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 setOperationAction(ISD::READ_REGISTER, MVT::i64, Custom);
430 setOperationAction(ISD::WRITE_REGISTER, MVT::i64, Custom);
432 if (Subtarget->hasNEON()) {
433 addDRTypeForNEON(MVT::v2f32);
434 addDRTypeForNEON(MVT::v8i8);
435 addDRTypeForNEON(MVT::v4i16);
436 addDRTypeForNEON(MVT::v2i32);
437 addDRTypeForNEON(MVT::v1i64);
439 addQRTypeForNEON(MVT::v4f32);
440 addQRTypeForNEON(MVT::v2f64);
441 addQRTypeForNEON(MVT::v16i8);
442 addQRTypeForNEON(MVT::v8i16);
443 addQRTypeForNEON(MVT::v4i32);
444 addQRTypeForNEON(MVT::v2i64);
446 // v2f64 is legal so that QR subregs can be extracted as f64 elements, but
447 // neither Neon nor VFP support any arithmetic operations on it.
448 // The same with v4f32. But keep in mind that vadd, vsub, vmul are natively
449 // supported for v4f32.
450 setOperationAction(ISD::FADD, MVT::v2f64, Expand);
451 setOperationAction(ISD::FSUB, MVT::v2f64, Expand);
452 setOperationAction(ISD::FMUL, MVT::v2f64, Expand);
453 // FIXME: Code duplication: FDIV and FREM are expanded always, see
454 // ARMTargetLowering::addTypeForNEON method for details.
455 setOperationAction(ISD::FDIV, MVT::v2f64, Expand);
456 setOperationAction(ISD::FREM, MVT::v2f64, Expand);
457 // FIXME: Create unittest.
458 // In another words, find a way when "copysign" appears in DAG with vector
460 setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand);
461 // FIXME: Code duplication: SETCC has custom operation action, see
462 // ARMTargetLowering::addTypeForNEON method for details.
463 setOperationAction(ISD::SETCC, MVT::v2f64, Expand);
464 // FIXME: Create unittest for FNEG and for FABS.
465 setOperationAction(ISD::FNEG, MVT::v2f64, Expand);
466 setOperationAction(ISD::FABS, MVT::v2f64, Expand);
467 setOperationAction(ISD::FSQRT, MVT::v2f64, Expand);
468 setOperationAction(ISD::FSIN, MVT::v2f64, Expand);
469 setOperationAction(ISD::FCOS, MVT::v2f64, Expand);
470 setOperationAction(ISD::FPOWI, MVT::v2f64, Expand);
471 setOperationAction(ISD::FPOW, MVT::v2f64, Expand);
472 setOperationAction(ISD::FLOG, MVT::v2f64, Expand);
473 setOperationAction(ISD::FLOG2, MVT::v2f64, Expand);
474 setOperationAction(ISD::FLOG10, MVT::v2f64, Expand);
475 setOperationAction(ISD::FEXP, MVT::v2f64, Expand);
476 setOperationAction(ISD::FEXP2, MVT::v2f64, Expand);
477 // FIXME: Create unittest for FCEIL, FTRUNC, FRINT, FNEARBYINT, FFLOOR.
478 setOperationAction(ISD::FCEIL, MVT::v2f64, Expand);
479 setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand);
480 setOperationAction(ISD::FRINT, MVT::v2f64, Expand);
481 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand);
482 setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand);
483 setOperationAction(ISD::FMA, MVT::v2f64, Expand);
485 setOperationAction(ISD::FSQRT, MVT::v4f32, Expand);
486 setOperationAction(ISD::FSIN, MVT::v4f32, Expand);
487 setOperationAction(ISD::FCOS, MVT::v4f32, Expand);
488 setOperationAction(ISD::FPOWI, MVT::v4f32, Expand);
489 setOperationAction(ISD::FPOW, MVT::v4f32, Expand);
490 setOperationAction(ISD::FLOG, MVT::v4f32, Expand);
491 setOperationAction(ISD::FLOG2, MVT::v4f32, Expand);
492 setOperationAction(ISD::FLOG10, MVT::v4f32, Expand);
493 setOperationAction(ISD::FEXP, MVT::v4f32, Expand);
494 setOperationAction(ISD::FEXP2, MVT::v4f32, Expand);
495 setOperationAction(ISD::FCEIL, MVT::v4f32, Expand);
496 setOperationAction(ISD::FTRUNC, MVT::v4f32, Expand);
497 setOperationAction(ISD::FRINT, MVT::v4f32, Expand);
498 setOperationAction(ISD::FNEARBYINT, MVT::v4f32, Expand);
499 setOperationAction(ISD::FFLOOR, MVT::v4f32, Expand);
501 // Mark v2f32 intrinsics.
502 setOperationAction(ISD::FSQRT, MVT::v2f32, Expand);
503 setOperationAction(ISD::FSIN, MVT::v2f32, Expand);
504 setOperationAction(ISD::FCOS, MVT::v2f32, Expand);
505 setOperationAction(ISD::FPOWI, MVT::v2f32, Expand);
506 setOperationAction(ISD::FPOW, MVT::v2f32, Expand);
507 setOperationAction(ISD::FLOG, MVT::v2f32, Expand);
508 setOperationAction(ISD::FLOG2, MVT::v2f32, Expand);
509 setOperationAction(ISD::FLOG10, MVT::v2f32, Expand);
510 setOperationAction(ISD::FEXP, MVT::v2f32, Expand);
511 setOperationAction(ISD::FEXP2, MVT::v2f32, Expand);
512 setOperationAction(ISD::FCEIL, MVT::v2f32, Expand);
513 setOperationAction(ISD::FTRUNC, MVT::v2f32, Expand);
514 setOperationAction(ISD::FRINT, MVT::v2f32, Expand);
515 setOperationAction(ISD::FNEARBYINT, MVT::v2f32, Expand);
516 setOperationAction(ISD::FFLOOR, MVT::v2f32, Expand);
518 // Neon does not support some operations on v1i64 and v2i64 types.
519 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
520 // Custom handling for some quad-vector types to detect VMULL.
521 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
522 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
523 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
524 // Custom handling for some vector types to avoid expensive expansions
525 setOperationAction(ISD::SDIV, MVT::v4i16, Custom);
526 setOperationAction(ISD::SDIV, MVT::v8i8, Custom);
527 setOperationAction(ISD::UDIV, MVT::v4i16, Custom);
528 setOperationAction(ISD::UDIV, MVT::v8i8, Custom);
529 setOperationAction(ISD::SETCC, MVT::v1i64, Expand);
530 setOperationAction(ISD::SETCC, MVT::v2i64, Expand);
531 // Neon does not have single instruction SINT_TO_FP and UINT_TO_FP with
532 // a destination type that is wider than the source, and nor does
533 // it have a FP_TO_[SU]INT instruction with a narrower destination than
535 setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Custom);
536 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
537 setOperationAction(ISD::FP_TO_UINT, MVT::v4i16, Custom);
538 setOperationAction(ISD::FP_TO_SINT, MVT::v4i16, Custom);
540 setOperationAction(ISD::FP_ROUND, MVT::v2f32, Expand);
541 setOperationAction(ISD::FP_EXTEND, MVT::v2f64, Expand);
543 // NEON does not have single instruction CTPOP for vectors with element
544 // types wider than 8-bits. However, custom lowering can leverage the
545 // v8i8/v16i8 vcnt instruction.
546 setOperationAction(ISD::CTPOP, MVT::v2i32, Custom);
547 setOperationAction(ISD::CTPOP, MVT::v4i32, Custom);
548 setOperationAction(ISD::CTPOP, MVT::v4i16, Custom);
549 setOperationAction(ISD::CTPOP, MVT::v8i16, Custom);
551 // NEON only has FMA instructions as of VFP4.
552 if (!Subtarget->hasVFP4()) {
553 setOperationAction(ISD::FMA, MVT::v2f32, Expand);
554 setOperationAction(ISD::FMA, MVT::v4f32, Expand);
557 setTargetDAGCombine(ISD::INTRINSIC_VOID);
558 setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN);
559 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
560 setTargetDAGCombine(ISD::SHL);
561 setTargetDAGCombine(ISD::SRL);
562 setTargetDAGCombine(ISD::SRA);
563 setTargetDAGCombine(ISD::SIGN_EXTEND);
564 setTargetDAGCombine(ISD::ZERO_EXTEND);
565 setTargetDAGCombine(ISD::ANY_EXTEND);
566 setTargetDAGCombine(ISD::SELECT_CC);
567 setTargetDAGCombine(ISD::BUILD_VECTOR);
568 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
569 setTargetDAGCombine(ISD::INSERT_VECTOR_ELT);
570 setTargetDAGCombine(ISD::STORE);
571 setTargetDAGCombine(ISD::FP_TO_SINT);
572 setTargetDAGCombine(ISD::FP_TO_UINT);
573 setTargetDAGCombine(ISD::FDIV);
574 setTargetDAGCombine(ISD::LOAD);
576 // It is legal to extload from v4i8 to v4i16 or v4i32.
577 for (MVT Ty : {MVT::v8i8, MVT::v4i8, MVT::v2i8, MVT::v4i16, MVT::v2i16,
579 for (MVT VT : MVT::integer_vector_valuetypes()) {
580 setLoadExtAction(ISD::EXTLOAD, VT, Ty, Legal);
581 setLoadExtAction(ISD::ZEXTLOAD, VT, Ty, Legal);
582 setLoadExtAction(ISD::SEXTLOAD, VT, Ty, Legal);
587 // ARM and Thumb2 support UMLAL/SMLAL.
588 if (!Subtarget->isThumb1Only())
589 setTargetDAGCombine(ISD::ADDC);
591 if (Subtarget->isFPOnlySP()) {
592 // When targetting a floating-point unit with only single-precision
593 // operations, f64 is legal for the few double-precision instructions which
594 // are present However, no double-precision operations other than moves,
595 // loads and stores are provided by the hardware.
596 setOperationAction(ISD::FADD, MVT::f64, Expand);
597 setOperationAction(ISD::FSUB, MVT::f64, Expand);
598 setOperationAction(ISD::FMUL, MVT::f64, Expand);
599 setOperationAction(ISD::FMA, MVT::f64, Expand);
600 setOperationAction(ISD::FDIV, MVT::f64, Expand);
601 setOperationAction(ISD::FREM, MVT::f64, Expand);
602 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
603 setOperationAction(ISD::FGETSIGN, MVT::f64, Expand);
604 setOperationAction(ISD::FNEG, MVT::f64, Expand);
605 setOperationAction(ISD::FABS, MVT::f64, Expand);
606 setOperationAction(ISD::FSQRT, MVT::f64, Expand);
607 setOperationAction(ISD::FSIN, MVT::f64, Expand);
608 setOperationAction(ISD::FCOS, MVT::f64, Expand);
609 setOperationAction(ISD::FPOWI, MVT::f64, Expand);
610 setOperationAction(ISD::FPOW, MVT::f64, Expand);
611 setOperationAction(ISD::FLOG, MVT::f64, Expand);
612 setOperationAction(ISD::FLOG2, MVT::f64, Expand);
613 setOperationAction(ISD::FLOG10, MVT::f64, Expand);
614 setOperationAction(ISD::FEXP, MVT::f64, Expand);
615 setOperationAction(ISD::FEXP2, MVT::f64, Expand);
616 setOperationAction(ISD::FCEIL, MVT::f64, Expand);
617 setOperationAction(ISD::FTRUNC, MVT::f64, Expand);
618 setOperationAction(ISD::FRINT, MVT::f64, Expand);
619 setOperationAction(ISD::FNEARBYINT, MVT::f64, Expand);
620 setOperationAction(ISD::FFLOOR, MVT::f64, Expand);
621 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
622 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
623 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
624 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
625 setOperationAction(ISD::FP_TO_SINT, MVT::f64, Custom);
626 setOperationAction(ISD::FP_TO_UINT, MVT::f64, Custom);
627 setOperationAction(ISD::FP_ROUND, MVT::f32, Custom);
628 setOperationAction(ISD::FP_EXTEND, MVT::f64, Custom);
631 computeRegisterProperties(Subtarget->getRegisterInfo());
633 // ARM does not have floating-point extending loads.
634 for (MVT VT : MVT::fp_valuetypes()) {
635 setLoadExtAction(ISD::EXTLOAD, VT, MVT::f32, Expand);
636 setLoadExtAction(ISD::EXTLOAD, VT, MVT::f16, Expand);
639 // ... or truncating stores
640 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
641 setTruncStoreAction(MVT::f32, MVT::f16, Expand);
642 setTruncStoreAction(MVT::f64, MVT::f16, Expand);
644 // ARM does not have i1 sign extending load.
645 for (MVT VT : MVT::integer_valuetypes())
646 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Promote);
648 // ARM supports all 4 flavors of integer indexed load / store.
649 if (!Subtarget->isThumb1Only()) {
650 for (unsigned im = (unsigned)ISD::PRE_INC;
651 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
652 setIndexedLoadAction(im, MVT::i1, Legal);
653 setIndexedLoadAction(im, MVT::i8, Legal);
654 setIndexedLoadAction(im, MVT::i16, Legal);
655 setIndexedLoadAction(im, MVT::i32, Legal);
656 setIndexedStoreAction(im, MVT::i1, Legal);
657 setIndexedStoreAction(im, MVT::i8, Legal);
658 setIndexedStoreAction(im, MVT::i16, Legal);
659 setIndexedStoreAction(im, MVT::i32, Legal);
663 setOperationAction(ISD::SADDO, MVT::i32, Custom);
664 setOperationAction(ISD::UADDO, MVT::i32, Custom);
665 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
666 setOperationAction(ISD::USUBO, MVT::i32, Custom);
668 // i64 operation support.
669 setOperationAction(ISD::MUL, MVT::i64, Expand);
670 setOperationAction(ISD::MULHU, MVT::i32, Expand);
671 if (Subtarget->isThumb1Only()) {
672 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
673 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
675 if (Subtarget->isThumb1Only() || !Subtarget->hasV6Ops()
676 || (Subtarget->isThumb2() && !Subtarget->hasThumb2DSP()))
677 setOperationAction(ISD::MULHS, MVT::i32, Expand);
679 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
680 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
681 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
682 setOperationAction(ISD::SRL, MVT::i64, Custom);
683 setOperationAction(ISD::SRA, MVT::i64, Custom);
685 if (!Subtarget->isThumb1Only()) {
686 // FIXME: We should do this for Thumb1 as well.
687 setOperationAction(ISD::ADDC, MVT::i32, Custom);
688 setOperationAction(ISD::ADDE, MVT::i32, Custom);
689 setOperationAction(ISD::SUBC, MVT::i32, Custom);
690 setOperationAction(ISD::SUBE, MVT::i32, Custom);
693 // ARM does not have ROTL.
694 setOperationAction(ISD::ROTL, MVT::i32, Expand);
695 setOperationAction(ISD::CTTZ, MVT::i32, Custom);
696 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
697 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
698 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
700 // These just redirect to CTTZ and CTLZ on ARM.
701 setOperationAction(ISD::CTTZ_ZERO_UNDEF , MVT::i32 , Expand);
702 setOperationAction(ISD::CTLZ_ZERO_UNDEF , MVT::i32 , Expand);
704 setOperationAction(ISD::READCYCLECOUNTER, MVT::i64, Custom);
706 // Only ARMv6 has BSWAP.
707 if (!Subtarget->hasV6Ops())
708 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
710 if (!(Subtarget->hasDivide() && Subtarget->isThumb2()) &&
711 !(Subtarget->hasDivideInARMMode() && !Subtarget->isThumb())) {
712 // These are expanded into libcalls if the cpu doesn't have HW divider.
713 setOperationAction(ISD::SDIV, MVT::i32, Expand);
714 setOperationAction(ISD::UDIV, MVT::i32, Expand);
717 // FIXME: Also set divmod for SREM on EABI
718 setOperationAction(ISD::SREM, MVT::i32, Expand);
719 setOperationAction(ISD::UREM, MVT::i32, Expand);
720 // Register based DivRem for AEABI (RTABI 4.2)
721 if (Subtarget->isTargetAEABI()) {
722 setLibcallName(RTLIB::SDIVREM_I8, "__aeabi_idivmod");
723 setLibcallName(RTLIB::SDIVREM_I16, "__aeabi_idivmod");
724 setLibcallName(RTLIB::SDIVREM_I32, "__aeabi_idivmod");
725 setLibcallName(RTLIB::SDIVREM_I64, "__aeabi_ldivmod");
726 setLibcallName(RTLIB::UDIVREM_I8, "__aeabi_uidivmod");
727 setLibcallName(RTLIB::UDIVREM_I16, "__aeabi_uidivmod");
728 setLibcallName(RTLIB::UDIVREM_I32, "__aeabi_uidivmod");
729 setLibcallName(RTLIB::UDIVREM_I64, "__aeabi_uldivmod");
731 setLibcallCallingConv(RTLIB::SDIVREM_I8, CallingConv::ARM_AAPCS);
732 setLibcallCallingConv(RTLIB::SDIVREM_I16, CallingConv::ARM_AAPCS);
733 setLibcallCallingConv(RTLIB::SDIVREM_I32, CallingConv::ARM_AAPCS);
734 setLibcallCallingConv(RTLIB::SDIVREM_I64, CallingConv::ARM_AAPCS);
735 setLibcallCallingConv(RTLIB::UDIVREM_I8, CallingConv::ARM_AAPCS);
736 setLibcallCallingConv(RTLIB::UDIVREM_I16, CallingConv::ARM_AAPCS);
737 setLibcallCallingConv(RTLIB::UDIVREM_I32, CallingConv::ARM_AAPCS);
738 setLibcallCallingConv(RTLIB::UDIVREM_I64, CallingConv::ARM_AAPCS);
740 setOperationAction(ISD::SDIVREM, MVT::i32, Custom);
741 setOperationAction(ISD::UDIVREM, MVT::i32, Custom);
743 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
744 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
747 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
748 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
749 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
750 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
751 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
753 setOperationAction(ISD::TRAP, MVT::Other, Legal);
755 // Use the default implementation.
756 setOperationAction(ISD::VASTART, MVT::Other, Custom);
757 setOperationAction(ISD::VAARG, MVT::Other, Expand);
758 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
759 setOperationAction(ISD::VAEND, MVT::Other, Expand);
760 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
761 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
763 if (!Subtarget->isTargetMachO()) {
764 // Non-MachO platforms may return values in these registers via the
765 // personality function.
766 setExceptionPointerRegister(ARM::R0);
767 setExceptionSelectorRegister(ARM::R1);
770 if (Subtarget->getTargetTriple().isWindowsItaniumEnvironment())
771 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
773 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
775 // ARMv6 Thumb1 (except for CPUs that support dmb / dsb) and earlier use
776 // the default expansion. If we are targeting a single threaded system,
777 // then set them all for expand so we can lower them later into their
779 if (TM.Options.ThreadModel == ThreadModel::Single)
780 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Expand);
781 else if (Subtarget->hasAnyDataBarrier() && !Subtarget->isThumb1Only()) {
782 // ATOMIC_FENCE needs custom lowering; the others should have been expanded
783 // to ldrex/strex loops already.
784 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Custom);
786 // On v8, we have particularly efficient implementations of atomic fences
787 // if they can be combined with nearby atomic loads and stores.
788 if (!Subtarget->hasV8Ops()) {
789 // Automatically insert fences (dmb ish) around ATOMIC_SWAP etc.
790 setInsertFencesForAtomic(true);
793 // If there's anything we can use as a barrier, go through custom lowering
795 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other,
796 Subtarget->hasAnyDataBarrier() ? Custom : Expand);
798 // Set them all for expansion, which will force libcalls.
799 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Expand);
800 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Expand);
801 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Expand);
802 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Expand);
803 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Expand);
804 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Expand);
805 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Expand);
806 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Expand);
807 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i32, Expand);
808 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i32, Expand);
809 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i32, Expand);
810 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i32, Expand);
811 // Mark ATOMIC_LOAD and ATOMIC_STORE custom so we can handle the
812 // Unordered/Monotonic case.
813 setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Custom);
814 setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Custom);
817 setOperationAction(ISD::PREFETCH, MVT::Other, Custom);
819 // Requires SXTB/SXTH, available on v6 and up in both ARM and Thumb modes.
820 if (!Subtarget->hasV6Ops()) {
821 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
822 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
824 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
826 if (!Subtarget->useSoftFloat() && Subtarget->hasVFP2() &&
827 !Subtarget->isThumb1Only()) {
828 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
829 // iff target supports vfp2.
830 setOperationAction(ISD::BITCAST, MVT::i64, Custom);
831 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
834 // We want to custom lower some of our intrinsics.
835 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
836 if (Subtarget->isTargetDarwin()) {
837 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
838 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
839 setLibcallName(RTLIB::UNWIND_RESUME, "_Unwind_SjLj_Resume");
842 setOperationAction(ISD::SETCC, MVT::i32, Expand);
843 setOperationAction(ISD::SETCC, MVT::f32, Expand);
844 setOperationAction(ISD::SETCC, MVT::f64, Expand);
845 setOperationAction(ISD::SELECT, MVT::i32, Custom);
846 setOperationAction(ISD::SELECT, MVT::f32, Custom);
847 setOperationAction(ISD::SELECT, MVT::f64, Custom);
848 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
849 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
850 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
852 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
853 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
854 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
855 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
856 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
858 // We don't support sin/cos/fmod/copysign/pow
859 setOperationAction(ISD::FSIN, MVT::f64, Expand);
860 setOperationAction(ISD::FSIN, MVT::f32, Expand);
861 setOperationAction(ISD::FCOS, MVT::f32, Expand);
862 setOperationAction(ISD::FCOS, MVT::f64, Expand);
863 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
864 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
865 setOperationAction(ISD::FREM, MVT::f64, Expand);
866 setOperationAction(ISD::FREM, MVT::f32, Expand);
867 if (!Subtarget->useSoftFloat() && Subtarget->hasVFP2() &&
868 !Subtarget->isThumb1Only()) {
869 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
870 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
872 setOperationAction(ISD::FPOW, MVT::f64, Expand);
873 setOperationAction(ISD::FPOW, MVT::f32, Expand);
875 if (!Subtarget->hasVFP4()) {
876 setOperationAction(ISD::FMA, MVT::f64, Expand);
877 setOperationAction(ISD::FMA, MVT::f32, Expand);
880 // Various VFP goodness
881 if (!Subtarget->useSoftFloat() && !Subtarget->isThumb1Only()) {
882 // FP-ARMv8 adds f64 <-> f16 conversion. Before that it should be expanded.
883 if (!Subtarget->hasFPARMv8() || Subtarget->isFPOnlySP()) {
884 setOperationAction(ISD::FP16_TO_FP, MVT::f64, Expand);
885 setOperationAction(ISD::FP_TO_FP16, MVT::f64, Expand);
888 // fp16 is a special v7 extension that adds f16 <-> f32 conversions.
889 if (!Subtarget->hasFP16()) {
890 setOperationAction(ISD::FP16_TO_FP, MVT::f32, Expand);
891 setOperationAction(ISD::FP_TO_FP16, MVT::f32, Expand);
895 // Combine sin / cos into one node or libcall if possible.
896 if (Subtarget->hasSinCos()) {
897 setLibcallName(RTLIB::SINCOS_F32, "sincosf");
898 setLibcallName(RTLIB::SINCOS_F64, "sincos");
899 if (Subtarget->getTargetTriple().isiOS()) {
900 // For iOS, we don't want to the normal expansion of a libcall to
901 // sincos. We want to issue a libcall to __sincos_stret.
902 setOperationAction(ISD::FSINCOS, MVT::f64, Custom);
903 setOperationAction(ISD::FSINCOS, MVT::f32, Custom);
907 // FP-ARMv8 implements a lot of rounding-like FP operations.
908 if (Subtarget->hasFPARMv8()) {
909 setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
910 setOperationAction(ISD::FCEIL, MVT::f32, Legal);
911 setOperationAction(ISD::FROUND, MVT::f32, Legal);
912 setOperationAction(ISD::FTRUNC, MVT::f32, Legal);
913 setOperationAction(ISD::FNEARBYINT, MVT::f32, Legal);
914 setOperationAction(ISD::FRINT, MVT::f32, Legal);
915 if (!Subtarget->isFPOnlySP()) {
916 setOperationAction(ISD::FFLOOR, MVT::f64, Legal);
917 setOperationAction(ISD::FCEIL, MVT::f64, Legal);
918 setOperationAction(ISD::FROUND, MVT::f64, Legal);
919 setOperationAction(ISD::FTRUNC, MVT::f64, Legal);
920 setOperationAction(ISD::FNEARBYINT, MVT::f64, Legal);
921 setOperationAction(ISD::FRINT, MVT::f64, Legal);
924 // We have target-specific dag combine patterns for the following nodes:
925 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
926 setTargetDAGCombine(ISD::ADD);
927 setTargetDAGCombine(ISD::SUB);
928 setTargetDAGCombine(ISD::MUL);
929 setTargetDAGCombine(ISD::AND);
930 setTargetDAGCombine(ISD::OR);
931 setTargetDAGCombine(ISD::XOR);
933 if (Subtarget->hasV6Ops())
934 setTargetDAGCombine(ISD::SRL);
936 setStackPointerRegisterToSaveRestore(ARM::SP);
938 if (Subtarget->useSoftFloat() || Subtarget->isThumb1Only() ||
939 !Subtarget->hasVFP2())
940 setSchedulingPreference(Sched::RegPressure);
942 setSchedulingPreference(Sched::Hybrid);
944 //// temporary - rewrite interface to use type
945 MaxStoresPerMemset = 8;
946 MaxStoresPerMemsetOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
947 MaxStoresPerMemcpy = 4; // For @llvm.memcpy -> sequence of stores
948 MaxStoresPerMemcpyOptSize = Subtarget->isTargetDarwin() ? 4 : 2;
949 MaxStoresPerMemmove = 4; // For @llvm.memmove -> sequence of stores
950 MaxStoresPerMemmoveOptSize = Subtarget->isTargetDarwin() ? 4 : 2;
952 // On ARM arguments smaller than 4 bytes are extended, so all arguments
953 // are at least 4 bytes aligned.
954 setMinStackArgumentAlignment(4);
956 // Prefer likely predicted branches to selects on out-of-order cores.
957 PredictableSelectIsExpensive = Subtarget->isLikeA9();
959 setMinFunctionAlignment(Subtarget->isThumb() ? 1 : 2);
962 bool ARMTargetLowering::useSoftFloat() const {
963 return Subtarget->useSoftFloat();
966 // FIXME: It might make sense to define the representative register class as the
967 // nearest super-register that has a non-null superset. For example, DPR_VFP2 is
968 // a super-register of SPR, and DPR is a superset if DPR_VFP2. Consequently,
969 // SPR's representative would be DPR_VFP2. This should work well if register
970 // pressure tracking were modified such that a register use would increment the
971 // pressure of the register class's representative and all of it's super
972 // classes' representatives transitively. We have not implemented this because
973 // of the difficulty prior to coalescing of modeling operand register classes
974 // due to the common occurrence of cross class copies and subregister insertions
976 std::pair<const TargetRegisterClass *, uint8_t>
977 ARMTargetLowering::findRepresentativeClass(const TargetRegisterInfo *TRI,
979 const TargetRegisterClass *RRC = nullptr;
981 switch (VT.SimpleTy) {
983 return TargetLowering::findRepresentativeClass(TRI, VT);
984 // Use DPR as representative register class for all floating point
985 // and vector types. Since there are 32 SPR registers and 32 DPR registers so
986 // the cost is 1 for both f32 and f64.
987 case MVT::f32: case MVT::f64: case MVT::v8i8: case MVT::v4i16:
988 case MVT::v2i32: case MVT::v1i64: case MVT::v2f32:
989 RRC = &ARM::DPRRegClass;
990 // When NEON is used for SP, only half of the register file is available
991 // because operations that define both SP and DP results will be constrained
992 // to the VFP2 class (D0-D15). We currently model this constraint prior to
993 // coalescing by double-counting the SP regs. See the FIXME above.
994 if (Subtarget->useNEONForSinglePrecisionFP())
997 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
998 case MVT::v4f32: case MVT::v2f64:
999 RRC = &ARM::DPRRegClass;
1003 RRC = &ARM::DPRRegClass;
1007 RRC = &ARM::DPRRegClass;
1011 return std::make_pair(RRC, Cost);
1014 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
1015 switch ((ARMISD::NodeType)Opcode) {
1016 case ARMISD::FIRST_NUMBER: break;
1017 case ARMISD::Wrapper: return "ARMISD::Wrapper";
1018 case ARMISD::WrapperPIC: return "ARMISD::WrapperPIC";
1019 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
1020 case ARMISD::COPY_STRUCT_BYVAL: return "ARMISD::COPY_STRUCT_BYVAL";
1021 case ARMISD::CALL: return "ARMISD::CALL";
1022 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
1023 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
1024 case ARMISD::tCALL: return "ARMISD::tCALL";
1025 case ARMISD::BRCOND: return "ARMISD::BRCOND";
1026 case ARMISD::BR_JT: return "ARMISD::BR_JT";
1027 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
1028 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
1029 case ARMISD::INTRET_FLAG: return "ARMISD::INTRET_FLAG";
1030 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
1031 case ARMISD::CMP: return "ARMISD::CMP";
1032 case ARMISD::CMN: return "ARMISD::CMN";
1033 case ARMISD::CMPZ: return "ARMISD::CMPZ";
1034 case ARMISD::CMPFP: return "ARMISD::CMPFP";
1035 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
1036 case ARMISD::BCC_i64: return "ARMISD::BCC_i64";
1037 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
1039 case ARMISD::CMOV: return "ARMISD::CMOV";
1041 case ARMISD::RBIT: return "ARMISD::RBIT";
1043 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
1044 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
1045 case ARMISD::RRX: return "ARMISD::RRX";
1047 case ARMISD::ADDC: return "ARMISD::ADDC";
1048 case ARMISD::ADDE: return "ARMISD::ADDE";
1049 case ARMISD::SUBC: return "ARMISD::SUBC";
1050 case ARMISD::SUBE: return "ARMISD::SUBE";
1052 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
1053 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
1055 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
1056 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
1058 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN";
1060 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
1062 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
1064 case ARMISD::MEMBARRIER_MCR: return "ARMISD::MEMBARRIER_MCR";
1066 case ARMISD::PRELOAD: return "ARMISD::PRELOAD";
1068 case ARMISD::WIN__CHKSTK: return "ARMISD:::WIN__CHKSTK";
1070 case ARMISD::VCEQ: return "ARMISD::VCEQ";
1071 case ARMISD::VCEQZ: return "ARMISD::VCEQZ";
1072 case ARMISD::VCGE: return "ARMISD::VCGE";
1073 case ARMISD::VCGEZ: return "ARMISD::VCGEZ";
1074 case ARMISD::VCLEZ: return "ARMISD::VCLEZ";
1075 case ARMISD::VCGEU: return "ARMISD::VCGEU";
1076 case ARMISD::VCGT: return "ARMISD::VCGT";
1077 case ARMISD::VCGTZ: return "ARMISD::VCGTZ";
1078 case ARMISD::VCLTZ: return "ARMISD::VCLTZ";
1079 case ARMISD::VCGTU: return "ARMISD::VCGTU";
1080 case ARMISD::VTST: return "ARMISD::VTST";
1082 case ARMISD::VSHL: return "ARMISD::VSHL";
1083 case ARMISD::VSHRs: return "ARMISD::VSHRs";
1084 case ARMISD::VSHRu: return "ARMISD::VSHRu";
1085 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
1086 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
1087 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
1088 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
1089 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
1090 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
1091 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
1092 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
1093 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
1094 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
1095 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
1096 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
1097 case ARMISD::VSLI: return "ARMISD::VSLI";
1098 case ARMISD::VSRI: return "ARMISD::VSRI";
1099 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
1100 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
1101 case ARMISD::VMOVIMM: return "ARMISD::VMOVIMM";
1102 case ARMISD::VMVNIMM: return "ARMISD::VMVNIMM";
1103 case ARMISD::VMOVFPIMM: return "ARMISD::VMOVFPIMM";
1104 case ARMISD::VDUP: return "ARMISD::VDUP";
1105 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
1106 case ARMISD::VEXT: return "ARMISD::VEXT";
1107 case ARMISD::VREV64: return "ARMISD::VREV64";
1108 case ARMISD::VREV32: return "ARMISD::VREV32";
1109 case ARMISD::VREV16: return "ARMISD::VREV16";
1110 case ARMISD::VZIP: return "ARMISD::VZIP";
1111 case ARMISD::VUZP: return "ARMISD::VUZP";
1112 case ARMISD::VTRN: return "ARMISD::VTRN";
1113 case ARMISD::VTBL1: return "ARMISD::VTBL1";
1114 case ARMISD::VTBL2: return "ARMISD::VTBL2";
1115 case ARMISD::VMULLs: return "ARMISD::VMULLs";
1116 case ARMISD::VMULLu: return "ARMISD::VMULLu";
1117 case ARMISD::UMLAL: return "ARMISD::UMLAL";
1118 case ARMISD::SMLAL: return "ARMISD::SMLAL";
1119 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
1120 case ARMISD::FMAX: return "ARMISD::FMAX";
1121 case ARMISD::FMIN: return "ARMISD::FMIN";
1122 case ARMISD::VMAXNM: return "ARMISD::VMAX";
1123 case ARMISD::VMINNM: return "ARMISD::VMIN";
1124 case ARMISD::BFI: return "ARMISD::BFI";
1125 case ARMISD::VORRIMM: return "ARMISD::VORRIMM";
1126 case ARMISD::VBICIMM: return "ARMISD::VBICIMM";
1127 case ARMISD::VBSL: return "ARMISD::VBSL";
1128 case ARMISD::MCOPY: return "ARMISD::MCOPY";
1129 case ARMISD::VLD2DUP: return "ARMISD::VLD2DUP";
1130 case ARMISD::VLD3DUP: return "ARMISD::VLD3DUP";
1131 case ARMISD::VLD4DUP: return "ARMISD::VLD4DUP";
1132 case ARMISD::VLD1_UPD: return "ARMISD::VLD1_UPD";
1133 case ARMISD::VLD2_UPD: return "ARMISD::VLD2_UPD";
1134 case ARMISD::VLD3_UPD: return "ARMISD::VLD3_UPD";
1135 case ARMISD::VLD4_UPD: return "ARMISD::VLD4_UPD";
1136 case ARMISD::VLD2LN_UPD: return "ARMISD::VLD2LN_UPD";
1137 case ARMISD::VLD3LN_UPD: return "ARMISD::VLD3LN_UPD";
1138 case ARMISD::VLD4LN_UPD: return "ARMISD::VLD4LN_UPD";
1139 case ARMISD::VLD2DUP_UPD: return "ARMISD::VLD2DUP_UPD";
1140 case ARMISD::VLD3DUP_UPD: return "ARMISD::VLD3DUP_UPD";
1141 case ARMISD::VLD4DUP_UPD: return "ARMISD::VLD4DUP_UPD";
1142 case ARMISD::VST1_UPD: return "ARMISD::VST1_UPD";
1143 case ARMISD::VST2_UPD: return "ARMISD::VST2_UPD";
1144 case ARMISD::VST3_UPD: return "ARMISD::VST3_UPD";
1145 case ARMISD::VST4_UPD: return "ARMISD::VST4_UPD";
1146 case ARMISD::VST2LN_UPD: return "ARMISD::VST2LN_UPD";
1147 case ARMISD::VST3LN_UPD: return "ARMISD::VST3LN_UPD";
1148 case ARMISD::VST4LN_UPD: return "ARMISD::VST4LN_UPD";
1153 EVT ARMTargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
1154 if (!VT.isVector()) return getPointerTy();
1155 return VT.changeVectorElementTypeToInteger();
1158 /// getRegClassFor - Return the register class that should be used for the
1159 /// specified value type.
1160 const TargetRegisterClass *ARMTargetLowering::getRegClassFor(MVT VT) const {
1161 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
1162 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
1163 // load / store 4 to 8 consecutive D registers.
1164 if (Subtarget->hasNEON()) {
1165 if (VT == MVT::v4i64)
1166 return &ARM::QQPRRegClass;
1167 if (VT == MVT::v8i64)
1168 return &ARM::QQQQPRRegClass;
1170 return TargetLowering::getRegClassFor(VT);
1173 // memcpy, and other memory intrinsics, typically tries to use LDM/STM if the
1174 // source/dest is aligned and the copy size is large enough. We therefore want
1175 // to align such objects passed to memory intrinsics.
1176 bool ARMTargetLowering::shouldAlignPointerArgs(CallInst *CI, unsigned &MinSize,
1177 unsigned &PrefAlign) const {
1178 if (!isa<MemIntrinsic>(CI))
1181 // On ARM11 onwards (excluding M class) 8-byte aligned LDM is typically 1
1182 // cycle faster than 4-byte aligned LDM.
1183 PrefAlign = (Subtarget->hasV6Ops() && !Subtarget->isMClass() ? 8 : 4);
1187 // Create a fast isel object.
1189 ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
1190 const TargetLibraryInfo *libInfo) const {
1191 return ARM::createFastISel(funcInfo, libInfo);
1194 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
1195 unsigned NumVals = N->getNumValues();
1197 return Sched::RegPressure;
1199 for (unsigned i = 0; i != NumVals; ++i) {
1200 EVT VT = N->getValueType(i);
1201 if (VT == MVT::Glue || VT == MVT::Other)
1203 if (VT.isFloatingPoint() || VT.isVector())
1207 if (!N->isMachineOpcode())
1208 return Sched::RegPressure;
1210 // Load are scheduled for latency even if there instruction itinerary
1211 // is not available.
1212 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
1213 const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
1215 if (MCID.getNumDefs() == 0)
1216 return Sched::RegPressure;
1217 if (!Itins->isEmpty() &&
1218 Itins->getOperandCycle(MCID.getSchedClass(), 0) > 2)
1221 return Sched::RegPressure;
1224 //===----------------------------------------------------------------------===//
1226 //===----------------------------------------------------------------------===//
1228 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
1229 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
1231 default: llvm_unreachable("Unknown condition code!");
1232 case ISD::SETNE: return ARMCC::NE;
1233 case ISD::SETEQ: return ARMCC::EQ;
1234 case ISD::SETGT: return ARMCC::GT;
1235 case ISD::SETGE: return ARMCC::GE;
1236 case ISD::SETLT: return ARMCC::LT;
1237 case ISD::SETLE: return ARMCC::LE;
1238 case ISD::SETUGT: return ARMCC::HI;
1239 case ISD::SETUGE: return ARMCC::HS;
1240 case ISD::SETULT: return ARMCC::LO;
1241 case ISD::SETULE: return ARMCC::LS;
1245 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
1246 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
1247 ARMCC::CondCodes &CondCode2) {
1248 CondCode2 = ARMCC::AL;
1250 default: llvm_unreachable("Unknown FP condition!");
1252 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
1254 case ISD::SETOGT: CondCode = ARMCC::GT; break;
1256 case ISD::SETOGE: CondCode = ARMCC::GE; break;
1257 case ISD::SETOLT: CondCode = ARMCC::MI; break;
1258 case ISD::SETOLE: CondCode = ARMCC::LS; break;
1259 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
1260 case ISD::SETO: CondCode = ARMCC::VC; break;
1261 case ISD::SETUO: CondCode = ARMCC::VS; break;
1262 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
1263 case ISD::SETUGT: CondCode = ARMCC::HI; break;
1264 case ISD::SETUGE: CondCode = ARMCC::PL; break;
1266 case ISD::SETULT: CondCode = ARMCC::LT; break;
1268 case ISD::SETULE: CondCode = ARMCC::LE; break;
1270 case ISD::SETUNE: CondCode = ARMCC::NE; break;
1274 //===----------------------------------------------------------------------===//
1275 // Calling Convention Implementation
1276 //===----------------------------------------------------------------------===//
1278 #include "ARMGenCallingConv.inc"
1280 /// getEffectiveCallingConv - Get the effective calling convention, taking into
1281 /// account presence of floating point hardware and calling convention
1282 /// limitations, such as support for variadic functions.
1284 ARMTargetLowering::getEffectiveCallingConv(CallingConv::ID CC,
1285 bool isVarArg) const {
1288 llvm_unreachable("Unsupported calling convention");
1289 case CallingConv::ARM_AAPCS:
1290 case CallingConv::ARM_APCS:
1291 case CallingConv::GHC:
1293 case CallingConv::ARM_AAPCS_VFP:
1294 return isVarArg ? CallingConv::ARM_AAPCS : CallingConv::ARM_AAPCS_VFP;
1295 case CallingConv::C:
1296 if (!Subtarget->isAAPCS_ABI())
1297 return CallingConv::ARM_APCS;
1298 else if (Subtarget->hasVFP2() && !Subtarget->isThumb1Only() &&
1299 getTargetMachine().Options.FloatABIType == FloatABI::Hard &&
1301 return CallingConv::ARM_AAPCS_VFP;
1303 return CallingConv::ARM_AAPCS;
1304 case CallingConv::Fast:
1305 if (!Subtarget->isAAPCS_ABI()) {
1306 if (Subtarget->hasVFP2() && !Subtarget->isThumb1Only() && !isVarArg)
1307 return CallingConv::Fast;
1308 return CallingConv::ARM_APCS;
1309 } else if (Subtarget->hasVFP2() && !Subtarget->isThumb1Only() && !isVarArg)
1310 return CallingConv::ARM_AAPCS_VFP;
1312 return CallingConv::ARM_AAPCS;
1316 /// CCAssignFnForNode - Selects the correct CCAssignFn for the given
1317 /// CallingConvention.
1318 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
1320 bool isVarArg) const {
1321 switch (getEffectiveCallingConv(CC, isVarArg)) {
1323 llvm_unreachable("Unsupported calling convention");
1324 case CallingConv::ARM_APCS:
1325 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1326 case CallingConv::ARM_AAPCS:
1327 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1328 case CallingConv::ARM_AAPCS_VFP:
1329 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1330 case CallingConv::Fast:
1331 return (Return ? RetFastCC_ARM_APCS : FastCC_ARM_APCS);
1332 case CallingConv::GHC:
1333 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS_GHC);
1337 /// LowerCallResult - Lower the result values of a call into the
1338 /// appropriate copies out of appropriate physical registers.
1340 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1341 CallingConv::ID CallConv, bool isVarArg,
1342 const SmallVectorImpl<ISD::InputArg> &Ins,
1343 SDLoc dl, SelectionDAG &DAG,
1344 SmallVectorImpl<SDValue> &InVals,
1345 bool isThisReturn, SDValue ThisVal) const {
1347 // Assign locations to each value returned by this call.
1348 SmallVector<CCValAssign, 16> RVLocs;
1349 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
1350 *DAG.getContext(), Call);
1351 CCInfo.AnalyzeCallResult(Ins,
1352 CCAssignFnForNode(CallConv, /* Return*/ true,
1355 // Copy all of the result registers out of their specified physreg.
1356 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1357 CCValAssign VA = RVLocs[i];
1359 // Pass 'this' value directly from the argument to return value, to avoid
1360 // reg unit interference
1361 if (i == 0 && isThisReturn) {
1362 assert(!VA.needsCustom() && VA.getLocVT() == MVT::i32 &&
1363 "unexpected return calling convention register assignment");
1364 InVals.push_back(ThisVal);
1369 if (VA.needsCustom()) {
1370 // Handle f64 or half of a v2f64.
1371 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1373 Chain = Lo.getValue(1);
1374 InFlag = Lo.getValue(2);
1375 VA = RVLocs[++i]; // skip ahead to next loc
1376 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1378 Chain = Hi.getValue(1);
1379 InFlag = Hi.getValue(2);
1380 if (!Subtarget->isLittle())
1382 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1384 if (VA.getLocVT() == MVT::v2f64) {
1385 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1386 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1387 DAG.getConstant(0, dl, MVT::i32));
1389 VA = RVLocs[++i]; // skip ahead to next loc
1390 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1391 Chain = Lo.getValue(1);
1392 InFlag = Lo.getValue(2);
1393 VA = RVLocs[++i]; // skip ahead to next loc
1394 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1395 Chain = Hi.getValue(1);
1396 InFlag = Hi.getValue(2);
1397 if (!Subtarget->isLittle())
1399 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1400 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1401 DAG.getConstant(1, dl, MVT::i32));
1404 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
1406 Chain = Val.getValue(1);
1407 InFlag = Val.getValue(2);
1410 switch (VA.getLocInfo()) {
1411 default: llvm_unreachable("Unknown loc info!");
1412 case CCValAssign::Full: break;
1413 case CCValAssign::BCvt:
1414 Val = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), Val);
1418 InVals.push_back(Val);
1424 /// LowerMemOpCallTo - Store the argument to the stack.
1426 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
1427 SDValue StackPtr, SDValue Arg,
1428 SDLoc dl, SelectionDAG &DAG,
1429 const CCValAssign &VA,
1430 ISD::ArgFlagsTy Flags) const {
1431 unsigned LocMemOffset = VA.getLocMemOffset();
1432 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset, dl);
1433 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1434 return DAG.getStore(Chain, dl, Arg, PtrOff,
1435 MachinePointerInfo::getStack(LocMemOffset),
1439 void ARMTargetLowering::PassF64ArgInRegs(SDLoc dl, SelectionDAG &DAG,
1440 SDValue Chain, SDValue &Arg,
1441 RegsToPassVector &RegsToPass,
1442 CCValAssign &VA, CCValAssign &NextVA,
1444 SmallVectorImpl<SDValue> &MemOpChains,
1445 ISD::ArgFlagsTy Flags) const {
1447 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1448 DAG.getVTList(MVT::i32, MVT::i32), Arg);
1449 unsigned id = Subtarget->isLittle() ? 0 : 1;
1450 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd.getValue(id)));
1452 if (NextVA.isRegLoc())
1453 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1-id)));
1455 assert(NextVA.isMemLoc());
1456 if (!StackPtr.getNode())
1457 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1459 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1-id),
1465 /// LowerCall - Lowering a call into a callseq_start <-
1466 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
1469 ARMTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
1470 SmallVectorImpl<SDValue> &InVals) const {
1471 SelectionDAG &DAG = CLI.DAG;
1473 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
1474 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
1475 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
1476 SDValue Chain = CLI.Chain;
1477 SDValue Callee = CLI.Callee;
1478 bool &isTailCall = CLI.IsTailCall;
1479 CallingConv::ID CallConv = CLI.CallConv;
1480 bool doesNotRet = CLI.DoesNotReturn;
1481 bool isVarArg = CLI.IsVarArg;
1483 MachineFunction &MF = DAG.getMachineFunction();
1484 bool isStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1485 bool isThisReturn = false;
1486 bool isSibCall = false;
1488 // Disable tail calls if they're not supported.
1489 if (!Subtarget->supportsTailCall() || MF.getTarget().Options.DisableTailCalls)
1493 // Check if it's really possible to do a tail call.
1494 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1495 isVarArg, isStructRet, MF.getFunction()->hasStructRetAttr(),
1496 Outs, OutVals, Ins, DAG);
1497 if (!isTailCall && CLI.CS && CLI.CS->isMustTailCall())
1498 report_fatal_error("failed to perform tail call elimination on a call "
1499 "site marked musttail");
1500 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1501 // detected sibcalls.
1508 // Analyze operands of the call, assigning locations to each operand.
1509 SmallVector<CCValAssign, 16> ArgLocs;
1510 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs,
1511 *DAG.getContext(), Call);
1512 CCInfo.AnalyzeCallOperands(Outs,
1513 CCAssignFnForNode(CallConv, /* Return*/ false,
1516 // Get a count of how many bytes are to be pushed on the stack.
1517 unsigned NumBytes = CCInfo.getNextStackOffset();
1519 // For tail calls, memory operands are available in our caller's stack.
1523 // Adjust the stack pointer for the new arguments...
1524 // These operations are automatically eliminated by the prolog/epilog pass
1526 Chain = DAG.getCALLSEQ_START(Chain,
1527 DAG.getIntPtrConstant(NumBytes, dl, true), dl);
1529 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1531 RegsToPassVector RegsToPass;
1532 SmallVector<SDValue, 8> MemOpChains;
1534 // Walk the register/memloc assignments, inserting copies/loads. In the case
1535 // of tail call optimization, arguments are handled later.
1536 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1538 ++i, ++realArgIdx) {
1539 CCValAssign &VA = ArgLocs[i];
1540 SDValue Arg = OutVals[realArgIdx];
1541 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1542 bool isByVal = Flags.isByVal();
1544 // Promote the value if needed.
