1 //===-- ARMISelLowering.cpp - ARM DAG Lowering Implementation -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the interfaces that ARM uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
15 #include "ARMISelLowering.h"
16 #include "ARMCallingConv.h"
17 #include "ARMConstantPoolValue.h"
18 #include "ARMMachineFunctionInfo.h"
19 #include "ARMPerfectShuffle.h"
20 #include "ARMSubtarget.h"
21 #include "ARMTargetMachine.h"
22 #include "ARMTargetObjectFile.h"
23 #include "MCTargetDesc/ARMAddressingModes.h"
24 #include "llvm/ADT/Statistic.h"
25 #include "llvm/ADT/StringExtras.h"
26 #include "llvm/ADT/StringSwitch.h"
27 #include "llvm/CodeGen/CallingConvLower.h"
28 #include "llvm/CodeGen/IntrinsicLowering.h"
29 #include "llvm/CodeGen/MachineBasicBlock.h"
30 #include "llvm/CodeGen/MachineFrameInfo.h"
31 #include "llvm/CodeGen/MachineFunction.h"
32 #include "llvm/CodeGen/MachineInstrBuilder.h"
33 #include "llvm/CodeGen/MachineJumpTableInfo.h"
34 #include "llvm/CodeGen/MachineModuleInfo.h"
35 #include "llvm/CodeGen/MachineRegisterInfo.h"
36 #include "llvm/CodeGen/SelectionDAG.h"
37 #include "llvm/IR/CallingConv.h"
38 #include "llvm/IR/Constants.h"
39 #include "llvm/IR/Function.h"
40 #include "llvm/IR/GlobalValue.h"
41 #include "llvm/IR/IRBuilder.h"
42 #include "llvm/IR/Instruction.h"
43 #include "llvm/IR/Instructions.h"
44 #include "llvm/IR/IntrinsicInst.h"
45 #include "llvm/IR/Intrinsics.h"
46 #include "llvm/IR/Type.h"
47 #include "llvm/MC/MCSectionMachO.h"
48 #include "llvm/Support/CommandLine.h"
49 #include "llvm/Support/Debug.h"
50 #include "llvm/Support/ErrorHandling.h"
51 #include "llvm/Support/MathExtras.h"
52 #include "llvm/Support/raw_ostream.h"
53 #include "llvm/Target/TargetOptions.h"
57 #define DEBUG_TYPE "arm-isel"
59 STATISTIC(NumTailCalls, "Number of tail calls");
60 STATISTIC(NumMovwMovt, "Number of GAs materialized with movw + movt");
61 STATISTIC(NumLoopByVals, "Number of loops generated for byval arguments");
64 EnableARMLongCalls("arm-long-calls", cl::Hidden,
65 cl::desc("Generate calls via indirect call instructions"),
69 ARMInterworking("arm-interworking", cl::Hidden,
70 cl::desc("Enable / disable ARM interworking (for debugging only)"),
74 class ARMCCState : public CCState {
76 ARMCCState(CallingConv::ID CC, bool isVarArg, MachineFunction &MF,
77 SmallVectorImpl<CCValAssign> &locs, LLVMContext &C,
79 : CCState(CC, isVarArg, MF, locs, C) {
80 assert(((PC == Call) || (PC == Prologue)) &&
81 "ARMCCState users must specify whether their context is call"
82 "or prologue generation.");
88 // The APCS parameter registers.
89 static const MCPhysReg GPRArgRegs[] = {
90 ARM::R0, ARM::R1, ARM::R2, ARM::R3
93 void ARMTargetLowering::addTypeForNEON(MVT VT, MVT PromotedLdStVT,
94 MVT PromotedBitwiseVT) {
95 if (VT != PromotedLdStVT) {
96 setOperationAction(ISD::LOAD, VT, Promote);
97 AddPromotedToType (ISD::LOAD, VT, PromotedLdStVT);
99 setOperationAction(ISD::STORE, VT, Promote);
100 AddPromotedToType (ISD::STORE, VT, PromotedLdStVT);
103 MVT ElemTy = VT.getVectorElementType();
104 if (ElemTy != MVT::i64 && ElemTy != MVT::f64)
105 setOperationAction(ISD::SETCC, VT, Custom);
106 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
107 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
108 if (ElemTy == MVT::i32) {
109 setOperationAction(ISD::SINT_TO_FP, VT, Custom);
110 setOperationAction(ISD::UINT_TO_FP, VT, Custom);
111 setOperationAction(ISD::FP_TO_SINT, VT, Custom);
112 setOperationAction(ISD::FP_TO_UINT, VT, Custom);
114 setOperationAction(ISD::SINT_TO_FP, VT, Expand);
115 setOperationAction(ISD::UINT_TO_FP, VT, Expand);
116 setOperationAction(ISD::FP_TO_SINT, VT, Expand);
117 setOperationAction(ISD::FP_TO_UINT, VT, Expand);
119 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
120 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
121 setOperationAction(ISD::CONCAT_VECTORS, VT, Legal);
122 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Legal);
123 setOperationAction(ISD::SELECT, VT, Expand);
124 setOperationAction(ISD::SELECT_CC, VT, Expand);
125 setOperationAction(ISD::VSELECT, VT, Expand);
126 setOperationAction(ISD::SIGN_EXTEND_INREG, VT, Expand);
127 if (VT.isInteger()) {
128 setOperationAction(ISD::SHL, VT, Custom);
129 setOperationAction(ISD::SRA, VT, Custom);
130 setOperationAction(ISD::SRL, VT, Custom);
133 // Promote all bit-wise operations.
134 if (VT.isInteger() && VT != PromotedBitwiseVT) {
135 setOperationAction(ISD::AND, VT, Promote);
136 AddPromotedToType (ISD::AND, VT, PromotedBitwiseVT);
137 setOperationAction(ISD::OR, VT, Promote);
138 AddPromotedToType (ISD::OR, VT, PromotedBitwiseVT);
139 setOperationAction(ISD::XOR, VT, Promote);
140 AddPromotedToType (ISD::XOR, VT, PromotedBitwiseVT);
143 // Neon does not support vector divide/remainder operations.
144 setOperationAction(ISD::SDIV, VT, Expand);
145 setOperationAction(ISD::UDIV, VT, Expand);
146 setOperationAction(ISD::FDIV, VT, Expand);
147 setOperationAction(ISD::SREM, VT, Expand);
148 setOperationAction(ISD::UREM, VT, Expand);
149 setOperationAction(ISD::FREM, VT, Expand);
152 void ARMTargetLowering::addDRTypeForNEON(MVT VT) {
153 addRegisterClass(VT, &ARM::DPRRegClass);
154 addTypeForNEON(VT, MVT::f64, MVT::v2i32);
157 void ARMTargetLowering::addQRTypeForNEON(MVT VT) {
158 addRegisterClass(VT, &ARM::DPairRegClass);
159 addTypeForNEON(VT, MVT::v2f64, MVT::v4i32);
162 ARMTargetLowering::ARMTargetLowering(const TargetMachine &TM,
163 const ARMSubtarget &STI)
164 : TargetLowering(TM), Subtarget(&STI) {
165 RegInfo = Subtarget->getRegisterInfo();
166 Itins = Subtarget->getInstrItineraryData();
168 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
170 if (Subtarget->isTargetMachO()) {
171 // Uses VFP for Thumb libfuncs if available.
172 if (Subtarget->isThumb() && Subtarget->hasVFP2() &&
173 Subtarget->hasARMOps() && !Subtarget->useSoftFloat()) {
174 // Single-precision floating-point arithmetic.
175 setLibcallName(RTLIB::ADD_F32, "__addsf3vfp");
176 setLibcallName(RTLIB::SUB_F32, "__subsf3vfp");
177 setLibcallName(RTLIB::MUL_F32, "__mulsf3vfp");
178 setLibcallName(RTLIB::DIV_F32, "__divsf3vfp");
180 // Double-precision floating-point arithmetic.
181 setLibcallName(RTLIB::ADD_F64, "__adddf3vfp");
182 setLibcallName(RTLIB::SUB_F64, "__subdf3vfp");
183 setLibcallName(RTLIB::MUL_F64, "__muldf3vfp");
184 setLibcallName(RTLIB::DIV_F64, "__divdf3vfp");
186 // Single-precision comparisons.
187 setLibcallName(RTLIB::OEQ_F32, "__eqsf2vfp");
188 setLibcallName(RTLIB::UNE_F32, "__nesf2vfp");
189 setLibcallName(RTLIB::OLT_F32, "__ltsf2vfp");
190 setLibcallName(RTLIB::OLE_F32, "__lesf2vfp");
191 setLibcallName(RTLIB::OGE_F32, "__gesf2vfp");
192 setLibcallName(RTLIB::OGT_F32, "__gtsf2vfp");
193 setLibcallName(RTLIB::UO_F32, "__unordsf2vfp");
194 setLibcallName(RTLIB::O_F32, "__unordsf2vfp");
196 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
197 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETNE);
198 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
199 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
200 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
201 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
202 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
203 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
205 // Double-precision comparisons.
206 setLibcallName(RTLIB::OEQ_F64, "__eqdf2vfp");
207 setLibcallName(RTLIB::UNE_F64, "__nedf2vfp");
208 setLibcallName(RTLIB::OLT_F64, "__ltdf2vfp");
209 setLibcallName(RTLIB::OLE_F64, "__ledf2vfp");
210 setLibcallName(RTLIB::OGE_F64, "__gedf2vfp");
211 setLibcallName(RTLIB::OGT_F64, "__gtdf2vfp");
212 setLibcallName(RTLIB::UO_F64, "__unorddf2vfp");
213 setLibcallName(RTLIB::O_F64, "__unorddf2vfp");
215 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
216 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETNE);
217 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
218 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
219 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
220 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
221 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
222 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
224 // Floating-point to integer conversions.
225 // i64 conversions are done via library routines even when generating VFP
226 // instructions, so use the same ones.
227 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__fixdfsivfp");
228 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__fixunsdfsivfp");
229 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__fixsfsivfp");
230 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__fixunssfsivfp");
232 // Conversions between floating types.
233 setLibcallName(RTLIB::FPROUND_F64_F32, "__truncdfsf2vfp");
234 setLibcallName(RTLIB::FPEXT_F32_F64, "__extendsfdf2vfp");
236 // Integer to floating-point conversions.
237 // i64 conversions are done via library routines even when generating VFP
238 // instructions, so use the same ones.
239 // FIXME: There appears to be some naming inconsistency in ARM libgcc:
240 // e.g., __floatunsidf vs. __floatunssidfvfp.
241 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__floatsidfvfp");
242 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__floatunssidfvfp");
243 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__floatsisfvfp");
244 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__floatunssisfvfp");
248 // These libcalls are not available in 32-bit.
249 setLibcallName(RTLIB::SHL_I128, nullptr);
250 setLibcallName(RTLIB::SRL_I128, nullptr);
251 setLibcallName(RTLIB::SRA_I128, nullptr);
253 if (Subtarget->isAAPCS_ABI() && !Subtarget->isTargetMachO() &&
254 !Subtarget->isTargetWindows()) {
255 static const struct {
256 const RTLIB::Libcall Op;
257 const char * const Name;
258 const CallingConv::ID CC;
259 const ISD::CondCode Cond;
261 // Double-precision floating-point arithmetic helper functions
262 // RTABI chapter 4.1.2, Table 2
263 { RTLIB::ADD_F64, "__aeabi_dadd", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
264 { RTLIB::DIV_F64, "__aeabi_ddiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
265 { RTLIB::MUL_F64, "__aeabi_dmul", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
266 { RTLIB::SUB_F64, "__aeabi_dsub", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
268 // Double-precision floating-point comparison helper functions
269 // RTABI chapter 4.1.2, Table 3
270 { RTLIB::OEQ_F64, "__aeabi_dcmpeq", CallingConv::ARM_AAPCS, ISD::SETNE },
271 { RTLIB::UNE_F64, "__aeabi_dcmpeq", CallingConv::ARM_AAPCS, ISD::SETEQ },
272 { RTLIB::OLT_F64, "__aeabi_dcmplt", CallingConv::ARM_AAPCS, ISD::SETNE },
273 { RTLIB::OLE_F64, "__aeabi_dcmple", CallingConv::ARM_AAPCS, ISD::SETNE },
274 { RTLIB::OGE_F64, "__aeabi_dcmpge", CallingConv::ARM_AAPCS, ISD::SETNE },
275 { RTLIB::OGT_F64, "__aeabi_dcmpgt", CallingConv::ARM_AAPCS, ISD::SETNE },
276 { RTLIB::UO_F64, "__aeabi_dcmpun", CallingConv::ARM_AAPCS, ISD::SETNE },
277 { RTLIB::O_F64, "__aeabi_dcmpun", CallingConv::ARM_AAPCS, ISD::SETEQ },
279 // Single-precision floating-point arithmetic helper functions
280 // RTABI chapter 4.1.2, Table 4
281 { RTLIB::ADD_F32, "__aeabi_fadd", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
282 { RTLIB::DIV_F32, "__aeabi_fdiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
283 { RTLIB::MUL_F32, "__aeabi_fmul", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
284 { RTLIB::SUB_F32, "__aeabi_fsub", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
286 // Single-precision floating-point comparison helper functions
287 // RTABI chapter 4.1.2, Table 5
288 { RTLIB::OEQ_F32, "__aeabi_fcmpeq", CallingConv::ARM_AAPCS, ISD::SETNE },
289 { RTLIB::UNE_F32, "__aeabi_fcmpeq", CallingConv::ARM_AAPCS, ISD::SETEQ },
290 { RTLIB::OLT_F32, "__aeabi_fcmplt", CallingConv::ARM_AAPCS, ISD::SETNE },
291 { RTLIB::OLE_F32, "__aeabi_fcmple", CallingConv::ARM_AAPCS, ISD::SETNE },
292 { RTLIB::OGE_F32, "__aeabi_fcmpge", CallingConv::ARM_AAPCS, ISD::SETNE },
293 { RTLIB::OGT_F32, "__aeabi_fcmpgt", CallingConv::ARM_AAPCS, ISD::SETNE },
294 { RTLIB::UO_F32, "__aeabi_fcmpun", CallingConv::ARM_AAPCS, ISD::SETNE },
295 { RTLIB::O_F32, "__aeabi_fcmpun", CallingConv::ARM_AAPCS, ISD::SETEQ },
297 // Floating-point to integer conversions.
298 // RTABI chapter 4.1.2, Table 6
299 { RTLIB::FPTOSINT_F64_I32, "__aeabi_d2iz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
300 { RTLIB::FPTOUINT_F64_I32, "__aeabi_d2uiz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
301 { RTLIB::FPTOSINT_F64_I64, "__aeabi_d2lz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
302 { RTLIB::FPTOUINT_F64_I64, "__aeabi_d2ulz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
303 { RTLIB::FPTOSINT_F32_I32, "__aeabi_f2iz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
304 { RTLIB::FPTOUINT_F32_I32, "__aeabi_f2uiz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
305 { RTLIB::FPTOSINT_F32_I64, "__aeabi_f2lz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
306 { RTLIB::FPTOUINT_F32_I64, "__aeabi_f2ulz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
308 // Conversions between floating types.
309 // RTABI chapter 4.1.2, Table 7
310 { RTLIB::FPROUND_F64_F32, "__aeabi_d2f", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
311 { RTLIB::FPROUND_F64_F16, "__aeabi_d2h", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
312 { RTLIB::FPEXT_F32_F64, "__aeabi_f2d", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
314 // Integer to floating-point conversions.
315 // RTABI chapter 4.1.2, Table 8
316 { RTLIB::SINTTOFP_I32_F64, "__aeabi_i2d", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
317 { RTLIB::UINTTOFP_I32_F64, "__aeabi_ui2d", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
318 { RTLIB::SINTTOFP_I64_F64, "__aeabi_l2d", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
319 { RTLIB::UINTTOFP_I64_F64, "__aeabi_ul2d", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
320 { RTLIB::SINTTOFP_I32_F32, "__aeabi_i2f", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
321 { RTLIB::UINTTOFP_I32_F32, "__aeabi_ui2f", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
322 { RTLIB::SINTTOFP_I64_F32, "__aeabi_l2f", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
323 { RTLIB::UINTTOFP_I64_F32, "__aeabi_ul2f", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
325 // Long long helper functions
326 // RTABI chapter 4.2, Table 9
327 { RTLIB::MUL_I64, "__aeabi_lmul", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
328 { RTLIB::SHL_I64, "__aeabi_llsl", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
329 { RTLIB::SRL_I64, "__aeabi_llsr", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
330 { RTLIB::SRA_I64, "__aeabi_lasr", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
332 // Integer division functions
333 // RTABI chapter 4.3.1
334 { RTLIB::SDIV_I8, "__aeabi_idiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
335 { RTLIB::SDIV_I16, "__aeabi_idiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
336 { RTLIB::SDIV_I32, "__aeabi_idiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
337 { RTLIB::SDIV_I64, "__aeabi_ldivmod", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
338 { RTLIB::UDIV_I8, "__aeabi_uidiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
339 { RTLIB::UDIV_I16, "__aeabi_uidiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
340 { RTLIB::UDIV_I32, "__aeabi_uidiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
341 { RTLIB::UDIV_I64, "__aeabi_uldivmod", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
344 // RTABI chapter 4.3.4
345 { RTLIB::MEMCPY, "__aeabi_memcpy", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
346 { RTLIB::MEMMOVE, "__aeabi_memmove", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
347 { RTLIB::MEMSET, "__aeabi_memset", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
350 for (const auto &LC : LibraryCalls) {
351 setLibcallName(LC.Op, LC.Name);
352 setLibcallCallingConv(LC.Op, LC.CC);
353 if (LC.Cond != ISD::SETCC_INVALID)
354 setCmpLibcallCC(LC.Op, LC.Cond);
358 if (Subtarget->isTargetWindows()) {
359 static const struct {
360 const RTLIB::Libcall Op;
361 const char * const Name;
362 const CallingConv::ID CC;
364 { RTLIB::FPTOSINT_F32_I64, "__stoi64", CallingConv::ARM_AAPCS_VFP },
365 { RTLIB::FPTOSINT_F64_I64, "__dtoi64", CallingConv::ARM_AAPCS_VFP },
366 { RTLIB::FPTOUINT_F32_I64, "__stou64", CallingConv::ARM_AAPCS_VFP },
367 { RTLIB::FPTOUINT_F64_I64, "__dtou64", CallingConv::ARM_AAPCS_VFP },
368 { RTLIB::SINTTOFP_I64_F32, "__i64tos", CallingConv::ARM_AAPCS_VFP },
369 { RTLIB::SINTTOFP_I64_F64, "__i64tod", CallingConv::ARM_AAPCS_VFP },
370 { RTLIB::UINTTOFP_I64_F32, "__u64tos", CallingConv::ARM_AAPCS_VFP },
371 { RTLIB::UINTTOFP_I64_F64, "__u64tod", CallingConv::ARM_AAPCS_VFP },
374 for (const auto &LC : LibraryCalls) {
375 setLibcallName(LC.Op, LC.Name);
376 setLibcallCallingConv(LC.Op, LC.CC);
380 // Use divmod compiler-rt calls for iOS 5.0 and later.
381 if (Subtarget->getTargetTriple().isiOS() &&
382 !Subtarget->getTargetTriple().isOSVersionLT(5, 0)) {
383 setLibcallName(RTLIB::SDIVREM_I32, "__divmodsi4");
384 setLibcallName(RTLIB::UDIVREM_I32, "__udivmodsi4");
387 // The half <-> float conversion functions are always soft-float, but are
388 // needed for some targets which use a hard-float calling convention by
390 if (Subtarget->isAAPCS_ABI()) {
391 setLibcallCallingConv(RTLIB::FPROUND_F32_F16, CallingConv::ARM_AAPCS);
392 setLibcallCallingConv(RTLIB::FPROUND_F64_F16, CallingConv::ARM_AAPCS);
393 setLibcallCallingConv(RTLIB::FPEXT_F16_F32, CallingConv::ARM_AAPCS);
395 setLibcallCallingConv(RTLIB::FPROUND_F32_F16, CallingConv::ARM_APCS);
396 setLibcallCallingConv(RTLIB::FPROUND_F64_F16, CallingConv::ARM_APCS);
397 setLibcallCallingConv(RTLIB::FPEXT_F16_F32, CallingConv::ARM_APCS);
400 if (Subtarget->isThumb1Only())
401 addRegisterClass(MVT::i32, &ARM::tGPRRegClass);
403 addRegisterClass(MVT::i32, &ARM::GPRRegClass);
404 if (!Subtarget->useSoftFloat() && Subtarget->hasVFP2() &&
405 !Subtarget->isThumb1Only()) {
406 addRegisterClass(MVT::f32, &ARM::SPRRegClass);
407 addRegisterClass(MVT::f64, &ARM::DPRRegClass);
410 for (MVT VT : MVT::vector_valuetypes()) {
411 for (MVT InnerVT : MVT::vector_valuetypes()) {
412 setTruncStoreAction(VT, InnerVT, Expand);
413 setLoadExtAction(ISD::SEXTLOAD, VT, InnerVT, Expand);
414 setLoadExtAction(ISD::ZEXTLOAD, VT, InnerVT, Expand);
415 setLoadExtAction(ISD::EXTLOAD, VT, InnerVT, Expand);
418 setOperationAction(ISD::MULHS, VT, Expand);
419 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
420 setOperationAction(ISD::MULHU, VT, Expand);
421 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
423 setOperationAction(ISD::BSWAP, VT, Expand);
426 setOperationAction(ISD::ConstantFP, MVT::f32, Custom);
427 setOperationAction(ISD::ConstantFP, MVT::f64, Custom);
429 if (Subtarget->hasNEON()) {
430 addDRTypeForNEON(MVT::v2f32);
431 addDRTypeForNEON(MVT::v8i8);
432 addDRTypeForNEON(MVT::v4i16);
433 addDRTypeForNEON(MVT::v2i32);
434 addDRTypeForNEON(MVT::v1i64);
436 addQRTypeForNEON(MVT::v4f32);
437 addQRTypeForNEON(MVT::v2f64);
438 addQRTypeForNEON(MVT::v16i8);
439 addQRTypeForNEON(MVT::v8i16);
440 addQRTypeForNEON(MVT::v4i32);
441 addQRTypeForNEON(MVT::v2i64);
443 // v2f64 is legal so that QR subregs can be extracted as f64 elements, but
444 // neither Neon nor VFP support any arithmetic operations on it.
445 // The same with v4f32. But keep in mind that vadd, vsub, vmul are natively
446 // supported for v4f32.
447 setOperationAction(ISD::FADD, MVT::v2f64, Expand);
448 setOperationAction(ISD::FSUB, MVT::v2f64, Expand);
449 setOperationAction(ISD::FMUL, MVT::v2f64, Expand);
450 // FIXME: Code duplication: FDIV and FREM are expanded always, see
451 // ARMTargetLowering::addTypeForNEON method for details.
452 setOperationAction(ISD::FDIV, MVT::v2f64, Expand);
453 setOperationAction(ISD::FREM, MVT::v2f64, Expand);
454 // FIXME: Create unittest.
455 // In another words, find a way when "copysign" appears in DAG with vector
457 setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand);
458 // FIXME: Code duplication: SETCC has custom operation action, see
459 // ARMTargetLowering::addTypeForNEON method for details.
460 setOperationAction(ISD::SETCC, MVT::v2f64, Expand);
461 // FIXME: Create unittest for FNEG and for FABS.
462 setOperationAction(ISD::FNEG, MVT::v2f64, Expand);
463 setOperationAction(ISD::FABS, MVT::v2f64, Expand);
464 setOperationAction(ISD::FSQRT, MVT::v2f64, Expand);
465 setOperationAction(ISD::FSIN, MVT::v2f64, Expand);
466 setOperationAction(ISD::FCOS, MVT::v2f64, Expand);
467 setOperationAction(ISD::FPOWI, MVT::v2f64, Expand);
468 setOperationAction(ISD::FPOW, MVT::v2f64, Expand);
469 setOperationAction(ISD::FLOG, MVT::v2f64, Expand);
470 setOperationAction(ISD::FLOG2, MVT::v2f64, Expand);
471 setOperationAction(ISD::FLOG10, MVT::v2f64, Expand);
472 setOperationAction(ISD::FEXP, MVT::v2f64, Expand);
473 setOperationAction(ISD::FEXP2, MVT::v2f64, Expand);
474 // FIXME: Create unittest for FCEIL, FTRUNC, FRINT, FNEARBYINT, FFLOOR.
475 setOperationAction(ISD::FCEIL, MVT::v2f64, Expand);
476 setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand);
477 setOperationAction(ISD::FRINT, MVT::v2f64, Expand);
478 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand);
479 setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand);
480 setOperationAction(ISD::FMA, MVT::v2f64, Expand);
482 setOperationAction(ISD::FSQRT, MVT::v4f32, Expand);
483 setOperationAction(ISD::FSIN, MVT::v4f32, Expand);
484 setOperationAction(ISD::FCOS, MVT::v4f32, Expand);
485 setOperationAction(ISD::FPOWI, MVT::v4f32, Expand);
486 setOperationAction(ISD::FPOW, MVT::v4f32, Expand);
487 setOperationAction(ISD::FLOG, MVT::v4f32, Expand);
488 setOperationAction(ISD::FLOG2, MVT::v4f32, Expand);
489 setOperationAction(ISD::FLOG10, MVT::v4f32, Expand);
490 setOperationAction(ISD::FEXP, MVT::v4f32, Expand);
491 setOperationAction(ISD::FEXP2, MVT::v4f32, Expand);
492 setOperationAction(ISD::FCEIL, MVT::v4f32, Expand);
493 setOperationAction(ISD::FTRUNC, MVT::v4f32, Expand);
494 setOperationAction(ISD::FRINT, MVT::v4f32, Expand);
495 setOperationAction(ISD::FNEARBYINT, MVT::v4f32, Expand);
496 setOperationAction(ISD::FFLOOR, MVT::v4f32, Expand);
498 // Mark v2f32 intrinsics.
499 setOperationAction(ISD::FSQRT, MVT::v2f32, Expand);
500 setOperationAction(ISD::FSIN, MVT::v2f32, Expand);
501 setOperationAction(ISD::FCOS, MVT::v2f32, Expand);
502 setOperationAction(ISD::FPOWI, MVT::v2f32, Expand);
503 setOperationAction(ISD::FPOW, MVT::v2f32, Expand);
504 setOperationAction(ISD::FLOG, MVT::v2f32, Expand);
505 setOperationAction(ISD::FLOG2, MVT::v2f32, Expand);
506 setOperationAction(ISD::FLOG10, MVT::v2f32, Expand);
507 setOperationAction(ISD::FEXP, MVT::v2f32, Expand);
508 setOperationAction(ISD::FEXP2, MVT::v2f32, Expand);
509 setOperationAction(ISD::FCEIL, MVT::v2f32, Expand);
510 setOperationAction(ISD::FTRUNC, MVT::v2f32, Expand);
511 setOperationAction(ISD::FRINT, MVT::v2f32, Expand);
512 setOperationAction(ISD::FNEARBYINT, MVT::v2f32, Expand);
513 setOperationAction(ISD::FFLOOR, MVT::v2f32, Expand);
515 // Neon does not support some operations on v1i64 and v2i64 types.
516 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
517 // Custom handling for some quad-vector types to detect VMULL.
518 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
519 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
520 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
521 // Custom handling for some vector types to avoid expensive expansions
522 setOperationAction(ISD::SDIV, MVT::v4i16, Custom);
523 setOperationAction(ISD::SDIV, MVT::v8i8, Custom);
524 setOperationAction(ISD::UDIV, MVT::v4i16, Custom);
525 setOperationAction(ISD::UDIV, MVT::v8i8, Custom);
526 setOperationAction(ISD::SETCC, MVT::v1i64, Expand);
527 setOperationAction(ISD::SETCC, MVT::v2i64, Expand);
528 // Neon does not have single instruction SINT_TO_FP and UINT_TO_FP with
529 // a destination type that is wider than the source, and nor does
530 // it have a FP_TO_[SU]INT instruction with a narrower destination than
532 setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Custom);
533 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
534 setOperationAction(ISD::FP_TO_UINT, MVT::v4i16, Custom);
535 setOperationAction(ISD::FP_TO_SINT, MVT::v4i16, Custom);
537 setOperationAction(ISD::FP_ROUND, MVT::v2f32, Expand);
538 setOperationAction(ISD::FP_EXTEND, MVT::v2f64, Expand);
540 // NEON does not have single instruction CTPOP for vectors with element
541 // types wider than 8-bits. However, custom lowering can leverage the
542 // v8i8/v16i8 vcnt instruction.
543 setOperationAction(ISD::CTPOP, MVT::v2i32, Custom);
544 setOperationAction(ISD::CTPOP, MVT::v4i32, Custom);
545 setOperationAction(ISD::CTPOP, MVT::v4i16, Custom);
546 setOperationAction(ISD::CTPOP, MVT::v8i16, Custom);
548 // NEON only has FMA instructions as of VFP4.
549 if (!Subtarget->hasVFP4()) {
550 setOperationAction(ISD::FMA, MVT::v2f32, Expand);
551 setOperationAction(ISD::FMA, MVT::v4f32, Expand);
554 setTargetDAGCombine(ISD::INTRINSIC_VOID);
555 setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN);
556 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
557 setTargetDAGCombine(ISD::SHL);
558 setTargetDAGCombine(ISD::SRL);
559 setTargetDAGCombine(ISD::SRA);
560 setTargetDAGCombine(ISD::SIGN_EXTEND);
561 setTargetDAGCombine(ISD::ZERO_EXTEND);
562 setTargetDAGCombine(ISD::ANY_EXTEND);
563 setTargetDAGCombine(ISD::SELECT_CC);
564 setTargetDAGCombine(ISD::BUILD_VECTOR);
565 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
566 setTargetDAGCombine(ISD::INSERT_VECTOR_ELT);
567 setTargetDAGCombine(ISD::STORE);
568 setTargetDAGCombine(ISD::FP_TO_SINT);
569 setTargetDAGCombine(ISD::FP_TO_UINT);
570 setTargetDAGCombine(ISD::FDIV);
571 setTargetDAGCombine(ISD::LOAD);
573 // It is legal to extload from v4i8 to v4i16 or v4i32.
574 for (MVT Ty : {MVT::v8i8, MVT::v4i8, MVT::v2i8, MVT::v4i16, MVT::v2i16,
576 for (MVT VT : MVT::integer_vector_valuetypes()) {
577 setLoadExtAction(ISD::EXTLOAD, VT, Ty, Legal);
578 setLoadExtAction(ISD::ZEXTLOAD, VT, Ty, Legal);
579 setLoadExtAction(ISD::SEXTLOAD, VT, Ty, Legal);
584 // ARM and Thumb2 support UMLAL/SMLAL.
585 if (!Subtarget->isThumb1Only())
586 setTargetDAGCombine(ISD::ADDC);
588 if (Subtarget->isFPOnlySP()) {
589 // When targetting a floating-point unit with only single-precision
590 // operations, f64 is legal for the few double-precision instructions which
591 // are present However, no double-precision operations other than moves,
592 // loads and stores are provided by the hardware.
593 setOperationAction(ISD::FADD, MVT::f64, Expand);
594 setOperationAction(ISD::FSUB, MVT::f64, Expand);
595 setOperationAction(ISD::FMUL, MVT::f64, Expand);
596 setOperationAction(ISD::FMA, MVT::f64, Expand);
597 setOperationAction(ISD::FDIV, MVT::f64, Expand);
598 setOperationAction(ISD::FREM, MVT::f64, Expand);
599 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
600 setOperationAction(ISD::FGETSIGN, MVT::f64, Expand);
601 setOperationAction(ISD::FNEG, MVT::f64, Expand);
602 setOperationAction(ISD::FABS, MVT::f64, Expand);
603 setOperationAction(ISD::FSQRT, MVT::f64, Expand);
604 setOperationAction(ISD::FSIN, MVT::f64, Expand);
605 setOperationAction(ISD::FCOS, MVT::f64, Expand);
606 setOperationAction(ISD::FPOWI, MVT::f64, Expand);
607 setOperationAction(ISD::FPOW, MVT::f64, Expand);
608 setOperationAction(ISD::FLOG, MVT::f64, Expand);
609 setOperationAction(ISD::FLOG2, MVT::f64, Expand);
610 setOperationAction(ISD::FLOG10, MVT::f64, Expand);
611 setOperationAction(ISD::FEXP, MVT::f64, Expand);
612 setOperationAction(ISD::FEXP2, MVT::f64, Expand);
613 setOperationAction(ISD::FCEIL, MVT::f64, Expand);
614 setOperationAction(ISD::FTRUNC, MVT::f64, Expand);
615 setOperationAction(ISD::FRINT, MVT::f64, Expand);
616 setOperationAction(ISD::FNEARBYINT, MVT::f64, Expand);
617 setOperationAction(ISD::FFLOOR, MVT::f64, Expand);
618 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
619 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
620 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
621 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
622 setOperationAction(ISD::FP_TO_SINT, MVT::f64, Custom);
623 setOperationAction(ISD::FP_TO_UINT, MVT::f64, Custom);
624 setOperationAction(ISD::FP_ROUND, MVT::f32, Custom);
625 setOperationAction(ISD::FP_EXTEND, MVT::f64, Custom);
628 computeRegisterProperties(Subtarget->getRegisterInfo());
630 // ARM does not have floating-point extending loads.
631 for (MVT VT : MVT::fp_valuetypes()) {
632 setLoadExtAction(ISD::EXTLOAD, VT, MVT::f32, Expand);
633 setLoadExtAction(ISD::EXTLOAD, VT, MVT::f16, Expand);
636 // ... or truncating stores
637 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
638 setTruncStoreAction(MVT::f32, MVT::f16, Expand);
639 setTruncStoreAction(MVT::f64, MVT::f16, Expand);
641 // ARM does not have i1 sign extending load.
642 for (MVT VT : MVT::integer_valuetypes())
643 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Promote);
645 // ARM supports all 4 flavors of integer indexed load / store.
646 if (!Subtarget->isThumb1Only()) {
647 for (unsigned im = (unsigned)ISD::PRE_INC;
648 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
649 setIndexedLoadAction(im, MVT::i1, Legal);
650 setIndexedLoadAction(im, MVT::i8, Legal);
651 setIndexedLoadAction(im, MVT::i16, Legal);
652 setIndexedLoadAction(im, MVT::i32, Legal);
653 setIndexedStoreAction(im, MVT::i1, Legal);
654 setIndexedStoreAction(im, MVT::i8, Legal);
655 setIndexedStoreAction(im, MVT::i16, Legal);
656 setIndexedStoreAction(im, MVT::i32, Legal);
660 setOperationAction(ISD::SADDO, MVT::i32, Custom);
661 setOperationAction(ISD::UADDO, MVT::i32, Custom);
662 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
663 setOperationAction(ISD::USUBO, MVT::i32, Custom);
665 // i64 operation support.
666 setOperationAction(ISD::MUL, MVT::i64, Expand);
667 setOperationAction(ISD::MULHU, MVT::i32, Expand);
668 if (Subtarget->isThumb1Only()) {
669 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
670 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
672 if (Subtarget->isThumb1Only() || !Subtarget->hasV6Ops()
673 || (Subtarget->isThumb2() && !Subtarget->hasThumb2DSP()))
674 setOperationAction(ISD::MULHS, MVT::i32, Expand);
676 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
677 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
678 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
679 setOperationAction(ISD::SRL, MVT::i64, Custom);
680 setOperationAction(ISD::SRA, MVT::i64, Custom);
682 if (!Subtarget->isThumb1Only()) {
683 // FIXME: We should do this for Thumb1 as well.
684 setOperationAction(ISD::ADDC, MVT::i32, Custom);
685 setOperationAction(ISD::ADDE, MVT::i32, Custom);
686 setOperationAction(ISD::SUBC, MVT::i32, Custom);
687 setOperationAction(ISD::SUBE, MVT::i32, Custom);
690 // ARM does not have ROTL.
691 setOperationAction(ISD::ROTL, MVT::i32, Expand);
692 setOperationAction(ISD::CTTZ, MVT::i32, Custom);
693 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
694 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
695 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
697 // These just redirect to CTTZ and CTLZ on ARM.
698 setOperationAction(ISD::CTTZ_ZERO_UNDEF , MVT::i32 , Expand);
699 setOperationAction(ISD::CTLZ_ZERO_UNDEF , MVT::i32 , Expand);
701 setOperationAction(ISD::READCYCLECOUNTER, MVT::i64, Custom);
703 // Only ARMv6 has BSWAP.
704 if (!Subtarget->hasV6Ops())
705 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
707 if (!(Subtarget->hasDivide() && Subtarget->isThumb2()) &&
708 !(Subtarget->hasDivideInARMMode() && !Subtarget->isThumb())) {
709 // These are expanded into libcalls if the cpu doesn't have HW divider.
710 setOperationAction(ISD::SDIV, MVT::i32, Expand);
711 setOperationAction(ISD::UDIV, MVT::i32, Expand);
714 // FIXME: Also set divmod for SREM on EABI
715 setOperationAction(ISD::SREM, MVT::i32, Expand);
716 setOperationAction(ISD::UREM, MVT::i32, Expand);
717 // Register based DivRem for AEABI (RTABI 4.2)
718 if (Subtarget->isTargetAEABI()) {
719 setLibcallName(RTLIB::SDIVREM_I8, "__aeabi_idivmod");
720 setLibcallName(RTLIB::SDIVREM_I16, "__aeabi_idivmod");
721 setLibcallName(RTLIB::SDIVREM_I32, "__aeabi_idivmod");
722 setLibcallName(RTLIB::SDIVREM_I64, "__aeabi_ldivmod");
723 setLibcallName(RTLIB::UDIVREM_I8, "__aeabi_uidivmod");
724 setLibcallName(RTLIB::UDIVREM_I16, "__aeabi_uidivmod");
725 setLibcallName(RTLIB::UDIVREM_I32, "__aeabi_uidivmod");
726 setLibcallName(RTLIB::UDIVREM_I64, "__aeabi_uldivmod");
728 setLibcallCallingConv(RTLIB::SDIVREM_I8, CallingConv::ARM_AAPCS);
729 setLibcallCallingConv(RTLIB::SDIVREM_I16, CallingConv::ARM_AAPCS);
730 setLibcallCallingConv(RTLIB::SDIVREM_I32, CallingConv::ARM_AAPCS);
731 setLibcallCallingConv(RTLIB::SDIVREM_I64, CallingConv::ARM_AAPCS);
732 setLibcallCallingConv(RTLIB::UDIVREM_I8, CallingConv::ARM_AAPCS);
733 setLibcallCallingConv(RTLIB::UDIVREM_I16, CallingConv::ARM_AAPCS);
734 setLibcallCallingConv(RTLIB::UDIVREM_I32, CallingConv::ARM_AAPCS);
735 setLibcallCallingConv(RTLIB::UDIVREM_I64, CallingConv::ARM_AAPCS);
737 setOperationAction(ISD::SDIVREM, MVT::i32, Custom);
738 setOperationAction(ISD::UDIVREM, MVT::i32, Custom);
740 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
741 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
744 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
745 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
746 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
747 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
748 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
750 setOperationAction(ISD::TRAP, MVT::Other, Legal);
752 // Use the default implementation.
753 setOperationAction(ISD::VASTART, MVT::Other, Custom);
754 setOperationAction(ISD::VAARG, MVT::Other, Expand);
755 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
756 setOperationAction(ISD::VAEND, MVT::Other, Expand);
757 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
758 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
760 if (!Subtarget->isTargetMachO()) {
761 // Non-MachO platforms may return values in these registers via the
762 // personality function.
763 setExceptionPointerRegister(ARM::R0);
764 setExceptionSelectorRegister(ARM::R1);
767 if (Subtarget->getTargetTriple().isWindowsItaniumEnvironment())
768 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
770 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
772 // ARMv6 Thumb1 (except for CPUs that support dmb / dsb) and earlier use
773 // the default expansion. If we are targeting a single threaded system,
774 // then set them all for expand so we can lower them later into their
776 if (TM.Options.ThreadModel == ThreadModel::Single)
777 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Expand);
778 else if (Subtarget->hasAnyDataBarrier() && !Subtarget->isThumb1Only()) {
779 // ATOMIC_FENCE needs custom lowering; the others should have been expanded
780 // to ldrex/strex loops already.
781 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Custom);
783 // On v8, we have particularly efficient implementations of atomic fences
784 // if they can be combined with nearby atomic loads and stores.
785 if (!Subtarget->hasV8Ops()) {
786 // Automatically insert fences (dmb ish) around ATOMIC_SWAP etc.
787 setInsertFencesForAtomic(true);
790 // If there's anything we can use as a barrier, go through custom lowering
792 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other,
793 Subtarget->hasAnyDataBarrier() ? Custom : Expand);
795 // Set them all for expansion, which will force libcalls.
796 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Expand);
797 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Expand);
798 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Expand);
799 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Expand);
800 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Expand);
801 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Expand);
802 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Expand);
803 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Expand);
804 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i32, Expand);
805 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i32, Expand);
806 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i32, Expand);
807 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i32, Expand);
808 // Mark ATOMIC_LOAD and ATOMIC_STORE custom so we can handle the
809 // Unordered/Monotonic case.
810 setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Custom);
811 setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Custom);
814 setOperationAction(ISD::PREFETCH, MVT::Other, Custom);
816 // Requires SXTB/SXTH, available on v6 and up in both ARM and Thumb modes.
817 if (!Subtarget->hasV6Ops()) {
818 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
819 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
821 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
823 if (!Subtarget->useSoftFloat() && Subtarget->hasVFP2() &&
824 !Subtarget->isThumb1Only()) {
825 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
826 // iff target supports vfp2.
827 setOperationAction(ISD::BITCAST, MVT::i64, Custom);
828 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
831 // We want to custom lower some of our intrinsics.
832 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
833 if (Subtarget->isTargetDarwin()) {
834 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
835 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
836 setLibcallName(RTLIB::UNWIND_RESUME, "_Unwind_SjLj_Resume");
839 setOperationAction(ISD::SETCC, MVT::i32, Expand);
840 setOperationAction(ISD::SETCC, MVT::f32, Expand);
841 setOperationAction(ISD::SETCC, MVT::f64, Expand);
842 setOperationAction(ISD::SELECT, MVT::i32, Custom);
843 setOperationAction(ISD::SELECT, MVT::f32, Custom);
844 setOperationAction(ISD::SELECT, MVT::f64, Custom);
845 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
846 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
847 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
849 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
850 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
851 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
852 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
853 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
855 // We don't support sin/cos/fmod/copysign/pow
856 setOperationAction(ISD::FSIN, MVT::f64, Expand);
857 setOperationAction(ISD::FSIN, MVT::f32, Expand);
858 setOperationAction(ISD::FCOS, MVT::f32, Expand);
859 setOperationAction(ISD::FCOS, MVT::f64, Expand);
860 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
861 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
862 setOperationAction(ISD::FREM, MVT::f64, Expand);
863 setOperationAction(ISD::FREM, MVT::f32, Expand);
864 if (!Subtarget->useSoftFloat() && Subtarget->hasVFP2() &&
865 !Subtarget->isThumb1Only()) {
866 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
867 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
869 setOperationAction(ISD::FPOW, MVT::f64, Expand);
870 setOperationAction(ISD::FPOW, MVT::f32, Expand);
872 if (!Subtarget->hasVFP4()) {
873 setOperationAction(ISD::FMA, MVT::f64, Expand);
874 setOperationAction(ISD::FMA, MVT::f32, Expand);
877 // Various VFP goodness
878 if (!Subtarget->useSoftFloat() && !Subtarget->isThumb1Only()) {
879 // FP-ARMv8 adds f64 <-> f16 conversion. Before that it should be expanded.
880 if (!Subtarget->hasFPARMv8() || Subtarget->isFPOnlySP()) {
881 setOperationAction(ISD::FP16_TO_FP, MVT::f64, Expand);
882 setOperationAction(ISD::FP_TO_FP16, MVT::f64, Expand);
885 // fp16 is a special v7 extension that adds f16 <-> f32 conversions.
886 if (!Subtarget->hasFP16()) {
887 setOperationAction(ISD::FP16_TO_FP, MVT::f32, Expand);
888 setOperationAction(ISD::FP_TO_FP16, MVT::f32, Expand);
892 // Combine sin / cos into one node or libcall if possible.
893 if (Subtarget->hasSinCos()) {
894 setLibcallName(RTLIB::SINCOS_F32, "sincosf");
895 setLibcallName(RTLIB::SINCOS_F64, "sincos");
896 if (Subtarget->getTargetTriple().isiOS()) {
897 // For iOS, we don't want to the normal expansion of a libcall to
898 // sincos. We want to issue a libcall to __sincos_stret.
899 setOperationAction(ISD::FSINCOS, MVT::f64, Custom);
900 setOperationAction(ISD::FSINCOS, MVT::f32, Custom);
904 // FP-ARMv8 implements a lot of rounding-like FP operations.
