base != HA_FLOAT)
11341 } else if (Ty->isDoubleTy()) {
11342 if (Base != HA_UNKNOWN && Base != HA_DOUBLE)
11346 } else if (const VectorType *VT = dyn_cast<VectorType>(Ty)) {
11353 return VT->getBitWidth() == 64;
11355 return VT->getBitWidth() == 128;
11357 switch (VT->getBitWidth()) {
11370 return (Members > 0 && Members <= 4);
11373 /// \brief Return true if a type is an AAPCS-VFP homogeneous aggregate or one of
11374 /// [N x i32] or [N x i64]. This allows front-ends to skip emitting padding when
11375 /// passing according to AAPCS rules.
11376 bool ARMTargetLowering::functionArgumentNeedsConsecutiveRegisters(
11377 Type *Ty, CallingConv::ID CallConv, bool isVarArg) const {
11378 if (getEffectiveCallingConv(CallConv, isVarArg) !=
11379 CallingConv::ARM_AAPCS_VFP)
11382 HABaseType Base = HA_UNKNOWN;
11383 uint64_t Members = 0;
11384 bool IsHA = isHomogeneousAggregate(Ty, Base, Members);
11385 DEBUG(dbgs() << "isHA: " << IsHA << " "; Ty->dump());
11387 bool IsIntArray = Ty->isArrayTy() && Ty->getArrayElementType()->isIntegerTy();
11388 return IsHA || IsIntArray;