1545 switch (VA.getLocInfo()) {
1546 default: llvm_unreachable("Unknown loc info!");
1547 case CCValAssign::Full: break;
1548 case CCValAssign::SExt:
1549 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1551 case CCValAssign::ZExt:
1552 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1554 case CCValAssign::AExt:
1555 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1557 case CCValAssign::BCvt:
1558 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1562 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1563 if (VA.needsCustom()) {
1564 if (VA.getLocVT() == MVT::v2f64) {
1565 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1566 DAG.getConstant(0, dl, MVT::i32));
1567 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1568 DAG.getConstant(1, dl, MVT::i32));
1570 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1571 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1573 VA = ArgLocs[++i]; // skip ahead to next loc
1574 if (VA.isRegLoc()) {
1575 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1576 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1578 assert(VA.isMemLoc());
1580 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1581 dl, DAG, VA, Flags));
1584 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1585 StackPtr, MemOpChains, Flags);
1587 } else if (VA.isRegLoc()) {
1588 if (realArgIdx == 0 && Flags.isReturned() && Outs[0].VT == MVT::i32) {
1589 assert(VA.getLocVT() == MVT::i32 &&
1590 "unexpected calling convention register assignment");
1591 assert(!Ins.empty() && Ins[0].VT == MVT::i32 &&
1592 "unexpected use of 'returned'");
1593 isThisReturn = true;
1595 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1596 } else if (isByVal) {
1597 assert(VA.isMemLoc());
1598 unsigned offset = 0;
1600 // True if this byval aggregate will be split between registers
1602 unsigned ByValArgsCount = CCInfo.getInRegsParamsCount();
1603 unsigned CurByValIdx = CCInfo.getInRegsParamsProcessed();
1605 if (CurByValIdx < ByValArgsCount) {
1607 unsigned RegBegin, RegEnd;
1608 CCInfo.getInRegsParamInfo(CurByValIdx, RegBegin, RegEnd);
1610 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1612 for (i = 0, j = RegBegin; j < RegEnd; i++, j++) {
1613 SDValue Const = DAG.getConstant(4*i, dl, MVT::i32);
1614 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
1615 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
1616 MachinePointerInfo(),
1617 false, false, false,
1618 DAG.InferPtrAlignment(AddArg));
1619 MemOpChains.push_back(Load.getValue(1));
1620 RegsToPass.push_back(std::make_pair(j, Load));
1623 // If parameter size outsides register area, "offset" value
1624 // helps us to calculate stack slot for remained part properly.
1625 offset = RegEnd - RegBegin;
1627 CCInfo.nextInRegsParam();
1630 if (Flags.getByValSize() > 4*offset) {
1631 unsigned LocMemOffset = VA.getLocMemOffset();
1632 SDValue StkPtrOff = DAG.getIntPtrConstant(LocMemOffset, dl);
1633 SDValue Dst = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr,
1635 SDValue SrcOffset = DAG.getIntPtrConstant(4*offset, dl);
1636 SDValue Src = DAG.getNode(ISD::ADD, dl, getPointerTy(), Arg, SrcOffset);
1637 SDValue SizeNode = DAG.getConstant(Flags.getByValSize() - 4*offset, dl,
1639 SDValue AlignNode = DAG.getConstant(Flags.getByValAlign(), dl,
1642 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
1643 SDValue Ops[] = { Chain, Dst, Src, SizeNode, AlignNode};
1644 MemOpChains.push_back(DAG.getNode(ARMISD::COPY_STRUCT_BYVAL, dl, VTs,
1647 } else if (!isSibCall) {
1648 assert(VA.isMemLoc());
1650 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1651 dl, DAG, VA, Flags));
1655 if (!MemOpChains.empty())
1656 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains);
1658 // Build a sequence of copy-to-reg nodes chained together with token chain
1659 // and flag operands which copy the outgoing args into the appropriate regs.
1661 // Tail call byval lowering might overwrite argument registers so in case of
1662 // tail call optimization the copies to registers are lowered later.
1664 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1665 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1666 RegsToPass[i].second, InFlag);
1667 InFlag = Chain.getValue(1);
1670 // For tail calls lower the arguments to the 'real' stack slot.
1672 // Force all the incoming stack arguments to be loaded from the stack
1673 // before any new outgoing arguments are stored to the stack, because the
1674 // outgoing stack slots may alias the incoming argument stack slots, and
1675 // the alias isn't otherwise explicit. This is slightly more conservative
1676 // than necessary, because it means that each store effectively depends
1677 // on every argument instead of just those arguments it would clobber.
1679 // Do not flag preceding copytoreg stuff together with the following stuff.
1681 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1682 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1683 RegsToPass[i].second, InFlag);
1684 InFlag = Chain.getValue(1);
1689 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1690 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1691 // node so that legalize doesn't hack it.
1692 bool isDirect = false;
1693 bool isARMFunc = false;
1694 bool isLocalARMFunc = false;
1695 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1697 if (EnableARMLongCalls) {
1698 assert((Subtarget->isTargetWindows() ||
1699 getTargetMachine().getRelocationModel() == Reloc::Static) &&
1700 "long-calls with non-static relocation model!");
1701 // Handle a global address or an external symbol. If it's not one of
1702 // those, the target's already in a register, so we don't need to do
1704 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1705 const GlobalValue *GV = G->getGlobal();
1706 // Create a constant pool entry for the callee address
1707 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1708 ARMConstantPoolValue *CPV =
1709 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 0);
1711 // Get the address of the callee into a register
1712 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1713 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1714 Callee = DAG.getLoad(getPointerTy(), dl,
1715 DAG.getEntryNode(), CPAddr,
1716 MachinePointerInfo::getConstantPool(),
1717 false, false, false, 0);
1718 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1719 const char *Sym = S->getSymbol();
1721 // Create a constant pool entry for the callee address
1722 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1723 ARMConstantPoolValue *CPV =
1724 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1725 ARMPCLabelIndex, 0);
1726 // Get the address of the callee into a register
1727 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1728 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1729 Callee = DAG.getLoad(getPointerTy(), dl,
1730 DAG.getEntryNode(), CPAddr,
1731 MachinePointerInfo::getConstantPool(),
1732 false, false, false, 0);
1734 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1735 const GlobalValue *GV = G->getGlobal();
1737 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1738 bool isStub = (isExt && Subtarget->isTargetMachO()) &&
1739 getTargetMachine().getRelocationModel() != Reloc::Static;
1740 isARMFunc = !Subtarget->isThumb() || (isStub && !Subtarget->isMClass());
1741 // ARM call to a local ARM function is predicable.
1742 isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking);
1743 // tBX takes a register source operand.
1744 if (isStub && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1745 assert(Subtarget->isTargetMachO() && "WrapperPIC use on non-MachO?");
1746 Callee = DAG.getNode(ARMISD::WrapperPIC, dl, getPointerTy(),
1747 DAG.getTargetGlobalAddress(GV, dl, getPointerTy(),
1748 0, ARMII::MO_NONLAZY));
1749 Callee = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Callee,
1750 MachinePointerInfo::getGOT(), false, false, true, 0);
1751 } else if (Subtarget->isTargetCOFF()) {
1752 assert(Subtarget->isTargetWindows() &&
1753 "Windows is the only supported COFF target");
1754 unsigned TargetFlags = GV->hasDLLImportStorageClass()
1755 ? ARMII::MO_DLLIMPORT
1756 : ARMII::MO_NO_FLAG;
1757 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), /*Offset=*/0,
1759 if (GV->hasDLLImportStorageClass())
1760 Callee = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
1761 DAG.getNode(ARMISD::Wrapper, dl, getPointerTy(),
1762 Callee), MachinePointerInfo::getGOT(),
1763 false, false, false, 0);
1765 // On ELF targets for PIC code, direct calls should go through the PLT
1766 unsigned OpFlags = 0;
1767 if (Subtarget->isTargetELF() &&
1768 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1769 OpFlags = ARMII::MO_PLT;
1770 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
1772 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1774 bool isStub = Subtarget->isTargetMachO() &&
1775 getTargetMachine().getRelocationModel() != Reloc::Static;
1776 isARMFunc = !Subtarget->isThumb() || (isStub && !Subtarget->isMClass());
1777 // tBX takes a register source operand.
1778 const char *Sym = S->getSymbol();
1779 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1780 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1781 ARMConstantPoolValue *CPV =
1782 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1783 ARMPCLabelIndex, 4);
1784 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1785 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1786 Callee = DAG.getLoad(getPointerTy(), dl,
1787 DAG.getEntryNode(), CPAddr,
1788 MachinePointerInfo::getConstantPool(),
1789 false, false, false, 0);
1790 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, dl, MVT::i32);
1791 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1792 getPointerTy(), Callee, PICLabel);
1794 unsigned OpFlags = 0;
1795 // On ELF targets for PIC code, direct calls should go through the PLT
1796 if (Subtarget->isTargetELF() &&
1797 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1798 OpFlags = ARMII::MO_PLT;
1799 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlags);
1803 // FIXME: handle tail calls differently.
1805 bool HasMinSizeAttr = MF.getFunction()->hasFnAttribute(Attribute::MinSize);
1806 if (Subtarget->isThumb()) {
1807 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1808 CallOpc = ARMISD::CALL_NOLINK;
1810 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1812 if (!isDirect && !Subtarget->hasV5TOps())
1813 CallOpc = ARMISD::CALL_NOLINK;
1814 else if (doesNotRet && isDirect && Subtarget->hasRAS() &&
1815 // Emit regular call when code size is the priority
1817 // "mov lr, pc; b _foo" to avoid confusing the RSP
1818 CallOpc = ARMISD::CALL_NOLINK;
1820 CallOpc = isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL;
1823 std::vector<SDValue> Ops;
1824 Ops.push_back(Chain);
1825 Ops.push_back(Callee);
1827 // Add argument registers to the end of the list so that they are known live
1829 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1830 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1831 RegsToPass[i].second.getValueType()));
1833 // Add a register mask operand representing the call-preserved registers.
1835 const uint32_t *Mask;
1836 const ARMBaseRegisterInfo *ARI = Subtarget->getRegisterInfo();
1838 // For 'this' returns, use the R0-preserving mask if applicable
1839 Mask = ARI->getThisReturnPreservedMask(MF, CallConv);
1841 // Set isThisReturn to false if the calling convention is not one that
1842 // allows 'returned' to be modeled in this way, so LowerCallResult does
1843 // not try to pass 'this' straight through
1844 isThisReturn = false;
1845 Mask = ARI->getCallPreservedMask(MF, CallConv);
1848 Mask = ARI->getCallPreservedMask(MF, CallConv);
1850 assert(Mask && "Missing call preserved mask for calling convention");
1851 Ops.push_back(DAG.getRegisterMask(Mask));
1854 if (InFlag.getNode())
1855 Ops.push_back(InFlag);
1857 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
1859 MF.getFrameInfo()->setHasTailCall();
1860 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, Ops);
1863 // Returns a chain and a flag for retval copy to use.
1864 Chain = DAG.getNode(CallOpc, dl, NodeTys, Ops);
1865 InFlag = Chain.getValue(1);
1867 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, dl, true),
1868 DAG.getIntPtrConstant(0, dl, true), InFlag, dl);
1870 InFlag = Chain.getValue(1);
1872 // Handle result values, copying them out of physregs into vregs that we
1874 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins, dl, DAG,
1875 InVals, isThisReturn,
1876 isThisReturn ? OutVals[0] : SDValue());
1879 /// HandleByVal - Every parameter *after* a byval parameter is passed
1880 /// on the stack. Remember the next parameter register to allocate,
1881 /// and then confiscate the rest of the parameter registers to insure
1883 void ARMTargetLowering::HandleByVal(CCState *State, unsigned &Size,
1884 unsigned Align) const {
1885 assert((State->getCallOrPrologue() == Prologue ||
1886 State->getCallOrPrologue() == Call) &&
1887 "unhandled ParmContext");
1889 // Byval (as with any stack) slots are always at least 4 byte aligned.
1890 Align = std::max(Align, 4U);
1892 unsigned Reg = State->AllocateReg(GPRArgRegs);
1896 unsigned AlignInRegs = Align / 4;
1897 unsigned Waste = (ARM::R4 - Reg) % AlignInRegs;
1898 for (unsigned i = 0; i < Waste; ++i)
1899 Reg = State->AllocateReg(GPRArgRegs);
1904 unsigned Excess = 4 * (ARM::R4 - Reg);
1906 // Special case when NSAA != SP and parameter size greater than size of
1907 // all remained GPR regs. In that case we can't split parameter, we must
1908 // send it to stack. We also must set NCRN to R4, so waste all
1909 // remained registers.
1910 const unsigned NSAAOffset = State->getNextStackOffset();
1911 if (NSAAOffset != 0 && Size > Excess) {
1912 while (State->AllocateReg(GPRArgRegs))
1917 // First register for byval parameter is the first register that wasn't
1918 // allocated before this method call, so it would be "reg".
1919 // If parameter is small enough to be saved in range [reg, r4), then
1920 // the end (first after last) register would be reg + param-size-in-regs,
1921 // else parameter would be splitted between registers and stack,
1922 // end register would be r4 in this case.
1923 unsigned ByValRegBegin = Reg;
1924 unsigned ByValRegEnd = std::min<unsigned>(Reg + Size / 4, ARM::R4);
1925 State->addInRegsParamInfo(ByValRegBegin, ByValRegEnd);
1926 // Note, first register is allocated in the beginning of function already,
1927 // allocate remained amount of registers we need.
1928 for (unsigned i = Reg + 1; i != ByValRegEnd; ++i)
1929 State->AllocateReg(GPRArgRegs);
1930 // A byval parameter that is split between registers and memory needs its
1931 // size truncated here.
1932 // In the case where the entire structure fits in registers, we set the
1933 // size in memory to zero.
1934 Size = std::max<int>(Size - Excess, 0);
1937 /// MatchingStackOffset - Return true if the given stack call argument is
1938 /// already available in the same position (relatively) of the caller's
1939 /// incoming argument stack.
1941 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1942 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1943 const TargetInstrInfo *TII) {
1944 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1946 if (Arg.getOpcode() == ISD::CopyFromReg) {
1947 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1948 if (!TargetRegisterInfo::isVirtualRegister(VR))
1950 MachineInstr *Def = MRI->getVRegDef(VR);
1953 if (!Flags.isByVal()) {
1954 if (!TII->isLoadFromStackSlot(Def, FI))
1959 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
1960 if (Flags.isByVal())
1961 // ByVal argument is passed in as a pointer but it's now being
1962 // dereferenced. e.g.
1963 // define @foo(%struct.X* %A) {
1964 // tail call @bar(%struct.X* byval %A)
1967 SDValue Ptr = Ld->getBasePtr();
1968 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
1971 FI = FINode->getIndex();
1975 assert(FI != INT_MAX);
1976 if (!MFI->isFixedObjectIndex(FI))
1978 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
1981 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
1982 /// for tail call optimization. Targets which want to do tail call
1983 /// optimization should implement this function.
1985 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1986 CallingConv::ID CalleeCC,
1988 bool isCalleeStructRet,
1989 bool isCallerStructRet,
1990 const SmallVectorImpl<ISD::OutputArg> &Outs,
1991 const SmallVectorImpl<SDValue> &OutVals,
1992 const SmallVectorImpl<ISD::InputArg> &Ins,
1993 SelectionDAG& DAG) const {
1994 const Function *CallerF = DAG.getMachineFunction().getFunction();
1995 CallingConv::ID CallerCC = CallerF->getCallingConv();
1996 bool CCMatch = CallerCC == CalleeCC;
1998 // Look for obvious safe cases to perform tail call optimization that do not
1999 // require ABI changes. This is what gcc calls sibcall.
2001 // Do not sibcall optimize vararg calls unless the call site is not passing
2003 if (isVarArg && !Outs.empty())
2006 // Exception-handling functions need a special set of instructions to indicate
2007 // a return to the hardware. Tail-calling another function would probably
2009 if (CallerF->hasFnAttribute("interrupt"))
2012 // Also avoid sibcall optimization if either caller or callee uses struct
2013 // return semantics.
2014 if (isCalleeStructRet || isCallerStructRet)
2017 // FIXME: Completely disable sibcall for Thumb1 since ThumbRegisterInfo::
2018 // emitEpilogue is not ready for them. Thumb tail calls also use t2B, as
2019 // the Thumb1 16-bit unconditional branch doesn't have sufficient relocation
2020 // support in the assembler and linker to be used. This would need to be
2021 // fixed to fully support tail calls in Thumb1.
2023 // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take
2024 // LR. This means if we need to reload LR, it takes an extra instructions,
2025 // which outweighs the value of the tail call; but here we don't know yet
2026 // whether LR is going to be used. Probably the right approach is to
2027 // generate the tail call here and turn it back into CALL/RET in
2028 // emitEpilogue if LR is used.
2030 // Thumb1 PIC calls to external symbols use BX, so they can be tail calls,
2031 // but we need to make sure there are enough registers; the only valid
2032 // registers are the 4 used for parameters. We don't currently do this
2034 if (Subtarget->isThumb1Only())
2037 // Externally-defined functions with weak linkage should not be
2038 // tail-called on ARM when the OS does not support dynamic
2039 // pre-emption of symbols, as the AAELF spec requires normal calls
2040 // to undefined weak functions to be replaced with a NOP or jump to the
2041 // next instruction. The behaviour of branch instructions in this
2042 // situation (as used for tail calls) is implementation-defined, so we
2043 // cannot rely on the linker replacing the tail call with a return.
2044 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2045 const GlobalValue *GV = G->getGlobal();
2046 const Triple TT(getTargetMachine().getTargetTriple());
2047 if (GV->hasExternalWeakLinkage() &&
2048 (!TT.isOSWindows() || TT.isOSBinFormatELF() || TT.isOSBinFormatMachO()))
2052 // If the calling conventions do not match, then we'd better make sure the
2053 // results are returned in the same way as what the caller expects.
2055 SmallVector<CCValAssign, 16> RVLocs1;
2056 ARMCCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(), RVLocs1,
2057 *DAG.getContext(), Call);
2058 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
2060 SmallVector<CCValAssign, 16> RVLocs2;
2061 ARMCCState CCInfo2(CallerCC, false, DAG.getMachineFunction(), RVLocs2,
2062 *DAG.getContext(), Call);
2063 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
2065 if (RVLocs1.size() != RVLocs2.size())
2067 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
2068 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
2070 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
2072 if (RVLocs1[i].isRegLoc()) {
2073 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
2076 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
2082 // If Caller's vararg or byval argument has been split between registers and
2083 // stack, do not perform tail call, since part of the argument is in caller's
2085 const ARMFunctionInfo *AFI_Caller = DAG.getMachineFunction().
2086 getInfo<ARMFunctionInfo>();
2087 if (AFI_Caller->getArgRegsSaveSize())
2090 // If the callee takes no arguments then go on to check the results of the
2092 if (!Outs.empty()) {
2093 // Check if stack adjustment is needed. For now, do not do this if any
2094 // argument is passed on the stack.
2095 SmallVector<CCValAssign, 16> ArgLocs;
2096 ARMCCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
2097 *DAG.getContext(), Call);
2098 CCInfo.AnalyzeCallOperands(Outs,
2099 CCAssignFnForNode(CalleeCC, false, isVarArg));
2100 if (CCInfo.getNextStackOffset()) {
2101 MachineFunction &MF = DAG.getMachineFunction();
2103 // Check if the arguments are already laid out in the right way as
2104 // the caller's fixed stack objects.
2105 MachineFrameInfo *MFI = MF.getFrameInfo();
2106 const MachineRegisterInfo *MRI = &MF.getRegInfo();
2107 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
2108 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
2110 ++i, ++realArgIdx) {
2111 CCValAssign &VA = ArgLocs[i];
2112 EVT RegVT = VA.getLocVT();
2113 SDValue Arg = OutVals[realArgIdx];
2114 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
2115 if (VA.getLocInfo() == CCValAssign::Indirect)
2117 if (VA.needsCustom()) {
2118 // f64 and vector types are split into multiple registers or
2119 // register/stack-slot combinations. The types will not match
2120 // the registers; give up on memory f64 refs until we figure
2121 // out what to do about this.
2124 if (!ArgLocs[++i].isRegLoc())
2126 if (RegVT == MVT::v2f64) {
2127 if (!ArgLocs[++i].isRegLoc())
2129 if (!ArgLocs[++i].isRegLoc())
2132 } else if (!VA.isRegLoc()) {
2133 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
2145 ARMTargetLowering::CanLowerReturn(CallingConv::ID CallConv,
2146 MachineFunction &MF, bool isVarArg,
2147 const SmallVectorImpl<ISD::OutputArg> &Outs,
2148 LLVMContext &Context) const {
2149 SmallVector<CCValAssign, 16> RVLocs;
2150 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, Context);
2151 return CCInfo.CheckReturn(Outs, CCAssignFnForNode(CallConv, /*Return=*/true,
2155 static SDValue LowerInterruptReturn(SmallVectorImpl<SDValue> &RetOps,
2156 SDLoc DL, SelectionDAG &DAG) {
2157 const MachineFunction &MF = DAG.getMachineFunction();
2158 const Function *F = MF.getFunction();
2160 StringRef IntKind = F->getFnAttribute("interrupt").getValueAsString();
2162 // See ARM ARM v7 B1.8.3. On exception entry LR is set to a possibly offset
2163 // version of the "preferred return address". These offsets affect the return
2164 // instruction if this is a return from PL1 without hypervisor extensions.
2165 // IRQ/FIQ: +4 "subs pc, lr, #4"
2166 // SWI: 0 "subs pc, lr, #0"
2167 // ABORT: +4 "subs pc, lr, #4"
2168 // UNDEF: +4/+2 "subs pc, lr, #0"
2169 // UNDEF varies depending on where the exception came from ARM or Thumb
2170 // mode. Alongside GCC, we throw our hands up in disgust and pretend it's 0.
2173 if (IntKind == "" || IntKind == "IRQ" || IntKind == "FIQ" ||
2176 else if (IntKind == "SWI" || IntKind == "UNDEF")
2179 report_fatal_error("Unsupported interrupt attribute. If present, value "
2180 "must be one of: IRQ, FIQ, SWI, ABORT or UNDEF");
2182 RetOps.insert(RetOps.begin() + 1,
2183 DAG.getConstant(LROffset, DL, MVT::i32, false));
2185 return DAG.getNode(ARMISD::INTRET_FLAG, DL, MVT::Other, RetOps);
2189 ARMTargetLowering::LowerReturn(SDValue Chain,
2190 CallingConv::ID CallConv, bool isVarArg,
2191 const SmallVectorImpl<ISD::OutputArg> &Outs,
2192 const SmallVectorImpl<SDValue> &OutVals,
2193 SDLoc dl, SelectionDAG &DAG) const {
2195 // CCValAssign - represent the assignment of the return value to a location.
2196 SmallVector<CCValAssign, 16> RVLocs;
2198 // CCState - Info about the registers and stack slots.
2199 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
2200 *DAG.getContext(), Call);
2202 // Analyze outgoing return values.
2203 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
2207 SmallVector<SDValue, 4> RetOps;
2208 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
2209 bool isLittleEndian = Subtarget->isLittle();
2211 MachineFunction &MF = DAG.getMachineFunction();
2212 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2213 AFI->setReturnRegsCount(RVLocs.size());
2215 // Copy the result values into the output registers.
2216 for (unsigned i = 0, realRVLocIdx = 0;
2218 ++i, ++realRVLocIdx) {
2219 CCValAssign &VA = RVLocs[i];
2220 assert(VA.isRegLoc() && "Can only return in registers!");
2222 SDValue Arg = OutVals[realRVLocIdx];
2224 switch (VA.getLocInfo()) {
2225 default: llvm_unreachable("Unknown loc info!");
2226 case CCValAssign::Full: break;
2227 case CCValAssign::BCvt:
2228 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
2232 if (VA.needsCustom()) {
2233 if (VA.getLocVT() == MVT::v2f64) {
2234 // Extract the first half and return it in two registers.
2235 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
2236 DAG.getConstant(0, dl, MVT::i32));
2237 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
2238 DAG.getVTList(MVT::i32, MVT::i32), Half);
2240 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2241 HalfGPRs.getValue(isLittleEndian ? 0 : 1),
2243 Flag = Chain.getValue(1);
2244 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2245 VA = RVLocs[++i]; // skip ahead to next loc
2246 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2247 HalfGPRs.getValue(isLittleEndian ? 1 : 0),
2249 Flag = Chain.getValue(1);
2250 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2251 VA = RVLocs[++i]; // skip ahead to next loc
2253 // Extract the 2nd half and fall through to handle it as an f64 value.
2254 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
2255 DAG.getConstant(1, dl, MVT::i32));
2257 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
2259 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
2260 DAG.getVTList(MVT::i32, MVT::i32), Arg);
2261 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2262 fmrrd.getValue(isLittleEndian ? 0 : 1),
2264 Flag = Chain.getValue(1);
2265 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2266 VA = RVLocs[++i]; // skip ahead to next loc
2267 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2268 fmrrd.getValue(isLittleEndian ? 1 : 0),
2271 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
2273 // Guarantee that all emitted copies are
2274 // stuck together, avoiding something bad.
2275 Flag = Chain.getValue(1);
2276 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2279 // Update chain and glue.
2282 RetOps.push_back(Flag);
2284 // CPUs which aren't M-class use a special sequence to return from
2285 // exceptions (roughly, any instruction setting pc and cpsr simultaneously,
2286 // though we use "subs pc, lr, #N").
2288 // M-class CPUs actually use a normal return sequence with a special
2289 // (hardware-provided) value in LR, so the normal code path works.
2290 if (DAG.getMachineFunction().getFunction()->hasFnAttribute("interrupt") &&
2291 !Subtarget->isMClass()) {
2292 if (Subtarget->isThumb1Only())
2293 report_fatal_error("interrupt attribute is not supported in Thumb1");
2294 return LowerInterruptReturn(RetOps, dl, DAG);
2297 return DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, RetOps);
2300 bool ARMTargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
2301 if (N->getNumValues() != 1)
2303 if (!N->hasNUsesOfValue(1, 0))
2306 SDValue TCChain = Chain;
2307 SDNode *Copy = *N->use_begin();
2308 if (Copy->getOpcode() == ISD::CopyToReg) {
2309 // If the copy has a glue operand, we conservatively assume it isn't safe to
2310 // perform a tail call.
2311 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
2313 TCChain = Copy->getOperand(0);
2314 } else if (Copy->getOpcode() == ARMISD::VMOVRRD) {
2315 SDNode *VMov = Copy;
2316 // f64 returned in a pair of GPRs.
2317 SmallPtrSet<SDNode*, 2> Copies;
2318 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
2320 if (UI->getOpcode() != ISD::CopyToReg)
2324 if (Copies.size() > 2)
2327 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
2329 SDValue UseChain = UI->getOperand(0);
2330 if (Copies.count(UseChain.getNode()))
2334 // We are at the top of this chain.
2335 // If the copy has a glue operand, we conservatively assume it
2336 // isn't safe to perform a tail call.
2337 if (UI->getOperand(UI->getNumOperands()-1).getValueType() == MVT::Glue)
2343 } else if (Copy->getOpcode() == ISD::BITCAST) {
2344 // f32 returned in a single GPR.
2345 if (!Copy->hasOneUse())
2347 Copy = *Copy->use_begin();
2348 if (Copy->getOpcode() != ISD::CopyToReg || !Copy->hasNUsesOfValue(1, 0))
2350 // If the copy has a glue operand, we conservatively assume it isn't safe to
2351 // perform a tail call.
2352 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
2354 TCChain = Copy->getOperand(0);
2359 bool HasRet = false;
2360 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
2362 if (UI->getOpcode() != ARMISD::RET_FLAG &&
2363 UI->getOpcode() != ARMISD::INTRET_FLAG)
2375 bool ARMTargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
2376 if (!Subtarget->supportsTailCall())
2379 if (!CI->isTailCall() || getTargetMachine().Options.DisableTailCalls)
2382 return !Subtarget->isThumb1Only();
2385 // Trying to write a 64 bit value so need to split into two 32 bit values first,
2386 // and pass the lower and high parts through.
2387 static SDValue LowerWRITE_REGISTER(SDValue Op, SelectionDAG &DAG) {
2389 SDValue WriteValue = Op->getOperand(2);
2391 // This function is only supposed to be called for i64 type argument.
2392 assert(WriteValue.getValueType() == MVT::i64
2393 && "LowerWRITE_REGISTER called for non-i64 type argument.");
2395 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, WriteValue,
2396 DAG.getConstant(0, DL, MVT::i32));
2397 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, WriteValue,
2398 DAG.getConstant(1, DL, MVT::i32));
2399 SDValue Ops[] = { Op->getOperand(0), Op->getOperand(1), Lo, Hi };
2400 return DAG.getNode(ISD::WRITE_REGISTER, DL, MVT::Other, Ops);
2403 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
2404 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
2405 // one of the above mentioned nodes. It has to be wrapped because otherwise
2406 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
2407 // be used to form addressing mode. These wrapped nodes will be selected
2409 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
2410 EVT PtrVT = Op.getValueType();
2411 // FIXME there is no actual debug info here
2413 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
2415 if (CP->isMachineConstantPoolEntry())
2416 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
2417 CP->getAlignment());
2419 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
2420 CP->getAlignment());
2421 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
2424 unsigned ARMTargetLowering::getJumpTableEncoding() const {
2425 return MachineJumpTableInfo::EK_Inline;
2428 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
2429 SelectionDAG &DAG) const {
2430 MachineFunction &MF = DAG.getMachineFunction();
2431 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2432 unsigned ARMPCLabelIndex = 0;
2434 EVT PtrVT = getPointerTy();
2435 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
2436 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2438 if (RelocM == Reloc::Static) {
2439 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
2441 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2442 ARMPCLabelIndex = AFI->createPICLabelUId();
2443 ARMConstantPoolValue *CPV =
2444 ARMConstantPoolConstant::Create(BA, ARMPCLabelIndex,
2445 ARMCP::CPBlockAddress, PCAdj);
2446 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2448 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
2449 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
2450 MachinePointerInfo::getConstantPool(),
2451 false, false, false, 0);
2452 if (RelocM == Reloc::Static)
2454 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, DL, MVT::i32);
2455 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
2458 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
2460 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
2461 SelectionDAG &DAG) const {
2463 EVT PtrVT = getPointerTy();
2464 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2465 MachineFunction &MF = DAG.getMachineFunction();
2466 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2467 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2468 ARMConstantPoolValue *CPV =
2469 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2470 ARMCP::CPValue, PCAdj, ARMCP::TLSGD, true);
2471 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2472 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
2473 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
2474 MachinePointerInfo::getConstantPool(),
2475 false, false, false, 0);
2476 SDValue Chain = Argument.getValue(1);
2478 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, dl, MVT::i32);
2479 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
2481 // call __tls_get_addr.
2484 Entry.Node = Argument;
2485 Entry.Ty = (Type *) Type::getInt32Ty(*DAG.getContext());
2486 Args.push_back(Entry);
2488 // FIXME: is there useful debug info available here?
2489 TargetLowering::CallLoweringInfo CLI(DAG);
2490 CLI.setDebugLoc(dl).setChain(Chain)
2491 .setCallee(CallingConv::C, Type::getInt32Ty(*DAG.getContext()),
2492 DAG.getExternalSymbol("__tls_get_addr", PtrVT), std::move(Args),
2495 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
2496 return CallResult.first;
2499 // Lower ISD::GlobalTLSAddress using the "initial exec" or
2500 // "local exec" model.
2502 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
2504 TLSModel::Model model) const {
2505 const GlobalValue *GV = GA->getGlobal();
2508 SDValue Chain = DAG.getEntryNode();
2509 EVT PtrVT = getPointerTy();
2510 // Get the Thread Pointer
2511 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2513 if (model == TLSModel::InitialExec) {
2514 MachineFunction &MF = DAG.getMachineFunction();
2515 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2516 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2517 // Initial exec model.
2518 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2519 ARMConstantPoolValue *CPV =
2520 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2521 ARMCP::CPValue, PCAdj, ARMCP::GOTTPOFF,
2523 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2524 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2525 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2526 MachinePointerInfo::getConstantPool(),
2527 false, false, false, 0);
2528 Chain = Offset.getValue(1);
2530 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, dl, MVT::i32);
2531 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
2533 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2534 MachinePointerInfo::getConstantPool(),
2535 false, false, false, 0);
2538 assert(model == TLSModel::LocalExec);
2539 ARMConstantPoolValue *CPV =
2540 ARMConstantPoolConstant::Create(GV, ARMCP::TPOFF);
2541 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2542 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2543 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2544 MachinePointerInfo::getConstantPool(),
2545 false, false, false, 0);
2548 // The address of the thread local variable is the add of the thread
2549 // pointer with the offset of the variable.
2550 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
2554 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
2555 // TODO: implement the "local dynamic" model
2556 assert(Subtarget->isTargetELF() &&
2557 "TLS not implemented for non-ELF targets");
2558 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
2560 TLSModel::Model model = getTargetMachine().getTLSModel(GA->getGlobal());
2563 case TLSModel::GeneralDynamic:
2564 case TLSModel::LocalDynamic:
2565 return LowerToTLSGeneralDynamicModel(GA, DAG);
2566 case TLSModel::InitialExec:
2567 case TLSModel::LocalExec:
2568 return LowerToTLSExecModels(GA, DAG, model);
2570 llvm_unreachable("bogus TLS model");
2573 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
2574 SelectionDAG &DAG) const {
2575 EVT PtrVT = getPointerTy();
2577 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2578 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2579 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
2580 ARMConstantPoolValue *CPV =
2581 ARMConstantPoolConstant::Create(GV,
2582 UseGOTOFF ? ARMCP::GOTOFF : ARMCP::GOT);
2583 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2584 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2585 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
2587 MachinePointerInfo::getConstantPool(),
2588 false, false, false, 0);
2589 SDValue Chain = Result.getValue(1);
2590 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
2591 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
2593 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
2594 MachinePointerInfo::getGOT(),
2595 false, false, false, 0);
2599 // If we have T2 ops, we can materialize the address directly via movt/movw
2600 // pair. This is always cheaper.
2601 if (Subtarget->useMovt(DAG.getMachineFunction())) {
2603 // FIXME: Once remat is capable of dealing with instructions with register
2604 // operands, expand this into two nodes.
2605 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2606 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2608 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2609 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2610 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2611 MachinePointerInfo::getConstantPool(),
2612 false, false, false, 0);
2616 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
2617 SelectionDAG &DAG) const {
2618 EVT PtrVT = getPointerTy();
2620 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2621 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2623 if (Subtarget->useMovt(DAG.getMachineFunction()))
2626 // FIXME: Once remat is capable of dealing with instructions with register
2627 // operands, expand this into multiple nodes
2629 RelocM == Reloc::PIC_ ? ARMISD::WrapperPIC : ARMISD::Wrapper;
2631 SDValue G = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, ARMII::MO_NONLAZY);
2632 SDValue Result = DAG.getNode(Wrapper, dl, PtrVT, G);
2634 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2635 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
2636 MachinePointerInfo::getGOT(), false, false, false, 0);
2640 SDValue ARMTargetLowering::LowerGlobalAddressWindows(SDValue Op,
2641 SelectionDAG &DAG) const {
2642 assert(Subtarget->isTargetWindows() && "non-Windows COFF is not supported");
2643 assert(Subtarget->useMovt(DAG.getMachineFunction()) &&
2644 "Windows on ARM expects to use movw/movt");
2646 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2647 const ARMII::TOF TargetFlags =
2648 (GV->hasDLLImportStorageClass() ? ARMII::MO_DLLIMPORT : ARMII::MO_NO_FLAG);
2649 EVT PtrVT = getPointerTy();
2655 // FIXME: Once remat is capable of dealing with instructions with register
2656 // operands, expand this into two nodes.
2657 Result = DAG.getNode(ARMISD::Wrapper, DL, PtrVT,
2658 DAG.getTargetGlobalAddress(GV, DL, PtrVT, /*Offset=*/0,
2660 if (GV->hasDLLImportStorageClass())
2661 Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), Result,
2662 MachinePointerInfo::getGOT(), false, false, false, 0);
2666 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
2667 SelectionDAG &DAG) const {
2668 assert(Subtarget->isTargetELF() &&
2669 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
2670 MachineFunction &MF = DAG.getMachineFunction();
2671 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2672 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2673 EVT PtrVT = getPointerTy();
2675 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2676 ARMConstantPoolValue *CPV =
2677 ARMConstantPoolSymbol::Create(*DAG.getContext(), "_GLOBAL_OFFSET_TABLE_",
2678 ARMPCLabelIndex, PCAdj);
2679 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2680 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2681 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2682 MachinePointerInfo::getConstantPool(),
2683 false, false, false, 0);
2684 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, dl, MVT::i32);
2685 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2689 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
2691 SDValue Val = DAG.getConstant(0, dl, MVT::i32);
2692 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl,
2693 DAG.getVTList(MVT::i32, MVT::Other), Op.getOperand(0),
2694 Op.getOperand(1), Val);
2698 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
2700 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
2701 Op.getOperand(1), DAG.getConstant(0, dl, MVT::i32));
2705 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
2706 const ARMSubtarget *Subtarget) const {
2707 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2710 default: return SDValue(); // Don't custom lower most intrinsics.
2711 case Intrinsic::arm_rbit: {
2712 assert(Op.getOperand(1).getValueType() == MVT::i32 &&
2713 "RBIT intrinsic must have i32 type!");
2714 return DAG.getNode(ARMISD::RBIT, dl, MVT::i32, Op.getOperand(1));
2716 case Intrinsic::arm_thread_pointer: {
2717 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2718 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2720 case Intrinsic::eh_sjlj_lsda: {
2721 MachineFunction &MF = DAG.getMachineFunction();
2722 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2723 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2724 EVT PtrVT = getPointerTy();
2725 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2727 unsigned PCAdj = (RelocM != Reloc::PIC_)
2728 ? 0 : (Subtarget->isThumb() ? 4 : 8);
2729 ARMConstantPoolValue *CPV =
2730 ARMConstantPoolConstant::Create(MF.getFunction(), ARMPCLabelIndex,
2731 ARMCP::CPLSDA, PCAdj);
2732 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2733 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2735 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2736 MachinePointerInfo::getConstantPool(),
2737 false, false, false, 0);
2739 if (RelocM == Reloc::PIC_) {
2740 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, dl, MVT::i32);
2741 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2745 case Intrinsic::arm_neon_vmulls:
2746 case Intrinsic::arm_neon_vmullu: {
2747 unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vmulls)
2748 ? ARMISD::VMULLs : ARMISD::VMULLu;
2749 return DAG.getNode(NewOpc, SDLoc(Op), Op.getValueType(),
2750 Op.getOperand(1), Op.getOperand(2));
2755 static SDValue LowerATOMIC_FENCE(SDValue Op, SelectionDAG &DAG,
2756 const ARMSubtarget *Subtarget) {
2757 // FIXME: handle "fence singlethread" more efficiently.
2759 if (!Subtarget->hasDataBarrier()) {
2760 // Some ARMv6 cpus can support data barriers with an mcr instruction.
2761 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2763 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2764 "Unexpected ISD::ATOMIC_FENCE encountered. Should be libcall!");
2765 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2766 DAG.getConstant(0, dl, MVT::i32));
2769 ConstantSDNode *OrdN = cast<ConstantSDNode>(Op.getOperand(1));
2770 AtomicOrdering Ord = static_cast<AtomicOrdering>(OrdN->getZExtValue());
2771 ARM_MB::MemBOpt Domain = ARM_MB::ISH;
2772 if (Subtarget->isMClass()) {
2773 // Only a full system barrier exists in the M-class architectures.
2774 Domain = ARM_MB::SY;
2775 } else if (Subtarget->isSwift() && Ord == Release) {
2776 // Swift happens to implement ISHST barriers in a way that's compatible with
2777 // Release semantics but weaker than ISH so we'd be fools not to use
2778 // it. Beware: other processors probably don't!
2779 Domain = ARM_MB::ISHST;
2782 return DAG.getNode(ISD::INTRINSIC_VOID, dl, MVT::Other, Op.getOperand(0),
2783 DAG.getConstant(Intrinsic::arm_dmb, dl, MVT::i32),
2784 DAG.getConstant(Domain, dl, MVT::i32));
2787 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG,
2788 const ARMSubtarget *Subtarget) {
2789 // ARM pre v5TE and Thumb1 does not have preload instructions.
2790 if (!(Subtarget->isThumb2() ||
2791 (!Subtarget->isThumb1Only() && Subtarget->hasV5TEOps())))
2792 // Just preserve the chain.
2793 return Op.getOperand(0);
2796 unsigned isRead = ~cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() & 1;
2798 (!Subtarget->hasV7Ops() || !Subtarget->hasMPExtension()))
2799 // ARMv7 with MP extension has PLDW.
2800 return Op.getOperand(0);
2802 unsigned isData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
2803 if (Subtarget->isThumb()) {
2805 isRead = ~isRead & 1;
2806 isData = ~isData & 1;
2809 return DAG.getNode(ARMISD::PRELOAD, dl, MVT::Other, Op.getOperand(0),
2810 Op.getOperand(1), DAG.getConstant(isRead, dl, MVT::i32),
2811 DAG.getConstant(isData, dl, MVT::i32));
2814 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
2815 MachineFunction &MF = DAG.getMachineFunction();
2816 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
2818 // vastart just stores the address of the VarArgsFrameIndex slot into the
2819 // memory location argument.
2821 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2822 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2823 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
2824 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
2825 MachinePointerInfo(SV), false, false, 0);
2829 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
2830 SDValue &Root, SelectionDAG &DAG,
2832 MachineFunction &MF = DAG.getMachineFunction();
2833 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2835 const TargetRegisterClass *RC;
2836 if (AFI->isThumb1OnlyFunction())
2837 RC = &ARM::tGPRRegClass;
2839 RC = &ARM::GPRRegClass;
2841 // Transform the arguments stored in physical registers into virtual ones.
2842 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2843 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2846 if (NextVA.isMemLoc()) {
2847 MachineFrameInfo *MFI = MF.getFrameInfo();
2848 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true);
2850 // Create load node to retrieve arguments from the stack.
2851 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2852 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
2853 MachinePointerInfo::getFixedStack(FI),
2854 false, false, false, 0);
2856 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
2857 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2859 if (!Subtarget->isLittle())
2860 std::swap (ArgValue, ArgValue2);
2861 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
2864 // The remaining GPRs hold either the beginning of variable-argument
2865 // data, or the beginning of an aggregate passed by value (usually
2866 // byval). Either way, we allocate stack slots adjacent to the data
2867 // provided by our caller, and store the unallocated registers there.
2868 // If this is a variadic function, the va_list pointer will begin with
2869 // these values; otherwise, this reassembles a (byval) structure that
2870 // was split between registers and memory.
2871 // Return: The frame index registers were stored into.
2873 ARMTargetLowering::StoreByValRegs(CCState &CCInfo, SelectionDAG &DAG,
2874 SDLoc dl, SDValue &Chain,
2875 const Value *OrigArg,
2876 unsigned InRegsParamRecordIdx,
2878 unsigned ArgSize) const {
2879 // Currently, two use-cases possible:
2880 // Case #1. Non-var-args function, and we meet first byval parameter.
2881 // Setup first unallocated register as first byval register;
2882 // eat all remained registers
2883 // (these two actions are performed by HandleByVal method).
2884 // Then, here, we initialize stack frame with
2885 // "store-reg" instructions.
2886 // Case #2. Var-args function, that doesn't contain byval parameters.
2887 // The same: eat all remained unallocated registers,
2888 // initialize stack frame.
2890 MachineFunction &MF = DAG.getMachineFunction();
2891 MachineFrameInfo *MFI = MF.getFrameInfo();
2892 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2893 unsigned RBegin, REnd;
2894 if (InRegsParamRecordIdx < CCInfo.getInRegsParamsCount()) {
2895 CCInfo.getInRegsParamInfo(InRegsParamRecordIdx, RBegin, REnd);
2897 unsigned RBeginIdx = CCInfo.getFirstUnallocated(GPRArgRegs);
2898 RBegin = RBeginIdx == 4 ? (unsigned)ARM::R4 : GPRArgRegs[RBeginIdx];
2903 ArgOffset = -4 * (ARM::R4 - RBegin);
2905 int FrameIndex = MFI->CreateFixedObject(ArgSize, ArgOffset, false);
2906 SDValue FIN = DAG.getFrameIndex(FrameIndex, getPointerTy());
2908 SmallVector<SDValue, 4> MemOps;
2909 const TargetRegisterClass *RC =
2910 AFI->isThumb1OnlyFunction() ? &ARM::tGPRRegClass : &ARM::GPRRegClass;
2912 for (unsigned Reg = RBegin, i = 0; Reg < REnd; ++Reg, ++i) {
2913 unsigned VReg = MF.addLiveIn(Reg, RC);
2914 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2916 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2917 MachinePointerInfo(OrigArg, 4 * i), false, false, 0);
2918 MemOps.push_back(Store);
2919 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
2920 DAG.getConstant(4, dl, getPointerTy()));
2923 if (!MemOps.empty())
2924 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOps);
2928 // Setup stack frame, the va_list pointer will start from.
2930 ARMTargetLowering::VarArgStyleRegisters(CCState &CCInfo, SelectionDAG &DAG,
2931 SDLoc dl, SDValue &Chain,
2933 unsigned TotalArgRegsSaveSize,
2934 bool ForceMutable) const {
2935 MachineFunction &MF = DAG.getMachineFunction();
2936 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2938 // Try to store any remaining integer argument regs
2939 // to their spots on the stack so that they may be loaded by deferencing
2940 // the result of va_next.
2941 // If there is no regs to be stored, just point address after last
2942 // argument passed via stack.
2943 int FrameIndex = StoreByValRegs(CCInfo, DAG, dl, Chain, nullptr,
2944 CCInfo.getInRegsParamsCount(),
2945 CCInfo.getNextStackOffset(), 4);
2946 AFI->setVarArgsFrameIndex(FrameIndex);
2950 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
2951 CallingConv::ID CallConv, bool isVarArg,
2952 const SmallVectorImpl<ISD::InputArg>
2954 SDLoc dl, SelectionDAG &DAG,
2955 SmallVectorImpl<SDValue> &InVals)
2957 MachineFunction &MF = DAG.getMachineFunction();
2958 MachineFrameInfo *MFI = MF.getFrameInfo();
2960 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2962 // Assign locations to all of the incoming arguments.
2963 SmallVector<CCValAssign, 16> ArgLocs;
2964 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs,
2965 *DAG.getContext(), Prologue);
2966 CCInfo.AnalyzeFormalArguments(Ins,
2967 CCAssignFnForNode(CallConv, /* Return*/ false,
2970 SmallVector<SDValue, 16> ArgValues;
2972 Function::const_arg_iterator CurOrigArg = MF.getFunction()->arg_begin();
2973 unsigned CurArgIdx = 0;
2975 // Initially ArgRegsSaveSize is zero.
2976 // Then we increase this value each time we meet byval parameter.
2977 // We also increase this value in case of varargs function.
2978 AFI->setArgRegsSaveSize(0);
2980 // Calculate the amount of stack space that we need to allocate to store
2981 // byval and variadic arguments that are passed in registers.
2982 // We need to know this before we allocate the first byval or variadic
2983 // argument, as they will be allocated a stack slot below the CFA (Canonical
2984 // Frame Address, the stack pointer at entry to the function).
2985 unsigned ArgRegBegin = ARM::R4;
2986 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2987 if (CCInfo.getInRegsParamsProcessed() >= CCInfo.getInRegsParamsCount())
2990 CCValAssign &VA = ArgLocs[i];
2991 unsigned Index = VA.getValNo();
2992 ISD::ArgFlagsTy Flags = Ins[Index].Flags;
2993 if (!Flags.isByVal())
2996 assert(VA.isMemLoc() && "unexpected byval pointer in reg");
2997 unsigned RBegin, REnd;
2998 CCInfo.getInRegsParamInfo(CCInfo.getInRegsParamsProcessed(), RBegin, REnd);
2999 ArgRegBegin = std::min(ArgRegBegin, RBegin);
3001 CCInfo.nextInRegsParam();
3003 CCInfo.rewindByValRegsInfo();
3005 int lastInsIndex = -1;
3006 if (isVarArg && MFI->hasVAStart()) {
3007 unsigned RegIdx = CCInfo.getFirstUnallocated(GPRArgRegs);
3008 if (RegIdx != array_lengthof(GPRArgRegs))
3009 ArgRegBegin = std::min(ArgRegBegin, (unsigned)GPRArgRegs[RegIdx]);
3012 unsigned TotalArgRegsSaveSize = 4 * (ARM::R4 - ArgRegBegin);
3013 AFI->setArgRegsSaveSize(TotalArgRegsSaveSize);
3015 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3016 CCValAssign &VA = ArgLocs[i];
3017 if (Ins[VA.getValNo()].isOrigArg()) {
3018 std::advance(CurOrigArg,
3019 Ins[VA.getValNo()].getOrigArgIndex() - CurArgIdx);
3020 CurArgIdx = Ins[VA.getValNo()].getOrigArgIndex();
3022 // Arguments stored in registers.
3023 if (VA.isRegLoc()) {
3024 EVT RegVT = VA.getLocVT();
3026 if (VA.needsCustom()) {
3027 // f64 and vector types are split up into multiple registers or
3028 // combinations of registers and stack slots.
3029 if (VA.getLocVT() == MVT::v2f64) {
3030 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
3032 VA = ArgLocs[++i]; // skip ahead to next loc
3034 if (VA.isMemLoc()) {
3035 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true);
3036 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
3037 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
3038 MachinePointerInfo::getFixedStack(FI),
3039 false, false, false, 0);
3041 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
3044 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
3045 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
3046 ArgValue, ArgValue1,
3047 DAG.getIntPtrConstant(0, dl));
3048 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
3049 ArgValue, ArgValue2,
3050 DAG.getIntPtrConstant(1, dl));
3052 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
3055 const TargetRegisterClass *RC;
3057 if (RegVT == MVT::f32)
3058 RC = &ARM::SPRRegClass;
3059 else if (RegVT == MVT::f64)
3060 RC = &ARM::DPRRegClass;
3061 else if (RegVT == MVT::v2f64)
3062 RC = &ARM::QPRRegClass;
3063 else if (RegVT == MVT::i32)
3064 RC = AFI->isThumb1OnlyFunction() ? &ARM::tGPRRegClass
3065 : &ARM::GPRRegClass;
3067 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
3069 // Transform the arguments in physical registers into virtual ones.