905 if (Subtarget->hasFPARMv8()) {
906 setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
907 setOperationAction(ISD::FCEIL, MVT::f32, Legal);
908 setOperationAction(ISD::FROUND, MVT::f32, Legal);
909 setOperationAction(ISD::FTRUNC, MVT::f32, Legal);
910 setOperationAction(ISD::FNEARBYINT, MVT::f32, Legal);
911 setOperationAction(ISD::FRINT, MVT::f32, Legal);
912 if (!Subtarget->isFPOnlySP()) {
913 setOperationAction(ISD::FFLOOR, MVT::f64, Legal);
914 setOperationAction(ISD::FCEIL, MVT::f64, Legal);
915 setOperationAction(ISD::FROUND, MVT::f64, Legal);
916 setOperationAction(ISD::FTRUNC, MVT::f64, Legal);
917 setOperationAction(ISD::FNEARBYINT, MVT::f64, Legal);
918 setOperationAction(ISD::FRINT, MVT::f64, Legal);
921 // We have target-specific dag combine patterns for the following nodes:
922 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
923 setTargetDAGCombine(ISD::ADD);
924 setTargetDAGCombine(ISD::SUB);
925 setTargetDAGCombine(ISD::MUL);
926 setTargetDAGCombine(ISD::AND);
927 setTargetDAGCombine(ISD::OR);
928 setTargetDAGCombine(ISD::XOR);
930 if (Subtarget->hasV6Ops())
931 setTargetDAGCombine(ISD::SRL);
933 setStackPointerRegisterToSaveRestore(ARM::SP);
935 if (Subtarget->useSoftFloat() || Subtarget->isThumb1Only() ||
936 !Subtarget->hasVFP2())
937 setSchedulingPreference(Sched::RegPressure);
939 setSchedulingPreference(Sched::Hybrid);
941 //// temporary - rewrite interface to use type
942 MaxStoresPerMemset = 8;
943 MaxStoresPerMemsetOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
944 MaxStoresPerMemcpy = 4; // For @llvm.memcpy -> sequence of stores
945 MaxStoresPerMemcpyOptSize = Subtarget->isTargetDarwin() ? 4 : 2;
946 MaxStoresPerMemmove = 4; // For @llvm.memmove -> sequence of stores
947 MaxStoresPerMemmoveOptSize = Subtarget->isTargetDarwin() ? 4 : 2;
949 // On ARM arguments smaller than 4 bytes are extended, so all arguments
950 // are at least 4 bytes aligned.
951 setMinStackArgumentAlignment(4);
953 // Prefer likely predicted branches to selects on out-of-order cores.
954 PredictableSelectIsExpensive = Subtarget->isLikeA9();
956 setMinFunctionAlignment(Subtarget->isThumb() ? 1 : 2);
959 bool ARMTargetLowering::useSoftFloat() const {
960 return Subtarget->useSoftFloat();
963 // FIXME: It might make sense to define the representative register class as the
964 // nearest super-register that has a non-null superset. For example, DPR_VFP2 is
965 // a super-register of SPR, and DPR is a superset if DPR_VFP2. Consequently,
966 // SPR's representative would be DPR_VFP2. This should work well if register
967 // pressure tracking were modified such that a register use would increment the
968 // pressure of the register class's representative and all of it's super
969 // classes' representatives transitively. We have not implemented this because
970 // of the difficulty prior to coalescing of modeling operand register classes
971 // due to the common occurrence of cross class copies and subregister insertions
973 std::pair<const TargetRegisterClass *, uint8_t>
974 ARMTargetLowering::findRepresentativeClass(const TargetRegisterInfo *TRI,
976 const TargetRegisterClass *RRC = nullptr;
978 switch (VT.SimpleTy) {
980 return TargetLowering::findRepresentativeClass(TRI, VT);
981 // Use DPR as representative register class for all floating point
982 // and vector types. Since there are 32 SPR registers and 32 DPR registers so
983 // the cost is 1 for both f32 and f64.
984 case MVT::f32: case MVT::f64: case MVT::v8i8: case MVT::v4i16:
985 case MVT::v2i32: case MVT::v1i64: case MVT::v2f32:
986 RRC = &ARM::DPRRegClass;
987 // When NEON is used for SP, only half of the register file is available
988 // because operations that define both SP and DP results will be constrained
989 // to the VFP2 class (D0-D15). We currently model this constraint prior to
990 // coalescing by double-counting the SP regs. See the FIXME above.
991 if (Subtarget->useNEONForSinglePrecisionFP())
994 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
995 case MVT::v4f32: case MVT::v2f64:
996 RRC = &ARM::DPRRegClass;
1000 RRC = &ARM::DPRRegClass;
1004 RRC = &ARM::DPRRegClass;
1008 return std::make_pair(RRC, Cost);
1011 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
1012 switch ((ARMISD::NodeType)Opcode) {
1013 case ARMISD::FIRST_NUMBER: break;
1014 case ARMISD::Wrapper: return "ARMISD::Wrapper";
1015 case ARMISD::WrapperPIC: return "ARMISD::WrapperPIC";
1016 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
1017 case ARMISD::COPY_STRUCT_BYVAL: return "ARMISD::COPY_STRUCT_BYVAL";
1018 case ARMISD::CALL: return "ARMISD::CALL";
1019 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
1020 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
1021 case ARMISD::tCALL: return "ARMISD::tCALL";
1022 case ARMISD::BRCOND: return "ARMISD::BRCOND";
1023 case ARMISD::BR_JT: return "ARMISD::BR_JT";
1024 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
1025 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
1026 case ARMISD::INTRET_FLAG: return "ARMISD::INTRET_FLAG";
1027 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
1028 case ARMISD::CMP: return "ARMISD::CMP";
1029 case ARMISD::CMN: return "ARMISD::CMN";
1030 case ARMISD::CMPZ: return "ARMISD::CMPZ";
1031 case ARMISD::CMPFP: return "ARMISD::CMPFP";
1032 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
1033 case ARMISD::BCC_i64: return "ARMISD::BCC_i64";
1034 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
1036 case ARMISD::CMOV: return "ARMISD::CMOV";
1038 case ARMISD::RBIT: return "ARMISD::RBIT";
1040 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
1041 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
1042 case ARMISD::RRX: return "ARMISD::RRX";
1044 case ARMISD::ADDC: return "ARMISD::ADDC";
1045 case ARMISD::ADDE: return "ARMISD::ADDE";
1046 case ARMISD::SUBC: return "ARMISD::SUBC";
1047 case ARMISD::SUBE: return "ARMISD::SUBE";
1049 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
1050 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
1052 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
1053 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
1055 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN";
1057 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
1059 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
1061 case ARMISD::MEMBARRIER_MCR: return "ARMISD::MEMBARRIER_MCR";
1063 case ARMISD::PRELOAD: return "ARMISD::PRELOAD";
1065 case ARMISD::WIN__CHKSTK: return "ARMISD:::WIN__CHKSTK";
1067 case ARMISD::VCEQ: return "ARMISD::VCEQ";
1068 case ARMISD::VCEQZ: return "ARMISD::VCEQZ";
1069 case ARMISD::VCGE: return "ARMISD::VCGE";
1070 case ARMISD::VCGEZ: return "ARMISD::VCGEZ";
1071 case ARMISD::VCLEZ: return "ARMISD::VCLEZ";
1072 case ARMISD::VCGEU: return "ARMISD::VCGEU";
1073 case ARMISD::VCGT: return "ARMISD::VCGT";
1074 case ARMISD::VCGTZ: return "ARMISD::VCGTZ";
1075 case ARMISD::VCLTZ: return "ARMISD::VCLTZ";
1076 case ARMISD::VCGTU: return "ARMISD::VCGTU";
1077 case ARMISD::VTST: return "ARMISD::VTST";
1079 case ARMISD::VSHL: return "ARMISD::VSHL";
1080 case ARMISD::VSHRs: return "ARMISD::VSHRs";
1081 case ARMISD::VSHRu: return "ARMISD::VSHRu";
1082 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
1083 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
1084 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
1085 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
1086 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
1087 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
1088 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
1089 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
1090 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
1091 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
1092 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
1093 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
1094 case ARMISD::VSLI: return "ARMISD::VSLI";
1095 case ARMISD::VSRI: return "ARMISD::VSRI";
1096 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
1097 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
1098 case ARMISD::VMOVIMM: return "ARMISD::VMOVIMM";
1099 case ARMISD::VMVNIMM: return "ARMISD::VMVNIMM";
1100 case ARMISD::VMOVFPIMM: return "ARMISD::VMOVFPIMM";
1101 case ARMISD::VDUP: return "ARMISD::VDUP";
1102 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
1103 case ARMISD::VEXT: return "ARMISD::VEXT";
1104 case ARMISD::VREV64: return "ARMISD::VREV64";
1105 case ARMISD::VREV32: return "ARMISD::VREV32";
1106 case ARMISD::VREV16: return "ARMISD::VREV16";
1107 case ARMISD::VZIP: return "ARMISD::VZIP";
1108 case ARMISD::VUZP: return "ARMISD::VUZP";
1109 case ARMISD::VTRN: return "ARMISD::VTRN";
1110 case ARMISD::VTBL1: return "ARMISD::VTBL1";
1111 case ARMISD::VTBL2: return "ARMISD::VTBL2";
1112 case ARMISD::VMULLs: return "ARMISD::VMULLs";
1113 case ARMISD::VMULLu: return "ARMISD::VMULLu";
1114 case ARMISD::UMLAL: return "ARMISD::UMLAL";
1115 case ARMISD::SMLAL: return "ARMISD::SMLAL";
1116 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
1117 case ARMISD::FMAX: return "ARMISD::FMAX";
1118 case ARMISD::FMIN: return "ARMISD::FMIN";
1119 case ARMISD::VMAXNM: return "ARMISD::VMAX";
1120 case ARMISD::VMINNM: return "ARMISD::VMIN";
1121 case ARMISD::BFI: return "ARMISD::BFI";
1122 case ARMISD::VORRIMM: return "ARMISD::VORRIMM";
1123 case ARMISD::VBICIMM: return "ARMISD::VBICIMM";
1124 case ARMISD::VBSL: return "ARMISD::VBSL";
1125 case ARMISD::VLD2DUP: return "ARMISD::VLD2DUP";
1126 case ARMISD::VLD3DUP: return "ARMISD::VLD3DUP";
1127 case ARMISD::VLD4DUP: return "ARMISD::VLD4DUP";
1128 case ARMISD::VLD1_UPD: return "ARMISD::VLD1_UPD";
1129 case ARMISD::VLD2_UPD: return "ARMISD::VLD2_UPD";
1130 case ARMISD::VLD3_UPD: return "ARMISD::VLD3_UPD";
1131 case ARMISD::VLD4_UPD: return "ARMISD::VLD4_UPD";
1132 case ARMISD::VLD2LN_UPD: return "ARMISD::VLD2LN_UPD";
1133 case ARMISD::VLD3LN_UPD: return "ARMISD::VLD3LN_UPD";
1134 case ARMISD::VLD4LN_UPD: return "ARMISD::VLD4LN_UPD";
1135 case ARMISD::VLD2DUP_UPD: return "ARMISD::VLD2DUP_UPD";
1136 case ARMISD::VLD3DUP_UPD: return "ARMISD::VLD3DUP_UPD";
1137 case ARMISD::VLD4DUP_UPD: return "ARMISD::VLD4DUP_UPD";
1138 case ARMISD::VST1_UPD: return "ARMISD::VST1_UPD";
1139 case ARMISD::VST2_UPD: return "ARMISD::VST2_UPD";
1140 case ARMISD::VST3_UPD: return "ARMISD::VST3_UPD";
1141 case ARMISD::VST4_UPD: return "ARMISD::VST4_UPD";
1142 case ARMISD::VST2LN_UPD: return "ARMISD::VST2LN_UPD";
1143 case ARMISD::VST3LN_UPD: return "ARMISD::VST3LN_UPD";
1144 case ARMISD::VST4LN_UPD: return "ARMISD::VST4LN_UPD";
1149 EVT ARMTargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
1150 if (!VT.isVector()) return getPointerTy();
1151 return VT.changeVectorElementTypeToInteger();
1154 /// getRegClassFor - Return the register class that should be used for the
1155 /// specified value type.
1156 const TargetRegisterClass *ARMTargetLowering::getRegClassFor(MVT VT) const {
1157 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
1158 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
1159 // load / store 4 to 8 consecutive D registers.
1160 if (Subtarget->hasNEON()) {
1161 if (VT == MVT::v4i64)
1162 return &ARM::QQPRRegClass;
1163 if (VT == MVT::v8i64)
1164 return &ARM::QQQQPRRegClass;
1166 return TargetLowering::getRegClassFor(VT);
1169 // memcpy, and other memory intrinsics, typically tries to use LDM/STM if the
1170 // source/dest is aligned and the copy size is large enough. We therefore want
1171 // to align such objects passed to memory intrinsics.
1172 bool ARMTargetLowering::shouldAlignPointerArgs(CallInst *CI, unsigned &MinSize,
1173 unsigned &PrefAlign) const {
1174 if (!isa<MemIntrinsic>(CI))
1177 // On ARM11 onwards (excluding M class) 8-byte aligned LDM is typically 1
1178 // cycle faster than 4-byte aligned LDM.
1179 PrefAlign = (Subtarget->hasV6Ops() && !Subtarget->isMClass() ? 8 : 4);
1183 // Create a fast isel object.
1185 ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
1186 const TargetLibraryInfo *libInfo) const {
1187 return ARM::createFastISel(funcInfo, libInfo);
1190 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
1191 unsigned NumVals = N->getNumValues();
1193 return Sched::RegPressure;
1195 for (unsigned i = 0; i != NumVals; ++i) {
1196 EVT VT = N->getValueType(i);
1197 if (VT == MVT::Glue || VT == MVT::Other)
1199 if (VT.isFloatingPoint() || VT.isVector())
1203 if (!N->isMachineOpcode())
1204 return Sched::RegPressure;
1206 // Load are scheduled for latency even if there instruction itinerary
1207 // is not available.
1208 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
1209 const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
1211 if (MCID.getNumDefs() == 0)
1212 return Sched::RegPressure;
1213 if (!Itins->isEmpty() &&
1214 Itins->getOperandCycle(MCID.getSchedClass(), 0) > 2)
1217 return Sched::RegPressure;
1220 //===----------------------------------------------------------------------===//
1222 //===----------------------------------------------------------------------===//
1224 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
1225 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
1227 default: llvm_unreachable("Unknown condition code!");
1228 case ISD::SETNE: return ARMCC::NE;
1229 case ISD::SETEQ: return ARMCC::EQ;
1230 case ISD::SETGT: return ARMCC::GT;
1231 case ISD::SETGE: return ARMCC::GE;
1232 case ISD::SETLT: return ARMCC::LT;
1233 case ISD::SETLE: return ARMCC::LE;
1234 case ISD::SETUGT: return ARMCC::HI;
1235 case ISD::SETUGE: return ARMCC::HS;
1236 case ISD::SETULT: return ARMCC::LO;
1237 case ISD::SETULE: return ARMCC::LS;
1241 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
1242 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
1243 ARMCC::CondCodes &CondCode2) {
1244 CondCode2 = ARMCC::AL;
1246 default: llvm_unreachable("Unknown FP condition!");
1248 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
1250 case ISD::SETOGT: CondCode = ARMCC::GT; break;
1252 case ISD::SETOGE: CondCode = ARMCC::GE; break;
1253 case ISD::SETOLT: CondCode = ARMCC::MI; break;
1254 case ISD::SETOLE: CondCode = ARMCC::LS; break;
1255 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
1256 case ISD::SETO: CondCode = ARMCC::VC; break;
1257 case ISD::SETUO: CondCode = ARMCC::VS; break;
1258 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
1259 case ISD::SETUGT: CondCode = ARMCC::HI; break;
1260 case ISD::SETUGE: CondCode = ARMCC::PL; break;
1262 case ISD::SETULT: CondCode = ARMCC::LT; break;
1264 case ISD::SETULE: CondCode = ARMCC::LE; break;
1266 case ISD::SETUNE: CondCode = ARMCC::NE; break;
1270 //===----------------------------------------------------------------------===//
1271 // Calling Convention Implementation
1272 //===----------------------------------------------------------------------===//
1274 #include "ARMGenCallingConv.inc"
1276 /// getEffectiveCallingConv - Get the effective calling convention, taking into
1277 /// account presence of floating point hardware and calling convention
1278 /// limitations, such as support for variadic functions.
1280 ARMTargetLowering::getEffectiveCallingConv(CallingConv::ID CC,
1281 bool isVarArg) const {
1284 llvm_unreachable("Unsupported calling convention");
1285 case CallingConv::ARM_AAPCS:
1286 case CallingConv::ARM_APCS:
1287 case CallingConv::GHC:
1289 case CallingConv::ARM_AAPCS_VFP:
1290 return isVarArg ? CallingConv::ARM_AAPCS : CallingConv::ARM_AAPCS_VFP;
1291 case CallingConv::C:
1292 if (!Subtarget->isAAPCS_ABI())
1293 return CallingConv::ARM_APCS;
1294 else if (Subtarget->hasVFP2() && !Subtarget->isThumb1Only() &&
1295 getTargetMachine().Options.FloatABIType == FloatABI::Hard &&
1297 return CallingConv::ARM_AAPCS_VFP;
1299 return CallingConv::ARM_AAPCS;
1300 case CallingConv::Fast:
1301 if (!Subtarget->isAAPCS_ABI()) {
1302 if (Subtarget->hasVFP2() && !Subtarget->isThumb1Only() && !isVarArg)
1303 return CallingConv::Fast;
1304 return CallingConv::ARM_APCS;
1305 } else if (Subtarget->hasVFP2() && !Subtarget->isThumb1Only() && !isVarArg)
1306 return CallingConv::ARM_AAPCS_VFP;
1308 return CallingConv::ARM_AAPCS;
1312 /// CCAssignFnForNode - Selects the correct CCAssignFn for the given
1313 /// CallingConvention.
1314 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
1316 bool isVarArg) const {
1317 switch (getEffectiveCallingConv(CC, isVarArg)) {
1319 llvm_unreachable("Unsupported calling convention");
1320 case CallingConv::ARM_APCS:
1321 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1322 case CallingConv::ARM_AAPCS:
1323 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1324 case CallingConv::ARM_AAPCS_VFP:
1325 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1326 case CallingConv::Fast:
1327 return (Return ? RetFastCC_ARM_APCS : FastCC_ARM_APCS);
1328 case CallingConv::GHC:
1329 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS_GHC);
1333 /// LowerCallResult - Lower the result values of a call into the
1334 /// appropriate copies out of appropriate physical registers.
1336 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1337 CallingConv::ID CallConv, bool isVarArg,
1338 const SmallVectorImpl<ISD::InputArg> &Ins,
1339 SDLoc dl, SelectionDAG &DAG,
1340 SmallVectorImpl<SDValue> &InVals,
1341 bool isThisReturn, SDValue ThisVal) const {
1343 // Assign locations to each value returned by this call.
1344 SmallVector<CCValAssign, 16> RVLocs;
1345 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
1346 *DAG.getContext(), Call);
1347 CCInfo.AnalyzeCallResult(Ins,
1348 CCAssignFnForNode(CallConv, /* Return*/ true,
1351 // Copy all of the result registers out of their specified physreg.
1352 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1353 CCValAssign VA = RVLocs[i];
1355 // Pass 'this' value directly from the argument to return value, to avoid
1356 // reg unit interference
1357 if (i == 0 && isThisReturn) {
1358 assert(!VA.needsCustom() && VA.getLocVT() == MVT::i32 &&
1359 "unexpected return calling convention register assignment");
1360 InVals.push_back(ThisVal);
1365 if (VA.needsCustom()) {
1366 // Handle f64 or half of a v2f64.
1367 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1369 Chain = Lo.getValue(1);
1370 InFlag = Lo.getValue(2);
1371 VA = RVLocs[++i]; // skip ahead to next loc
1372 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1374 Chain = Hi.getValue(1);
1375 InFlag = Hi.getValue(2);
1376 if (!Subtarget->isLittle())
1378 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1380 if (VA.getLocVT() == MVT::v2f64) {
1381 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1382 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1383 DAG.getConstant(0, dl, MVT::i32));
1385 VA = RVLocs[++i]; // skip ahead to next loc
1386 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1387 Chain = Lo.getValue(1);
1388 InFlag = Lo.getValue(2);
1389 VA = RVLocs[++i]; // skip ahead to next loc
1390 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1391 Chain = Hi.getValue(1);
1392 InFlag = Hi.getValue(2);
1393 if (!Subtarget->isLittle())
1395 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1396 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1397 DAG.getConstant(1, dl, MVT::i32));
1400 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
1402 Chain = Val.getValue(1);
1403 InFlag = Val.getValue(2);
1406 switch (VA.getLocInfo()) {
1407 default: llvm_unreachable("Unknown loc info!");
1408 case CCValAssign::Full: break;
1409 case CCValAssign::BCvt:
1410 Val = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), Val);
1414 InVals.push_back(Val);
1420 /// LowerMemOpCallTo - Store the argument to the stack.
1422 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
1423 SDValue StackPtr, SDValue Arg,
1424 SDLoc dl, SelectionDAG &DAG,
1425 const CCValAssign &VA,
1426 ISD::ArgFlagsTy Flags) const {
1427 unsigned LocMemOffset = VA.getLocMemOffset();
1428 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset, dl);
1429 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1430 return DAG.getStore(Chain, dl, Arg, PtrOff,
1431 MachinePointerInfo::getStack(LocMemOffset),
1435 void ARMTargetLowering::PassF64ArgInRegs(SDLoc dl, SelectionDAG &DAG,
1436 SDValue Chain, SDValue &Arg,
1437 RegsToPassVector &RegsToPass,
1438 CCValAssign &VA, CCValAssign &NextVA,
1440 SmallVectorImpl<SDValue> &MemOpChains,
1441 ISD::ArgFlagsTy Flags) const {
1443 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1444 DAG.getVTList(MVT::i32, MVT::i32), Arg);
1445 unsigned id = Subtarget->isLittle() ? 0 : 1;
1446 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd.getValue(id)));
1448 if (NextVA.isRegLoc())
1449 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1-id)));
1451 assert(NextVA.isMemLoc());
1452 if (!StackPtr.getNode())
1453 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1455 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1-id),
1461 /// LowerCall - Lowering a call into a callseq_start <-
1462 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
1465 ARMTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
1466 SmallVectorImpl<SDValue> &InVals) const {
1467 SelectionDAG &DAG = CLI.DAG;
1469 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
1470 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
1471 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
1472 SDValue Chain = CLI.Chain;
1473 SDValue Callee = CLI.Callee;
1474 bool &isTailCall = CLI.IsTailCall;
1475 CallingConv::ID CallConv = CLI.CallConv;
1476 bool doesNotRet = CLI.DoesNotReturn;
1477 bool isVarArg = CLI.IsVarArg;
1479 MachineFunction &MF = DAG.getMachineFunction();
1480 bool isStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1481 bool isThisReturn = false;
1482 bool isSibCall = false;
1484 // Disable tail calls if they're not supported.
1485 if (!Subtarget->supportsTailCall() || MF.getTarget().Options.DisableTailCalls)
1489 // Check if it's really possible to do a tail call.
1490 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1491 isVarArg, isStructRet, MF.getFunction()->hasStructRetAttr(),
1492 Outs, OutVals, Ins, DAG);
1493 if (!isTailCall && CLI.CS && CLI.CS->isMustTailCall())
1494 report_fatal_error("failed to perform tail call elimination on a call "
1495 "site marked musttail");
1496 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1497 // detected sibcalls.
1504 // Analyze operands of the call, assigning locations to each operand.
1505 SmallVector<CCValAssign, 16> ArgLocs;
1506 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs,
1507 *DAG.getContext(), Call);
1508 CCInfo.AnalyzeCallOperands(Outs,
1509 CCAssignFnForNode(CallConv, /* Return*/ false,
1512 // Get a count of how many bytes are to be pushed on the stack.
1513 unsigned NumBytes = CCInfo.getNextStackOffset();
1515 // For tail calls, memory operands are available in our caller's stack.
1519 // Adjust the stack pointer for the new arguments...
1520 // These operations are automatically eliminated by the prolog/epilog pass
1522 Chain = DAG.getCALLSEQ_START(Chain,
1523 DAG.getIntPtrConstant(NumBytes, dl, true), dl);
1525 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1527 RegsToPassVector RegsToPass;
1528 SmallVector<SDValue, 8> MemOpChains;
1530 // Walk the register/memloc assignments, inserting copies/loads. In the case
1531 // of tail call optimization, arguments are handled later.
1532 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1534 ++i, ++realArgIdx) {
1535 CCValAssign &VA = ArgLocs[i];
1536 SDValue Arg = OutVals[realArgIdx];
1537 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1538 bool isByVal = Flags.isByVal();
1540 // Promote the value if needed.
1541 switch (VA.getLocInfo()) {
1542 default: llvm_unreachable("Unknown loc info!");
1543 case CCValAssign::Full: break;
1544 case CCValAssign::SExt:
1545 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1547 case CCValAssign::ZExt:
1548 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1550 case CCValAssign::AExt:
1551 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1553 case CCValAssign::BCvt:
1554 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1558 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1559 if (VA.needsCustom()) {
1560 if (VA.getLocVT() == MVT::v2f64) {
1561 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1562 DAG.getConstant(0, dl, MVT::i32));
1563 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1564 DAG.getConstant(1, dl, MVT::i32));
1566 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1567 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1569 VA = ArgLocs[++i]; // skip ahead to next loc
1570 if (VA.isRegLoc()) {
1571 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1572 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1574 assert(VA.isMemLoc());
1576 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1577 dl, DAG, VA, Flags));
1580 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1581 StackPtr, MemOpChains, Flags);
1583 } else if (VA.isRegLoc()) {
1584 if (realArgIdx == 0 && Flags.isReturned() && Outs[0].VT == MVT::i32) {
1585 assert(VA.getLocVT() == MVT::i32 &&
1586 "unexpected calling convention register assignment");
1587 assert(!Ins.empty() && Ins[0].VT == MVT::i32 &&
1588 "unexpected use of 'returned'");
1589 isThisReturn = true;
1591 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1592 } else if (isByVal) {
1593 assert(VA.isMemLoc());
1594 unsigned offset = 0;
1596 // True if this byval aggregate will be split between registers
1598 unsigned ByValArgsCount = CCInfo.getInRegsParamsCount();
1599 unsigned CurByValIdx = CCInfo.getInRegsParamsProcessed();
1601 if (CurByValIdx < ByValArgsCount) {
1603 unsigned RegBegin, RegEnd;
1604 CCInfo.getInRegsParamInfo(CurByValIdx, RegBegin, RegEnd);
1606 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1608 for (i = 0, j = RegBegin; j < RegEnd; i++, j++) {
1609 SDValue Const = DAG.getConstant(4*i, dl, MVT::i32);
1610 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
1611 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
1612 MachinePointerInfo(),
1613 false, false, false,
1614 DAG.InferPtrAlignment(AddArg));
1615 MemOpChains.push_back(Load.getValue(1));
1616 RegsToPass.push_back(std::make_pair(j, Load));
1619 // If parameter size outsides register area, "offset" value
1620 // helps us to calculate stack slot for remained part properly.
1621 offset = RegEnd - RegBegin;
1623 CCInfo.nextInRegsParam();
1626 if (Flags.getByValSize() > 4*offset) {
1627 unsigned LocMemOffset = VA.getLocMemOffset();
1628 SDValue StkPtrOff = DAG.getIntPtrConstant(LocMemOffset, dl);
1629 SDValue Dst = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr,
1631 SDValue SrcOffset = DAG.getIntPtrConstant(4*offset, dl);
1632 SDValue Src = DAG.getNode(ISD::ADD, dl, getPointerTy(), Arg, SrcOffset);
1633 SDValue SizeNode = DAG.getConstant(Flags.getByValSize() - 4*offset, dl,
1635 SDValue AlignNode = DAG.getConstant(Flags.getByValAlign(), dl,
1638 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
1639 SDValue Ops[] = { Chain, Dst, Src, SizeNode, AlignNode};
1640 MemOpChains.push_back(DAG.getNode(ARMISD::COPY_STRUCT_BYVAL, dl, VTs,
1643 } else if (!isSibCall) {
1644 assert(VA.isMemLoc());
1646 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1647 dl, DAG, VA, Flags));
1651 if (!MemOpChains.empty())
1652 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains);
1654 // Build a sequence of copy-to-reg nodes chained together with token chain
1655 // and flag operands which copy the outgoing args into the appropriate regs.
1657 // Tail call byval lowering might overwrite argument registers so in case of
1658 // tail call optimization the copies to registers are lowered later.
1660 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1661 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1662 RegsToPass[i].second, InFlag);
1663 InFlag = Chain.getValue(1);
1666 // For tail calls lower the arguments to the 'real' stack slot.
1668 // Force all the incoming stack arguments to be loaded from the stack
1669 // before any new outgoing arguments are stored to the stack, because the
1670 // outgoing stack slots may alias the incoming argument stack slots, and
1671 // the alias isn't otherwise explicit. This is slightly more conservative
1672 // than necessary, because it means that each store effectively depends
1673 // on every argument instead of just those arguments it would clobber.
1675 // Do not flag preceding copytoreg stuff together with the following stuff.
1677 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1678 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1679 RegsToPass[i].second, InFlag);
1680 InFlag = Chain.getValue(1);
1685 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1686 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1687 // node so that legalize doesn't hack it.
1688 bool isDirect = false;
1689 bool isARMFunc = false;
1690 bool isLocalARMFunc = false;
1691 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1693 if (EnableARMLongCalls) {
1694 assert((Subtarget->isTargetWindows() ||
1695 getTargetMachine().getRelocationModel() == Reloc::Static) &&
1696 "long-calls with non-static relocation model!");
1697 // Handle a global address or an external symbol. If it's not one of
1698 // those, the target's already in a register, so we don't need to do
1700 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1701 const GlobalValue *GV = G->getGlobal();
1702 // Create a constant pool entry for the callee address
1703 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1704 ARMConstantPoolValue *CPV =
1705 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 0);
1707 // Get the address of the callee into a register
1708 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1709 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1710 Callee = DAG.getLoad(getPointerTy(), dl,
1711 DAG.getEntryNode(), CPAddr,
1712 MachinePointerInfo::getConstantPool(),
1713 false, false, false, 0);
1714 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1715 const char *Sym = S->getSymbol();
1717 // Create a constant pool entry for the callee address
1718 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1719 ARMConstantPoolValue *CPV =
1720 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1721 ARMPCLabelIndex, 0);
1722 // Get the address of the callee into a register
1723 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1724 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1725 Callee = DAG.getLoad(getPointerTy(), dl,
1726 DAG.getEntryNode(), CPAddr,
1727 MachinePointerInfo::getConstantPool(),
1728 false, false, false, 0);
1730 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1731 const GlobalValue *GV = G->getGlobal();
1733 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1734 bool isStub = (isExt && Subtarget->isTargetMachO()) &&
1735 getTargetMachine().getRelocationModel() != Reloc::Static;
1736 isARMFunc = !Subtarget->isThumb() || (isStub && !Subtarget->isMClass());
1737 // ARM call to a local ARM function is predicable.
1738 isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking);
1739 // tBX takes a register source operand.
1740 if (isStub && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1741 assert(Subtarget->isTargetMachO() && "WrapperPIC use on non-MachO?");
1742 Callee = DAG.getNode(ARMISD::WrapperPIC, dl, getPointerTy(),
1743 DAG.getTargetGlobalAddress(GV, dl, getPointerTy(),
1744 0, ARMII::MO_NONLAZY));
1745 Callee = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Callee,
1746 MachinePointerInfo::getGOT(), false, false, true, 0);
1747 } else if (Subtarget->isTargetCOFF()) {
1748 assert(Subtarget->isTargetWindows() &&
1749 "Windows is the only supported COFF target");
1750 unsigned TargetFlags = GV->hasDLLImportStorageClass()
1751 ? ARMII::MO_DLLIMPORT
1752 : ARMII::MO_NO_FLAG;
1753 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), /*Offset=*/0,
1755 if (GV->hasDLLImportStorageClass())
1756 Callee = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
1757 DAG.getNode(ARMISD::Wrapper, dl, getPointerTy(),
1758 Callee), MachinePointerInfo::getGOT(),
1759 false, false, false, 0);
1761 // On ELF targets for PIC code, direct calls should go through the PLT
1762 unsigned OpFlags = 0;
1763 if (Subtarget->isTargetELF() &&
1764 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1765 OpFlags = ARMII::MO_PLT;
1766 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
1768 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1770 bool isStub = Subtarget->isTargetMachO() &&
1771 getTargetMachine().getRelocationModel() != Reloc::Static;
1772 isARMFunc = !Subtarget->isThumb() || (isStub && !Subtarget->isMClass());
1773 // tBX takes a register source operand.
1774 const char *Sym = S->getSymbol();
1775 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1776 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1777 ARMConstantPoolValue *CPV =
1778 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1779 ARMPCLabelIndex, 4);
1780 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1781 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1782 Callee = DAG.getLoad(getPointerTy(), dl,
1783 DAG.getEntryNode(), CPAddr,
1784 MachinePointerInfo::getConstantPool(),
1785 false, false, false, 0);
1786 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, dl, MVT::i32);
1787 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1788 getPointerTy(), Callee, PICLabel);
1790 unsigned OpFlags = 0;
1791 // On ELF targets for PIC code, direct calls should go through the PLT
1792 if (Subtarget->isTargetELF() &&
1793 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1794 OpFlags = ARMII::MO_PLT;
1795 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlags);
1799 // FIXME: handle tail calls differently.
1801 bool HasMinSizeAttr = MF.getFunction()->hasFnAttribute(Attribute::MinSize);
1802 if (Subtarget->isThumb()) {
1803 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1804 CallOpc = ARMISD::CALL_NOLINK;
1806 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1808 if (!isDirect && !Subtarget->hasV5TOps())
1809 CallOpc = ARMISD::CALL_NOLINK;
1810 else if (doesNotRet && isDirect && Subtarget->hasRAS() &&
1811 // Emit regular call when code size is the priority
1813 // "mov lr, pc; b _foo" to avoid confusing the RSP
1814 CallOpc = ARMISD::CALL_NOLINK;
1816 CallOpc = isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL;
1819 std::vector<SDValue> Ops;
1820 Ops.push_back(Chain);
1821 Ops.push_back(Callee);
1823 // Add argument registers to the end of the list so that they are known live
1825 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1826 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1827 RegsToPass[i].second.getValueType()));
1829 // Add a register mask operand representing the call-preserved registers.
1831 const uint32_t *Mask;
1832 const ARMBaseRegisterInfo *ARI = Subtarget->getRegisterInfo();
1834 // For 'this' returns, use the R0-preserving mask if applicable
1835 Mask = ARI->getThisReturnPreservedMask(MF, CallConv);
1837 // Set isThisReturn to false if the calling convention is not one that
1838 // allows 'returned' to be modeled in this way, so LowerCallResult does
1839 // not try to pass 'this' straight through
1840 isThisReturn = false;
1841 Mask = ARI->getCallPreservedMask(MF, CallConv);
1844 Mask = ARI->getCallPreservedMask(MF, CallConv);
1846 assert(Mask && "Missing call preserved mask for calling convention");
1847 Ops.push_back(DAG.getRegisterMask(Mask));
1850 if (InFlag.getNode())
1851 Ops.push_back(InFlag);
1853 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
1855 MF.getFrameInfo()->setHasTailCall();
1856 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, Ops);
1859 // Returns a chain and a flag for retval copy to use.
1860 Chain = DAG.getNode(CallOpc, dl, NodeTys, Ops);
1861 InFlag = Chain.getValue(1);
1863 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, dl, true),
1864 DAG.getIntPtrConstant(0, dl, true), InFlag, dl);
1866 InFlag = Chain.getValue(1);
1868 // Handle result values, copying them out of physregs into vregs that we
1870 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins, dl, DAG,
1871 InVals, isThisReturn,
1872 isThisReturn ? OutVals[0] : SDValue());
1875 /// HandleByVal - Every parameter *after* a byval parameter is passed
1876 /// on the stack. Remember the next parameter register to allocate,
1877 /// and then confiscate the rest of the parameter registers to insure
1879 void ARMTargetLowering::HandleByVal(CCState *State, unsigned &Size,
1880 unsigned Align) const {
1881 assert((State->getCallOrPrologue() == Prologue ||
1882 State->getCallOrPrologue() == Call) &&
1883 "unhandled ParmContext");
1885 // Byval (as with any stack) slots are always at least 4 byte aligned.
1886 Align = std::max(Align, 4U);
1888 unsigned Reg = State->AllocateReg(GPRArgRegs);
1892 unsigned AlignInRegs = Align / 4;
1893 unsigned Waste = (ARM::R4 - Reg) % AlignInRegs;
1894 for (unsigned i = 0; i < Waste; ++i)
1895 Reg = State->AllocateReg(GPRArgRegs);
1900 unsigned Excess = 4 * (ARM::R4 - Reg);
1902 // Special case when NSAA != SP and parameter size greater than size of
1903 // all remained GPR regs. In that case we can't split parameter, we must
1904 // send it to stack. We also must set NCRN to R4, so waste all
1905 // remained registers.
1906 const unsigned NSAAOffset = State->getNextStackOffset();
1907 if (NSAAOffset != 0 && Size > Excess) {
1908 while (State->AllocateReg(GPRArgRegs))
1913 // First register for byval parameter is the first register that wasn't
1914 // allocated before this method call, so it would be "reg".
1915 // If parameter is small enough to be saved in range [reg, r4), then
1916 // the end (first after last) register would be reg + param-size-in-regs,
1917 // else parameter would be splitted between registers and stack,
1918 // end register would be r4 in this case.
1919 unsigned ByValRegBegin = Reg;
1920 unsigned ByValRegEnd = std::min<unsigned>(Reg + Size / 4, ARM::R4);
1921 State->addInRegsParamInfo(ByValRegBegin, ByValRegEnd);
1922 // Note, first register is allocated in the beginning of function already,
1923 // allocate remained amount of registers we need.
1924 for (unsigned i = Reg + 1; i != ByValRegEnd; ++i)
1925 State->AllocateReg(GPRArgRegs);
1926 // A byval parameter that is split between registers and memory needs its
1927 // size truncated here.
1928 // In the case where the entire structure fits in registers, we set the
1929 // size in memory to zero.
1930 Size = std::max<int>(Size - Excess, 0);
1933 /// MatchingStackOffset - Return true if the given stack call argument is
1934 /// already available in the same position (relatively) of the caller's
1935 /// incoming argument stack.
1937 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1938 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1939 const TargetInstrInfo *TII) {
1940 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1942 if (Arg.getOpcode() == ISD::CopyFromReg) {
1943 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1944 if (!TargetRegisterInfo::isVirtualRegister(VR))
1946 MachineInstr *Def = MRI->getVRegDef(VR);
1949 if (!Flags.isByVal()) {
1950 if (!TII->isLoadFromStackSlot(Def, FI))
1955 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
1956 if (Flags.isByVal())
1957 // ByVal argument is passed in as a pointer but it's now being
1958 // dereferenced. e.g.
1959 // define @foo(%struct.X* %A) {
1960 // tail call @bar(%struct.X* byval %A)
1963 SDValue Ptr = Ld->getBasePtr();
1964 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
1967 FI = FINode->getIndex();
1971 assert(FI != INT_MAX);
1972 if (!MFI->isFixedObjectIndex(FI))
1974 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
1977 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
1978 /// for tail call optimization. Targets which want to do tail call
1979 /// optimization should implement this function.
1981 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1982 CallingConv::ID CalleeCC,
1984 bool isCalleeStructRet,
1985 bool isCallerStructRet,
1986 const SmallVectorImpl<ISD::OutputArg> &Outs,
1987 const SmallVectorImpl<SDValue> &OutVals,
1988 const SmallVectorImpl<ISD::InputArg> &Ins,
1989 SelectionDAG& DAG) const {
1990 const Function *CallerF = DAG.getMachineFunction().getFunction();
1991 CallingConv::ID CallerCC = CallerF->getCallingConv();
1992 bool CCMatch = CallerCC == CalleeCC;
1994 // Look for obvious safe cases to perform tail call optimization that do not
1995 // require ABI changes. This is what gcc calls sibcall.
1997 // Do not sibcall optimize vararg calls unless the call site is not passing
1999 if (isVarArg && !Outs.empty())
2002 // Exception-handling functions need a special set of instructions to indicate
2003 // a return to the hardware. Tail-calling another function would probably
2005 if (CallerF->hasFnAttribute("interrupt"))
2008 // Also avoid sibcall optimization if either caller or callee uses struct
2009 // return semantics.
2010 if (isCalleeStructRet || isCallerStructRet)
2013 // FIXME: Completely disable sibcall for Thumb1 since ThumbRegisterInfo::
2014 // emitEpilogue is not ready for them. Thumb tail calls also use t2B, as
2015 // the Thumb1 16-bit unconditional branch doesn't have sufficient relocation
2016 // support in the assembler and linker to be used. This would need to be
2017 // fixed to fully support tail calls in Thumb1.
2019 // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take
2020 // LR. This means if we need to reload LR, it takes an extra instructions,
2021 // which outweighs the value of the tail call; but here we don't know yet
2022 // whether LR is going to be used. Probably the right approach is to
2023 // generate the tail call here and turn it back into CALL/RET in
2024 // emitEpilogue if LR is used.
2026 // Thumb1 PIC calls to external symbols use BX, so they can be tail calls,
2027 // but we need to make sure there are enough registers; the only valid
2028 // registers are the 4 used for parameters. We don't currently do this
2030 if (Subtarget->isThumb1Only())
2033 // Externally-defined functions with weak linkage should not be
2034 // tail-called on ARM when the OS does not support dynamic
2035 // pre-emption of symbols, as the AAELF spec requires normal calls
2036 // to undefined weak functions to be replaced with a NOP or jump to the
2037 // next instruction. The behaviour of branch instructions in this
2038 // situation (as used for tail calls) is implementation-defined, so we
2039 // cannot rely on the linker replacing the tail call with a return.
2040 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2041 const GlobalValue *GV = G->getGlobal();
2042 const Triple TT(getTargetMachine().getTargetTriple());
2043 if (GV->hasExternalWeakLinkage() &&
2044 (!TT.isOSWindows() || TT.isOSBinFormatELF() || TT.isOSBinFormatMachO()))
2048 // If the calling conventions do not match, then we'd better make sure the
2049 // results are returned in the same way as what the caller expects.
2051 SmallVector<CCValAssign, 16> RVLocs1;
2052 ARMCCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(), RVLocs1,
2053 *DAG.getContext(), Call);
2054 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
2056 SmallVector<CCValAssign, 16> RVLocs2;
2057 ARMCCState CCInfo2(CallerCC, false, DAG.getMachineFunction(), RVLocs2,
2058 *DAG.getContext(), Call);
2059 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
2061 if (RVLocs1.size() != RVLocs2.size())
2063 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
2064 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
2066 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
2068 if (RVLocs1[i].isRegLoc()) {
2069 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
2072 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
2078 // If Caller's vararg or byval argument has been split between registers and
2079 // stack, do not perform tail call, since part of the argument is in caller's
2081 const ARMFunctionInfo *AFI_Caller = DAG.getMachineFunction().
2082 getInfo<ARMFunctionInfo>();
2083 if (AFI_Caller->getArgRegsSaveSize())
2086 // If the callee takes no arguments then go on to check the results of the
2088 if (!Outs.empty()) {
2089 // Check if stack adjustment is needed. For now, do not do this if any
2090 // argument is passed on the stack.
2091 SmallVector<CCValAssign, 16> ArgLocs;
2092 ARMCCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
2093 *DAG.getContext(), Call);
2094 CCInfo.AnalyzeCallOperands(Outs,
2095 CCAssignFnForNode(CalleeCC, false, isVarArg));
2096 if (CCInfo.getNextStackOffset()) {
2097 MachineFunction &MF = DAG.getMachineFunction();
2099 // Check if the arguments are already laid out in the right way as
2100 // the caller's fixed stack objects.
2101 MachineFrameInfo *MFI = MF.getFrameInfo();
2102 const MachineRegisterInfo *MRI = &MF.getRegInfo();
2103 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
2104 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
2106 ++i, ++realArgIdx) {
2107 CCValAssign &VA = ArgLocs[i];
2108 EVT RegVT = VA.getLocVT();
2109 SDValue Arg = OutVals[realArgIdx];
2110 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
2111 if (VA.getLocInfo() == CCValAssign::Indirect)
2113 if (VA.needsCustom()) {
2114 // f64 and vector types are split into multiple registers or
2115 // register/stack-slot combinations. The types will not match
2116 // the registers; give up on memory f64 refs until we figure
2117 // out what to do about this.
2120 if (!ArgLocs[++i].isRegLoc())
2122 if (RegVT == MVT::v2f64) {
2123 if (!ArgLocs[++i].isRegLoc())
2125 if (!ArgLocs[++i].isRegLoc())
2128 } else if (!VA.isRegLoc()) {
2129 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
2141 ARMTargetLowering::CanLowerReturn(CallingConv::ID CallConv,
2142 MachineFunction &MF, bool isVarArg,
2143 const SmallVectorImpl<ISD::OutputArg> &Outs,
2144 LLVMContext &Context) const {
2145 SmallVector<CCValAssign, 16> RVLocs;
2146 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, Context);
2147 return CCInfo.CheckReturn(Outs, CCAssignFnForNode(CallConv, /*Return=*/true,
2151 static SDValue LowerInterruptReturn(SmallVectorImpl<SDValue> &RetOps,
2152 SDLoc DL, SelectionDAG &DAG) {
2153 const MachineFunction &MF = DAG.getMachineFunction();
2154 const Function *F = MF.getFunction();
2156 StringRef IntKind = F->getFnAttribute("interrupt").getValueAsString();
2158 // See ARM ARM v7 B1.8.3. On exception entry LR is set to a possibly offset
2159 // version of the "preferred return address". These offsets affect the return
2160 // instruction if this is a return from PL1 without hypervisor extensions.
2161 // IRQ/FIQ: +4 "subs pc, lr, #4"
2162 // SWI: 0 "subs pc, lr, #0"
2163 // ABORT: +4 "subs pc, lr, #4"
2164 // UNDEF: +4/+2 "subs pc, lr, #0"
2165 // UNDEF varies depending on where the exception came from ARM or Thumb
2166 // mode. Alongside GCC, we throw our hands up in disgust and pretend it's 0.
2169 if (IntKind == "" || IntKind == "IRQ" || IntKind == "FIQ" ||
2172 else if (IntKind == "SWI" || IntKind == "UNDEF")
2175 report_fatal_error("Unsupported interrupt attribute. If present, value "
2176 "must be one of: IRQ, FIQ, SWI, ABORT or UNDEF");
2178 RetOps.insert(RetOps.begin() + 1,
2179 DAG.getConstant(LROffset, DL, MVT::i32, false));
2181 return DAG.getNode(ARMISD::INTRET_FLAG, DL, MVT::Other, RetOps);
2185 ARMTargetLowering::LowerReturn(SDValue Chain,
2186 CallingConv::ID CallConv, bool isVarArg,
2187 const SmallVectorImpl<ISD::OutputArg> &Outs,
2188 const SmallVectorImpl<SDValue> &OutVals,
2189 SDLoc dl, SelectionDAG &DAG) const {
2191 // CCValAssign - represent the assignment of the return value to a location.
2192 SmallVector<CCValAssign, 16> RVLocs;
2194 // CCState - Info about the registers and stack slots.
2195 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
2196 *DAG.getContext(), Call);
2198 // Analyze outgoing return values.
2199 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
2203 SmallVector<SDValue, 4> RetOps;
2204 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
2205 bool isLittleEndian = Subtarget->isLittle();
2207 MachineFunction &MF = DAG.getMachineFunction();
2208 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2209 AFI->setReturnRegsCount(RVLocs.size());
2211 // Copy the result values into the output registers.
2212 for (unsigned i = 0, realRVLocIdx = 0;
2214 ++i, ++realRVLocIdx) {
2215 CCValAssign &VA = RVLocs[i];
2216 assert(VA.isRegLoc() && "Can only return in registers!");
2218 SDValue Arg = OutVals[realRVLocIdx];
2220 switch (VA.getLocInfo()) {
2221 default: llvm_unreachable("Unknown loc info!");
2222 case CCValAssign::Full: break;
2223 case CCValAssign::BCvt:
2224 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
2228 if (VA.needsCustom()) {
2229 if (VA.getLocVT() == MVT::v2f64) {
2230 // Extract the first half and return it in two registers.
2231 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
2232 DAG.getConstant(0, dl, MVT::i32));
2233 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
2234 DAG.getVTList(MVT::i32, MVT::i32), Half);
2236 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2237 HalfGPRs.getValue(isLittleEndian ? 0 : 1),
2239 Flag = Chain.getValue(1);
2240 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2241 VA = RVLocs[++i]; // skip ahead to next loc
2242 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2243 HalfGPRs.getValue(isLittleEndian ? 1 : 0),
2245 Flag = Chain.getValue(1);
2246 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2247 VA = RVLocs[++i]; // skip ahead to next loc
2249 // Extract the 2nd half and fall through to handle it as an f64 value.
2250 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
2251 DAG.getConstant(1, dl, MVT::i32));
2253 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
2255 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
2256 DAG.getVTList(MVT::i32, MVT::i32), Arg);
2257 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2258 fmrrd.getValue(isLittleEndian ? 0 : 1),
2260 Flag = Chain.getValue(1);
2261 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2262 VA = RVLocs[++i]; // skip ahead to next loc
2263 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2264 fmrrd.getValue(isLittleEndian ? 1 : 0),
2267 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
2269 // Guarantee that all emitted copies are
2270 // stuck together, avoiding something bad.
2271 Flag = Chain.getValue(1);
2272 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2275 // Update chain and glue.
2278 RetOps.push_back(Flag);
2280 // CPUs which aren't M-class use a special sequence to return from
2281 // exceptions (roughly, any instruction setting pc and cpsr simultaneously,
2282 // though we use "subs pc, lr, #N").
2284 // M-class CPUs actually use a normal return sequence with a special
2285 // (hardware-provided) value in LR, so the normal code path works.
2286 if (DAG.getMachineFunction().getFunction()->hasFnAttribute("interrupt") &&
2287 !Subtarget->isMClass()) {
2288 if (Subtarget->isThumb1Only())
2289 report_fatal_error("interrupt attribute is not supported in Thumb1");
2290 return LowerInterruptReturn(RetOps, dl, DAG);
2293 return DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, RetOps);
2296 bool ARMTargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
2297 if (N->getNumValues() != 1)
2299 if (!N->hasNUsesOfValue(1, 0))
2302 SDValue TCChain = Chain;
2303 SDNode *Copy = *N->use_begin();
2304 if (Copy->getOpcode() == ISD::CopyToReg) {
2305 // If the copy has a glue operand, we conservatively assume it isn't safe to
2306 // perform a tail call.
2307 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
2309 TCChain = Copy->getOperand(0);
2310 } else if (Copy->getOpcode() == ARMISD::VMOVRRD) {
2311 SDNode *VMov = Copy;
2312 // f64 returned in a pair of GPRs.
2313 SmallPtrSet<SDNode*, 2> Copies;
2314 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
2316 if (UI->getOpcode() != ISD::CopyToReg)
2320 if (Copies.size() > 2)
2323 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
2325 SDValue UseChain = UI->getOperand(0);
2326 if (Copies.count(UseChain.getNode()))
2330 // We are at the top of this chain.
2331 // If the copy has a glue operand, we conservatively assume it
2332 // isn't safe to perform a tail call.
2333 if (UI->getOperand(UI->getNumOperands()-1).getValueType() == MVT::Glue)
2339 } else if (Copy->getOpcode() == ISD::BITCAST) {
2340 // f32 returned in a single GPR.
2341 if (!Copy->hasOneUse())
2343 Copy = *Copy->use_begin();
2344 if (Copy->getOpcode() != ISD::CopyToReg || !Copy->hasNUsesOfValue(1, 0))
2346 // If the copy has a glue operand, we conservatively assume it isn't safe to
2347 // perform a tail call.
2348 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
2350 TCChain = Copy->getOperand(0);
2355 bool HasRet = false;
2356 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
2358 if (UI->getOpcode() != ARMISD::RET_FLAG &&
2359 UI->getOpcode() != ARMISD::INTRET_FLAG)
2371 bool ARMTargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
2372 if (!Subtarget->supportsTailCall())
2375 if (!CI->isTailCall() || getTargetMachine().Options.DisableTailCalls)
2378 return !Subtarget->isThumb1Only();
2381 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
2382 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
2383 // one of the above mentioned nodes. It has to be wrapped because otherwise
2384 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
2385 // be used to form addressing mode. These wrapped nodes will be selected
2387 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
2388 EVT PtrVT = Op.getValueType();
2389 // FIXME there is no actual debug info here
2391 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
2393 if (CP->isMachineConstantPoolEntry())
2394 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
2395 CP->getAlignment());
2397 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
2398 CP->getAlignment());
2399 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
2402 unsigned ARMTargetLowering::getJumpTableEncoding() const {
2403 return MachineJumpTableInfo::EK_Inline;
2406 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
2407 SelectionDAG &DAG) const {
2408 MachineFunction &MF = DAG.getMachineFunction();
2409 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2410 unsigned ARMPCLabelIndex = 0;
2412 EVT PtrVT = getPointerTy();
2413 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
2414 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2416 if (RelocM == Reloc::Static) {
2417 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
2419 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2420 ARMPCLabelIndex = AFI->createPICLabelUId();
2421 ARMConstantPoolValue *CPV =
2422 ARMConstantPoolConstant::Create(BA, ARMPCLabelIndex,
2423 ARMCP::CPBlockAddress, PCAdj);
2424 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2426 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
2427 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
2428 MachinePointerInfo::getConstantPool(),
2429 false, false, false, 0);
2430 if (RelocM == Reloc::Static)
2432 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, DL, MVT::i32);
2433 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
2436 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
2438 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
2439 SelectionDAG &DAG) const {
2441 EVT PtrVT = getPointerTy();
2442 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2443 MachineFunction &MF = DAG.getMachineFunction();
2444 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2445 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2446 ARMConstantPoolValue *CPV =
2447 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2448 ARMCP::CPValue, PCAdj, ARMCP::TLSGD, true);
2449 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2450 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
2451 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
2452 MachinePointerInfo::getConstantPool(),
2453 false, false, false, 0);
2454 SDValue Chain = Argument.getValue(1);
2456 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, dl, MVT::i32);
2457 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
2459 // call __tls_get_addr.
2462 Entry.Node = Argument;
2463 Entry.Ty = (Type *) Type::getInt32Ty(*DAG.getContext());
2464 Args.push_back(Entry);
2466 // FIXME: is there useful debug info available here?
2467 TargetLowering::CallLoweringInfo CLI(DAG);
2468 CLI.setDebugLoc(dl).setChain(Chain)
2469 .setCallee(CallingConv::C, Type::getInt32Ty(*DAG.getContext()),
2470 DAG.getExternalSymbol("__tls_get_addr", PtrVT), std::move(Args),
2473 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
2474 return CallResult.first;
2477 // Lower ISD::GlobalTLSAddress using the "initial exec" or
2478 // "local exec" model.
2480 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
2482 TLSModel::Model model) const {
2483 const GlobalValue *GV = GA->getGlobal();
2486 SDValue Chain = DAG.getEntryNode();
2487 EVT PtrVT = getPointerTy();
2488 // Get the Thread Pointer
2489 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2491 if (model == TLSModel::InitialExec) {
2492 MachineFunction &MF = DAG.getMachineFunction();
2493 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2494 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2495 // Initial exec model.
2496 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2497 ARMConstantPoolValue *CPV =
2498 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2499 ARMCP::CPValue, PCAdj, ARMCP::GOTTPOFF,
2501 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2502 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2503 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2504 MachinePointerInfo::getConstantPool(),
2505 false, false, false, 0);
2506 Chain = Offset.getValue(1);
2508 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, dl, MVT::i32);
2509 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
2511 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2512 MachinePointerInfo::getConstantPool(),
2513 false, false, false, 0);
2516 assert(model == TLSModel::LocalExec);
2517 ARMConstantPoolValue *CPV =
2518 ARMConstantPoolConstant::Create(GV, ARMCP::TPOFF);
2519 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2520 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2521 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2522 MachinePointerInfo::getConstantPool(),
2523 false, false, false, 0);
2526 // The address of the thread local variable is the add of the thread
2527 // pointer with the offset of the variable.
2528 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
2532 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
2533 // TODO: implement the "local dynamic" model
2534 assert(Subtarget->isTargetELF() &&
2535 "TLS not implemented for non-ELF targets");
2536 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
2538 TLSModel::Model model = getTargetMachine().getTLSModel(GA->getGlobal());
2541 case TLSModel::GeneralDynamic:
2542 case TLSModel::LocalDynamic:
2543 return LowerToTLSGeneralDynamicModel(GA, DAG);
2544 case TLSModel::InitialExec:
2545 case TLSModel::LocalExec:
2546 return LowerToTLSExecModels(GA, DAG, model);
2548 llvm_unreachable("bogus TLS model");
2551 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
2552 SelectionDAG &DAG) const {
2553 EVT PtrVT = getPointerTy();
2555 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2556 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2557 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
2558 ARMConstantPoolValue *CPV =
2559 ARMConstantPoolConstant::Create(GV,
2560 UseGOTOFF ? ARMCP::GOTOFF : ARMCP::GOT);
2561 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2562 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2563 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
2565 MachinePointerInfo::getConstantPool(),
2566 false, false, false, 0);
2567 SDValue Chain = Result.getValue(1);
2568 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
2569 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
2571 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
2572 MachinePointerInfo::getGOT(),
2573 false, false, false, 0);
2577 // If we have T2 ops, we can materialize the address directly via movt/movw
2578 // pair. This is always cheaper.
2579 if (Subtarget->useMovt(DAG.getMachineFunction())) {
2581 // FIXME: Once remat is capable of dealing with instructions with register
2582 // operands, expand this into two nodes.
2583 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2584 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2586 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2587 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2588 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2589 MachinePointerInfo::getConstantPool(),
2590 false, false, false, 0);
2594 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
2595 SelectionDAG &DAG) const {
2596 EVT PtrVT = getPointerTy();
2598 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2599 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2601 if (Subtarget->useMovt(DAG.getMachineFunction()))
2604 // FIXME: Once remat is capable of dealing with instructions with register
2605 // operands, expand this into multiple nodes
2607 RelocM == Reloc::PIC_ ? ARMISD::WrapperPIC : ARMISD::Wrapper;
2609 SDValue G = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, ARMII::MO_NONLAZY);
2610 SDValue Result = DAG.getNode(Wrapper, dl, PtrVT, G);
2612 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2613 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
2614 MachinePointerInfo::getGOT(), false, false, false, 0);
2618 SDValue ARMTargetLowering::LowerGlobalAddressWindows(SDValue Op,
2619 SelectionDAG &DAG) const {
2620 assert(Subtarget->isTargetWindows() && "non-Windows COFF is not supported");
2621 assert(Subtarget->useMovt(DAG.getMachineFunction()) &&
2622 "Windows on ARM expects to use movw/movt");
2624 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2625 const ARMII::TOF TargetFlags =
2626 (GV->hasDLLImportStorageClass() ? ARMII::MO_DLLIMPORT : ARMII::MO_NO_FLAG);
2627 EVT PtrVT = getPointerTy();
2633 // FIXME: Once remat is capable of dealing with instructions with register
2634 // operands, expand this into two nodes.
2635 Result = DAG.getNode(ARMISD::Wrapper, DL, PtrVT,
2636 DAG.getTargetGlobalAddress(GV, DL, PtrVT, /*Offset=*/0,
2638 if (GV->hasDLLImportStorageClass())
2639 Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), Result,
2640 MachinePointerInfo::getGOT(), false, false, false, 0);
2644 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
2645 SelectionDAG &DAG) const {
2646 assert(Subtarget->isTargetELF() &&
2647 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
2648 MachineFunction &MF = DAG.getMachineFunction();
2649 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2650 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2651 EVT PtrVT = getPointerTy();
2653 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2654 ARMConstantPoolValue *CPV =
2655 ARMConstantPoolSymbol::Create(*DAG.getContext(), "_GLOBAL_OFFSET_TABLE_",
2656 ARMPCLabelIndex, PCAdj);
2657 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2658 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2659 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2660 MachinePointerInfo::getConstantPool(),
2661 false, false, false, 0);
2662 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, dl, MVT::i32);
2663 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2667 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
2669 SDValue Val = DAG.getConstant(0, dl, MVT::i32);
2670 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl,
2671 DAG.getVTList(MVT::i32, MVT::Other), Op.getOperand(0),
2672 Op.getOperand(1), Val);
2676 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
2678 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
2679 Op.getOperand(1), DAG.getConstant(0, dl, MVT::i32));
2683 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
2684 const ARMSubtarget *Subtarget) const {
2685 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2688 default: return SDValue(); // Don't custom lower most intrinsics.
2689 case Intrinsic::arm_rbit: {
2690 assert(Op.getOperand(1).getValueType() == MVT::i32 &&
2691 "RBIT intrinsic must have i32 type!");
2692 return DAG.getNode(ARMISD::RBIT, dl, MVT::i32, Op.getOperand(1));
2694 case Intrinsic::arm_thread_pointer: {
2695 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2696 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2698 case Intrinsic::eh_sjlj_lsda: {
2699 MachineFunction &MF = DAG.getMachineFunction();
2700 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2701 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2702 EVT PtrVT = getPointerTy();
2703 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2705 unsigned PCAdj = (RelocM != Reloc::PIC_)
2706 ? 0 : (Subtarget->isThumb() ? 4 : 8);
2707 ARMConstantPoolValue *CPV =
2708 ARMConstantPoolConstant::Create(MF.getFunction(), ARMPCLabelIndex,
2709 ARMCP::CPLSDA, PCAdj);
2710 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2711 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2713 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2714 MachinePointerInfo::getConstantPool(),
2715 false, false, false, 0);
2717 if (RelocM == Reloc::PIC_) {
2718 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, dl, MVT::i32);
2719 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2723 case Intrinsic::arm_neon_vmulls:
2724 case Intrinsic::arm_neon_vmullu: {
2725 unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vmulls)
2726 ? ARMISD::VMULLs : ARMISD::VMULLu;
2727 return DAG.getNode(NewOpc, SDLoc(Op), Op.getValueType(),
2728 Op.getOperand(1), Op.getOperand(2));
2733 static SDValue LowerATOMIC_FENCE(SDValue Op, SelectionDAG &DAG,
2734 const ARMSubtarget *Subtarget) {
2735 // FIXME: handle "fence singlethread" more efficiently.
2737 if (!Subtarget->hasDataBarrier()) {
2738 // Some ARMv6 cpus can support data barriers with an mcr instruction.
2739 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2741 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2742 "Unexpected ISD::ATOMIC_FENCE encountered. Should be libcall!");
2743 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2744 DAG.getConstant(0, dl, MVT::i32));
2747 ConstantSDNode *OrdN = cast<ConstantSDNode>(Op.getOperand(1));
2748 AtomicOrdering Ord = static_cast<AtomicOrdering>(OrdN->getZExtValue());
2749 ARM_MB::MemBOpt Domain = ARM_MB::ISH;
2750 if (Subtarget->isMClass()) {
2751 // Only a full system barrier exists in the M-class architectures.
2752 Domain = ARM_MB::SY;
2753 } else if (Subtarget->isSwift() && Ord == Release) {
2754 // Swift happens to implement ISHST barriers in a way that's compatible with
2755 // Release semantics but weaker than ISH so we'd be fools not to use
2756 // it. Beware: other processors probably don't!
2757 Domain = ARM_MB::ISHST;
2760 return DAG.getNode(ISD::INTRINSIC_VOID, dl, MVT::Other, Op.getOperand(0),
2761 DAG.getConstant(Intrinsic::arm_dmb, dl, MVT::i32),
2762 DAG.getConstant(Domain, dl, MVT::i32));
2765 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG,
2766 const ARMSubtarget *Subtarget) {
2767 // ARM pre v5TE and Thumb1 does not have preload instructions.
2768 if (!(Subtarget->isThumb2() ||
2769 (!Subtarget->isThumb1Only() && Subtarget->hasV5TEOps())))
2770 // Just preserve the chain.
2771 return Op.getOperand(0);
2774 unsigned isRead = ~cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() & 1;
2776 (!Subtarget->hasV7Ops() || !Subtarget->hasMPExtension()))
2777 // ARMv7 with MP extension has PLDW.
2778 return Op.getOperand(0);
2780 unsigned isData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
2781 if (Subtarget->isThumb()) {
2783 isRead = ~isRead & 1;
2784 isData = ~isData & 1;
2787 return DAG.getNode(ARMISD::PRELOAD, dl, MVT::Other, Op.getOperand(0),
2788 Op.getOperand(1), DAG.getConstant(isRead, dl, MVT::i32),
2789 DAG.getConstant(isData, dl, MVT::i32));
2792 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
2793 MachineFunction &MF = DAG.getMachineFunction();
2794 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
2796 // vastart just stores the address of the VarArgsFrameIndex slot into the
2797 // memory location argument.
2799 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2800 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2801 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
2802 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
2803 MachinePointerInfo(SV), false, false, 0);
2807 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
2808 SDValue &Root, SelectionDAG &DAG,
2810 MachineFunction &MF = DAG.getMachineFunction();
2811 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2813 const TargetRegisterClass *RC;
2814 if (AFI->isThumb1OnlyFunction())
2815 RC = &ARM::tGPRRegClass;
2817 RC = &ARM::GPRRegClass;
2819 // Transform the arguments stored in physical registers into virtual ones.
2820 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2821 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2824 if (NextVA.isMemLoc()) {
2825 MachineFrameInfo *MFI = MF.getFrameInfo();
2826 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true);
2828 // Create load node to retrieve arguments from the stack.
2829 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2830 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
2831 MachinePointerInfo::getFixedStack(FI),
2832 false, false, false, 0);
2834 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
2835 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2837 if (!Subtarget->isLittle())
2838 std::swap (ArgValue, ArgValue2);
2839 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
2842 // The remaining GPRs hold either the beginning of variable-argument
2843 // data, or the beginning of an aggregate passed by value (usually
2844 // byval). Either way, we allocate stack slots adjacent to the data
2845 // provided by our caller, and store the unallocated registers there.
2846 // If this is a variadic function, the va_list pointer will begin with
2847 // these values; otherwise, this reassembles a (byval) structure that
2848 // was split between registers and memory.
2849 // Return: The frame index registers were stored into.
2851 ARMTargetLowering::StoreByValRegs(CCState &CCInfo, SelectionDAG &DAG,
2852 SDLoc dl, SDValue &Chain,
2853 const Value *OrigArg,
2854 unsigned InRegsParamRecordIdx,
2856 unsigned ArgSize) const {
2857 // Currently, two use-cases possible:
2858 // Case #1. Non-var-args function, and we meet first byval parameter.
2859 // Setup first unallocated register as first byval register;
2860 // eat all remained registers
2861 // (these two actions are performed by HandleByVal method).
2862 // Then, here, we initialize stack frame with
2863 // "store-reg" instructions.
2864 // Case #2. Var-args function, that doesn't contain byval parameters.
2865 // The same: eat all remained unallocated registers,
2866 // initialize stack frame.
2868 MachineFunction &MF = DAG.getMachineFunction();
2869 MachineFrameInfo *MFI = MF.getFrameInfo();
2870 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2871 unsigned RBegin, REnd;
2872 if (InRegsParamRecordIdx < CCInfo.getInRegsParamsCount()) {
2873 CCInfo.getInRegsParamInfo(InRegsParamRecordIdx, RBegin, REnd);
2875 unsigned RBeginIdx = CCInfo.getFirstUnallocated(GPRArgRegs);
2876 RBegin = RBeginIdx == 4 ? (unsigned)ARM::R4 : GPRArgRegs[RBeginIdx];
2881 ArgOffset = -4 * (ARM::R4 - RBegin);
2883 int FrameIndex = MFI->CreateFixedObject(ArgSize, ArgOffset, false);
2884 SDValue FIN = DAG.getFrameIndex(FrameIndex, getPointerTy());
2886 SmallVector<SDValue, 4> MemOps;
2887 const TargetRegisterClass *RC =
2888 AFI->isThumb1OnlyFunction() ? &ARM::tGPRRegClass : &ARM::GPRRegClass;
2890 for (unsigned Reg = RBegin, i = 0; Reg < REnd; ++Reg, ++i) {
2891 unsigned VReg = MF.addLiveIn(Reg, RC);
2892 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2894 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2895 MachinePointerInfo(OrigArg, 4 * i), false, false, 0);
2896 MemOps.push_back(Store);
2897 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
2898 DAG.getConstant(4, dl, getPointerTy()));
2901 if (!MemOps.empty())
2902 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOps);
2906 // Setup stack frame, the va_list pointer will start from.
2908 ARMTargetLowering::VarArgStyleRegisters(CCState &CCInfo, SelectionDAG &DAG,
2909 SDLoc dl, SDValue &Chain,
2911 unsigned TotalArgRegsSaveSize,
2912 bool ForceMutable) const {
2913 MachineFunction &MF = DAG.getMachineFunction();
2914 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2916 // Try to store any remaining integer argument regs
2917 // to their spots on the stack so that they may be loaded by deferencing
2918 // the result of va_next.
2919 // If there is no regs to be stored, just point address after last
2920 // argument passed via stack.
2921 int FrameIndex = StoreByValRegs(CCInfo, DAG, dl, Chain, nullptr,
2922 CCInfo.getInRegsParamsCount(),
2923 CCInfo.getNextStackOffset(), 4);
2924 AFI->setVarArgsFrameIndex(FrameIndex);
2928 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
2929 CallingConv::ID CallConv, bool isVarArg,
2930 const SmallVectorImpl<ISD::InputArg>
2932 SDLoc dl, SelectionDAG &DAG,
2933 SmallVectorImpl<SDValue> &InVals)
2935 MachineFunction &MF = DAG.getMachineFunction();
2936 MachineFrameInfo *MFI = MF.getFrameInfo();
2938 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2940 // Assign locations to all of the incoming arguments.
2941 SmallVector<CCValAssign, 16> ArgLocs;
2942 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs,
2943 *DAG.getContext(), Prologue);
2944 CCInfo.AnalyzeFormalArguments(Ins,
2945 CCAssignFnForNode(CallConv, /* Return*/ false,
2948 SmallVector<SDValue, 16> ArgValues;
2950 Function::const_arg_iterator CurOrigArg = MF.getFunction()->arg_begin();
2951 unsigned CurArgIdx = 0;
2953 // Initially ArgRegsSaveSize is zero.
2954 // Then we increase this value each time we meet byval parameter.
2955 // We also increase this value in case of varargs function.
2956 AFI->setArgRegsSaveSize(0);
2958 // Calculate the amount of stack space that we need to allocate to store
2959 // byval and variadic arguments that are passed in registers.
2960 // We need to know this before we allocate the first byval or variadic
2961 // argument, as they will be allocated a stack slot below the CFA (Canonical
2962 // Frame Address, the stack pointer at entry to the function).
2963 unsigned ArgRegBegin = ARM::R4;
2964 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2965 if (CCInfo.getInRegsParamsProcessed() >= CCInfo.getInRegsParamsCount())
2968 CCValAssign &VA = ArgLocs[i];
2969 unsigned Index = VA.getValNo();
2970 ISD::ArgFlagsTy Flags = Ins[Index].Flags;
2971 if (!Flags.isByVal())
2974 assert(VA.isMemLoc() && "unexpected byval pointer in reg");
2975 unsigned RBegin, REnd;
2976 CCInfo.getInRegsParamInfo(CCInfo.getInRegsParamsProcessed(), RBegin, REnd);
2977 ArgRegBegin = std::min(ArgRegBegin, RBegin);
2979 CCInfo.nextInRegsParam();
2981 CCInfo.rewindByValRegsInfo();
2983 int lastInsIndex = -1;
2984 if (isVarArg && MFI->hasVAStart()) {
2985 unsigned RegIdx = CCInfo.getFirstUnallocated(GPRArgRegs);
2986 if (RegIdx != array_lengthof(GPRArgRegs))
2987 ArgRegBegin = std::min(ArgRegBegin, (unsigned)GPRArgRegs[RegIdx]);
2990 unsigned TotalArgRegsSaveSize = 4 * (ARM::R4 - ArgRegBegin);
2991 AFI->setArgRegsSaveSize(TotalArgRegsSaveSize);
2993 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2994 CCValAssign &VA = ArgLocs[i];
2995 if (Ins[VA.getValNo()].isOrigArg()) {
2996 std::advance(CurOrigArg,
2997 Ins[VA.getValNo()].getOrigArgIndex() - CurArgIdx);
2998 CurArgIdx = Ins[VA.getValNo()].getOrigArgIndex();
3000 // Arguments stored in registers.
3001 if (VA.isRegLoc()) {
3002 EVT RegVT = VA.getLocVT();
3004 if (VA.needsCustom()) {
3005 // f64 and vector types are split up into multiple registers or
3006 // combinations of registers and stack slots.
3007 if (VA.getLocVT() == MVT::v2f64) {
3008 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
3010 VA = ArgLocs[++i]; // skip ahead to next loc
3012 if (VA.isMemLoc()) {
3013 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true);
3014 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
3015 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
3016 MachinePointerInfo::getFixedStack(FI),
3017 false, false, false, 0);
3019 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
3022 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
3023 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
3024 ArgValue, ArgValue1,
3025 DAG.getIntPtrConstant(0, dl));
3026 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
3027 ArgValue, ArgValue2,
3028 DAG.getIntPtrConstant(1, dl));
3030 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
3033 const TargetRegisterClass *RC;
3035 if (RegVT == MVT::f32)
3036 RC = &ARM::SPRRegClass;
3037 else if (RegVT == MVT::f64)
3038 RC = &ARM::DPRRegClass;
3039 else if (RegVT == MVT::v2f64)
3040 RC = &ARM::QPRRegClass;
3041 else if (RegVT == MVT::i32)
3042 RC = AFI->isThumb1OnlyFunction() ? &ARM::tGPRRegClass
3043 : &ARM::GPRRegClass;
3045 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
3047 // Transform the arguments in physical registers into virtual ones.
3048 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
3049 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
3052 // If this is an 8 or 16-bit value, it is really passed promoted
3053 // to 32 bits. Insert an assert[sz]ext to capture this, then
3054 // truncate to the right size.
3055 switch (VA.getLocInfo()) {
3056 default: llvm_unreachable("Unknown loc info!");
3057 case CCValAssign::Full: break;
3058 case CCValAssign::BCvt:
3059 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
3061 case CCValAssign::SExt:
3062 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
3063 DAG.getValueType(VA.getValVT()));
3064 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
3066 case CCValAssign::ZExt:
3067 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
3068 DAG.getValueType(VA.getValVT()));
3069 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
3073 InVals.push_back(ArgValue);
3075 } else { // VA.isRegLoc()
3078 assert(VA.isMemLoc());
3079 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
3081 int index = VA.getValNo();
3083 // Some Ins[] entries become multiple ArgLoc[] entries.
3084 // Process them only once.
3085 if (index != lastInsIndex)
3087 ISD::ArgFlagsTy Flags = Ins[index].Flags;
3088 // FIXME: For now, all byval parameter objects are marked mutable.
3089 // This can be changed with more analysis.
3090 // In case of tail call optimization mark all arguments mutable.
3091 // Since they could be overwritten by lowering of arguments in case of
3093 if (Flags.isByVal()) {
3094 assert(Ins[index].isOrigArg() &&
3095 "Byval arguments cannot be implicit");
3096 unsigned CurByValIndex = CCInfo.getInRegsParamsProcessed();
3098 int FrameIndex = StoreByValRegs(CCInfo, DAG, dl, Chain, CurOrigArg,
3099 CurByValIndex, VA.getLocMemOffset(),
3100 Flags.getByValSize());
3101 InVals.push_back(DAG.getFrameIndex(FrameIndex, getPointerTy()));
3102 CCInfo.nextInRegsParam();
3104 unsigned FIOffset = VA.getLocMemOffset();
3105 int FI = MFI->CreateFixedObject(VA.getLocVT().getSizeInBits()/8,
3108 // Create load nodes to retrieve arguments from the stack.
3109 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
3110 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
3111 MachinePointerInfo::getFixedStack(FI),
3112 false, false, false, 0));
3114 lastInsIndex = index;
3120 if (isVarArg && MFI->hasVAStart())
3121 VarArgStyleRegisters(CCInfo, DAG, dl, Chain,
3122 CCInfo.getNextStackOffset(),
3123 TotalArgRegsSaveSize);
3125 AFI->setArgumentStackSize(CCInfo.getNextStackOffset());
3130 /// isFloatingPointZero - Return true if this is +0.0.
3131 static bool isFloatingPointZero(SDValue Op) {
3132 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
3133 return CFP->getValueAPF().isPosZero();
3134 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
3135 // Maybe this has already been legalized into the constant pool?
3136 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
3137 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
3138 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
3139 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
3140 return CFP->getValueAPF().isPosZero();
3142 } else if (Op->getOpcode() == ISD::BITCAST &&
3143 Op->getValueType(0) == MVT::f64) {
3144 // Handle (ISD::BITCAST (ARMISD::VMOVIMM (ISD::TargetConstant 0)) MVT::f64)
3145 // created by LowerConstantFP().
3146 SDValue BitcastOp = Op->getOperand(0);
3147 if (BitcastOp->getOpcode() == ARMISD::VMOVIMM) {
3148 SDValue MoveOp = BitcastOp->getOperand(0);
3149 if (MoveOp->getOpcode() == ISD::TargetConstant &&
3150 cast<ConstantSDNode>(MoveOp)->getZExtValue() == 0) {
3158 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
3159 /// the given operands.
3161 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
3162 SDValue &ARMcc, SelectionDAG &DAG,
3164 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
3165 unsigned C = RHSC->getZExtValue();
3166 if (!isLegalICmpImmediate(C)) {
3167 // Constant does not fit, try adjusting it by one?
3172 if (C != 0x80000000 && isLegalICmpImmediate(C-1)) {
3173 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
3174 RHS = DAG.getConstant(C - 1, dl, MVT::i32);
3179 if (C != 0 && isLegalICmpImmediate(C-1)) {
3180 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
3181 RHS = DAG.getConstant(C - 1, dl, MVT::i32);
3186 if (C != 0x7fffffff && isLegalICmpImmediate(C+1)) {
3187 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
3188 RHS = DAG.getConstant(C + 1, dl, MVT::i32);
3193 if (C != 0xffffffff && isLegalICmpImmediate(C+1)) {
3194 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
3195 RHS = DAG.getConstant(C + 1, dl, MVT::i32);
3202 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3203 ARMISD::NodeType CompareType;
3206 CompareType = ARMISD::CMP;
3211 CompareType = ARMISD::CMPZ;
3214 ARMcc = DAG.getConstant(CondCode, dl, MVT::i32);
3215 return DAG.getNode(CompareType, dl, MVT::Glue, LHS, RHS);
3218 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
3220 ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
3222 assert(!Subtarget->isFPOnlySP() || RHS.getValueType() != MVT::f64);
3224 if (!isFloatingPointZero(RHS))
3225 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Glue, LHS, RHS);
3227 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Glue, LHS);
3228 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Glue, Cmp);
3231 /// duplicateCmp - Glue values can have only one use, so this function
3232 /// duplicates a comparison node.
3234 ARMTargetLowering::duplicateCmp(SDValue Cmp, SelectionDAG &DAG) const {
3235 unsigned Opc = Cmp.getOpcode();
3237 if (Opc == ARMISD::CMP || Opc == ARMISD::CMPZ)
3238 return DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
3240 assert(Opc == ARMISD::FMSTAT && "unexpected comparison operation");
3241 Cmp = Cmp.getOperand(0);
3242 Opc = Cmp.getOpcode();
3243 if (Opc == ARMISD::CMPFP)
3244 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
3246 assert(Opc == ARMISD::CMPFPw0 && "unexpected operand of FMSTAT");
3247 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0));
3249 return DAG.getNode(ARMISD::FMSTAT, DL, MVT::Glue, Cmp);
3252 std::pair<SDValue, SDValue>
3253 ARMTargetLowering::getARMXALUOOp(SDValue Op, SelectionDAG &DAG,
3254 SDValue &ARMcc) const {
3255 assert(Op.getValueType() == MVT::i32 && "Unsupported value type");
3257 SDValue Value, OverflowCmp;
3258 SDValue LHS = Op.getOperand(0);
3259 SDValue RHS = Op.getOperand(1);
3262 // FIXME: We are currently always generating CMPs because we don't support
3263 // generating CMN through the backend. This is not as good as the natural
3264 // CMP case because it causes a register dependency and cannot be folded
3267 switch (Op.getOpcode()) {
3269 llvm_unreachable("Unknown overflow instruction!");
3271 ARMcc = DAG.getConstant(ARMCC::VC, dl, MVT::i32);
3272 Value = DAG.getNode(ISD::ADD, dl, Op.getValueType(), LHS, RHS);
3273 OverflowCmp = DAG.getNode(ARMISD::CMP, dl, MVT::Glue, Value, LHS);
3276 ARMcc = DAG.getConstant(ARMCC::HS, dl, MVT::i32);
3277 Value = DAG.getNode(ISD::ADD, dl, Op.getValueType(), LHS, RHS);
3278 OverflowCmp = DAG.getNode(ARMISD::CMP, dl, MVT::Glue, Value, LHS);
3281 ARMcc = DAG.getConstant(ARMCC::VC, dl, MVT::i32);
3282 Value = DAG.getNode(ISD::SUB, dl, Op.getValueType(), LHS, RHS);
3283 OverflowCmp = DAG.getNode(ARMISD::CMP, dl, MVT::Glue, LHS, RHS);
3286 ARMcc = DAG.getConstant(ARMCC::HS, dl, MVT::i32);
3287 Value = DAG.getNode(ISD::SUB, dl, Op.getValueType(), LHS, RHS);
3288 OverflowCmp = DAG.getNode(ARMISD::CMP, dl, MVT::Glue, LHS, RHS);
3292 return std::make_pair(Value, OverflowCmp);
3297 ARMTargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) const {
3298 // Let legalize expand this if it isn't a legal type yet.
3299 if (!DAG.getTargetLoweringInfo().isTypeLegal(Op.getValueType()))
3302 SDValue Value, OverflowCmp;
3304 std::tie(Value, OverflowCmp) = getARMXALUOOp(Op, DAG, ARMcc);
3305 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3307 // We use 0 and 1 as false and true values.
3308 SDValue TVal = DAG.getConstant(1, dl, MVT::i32);
3309 SDValue FVal = DAG.getConstant(0, dl, MVT::i32);
3310 EVT VT = Op.getValueType();
3312 SDValue Overflow = DAG.getNode(ARMISD::CMOV, dl, VT, TVal, FVal,
3313 ARMcc, CCR, OverflowCmp);
3315 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
3316 return DAG.getNode(ISD::MERGE_VALUES, dl, VTs, Value, Overflow);
3320 SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
3321 SDValue Cond = Op.getOperand(0);
3322 SDValue SelectTrue = Op.getOperand(1);
3323 SDValue SelectFalse = Op.getOperand(2);
3325 unsigned Opc = Cond.getOpcode();
3327 if (Cond.getResNo() == 1 &&
3328 (Opc == ISD::SADDO || Opc == ISD::UADDO || Opc == ISD::SSUBO ||
3329 Opc == ISD::USUBO)) {
3330 if (!DAG.getTargetLoweringInfo().isTypeLegal(Cond->getValueType(0)))
3333 SDValue Value, OverflowCmp;
3335 std::tie(Value, OverflowCmp) = getARMXALUOOp(Cond, DAG, ARMcc);
3336 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3337 EVT VT = Op.getValueType();
3339 return getCMOV(dl, VT, SelectTrue, SelectFalse, ARMcc, CCR,
3345 // (select (cmov 1, 0, cond), t, f) -> (cmov t, f, cond)
3346 // (select (cmov 0, 1, cond), t, f) -> (cmov f, t, cond)
3348 if (Cond.getOpcode() == ARMISD::CMOV && Cond.hasOneUse()) {
3349 const ConstantSDNode *CMOVTrue =
3350 dyn_cast<ConstantSDNode>(Cond.getOperand(0));
3351 const ConstantSDNode *CMOVFalse =
3352 dyn_cast<ConstantSDNode>(Cond.getOperand(1));
3354 if (CMOVTrue && CMOVFalse) {
3355 unsigned CMOVTrueVal = CMOVTrue->getZExtValue();
3356 unsigned CMOVFalseVal = CMOVFalse->getZExtValue();
3360 if (CMOVTrueVal == 1 && CMOVFalseVal == 0) {
3362 False = SelectFalse;
3363 } else if (CMOVTrueVal == 0 && CMOVFalseVal == 1) {
3368 if (True.getNode() && False.getNode()) {
3369 EVT VT = Op.getValueType();
3370 SDValue ARMcc = Cond.getOperand(2);
3371 SDValue CCR = Cond.getOperand(3);
3372 SDValue Cmp = duplicateCmp(Cond.getOperand(4), DAG);
3373 assert(True.getValueType() == VT);
3374 return getCMOV(dl, VT, True, False, ARMcc, CCR, Cmp, DAG);
3379 // ARM's BooleanContents value is UndefinedBooleanContent. Mask out the
3380 // undefined bits before doing a full-word comparison with zero.
3381 Cond = DAG.getNode(ISD::AND, dl, Cond.getValueType(), Cond,
3382 DAG.getConstant(1, dl, Cond.getValueType()));
3384 return DAG.getSelectCC(dl, Cond,
3385 DAG.getConstant(0, dl, Cond.getValueType()),
3386 SelectTrue, SelectFalse, ISD::SETNE);
3389 static void checkVSELConstraints(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
3390 bool &swpCmpOps, bool &swpVselOps) {
3391 // Start by selecting the GE condition code for opcodes that return true for
3393 if (CC == ISD::SETUGE || CC == ISD::SETOGE || CC == ISD::SETOLE ||
3395 CondCode = ARMCC::GE;
3397 // and GT for opcodes that return false for 'equality'.
3398 else if (CC == ISD::SETUGT || CC == ISD::SETOGT || CC == ISD::SETOLT ||
3400 CondCode = ARMCC::GT;
3402 // Since we are constrained to GE/GT, if the opcode contains 'less', we need
3403 // to swap the compare operands.
3404 if (CC == ISD::SETOLE || CC == ISD::SETULE || CC == ISD::SETOLT ||
3408 // Both GT and GE are ordered comparisons, and return false for 'unordered'.
3409 // If we have an unordered opcode, we need to swap the operands to the VSEL
3410 // instruction (effectively negating the condition).
3412 // This also has the effect of swapping which one of 'less' or 'greater'
3413 // returns true, so we also swap the compare operands. It also switches
3414 // whether we return true for 'equality', so we compensate by picking the
3415 // opposite condition code to our original choice.
3416 if (CC == ISD::SETULE || CC == ISD::SETULT || CC == ISD::SETUGE ||
3417 CC == ISD::SETUGT) {
3418 swpCmpOps = !swpCmpOps;
3419 swpVselOps = !swpVselOps;
3420 CondCode = CondCode == ARMCC::GT ? ARMCC::GE : ARMCC::GT;
3423 // 'ordered' is 'anything but unordered', so use the VS condition code and
3424 // swap the VSEL operands.
3425 if (CC == ISD::SETO) {
3426 CondCode = ARMCC::VS;
3430 // 'unordered or not equal' is 'anything but equal', so use the EQ condition
3431 // code and swap the VSEL operands.
3432 if (CC == ISD::SETUNE) {
3433 CondCode = ARMCC::EQ;
3438 SDValue ARMTargetLowering::getCMOV(SDLoc dl, EVT VT, SDValue FalseVal,
3439 SDValue TrueVal, SDValue ARMcc, SDValue CCR,
3440 SDValue Cmp, SelectionDAG &DAG) const {
3441 if (Subtarget->isFPOnlySP() && VT == MVT::f64) {
3442 FalseVal = DAG.getNode(ARMISD::VMOVRRD, dl,
3443 DAG.getVTList(MVT::i32, MVT::i32), FalseVal);
3444 TrueVal = DAG.getNode(ARMISD::VMOVRRD, dl,
3445 DAG.getVTList(MVT::i32, MVT::i32), TrueVal);
3447 SDValue TrueLow = TrueVal.getValue(0);
3448 SDValue TrueHigh = TrueVal.getValue(1);
3449 SDValue FalseLow = FalseVal.getValue(0);
3450 SDValue FalseHigh = FalseVal.getValue(1);
3452 SDValue Low = DAG.getNode(ARMISD::CMOV, dl, MVT::i32, FalseLow, TrueLow,
3454 SDValue High = DAG.getNode(ARMISD::CMOV, dl, MVT::i32, FalseHigh, TrueHigh,
3455 ARMcc, CCR, duplicateCmp(Cmp, DAG));
3457 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Low, High);
3459 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,
3464 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
3465 EVT VT = Op.getValueType();
3466 SDValue LHS = Op.getOperand(0);
3467 SDValue RHS = Op.getOperand(1);
3468 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
3469 SDValue TrueVal = Op.getOperand(2);
3470 SDValue FalseVal = Op.getOperand(3);
3473 if (Subtarget->isFPOnlySP() && LHS.getValueType() == MVT::f64) {
3474 DAG.getTargetLoweringInfo().softenSetCCOperands(DAG, MVT::f64, LHS, RHS, CC,
3477 // If softenSetCCOperands only returned one value, we should compare it to
3479 if (!RHS.getNode()) {
3480 RHS = DAG.getConstant(0, dl, LHS.getValueType());
3485 if (LHS.getValueType() == MVT::i32) {
3486 // Try to generate VSEL on ARMv8.
3487 // The VSEL instruction can't use all the usual ARM condition
3488 // codes: it only has two bits to select the condition code, so it's
3489 // constrained to use only GE, GT, VS and EQ.