3070 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
3071 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
3074 // If this is an 8 or 16-bit value, it is really passed promoted
3075 // to 32 bits. Insert an assert[sz]ext to capture this, then
3076 // truncate to the right size.
3077 switch (VA.getLocInfo()) {
3078 default: llvm_unreachable("Unknown loc info!");
3079 case CCValAssign::Full: break;
3080 case CCValAssign::BCvt:
3081 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
3083 case CCValAssign::SExt:
3084 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
3085 DAG.getValueType(VA.getValVT()));
3086 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
3088 case CCValAssign::ZExt:
3089 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
3090 DAG.getValueType(VA.getValVT()));
3091 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
3095 InVals.push_back(ArgValue);
3097 } else { // VA.isRegLoc()
3100 assert(VA.isMemLoc());
3101 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
3103 int index = VA.getValNo();
3105 // Some Ins[] entries become multiple ArgLoc[] entries.
3106 // Process them only once.
3107 if (index != lastInsIndex)
3109 ISD::ArgFlagsTy Flags = Ins[index].Flags;
3110 // FIXME: For now, all byval parameter objects are marked mutable.
3111 // This can be changed with more analysis.
3112 // In case of tail call optimization mark all arguments mutable.
3113 // Since they could be overwritten by lowering of arguments in case of
3115 if (Flags.isByVal()) {
3116 assert(Ins[index].isOrigArg() &&
3117 "Byval arguments cannot be implicit");
3118 unsigned CurByValIndex = CCInfo.getInRegsParamsProcessed();
3120 int FrameIndex = StoreByValRegs(CCInfo, DAG, dl, Chain, CurOrigArg,
3121 CurByValIndex, VA.getLocMemOffset(),
3122 Flags.getByValSize());
3123 InVals.push_back(DAG.getFrameIndex(FrameIndex, getPointerTy()));
3124 CCInfo.nextInRegsParam();
3126 unsigned FIOffset = VA.getLocMemOffset();
3127 int FI = MFI->CreateFixedObject(VA.getLocVT().getSizeInBits()/8,
3130 // Create load nodes to retrieve arguments from the stack.
3131 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
3132 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
3133 MachinePointerInfo::getFixedStack(FI),
3134 false, false, false, 0));
3136 lastInsIndex = index;
3142 if (isVarArg && MFI->hasVAStart())
3143 VarArgStyleRegisters(CCInfo, DAG, dl, Chain,
3144 CCInfo.getNextStackOffset(),
3145 TotalArgRegsSaveSize);
3147 AFI->setArgumentStackSize(CCInfo.getNextStackOffset());
3152 /// isFloatingPointZero - Return true if this is +0.0.
3153 static bool isFloatingPointZero(SDValue Op) {
3154 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
3155 return CFP->getValueAPF().isPosZero();
3156 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
3157 // Maybe this has already been legalized into the constant pool?
3158 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
3159 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
3160 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
3161 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
3162 return CFP->getValueAPF().isPosZero();
3164 } else if (Op->getOpcode() == ISD::BITCAST &&
3165 Op->getValueType(0) == MVT::f64) {
3166 // Handle (ISD::BITCAST (ARMISD::VMOVIMM (ISD::TargetConstant 0)) MVT::f64)
3167 // created by LowerConstantFP().
3168 SDValue BitcastOp = Op->getOperand(0);
3169 if (BitcastOp->getOpcode() == ARMISD::VMOVIMM) {
3170 SDValue MoveOp = BitcastOp->getOperand(0);
3171 if (MoveOp->getOpcode() == ISD::TargetConstant &&
3172 cast<ConstantSDNode>(MoveOp)->getZExtValue() == 0) {
3180 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
3181 /// the given operands.
3183 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
3184 SDValue &ARMcc, SelectionDAG &DAG,
3186 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
3187 unsigned C = RHSC->getZExtValue();
3188 if (!isLegalICmpImmediate(C)) {
3189 // Constant does not fit, try adjusting it by one?
3194 if (C != 0x80000000 && isLegalICmpImmediate(C-1)) {
3195 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
3196 RHS = DAG.getConstant(C - 1, dl, MVT::i32);
3201 if (C != 0 && isLegalICmpImmediate(C-1)) {
3202 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
3203 RHS = DAG.getConstant(C - 1, dl, MVT::i32);
3208 if (C != 0x7fffffff && isLegalICmpImmediate(C+1)) {
3209 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
3210 RHS = DAG.getConstant(C + 1, dl, MVT::i32);
3215 if (C != 0xffffffff && isLegalICmpImmediate(C+1)) {
3216 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
3217 RHS = DAG.getConstant(C + 1, dl, MVT::i32);
3224 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3225 ARMISD::NodeType CompareType;
3228 CompareType = ARMISD::CMP;
3233 CompareType = ARMISD::CMPZ;
3236 ARMcc = DAG.getConstant(CondCode, dl, MVT::i32);
3237 return DAG.getNode(CompareType, dl, MVT::Glue, LHS, RHS);
3240 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
3242 ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
3244 assert(!Subtarget->isFPOnlySP() || RHS.getValueType() != MVT::f64);
3246 if (!isFloatingPointZero(RHS))
3247 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Glue, LHS, RHS);
3249 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Glue, LHS);
3250 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Glue, Cmp);
3253 /// duplicateCmp - Glue values can have only one use, so this function
3254 /// duplicates a comparison node.
3256 ARMTargetLowering::duplicateCmp(SDValue Cmp, SelectionDAG &DAG) const {
3257 unsigned Opc = Cmp.getOpcode();
3259 if (Opc == ARMISD::CMP || Opc == ARMISD::CMPZ)
3260 return DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
3262 assert(Opc == ARMISD::FMSTAT && "unexpected comparison operation");
3263 Cmp = Cmp.getOperand(0);
3264 Opc = Cmp.getOpcode();
3265 if (Opc == ARMISD::CMPFP)
3266 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
3268 assert(Opc == ARMISD::CMPFPw0 && "unexpected operand of FMSTAT");
3269 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0));
3271 return DAG.getNode(ARMISD::FMSTAT, DL, MVT::Glue, Cmp);
3274 std::pair<SDValue, SDValue>
3275 ARMTargetLowering::getARMXALUOOp(SDValue Op, SelectionDAG &DAG,
3276 SDValue &ARMcc) const {
3277 assert(Op.getValueType() == MVT::i32 && "Unsupported value type");
3279 SDValue Value, OverflowCmp;
3280 SDValue LHS = Op.getOperand(0);
3281 SDValue RHS = Op.getOperand(1);
3284 // FIXME: We are currently always generating CMPs because we don't support
3285 // generating CMN through the backend. This is not as good as the natural
3286 // CMP case because it causes a register dependency and cannot be folded
3289 switch (Op.getOpcode()) {
3291 llvm_unreachable("Unknown overflow instruction!");
3293 ARMcc = DAG.getConstant(ARMCC::VC, dl, MVT::i32);
3294 Value = DAG.getNode(ISD::ADD, dl, Op.getValueType(), LHS, RHS);
3295 OverflowCmp = DAG.getNode(ARMISD::CMP, dl, MVT::Glue, Value, LHS);
3298 ARMcc = DAG.getConstant(ARMCC::HS, dl, MVT::i32);
3299 Value = DAG.getNode(ISD::ADD, dl, Op.getValueType(), LHS, RHS);
3300 OverflowCmp = DAG.getNode(ARMISD::CMP, dl, MVT::Glue, Value, LHS);
3303 ARMcc = DAG.getConstant(ARMCC::VC, dl, MVT::i32);
3304 Value = DAG.getNode(ISD::SUB, dl, Op.getValueType(), LHS, RHS);
3305 OverflowCmp = DAG.getNode(ARMISD::CMP, dl, MVT::Glue, LHS, RHS);
3308 ARMcc = DAG.getConstant(ARMCC::HS, dl, MVT::i32);
3309 Value = DAG.getNode(ISD::SUB, dl, Op.getValueType(), LHS, RHS);
3310 OverflowCmp = DAG.getNode(ARMISD::CMP, dl, MVT::Glue, LHS, RHS);
3314 return std::make_pair(Value, OverflowCmp);
3319 ARMTargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) const {
3320 // Let legalize expand this if it isn't a legal type yet.
3321 if (!DAG.getTargetLoweringInfo().isTypeLegal(Op.getValueType()))
3324 SDValue Value, OverflowCmp;
3326 std::tie(Value, OverflowCmp) = getARMXALUOOp(Op, DAG, ARMcc);
3327 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3329 // We use 0 and 1 as false and true values.
3330 SDValue TVal = DAG.getConstant(1, dl, MVT::i32);
3331 SDValue FVal = DAG.getConstant(0, dl, MVT::i32);
3332 EVT VT = Op.getValueType();
3334 SDValue Overflow = DAG.getNode(ARMISD::CMOV, dl, VT, TVal, FVal,
3335 ARMcc, CCR, OverflowCmp);
3337 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
3338 return DAG.getNode(ISD::MERGE_VALUES, dl, VTs, Value, Overflow);
3342 SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
3343 SDValue Cond = Op.getOperand(0);
3344 SDValue SelectTrue = Op.getOperand(1);
3345 SDValue SelectFalse = Op.getOperand(2);
3347 unsigned Opc = Cond.getOpcode();
3349 if (Cond.getResNo() == 1 &&
3350 (Opc == ISD::SADDO || Opc == ISD::UADDO || Opc == ISD::SSUBO ||
3351 Opc == ISD::USUBO)) {
3352 if (!DAG.getTargetLoweringInfo().isTypeLegal(Cond->getValueType(0)))
3355 SDValue Value, OverflowCmp;
3357 std::tie(Value, OverflowCmp) = getARMXALUOOp(Cond, DAG, ARMcc);
3358 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3359 EVT VT = Op.getValueType();
3361 return getCMOV(dl, VT, SelectTrue, SelectFalse, ARMcc, CCR,
3367 // (select (cmov 1, 0, cond), t, f) -> (cmov t, f, cond)
3368 // (select (cmov 0, 1, cond), t, f) -> (cmov f, t, cond)
3370 if (Cond.getOpcode() == ARMISD::CMOV && Cond.hasOneUse()) {
3371 const ConstantSDNode *CMOVTrue =
3372 dyn_cast<ConstantSDNode>(Cond.getOperand(0));
3373 const ConstantSDNode *CMOVFalse =
3374 dyn_cast<ConstantSDNode>(Cond.getOperand(1));
3376 if (CMOVTrue && CMOVFalse) {
3377 unsigned CMOVTrueVal = CMOVTrue->getZExtValue();
3378 unsigned CMOVFalseVal = CMOVFalse->getZExtValue();
3382 if (CMOVTrueVal == 1 && CMOVFalseVal == 0) {
3384 False = SelectFalse;
3385 } else if (CMOVTrueVal == 0 && CMOVFalseVal == 1) {
3390 if (True.getNode() && False.getNode()) {
3391 EVT VT = Op.getValueType();
3392 SDValue ARMcc = Cond.getOperand(2);
3393 SDValue CCR = Cond.getOperand(3);
3394 SDValue Cmp = duplicateCmp(Cond.getOperand(4), DAG);
3395 assert(True.getValueType() == VT);
3396 return getCMOV(dl, VT, True, False, ARMcc, CCR, Cmp, DAG);
3401 // ARM's BooleanContents value is UndefinedBooleanContent. Mask out the
3402 // undefined bits before doing a full-word comparison with zero.
3403 Cond = DAG.getNode(ISD::AND, dl, Cond.getValueType(), Cond,
3404 DAG.getConstant(1, dl, Cond.getValueType()));
3406 return DAG.getSelectCC(dl, Cond,
3407 DAG.getConstant(0, dl, Cond.getValueType()),
3408 SelectTrue, SelectFalse, ISD::SETNE);
3411 static void checkVSELConstraints(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
3412 bool &swpCmpOps, bool &swpVselOps) {
3413 // Start by selecting the GE condition code for opcodes that return true for
3415 if (CC == ISD::SETUGE || CC == ISD::SETOGE || CC == ISD::SETOLE ||
3417 CondCode = ARMCC::GE;
3419 // and GT for opcodes that return false for 'equality'.
3420 else if (CC == ISD::SETUGT || CC == ISD::SETOGT || CC == ISD::SETOLT ||
3422 CondCode = ARMCC::GT;
3424 // Since we are constrained to GE/GT, if the opcode contains 'less', we need
3425 // to swap the compare operands.
3426 if (CC == ISD::SETOLE || CC == ISD::SETULE || CC == ISD::SETOLT ||
3430 // Both GT and GE are ordered comparisons, and return false for 'unordered'.
3431 // If we have an unordered opcode, we need to swap the operands to the VSEL
3432 // instruction (effectively negating the condition).
3434 // This also has the effect of swapping which one of 'less' or 'greater'
3435 // returns true, so we also swap the compare operands. It also switches
3436 // whether we return true for 'equality', so we compensate by picking the
3437 // opposite condition code to our original choice.
3438 if (CC == ISD::SETULE || CC == ISD::SETULT || CC == ISD::SETUGE ||
3439 CC == ISD::SETUGT) {
3440 swpCmpOps = !swpCmpOps;
3441 swpVselOps = !swpVselOps;
3442 CondCode = CondCode == ARMCC::GT ? ARMCC::GE : ARMCC::GT;
3445 // 'ordered' is 'anything but unordered', so use the VS condition code and
3446 // swap the VSEL operands.
3447 if (CC == ISD::SETO) {
3448 CondCode = ARMCC::VS;
3452 // 'unordered or not equal' is 'anything but equal', so use the EQ condition
3453 // code and swap the VSEL operands.
3454 if (CC == ISD::SETUNE) {
3455 CondCode = ARMCC::EQ;
3460 SDValue ARMTargetLowering::getCMOV(SDLoc dl, EVT VT, SDValue FalseVal,
3461 SDValue TrueVal, SDValue ARMcc, SDValue CCR,
3462 SDValue Cmp, SelectionDAG &DAG) const {
3463 if (Subtarget->isFPOnlySP() && VT == MVT::f64) {
3464 FalseVal = DAG.getNode(ARMISD::VMOVRRD, dl,
3465 DAG.getVTList(MVT::i32, MVT::i32), FalseVal);
3466 TrueVal = DAG.getNode(ARMISD::VMOVRRD, dl,
3467 DAG.getVTList(MVT::i32, MVT::i32), TrueVal);
3469 SDValue TrueLow = TrueVal.getValue(0);
3470 SDValue TrueHigh = TrueVal.getValue(1);
3471 SDValue FalseLow = FalseVal.getValue(0);
3472 SDValue FalseHigh = FalseVal.getValue(1);
3474 SDValue Low = DAG.getNode(ARMISD::CMOV, dl, MVT::i32, FalseLow, TrueLow,
3476 SDValue High = DAG.getNode(ARMISD::CMOV, dl, MVT::i32, FalseHigh, TrueHigh,
3477 ARMcc, CCR, duplicateCmp(Cmp, DAG));
3479 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Low, High);
3481 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,
3486 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
3487 EVT VT = Op.getValueType();
3488 SDValue LHS = Op.getOperand(0);
3489 SDValue RHS = Op.getOperand(1);
3490 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
3491 SDValue TrueVal = Op.getOperand(2);
3492 SDValue FalseVal = Op.getOperand(3);
3495 if (Subtarget->isFPOnlySP() && LHS.getValueType() == MVT::f64) {
3496 DAG.getTargetLoweringInfo().softenSetCCOperands(DAG, MVT::f64, LHS, RHS, CC,
3499 // If softenSetCCOperands only returned one value, we should compare it to
3501 if (!RHS.getNode()) {
3502 RHS = DAG.getConstant(0, dl, LHS.getValueType());
3507 if (LHS.getValueType() == MVT::i32) {
3508 // Try to generate VSEL on ARMv8.
3509 // The VSEL instruction can't use all the usual ARM condition
3510 // codes: it only has two bits to select the condition code, so it's
3511 // constrained to use only GE, GT, VS and EQ.
3513 // To implement all the various ISD::SETXXX opcodes, we sometimes need to
3514 // swap the operands of the previous compare instruction (effectively
3515 // inverting the compare condition, swapping 'less' and 'greater') and
3516 // sometimes need to swap the operands to the VSEL (which inverts the
3517 // condition in the sense of firing whenever the previous condition didn't)
3518 if (Subtarget->hasFPARMv8() && (TrueVal.getValueType() == MVT::f32 ||
3519 TrueVal.getValueType() == MVT::f64)) {
3520 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3521 if (CondCode == ARMCC::LT || CondCode == ARMCC::LE ||
3522 CondCode == ARMCC::VC || CondCode == ARMCC::NE) {
3523 CC = ISD::getSetCCInverse(CC, true);
3524 std::swap(TrueVal, FalseVal);
3529 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3530 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3531 return getCMOV(dl, VT, FalseVal, TrueVal, ARMcc, CCR, Cmp, DAG);
3534 ARMCC::CondCodes CondCode, CondCode2;
3535 FPCCToARMCC(CC, CondCode, CondCode2);
3537 // Try to generate VMAXNM/VMINNM on ARMv8.
3538 if (Subtarget->hasFPARMv8() && (TrueVal.getValueType() == MVT::f32 ||
3539 TrueVal.getValueType() == MVT::f64)) {
3540 // We can use VMAXNM/VMINNM for a compare followed by a select with the
3541 // same operands, as follows:
3542 // c = fcmp [?gt, ?ge, ?lt, ?le] a, b
3544 // In NoNaNsFPMath the CC will have been changed from, e.g., 'ogt' to 'gt'.
3545 bool swapSides = false;
3546 if (!getTargetMachine().Options.NoNaNsFPMath) {
3547 // transformability may depend on which way around we compare
3555 // the non-NaN should be RHS
3556 swapSides = DAG.isKnownNeverNaN(LHS) && !DAG.isKnownNeverNaN(RHS);
3562 // the non-NaN should be LHS
3563 swapSides = DAG.isKnownNeverNaN(RHS) && !DAG.isKnownNeverNaN(LHS);
3567 swapSides = swapSides || (LHS == FalseVal && RHS == TrueVal);
3569 CC = ISD::getSetCCSwappedOperands(CC);
3570 std::swap(LHS, RHS);
3572 if (LHS == TrueVal && RHS == FalseVal) {
3573 bool canTransform = true;
3574 // FIXME: FastMathFlags::noSignedZeros() doesn't appear reachable from here
3575 if (!getTargetMachine().Options.UnsafeFPMath &&
3576 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
3577 const ConstantFPSDNode *Zero;
3584 // RHS must not be -0
3585 canTransform = (Zero = dyn_cast<ConstantFPSDNode>(RHS)) &&
3586 !Zero->isNegative();
3591 // LHS must not be -0
3592 canTransform = (Zero = dyn_cast<ConstantFPSDNode>(LHS)) &&
3593 !Zero->isNegative();
3598 // RHS must not be +0
3599 canTransform = (Zero = dyn_cast<ConstantFPSDNode>(RHS)) &&
3605 // LHS must not be +0
3606 canTransform = (Zero = dyn_cast<ConstantFPSDNode>(LHS)) &&
3612 // Note: If one of the elements in a pair is a number and the other
3613 // element is NaN, the corresponding result element is the number.
3614 // This is consistent with the IEEE 754-2008 standard.
3615 // Therefore, a > b ? a : b <=> vmax(a,b), if b is constant and a is NaN
3621 if (!DAG.isKnownNeverNaN(RHS))
3623 return DAG.getNode(ARMISD::VMAXNM, dl, VT, LHS, RHS);
3626 if (!DAG.isKnownNeverNaN(LHS))
3630 return DAG.getNode(ARMISD::VMAXNM, dl, VT, LHS, RHS);
3633 if (!DAG.isKnownNeverNaN(RHS))
3635 return DAG.getNode(ARMISD::VMINNM, dl, VT, LHS, RHS);
3638 if (!DAG.isKnownNeverNaN(LHS))
3642 return DAG.getNode(ARMISD::VMINNM, dl, VT, LHS, RHS);
3647 bool swpCmpOps = false;
3648 bool swpVselOps = false;
3649 checkVSELConstraints(CC, CondCode, swpCmpOps, swpVselOps);
3651 if (CondCode == ARMCC::GT || CondCode == ARMCC::GE ||
3652 CondCode == ARMCC::VS || CondCode == ARMCC::EQ) {
3654 std::swap(LHS, RHS);
3656 std::swap(TrueVal, FalseVal);
3660 SDValue ARMcc = DAG.getConstant(CondCode, dl, MVT::i32);
3661 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
3662 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3663 SDValue Result = getCMOV(dl, VT, FalseVal, TrueVal, ARMcc, CCR, Cmp, DAG);
3664 if (CondCode2 != ARMCC::AL) {
3665 SDValue ARMcc2 = DAG.getConstant(CondCode2, dl, MVT::i32);
3666 // FIXME: Needs another CMP because flag can have but one use.
3667 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
3668 Result = getCMOV(dl, VT, Result, TrueVal, ARMcc2, CCR, Cmp2, DAG);
3673 /// canChangeToInt - Given the fp compare operand, return true if it is suitable
3674 /// to morph to an integer compare sequence.
3675 static bool canChangeToInt(SDValue Op, bool &SeenZero,
3676 const ARMSubtarget *Subtarget) {
3677 SDNode *N = Op.getNode();
3678 if (!N->hasOneUse())
3679 // Otherwise it requires moving the value from fp to integer registers.
3681 if (!N->getNumValues())
3683 EVT VT = Op.getValueType();
3684 if (VT != MVT::f32 && !Subtarget->isFPBrccSlow())
3685 // f32 case is generally profitable. f64 case only makes sense when vcmpe +
3686 // vmrs are very slow, e.g. cortex-a8.
3689 if (isFloatingPointZero(Op)) {
3693 return ISD::isNormalLoad(N);
3696 static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) {
3697 if (isFloatingPointZero(Op))
3698 return DAG.getConstant(0, SDLoc(Op), MVT::i32);
3700 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op))
3701 return DAG.getLoad(MVT::i32, SDLoc(Op),
3702 Ld->getChain(), Ld->getBasePtr(), Ld->getPointerInfo(),
3703 Ld->isVolatile(), Ld->isNonTemporal(),
3704 Ld->isInvariant(), Ld->getAlignment());
3706 llvm_unreachable("Unknown VFP cmp argument!");
3709 static void expandf64Toi32(SDValue Op, SelectionDAG &DAG,
3710 SDValue &RetVal1, SDValue &RetVal2) {
3713 if (isFloatingPointZero(Op)) {
3714 RetVal1 = DAG.getConstant(0, dl, MVT::i32);
3715 RetVal2 = DAG.getConstant(0, dl, MVT::i32);
3719 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) {
3720 SDValue Ptr = Ld->getBasePtr();
3721 RetVal1 = DAG.getLoad(MVT::i32, dl,
3722 Ld->getChain(), Ptr,
3723 Ld->getPointerInfo(),
3724 Ld->isVolatile(), Ld->isNonTemporal(),
3725 Ld->isInvariant(), Ld->getAlignment());
3727 EVT PtrType = Ptr.getValueType();
3728 unsigned NewAlign = MinAlign(Ld->getAlignment(), 4);
3729 SDValue NewPtr = DAG.getNode(ISD::ADD, dl,
3730 PtrType, Ptr, DAG.getConstant(4, dl, PtrType));
3731 RetVal2 = DAG.getLoad(MVT::i32, dl,
3732 Ld->getChain(), NewPtr,
3733 Ld->getPointerInfo().getWithOffset(4),
3734 Ld->isVolatile(), Ld->isNonTemporal(),
3735 Ld->isInvariant(), NewAlign);
3739 llvm_unreachable("Unknown VFP cmp argument!");
3742 /// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some
3743 /// f32 and even f64 comparisons to integer ones.
3745 ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const {
3746 SDValue Chain = Op.getOperand(0);
3747 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3748 SDValue LHS = Op.getOperand(2);
3749 SDValue RHS = Op.getOperand(3);
3750 SDValue Dest = Op.getOperand(4);
3753 bool LHSSeenZero = false;
3754 bool LHSOk = canChangeToInt(LHS, LHSSeenZero, Subtarget);
3755 bool RHSSeenZero = false;
3756 bool RHSOk = canChangeToInt(RHS, RHSSeenZero, Subtarget);
3757 if (LHSOk && RHSOk && (LHSSeenZero || RHSSeenZero)) {
3758 // If unsafe fp math optimization is enabled and there are no other uses of
3759 // the CMP operands, and the condition code is EQ or NE, we can optimize it
3760 // to an integer comparison.
3761 if (CC == ISD::SETOEQ)
3763 else if (CC == ISD::SETUNE)
3766 SDValue Mask = DAG.getConstant(0x7fffffff, dl, MVT::i32);
3768 if (LHS.getValueType() == MVT::f32) {
3769 LHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3770 bitcastf32Toi32(LHS, DAG), Mask);
3771 RHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3772 bitcastf32Toi32(RHS, DAG), Mask);
3773 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3774 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3775 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3776 Chain, Dest, ARMcc, CCR, Cmp);
3781 expandf64Toi32(LHS, DAG, LHS1, LHS2);
3782 expandf64Toi32(RHS, DAG, RHS1, RHS2);
3783 LHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, LHS2, Mask);
3784 RHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, RHS2, Mask);
3785 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3786 ARMcc = DAG.getConstant(CondCode, dl, MVT::i32);
3787 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3788 SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest };
3789 return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops);
3795 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
3796 SDValue Chain = Op.getOperand(0);
3797 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3798 SDValue LHS = Op.getOperand(2);
3799 SDValue RHS = Op.getOperand(3);
3800 SDValue Dest = Op.getOperand(4);
3803 if (Subtarget->isFPOnlySP() && LHS.getValueType() == MVT::f64) {
3804 DAG.getTargetLoweringInfo().softenSetCCOperands(DAG, MVT::f64, LHS, RHS, CC,
3807 // If softenSetCCOperands only returned one value, we should compare it to
3809 if (!RHS.getNode()) {
3810 RHS = DAG.getConstant(0, dl, LHS.getValueType());
3815 if (LHS.getValueType() == MVT::i32) {
3817 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3818 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3819 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3820 Chain, Dest, ARMcc, CCR, Cmp);
3823 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3825 if (getTargetMachine().Options.UnsafeFPMath &&
3826 (CC == ISD::SETEQ || CC == ISD::SETOEQ ||
3827 CC == ISD::SETNE || CC == ISD::SETUNE)) {
3828 SDValue Result = OptimizeVFPBrcond(Op, DAG);
3829 if (Result.getNode())
3833 ARMCC::CondCodes CondCode, CondCode2;
3834 FPCCToARMCC(CC, CondCode, CondCode2);
3836 SDValue ARMcc = DAG.getConstant(CondCode, dl, MVT::i32);
3837 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
3838 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3839 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3840 SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp };
3841 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops);
3842 if (CondCode2 != ARMCC::AL) {
3843 ARMcc = DAG.getConstant(CondCode2, dl, MVT::i32);
3844 SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) };
3845 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops);
3850 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
3851 SDValue Chain = Op.getOperand(0);
3852 SDValue Table = Op.getOperand(1);
3853 SDValue Index = Op.getOperand(2);
3856 EVT PTy = getPointerTy();
3857 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
3858 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
3859 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI);
3860 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, dl, PTy));
3861 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
3862 if (Subtarget->isThumb2()) {
3863 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
3864 // which does another jump to the destination. This also makes it easier
3865 // to translate it to TBB / TBH later.
3866 // FIXME: This might not work if the function is extremely large.
3867 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
3868 Addr, Op.getOperand(2), JTI);
3870 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
3871 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
3872 MachinePointerInfo::getJumpTable(),
3873 false, false, false, 0);
3874 Chain = Addr.getValue(1);
3875 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
3876 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI);
3878 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
3879 MachinePointerInfo::getJumpTable(),
3880 false, false, false, 0);
3881 Chain = Addr.getValue(1);
3882 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI);
3886 static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3887 EVT VT = Op.getValueType();
3890 if (Op.getValueType().getVectorElementType() == MVT::i32) {
3891 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::f32)
3893 return DAG.UnrollVectorOp(Op.getNode());
3896 assert(Op.getOperand(0).getValueType() == MVT::v4f32 &&
3897 "Invalid type for custom lowering!");
3898 if (VT != MVT::v4i16)
3899 return DAG.UnrollVectorOp(Op.getNode());
3901 Op = DAG.getNode(Op.getOpcode(), dl, MVT::v4i32, Op.getOperand(0));
3902 return DAG.getNode(ISD::TRUNCATE, dl, VT, Op);
3905 SDValue ARMTargetLowering::LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) const {
3906 EVT VT = Op.getValueType();
3908 return LowerVectorFP_TO_INT(Op, DAG);
3909 if (Subtarget->isFPOnlySP() && Op.getOperand(0).getValueType() == MVT::f64) {
3911 if (Op.getOpcode() == ISD::FP_TO_SINT)
3912 LC = RTLIB::getFPTOSINT(Op.getOperand(0).getValueType(),
3915 LC = RTLIB::getFPTOUINT(Op.getOperand(0).getValueType(),
3917 return makeLibCall(DAG, LC, Op.getValueType(), &Op.getOperand(0), 1,
3918 /*isSigned*/ false, SDLoc(Op)).first;
3924 static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3925 EVT VT = Op.getValueType();
3928 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::i32) {
3929 if (VT.getVectorElementType() == MVT::f32)
3931 return DAG.UnrollVectorOp(Op.getNode());
3934 assert(Op.getOperand(0).getValueType() == MVT::v4i16 &&
3935 "Invalid type for custom lowering!");
3936 if (VT != MVT::v4f32)
3937 return DAG.UnrollVectorOp(Op.getNode());
3941 switch (Op.getOpcode()) {
3942 default: llvm_unreachable("Invalid opcode!");
3943 case ISD::SINT_TO_FP:
3944 CastOpc = ISD::SIGN_EXTEND;
3945 Opc = ISD::SINT_TO_FP;
3947 case ISD::UINT_TO_FP:
3948 CastOpc = ISD::ZERO_EXTEND;
3949 Opc = ISD::UINT_TO_FP;
3953 Op = DAG.getNode(CastOpc, dl, MVT::v4i32, Op.getOperand(0));
3954 return DAG.getNode(Opc, dl, VT, Op);
3957 SDValue ARMTargetLowering::LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) const {
3958 EVT VT = Op.getValueType();
3960 return LowerVectorINT_TO_FP(Op, DAG);
3961 if (Subtarget->isFPOnlySP() && Op.getValueType() == MVT::f64) {
3963 if (Op.getOpcode() == ISD::SINT_TO_FP)
3964 LC = RTLIB::getSINTTOFP(Op.getOperand(0).getValueType(),
3967 LC = RTLIB::getUINTTOFP(Op.getOperand(0).getValueType(),
3969 return makeLibCall(DAG, LC, Op.getValueType(), &Op.getOperand(0), 1,
3970 /*isSigned*/ false, SDLoc(Op)).first;
3976 SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
3977 // Implement fcopysign with a fabs and a conditional fneg.
3978 SDValue Tmp0 = Op.getOperand(0);
3979 SDValue Tmp1 = Op.getOperand(1);
3981 EVT VT = Op.getValueType();
3982 EVT SrcVT = Tmp1.getValueType();
3983 bool InGPR = Tmp0.getOpcode() == ISD::BITCAST ||
3984 Tmp0.getOpcode() == ARMISD::VMOVDRR;
3985 bool UseNEON = !InGPR && Subtarget->hasNEON();
3988 // Use VBSL to copy the sign bit.
3989 unsigned EncodedVal = ARM_AM::createNEONModImm(0x6, 0x80);
3990 SDValue Mask = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v2i32,
3991 DAG.getTargetConstant(EncodedVal, dl, MVT::i32));
3992 EVT OpVT = (VT == MVT::f32) ? MVT::v2i32 : MVT::v1i64;
3994 Mask = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3995 DAG.getNode(ISD::BITCAST, dl, OpVT, Mask),
3996 DAG.getConstant(32, dl, MVT::i32));
3997 else /*if (VT == MVT::f32)*/
3998 Tmp0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp0);
3999 if (SrcVT == MVT::f32) {
4000 Tmp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp1);
4002 Tmp1 = DAG.getNode(ARMISD::VSHL, dl, OpVT,
4003 DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1),
4004 DAG.getConstant(32, dl, MVT::i32));
4005 } else if (VT == MVT::f32)
4006 Tmp1 = DAG.getNode(ARMISD::VSHRu, dl, MVT::v1i64,
4007 DAG.getNode(ISD::BITCAST, dl, MVT::v1i64, Tmp1),
4008 DAG.getConstant(32, dl, MVT::i32));
4009 Tmp0 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp0);
4010 Tmp1 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1);
4012 SDValue AllOnes = DAG.getTargetConstant(ARM_AM::createNEONModImm(0xe, 0xff),
4014 AllOnes = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v8i8, AllOnes);
4015 SDValue MaskNot = DAG.getNode(ISD::XOR, dl, OpVT, Mask,
4016 DAG.getNode(ISD::BITCAST, dl, OpVT, AllOnes));
4018 SDValue Res = DAG.getNode(ISD::OR, dl, OpVT,
4019 DAG.getNode(ISD::AND, dl, OpVT, Tmp1, Mask),
4020 DAG.getNode(ISD::AND, dl, OpVT, Tmp0, MaskNot));
4021 if (VT == MVT::f32) {
4022 Res = DAG.getNode(ISD::BITCAST, dl, MVT::v2f32, Res);
4023 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, Res,
4024 DAG.getConstant(0, dl, MVT::i32));
4026 Res = DAG.getNode(ISD::BITCAST, dl, MVT::f64, Res);
4032 // Bitcast operand 1 to i32.
4033 if (SrcVT == MVT::f64)
4034 Tmp1 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
4036 Tmp1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp1);
4038 // Or in the signbit with integer operations.
4039 SDValue Mask1 = DAG.getConstant(0x80000000, dl, MVT::i32);
4040 SDValue Mask2 = DAG.getConstant(0x7fffffff, dl, MVT::i32);
4041 Tmp1 = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp1, Mask1);
4042 if (VT == MVT::f32) {
4043 Tmp0 = DAG.getNode(ISD::AND, dl, MVT::i32,
4044 DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp0), Mask2);
4045 return DAG.getNode(ISD::BITCAST, dl, MVT::f32,
4046 DAG.getNode(ISD::OR, dl, MVT::i32, Tmp0, Tmp1));
4049 // f64: Or the high part with signbit and then combine two parts.
4050 Tmp0 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
4052 SDValue Lo = Tmp0.getValue(0);
4053 SDValue Hi = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp0.getValue(1), Mask2);
4054 Hi = DAG.getNode(ISD::OR, dl, MVT::i32, Hi, Tmp1);
4055 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
4058 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
4059 MachineFunction &MF = DAG.getMachineFunction();
4060 MachineFrameInfo *MFI = MF.getFrameInfo();
4061 MFI->setReturnAddressIsTaken(true);
4063 if (verifyReturnAddressArgumentIsConstant(Op, DAG))
4066 EVT VT = Op.getValueType();
4068 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
4070 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
4071 SDValue Offset = DAG.getConstant(4, dl, MVT::i32);
4072 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
4073 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
4074 MachinePointerInfo(), false, false, false, 0);
4077 // Return LR, which contains the return address. Mark it an implicit live-in.
4078 unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32));
4079 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
4082 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
4083 const ARMBaseRegisterInfo &ARI =
4084 *static_cast<const ARMBaseRegisterInfo*>(RegInfo);
4085 MachineFunction &MF = DAG.getMachineFunction();
4086 MachineFrameInfo *MFI = MF.getFrameInfo();
4087 MFI->setFrameAddressIsTaken(true);
4089 EVT VT = Op.getValueType();
4090 SDLoc dl(Op); // FIXME probably not meaningful
4091 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
4092 unsigned FrameReg = ARI.getFrameRegister(MF);
4093 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
4095 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
4096 MachinePointerInfo(),
4097 false, false, false, 0);
4101 // FIXME? Maybe this could be a TableGen attribute on some registers and
4102 // this table could be generated automatically from RegInfo.
4103 unsigned ARMTargetLowering::getRegisterByName(const char* RegName,
4105 unsigned Reg = StringSwitch<unsigned>(RegName)
4106 .Case("sp", ARM::SP)
4110 report_fatal_error(Twine("Invalid register name \""
4111 + StringRef(RegName) + "\"."));
4114 // Result is 64 bit value so split into two 32 bit values and return as a
4116 static void ExpandREAD_REGISTER(SDNode *N, SmallVectorImpl<SDValue> &Results,
4117 SelectionDAG &DAG) {
4120 // This function is only supposed to be called for i64 type destination.
4121 assert(N->getValueType(0) == MVT::i64
4122 && "ExpandREAD_REGISTER called for non-i64 type result.");
4124 SDValue Read = DAG.getNode(ISD::READ_REGISTER, DL,
4125 DAG.getVTList(MVT::i32, MVT::i32, MVT::Other),
4129 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Read.getValue(0),
4131 Results.push_back(Read.getOperand(0));
4134 /// ExpandBITCAST - If the target supports VFP, this function is called to
4135 /// expand a bit convert where either the source or destination type is i64 to
4136 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
4137 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
4138 /// vectors), since the legalizer won't know what to do with that.
4139 static SDValue ExpandBITCAST(SDNode *N, SelectionDAG &DAG) {
4140 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4142 SDValue Op = N->getOperand(0);
4144 // This function is only supposed to be called for i64 types, either as the
4145 // source or destination of the bit convert.
4146 EVT SrcVT = Op.getValueType();
4147 EVT DstVT = N->getValueType(0);
4148 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
4149 "ExpandBITCAST called for non-i64 type");
4151 // Turn i64->f64 into VMOVDRR.
4152 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
4153 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
4154 DAG.getConstant(0, dl, MVT::i32));
4155 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
4156 DAG.getConstant(1, dl, MVT::i32));
4157 return DAG.getNode(ISD::BITCAST, dl, DstVT,
4158 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
4161 // Turn f64->i64 into VMOVRRD.
4162 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
4164 if (TLI.isBigEndian() && SrcVT.isVector() &&
4165 SrcVT.getVectorNumElements() > 1)
4166 Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
4167 DAG.getVTList(MVT::i32, MVT::i32),
4168 DAG.getNode(ARMISD::VREV64, dl, SrcVT, Op));
4170 Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
4171 DAG.getVTList(MVT::i32, MVT::i32), Op);
4172 // Merge the pieces into a single i64 value.
4173 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
4179 /// getZeroVector - Returns a vector of specified type with all zero elements.
4180 /// Zero vectors are used to represent vector negation and in those cases
4181 /// will be implemented with the NEON VNEG instruction. However, VNEG does
4182 /// not support i64 elements, so sometimes the zero vectors will need to be
4183 /// explicitly constructed. Regardless, use a canonical VMOV to create the
4185 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, SDLoc dl) {
4186 assert(VT.isVector() && "Expected a vector type");
4187 // The canonical modified immediate encoding of a zero vector is....0!
4188 SDValue EncodedVal = DAG.getTargetConstant(0, dl, MVT::i32);
4189 EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
4190 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal);
4191 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4194 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
4195 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
4196 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
4197 SelectionDAG &DAG) const {
4198 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4199 EVT VT = Op.getValueType();
4200 unsigned VTBits = VT.getSizeInBits();
4202 SDValue ShOpLo = Op.getOperand(0);
4203 SDValue ShOpHi = Op.getOperand(1);
4204 SDValue ShAmt = Op.getOperand(2);
4206 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
4208 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
4210 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
4211 DAG.getConstant(VTBits, dl, MVT::i32), ShAmt);
4212 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
4213 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
4214 DAG.getConstant(VTBits, dl, MVT::i32));
4215 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
4216 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
4217 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
4219 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
4220 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, dl, MVT::i32),
4221 ISD::SETGE, ARMcc, DAG, dl);
4222 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
4223 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc,
4226 SDValue Ops[2] = { Lo, Hi };
4227 return DAG.getMergeValues(Ops, dl);
4230 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
4231 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
4232 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
4233 SelectionDAG &DAG) const {
4234 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4235 EVT VT = Op.getValueType();
4236 unsigned VTBits = VT.getSizeInBits();
4238 SDValue ShOpLo = Op.getOperand(0);
4239 SDValue ShOpHi = Op.getOperand(1);
4240 SDValue ShAmt = Op.getOperand(2);
4243 assert(Op.getOpcode() == ISD::SHL_PARTS);
4244 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
4245 DAG.getConstant(VTBits, dl, MVT::i32), ShAmt);
4246 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
4247 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
4248 DAG.getConstant(VTBits, dl, MVT::i32));
4249 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
4250 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
4252 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
4253 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
4254 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, dl, MVT::i32),
4255 ISD::SETGE, ARMcc, DAG, dl);
4256 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
4257 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc,
4260 SDValue Ops[2] = { Lo, Hi };
4261 return DAG.getMergeValues(Ops, dl);
4264 SDValue ARMTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
4265 SelectionDAG &DAG) const {
4266 // The rounding mode is in bits 23:22 of the FPSCR.
4267 // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0
4268 // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3)
4269 // so that the shift + and get folded into a bitfield extract.
4271 SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::i32,
4272 DAG.getConstant(Intrinsic::arm_get_fpscr, dl,
4274 SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPSCR,
4275 DAG.getConstant(1U << 22, dl, MVT::i32));
4276 SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds,
4277 DAG.getConstant(22, dl, MVT::i32));
4278 return DAG.getNode(ISD::AND, dl, MVT::i32, RMODE,
4279 DAG.getConstant(3, dl, MVT::i32));
4282 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
4283 const ARMSubtarget *ST) {
4284 EVT VT = N->getValueType(0);
4287 if (!ST->hasV6T2Ops())
4290 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
4291 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
4294 /// getCTPOP16BitCounts - Returns a v8i8/v16i8 vector containing the bit-count
4295 /// for each 16-bit element from operand, repeated. The basic idea is to
4296 /// leverage vcnt to get the 8-bit counts, gather and add the results.
4298 /// Trace for v4i16:
4299 /// input = [v0 v1 v2 v3 ] (vi 16-bit element)
4300 /// cast: N0 = [w0 w1 w2 w3 w4 w5 w6 w7] (v0 = [w0 w1], wi 8-bit element)
4301 /// vcnt: N1 = [b0 b1 b2 b3 b4 b5 b6 b7] (bi = bit-count of 8-bit element wi)
4302 /// vrev: N2 = [b1 b0 b3 b2 b5 b4 b7 b6]
4303 /// [b0 b1 b2 b3 b4 b5 b6 b7]
4304 /// +[b1 b0 b3 b2 b5 b4 b7 b6]
4305 /// N3=N1+N2 = [k0 k0 k1 k1 k2 k2 k3 k3] (k0 = b0+b1 = bit-count of 16-bit v0,
4306 /// vuzp: = [k0 k1 k2 k3 k0 k1 k2 k3] each ki is 8-bits)
4307 static SDValue getCTPOP16BitCounts(SDNode *N, SelectionDAG &DAG) {
4308 EVT VT = N->getValueType(0);
4311 EVT VT8Bit = VT.is64BitVector() ? MVT::v8i8 : MVT::v16i8;
4312 SDValue N0 = DAG.getNode(ISD::BITCAST, DL, VT8Bit, N->getOperand(0));
4313 SDValue N1 = DAG.getNode(ISD::CTPOP, DL, VT8Bit, N0);
4314 SDValue N2 = DAG.getNode(ARMISD::VREV16, DL, VT8Bit, N1);
4315 SDValue N3 = DAG.getNode(ISD::ADD, DL, VT8Bit, N1, N2);
4316 return DAG.getNode(ARMISD::VUZP, DL, VT8Bit, N3, N3);
4319 /// lowerCTPOP16BitElements - Returns a v4i16/v8i16 vector containing the
4320 /// bit-count for each 16-bit element from the operand. We need slightly
4321 /// different sequencing for v4i16 and v8i16 to stay within NEON's available
4322 /// 64/128-bit registers.
4324 /// Trace for v4i16:
4325 /// input = [v0 v1 v2 v3 ] (vi 16-bit element)
4326 /// v8i8: BitCounts = [k0 k1 k2 k3 k0 k1 k2 k3 ] (ki is the bit-count of vi)
4327 /// v8i16:Extended = [k0 k1 k2 k3 k0 k1 k2 k3 ]
4328 /// v4i16:Extracted = [k0 k1 k2 k3 ]
4329 static SDValue lowerCTPOP16BitElements(SDNode *N, SelectionDAG &DAG) {
4330 EVT VT = N->getValueType(0);
4333 SDValue BitCounts = getCTPOP16BitCounts(N, DAG);
4334 if (VT.is64BitVector()) {
4335 SDValue Extended = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v8i16, BitCounts);
4336 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i16, Extended,
4337 DAG.getIntPtrConstant(0, DL));
4339 SDValue Extracted = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v8i8,
4340 BitCounts, DAG.getIntPtrConstant(0, DL));
4341 return DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v8i16, Extracted);
4345 /// lowerCTPOP32BitElements - Returns a v2i32/v4i32 vector containing the
4346 /// bit-count for each 32-bit element from the operand. The idea here is
4347 /// to split the vector into 16-bit elements, leverage the 16-bit count
4348 /// routine, and then combine the results.
4350 /// Trace for v2i32 (v4i32 similar with Extracted/Extended exchanged):
4351 /// input = [v0 v1 ] (vi: 32-bit elements)
4352 /// Bitcast = [w0 w1 w2 w3 ] (wi: 16-bit elements, v0 = [w0 w1])
4353 /// Counts16 = [k0 k1 k2 k3 ] (ki: 16-bit elements, bit-count of wi)
4354 /// vrev: N0 = [k1 k0 k3 k2 ]
4356 /// N1 =+[k1 k0 k3 k2 ]
4358 /// N2 =+[k1 k3 k0 k2 ]
4360 /// Extended =+[k1 k3 k0 k2 ]
4362 /// Extracted=+[k1 k3 ]
4364 static SDValue lowerCTPOP32BitElements(SDNode *N, SelectionDAG &DAG) {
4365 EVT VT = N->getValueType(0);
4368 EVT VT16Bit = VT.is64BitVector() ? MVT::v4i16 : MVT::v8i16;
4370 SDValue Bitcast = DAG.getNode(ISD::BITCAST, DL, VT16Bit, N->getOperand(0));
4371 SDValue Counts16 = lowerCTPOP16BitElements(Bitcast.getNode(), DAG);
4372 SDValue N0 = DAG.getNode(ARMISD::VREV32, DL, VT16Bit, Counts16);
4373 SDValue N1 = DAG.getNode(ISD::ADD, DL, VT16Bit, Counts16, N0);
4374 SDValue N2 = DAG.getNode(ARMISD::VUZP, DL, VT16Bit, N1, N1);
4376 if (VT.is64BitVector()) {
4377 SDValue Extended = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v4i32, N2);
4378 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i32, Extended,
4379 DAG.getIntPtrConstant(0, DL));
4381 SDValue Extracted = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i16, N2,
4382 DAG.getIntPtrConstant(0, DL));
4383 return DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v4i32, Extracted);
4387 static SDValue LowerCTPOP(SDNode *N, SelectionDAG &DAG,
4388 const ARMSubtarget *ST) {
4389 EVT VT = N->getValueType(0);
4391 assert(ST->hasNEON() && "Custom ctpop lowering requires NEON.");
4392 assert((VT == MVT::v2i32 || VT == MVT::v4i32 ||
4393 VT == MVT::v4i16 || VT == MVT::v8i16) &&
4394 "Unexpected type for custom ctpop lowering");
4396 if (VT.getVectorElementType() == MVT::i32)
4397 return lowerCTPOP32BitElements(N, DAG);
4399 return lowerCTPOP16BitElements(N, DAG);
4402 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
4403 const ARMSubtarget *ST) {
4404 EVT VT = N->getValueType(0);
4410 // Lower vector shifts on NEON to use VSHL.
4411 assert(ST->hasNEON() && "unexpected vector shift");
4413 // Left shifts translate directly to the vshiftu intrinsic.
4414 if (N->getOpcode() == ISD::SHL)
4415 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
4416 DAG.getConstant(Intrinsic::arm_neon_vshiftu, dl,
4418 N->getOperand(0), N->getOperand(1));
4420 assert((N->getOpcode() == ISD::SRA ||
4421 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
4423 // NEON uses the same intrinsics for both left and right shifts. For
4424 // right shifts, the shift amounts are negative, so negate the vector of
4426 EVT ShiftVT = N->getOperand(1).getValueType();
4427 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
4428 getZeroVector(ShiftVT, DAG, dl),
4430 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
4431 Intrinsic::arm_neon_vshifts :
4432 Intrinsic::arm_neon_vshiftu);
4433 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
4434 DAG.getConstant(vshiftInt, dl, MVT::i32),
4435 N->getOperand(0), NegatedCount);
4438 static SDValue Expand64BitShift(SDNode *N, SelectionDAG &DAG,
4439 const ARMSubtarget *ST) {
4440 EVT VT = N->getValueType(0);
4443 // We can get here for a node like i32 = ISD::SHL i32, i64
4447 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
4448 "Unknown shift to lower!");
4450 // We only lower SRA, SRL of 1 here, all others use generic lowering.