3491 // To implement all the various ISD::SETXXX opcodes, we sometimes need to
3492 // swap the operands of the previous compare instruction (effectively
3493 // inverting the compare condition, swapping 'less' and 'greater') and
3494 // sometimes need to swap the operands to the VSEL (which inverts the
3495 // condition in the sense of firing whenever the previous condition didn't)
3496 if (Subtarget->hasFPARMv8() && (TrueVal.getValueType() == MVT::f32 ||
3497 TrueVal.getValueType() == MVT::f64)) {
3498 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3499 if (CondCode == ARMCC::LT || CondCode == ARMCC::LE ||
3500 CondCode == ARMCC::VC || CondCode == ARMCC::NE) {
3501 CC = ISD::getSetCCInverse(CC, true);
3502 std::swap(TrueVal, FalseVal);
3507 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3508 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3509 return getCMOV(dl, VT, FalseVal, TrueVal, ARMcc, CCR, Cmp, DAG);
3512 ARMCC::CondCodes CondCode, CondCode2;
3513 FPCCToARMCC(CC, CondCode, CondCode2);
3515 // Try to generate VMAXNM/VMINNM on ARMv8.
3516 if (Subtarget->hasFPARMv8() && (TrueVal.getValueType() == MVT::f32 ||
3517 TrueVal.getValueType() == MVT::f64)) {
3518 // We can use VMAXNM/VMINNM for a compare followed by a select with the
3519 // same operands, as follows:
3520 // c = fcmp [?gt, ?ge, ?lt, ?le] a, b
3522 // In NoNaNsFPMath the CC will have been changed from, e.g., 'ogt' to 'gt'.
3523 bool swapSides = false;
3524 if (!getTargetMachine().Options.NoNaNsFPMath) {
3525 // transformability may depend on which way around we compare
3533 // the non-NaN should be RHS
3534 swapSides = DAG.isKnownNeverNaN(LHS) && !DAG.isKnownNeverNaN(RHS);
3540 // the non-NaN should be LHS
3541 swapSides = DAG.isKnownNeverNaN(RHS) && !DAG.isKnownNeverNaN(LHS);
3545 swapSides = swapSides || (LHS == FalseVal && RHS == TrueVal);
3547 CC = ISD::getSetCCSwappedOperands(CC);
3548 std::swap(LHS, RHS);
3550 if (LHS == TrueVal && RHS == FalseVal) {
3551 bool canTransform = true;
3552 // FIXME: FastMathFlags::noSignedZeros() doesn't appear reachable from here
3553 if (!getTargetMachine().Options.UnsafeFPMath &&
3554 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
3555 const ConstantFPSDNode *Zero;
3562 // RHS must not be -0
3563 canTransform = (Zero = dyn_cast<ConstantFPSDNode>(RHS)) &&
3564 !Zero->isNegative();
3569 // LHS must not be -0
3570 canTransform = (Zero = dyn_cast<ConstantFPSDNode>(LHS)) &&
3571 !Zero->isNegative();
3576 // RHS must not be +0
3577 canTransform = (Zero = dyn_cast<ConstantFPSDNode>(RHS)) &&
3583 // LHS must not be +0
3584 canTransform = (Zero = dyn_cast<ConstantFPSDNode>(LHS)) &&
3590 // Note: If one of the elements in a pair is a number and the other
3591 // element is NaN, the corresponding result element is the number.
3592 // This is consistent with the IEEE 754-2008 standard.
3593 // Therefore, a > b ? a : b <=> vmax(a,b), if b is constant and a is NaN
3599 if (!DAG.isKnownNeverNaN(RHS))
3601 return DAG.getNode(ARMISD::VMAXNM, dl, VT, LHS, RHS);
3604 if (!DAG.isKnownNeverNaN(LHS))
3608 return DAG.getNode(ARMISD::VMAXNM, dl, VT, LHS, RHS);
3611 if (!DAG.isKnownNeverNaN(RHS))
3613 return DAG.getNode(ARMISD::VMINNM, dl, VT, LHS, RHS);
3616 if (!DAG.isKnownNeverNaN(LHS))
3620 return DAG.getNode(ARMISD::VMINNM, dl, VT, LHS, RHS);
3625 bool swpCmpOps = false;
3626 bool swpVselOps = false;
3627 checkVSELConstraints(CC, CondCode, swpCmpOps, swpVselOps);
3629 if (CondCode == ARMCC::GT || CondCode == ARMCC::GE ||
3630 CondCode == ARMCC::VS || CondCode == ARMCC::EQ) {
3632 std::swap(LHS, RHS);
3634 std::swap(TrueVal, FalseVal);
3638 SDValue ARMcc = DAG.getConstant(CondCode, dl, MVT::i32);
3639 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
3640 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3641 SDValue Result = getCMOV(dl, VT, FalseVal, TrueVal, ARMcc, CCR, Cmp, DAG);
3642 if (CondCode2 != ARMCC::AL) {
3643 SDValue ARMcc2 = DAG.getConstant(CondCode2, dl, MVT::i32);
3644 // FIXME: Needs another CMP because flag can have but one use.
3645 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
3646 Result = getCMOV(dl, VT, Result, TrueVal, ARMcc2, CCR, Cmp2, DAG);
3651 /// canChangeToInt - Given the fp compare operand, return true if it is suitable
3652 /// to morph to an integer compare sequence.
3653 static bool canChangeToInt(SDValue Op, bool &SeenZero,
3654 const ARMSubtarget *Subtarget) {
3655 SDNode *N = Op.getNode();
3656 if (!N->hasOneUse())
3657 // Otherwise it requires moving the value from fp to integer registers.
3659 if (!N->getNumValues())
3661 EVT VT = Op.getValueType();
3662 if (VT != MVT::f32 && !Subtarget->isFPBrccSlow())
3663 // f32 case is generally profitable. f64 case only makes sense when vcmpe +
3664 // vmrs are very slow, e.g. cortex-a8.
3667 if (isFloatingPointZero(Op)) {
3671 return ISD::isNormalLoad(N);
3674 static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) {
3675 if (isFloatingPointZero(Op))
3676 return DAG.getConstant(0, SDLoc(Op), MVT::i32);
3678 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op))
3679 return DAG.getLoad(MVT::i32, SDLoc(Op),
3680 Ld->getChain(), Ld->getBasePtr(), Ld->getPointerInfo(),
3681 Ld->isVolatile(), Ld->isNonTemporal(),
3682 Ld->isInvariant(), Ld->getAlignment());
3684 llvm_unreachable("Unknown VFP cmp argument!");
3687 static void expandf64Toi32(SDValue Op, SelectionDAG &DAG,
3688 SDValue &RetVal1, SDValue &RetVal2) {
3691 if (isFloatingPointZero(Op)) {
3692 RetVal1 = DAG.getConstant(0, dl, MVT::i32);
3693 RetVal2 = DAG.getConstant(0, dl, MVT::i32);
3697 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) {
3698 SDValue Ptr = Ld->getBasePtr();
3699 RetVal1 = DAG.getLoad(MVT::i32, dl,
3700 Ld->getChain(), Ptr,
3701 Ld->getPointerInfo(),
3702 Ld->isVolatile(), Ld->isNonTemporal(),
3703 Ld->isInvariant(), Ld->getAlignment());
3705 EVT PtrType = Ptr.getValueType();
3706 unsigned NewAlign = MinAlign(Ld->getAlignment(), 4);
3707 SDValue NewPtr = DAG.getNode(ISD::ADD, dl,
3708 PtrType, Ptr, DAG.getConstant(4, dl, PtrType));
3709 RetVal2 = DAG.getLoad(MVT::i32, dl,
3710 Ld->getChain(), NewPtr,
3711 Ld->getPointerInfo().getWithOffset(4),
3712 Ld->isVolatile(), Ld->isNonTemporal(),
3713 Ld->isInvariant(), NewAlign);
3717 llvm_unreachable("Unknown VFP cmp argument!");
3720 /// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some
3721 /// f32 and even f64 comparisons to integer ones.
3723 ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const {
3724 SDValue Chain = Op.getOperand(0);
3725 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3726 SDValue LHS = Op.getOperand(2);
3727 SDValue RHS = Op.getOperand(3);
3728 SDValue Dest = Op.getOperand(4);
3731 bool LHSSeenZero = false;
3732 bool LHSOk = canChangeToInt(LHS, LHSSeenZero, Subtarget);
3733 bool RHSSeenZero = false;
3734 bool RHSOk = canChangeToInt(RHS, RHSSeenZero, Subtarget);
3735 if (LHSOk && RHSOk && (LHSSeenZero || RHSSeenZero)) {
3736 // If unsafe fp math optimization is enabled and there are no other uses of
3737 // the CMP operands, and the condition code is EQ or NE, we can optimize it
3738 // to an integer comparison.
3739 if (CC == ISD::SETOEQ)
3741 else if (CC == ISD::SETUNE)
3744 SDValue Mask = DAG.getConstant(0x7fffffff, dl, MVT::i32);
3746 if (LHS.getValueType() == MVT::f32) {
3747 LHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3748 bitcastf32Toi32(LHS, DAG), Mask);
3749 RHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3750 bitcastf32Toi32(RHS, DAG), Mask);
3751 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3752 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3753 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3754 Chain, Dest, ARMcc, CCR, Cmp);
3759 expandf64Toi32(LHS, DAG, LHS1, LHS2);
3760 expandf64Toi32(RHS, DAG, RHS1, RHS2);
3761 LHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, LHS2, Mask);
3762 RHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, RHS2, Mask);
3763 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3764 ARMcc = DAG.getConstant(CondCode, dl, MVT::i32);
3765 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3766 SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest };
3767 return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops);
3773 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
3774 SDValue Chain = Op.getOperand(0);
3775 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3776 SDValue LHS = Op.getOperand(2);
3777 SDValue RHS = Op.getOperand(3);
3778 SDValue Dest = Op.getOperand(4);
3781 if (Subtarget->isFPOnlySP() && LHS.getValueType() == MVT::f64) {
3782 DAG.getTargetLoweringInfo().softenSetCCOperands(DAG, MVT::f64, LHS, RHS, CC,
3785 // If softenSetCCOperands only returned one value, we should compare it to
3787 if (!RHS.getNode()) {
3788 RHS = DAG.getConstant(0, dl, LHS.getValueType());
3793 if (LHS.getValueType() == MVT::i32) {
3795 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3796 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3797 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3798 Chain, Dest, ARMcc, CCR, Cmp);
3801 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3803 if (getTargetMachine().Options.UnsafeFPMath &&
3804 (CC == ISD::SETEQ || CC == ISD::SETOEQ ||
3805 CC == ISD::SETNE || CC == ISD::SETUNE)) {
3806 SDValue Result = OptimizeVFPBrcond(Op, DAG);
3807 if (Result.getNode())
3811 ARMCC::CondCodes CondCode, CondCode2;
3812 FPCCToARMCC(CC, CondCode, CondCode2);
3814 SDValue ARMcc = DAG.getConstant(CondCode, dl, MVT::i32);
3815 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
3816 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3817 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3818 SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp };
3819 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops);
3820 if (CondCode2 != ARMCC::AL) {
3821 ARMcc = DAG.getConstant(CondCode2, dl, MVT::i32);
3822 SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) };
3823 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops);
3828 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
3829 SDValue Chain = Op.getOperand(0);
3830 SDValue Table = Op.getOperand(1);
3831 SDValue Index = Op.getOperand(2);
3834 EVT PTy = getPointerTy();
3835 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
3836 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
3837 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI);
3838 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, dl, PTy));
3839 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
3840 if (Subtarget->isThumb2()) {
3841 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
3842 // which does another jump to the destination. This also makes it easier
3843 // to translate it to TBB / TBH later.
3844 // FIXME: This might not work if the function is extremely large.
3845 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
3846 Addr, Op.getOperand(2), JTI);
3848 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
3849 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
3850 MachinePointerInfo::getJumpTable(),
3851 false, false, false, 0);
3852 Chain = Addr.getValue(1);
3853 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
3854 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI);
3856 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
3857 MachinePointerInfo::getJumpTable(),
3858 false, false, false, 0);
3859 Chain = Addr.getValue(1);
3860 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI);
3864 static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3865 EVT VT = Op.getValueType();
3868 if (Op.getValueType().getVectorElementType() == MVT::i32) {
3869 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::f32)
3871 return DAG.UnrollVectorOp(Op.getNode());
3874 assert(Op.getOperand(0).getValueType() == MVT::v4f32 &&
3875 "Invalid type for custom lowering!");
3876 if (VT != MVT::v4i16)
3877 return DAG.UnrollVectorOp(Op.getNode());
3879 Op = DAG.getNode(Op.getOpcode(), dl, MVT::v4i32, Op.getOperand(0));
3880 return DAG.getNode(ISD::TRUNCATE, dl, VT, Op);
3883 SDValue ARMTargetLowering::LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) const {
3884 EVT VT = Op.getValueType();
3886 return LowerVectorFP_TO_INT(Op, DAG);
3887 if (Subtarget->isFPOnlySP() && Op.getOperand(0).getValueType() == MVT::f64) {
3889 if (Op.getOpcode() == ISD::FP_TO_SINT)
3890 LC = RTLIB::getFPTOSINT(Op.getOperand(0).getValueType(),
3893 LC = RTLIB::getFPTOUINT(Op.getOperand(0).getValueType(),
3895 return makeLibCall(DAG, LC, Op.getValueType(), &Op.getOperand(0), 1,
3896 /*isSigned*/ false, SDLoc(Op)).first;
3902 static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3903 EVT VT = Op.getValueType();
3906 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::i32) {
3907 if (VT.getVectorElementType() == MVT::f32)
3909 return DAG.UnrollVectorOp(Op.getNode());
3912 assert(Op.getOperand(0).getValueType() == MVT::v4i16 &&
3913 "Invalid type for custom lowering!");
3914 if (VT != MVT::v4f32)
3915 return DAG.UnrollVectorOp(Op.getNode());
3919 switch (Op.getOpcode()) {
3920 default: llvm_unreachable("Invalid opcode!");
3921 case ISD::SINT_TO_FP:
3922 CastOpc = ISD::SIGN_EXTEND;
3923 Opc = ISD::SINT_TO_FP;
3925 case ISD::UINT_TO_FP:
3926 CastOpc = ISD::ZERO_EXTEND;
3927 Opc = ISD::UINT_TO_FP;
3931 Op = DAG.getNode(CastOpc, dl, MVT::v4i32, Op.getOperand(0));
3932 return DAG.getNode(Opc, dl, VT, Op);
3935 SDValue ARMTargetLowering::LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) const {
3936 EVT VT = Op.getValueType();
3938 return LowerVectorINT_TO_FP(Op, DAG);
3939 if (Subtarget->isFPOnlySP() && Op.getValueType() == MVT::f64) {
3941 if (Op.getOpcode() == ISD::SINT_TO_FP)
3942 LC = RTLIB::getSINTTOFP(Op.getOperand(0).getValueType(),
3945 LC = RTLIB::getUINTTOFP(Op.getOperand(0).getValueType(),
3947 return makeLibCall(DAG, LC, Op.getValueType(), &Op.getOperand(0), 1,
3948 /*isSigned*/ false, SDLoc(Op)).first;
3954 SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
3955 // Implement fcopysign with a fabs and a conditional fneg.
3956 SDValue Tmp0 = Op.getOperand(0);
3957 SDValue Tmp1 = Op.getOperand(1);
3959 EVT VT = Op.getValueType();
3960 EVT SrcVT = Tmp1.getValueType();
3961 bool InGPR = Tmp0.getOpcode() == ISD::BITCAST ||
3962 Tmp0.getOpcode() == ARMISD::VMOVDRR;
3963 bool UseNEON = !InGPR && Subtarget->hasNEON();
3966 // Use VBSL to copy the sign bit.
3967 unsigned EncodedVal = ARM_AM::createNEONModImm(0x6, 0x80);
3968 SDValue Mask = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v2i32,
3969 DAG.getTargetConstant(EncodedVal, dl, MVT::i32));
3970 EVT OpVT = (VT == MVT::f32) ? MVT::v2i32 : MVT::v1i64;
3972 Mask = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3973 DAG.getNode(ISD::BITCAST, dl, OpVT, Mask),
3974 DAG.getConstant(32, dl, MVT::i32));
3975 else /*if (VT == MVT::f32)*/
3976 Tmp0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp0);
3977 if (SrcVT == MVT::f32) {
3978 Tmp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp1);
3980 Tmp1 = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3981 DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1),
3982 DAG.getConstant(32, dl, MVT::i32));
3983 } else if (VT == MVT::f32)
3984 Tmp1 = DAG.getNode(ARMISD::VSHRu, dl, MVT::v1i64,
3985 DAG.getNode(ISD::BITCAST, dl, MVT::v1i64, Tmp1),
3986 DAG.getConstant(32, dl, MVT::i32));
3987 Tmp0 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp0);
3988 Tmp1 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1);
3990 SDValue AllOnes = DAG.getTargetConstant(ARM_AM::createNEONModImm(0xe, 0xff),
3992 AllOnes = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v8i8, AllOnes);
3993 SDValue MaskNot = DAG.getNode(ISD::XOR, dl, OpVT, Mask,
3994 DAG.getNode(ISD::BITCAST, dl, OpVT, AllOnes));
3996 SDValue Res = DAG.getNode(ISD::OR, dl, OpVT,
3997 DAG.getNode(ISD::AND, dl, OpVT, Tmp1, Mask),
3998 DAG.getNode(ISD::AND, dl, OpVT, Tmp0, MaskNot));
3999 if (VT == MVT::f32) {
4000 Res = DAG.getNode(ISD::BITCAST, dl, MVT::v2f32, Res);
4001 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, Res,
4002 DAG.getConstant(0, dl, MVT::i32));
4004 Res = DAG.getNode(ISD::BITCAST, dl, MVT::f64, Res);
4010 // Bitcast operand 1 to i32.
4011 if (SrcVT == MVT::f64)
4012 Tmp1 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
4014 Tmp1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp1);
4016 // Or in the signbit with integer operations.
4017 SDValue Mask1 = DAG.getConstant(0x80000000, dl, MVT::i32);
4018 SDValue Mask2 = DAG.getConstant(0x7fffffff, dl, MVT::i32);
4019 Tmp1 = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp1, Mask1);
4020 if (VT == MVT::f32) {
4021 Tmp0 = DAG.getNode(ISD::AND, dl, MVT::i32,
4022 DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp0), Mask2);
4023 return DAG.getNode(ISD::BITCAST, dl, MVT::f32,
4024 DAG.getNode(ISD::OR, dl, MVT::i32, Tmp0, Tmp1));
4027 // f64: Or the high part with signbit and then combine two parts.
4028 Tmp0 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
4030 SDValue Lo = Tmp0.getValue(0);
4031 SDValue Hi = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp0.getValue(1), Mask2);
4032 Hi = DAG.getNode(ISD::OR, dl, MVT::i32, Hi, Tmp1);
4033 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
4036 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
4037 MachineFunction &MF = DAG.getMachineFunction();
4038 MachineFrameInfo *MFI = MF.getFrameInfo();
4039 MFI->setReturnAddressIsTaken(true);
4041 if (verifyReturnAddressArgumentIsConstant(Op, DAG))
4044 EVT VT = Op.getValueType();
4046 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
4048 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
4049 SDValue Offset = DAG.getConstant(4, dl, MVT::i32);
4050 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
4051 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
4052 MachinePointerInfo(), false, false, false, 0);
4055 // Return LR, which contains the return address. Mark it an implicit live-in.
4056 unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32));
4057 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
4060 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
4061 const ARMBaseRegisterInfo &ARI =
4062 *static_cast<const ARMBaseRegisterInfo*>(RegInfo);
4063 MachineFunction &MF = DAG.getMachineFunction();
4064 MachineFrameInfo *MFI = MF.getFrameInfo();
4065 MFI->setFrameAddressIsTaken(true);
4067 EVT VT = Op.getValueType();
4068 SDLoc dl(Op); // FIXME probably not meaningful
4069 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
4070 unsigned FrameReg = ARI.getFrameRegister(MF);
4071 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
4073 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
4074 MachinePointerInfo(),
4075 false, false, false, 0);
4079 // FIXME? Maybe this could be a TableGen attribute on some registers and
4080 // this table could be generated automatically from RegInfo.
4081 unsigned ARMTargetLowering::getRegisterByName(const char* RegName,
4083 unsigned Reg = StringSwitch<unsigned>(RegName)
4084 .Case("sp", ARM::SP)
4088 report_fatal_error("Invalid register name global variable");
4091 /// ExpandBITCAST - If the target supports VFP, this function is called to
4092 /// expand a bit convert where either the source or destination type is i64 to
4093 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
4094 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
4095 /// vectors), since the legalizer won't know what to do with that.
4096 static SDValue ExpandBITCAST(SDNode *N, SelectionDAG &DAG) {
4097 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4099 SDValue Op = N->getOperand(0);
4101 // This function is only supposed to be called for i64 types, either as the
4102 // source or destination of the bit convert.
4103 EVT SrcVT = Op.getValueType();
4104 EVT DstVT = N->getValueType(0);
4105 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
4106 "ExpandBITCAST called for non-i64 type");
4108 // Turn i64->f64 into VMOVDRR.
4109 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
4110 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
4111 DAG.getConstant(0, dl, MVT::i32));
4112 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
4113 DAG.getConstant(1, dl, MVT::i32));
4114 return DAG.getNode(ISD::BITCAST, dl, DstVT,
4115 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
4118 // Turn f64->i64 into VMOVRRD.
4119 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
4121 if (TLI.isBigEndian() && SrcVT.isVector() &&
4122 SrcVT.getVectorNumElements() > 1)
4123 Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
4124 DAG.getVTList(MVT::i32, MVT::i32),
4125 DAG.getNode(ARMISD::VREV64, dl, SrcVT, Op));
4127 Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
4128 DAG.getVTList(MVT::i32, MVT::i32), Op);
4129 // Merge the pieces into a single i64 value.
4130 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
4136 /// getZeroVector - Returns a vector of specified type with all zero elements.
4137 /// Zero vectors are used to represent vector negation and in those cases
4138 /// will be implemented with the NEON VNEG instruction. However, VNEG does
4139 /// not support i64 elements, so sometimes the zero vectors will need to be
4140 /// explicitly constructed. Regardless, use a canonical VMOV to create the
4142 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, SDLoc dl) {
4143 assert(VT.isVector() && "Expected a vector type");
4144 // The canonical modified immediate encoding of a zero vector is....0!
4145 SDValue EncodedVal = DAG.getTargetConstant(0, dl, MVT::i32);
4146 EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
4147 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal);
4148 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4151 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
4152 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
4153 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
4154 SelectionDAG &DAG) const {
4155 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4156 EVT VT = Op.getValueType();
4157 unsigned VTBits = VT.getSizeInBits();
4159 SDValue ShOpLo = Op.getOperand(0);
4160 SDValue ShOpHi = Op.getOperand(1);
4161 SDValue ShAmt = Op.getOperand(2);
4163 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
4165 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
4167 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
4168 DAG.getConstant(VTBits, dl, MVT::i32), ShAmt);
4169 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
4170 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
4171 DAG.getConstant(VTBits, dl, MVT::i32));
4172 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
4173 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
4174 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
4176 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
4177 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, dl, MVT::i32),
4178 ISD::SETGE, ARMcc, DAG, dl);
4179 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
4180 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc,
4183 SDValue Ops[2] = { Lo, Hi };
4184 return DAG.getMergeValues(Ops, dl);
4187 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
4188 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
4189 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
4190 SelectionDAG &DAG) const {
4191 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4192 EVT VT = Op.getValueType();
4193 unsigned VTBits = VT.getSizeInBits();
4195 SDValue ShOpLo = Op.getOperand(0);
4196 SDValue ShOpHi = Op.getOperand(1);
4197 SDValue ShAmt = Op.getOperand(2);
4200 assert(Op.getOpcode() == ISD::SHL_PARTS);
4201 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
4202 DAG.getConstant(VTBits, dl, MVT::i32), ShAmt);
4203 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
4204 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
4205 DAG.getConstant(VTBits, dl, MVT::i32));
4206 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
4207 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
4209 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
4210 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
4211 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, dl, MVT::i32),
4212 ISD::SETGE, ARMcc, DAG, dl);
4213 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
4214 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc,
4217 SDValue Ops[2] = { Lo, Hi };
4218 return DAG.getMergeValues(Ops, dl);
4221 SDValue ARMTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
4222 SelectionDAG &DAG) const {
4223 // The rounding mode is in bits 23:22 of the FPSCR.
4224 // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0
4225 // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3)
4226 // so that the shift + and get folded into a bitfield extract.
4228 SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::i32,
4229 DAG.getConstant(Intrinsic::arm_get_fpscr, dl,
4231 SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPSCR,
4232 DAG.getConstant(1U << 22, dl, MVT::i32));
4233 SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds,
4234 DAG.getConstant(22, dl, MVT::i32));
4235 return DAG.getNode(ISD::AND, dl, MVT::i32, RMODE,
4236 DAG.getConstant(3, dl, MVT::i32));
4239 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
4240 const ARMSubtarget *ST) {
4241 EVT VT = N->getValueType(0);
4244 if (!ST->hasV6T2Ops())
4247 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
4248 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
4251 /// getCTPOP16BitCounts - Returns a v8i8/v16i8 vector containing the bit-count
4252 /// for each 16-bit element from operand, repeated. The basic idea is to
4253 /// leverage vcnt to get the 8-bit counts, gather and add the results.
4255 /// Trace for v4i16:
4256 /// input = [v0 v1 v2 v3 ] (vi 16-bit element)
4257 /// cast: N0 = [w0 w1 w2 w3 w4 w5 w6 w7] (v0 = [w0 w1], wi 8-bit element)
4258 /// vcnt: N1 = [b0 b1 b2 b3 b4 b5 b6 b7] (bi = bit-count of 8-bit element wi)
4259 /// vrev: N2 = [b1 b0 b3 b2 b5 b4 b7 b6]
4260 /// [b0 b1 b2 b3 b4 b5 b6 b7]
4261 /// +[b1 b0 b3 b2 b5 b4 b7 b6]
4262 /// N3=N1+N2 = [k0 k0 k1 k1 k2 k2 k3 k3] (k0 = b0+b1 = bit-count of 16-bit v0,
4263 /// vuzp: = [k0 k1 k2 k3 k0 k1 k2 k3] each ki is 8-bits)
4264 static SDValue getCTPOP16BitCounts(SDNode *N, SelectionDAG &DAG) {
4265 EVT VT = N->getValueType(0);
4268 EVT VT8Bit = VT.is64BitVector() ? MVT::v8i8 : MVT::v16i8;
4269 SDValue N0 = DAG.getNode(ISD::BITCAST, DL, VT8Bit, N->getOperand(0));
4270 SDValue N1 = DAG.getNode(ISD::CTPOP, DL, VT8Bit, N0);
4271 SDValue N2 = DAG.getNode(ARMISD::VREV16, DL, VT8Bit, N1);
4272 SDValue N3 = DAG.getNode(ISD::ADD, DL, VT8Bit, N1, N2);
4273 return DAG.getNode(ARMISD::VUZP, DL, VT8Bit, N3, N3);
4276 /// lowerCTPOP16BitElements - Returns a v4i16/v8i16 vector containing the
4277 /// bit-count for each 16-bit element from the operand. We need slightly
4278 /// different sequencing for v4i16 and v8i16 to stay within NEON's available
4279 /// 64/128-bit registers.
4281 /// Trace for v4i16:
4282 /// input = [v0 v1 v2 v3 ] (vi 16-bit element)
4283 /// v8i8: BitCounts = [k0 k1 k2 k3 k0 k1 k2 k3 ] (ki is the bit-count of vi)
4284 /// v8i16:Extended = [k0 k1 k2 k3 k0 k1 k2 k3 ]
4285 /// v4i16:Extracted = [k0 k1 k2 k3 ]
4286 static SDValue lowerCTPOP16BitElements(SDNode *N, SelectionDAG &DAG) {
4287 EVT VT = N->getValueType(0);
4290 SDValue BitCounts = getCTPOP16BitCounts(N, DAG);
4291 if (VT.is64BitVector()) {
4292 SDValue Extended = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v8i16, BitCounts);
4293 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i16, Extended,
4294 DAG.getIntPtrConstant(0, DL));
4296 SDValue Extracted = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v8i8,
4297 BitCounts, DAG.getIntPtrConstant(0, DL));
4298 return DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v8i16, Extracted);
4302 /// lowerCTPOP32BitElements - Returns a v2i32/v4i32 vector containing the
4303 /// bit-count for each 32-bit element from the operand. The idea here is
4304 /// to split the vector into 16-bit elements, leverage the 16-bit count
4305 /// routine, and then combine the results.
4307 /// Trace for v2i32 (v4i32 similar with Extracted/Extended exchanged):
4308 /// input = [v0 v1 ] (vi: 32-bit elements)
4309 /// Bitcast = [w0 w1 w2 w3 ] (wi: 16-bit elements, v0 = [w0 w1])
4310 /// Counts16 = [k0 k1 k2 k3 ] (ki: 16-bit elements, bit-count of wi)
4311 /// vrev: N0 = [k1 k0 k3 k2 ]
4313 /// N1 =+[k1 k0 k3 k2 ]
4315 /// N2 =+[k1 k3 k0 k2 ]
4317 /// Extended =+[k1 k3 k0 k2 ]
4319 /// Extracted=+[k1 k3 ]
4321 static SDValue lowerCTPOP32BitElements(SDNode *N, SelectionDAG &DAG) {
4322 EVT VT = N->getValueType(0);
4325 EVT VT16Bit = VT.is64BitVector() ? MVT::v4i16 : MVT::v8i16;
4327 SDValue Bitcast = DAG.getNode(ISD::BITCAST, DL, VT16Bit, N->getOperand(0));
4328 SDValue Counts16 = lowerCTPOP16BitElements(Bitcast.getNode(), DAG);
4329 SDValue N0 = DAG.getNode(ARMISD::VREV32, DL, VT16Bit, Counts16);
4330 SDValue N1 = DAG.getNode(ISD::ADD, DL, VT16Bit, Counts16, N0);
4331 SDValue N2 = DAG.getNode(ARMISD::VUZP, DL, VT16Bit, N1, N1);
4333 if (VT.is64BitVector()) {
4334 SDValue Extended = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v4i32, N2);
4335 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i32, Extended,
4336 DAG.getIntPtrConstant(0, DL));
4338 SDValue Extracted = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i16, N2,
4339 DAG.getIntPtrConstant(0, DL));
4340 return DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v4i32, Extracted);
4344 static SDValue LowerCTPOP(SDNode *N, SelectionDAG &DAG,
4345 const ARMSubtarget *ST) {
4346 EVT VT = N->getValueType(0);
4348 assert(ST->hasNEON() && "Custom ctpop lowering requires NEON.");
4349 assert((VT == MVT::v2i32 || VT == MVT::v4i32 ||
4350 VT == MVT::v4i16 || VT == MVT::v8i16) &&
4351 "Unexpected type for custom ctpop lowering");
4353 if (VT.getVectorElementType() == MVT::i32)
4354 return lowerCTPOP32BitElements(N, DAG);
4356 return lowerCTPOP16BitElements(N, DAG);
4359 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
4360 const ARMSubtarget *ST) {
4361 EVT VT = N->getValueType(0);
4367 // Lower vector shifts on NEON to use VSHL.
4368 assert(ST->hasNEON() && "unexpected vector shift");
4370 // Left shifts translate directly to the vshiftu intrinsic.
4371 if (N->getOpcode() == ISD::SHL)
4372 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
4373 DAG.getConstant(Intrinsic::arm_neon_vshiftu, dl,
4375 N->getOperand(0), N->getOperand(1));
4377 assert((N->getOpcode() == ISD::SRA ||
4378 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
4380 // NEON uses the same intrinsics for both left and right shifts. For
4381 // right shifts, the shift amounts are negative, so negate the vector of
4383 EVT ShiftVT = N->getOperand(1).getValueType();
4384 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
4385 getZeroVector(ShiftVT, DAG, dl),
4387 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
4388 Intrinsic::arm_neon_vshifts :
4389 Intrinsic::arm_neon_vshiftu);
4390 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
4391 DAG.getConstant(vshiftInt, dl, MVT::i32),
4392 N->getOperand(0), NegatedCount);
4395 static SDValue Expand64BitShift(SDNode *N, SelectionDAG &DAG,
4396 const ARMSubtarget *ST) {
4397 EVT VT = N->getValueType(0);
4400 // We can get here for a node like i32 = ISD::SHL i32, i64
4404 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
4405 "Unknown shift to lower!");
4407 // We only lower SRA, SRL of 1 here, all others use generic lowering.
4408 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
4409 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
4412 // If we are in thumb mode, we don't have RRX.
4413 if (ST->isThumb1Only()) return SDValue();
4415 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
4416 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
4417 DAG.getConstant(0, dl, MVT::i32));
4418 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
4419 DAG.getConstant(1, dl, MVT::i32));
4421 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
4422 // captures the result into a carry flag.
4423 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
4424 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Glue), Hi);
4426 // The low part is an ARMISD::RRX operand, which shifts the carry in.
4427 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
4429 // Merge the pieces into a single i64 value.
4430 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
4433 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
4434 SDValue TmpOp0, TmpOp1;
4435 bool Invert = false;
4439 SDValue Op0 = Op.getOperand(0);
4440 SDValue Op1 = Op.getOperand(1);
4441 SDValue CC = Op.getOperand(2);
4442 EVT CmpVT = Op0.getValueType().changeVectorElementTypeToInteger();
4443 EVT VT = Op.getValueType();
4444 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
4447 if (Op1.getValueType().isFloatingPoint()) {
4448 switch (SetCCOpcode) {
4449 default: llvm_unreachable("Illegal FP comparison");
4451 case ISD::SETNE: Invert = true; // Fallthrough
4453 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
4455 case ISD::SETLT: Swap = true; // Fallthrough
4457 case ISD::SETGT: Opc = ARMISD::VCGT; break;
4459 case ISD::SETLE: Swap = true; // Fallthrough
4461 case ISD::SETGE: Opc = ARMISD::VCGE; break;
4462 case ISD::SETUGE: Swap = true; // Fallthrough
4463 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
4464 case ISD::SETUGT: Swap = true; // Fallthrough
4465 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
4466 case ISD::SETUEQ: Invert = true; // Fallthrough
4468 // Expand this to (OLT | OGT).
4472 Op0 = DAG.getNode(ARMISD::VCGT, dl, CmpVT, TmpOp1, TmpOp0);
4473 Op1 = DAG.getNode(ARMISD::VCGT, dl, CmpVT, TmpOp0, TmpOp1);
4475 case ISD::SETUO: Invert = true; // Fallthrough
4477 // Expand this to (OLT | OGE).
4481 Op0 = DAG.getNode(ARMISD::VCGT, dl, CmpVT, TmpOp1, TmpOp0);
4482 Op1 = DAG.getNode(ARMISD::VCGE, dl, CmpVT, TmpOp0, TmpOp1);
4486 // Integer comparisons.
4487 switch (SetCCOpcode) {
4488 default: llvm_unreachable("Illegal integer comparison");
4489 case ISD::SETNE: Invert = true;
4490 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
4491 case ISD::SETLT: Swap = true;
4492 case ISD::SETGT: Opc = ARMISD::VCGT; break;
4493 case ISD::SETLE: Swap = true;
4494 case ISD::SETGE: Opc = ARMISD::VCGE; break;
4495 case ISD::SETULT: Swap = true;
4496 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
4497 case ISD::SETULE: Swap = true;
4498 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
4501 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
4502 if (Opc == ARMISD::VCEQ) {
4505 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
4507 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
4510 // Ignore bitconvert.
4511 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BITCAST)
4512 AndOp = AndOp.getOperand(0);
4514 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
4516 Op0 = DAG.getNode(ISD::BITCAST, dl, CmpVT, AndOp.getOperand(0));
4517 Op1 = DAG.getNode(ISD::BITCAST, dl, CmpVT, AndOp.getOperand(1));
4524 std::swap(Op0, Op1);
4526 // If one of the operands is a constant vector zero, attempt to fold the
4527 // comparison to a specialized compare-against-zero form.
4529 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
4531 else if (ISD::isBuildVectorAllZeros(Op0.getNode())) {
4532 if (Opc == ARMISD::VCGE)
4533 Opc = ARMISD::VCLEZ;
4534 else if (Opc == ARMISD::VCGT)
4535 Opc = ARMISD::VCLTZ;
4540 if (SingleOp.getNode()) {
4543 Result = DAG.getNode(ARMISD::VCEQZ, dl, CmpVT, SingleOp); break;
4545 Result = DAG.getNode(ARMISD::VCGEZ, dl, CmpVT, SingleOp); break;
4547 Result = DAG.getNode(ARMISD::VCLEZ, dl, CmpVT, SingleOp); break;
4549 Result = DAG.getNode(ARMISD::VCGTZ, dl, CmpVT, SingleOp); break;
4551 Result = DAG.getNode(ARMISD::VCLTZ, dl, CmpVT, SingleOp); break;
4553 Result = DAG.getNode(Opc, dl, CmpVT, Op0, Op1);
4556 Result = DAG.getNode(Opc, dl, CmpVT, Op0, Op1);
4559 Result = DAG.getSExtOrTrunc(Result, dl, VT);
4562 Result = DAG.getNOT(dl, Result, VT);
4567 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
4568 /// valid vector constant for a NEON instruction with a "modified immediate"
4569 /// operand (e.g., VMOV). If so, return the encoded value.
4570 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
4571 unsigned SplatBitSize, SelectionDAG &DAG,
4572 SDLoc dl, EVT &VT, bool is128Bits,
4573 NEONModImmType type) {
4574 unsigned OpCmode, Imm;
4576 // SplatBitSize is set to the smallest size that splats the vector, so a
4577 // zero vector will always have SplatBitSize == 8. However, NEON modified
4578 // immediate instructions others than VMOV do not support the 8-bit encoding
4579 // of a zero vector, and the default encoding of zero is supposed to be the
4584 switch (SplatBitSize) {
4586 if (type != VMOVModImm)
4588 // Any 1-byte value is OK. Op=0, Cmode=1110.
4589 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
4592 VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
4596 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
4597 VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
4598 if ((SplatBits & ~0xff) == 0) {
4599 // Value = 0x00nn: Op=x, Cmode=100x.
4604 if ((SplatBits & ~0xff00) == 0) {
4605 // Value = 0xnn00: Op=x, Cmode=101x.
4607 Imm = SplatBits >> 8;
4613 // NEON's 32-bit VMOV supports splat values where:
4614 // * only one byte is nonzero, or
4615 // * the least significant byte is 0xff and the second byte is nonzero, or
4616 // * the least significant 2 bytes are 0xff and the third is nonzero.
4617 VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
4618 if ((SplatBits & ~0xff) == 0) {
4619 // Value = 0x000000nn: Op=x, Cmode=000x.
4624 if ((SplatBits & ~0xff00) == 0) {
4625 // Value = 0x0000nn00: Op=x, Cmode=001x.
4627 Imm = SplatBits >> 8;
4630 if ((SplatBits & ~0xff0000) == 0) {
4631 // Value = 0x00nn0000: Op=x, Cmode=010x.
4633 Imm = SplatBits >> 16;
4636 if ((SplatBits & ~0xff000000) == 0) {
4637 // Value = 0xnn000000: Op=x, Cmode=011x.
4639 Imm = SplatBits >> 24;
4643 // cmode == 0b1100 and cmode == 0b1101 are not supported for VORR or VBIC
4644 if (type == OtherModImm) return SDValue();
4646 if ((SplatBits & ~0xffff) == 0 &&
4647 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
4648 // Value = 0x0000nnff: Op=x, Cmode=1100.
4650 Imm = SplatBits >> 8;
4654 if ((SplatBits & ~0xffffff) == 0 &&
4655 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
4656 // Value = 0x00nnffff: Op=x, Cmode=1101.
4658 Imm = SplatBits >> 16;
4662 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
4663 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
4664 // VMOV.I32. A (very) minor optimization would be to replicate the value
4665 // and fall through here to test for a valid 64-bit splat. But, then the
4666 // caller would also need to check and handle the change in size.
4670 if (type != VMOVModImm)
4672 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
4673 uint64_t BitMask = 0xff;
4675 unsigned ImmMask = 1;
4677 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
4678 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
4681 } else if ((SplatBits & BitMask) != 0) {
4688 if (DAG.getTargetLoweringInfo().isBigEndian())
4689 // swap higher and lower 32 bit word
4690 Imm = ((Imm & 0xf) << 4) | ((Imm & 0xf0) >> 4);
4692 // Op=1, Cmode=1110.
4694 VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
4699 llvm_unreachable("unexpected size for isNEONModifiedImm");
4702 unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm);
4703 return DAG.getTargetConstant(EncodedVal, dl, MVT::i32);
4706 SDValue ARMTargetLowering::LowerConstantFP(SDValue Op, SelectionDAG &DAG,
4707 const ARMSubtarget *ST) const {
4711 bool IsDouble = Op.getValueType() == MVT::f64;
4712 ConstantFPSDNode *CFP = cast<ConstantFPSDNode>(Op);
4714 // Use the default (constant pool) lowering for double constants when we have
4716 if (IsDouble && Subtarget->isFPOnlySP())
4719 // Try splatting with a VMOV.f32...
4720 APFloat FPVal = CFP->getValueAPF();
4721 int ImmVal = IsDouble ? ARM_AM::getFP64Imm(FPVal) : ARM_AM::getFP32Imm(FPVal);
4724 if (IsDouble || !ST->useNEONForSinglePrecisionFP()) {
4725 // We have code in place to select a valid ConstantFP already, no need to
4730 // It's a float and we are trying to use NEON operations where
4731 // possible. Lower it to a splat followed by an extract.
4733 SDValue NewVal = DAG.getTargetConstant(ImmVal, DL, MVT::i32);
4734 SDValue VecConstant = DAG.getNode(ARMISD::VMOVFPIMM, DL, MVT::v2f32,
4736 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecConstant,
4737 DAG.getConstant(0, DL, MVT::i32));
4740 // The rest of our options are NEON only, make sure that's allowed before
4742 if (!ST->hasNEON() || (!IsDouble && !ST->useNEONForSinglePrecisionFP()))
4746 uint64_t iVal = FPVal.bitcastToAPInt().getZExtValue();
4748 // It wouldn't really be worth bothering for doubles except for one very
4749 // important value, which does happen to match: 0.0. So make sure we don't do
4751 if (IsDouble && (iVal & 0xffffffff) != (iVal >> 32))
4754 // Try a VMOV.i32 (FIXME: i8, i16, or i64 could work too).
4755 SDValue NewVal = isNEONModifiedImm(iVal & 0xffffffffU, 0, 32, DAG, SDLoc(Op),
4756 VMovVT, false, VMOVModImm);
4757 if (NewVal != SDValue()) {
4759 SDValue VecConstant = DAG.getNode(ARMISD::VMOVIMM, DL, VMovVT,
4762 return DAG.getNode(ISD::BITCAST, DL, MVT::f64, VecConstant);
4764 // It's a float: cast and extract a vector element.
4765 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
4767 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
4768 DAG.getConstant(0, DL, MVT::i32));
4771 // Finally, try a VMVN.i32
4772 NewVal = isNEONModifiedImm(~iVal & 0xffffffffU, 0, 32, DAG, SDLoc(Op), VMovVT,
4774 if (NewVal != SDValue()) {
4776 SDValue VecConstant = DAG.getNode(ARMISD::VMVNIMM, DL, VMovVT, NewVal);
4779 return DAG.getNode(ISD::BITCAST, DL, MVT::f64, VecConstant);
4781 // It's a float: cast and extract a vector element.
4782 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
4784 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
4785 DAG.getConstant(0, DL, MVT::i32));
4791 // check if an VEXT instruction can handle the shuffle mask when the
4792 // vector sources of the shuffle are the same.
4793 static bool isSingletonVEXTMask(ArrayRef<int> M, EVT VT, unsigned &Imm) {
4794 unsigned NumElts = VT.getVectorNumElements();
4796 // Assume that the first shuffle index is not UNDEF. Fail if it is.
4802 // If this is a VEXT shuffle, the immediate value is the index of the first
4803 // element. The other shuffle indices must be the successive elements after
4805 unsigned ExpectedElt = Imm;
4806 for (unsigned i = 1; i < NumElts; ++i) {
4807 // Increment the expected index. If it wraps around, just follow it
4808 // back to index zero and keep going.
4810 if (ExpectedElt == NumElts)
4813 if (M[i] < 0) continue; // ignore UNDEF indices
4814 if (ExpectedElt != static_cast<unsigned>(M[i]))
4822 static bool isVEXTMask(ArrayRef<int> M, EVT VT,
4823 bool &ReverseVEXT, unsigned &Imm) {
4824 unsigned NumElts = VT.getVectorNumElements();
4825 ReverseVEXT = false;
4827 // Assume that the first shuffle index is not UNDEF. Fail if it is.
4833 // If this is a VEXT shuffle, the immediate value is the index of the first
4834 // element. The other shuffle indices must be the successive elements after
4836 unsigned ExpectedElt = Imm;
4837 for (unsigned i = 1; i < NumElts; ++i) {
4838 // Increment the expected index. If it wraps around, it may still be
4839 // a VEXT but the source vectors must be swapped.
4841 if (ExpectedElt == NumElts * 2) {
4846 if (M[i] < 0) continue; // ignore UNDEF indices
4847 if (ExpectedElt != static_cast<unsigned>(M[i]))
4851 // Adjust the index value if the source operands will be swapped.
4858 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
4859 /// instruction with the specified blocksize. (The order of the elements
4860 /// within each block of the vector is reversed.)