4451 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
4452 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
4455 // If we are in thumb mode, we don't have RRX.
4456 if (ST->isThumb1Only()) return SDValue();
4458 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
4459 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
4460 DAG.getConstant(0, dl, MVT::i32));
4461 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
4462 DAG.getConstant(1, dl, MVT::i32));
4464 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
4465 // captures the result into a carry flag.
4466 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
4467 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Glue), Hi);
4469 // The low part is an ARMISD::RRX operand, which shifts the carry in.
4470 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
4472 // Merge the pieces into a single i64 value.
4473 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
4476 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
4477 SDValue TmpOp0, TmpOp1;
4478 bool Invert = false;
4482 SDValue Op0 = Op.getOperand(0);
4483 SDValue Op1 = Op.getOperand(1);
4484 SDValue CC = Op.getOperand(2);
4485 EVT CmpVT = Op0.getValueType().changeVectorElementTypeToInteger();
4486 EVT VT = Op.getValueType();
4487 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
4490 if (Op1.getValueType().isFloatingPoint()) {
4491 switch (SetCCOpcode) {
4492 default: llvm_unreachable("Illegal FP comparison");
4494 case ISD::SETNE: Invert = true; // Fallthrough
4496 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
4498 case ISD::SETLT: Swap = true; // Fallthrough
4500 case ISD::SETGT: Opc = ARMISD::VCGT; break;
4502 case ISD::SETLE: Swap = true; // Fallthrough
4504 case ISD::SETGE: Opc = ARMISD::VCGE; break;
4505 case ISD::SETUGE: Swap = true; // Fallthrough
4506 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
4507 case ISD::SETUGT: Swap = true; // Fallthrough
4508 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
4509 case ISD::SETUEQ: Invert = true; // Fallthrough
4511 // Expand this to (OLT | OGT).
4515 Op0 = DAG.getNode(ARMISD::VCGT, dl, CmpVT, TmpOp1, TmpOp0);
4516 Op1 = DAG.getNode(ARMISD::VCGT, dl, CmpVT, TmpOp0, TmpOp1);
4518 case ISD::SETUO: Invert = true; // Fallthrough
4520 // Expand this to (OLT | OGE).
4524 Op0 = DAG.getNode(ARMISD::VCGT, dl, CmpVT, TmpOp1, TmpOp0);
4525 Op1 = DAG.getNode(ARMISD::VCGE, dl, CmpVT, TmpOp0, TmpOp1);
4529 // Integer comparisons.
4530 switch (SetCCOpcode) {
4531 default: llvm_unreachable("Illegal integer comparison");
4532 case ISD::SETNE: Invert = true;
4533 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
4534 case ISD::SETLT: Swap = true;
4535 case ISD::SETGT: Opc = ARMISD::VCGT; break;
4536 case ISD::SETLE: Swap = true;
4537 case ISD::SETGE: Opc = ARMISD::VCGE; break;
4538 case ISD::SETULT: Swap = true;
4539 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
4540 case ISD::SETULE: Swap = true;
4541 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
4544 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
4545 if (Opc == ARMISD::VCEQ) {
4548 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
4550 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
4553 // Ignore bitconvert.
4554 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BITCAST)
4555 AndOp = AndOp.getOperand(0);
4557 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
4559 Op0 = DAG.getNode(ISD::BITCAST, dl, CmpVT, AndOp.getOperand(0));
4560 Op1 = DAG.getNode(ISD::BITCAST, dl, CmpVT, AndOp.getOperand(1));
4567 std::swap(Op0, Op1);
4569 // If one of the operands is a constant vector zero, attempt to fold the
4570 // comparison to a specialized compare-against-zero form.
4572 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
4574 else if (ISD::isBuildVectorAllZeros(Op0.getNode())) {
4575 if (Opc == ARMISD::VCGE)
4576 Opc = ARMISD::VCLEZ;
4577 else if (Opc == ARMISD::VCGT)
4578 Opc = ARMISD::VCLTZ;
4583 if (SingleOp.getNode()) {
4586 Result = DAG.getNode(ARMISD::VCEQZ, dl, CmpVT, SingleOp); break;
4588 Result = DAG.getNode(ARMISD::VCGEZ, dl, CmpVT, SingleOp); break;
4590 Result = DAG.getNode(ARMISD::VCLEZ, dl, CmpVT, SingleOp); break;
4592 Result = DAG.getNode(ARMISD::VCGTZ, dl, CmpVT, SingleOp); break;
4594 Result = DAG.getNode(ARMISD::VCLTZ, dl, CmpVT, SingleOp); break;
4596 Result = DAG.getNode(Opc, dl, CmpVT, Op0, Op1);
4599 Result = DAG.getNode(Opc, dl, CmpVT, Op0, Op1);
4602 Result = DAG.getSExtOrTrunc(Result, dl, VT);
4605 Result = DAG.getNOT(dl, Result, VT);
4610 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
4611 /// valid vector constant for a NEON instruction with a "modified immediate"
4612 /// operand (e.g., VMOV). If so, return the encoded value.
4613 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
4614 unsigned SplatBitSize, SelectionDAG &DAG,
4615 SDLoc dl, EVT &VT, bool is128Bits,
4616 NEONModImmType type) {
4617 unsigned OpCmode, Imm;
4619 // SplatBitSize is set to the smallest size that splats the vector, so a
4620 // zero vector will always have SplatBitSize == 8. However, NEON modified
4621 // immediate instructions others than VMOV do not support the 8-bit encoding
4622 // of a zero vector, and the default encoding of zero is supposed to be the
4627 switch (SplatBitSize) {
4629 if (type != VMOVModImm)
4631 // Any 1-byte value is OK. Op=0, Cmode=1110.
4632 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
4635 VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
4639 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
4640 VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
4641 if ((SplatBits & ~0xff) == 0) {
4642 // Value = 0x00nn: Op=x, Cmode=100x.
4647 if ((SplatBits & ~0xff00) == 0) {
4648 // Value = 0xnn00: Op=x, Cmode=101x.
4650 Imm = SplatBits >> 8;
4656 // NEON's 32-bit VMOV supports splat values where:
4657 // * only one byte is nonzero, or
4658 // * the least significant byte is 0xff and the second byte is nonzero, or
4659 // * the least significant 2 bytes are 0xff and the third is nonzero.
4660 VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
4661 if ((SplatBits & ~0xff) == 0) {
4662 // Value = 0x000000nn: Op=x, Cmode=000x.
4667 if ((SplatBits & ~0xff00) == 0) {
4668 // Value = 0x0000nn00: Op=x, Cmode=001x.
4670 Imm = SplatBits >> 8;
4673 if ((SplatBits & ~0xff0000) == 0) {
4674 // Value = 0x00nn0000: Op=x, Cmode=010x.
4676 Imm = SplatBits >> 16;
4679 if ((SplatBits & ~0xff000000) == 0) {
4680 // Value = 0xnn000000: Op=x, Cmode=011x.
4682 Imm = SplatBits >> 24;
4686 // cmode == 0b1100 and cmode == 0b1101 are not supported for VORR or VBIC
4687 if (type == OtherModImm) return SDValue();
4689 if ((SplatBits & ~0xffff) == 0 &&
4690 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
4691 // Value = 0x0000nnff: Op=x, Cmode=1100.
4693 Imm = SplatBits >> 8;
4697 if ((SplatBits & ~0xffffff) == 0 &&
4698 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
4699 // Value = 0x00nnffff: Op=x, Cmode=1101.
4701 Imm = SplatBits >> 16;
4705 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
4706 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
4707 // VMOV.I32. A (very) minor optimization would be to replicate the value
4708 // and fall through here to test for a valid 64-bit splat. But, then the
4709 // caller would also need to check and handle the change in size.
4713 if (type != VMOVModImm)
4715 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
4716 uint64_t BitMask = 0xff;
4718 unsigned ImmMask = 1;
4720 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
4721 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
4724 } else if ((SplatBits & BitMask) != 0) {
4731 if (DAG.getTargetLoweringInfo().isBigEndian())
4732 // swap higher and lower 32 bit word
4733 Imm = ((Imm & 0xf) << 4) | ((Imm & 0xf0) >> 4);
4735 // Op=1, Cmode=1110.
4737 VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
4742 llvm_unreachable("unexpected size for isNEONModifiedImm");
4745 unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm);
4746 return DAG.getTargetConstant(EncodedVal, dl, MVT::i32);
4749 SDValue ARMTargetLowering::LowerConstantFP(SDValue Op, SelectionDAG &DAG,
4750 const ARMSubtarget *ST) const {
4754 bool IsDouble = Op.getValueType() == MVT::f64;
4755 ConstantFPSDNode *CFP = cast<ConstantFPSDNode>(Op);
4757 // Use the default (constant pool) lowering for double constants when we have
4759 if (IsDouble && Subtarget->isFPOnlySP())
4762 // Try splatting with a VMOV.f32...
4763 APFloat FPVal = CFP->getValueAPF();
4764 int ImmVal = IsDouble ? ARM_AM::getFP64Imm(FPVal) : ARM_AM::getFP32Imm(FPVal);
4767 if (IsDouble || !ST->useNEONForSinglePrecisionFP()) {
4768 // We have code in place to select a valid ConstantFP already, no need to
4773 // It's a float and we are trying to use NEON operations where
4774 // possible. Lower it to a splat followed by an extract.
4776 SDValue NewVal = DAG.getTargetConstant(ImmVal, DL, MVT::i32);
4777 SDValue VecConstant = DAG.getNode(ARMISD::VMOVFPIMM, DL, MVT::v2f32,
4779 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecConstant,
4780 DAG.getConstant(0, DL, MVT::i32));
4783 // The rest of our options are NEON only, make sure that's allowed before
4785 if (!ST->hasNEON() || (!IsDouble && !ST->useNEONForSinglePrecisionFP()))
4789 uint64_t iVal = FPVal.bitcastToAPInt().getZExtValue();
4791 // It wouldn't really be worth bothering for doubles except for one very
4792 // important value, which does happen to match: 0.0. So make sure we don't do
4794 if (IsDouble && (iVal & 0xffffffff) != (iVal >> 32))
4797 // Try a VMOV.i32 (FIXME: i8, i16, or i64 could work too).
4798 SDValue NewVal = isNEONModifiedImm(iVal & 0xffffffffU, 0, 32, DAG, SDLoc(Op),
4799 VMovVT, false, VMOVModImm);
4800 if (NewVal != SDValue()) {
4802 SDValue VecConstant = DAG.getNode(ARMISD::VMOVIMM, DL, VMovVT,
4805 return DAG.getNode(ISD::BITCAST, DL, MVT::f64, VecConstant);
4807 // It's a float: cast and extract a vector element.
4808 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
4810 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
4811 DAG.getConstant(0, DL, MVT::i32));
4814 // Finally, try a VMVN.i32
4815 NewVal = isNEONModifiedImm(~iVal & 0xffffffffU, 0, 32, DAG, SDLoc(Op), VMovVT,
4817 if (NewVal != SDValue()) {
4819 SDValue VecConstant = DAG.getNode(ARMISD::VMVNIMM, DL, VMovVT, NewVal);
4822 return DAG.getNode(ISD::BITCAST, DL, MVT::f64, VecConstant);
4824 // It's a float: cast and extract a vector element.
4825 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
4827 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
4828 DAG.getConstant(0, DL, MVT::i32));
4834 // check if an VEXT instruction can handle the shuffle mask when the
4835 // vector sources of the shuffle are the same.
4836 static bool isSingletonVEXTMask(ArrayRef<int> M, EVT VT, unsigned &Imm) {
4837 unsigned NumElts = VT.getVectorNumElements();
4839 // Assume that the first shuffle index is not UNDEF. Fail if it is.
4845 // If this is a VEXT shuffle, the immediate value is the index of the first
4846 // element. The other shuffle indices must be the successive elements after
4848 unsigned ExpectedElt = Imm;
4849 for (unsigned i = 1; i < NumElts; ++i) {
4850 // Increment the expected index. If it wraps around, just follow it
4851 // back to index zero and keep going.
4853 if (ExpectedElt == NumElts)
4856 if (M[i] < 0) continue; // ignore UNDEF indices
4857 if (ExpectedElt != static_cast<unsigned>(M[i]))
4865 static bool isVEXTMask(ArrayRef<int> M, EVT VT,
4866 bool &ReverseVEXT, unsigned &Imm) {
4867 unsigned NumElts = VT.getVectorNumElements();
4868 ReverseVEXT = false;
4870 // Assume that the first shuffle index is not UNDEF. Fail if it is.
4876 // If this is a VEXT shuffle, the immediate value is the index of the first
4877 // element. The other shuffle indices must be the successive elements after
4879 unsigned ExpectedElt = Imm;
4880 for (unsigned i = 1; i < NumElts; ++i) {
4881 // Increment the expected index. If it wraps around, it may still be
4882 // a VEXT but the source vectors must be swapped.
4884 if (ExpectedElt == NumElts * 2) {
4889 if (M[i] < 0) continue; // ignore UNDEF indices
4890 if (ExpectedElt != static_cast<unsigned>(M[i]))
4894 // Adjust the index value if the source operands will be swapped.
4901 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
4902 /// instruction with the specified blocksize. (The order of the elements
4903 /// within each block of the vector is reversed.)
4904 static bool isVREVMask(ArrayRef<int> M, EVT VT, unsigned BlockSize) {
4905 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
4906 "Only possible block sizes for VREV are: 16, 32, 64");
4908 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4912 unsigned NumElts = VT.getVectorNumElements();
4913 unsigned BlockElts = M[0] + 1;
4914 // If the first shuffle index is UNDEF, be optimistic.
4916 BlockElts = BlockSize / EltSz;
4918 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
4921 for (unsigned i = 0; i < NumElts; ++i) {
4922 if (M[i] < 0) continue; // ignore UNDEF indices
4923 if ((unsigned) M[i] != (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
4930 static bool isVTBLMask(ArrayRef<int> M, EVT VT) {
4931 // We can handle <8 x i8> vector shuffles. If the index in the mask is out of
4932 // range, then 0 is placed into the resulting vector. So pretty much any mask
4933 // of 8 elements can work here.
4934 return VT == MVT::v8i8 && M.size() == 8;
4937 static bool isVTRNMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4938 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4942 unsigned NumElts = VT.getVectorNumElements();
4943 WhichResult = (M[0] == 0 ? 0 : 1);
4944 for (unsigned i = 0; i < NumElts; i += 2) {
4945 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
4946 (M[i+1] >= 0 && (unsigned) M[i+1] != i + NumElts + WhichResult))
4952 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
4953 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4954 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
4955 static bool isVTRN_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4956 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4960 unsigned NumElts = VT.getVectorNumElements();
4961 WhichResult = (M[0] == 0 ? 0 : 1);
4962 for (unsigned i = 0; i < NumElts; i += 2) {
4963 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
4964 (M[i+1] >= 0 && (unsigned) M[i+1] != i + WhichResult))
4970 static bool isVUZPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4971 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4975 unsigned NumElts = VT.getVectorNumElements();
4976 WhichResult = (M[0] == 0 ? 0 : 1);
4977 for (unsigned i = 0; i != NumElts; ++i) {
4978 if (M[i] < 0) continue; // ignore UNDEF indices
4979 if ((unsigned) M[i] != 2 * i + WhichResult)
4983 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4984 if (VT.is64BitVector() && EltSz == 32)
4990 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
4991 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4992 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
4993 static bool isVUZP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4994 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4998 unsigned Half = VT.getVectorNumElements() / 2;
4999 WhichResult = (M[0] == 0 ? 0 : 1);
5000 for (unsigned j = 0; j != 2; ++j) {
5001 unsigned Idx = WhichResult;
5002 for (unsigned i = 0; i != Half; ++i) {
5003 int MIdx = M[i + j * Half];
5004 if (MIdx >= 0 && (unsigned) MIdx != Idx)
5010 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
5011 if (VT.is64BitVector() && EltSz == 32)
5017 static bool isVZIPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
5018 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
5022 unsigned NumElts = VT.getVectorNumElements();
5023 WhichResult = (M[0] == 0 ? 0 : 1);
5024 unsigned Idx = WhichResult * NumElts / 2;
5025 for (unsigned i = 0; i != NumElts; i += 2) {
5026 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
5027 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx + NumElts))
5032 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
5033 if (VT.is64BitVector() && EltSz == 32)
5039 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
5040 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
5041 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
5042 static bool isVZIP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
5043 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
5047 unsigned NumElts = VT.getVectorNumElements();
5048 WhichResult = (M[0] == 0 ? 0 : 1);
5049 unsigned Idx = WhichResult * NumElts / 2;
5050 for (unsigned i = 0; i != NumElts; i += 2) {
5051 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
5052 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx))
5057 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
5058 if (VT.is64BitVector() && EltSz == 32)
5064 /// \return true if this is a reverse operation on an vector.
5065 static bool isReverseMask(ArrayRef<int> M, EVT VT) {
5066 unsigned NumElts = VT.getVectorNumElements();
5067 // Make sure the mask has the right size.
5068 if (NumElts != M.size())
5071 // Look for <15, ..., 3, -1, 1, 0>.
5072 for (unsigned i = 0; i != NumElts; ++i)
5073 if (M[i] >= 0 && M[i] != (int) (NumElts - 1 - i))
5079 // If N is an integer constant that can be moved into a register in one
5080 // instruction, return an SDValue of such a constant (will become a MOV
5081 // instruction). Otherwise return null.
5082 static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG,
5083 const ARMSubtarget *ST, SDLoc dl) {
5085 if (!isa<ConstantSDNode>(N))
5087 Val = cast<ConstantSDNode>(N)->getZExtValue();
5089 if (ST->isThumb1Only()) {
5090 if (Val <= 255 || ~Val <= 255)
5091 return DAG.getConstant(Val, dl, MVT::i32);
5093 if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1)
5094 return DAG.getConstant(Val, dl, MVT::i32);
5099 // If this is a case we can't handle, return null and let the default
5100 // expansion code take care of it.
5101 SDValue ARMTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
5102 const ARMSubtarget *ST) const {
5103 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
5105 EVT VT = Op.getValueType();
5107 APInt SplatBits, SplatUndef;
5108 unsigned SplatBitSize;
5110 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
5111 if (SplatBitSize <= 64) {
5112 // Check if an immediate VMOV works.
5114 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
5115 SplatUndef.getZExtValue(), SplatBitSize,
5116 DAG, dl, VmovVT, VT.is128BitVector(),
5118 if (Val.getNode()) {
5119 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
5120 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
5123 // Try an immediate VMVN.
5124 uint64_t NegatedImm = (~SplatBits).getZExtValue();
5125 Val = isNEONModifiedImm(NegatedImm,
5126 SplatUndef.getZExtValue(), SplatBitSize,
5127 DAG, dl, VmovVT, VT.is128BitVector(),
5129 if (Val.getNode()) {
5130 SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
5131 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
5134 // Use vmov.f32 to materialize other v2f32 and v4f32 splats.
5135 if ((VT == MVT::v2f32 || VT == MVT::v4f32) && SplatBitSize == 32) {
5136 int ImmVal = ARM_AM::getFP32Imm(SplatBits);
5138 SDValue Val = DAG.getTargetConstant(ImmVal, dl, MVT::i32);
5139 return DAG.getNode(ARMISD::VMOVFPIMM, dl, VT, Val);
5145 // Scan through the operands to see if only one value is used.
5147 // As an optimisation, even if more than one value is used it may be more
5148 // profitable to splat with one value then change some lanes.
5150 // Heuristically we decide to do this if the vector has a "dominant" value,
5151 // defined as splatted to more than half of the lanes.
5152 unsigned NumElts = VT.getVectorNumElements();
5153 bool isOnlyLowElement = true;
5154 bool usesOnlyOneValue = true;
5155 bool hasDominantValue = false;
5156 bool isConstant = true;
5158 // Map of the number of times a particular SDValue appears in the
5160 DenseMap<SDValue, unsigned> ValueCounts;
5162 for (unsigned i = 0; i < NumElts; ++i) {
5163 SDValue V = Op.getOperand(i);
5164 if (V.getOpcode() == ISD::UNDEF)
5167 isOnlyLowElement = false;
5168 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
5171 ValueCounts.insert(std::make_pair(V, 0));
5172 unsigned &Count = ValueCounts[V];
5174 // Is this value dominant? (takes up more than half of the lanes)
5175 if (++Count > (NumElts / 2)) {
5176 hasDominantValue = true;
5180 if (ValueCounts.size() != 1)
5181 usesOnlyOneValue = false;
5182 if (!Value.getNode() && ValueCounts.size() > 0)
5183 Value = ValueCounts.begin()->first;
5185 if (ValueCounts.size() == 0)
5186 return DAG.getUNDEF(VT);
5188 // Loads are better lowered with insert_vector_elt/ARMISD::BUILD_VECTOR.
5189 // Keep going if we are hitting this case.
5190 if (isOnlyLowElement && !ISD::isNormalLoad(Value.getNode()))
5191 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
5193 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5195 // Use VDUP for non-constant splats. For f32 constant splats, reduce to
5196 // i32 and try again.
5197 if (hasDominantValue && EltSize <= 32) {
5201 // If we are VDUPing a value that comes directly from a vector, that will
5202 // cause an unnecessary move to and from a GPR, where instead we could
5203 // just use VDUPLANE. We can only do this if the lane being extracted
5204 // is at a constant index, as the VDUP from lane instructions only have
5205 // constant-index forms.
5206 if (Value->getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5207 isa<ConstantSDNode>(Value->getOperand(1))) {
5208 // We need to create a new undef vector to use for the VDUPLANE if the
5209 // size of the vector from which we get the value is different than the
5210 // size of the vector that we need to create. We will insert the element
5211 // such that the register coalescer will remove unnecessary copies.
5212 if (VT != Value->getOperand(0).getValueType()) {
5213 ConstantSDNode *constIndex;
5214 constIndex = dyn_cast<ConstantSDNode>(Value->getOperand(1));
5215 assert(constIndex && "The index is not a constant!");
5216 unsigned index = constIndex->getAPIntValue().getLimitedValue() %
5217 VT.getVectorNumElements();
5218 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
5219 DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DAG.getUNDEF(VT),
5220 Value, DAG.getConstant(index, dl, MVT::i32)),
5221 DAG.getConstant(index, dl, MVT::i32));
5223 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
5224 Value->getOperand(0), Value->getOperand(1));
5226 N = DAG.getNode(ARMISD::VDUP, dl, VT, Value);
5228 if (!usesOnlyOneValue) {
5229 // The dominant value was splatted as 'N', but we now have to insert
5230 // all differing elements.
5231 for (unsigned I = 0; I < NumElts; ++I) {
5232 if (Op.getOperand(I) == Value)
5234 SmallVector<SDValue, 3> Ops;
5236 Ops.push_back(Op.getOperand(I));
5237 Ops.push_back(DAG.getConstant(I, dl, MVT::i32));
5238 N = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Ops);
5243 if (VT.getVectorElementType().isFloatingPoint()) {
5244 SmallVector<SDValue, 8> Ops;
5245 for (unsigned i = 0; i < NumElts; ++i)
5246 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, MVT::i32,
5248 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
5249 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, Ops);
5250 Val = LowerBUILD_VECTOR(Val, DAG, ST);
5252 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
5254 if (usesOnlyOneValue) {
5255 SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
5256 if (isConstant && Val.getNode())
5257 return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
5261 // If all elements are constants and the case above didn't get hit, fall back
5262 // to the default expansion, which will generate a load from the constant
5267 // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
5269 SDValue shuffle = ReconstructShuffle(Op, DAG);
5270 if (shuffle != SDValue())
5274 // Vectors with 32- or 64-bit elements can be built by directly assigning
5275 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
5276 // will be legalized.
5277 if (EltSize >= 32) {
5278 // Do the expansion with floating-point types, since that is what the VFP
5279 // registers are defined to use, and since i64 is not legal.
5280 EVT EltVT = EVT::getFloatingPointVT(EltSize);
5281 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
5282 SmallVector<SDValue, 8> Ops;
5283 for (unsigned i = 0; i < NumElts; ++i)
5284 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, EltVT, Op.getOperand(i)));
5285 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, Ops);
5286 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
5289 // If all else fails, just use a sequence of INSERT_VECTOR_ELT when we
5290 // know the default expansion would otherwise fall back on something even
5291 // worse. For a vector with one or two non-undef values, that's
5292 // scalar_to_vector for the elements followed by a shuffle (provided the
5293 // shuffle is valid for the target) and materialization element by element
5294 // on the stack followed by a load for everything else.
5295 if (!isConstant && !usesOnlyOneValue) {
5296 SDValue Vec = DAG.getUNDEF(VT);
5297 for (unsigned i = 0 ; i < NumElts; ++i) {
5298 SDValue V = Op.getOperand(i);
5299 if (V.getOpcode() == ISD::UNDEF)
5301 SDValue LaneIdx = DAG.getConstant(i, dl, MVT::i32);
5302 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Vec, V, LaneIdx);
5310 // Gather data to see if the operation can be modelled as a
5311 // shuffle in combination with VEXTs.
5312 SDValue ARMTargetLowering::ReconstructShuffle(SDValue Op,
5313 SelectionDAG &DAG) const {
5315 EVT VT = Op.getValueType();
5316 unsigned NumElts = VT.getVectorNumElements();
5318 SmallVector<SDValue, 2> SourceVecs;
5319 SmallVector<unsigned, 2> MinElts;
5320 SmallVector<unsigned, 2> MaxElts;
5322 for (unsigned i = 0; i < NumElts; ++i) {
5323 SDValue V = Op.getOperand(i);
5324 if (V.getOpcode() == ISD::UNDEF)
5326 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
5327 // A shuffle can only come from building a vector from various
5328 // elements of other vectors.
5330 } else if (V.getOperand(0).getValueType().getVectorElementType() !=
5331 VT.getVectorElementType()) {
5332 // This code doesn't know how to handle shuffles where the vector
5333 // element types do not match (this happens because type legalization
5334 // promotes the return type of EXTRACT_VECTOR_ELT).
5335 // FIXME: It might be appropriate to extend this code to handle
5336 // mismatched types.
5340 // Record this extraction against the appropriate vector if possible...
5341 SDValue SourceVec = V.getOperand(0);
5342 // If the element number isn't a constant, we can't effectively
5343 // analyze what's going on.
5344 if (!isa<ConstantSDNode>(V.getOperand(1)))
5346 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
5347 bool FoundSource = false;
5348 for (unsigned j = 0; j < SourceVecs.size(); ++j) {
5349 if (SourceVecs[j] == SourceVec) {
5350 if (MinElts[j] > EltNo)
5352 if (MaxElts[j] < EltNo)
5359 // Or record a new source if not...
5361 SourceVecs.push_back(SourceVec);
5362 MinElts.push_back(EltNo);
5363 MaxElts.push_back(EltNo);
5367 // Currently only do something sane when at most two source vectors
5369 if (SourceVecs.size() > 2)
5372 SDValue ShuffleSrcs[2] = {DAG.getUNDEF(VT), DAG.getUNDEF(VT) };
5373 int VEXTOffsets[2] = {0, 0};
5375 // This loop extracts the usage patterns of the source vectors
5376 // and prepares appropriate SDValues for a shuffle if possible.
5377 for (unsigned i = 0; i < SourceVecs.size(); ++i) {
5378 if (SourceVecs[i].getValueType() == VT) {
5379 // No VEXT necessary
5380 ShuffleSrcs[i] = SourceVecs[i];
5383 } else if (SourceVecs[i].getValueType().getVectorNumElements() < NumElts) {
5384 // It probably isn't worth padding out a smaller vector just to
5385 // break it down again in a shuffle.
5389 // Since only 64-bit and 128-bit vectors are legal on ARM and
5390 // we've eliminated the other cases...
5391 assert(SourceVecs[i].getValueType().getVectorNumElements() == 2*NumElts &&
5392 "unexpected vector sizes in ReconstructShuffle");
5394 if (MaxElts[i] - MinElts[i] >= NumElts) {
5395 // Span too large for a VEXT to cope
5399 if (MinElts[i] >= NumElts) {
5400 // The extraction can just take the second half
5401 VEXTOffsets[i] = NumElts;
5402 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5404 DAG.getIntPtrConstant(NumElts, dl));
5405 } else if (MaxElts[i] < NumElts) {
5406 // The extraction can just take the first half
5408 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5410 DAG.getIntPtrConstant(0, dl));
5412 // An actual VEXT is needed
5413 VEXTOffsets[i] = MinElts[i];
5414 SDValue VEXTSrc1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5416 DAG.getIntPtrConstant(0, dl));
5417 SDValue VEXTSrc2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5419 DAG.getIntPtrConstant(NumElts, dl));
5420 ShuffleSrcs[i] = DAG.getNode(ARMISD::VEXT, dl, VT, VEXTSrc1, VEXTSrc2,
5421 DAG.getConstant(VEXTOffsets[i], dl,
5426 SmallVector<int, 8> Mask;
5428 for (unsigned i = 0; i < NumElts; ++i) {
5429 SDValue Entry = Op.getOperand(i);
5430 if (Entry.getOpcode() == ISD::UNDEF) {
5435 SDValue ExtractVec = Entry.getOperand(0);
5436 int ExtractElt = cast<ConstantSDNode>(Op.getOperand(i)
5437 .getOperand(1))->getSExtValue();
5438 if (ExtractVec == SourceVecs[0]) {
5439 Mask.push_back(ExtractElt - VEXTOffsets[0]);
5441 Mask.push_back(ExtractElt + NumElts - VEXTOffsets[1]);
5445 // Final check before we try to produce nonsense...
5446 if (isShuffleMaskLegal(Mask, VT))
5447 return DAG.getVectorShuffle(VT, dl, ShuffleSrcs[0], ShuffleSrcs[1],
5453 /// isShuffleMaskLegal - Targets can use this to indicate that they only
5454 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
5455 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
5456 /// are assumed to be legal.
5458 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
5460 if (VT.getVectorNumElements() == 4 &&
5461 (VT.is128BitVector() || VT.is64BitVector())) {
5462 unsigned PFIndexes[4];
5463 for (unsigned i = 0; i != 4; ++i) {
5467 PFIndexes[i] = M[i];
5470 // Compute the index in the perfect shuffle table.
5471 unsigned PFTableIndex =
5472 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
5473 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
5474 unsigned Cost = (PFEntry >> 30);
5481 unsigned Imm, WhichResult;
5483 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5484 return (EltSize >= 32 ||
5485 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
5486 isVREVMask(M, VT, 64) ||
5487 isVREVMask(M, VT, 32) ||
5488 isVREVMask(M, VT, 16) ||
5489 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
5490 isVTBLMask(M, VT) ||
5491 isVTRNMask(M, VT, WhichResult) ||
5492 isVUZPMask(M, VT, WhichResult) ||
5493 isVZIPMask(M, VT, WhichResult) ||
5494 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
5495 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
5496 isVZIP_v_undef_Mask(M, VT, WhichResult) ||
5497 ((VT == MVT::v8i16 || VT == MVT::v16i8) && isReverseMask(M, VT)));
5500 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
5501 /// the specified operations to build the shuffle.
5502 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
5503 SDValue RHS, SelectionDAG &DAG,
5505 unsigned OpNum = (PFEntry >> 26) & 0x0F;
5506 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
5507 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
5510 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
5519 OP_VUZPL, // VUZP, left result
5520 OP_VUZPR, // VUZP, right result
5521 OP_VZIPL, // VZIP, left result
5522 OP_VZIPR, // VZIP, right result
5523 OP_VTRNL, // VTRN, left result
5524 OP_VTRNR // VTRN, right result
5527 if (OpNum == OP_COPY) {
5528 if (LHSID == (1*9+2)*9+3) return LHS;
5529 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
5533 SDValue OpLHS, OpRHS;
5534 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
5535 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
5536 EVT VT = OpLHS.getValueType();
5539 default: llvm_unreachable("Unknown shuffle opcode!");
5541 // VREV divides the vector in half and swaps within the half.
5542 if (VT.getVectorElementType() == MVT::i32 ||
5543 VT.getVectorElementType() == MVT::f32)
5544 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
5545 // vrev <4 x i16> -> VREV32
5546 if (VT.getVectorElementType() == MVT::i16)
5547 return DAG.getNode(ARMISD::VREV32, dl, VT, OpLHS);
5548 // vrev <4 x i8> -> VREV16
5549 assert(VT.getVectorElementType() == MVT::i8);
5550 return DAG.getNode(ARMISD::VREV16, dl, VT, OpLHS);
5555 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
5556 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, dl, MVT::i32));
5560 return DAG.getNode(ARMISD::VEXT, dl, VT,
5562 DAG.getConstant(OpNum - OP_VEXT1 + 1, dl, MVT::i32));
5565 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
5566 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
5569 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
5570 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
5573 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
5574 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
5578 static SDValue LowerVECTOR_SHUFFLEv8i8(SDValue Op,
5579 ArrayRef<int> ShuffleMask,
5580 SelectionDAG &DAG) {
5581 // Check to see if we can use the VTBL instruction.
5582 SDValue V1 = Op.getOperand(0);
5583 SDValue V2 = Op.getOperand(1);
5586 SmallVector<SDValue, 8> VTBLMask;
5587 for (ArrayRef<int>::iterator
5588 I = ShuffleMask.begin(), E = ShuffleMask.end(); I != E; ++I)
5589 VTBLMask.push_back(DAG.getConstant(*I, DL, MVT::i32));
5591 if (V2.getNode()->getOpcode() == ISD::UNDEF)
5592 return DAG.getNode(ARMISD::VTBL1, DL, MVT::v8i8, V1,
5593 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8, VTBLMask));
5595 return DAG.getNode(ARMISD::VTBL2, DL, MVT::v8i8, V1, V2,
5596 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8, VTBLMask));
5599 static SDValue LowerReverse_VECTOR_SHUFFLEv16i8_v8i16(SDValue Op,
5600 SelectionDAG &DAG) {
5602 SDValue OpLHS = Op.getOperand(0);
5603 EVT VT = OpLHS.getValueType();
5605 assert((VT == MVT::v8i16 || VT == MVT::v16i8) &&
5606 "Expect an v8i16/v16i8 type");
5607 OpLHS = DAG.getNode(ARMISD::VREV64, DL, VT, OpLHS);
5608 // For a v16i8 type: After the VREV, we have got <8, ...15, 8, ..., 0>. Now,
5609 // extract the first 8 bytes into the top double word and the last 8 bytes
5610 // into the bottom double word. The v8i16 case is similar.
5611 unsigned ExtractNum = (VT == MVT::v16i8) ? 8 : 4;
5612 return DAG.getNode(ARMISD::VEXT, DL, VT, OpLHS, OpLHS,
5613 DAG.getConstant(ExtractNum, DL, MVT::i32));
5616 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
5617 SDValue V1 = Op.getOperand(0);
5618 SDValue V2 = Op.getOperand(1);
5620 EVT VT = Op.getValueType();
5621 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
5623 // Convert shuffles that are directly supported on NEON to target-specific
5624 // DAG nodes, instead of keeping them as shuffles and matching them again
5625 // during code selection. This is more efficient and avoids the possibility
5626 // of inconsistencies between legalization and selection.
5627 // FIXME: floating-point vectors should be canonicalized to integer vectors
5628 // of the same time so that they get CSEd properly.
5629 ArrayRef<int> ShuffleMask = SVN->getMask();
5631 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5632 if (EltSize <= 32) {
5633 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
5634 int Lane = SVN->getSplatIndex();
5635 // If this is undef splat, generate it via "just" vdup, if possible.
5636 if (Lane == -1) Lane = 0;
5638 // Test if V1 is a SCALAR_TO_VECTOR.
5639 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
5640 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
5642 // Test if V1 is a BUILD_VECTOR which is equivalent to a SCALAR_TO_VECTOR
5643 // (and probably will turn into a SCALAR_TO_VECTOR once legalization
5645 if (Lane == 0 && V1.getOpcode() == ISD::BUILD_VECTOR &&
5646 !isa<ConstantSDNode>(V1.getOperand(0))) {
5647 bool IsScalarToVector = true;
5648 for (unsigned i = 1, e = V1.getNumOperands(); i != e; ++i)
5649 if (V1.getOperand(i).getOpcode() != ISD::UNDEF) {
5650 IsScalarToVector = false;
5653 if (IsScalarToVector)
5654 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
5656 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
5657 DAG.getConstant(Lane, dl, MVT::i32));
5662 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
5665 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
5666 DAG.getConstant(Imm, dl, MVT::i32));
5669 if (isVREVMask(ShuffleMask, VT, 64))
5670 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
5671 if (isVREVMask(ShuffleMask, VT, 32))
5672 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
5673 if (isVREVMask(ShuffleMask, VT, 16))
5674 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
5676 if (V2->getOpcode() == ISD::UNDEF &&
5677 isSingletonVEXTMask(ShuffleMask, VT, Imm)) {
5678 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V1,
5679 DAG.getConstant(Imm, dl, MVT::i32));
5682 // Check for Neon shuffles that modify both input vectors in place.
5683 // If both results are used, i.e., if there are two shuffles with the same
5684 // source operands and with masks corresponding to both results of one of
5685 // these operations, DAG memoization will ensure that a single node is
5686 // used for both shuffles.
5687 unsigned WhichResult;
5688 if (isVTRNMask(ShuffleMask, VT, WhichResult))
5689 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
5690 V1, V2).getValue(WhichResult);
5691 if (isVUZPMask(ShuffleMask, VT, WhichResult))
5692 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
5693 V1, V2).getValue(WhichResult);
5694 if (isVZIPMask(ShuffleMask, VT, WhichResult))
5695 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
5696 V1, V2).getValue(WhichResult);
5698 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
5699 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
5700 V1, V1).getValue(WhichResult);
5701 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
5702 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
5703 V1, V1).getValue(WhichResult);
5704 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
5705 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
5706 V1, V1).getValue(WhichResult);
5709 // If the shuffle is not directly supported and it has 4 elements, use
5710 // the PerfectShuffle-generated table to synthesize it from other shuffles.
5711 unsigned NumElts = VT.getVectorNumElements();
5713 unsigned PFIndexes[4];
5714 for (unsigned i = 0; i != 4; ++i) {
5715 if (ShuffleMask[i] < 0)
5718 PFIndexes[i] = ShuffleMask[i];
5721 // Compute the index in the perfect shuffle table.
5722 unsigned PFTableIndex =
5723 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
5724 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
5725 unsigned Cost = (PFEntry >> 30);
5728 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
5731 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
5732 if (EltSize >= 32) {
5733 // Do the expansion with floating-point types, since that is what the VFP
5734 // registers are defined to use, and since i64 is not legal.
5735 EVT EltVT = EVT::getFloatingPointVT(EltSize);
5736 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
5737 V1 = DAG.getNode(ISD::BITCAST, dl, VecVT, V1);
5738 V2 = DAG.getNode(ISD::BITCAST, dl, VecVT, V2);
5739 SmallVector<SDValue, 8> Ops;
5740 for (unsigned i = 0; i < NumElts; ++i) {
5741 if (ShuffleMask[i] < 0)
5742 Ops.push_back(DAG.getUNDEF(EltVT));
5744 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
5745 ShuffleMask[i] < (int)NumElts ? V1 : V2,
5746 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
5749 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, Ops);
5750 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
5753 if ((VT == MVT::v8i16 || VT == MVT::v16i8) && isReverseMask(ShuffleMask, VT))
5754 return LowerReverse_VECTOR_SHUFFLEv16i8_v8i16(Op, DAG);
5756 if (VT == MVT::v8i8) {
5757 SDValue NewOp = LowerVECTOR_SHUFFLEv8i8(Op, ShuffleMask, DAG);
5758 if (NewOp.getNode())
5765 static SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
5766 // INSERT_VECTOR_ELT is legal only for immediate indexes.
5767 SDValue Lane = Op.getOperand(2);
5768 if (!isa<ConstantSDNode>(Lane))
5774 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
5775 // EXTRACT_VECTOR_ELT is legal only for immediate indexes.
5776 SDValue Lane = Op.getOperand(1);
5777 if (!isa<ConstantSDNode>(Lane))
5780 SDValue Vec = Op.getOperand(0);
5781 if (Op.getValueType() == MVT::i32 &&
5782 Vec.getValueType().getVectorElementType().getSizeInBits() < 32) {
5784 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
5790 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
5791 // The only time a CONCAT_VECTORS operation can have legal types is when
5792 // two 64-bit vectors are concatenated to a 128-bit vector.
5793 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
5794 "unexpected CONCAT_VECTORS");
5796 SDValue Val = DAG.getUNDEF(MVT::v2f64);
5797 SDValue Op0 = Op.getOperand(0);
5798 SDValue Op1 = Op.getOperand(1);
5799 if (Op0.getOpcode() != ISD::UNDEF)
5800 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
5801 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op0),
5802 DAG.getIntPtrConstant(0, dl));
5803 if (Op1.getOpcode() != ISD::UNDEF)
5804 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
5805 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op1),
5806 DAG.getIntPtrConstant(1, dl));
5807 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Val);
5810 /// isExtendedBUILD_VECTOR - Check if N is a constant BUILD_VECTOR where each
5811 /// element has been zero/sign-extended, depending on the isSigned parameter,
5812 /// from an integer type half its size.
5813 static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG,
5815 // A v2i64 BUILD_VECTOR will have been legalized to a BITCAST from v4i32.
5816 EVT VT = N->getValueType(0);
5817 if (VT == MVT::v2i64 && N->getOpcode() == ISD::BITCAST) {
5818 SDNode *BVN = N->getOperand(0).getNode();
5819 if (BVN->getValueType(0) != MVT::v4i32 ||
5820 BVN->getOpcode() != ISD::BUILD_VECTOR)
5822 unsigned LoElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
5823 unsigned HiElt = 1 - LoElt;
5824 ConstantSDNode *Lo0 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt));
5825 ConstantSDNode *Hi0 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt));
5826 ConstantSDNode *Lo1 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt+2));
5827 ConstantSDNode *Hi1 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt+2));
5828 if (!Lo0 || !Hi0 || !Lo1 || !Hi1)
5831 if (Hi0->getSExtValue() == Lo0->getSExtValue() >> 32 &&
5832 Hi1->getSExtValue() == Lo1->getSExtValue() >> 32)
5835 if (Hi0->isNullValue() && Hi1->isNullValue())
5841 if (N->getOpcode() != ISD::BUILD_VECTOR)
5844 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
5845 SDNode *Elt = N->getOperand(i).getNode();
5846 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
5847 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5848 unsigned HalfSize = EltSize / 2;
5850 if (!isIntN(HalfSize, C->getSExtValue()))
5853 if (!isUIntN(HalfSize, C->getZExtValue()))
5864 /// isSignExtended - Check if a node is a vector value that is sign-extended
5865 /// or a constant BUILD_VECTOR with sign-extended elements.
5866 static bool isSignExtended(SDNode *N, SelectionDAG &DAG) {
5867 if (N->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N))
5869 if (isExtendedBUILD_VECTOR(N, DAG, true))
5874 /// isZeroExtended - Check if a node is a vector value that is zero-extended
5875 /// or a constant BUILD_VECTOR with zero-extended elements.
5876 static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) {
5877 if (N->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N))
5879 if (isExtendedBUILD_VECTOR(N, DAG, false))
5884 static EVT getExtensionTo64Bits(const EVT &OrigVT) {
5885 if (OrigVT.getSizeInBits() >= 64)
5888 assert(OrigVT.isSimple() && "Expecting a simple value type");
5890 MVT::SimpleValueType OrigSimpleTy = OrigVT.getSimpleVT().SimpleTy;
5891 switch (OrigSimpleTy) {
5892 default: llvm_unreachable("Unexpected Vector Type");
5901 /// AddRequiredExtensionForVMULL - Add a sign/zero extension to extend the total
5902 /// value size to 64 bits. We need a 64-bit D register as an operand to VMULL.
5903 /// We insert the required extension here to get the vector to fill a D register.
5904 static SDValue AddRequiredExtensionForVMULL(SDValue N, SelectionDAG &DAG,
5907 unsigned ExtOpcode) {
5908 // The vector originally had a size of OrigTy. It was then extended to ExtTy.
5909 // We expect the ExtTy to be 128-bits total. If the OrigTy is less than
5910 // 64-bits we need to insert a new extension so that it will be 64-bits.
5911 assert(ExtTy.is128BitVector() && "Unexpected extension size");
5912 if (OrigTy.getSizeInBits() >= 64)
5915 // Must extend size to at least 64 bits to be used as an operand for VMULL.
5916 EVT NewVT = getExtensionTo64Bits(OrigTy);
5918 return DAG.getNode(ExtOpcode, SDLoc(N), NewVT, N);
5921 /// SkipLoadExtensionForVMULL - return a load of the original vector size that
5922 /// does not do any sign/zero extension. If the original vector is less
5923 /// than 64 bits, an appropriate extension will be added after the load to
5924 /// reach a total size of 64 bits. We have to add the extension separately
5925 /// because ARM does not have a sign/zero extending load for vectors.
5926 static SDValue SkipLoadExtensionForVMULL(LoadSDNode *LD, SelectionDAG& DAG) {
5927 EVT ExtendedTy = getExtensionTo64Bits(LD->getMemoryVT());
5929 // The load already has the right type.
5930 if (ExtendedTy == LD->getMemoryVT())
5931 return DAG.getLoad(LD->getMemoryVT(), SDLoc(LD), LD->getChain(),
5932 LD->getBasePtr(), LD->getPointerInfo(), LD->isVolatile(),
5933 LD->isNonTemporal(), LD->isInvariant(),
5934 LD->getAlignment());
5936 // We need to create a zextload/sextload. We cannot just create a load
5937 // followed by a zext/zext node because LowerMUL is also run during normal
5938 // operation legalization where we can't create illegal types.
5939 return DAG.getExtLoad(LD->getExtensionType(), SDLoc(LD), ExtendedTy,
5940 LD->getChain(), LD->getBasePtr(), LD->getPointerInfo(),
5941 LD->getMemoryVT(), LD->isVolatile(), LD->isInvariant(),
5942 LD->isNonTemporal(), LD->getAlignment());
5945 /// SkipExtensionForVMULL - For a node that is a SIGN_EXTEND, ZERO_EXTEND,
5946 /// extending load, or BUILD_VECTOR with extended elements, return the
5947 /// unextended value. The unextended vector should be 64 bits so that it can
5948 /// be used as an operand to a VMULL instruction. If the original vector size
5949 /// before extension is less than 64 bits we add a an extension to resize
5950 /// the vector to 64 bits.
5951 static SDValue SkipExtensionForVMULL(SDNode *N, SelectionDAG &DAG) {
5952 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
5953 return AddRequiredExtensionForVMULL(N->getOperand(0), DAG,
5954 N->getOperand(0)->getValueType(0),
5958 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
5959 return SkipLoadExtensionForVMULL(LD, DAG);
5961 // Otherwise, the value must be a BUILD_VECTOR. For v2i64, it will
5962 // have been legalized as a BITCAST from v4i32.
5963 if (N->getOpcode() == ISD::BITCAST) {
5964 SDNode *BVN = N->getOperand(0).getNode();
5965 assert(BVN->getOpcode() == ISD::BUILD_VECTOR &&
5966 BVN->getValueType(0) == MVT::v4i32 && "expected v4i32 BUILD_VECTOR");
5967 unsigned LowElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
5968 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), MVT::v2i32,
5969 BVN->getOperand(LowElt), BVN->getOperand(LowElt+2));
5971 // Construct a new BUILD_VECTOR with elements truncated to half the size.
5972 assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR");
5973 EVT VT = N->getValueType(0);
5974 unsigned EltSize = VT.getVectorElementType().getSizeInBits() / 2;
5975 unsigned NumElts = VT.getVectorNumElements();
5976 MVT TruncVT = MVT::getIntegerVT(EltSize);
5977 SmallVector<SDValue, 8> Ops;
5979 for (unsigned i = 0; i != NumElts; ++i) {
5980 ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i));
5981 const APInt &CInt = C->getAPIntValue();
5982 // Element types smaller than 32 bits are not legal, so use i32 elements.
5983 // The values are implicitly truncated so sext vs. zext doesn't matter.
5984 Ops.push_back(DAG.getConstant(CInt.zextOrTrunc(32), dl, MVT::i32));
5986 return DAG.getNode(ISD::BUILD_VECTOR, dl,
5987 MVT::getVectorVT(TruncVT, NumElts), Ops);
5990 static bool isAddSubSExt(SDNode *N, SelectionDAG &DAG) {
5991 unsigned Opcode = N->getOpcode();
5992 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
5993 SDNode *N0 = N->getOperand(0).getNode();
5994 SDNode *N1 = N->getOperand(1).getNode();
5995 return N0->hasOneUse() && N1->hasOneUse() &&
5996 isSignExtended(N0, DAG) && isSignExtended(N1, DAG);
6001 static bool isAddSubZExt(SDNode *N, SelectionDAG &DAG) {
6002 unsigned Opcode = N->getOpcode();
6003 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
6004 SDNode *N0 = N->getOperand(0).getNode();
6005 SDNode *N1 = N->getOperand(1).getNode();
6006 return N0->hasOneUse() && N1->hasOneUse() &&
6007 isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG);
6012 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
6013 // Multiplications are only custom-lowered for 128-bit vectors so that
6014 // VMULL can be detected. Otherwise v2i64 multiplications are not legal.