4861 static bool isVREVMask(ArrayRef<int> M, EVT VT, unsigned BlockSize) {
4862 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
4863 "Only possible block sizes for VREV are: 16, 32, 64");
4865 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4869 unsigned NumElts = VT.getVectorNumElements();
4870 unsigned BlockElts = M[0] + 1;
4871 // If the first shuffle index is UNDEF, be optimistic.
4873 BlockElts = BlockSize / EltSz;
4875 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
4878 for (unsigned i = 0; i < NumElts; ++i) {
4879 if (M[i] < 0) continue; // ignore UNDEF indices
4880 if ((unsigned) M[i] != (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
4887 static bool isVTBLMask(ArrayRef<int> M, EVT VT) {
4888 // We can handle <8 x i8> vector shuffles. If the index in the mask is out of
4889 // range, then 0 is placed into the resulting vector. So pretty much any mask
4890 // of 8 elements can work here.
4891 return VT == MVT::v8i8 && M.size() == 8;
4894 static bool isVTRNMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4895 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4899 unsigned NumElts = VT.getVectorNumElements();
4900 WhichResult = (M[0] == 0 ? 0 : 1);
4901 for (unsigned i = 0; i < NumElts; i += 2) {
4902 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
4903 (M[i+1] >= 0 && (unsigned) M[i+1] != i + NumElts + WhichResult))
4909 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
4910 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4911 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
4912 static bool isVTRN_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4913 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4917 unsigned NumElts = VT.getVectorNumElements();
4918 WhichResult = (M[0] == 0 ? 0 : 1);
4919 for (unsigned i = 0; i < NumElts; i += 2) {
4920 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
4921 (M[i+1] >= 0 && (unsigned) M[i+1] != i + WhichResult))
4927 static bool isVUZPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4928 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4932 unsigned NumElts = VT.getVectorNumElements();
4933 WhichResult = (M[0] == 0 ? 0 : 1);
4934 for (unsigned i = 0; i != NumElts; ++i) {
4935 if (M[i] < 0) continue; // ignore UNDEF indices
4936 if ((unsigned) M[i] != 2 * i + WhichResult)
4940 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4941 if (VT.is64BitVector() && EltSz == 32)
4947 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
4948 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4949 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
4950 static bool isVUZP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4951 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4955 unsigned Half = VT.getVectorNumElements() / 2;
4956 WhichResult = (M[0] == 0 ? 0 : 1);
4957 for (unsigned j = 0; j != 2; ++j) {
4958 unsigned Idx = WhichResult;
4959 for (unsigned i = 0; i != Half; ++i) {
4960 int MIdx = M[i + j * Half];
4961 if (MIdx >= 0 && (unsigned) MIdx != Idx)
4967 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4968 if (VT.is64BitVector() && EltSz == 32)
4974 static bool isVZIPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4975 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4979 unsigned NumElts = VT.getVectorNumElements();
4980 WhichResult = (M[0] == 0 ? 0 : 1);
4981 unsigned Idx = WhichResult * NumElts / 2;
4982 for (unsigned i = 0; i != NumElts; i += 2) {
4983 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
4984 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx + NumElts))
4989 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4990 if (VT.is64BitVector() && EltSz == 32)
4996 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
4997 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4998 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
4999 static bool isVZIP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
5000 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
5004 unsigned NumElts = VT.getVectorNumElements();
5005 WhichResult = (M[0] == 0 ? 0 : 1);
5006 unsigned Idx = WhichResult * NumElts / 2;
5007 for (unsigned i = 0; i != NumElts; i += 2) {
5008 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
5009 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx))
5014 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
5015 if (VT.is64BitVector() && EltSz == 32)
5021 /// \return true if this is a reverse operation on an vector.
5022 static bool isReverseMask(ArrayRef<int> M, EVT VT) {
5023 unsigned NumElts = VT.getVectorNumElements();
5024 // Make sure the mask has the right size.
5025 if (NumElts != M.size())
5028 // Look for <15, ..., 3, -1, 1, 0>.
5029 for (unsigned i = 0; i != NumElts; ++i)
5030 if (M[i] >= 0 && M[i] != (int) (NumElts - 1 - i))
5036 // If N is an integer constant that can be moved into a register in one
5037 // instruction, return an SDValue of such a constant (will become a MOV
5038 // instruction). Otherwise return null.
5039 static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG,
5040 const ARMSubtarget *ST, SDLoc dl) {
5042 if (!isa<ConstantSDNode>(N))
5044 Val = cast<ConstantSDNode>(N)->getZExtValue();
5046 if (ST->isThumb1Only()) {
5047 if (Val <= 255 || ~Val <= 255)
5048 return DAG.getConstant(Val, dl, MVT::i32);
5050 if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1)
5051 return DAG.getConstant(Val, dl, MVT::i32);
5056 // If this is a case we can't handle, return null and let the default
5057 // expansion code take care of it.
5058 SDValue ARMTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
5059 const ARMSubtarget *ST) const {
5060 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
5062 EVT VT = Op.getValueType();
5064 APInt SplatBits, SplatUndef;
5065 unsigned SplatBitSize;
5067 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
5068 if (SplatBitSize <= 64) {
5069 // Check if an immediate VMOV works.
5071 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
5072 SplatUndef.getZExtValue(), SplatBitSize,
5073 DAG, dl, VmovVT, VT.is128BitVector(),
5075 if (Val.getNode()) {
5076 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
5077 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
5080 // Try an immediate VMVN.
5081 uint64_t NegatedImm = (~SplatBits).getZExtValue();
5082 Val = isNEONModifiedImm(NegatedImm,
5083 SplatUndef.getZExtValue(), SplatBitSize,
5084 DAG, dl, VmovVT, VT.is128BitVector(),
5086 if (Val.getNode()) {
5087 SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
5088 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
5091 // Use vmov.f32 to materialize other v2f32 and v4f32 splats.
5092 if ((VT == MVT::v2f32 || VT == MVT::v4f32) && SplatBitSize == 32) {
5093 int ImmVal = ARM_AM::getFP32Imm(SplatBits);
5095 SDValue Val = DAG.getTargetConstant(ImmVal, dl, MVT::i32);
5096 return DAG.getNode(ARMISD::VMOVFPIMM, dl, VT, Val);
5102 // Scan through the operands to see if only one value is used.
5104 // As an optimisation, even if more than one value is used it may be more
5105 // profitable to splat with one value then change some lanes.
5107 // Heuristically we decide to do this if the vector has a "dominant" value,
5108 // defined as splatted to more than half of the lanes.
5109 unsigned NumElts = VT.getVectorNumElements();
5110 bool isOnlyLowElement = true;
5111 bool usesOnlyOneValue = true;
5112 bool hasDominantValue = false;
5113 bool isConstant = true;
5115 // Map of the number of times a particular SDValue appears in the
5117 DenseMap<SDValue, unsigned> ValueCounts;
5119 for (unsigned i = 0; i < NumElts; ++i) {
5120 SDValue V = Op.getOperand(i);
5121 if (V.getOpcode() == ISD::UNDEF)
5124 isOnlyLowElement = false;
5125 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
5128 ValueCounts.insert(std::make_pair(V, 0));
5129 unsigned &Count = ValueCounts[V];
5131 // Is this value dominant? (takes up more than half of the lanes)
5132 if (++Count > (NumElts / 2)) {
5133 hasDominantValue = true;
5137 if (ValueCounts.size() != 1)
5138 usesOnlyOneValue = false;
5139 if (!Value.getNode() && ValueCounts.size() > 0)
5140 Value = ValueCounts.begin()->first;
5142 if (ValueCounts.size() == 0)
5143 return DAG.getUNDEF(VT);
5145 // Loads are better lowered with insert_vector_elt/ARMISD::BUILD_VECTOR.
5146 // Keep going if we are hitting this case.
5147 if (isOnlyLowElement && !ISD::isNormalLoad(Value.getNode()))
5148 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
5150 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5152 // Use VDUP for non-constant splats. For f32 constant splats, reduce to
5153 // i32 and try again.
5154 if (hasDominantValue && EltSize <= 32) {
5158 // If we are VDUPing a value that comes directly from a vector, that will
5159 // cause an unnecessary move to and from a GPR, where instead we could
5160 // just use VDUPLANE. We can only do this if the lane being extracted
5161 // is at a constant index, as the VDUP from lane instructions only have
5162 // constant-index forms.
5163 if (Value->getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5164 isa<ConstantSDNode>(Value->getOperand(1))) {
5165 // We need to create a new undef vector to use for the VDUPLANE if the
5166 // size of the vector from which we get the value is different than the
5167 // size of the vector that we need to create. We will insert the element
5168 // such that the register coalescer will remove unnecessary copies.
5169 if (VT != Value->getOperand(0).getValueType()) {
5170 ConstantSDNode *constIndex;
5171 constIndex = dyn_cast<ConstantSDNode>(Value->getOperand(1));
5172 assert(constIndex && "The index is not a constant!");
5173 unsigned index = constIndex->getAPIntValue().getLimitedValue() %
5174 VT.getVectorNumElements();
5175 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
5176 DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DAG.getUNDEF(VT),
5177 Value, DAG.getConstant(index, dl, MVT::i32)),
5178 DAG.getConstant(index, dl, MVT::i32));
5180 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
5181 Value->getOperand(0), Value->getOperand(1));
5183 N = DAG.getNode(ARMISD::VDUP, dl, VT, Value);
5185 if (!usesOnlyOneValue) {
5186 // The dominant value was splatted as 'N', but we now have to insert
5187 // all differing elements.
5188 for (unsigned I = 0; I < NumElts; ++I) {
5189 if (Op.getOperand(I) == Value)
5191 SmallVector<SDValue, 3> Ops;
5193 Ops.push_back(Op.getOperand(I));
5194 Ops.push_back(DAG.getConstant(I, dl, MVT::i32));
5195 N = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Ops);
5200 if (VT.getVectorElementType().isFloatingPoint()) {
5201 SmallVector<SDValue, 8> Ops;
5202 for (unsigned i = 0; i < NumElts; ++i)
5203 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, MVT::i32,
5205 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
5206 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, Ops);
5207 Val = LowerBUILD_VECTOR(Val, DAG, ST);
5209 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
5211 if (usesOnlyOneValue) {
5212 SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
5213 if (isConstant && Val.getNode())
5214 return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
5218 // If all elements are constants and the case above didn't get hit, fall back
5219 // to the default expansion, which will generate a load from the constant
5224 // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
5226 SDValue shuffle = ReconstructShuffle(Op, DAG);
5227 if (shuffle != SDValue())
5231 // Vectors with 32- or 64-bit elements can be built by directly assigning
5232 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
5233 // will be legalized.
5234 if (EltSize >= 32) {
5235 // Do the expansion with floating-point types, since that is what the VFP
5236 // registers are defined to use, and since i64 is not legal.
5237 EVT EltVT = EVT::getFloatingPointVT(EltSize);
5238 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
5239 SmallVector<SDValue, 8> Ops;
5240 for (unsigned i = 0; i < NumElts; ++i)
5241 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, EltVT, Op.getOperand(i)));
5242 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, Ops);
5243 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
5246 // If all else fails, just use a sequence of INSERT_VECTOR_ELT when we
5247 // know the default expansion would otherwise fall back on something even
5248 // worse. For a vector with one or two non-undef values, that's
5249 // scalar_to_vector for the elements followed by a shuffle (provided the
5250 // shuffle is valid for the target) and materialization element by element
5251 // on the stack followed by a load for everything else.
5252 if (!isConstant && !usesOnlyOneValue) {
5253 SDValue Vec = DAG.getUNDEF(VT);
5254 for (unsigned i = 0 ; i < NumElts; ++i) {
5255 SDValue V = Op.getOperand(i);
5256 if (V.getOpcode() == ISD::UNDEF)
5258 SDValue LaneIdx = DAG.getConstant(i, dl, MVT::i32);
5259 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Vec, V, LaneIdx);
5267 // Gather data to see if the operation can be modelled as a
5268 // shuffle in combination with VEXTs.
5269 SDValue ARMTargetLowering::ReconstructShuffle(SDValue Op,
5270 SelectionDAG &DAG) const {
5272 EVT VT = Op.getValueType();
5273 unsigned NumElts = VT.getVectorNumElements();
5275 SmallVector<SDValue, 2> SourceVecs;
5276 SmallVector<unsigned, 2> MinElts;
5277 SmallVector<unsigned, 2> MaxElts;
5279 for (unsigned i = 0; i < NumElts; ++i) {
5280 SDValue V = Op.getOperand(i);
5281 if (V.getOpcode() == ISD::UNDEF)
5283 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
5284 // A shuffle can only come from building a vector from various
5285 // elements of other vectors.
5287 } else if (V.getOperand(0).getValueType().getVectorElementType() !=
5288 VT.getVectorElementType()) {
5289 // This code doesn't know how to handle shuffles where the vector
5290 // element types do not match (this happens because type legalization
5291 // promotes the return type of EXTRACT_VECTOR_ELT).
5292 // FIXME: It might be appropriate to extend this code to handle
5293 // mismatched types.
5297 // Record this extraction against the appropriate vector if possible...
5298 SDValue SourceVec = V.getOperand(0);
5299 // If the element number isn't a constant, we can't effectively
5300 // analyze what's going on.
5301 if (!isa<ConstantSDNode>(V.getOperand(1)))
5303 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
5304 bool FoundSource = false;
5305 for (unsigned j = 0; j < SourceVecs.size(); ++j) {
5306 if (SourceVecs[j] == SourceVec) {
5307 if (MinElts[j] > EltNo)
5309 if (MaxElts[j] < EltNo)
5316 // Or record a new source if not...
5318 SourceVecs.push_back(SourceVec);
5319 MinElts.push_back(EltNo);
5320 MaxElts.push_back(EltNo);
5324 // Currently only do something sane when at most two source vectors
5326 if (SourceVecs.size() > 2)
5329 SDValue ShuffleSrcs[2] = {DAG.getUNDEF(VT), DAG.getUNDEF(VT) };
5330 int VEXTOffsets[2] = {0, 0};
5332 // This loop extracts the usage patterns of the source vectors
5333 // and prepares appropriate SDValues for a shuffle if possible.
5334 for (unsigned i = 0; i < SourceVecs.size(); ++i) {
5335 if (SourceVecs[i].getValueType() == VT) {
5336 // No VEXT necessary
5337 ShuffleSrcs[i] = SourceVecs[i];
5340 } else if (SourceVecs[i].getValueType().getVectorNumElements() < NumElts) {
5341 // It probably isn't worth padding out a smaller vector just to
5342 // break it down again in a shuffle.
5346 // Since only 64-bit and 128-bit vectors are legal on ARM and
5347 // we've eliminated the other cases...
5348 assert(SourceVecs[i].getValueType().getVectorNumElements() == 2*NumElts &&
5349 "unexpected vector sizes in ReconstructShuffle");
5351 if (MaxElts[i] - MinElts[i] >= NumElts) {
5352 // Span too large for a VEXT to cope
5356 if (MinElts[i] >= NumElts) {
5357 // The extraction can just take the second half
5358 VEXTOffsets[i] = NumElts;
5359 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5361 DAG.getIntPtrConstant(NumElts, dl));
5362 } else if (MaxElts[i] < NumElts) {
5363 // The extraction can just take the first half
5365 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5367 DAG.getIntPtrConstant(0, dl));
5369 // An actual VEXT is needed
5370 VEXTOffsets[i] = MinElts[i];
5371 SDValue VEXTSrc1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5373 DAG.getIntPtrConstant(0, dl));
5374 SDValue VEXTSrc2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5376 DAG.getIntPtrConstant(NumElts, dl));
5377 ShuffleSrcs[i] = DAG.getNode(ARMISD::VEXT, dl, VT, VEXTSrc1, VEXTSrc2,
5378 DAG.getConstant(VEXTOffsets[i], dl,
5383 SmallVector<int, 8> Mask;
5385 for (unsigned i = 0; i < NumElts; ++i) {
5386 SDValue Entry = Op.getOperand(i);
5387 if (Entry.getOpcode() == ISD::UNDEF) {
5392 SDValue ExtractVec = Entry.getOperand(0);
5393 int ExtractElt = cast<ConstantSDNode>(Op.getOperand(i)
5394 .getOperand(1))->getSExtValue();
5395 if (ExtractVec == SourceVecs[0]) {
5396 Mask.push_back(ExtractElt - VEXTOffsets[0]);
5398 Mask.push_back(ExtractElt + NumElts - VEXTOffsets[1]);
5402 // Final check before we try to produce nonsense...
5403 if (isShuffleMaskLegal(Mask, VT))
5404 return DAG.getVectorShuffle(VT, dl, ShuffleSrcs[0], ShuffleSrcs[1],
5410 /// isShuffleMaskLegal - Targets can use this to indicate that they only
5411 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
5412 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
5413 /// are assumed to be legal.
5415 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
5417 if (VT.getVectorNumElements() == 4 &&
5418 (VT.is128BitVector() || VT.is64BitVector())) {
5419 unsigned PFIndexes[4];
5420 for (unsigned i = 0; i != 4; ++i) {
5424 PFIndexes[i] = M[i];
5427 // Compute the index in the perfect shuffle table.
5428 unsigned PFTableIndex =
5429 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
5430 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
5431 unsigned Cost = (PFEntry >> 30);
5438 unsigned Imm, WhichResult;
5440 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5441 return (EltSize >= 32 ||
5442 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
5443 isVREVMask(M, VT, 64) ||
5444 isVREVMask(M, VT, 32) ||
5445 isVREVMask(M, VT, 16) ||
5446 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
5447 isVTBLMask(M, VT) ||
5448 isVTRNMask(M, VT, WhichResult) ||
5449 isVUZPMask(M, VT, WhichResult) ||
5450 isVZIPMask(M, VT, WhichResult) ||
5451 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
5452 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
5453 isVZIP_v_undef_Mask(M, VT, WhichResult) ||
5454 ((VT == MVT::v8i16 || VT == MVT::v16i8) && isReverseMask(M, VT)));
5457 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
5458 /// the specified operations to build the shuffle.
5459 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
5460 SDValue RHS, SelectionDAG &DAG,
5462 unsigned OpNum = (PFEntry >> 26) & 0x0F;
5463 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
5464 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
5467 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
5476 OP_VUZPL, // VUZP, left result
5477 OP_VUZPR, // VUZP, right result
5478 OP_VZIPL, // VZIP, left result
5479 OP_VZIPR, // VZIP, right result
5480 OP_VTRNL, // VTRN, left result
5481 OP_VTRNR // VTRN, right result
5484 if (OpNum == OP_COPY) {
5485 if (LHSID == (1*9+2)*9+3) return LHS;
5486 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
5490 SDValue OpLHS, OpRHS;
5491 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
5492 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
5493 EVT VT = OpLHS.getValueType();
5496 default: llvm_unreachable("Unknown shuffle opcode!");
5498 // VREV divides the vector in half and swaps within the half.
5499 if (VT.getVectorElementType() == MVT::i32 ||
5500 VT.getVectorElementType() == MVT::f32)
5501 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
5502 // vrev <4 x i16> -> VREV32
5503 if (VT.getVectorElementType() == MVT::i16)
5504 return DAG.getNode(ARMISD::VREV32, dl, VT, OpLHS);
5505 // vrev <4 x i8> -> VREV16
5506 assert(VT.getVectorElementType() == MVT::i8);
5507 return DAG.getNode(ARMISD::VREV16, dl, VT, OpLHS);
5512 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
5513 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, dl, MVT::i32));
5517 return DAG.getNode(ARMISD::VEXT, dl, VT,
5519 DAG.getConstant(OpNum - OP_VEXT1 + 1, dl, MVT::i32));
5522 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
5523 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
5526 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
5527 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
5530 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
5531 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
5535 static SDValue LowerVECTOR_SHUFFLEv8i8(SDValue Op,
5536 ArrayRef<int> ShuffleMask,
5537 SelectionDAG &DAG) {
5538 // Check to see if we can use the VTBL instruction.
5539 SDValue V1 = Op.getOperand(0);
5540 SDValue V2 = Op.getOperand(1);
5543 SmallVector<SDValue, 8> VTBLMask;
5544 for (ArrayRef<int>::iterator
5545 I = ShuffleMask.begin(), E = ShuffleMask.end(); I != E; ++I)
5546 VTBLMask.push_back(DAG.getConstant(*I, DL, MVT::i32));
5548 if (V2.getNode()->getOpcode() == ISD::UNDEF)
5549 return DAG.getNode(ARMISD::VTBL1, DL, MVT::v8i8, V1,
5550 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8, VTBLMask));
5552 return DAG.getNode(ARMISD::VTBL2, DL, MVT::v8i8, V1, V2,
5553 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8, VTBLMask));
5556 static SDValue LowerReverse_VECTOR_SHUFFLEv16i8_v8i16(SDValue Op,
5557 SelectionDAG &DAG) {
5559 SDValue OpLHS = Op.getOperand(0);
5560 EVT VT = OpLHS.getValueType();
5562 assert((VT == MVT::v8i16 || VT == MVT::v16i8) &&
5563 "Expect an v8i16/v16i8 type");
5564 OpLHS = DAG.getNode(ARMISD::VREV64, DL, VT, OpLHS);
5565 // For a v16i8 type: After the VREV, we have got <8, ...15, 8, ..., 0>. Now,
5566 // extract the first 8 bytes into the top double word and the last 8 bytes
5567 // into the bottom double word. The v8i16 case is similar.
5568 unsigned ExtractNum = (VT == MVT::v16i8) ? 8 : 4;
5569 return DAG.getNode(ARMISD::VEXT, DL, VT, OpLHS, OpLHS,
5570 DAG.getConstant(ExtractNum, DL, MVT::i32));
5573 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
5574 SDValue V1 = Op.getOperand(0);
5575 SDValue V2 = Op.getOperand(1);
5577 EVT VT = Op.getValueType();
5578 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
5580 // Convert shuffles that are directly supported on NEON to target-specific
5581 // DAG nodes, instead of keeping them as shuffles and matching them again
5582 // during code selection. This is more efficient and avoids the possibility
5583 // of inconsistencies between legalization and selection.
5584 // FIXME: floating-point vectors should be canonicalized to integer vectors
5585 // of the same time so that they get CSEd properly.
5586 ArrayRef<int> ShuffleMask = SVN->getMask();
5588 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5589 if (EltSize <= 32) {
5590 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
5591 int Lane = SVN->getSplatIndex();
5592 // If this is undef splat, generate it via "just" vdup, if possible.
5593 if (Lane == -1) Lane = 0;
5595 // Test if V1 is a SCALAR_TO_VECTOR.
5596 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
5597 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
5599 // Test if V1 is a BUILD_VECTOR which is equivalent to a SCALAR_TO_VECTOR
5600 // (and probably will turn into a SCALAR_TO_VECTOR once legalization
5602 if (Lane == 0 && V1.getOpcode() == ISD::BUILD_VECTOR &&
5603 !isa<ConstantSDNode>(V1.getOperand(0))) {
5604 bool IsScalarToVector = true;
5605 for (unsigned i = 1, e = V1.getNumOperands(); i != e; ++i)
5606 if (V1.getOperand(i).getOpcode() != ISD::UNDEF) {
5607 IsScalarToVector = false;
5610 if (IsScalarToVector)
5611 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
5613 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
5614 DAG.getConstant(Lane, dl, MVT::i32));
5619 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
5622 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
5623 DAG.getConstant(Imm, dl, MVT::i32));
5626 if (isVREVMask(ShuffleMask, VT, 64))
5627 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
5628 if (isVREVMask(ShuffleMask, VT, 32))
5629 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
5630 if (isVREVMask(ShuffleMask, VT, 16))
5631 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
5633 if (V2->getOpcode() == ISD::UNDEF &&
5634 isSingletonVEXTMask(ShuffleMask, VT, Imm)) {
5635 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V1,
5636 DAG.getConstant(Imm, dl, MVT::i32));
5639 // Check for Neon shuffles that modify both input vectors in place.
5640 // If both results are used, i.e., if there are two shuffles with the same
5641 // source operands and with masks corresponding to both results of one of
5642 // these operations, DAG memoization will ensure that a single node is
5643 // used for both shuffles.
5644 unsigned WhichResult;
5645 if (isVTRNMask(ShuffleMask, VT, WhichResult))
5646 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
5647 V1, V2).getValue(WhichResult);
5648 if (isVUZPMask(ShuffleMask, VT, WhichResult))
5649 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
5650 V1, V2).getValue(WhichResult);
5651 if (isVZIPMask(ShuffleMask, VT, WhichResult))
5652 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
5653 V1, V2).getValue(WhichResult);
5655 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
5656 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
5657 V1, V1).getValue(WhichResult);
5658 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
5659 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
5660 V1, V1).getValue(WhichResult);
5661 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
5662 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
5663 V1, V1).getValue(WhichResult);
5666 // If the shuffle is not directly supported and it has 4 elements, use
5667 // the PerfectShuffle-generated table to synthesize it from other shuffles.
5668 unsigned NumElts = VT.getVectorNumElements();
5670 unsigned PFIndexes[4];
5671 for (unsigned i = 0; i != 4; ++i) {
5672 if (ShuffleMask[i] < 0)
5675 PFIndexes[i] = ShuffleMask[i];
5678 // Compute the index in the perfect shuffle table.
5679 unsigned PFTableIndex =
5680 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
5681 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
5682 unsigned Cost = (PFEntry >> 30);
5685 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
5688 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
5689 if (EltSize >= 32) {
5690 // Do the expansion with floating-point types, since that is what the VFP
5691 // registers are defined to use, and since i64 is not legal.
5692 EVT EltVT = EVT::getFloatingPointVT(EltSize);
5693 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
5694 V1 = DAG.getNode(ISD::BITCAST, dl, VecVT, V1);
5695 V2 = DAG.getNode(ISD::BITCAST, dl, VecVT, V2);
5696 SmallVector<SDValue, 8> Ops;
5697 for (unsigned i = 0; i < NumElts; ++i) {
5698 if (ShuffleMask[i] < 0)
5699 Ops.push_back(DAG.getUNDEF(EltVT));
5701 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
5702 ShuffleMask[i] < (int)NumElts ? V1 : V2,
5703 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
5706 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, Ops);
5707 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
5710 if ((VT == MVT::v8i16 || VT == MVT::v16i8) && isReverseMask(ShuffleMask, VT))
5711 return LowerReverse_VECTOR_SHUFFLEv16i8_v8i16(Op, DAG);
5713 if (VT == MVT::v8i8) {
5714 SDValue NewOp = LowerVECTOR_SHUFFLEv8i8(Op, ShuffleMask, DAG);
5715 if (NewOp.getNode())
5722 static SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
5723 // INSERT_VECTOR_ELT is legal only for immediate indexes.
5724 SDValue Lane = Op.getOperand(2);
5725 if (!isa<ConstantSDNode>(Lane))
5731 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
5732 // EXTRACT_VECTOR_ELT is legal only for immediate indexes.
5733 SDValue Lane = Op.getOperand(1);
5734 if (!isa<ConstantSDNode>(Lane))
5737 SDValue Vec = Op.getOperand(0);
5738 if (Op.getValueType() == MVT::i32 &&
5739 Vec.getValueType().getVectorElementType().getSizeInBits() < 32) {
5741 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
5747 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
5748 // The only time a CONCAT_VECTORS operation can have legal types is when
5749 // two 64-bit vectors are concatenated to a 128-bit vector.
5750 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
5751 "unexpected CONCAT_VECTORS");
5753 SDValue Val = DAG.getUNDEF(MVT::v2f64);
5754 SDValue Op0 = Op.getOperand(0);
5755 SDValue Op1 = Op.getOperand(1);
5756 if (Op0.getOpcode() != ISD::UNDEF)
5757 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
5758 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op0),
5759 DAG.getIntPtrConstant(0, dl));
5760 if (Op1.getOpcode() != ISD::UNDEF)
5761 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
5762 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op1),
5763 DAG.getIntPtrConstant(1, dl));
5764 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Val);
5767 /// isExtendedBUILD_VECTOR - Check if N is a constant BUILD_VECTOR where each
5768 /// element has been zero/sign-extended, depending on the isSigned parameter,
5769 /// from an integer type half its size.
5770 static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG,
5772 // A v2i64 BUILD_VECTOR will have been legalized to a BITCAST from v4i32.
5773 EVT VT = N->getValueType(0);
5774 if (VT == MVT::v2i64 && N->getOpcode() == ISD::BITCAST) {
5775 SDNode *BVN = N->getOperand(0).getNode();
5776 if (BVN->getValueType(0) != MVT::v4i32 ||
5777 BVN->getOpcode() != ISD::BUILD_VECTOR)
5779 unsigned LoElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
5780 unsigned HiElt = 1 - LoElt;
5781 ConstantSDNode *Lo0 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt));
5782 ConstantSDNode *Hi0 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt));
5783 ConstantSDNode *Lo1 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt+2));
5784 ConstantSDNode *Hi1 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt+2));
5785 if (!Lo0 || !Hi0 || !Lo1 || !Hi1)
5788 if (Hi0->getSExtValue() == Lo0->getSExtValue() >> 32 &&
5789 Hi1->getSExtValue() == Lo1->getSExtValue() >> 32)
5792 if (Hi0->isNullValue() && Hi1->isNullValue())
5798 if (N->getOpcode() != ISD::BUILD_VECTOR)
5801 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
5802 SDNode *Elt = N->getOperand(i).getNode();
5803 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
5804 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5805 unsigned HalfSize = EltSize / 2;
5807 if (!isIntN(HalfSize, C->getSExtValue()))
5810 if (!isUIntN(HalfSize, C->getZExtValue()))
5821 /// isSignExtended - Check if a node is a vector value that is sign-extended
5822 /// or a constant BUILD_VECTOR with sign-extended elements.
5823 static bool isSignExtended(SDNode *N, SelectionDAG &DAG) {
5824 if (N->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N))
5826 if (isExtendedBUILD_VECTOR(N, DAG, true))
5831 /// isZeroExtended - Check if a node is a vector value that is zero-extended
5832 /// or a constant BUILD_VECTOR with zero-extended elements.
5833 static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) {
5834 if (N->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N))
5836 if (isExtendedBUILD_VECTOR(N, DAG, false))
5841 static EVT getExtensionTo64Bits(const EVT &OrigVT) {
5842 if (OrigVT.getSizeInBits() >= 64)
5845 assert(OrigVT.isSimple() && "Expecting a simple value type");
5847 MVT::SimpleValueType OrigSimpleTy = OrigVT.getSimpleVT().SimpleTy;
5848 switch (OrigSimpleTy) {
5849 default: llvm_unreachable("Unexpected Vector Type");
5858 /// AddRequiredExtensionForVMULL - Add a sign/zero extension to extend the total
5859 /// value size to 64 bits. We need a 64-bit D register as an operand to VMULL.
5860 /// We insert the required extension here to get the vector to fill a D register.
5861 static SDValue AddRequiredExtensionForVMULL(SDValue N, SelectionDAG &DAG,
5864 unsigned ExtOpcode) {
5865 // The vector originally had a size of OrigTy. It was then extended to ExtTy.
5866 // We expect the ExtTy to be 128-bits total. If the OrigTy is less than
5867 // 64-bits we need to insert a new extension so that it will be 64-bits.
5868 assert(ExtTy.is128BitVector() && "Unexpected extension size");
5869 if (OrigTy.getSizeInBits() >= 64)
5872 // Must extend size to at least 64 bits to be used as an operand for VMULL.
5873 EVT NewVT = getExtensionTo64Bits(OrigTy);
5875 return DAG.getNode(ExtOpcode, SDLoc(N), NewVT, N);
5878 /// SkipLoadExtensionForVMULL - return a load of the original vector size that
5879 /// does not do any sign/zero extension. If the original vector is less
5880 /// than 64 bits, an appropriate extension will be added after the load to
5881 /// reach a total size of 64 bits. We have to add the extension separately
5882 /// because ARM does not have a sign/zero extending load for vectors.
5883 static SDValue SkipLoadExtensionForVMULL(LoadSDNode *LD, SelectionDAG& DAG) {
5884 EVT ExtendedTy = getExtensionTo64Bits(LD->getMemoryVT());
5886 // The load already has the right type.
5887 if (ExtendedTy == LD->getMemoryVT())
5888 return DAG.getLoad(LD->getMemoryVT(), SDLoc(LD), LD->getChain(),
5889 LD->getBasePtr(), LD->getPointerInfo(), LD->isVolatile(),
5890 LD->isNonTemporal(), LD->isInvariant(),
5891 LD->getAlignment());
5893 // We need to create a zextload/sextload. We cannot just create a load
5894 // followed by a zext/zext node because LowerMUL is also run during normal
5895 // operation legalization where we can't create illegal types.
5896 return DAG.getExtLoad(LD->getExtensionType(), SDLoc(LD), ExtendedTy,
5897 LD->getChain(), LD->getBasePtr(), LD->getPointerInfo(),
5898 LD->getMemoryVT(), LD->isVolatile(), LD->isInvariant(),
5899 LD->isNonTemporal(), LD->getAlignment());
5902 /// SkipExtensionForVMULL - For a node that is a SIGN_EXTEND, ZERO_EXTEND,
5903 /// extending load, or BUILD_VECTOR with extended elements, return the
5904 /// unextended value. The unextended vector should be 64 bits so that it can
5905 /// be used as an operand to a VMULL instruction. If the original vector size
5906 /// before extension is less than 64 bits we add a an extension to resize
5907 /// the vector to 64 bits.
5908 static SDValue SkipExtensionForVMULL(SDNode *N, SelectionDAG &DAG) {
5909 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
5910 return AddRequiredExtensionForVMULL(N->getOperand(0), DAG,
5911 N->getOperand(0)->getValueType(0),
5915 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
5916 return SkipLoadExtensionForVMULL(LD, DAG);
5918 // Otherwise, the value must be a BUILD_VECTOR. For v2i64, it will
5919 // have been legalized as a BITCAST from v4i32.
5920 if (N->getOpcode() == ISD::BITCAST) {
5921 SDNode *BVN = N->getOperand(0).getNode();
5922 assert(BVN->getOpcode() == ISD::BUILD_VECTOR &&
5923 BVN->getValueType(0) == MVT::v4i32 && "expected v4i32 BUILD_VECTOR");
5924 unsigned LowElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
5925 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), MVT::v2i32,
5926 BVN->getOperand(LowElt), BVN->getOperand(LowElt+2));
5928 // Construct a new BUILD_VECTOR with elements truncated to half the size.
5929 assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR");
5930 EVT VT = N->getValueType(0);
5931 unsigned EltSize = VT.getVectorElementType().getSizeInBits() / 2;
5932 unsigned NumElts = VT.getVectorNumElements();
5933 MVT TruncVT = MVT::getIntegerVT(EltSize);
5934 SmallVector<SDValue, 8> Ops;
5936 for (unsigned i = 0; i != NumElts; ++i) {
5937 ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i));
5938 const APInt &CInt = C->getAPIntValue();
5939 // Element types smaller than 32 bits are not legal, so use i32 elements.
5940 // The values are implicitly truncated so sext vs. zext doesn't matter.
5941 Ops.push_back(DAG.getConstant(CInt.zextOrTrunc(32), dl, MVT::i32));
5943 return DAG.getNode(ISD::BUILD_VECTOR, dl,
5944 MVT::getVectorVT(TruncVT, NumElts), Ops);
5947 static bool isAddSubSExt(SDNode *N, SelectionDAG &DAG) {
5948 unsigned Opcode = N->getOpcode();
5949 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
5950 SDNode *N0 = N->getOperand(0).getNode();
5951 SDNode *N1 = N->getOperand(1).getNode();
5952 return N0->hasOneUse() && N1->hasOneUse() &&
5953 isSignExtended(N0, DAG) && isSignExtended(N1, DAG);
5958 static bool isAddSubZExt(SDNode *N, SelectionDAG &DAG) {
5959 unsigned Opcode = N->getOpcode();
5960 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
5961 SDNode *N0 = N->getOperand(0).getNode();
5962 SDNode *N1 = N->getOperand(1).getNode();
5963 return N0->hasOneUse() && N1->hasOneUse() &&
5964 isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG);
5969 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
5970 // Multiplications are only custom-lowered for 128-bit vectors so that
5971 // VMULL can be detected. Otherwise v2i64 multiplications are not legal.
5972 EVT VT = Op.getValueType();
5973 assert(VT.is128BitVector() && VT.isInteger() &&
5974 "unexpected type for custom-lowering ISD::MUL");
5975 SDNode *N0 = Op.getOperand(0).getNode();
5976 SDNode *N1 = Op.getOperand(1).getNode();
5977 unsigned NewOpc = 0;
5979 bool isN0SExt = isSignExtended(N0, DAG);
5980 bool isN1SExt = isSignExtended(N1, DAG);
5981 if (isN0SExt && isN1SExt)
5982 NewOpc = ARMISD::VMULLs;
5984 bool isN0ZExt = isZeroExtended(N0, DAG);
5985 bool isN1ZExt = isZeroExtended(N1, DAG);
5986 if (isN0ZExt && isN1ZExt)
5987 NewOpc = ARMISD::VMULLu;
5988 else if (isN1SExt || isN1ZExt) {
5989 // Look for (s/zext A + s/zext B) * (s/zext C). We want to turn these
5990 // into (s/zext A * s/zext C) + (s/zext B * s/zext C)
5991 if (isN1SExt && isAddSubSExt(N0, DAG)) {
5992 NewOpc = ARMISD::VMULLs;
5994 } else if (isN1ZExt && isAddSubZExt(N0, DAG)) {
5995 NewOpc = ARMISD::VMULLu;
5997 } else if (isN0ZExt && isAddSubZExt(N1, DAG)) {
5999 NewOpc = ARMISD::VMULLu;
6005 if (VT == MVT::v2i64)
6006 // Fall through to expand this. It is not legal.
6009 // Other vector multiplications are legal.
6014 // Legalize to a VMULL instruction.
6017 SDValue Op1 = SkipExtensionForVMULL(N1, DAG);
6019 Op0 = SkipExtensionForVMULL(N0, DAG);
6020 assert(Op0.getValueType().is64BitVector() &&
6021 Op1.getValueType().is64BitVector() &&
6022 "unexpected types for extended operands to VMULL");
6023 return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
6026 // Optimizing (zext A + zext B) * C, to (VMULL A, C) + (VMULL B, C) during
6027 // isel lowering to take advantage of no-stall back to back vmul + vmla.
6034 SDValue N00 = SkipExtensionForVMULL(N0->getOperand(0).getNode(), DAG);
6035 SDValue N01 = SkipExtensionForVMULL(N0->getOperand(1).getNode(), DAG);
6036 EVT Op1VT = Op1.getValueType();
6037 return DAG.getNode(N0->getOpcode(), DL, VT,
6038 DAG.getNode(NewOpc, DL, VT,
6039 DAG.getNode(ISD::BITCAST, DL, Op1VT, N00), Op1),
6040 DAG.getNode(NewOpc, DL, VT,
6041 DAG.getNode(ISD::BITCAST, DL, Op1VT, N01), Op1));
6045 LowerSDIV_v4i8(SDValue X, SDValue Y, SDLoc dl, SelectionDAG &DAG) {
6047 // float4 xf = vcvt_f32_s32(vmovl_s16(a.lo));
6048 // float4 yf = vcvt_f32_s32(vmovl_s16(b.lo));
6049 X = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, X);
6050 Y = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, Y);
6051 X = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, X);
6052 Y = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, Y);
6053 // Get reciprocal estimate.
6054 // float4 recip = vrecpeq_f32(yf);
6055 Y = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6056 DAG.getConstant(Intrinsic::arm_neon_vrecpe, dl, MVT::i32),
6058 // Because char has a smaller range than uchar, we can actually get away
6059 // without any newton steps. This requires that we use a weird bias
6060 // of 0xb000, however (again, this has been exhaustively tested).
6061 // float4 result = as_float4(as_int4(xf*recip) + 0xb000);
6062 X = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, X, Y);
6063 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, X);
6064 Y = DAG.getConstant(0xb000, dl, MVT::i32);
6065 Y = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Y, Y, Y, Y);
6066 X = DAG.getNode(ISD::ADD, dl, MVT::v4i32, X, Y);
6067 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, X);
6068 // Convert back to short.
6069 X = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, X);
6070 X = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, X);
6075 LowerSDIV_v4i16(SDValue N0, SDValue N1, SDLoc dl, SelectionDAG &DAG) {
6077 // Convert to float.
6078 // float4 yf = vcvt_f32_s32(vmovl_s16(y));
6079 // float4 xf = vcvt_f32_s32(vmovl_s16(x));
6080 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N0);
6081 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N1);
6082 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
6083 N1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
6085 // Use reciprocal estimate and one refinement step.
6086 // float4 recip = vrecpeq_f32(yf);
6087 // recip *= vrecpsq_f32(yf, recip);
6088 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6089 DAG.getConstant(Intrinsic::arm_neon_vrecpe, dl, MVT::i32),
6091 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6092 DAG.getConstant(Intrinsic::arm_neon_vrecps, dl, MVT::i32),
6094 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
6095 // Because short has a smaller range than ushort, we can actually get away
6096 // with only a single newton step. This requires that we use a weird bias
6097 // of 89, however (again, this has been exhaustively tested).
6098 // float4 result = as_float4(as_int4(xf*recip) + 0x89);
6099 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
6100 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
6101 N1 = DAG.getConstant(0x89, dl, MVT::i32);
6102 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
6103 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
6104 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
6105 // Convert back to integer and return.
6106 // return vmovn_s32(vcvt_s32_f32(result));
6107 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
6108 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
6112 static SDValue LowerSDIV(SDValue Op, SelectionDAG &DAG) {
6113 EVT VT = Op.getValueType();
6114 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
6115 "unexpected type for custom-lowering ISD::SDIV");
6118 SDValue N0 = Op.getOperand(0);
6119 SDValue N1 = Op.getOperand(1);
6122 if (VT == MVT::v8i8) {
6123 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N0);
6124 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N1);
6126 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
6127 DAG.getIntPtrConstant(4, dl));
6128 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
6129 DAG.getIntPtrConstant(4, dl));
6130 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
6131 DAG.getIntPtrConstant(0, dl));
6132 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
6133 DAG.getIntPtrConstant(0, dl));
6135 N0 = LowerSDIV_v4i8(N0, N1, dl, DAG); // v4i16
6136 N2 = LowerSDIV_v4i8(N2, N3, dl, DAG); // v4i16
6138 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
6139 N0 = LowerCONCAT_VECTORS(N0, DAG);
6141 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v8i8, N0);
6144 return LowerSDIV_v4i16(N0, N1, dl, DAG);
6147 static SDValue LowerUDIV(SDValue Op, SelectionDAG &DAG) {
6148 EVT VT = Op.getValueType();
6149 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
6150 "unexpected type for custom-lowering ISD::UDIV");
6153 SDValue N0 = Op.getOperand(0);
6154 SDValue N1 = Op.getOperand(1);
6157 if (VT == MVT::v8i8) {
6158 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N0);
6159 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N1);
6161 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
6162 DAG.getIntPtrConstant(4, dl));
6163 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
6164 DAG.getIntPtrConstant(4, dl));
6165 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
6166 DAG.getIntPtrConstant(0, dl));
6167 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
6168 DAG.getIntPtrConstant(0, dl));
6170 N0 = LowerSDIV_v4i16(N0, N1, dl, DAG); // v4i16
6171 N2 = LowerSDIV_v4i16(N2, N3, dl, DAG); // v4i16
6173 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
6174 N0 = LowerCONCAT_VECTORS(N0, DAG);
6176 N0 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v8i8,
6177 DAG.getConstant(Intrinsic::arm_neon_vqmovnsu, dl,
6183 // v4i16 sdiv ... Convert to float.
6184 // float4 yf = vcvt_f32_s32(vmovl_u16(y));
6185 // float4 xf = vcvt_f32_s32(vmovl_u16(x));
6186 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N0);
6187 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N1);
6188 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
6189 SDValue BN1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
6191 // Use reciprocal estimate and two refinement steps.
6192 // float4 recip = vrecpeq_f32(yf);
6193 // recip *= vrecpsq_f32(yf, recip);
6194 // recip *= vrecpsq_f32(yf, recip);
6195 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6196 DAG.getConstant(Intrinsic::arm_neon_vrecpe, dl, MVT::i32),
6198 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6199 DAG.getConstant(Intrinsic::arm_neon_vrecps, dl, MVT::i32),
6201 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
6202 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6203 DAG.getConstant(Intrinsic::arm_neon_vrecps, dl, MVT::i32),
6205 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
6206 // Simply multiplying by the reciprocal estimate can leave us a few ulps
6207 // too low, so we add 2 ulps (exhaustive testing shows that this is enough,
6208 // and that it will never cause us to return an answer too large).
6209 // float4 result = as_float4(as_int4(xf*recip) + 2);
6210 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
6211 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
6212 N1 = DAG.getConstant(2, dl, MVT::i32);
6213 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
6214 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
6215 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
6216 // Convert back to integer and return.