6015 EVT VT = Op.getValueType();
6016 assert(VT.is128BitVector() && VT.isInteger() &&
6017 "unexpected type for custom-lowering ISD::MUL");
6018 SDNode *N0 = Op.getOperand(0).getNode();
6019 SDNode *N1 = Op.getOperand(1).getNode();
6020 unsigned NewOpc = 0;
6022 bool isN0SExt = isSignExtended(N0, DAG);
6023 bool isN1SExt = isSignExtended(N1, DAG);
6024 if (isN0SExt && isN1SExt)
6025 NewOpc = ARMISD::VMULLs;
6027 bool isN0ZExt = isZeroExtended(N0, DAG);
6028 bool isN1ZExt = isZeroExtended(N1, DAG);
6029 if (isN0ZExt && isN1ZExt)
6030 NewOpc = ARMISD::VMULLu;
6031 else if (isN1SExt || isN1ZExt) {
6032 // Look for (s/zext A + s/zext B) * (s/zext C). We want to turn these
6033 // into (s/zext A * s/zext C) + (s/zext B * s/zext C)
6034 if (isN1SExt && isAddSubSExt(N0, DAG)) {
6035 NewOpc = ARMISD::VMULLs;
6037 } else if (isN1ZExt && isAddSubZExt(N0, DAG)) {
6038 NewOpc = ARMISD::VMULLu;
6040 } else if (isN0ZExt && isAddSubZExt(N1, DAG)) {
6042 NewOpc = ARMISD::VMULLu;
6048 if (VT == MVT::v2i64)
6049 // Fall through to expand this. It is not legal.
6052 // Other vector multiplications are legal.
6057 // Legalize to a VMULL instruction.
6060 SDValue Op1 = SkipExtensionForVMULL(N1, DAG);
6062 Op0 = SkipExtensionForVMULL(N0, DAG);
6063 assert(Op0.getValueType().is64BitVector() &&
6064 Op1.getValueType().is64BitVector() &&
6065 "unexpected types for extended operands to VMULL");
6066 return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
6069 // Optimizing (zext A + zext B) * C, to (VMULL A, C) + (VMULL B, C) during
6070 // isel lowering to take advantage of no-stall back to back vmul + vmla.
6077 SDValue N00 = SkipExtensionForVMULL(N0->getOperand(0).getNode(), DAG);
6078 SDValue N01 = SkipExtensionForVMULL(N0->getOperand(1).getNode(), DAG);
6079 EVT Op1VT = Op1.getValueType();
6080 return DAG.getNode(N0->getOpcode(), DL, VT,
6081 DAG.getNode(NewOpc, DL, VT,
6082 DAG.getNode(ISD::BITCAST, DL, Op1VT, N00), Op1),
6083 DAG.getNode(NewOpc, DL, VT,
6084 DAG.getNode(ISD::BITCAST, DL, Op1VT, N01), Op1));
6088 LowerSDIV_v4i8(SDValue X, SDValue Y, SDLoc dl, SelectionDAG &DAG) {
6090 // float4 xf = vcvt_f32_s32(vmovl_s16(a.lo));
6091 // float4 yf = vcvt_f32_s32(vmovl_s16(b.lo));
6092 X = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, X);
6093 Y = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, Y);
6094 X = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, X);
6095 Y = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, Y);
6096 // Get reciprocal estimate.
6097 // float4 recip = vrecpeq_f32(yf);
6098 Y = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6099 DAG.getConstant(Intrinsic::arm_neon_vrecpe, dl, MVT::i32),
6101 // Because char has a smaller range than uchar, we can actually get away
6102 // without any newton steps. This requires that we use a weird bias
6103 // of 0xb000, however (again, this has been exhaustively tested).
6104 // float4 result = as_float4(as_int4(xf*recip) + 0xb000);
6105 X = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, X, Y);
6106 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, X);
6107 Y = DAG.getConstant(0xb000, dl, MVT::i32);
6108 Y = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Y, Y, Y, Y);
6109 X = DAG.getNode(ISD::ADD, dl, MVT::v4i32, X, Y);
6110 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, X);
6111 // Convert back to short.
6112 X = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, X);
6113 X = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, X);
6118 LowerSDIV_v4i16(SDValue N0, SDValue N1, SDLoc dl, SelectionDAG &DAG) {
6120 // Convert to float.
6121 // float4 yf = vcvt_f32_s32(vmovl_s16(y));
6122 // float4 xf = vcvt_f32_s32(vmovl_s16(x));
6123 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N0);
6124 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N1);
6125 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
6126 N1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
6128 // Use reciprocal estimate and one refinement step.
6129 // float4 recip = vrecpeq_f32(yf);
6130 // recip *= vrecpsq_f32(yf, recip);
6131 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6132 DAG.getConstant(Intrinsic::arm_neon_vrecpe, dl, MVT::i32),
6134 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6135 DAG.getConstant(Intrinsic::arm_neon_vrecps, dl, MVT::i32),
6137 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
6138 // Because short has a smaller range than ushort, we can actually get away
6139 // with only a single newton step. This requires that we use a weird bias
6140 // of 89, however (again, this has been exhaustively tested).
6141 // float4 result = as_float4(as_int4(xf*recip) + 0x89);
6142 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
6143 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
6144 N1 = DAG.getConstant(0x89, dl, MVT::i32);
6145 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
6146 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
6147 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
6148 // Convert back to integer and return.
6149 // return vmovn_s32(vcvt_s32_f32(result));
6150 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
6151 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
6155 static SDValue LowerSDIV(SDValue Op, SelectionDAG &DAG) {
6156 EVT VT = Op.getValueType();
6157 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
6158 "unexpected type for custom-lowering ISD::SDIV");
6161 SDValue N0 = Op.getOperand(0);
6162 SDValue N1 = Op.getOperand(1);
6165 if (VT == MVT::v8i8) {
6166 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N0);
6167 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N1);
6169 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
6170 DAG.getIntPtrConstant(4, dl));
6171 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
6172 DAG.getIntPtrConstant(4, dl));
6173 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
6174 DAG.getIntPtrConstant(0, dl));
6175 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
6176 DAG.getIntPtrConstant(0, dl));
6178 N0 = LowerSDIV_v4i8(N0, N1, dl, DAG); // v4i16
6179 N2 = LowerSDIV_v4i8(N2, N3, dl, DAG); // v4i16
6181 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
6182 N0 = LowerCONCAT_VECTORS(N0, DAG);
6184 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v8i8, N0);
6187 return LowerSDIV_v4i16(N0, N1, dl, DAG);
6190 static SDValue LowerUDIV(SDValue Op, SelectionDAG &DAG) {
6191 EVT VT = Op.getValueType();
6192 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
6193 "unexpected type for custom-lowering ISD::UDIV");
6196 SDValue N0 = Op.getOperand(0);
6197 SDValue N1 = Op.getOperand(1);
6200 if (VT == MVT::v8i8) {
6201 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N0);
6202 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N1);
6204 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
6205 DAG.getIntPtrConstant(4, dl));
6206 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
6207 DAG.getIntPtrConstant(4, dl));
6208 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
6209 DAG.getIntPtrConstant(0, dl));
6210 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
6211 DAG.getIntPtrConstant(0, dl));
6213 N0 = LowerSDIV_v4i16(N0, N1, dl, DAG); // v4i16
6214 N2 = LowerSDIV_v4i16(N2, N3, dl, DAG); // v4i16
6216 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
6217 N0 = LowerCONCAT_VECTORS(N0, DAG);
6219 N0 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v8i8,
6220 DAG.getConstant(Intrinsic::arm_neon_vqmovnsu, dl,
6226 // v4i16 sdiv ... Convert to float.
6227 // float4 yf = vcvt_f32_s32(vmovl_u16(y));
6228 // float4 xf = vcvt_f32_s32(vmovl_u16(x));
6229 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N0);
6230 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N1);
6231 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
6232 SDValue BN1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
6234 // Use reciprocal estimate and two refinement steps.
6235 // float4 recip = vrecpeq_f32(yf);
6236 // recip *= vrecpsq_f32(yf, recip);
6237 // recip *= vrecpsq_f32(yf, recip);
6238 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6239 DAG.getConstant(Intrinsic::arm_neon_vrecpe, dl, MVT::i32),
6241 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6242 DAG.getConstant(Intrinsic::arm_neon_vrecps, dl, MVT::i32),
6244 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
6245 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6246 DAG.getConstant(Intrinsic::arm_neon_vrecps, dl, MVT::i32),
6248 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
6249 // Simply multiplying by the reciprocal estimate can leave us a few ulps
6250 // too low, so we add 2 ulps (exhaustive testing shows that this is enough,
6251 // and that it will never cause us to return an answer too large).
6252 // float4 result = as_float4(as_int4(xf*recip) + 2);
6253 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
6254 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
6255 N1 = DAG.getConstant(2, dl, MVT::i32);
6256 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
6257 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
6258 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
6259 // Convert back to integer and return.
6260 // return vmovn_u32(vcvt_s32_f32(result));
6261 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
6262 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
6266 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
6267 EVT VT = Op.getNode()->getValueType(0);
6268 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
6271 bool ExtraOp = false;
6272 switch (Op.getOpcode()) {
6273 default: llvm_unreachable("Invalid code");
6274 case ISD::ADDC: Opc = ARMISD::ADDC; break;
6275 case ISD::ADDE: Opc = ARMISD::ADDE; ExtraOp = true; break;
6276 case ISD::SUBC: Opc = ARMISD::SUBC; break;
6277 case ISD::SUBE: Opc = ARMISD::SUBE; ExtraOp = true; break;
6281 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
6283 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
6284 Op.getOperand(1), Op.getOperand(2));
6287 SDValue ARMTargetLowering::LowerFSINCOS(SDValue Op, SelectionDAG &DAG) const {
6288 assert(Subtarget->isTargetDarwin());
6290 // For iOS, we want to call an alternative entry point: __sincos_stret,
6291 // return values are passed via sret.
6293 SDValue Arg = Op.getOperand(0);
6294 EVT ArgVT = Arg.getValueType();
6295 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
6297 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
6298 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6300 // Pair of floats / doubles used to pass the result.
6301 StructType *RetTy = StructType::get(ArgTy, ArgTy, nullptr);
6303 // Create stack object for sret.
6304 const uint64_t ByteSize = TLI.getDataLayout()->getTypeAllocSize(RetTy);
6305 const unsigned StackAlign = TLI.getDataLayout()->getPrefTypeAlignment(RetTy);
6306 int FrameIdx = FrameInfo->CreateStackObject(ByteSize, StackAlign, false);
6307 SDValue SRet = DAG.getFrameIndex(FrameIdx, TLI.getPointerTy());
6313 Entry.Ty = RetTy->getPointerTo();
6314 Entry.isSExt = false;
6315 Entry.isZExt = false;
6316 Entry.isSRet = true;
6317 Args.push_back(Entry);
6321 Entry.isSExt = false;
6322 Entry.isZExt = false;
6323 Args.push_back(Entry);
6325 const char *LibcallName = (ArgVT == MVT::f64)
6326 ? "__sincos_stret" : "__sincosf_stret";
6327 SDValue Callee = DAG.getExternalSymbol(LibcallName, getPointerTy());
6329 TargetLowering::CallLoweringInfo CLI(DAG);
6330 CLI.setDebugLoc(dl).setChain(DAG.getEntryNode())
6331 .setCallee(CallingConv::C, Type::getVoidTy(*DAG.getContext()), Callee,
6333 .setDiscardResult();
6335 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
6337 SDValue LoadSin = DAG.getLoad(ArgVT, dl, CallResult.second, SRet,
6338 MachinePointerInfo(), false, false, false, 0);
6340 // Address of cos field.
6341 SDValue Add = DAG.getNode(ISD::ADD, dl, getPointerTy(), SRet,
6342 DAG.getIntPtrConstant(ArgVT.getStoreSize(), dl));
6343 SDValue LoadCos = DAG.getLoad(ArgVT, dl, LoadSin.getValue(1), Add,
6344 MachinePointerInfo(), false, false, false, 0);
6346 SDVTList Tys = DAG.getVTList(ArgVT, ArgVT);
6347 return DAG.getNode(ISD::MERGE_VALUES, dl, Tys,
6348 LoadSin.getValue(0), LoadCos.getValue(0));
6351 static SDValue LowerAtomicLoadStore(SDValue Op, SelectionDAG &DAG) {
6352 // Monotonic load/store is legal for all targets
6353 if (cast<AtomicSDNode>(Op)->getOrdering() <= Monotonic)
6356 // Acquire/Release load/store is not legal for targets without a
6357 // dmb or equivalent available.
6361 static void ReplaceREADCYCLECOUNTER(SDNode *N,
6362 SmallVectorImpl<SDValue> &Results,
6364 const ARMSubtarget *Subtarget) {
6366 SDValue Cycles32, OutChain;
6368 if (Subtarget->hasPerfMon()) {
6369 // Under Power Management extensions, the cycle-count is:
6370 // mrc p15, #0, <Rt>, c9, c13, #0
6371 SDValue Ops[] = { N->getOperand(0), // Chain
6372 DAG.getConstant(Intrinsic::arm_mrc, DL, MVT::i32),
6373 DAG.getConstant(15, DL, MVT::i32),
6374 DAG.getConstant(0, DL, MVT::i32),
6375 DAG.getConstant(9, DL, MVT::i32),
6376 DAG.getConstant(13, DL, MVT::i32),
6377 DAG.getConstant(0, DL, MVT::i32)
6380 Cycles32 = DAG.getNode(ISD::INTRINSIC_W_CHAIN, DL,
6381 DAG.getVTList(MVT::i32, MVT::Other), Ops);
6382 OutChain = Cycles32.getValue(1);
6384 // Intrinsic is defined to return 0 on unsupported platforms. Technically
6385 // there are older ARM CPUs that have implementation-specific ways of
6386 // obtaining this information (FIXME!).
6387 Cycles32 = DAG.getConstant(0, DL, MVT::i32);
6388 OutChain = DAG.getEntryNode();
6392 SDValue Cycles64 = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64,
6393 Cycles32, DAG.getConstant(0, DL, MVT::i32));
6394 Results.push_back(Cycles64);
6395 Results.push_back(OutChain);
6398 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
6399 switch (Op.getOpcode()) {
6400 default: llvm_unreachable("Don't know how to custom lower this!");
6401 case ISD::WRITE_REGISTER: return LowerWRITE_REGISTER(Op, DAG);
6402 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
6403 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
6404 case ISD::GlobalAddress:
6405 switch (Subtarget->getTargetTriple().getObjectFormat()) {
6406 default: llvm_unreachable("unknown object format");
6408 return LowerGlobalAddressWindows(Op, DAG);
6410 return LowerGlobalAddressELF(Op, DAG);
6412 return LowerGlobalAddressDarwin(Op, DAG);
6414 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
6415 case ISD::SELECT: return LowerSELECT(Op, DAG);
6416 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
6417 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
6418 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
6419 case ISD::VASTART: return LowerVASTART(Op, DAG);
6420 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, DAG, Subtarget);
6421 case ISD::PREFETCH: return LowerPREFETCH(Op, DAG, Subtarget);
6422 case ISD::SINT_TO_FP:
6423 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
6424 case ISD::FP_TO_SINT:
6425 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
6426 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
6427 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
6428 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
6429 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
6430 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
6431 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
6432 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
6434 case ISD::BITCAST: return ExpandBITCAST(Op.getNode(), DAG);
6437 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
6438 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
6439 case ISD::SRL_PARTS:
6440 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
6441 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
6442 case ISD::CTPOP: return LowerCTPOP(Op.getNode(), DAG, Subtarget);
6443 case ISD::SETCC: return LowerVSETCC(Op, DAG);
6444 case ISD::ConstantFP: return LowerConstantFP(Op, DAG, Subtarget);
6445 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG, Subtarget);
6446 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
6447 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
6448 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
6449 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
6450 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
6451 case ISD::MUL: return LowerMUL(Op, DAG);
6452 case ISD::SDIV: return LowerSDIV(Op, DAG);
6453 case ISD::UDIV: return LowerUDIV(Op, DAG);
6457 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
6462 return LowerXALUO(Op, DAG);
6463 case ISD::ATOMIC_LOAD:
6464 case ISD::ATOMIC_STORE: return LowerAtomicLoadStore(Op, DAG);
6465 case ISD::FSINCOS: return LowerFSINCOS(Op, DAG);
6467 case ISD::UDIVREM: return LowerDivRem(Op, DAG);
6468 case ISD::DYNAMIC_STACKALLOC:
6469 if (Subtarget->getTargetTriple().isWindowsItaniumEnvironment())
6470 return LowerDYNAMIC_STACKALLOC(Op, DAG);
6471 llvm_unreachable("Don't know how to custom lower this!");
6472 case ISD::FP_ROUND: return LowerFP_ROUND(Op, DAG);
6473 case ISD::FP_EXTEND: return LowerFP_EXTEND(Op, DAG);
6477 /// ReplaceNodeResults - Replace the results of node with an illegal result
6478 /// type with new values built out of custom code.
6479 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
6480 SmallVectorImpl<SDValue>&Results,
6481 SelectionDAG &DAG) const {
6483 switch (N->getOpcode()) {
6485 llvm_unreachable("Don't know how to custom expand this!");
6486 case ISD::READ_REGISTER:
6487 ExpandREAD_REGISTER(N, Results, DAG);
6490 Res = ExpandBITCAST(N, DAG);
6494 Res = Expand64BitShift(N, DAG, Subtarget);
6496 case ISD::READCYCLECOUNTER:
6497 ReplaceREADCYCLECOUNTER(N, Results, DAG, Subtarget);
6501 Results.push_back(Res);
6504 //===----------------------------------------------------------------------===//
6505 // ARM Scheduler Hooks
6506 //===----------------------------------------------------------------------===//
6508 /// SetupEntryBlockForSjLj - Insert code into the entry block that creates and
6509 /// registers the function context.
6510 void ARMTargetLowering::
6511 SetupEntryBlockForSjLj(MachineInstr *MI, MachineBasicBlock *MBB,
6512 MachineBasicBlock *DispatchBB, int FI) const {
6513 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
6514 DebugLoc dl = MI->getDebugLoc();
6515 MachineFunction *MF = MBB->getParent();
6516 MachineRegisterInfo *MRI = &MF->getRegInfo();
6517 MachineConstantPool *MCP = MF->getConstantPool();
6518 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
6519 const Function *F = MF->getFunction();
6521 bool isThumb = Subtarget->isThumb();
6522 bool isThumb2 = Subtarget->isThumb2();
6524 unsigned PCLabelId = AFI->createPICLabelUId();
6525 unsigned PCAdj = (isThumb || isThumb2) ? 4 : 8;
6526 ARMConstantPoolValue *CPV =
6527 ARMConstantPoolMBB::Create(F->getContext(), DispatchBB, PCLabelId, PCAdj);
6528 unsigned CPI = MCP->getConstantPoolIndex(CPV, 4);
6530 const TargetRegisterClass *TRC = isThumb ? &ARM::tGPRRegClass
6531 : &ARM::GPRRegClass;
6533 // Grab constant pool and fixed stack memory operands.
6534 MachineMemOperand *CPMMO =
6535 MF->getMachineMemOperand(MachinePointerInfo::getConstantPool(),
6536 MachineMemOperand::MOLoad, 4, 4);
6538 MachineMemOperand *FIMMOSt =
6539 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
6540 MachineMemOperand::MOStore, 4, 4);
6542 // Load the address of the dispatch MBB into the jump buffer.
6544 // Incoming value: jbuf
6545 // ldr.n r5, LCPI1_1
6548 // str r5, [$jbuf, #+4] ; &jbuf[1]
6549 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6550 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2LDRpci), NewVReg1)
6551 .addConstantPoolIndex(CPI)
6552 .addMemOperand(CPMMO));
6553 // Set the low bit because of thumb mode.
6554 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6556 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2ORRri), NewVReg2)
6557 .addReg(NewVReg1, RegState::Kill)
6559 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6560 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg3)
6561 .addReg(NewVReg2, RegState::Kill)
6563 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2STRi12))
6564 .addReg(NewVReg3, RegState::Kill)
6566 .addImm(36) // &jbuf[1] :: pc
6567 .addMemOperand(FIMMOSt));
6568 } else if (isThumb) {
6569 // Incoming value: jbuf
6570 // ldr.n r1, LCPI1_4
6574 // add r2, $jbuf, #+4 ; &jbuf[1]
6576 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6577 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tLDRpci), NewVReg1)
6578 .addConstantPoolIndex(CPI)
6579 .addMemOperand(CPMMO));
6580 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6581 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg2)
6582 .addReg(NewVReg1, RegState::Kill)
6584 // Set the low bit because of thumb mode.
6585 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6586 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tMOVi8), NewVReg3)
6587 .addReg(ARM::CPSR, RegState::Define)
6589 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6590 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tORR), NewVReg4)
6591 .addReg(ARM::CPSR, RegState::Define)
6592 .addReg(NewVReg2, RegState::Kill)
6593 .addReg(NewVReg3, RegState::Kill));
6594 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6595 BuildMI(*MBB, MI, dl, TII->get(ARM::tADDframe), NewVReg5)
6597 .addImm(36); // &jbuf[1] :: pc
6598 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tSTRi))
6599 .addReg(NewVReg4, RegState::Kill)
6600 .addReg(NewVReg5, RegState::Kill)
6602 .addMemOperand(FIMMOSt));
6604 // Incoming value: jbuf
6607 // str r1, [$jbuf, #+4] ; &jbuf[1]
6608 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6609 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::LDRi12), NewVReg1)
6610 .addConstantPoolIndex(CPI)
6612 .addMemOperand(CPMMO));
6613 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6614 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::PICADD), NewVReg2)
6615 .addReg(NewVReg1, RegState::Kill)
6616 .addImm(PCLabelId));
6617 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::STRi12))
6618 .addReg(NewVReg2, RegState::Kill)
6620 .addImm(36) // &jbuf[1] :: pc
6621 .addMemOperand(FIMMOSt));
6625 void ARMTargetLowering::EmitSjLjDispatchBlock(MachineInstr *MI,
6626 MachineBasicBlock *MBB) const {
6627 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
6628 DebugLoc dl = MI->getDebugLoc();
6629 MachineFunction *MF = MBB->getParent();
6630 MachineRegisterInfo *MRI = &MF->getRegInfo();
6631 MachineFrameInfo *MFI = MF->getFrameInfo();
6632 int FI = MFI->getFunctionContextIndex();
6634 const TargetRegisterClass *TRC = Subtarget->isThumb() ? &ARM::tGPRRegClass
6635 : &ARM::GPRnopcRegClass;
6637 // Get a mapping of the call site numbers to all of the landing pads they're
6639 DenseMap<unsigned, SmallVector<MachineBasicBlock*, 2> > CallSiteNumToLPad;
6640 unsigned MaxCSNum = 0;
6641 MachineModuleInfo &MMI = MF->getMMI();
6642 for (MachineFunction::iterator BB = MF->begin(), E = MF->end(); BB != E;
6644 if (!BB->isLandingPad()) continue;
6646 // FIXME: We should assert that the EH_LABEL is the first MI in the landing
6648 for (MachineBasicBlock::iterator
6649 II = BB->begin(), IE = BB->end(); II != IE; ++II) {
6650 if (!II->isEHLabel()) continue;
6652 MCSymbol *Sym = II->getOperand(0).getMCSymbol();
6653 if (!MMI.hasCallSiteLandingPad(Sym)) continue;
6655 SmallVectorImpl<unsigned> &CallSiteIdxs = MMI.getCallSiteLandingPad(Sym);
6656 for (SmallVectorImpl<unsigned>::iterator
6657 CSI = CallSiteIdxs.begin(), CSE = CallSiteIdxs.end();
6658 CSI != CSE; ++CSI) {
6659 CallSiteNumToLPad[*CSI].push_back(BB);
6660 MaxCSNum = std::max(MaxCSNum, *CSI);
6666 // Get an ordered list of the machine basic blocks for the jump table.
6667 std::vector<MachineBasicBlock*> LPadList;
6668 SmallPtrSet<MachineBasicBlock*, 64> InvokeBBs;
6669 LPadList.reserve(CallSiteNumToLPad.size());
6670 for (unsigned I = 1; I <= MaxCSNum; ++I) {
6671 SmallVectorImpl<MachineBasicBlock*> &MBBList = CallSiteNumToLPad[I];
6672 for (SmallVectorImpl<MachineBasicBlock*>::iterator
6673 II = MBBList.begin(), IE = MBBList.end(); II != IE; ++II) {
6674 LPadList.push_back(*II);
6675 InvokeBBs.insert((*II)->pred_begin(), (*II)->pred_end());
6679 assert(!LPadList.empty() &&
6680 "No landing pad destinations for the dispatch jump table!");
6682 // Create the jump table and associated information.
6683 MachineJumpTableInfo *JTI =
6684 MF->getOrCreateJumpTableInfo(MachineJumpTableInfo::EK_Inline);
6685 unsigned MJTI = JTI->createJumpTableIndex(LPadList);
6686 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
6688 // Create the MBBs for the dispatch code.
6690 // Shove the dispatch's address into the return slot in the function context.
6691 MachineBasicBlock *DispatchBB = MF->CreateMachineBasicBlock();
6692 DispatchBB->setIsLandingPad();
6694 MachineBasicBlock *TrapBB = MF->CreateMachineBasicBlock();
6695 unsigned trap_opcode;
6696 if (Subtarget->isThumb())
6697 trap_opcode = ARM::tTRAP;
6699 trap_opcode = Subtarget->useNaClTrap() ? ARM::TRAPNaCl : ARM::TRAP;
6701 BuildMI(TrapBB, dl, TII->get(trap_opcode));
6702 DispatchBB->addSuccessor(TrapBB);
6704 MachineBasicBlock *DispContBB = MF->CreateMachineBasicBlock();
6705 DispatchBB->addSuccessor(DispContBB);
6708 MF->insert(MF->end(), DispatchBB);
6709 MF->insert(MF->end(), DispContBB);
6710 MF->insert(MF->end(), TrapBB);
6712 // Insert code into the entry block that creates and registers the function
6714 SetupEntryBlockForSjLj(MI, MBB, DispatchBB, FI);
6716 MachineMemOperand *FIMMOLd =
6717 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
6718 MachineMemOperand::MOLoad |
6719 MachineMemOperand::MOVolatile, 4, 4);
6721 MachineInstrBuilder MIB;
6722 MIB = BuildMI(DispatchBB, dl, TII->get(ARM::Int_eh_sjlj_dispatchsetup));
6724 const ARMBaseInstrInfo *AII = static_cast<const ARMBaseInstrInfo*>(TII);
6725 const ARMBaseRegisterInfo &RI = AII->getRegisterInfo();
6727 // Add a register mask with no preserved registers. This results in all
6728 // registers being marked as clobbered.
6729 MIB.addRegMask(RI.getNoPreservedMask());
6731 unsigned NumLPads = LPadList.size();
6732 if (Subtarget->isThumb2()) {
6733 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6734 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2LDRi12), NewVReg1)
6737 .addMemOperand(FIMMOLd));
6739 if (NumLPads < 256) {
6740 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPri))
6742 .addImm(LPadList.size()));
6744 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6745 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVi16), VReg1)
6746 .addImm(NumLPads & 0xFFFF));
6748 unsigned VReg2 = VReg1;
6749 if ((NumLPads & 0xFFFF0000) != 0) {
6750 VReg2 = MRI->createVirtualRegister(TRC);
6751 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVTi16), VReg2)
6753 .addImm(NumLPads >> 16));
6756 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPrr))
6761 BuildMI(DispatchBB, dl, TII->get(ARM::t2Bcc))
6766 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6767 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::t2LEApcrelJT),NewVReg3)
6768 .addJumpTableIndex(MJTI));
6770 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6773 BuildMI(DispContBB, dl, TII->get(ARM::t2ADDrs), NewVReg4)
6774 .addReg(NewVReg3, RegState::Kill)
6776 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
6778 BuildMI(DispContBB, dl, TII->get(ARM::t2BR_JT))
6779 .addReg(NewVReg4, RegState::Kill)
6781 .addJumpTableIndex(MJTI);
6782 } else if (Subtarget->isThumb()) {
6783 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6784 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRspi), NewVReg1)
6787 .addMemOperand(FIMMOLd));
6789 if (NumLPads < 256) {
6790 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPi8))
6794 MachineConstantPool *ConstantPool = MF->getConstantPool();
6795 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
6796 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
6798 // MachineConstantPool wants an explicit alignment.
6799 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
6801 Align = getDataLayout()->getTypeAllocSize(C->getType());
6802 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6804 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6805 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRpci))
6806 .addReg(VReg1, RegState::Define)
6807 .addConstantPoolIndex(Idx));
6808 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPr))
6813 BuildMI(DispatchBB, dl, TII->get(ARM::tBcc))
6818 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6819 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLSLri), NewVReg2)
6820 .addReg(ARM::CPSR, RegState::Define)
6824 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6825 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLEApcrelJT), NewVReg3)
6826 .addJumpTableIndex(MJTI));
6828 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6829 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg4)
6830 .addReg(ARM::CPSR, RegState::Define)
6831 .addReg(NewVReg2, RegState::Kill)
6834 MachineMemOperand *JTMMOLd =
6835 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
6836 MachineMemOperand::MOLoad, 4, 4);
6838 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6839 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLDRi), NewVReg5)
6840 .addReg(NewVReg4, RegState::Kill)
6842 .addMemOperand(JTMMOLd));
6844 unsigned NewVReg6 = NewVReg5;
6845 if (RelocM == Reloc::PIC_) {
6846 NewVReg6 = MRI->createVirtualRegister(TRC);
6847 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg6)
6848 .addReg(ARM::CPSR, RegState::Define)
6849 .addReg(NewVReg5, RegState::Kill)
6853 BuildMI(DispContBB, dl, TII->get(ARM::tBR_JTr))
6854 .addReg(NewVReg6, RegState::Kill)
6855 .addJumpTableIndex(MJTI);
6857 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6858 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRi12), NewVReg1)
6861 .addMemOperand(FIMMOLd));
6863 if (NumLPads < 256) {
6864 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPri))
6867 } else if (Subtarget->hasV6T2Ops() && isUInt<16>(NumLPads)) {
6868 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6869 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVi16), VReg1)
6870 .addImm(NumLPads & 0xFFFF));
6872 unsigned VReg2 = VReg1;
6873 if ((NumLPads & 0xFFFF0000) != 0) {
6874 VReg2 = MRI->createVirtualRegister(TRC);
6875 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVTi16), VReg2)
6877 .addImm(NumLPads >> 16));
6880 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
6884 MachineConstantPool *ConstantPool = MF->getConstantPool();
6885 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
6886 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
6888 // MachineConstantPool wants an explicit alignment.
6889 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
6891 Align = getDataLayout()->getTypeAllocSize(C->getType());
6892 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6894 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6895 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRcp))
6896 .addReg(VReg1, RegState::Define)
6897 .addConstantPoolIndex(Idx)
6899 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
6901 .addReg(VReg1, RegState::Kill));
6904 BuildMI(DispatchBB, dl, TII->get(ARM::Bcc))
6909 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6911 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::MOVsi), NewVReg3)
6913 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
6914 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6915 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::LEApcrelJT), NewVReg4)
6916 .addJumpTableIndex(MJTI));
6918 MachineMemOperand *JTMMOLd =
6919 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
6920 MachineMemOperand::MOLoad, 4, 4);
6921 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6923 BuildMI(DispContBB, dl, TII->get(ARM::LDRrs), NewVReg5)
6924 .addReg(NewVReg3, RegState::Kill)
6927 .addMemOperand(JTMMOLd));
6929 if (RelocM == Reloc::PIC_) {
6930 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTadd))
6931 .addReg(NewVReg5, RegState::Kill)
6933 .addJumpTableIndex(MJTI);
6935 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTr))
6936 .addReg(NewVReg5, RegState::Kill)
6937 .addJumpTableIndex(MJTI);
6941 // Add the jump table entries as successors to the MBB.
6942 SmallPtrSet<MachineBasicBlock*, 8> SeenMBBs;
6943 for (std::vector<MachineBasicBlock*>::iterator
6944 I = LPadList.begin(), E = LPadList.end(); I != E; ++I) {
6945 MachineBasicBlock *CurMBB = *I;
6946 if (SeenMBBs.insert(CurMBB).second)
6947 DispContBB->addSuccessor(CurMBB);
6950 // N.B. the order the invoke BBs are processed in doesn't matter here.
6951 const MCPhysReg *SavedRegs = RI.getCalleeSavedRegs(MF);
6952 SmallVector<MachineBasicBlock*, 64> MBBLPads;
6953 for (MachineBasicBlock *BB : InvokeBBs) {
6955 // Remove the landing pad successor from the invoke block and replace it
6956 // with the new dispatch block.
6957 SmallVector<MachineBasicBlock*, 4> Successors(BB->succ_begin(),
6959 while (!Successors.empty()) {
6960 MachineBasicBlock *SMBB = Successors.pop_back_val();
6961 if (SMBB->isLandingPad()) {
6962 BB->removeSuccessor(SMBB);
6963 MBBLPads.push_back(SMBB);
6967 BB->addSuccessor(DispatchBB);
6969 // Find the invoke call and mark all of the callee-saved registers as
6970 // 'implicit defined' so that they're spilled. This prevents code from
6971 // moving instructions to before the EH block, where they will never be
6973 for (MachineBasicBlock::reverse_iterator
6974 II = BB->rbegin(), IE = BB->rend(); II != IE; ++II) {
6975 if (!II->isCall()) continue;
6977 DenseMap<unsigned, bool> DefRegs;
6978 for (MachineInstr::mop_iterator
6979 OI = II->operands_begin(), OE = II->operands_end();
6981 if (!OI->isReg()) continue;
6982 DefRegs[OI->getReg()] = true;
6985 MachineInstrBuilder MIB(*MF, &*II);
6987 for (unsigned i = 0; SavedRegs[i] != 0; ++i) {
6988 unsigned Reg = SavedRegs[i];
6989 if (Subtarget->isThumb2() &&
6990 !ARM::tGPRRegClass.contains(Reg) &&
6991 !ARM::hGPRRegClass.contains(Reg))
6993 if (Subtarget->isThumb1Only() && !ARM::tGPRRegClass.contains(Reg))
6995 if (!Subtarget->isThumb() && !ARM::GPRRegClass.contains(Reg))
6998 MIB.addReg(Reg, RegState::ImplicitDefine | RegState::Dead);
7005 // Mark all former landing pads as non-landing pads. The dispatch is the only
7007 for (SmallVectorImpl<MachineBasicBlock*>::iterator
7008 I = MBBLPads.begin(), E = MBBLPads.end(); I != E; ++I)
7009 (*I)->setIsLandingPad(false);
7011 // The instruction is gone now.
7012 MI->eraseFromParent();
7016 MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) {
7017 for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
7018 E = MBB->succ_end(); I != E; ++I)
7021 llvm_unreachable("Expecting a BB with two successors!");
7024 /// Return the load opcode for a given load size. If load size >= 8,
7025 /// neon opcode will be returned.
7026 static unsigned getLdOpcode(unsigned LdSize, bool IsThumb1, bool IsThumb2) {
7028 return LdSize == 16 ? ARM::VLD1q32wb_fixed
7029 : LdSize == 8 ? ARM::VLD1d32wb_fixed : 0;
7031 return LdSize == 4 ? ARM::tLDRi
7032 : LdSize == 2 ? ARM::tLDRHi
7033 : LdSize == 1 ? ARM::tLDRBi : 0;
7035 return LdSize == 4 ? ARM::t2LDR_POST
7036 : LdSize == 2 ? ARM::t2LDRH_POST
7037 : LdSize == 1 ? ARM::t2LDRB_POST : 0;
7038 return LdSize == 4 ? ARM::LDR_POST_IMM
7039 : LdSize == 2 ? ARM::LDRH_POST
7040 : LdSize == 1 ? ARM::LDRB_POST_IMM : 0;
7043 /// Return the store opcode for a given store size. If store size >= 8,
7044 /// neon opcode will be returned.
7045 static unsigned getStOpcode(unsigned StSize, bool IsThumb1, bool IsThumb2) {
7047 return StSize == 16 ? ARM::VST1q32wb_fixed
7048 : StSize == 8 ? ARM::VST1d32wb_fixed : 0;
7050 return StSize == 4 ? ARM::tSTRi
7051 : StSize == 2 ? ARM::tSTRHi
7052 : StSize == 1 ? ARM::tSTRBi : 0;
7054 return StSize == 4 ? ARM::t2STR_POST
7055 : StSize == 2 ? ARM::t2STRH_POST
7056 : StSize == 1 ? ARM::t2STRB_POST : 0;
7057 return StSize == 4 ? ARM::STR_POST_IMM
7058 : StSize == 2 ? ARM::STRH_POST
7059 : StSize == 1 ? ARM::STRB_POST_IMM : 0;
7062 /// Emit a post-increment load operation with given size. The instructions
7063 /// will be added to BB at Pos.
7064 static void emitPostLd(MachineBasicBlock *BB, MachineInstr *Pos,
7065 const TargetInstrInfo *TII, DebugLoc dl,
7066 unsigned LdSize, unsigned Data, unsigned AddrIn,
7067 unsigned AddrOut, bool IsThumb1, bool IsThumb2) {
7068 unsigned LdOpc = getLdOpcode(LdSize, IsThumb1, IsThumb2);
7069 assert(LdOpc != 0 && "Should have a load opcode");
7071 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
7072 .addReg(AddrOut, RegState::Define).addReg(AddrIn)
7074 } else if (IsThumb1) {
7075 // load + update AddrIn
7076 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
7077 .addReg(AddrIn).addImm(0));
7078 MachineInstrBuilder MIB =
7079 BuildMI(*BB, Pos, dl, TII->get(ARM::tADDi8), AddrOut);
7080 MIB = AddDefaultT1CC(MIB);
7081 MIB.addReg(AddrIn).addImm(LdSize);
7082 AddDefaultPred(MIB);
7083 } else if (IsThumb2) {
7084 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
7085 .addReg(AddrOut, RegState::Define).addReg(AddrIn)
7088 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
7089 .addReg(AddrOut, RegState::Define).addReg(AddrIn)
7090 .addReg(0).addImm(LdSize));
7094 /// Emit a post-increment store operation with given size. The instructions
7095 /// will be added to BB at Pos.
7096 static void emitPostSt(MachineBasicBlock *BB, MachineInstr *Pos,
7097 const TargetInstrInfo *TII, DebugLoc dl,
7098 unsigned StSize, unsigned Data, unsigned AddrIn,
7099 unsigned AddrOut, bool IsThumb1, bool IsThumb2) {
7100 unsigned StOpc = getStOpcode(StSize, IsThumb1, IsThumb2);
7101 assert(StOpc != 0 && "Should have a store opcode");
7103 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc), AddrOut)
7104 .addReg(AddrIn).addImm(0).addReg(Data));
7105 } else if (IsThumb1) {
7106 // store + update AddrIn
7107 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc)).addReg(Data)
7108 .addReg(AddrIn).addImm(0));
7109 MachineInstrBuilder MIB =
7110 BuildMI(*BB, Pos, dl, TII->get(ARM::tADDi8), AddrOut);
7111 MIB = AddDefaultT1CC(MIB);
7112 MIB.addReg(AddrIn).addImm(StSize);
7113 AddDefaultPred(MIB);
7114 } else if (IsThumb2) {
7115 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc), AddrOut)
7116 .addReg(Data).addReg(AddrIn).addImm(StSize));
7118 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc), AddrOut)
7119 .addReg(Data).addReg(AddrIn).addReg(0)
7125 ARMTargetLowering::EmitStructByval(MachineInstr *MI,
7126 MachineBasicBlock *BB) const {
7127 // This pseudo instruction has 3 operands: dst, src, size
7128 // We expand it to a loop if size > Subtarget->getMaxInlineSizeThreshold().
7129 // Otherwise, we will generate unrolled scalar copies.
7130 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
7131 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7132 MachineFunction::iterator It = BB;
7135 unsigned dest = MI->getOperand(0).getReg();
7136 unsigned src = MI->getOperand(1).getReg();
7137 unsigned SizeVal = MI->getOperand(2).getImm();
7138 unsigned Align = MI->getOperand(3).getImm();
7139 DebugLoc dl = MI->getDebugLoc();
7141 MachineFunction *MF = BB->getParent();
7142 MachineRegisterInfo &MRI = MF->getRegInfo();
7143 unsigned UnitSize = 0;
7144 const TargetRegisterClass *TRC = nullptr;
7145 const TargetRegisterClass *VecTRC = nullptr;
7147 bool IsThumb1 = Subtarget->isThumb1Only();
7148 bool IsThumb2 = Subtarget->isThumb2();
7152 } else if (Align & 2) {
7155 // Check whether we can use NEON instructions.
7156 if (!MF->getFunction()->hasFnAttribute(Attribute::NoImplicitFloat) &&
7157 Subtarget->hasNEON()) {
7158 if ((Align % 16 == 0) && SizeVal >= 16)
7160 else if ((Align % 8 == 0) && SizeVal >= 8)
7163 // Can't use NEON instructions.
7168 // Select the correct opcode and register class for unit size load/store
7169 bool IsNeon = UnitSize >= 8;
7170 TRC = (IsThumb1 || IsThumb2) ? &ARM::tGPRRegClass : &ARM::GPRRegClass;
7172 VecTRC = UnitSize == 16 ? &ARM::DPairRegClass
7173 : UnitSize == 8 ? &ARM::DPRRegClass
7176 unsigned BytesLeft = SizeVal % UnitSize;
7177 unsigned LoopSize = SizeVal - BytesLeft;
7179 if (SizeVal <= Subtarget->getMaxInlineSizeThreshold()) {
7180 // Use LDR and STR to copy.
7181 // [scratch, srcOut] = LDR_POST(srcIn, UnitSize)
7182 // [destOut] = STR_POST(scratch, destIn, UnitSize)
7183 unsigned srcIn = src;
7184 unsigned destIn = dest;
7185 for (unsigned i = 0; i < LoopSize; i+=UnitSize) {
7186 unsigned srcOut = MRI.createVirtualRegister(TRC);
7187 unsigned destOut = MRI.createVirtualRegister(TRC);
7188 unsigned scratch = MRI.createVirtualRegister(IsNeon ? VecTRC : TRC);
7189 emitPostLd(BB, MI, TII, dl, UnitSize, scratch, srcIn, srcOut,
7190 IsThumb1, IsThumb2);
7191 emitPostSt(BB, MI, TII, dl, UnitSize, scratch, destIn, destOut,
7192 IsThumb1, IsThumb2);
7197 // Handle the leftover bytes with LDRB and STRB.
7198 // [scratch, srcOut] = LDRB_POST(srcIn, 1)
7199 // [destOut] = STRB_POST(scratch, destIn, 1)
7200 for (unsigned i = 0; i < BytesLeft; i++) {
7201 unsigned srcOut = MRI.createVirtualRegister(TRC);
7202 unsigned destOut = MRI.createVirtualRegister(TRC);
7203 unsigned scratch = MRI.createVirtualRegister(TRC);
7204 emitPostLd(BB, MI, TII, dl, 1, scratch, srcIn, srcOut,
7205 IsThumb1, IsThumb2);
7206 emitPostSt(BB, MI, TII, dl, 1, scratch, destIn, destOut,
7207 IsThumb1, IsThumb2);
7211 MI->eraseFromParent(); // The instruction is gone now.
7215 // Expand the pseudo op to a loop.
7218 // movw varEnd, # --> with thumb2
7220 // ldrcp varEnd, idx --> without thumb2
7221 // fallthrough --> loopMBB
7223 // PHI varPhi, varEnd, varLoop
7224 // PHI srcPhi, src, srcLoop
7225 // PHI destPhi, dst, destLoop
7226 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
7227 // [destLoop] = STR_POST(scratch, destPhi, UnitSize)
7228 // subs varLoop, varPhi, #UnitSize
7230 // fallthrough --> exitMBB
7232 // epilogue to handle left-over bytes
7233 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
7234 // [destOut] = STRB_POST(scratch, destLoop, 1)
7235 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
7236 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
7237 MF->insert(It, loopMBB);
7238 MF->insert(It, exitMBB);
7240 // Transfer the remainder of BB and its successor edges to exitMBB.
7241 exitMBB->splice(exitMBB->begin(), BB,
7242 std::next(MachineBasicBlock::iterator(MI)), BB->end());
7243 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
7245 // Load an immediate to varEnd.
7246 unsigned varEnd = MRI.createVirtualRegister(TRC);
7247 if (Subtarget->useMovt(*MF)) {
7248 unsigned Vtmp = varEnd;
7249 if ((LoopSize & 0xFFFF0000) != 0)
7250 Vtmp = MRI.createVirtualRegister(TRC);
7251 AddDefaultPred(BuildMI(BB, dl,
7252 TII->get(IsThumb2 ? ARM::t2MOVi16 : ARM::MOVi16),
7253 Vtmp).addImm(LoopSize & 0xFFFF));
7255 if ((LoopSize & 0xFFFF0000) != 0)
7256 AddDefaultPred(BuildMI(BB, dl,
7257 TII->get(IsThumb2 ? ARM::t2MOVTi16 : ARM::MOVTi16),
7260 .addImm(LoopSize >> 16));
7262 MachineConstantPool *ConstantPool = MF->getConstantPool();
7263 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
7264 const Constant *C = ConstantInt::get(Int32Ty, LoopSize);
7266 // MachineConstantPool wants an explicit alignment.
7267 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
7269 Align = getDataLayout()->getTypeAllocSize(C->getType());
7270 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
7273 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(ARM::tLDRpci)).addReg(
7274 varEnd, RegState::Define).addConstantPoolIndex(Idx));
7276 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(ARM::LDRcp)).addReg(
7277 varEnd, RegState::Define).addConstantPoolIndex(Idx).addImm(0));
7279 BB->addSuccessor(loopMBB);
7281 // Generate the loop body:
7282 // varPhi = PHI(varLoop, varEnd)
7283 // srcPhi = PHI(srcLoop, src)
7284 // destPhi = PHI(destLoop, dst)
7285 MachineBasicBlock *entryBB = BB;
7287 unsigned varLoop = MRI.createVirtualRegister(TRC);
7288 unsigned varPhi = MRI.createVirtualRegister(TRC);
7289 unsigned srcLoop = MRI.createVirtualRegister(TRC);
7290 unsigned srcPhi = MRI.createVirtualRegister(TRC);
7291 unsigned destLoop = MRI.createVirtualRegister(TRC);
7292 unsigned destPhi = MRI.createVirtualRegister(TRC);
7294 BuildMI(*BB, BB->begin(), dl, TII->get(ARM::PHI), varPhi)
7295 .addReg(varLoop).addMBB(loopMBB)
7296 .addReg(varEnd).addMBB(entryBB);
7297 BuildMI(BB, dl, TII->get(ARM::PHI), srcPhi)
7298 .addReg(srcLoop).addMBB(loopMBB)
7299 .addReg(src).addMBB(entryBB);
7300 BuildMI(BB, dl, TII->get(ARM::PHI), destPhi)
7301 .addReg(destLoop).addMBB(loopMBB)
7302 .addReg(dest).addMBB(entryBB);
7304 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
7305 // [destLoop] = STR_POST(scratch, destPhi, UnitSiz)
7306 unsigned scratch = MRI.createVirtualRegister(IsNeon ? VecTRC : TRC);
7307 emitPostLd(BB, BB->end(), TII, dl, UnitSize, scratch, srcPhi, srcLoop,
7308 IsThumb1, IsThumb2);
7309 emitPostSt(BB, BB->end(), TII, dl, UnitSize, scratch, destPhi, destLoop,
7310 IsThumb1, IsThumb2);
7312 // Decrement loop variable by UnitSize.
7314 MachineInstrBuilder MIB =
7315 BuildMI(*BB, BB->end(), dl, TII->get(ARM::tSUBi8), varLoop);
7316 MIB = AddDefaultT1CC(MIB);
7317 MIB.addReg(varPhi).addImm(UnitSize);
7318 AddDefaultPred(MIB);
7320 MachineInstrBuilder MIB =
7321 BuildMI(*BB, BB->end(), dl,
7322 TII->get(IsThumb2 ? ARM::t2SUBri : ARM::SUBri), varLoop);
7323 AddDefaultCC(AddDefaultPred(MIB.addReg(varPhi).addImm(UnitSize)));
7324 MIB->getOperand(5).setReg(ARM::CPSR);
7325 MIB->getOperand(5).setIsDef(true);
7327 BuildMI(*BB, BB->end(), dl,
7328 TII->get(IsThumb1 ? ARM::tBcc : IsThumb2 ? ARM::t2Bcc : ARM::Bcc))
7329 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
7331 // loopMBB can loop back to loopMBB or fall through to exitMBB.
7332 BB->addSuccessor(loopMBB);
7333 BB->addSuccessor(exitMBB);
7335 // Add epilogue to handle BytesLeft.