6217 // return vmovn_u32(vcvt_s32_f32(result));
6218 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
6219 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
6223 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
6224 EVT VT = Op.getNode()->getValueType(0);
6225 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
6228 bool ExtraOp = false;
6229 switch (Op.getOpcode()) {
6230 default: llvm_unreachable("Invalid code");
6231 case ISD::ADDC: Opc = ARMISD::ADDC; break;
6232 case ISD::ADDE: Opc = ARMISD::ADDE; ExtraOp = true; break;
6233 case ISD::SUBC: Opc = ARMISD::SUBC; break;
6234 case ISD::SUBE: Opc = ARMISD::SUBE; ExtraOp = true; break;
6238 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
6240 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
6241 Op.getOperand(1), Op.getOperand(2));
6244 SDValue ARMTargetLowering::LowerFSINCOS(SDValue Op, SelectionDAG &DAG) const {
6245 assert(Subtarget->isTargetDarwin());
6247 // For iOS, we want to call an alternative entry point: __sincos_stret,
6248 // return values are passed via sret.
6250 SDValue Arg = Op.getOperand(0);
6251 EVT ArgVT = Arg.getValueType();
6252 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
6254 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
6255 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6257 // Pair of floats / doubles used to pass the result.
6258 StructType *RetTy = StructType::get(ArgTy, ArgTy, nullptr);
6260 // Create stack object for sret.
6261 const uint64_t ByteSize = TLI.getDataLayout()->getTypeAllocSize(RetTy);
6262 const unsigned StackAlign = TLI.getDataLayout()->getPrefTypeAlignment(RetTy);
6263 int FrameIdx = FrameInfo->CreateStackObject(ByteSize, StackAlign, false);
6264 SDValue SRet = DAG.getFrameIndex(FrameIdx, TLI.getPointerTy());
6270 Entry.Ty = RetTy->getPointerTo();
6271 Entry.isSExt = false;
6272 Entry.isZExt = false;
6273 Entry.isSRet = true;
6274 Args.push_back(Entry);
6278 Entry.isSExt = false;
6279 Entry.isZExt = false;
6280 Args.push_back(Entry);
6282 const char *LibcallName = (ArgVT == MVT::f64)
6283 ? "__sincos_stret" : "__sincosf_stret";
6284 SDValue Callee = DAG.getExternalSymbol(LibcallName, getPointerTy());
6286 TargetLowering::CallLoweringInfo CLI(DAG);
6287 CLI.setDebugLoc(dl).setChain(DAG.getEntryNode())
6288 .setCallee(CallingConv::C, Type::getVoidTy(*DAG.getContext()), Callee,
6290 .setDiscardResult();
6292 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
6294 SDValue LoadSin = DAG.getLoad(ArgVT, dl, CallResult.second, SRet,
6295 MachinePointerInfo(), false, false, false, 0);
6297 // Address of cos field.
6298 SDValue Add = DAG.getNode(ISD::ADD, dl, getPointerTy(), SRet,
6299 DAG.getIntPtrConstant(ArgVT.getStoreSize(), dl));
6300 SDValue LoadCos = DAG.getLoad(ArgVT, dl, LoadSin.getValue(1), Add,
6301 MachinePointerInfo(), false, false, false, 0);
6303 SDVTList Tys = DAG.getVTList(ArgVT, ArgVT);
6304 return DAG.getNode(ISD::MERGE_VALUES, dl, Tys,
6305 LoadSin.getValue(0), LoadCos.getValue(0));
6308 static SDValue LowerAtomicLoadStore(SDValue Op, SelectionDAG &DAG) {
6309 // Monotonic load/store is legal for all targets
6310 if (cast<AtomicSDNode>(Op)->getOrdering() <= Monotonic)
6313 // Acquire/Release load/store is not legal for targets without a
6314 // dmb or equivalent available.
6318 static void ReplaceREADCYCLECOUNTER(SDNode *N,
6319 SmallVectorImpl<SDValue> &Results,
6321 const ARMSubtarget *Subtarget) {
6323 SDValue Cycles32, OutChain;
6325 if (Subtarget->hasPerfMon()) {
6326 // Under Power Management extensions, the cycle-count is:
6327 // mrc p15, #0, <Rt>, c9, c13, #0
6328 SDValue Ops[] = { N->getOperand(0), // Chain
6329 DAG.getConstant(Intrinsic::arm_mrc, DL, MVT::i32),
6330 DAG.getConstant(15, DL, MVT::i32),
6331 DAG.getConstant(0, DL, MVT::i32),
6332 DAG.getConstant(9, DL, MVT::i32),
6333 DAG.getConstant(13, DL, MVT::i32),
6334 DAG.getConstant(0, DL, MVT::i32)
6337 Cycles32 = DAG.getNode(ISD::INTRINSIC_W_CHAIN, DL,
6338 DAG.getVTList(MVT::i32, MVT::Other), Ops);
6339 OutChain = Cycles32.getValue(1);
6341 // Intrinsic is defined to return 0 on unsupported platforms. Technically
6342 // there are older ARM CPUs that have implementation-specific ways of
6343 // obtaining this information (FIXME!).
6344 Cycles32 = DAG.getConstant(0, DL, MVT::i32);
6345 OutChain = DAG.getEntryNode();
6349 SDValue Cycles64 = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64,
6350 Cycles32, DAG.getConstant(0, DL, MVT::i32));
6351 Results.push_back(Cycles64);
6352 Results.push_back(OutChain);
6355 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
6356 switch (Op.getOpcode()) {
6357 default: llvm_unreachable("Don't know how to custom lower this!");
6358 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
6359 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
6360 case ISD::GlobalAddress:
6361 switch (Subtarget->getTargetTriple().getObjectFormat()) {
6362 default: llvm_unreachable("unknown object format");
6364 return LowerGlobalAddressWindows(Op, DAG);
6366 return LowerGlobalAddressELF(Op, DAG);
6368 return LowerGlobalAddressDarwin(Op, DAG);
6370 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
6371 case ISD::SELECT: return LowerSELECT(Op, DAG);
6372 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
6373 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
6374 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
6375 case ISD::VASTART: return LowerVASTART(Op, DAG);
6376 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, DAG, Subtarget);
6377 case ISD::PREFETCH: return LowerPREFETCH(Op, DAG, Subtarget);
6378 case ISD::SINT_TO_FP:
6379 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
6380 case ISD::FP_TO_SINT:
6381 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
6382 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
6383 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
6384 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
6385 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
6386 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
6387 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
6388 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
6390 case ISD::BITCAST: return ExpandBITCAST(Op.getNode(), DAG);
6393 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
6394 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
6395 case ISD::SRL_PARTS:
6396 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
6397 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
6398 case ISD::CTPOP: return LowerCTPOP(Op.getNode(), DAG, Subtarget);
6399 case ISD::SETCC: return LowerVSETCC(Op, DAG);
6400 case ISD::ConstantFP: return LowerConstantFP(Op, DAG, Subtarget);
6401 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG, Subtarget);
6402 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
6403 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
6404 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
6405 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
6406 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
6407 case ISD::MUL: return LowerMUL(Op, DAG);
6408 case ISD::SDIV: return LowerSDIV(Op, DAG);
6409 case ISD::UDIV: return LowerUDIV(Op, DAG);
6413 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
6418 return LowerXALUO(Op, DAG);
6419 case ISD::ATOMIC_LOAD:
6420 case ISD::ATOMIC_STORE: return LowerAtomicLoadStore(Op, DAG);
6421 case ISD::FSINCOS: return LowerFSINCOS(Op, DAG);
6423 case ISD::UDIVREM: return LowerDivRem(Op, DAG);
6424 case ISD::DYNAMIC_STACKALLOC:
6425 if (Subtarget->getTargetTriple().isWindowsItaniumEnvironment())
6426 return LowerDYNAMIC_STACKALLOC(Op, DAG);
6427 llvm_unreachable("Don't know how to custom lower this!");
6428 case ISD::FP_ROUND: return LowerFP_ROUND(Op, DAG);
6429 case ISD::FP_EXTEND: return LowerFP_EXTEND(Op, DAG);
6433 /// ReplaceNodeResults - Replace the results of node with an illegal result
6434 /// type with new values built out of custom code.
6435 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
6436 SmallVectorImpl<SDValue>&Results,
6437 SelectionDAG &DAG) const {
6439 switch (N->getOpcode()) {
6441 llvm_unreachable("Don't know how to custom expand this!");
6443 Res = ExpandBITCAST(N, DAG);
6447 Res = Expand64BitShift(N, DAG, Subtarget);
6449 case ISD::READCYCLECOUNTER:
6450 ReplaceREADCYCLECOUNTER(N, Results, DAG, Subtarget);
6454 Results.push_back(Res);
6457 //===----------------------------------------------------------------------===//
6458 // ARM Scheduler Hooks
6459 //===----------------------------------------------------------------------===//
6461 /// SetupEntryBlockForSjLj - Insert code into the entry block that creates and
6462 /// registers the function context.
6463 void ARMTargetLowering::
6464 SetupEntryBlockForSjLj(MachineInstr *MI, MachineBasicBlock *MBB,
6465 MachineBasicBlock *DispatchBB, int FI) const {
6466 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
6467 DebugLoc dl = MI->getDebugLoc();
6468 MachineFunction *MF = MBB->getParent();
6469 MachineRegisterInfo *MRI = &MF->getRegInfo();
6470 MachineConstantPool *MCP = MF->getConstantPool();
6471 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
6472 const Function *F = MF->getFunction();
6474 bool isThumb = Subtarget->isThumb();
6475 bool isThumb2 = Subtarget->isThumb2();
6477 unsigned PCLabelId = AFI->createPICLabelUId();
6478 unsigned PCAdj = (isThumb || isThumb2) ? 4 : 8;
6479 ARMConstantPoolValue *CPV =
6480 ARMConstantPoolMBB::Create(F->getContext(), DispatchBB, PCLabelId, PCAdj);
6481 unsigned CPI = MCP->getConstantPoolIndex(CPV, 4);
6483 const TargetRegisterClass *TRC = isThumb ? &ARM::tGPRRegClass
6484 : &ARM::GPRRegClass;
6486 // Grab constant pool and fixed stack memory operands.
6487 MachineMemOperand *CPMMO =
6488 MF->getMachineMemOperand(MachinePointerInfo::getConstantPool(),
6489 MachineMemOperand::MOLoad, 4, 4);
6491 MachineMemOperand *FIMMOSt =
6492 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
6493 MachineMemOperand::MOStore, 4, 4);
6495 // Load the address of the dispatch MBB into the jump buffer.
6497 // Incoming value: jbuf
6498 // ldr.n r5, LCPI1_1
6501 // str r5, [$jbuf, #+4] ; &jbuf[1]
6502 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6503 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2LDRpci), NewVReg1)
6504 .addConstantPoolIndex(CPI)
6505 .addMemOperand(CPMMO));
6506 // Set the low bit because of thumb mode.
6507 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6509 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2ORRri), NewVReg2)
6510 .addReg(NewVReg1, RegState::Kill)
6512 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6513 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg3)
6514 .addReg(NewVReg2, RegState::Kill)
6516 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2STRi12))
6517 .addReg(NewVReg3, RegState::Kill)
6519 .addImm(36) // &jbuf[1] :: pc
6520 .addMemOperand(FIMMOSt));
6521 } else if (isThumb) {
6522 // Incoming value: jbuf
6523 // ldr.n r1, LCPI1_4
6527 // add r2, $jbuf, #+4 ; &jbuf[1]
6529 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6530 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tLDRpci), NewVReg1)
6531 .addConstantPoolIndex(CPI)
6532 .addMemOperand(CPMMO));
6533 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6534 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg2)
6535 .addReg(NewVReg1, RegState::Kill)
6537 // Set the low bit because of thumb mode.
6538 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6539 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tMOVi8), NewVReg3)
6540 .addReg(ARM::CPSR, RegState::Define)
6542 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6543 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tORR), NewVReg4)
6544 .addReg(ARM::CPSR, RegState::Define)
6545 .addReg(NewVReg2, RegState::Kill)
6546 .addReg(NewVReg3, RegState::Kill));
6547 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6548 BuildMI(*MBB, MI, dl, TII->get(ARM::tADDframe), NewVReg5)
6550 .addImm(36); // &jbuf[1] :: pc
6551 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tSTRi))
6552 .addReg(NewVReg4, RegState::Kill)
6553 .addReg(NewVReg5, RegState::Kill)
6555 .addMemOperand(FIMMOSt));
6557 // Incoming value: jbuf
6560 // str r1, [$jbuf, #+4] ; &jbuf[1]
6561 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6562 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::LDRi12), NewVReg1)
6563 .addConstantPoolIndex(CPI)
6565 .addMemOperand(CPMMO));
6566 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6567 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::PICADD), NewVReg2)
6568 .addReg(NewVReg1, RegState::Kill)
6569 .addImm(PCLabelId));
6570 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::STRi12))
6571 .addReg(NewVReg2, RegState::Kill)
6573 .addImm(36) // &jbuf[1] :: pc
6574 .addMemOperand(FIMMOSt));
6578 void ARMTargetLowering::EmitSjLjDispatchBlock(MachineInstr *MI,
6579 MachineBasicBlock *MBB) const {
6580 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
6581 DebugLoc dl = MI->getDebugLoc();
6582 MachineFunction *MF = MBB->getParent();
6583 MachineRegisterInfo *MRI = &MF->getRegInfo();
6584 MachineFrameInfo *MFI = MF->getFrameInfo();
6585 int FI = MFI->getFunctionContextIndex();
6587 const TargetRegisterClass *TRC = Subtarget->isThumb() ? &ARM::tGPRRegClass
6588 : &ARM::GPRnopcRegClass;
6590 // Get a mapping of the call site numbers to all of the landing pads they're
6592 DenseMap<unsigned, SmallVector<MachineBasicBlock*, 2> > CallSiteNumToLPad;
6593 unsigned MaxCSNum = 0;
6594 MachineModuleInfo &MMI = MF->getMMI();
6595 for (MachineFunction::iterator BB = MF->begin(), E = MF->end(); BB != E;
6597 if (!BB->isLandingPad()) continue;
6599 // FIXME: We should assert that the EH_LABEL is the first MI in the landing
6601 for (MachineBasicBlock::iterator
6602 II = BB->begin(), IE = BB->end(); II != IE; ++II) {
6603 if (!II->isEHLabel()) continue;
6605 MCSymbol *Sym = II->getOperand(0).getMCSymbol();
6606 if (!MMI.hasCallSiteLandingPad(Sym)) continue;
6608 SmallVectorImpl<unsigned> &CallSiteIdxs = MMI.getCallSiteLandingPad(Sym);
6609 for (SmallVectorImpl<unsigned>::iterator
6610 CSI = CallSiteIdxs.begin(), CSE = CallSiteIdxs.end();
6611 CSI != CSE; ++CSI) {
6612 CallSiteNumToLPad[*CSI].push_back(BB);
6613 MaxCSNum = std::max(MaxCSNum, *CSI);
6619 // Get an ordered list of the machine basic blocks for the jump table.
6620 std::vector<MachineBasicBlock*> LPadList;
6621 SmallPtrSet<MachineBasicBlock*, 64> InvokeBBs;
6622 LPadList.reserve(CallSiteNumToLPad.size());
6623 for (unsigned I = 1; I <= MaxCSNum; ++I) {
6624 SmallVectorImpl<MachineBasicBlock*> &MBBList = CallSiteNumToLPad[I];
6625 for (SmallVectorImpl<MachineBasicBlock*>::iterator
6626 II = MBBList.begin(), IE = MBBList.end(); II != IE; ++II) {
6627 LPadList.push_back(*II);
6628 InvokeBBs.insert((*II)->pred_begin(), (*II)->pred_end());
6632 assert(!LPadList.empty() &&
6633 "No landing pad destinations for the dispatch jump table!");
6635 // Create the jump table and associated information.
6636 MachineJumpTableInfo *JTI =
6637 MF->getOrCreateJumpTableInfo(MachineJumpTableInfo::EK_Inline);
6638 unsigned MJTI = JTI->createJumpTableIndex(LPadList);
6639 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
6641 // Create the MBBs for the dispatch code.
6643 // Shove the dispatch's address into the return slot in the function context.
6644 MachineBasicBlock *DispatchBB = MF->CreateMachineBasicBlock();
6645 DispatchBB->setIsLandingPad();
6647 MachineBasicBlock *TrapBB = MF->CreateMachineBasicBlock();
6648 unsigned trap_opcode;
6649 if (Subtarget->isThumb())
6650 trap_opcode = ARM::tTRAP;
6652 trap_opcode = Subtarget->useNaClTrap() ? ARM::TRAPNaCl : ARM::TRAP;
6654 BuildMI(TrapBB, dl, TII->get(trap_opcode));
6655 DispatchBB->addSuccessor(TrapBB);
6657 MachineBasicBlock *DispContBB = MF->CreateMachineBasicBlock();
6658 DispatchBB->addSuccessor(DispContBB);
6661 MF->insert(MF->end(), DispatchBB);
6662 MF->insert(MF->end(), DispContBB);
6663 MF->insert(MF->end(), TrapBB);
6665 // Insert code into the entry block that creates and registers the function
6667 SetupEntryBlockForSjLj(MI, MBB, DispatchBB, FI);
6669 MachineMemOperand *FIMMOLd =
6670 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
6671 MachineMemOperand::MOLoad |
6672 MachineMemOperand::MOVolatile, 4, 4);
6674 MachineInstrBuilder MIB;
6675 MIB = BuildMI(DispatchBB, dl, TII->get(ARM::Int_eh_sjlj_dispatchsetup));
6677 const ARMBaseInstrInfo *AII = static_cast<const ARMBaseInstrInfo*>(TII);
6678 const ARMBaseRegisterInfo &RI = AII->getRegisterInfo();
6680 // Add a register mask with no preserved registers. This results in all
6681 // registers being marked as clobbered.
6682 MIB.addRegMask(RI.getNoPreservedMask());
6684 unsigned NumLPads = LPadList.size();
6685 if (Subtarget->isThumb2()) {
6686 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6687 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2LDRi12), NewVReg1)
6690 .addMemOperand(FIMMOLd));
6692 if (NumLPads < 256) {
6693 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPri))
6695 .addImm(LPadList.size()));
6697 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6698 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVi16), VReg1)
6699 .addImm(NumLPads & 0xFFFF));
6701 unsigned VReg2 = VReg1;
6702 if ((NumLPads & 0xFFFF0000) != 0) {
6703 VReg2 = MRI->createVirtualRegister(TRC);
6704 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVTi16), VReg2)
6706 .addImm(NumLPads >> 16));
6709 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPrr))
6714 BuildMI(DispatchBB, dl, TII->get(ARM::t2Bcc))
6719 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6720 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::t2LEApcrelJT),NewVReg3)
6721 .addJumpTableIndex(MJTI));
6723 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6726 BuildMI(DispContBB, dl, TII->get(ARM::t2ADDrs), NewVReg4)
6727 .addReg(NewVReg3, RegState::Kill)
6729 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
6731 BuildMI(DispContBB, dl, TII->get(ARM::t2BR_JT))
6732 .addReg(NewVReg4, RegState::Kill)
6734 .addJumpTableIndex(MJTI);
6735 } else if (Subtarget->isThumb()) {
6736 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6737 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRspi), NewVReg1)
6740 .addMemOperand(FIMMOLd));
6742 if (NumLPads < 256) {
6743 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPi8))
6747 MachineConstantPool *ConstantPool = MF->getConstantPool();
6748 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
6749 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
6751 // MachineConstantPool wants an explicit alignment.
6752 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
6754 Align = getDataLayout()->getTypeAllocSize(C->getType());
6755 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6757 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6758 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRpci))
6759 .addReg(VReg1, RegState::Define)
6760 .addConstantPoolIndex(Idx));
6761 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPr))
6766 BuildMI(DispatchBB, dl, TII->get(ARM::tBcc))
6771 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6772 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLSLri), NewVReg2)
6773 .addReg(ARM::CPSR, RegState::Define)
6777 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6778 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLEApcrelJT), NewVReg3)
6779 .addJumpTableIndex(MJTI));
6781 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6782 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg4)
6783 .addReg(ARM::CPSR, RegState::Define)
6784 .addReg(NewVReg2, RegState::Kill)
6787 MachineMemOperand *JTMMOLd =
6788 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
6789 MachineMemOperand::MOLoad, 4, 4);
6791 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6792 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLDRi), NewVReg5)
6793 .addReg(NewVReg4, RegState::Kill)
6795 .addMemOperand(JTMMOLd));
6797 unsigned NewVReg6 = NewVReg5;
6798 if (RelocM == Reloc::PIC_) {
6799 NewVReg6 = MRI->createVirtualRegister(TRC);
6800 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg6)
6801 .addReg(ARM::CPSR, RegState::Define)
6802 .addReg(NewVReg5, RegState::Kill)
6806 BuildMI(DispContBB, dl, TII->get(ARM::tBR_JTr))
6807 .addReg(NewVReg6, RegState::Kill)
6808 .addJumpTableIndex(MJTI);
6810 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6811 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRi12), NewVReg1)
6814 .addMemOperand(FIMMOLd));
6816 if (NumLPads < 256) {
6817 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPri))
6820 } else if (Subtarget->hasV6T2Ops() && isUInt<16>(NumLPads)) {
6821 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6822 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVi16), VReg1)
6823 .addImm(NumLPads & 0xFFFF));
6825 unsigned VReg2 = VReg1;
6826 if ((NumLPads & 0xFFFF0000) != 0) {
6827 VReg2 = MRI->createVirtualRegister(TRC);
6828 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVTi16), VReg2)
6830 .addImm(NumLPads >> 16));
6833 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
6837 MachineConstantPool *ConstantPool = MF->getConstantPool();
6838 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
6839 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
6841 // MachineConstantPool wants an explicit alignment.
6842 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
6844 Align = getDataLayout()->getTypeAllocSize(C->getType());
6845 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6847 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6848 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRcp))
6849 .addReg(VReg1, RegState::Define)
6850 .addConstantPoolIndex(Idx)
6852 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
6854 .addReg(VReg1, RegState::Kill));
6857 BuildMI(DispatchBB, dl, TII->get(ARM::Bcc))
6862 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6864 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::MOVsi), NewVReg3)
6866 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
6867 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6868 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::LEApcrelJT), NewVReg4)
6869 .addJumpTableIndex(MJTI));
6871 MachineMemOperand *JTMMOLd =
6872 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
6873 MachineMemOperand::MOLoad, 4, 4);
6874 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6876 BuildMI(DispContBB, dl, TII->get(ARM::LDRrs), NewVReg5)
6877 .addReg(NewVReg3, RegState::Kill)
6880 .addMemOperand(JTMMOLd));
6882 if (RelocM == Reloc::PIC_) {
6883 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTadd))
6884 .addReg(NewVReg5, RegState::Kill)
6886 .addJumpTableIndex(MJTI);
6888 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTr))
6889 .addReg(NewVReg5, RegState::Kill)
6890 .addJumpTableIndex(MJTI);
6894 // Add the jump table entries as successors to the MBB.
6895 SmallPtrSet<MachineBasicBlock*, 8> SeenMBBs;
6896 for (std::vector<MachineBasicBlock*>::iterator
6897 I = LPadList.begin(), E = LPadList.end(); I != E; ++I) {
6898 MachineBasicBlock *CurMBB = *I;
6899 if (SeenMBBs.insert(CurMBB).second)
6900 DispContBB->addSuccessor(CurMBB);
6903 // N.B. the order the invoke BBs are processed in doesn't matter here.
6904 const MCPhysReg *SavedRegs = RI.getCalleeSavedRegs(MF);
6905 SmallVector<MachineBasicBlock*, 64> MBBLPads;
6906 for (MachineBasicBlock *BB : InvokeBBs) {
6908 // Remove the landing pad successor from the invoke block and replace it
6909 // with the new dispatch block.
6910 SmallVector<MachineBasicBlock*, 4> Successors(BB->succ_begin(),
6912 while (!Successors.empty()) {
6913 MachineBasicBlock *SMBB = Successors.pop_back_val();
6914 if (SMBB->isLandingPad()) {
6915 BB->removeSuccessor(SMBB);
6916 MBBLPads.push_back(SMBB);
6920 BB->addSuccessor(DispatchBB);
6922 // Find the invoke call and mark all of the callee-saved registers as
6923 // 'implicit defined' so that they're spilled. This prevents code from
6924 // moving instructions to before the EH block, where they will never be
6926 for (MachineBasicBlock::reverse_iterator
6927 II = BB->rbegin(), IE = BB->rend(); II != IE; ++II) {
6928 if (!II->isCall()) continue;
6930 DenseMap<unsigned, bool> DefRegs;
6931 for (MachineInstr::mop_iterator
6932 OI = II->operands_begin(), OE = II->operands_end();
6934 if (!OI->isReg()) continue;
6935 DefRegs[OI->getReg()] = true;
6938 MachineInstrBuilder MIB(*MF, &*II);
6940 for (unsigned i = 0; SavedRegs[i] != 0; ++i) {
6941 unsigned Reg = SavedRegs[i];
6942 if (Subtarget->isThumb2() &&
6943 !ARM::tGPRRegClass.contains(Reg) &&
6944 !ARM::hGPRRegClass.contains(Reg))
6946 if (Subtarget->isThumb1Only() && !ARM::tGPRRegClass.contains(Reg))
6948 if (!Subtarget->isThumb() && !ARM::GPRRegClass.contains(Reg))
6951 MIB.addReg(Reg, RegState::ImplicitDefine | RegState::Dead);
6958 // Mark all former landing pads as non-landing pads. The dispatch is the only
6960 for (SmallVectorImpl<MachineBasicBlock*>::iterator
6961 I = MBBLPads.begin(), E = MBBLPads.end(); I != E; ++I)
6962 (*I)->setIsLandingPad(false);
6964 // The instruction is gone now.
6965 MI->eraseFromParent();
6969 MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) {
6970 for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
6971 E = MBB->succ_end(); I != E; ++I)
6974 llvm_unreachable("Expecting a BB with two successors!");
6977 /// Return the load opcode for a given load size. If load size >= 8,
6978 /// neon opcode will be returned.
6979 static unsigned getLdOpcode(unsigned LdSize, bool IsThumb1, bool IsThumb2) {
6981 return LdSize == 16 ? ARM::VLD1q32wb_fixed
6982 : LdSize == 8 ? ARM::VLD1d32wb_fixed : 0;
6984 return LdSize == 4 ? ARM::tLDRi
6985 : LdSize == 2 ? ARM::tLDRHi
6986 : LdSize == 1 ? ARM::tLDRBi : 0;
6988 return LdSize == 4 ? ARM::t2LDR_POST
6989 : LdSize == 2 ? ARM::t2LDRH_POST
6990 : LdSize == 1 ? ARM::t2LDRB_POST : 0;
6991 return LdSize == 4 ? ARM::LDR_POST_IMM
6992 : LdSize == 2 ? ARM::LDRH_POST
6993 : LdSize == 1 ? ARM::LDRB_POST_IMM : 0;
6996 /// Return the store opcode for a given store size. If store size >= 8,
6997 /// neon opcode will be returned.
6998 static unsigned getStOpcode(unsigned StSize, bool IsThumb1, bool IsThumb2) {
7000 return StSize == 16 ? ARM::VST1q32wb_fixed
7001 : StSize == 8 ? ARM::VST1d32wb_fixed : 0;
7003 return StSize == 4 ? ARM::tSTRi
7004 : StSize == 2 ? ARM::tSTRHi
7005 : StSize == 1 ? ARM::tSTRBi : 0;
7007 return StSize == 4 ? ARM::t2STR_POST
7008 : StSize == 2 ? ARM::t2STRH_POST
7009 : StSize == 1 ? ARM::t2STRB_POST : 0;
7010 return StSize == 4 ? ARM::STR_POST_IMM
7011 : StSize == 2 ? ARM::STRH_POST
7012 : StSize == 1 ? ARM::STRB_POST_IMM : 0;
7015 /// Emit a post-increment load operation with given size. The instructions
7016 /// will be added to BB at Pos.
7017 static void emitPostLd(MachineBasicBlock *BB, MachineInstr *Pos,
7018 const TargetInstrInfo *TII, DebugLoc dl,
7019 unsigned LdSize, unsigned Data, unsigned AddrIn,
7020 unsigned AddrOut, bool IsThumb1, bool IsThumb2) {
7021 unsigned LdOpc = getLdOpcode(LdSize, IsThumb1, IsThumb2);
7022 assert(LdOpc != 0 && "Should have a load opcode");
7024 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
7025 .addReg(AddrOut, RegState::Define).addReg(AddrIn)
7027 } else if (IsThumb1) {
7028 // load + update AddrIn
7029 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
7030 .addReg(AddrIn).addImm(0));
7031 MachineInstrBuilder MIB =
7032 BuildMI(*BB, Pos, dl, TII->get(ARM::tADDi8), AddrOut);
7033 MIB = AddDefaultT1CC(MIB);
7034 MIB.addReg(AddrIn).addImm(LdSize);
7035 AddDefaultPred(MIB);
7036 } else if (IsThumb2) {
7037 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
7038 .addReg(AddrOut, RegState::Define).addReg(AddrIn)
7041 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
7042 .addReg(AddrOut, RegState::Define).addReg(AddrIn)
7043 .addReg(0).addImm(LdSize));
7047 /// Emit a post-increment store operation with given size. The instructions
7048 /// will be added to BB at Pos.
7049 static void emitPostSt(MachineBasicBlock *BB, MachineInstr *Pos,
7050 const TargetInstrInfo *TII, DebugLoc dl,
7051 unsigned StSize, unsigned Data, unsigned AddrIn,
7052 unsigned AddrOut, bool IsThumb1, bool IsThumb2) {
7053 unsigned StOpc = getStOpcode(StSize, IsThumb1, IsThumb2);
7054 assert(StOpc != 0 && "Should have a store opcode");
7056 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc), AddrOut)
7057 .addReg(AddrIn).addImm(0).addReg(Data));
7058 } else if (IsThumb1) {
7059 // store + update AddrIn
7060 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc)).addReg(Data)
7061 .addReg(AddrIn).addImm(0));
7062 MachineInstrBuilder MIB =
7063 BuildMI(*BB, Pos, dl, TII->get(ARM::tADDi8), AddrOut);
7064 MIB = AddDefaultT1CC(MIB);
7065 MIB.addReg(AddrIn).addImm(StSize);
7066 AddDefaultPred(MIB);
7067 } else if (IsThumb2) {
7068 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc), AddrOut)
7069 .addReg(Data).addReg(AddrIn).addImm(StSize));
7071 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc), AddrOut)
7072 .addReg(Data).addReg(AddrIn).addReg(0)
7078 ARMTargetLowering::EmitStructByval(MachineInstr *MI,
7079 MachineBasicBlock *BB) const {
7080 // This pseudo instruction has 3 operands: dst, src, size
7081 // We expand it to a loop if size > Subtarget->getMaxInlineSizeThreshold().
7082 // Otherwise, we will generate unrolled scalar copies.
7083 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
7084 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7085 MachineFunction::iterator It = BB;
7088 unsigned dest = MI->getOperand(0).getReg();
7089 unsigned src = MI->getOperand(1).getReg();
7090 unsigned SizeVal = MI->getOperand(2).getImm();
7091 unsigned Align = MI->getOperand(3).getImm();
7092 DebugLoc dl = MI->getDebugLoc();
7094 MachineFunction *MF = BB->getParent();
7095 MachineRegisterInfo &MRI = MF->getRegInfo();
7096 unsigned UnitSize = 0;
7097 const TargetRegisterClass *TRC = nullptr;
7098 const TargetRegisterClass *VecTRC = nullptr;
7100 bool IsThumb1 = Subtarget->isThumb1Only();
7101 bool IsThumb2 = Subtarget->isThumb2();
7105 } else if (Align & 2) {
7108 // Check whether we can use NEON instructions.
7109 if (!MF->getFunction()->hasFnAttribute(Attribute::NoImplicitFloat) &&
7110 Subtarget->hasNEON()) {
7111 if ((Align % 16 == 0) && SizeVal >= 16)
7113 else if ((Align % 8 == 0) && SizeVal >= 8)
7116 // Can't use NEON instructions.
7121 // Select the correct opcode and register class for unit size load/store
7122 bool IsNeon = UnitSize >= 8;
7123 TRC = (IsThumb1 || IsThumb2) ? &ARM::tGPRRegClass : &ARM::GPRRegClass;
7125 VecTRC = UnitSize == 16 ? &ARM::DPairRegClass
7126 : UnitSize == 8 ? &ARM::DPRRegClass
7129 unsigned BytesLeft = SizeVal % UnitSize;
7130 unsigned LoopSize = SizeVal - BytesLeft;
7132 if (SizeVal <= Subtarget->getMaxInlineSizeThreshold()) {
7133 // Use LDR and STR to copy.
7134 // [scratch, srcOut] = LDR_POST(srcIn, UnitSize)
7135 // [destOut] = STR_POST(scratch, destIn, UnitSize)
7136 unsigned srcIn = src;
7137 unsigned destIn = dest;
7138 for (unsigned i = 0; i < LoopSize; i+=UnitSize) {
7139 unsigned srcOut = MRI.createVirtualRegister(TRC);
7140 unsigned destOut = MRI.createVirtualRegister(TRC);
7141 unsigned scratch = MRI.createVirtualRegister(IsNeon ? VecTRC : TRC);
7142 emitPostLd(BB, MI, TII, dl, UnitSize, scratch, srcIn, srcOut,
7143 IsThumb1, IsThumb2);
7144 emitPostSt(BB, MI, TII, dl, UnitSize, scratch, destIn, destOut,
7145 IsThumb1, IsThumb2);
7150 // Handle the leftover bytes with LDRB and STRB.
7151 // [scratch, srcOut] = LDRB_POST(srcIn, 1)
7152 // [destOut] = STRB_POST(scratch, destIn, 1)
7153 for (unsigned i = 0; i < BytesLeft; i++) {
7154 unsigned srcOut = MRI.createVirtualRegister(TRC);
7155 unsigned destOut = MRI.createVirtualRegister(TRC);
7156 unsigned scratch = MRI.createVirtualRegister(TRC);
7157 emitPostLd(BB, MI, TII, dl, 1, scratch, srcIn, srcOut,
7158 IsThumb1, IsThumb2);
7159 emitPostSt(BB, MI, TII, dl, 1, scratch, destIn, destOut,
7160 IsThumb1, IsThumb2);
7164 MI->eraseFromParent(); // The instruction is gone now.
7168 // Expand the pseudo op to a loop.
7171 // movw varEnd, # --> with thumb2
7173 // ldrcp varEnd, idx --> without thumb2
7174 // fallthrough --> loopMBB
7176 // PHI varPhi, varEnd, varLoop
7177 // PHI srcPhi, src, srcLoop
7178 // PHI destPhi, dst, destLoop
7179 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
7180 // [destLoop] = STR_POST(scratch, destPhi, UnitSize)
7181 // subs varLoop, varPhi, #UnitSize
7183 // fallthrough --> exitMBB
7185 // epilogue to handle left-over bytes
7186 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
7187 // [destOut] = STRB_POST(scratch, destLoop, 1)
7188 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
7189 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
7190 MF->insert(It, loopMBB);
7191 MF->insert(It, exitMBB);
7193 // Transfer the remainder of BB and its successor edges to exitMBB.
7194 exitMBB->splice(exitMBB->begin(), BB,
7195 std::next(MachineBasicBlock::iterator(MI)), BB->end());
7196 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
7198 // Load an immediate to varEnd.
7199 unsigned varEnd = MRI.createVirtualRegister(TRC);
7200 if (Subtarget->useMovt(*MF)) {
7201 unsigned Vtmp = varEnd;
7202 if ((LoopSize & 0xFFFF0000) != 0)
7203 Vtmp = MRI.createVirtualRegister(TRC);
7204 AddDefaultPred(BuildMI(BB, dl,
7205 TII->get(IsThumb2 ? ARM::t2MOVi16 : ARM::MOVi16),
7206 Vtmp).addImm(LoopSize & 0xFFFF));
7208 if ((LoopSize & 0xFFFF0000) != 0)
7209 AddDefaultPred(BuildMI(BB, dl,
7210 TII->get(IsThumb2 ? ARM::t2MOVTi16 : ARM::MOVTi16),
7213 .addImm(LoopSize >> 16));
7215 MachineConstantPool *ConstantPool = MF->getConstantPool();
7216 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
7217 const Constant *C = ConstantInt::get(Int32Ty, LoopSize);
7219 // MachineConstantPool wants an explicit alignment.
7220 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
7222 Align = getDataLayout()->getTypeAllocSize(C->getType());
7223 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
7226 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(ARM::tLDRpci)).addReg(
7227 varEnd, RegState::Define).addConstantPoolIndex(Idx));
7229 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(ARM::LDRcp)).addReg(
7230 varEnd, RegState::Define).addConstantPoolIndex(Idx).addImm(0));
7232 BB->addSuccessor(loopMBB);
7234 // Generate the loop body:
7235 // varPhi = PHI(varLoop, varEnd)
7236 // srcPhi = PHI(srcLoop, src)
7237 // destPhi = PHI(destLoop, dst)
7238 MachineBasicBlock *entryBB = BB;
7240 unsigned varLoop = MRI.createVirtualRegister(TRC);
7241 unsigned varPhi = MRI.createVirtualRegister(TRC);
7242 unsigned srcLoop = MRI.createVirtualRegister(TRC);
7243 unsigned srcPhi = MRI.createVirtualRegister(TRC);
7244 unsigned destLoop = MRI.createVirtualRegister(TRC);
7245 unsigned destPhi = MRI.createVirtualRegister(TRC);
7247 BuildMI(*BB, BB->begin(), dl, TII->get(ARM::PHI), varPhi)
7248 .addReg(varLoop).addMBB(loopMBB)
7249 .addReg(varEnd).addMBB(entryBB);
7250 BuildMI(BB, dl, TII->get(ARM::PHI), srcPhi)
7251 .addReg(srcLoop).addMBB(loopMBB)
7252 .addReg(src).addMBB(entryBB);
7253 BuildMI(BB, dl, TII->get(ARM::PHI), destPhi)
7254 .addReg(destLoop).addMBB(loopMBB)
7255 .addReg(dest).addMBB(entryBB);
7257 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
7258 // [destLoop] = STR_POST(scratch, destPhi, UnitSiz)
7259 unsigned scratch = MRI.createVirtualRegister(IsNeon ? VecTRC : TRC);
7260 emitPostLd(BB, BB->end(), TII, dl, UnitSize, scratch, srcPhi, srcLoop,
7261 IsThumb1, IsThumb2);
7262 emitPostSt(BB, BB->end(), TII, dl, UnitSize, scratch, destPhi, destLoop,
7263 IsThumb1, IsThumb2);
7265 // Decrement loop variable by UnitSize.
7267 MachineInstrBuilder MIB =
7268 BuildMI(*BB, BB->end(), dl, TII->get(ARM::tSUBi8), varLoop);
7269 MIB = AddDefaultT1CC(MIB);
7270 MIB.addReg(varPhi).addImm(UnitSize);
7271 AddDefaultPred(MIB);
7273 MachineInstrBuilder MIB =
7274 BuildMI(*BB, BB->end(), dl,
7275 TII->get(IsThumb2 ? ARM::t2SUBri : ARM::SUBri), varLoop);
7276 AddDefaultCC(AddDefaultPred(MIB.addReg(varPhi).addImm(UnitSize)));
7277 MIB->getOperand(5).setReg(ARM::CPSR);
7278 MIB->getOperand(5).setIsDef(true);
7280 BuildMI(*BB, BB->end(), dl,
7281 TII->get(IsThumb1 ? ARM::tBcc : IsThumb2 ? ARM::t2Bcc : ARM::Bcc))
7282 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
7284 // loopMBB can loop back to loopMBB or fall through to exitMBB.
7285 BB->addSuccessor(loopMBB);
7286 BB->addSuccessor(exitMBB);
7288 // Add epilogue to handle BytesLeft.
7290 MachineInstr *StartOfExit = exitMBB->begin();
7292 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
7293 // [destOut] = STRB_POST(scratch, destLoop, 1)
7294 unsigned srcIn = srcLoop;
7295 unsigned destIn = destLoop;
7296 for (unsigned i = 0; i < BytesLeft; i++) {
7297 unsigned srcOut = MRI.createVirtualRegister(TRC);
7298 unsigned destOut = MRI.createVirtualRegister(TRC);
7299 unsigned scratch = MRI.createVirtualRegister(TRC);
7300 emitPostLd(BB, StartOfExit, TII, dl, 1, scratch, srcIn, srcOut,
7301 IsThumb1, IsThumb2);
7302 emitPostSt(BB, StartOfExit, TII, dl, 1, scratch, destIn, destOut,
7303 IsThumb1, IsThumb2);
7308 MI->eraseFromParent(); // The instruction is gone now.
7313 ARMTargetLowering::EmitLowered__chkstk(MachineInstr *MI,
7314 MachineBasicBlock *MBB) const {
7315 const TargetMachine &TM = getTargetMachine();
7316 const TargetInstrInfo &TII = *Subtarget->getInstrInfo();
7317 DebugLoc DL = MI->getDebugLoc();
7319 assert(Subtarget->isTargetWindows() &&
7320 "__chkstk is only supported on Windows");
7321 assert(Subtarget->isThumb2() && "Windows on ARM requires Thumb-2 mode");
7323 // __chkstk takes the number of words to allocate on the stack in R4, and
7324 // returns the stack adjustment in number of bytes in R4. This will not
7325 // clober any other registers (other than the obvious lr).
7327 // Although, technically, IP should be considered a register which may be
7328 // clobbered, the call itself will not touch it. Windows on ARM is a pure
7329 // thumb-2 environment, so there is no interworking required. As a result, we
7330 // do not expect a veneer to be emitted by the linker, clobbering IP.
7332 // Each module receives its own copy of __chkstk, so no import thunk is
7333 // required, again, ensuring that IP is not clobbered.
7335 // Finally, although some linkers may theoretically provide a trampoline for
7336 // out of range calls (which is quite common due to a 32M range limitation of
7337 // branches for Thumb), we can generate the long-call version via
7338 // -mcmodel=large, alleviating the need for the trampoline which may clobber
7341 switch (TM.getCodeModel()) {
7342 case CodeModel::Small:
7343 case CodeModel::Medium:
7344 case CodeModel::Default:
7345 case CodeModel::Kernel:
7346 BuildMI(*MBB, MI, DL, TII.get(ARM::tBL))
7347 .addImm((unsigned)ARMCC::AL).addReg(0)
7348 .addExternalSymbol("__chkstk")
7349 .addReg(ARM::R4, RegState::Implicit | RegState::Kill)
7350 .addReg(ARM::R4, RegState::Implicit | RegState::Define)
7351 .addReg(ARM::R12, RegState::Implicit | RegState::Define | RegState::Dead);
7353 case CodeModel::Large:
7354 case CodeModel::JITDefault: {
7355 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
7356 unsigned Reg = MRI.createVirtualRegister(&ARM::rGPRRegClass);
7358 BuildMI(*MBB, MI, DL, TII.get(ARM::t2MOVi32imm), Reg)
7359 .addExternalSymbol("__chkstk");
7360 BuildMI(*MBB, MI, DL, TII.get(ARM::tBLXr))
7361 .addImm((unsigned)ARMCC::AL).addReg(0)
7362 .addReg(Reg, RegState::Kill)
7363 .addReg(ARM::R4, RegState::Implicit | RegState::Kill)
7364 .addReg(ARM::R4, RegState::Implicit | RegState::Define)
7365 .addReg(ARM::R12, RegState::Implicit | RegState::Define | RegState::Dead);
7370 AddDefaultCC(AddDefaultPred(BuildMI(*MBB, MI, DL, TII.get(ARM::t2SUBrr),
7372 .addReg(ARM::SP).addReg(ARM::R4)));
7374 MI->eraseFromParent();
7379 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
7380 MachineBasicBlock *BB) const {
7381 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
7382 DebugLoc dl = MI->getDebugLoc();
7383 bool isThumb2 = Subtarget->isThumb2();
7384 switch (MI->getOpcode()) {
7387 llvm_unreachable("Unexpected instr type to insert");
7389 // The Thumb2 pre-indexed stores have the same MI operands, they just
7390 // define them differently in the .td files from the isel patterns, so
7391 // they need pseudos.
7392 case ARM::t2STR_preidx:
7393 MI->setDesc(TII->get(ARM::t2STR_PRE));
7395 case ARM::t2STRB_preidx:
7396 MI->setDesc(TII->get(ARM::t2STRB_PRE));
7398 case ARM::t2STRH_preidx:
7399 MI->setDesc(TII->get(ARM::t2STRH_PRE));
7402 case ARM::STRi_preidx:
7403 case ARM::STRBi_preidx: {
7404 unsigned NewOpc = MI->getOpcode() == ARM::STRi_preidx ?
7405 ARM::STR_PRE_IMM : ARM::STRB_PRE_IMM;
7406 // Decode the offset.
7407 unsigned Offset = MI->getOperand(4).getImm();
7408 bool isSub = ARM_AM::getAM2Op(Offset) == ARM_AM::sub;
7409 Offset = ARM_AM::getAM2Offset(Offset);
7413 MachineMemOperand *MMO = *MI->memoperands_begin();
7414 BuildMI(*BB, MI, dl, TII->get(NewOpc))
7415 .addOperand(MI->getOperand(0)) // Rn_wb
7416 .addOperand(MI->getOperand(1)) // Rt
7417 .addOperand(MI->getOperand(2)) // Rn
7418 .addImm(Offset) // offset (skip GPR==zero_reg)
7419 .addOperand(MI->getOperand(5)) // pred
7420 .addOperand(MI->getOperand(6))
7421 .addMemOperand(MMO);
7422 MI->eraseFromParent();
7425 case ARM::STRr_preidx:
7426 case ARM::STRBr_preidx:
7427 case ARM::STRH_preidx: {
7429 switch (MI->getOpcode()) {
7430 default: llvm_unreachable("unexpected opcode!");
7431 case ARM::STRr_preidx: NewOpc = ARM::STR_PRE_REG; break;
7432 case ARM::STRBr_preidx: NewOpc = ARM::STRB_PRE_REG; break;
7433 case ARM::STRH_preidx: NewOpc = ARM::STRH_PRE; break;
7435 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(NewOpc));
7436 for (unsigned i = 0; i < MI->getNumOperands(); ++i)
7437 MIB.addOperand(MI->getOperand(i));
7438 MI->eraseFromParent();
7442 case ARM::tMOVCCr_pseudo: {
7443 // To "insert" a SELECT_CC instruction, we actually have to insert the
7444 // diamond control-flow pattern. The incoming instruction knows the
7445 // destination vreg to set, the condition code register to branch on, the
7446 // true/false values to select between, and a branch opcode to use.
7447 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7448 MachineFunction::iterator It = BB;
7454 // cmpTY ccX, r1, r2
7456 // fallthrough --> copy0MBB
7457 MachineBasicBlock *thisMBB = BB;
7458 MachineFunction *F = BB->getParent();
7459 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
7460 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
7461 F->insert(It, copy0MBB);
7462 F->insert(It, sinkMBB);
7464 // Transfer the remainder of BB and its successor edges to sinkMBB.
7465 sinkMBB->splice(sinkMBB->begin(), BB,
7466 std::next(MachineBasicBlock::iterator(MI)), BB->end());
7467 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
7469 BB->addSuccessor(copy0MBB);
7470 BB->addSuccessor(sinkMBB);
7472 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
7473 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
7476 // %FalseValue = ...
7477 // # fallthrough to sinkMBB
7480 // Update machine-CFG edges
7481 BB->addSuccessor(sinkMBB);
7484 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
7487 BuildMI(*BB, BB->begin(), dl,
7488 TII->get(ARM::PHI), MI->getOperand(0).getReg())
7489 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
7490 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
7492 MI->eraseFromParent(); // The pseudo instruction is gone now.
7497 case ARM::BCCZi64: {
7498 // If there is an unconditional branch to the other successor, remove it.
7499 BB->erase(std::next(MachineBasicBlock::iterator(MI)), BB->end());
7501 // Compare both parts that make up the double comparison separately for
7503 bool RHSisZero = MI->getOpcode() == ARM::BCCZi64;
7505 unsigned LHS1 = MI->getOperand(1).getReg();
7506 unsigned LHS2 = MI->getOperand(2).getReg();
7508 AddDefaultPred(BuildMI(BB, dl,
7509 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7510 .addReg(LHS1).addImm(0));
7511 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7512 .addReg(LHS2).addImm(0)
7513 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
7515 unsigned RHS1 = MI->getOperand(3).getReg();
7516 unsigned RHS2 = MI->getOperand(4).getReg();
7517 AddDefaultPred(BuildMI(BB, dl,
7518 TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
7519 .addReg(LHS1).addReg(RHS1));
7520 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
7521 .addReg(LHS2).addReg(RHS2)
7522 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
7525 MachineBasicBlock *destMBB = MI->getOperand(RHSisZero ? 3 : 5).getMBB();
7526 MachineBasicBlock *exitMBB = OtherSucc(BB, destMBB);
7527 if (MI->getOperand(0).getImm() == ARMCC::NE)
7528 std::swap(destMBB, exitMBB);
7530 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
7531 .addMBB(destMBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
7533 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2B)).addMBB(exitMBB));
7535 BuildMI(BB, dl, TII->get(ARM::B)) .addMBB(exitMBB);
7537 MI->eraseFromParent(); // The pseudo instruction is gone now.
7541 case ARM::Int_eh_sjlj_setjmp:
7542 case ARM::Int_eh_sjlj_setjmp_nofp:
7543 case ARM::tInt_eh_sjlj_setjmp:
7544 case ARM::t2Int_eh_sjlj_setjmp:
7545 case ARM::t2Int_eh_sjlj_setjmp_nofp:
7546 EmitSjLjDispatchBlock(MI, BB);
7551 // To insert an ABS instruction, we have to insert the
7552 // diamond control-flow pattern. The incoming instruction knows the
7553 // source vreg to test against 0, the destination vreg to set,
7554 // the condition code register to branch on, the
7555 // true/false values to select between, and a branch opcode to use.
7560 // BCC (branch to SinkBB if V0 >= 0)
7561 // RSBBB: V3 = RSBri V2, 0 (compute ABS if V2 < 0)
7562 // SinkBB: V1 = PHI(V2, V3)
7563 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7564 MachineFunction::iterator BBI = BB;
7566 MachineFunction *Fn = BB->getParent();
7567 MachineBasicBlock *RSBBB = Fn->CreateMachineBasicBlock(LLVM_BB);
7568 MachineBasicBlock *SinkBB = Fn->CreateMachineBasicBlock(LLVM_BB);
7569 Fn->insert(BBI, RSBBB);
7570 Fn->insert(BBI, SinkBB);
7572 unsigned int ABSSrcReg = MI->getOperand(1).getReg();
7573 unsigned int ABSDstReg = MI->getOperand(0).getReg();
7574 bool ABSSrcKIll = MI->getOperand(1).isKill();
7575 bool isThumb2 = Subtarget->isThumb2();
7576 MachineRegisterInfo &MRI = Fn->getRegInfo();
7577 // In Thumb mode S must not be specified if source register is the SP or
7578 // PC and if destination register is the SP, so restrict register class
7579 unsigned NewRsbDstReg =
7580 MRI.createVirtualRegister(isThumb2 ? &ARM::rGPRRegClass : &ARM::GPRRegClass);
7582 // Transfer the remainder of BB and its successor edges to sinkMBB.
7583 SinkBB->splice(SinkBB->begin(), BB,
7584 std::next(MachineBasicBlock::iterator(MI)), BB->end());
7585 SinkBB->transferSuccessorsAndUpdatePHIs(BB);
7587 BB->addSuccessor(RSBBB);
7588 BB->addSuccessor(SinkBB);
7590 // fall through to SinkMBB
7591 RSBBB->addSuccessor(SinkBB);
7593 // insert a cmp at the end of BB
7594 AddDefaultPred(BuildMI(BB, dl,
7595 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7596 .addReg(ABSSrcReg).addImm(0));
7598 // insert a bcc with opposite CC to ARMCC::MI at the end of BB
7600 TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)).addMBB(SinkBB)
7601 .addImm(ARMCC::getOppositeCondition(ARMCC::MI)).addReg(ARM::CPSR);
7603 // insert rsbri in RSBBB
7604 // Note: BCC and rsbri will be converted into predicated rsbmi
7605 // by if-conversion pass
7606 BuildMI(*RSBBB, RSBBB->begin(), dl,
7607 TII->get(isThumb2 ? ARM::t2RSBri : ARM::RSBri), NewRsbDstReg)
7608 .addReg(ABSSrcReg, ABSSrcKIll ? RegState::Kill : 0)
7609 .addImm(0).addImm((unsigned)ARMCC::AL).addReg(0).addReg(0);
7611 // insert PHI in SinkBB,
7612 // reuse ABSDstReg to not change uses of ABS instruction
7613 BuildMI(*SinkBB, SinkBB->begin(), dl,
7614 TII->get(ARM::PHI), ABSDstReg)
7615 .addReg(NewRsbDstReg).addMBB(RSBBB)
7616 .addReg(ABSSrcReg).addMBB(BB);
7618 // remove ABS instruction
7619 MI->eraseFromParent();
7621 // return last added BB
7624 case ARM::COPY_STRUCT_BYVAL_I32:
7626 return EmitStructByval(MI, BB);
7627 case ARM::WIN__CHKSTK:
7628 return EmitLowered__chkstk(MI, BB);
7632 void ARMTargetLowering::AdjustInstrPostInstrSelection(MachineInstr *MI,
7633 SDNode *Node) const {
7634 const MCInstrDesc *MCID = &MI->getDesc();
7635 // Adjust potentially 's' setting instructions after isel, i.e. ADC, SBC, RSB,
7636 // RSC. Coming out of isel, they have an implicit CPSR def, but the optional
7637 // operand is still set to noreg. If needed, set the optional operand's
7638 // register to CPSR, and remove the redundant implicit def.
7640 // e.g. ADCS (..., CPSR<imp-def>) -> ADC (... opt:CPSR<def>).
7642 // Rename pseudo opcodes.
7643 unsigned NewOpc = convertAddSubFlagsOpcode(MI->getOpcode());
7645 const ARMBaseInstrInfo *TII = Subtarget->getInstrInfo();
7646 MCID = &TII->get(NewOpc);
7648 assert(MCID->getNumOperands() == MI->getDesc().getNumOperands() + 1 &&
7649 "converted opcode should be the same except for cc_out");
7653 // Add the optional cc_out operand
7654 MI->addOperand(MachineOperand::CreateReg(0, /*isDef=*/true));
7656 unsigned ccOutIdx = MCID->getNumOperands() - 1;
7658 // Any ARM instruction that sets the 's' bit should specify an optional
7659 // "cc_out" operand in the last operand position.
7660 if (!MI->hasOptionalDef() || !MCID->OpInfo[ccOutIdx].isOptionalDef()) {
7661 assert(!NewOpc && "Optional cc_out operand required");
7664 // Look for an implicit def of CPSR added by MachineInstr ctor. Remove it
7665 // since we already have an optional CPSR def.
7666 bool definesCPSR = false;
7667 bool deadCPSR = false;
7668 for (unsigned i = MCID->getNumOperands(), e = MI->getNumOperands();
7670 const MachineOperand &MO = MI->getOperand(i);
7671 if (MO.isReg() && MO.isDef() && MO.getReg() == ARM::CPSR) {
7675 MI->RemoveOperand(i);
7680 assert(!NewOpc && "Optional cc_out operand required");
7683 assert(deadCPSR == !Node->hasAnyUseOfValue(1) && "inconsistent dead flag");
7685 assert(!MI->getOperand(ccOutIdx).getReg() &&
7686 "expect uninitialized optional cc_out operand");
7690 // If this instruction was defined with an optional CPSR def and its dag node
7691 // had a live implicit CPSR def, then activate the optional CPSR def.
7692 MachineOperand &MO = MI->getOperand(ccOutIdx);
7693 MO.setReg(ARM::CPSR);
7697 //===----------------------------------------------------------------------===//
7698 // ARM Optimization Hooks
7699 //===----------------------------------------------------------------------===//
7701 // Helper function that checks if N is a null or all ones constant.
7702 static inline bool isZeroOrAllOnes(SDValue N, bool AllOnes) {
7703 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N);
7706 return AllOnes ? C->isAllOnesValue() : C->isNullValue();
7709 // Return true if N is conditionally 0 or all ones.
7710 // Detects these expressions where cc is an i1 value:
7712 // (select cc 0, y) [AllOnes=0]
7713 // (select cc y, 0) [AllOnes=0]
7714 // (zext cc) [AllOnes=0]
7715 // (sext cc) [AllOnes=0/1]
7716 // (select cc -1, y) [AllOnes=1]
7717 // (select cc y, -1) [AllOnes=1]
7719 // Invert is set when N is the null/all ones constant when CC is false.
7720 // OtherOp is set to the alternative value of N.
7721 static bool isConditionalZeroOrAllOnes(SDNode *N, bool AllOnes,
7722 SDValue &CC, bool &Invert,
7724 SelectionDAG &DAG) {
7725 switch (N->getOpcode()) {
7726 default: return false;
7728 CC = N->getOperand(0);
7729 SDValue N1 = N->getOperand(1);
7730 SDValue N2 = N->getOperand(2);
7731 if (isZeroOrAllOnes(N1, AllOnes)) {
7736 if (isZeroOrAllOnes(N2, AllOnes)) {
7743 case ISD::ZERO_EXTEND:
7744 // (zext cc) can never be the all ones value.
7748 case ISD::SIGN_EXTEND: {
7750 EVT VT = N->getValueType(0);
7751 CC = N->getOperand(0);
7752 if (CC.getValueType() != MVT::i1)
7756 // When looking for an AllOnes constant, N is an sext, and the 'other'
7758 OtherOp = DAG.getConstant(0, dl, VT);
7759 else if (N->getOpcode() == ISD::ZERO_EXTEND)
7760 // When looking for a 0 constant, N can be zext or sext.
7761 OtherOp = DAG.getConstant(1, dl, VT);
7763 OtherOp = DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), dl,
7770 // Combine a constant select operand into its use:
7772 // (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
7773 // (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
7774 // (and (select cc, -1, c), x) -> (select cc, x, (and, x, c)) [AllOnes=1]
7775 // (or (select cc, 0, c), x) -> (select cc, x, (or, x, c))
7776 // (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c))
7778 // The transform is rejected if the select doesn't have a constant operand that
7779 // is null, or all ones when AllOnes is set.
7781 // Also recognize sext/zext from i1:
7783 // (add (zext cc), x) -> (select cc (add x, 1), x)
7784 // (add (sext cc), x) -> (select cc (add x, -1), x)
7786 // These transformations eventually create predicated instructions.
7788 // @param N The node to transform.
7789 // @param Slct The N operand that is a select.
7790 // @param OtherOp The other N operand (x above).
7791 // @param DCI Context.
7792 // @param AllOnes Require the select constant to be all ones instead of null.
7793 // @returns The new node, or SDValue() on failure.
7795 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
7796 TargetLowering::DAGCombinerInfo &DCI,
7797 bool AllOnes = false) {
7798 SelectionDAG &DAG = DCI.DAG;
7799 EVT VT = N->getValueType(0);
7800 SDValue NonConstantVal;
7803 if (!isConditionalZeroOrAllOnes(Slct.getNode(), AllOnes, CCOp, SwapSelectOps,
7804 NonConstantVal, DAG))
7807 // Slct is now know to be the desired identity constant when CC is true.
7808 SDValue TrueVal = OtherOp;
7809 SDValue FalseVal = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
7810 OtherOp, NonConstantVal);
7811 // Unless SwapSelectOps says CC should be false.
7813 std::swap(TrueVal, FalseVal);
7815 return DAG.getNode(ISD::SELECT, SDLoc(N), VT,
7816 CCOp, TrueVal, FalseVal);
7819 // Attempt combineSelectAndUse on each operand of a commutative operator N.
7821 SDValue combineSelectAndUseCommutative(SDNode *N, bool AllOnes,
7822 TargetLowering::DAGCombinerInfo &DCI) {
7823 SDValue N0 = N->getOperand(0);
7824 SDValue N1 = N->getOperand(1);
7825 if (N0.getNode()->hasOneUse()) {
7826 SDValue Result = combineSelectAndUse(N, N0, N1, DCI, AllOnes);
7827 if (Result.getNode())
7830 if (N1.getNode()->hasOneUse()) {
7831 SDValue Result = combineSelectAndUse(N, N1, N0, DCI, AllOnes);
7832 if (Result.getNode())
7838 // AddCombineToVPADDL- For pair-wise add on neon, use the vpaddl instruction
7839 // (only after legalization).
7840 static SDValue AddCombineToVPADDL(SDNode *N, SDValue N0, SDValue N1,
7841 TargetLowering::DAGCombinerInfo &DCI,
7842 const ARMSubtarget *Subtarget) {
7844 // Only perform optimization if after legalize, and if NEON is available. We
7845 // also expected both operands to be BUILD_VECTORs.
7846 if (DCI.isBeforeLegalize() || !Subtarget->hasNEON()
7847 || N0.getOpcode() != ISD::BUILD_VECTOR
7848 || N1.getOpcode() != ISD::BUILD_VECTOR)
7851 // Check output type since VPADDL operand elements can only be 8, 16, or 32.
7852 EVT VT = N->getValueType(0);
7853 if (!VT.isInteger() || VT.getVectorElementType() == MVT::i64)
7856 // Check that the vector operands are of the right form.
7857 // N0 and N1 are BUILD_VECTOR nodes with N number of EXTRACT_VECTOR
7858 // operands, where N is the size of the formed vector.
7859 // Each EXTRACT_VECTOR should have the same input vector and odd or even
7860 // index such that we have a pair wise add pattern.
7862 // Grab the vector that all EXTRACT_VECTOR nodes should be referencing.
7863 if (N0->getOperand(0)->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
7865 SDValue Vec = N0->getOperand(0)->getOperand(0);
7866 SDNode *V = Vec.getNode();
7867 unsigned nextIndex = 0;
7869 // For each operands to the ADD which are BUILD_VECTORs,
7870 // check to see if each of their operands are an EXTRACT_VECTOR with
7871 // the same vector and appropriate index.
7872 for (unsigned i = 0, e = N0->getNumOperands(); i != e; ++i) {
7873 if (N0->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT
7874 && N1->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
7876 SDValue ExtVec0 = N0->getOperand(i);
7877 SDValue ExtVec1 = N1->getOperand(i);
7879 // First operand is the vector, verify its the same.
7880 if (V != ExtVec0->getOperand(0).getNode() ||
7881 V != ExtVec1->getOperand(0).getNode())
7884 // Second is the constant, verify its correct.
7885 ConstantSDNode *C0 = dyn_cast<ConstantSDNode>(ExtVec0->getOperand(1));
7886 ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(ExtVec1->getOperand(1));
7888 // For the constant, we want to see all the even or all the odd.
7889 if (!C0 || !C1 || C0->getZExtValue() != nextIndex
7890 || C1->getZExtValue() != nextIndex+1)
7899 // Create VPADDL node.
7900 SelectionDAG &DAG = DCI.DAG;
7901 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7905 // Build operand list.
7906 SmallVector<SDValue, 8> Ops;
7907 Ops.push_back(DAG.getConstant(Intrinsic::arm_neon_vpaddls, dl,
7908 TLI.getPointerTy()));
7910 // Input is the vector.
7913 // Get widened type and narrowed type.
7915 unsigned numElem = VT.getVectorNumElements();
7917 EVT inputLaneType = Vec.getValueType().getVectorElementType();
7918 switch (inputLaneType.getSimpleVT().SimpleTy) {
7919 case MVT::i8: widenType = MVT::getVectorVT(MVT::i16, numElem); break;
7920 case MVT::i16: widenType = MVT::getVectorVT(MVT::i32, numElem); break;
7921 case MVT::i32: widenType = MVT::getVectorVT(MVT::i64, numElem); break;
7923 llvm_unreachable("Invalid vector element type for padd optimization.");
7926 SDValue tmp = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, widenType, Ops);
7927 unsigned ExtOp = VT.bitsGT(tmp.getValueType()) ? ISD::ANY_EXTEND : ISD::TRUNCATE;
7928 return DAG.getNode(ExtOp, dl, VT, tmp);
7931 static SDValue findMUL_LOHI(SDValue V) {
7932 if (V->getOpcode() == ISD::UMUL_LOHI ||
7933 V->getOpcode() == ISD::SMUL_LOHI)
7938 static SDValue AddCombineTo64bitMLAL(SDNode *AddcNode,
7939 TargetLowering::DAGCombinerInfo &DCI,
7940 const ARMSubtarget *Subtarget) {
7942 if (Subtarget->isThumb1Only()) return SDValue();
7944 // Only perform the checks after legalize when the pattern is available.
7945 if (DCI.isBeforeLegalize()) return SDValue();
7947 // Look for multiply add opportunities.
7948 // The pattern is a ISD::UMUL_LOHI followed by two add nodes, where
7949 // each add nodes consumes a value from ISD::UMUL_LOHI and there is
7950 // a glue link from the first add to the second add.
7951 // If we find this pattern, we can replace the U/SMUL_LOHI, ADDC, and ADDE by
7952 // a S/UMLAL instruction.
7955 // / \ [no multiline comment]
7961 assert(AddcNode->getOpcode() == ISD::ADDC && "Expect an ADDC");
7962 SDValue AddcOp0 = AddcNode->getOperand(0);
7963 SDValue AddcOp1 = AddcNode->getOperand(1);
7965 // Check if the two operands are from the same mul_lohi node.
7966 if (AddcOp0.getNode() == AddcOp1.getNode())
7969 assert(AddcNode->getNumValues() == 2 &&
7970 AddcNode->getValueType(0) == MVT::i32 &&
7971 "Expect ADDC with two result values. First: i32");
7973 // Check that we have a glued ADDC node.
7974 if (AddcNode->getValueType(1) != MVT::Glue)
7977 // Check that the ADDC adds the low result of the S/UMUL_LOHI.
7978 if (AddcOp0->getOpcode() != ISD::UMUL_LOHI &&
7979 AddcOp0->getOpcode() != ISD::SMUL_LOHI &&
7980 AddcOp1->getOpcode() != ISD::UMUL_LOHI &&
7981 AddcOp1->getOpcode() != ISD::SMUL_LOHI)
7984 // Look for the glued ADDE.
7985 SDNode* AddeNode = AddcNode->getGluedUser();
7989 // Make sure it is really an ADDE.
7990 if (AddeNode->getOpcode() != ISD::ADDE)
7993 assert(AddeNode->getNumOperands() == 3 &&
7994 AddeNode->getOperand(2).getValueType() == MVT::Glue &&
7995 "ADDE node has the wrong inputs");
7997 // Check for the triangle shape.
7998 SDValue AddeOp0 = AddeNode->getOperand(0);
7999 SDValue AddeOp1 = AddeNode->getOperand(1);
8001 // Make sure that the ADDE operands are not coming from the same node.
8002 if (AddeOp0.getNode() == AddeOp1.getNode())
8005 // Find the MUL_LOHI node walking up ADDE's operands.
8006 bool IsLeftOperandMUL = false;
8007 SDValue MULOp = findMUL_LOHI(AddeOp0);
8008 if (MULOp == SDValue())
8009 MULOp = findMUL_LOHI(AddeOp1);
8011 IsLeftOperandMUL = true;
8012 if (MULOp == SDValue())
8015 // Figure out the right opcode.
8016 unsigned Opc = MULOp->getOpcode();
8017 unsigned FinalOpc = (Opc == ISD::SMUL_LOHI) ? ARMISD::SMLAL : ARMISD::UMLAL;
8019 // Figure out the high and low input values to the MLAL node.
8020 SDValue* HiAdd = nullptr;
8021 SDValue* LoMul = nullptr;
8022 SDValue* LowAdd = nullptr;
8024 // Ensure that ADDE is from high result of ISD::SMUL_LOHI.
8025 if ((AddeOp0 != MULOp.getValue(1)) && (AddeOp1 != MULOp.getValue(1)))
8028 if (IsLeftOperandMUL)
8034 // Ensure that LoMul and LowAdd are taken from correct ISD::SMUL_LOHI node
8035 // whose low result is fed to the ADDC we are checking.
8037 if (AddcOp0 == MULOp.getValue(0)) {
8041 if (AddcOp1 == MULOp.getValue(0)) {
8049 // Create the merged node.
8050 SelectionDAG &DAG = DCI.DAG;
8052 // Build operand list.
8053 SmallVector<SDValue, 8> Ops;
8054 Ops.push_back(LoMul->getOperand(0));
8055 Ops.push_back(LoMul->getOperand(1));
8056 Ops.push_back(*LowAdd);
8057 Ops.push_back(*HiAdd);
8059 SDValue MLALNode = DAG.getNode(FinalOpc, SDLoc(AddcNode),
8060 DAG.getVTList(MVT::i32, MVT::i32), Ops);
8062 // Replace the ADDs' nodes uses by the MLA node's values.
8063 SDValue HiMLALResult(MLALNode.getNode(), 1);
8064 DAG.ReplaceAllUsesOfValueWith(SDValue(AddeNode, 0), HiMLALResult);
8066 SDValue LoMLALResult(MLALNode.getNode(), 0);
8067 DAG.ReplaceAllUsesOfValueWith(SDValue(AddcNode, 0), LoMLALResult);
8069 // Return original node to notify the driver to stop replacing.
8070 SDValue resNode(AddcNode, 0);
8074 /// PerformADDCCombine - Target-specific dag combine transform from
8075 /// ISD::ADDC, ISD::ADDE, and ISD::MUL_LOHI to MLAL.
8076 static SDValue PerformADDCCombine(SDNode *N,
8077 TargetLowering::DAGCombinerInfo &DCI,
8078 const ARMSubtarget *Subtarget) {
8080 return AddCombineTo64bitMLAL(N, DCI, Subtarget);
8084 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
8085 /// operands N0 and N1. This is a helper for PerformADDCombine that is
8086 /// called with the default operands, and if that fails, with commuted
8088 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
8089 TargetLowering::DAGCombinerInfo &DCI,
8090 const ARMSubtarget *Subtarget){
8092 // Attempt to create vpaddl for this add.
8093 SDValue Result = AddCombineToVPADDL(N, N0, N1, DCI, Subtarget);
8094 if (Result.getNode())
8097 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
8098 if (N0.getNode()->hasOneUse()) {
8099 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
8100 if (Result.getNode()) return Result;
8105 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
8107 static SDValue PerformADDCombine(SDNode *N,
8108 TargetLowering::DAGCombinerInfo &DCI,
8109 const ARMSubtarget *Subtarget) {
8110 SDValue N0 = N->getOperand(0);
8111 SDValue N1 = N->getOperand(1);
8113 // First try with the default operand order.
8114 SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI, Subtarget);
8115 if (Result.getNode())
8118 // If that didn't work, try again with the operands commuted.
8119 return PerformADDCombineWithOperands(N, N1, N0, DCI, Subtarget);
8122 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
8124 static SDValue PerformSUBCombine(SDNode *N,
8125 TargetLowering::DAGCombinerInfo &DCI) {
8126 SDValue N0 = N->getOperand(0);
8127 SDValue N1 = N->getOperand(1);
8129 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
8130 if (N1.getNode()->hasOneUse()) {
8131 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
8132 if (Result.getNode()) return Result;
8138 /// PerformVMULCombine
8139 /// Distribute (A + B) * C to (A * C) + (B * C) to take advantage of the
8140 /// special multiplier accumulator forwarding.
8146 // However, for (A + B) * (A + B),
8153 static SDValue PerformVMULCombine(SDNode *N,
8154 TargetLowering::DAGCombinerInfo &DCI,
8155 const ARMSubtarget *Subtarget) {
8156 if (!Subtarget->hasVMLxForwarding())
8159 SelectionDAG &DAG = DCI.DAG;
8160 SDValue N0 = N->getOperand(0);
8161 SDValue N1 = N->getOperand(1);
8162 unsigned Opcode = N0.getOpcode();
8163 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
8164 Opcode != ISD::FADD && Opcode != ISD::FSUB) {
8165 Opcode = N1.getOpcode();
8166 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
8167 Opcode != ISD::FADD && Opcode != ISD::FSUB)
8175 EVT VT = N->getValueType(0);
8177 SDValue N00 = N0->getOperand(0);
8178 SDValue N01 = N0->getOperand(1);
8179 return DAG.getNode(Opcode, DL, VT,
8180 DAG.getNode(ISD::MUL, DL, VT, N00, N1),
8181 DAG.getNode(ISD::MUL, DL, VT, N01, N1));
8184 static SDValue PerformMULCombine(SDNode *N,
8185 TargetLowering::DAGCombinerInfo &DCI,
8186 const ARMSubtarget *Subtarget) {
8187 SelectionDAG &DAG = DCI.DAG;
8189 if (Subtarget->isThumb1Only())
8192 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
8195 EVT VT = N->getValueType(0);
8196 if (VT.is64BitVector() || VT.is128BitVector())
8197 return PerformVMULCombine(N, DCI, Subtarget);
8201 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
8205 int64_t MulAmt = C->getSExtValue();
8206 unsigned ShiftAmt = countTrailingZeros<uint64_t>(MulAmt);
8208 ShiftAmt = ShiftAmt & (32 - 1);
8209 SDValue V = N->getOperand(0);
8213 MulAmt >>= ShiftAmt;
8216 if (isPowerOf2_32(MulAmt - 1)) {
8217 // (mul x, 2^N + 1) => (add (shl x, N), x)
8218 Res = DAG.getNode(ISD::ADD, DL, VT,
8220 DAG.getNode(ISD::SHL, DL, VT,
8222 DAG.getConstant(Log2_32(MulAmt - 1), DL,
8224 } else if (isPowerOf2_32(MulAmt + 1)) {
8225 // (mul x, 2^N - 1) => (sub (shl x, N), x)
8226 Res = DAG.getNode(ISD::SUB, DL, VT,
8227 DAG.getNode(ISD::SHL, DL, VT,
8229 DAG.getConstant(Log2_32(MulAmt + 1), DL,
8235 uint64_t MulAmtAbs = -MulAmt;
8236 if (isPowerOf2_32(MulAmtAbs + 1)) {
8237 // (mul x, -(2^N - 1)) => (sub x, (shl x, N))
8238 Res = DAG.getNode(ISD::SUB, DL, VT,
8240 DAG.getNode(ISD::SHL, DL, VT,
8242 DAG.getConstant(Log2_32(MulAmtAbs + 1), DL,
8244 } else if (isPowerOf2_32(MulAmtAbs - 1)) {
8245 // (mul x, -(2^N + 1)) => - (add (shl x, N), x)
8246 Res = DAG.getNode(ISD::ADD, DL, VT,
8248 DAG.getNode(ISD::SHL, DL, VT,
8250 DAG.getConstant(Log2_32(MulAmtAbs - 1), DL,
8252 Res = DAG.getNode(ISD::SUB, DL, VT,
8253 DAG.getConstant(0, DL, MVT::i32), Res);
8260 Res = DAG.getNode(ISD::SHL, DL, VT,
8261 Res, DAG.getConstant(ShiftAmt, DL, MVT::i32));
8263 // Do not add new nodes to DAG combiner worklist.
8264 DCI.CombineTo(N, Res, false);
8268 static SDValue PerformANDCombine(SDNode *N,
8269 TargetLowering::DAGCombinerInfo &DCI,
8270 const ARMSubtarget *Subtarget) {
8272 // Attempt to use immediate-form VBIC
8273 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
8275 EVT VT = N->getValueType(0);
8276 SelectionDAG &DAG = DCI.DAG;
8278 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8281 APInt SplatBits, SplatUndef;
8282 unsigned SplatBitSize;
8285 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
8286 if (SplatBitSize <= 64) {
8288 SDValue Val = isNEONModifiedImm((~SplatBits).getZExtValue(),
8289 SplatUndef.getZExtValue(), SplatBitSize,
8290 DAG, dl, VbicVT, VT.is128BitVector(),
8292 if (Val.getNode()) {
8294 DAG.getNode(ISD::BITCAST, dl, VbicVT, N->getOperand(0));
8295 SDValue Vbic = DAG.getNode(ARMISD::VBICIMM, dl, VbicVT, Input, Val);
8296 return DAG.getNode(ISD::BITCAST, dl, VT, Vbic);
8301 if (!Subtarget->isThumb1Only()) {
8302 // fold (and (select cc, -1, c), x) -> (select cc, x, (and, x, c))
8303 SDValue Result = combineSelectAndUseCommutative(N, true, DCI);
8304 if (Result.getNode())
8311 /// PerformORCombine - Target-specific dag combine xforms for ISD::OR
8312 static SDValue PerformORCombine(SDNode *N,
8313 TargetLowering::DAGCombinerInfo &DCI,
8314 const ARMSubtarget *Subtarget) {
8315 // Attempt to use immediate-form VORR
8316 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
8318 EVT VT = N->getValueType(0);
8319 SelectionDAG &DAG = DCI.DAG;
8321 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8324 APInt SplatBits, SplatUndef;
8325 unsigned SplatBitSize;
8327 if (BVN && Subtarget->hasNEON() &&
8328 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
8329 if (SplatBitSize <= 64) {
8331 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
8332 SplatUndef.getZExtValue(), SplatBitSize,
8333 DAG, dl, VorrVT, VT.is128BitVector(),
8335 if (Val.getNode()) {
8337 DAG.getNode(ISD::BITCAST, dl, VorrVT, N->getOperand(0));
8338 SDValue Vorr = DAG.getNode(ARMISD::VORRIMM, dl, VorrVT, Input, Val);
8339 return DAG.getNode(ISD::BITCAST, dl, VT, Vorr);
8344 if (!Subtarget->isThumb1Only()) {
8345 // fold (or (select cc, 0, c), x) -> (select cc, x, (or, x, c))
8346 SDValue Result = combineSelectAndUseCommutative(N, false, DCI);
8347 if (Result.getNode())
8351 // The code below optimizes (or (and X, Y), Z).
8352 // The AND operand needs to have a single user to make these optimizations
8354 SDValue N0 = N->getOperand(0);
8355 if (N0.getOpcode() != ISD::AND || !N0.hasOneUse())
8357 SDValue N1 = N->getOperand(1);
8359 // (or (and B, A), (and C, ~A)) => (VBSL A, B, C) when A is a constant.
8360 if (Subtarget->hasNEON() && N1.getOpcode() == ISD::AND && VT.isVector() &&
8361 DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
8363 unsigned SplatBitSize;
8366 APInt SplatBits0, SplatBits1;
8367 BuildVectorSDNode *BVN0 = dyn_cast<BuildVectorSDNode>(N0->getOperand(1));
8368 BuildVectorSDNode *BVN1 = dyn_cast<BuildVectorSDNode>(N1->getOperand(1));
8369 // Ensure that the second operand of both ands are constants
8370 if (BVN0 && BVN0->isConstantSplat(SplatBits0, SplatUndef, SplatBitSize,
8371 HasAnyUndefs) && !HasAnyUndefs) {
8372 if (BVN1 && BVN1->isConstantSplat(SplatBits1, SplatUndef, SplatBitSize,
8373 HasAnyUndefs) && !HasAnyUndefs) {
8374 // Ensure that the bit width of the constants are the same and that
8375 // the splat arguments are logical inverses as per the pattern we
8376 // are trying to simplify.
8377 if (SplatBits0.getBitWidth() == SplatBits1.getBitWidth() &&
8378 SplatBits0 == ~SplatBits1) {
8379 // Canonicalize the vector type to make instruction selection
8381 EVT CanonicalVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
8382 SDValue Result = DAG.getNode(ARMISD::VBSL, dl, CanonicalVT,
8386 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
8392 // Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when
8395 // BFI is only available on V6T2+
8396 if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops())
8400 // 1) or (and A, mask), val => ARMbfi A, val, mask
8401 // iff (val & mask) == val
8403 // 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
8404 // 2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2)
8405 // && mask == ~mask2
8406 // 2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2)
8407 // && ~mask == mask2
8408 // (i.e., copy a bitfield value into another bitfield of the same width)
8413 SDValue N00 = N0.getOperand(0);
8415 // The value and the mask need to be constants so we can verify this is
8416 // actually a bitfield set. If the mask is 0xffff, we can do better
8417 // via a movt instruction, so don't use BFI in that case.
8418 SDValue MaskOp = N0.getOperand(1);
8419 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(MaskOp);
8422 unsigned Mask = MaskC->getZExtValue();
8426 // Case (1): or (and A, mask), val => ARMbfi A, val, mask
8427 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
8429 unsigned Val = N1C->getZExtValue();
8430 if ((Val & ~Mask) != Val)
8433 if (ARM::isBitFieldInvertedMask(Mask)) {
8434 Val >>= countTrailingZeros(~Mask);
8436 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00,
8437 DAG.getConstant(Val, DL, MVT::i32),
8438 DAG.getConstant(Mask, DL, MVT::i32));
8440 // Do not add new nodes to DAG combiner worklist.
8441 DCI.CombineTo(N, Res, false);
8444 } else if (N1.getOpcode() == ISD::AND) {
8445 // case (2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
8446 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
8449 unsigned Mask2 = N11C->getZExtValue();
8451 // Mask and ~Mask2 (or reverse) must be equivalent for the BFI pattern
8453 if (ARM::isBitFieldInvertedMask(Mask) &&
8455 // The pack halfword instruction works better for masks that fit it,
8456 // so use that when it's available.
8457 if (Subtarget->hasT2ExtractPack() &&
8458 (Mask == 0xffff || Mask == 0xffff0000))
8461 unsigned amt = countTrailingZeros(Mask2);
8462 Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0),
8463 DAG.getConstant(amt, DL, MVT::i32));
8464 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00, Res,
8465 DAG.getConstant(Mask, DL, MVT::i32));
8466 // Do not add new nodes to DAG combiner worklist.
8467 DCI.CombineTo(N, Res, false);
8469 } else if (ARM::isBitFieldInvertedMask(~Mask) &&
8471 // The pack halfword instruction works better for masks that fit it,
8472 // so use that when it's available.
8473 if (Subtarget->hasT2ExtractPack() &&
8474 (Mask2 == 0xffff || Mask2 == 0xffff0000))
8477 unsigned lsb = countTrailingZeros(Mask);
8478 Res = DAG.getNode(ISD::SRL, DL, VT, N00,
8479 DAG.getConstant(lsb, DL, MVT::i32));
8480 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res,
8481 DAG.getConstant(Mask2, DL, MVT::i32));
8482 // Do not add new nodes to DAG combiner worklist.
8483 DCI.CombineTo(N, Res, false);
8488 if (DAG.MaskedValueIsZero(N1, MaskC->getAPIntValue()) &&
8489 N00.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N00.getOperand(1)) &&
8490 ARM::isBitFieldInvertedMask(~Mask)) {
8491 // Case (3): or (and (shl A, #shamt), mask), B => ARMbfi B, A, ~mask
8492 // where lsb(mask) == #shamt and masked bits of B are known zero.
8493 SDValue ShAmt = N00.getOperand(1);
8494 unsigned ShAmtC = cast<ConstantSDNode>(ShAmt)->getZExtValue();
8495 unsigned LSB = countTrailingZeros(Mask);
8499 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1, N00.getOperand(0),
8500 DAG.getConstant(~Mask, DL, MVT::i32));
8502 // Do not add new nodes to DAG combiner worklist.
8503 DCI.CombineTo(N, Res, false);
8509 static SDValue PerformXORCombine(SDNode *N,
8510 TargetLowering::DAGCombinerInfo &DCI,
8511 const ARMSubtarget *Subtarget) {
8512 EVT VT = N->getValueType(0);
8513 SelectionDAG &DAG = DCI.DAG;
8515 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8518 if (!Subtarget->isThumb1Only()) {
8519 // fold (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c))
8520 SDValue Result = combineSelectAndUseCommutative(N, false, DCI);
8521 if (Result.getNode())
8528 /// PerformBFICombine - (bfi A, (and B, Mask1), Mask2) -> (bfi A, B, Mask2) iff
8529 /// the bits being cleared by the AND are not demanded by the BFI.
8530 static SDValue PerformBFICombine(SDNode *N,
8531 TargetLowering::DAGCombinerInfo &DCI) {
8532 SDValue N1 = N->getOperand(1);
8533 if (N1.getOpcode() == ISD::AND) {
8534 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
8537 unsigned InvMask = cast<ConstantSDNode>(N->getOperand(2))->getZExtValue();
8538 unsigned LSB = countTrailingZeros(~InvMask);
8539 unsigned Width = (32 - countLeadingZeros(~InvMask)) - LSB;
8541 static_cast<unsigned>(std::numeric_limits<unsigned>::digits) &&
8542 "undefined behavior");
8543 unsigned Mask = (1u << Width) - 1;
8544 unsigned Mask2 = N11C->getZExtValue();
8545 if ((Mask & (~Mask2)) == 0)
8546 return DCI.DAG.getNode(ARMISD::BFI, SDLoc(N), N->getValueType(0),
8547 N->getOperand(0), N1.getOperand(0),
8553 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
8554 /// ARMISD::VMOVRRD.
8555 static SDValue PerformVMOVRRDCombine(SDNode *N,
8556 TargetLowering::DAGCombinerInfo &DCI,
8557 const ARMSubtarget *Subtarget) {
8558 // vmovrrd(vmovdrr x, y) -> x,y
8559 SDValue InDouble = N->getOperand(0);
8560 if (InDouble.getOpcode() == ARMISD::VMOVDRR && !Subtarget->isFPOnlySP())
8561 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
8563 // vmovrrd(load f64) -> (load i32), (load i32)
8564 SDNode *InNode = InDouble.getNode();
8565 if (ISD::isNormalLoad(InNode) && InNode->hasOneUse() &&
8566 InNode->getValueType(0) == MVT::f64 &&
8567 InNode->getOperand(1).getOpcode() == ISD::FrameIndex &&
8568 !cast<LoadSDNode>(InNode)->isVolatile()) {
8569 // TODO: Should this be done for non-FrameIndex operands?
8570 LoadSDNode *LD = cast<LoadSDNode>(InNode);
8572 SelectionDAG &DAG = DCI.DAG;
8574 SDValue BasePtr = LD->getBasePtr();
8575 SDValue NewLD1 = DAG.getLoad(MVT::i32, DL, LD->getChain(), BasePtr,
8576 LD->getPointerInfo(), LD->isVolatile(),
8577 LD->isNonTemporal(), LD->isInvariant(),
8578 LD->getAlignment());
8580 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
8581 DAG.getConstant(4, DL, MVT::i32));
8582 SDValue NewLD2 = DAG.getLoad(MVT::i32, DL, NewLD1.getValue(1), OffsetPtr,
8583 LD->getPointerInfo(), LD->isVolatile(),
8584 LD->isNonTemporal(), LD->isInvariant(),
8585 std::min(4U, LD->getAlignment() / 2));
8587 DAG.ReplaceAllUsesOfValueWith(SDValue(LD, 1), NewLD2.getValue(1));
8588 if (DCI.DAG.getTargetLoweringInfo().isBigEndian())
8589 std::swap (NewLD1, NewLD2);
8590 SDValue Result = DCI.CombineTo(N, NewLD1, NewLD2);
8597 /// PerformVMOVDRRCombine - Target-specific dag combine xforms for
8598 /// ARMISD::VMOVDRR. This is also used for BUILD_VECTORs with 2 operands.
8599 static SDValue PerformVMOVDRRCombine(SDNode *N, SelectionDAG &DAG) {
8600 // N=vmovrrd(X); vmovdrr(N:0, N:1) -> bit_convert(X)
8601 SDValue Op0 = N->getOperand(0);
8602 SDValue Op1 = N->getOperand(1);
8603 if (Op0.getOpcode() == ISD::BITCAST)
8604 Op0 = Op0.getOperand(0);
8605 if (Op1.getOpcode() == ISD::BITCAST)
8606 Op1 = Op1.getOperand(0);
8607 if (Op0.getOpcode() == ARMISD::VMOVRRD &&
8608 Op0.getNode() == Op1.getNode() &&
8609 Op0.getResNo() == 0 && Op1.getResNo() == 1)
8610 return DAG.getNode(ISD::BITCAST, SDLoc(N),
8611 N->getValueType(0), Op0.getOperand(0));
8615 /// hasNormalLoadOperand - Check if any of the operands of a BUILD_VECTOR node
8616 /// are normal, non-volatile loads. If so, it is profitable to bitcast an
8617 /// i64 vector to have f64 elements, since the value can then be loaded
8618 /// directly into a VFP register.
8619 static bool hasNormalLoadOperand(SDNode *N) {
8620 unsigned NumElts = N->getValueType(0).getVectorNumElements();
8621 for (unsigned i = 0; i < NumElts; ++i) {
8622 SDNode *Elt = N->getOperand(i).getNode();
8623 if (ISD::isNormalLoad(Elt) && !cast<LoadSDNode>(Elt)->isVolatile())
8629 /// PerformBUILD_VECTORCombine - Target-specific dag combine xforms for
8630 /// ISD::BUILD_VECTOR.
8631 static SDValue PerformBUILD_VECTORCombine(SDNode *N,
8632 TargetLowering::DAGCombinerInfo &DCI,
8633 const ARMSubtarget *Subtarget) {
8634 // build_vector(N=ARMISD::VMOVRRD(X), N:1) -> bit_convert(X):
8635 // VMOVRRD is introduced when legalizing i64 types. It forces the i64 value
8636 // into a pair of GPRs, which is fine when the value is used as a scalar,
8637 // but if the i64 value is converted to a vector, we need to undo the VMOVRRD.
8638 SelectionDAG &DAG = DCI.DAG;
8639 if (N->getNumOperands() == 2) {
8640 SDValue RV = PerformVMOVDRRCombine(N, DAG);
8645 // Load i64 elements as f64 values so that type legalization does not split
8646 // them up into i32 values.
8647 EVT VT = N->getValueType(0);
8648 if (VT.getVectorElementType() != MVT::i64 || !hasNormalLoadOperand(N))
8651 SmallVector<SDValue, 8> Ops;
8652 unsigned NumElts = VT.getVectorNumElements();
8653 for (unsigned i = 0; i < NumElts; ++i) {
8654 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(i));
8656 // Make the DAGCombiner fold the bitcast.
8657 DCI.AddToWorklist(V.getNode());
8659 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, NumElts);
8660 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, FloatVT, Ops);
8661 return DAG.getNode(ISD::BITCAST, dl, VT, BV);
8664 /// \brief Target-specific dag combine xforms for ARMISD::BUILD_VECTOR.