7337 MachineInstr *StartOfExit = exitMBB->begin();
7339 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
7340 // [destOut] = STRB_POST(scratch, destLoop, 1)
7341 unsigned srcIn = srcLoop;
7342 unsigned destIn = destLoop;
7343 for (unsigned i = 0; i < BytesLeft; i++) {
7344 unsigned srcOut = MRI.createVirtualRegister(TRC);
7345 unsigned destOut = MRI.createVirtualRegister(TRC);
7346 unsigned scratch = MRI.createVirtualRegister(TRC);
7347 emitPostLd(BB, StartOfExit, TII, dl, 1, scratch, srcIn, srcOut,
7348 IsThumb1, IsThumb2);
7349 emitPostSt(BB, StartOfExit, TII, dl, 1, scratch, destIn, destOut,
7350 IsThumb1, IsThumb2);
7355 MI->eraseFromParent(); // The instruction is gone now.
7360 ARMTargetLowering::EmitLowered__chkstk(MachineInstr *MI,
7361 MachineBasicBlock *MBB) const {
7362 const TargetMachine &TM = getTargetMachine();
7363 const TargetInstrInfo &TII = *Subtarget->getInstrInfo();
7364 DebugLoc DL = MI->getDebugLoc();
7366 assert(Subtarget->isTargetWindows() &&
7367 "__chkstk is only supported on Windows");
7368 assert(Subtarget->isThumb2() && "Windows on ARM requires Thumb-2 mode");
7370 // __chkstk takes the number of words to allocate on the stack in R4, and
7371 // returns the stack adjustment in number of bytes in R4. This will not
7372 // clober any other registers (other than the obvious lr).
7374 // Although, technically, IP should be considered a register which may be
7375 // clobbered, the call itself will not touch it. Windows on ARM is a pure
7376 // thumb-2 environment, so there is no interworking required. As a result, we
7377 // do not expect a veneer to be emitted by the linker, clobbering IP.
7379 // Each module receives its own copy of __chkstk, so no import thunk is
7380 // required, again, ensuring that IP is not clobbered.
7382 // Finally, although some linkers may theoretically provide a trampoline for
7383 // out of range calls (which is quite common due to a 32M range limitation of
7384 // branches for Thumb), we can generate the long-call version via
7385 // -mcmodel=large, alleviating the need for the trampoline which may clobber
7388 switch (TM.getCodeModel()) {
7389 case CodeModel::Small:
7390 case CodeModel::Medium:
7391 case CodeModel::Default:
7392 case CodeModel::Kernel:
7393 BuildMI(*MBB, MI, DL, TII.get(ARM::tBL))
7394 .addImm((unsigned)ARMCC::AL).addReg(0)
7395 .addExternalSymbol("__chkstk")
7396 .addReg(ARM::R4, RegState::Implicit | RegState::Kill)
7397 .addReg(ARM::R4, RegState::Implicit | RegState::Define)
7398 .addReg(ARM::R12, RegState::Implicit | RegState::Define | RegState::Dead);
7400 case CodeModel::Large:
7401 case CodeModel::JITDefault: {
7402 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
7403 unsigned Reg = MRI.createVirtualRegister(&ARM::rGPRRegClass);
7405 BuildMI(*MBB, MI, DL, TII.get(ARM::t2MOVi32imm), Reg)
7406 .addExternalSymbol("__chkstk");
7407 BuildMI(*MBB, MI, DL, TII.get(ARM::tBLXr))
7408 .addImm((unsigned)ARMCC::AL).addReg(0)
7409 .addReg(Reg, RegState::Kill)
7410 .addReg(ARM::R4, RegState::Implicit | RegState::Kill)
7411 .addReg(ARM::R4, RegState::Implicit | RegState::Define)
7412 .addReg(ARM::R12, RegState::Implicit | RegState::Define | RegState::Dead);
7417 AddDefaultCC(AddDefaultPred(BuildMI(*MBB, MI, DL, TII.get(ARM::t2SUBrr),
7419 .addReg(ARM::SP).addReg(ARM::R4)));
7421 MI->eraseFromParent();
7426 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
7427 MachineBasicBlock *BB) const {
7428 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
7429 DebugLoc dl = MI->getDebugLoc();
7430 bool isThumb2 = Subtarget->isThumb2();
7431 switch (MI->getOpcode()) {
7434 llvm_unreachable("Unexpected instr type to insert");
7436 // The Thumb2 pre-indexed stores have the same MI operands, they just
7437 // define them differently in the .td files from the isel patterns, so
7438 // they need pseudos.
7439 case ARM::t2STR_preidx:
7440 MI->setDesc(TII->get(ARM::t2STR_PRE));
7442 case ARM::t2STRB_preidx:
7443 MI->setDesc(TII->get(ARM::t2STRB_PRE));
7445 case ARM::t2STRH_preidx:
7446 MI->setDesc(TII->get(ARM::t2STRH_PRE));
7449 case ARM::STRi_preidx:
7450 case ARM::STRBi_preidx: {
7451 unsigned NewOpc = MI->getOpcode() == ARM::STRi_preidx ?
7452 ARM::STR_PRE_IMM : ARM::STRB_PRE_IMM;
7453 // Decode the offset.
7454 unsigned Offset = MI->getOperand(4).getImm();
7455 bool isSub = ARM_AM::getAM2Op(Offset) == ARM_AM::sub;
7456 Offset = ARM_AM::getAM2Offset(Offset);
7460 MachineMemOperand *MMO = *MI->memoperands_begin();
7461 BuildMI(*BB, MI, dl, TII->get(NewOpc))
7462 .addOperand(MI->getOperand(0)) // Rn_wb
7463 .addOperand(MI->getOperand(1)) // Rt
7464 .addOperand(MI->getOperand(2)) // Rn
7465 .addImm(Offset) // offset (skip GPR==zero_reg)
7466 .addOperand(MI->getOperand(5)) // pred
7467 .addOperand(MI->getOperand(6))
7468 .addMemOperand(MMO);
7469 MI->eraseFromParent();
7472 case ARM::STRr_preidx:
7473 case ARM::STRBr_preidx:
7474 case ARM::STRH_preidx: {
7476 switch (MI->getOpcode()) {
7477 default: llvm_unreachable("unexpected opcode!");
7478 case ARM::STRr_preidx: NewOpc = ARM::STR_PRE_REG; break;
7479 case ARM::STRBr_preidx: NewOpc = ARM::STRB_PRE_REG; break;
7480 case ARM::STRH_preidx: NewOpc = ARM::STRH_PRE; break;
7482 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(NewOpc));
7483 for (unsigned i = 0; i < MI->getNumOperands(); ++i)
7484 MIB.addOperand(MI->getOperand(i));
7485 MI->eraseFromParent();
7489 case ARM::tMOVCCr_pseudo: {
7490 // To "insert" a SELECT_CC instruction, we actually have to insert the
7491 // diamond control-flow pattern. The incoming instruction knows the
7492 // destination vreg to set, the condition code register to branch on, the
7493 // true/false values to select between, and a branch opcode to use.
7494 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7495 MachineFunction::iterator It = BB;
7501 // cmpTY ccX, r1, r2
7503 // fallthrough --> copy0MBB
7504 MachineBasicBlock *thisMBB = BB;
7505 MachineFunction *F = BB->getParent();
7506 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
7507 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
7508 F->insert(It, copy0MBB);
7509 F->insert(It, sinkMBB);
7511 // Transfer the remainder of BB and its successor edges to sinkMBB.
7512 sinkMBB->splice(sinkMBB->begin(), BB,
7513 std::next(MachineBasicBlock::iterator(MI)), BB->end());
7514 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
7516 BB->addSuccessor(copy0MBB);
7517 BB->addSuccessor(sinkMBB);
7519 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
7520 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
7523 // %FalseValue = ...
7524 // # fallthrough to sinkMBB
7527 // Update machine-CFG edges
7528 BB->addSuccessor(sinkMBB);
7531 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
7534 BuildMI(*BB, BB->begin(), dl,
7535 TII->get(ARM::PHI), MI->getOperand(0).getReg())
7536 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
7537 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
7539 MI->eraseFromParent(); // The pseudo instruction is gone now.
7544 case ARM::BCCZi64: {
7545 // If there is an unconditional branch to the other successor, remove it.
7546 BB->erase(std::next(MachineBasicBlock::iterator(MI)), BB->end());
7548 // Compare both parts that make up the double comparison separately for
7550 bool RHSisZero = MI->getOpcode() == ARM::BCCZi64;
7552 unsigned LHS1 = MI->getOperand(1).getReg();
7553 unsigned LHS2 = MI->getOperand(2).getReg();
7555 AddDefaultPred(BuildMI(BB, dl,
7556 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7557 .addReg(LHS1).addImm(0));
7558 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7559 .addReg(LHS2).addImm(0)
7560 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
7562 unsigned RHS1 = MI->getOperand(3).getReg();
7563 unsigned RHS2 = MI->getOperand(4).getReg();
7564 AddDefaultPred(BuildMI(BB, dl,
7565 TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
7566 .addReg(LHS1).addReg(RHS1));
7567 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
7568 .addReg(LHS2).addReg(RHS2)
7569 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
7572 MachineBasicBlock *destMBB = MI->getOperand(RHSisZero ? 3 : 5).getMBB();
7573 MachineBasicBlock *exitMBB = OtherSucc(BB, destMBB);
7574 if (MI->getOperand(0).getImm() == ARMCC::NE)
7575 std::swap(destMBB, exitMBB);
7577 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
7578 .addMBB(destMBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
7580 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2B)).addMBB(exitMBB));
7582 BuildMI(BB, dl, TII->get(ARM::B)) .addMBB(exitMBB);
7584 MI->eraseFromParent(); // The pseudo instruction is gone now.
7588 case ARM::Int_eh_sjlj_setjmp:
7589 case ARM::Int_eh_sjlj_setjmp_nofp:
7590 case ARM::tInt_eh_sjlj_setjmp:
7591 case ARM::t2Int_eh_sjlj_setjmp:
7592 case ARM::t2Int_eh_sjlj_setjmp_nofp:
7593 EmitSjLjDispatchBlock(MI, BB);
7598 // To insert an ABS instruction, we have to insert the
7599 // diamond control-flow pattern. The incoming instruction knows the
7600 // source vreg to test against 0, the destination vreg to set,
7601 // the condition code register to branch on, the
7602 // true/false values to select between, and a branch opcode to use.
7607 // BCC (branch to SinkBB if V0 >= 0)
7608 // RSBBB: V3 = RSBri V2, 0 (compute ABS if V2 < 0)
7609 // SinkBB: V1 = PHI(V2, V3)
7610 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7611 MachineFunction::iterator BBI = BB;
7613 MachineFunction *Fn = BB->getParent();
7614 MachineBasicBlock *RSBBB = Fn->CreateMachineBasicBlock(LLVM_BB);
7615 MachineBasicBlock *SinkBB = Fn->CreateMachineBasicBlock(LLVM_BB);
7616 Fn->insert(BBI, RSBBB);
7617 Fn->insert(BBI, SinkBB);
7619 unsigned int ABSSrcReg = MI->getOperand(1).getReg();
7620 unsigned int ABSDstReg = MI->getOperand(0).getReg();
7621 bool ABSSrcKIll = MI->getOperand(1).isKill();
7622 bool isThumb2 = Subtarget->isThumb2();
7623 MachineRegisterInfo &MRI = Fn->getRegInfo();
7624 // In Thumb mode S must not be specified if source register is the SP or
7625 // PC and if destination register is the SP, so restrict register class
7626 unsigned NewRsbDstReg =
7627 MRI.createVirtualRegister(isThumb2 ? &ARM::rGPRRegClass : &ARM::GPRRegClass);
7629 // Transfer the remainder of BB and its successor edges to sinkMBB.
7630 SinkBB->splice(SinkBB->begin(), BB,
7631 std::next(MachineBasicBlock::iterator(MI)), BB->end());
7632 SinkBB->transferSuccessorsAndUpdatePHIs(BB);
7634 BB->addSuccessor(RSBBB);
7635 BB->addSuccessor(SinkBB);
7637 // fall through to SinkMBB
7638 RSBBB->addSuccessor(SinkBB);
7640 // insert a cmp at the end of BB
7641 AddDefaultPred(BuildMI(BB, dl,
7642 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7643 .addReg(ABSSrcReg).addImm(0));
7645 // insert a bcc with opposite CC to ARMCC::MI at the end of BB
7647 TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)).addMBB(SinkBB)
7648 .addImm(ARMCC::getOppositeCondition(ARMCC::MI)).addReg(ARM::CPSR);
7650 // insert rsbri in RSBBB
7651 // Note: BCC and rsbri will be converted into predicated rsbmi
7652 // by if-conversion pass
7653 BuildMI(*RSBBB, RSBBB->begin(), dl,
7654 TII->get(isThumb2 ? ARM::t2RSBri : ARM::RSBri), NewRsbDstReg)
7655 .addReg(ABSSrcReg, ABSSrcKIll ? RegState::Kill : 0)
7656 .addImm(0).addImm((unsigned)ARMCC::AL).addReg(0).addReg(0);
7658 // insert PHI in SinkBB,
7659 // reuse ABSDstReg to not change uses of ABS instruction
7660 BuildMI(*SinkBB, SinkBB->begin(), dl,
7661 TII->get(ARM::PHI), ABSDstReg)
7662 .addReg(NewRsbDstReg).addMBB(RSBBB)
7663 .addReg(ABSSrcReg).addMBB(BB);
7665 // remove ABS instruction
7666 MI->eraseFromParent();
7668 // return last added BB
7671 case ARM::COPY_STRUCT_BYVAL_I32:
7673 return EmitStructByval(MI, BB);
7674 case ARM::WIN__CHKSTK:
7675 return EmitLowered__chkstk(MI, BB);
7679 /// \brief Lowers MCOPY to either LDMIA/STMIA or LDMIA_UPD/STMID_UPD depending
7680 /// on whether the result is used. This is done as a post-isel lowering instead
7681 /// of as a custom inserter because we need the use list from the SDNode.
7682 static void LowerMCOPY(const ARMSubtarget *Subtarget, MachineInstr *MI,
7684 bool isThumb1 = Subtarget->isThumb1Only();
7685 bool isThumb2 = Subtarget->isThumb2();
7686 const ARMBaseInstrInfo *TII = Subtarget->getInstrInfo();
7688 DebugLoc dl = MI->getDebugLoc();
7689 MachineBasicBlock *BB = MI->getParent();
7690 MachineFunction *MF = BB->getParent();
7691 MachineRegisterInfo &MRI = MF->getRegInfo();
7693 MachineInstrBuilder LD, ST;
7694 if (isThumb1 || Node->hasAnyUseOfValue(1)) {
7695 LD = BuildMI(*BB, MI, dl, TII->get(isThumb2 ? ARM::t2LDMIA_UPD
7696 : isThumb1 ? ARM::tLDMIA_UPD
7698 .addOperand(MI->getOperand(1));
7700 LD = BuildMI(*BB, MI, dl, TII->get(isThumb2 ? ARM::t2LDMIA : ARM::LDMIA));
7703 if (isThumb1 || Node->hasAnyUseOfValue(0)) {
7704 ST = BuildMI(*BB, MI, dl, TII->get(isThumb2 ? ARM::t2STMIA_UPD
7705 : isThumb1 ? ARM::tSTMIA_UPD
7707 .addOperand(MI->getOperand(0));
7709 ST = BuildMI(*BB, MI, dl, TII->get(isThumb2 ? ARM::t2STMIA : ARM::STMIA));
7712 LD.addOperand(MI->getOperand(3)).addImm(ARMCC::AL).addReg(0);
7713 ST.addOperand(MI->getOperand(2)).addImm(ARMCC::AL).addReg(0);
7715 for (unsigned I = 0; I != MI->getOperand(4).getImm(); ++I) {
7716 unsigned TmpReg = MRI.createVirtualRegister(isThumb1 ? &ARM::tGPRRegClass
7717 : &ARM::GPRRegClass);
7718 LD.addReg(TmpReg, RegState::Define);
7719 ST.addReg(TmpReg, RegState::Kill);
7722 MI->eraseFromParent();
7725 void ARMTargetLowering::AdjustInstrPostInstrSelection(MachineInstr *MI,
7726 SDNode *Node) const {
7727 if (MI->getOpcode() == ARM::MCOPY) {
7728 LowerMCOPY(Subtarget, MI, Node);
7732 const MCInstrDesc *MCID = &MI->getDesc();
7733 // Adjust potentially 's' setting instructions after isel, i.e. ADC, SBC, RSB,
7734 // RSC. Coming out of isel, they have an implicit CPSR def, but the optional
7735 // operand is still set to noreg. If needed, set the optional operand's
7736 // register to CPSR, and remove the redundant implicit def.
7738 // e.g. ADCS (..., CPSR<imp-def>) -> ADC (... opt:CPSR<def>).
7740 // Rename pseudo opcodes.
7741 unsigned NewOpc = convertAddSubFlagsOpcode(MI->getOpcode());
7743 const ARMBaseInstrInfo *TII = Subtarget->getInstrInfo();
7744 MCID = &TII->get(NewOpc);
7746 assert(MCID->getNumOperands() == MI->getDesc().getNumOperands() + 1 &&
7747 "converted opcode should be the same except for cc_out");
7751 // Add the optional cc_out operand
7752 MI->addOperand(MachineOperand::CreateReg(0, /*isDef=*/true));
7754 unsigned ccOutIdx = MCID->getNumOperands() - 1;
7756 // Any ARM instruction that sets the 's' bit should specify an optional
7757 // "cc_out" operand in the last operand position.
7758 if (!MI->hasOptionalDef() || !MCID->OpInfo[ccOutIdx].isOptionalDef()) {
7759 assert(!NewOpc && "Optional cc_out operand required");
7762 // Look for an implicit def of CPSR added by MachineInstr ctor. Remove it
7763 // since we already have an optional CPSR def.
7764 bool definesCPSR = false;
7765 bool deadCPSR = false;
7766 for (unsigned i = MCID->getNumOperands(), e = MI->getNumOperands();
7768 const MachineOperand &MO = MI->getOperand(i);
7769 if (MO.isReg() && MO.isDef() && MO.getReg() == ARM::CPSR) {
7773 MI->RemoveOperand(i);
7778 assert(!NewOpc && "Optional cc_out operand required");
7781 assert(deadCPSR == !Node->hasAnyUseOfValue(1) && "inconsistent dead flag");
7783 assert(!MI->getOperand(ccOutIdx).getReg() &&
7784 "expect uninitialized optional cc_out operand");
7788 // If this instruction was defined with an optional CPSR def and its dag node
7789 // had a live implicit CPSR def, then activate the optional CPSR def.
7790 MachineOperand &MO = MI->getOperand(ccOutIdx);
7791 MO.setReg(ARM::CPSR);
7795 //===----------------------------------------------------------------------===//
7796 // ARM Optimization Hooks
7797 //===----------------------------------------------------------------------===//
7799 // Helper function that checks if N is a null or all ones constant.
7800 static inline bool isZeroOrAllOnes(SDValue N, bool AllOnes) {
7801 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N);
7804 return AllOnes ? C->isAllOnesValue() : C->isNullValue();
7807 // Return true if N is conditionally 0 or all ones.
7808 // Detects these expressions where cc is an i1 value:
7810 // (select cc 0, y) [AllOnes=0]
7811 // (select cc y, 0) [AllOnes=0]
7812 // (zext cc) [AllOnes=0]
7813 // (sext cc) [AllOnes=0/1]
7814 // (select cc -1, y) [AllOnes=1]
7815 // (select cc y, -1) [AllOnes=1]
7817 // Invert is set when N is the null/all ones constant when CC is false.
7818 // OtherOp is set to the alternative value of N.
7819 static bool isConditionalZeroOrAllOnes(SDNode *N, bool AllOnes,
7820 SDValue &CC, bool &Invert,
7822 SelectionDAG &DAG) {
7823 switch (N->getOpcode()) {
7824 default: return false;
7826 CC = N->getOperand(0);
7827 SDValue N1 = N->getOperand(1);
7828 SDValue N2 = N->getOperand(2);
7829 if (isZeroOrAllOnes(N1, AllOnes)) {
7834 if (isZeroOrAllOnes(N2, AllOnes)) {
7841 case ISD::ZERO_EXTEND:
7842 // (zext cc) can never be the all ones value.
7846 case ISD::SIGN_EXTEND: {
7848 EVT VT = N->getValueType(0);
7849 CC = N->getOperand(0);
7850 if (CC.getValueType() != MVT::i1)
7854 // When looking for an AllOnes constant, N is an sext, and the 'other'
7856 OtherOp = DAG.getConstant(0, dl, VT);
7857 else if (N->getOpcode() == ISD::ZERO_EXTEND)
7858 // When looking for a 0 constant, N can be zext or sext.
7859 OtherOp = DAG.getConstant(1, dl, VT);
7861 OtherOp = DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), dl,
7868 // Combine a constant select operand into its use:
7870 // (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
7871 // (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
7872 // (and (select cc, -1, c), x) -> (select cc, x, (and, x, c)) [AllOnes=1]
7873 // (or (select cc, 0, c), x) -> (select cc, x, (or, x, c))
7874 // (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c))
7876 // The transform is rejected if the select doesn't have a constant operand that
7877 // is null, or all ones when AllOnes is set.
7879 // Also recognize sext/zext from i1:
7881 // (add (zext cc), x) -> (select cc (add x, 1), x)
7882 // (add (sext cc), x) -> (select cc (add x, -1), x)
7884 // These transformations eventually create predicated instructions.
7886 // @param N The node to transform.
7887 // @param Slct The N operand that is a select.
7888 // @param OtherOp The other N operand (x above).
7889 // @param DCI Context.
7890 // @param AllOnes Require the select constant to be all ones instead of null.
7891 // @returns The new node, or SDValue() on failure.
7893 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
7894 TargetLowering::DAGCombinerInfo &DCI,
7895 bool AllOnes = false) {
7896 SelectionDAG &DAG = DCI.DAG;
7897 EVT VT = N->getValueType(0);
7898 SDValue NonConstantVal;
7901 if (!isConditionalZeroOrAllOnes(Slct.getNode(), AllOnes, CCOp, SwapSelectOps,
7902 NonConstantVal, DAG))
7905 // Slct is now know to be the desired identity constant when CC is true.
7906 SDValue TrueVal = OtherOp;
7907 SDValue FalseVal = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
7908 OtherOp, NonConstantVal);
7909 // Unless SwapSelectOps says CC should be false.
7911 std::swap(TrueVal, FalseVal);
7913 return DAG.getNode(ISD::SELECT, SDLoc(N), VT,
7914 CCOp, TrueVal, FalseVal);
7917 // Attempt combineSelectAndUse on each operand of a commutative operator N.
7919 SDValue combineSelectAndUseCommutative(SDNode *N, bool AllOnes,
7920 TargetLowering::DAGCombinerInfo &DCI) {
7921 SDValue N0 = N->getOperand(0);
7922 SDValue N1 = N->getOperand(1);
7923 if (N0.getNode()->hasOneUse()) {
7924 SDValue Result = combineSelectAndUse(N, N0, N1, DCI, AllOnes);
7925 if (Result.getNode())
7928 if (N1.getNode()->hasOneUse()) {
7929 SDValue Result = combineSelectAndUse(N, N1, N0, DCI, AllOnes);
7930 if (Result.getNode())
7936 // AddCombineToVPADDL- For pair-wise add on neon, use the vpaddl instruction
7937 // (only after legalization).
7938 static SDValue AddCombineToVPADDL(SDNode *N, SDValue N0, SDValue N1,
7939 TargetLowering::DAGCombinerInfo &DCI,
7940 const ARMSubtarget *Subtarget) {
7942 // Only perform optimization if after legalize, and if NEON is available. We
7943 // also expected both operands to be BUILD_VECTORs.
7944 if (DCI.isBeforeLegalize() || !Subtarget->hasNEON()
7945 || N0.getOpcode() != ISD::BUILD_VECTOR
7946 || N1.getOpcode() != ISD::BUILD_VECTOR)
7949 // Check output type since VPADDL operand elements can only be 8, 16, or 32.
7950 EVT VT = N->getValueType(0);
7951 if (!VT.isInteger() || VT.getVectorElementType() == MVT::i64)
7954 // Check that the vector operands are of the right form.
7955 // N0 and N1 are BUILD_VECTOR nodes with N number of EXTRACT_VECTOR
7956 // operands, where N is the size of the formed vector.
7957 // Each EXTRACT_VECTOR should have the same input vector and odd or even
7958 // index such that we have a pair wise add pattern.
7960 // Grab the vector that all EXTRACT_VECTOR nodes should be referencing.
7961 if (N0->getOperand(0)->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
7963 SDValue Vec = N0->getOperand(0)->getOperand(0);
7964 SDNode *V = Vec.getNode();
7965 unsigned nextIndex = 0;
7967 // For each operands to the ADD which are BUILD_VECTORs,
7968 // check to see if each of their operands are an EXTRACT_VECTOR with
7969 // the same vector and appropriate index.
7970 for (unsigned i = 0, e = N0->getNumOperands(); i != e; ++i) {
7971 if (N0->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT
7972 && N1->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
7974 SDValue ExtVec0 = N0->getOperand(i);
7975 SDValue ExtVec1 = N1->getOperand(i);
7977 // First operand is the vector, verify its the same.
7978 if (V != ExtVec0->getOperand(0).getNode() ||
7979 V != ExtVec1->getOperand(0).getNode())
7982 // Second is the constant, verify its correct.
7983 ConstantSDNode *C0 = dyn_cast<ConstantSDNode>(ExtVec0->getOperand(1));
7984 ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(ExtVec1->getOperand(1));
7986 // For the constant, we want to see all the even or all the odd.
7987 if (!C0 || !C1 || C0->getZExtValue() != nextIndex
7988 || C1->getZExtValue() != nextIndex+1)
7997 // Create VPADDL node.
7998 SelectionDAG &DAG = DCI.DAG;
7999 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8003 // Build operand list.
8004 SmallVector<SDValue, 8> Ops;
8005 Ops.push_back(DAG.getConstant(Intrinsic::arm_neon_vpaddls, dl,
8006 TLI.getPointerTy()));
8008 // Input is the vector.
8011 // Get widened type and narrowed type.
8013 unsigned numElem = VT.getVectorNumElements();
8015 EVT inputLaneType = Vec.getValueType().getVectorElementType();
8016 switch (inputLaneType.getSimpleVT().SimpleTy) {
8017 case MVT::i8: widenType = MVT::getVectorVT(MVT::i16, numElem); break;
8018 case MVT::i16: widenType = MVT::getVectorVT(MVT::i32, numElem); break;
8019 case MVT::i32: widenType = MVT::getVectorVT(MVT::i64, numElem); break;
8021 llvm_unreachable("Invalid vector element type for padd optimization.");
8024 SDValue tmp = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, widenType, Ops);
8025 unsigned ExtOp = VT.bitsGT(tmp.getValueType()) ? ISD::ANY_EXTEND : ISD::TRUNCATE;
8026 return DAG.getNode(ExtOp, dl, VT, tmp);
8029 static SDValue findMUL_LOHI(SDValue V) {
8030 if (V->getOpcode() == ISD::UMUL_LOHI ||
8031 V->getOpcode() == ISD::SMUL_LOHI)
8036 static SDValue AddCombineTo64bitMLAL(SDNode *AddcNode,
8037 TargetLowering::DAGCombinerInfo &DCI,
8038 const ARMSubtarget *Subtarget) {
8040 if (Subtarget->isThumb1Only()) return SDValue();
8042 // Only perform the checks after legalize when the pattern is available.
8043 if (DCI.isBeforeLegalize()) return SDValue();
8045 // Look for multiply add opportunities.
8046 // The pattern is a ISD::UMUL_LOHI followed by two add nodes, where
8047 // each add nodes consumes a value from ISD::UMUL_LOHI and there is
8048 // a glue link from the first add to the second add.
8049 // If we find this pattern, we can replace the U/SMUL_LOHI, ADDC, and ADDE by
8050 // a S/UMLAL instruction.
8053 // / \ [no multiline comment]
8059 assert(AddcNode->getOpcode() == ISD::ADDC && "Expect an ADDC");
8060 SDValue AddcOp0 = AddcNode->getOperand(0);
8061 SDValue AddcOp1 = AddcNode->getOperand(1);
8063 // Check if the two operands are from the same mul_lohi node.
8064 if (AddcOp0.getNode() == AddcOp1.getNode())
8067 assert(AddcNode->getNumValues() == 2 &&
8068 AddcNode->getValueType(0) == MVT::i32 &&
8069 "Expect ADDC with two result values. First: i32");
8071 // Check that we have a glued ADDC node.
8072 if (AddcNode->getValueType(1) != MVT::Glue)
8075 // Check that the ADDC adds the low result of the S/UMUL_LOHI.
8076 if (AddcOp0->getOpcode() != ISD::UMUL_LOHI &&
8077 AddcOp0->getOpcode() != ISD::SMUL_LOHI &&
8078 AddcOp1->getOpcode() != ISD::UMUL_LOHI &&
8079 AddcOp1->getOpcode() != ISD::SMUL_LOHI)
8082 // Look for the glued ADDE.
8083 SDNode* AddeNode = AddcNode->getGluedUser();
8087 // Make sure it is really an ADDE.
8088 if (AddeNode->getOpcode() != ISD::ADDE)
8091 assert(AddeNode->getNumOperands() == 3 &&
8092 AddeNode->getOperand(2).getValueType() == MVT::Glue &&
8093 "ADDE node has the wrong inputs");
8095 // Check for the triangle shape.
8096 SDValue AddeOp0 = AddeNode->getOperand(0);
8097 SDValue AddeOp1 = AddeNode->getOperand(1);
8099 // Make sure that the ADDE operands are not coming from the same node.
8100 if (AddeOp0.getNode() == AddeOp1.getNode())
8103 // Find the MUL_LOHI node walking up ADDE's operands.
8104 bool IsLeftOperandMUL = false;
8105 SDValue MULOp = findMUL_LOHI(AddeOp0);
8106 if (MULOp == SDValue())
8107 MULOp = findMUL_LOHI(AddeOp1);
8109 IsLeftOperandMUL = true;
8110 if (MULOp == SDValue())
8113 // Figure out the right opcode.
8114 unsigned Opc = MULOp->getOpcode();
8115 unsigned FinalOpc = (Opc == ISD::SMUL_LOHI) ? ARMISD::SMLAL : ARMISD::UMLAL;
8117 // Figure out the high and low input values to the MLAL node.
8118 SDValue* HiAdd = nullptr;
8119 SDValue* LoMul = nullptr;
8120 SDValue* LowAdd = nullptr;
8122 // Ensure that ADDE is from high result of ISD::SMUL_LOHI.
8123 if ((AddeOp0 != MULOp.getValue(1)) && (AddeOp1 != MULOp.getValue(1)))
8126 if (IsLeftOperandMUL)
8132 // Ensure that LoMul and LowAdd are taken from correct ISD::SMUL_LOHI node
8133 // whose low result is fed to the ADDC we are checking.
8135 if (AddcOp0 == MULOp.getValue(0)) {
8139 if (AddcOp1 == MULOp.getValue(0)) {
8147 // Create the merged node.
8148 SelectionDAG &DAG = DCI.DAG;
8150 // Build operand list.
8151 SmallVector<SDValue, 8> Ops;
8152 Ops.push_back(LoMul->getOperand(0));
8153 Ops.push_back(LoMul->getOperand(1));
8154 Ops.push_back(*LowAdd);
8155 Ops.push_back(*HiAdd);
8157 SDValue MLALNode = DAG.getNode(FinalOpc, SDLoc(AddcNode),
8158 DAG.getVTList(MVT::i32, MVT::i32), Ops);
8160 // Replace the ADDs' nodes uses by the MLA node's values.
8161 SDValue HiMLALResult(MLALNode.getNode(), 1);
8162 DAG.ReplaceAllUsesOfValueWith(SDValue(AddeNode, 0), HiMLALResult);
8164 SDValue LoMLALResult(MLALNode.getNode(), 0);
8165 DAG.ReplaceAllUsesOfValueWith(SDValue(AddcNode, 0), LoMLALResult);
8167 // Return original node to notify the driver to stop replacing.
8168 SDValue resNode(AddcNode, 0);
8172 /// PerformADDCCombine - Target-specific dag combine transform from
8173 /// ISD::ADDC, ISD::ADDE, and ISD::MUL_LOHI to MLAL.
8174 static SDValue PerformADDCCombine(SDNode *N,
8175 TargetLowering::DAGCombinerInfo &DCI,
8176 const ARMSubtarget *Subtarget) {
8178 return AddCombineTo64bitMLAL(N, DCI, Subtarget);
8182 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
8183 /// operands N0 and N1. This is a helper for PerformADDCombine that is
8184 /// called with the default operands, and if that fails, with commuted
8186 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
8187 TargetLowering::DAGCombinerInfo &DCI,
8188 const ARMSubtarget *Subtarget){
8190 // Attempt to create vpaddl for this add.
8191 SDValue Result = AddCombineToVPADDL(N, N0, N1, DCI, Subtarget);
8192 if (Result.getNode())
8195 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
8196 if (N0.getNode()->hasOneUse()) {
8197 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
8198 if (Result.getNode()) return Result;
8203 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
8205 static SDValue PerformADDCombine(SDNode *N,
8206 TargetLowering::DAGCombinerInfo &DCI,
8207 const ARMSubtarget *Subtarget) {
8208 SDValue N0 = N->getOperand(0);
8209 SDValue N1 = N->getOperand(1);
8211 // First try with the default operand order.
8212 SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI, Subtarget);
8213 if (Result.getNode())
8216 // If that didn't work, try again with the operands commuted.
8217 return PerformADDCombineWithOperands(N, N1, N0, DCI, Subtarget);
8220 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
8222 static SDValue PerformSUBCombine(SDNode *N,
8223 TargetLowering::DAGCombinerInfo &DCI) {
8224 SDValue N0 = N->getOperand(0);
8225 SDValue N1 = N->getOperand(1);
8227 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
8228 if (N1.getNode()->hasOneUse()) {
8229 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
8230 if (Result.getNode()) return Result;
8236 /// PerformVMULCombine
8237 /// Distribute (A + B) * C to (A * C) + (B * C) to take advantage of the
8238 /// special multiplier accumulator forwarding.
8244 // However, for (A + B) * (A + B),
8251 static SDValue PerformVMULCombine(SDNode *N,
8252 TargetLowering::DAGCombinerInfo &DCI,
8253 const ARMSubtarget *Subtarget) {
8254 if (!Subtarget->hasVMLxForwarding())
8257 SelectionDAG &DAG = DCI.DAG;
8258 SDValue N0 = N->getOperand(0);
8259 SDValue N1 = N->getOperand(1);
8260 unsigned Opcode = N0.getOpcode();
8261 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
8262 Opcode != ISD::FADD && Opcode != ISD::FSUB) {
8263 Opcode = N1.getOpcode();
8264 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
8265 Opcode != ISD::FADD && Opcode != ISD::FSUB)
8273 EVT VT = N->getValueType(0);
8275 SDValue N00 = N0->getOperand(0);
8276 SDValue N01 = N0->getOperand(1);
8277 return DAG.getNode(Opcode, DL, VT,
8278 DAG.getNode(ISD::MUL, DL, VT, N00, N1),
8279 DAG.getNode(ISD::MUL, DL, VT, N01, N1));
8282 static SDValue PerformMULCombine(SDNode *N,
8283 TargetLowering::DAGCombinerInfo &DCI,
8284 const ARMSubtarget *Subtarget) {
8285 SelectionDAG &DAG = DCI.DAG;
8287 if (Subtarget->isThumb1Only())
8290 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
8293 EVT VT = N->getValueType(0);
8294 if (VT.is64BitVector() || VT.is128BitVector())
8295 return PerformVMULCombine(N, DCI, Subtarget);
8299 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
8303 int64_t MulAmt = C->getSExtValue();
8304 unsigned ShiftAmt = countTrailingZeros<uint64_t>(MulAmt);
8306 ShiftAmt = ShiftAmt & (32 - 1);
8307 SDValue V = N->getOperand(0);
8311 MulAmt >>= ShiftAmt;
8314 if (isPowerOf2_32(MulAmt - 1)) {
8315 // (mul x, 2^N + 1) => (add (shl x, N), x)
8316 Res = DAG.getNode(ISD::ADD, DL, VT,
8318 DAG.getNode(ISD::SHL, DL, VT,
8320 DAG.getConstant(Log2_32(MulAmt - 1), DL,
8322 } else if (isPowerOf2_32(MulAmt + 1)) {
8323 // (mul x, 2^N - 1) => (sub (shl x, N), x)
8324 Res = DAG.getNode(ISD::SUB, DL, VT,
8325 DAG.getNode(ISD::SHL, DL, VT,
8327 DAG.getConstant(Log2_32(MulAmt + 1), DL,
8333 uint64_t MulAmtAbs = -MulAmt;
8334 if (isPowerOf2_32(MulAmtAbs + 1)) {
8335 // (mul x, -(2^N - 1)) => (sub x, (shl x, N))
8336 Res = DAG.getNode(ISD::SUB, DL, VT,
8338 DAG.getNode(ISD::SHL, DL, VT,
8340 DAG.getConstant(Log2_32(MulAmtAbs + 1), DL,
8342 } else if (isPowerOf2_32(MulAmtAbs - 1)) {
8343 // (mul x, -(2^N + 1)) => - (add (shl x, N), x)
8344 Res = DAG.getNode(ISD::ADD, DL, VT,
8346 DAG.getNode(ISD::SHL, DL, VT,
8348 DAG.getConstant(Log2_32(MulAmtAbs - 1), DL,
8350 Res = DAG.getNode(ISD::SUB, DL, VT,
8351 DAG.getConstant(0, DL, MVT::i32), Res);
8358 Res = DAG.getNode(ISD::SHL, DL, VT,
8359 Res, DAG.getConstant(ShiftAmt, DL, MVT::i32));
8361 // Do not add new nodes to DAG combiner worklist.
8362 DCI.CombineTo(N, Res, false);
8366 static SDValue PerformANDCombine(SDNode *N,
8367 TargetLowering::DAGCombinerInfo &DCI,
8368 const ARMSubtarget *Subtarget) {
8370 // Attempt to use immediate-form VBIC
8371 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
8373 EVT VT = N->getValueType(0);
8374 SelectionDAG &DAG = DCI.DAG;
8376 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8379 APInt SplatBits, SplatUndef;
8380 unsigned SplatBitSize;
8383 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
8384 if (SplatBitSize <= 64) {
8386 SDValue Val = isNEONModifiedImm((~SplatBits).getZExtValue(),
8387 SplatUndef.getZExtValue(), SplatBitSize,
8388 DAG, dl, VbicVT, VT.is128BitVector(),
8390 if (Val.getNode()) {
8392 DAG.getNode(ISD::BITCAST, dl, VbicVT, N->getOperand(0));
8393 SDValue Vbic = DAG.getNode(ARMISD::VBICIMM, dl, VbicVT, Input, Val);
8394 return DAG.getNode(ISD::BITCAST, dl, VT, Vbic);
8399 if (!Subtarget->isThumb1Only()) {
8400 // fold (and (select cc, -1, c), x) -> (select cc, x, (and, x, c))
8401 SDValue Result = combineSelectAndUseCommutative(N, true, DCI);
8402 if (Result.getNode())
8409 /// PerformORCombine - Target-specific dag combine xforms for ISD::OR
8410 static SDValue PerformORCombine(SDNode *N,
8411 TargetLowering::DAGCombinerInfo &DCI,
8412 const ARMSubtarget *Subtarget) {
8413 // Attempt to use immediate-form VORR
8414 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
8416 EVT VT = N->getValueType(0);
8417 SelectionDAG &DAG = DCI.DAG;
8419 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8422 APInt SplatBits, SplatUndef;
8423 unsigned SplatBitSize;
8425 if (BVN && Subtarget->hasNEON() &&
8426 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
8427 if (SplatBitSize <= 64) {
8429 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
8430 SplatUndef.getZExtValue(), SplatBitSize,
8431 DAG, dl, VorrVT, VT.is128BitVector(),
8433 if (Val.getNode()) {
8435 DAG.getNode(ISD::BITCAST, dl, VorrVT, N->getOperand(0));
8436 SDValue Vorr = DAG.getNode(ARMISD::VORRIMM, dl, VorrVT, Input, Val);
8437 return DAG.getNode(ISD::BITCAST, dl, VT, Vorr);
8442 if (!Subtarget->isThumb1Only()) {
8443 // fold (or (select cc, 0, c), x) -> (select cc, x, (or, x, c))
8444 SDValue Result = combineSelectAndUseCommutative(N, false, DCI);
8445 if (Result.getNode())
8449 // The code below optimizes (or (and X, Y), Z).
8450 // The AND operand needs to have a single user to make these optimizations
8452 SDValue N0 = N->getOperand(0);
8453 if (N0.getOpcode() != ISD::AND || !N0.hasOneUse())
8455 SDValue N1 = N->getOperand(1);
8457 // (or (and B, A), (and C, ~A)) => (VBSL A, B, C) when A is a constant.
8458 if (Subtarget->hasNEON() && N1.getOpcode() == ISD::AND && VT.isVector() &&
8459 DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
8461 unsigned SplatBitSize;
8464 APInt SplatBits0, SplatBits1;
8465 BuildVectorSDNode *BVN0 = dyn_cast<BuildVectorSDNode>(N0->getOperand(1));
8466 BuildVectorSDNode *BVN1 = dyn_cast<BuildVectorSDNode>(N1->getOperand(1));
8467 // Ensure that the second operand of both ands are constants
8468 if (BVN0 && BVN0->isConstantSplat(SplatBits0, SplatUndef, SplatBitSize,
8469 HasAnyUndefs) && !HasAnyUndefs) {
8470 if (BVN1 && BVN1->isConstantSplat(SplatBits1, SplatUndef, SplatBitSize,
8471 HasAnyUndefs) && !HasAnyUndefs) {
8472 // Ensure that the bit width of the constants are the same and that
8473 // the splat arguments are logical inverses as per the pattern we
8474 // are trying to simplify.
8475 if (SplatBits0.getBitWidth() == SplatBits1.getBitWidth() &&
8476 SplatBits0 == ~SplatBits1) {
8477 // Canonicalize the vector type to make instruction selection
8479 EVT CanonicalVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
8480 SDValue Result = DAG.getNode(ARMISD::VBSL, dl, CanonicalVT,
8484 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
8490 // Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when
8493 // BFI is only available on V6T2+
8494 if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops())
8498 // 1) or (and A, mask), val => ARMbfi A, val, mask
8499 // iff (val & mask) == val
8501 // 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
8502 // 2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2)
8503 // && mask == ~mask2
8504 // 2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2)
8505 // && ~mask == mask2
8506 // (i.e., copy a bitfield value into another bitfield of the same width)
8511 SDValue N00 = N0.getOperand(0);
8513 // The value and the mask need to be constants so we can verify this is
8514 // actually a bitfield set. If the mask is 0xffff, we can do better
8515 // via a movt instruction, so don't use BFI in that case.
8516 SDValue MaskOp = N0.getOperand(1);
8517 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(MaskOp);
8520 unsigned Mask = MaskC->getZExtValue();
8524 // Case (1): or (and A, mask), val => ARMbfi A, val, mask
8525 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
8527 unsigned Val = N1C->getZExtValue();
8528 if ((Val & ~Mask) != Val)
8531 if (ARM::isBitFieldInvertedMask(Mask)) {
8532 Val >>= countTrailingZeros(~Mask);
8534 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00,
8535 DAG.getConstant(Val, DL, MVT::i32),
8536 DAG.getConstant(Mask, DL, MVT::i32));
8538 // Do not add new nodes to DAG combiner worklist.
8539 DCI.CombineTo(N, Res, false);
8542 } else if (N1.getOpcode() == ISD::AND) {
8543 // case (2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
8544 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
8547 unsigned Mask2 = N11C->getZExtValue();
8549 // Mask and ~Mask2 (or reverse) must be equivalent for the BFI pattern
8551 if (ARM::isBitFieldInvertedMask(Mask) &&
8553 // The pack halfword instruction works better for masks that fit it,
8554 // so use that when it's available.
8555 if (Subtarget->hasT2ExtractPack() &&
8556 (Mask == 0xffff || Mask == 0xffff0000))
8559 unsigned amt = countTrailingZeros(Mask2);
8560 Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0),
8561 DAG.getConstant(amt, DL, MVT::i32));
8562 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00, Res,
8563 DAG.getConstant(Mask, DL, MVT::i32));
8564 // Do not add new nodes to DAG combiner worklist.
8565 DCI.CombineTo(N, Res, false);
8567 } else if (ARM::isBitFieldInvertedMask(~Mask) &&
8569 // The pack halfword instruction works better for masks that fit it,
8570 // so use that when it's available.
8571 if (Subtarget->hasT2ExtractPack() &&
8572 (Mask2 == 0xffff || Mask2 == 0xffff0000))
8575 unsigned lsb = countTrailingZeros(Mask);
8576 Res = DAG.getNode(ISD::SRL, DL, VT, N00,
8577 DAG.getConstant(lsb, DL, MVT::i32));
8578 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res,
8579 DAG.getConstant(Mask2, DL, MVT::i32));
8580 // Do not add new nodes to DAG combiner worklist.
8581 DCI.CombineTo(N, Res, false);
8586 if (DAG.MaskedValueIsZero(N1, MaskC->getAPIntValue()) &&
8587 N00.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N00.getOperand(1)) &&
8588 ARM::isBitFieldInvertedMask(~Mask)) {
8589 // Case (3): or (and (shl A, #shamt), mask), B => ARMbfi B, A, ~mask
8590 // where lsb(mask) == #shamt and masked bits of B are known zero.
8591 SDValue ShAmt = N00.getOperand(1);
8592 unsigned ShAmtC = cast<ConstantSDNode>(ShAmt)->getZExtValue();
8593 unsigned LSB = countTrailingZeros(Mask);
8597 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1, N00.getOperand(0),
8598 DAG.getConstant(~Mask, DL, MVT::i32));
8600 // Do not add new nodes to DAG combiner worklist.
8601 DCI.CombineTo(N, Res, false);
8607 static SDValue PerformXORCombine(SDNode *N,
8608 TargetLowering::DAGCombinerInfo &DCI,
8609 const ARMSubtarget *Subtarget) {
8610 EVT VT = N->getValueType(0);
8611 SelectionDAG &DAG = DCI.DAG;
8613 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8616 if (!Subtarget->isThumb1Only()) {
8617 // fold (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c))
8618 SDValue Result = combineSelectAndUseCommutative(N, false, DCI);
8619 if (Result.getNode())
8626 /// PerformBFICombine - (bfi A, (and B, Mask1), Mask2) -> (bfi A, B, Mask2) iff
8627 /// the bits being cleared by the AND are not demanded by the BFI.
8628 static SDValue PerformBFICombine(SDNode *N,
8629 TargetLowering::DAGCombinerInfo &DCI) {
8630 SDValue N1 = N->getOperand(1);
8631 if (N1.getOpcode() == ISD::AND) {
8632 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
8635 unsigned InvMask = cast<ConstantSDNode>(N->getOperand(2))->getZExtValue();
8636 unsigned LSB = countTrailingZeros(~InvMask);
8637 unsigned Width = (32 - countLeadingZeros(~InvMask)) - LSB;
8639 static_cast<unsigned>(std::numeric_limits<unsigned>::digits) &&
8640 "undefined behavior");
8641 unsigned Mask = (1u << Width) - 1;
8642 unsigned Mask2 = N11C->getZExtValue();
8643 if ((Mask & (~Mask2)) == 0)
8644 return DCI.DAG.getNode(ARMISD::BFI, SDLoc(N), N->getValueType(0),
8645 N->getOperand(0), N1.getOperand(0),
8651 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
8652 /// ARMISD::VMOVRRD.
8653 static SDValue PerformVMOVRRDCombine(SDNode *N,
8654 TargetLowering::DAGCombinerInfo &DCI,
8655 const ARMSubtarget *Subtarget) {
8656 // vmovrrd(vmovdrr x, y) -> x,y
8657 SDValue InDouble = N->getOperand(0);
8658 if (InDouble.getOpcode() == ARMISD::VMOVDRR && !Subtarget->isFPOnlySP())
8659 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
8661 // vmovrrd(load f64) -> (load i32), (load i32)
8662 SDNode *InNode = InDouble.getNode();
8663 if (ISD::isNormalLoad(InNode) && InNode->hasOneUse() &&
8664 InNode->getValueType(0) == MVT::f64 &&
8665 InNode->getOperand(1).getOpcode() == ISD::FrameIndex &&
8666 !cast<LoadSDNode>(InNode)->isVolatile()) {
8667 // TODO: Should this be done for non-FrameIndex operands?
8668 LoadSDNode *LD = cast<LoadSDNode>(InNode);
8670 SelectionDAG &DAG = DCI.DAG;
8672 SDValue BasePtr = LD->getBasePtr();
8673 SDValue NewLD1 = DAG.getLoad(MVT::i32, DL, LD->getChain(), BasePtr,
8674 LD->getPointerInfo(), LD->isVolatile(),
8675 LD->isNonTemporal(), LD->isInvariant(),
8676 LD->getAlignment());
8678 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
8679 DAG.getConstant(4, DL, MVT::i32));
8680 SDValue NewLD2 = DAG.getLoad(MVT::i32, DL, NewLD1.getValue(1), OffsetPtr,
8681 LD->getPointerInfo(), LD->isVolatile(),
8682 LD->isNonTemporal(), LD->isInvariant(),
8683 std::min(4U, LD->getAlignment() / 2));
8685 DAG.ReplaceAllUsesOfValueWith(SDValue(LD, 1), NewLD2.getValue(1));
8686 if (DCI.DAG.getTargetLoweringInfo().isBigEndian())
8687 std::swap (NewLD1, NewLD2);
8688 SDValue Result = DCI.CombineTo(N, NewLD1, NewLD2);
8695 /// PerformVMOVDRRCombine - Target-specific dag combine xforms for
8696 /// ARMISD::VMOVDRR. This is also used for BUILD_VECTORs with 2 operands.
8697 static SDValue PerformVMOVDRRCombine(SDNode *N, SelectionDAG &DAG) {
8698 // N=vmovrrd(X); vmovdrr(N:0, N:1) -> bit_convert(X)
8699 SDValue Op0 = N->getOperand(0);
8700 SDValue Op1 = N->getOperand(1);
8701 if (Op0.getOpcode() == ISD::BITCAST)
8702 Op0 = Op0.getOperand(0);
8703 if (Op1.getOpcode() == ISD::BITCAST)
8704 Op1 = Op1.getOperand(0);
8705 if (Op0.getOpcode() == ARMISD::VMOVRRD &&
8706 Op0.getNode() == Op1.getNode() &&
8707 Op0.getResNo() == 0 && Op1.getResNo() == 1)
8708 return DAG.getNode(ISD::BITCAST, SDLoc(N),
8709 N->getValueType(0), Op0.getOperand(0));
8713 /// hasNormalLoadOperand - Check if any of the operands of a BUILD_VECTOR node
8714 /// are normal, non-volatile loads. If so, it is profitable to bitcast an
8715 /// i64 vector to have f64 elements, since the value can then be loaded
8716 /// directly into a VFP register.
8717 static bool hasNormalLoadOperand(SDNode *N) {
8718 unsigned NumElts = N->getValueType(0).getVectorNumElements();
8719 for (unsigned i = 0; i < NumElts; ++i) {
8720 SDNode *Elt = N->getOperand(i).getNode();
8721 if (ISD::isNormalLoad(Elt) && !cast<LoadSDNode>(Elt)->isVolatile())
8727 /// PerformBUILD_VECTORCombine - Target-specific dag combine xforms for
8728 /// ISD::BUILD_VECTOR.
8729 static SDValue PerformBUILD_VECTORCombine(SDNode *N,
8730 TargetLowering::DAGCombinerInfo &DCI,
8731 const ARMSubtarget *Subtarget) {
8732 // build_vector(N=ARMISD::VMOVRRD(X), N:1) -> bit_convert(X):
8733 // VMOVRRD is introduced when legalizing i64 types. It forces the i64 value
8734 // into a pair of GPRs, which is fine when the value is used as a scalar,
8735 // but if the i64 value is converted to a vector, we need to undo the VMOVRRD.
8736 SelectionDAG &DAG = DCI.DAG;
8737 if (N->getNumOperands() == 2) {
8738 SDValue RV = PerformVMOVDRRCombine(N, DAG);
8743 // Load i64 elements as f64 values so that type legalization does not split
8744 // them up into i32 values.
8745 EVT VT = N->getValueType(0);
8746 if (VT.getVectorElementType() != MVT::i64 || !hasNormalLoadOperand(N))
8749 SmallVector<SDValue, 8> Ops;
8750 unsigned NumElts = VT.getVectorNumElements();
8751 for (unsigned i = 0; i < NumElts; ++i) {
8752 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(i));
8754 // Make the DAGCombiner fold the bitcast.
8755 DCI.AddToWorklist(V.getNode());
8757 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, NumElts);
8758 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, FloatVT, Ops);
8759 return DAG.getNode(ISD::BITCAST, dl, VT, BV);
8762 /// \brief Target-specific dag combine xforms for ARMISD::BUILD_VECTOR.
8764 PerformARMBUILD_VECTORCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
8765 // ARMISD::BUILD_VECTOR is introduced when legalizing ISD::BUILD_VECTOR.
8766 // At that time, we may have inserted bitcasts from integer to float.
8767 // If these bitcasts have survived DAGCombine, change the lowering of this
8768 // BUILD_VECTOR in something more vector friendly, i.e., that does not
8769 // force to use floating point types.
8771 // Make sure we can change the type of the vector.
8772 // This is possible iff:
8773 // 1. The vector is only used in a bitcast to a integer type. I.e.,
8774 // 1.1. Vector is used only once.
8775 // 1.2. Use is a bit convert to an integer type.
8776 // 2. The size of its operands are 32-bits (64-bits are not legal).
8777 EVT VT = N->getValueType(0);
8778 EVT EltVT = VT.getVectorElementType();
8780 // Check 1.1. and 2.
8781 if (EltVT.getSizeInBits() != 32 || !N->hasOneUse())
8784 // By construction, the input type must be float.
8785 assert(EltVT == MVT::f32 && "Unexpected type!");
8788 SDNode *Use = *N->use_begin();
8789 if (Use->getOpcode() != ISD::BITCAST ||
8790 Use->getValueType(0).isFloatingPoint())
8793 // Check profitability.
8794 // Model is, if more than half of the relevant operands are bitcast from
8795 // i32, turn the build_vector into a sequence of insert_vector_elt.
8796 // Relevant operands are everything that is not statically
8797 // (i.e., at compile time) bitcasted.
8798 unsigned NumOfBitCastedElts = 0;
8799 unsigned NumElts = VT.getVectorNumElements();
8800 unsigned NumOfRelevantElts = NumElts;
8801 for (unsigned Idx = 0; Idx < NumElts; ++Idx) {
8802 SDValue Elt = N->getOperand(Idx);
8803 if (Elt->getOpcode() == ISD::BITCAST) {
8804 // Assume only bit cast to i32 will go away.
8805 if (Elt->getOperand(0).getValueType() == MVT::i32)
8806 ++NumOfBitCastedElts;
8807 } else if (Elt.getOpcode() == ISD::UNDEF || isa<ConstantSDNode>(Elt))
8808 // Constants are statically casted, thus do not count them as
8809 // relevant operands.
8810 --NumOfRelevantElts;
8813 // Check if more than half of the elements require a non-free bitcast.
8814 if (NumOfBitCastedElts <= NumOfRelevantElts / 2)
8817 SelectionDAG &DAG = DCI.DAG;
8818 // Create the new vector type.
8819 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
8820 // Check if the type is legal.
8821 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8822 if (!TLI.isTypeLegal(VecVT))
8826 // ARMISD::BUILD_VECTOR E1, E2, ..., EN.
8827 // => BITCAST INSERT_VECTOR_ELT
8828 // (INSERT_VECTOR_ELT (...), (BITCAST EN-1), N-1),
8830 SDValue Vec = DAG.getUNDEF(VecVT);
8832 for (unsigned Idx = 0 ; Idx < NumElts; ++Idx) {
8833 SDValue V = N->getOperand(Idx);
8834 if (V.getOpcode() == ISD::UNDEF)
8836 if (V.getOpcode() == ISD::BITCAST &&
8837 V->getOperand(0).getValueType() == MVT::i32)
8838 // Fold obvious case.
8839 V = V.getOperand(0);
8841 V = DAG.getNode(ISD::BITCAST, SDLoc(V), MVT::i32, V);
8842 // Make the DAGCombiner fold the bitcasts.
8843 DCI.AddToWorklist(V.getNode());
8845 SDValue LaneIdx = DAG.getConstant(Idx, dl, MVT::i32);
8846 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VecVT, Vec, V, LaneIdx);
8848 Vec = DAG.getNode(ISD::BITCAST, dl, VT, Vec);
8849 // Make the DAGCombiner fold the bitcasts.
8850 DCI.AddToWorklist(Vec.getNode());
8854 /// PerformInsertEltCombine - Target-specific dag combine xforms for
8855 /// ISD::INSERT_VECTOR_ELT.
8856 static SDValue PerformInsertEltCombine(SDNode *N,
8857 TargetLowering::DAGCombinerInfo &DCI) {
8858 // Bitcast an i64 load inserted into a vector to f64.
8859 // Otherwise, the i64 value will be legalized to a pair of i32 values.
8860 EVT VT = N->getValueType(0);
8861 SDNode *Elt = N->getOperand(1).getNode();
8862 if (VT.getVectorElementType() != MVT::i64 ||
8863 !ISD::isNormalLoad(Elt) || cast<LoadSDNode>(Elt)->isVolatile())
8866 SelectionDAG &DAG = DCI.DAG;
8868 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
8869 VT.getVectorNumElements());
8870 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, N->getOperand(0));
8871 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(1));
8872 // Make the DAGCombiner fold the bitcasts.
8873 DCI.AddToWorklist(Vec.getNode());
8874 DCI.AddToWorklist(V.getNode());
8875 SDValue InsElt = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, FloatVT,
8876 Vec, V, N->getOperand(2));
8877 return DAG.getNode(ISD::BITCAST, dl, VT, InsElt);
8880 /// PerformVECTOR_SHUFFLECombine - Target-specific dag combine xforms for
8881 /// ISD::VECTOR_SHUFFLE.
8882 static SDValue PerformVECTOR_SHUFFLECombine(SDNode *N, SelectionDAG &DAG) {
8883 // The LLVM shufflevector instruction does not require the shuffle mask
8884 // length to match the operand vector length, but ISD::VECTOR_SHUFFLE does
8885 // have that requirement. When translating to ISD::VECTOR_SHUFFLE, if the
8886 // operands do not match the mask length, they are extended by concatenating
8887 // them with undef vectors. That is probably the right thing for other
8888 // targets, but for NEON it is better to concatenate two double-register
8889 // size vector operands into a single quad-register size vector. Do that
8890 // transformation here:
8891 // shuffle(concat(v1, undef), concat(v2, undef)) ->
8892 // shuffle(concat(v1, v2), undef)
8893 SDValue Op0 = N->getOperand(0);
8894 SDValue Op1 = N->getOperand(1);
8895 if (Op0.getOpcode() != ISD::CONCAT_VECTORS ||
8896 Op1.getOpcode() != ISD::CONCAT_VECTORS ||
8897 Op0.getNumOperands() != 2 ||
8898 Op1.getNumOperands() != 2)
8900 SDValue Concat0Op1 = Op0.getOperand(1);
8901 SDValue Concat1Op1 = Op1.getOperand(1);
8902 if (Concat0Op1.getOpcode() != ISD::UNDEF ||
8903 Concat1Op1.getOpcode() != ISD::UNDEF)
8905 // Skip the transformation if any of the types are illegal.
8906 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8907 EVT VT = N->getValueType(0);
8908 if (!TLI.isTypeLegal(VT) ||
8909 !TLI.isTypeLegal(Concat0Op1.getValueType()) ||
8910 !TLI.isTypeLegal(Concat1Op1.getValueType()))
8913 SDValue NewConcat = DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT,
8914 Op0.getOperand(0), Op1.getOperand(0));
8915 // Translate the shuffle mask.
8916 SmallVector<int, 16> NewMask;
8917 unsigned NumElts = VT.getVectorNumElements();
8918 unsigned HalfElts = NumElts/2;
8919 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
8920 for (unsigned n = 0; n < NumElts; ++n) {
8921 int MaskElt = SVN->getMaskElt(n);
8923 if (MaskElt < (int)HalfElts)
8925 else if (MaskElt >= (int)NumElts && MaskElt < (int)(NumElts + HalfElts))
8926 NewElt = HalfElts + MaskElt - NumElts;
8927 NewMask.push_back(NewElt);
8929 return DAG.getVectorShuffle(VT, SDLoc(N), NewConcat,
8930 DAG.getUNDEF(VT), NewMask.data());
8933 /// CombineBaseUpdate - Target-specific DAG combine function for VLDDUP,
8934 /// NEON load/store intrinsics, and generic vector load/stores, to merge
8935 /// base address updates.
8936 /// For generic load/stores, the memory type is assumed to be a vector.
8937 /// The caller is assumed to have checked legality.
8938 static SDValue CombineBaseUpdate(SDNode *N,
8939 TargetLowering::DAGCombinerInfo &DCI) {
8940 SelectionDAG &DAG = DCI.DAG;
8941 const bool isIntrinsic = (N->getOpcode() == ISD::INTRINSIC_VOID ||
8942 N->getOpcode() == ISD::INTRINSIC_W_CHAIN);
8943 const bool isStore = N->getOpcode() == ISD::STORE;
8944 const unsigned AddrOpIdx = ((isIntrinsic || isStore) ? 2 : 1);
8945 SDValue Addr = N->getOperand(AddrOpIdx);
8946 MemSDNode *MemN = cast<MemSDNode>(N);
8949 // Search for a use of the address operand that is an increment.
8950 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(),
8951 UE = Addr.getNode()->use_end(); UI != UE; ++UI) {
8953 if (User->getOpcode() != ISD::ADD ||
8954 UI.getUse().getResNo() != Addr.getResNo())
8957 // Check that the add is independent of the load/store. Otherwise, folding
8958 // it would create a cycle.
8959 if (User->isPredecessorOf(N) || N->isPredecessorOf(User))
8962 // Find the new opcode for the updating load/store.
8963 bool isLoadOp = true;
8964 bool isLaneOp = false;
8965 unsigned NewOpc = 0;
8966 unsigned NumVecs = 0;
8968 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
8970 default: llvm_unreachable("unexpected intrinsic for Neon base update");
8971 case Intrinsic::arm_neon_vld1: NewOpc = ARMISD::VLD1_UPD;
8973 case Intrinsic::arm_neon_vld2: NewOpc = ARMISD::VLD2_UPD;
8975 case Intrinsic::arm_neon_vld3: NewOpc = ARMISD::VLD3_UPD;
8977 case Intrinsic::arm_neon_vld4: NewOpc = ARMISD::VLD4_UPD;
8979 case Intrinsic::arm_neon_vld2lane: NewOpc = ARMISD::VLD2LN_UPD;
8980 NumVecs = 2; isLaneOp = true; break;
8981 case Intrinsic::arm_neon_vld3lane: NewOpc = ARMISD::VLD3LN_UPD;
8982 NumVecs = 3; isLaneOp = true; break;
8983 case Intrinsic::arm_neon_vld4lane: NewOpc = ARMISD::VLD4LN_UPD;
8984 NumVecs = 4; isLaneOp = true; break;
8985 case Intrinsic::arm_neon_vst1: NewOpc = ARMISD::VST1_UPD;
8986 NumVecs = 1; isLoadOp = false; break;
8987 case Intrinsic::arm_neon_vst2: NewOpc = ARMISD::VST2_UPD;
8988 NumVecs = 2; isLoadOp = false; break;
8989 case Intrinsic::arm_neon_vst3: NewOpc = ARMISD::VST3_UPD;
8990 NumVecs = 3; isLoadOp = false; break;
8991 case Intrinsic::arm_neon_vst4: NewOpc = ARMISD::VST4_UPD;
8992 NumVecs = 4; isLoadOp = false; break;
8993 case Intrinsic::arm_neon_vst2lane: NewOpc = ARMISD::VST2LN_UPD;
8994 NumVecs = 2; isLoadOp = false; isLaneOp = true; break;
8995 case Intrinsic::arm_neon_vst3lane: NewOpc = ARMISD::VST3LN_UPD;
8996 NumVecs = 3; isLoadOp = false; isLaneOp = true; break;
8997 case Intrinsic::arm_neon_vst4lane: NewOpc = ARMISD::VST4LN_UPD;
8998 NumVecs = 4; isLoadOp = false; isLaneOp = true; break;
9002 switch (N->getOpcode()) {
9003 default: llvm_unreachable("unexpected opcode for Neon base update");
9004 case ARMISD::VLD2DUP: NewOpc = ARMISD::VLD2DUP_UPD; NumVecs = 2; break;
9005 case ARMISD::VLD3DUP: NewOpc = ARMISD::VLD3DUP_UPD; NumVecs = 3; break;
9006 case ARMISD::VLD4DUP: NewOpc = ARMISD::VLD4DUP_UPD; NumVecs = 4; break;
9007 case ISD::LOAD: NewOpc = ARMISD::VLD1_UPD;
9008 NumVecs = 1; isLaneOp = false; break;
9009 case ISD::STORE: NewOpc = ARMISD::VST1_UPD;
9010 NumVecs = 1; isLaneOp = false; isLoadOp = false; break;
9014 // Find the size of memory referenced by the load/store.
9017 VecTy = N->getValueType(0);
9018 } else if (isIntrinsic) {
9019 VecTy = N->getOperand(AddrOpIdx+1).getValueType();
9021 assert(isStore && "Node has to be a load, a store, or an intrinsic!");
9022 VecTy = N->getOperand(1).getValueType();
9025 unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8;
9027 NumBytes /= VecTy.getVectorNumElements();
9029 // If the increment is a constant, it must match the memory ref size.
9030 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
9031 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
9032 uint64_t IncVal = CInc->getZExtValue();
9033 if (IncVal != NumBytes)
9035 } else if (NumBytes >= 3 * 16) {
9036 // VLD3/4 and VST3/4 for 128-bit vectors are implemented with two
9037 // separate instructions that make it harder to use a non-constant update.
9041 // OK, we found an ADD we can fold into the base update.
9042 // Now, create a _UPD node, taking care of not breaking alignment.
9044 EVT AlignedVecTy = VecTy;
9045 unsigned Alignment = MemN->getAlignment();
9047 // If this is a less-than-standard-aligned load/store, change the type to
9048 // match the standard alignment.
9049 // The alignment is overlooked when selecting _UPD variants; and it's
9050 // easier to introduce bitcasts here than fix that.
9051 // There are 3 ways to get to this base-update combine:
9052 // - intrinsics: they are assumed to be properly aligned (to the standard
9053 // alignment of the memory type), so we don't need to do anything.
9054 // - ARMISD::VLDx nodes: they are only generated from the aforementioned
9055 // intrinsics, so, likewise, there's nothing to do.
9056 // - generic load/store instructions: the alignment is specified as an
9057 // explicit operand, rather than implicitly as the standard alignment
9058 // of the memory type (like the intrisics). We need to change the
9059 // memory type to match the explicit alignment. That way, we don't
9060 // generate non-standard-aligned ARMISD::VLDx nodes.
9061 if (isa<LSBaseSDNode>(N)) {
9064 if (Alignment < VecTy.getScalarSizeInBits() / 8) {
9065 MVT EltTy = MVT::getIntegerVT(Alignment * 8);
9066 assert(NumVecs == 1 && "Unexpected multi-element generic load/store.");
9067 assert(!isLaneOp && "Unexpected generic load/store lane.");
9068 unsigned NumElts = NumBytes / (EltTy.getSizeInBits() / 8);
9069 AlignedVecTy = MVT::getVectorVT(EltTy, NumElts);
9071 // Don't set an explicit alignment on regular load/stores that we want
9072 // to transform to VLD/VST 1_UPD nodes.
9073 // This matches the behavior of regular load/stores, which only get an
9074 // explicit alignment if the MMO alignment is larger than the standard
9075 // alignment of the memory type.
9076 // Intrinsics, however, always get an explicit alignment, set to the
9077 // alignment of the MMO.
9081 // Create the new updating load/store node.
9082 // First, create an SDVTList for the new updating node's results.
9084 unsigned NumResultVecs = (isLoadOp ? NumVecs : 0);
9086 for (n = 0; n < NumResultVecs; ++n)
9087 Tys[n] = AlignedVecTy;
9088 Tys[n++] = MVT::i32;
9089 Tys[n] = MVT::Other;
9090 SDVTList SDTys = DAG.getVTList(makeArrayRef(Tys, NumResultVecs+2));
9092 // Then, gather the new node's operands.
9093 SmallVector<SDValue, 8> Ops;
9094 Ops.push_back(N->getOperand(0)); // incoming chain
9095 Ops.push_back(N->getOperand(AddrOpIdx));
9098 if (StoreSDNode *StN = dyn_cast<StoreSDNode>(N)) {
9099 // Try to match the intrinsic's signature
9100 Ops.push_back(StN->getValue());
9102 // Loads (and of course intrinsics) match the intrinsics' signature,
9103 // so just add all but the alignment operand.
9104 for (unsigned i = AddrOpIdx + 1; i < N->getNumOperands() - 1; ++i)
9105 Ops.push_back(N->getOperand(i));
9108 // For all node types, the alignment operand is always the last one.
9109 Ops.push_back(DAG.getConstant(Alignment, dl, MVT::i32));
9111 // If this is a non-standard-aligned STORE, the penultimate operand is the
9112 // stored value. Bitcast it to the aligned type.
9113 if (AlignedVecTy != VecTy && N->getOpcode() == ISD::STORE) {
9114 SDValue &StVal = Ops[Ops.size()-2];
9115 StVal = DAG.getNode(ISD::BITCAST, dl, AlignedVecTy, StVal);
9118 SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, dl, SDTys,
9120 MemN->getMemOperand());
9123 SmallVector<SDValue, 5> NewResults;
9124 for (unsigned i = 0; i < NumResultVecs; ++i)
9125 NewResults.push_back(SDValue(UpdN.getNode(), i));
9127 // If this is an non-standard-aligned LOAD, the first result is the loaded
9128 // value. Bitcast it to the expected result type.
9129 if (AlignedVecTy != VecTy && N->getOpcode() == ISD::LOAD) {
9130 SDValue &LdVal = NewResults[0];
9131 LdVal = DAG.getNode(ISD::BITCAST, dl, VecTy, LdVal);
9134 NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs+1)); // chain
9135 DCI.CombineTo(N, NewResults);
9136 DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs));
9143 static SDValue PerformVLDCombine(SDNode *N,
9144 TargetLowering::DAGCombinerInfo &DCI) {
9145 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
9148 return CombineBaseUpdate(N, DCI);
9151 /// CombineVLDDUP - For a VDUPLANE node N, check if its source operand is a
9152 /// vldN-lane (N > 1) intrinsic, and if all the other uses of that intrinsic
9153 /// are also VDUPLANEs. If so, combine them to a vldN-dup operation and
9155 static bool CombineVLDDUP(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
9156 SelectionDAG &DAG = DCI.DAG;
9157 EVT VT = N->getValueType(0);
9158 // vldN-dup instructions only support 64-bit vectors for N > 1.
9159 if (!VT.is64BitVector())
9162 // Check if the VDUPLANE operand is a vldN-dup intrinsic.
9163 SDNode *VLD = N->getOperand(0).getNode();
9164 if (VLD->getOpcode() != ISD::INTRINSIC_W_CHAIN)
9166 unsigned NumVecs = 0;
9167 unsigned NewOpc = 0;
9168 unsigned IntNo = cast<ConstantSDNode>(VLD->getOperand(1))->getZExtValue();
9169 if (IntNo == Intrinsic::arm_neon_vld2lane) {
9171 NewOpc = ARMISD::VLD2DUP;
9172 } else if (IntNo == Intrinsic::arm_neon_vld3lane) {
9174 NewOpc = ARMISD::VLD3DUP;
9175 } else if (IntNo == Intrinsic::arm_neon_vld4lane) {
9177 NewOpc = ARMISD::VLD4DUP;
9182 // First check that all the vldN-lane uses are VDUPLANEs and that the lane
9183 // numbers match the load.
9184 unsigned VLDLaneNo =
9185 cast<ConstantSDNode>(VLD->getOperand(NumVecs+3))->getZExtValue();
9186 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
9188 // Ignore uses of the chain result.
9189 if (UI.getUse().getResNo() == NumVecs)
9192 if (User->getOpcode() != ARMISD::VDUPLANE ||
9193 VLDLaneNo != cast<ConstantSDNode>(User->getOperand(1))->getZExtValue())
9197 // Create the vldN-dup node.
9200 for (n = 0; n < NumVecs; ++n)
9202 Tys[n] = MVT::Other;
9203 SDVTList SDTys = DAG.getVTList(makeArrayRef(Tys, NumVecs+1));
9204 SDValue Ops[] = { VLD->getOperand(0), VLD->getOperand(2) };
9205 MemIntrinsicSDNode *VLDMemInt = cast<MemIntrinsicSDNode>(VLD);
9206 SDValue VLDDup = DAG.getMemIntrinsicNode(NewOpc, SDLoc(VLD), SDTys,
9207 Ops, VLDMemInt->getMemoryVT(),
9208 VLDMemInt->getMemOperand());
9211 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
9213 unsigned ResNo = UI.getUse().getResNo();
9214 // Ignore uses of the chain result.
9215 if (ResNo == NumVecs)
9218 DCI.CombineTo(User, SDValue(VLDDup.getNode(), ResNo));
9221 // Now the vldN-lane intrinsic is dead except for its chain result.
9222 // Update uses of the chain.
9223 std::vector<SDValue> VLDDupResults;
9224 for (unsigned n = 0; n < NumVecs; ++n)
9225 VLDDupResults.push_back(SDValue(VLDDup.getNode(), n));
9226 VLDDupResults.push_back(SDValue(VLDDup.getNode(), NumVecs));
9227 DCI.CombineTo(VLD, VLDDupResults);
9232 /// PerformVDUPLANECombine - Target-specific dag combine xforms for
9233 /// ARMISD::VDUPLANE.
9234 static SDValue PerformVDUPLANECombine(SDNode *N,
9235 TargetLowering::DAGCombinerInfo &DCI) {
9236 SDValue Op = N->getOperand(0);
9238 // If the source is a vldN-lane (N > 1) intrinsic, and all the other uses
9239 // of that intrinsic are also VDUPLANEs, combine them to a vldN-dup operation.
9240 if (CombineVLDDUP(N, DCI))
9241 return SDValue(N, 0);
9243 // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is
9244 // redundant. Ignore bit_converts for now; element sizes are checked below.
9245 while (Op.getOpcode() == ISD::BITCAST)
9246 Op = Op.getOperand(0);
9247 if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM)
9250 // Make sure the VMOV element size is not bigger than the VDUPLANE elements.
9251 unsigned EltSize = Op.getValueType().getVectorElementType().getSizeInBits();
9252 // The canonical VMOV for a zero vector uses a 32-bit element size.
9253 unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
9255 if (ARM_AM::decodeNEONModImm(Imm, EltBits) == 0)
9257 EVT VT = N->getValueType(0);
9258 if (EltSize > VT.getVectorElementType().getSizeInBits())
9261 return DCI.DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Op);
9264 static SDValue PerformLOADCombine(SDNode *N,
9265 TargetLowering::DAGCombinerInfo &DCI) {
9266 EVT VT = N->getValueType(0);
9268 // If this is a legal vector load, try to combine it into a VLD1_UPD.
9269 if (ISD::isNormalLoad(N) && VT.isVector() &&
9270 DCI.DAG.getTargetLoweringInfo().isTypeLegal(VT))
9271 return CombineBaseUpdate(N, DCI);
9276 /// PerformSTORECombine - Target-specific dag combine xforms for
9278 static SDValue PerformSTORECombine(SDNode *N,
9279 TargetLowering::DAGCombinerInfo &DCI) {
9280 StoreSDNode *St = cast<StoreSDNode>(N);
9281 if (St->isVolatile())
9284 // Optimize trunc store (of multiple scalars) to shuffle and store. First,
9285 // pack all of the elements in one place. Next, store to memory in fewer
9287 SDValue StVal = St->getValue();
9288 EVT VT = StVal.getValueType();
9289 if (St->isTruncatingStore() && VT.isVector()) {
9290 SelectionDAG &DAG = DCI.DAG;
9291 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9292 EVT StVT = St->getMemoryVT();
9293 unsigned NumElems = VT.getVectorNumElements();
9294 assert(StVT != VT && "Cannot truncate to the same type");
9295 unsigned FromEltSz = VT.getVectorElementType().getSizeInBits();
9296 unsigned ToEltSz = StVT.getVectorElementType().getSizeInBits();
9298 // From, To sizes and ElemCount must be pow of two
9299 if (!isPowerOf2_32(NumElems * FromEltSz * ToEltSz)) return SDValue();
9301 // We are going to use the original vector elt for storing.
9302 // Accumulated smaller vector elements must be a multiple of the store size.
9303 if (0 != (NumElems * FromEltSz) % ToEltSz) return SDValue();
9305 unsigned SizeRatio = FromEltSz / ToEltSz;
9306 assert(SizeRatio * NumElems * ToEltSz == VT.getSizeInBits());
9308 // Create a type on which we perform the shuffle.
9309 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(), StVT.getScalarType(),
9310 NumElems*SizeRatio);
9311 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
9314 SDValue WideVec = DAG.getNode(ISD::BITCAST, DL, WideVecVT, StVal);
9315 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
9316 for (unsigned i = 0; i < NumElems; ++i)
9317 ShuffleVec[i] = TLI.isBigEndian() ? (i+1) * SizeRatio - 1 : i * SizeRatio;
9319 // Can't shuffle using an illegal type.
9320 if (!TLI.isTypeLegal(WideVecVT)) return SDValue();
9322 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, DL, WideVec,
9323 DAG.getUNDEF(WideVec.getValueType()),
9325 // At this point all of the data is stored at the bottom of the
9326 // register. We now need to save it to mem.
9328 // Find the largest store unit
9329 MVT StoreType = MVT::i8;
9330 for (MVT Tp : MVT::integer_valuetypes()) {
9331 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToEltSz)
9334 // Didn't find a legal store type.
9335 if (!TLI.isTypeLegal(StoreType))
9338 // Bitcast the original vector into a vector of store-size units
9339 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
9340 StoreType, VT.getSizeInBits()/EVT(StoreType).getSizeInBits());
9341 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
9342 SDValue ShuffWide = DAG.getNode(ISD::BITCAST, DL, StoreVecVT, Shuff);
9343 SmallVector<SDValue, 8> Chains;
9344 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits()/8, DL,
9345 TLI.getPointerTy());
9346 SDValue BasePtr = St->getBasePtr();
9348 // Perform one or more big stores into memory.
9349 unsigned E = (ToEltSz*NumElems)/StoreType.getSizeInBits();
9350 for (unsigned I = 0; I < E; I++) {
9351 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
9352 StoreType, ShuffWide,
9353 DAG.getIntPtrConstant(I, DL));
9354 SDValue Ch = DAG.getStore(St->getChain(), DL, SubVec, BasePtr,
9355 St->getPointerInfo(), St->isVolatile(),
9356 St->isNonTemporal(), St->getAlignment());
9357 BasePtr = DAG.getNode(ISD::ADD, DL, BasePtr.getValueType(), BasePtr,
9359 Chains.push_back(Ch);
9361 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
9364 if (!ISD::isNormalStore(St))
9367 // Split a store of a VMOVDRR into two integer stores to avoid mixing NEON and
9368 // ARM stores of arguments in the same cache line.
9369 if (StVal.getNode()->getOpcode() == ARMISD::VMOVDRR &&
9370 StVal.getNode()->hasOneUse()) {
9371 SelectionDAG &DAG = DCI.DAG;
9372 bool isBigEndian = DAG.getTargetLoweringInfo().isBigEndian();
9374 SDValue BasePtr = St->getBasePtr();
9375 SDValue NewST1 = DAG.getStore(St->getChain(), DL,
9376 StVal.getNode()->getOperand(isBigEndian ? 1 : 0 ),
9377 BasePtr, St->getPointerInfo(), St->isVolatile(),
9378 St->isNonTemporal(), St->getAlignment());
9380 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
9381 DAG.getConstant(4, DL, MVT::i32));
9382 return DAG.getStore(NewST1.getValue(0), DL,
9383 StVal.getNode()->getOperand(isBigEndian ? 0 : 1),
9384 OffsetPtr, St->getPointerInfo(), St->isVolatile(),
9385 St->isNonTemporal(),
9386 std::min(4U, St->getAlignment() / 2));
9389 if (StVal.getValueType() == MVT::i64 &&
9390 StVal.getNode()->getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
9392 // Bitcast an i64 store extracted from a vector to f64.
9393 // Otherwise, the i64 value will be legalized to a pair of i32 values.
9394 SelectionDAG &DAG = DCI.DAG;
9396 SDValue IntVec = StVal.getOperand(0);
9397 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
9398 IntVec.getValueType().getVectorNumElements());
9399 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, IntVec);
9400 SDValue ExtElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
9401 Vec, StVal.getOperand(1));
9403 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ExtElt);
9404 // Make the DAGCombiner fold the bitcasts.
9405 DCI.AddToWorklist(Vec.getNode());
9406 DCI.AddToWorklist(ExtElt.getNode());
9407 DCI.AddToWorklist(V.getNode());
9408 return DAG.getStore(St->getChain(), dl, V, St->getBasePtr(),
9409 St->getPointerInfo(), St->isVolatile(),
9410 St->isNonTemporal(), St->getAlignment(),
9414 // If this is a legal vector store, try to combine it into a VST1_UPD.
9415 if (ISD::isNormalStore(N) && VT.isVector() &&
9416 DCI.DAG.getTargetLoweringInfo().isTypeLegal(VT))
9417 return CombineBaseUpdate(N, DCI);
9422 // isConstVecPow2 - Return true if each vector element is a power of 2, all
9423 // elements are the same constant, C, and Log2(C) ranges from 1 to 32.
9424 static bool isConstVecPow2(SDValue ConstVec, bool isSigned, uint64_t &C)
9428 for (unsigned I = 0, E = ConstVec.getValueType().getVectorNumElements();
9430 ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(ConstVec.getOperand(I));
9435 APFloat APF = C->getValueAPF();
9436 if (APF.convertToInteger(&cN, 64, isSigned, APFloat::rmTowardZero, &isExact)
9437 != APFloat::opOK || !isExact)
9440 c0 = (I == 0) ? cN : c0;
9441 if (!isPowerOf2_64(cN) || c0 != cN || Log2_64(c0) < 1 || Log2_64(c0) > 32)
9448 /// PerformVCVTCombine - VCVT (floating-point to fixed-point, Advanced SIMD)
9449 /// can replace combinations of VMUL and VCVT (floating-point to integer)
9450 /// when the VMUL has a constant operand that is a power of 2.
9452 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
9453 /// vmul.f32 d16, d17, d16
9454 /// vcvt.s32.f32 d16, d16
9456 /// vcvt.s32.f32 d16, d16, #3
9457 static SDValue PerformVCVTCombine(SDNode *N,
9458 TargetLowering::DAGCombinerInfo &DCI,
9459 const ARMSubtarget *Subtarget) {
9460 SelectionDAG &DAG = DCI.DAG;
9461 SDValue Op = N->getOperand(0);
9463 if (!Subtarget->hasNEON() || !Op.getValueType().isVector() ||
9464 Op.getOpcode() != ISD::FMUL)
9468 SDValue N0 = Op->getOperand(0);
9469 SDValue ConstVec = Op->getOperand(1);
9470 bool isSigned = N->getOpcode() == ISD::FP_TO_SINT;
9472 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
9473 !isConstVecPow2(ConstVec, isSigned, C))
9476 MVT FloatTy = Op.getSimpleValueType().getVectorElementType();
9477 MVT IntTy = N->getSimpleValueType(0).getVectorElementType();
9478 unsigned NumLanes = Op.getValueType().getVectorNumElements();
9479 if (FloatTy.getSizeInBits() != 32 || IntTy.getSizeInBits() > 32 ||
9481 // These instructions only exist converting from f32 to i32. We can handle
9482 // smaller integers by generating an extra truncate, but larger ones would
9483 // be lossy. We also can't handle more then 4 lanes, since these intructions
9484 // only support v2i32/v4i32 types.
9489 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfp2fxs :
9490 Intrinsic::arm_neon_vcvtfp2fxu;
9491 SDValue FixConv = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl,
9492 NumLanes == 2 ? MVT::v2i32 : MVT::v4i32,
9493 DAG.getConstant(IntrinsicOpcode, dl, MVT::i32),
9495 DAG.getConstant(Log2_64(C), dl, MVT::i32));
9497 if (IntTy.getSizeInBits() < FloatTy.getSizeInBits())
9498 FixConv = DAG.getNode(ISD::TRUNCATE, dl, N->getValueType(0), FixConv);
9503 /// PerformVDIVCombine - VCVT (fixed-point to floating-point, Advanced SIMD)
9504 /// can replace combinations of VCVT (integer to floating-point) and VDIV
9505 /// when the VDIV has a constant operand that is a power of 2.
9507 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
9508 /// vcvt.f32.s32 d16, d16
9509 /// vdiv.f32 d16, d17, d16
9511 /// vcvt.f32.s32 d16, d16, #3
9512 static SDValue PerformVDIVCombine(SDNode *N,
9513 TargetLowering::DAGCombinerInfo &DCI,
9514 const ARMSubtarget *Subtarget) {
9515 SelectionDAG &DAG = DCI.DAG;
9516 SDValue Op = N->getOperand(0);
9517 unsigned OpOpcode = Op.getNode()->getOpcode();
9519 if (!Subtarget->hasNEON() || !N->getValueType(0).isVector() ||
9520 (OpOpcode != ISD::SINT_TO_FP && OpOpcode != ISD::UINT_TO_FP))
9524 SDValue ConstVec = N->getOperand(1);
9525 bool isSigned = OpOpcode == ISD::SINT_TO_FP;
9527 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
9528 !isConstVecPow2(ConstVec, isSigned, C))
9531 MVT FloatTy = N->getSimpleValueType(0).getVectorElementType();
9532 MVT IntTy = Op.getOperand(0).getSimpleValueType().getVectorElementType();
9533 if (FloatTy.getSizeInBits() != 32 || IntTy.getSizeInBits() > 32) {
9534 // These instructions only exist converting from i32 to f32. We can handle
9535 // smaller integers by generating an extra extend, but larger ones would
9541 SDValue ConvInput = Op.getOperand(0);
9542 unsigned NumLanes = Op.getValueType().getVectorNumElements();
9543 if (IntTy.getSizeInBits() < FloatTy.getSizeInBits())
9544 ConvInput = DAG.getNode(isSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND,
9545 dl, NumLanes == 2 ? MVT::v2i32 : MVT::v4i32,
9548 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfxs2fp :
9549 Intrinsic::arm_neon_vcvtfxu2fp;
9550 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl,
9552 DAG.getConstant(IntrinsicOpcode, dl, MVT::i32),
9553 ConvInput, DAG.getConstant(Log2_64(C), dl, MVT::i32));
9556 /// Getvshiftimm - Check if this is a valid build_vector for the immediate
9557 /// operand of a vector shift operation, where all the elements of the
9558 /// build_vector must have the same constant integer value.
9559 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
9560 // Ignore bit_converts.
9561 while (Op.getOpcode() == ISD::BITCAST)
9562 Op = Op.getOperand(0);
9563 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
9564 APInt SplatBits, SplatUndef;
9565 unsigned SplatBitSize;
9567 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
9568 HasAnyUndefs, ElementBits) ||
9569 SplatBitSize > ElementBits)
9571 Cnt = SplatBits.getSExtValue();
9575 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
9576 /// operand of a vector shift left operation. That value must be in the range:
9577 /// 0 <= Value < ElementBits for a left shift; or
9578 /// 0 <= Value <= ElementBits for a long left shift.
9579 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
9580 assert(VT.isVector() && "vector shift count is not a vector type");
9581 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
9582 if (! getVShiftImm(Op, ElementBits, Cnt))
9584 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
9587 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
9588 /// operand of a vector shift right operation. For a shift opcode, the value
9589 /// is positive, but for an intrinsic the value count must be negative. The
9590 /// absolute value must be in the range:
9591 /// 1 <= |Value| <= ElementBits for a right shift; or
9592 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
9593 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
9595 assert(VT.isVector() && "vector shift count is not a vector type");
9596 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
9597 if (! getVShiftImm(Op, ElementBits, Cnt))
9601 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
9604 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
9605 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
9606 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
9609 // Don't do anything for most intrinsics.
9612 // Vector shifts: check for immediate versions and lower them.
9613 // Note: This is done during DAG combining instead of DAG legalizing because
9614 // the build_vectors for 64-bit vector element shift counts are generally
9615 // not legal, and it is hard to see their values after they get legalized to
9616 // loads from a constant pool.
9617 case Intrinsic::arm_neon_vshifts:
9618 case Intrinsic::arm_neon_vshiftu:
9619 case Intrinsic::arm_neon_vrshifts:
9620 case Intrinsic::arm_neon_vrshiftu:
9621 case Intrinsic::arm_neon_vrshiftn:
9622 case Intrinsic::arm_neon_vqshifts:
9623 case Intrinsic::arm_neon_vqshiftu:
9624 case Intrinsic::arm_neon_vqshiftsu:
9625 case Intrinsic::arm_neon_vqshiftns:
9626 case Intrinsic::arm_neon_vqshiftnu:
9627 case Intrinsic::arm_neon_vqshiftnsu:
9628 case Intrinsic::arm_neon_vqrshiftns:
9629 case Intrinsic::arm_neon_vqrshiftnu:
9630 case Intrinsic::arm_neon_vqrshiftnsu: {
9631 EVT VT = N->getOperand(1).getValueType();
9633 unsigned VShiftOpc = 0;
9636 case Intrinsic::arm_neon_vshifts:
9637 case Intrinsic::arm_neon_vshiftu:
9638 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
9639 VShiftOpc = ARMISD::VSHL;
9642 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
9643 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
9644 ARMISD::VSHRs : ARMISD::VSHRu);
9649 case Intrinsic::arm_neon_vrshifts:
9650 case Intrinsic::arm_neon_vrshiftu:
9651 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
9655 case Intrinsic::arm_neon_vqshifts:
9656 case Intrinsic::arm_neon_vqshiftu:
9657 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
9661 case Intrinsic::arm_neon_vqshiftsu:
9662 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
9664 llvm_unreachable("invalid shift count for vqshlu intrinsic");
9666 case Intrinsic::arm_neon_vrshiftn:
9667 case Intrinsic::arm_neon_vqshiftns:
9668 case Intrinsic::arm_neon_vqshiftnu:
9669 case Intrinsic::arm_neon_vqshiftnsu:
9670 case Intrinsic::arm_neon_vqrshiftns:
9671 case Intrinsic::arm_neon_vqrshiftnu:
9672 case Intrinsic::arm_neon_vqrshiftnsu:
9673 // Narrowing shifts require an immediate right shift.
9674 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
9676 llvm_unreachable("invalid shift count for narrowing vector shift "
9680 llvm_unreachable("unhandled vector shift");
9684 case Intrinsic::arm_neon_vshifts:
9685 case Intrinsic::arm_neon_vshiftu:
9686 // Opcode already set above.
9688 case Intrinsic::arm_neon_vrshifts:
9689 VShiftOpc = ARMISD::VRSHRs; break;
9690 case Intrinsic::arm_neon_vrshiftu:
9691 VShiftOpc = ARMISD::VRSHRu; break;
9692 case Intrinsic::arm_neon_vrshiftn:
9693 VShiftOpc = ARMISD::VRSHRN; break;
9694 case Intrinsic::arm_neon_vqshifts:
9695 VShiftOpc = ARMISD::VQSHLs; break;
9696 case Intrinsic::arm_neon_vqshiftu:
9697 VShiftOpc = ARMISD::VQSHLu; break;
9698 case Intrinsic::arm_neon_vqshiftsu:
9699 VShiftOpc = ARMISD::VQSHLsu; break;
9700 case Intrinsic::arm_neon_vqshiftns:
9701 VShiftOpc = ARMISD::VQSHRNs; break;
9702 case Intrinsic::arm_neon_vqshiftnu:
9703 VShiftOpc = ARMISD::VQSHRNu; break;
9704 case Intrinsic::arm_neon_vqshiftnsu:
9705 VShiftOpc = ARMISD::VQSHRNsu; break;
9706 case Intrinsic::arm_neon_vqrshiftns:
9707 VShiftOpc = ARMISD::VQRSHRNs; break;
9708 case Intrinsic::arm_neon_vqrshiftnu:
9709 VShiftOpc = ARMISD::VQRSHRNu; break;
9710 case Intrinsic::arm_neon_vqrshiftnsu:
9711 VShiftOpc = ARMISD::VQRSHRNsu; break;
9715 return DAG.getNode(VShiftOpc, dl, N->getValueType(0),
9716 N->getOperand(1), DAG.getConstant(Cnt, dl, MVT::i32));
9719 case Intrinsic::arm_neon_vshiftins: {
9720 EVT VT = N->getOperand(1).getValueType();
9722 unsigned VShiftOpc = 0;
9724 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
9725 VShiftOpc = ARMISD::VSLI;
9726 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
9727 VShiftOpc = ARMISD::VSRI;
9729 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
9733 return DAG.getNode(VShiftOpc, dl, N->getValueType(0),
9734 N->getOperand(1), N->getOperand(2),
9735 DAG.getConstant(Cnt, dl, MVT::i32));
9738 case Intrinsic::arm_neon_vqrshifts:
9739 case Intrinsic::arm_neon_vqrshiftu:
9740 // No immediate versions of these to check for.
9747 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
9748 /// lowers them. As with the vector shift intrinsics, this is done during DAG
9749 /// combining instead of DAG legalizing because the build_vectors for 64-bit
9750 /// vector element shift counts are generally not legal, and it is hard to see
9751 /// their values after they get legalized to loads from a constant pool.
9752 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
9753 const ARMSubtarget *ST) {
9754 EVT VT = N->getValueType(0);
9755 if (N->getOpcode() == ISD::SRL && VT == MVT::i32 && ST->hasV6Ops()) {
9756 // Canonicalize (srl (bswap x), 16) to (rotr (bswap x), 16) if the high
9757 // 16-bits of x is zero. This optimizes rev + lsr 16 to rev16.
9758 SDValue N1 = N->getOperand(1);
9759 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) {
9760 SDValue N0 = N->getOperand(0);
9761 if (C->getZExtValue() == 16 && N0.getOpcode() == ISD::BSWAP &&
9762 DAG.MaskedValueIsZero(N0.getOperand(0),
9763 APInt::getHighBitsSet(32, 16)))
9764 return DAG.getNode(ISD::ROTR, SDLoc(N), VT, N0, N1);
9768 // Nothing to be done for scalar shifts.
9769 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9770 if (!VT.isVector() || !TLI.isTypeLegal(VT))
9773 assert(ST->hasNEON() && "unexpected vector shift");
9776 switch (N->getOpcode()) {
9777 default: llvm_unreachable("unexpected shift opcode");
9780 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt)) {
9782 return DAG.getNode(ARMISD::VSHL, dl, VT, N->getOperand(0),
9783 DAG.getConstant(Cnt, dl, MVT::i32));
9789 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
9790 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
9791 ARMISD::VSHRs : ARMISD::VSHRu);
9793 return DAG.getNode(VShiftOpc, dl, VT, N->getOperand(0),
9794 DAG.getConstant(Cnt, dl, MVT::i32));
9800 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
9801 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
9802 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
9803 const ARMSubtarget *ST) {
9804 SDValue N0 = N->getOperand(0);
9806 // Check for sign- and zero-extensions of vector extract operations of 8-
9807 // and 16-bit vector elements. NEON supports these directly. They are
9808 // handled during DAG combining because type legalization will promote them
9809 // to 32-bit types and it is messy to recognize the operations after that.
9810 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
9811 SDValue Vec = N0.getOperand(0);
9812 SDValue Lane = N0.getOperand(1);
9813 EVT VT = N->getValueType(0);
9814 EVT EltVT = N0.getValueType();
9815 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9817 if (VT == MVT::i32 &&
9818 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
9819 TLI.isTypeLegal(Vec.getValueType()) &&
9820 isa<ConstantSDNode>(Lane)) {
9823 switch (N->getOpcode()) {
9824 default: llvm_unreachable("unexpected opcode");
9825 case ISD::SIGN_EXTEND:
9826 Opc = ARMISD::VGETLANEs;
9828 case ISD::ZERO_EXTEND:
9829 case ISD::ANY_EXTEND:
9830 Opc = ARMISD::VGETLANEu;
9833 return DAG.getNode(Opc, SDLoc(N), VT, Vec, Lane);
9840 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
9841 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
9842 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
9843 const ARMSubtarget *ST) {
9844 // If the target supports NEON, try to use vmax/vmin instructions for f32
9845 // selects like "x < y ? x : y". Unless the NoNaNsFPMath option is set,
9846 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
9847 // a NaN; only do the transformation when it matches that behavior.
9849 // For now only do this when using NEON for FP operations; if using VFP, it
9850 // is not obvious that the benefit outweighs the cost of switching to the
9852 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
9853 N->getValueType(0) != MVT::f32)
9856 SDValue CondLHS = N->getOperand(0);
9857 SDValue CondRHS = N->getOperand(1);
9858 SDValue LHS = N->getOperand(2);
9859 SDValue RHS = N->getOperand(3);
9860 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
9862 unsigned Opcode = 0;
9864 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
9865 IsReversed = false; // x CC y ? x : y
9866 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
9867 IsReversed = true ; // x CC y ? y : x
9881 // If LHS is NaN, an ordered comparison will be false and the result will
9882 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
9883 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
9884 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
9885 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
9887 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
9888 // will return -0, so vmin can only be used for unsafe math or if one of
9889 // the operands is known to be nonzero.
9890 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
9891 !DAG.getTarget().Options.UnsafeFPMath &&
9892 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9894 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
9903 // If LHS is NaN, an ordered comparison will be false and the result will
9904 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
9905 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
9906 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
9907 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
9909 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
9910 // will return +0, so vmax can only be used for unsafe math or if one of
9911 // the operands is known to be nonzero.
9912 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
9913 !DAG.getTarget().Options.UnsafeFPMath &&
9914 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9916 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
9922 return DAG.getNode(Opcode, SDLoc(N), N->getValueType(0), LHS, RHS);
9925 /// PerformCMOVCombine - Target-specific DAG combining for ARMISD::CMOV.
9927 ARMTargetLowering::PerformCMOVCombine(SDNode *N, SelectionDAG &DAG) const {
9928 SDValue Cmp = N->getOperand(4);
9929 if (Cmp.getOpcode() != ARMISD::CMPZ)
9930 // Only looking at EQ and NE cases.
9933 EVT VT = N->getValueType(0);
9935 SDValue LHS = Cmp.getOperand(0);
9936 SDValue RHS = Cmp.getOperand(1);
9937 SDValue FalseVal = N->getOperand(0);
9938 SDValue TrueVal = N->getOperand(1);
9939 SDValue ARMcc = N->getOperand(2);
9940 ARMCC::CondCodes CC =
9941 (ARMCC::CondCodes)cast<ConstantSDNode>(ARMcc)->getZExtValue();
9959 /// FIXME: Turn this into a target neutral optimization?
9961 if (CC == ARMCC::NE && FalseVal == RHS && FalseVal != LHS) {
9962 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, TrueVal, ARMcc,
9963 N->getOperand(3), Cmp);
9964 } else if (CC == ARMCC::EQ && TrueVal == RHS) {
9966 SDValue NewCmp = getARMCmp(LHS, RHS, ISD::SETNE, ARMcc, DAG, dl);
9967 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, FalseVal, ARMcc,
9968 N->getOperand(3), NewCmp);
9971 if (Res.getNode()) {
9972 APInt KnownZero, KnownOne;
9973 DAG.computeKnownBits(SDValue(N,0), KnownZero, KnownOne);
9974 // Capture demanded bits information that would be otherwise lost.
9975 if (KnownZero == 0xfffffffe)
9976 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
9977 DAG.getValueType(MVT::i1));
9978 else if (KnownZero == 0xffffff00)
9979 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
9980 DAG.getValueType(MVT::i8));
9981 else if (KnownZero == 0xffff0000)
9982 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
9983 DAG.getValueType(MVT::i16));
9989 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
9990 DAGCombinerInfo &DCI) const {
9991 switch (N->getOpcode()) {
9993 case ISD::ADDC: return PerformADDCCombine(N, DCI, Subtarget);
9994 case ISD::ADD: return PerformADDCombine(N, DCI, Subtarget);
9995 case ISD::SUB: return PerformSUBCombine(N, DCI);
9996 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
9997 case ISD::OR: return PerformORCombine(N, DCI, Subtarget);
9998 case ISD::XOR: return PerformXORCombine(N, DCI, Subtarget);
9999 case ISD::AND: return PerformANDCombine(N, DCI, Subtarget);
10000 case ARMISD::BFI: return PerformBFICombine(N, DCI);
10001 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI, Subtarget);
10002 case ARMISD::VMOVDRR: return PerformVMOVDRRCombine(N, DCI.DAG);
10003 case ISD::STORE: return PerformSTORECombine(N, DCI);
10004 case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DCI, Subtarget);
10005 case ISD::INSERT_VECTOR_ELT: return PerformInsertEltCombine(N, DCI);
10006 case ISD::VECTOR_SHUFFLE: return PerformVECTOR_SHUFFLECombine(N, DCI.DAG);
10007 case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI);
10008 case ISD::FP_TO_SINT:
10009 case ISD::FP_TO_UINT: return PerformVCVTCombine(N, DCI, Subtarget);
10010 case ISD::FDIV: return PerformVDIVCombine(N, DCI, Subtarget);
10011 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
10014 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
10015 case ISD::SIGN_EXTEND:
10016 case ISD::ZERO_EXTEND:
10017 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
10018 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
10019 case ARMISD::CMOV: return PerformCMOVCombine(N, DCI.DAG);
10020 case ISD::LOAD: return PerformLOADCombine(N, DCI);
10021 case ARMISD::VLD2DUP:
10022 case ARMISD::VLD3DUP:
10023 case ARMISD::VLD4DUP:
10024 return PerformVLDCombine(N, DCI);
10025 case ARMISD::BUILD_VECTOR:
10026 return PerformARMBUILD_VECTORCombine(N, DCI);
10027 case ISD::INTRINSIC_VOID:
10028 case ISD::INTRINSIC_W_CHAIN:
10029 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
10030 case Intrinsic::arm_neon_vld1:
10031 case Intrinsic::arm_neon_vld2:
10032 case Intrinsic::arm_neon_vld3:
10033 case Intrinsic::arm_neon_vld4:
10034 case Intrinsic::arm_neon_vld2lane:
10035 case Intrinsic::arm_neon_vld3lane:
10036 case Intrinsic::arm_neon_vld4lane:
10037 case Intrinsic::arm_neon_vst1:
10038 case Intrinsic::arm_neon_vst2:
10039 case Intrinsic::arm_neon_vst3:
10040 case Intrinsic::arm_neon_vst4:
10041 case Intrinsic::arm_neon_vst2lane:
10042 case Intrinsic::arm_neon_vst3lane:
10043 case Intrinsic::arm_neon_vst4lane:
10044 return PerformVLDCombine(N, DCI);
10052 bool ARMTargetLowering::isDesirableToTransformToIntegerOp(unsigned Opc,
10054 return (VT == MVT::f32) && (Opc == ISD::LOAD || Opc == ISD::STORE);
10057 bool ARMTargetLowering::allowsMisalignedMemoryAccesses(EVT VT,
10060 bool *Fast) const {
10061 // The AllowsUnaliged flag models the SCTLR.A setting in ARM cpus
10062 bool AllowsUnaligned = Subtarget->allowsUnalignedMem();
10064 switch (VT.getSimpleVT().SimpleTy) {
10070 // Unaligned access can use (for example) LRDB, LRDH, LDR
10071 if (AllowsUnaligned) {
10073 *Fast = Subtarget->hasV7Ops();
10080 // For any little-endian targets with neon, we can support unaligned ld/st
10081 // of D and Q (e.g. {D0,D1}) registers by using vld1.i8/vst1.i8.
10082 // A big-endian target may also explicitly support unaligned accesses
10083 if (Subtarget->hasNEON() && (AllowsUnaligned || isLittleEndian())) {
10093 static bool memOpAlign(unsigned DstAlign, unsigned SrcAlign,
10094 unsigned AlignCheck) {
10095 return ((SrcAlign == 0 || SrcAlign % AlignCheck == 0) &&
10096 (DstAlign == 0 || DstAlign % AlignCheck == 0));
10099 EVT ARMTargetLowering::getOptimalMemOpType(uint64_t Size,
10100 unsigned DstAlign, unsigned SrcAlign,
10101 bool IsMemset, bool ZeroMemset,
10103 MachineFunction &MF) const {
10104 const Function *F = MF.getFunction();
10106 // See if we can use NEON instructions for this...
10107 if ((!IsMemset || ZeroMemset) && Subtarget->hasNEON() &&
10108 !F->hasFnAttribute(Attribute::NoImplicitFloat)) {
10111 (memOpAlign(SrcAlign, DstAlign, 16) ||
10112 (allowsMisalignedMemoryAccesses(MVT::v2f64, 0, 1, &Fast) && Fast))) {
10114 } else if (Size >= 8 &&
10115 (memOpAlign(SrcAlign, DstAlign, 8) ||
10116 (allowsMisalignedMemoryAccesses(MVT::f64, 0, 1, &Fast) &&
10122 // Lowering to i32/i16 if the size permits.
10125 else if (Size >= 2)
10128 // Let the target-independent logic figure it out.
10132 bool ARMTargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
10133 if (Val.getOpcode() != ISD::LOAD)
10136 EVT VT1 = Val.getValueType();
10137 if (!VT1.isSimple() || !VT1.isInteger() ||
10138 !VT2.isSimple() || !VT2.isInteger())
10141 switch (VT1.getSimpleVT().SimpleTy) {
10146 // 8-bit and 16-bit loads implicitly zero-extend to 32-bits.
10153 bool ARMTargetLowering::isVectorLoadExtDesirable(SDValue ExtVal) const {
10154 EVT VT = ExtVal.getValueType();
10156 if (!isTypeLegal(VT))
10159 // Don't create a loadext if we can fold the extension into a wide/long
10161 // If there's more than one user instruction, the loadext is desirable no
10162 // matter what. There can be two uses by the same instruction.
10163 if (ExtVal->use_empty() ||
10164 !ExtVal->use_begin()->isOnlyUserOf(ExtVal.getNode()))
10167 SDNode *U = *ExtVal->use_begin();
10168 if ((U->getOpcode() == ISD::ADD || U->getOpcode() == ISD::SUB ||
10169 U->getOpcode() == ISD::SHL || U->getOpcode() == ARMISD::VSHL))
10175 bool ARMTargetLowering::allowTruncateForTailCall(Type *Ty1, Type *Ty2) const {
10176 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
10179 if (!isTypeLegal(EVT::getEVT(Ty1)))
10182 assert(Ty1->getPrimitiveSizeInBits() <= 64 && "i128 is probably not a noop");
10184 // Assuming the caller doesn't have a zeroext or signext return parameter,
10185 // truncation all the way down to i1 is valid.
10190 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
10194 unsigned Scale = 1;
10195 switch (VT.getSimpleVT().SimpleTy) {
10196 default: return false;
10211 if ((V & (Scale - 1)) != 0)
10214 return V == (V & ((1LL << 5) - 1));
10217 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
10218 const ARMSubtarget *Subtarget) {
10219 bool isNeg = false;
10225 switch (VT.getSimpleVT().SimpleTy) {
10226 default: return false;
10231 // + imm12 or - imm8
10233 return V == (V & ((1LL << 8) - 1));
10234 return V == (V & ((1LL << 12) - 1));
10237 // Same as ARM mode. FIXME: NEON?
10238 if (!Subtarget->hasVFP2())
10243 return V == (V & ((1LL << 8) - 1));
10247 /// isLegalAddressImmediate - Return true if the integer value can be used
10248 /// as the offset of the target addressing mode for load / store of the
10250 static bool isLegalAddressImmediate(int64_t V, EVT VT,
10251 const ARMSubtarget *Subtarget) {
10255 if (!VT.isSimple())
10258 if (Subtarget->isThumb1Only())
10259 return isLegalT1AddressImmediate(V, VT);
10260 else if (Subtarget->isThumb2())
10261 return isLegalT2AddressImmediate(V, VT, Subtarget);
10266 switch (VT.getSimpleVT().SimpleTy) {
10267 default: return false;
10272 return V == (V & ((1LL << 12) - 1));
10275 return V == (V & ((1LL << 8) - 1));
10278 if (!Subtarget->hasVFP2()) // FIXME: NEON?
10283 return V == (V & ((1LL << 8) - 1));
10287 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
10289 int Scale = AM.Scale;
10293 switch (VT.getSimpleVT().SimpleTy) {
10294 default: return false;
10302 Scale = Scale & ~1;
10303 return Scale == 2 || Scale == 4 || Scale == 8;
10306 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
10310 // Note, we allow "void" uses (basically, uses that aren't loads or
10311 // stores), because arm allows folding a scale into many arithmetic
10312 // operations. This should be made more precise and revisited later.
10314 // Allow r << imm, but the imm has to be a multiple of two.
10315 if (Scale & 1) return false;
10316 return isPowerOf2_32(Scale);
10320 /// isLegalAddressingMode - Return true if the addressing mode represented
10321 /// by AM is legal for this target, for a load/store of the specified type.
10322 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
10324 unsigned AS) const {
10325 EVT VT = getValueType(Ty, true);
10326 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
10329 // Can never fold addr of global into load/store.
10333 switch (AM.Scale) {
10334 case 0: // no scale reg, must be "r+i" or "r", or "i".
10337 if (Subtarget->isThumb1Only())
10341 // ARM doesn't support any R+R*scale+imm addr modes.
10345 if (!VT.isSimple())
10348 if (Subtarget->isThumb2())
10349 return isLegalT2ScaledAddressingMode(AM, VT);
10351 int Scale = AM.Scale;
10352 switch (VT.getSimpleVT().SimpleTy) {
10353 default: return false;
10357 if (Scale < 0) Scale = -Scale;
10361 return isPowerOf2_32(Scale & ~1);
10365 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
10370 // Note, we allow "void" uses (basically, uses that aren't loads or
10371 // stores), because arm allows folding a scale into many arithmetic
10372 // operations. This should be made more precise and revisited later.
10374 // Allow r << imm, but the imm has to be a multiple of two.
10375 if (Scale & 1) return false;
10376 return isPowerOf2_32(Scale);
10382 /// isLegalICmpImmediate - Return true if the specified immediate is legal
10383 /// icmp immediate, that is the target has icmp instructions which can compare
10384 /// a register against the immediate without having to materialize the
10385 /// immediate into a register.
10386 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
10387 // Thumb2 and ARM modes can use cmn for negative immediates.
10388 if (!Subtarget->isThumb())
10389 return ARM_AM::getSOImmVal(std::abs(Imm)) != -1;
10390 if (Subtarget->isThumb2())
10391 return ARM_AM::getT2SOImmVal(std::abs(Imm)) != -1;
10392 // Thumb1 doesn't have cmn, and only 8-bit immediates.
10393 return Imm >= 0 && Imm <= 255;
10396 /// isLegalAddImmediate - Return true if the specified immediate is a legal add
10397 /// *or sub* immediate, that is the target has add or sub instructions which can
10398 /// add a register with the immediate without having to materialize the
10399 /// immediate into a register.
10400 bool ARMTargetLowering::isLegalAddImmediate(int64_t Imm) const {
10401 // Same encoding for add/sub, just flip the sign.
10402 int64_t AbsImm = std::abs(Imm);
10403 if (!Subtarget->isThumb())
10404 return ARM_AM::getSOImmVal(AbsImm) != -1;
10405 if (Subtarget->isThumb2())
10406 return ARM_AM::getT2SOImmVal(AbsImm) != -1;
10407 // Thumb1 only has 8-bit unsigned immediate.
10408 return AbsImm >= 0 && AbsImm <= 255;
10411 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
10412 bool isSEXTLoad, SDValue &Base,
10413 SDValue &Offset, bool &isInc,
10414 SelectionDAG &DAG) {
10415 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
10418 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
10419 // AddressingMode 3
10420 Base = Ptr->getOperand(0);
10421 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
10422 int RHSC = (int)RHS->getZExtValue();
10423 if (RHSC < 0 && RHSC > -256) {
10424 assert(Ptr->getOpcode() == ISD::ADD);
10426 Offset = DAG.getConstant(-RHSC, SDLoc(Ptr), RHS->getValueType(0));
10430 isInc = (Ptr->getOpcode() == ISD::ADD);
10431 Offset = Ptr->getOperand(1);
10433 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
10434 // AddressingMode 2
10435 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
10436 int RHSC = (int)RHS->getZExtValue();
10437 if (RHSC < 0 && RHSC > -0x1000) {
10438 assert(Ptr->getOpcode() == ISD::ADD);
10440 Offset = DAG.getConstant(-RHSC, SDLoc(Ptr), RHS->getValueType(0));
10441 Base = Ptr->getOperand(0);
10446 if (Ptr->getOpcode() == ISD::ADD) {
10448 ARM_AM::ShiftOpc ShOpcVal=
10449 ARM_AM::getShiftOpcForNode(Ptr->getOperand(0).getOpcode());
10450 if (ShOpcVal != ARM_AM::no_shift) {
10451 Base = Ptr->getOperand(1);
10452 Offset = Ptr->getOperand(0);
10454 Base = Ptr->getOperand(0);
10455 Offset = Ptr->getOperand(1);
10460 isInc = (Ptr->getOpcode() == ISD::ADD);
10461 Base = Ptr->getOperand(0);
10462 Offset = Ptr->getOperand(1);
10466 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
10470 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
10471 bool isSEXTLoad, SDValue &Base,
10472 SDValue &Offset, bool &isInc,
10473 SelectionDAG &DAG) {
10474 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
10477 Base = Ptr->getOperand(0);
10478 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
10479 int RHSC = (int)RHS->getZExtValue();
10480 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
10481 assert(Ptr->getOpcode() == ISD::ADD);
10483 Offset = DAG.getConstant(-RHSC, SDLoc(Ptr), RHS->getValueType(0));
10485 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
10486 isInc = Ptr->getOpcode() == ISD::ADD;
10487 Offset = DAG.getConstant(RHSC, SDLoc(Ptr), RHS->getValueType(0));
10495 /// getPreIndexedAddressParts - returns true by value, base pointer and
10496 /// offset pointer and addressing mode by reference if the node's address
10497 /// can be legally represented as pre-indexed load / store address.
10499 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
10501 ISD::MemIndexedMode &AM,
10502 SelectionDAG &DAG) const {
10503 if (Subtarget->isThumb1Only())
10508 bool isSEXTLoad = false;
10509 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
10510 Ptr = LD->getBasePtr();
10511 VT = LD->getMemoryVT();
10512 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
10513 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
10514 Ptr = ST->getBasePtr();
10515 VT = ST->getMemoryVT();
10520 bool isLegal = false;
10521 if (Subtarget->isThumb2())
10522 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
10523 Offset, isInc, DAG);
10525 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
10526 Offset, isInc, DAG);
10530 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
10534 /// getPostIndexedAddressParts - returns true by value, base pointer and
10535 /// offset pointer and addressing mode by reference if this node can be
10536 /// combined with a load / store to form a post-indexed load / store.
10537 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
10540 ISD::MemIndexedMode &AM,
10541 SelectionDAG &DAG) const {
10542 if (Subtarget->isThumb1Only())
10547 bool isSEXTLoad = false;
10548 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
10549 VT = LD->getMemoryVT();
10550 Ptr = LD->getBasePtr();
10551 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
10552 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
10553 VT = ST->getMemoryVT();
10554 Ptr = ST->getBasePtr();
10559 bool isLegal = false;
10560 if (Subtarget->isThumb2())
10561 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
10564 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
10570 // Swap base ptr and offset to catch more post-index load / store when
10571 // it's legal. In Thumb2 mode, offset must be an immediate.
10572 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
10573 !Subtarget->isThumb2())
10574 std::swap(Base, Offset);
10576 // Post-indexed load / store update the base pointer.
10581 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
10585 void ARMTargetLowering::computeKnownBitsForTargetNode(const SDValue Op,
10588 const SelectionDAG &DAG,
10589 unsigned Depth) const {
10590 unsigned BitWidth = KnownOne.getBitWidth();
10591 KnownZero = KnownOne = APInt(BitWidth, 0);
10592 switch (Op.getOpcode()) {
10598 // These nodes' second result is a boolean
10599 if (Op.getResNo() == 0)
10601 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - 1);
10603 case ARMISD::CMOV: {
10604 // Bits are known zero/one if known on the LHS and RHS.
10605 DAG.computeKnownBits(Op.getOperand(0), KnownZero, KnownOne, Depth+1);
10606 if (KnownZero == 0 && KnownOne == 0) return;
10608 APInt KnownZeroRHS, KnownOneRHS;
10609 DAG.computeKnownBits(Op.getOperand(1), KnownZeroRHS, KnownOneRHS, Depth+1);
10610 KnownZero &= KnownZeroRHS;
10611 KnownOne &= KnownOneRHS;
10614 case ISD::INTRINSIC_W_CHAIN: {
10615 ConstantSDNode *CN = cast<ConstantSDNode>(Op->getOperand(1));
10616 Intrinsic::ID IntID = static_cast<Intrinsic::ID>(CN->getZExtValue());
10619 case Intrinsic::arm_ldaex:
10620 case Intrinsic::arm_ldrex: {
10621 EVT VT = cast<MemIntrinsicSDNode>(Op)->getMemoryVT();
10622 unsigned MemBits = VT.getScalarType().getSizeInBits();
10623 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - MemBits);
10631 //===----------------------------------------------------------------------===//
10632 // ARM Inline Assembly Support
10633 //===----------------------------------------------------------------------===//
10635 bool ARMTargetLowering::ExpandInlineAsm(CallInst *CI) const {
10636 // Looking for "rev" which is V6+.
10637 if (!Subtarget->hasV6Ops())
10640 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
10641 std::string AsmStr = IA->getAsmString();
10642 SmallVector<StringRef, 4> AsmPieces;
10643 SplitString(AsmStr, AsmPieces, ";\n");
10645 switch (AsmPieces.size()) {
10646 default: return false;
10648 AsmStr = AsmPieces[0];
10650 SplitString(AsmStr, AsmPieces, " \t,");
10653 if (AsmPieces.size() == 3 &&
10654 AsmPieces[0] == "rev" && AsmPieces[1] == "$0" && AsmPieces[2] == "$1" &&
10655 IA->getConstraintString().compare(0, 4, "=l,l") == 0) {
10656 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
10657 if (Ty && Ty->getBitWidth() == 32)
10658 return IntrinsicLowering::LowerToByteSwap(CI);
10666 /// getConstraintType - Given a constraint letter, return the type of
10667 /// constraint it is for this target.
10668 ARMTargetLowering::ConstraintType
10669 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
10670 if (Constraint.size() == 1) {
10671 switch (Constraint[0]) {
10673 case 'l': return C_RegisterClass;
10674 case 'w': return C_RegisterClass;
10675 case 'h': return C_RegisterClass;
10676 case 'x': return C_RegisterClass;
10677 case 't': return C_RegisterClass;
10678 case 'j': return C_Other; // Constant for movw.
10679 // An address with a single base register. Due to the way we
10680 // currently handle addresses it is the same as an 'r' memory constraint.
10681 case 'Q': return C_Memory;
10683 } else if (Constraint.size() == 2) {
10684 switch (Constraint[0]) {
10686 // All 'U+' constraints are addresses.
10687 case 'U': return C_Memory;
10690 return TargetLowering::getConstraintType(Constraint);
10693 /// Examine constraint type and operand type and determine a weight value.
10694 /// This object must already have been set up with the operand type
10695 /// and the current alternative constraint selected.
10696 TargetLowering::ConstraintWeight
10697 ARMTargetLowering::getSingleConstraintMatchWeight(
10698 AsmOperandInfo &info, const char *constraint) const {
10699 ConstraintWeight weight = CW_Invalid;
10700 Value *CallOperandVal = info.CallOperandVal;
10701 // If we don't have a value, we can't do a match,
10702 // but allow it at the lowest weight.
10703 if (!CallOperandVal)
10705 Type *type = CallOperandVal->getType();
10706 // Look at the constraint type.
10707 switch (*constraint) {
10709 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
10712 if (type->isIntegerTy()) {
10713 if (Subtarget->isThumb())
10714 weight = CW_SpecificReg;
10716 weight = CW_Register;
10720 if (type->isFloatingPointTy())
10721 weight = CW_Register;
10727 typedef std::pair<unsigned, const TargetRegisterClass*> RCPair;
10729 ARMTargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
10730 const std::string &Constraint,
10732 if (Constraint.size() == 1) {
10733 // GCC ARM Constraint Letters
10734 switch (Constraint[0]) {
10735 case 'l': // Low regs or general regs.
10736 if (Subtarget->isThumb())
10737 return RCPair(0U, &ARM::tGPRRegClass);
10738 return RCPair(0U, &ARM::GPRRegClass);
10739 case 'h': // High regs or no regs.
10740 if (Subtarget->isThumb())
10741 return RCPair(0U, &ARM::hGPRRegClass);
10744 if (Subtarget->isThumb1Only())
10745 return RCPair(0U, &ARM::tGPRRegClass);
10746 return RCPair(0U, &ARM::GPRRegClass);
10748 if (VT == MVT::Other)
10750 if (VT == MVT::f32)
10751 return RCPair(0U, &ARM::SPRRegClass);
10752 if (VT.getSizeInBits() == 64)
10753 return RCPair(0U, &ARM::DPRRegClass);
10754 if (VT.getSizeInBits() == 128)
10755 return RCPair(0U, &ARM::QPRRegClass);
10758 if (VT == MVT::Other)
10760 if (VT == MVT::f32)
10761 return RCPair(0U, &ARM::SPR_8RegClass);
10762 if (VT.getSizeInBits() == 64)
10763 return RCPair(0U, &ARM::DPR_8RegClass);
10764 if (VT.getSizeInBits() == 128)
10765 return RCPair(0U, &ARM::QPR_8RegClass);
10768 if (VT == MVT::f32)
10769 return RCPair(0U, &ARM::SPRRegClass);
10773 if (StringRef("{cc}").equals_lower(Constraint))
10774 return std::make_pair(unsigned(ARM::CPSR), &ARM::CCRRegClass);
10776 return TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
10779 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
10780 /// vector. If it is invalid, don't add anything to Ops.
10781 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
10782 std::string &Constraint,
10783 std::vector<SDValue>&Ops,
10784 SelectionDAG &DAG) const {
10787 // Currently only support length 1 constraints.
10788 if (Constraint.length() != 1) return;
10790 char ConstraintLetter = Constraint[0];
10791 switch (ConstraintLetter) {
10794 case 'I': case 'J': case 'K': case 'L':
10795 case 'M': case 'N': case 'O':
10796 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
10800 int64_t CVal64 = C->getSExtValue();
10801 int CVal = (int) CVal64;
10802 // None of these constraints allow values larger than 32 bits. Check
10803 // that the value fits in an int.
10804 if (CVal != CVal64)
10807 switch (ConstraintLetter) {
10809 // Constant suitable for movw, must be between 0 and
10811 if (Subtarget->hasV6T2Ops())
10812 if (CVal >= 0 && CVal <= 65535)
10816 if (Subtarget->isThumb1Only()) {
10817 // This must be a constant between 0 and 255, for ADD
10819 if (CVal >= 0 && CVal <= 255)
10821 } else if (Subtarget->isThumb2()) {
10822 // A constant that can be used as an immediate value in a
10823 // data-processing instruction.
10824 if (ARM_AM::getT2SOImmVal(CVal) != -1)
10827 // A constant that can be used as an immediate value in a
10828 // data-processing instruction.
10829 if (ARM_AM::getSOImmVal(CVal) != -1)
10835 if (Subtarget->isThumb()) { // FIXME thumb2
10836 // This must be a constant between -255 and -1, for negated ADD
10837 // immediates. This can be used in GCC with an "n" modifier that
10838 // prints the negated value, for use with SUB instructions. It is
10839 // not useful otherwise but is implemented for compatibility.
10840 if (CVal >= -255 && CVal <= -1)
10843 // This must be a constant between -4095 and 4095. It is not clear
10844 // what this constraint is intended for. Implemented for
10845 // compatibility with GCC.
10846 if (CVal >= -4095 && CVal <= 4095)
10852 if (Subtarget->isThumb1Only()) {
10853 // A 32-bit value where only one byte has a nonzero value. Exclude
10854 // zero to match GCC. This constraint is used by GCC internally for
10855 // constants that can be loaded with a move/shift combination.
10856 // It is not useful otherwise but is implemented for compatibility.
10857 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
10859 } else if (Subtarget->isThumb2()) {
10860 // A constant whose bitwise inverse can be used as an immediate
10861 // value in a data-processing instruction. This can be used in GCC
10862 // with a "B" modifier that prints the inverted value, for use with
10863 // BIC and MVN instructions. It is not useful otherwise but is
10864 // implemented for compatibility.
10865 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
10868 // A constant whose bitwise inverse can be used as an immediate
10869 // value in a data-processing instruction. This can be used in GCC
10870 // with a "B" modifier that prints the inverted value, for use with
10871 // BIC and MVN instructions. It is not useful otherwise but is
10872 // implemented for compatibility.
10873 if (ARM_AM::getSOImmVal(~CVal) != -1)
10879 if (Subtarget->isThumb1Only()) {
10880 // This must be a constant between -7 and 7,
10881 // for 3-operand ADD/SUB immediate instructions.
10882 if (CVal >= -7 && CVal < 7)
10884 } else if (Subtarget->isThumb2()) {
10885 // A constant whose negation can be used as an immediate value in a
10886 // data-processing instruction. This can be used in GCC with an "n"
10887 // modifier that prints the negated value, for use with SUB
10888 // instructions. It is not useful otherwise but is implemented for
10890 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
10893 // A constant whose negation can be used as an immediate value in a
10894 // data-processing instruction. This can be used in GCC with an "n"
10895 // modifier that prints the negated value, for use with SUB
10896 // instructions. It is not useful otherwise but is implemented for
10898 if (ARM_AM::getSOImmVal(-CVal) != -1)
10904 if (Subtarget->isThumb()) { // FIXME thumb2
10905 // This must be a multiple of 4 between 0 and 1020, for
10906 // ADD sp + immediate.
10907 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
10910 // A power of two or a constant between 0 and 32. This is used in
10911 // GCC for the shift amount on shifted register operands, but it is
10912 // useful in general for any shift amounts.
10913 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
10919 if (Subtarget->isThumb()) { // FIXME thumb2
10920 // This must be a constant between 0 and 31, for shift amounts.
10921 if (CVal >= 0 && CVal <= 31)
10927 if (Subtarget->isThumb()) { // FIXME thumb2
10928 // This must be a multiple of 4 between -508 and 508, for
10929 // ADD/SUB sp = sp + immediate.
10930 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
10935 Result = DAG.getTargetConstant(CVal, SDLoc(Op), Op.getValueType());
10939 if (Result.getNode()) {
10940 Ops.push_back(Result);
10943 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
10946 SDValue ARMTargetLowering::LowerDivRem(SDValue Op, SelectionDAG &DAG) const {
10947 assert(Subtarget->isTargetAEABI() && "Register-based DivRem lowering only");
10948 unsigned Opcode = Op->getOpcode();
10949 assert((Opcode == ISD::SDIVREM || Opcode == ISD::UDIVREM) &&
10950 "Invalid opcode for Div/Rem lowering");
10951 bool isSigned = (Opcode == ISD::SDIVREM);
10952 EVT VT = Op->getValueType(0);
10953 Type *Ty = VT.getTypeForEVT(*DAG.getContext());
10956 switch (VT.getSimpleVT().SimpleTy) {
10957 default: llvm_unreachable("Unexpected request for libcall!");
10958 case MVT::i8: LC = isSigned ? RTLIB::SDIVREM_I8 : RTLIB::UDIVREM_I8; break;
10959 case MVT::i16: LC = isSigned ? RTLIB::SDIVREM_I16 : RTLIB::UDIVREM_I16; break;
10960 case MVT::i32: LC = isSigned ? RTLIB::SDIVREM_I32 : RTLIB::UDIVREM_I32; break;
10961 case MVT::i64: LC = isSigned ? RTLIB::SDIVREM_I64 : RTLIB::UDIVREM_I64; break;
10964 SDValue InChain = DAG.getEntryNode();
10966 TargetLowering::ArgListTy Args;
10967 TargetLowering::ArgListEntry Entry;
10968 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i) {
10969 EVT ArgVT = Op->getOperand(i).getValueType();
10970 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
10971 Entry.Node = Op->getOperand(i);
10973 Entry.isSExt = isSigned;
10974 Entry.isZExt = !isSigned;
10975 Args.push_back(Entry);
10978 SDValue Callee = DAG.getExternalSymbol(getLibcallName(LC),
10981 Type *RetTy = (Type*)StructType::get(Ty, Ty, nullptr);
10984 TargetLowering::CallLoweringInfo CLI(DAG);
10985 CLI.setDebugLoc(dl).setChain(InChain)
10986 .setCallee(getLibcallCallingConv(LC), RetTy, Callee, std::move(Args), 0)
10987 .setInRegister().setSExtResult(isSigned).setZExtResult(!isSigned);
10989 std::pair<SDValue, SDValue> CallInfo = LowerCallTo(CLI);
10990 return CallInfo.first;
10994 ARMTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const {
10995 assert(Subtarget->isTargetWindows() && "unsupported target platform");
10999 SDValue Chain = Op.getOperand(0);
11000 SDValue Size = Op.getOperand(1);
11002 SDValue Words = DAG.getNode(ISD::SRL, DL, MVT::i32, Size,
11003 DAG.getConstant(2, DL, MVT::i32));
11006 Chain = DAG.getCopyToReg(Chain, DL, ARM::R4, Words, Flag);
11007 Flag = Chain.getValue(1);
11009 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
11010 Chain = DAG.getNode(ARMISD::WIN__CHKSTK, DL, NodeTys, Chain, Flag);
11012 SDValue NewSP = DAG.getCopyFromReg(Chain, DL, ARM::SP, MVT::i32);
11013 Chain = NewSP.getValue(1);
11015 SDValue Ops[2] = { NewSP, Chain };
11016 return DAG.getMergeValues(Ops, DL);
11019 SDValue ARMTargetLowering::LowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) const {
11020 assert(Op.getValueType() == MVT::f64 && Subtarget->isFPOnlySP() &&
11021 "Unexpected type for custom-lowering FP_EXTEND");
11024 LC = RTLIB::getFPEXT(Op.getOperand(0).getValueType(), Op.getValueType());
11026 SDValue SrcVal = Op.getOperand(0);
11027 return makeLibCall(DAG, LC, Op.getValueType(), &SrcVal, 1,
11028 /*isSigned*/ false, SDLoc(Op)).first;
11031 SDValue ARMTargetLowering::LowerFP_ROUND(SDValue Op, SelectionDAG &DAG) const {
11032 assert(Op.getOperand(0).getValueType() == MVT::f64 &&
11033 Subtarget->isFPOnlySP() &&
11034 "Unexpected type for custom-lowering FP_ROUND");
11037 LC = RTLIB::getFPROUND(Op.getOperand(0).getValueType(), Op.getValueType());
11039 SDValue SrcVal = Op.getOperand(0);
11040 return makeLibCall(DAG, LC, Op.getValueType(), &SrcVal, 1,
11041 /*isSigned*/ false, SDLoc(Op)).first;
11045 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
11046 // The ARM target isn't yet aware of offsets.
11050 bool ARM::isBitFieldInvertedMask(unsigned v) {
11051 if (v == 0xffffffff)
11054 // there can be 1's on either or both "outsides", all the "inside"
11055 // bits must be 0's
11056 return isShiftedMask_32(~v);
11059 /// isFPImmLegal - Returns true if the target can instruction select the
11060 /// specified FP immediate natively. If false, the legalizer will
11061 /// materialize the FP immediate as a load from a constant pool.
11062 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
11063 if (!Subtarget->hasVFP3())
11065 if (VT == MVT::f32)
11066 return ARM_AM::getFP32Imm(Imm) != -1;
11067 if (VT == MVT::f64 && !Subtarget->isFPOnlySP())
11068 return ARM_AM::getFP64Imm(Imm) != -1;
11072 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
11073 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
11074 /// specified in the intrinsic calls.
11075 bool ARMTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
11077 unsigned Intrinsic) const {
11078 switch (Intrinsic) {
11079 case Intrinsic::arm_neon_vld1:
11080 case Intrinsic::arm_neon_vld2:
11081 case Intrinsic::arm_neon_vld3:
11082 case Intrinsic::arm_neon_vld4:
11083 case Intrinsic::arm_neon_vld2lane:
11084 case Intrinsic::arm_neon_vld3lane:
11085 case Intrinsic::arm_neon_vld4lane: {
11086 Info.opc = ISD::INTRINSIC_W_CHAIN;
11087 // Conservatively set memVT to the entire set of vectors loaded.
11088 uint64_t NumElts = getDataLayout()->getTypeAllocSize(I.getType()) / 8;
11089 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
11090 Info.ptrVal = I.getArgOperand(0);
11092 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
11093 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
11094 Info.vol = false; // volatile loads with NEON intrinsics not supported
11095 Info.readMem = true;
11096 Info.writeMem = false;
11099 case Intrinsic::arm_neon_vst1:
11100 case Intrinsic::arm_neon_vst2:
11101 case Intrinsic::arm_neon_vst3:
11102 case Intrinsic::arm_neon_vst4:
11103 case Intrinsic::arm_neon_vst2lane:
11104 case Intrinsic::arm_neon_vst3lane:
11105 case Intrinsic::arm_neon_vst4lane: {
11106 Info.opc = ISD::INTRINSIC_VOID;
11107 // Conservatively set memVT to the entire set of vectors stored.
11108 unsigned NumElts = 0;
11109 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
11110 Type *ArgTy = I.getArgOperand(ArgI)->getType();
11111 if (!ArgTy->isVectorTy())
11113 NumElts += getDataLayout()->getTypeAllocSize(ArgTy) / 8;
11115 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
11116 Info.ptrVal = I.getArgOperand(0);
11118 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
11119 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
11120 Info.vol = false; // volatile stores with NEON intrinsics not supported
11121 Info.readMem = false;
11122 Info.writeMem = true;
11125 case Intrinsic::arm_ldaex:
11126 case Intrinsic::arm_ldrex: {
11127 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(0)->getType());
11128 Info.opc = ISD::INTRINSIC_W_CHAIN;
11129 Info.memVT = MVT::getVT(PtrTy->getElementType());
11130 Info.ptrVal = I.getArgOperand(0);
11132 Info.align = getDataLayout()->getABITypeAlignment(PtrTy->getElementType());
11134 Info.readMem = true;
11135 Info.writeMem = false;
11138 case Intrinsic::arm_stlex:
11139 case Intrinsic::arm_strex: {
11140 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(1)->getType());
11141 Info.opc = ISD::INTRINSIC_W_CHAIN;
11142 Info.memVT = MVT::getVT(PtrTy->getElementType());
11143 Info.ptrVal = I.getArgOperand(1);
11145 Info.align = getDataLayout()->getABITypeAlignment(PtrTy->getElementType());
11147 Info.readMem = false;
11148 Info.writeMem = true;
11151 case Intrinsic::arm_stlexd:
11152 case Intrinsic::arm_strexd: {
11153 Info.opc = ISD::INTRINSIC_W_CHAIN;
11154 Info.memVT = MVT::i64;
11155 Info.ptrVal = I.getArgOperand(2);
11159 Info.readMem = false;
11160 Info.writeMem = true;
11163 case Intrinsic::arm_ldaexd:
11164 case Intrinsic::arm_ldrexd: {
11165 Info.opc = ISD::INTRINSIC_W_CHAIN;
11166 Info.memVT = MVT::i64;
11167 Info.ptrVal = I.getArgOperand(0);
11171 Info.readMem = true;
11172 Info.writeMem = false;
11182 /// \brief Returns true if it is beneficial to convert a load of a constant
11183 /// to just the constant itself.
11184 bool ARMTargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
11186 assert(Ty->isIntegerTy());
11188 unsigned Bits = Ty->getPrimitiveSizeInBits();
11189 if (Bits == 0 || Bits > 32)
11194 bool ARMTargetLowering::hasLoadLinkedStoreConditional() const { return true; }
11196 Instruction* ARMTargetLowering::makeDMB(IRBuilder<> &Builder,
11197 ARM_MB::MemBOpt Domain) const {
11198 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
11200 // First, if the target has no DMB, see what fallback we can use.
11201 if (!Subtarget->hasDataBarrier()) {
11202 // Some ARMv6 cpus can support data barriers with an mcr instruction.
11203 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
11205 if (Subtarget->hasV6Ops() && !Subtarget->isThumb()) {
11206 Function *MCR = llvm::Intrinsic::getDeclaration(M, Intrinsic::arm_mcr);
11207 Value* args[6] = {Builder.getInt32(15), Builder.getInt32(0),
11208 Builder.getInt32(0), Builder.getInt32(7),
11209 Builder.getInt32(10), Builder.getInt32(5)};
11210 return Builder.CreateCall(MCR, args);
11212 // Instead of using barriers, atomic accesses on these subtargets use
11214 llvm_unreachable("makeDMB on a target so old that it has no barriers");
11217 Function *DMB = llvm::Intrinsic::getDeclaration(M, Intrinsic::arm_dmb);
11218 // Only a full system barrier exists in the M-class architectures.
11219 Domain = Subtarget->isMClass() ? ARM_MB::SY : Domain;
11220 Constant *CDomain = Builder.getInt32(Domain);
11221 return Builder.CreateCall(DMB, CDomain);
11225 // Based on http://www.cl.cam.ac.uk/~pes20/cpp/cpp0xmappings.html
11226 Instruction* ARMTargetLowering::emitLeadingFence(IRBuilder<> &Builder,
11227 AtomicOrdering Ord, bool IsStore,
11228 bool IsLoad) const {
11229 if (!getInsertFencesForAtomic())
11235 llvm_unreachable("Invalid fence: unordered/non-atomic");
11238 return nullptr; // Nothing to do
11239 case SequentiallyConsistent:
11241 return nullptr; // Nothing to do
11244 case AcquireRelease:
11245 if (Subtarget->isSwift())
11246 return makeDMB(Builder, ARM_MB::ISHST);
11247 // FIXME: add a comment with a link to documentation justifying this.
11249 return makeDMB(Builder, ARM_MB::ISH);
11251 llvm_unreachable("Unknown fence ordering in emitLeadingFence");
11254 Instruction* ARMTargetLowering::emitTrailingFence(IRBuilder<> &Builder,
11255 AtomicOrdering Ord, bool IsStore,
11256 bool IsLoad) const {
11257 if (!getInsertFencesForAtomic())
11263 llvm_unreachable("Invalid fence: unordered/not-atomic");
11266 return nullptr; // Nothing to do
11268 case AcquireRelease:
11269 case SequentiallyConsistent:
11270 return makeDMB(Builder, ARM_MB::ISH);
11272 llvm_unreachable("Unknown fence ordering in emitTrailingFence");
11275 // Loads and stores less than 64-bits are already atomic; ones above that
11276 // are doomed anyway, so defer to the default libcall and blame the OS when
11277 // things go wrong. Cortex M doesn't have ldrexd/strexd though, so don't emit
11278 // anything for those.
11279 bool ARMTargetLowering::shouldExpandAtomicStoreInIR(StoreInst *SI) const {
11280 unsigned Size = SI->getValueOperand()->getType()->getPrimitiveSizeInBits();
11281 return (Size == 64) && !Subtarget->isMClass();
11284 // Loads and stores less than 64-bits are already atomic; ones above that
11285 // are doomed anyway, so defer to the default libcall and blame the OS when
11286 // things go wrong. Cortex M doesn't have ldrexd/strexd though, so don't emit
11287 // anything for those.
11288 // FIXME: ldrd and strd are atomic if the CPU has LPAE (e.g. A15 has that
11289 // guarantee, see DDI0406C ARM architecture reference manual,
11290 // sections A8.8.72-74 LDRD)
11291 bool ARMTargetLowering::shouldExpandAtomicLoadInIR(LoadInst *LI) const {
11292 unsigned Size = LI->getType()->getPrimitiveSizeInBits();
11293 return (Size == 64) && !Subtarget->isMClass();
11296 // For the real atomic operations, we have ldrex/strex up to 32 bits,
11297 // and up to 64 bits on the non-M profiles
11298 TargetLoweringBase::AtomicRMWExpansionKind
11299 ARMTargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
11300 unsigned Size = AI->getType()->getPrimitiveSizeInBits();
11301 return (Size <= (Subtarget->isMClass() ? 32U : 64U))
11302 ? AtomicRMWExpansionKind::LLSC
11303 : AtomicRMWExpansionKind::None;
11306 // This has so far only been implemented for MachO.
11307 bool ARMTargetLowering::useLoadStackGuardNode() const {
11308 return Subtarget->isTargetMachO();
11311 bool ARMTargetLowering::canCombineStoreAndExtract(Type *VectorTy, Value *Idx,
11312 unsigned &Cost) const {
11313 // If we do not have NEON, vector types are not natively supported.
11314 if (!Subtarget->hasNEON())
11317 // Floating point values and vector values map to the same register file.
11318 // Therefore, althought we could do a store extract of a vector type, this is
11319 // better to leave at float as we have more freedom in the addressing mode for
11321 if (VectorTy->isFPOrFPVectorTy())
11324 // If the index is unknown at compile time, this is very expensive to lower
11325 // and it is not possible to combine the store with the extract.
11326 if (!isa<ConstantInt>(Idx))
11329 assert(VectorTy->isVectorTy() && "VectorTy is not a vector type");
11330 unsigned BitWidth = cast<VectorType>(VectorTy)->getBitWidth();
11331 // We can do a store + vector extract on any vector that fits perfectly in a D
11333 if (BitWidth == 64 || BitWidth == 128) {
11340 Value *ARMTargetLowering::emitLoadLinked(IRBuilder<> &Builder, Value *Addr,
11341 AtomicOrdering Ord) const {
11342 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
11343 Type *ValTy = cast<PointerType>(Addr->getType())->getElementType();
11344 bool IsAcquire = isAtLeastAcquire(Ord);
11346 // Since i64 isn't legal and intrinsics don't get type-lowered, the ldrexd
11347 // intrinsic must return {i32, i32} and we have to recombine them into a
11348 // single i64 here.
11349 if (ValTy->getPrimitiveSizeInBits() == 64) {
11350 Intrinsic::ID Int =
11351 IsAcquire ? Intrinsic::arm_ldaexd : Intrinsic::arm_ldrexd;
11352 Function *Ldrex = llvm::Intrinsic::getDeclaration(M, Int);
11354 Addr = Builder.CreateBitCast(Addr, Type::getInt8PtrTy(M->getContext()));
11355 Value *LoHi = Builder.CreateCall(Ldrex, Addr, "lohi");
11357 Value *Lo = Builder.CreateExtractValue(LoHi, 0, "lo");
11358 Value *Hi = Builder.CreateExtractValue(LoHi, 1, "hi");
11359 if (!Subtarget->isLittle())
11360 std::swap (Lo, Hi);
11361 Lo = Builder.CreateZExt(Lo, ValTy, "lo64");
11362 Hi = Builder.CreateZExt(Hi, ValTy, "hi64");
11363 return Builder.CreateOr(
11364 Lo, Builder.CreateShl(Hi, ConstantInt::get(ValTy, 32)), "val64");
11367 Type *Tys[] = { Addr->getType() };
11368 Intrinsic::ID Int = IsAcquire ? Intrinsic::arm_ldaex : Intrinsic::arm_ldrex;
11369 Function *Ldrex = llvm::Intrinsic::getDeclaration(M, Int, Tys);
11371 return Builder.CreateTruncOrBitCast(
11372 Builder.CreateCall(Ldrex, Addr),
11373 cast<PointerType>(Addr->getType())->getElementType());
11376 Value *ARMTargetLowering::emitStoreConditional(IRBuilder<> &Builder, Value *Val,
11378 AtomicOrdering Ord) const {
11379 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
11380 bool IsRelease = isAtLeastRelease(Ord);
11382 // Since the intrinsics must have legal type, the i64 intrinsics take two
11383 // parameters: "i32, i32". We must marshal Val into the appropriate form
11384 // before the call.
11385 if (Val->getType()->getPrimitiveSizeInBits() == 64) {
11386 Intrinsic::ID Int =
11387 IsRelease ? Intrinsic::arm_stlexd : Intrinsic::arm_strexd;
11388 Function *Strex = Intrinsic::getDeclaration(M, Int);
11389 Type *Int32Ty = Type::getInt32Ty(M->getContext());
11391 Value *Lo = Builder.CreateTrunc(Val, Int32Ty, "lo");
11392 Value *Hi = Builder.CreateTrunc(Builder.CreateLShr(Val, 32), Int32Ty, "hi");
11393 if (!Subtarget->isLittle())
11394 std::swap (Lo, Hi);
11395 Addr = Builder.CreateBitCast(Addr, Type::getInt8PtrTy(M->getContext()));
11396 return Builder.CreateCall(Strex, {Lo, Hi, Addr});
11399 Intrinsic::ID Int = IsRelease ? Intrinsic::arm_stlex : Intrinsic::arm_strex;
11400 Type *Tys[] = { Addr->getType() };
11401 Function *Strex = Intrinsic::getDeclaration(M, Int, Tys);
11403 return Builder.CreateCall(
11404 Strex, {Builder.CreateZExtOrBitCast(
11405 Val, Strex->getFunctionType()->getParamType(0)),
11417 static bool isHomogeneousAggregate(Type *Ty, HABaseType &Base,
11418 uint64_t &Members) {
11419 if (const StructType *ST = dyn_cast<StructType>(Ty)) {
11420 for (unsigned i = 0; i < ST->getNumElements(); ++i) {
11421 uint64_t SubMembers = 0;
11422 if (!isHomogeneousAggregate(ST->getElementType(i), Base, SubMembers))
11424 Members += SubMembers;
11426 } else if (const ArrayType *AT = dyn_cast<ArrayType>(Ty)) {
11427 uint64_t SubMembers = 0;
11428 if (!isHomogeneousAggregate(AT->getElementType(), Base, SubMembers))
11430 Members += SubMembers * AT->getNumElements();
11431 } else if (Ty->isFloatTy()) {
11432 if (Base != HA_UNKNOWN && Base != HA_FLOAT)
11436 } else if (Ty->isDoubleTy()) {
11437 if (Base != HA_UNKNOWN && Base != HA_DOUBLE)
11441 } else if (const VectorType *VT = dyn_cast<VectorType>(Ty)) {
11448 return VT->getBitWidth() == 64;
11450 return VT->getBitWidth() == 128;
11452 switch (VT->getBitWidth()) {
11465 return (Members > 0 && Members <= 4);
11468 /// \brief Return true if a type is an AAPCS-VFP homogeneous aggregate or one of
11469 /// [N x i32] or [N x i64]. This allows front-ends to skip emitting padding when
11470 /// passing according to AAPCS rules.
11471 bool ARMTargetLowering::functionArgumentNeedsConsecutiveRegisters(
11472 Type *Ty, CallingConv::ID CallConv, bool isVarArg) const {
11473 if (getEffectiveCallingConv(CallConv, isVarArg) !=
11474 CallingConv::ARM_AAPCS_VFP)
11477 HABaseType Base = HA_UNKNOWN;
11478 uint64_t Members = 0;
11479 bool IsHA = isHomogeneousAggregate(Ty, Base, Members);
11480 DEBUG(dbgs() << "isHA: " << IsHA << " "; Ty->dump());
11482 bool IsIntArray = Ty->isArrayTy() && Ty->getArrayElementType()->isIntegerTy();
11483 return IsHA || IsIntArray;