8666 PerformARMBUILD_VECTORCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
8667 // ARMISD::BUILD_VECTOR is introduced when legalizing ISD::BUILD_VECTOR.
8668 // At that time, we may have inserted bitcasts from integer to float.
8669 // If these bitcasts have survived DAGCombine, change the lowering of this
8670 // BUILD_VECTOR in something more vector friendly, i.e., that does not
8671 // force to use floating point types.
8673 // Make sure we can change the type of the vector.
8674 // This is possible iff:
8675 // 1. The vector is only used in a bitcast to a integer type. I.e.,
8676 // 1.1. Vector is used only once.
8677 // 1.2. Use is a bit convert to an integer type.
8678 // 2. The size of its operands are 32-bits (64-bits are not legal).
8679 EVT VT = N->getValueType(0);
8680 EVT EltVT = VT.getVectorElementType();
8682 // Check 1.1. and 2.
8683 if (EltVT.getSizeInBits() != 32 || !N->hasOneUse())
8686 // By construction, the input type must be float.
8687 assert(EltVT == MVT::f32 && "Unexpected type!");
8690 SDNode *Use = *N->use_begin();
8691 if (Use->getOpcode() != ISD::BITCAST ||
8692 Use->getValueType(0).isFloatingPoint())
8695 // Check profitability.
8696 // Model is, if more than half of the relevant operands are bitcast from
8697 // i32, turn the build_vector into a sequence of insert_vector_elt.
8698 // Relevant operands are everything that is not statically
8699 // (i.e., at compile time) bitcasted.
8700 unsigned NumOfBitCastedElts = 0;
8701 unsigned NumElts = VT.getVectorNumElements();
8702 unsigned NumOfRelevantElts = NumElts;
8703 for (unsigned Idx = 0; Idx < NumElts; ++Idx) {
8704 SDValue Elt = N->getOperand(Idx);
8705 if (Elt->getOpcode() == ISD::BITCAST) {
8706 // Assume only bit cast to i32 will go away.
8707 if (Elt->getOperand(0).getValueType() == MVT::i32)
8708 ++NumOfBitCastedElts;
8709 } else if (Elt.getOpcode() == ISD::UNDEF || isa<ConstantSDNode>(Elt))
8710 // Constants are statically casted, thus do not count them as
8711 // relevant operands.
8712 --NumOfRelevantElts;
8715 // Check if more than half of the elements require a non-free bitcast.
8716 if (NumOfBitCastedElts <= NumOfRelevantElts / 2)
8719 SelectionDAG &DAG = DCI.DAG;
8720 // Create the new vector type.
8721 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
8722 // Check if the type is legal.
8723 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8724 if (!TLI.isTypeLegal(VecVT))
8728 // ARMISD::BUILD_VECTOR E1, E2, ..., EN.
8729 // => BITCAST INSERT_VECTOR_ELT
8730 // (INSERT_VECTOR_ELT (...), (BITCAST EN-1), N-1),
8732 SDValue Vec = DAG.getUNDEF(VecVT);
8734 for (unsigned Idx = 0 ; Idx < NumElts; ++Idx) {
8735 SDValue V = N->getOperand(Idx);
8736 if (V.getOpcode() == ISD::UNDEF)
8738 if (V.getOpcode() == ISD::BITCAST &&
8739 V->getOperand(0).getValueType() == MVT::i32)
8740 // Fold obvious case.
8741 V = V.getOperand(0);
8743 V = DAG.getNode(ISD::BITCAST, SDLoc(V), MVT::i32, V);
8744 // Make the DAGCombiner fold the bitcasts.
8745 DCI.AddToWorklist(V.getNode());
8747 SDValue LaneIdx = DAG.getConstant(Idx, dl, MVT::i32);
8748 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VecVT, Vec, V, LaneIdx);
8750 Vec = DAG.getNode(ISD::BITCAST, dl, VT, Vec);
8751 // Make the DAGCombiner fold the bitcasts.
8752 DCI.AddToWorklist(Vec.getNode());
8756 /// PerformInsertEltCombine - Target-specific dag combine xforms for
8757 /// ISD::INSERT_VECTOR_ELT.
8758 static SDValue PerformInsertEltCombine(SDNode *N,
8759 TargetLowering::DAGCombinerInfo &DCI) {
8760 // Bitcast an i64 load inserted into a vector to f64.
8761 // Otherwise, the i64 value will be legalized to a pair of i32 values.
8762 EVT VT = N->getValueType(0);
8763 SDNode *Elt = N->getOperand(1).getNode();
8764 if (VT.getVectorElementType() != MVT::i64 ||
8765 !ISD::isNormalLoad(Elt) || cast<LoadSDNode>(Elt)->isVolatile())
8768 SelectionDAG &DAG = DCI.DAG;
8770 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
8771 VT.getVectorNumElements());
8772 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, N->getOperand(0));
8773 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(1));
8774 // Make the DAGCombiner fold the bitcasts.
8775 DCI.AddToWorklist(Vec.getNode());
8776 DCI.AddToWorklist(V.getNode());
8777 SDValue InsElt = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, FloatVT,
8778 Vec, V, N->getOperand(2));
8779 return DAG.getNode(ISD::BITCAST, dl, VT, InsElt);
8782 /// PerformVECTOR_SHUFFLECombine - Target-specific dag combine xforms for
8783 /// ISD::VECTOR_SHUFFLE.
8784 static SDValue PerformVECTOR_SHUFFLECombine(SDNode *N, SelectionDAG &DAG) {
8785 // The LLVM shufflevector instruction does not require the shuffle mask
8786 // length to match the operand vector length, but ISD::VECTOR_SHUFFLE does
8787 // have that requirement. When translating to ISD::VECTOR_SHUFFLE, if the
8788 // operands do not match the mask length, they are extended by concatenating
8789 // them with undef vectors. That is probably the right thing for other
8790 // targets, but for NEON it is better to concatenate two double-register
8791 // size vector operands into a single quad-register size vector. Do that
8792 // transformation here:
8793 // shuffle(concat(v1, undef), concat(v2, undef)) ->
8794 // shuffle(concat(v1, v2), undef)
8795 SDValue Op0 = N->getOperand(0);
8796 SDValue Op1 = N->getOperand(1);
8797 if (Op0.getOpcode() != ISD::CONCAT_VECTORS ||
8798 Op1.getOpcode() != ISD::CONCAT_VECTORS ||
8799 Op0.getNumOperands() != 2 ||
8800 Op1.getNumOperands() != 2)
8802 SDValue Concat0Op1 = Op0.getOperand(1);
8803 SDValue Concat1Op1 = Op1.getOperand(1);
8804 if (Concat0Op1.getOpcode() != ISD::UNDEF ||
8805 Concat1Op1.getOpcode() != ISD::UNDEF)
8807 // Skip the transformation if any of the types are illegal.
8808 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8809 EVT VT = N->getValueType(0);
8810 if (!TLI.isTypeLegal(VT) ||
8811 !TLI.isTypeLegal(Concat0Op1.getValueType()) ||
8812 !TLI.isTypeLegal(Concat1Op1.getValueType()))
8815 SDValue NewConcat = DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT,
8816 Op0.getOperand(0), Op1.getOperand(0));
8817 // Translate the shuffle mask.
8818 SmallVector<int, 16> NewMask;
8819 unsigned NumElts = VT.getVectorNumElements();
8820 unsigned HalfElts = NumElts/2;
8821 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
8822 for (unsigned n = 0; n < NumElts; ++n) {
8823 int MaskElt = SVN->getMaskElt(n);
8825 if (MaskElt < (int)HalfElts)
8827 else if (MaskElt >= (int)NumElts && MaskElt < (int)(NumElts + HalfElts))
8828 NewElt = HalfElts + MaskElt - NumElts;
8829 NewMask.push_back(NewElt);
8831 return DAG.getVectorShuffle(VT, SDLoc(N), NewConcat,
8832 DAG.getUNDEF(VT), NewMask.data());
8835 /// CombineBaseUpdate - Target-specific DAG combine function for VLDDUP,
8836 /// NEON load/store intrinsics, and generic vector load/stores, to merge
8837 /// base address updates.
8838 /// For generic load/stores, the memory type is assumed to be a vector.
8839 /// The caller is assumed to have checked legality.
8840 static SDValue CombineBaseUpdate(SDNode *N,
8841 TargetLowering::DAGCombinerInfo &DCI) {
8842 SelectionDAG &DAG = DCI.DAG;
8843 const bool isIntrinsic = (N->getOpcode() == ISD::INTRINSIC_VOID ||
8844 N->getOpcode() == ISD::INTRINSIC_W_CHAIN);
8845 const bool isStore = N->getOpcode() == ISD::STORE;
8846 const unsigned AddrOpIdx = ((isIntrinsic || isStore) ? 2 : 1);
8847 SDValue Addr = N->getOperand(AddrOpIdx);
8848 MemSDNode *MemN = cast<MemSDNode>(N);
8851 // Search for a use of the address operand that is an increment.
8852 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(),
8853 UE = Addr.getNode()->use_end(); UI != UE; ++UI) {
8855 if (User->getOpcode() != ISD::ADD ||
8856 UI.getUse().getResNo() != Addr.getResNo())
8859 // Check that the add is independent of the load/store. Otherwise, folding
8860 // it would create a cycle.
8861 if (User->isPredecessorOf(N) || N->isPredecessorOf(User))
8864 // Find the new opcode for the updating load/store.
8865 bool isLoadOp = true;
8866 bool isLaneOp = false;
8867 unsigned NewOpc = 0;
8868 unsigned NumVecs = 0;
8870 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
8872 default: llvm_unreachable("unexpected intrinsic for Neon base update");
8873 case Intrinsic::arm_neon_vld1: NewOpc = ARMISD::VLD1_UPD;
8875 case Intrinsic::arm_neon_vld2: NewOpc = ARMISD::VLD2_UPD;
8877 case Intrinsic::arm_neon_vld3: NewOpc = ARMISD::VLD3_UPD;
8879 case Intrinsic::arm_neon_vld4: NewOpc = ARMISD::VLD4_UPD;
8881 case Intrinsic::arm_neon_vld2lane: NewOpc = ARMISD::VLD2LN_UPD;
8882 NumVecs = 2; isLaneOp = true; break;
8883 case Intrinsic::arm_neon_vld3lane: NewOpc = ARMISD::VLD3LN_UPD;
8884 NumVecs = 3; isLaneOp = true; break;
8885 case Intrinsic::arm_neon_vld4lane: NewOpc = ARMISD::VLD4LN_UPD;
8886 NumVecs = 4; isLaneOp = true; break;
8887 case Intrinsic::arm_neon_vst1: NewOpc = ARMISD::VST1_UPD;
8888 NumVecs = 1; isLoadOp = false; break;
8889 case Intrinsic::arm_neon_vst2: NewOpc = ARMISD::VST2_UPD;
8890 NumVecs = 2; isLoadOp = false; break;
8891 case Intrinsic::arm_neon_vst3: NewOpc = ARMISD::VST3_UPD;
8892 NumVecs = 3; isLoadOp = false; break;
8893 case Intrinsic::arm_neon_vst4: NewOpc = ARMISD::VST4_UPD;
8894 NumVecs = 4; isLoadOp = false; break;
8895 case Intrinsic::arm_neon_vst2lane: NewOpc = ARMISD::VST2LN_UPD;
8896 NumVecs = 2; isLoadOp = false; isLaneOp = true; break;
8897 case Intrinsic::arm_neon_vst3lane: NewOpc = ARMISD::VST3LN_UPD;
8898 NumVecs = 3; isLoadOp = false; isLaneOp = true; break;
8899 case Intrinsic::arm_neon_vst4lane: NewOpc = ARMISD::VST4LN_UPD;
8900 NumVecs = 4; isLoadOp = false; isLaneOp = true; break;
8904 switch (N->getOpcode()) {
8905 default: llvm_unreachable("unexpected opcode for Neon base update");
8906 case ARMISD::VLD2DUP: NewOpc = ARMISD::VLD2DUP_UPD; NumVecs = 2; break;
8907 case ARMISD::VLD3DUP: NewOpc = ARMISD::VLD3DUP_UPD; NumVecs = 3; break;
8908 case ARMISD::VLD4DUP: NewOpc = ARMISD::VLD4DUP_UPD; NumVecs = 4; break;
8909 case ISD::LOAD: NewOpc = ARMISD::VLD1_UPD;
8910 NumVecs = 1; isLaneOp = false; break;
8911 case ISD::STORE: NewOpc = ARMISD::VST1_UPD;
8912 NumVecs = 1; isLaneOp = false; isLoadOp = false; break;
8916 // Find the size of memory referenced by the load/store.
8919 VecTy = N->getValueType(0);
8920 } else if (isIntrinsic) {
8921 VecTy = N->getOperand(AddrOpIdx+1).getValueType();
8923 assert(isStore && "Node has to be a load, a store, or an intrinsic!");
8924 VecTy = N->getOperand(1).getValueType();
8927 unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8;
8929 NumBytes /= VecTy.getVectorNumElements();
8931 // If the increment is a constant, it must match the memory ref size.
8932 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
8933 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
8934 uint64_t IncVal = CInc->getZExtValue();
8935 if (IncVal != NumBytes)
8937 } else if (NumBytes >= 3 * 16) {
8938 // VLD3/4 and VST3/4 for 128-bit vectors are implemented with two
8939 // separate instructions that make it harder to use a non-constant update.
8943 // OK, we found an ADD we can fold into the base update.
8944 // Now, create a _UPD node, taking care of not breaking alignment.
8946 EVT AlignedVecTy = VecTy;
8947 unsigned Alignment = MemN->getAlignment();
8949 // If this is a less-than-standard-aligned load/store, change the type to
8950 // match the standard alignment.
8951 // The alignment is overlooked when selecting _UPD variants; and it's
8952 // easier to introduce bitcasts here than fix that.
8953 // There are 3 ways to get to this base-update combine:
8954 // - intrinsics: they are assumed to be properly aligned (to the standard
8955 // alignment of the memory type), so we don't need to do anything.
8956 // - ARMISD::VLDx nodes: they are only generated from the aforementioned
8957 // intrinsics, so, likewise, there's nothing to do.
8958 // - generic load/store instructions: the alignment is specified as an
8959 // explicit operand, rather than implicitly as the standard alignment
8960 // of the memory type (like the intrisics). We need to change the
8961 // memory type to match the explicit alignment. That way, we don't
8962 // generate non-standard-aligned ARMISD::VLDx nodes.
8963 if (isa<LSBaseSDNode>(N)) {
8966 if (Alignment < VecTy.getScalarSizeInBits() / 8) {
8967 MVT EltTy = MVT::getIntegerVT(Alignment * 8);
8968 assert(NumVecs == 1 && "Unexpected multi-element generic load/store.");
8969 assert(!isLaneOp && "Unexpected generic load/store lane.");
8970 unsigned NumElts = NumBytes / (EltTy.getSizeInBits() / 8);
8971 AlignedVecTy = MVT::getVectorVT(EltTy, NumElts);
8973 // Don't set an explicit alignment on regular load/stores that we want
8974 // to transform to VLD/VST 1_UPD nodes.
8975 // This matches the behavior of regular load/stores, which only get an
8976 // explicit alignment if the MMO alignment is larger than the standard
8977 // alignment of the memory type.
8978 // Intrinsics, however, always get an explicit alignment, set to the
8979 // alignment of the MMO.
8983 // Create the new updating load/store node.
8984 // First, create an SDVTList for the new updating node's results.
8986 unsigned NumResultVecs = (isLoadOp ? NumVecs : 0);
8988 for (n = 0; n < NumResultVecs; ++n)
8989 Tys[n] = AlignedVecTy;
8990 Tys[n++] = MVT::i32;
8991 Tys[n] = MVT::Other;
8992 SDVTList SDTys = DAG.getVTList(makeArrayRef(Tys, NumResultVecs+2));
8994 // Then, gather the new node's operands.
8995 SmallVector<SDValue, 8> Ops;
8996 Ops.push_back(N->getOperand(0)); // incoming chain
8997 Ops.push_back(N->getOperand(AddrOpIdx));
9000 if (StoreSDNode *StN = dyn_cast<StoreSDNode>(N)) {
9001 // Try to match the intrinsic's signature
9002 Ops.push_back(StN->getValue());
9004 // Loads (and of course intrinsics) match the intrinsics' signature,
9005 // so just add all but the alignment operand.
9006 for (unsigned i = AddrOpIdx + 1; i < N->getNumOperands() - 1; ++i)
9007 Ops.push_back(N->getOperand(i));
9010 // For all node types, the alignment operand is always the last one.
9011 Ops.push_back(DAG.getConstant(Alignment, dl, MVT::i32));
9013 // If this is a non-standard-aligned STORE, the penultimate operand is the
9014 // stored value. Bitcast it to the aligned type.
9015 if (AlignedVecTy != VecTy && N->getOpcode() == ISD::STORE) {
9016 SDValue &StVal = Ops[Ops.size()-2];
9017 StVal = DAG.getNode(ISD::BITCAST, dl, AlignedVecTy, StVal);
9020 SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, dl, SDTys,
9022 MemN->getMemOperand());
9025 SmallVector<SDValue, 5> NewResults;
9026 for (unsigned i = 0; i < NumResultVecs; ++i)
9027 NewResults.push_back(SDValue(UpdN.getNode(), i));
9029 // If this is an non-standard-aligned LOAD, the first result is the loaded
9030 // value. Bitcast it to the expected result type.
9031 if (AlignedVecTy != VecTy && N->getOpcode() == ISD::LOAD) {
9032 SDValue &LdVal = NewResults[0];
9033 LdVal = DAG.getNode(ISD::BITCAST, dl, VecTy, LdVal);
9036 NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs+1)); // chain
9037 DCI.CombineTo(N, NewResults);
9038 DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs));
9045 static SDValue PerformVLDCombine(SDNode *N,
9046 TargetLowering::DAGCombinerInfo &DCI) {
9047 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
9050 return CombineBaseUpdate(N, DCI);
9053 /// CombineVLDDUP - For a VDUPLANE node N, check if its source operand is a
9054 /// vldN-lane (N > 1) intrinsic, and if all the other uses of that intrinsic
9055 /// are also VDUPLANEs. If so, combine them to a vldN-dup operation and
9057 static bool CombineVLDDUP(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
9058 SelectionDAG &DAG = DCI.DAG;
9059 EVT VT = N->getValueType(0);
9060 // vldN-dup instructions only support 64-bit vectors for N > 1.
9061 if (!VT.is64BitVector())
9064 // Check if the VDUPLANE operand is a vldN-dup intrinsic.
9065 SDNode *VLD = N->getOperand(0).getNode();
9066 if (VLD->getOpcode() != ISD::INTRINSIC_W_CHAIN)
9068 unsigned NumVecs = 0;
9069 unsigned NewOpc = 0;
9070 unsigned IntNo = cast<ConstantSDNode>(VLD->getOperand(1))->getZExtValue();
9071 if (IntNo == Intrinsic::arm_neon_vld2lane) {
9073 NewOpc = ARMISD::VLD2DUP;
9074 } else if (IntNo == Intrinsic::arm_neon_vld3lane) {
9076 NewOpc = ARMISD::VLD3DUP;
9077 } else if (IntNo == Intrinsic::arm_neon_vld4lane) {
9079 NewOpc = ARMISD::VLD4DUP;
9084 // First check that all the vldN-lane uses are VDUPLANEs and that the lane
9085 // numbers match the load.
9086 unsigned VLDLaneNo =
9087 cast<ConstantSDNode>(VLD->getOperand(NumVecs+3))->getZExtValue();
9088 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
9090 // Ignore uses of the chain result.
9091 if (UI.getUse().getResNo() == NumVecs)
9094 if (User->getOpcode() != ARMISD::VDUPLANE ||
9095 VLDLaneNo != cast<ConstantSDNode>(User->getOperand(1))->getZExtValue())
9099 // Create the vldN-dup node.
9102 for (n = 0; n < NumVecs; ++n)
9104 Tys[n] = MVT::Other;
9105 SDVTList SDTys = DAG.getVTList(makeArrayRef(Tys, NumVecs+1));
9106 SDValue Ops[] = { VLD->getOperand(0), VLD->getOperand(2) };
9107 MemIntrinsicSDNode *VLDMemInt = cast<MemIntrinsicSDNode>(VLD);
9108 SDValue VLDDup = DAG.getMemIntrinsicNode(NewOpc, SDLoc(VLD), SDTys,
9109 Ops, VLDMemInt->getMemoryVT(),
9110 VLDMemInt->getMemOperand());
9113 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
9115 unsigned ResNo = UI.getUse().getResNo();
9116 // Ignore uses of the chain result.
9117 if (ResNo == NumVecs)
9120 DCI.CombineTo(User, SDValue(VLDDup.getNode(), ResNo));
9123 // Now the vldN-lane intrinsic is dead except for its chain result.
9124 // Update uses of the chain.
9125 std::vector<SDValue> VLDDupResults;
9126 for (unsigned n = 0; n < NumVecs; ++n)
9127 VLDDupResults.push_back(SDValue(VLDDup.getNode(), n));
9128 VLDDupResults.push_back(SDValue(VLDDup.getNode(), NumVecs));
9129 DCI.CombineTo(VLD, VLDDupResults);
9134 /// PerformVDUPLANECombine - Target-specific dag combine xforms for
9135 /// ARMISD::VDUPLANE.
9136 static SDValue PerformVDUPLANECombine(SDNode *N,
9137 TargetLowering::DAGCombinerInfo &DCI) {
9138 SDValue Op = N->getOperand(0);
9140 // If the source is a vldN-lane (N > 1) intrinsic, and all the other uses
9141 // of that intrinsic are also VDUPLANEs, combine them to a vldN-dup operation.
9142 if (CombineVLDDUP(N, DCI))
9143 return SDValue(N, 0);
9145 // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is
9146 // redundant. Ignore bit_converts for now; element sizes are checked below.
9147 while (Op.getOpcode() == ISD::BITCAST)
9148 Op = Op.getOperand(0);
9149 if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM)
9152 // Make sure the VMOV element size is not bigger than the VDUPLANE elements.
9153 unsigned EltSize = Op.getValueType().getVectorElementType().getSizeInBits();
9154 // The canonical VMOV for a zero vector uses a 32-bit element size.
9155 unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
9157 if (ARM_AM::decodeNEONModImm(Imm, EltBits) == 0)
9159 EVT VT = N->getValueType(0);
9160 if (EltSize > VT.getVectorElementType().getSizeInBits())
9163 return DCI.DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Op);
9166 static SDValue PerformLOADCombine(SDNode *N,
9167 TargetLowering::DAGCombinerInfo &DCI) {
9168 EVT VT = N->getValueType(0);
9170 // If this is a legal vector load, try to combine it into a VLD1_UPD.
9171 if (ISD::isNormalLoad(N) && VT.isVector() &&
9172 DCI.DAG.getTargetLoweringInfo().isTypeLegal(VT))
9173 return CombineBaseUpdate(N, DCI);
9178 /// PerformSTORECombine - Target-specific dag combine xforms for
9180 static SDValue PerformSTORECombine(SDNode *N,
9181 TargetLowering::DAGCombinerInfo &DCI) {
9182 StoreSDNode *St = cast<StoreSDNode>(N);
9183 if (St->isVolatile())
9186 // Optimize trunc store (of multiple scalars) to shuffle and store. First,
9187 // pack all of the elements in one place. Next, store to memory in fewer
9189 SDValue StVal = St->getValue();
9190 EVT VT = StVal.getValueType();
9191 if (St->isTruncatingStore() && VT.isVector()) {
9192 SelectionDAG &DAG = DCI.DAG;
9193 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9194 EVT StVT = St->getMemoryVT();
9195 unsigned NumElems = VT.getVectorNumElements();
9196 assert(StVT != VT && "Cannot truncate to the same type");
9197 unsigned FromEltSz = VT.getVectorElementType().getSizeInBits();
9198 unsigned ToEltSz = StVT.getVectorElementType().getSizeInBits();
9200 // From, To sizes and ElemCount must be pow of two
9201 if (!isPowerOf2_32(NumElems * FromEltSz * ToEltSz)) return SDValue();
9203 // We are going to use the original vector elt for storing.
9204 // Accumulated smaller vector elements must be a multiple of the store size.
9205 if (0 != (NumElems * FromEltSz) % ToEltSz) return SDValue();
9207 unsigned SizeRatio = FromEltSz / ToEltSz;
9208 assert(SizeRatio * NumElems * ToEltSz == VT.getSizeInBits());
9210 // Create a type on which we perform the shuffle.
9211 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(), StVT.getScalarType(),
9212 NumElems*SizeRatio);
9213 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
9216 SDValue WideVec = DAG.getNode(ISD::BITCAST, DL, WideVecVT, StVal);
9217 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
9218 for (unsigned i = 0; i < NumElems; ++i)
9219 ShuffleVec[i] = TLI.isBigEndian() ? (i+1) * SizeRatio - 1 : i * SizeRatio;
9221 // Can't shuffle using an illegal type.
9222 if (!TLI.isTypeLegal(WideVecVT)) return SDValue();
9224 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, DL, WideVec,
9225 DAG.getUNDEF(WideVec.getValueType()),
9227 // At this point all of the data is stored at the bottom of the
9228 // register. We now need to save it to mem.
9230 // Find the largest store unit
9231 MVT StoreType = MVT::i8;
9232 for (MVT Tp : MVT::integer_valuetypes()) {
9233 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToEltSz)
9236 // Didn't find a legal store type.
9237 if (!TLI.isTypeLegal(StoreType))
9240 // Bitcast the original vector into a vector of store-size units
9241 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
9242 StoreType, VT.getSizeInBits()/EVT(StoreType).getSizeInBits());
9243 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
9244 SDValue ShuffWide = DAG.getNode(ISD::BITCAST, DL, StoreVecVT, Shuff);
9245 SmallVector<SDValue, 8> Chains;
9246 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits()/8, DL,
9247 TLI.getPointerTy());
9248 SDValue BasePtr = St->getBasePtr();
9250 // Perform one or more big stores into memory.
9251 unsigned E = (ToEltSz*NumElems)/StoreType.getSizeInBits();
9252 for (unsigned I = 0; I < E; I++) {
9253 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
9254 StoreType, ShuffWide,
9255 DAG.getIntPtrConstant(I, DL));
9256 SDValue Ch = DAG.getStore(St->getChain(), DL, SubVec, BasePtr,
9257 St->getPointerInfo(), St->isVolatile(),
9258 St->isNonTemporal(), St->getAlignment());
9259 BasePtr = DAG.getNode(ISD::ADD, DL, BasePtr.getValueType(), BasePtr,
9261 Chains.push_back(Ch);
9263 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
9266 if (!ISD::isNormalStore(St))
9269 // Split a store of a VMOVDRR into two integer stores to avoid mixing NEON and
9270 // ARM stores of arguments in the same cache line.
9271 if (StVal.getNode()->getOpcode() == ARMISD::VMOVDRR &&
9272 StVal.getNode()->hasOneUse()) {
9273 SelectionDAG &DAG = DCI.DAG;
9274 bool isBigEndian = DAG.getTargetLoweringInfo().isBigEndian();
9276 SDValue BasePtr = St->getBasePtr();
9277 SDValue NewST1 = DAG.getStore(St->getChain(), DL,
9278 StVal.getNode()->getOperand(isBigEndian ? 1 : 0 ),
9279 BasePtr, St->getPointerInfo(), St->isVolatile(),
9280 St->isNonTemporal(), St->getAlignment());
9282 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
9283 DAG.getConstant(4, DL, MVT::i32));
9284 return DAG.getStore(NewST1.getValue(0), DL,
9285 StVal.getNode()->getOperand(isBigEndian ? 0 : 1),
9286 OffsetPtr, St->getPointerInfo(), St->isVolatile(),
9287 St->isNonTemporal(),
9288 std::min(4U, St->getAlignment() / 2));
9291 if (StVal.getValueType() == MVT::i64 &&
9292 StVal.getNode()->getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
9294 // Bitcast an i64 store extracted from a vector to f64.
9295 // Otherwise, the i64 value will be legalized to a pair of i32 values.
9296 SelectionDAG &DAG = DCI.DAG;
9298 SDValue IntVec = StVal.getOperand(0);
9299 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
9300 IntVec.getValueType().getVectorNumElements());
9301 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, IntVec);
9302 SDValue ExtElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
9303 Vec, StVal.getOperand(1));
9305 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ExtElt);
9306 // Make the DAGCombiner fold the bitcasts.
9307 DCI.AddToWorklist(Vec.getNode());
9308 DCI.AddToWorklist(ExtElt.getNode());
9309 DCI.AddToWorklist(V.getNode());
9310 return DAG.getStore(St->getChain(), dl, V, St->getBasePtr(),
9311 St->getPointerInfo(), St->isVolatile(),
9312 St->isNonTemporal(), St->getAlignment(),
9316 // If this is a legal vector store, try to combine it into a VST1_UPD.
9317 if (ISD::isNormalStore(N) && VT.isVector() &&
9318 DCI.DAG.getTargetLoweringInfo().isTypeLegal(VT))
9319 return CombineBaseUpdate(N, DCI);
9324 // isConstVecPow2 - Return true if each vector element is a power of 2, all
9325 // elements are the same constant, C, and Log2(C) ranges from 1 to 32.
9326 static bool isConstVecPow2(SDValue ConstVec, bool isSigned, uint64_t &C)
9330 for (unsigned I = 0, E = ConstVec.getValueType().getVectorNumElements();
9332 ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(ConstVec.getOperand(I));
9337 APFloat APF = C->getValueAPF();
9338 if (APF.convertToInteger(&cN, 64, isSigned, APFloat::rmTowardZero, &isExact)
9339 != APFloat::opOK || !isExact)
9342 c0 = (I == 0) ? cN : c0;
9343 if (!isPowerOf2_64(cN) || c0 != cN || Log2_64(c0) < 1 || Log2_64(c0) > 32)
9350 /// PerformVCVTCombine - VCVT (floating-point to fixed-point, Advanced SIMD)
9351 /// can replace combinations of VMUL and VCVT (floating-point to integer)
9352 /// when the VMUL has a constant operand that is a power of 2.
9354 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
9355 /// vmul.f32 d16, d17, d16
9356 /// vcvt.s32.f32 d16, d16
9358 /// vcvt.s32.f32 d16, d16, #3
9359 static SDValue PerformVCVTCombine(SDNode *N,
9360 TargetLowering::DAGCombinerInfo &DCI,
9361 const ARMSubtarget *Subtarget) {
9362 SelectionDAG &DAG = DCI.DAG;
9363 SDValue Op = N->getOperand(0);
9365 if (!Subtarget->hasNEON() || !Op.getValueType().isVector() ||
9366 Op.getOpcode() != ISD::FMUL)
9370 SDValue N0 = Op->getOperand(0);
9371 SDValue ConstVec = Op->getOperand(1);
9372 bool isSigned = N->getOpcode() == ISD::FP_TO_SINT;
9374 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
9375 !isConstVecPow2(ConstVec, isSigned, C))
9378 MVT FloatTy = Op.getSimpleValueType().getVectorElementType();
9379 MVT IntTy = N->getSimpleValueType(0).getVectorElementType();
9380 unsigned NumLanes = Op.getValueType().getVectorNumElements();
9381 if (FloatTy.getSizeInBits() != 32 || IntTy.getSizeInBits() > 32 ||
9383 // These instructions only exist converting from f32 to i32. We can handle
9384 // smaller integers by generating an extra truncate, but larger ones would
9385 // be lossy. We also can't handle more then 4 lanes, since these intructions
9386 // only support v2i32/v4i32 types.
9391 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfp2fxs :
9392 Intrinsic::arm_neon_vcvtfp2fxu;
9393 SDValue FixConv = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl,
9394 NumLanes == 2 ? MVT::v2i32 : MVT::v4i32,
9395 DAG.getConstant(IntrinsicOpcode, dl, MVT::i32),
9397 DAG.getConstant(Log2_64(C), dl, MVT::i32));
9399 if (IntTy.getSizeInBits() < FloatTy.getSizeInBits())
9400 FixConv = DAG.getNode(ISD::TRUNCATE, dl, N->getValueType(0), FixConv);
9405 /// PerformVDIVCombine - VCVT (fixed-point to floating-point, Advanced SIMD)
9406 /// can replace combinations of VCVT (integer to floating-point) and VDIV
9407 /// when the VDIV has a constant operand that is a power of 2.
9409 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
9410 /// vcvt.f32.s32 d16, d16
9411 /// vdiv.f32 d16, d17, d16
9413 /// vcvt.f32.s32 d16, d16, #3
9414 static SDValue PerformVDIVCombine(SDNode *N,
9415 TargetLowering::DAGCombinerInfo &DCI,
9416 const ARMSubtarget *Subtarget) {
9417 SelectionDAG &DAG = DCI.DAG;
9418 SDValue Op = N->getOperand(0);
9419 unsigned OpOpcode = Op.getNode()->getOpcode();
9421 if (!Subtarget->hasNEON() || !N->getValueType(0).isVector() ||
9422 (OpOpcode != ISD::SINT_TO_FP && OpOpcode != ISD::UINT_TO_FP))
9426 SDValue ConstVec = N->getOperand(1);
9427 bool isSigned = OpOpcode == ISD::SINT_TO_FP;
9429 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
9430 !isConstVecPow2(ConstVec, isSigned, C))
9433 MVT FloatTy = N->getSimpleValueType(0).getVectorElementType();
9434 MVT IntTy = Op.getOperand(0).getSimpleValueType().getVectorElementType();
9435 if (FloatTy.getSizeInBits() != 32 || IntTy.getSizeInBits() > 32) {
9436 // These instructions only exist converting from i32 to f32. We can handle
9437 // smaller integers by generating an extra extend, but larger ones would
9443 SDValue ConvInput = Op.getOperand(0);
9444 unsigned NumLanes = Op.getValueType().getVectorNumElements();
9445 if (IntTy.getSizeInBits() < FloatTy.getSizeInBits())
9446 ConvInput = DAG.getNode(isSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND,
9447 dl, NumLanes == 2 ? MVT::v2i32 : MVT::v4i32,
9450 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfxs2fp :
9451 Intrinsic::arm_neon_vcvtfxu2fp;
9452 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl,
9454 DAG.getConstant(IntrinsicOpcode, dl, MVT::i32),
9455 ConvInput, DAG.getConstant(Log2_64(C), dl, MVT::i32));
9458 /// Getvshiftimm - Check if this is a valid build_vector for the immediate
9459 /// operand of a vector shift operation, where all the elements of the
9460 /// build_vector must have the same constant integer value.
9461 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
9462 // Ignore bit_converts.
9463 while (Op.getOpcode() == ISD::BITCAST)
9464 Op = Op.getOperand(0);
9465 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
9466 APInt SplatBits, SplatUndef;
9467 unsigned SplatBitSize;
9469 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
9470 HasAnyUndefs, ElementBits) ||
9471 SplatBitSize > ElementBits)
9473 Cnt = SplatBits.getSExtValue();
9477 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
9478 /// operand of a vector shift left operation. That value must be in the range:
9479 /// 0 <= Value < ElementBits for a left shift; or
9480 /// 0 <= Value <= ElementBits for a long left shift.
9481 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
9482 assert(VT.isVector() && "vector shift count is not a vector type");
9483 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
9484 if (! getVShiftImm(Op, ElementBits, Cnt))
9486 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
9489 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
9490 /// operand of a vector shift right operation. For a shift opcode, the value
9491 /// is positive, but for an intrinsic the value count must be negative. The
9492 /// absolute value must be in the range:
9493 /// 1 <= |Value| <= ElementBits for a right shift; or
9494 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
9495 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
9497 assert(VT.isVector() && "vector shift count is not a vector type");
9498 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
9499 if (! getVShiftImm(Op, ElementBits, Cnt))
9503 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
9506 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
9507 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
9508 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
9511 // Don't do anything for most intrinsics.
9514 // Vector shifts: check for immediate versions and lower them.
9515 // Note: This is done during DAG combining instead of DAG legalizing because
9516 // the build_vectors for 64-bit vector element shift counts are generally
9517 // not legal, and it is hard to see their values after they get legalized to
9518 // loads from a constant pool.
9519 case Intrinsic::arm_neon_vshifts:
9520 case Intrinsic::arm_neon_vshiftu:
9521 case Intrinsic::arm_neon_vrshifts:
9522 case Intrinsic::arm_neon_vrshiftu:
9523 case Intrinsic::arm_neon_vrshiftn:
9524 case Intrinsic::arm_neon_vqshifts:
9525 case Intrinsic::arm_neon_vqshiftu:
9526 case Intrinsic::arm_neon_vqshiftsu:
9527 case Intrinsic::arm_neon_vqshiftns:
9528 case Intrinsic::arm_neon_vqshiftnu:
9529 case Intrinsic::arm_neon_vqshiftnsu:
9530 case Intrinsic::arm_neon_vqrshiftns:
9531 case Intrinsic::arm_neon_vqrshiftnu:
9532 case Intrinsic::arm_neon_vqrshiftnsu: {
9533 EVT VT = N->getOperand(1).getValueType();
9535 unsigned VShiftOpc = 0;
9538 case Intrinsic::arm_neon_vshifts:
9539 case Intrinsic::arm_neon_vshiftu:
9540 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
9541 VShiftOpc = ARMISD::VSHL;
9544 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
9545 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
9546 ARMISD::VSHRs : ARMISD::VSHRu);
9551 case Intrinsic::arm_neon_vrshifts:
9552 case Intrinsic::arm_neon_vrshiftu:
9553 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
9557 case Intrinsic::arm_neon_vqshifts:
9558 case Intrinsic::arm_neon_vqshiftu:
9559 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
9563 case Intrinsic::arm_neon_vqshiftsu:
9564 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
9566 llvm_unreachable("invalid shift count for vqshlu intrinsic");
9568 case Intrinsic::arm_neon_vrshiftn:
9569 case Intrinsic::arm_neon_vqshiftns:
9570 case Intrinsic::arm_neon_vqshiftnu:
9571 case Intrinsic::arm_neon_vqshiftnsu:
9572 case Intrinsic::arm_neon_vqrshiftns:
9573 case Intrinsic::arm_neon_vqrshiftnu:
9574 case Intrinsic::arm_neon_vqrshiftnsu:
9575 // Narrowing shifts require an immediate right shift.
9576 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
9578 llvm_unreachable("invalid shift count for narrowing vector shift "
9582 llvm_unreachable("unhandled vector shift");
9586 case Intrinsic::arm_neon_vshifts:
9587 case Intrinsic::arm_neon_vshiftu:
9588 // Opcode already set above.
9590 case Intrinsic::arm_neon_vrshifts:
9591 VShiftOpc = ARMISD::VRSHRs; break;
9592 case Intrinsic::arm_neon_vrshiftu:
9593 VShiftOpc = ARMISD::VRSHRu; break;
9594 case Intrinsic::arm_neon_vrshiftn:
9595 VShiftOpc = ARMISD::VRSHRN; break;
9596 case Intrinsic::arm_neon_vqshifts:
9597 VShiftOpc = ARMISD::VQSHLs; break;
9598 case Intrinsic::arm_neon_vqshiftu:
9599 VShiftOpc = ARMISD::VQSHLu; break;
9600 case Intrinsic::arm_neon_vqshiftsu:
9601 VShiftOpc = ARMISD::VQSHLsu; break;
9602 case Intrinsic::arm_neon_vqshiftns:
9603 VShiftOpc = ARMISD::VQSHRNs; break;
9604 case Intrinsic::arm_neon_vqshiftnu:
9605 VShiftOpc = ARMISD::VQSHRNu; break;
9606 case Intrinsic::arm_neon_vqshiftnsu:
9607 VShiftOpc = ARMISD::VQSHRNsu; break;
9608 case Intrinsic::arm_neon_vqrshiftns:
9609 VShiftOpc = ARMISD::VQRSHRNs; break;
9610 case Intrinsic::arm_neon_vqrshiftnu:
9611 VShiftOpc = ARMISD::VQRSHRNu; break;
9612 case Intrinsic::arm_neon_vqrshiftnsu:
9613 VShiftOpc = ARMISD::VQRSHRNsu; break;
9617 return DAG.getNode(VShiftOpc, dl, N->getValueType(0),
9618 N->getOperand(1), DAG.getConstant(Cnt, dl, MVT::i32));
9621 case Intrinsic::arm_neon_vshiftins: {
9622 EVT VT = N->getOperand(1).getValueType();
9624 unsigned VShiftOpc = 0;
9626 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
9627 VShiftOpc = ARMISD::VSLI;
9628 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
9629 VShiftOpc = ARMISD::VSRI;
9631 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
9635 return DAG.getNode(VShiftOpc, dl, N->getValueType(0),
9636 N->getOperand(1), N->getOperand(2),
9637 DAG.getConstant(Cnt, dl, MVT::i32));
9640 case Intrinsic::arm_neon_vqrshifts:
9641 case Intrinsic::arm_neon_vqrshiftu:
9642 // No immediate versions of these to check for.
9649 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
9650 /// lowers them. As with the vector shift intrinsics, this is done during DAG
9651 /// combining instead of DAG legalizing because the build_vectors for 64-bit
9652 /// vector element shift counts are generally not legal, and it is hard to see
9653 /// their values after they get legalized to loads from a constant pool.
9654 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
9655 const ARMSubtarget *ST) {
9656 EVT VT = N->getValueType(0);
9657 if (N->getOpcode() == ISD::SRL && VT == MVT::i32 && ST->hasV6Ops()) {
9658 // Canonicalize (srl (bswap x), 16) to (rotr (bswap x), 16) if the high
9659 // 16-bits of x is zero. This optimizes rev + lsr 16 to rev16.
9660 SDValue N1 = N->getOperand(1);
9661 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) {
9662 SDValue N0 = N->getOperand(0);
9663 if (C->getZExtValue() == 16 && N0.getOpcode() == ISD::BSWAP &&
9664 DAG.MaskedValueIsZero(N0.getOperand(0),
9665 APInt::getHighBitsSet(32, 16)))
9666 return DAG.getNode(ISD::ROTR, SDLoc(N), VT, N0, N1);
9670 // Nothing to be done for scalar shifts.
9671 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9672 if (!VT.isVector() || !TLI.isTypeLegal(VT))
9675 assert(ST->hasNEON() && "unexpected vector shift");
9678 switch (N->getOpcode()) {
9679 default: llvm_unreachable("unexpected shift opcode");
9682 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt)) {
9684 return DAG.getNode(ARMISD::VSHL, dl, VT, N->getOperand(0),
9685 DAG.getConstant(Cnt, dl, MVT::i32));
9691 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
9692 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
9693 ARMISD::VSHRs : ARMISD::VSHRu);
9695 return DAG.getNode(VShiftOpc, dl, VT, N->getOperand(0),
9696 DAG.getConstant(Cnt, dl, MVT::i32));
9702 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
9703 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
9704 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
9705 const ARMSubtarget *ST) {
9706 SDValue N0 = N->getOperand(0);
9708 // Check for sign- and zero-extensions of vector extract operations of 8-
9709 // and 16-bit vector elements. NEON supports these directly. They are
9710 // handled during DAG combining because type legalization will promote them
9711 // to 32-bit types and it is messy to recognize the operations after that.
9712 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
9713 SDValue Vec = N0.getOperand(0);
9714 SDValue Lane = N0.getOperand(1);
9715 EVT VT = N->getValueType(0);
9716 EVT EltVT = N0.getValueType();
9717 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9719 if (VT == MVT::i32 &&
9720 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
9721 TLI.isTypeLegal(Vec.getValueType()) &&
9722 isa<ConstantSDNode>(Lane)) {
9725 switch (N->getOpcode()) {
9726 default: llvm_unreachable("unexpected opcode");
9727 case ISD::SIGN_EXTEND:
9728 Opc = ARMISD::VGETLANEs;
9730 case ISD::ZERO_EXTEND:
9731 case ISD::ANY_EXTEND:
9732 Opc = ARMISD::VGETLANEu;
9735 return DAG.getNode(Opc, SDLoc(N), VT, Vec, Lane);
9742 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
9743 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
9744 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
9745 const ARMSubtarget *ST) {
9746 // If the target supports NEON, try to use vmax/vmin instructions for f32
9747 // selects like "x < y ? x : y". Unless the NoNaNsFPMath option is set,
9748 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
9749 // a NaN; only do the transformation when it matches that behavior.
9751 // For now only do this when using NEON for FP operations; if using VFP, it
9752 // is not obvious that the benefit outweighs the cost of switching to the
9754 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
9755 N->getValueType(0) != MVT::f32)
9758 SDValue CondLHS = N->getOperand(0);
9759 SDValue CondRHS = N->getOperand(1);
9760 SDValue LHS = N->getOperand(2);
9761 SDValue RHS = N->getOperand(3);
9762 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
9764 unsigned Opcode = 0;
9766 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
9767 IsReversed = false; // x CC y ? x : y
9768 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
9769 IsReversed = true ; // x CC y ? y : x
9783 // If LHS is NaN, an ordered comparison will be false and the result will
9784 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
9785 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
9786 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
9787 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
9789 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
9790 // will return -0, so vmin can only be used for unsafe math or if one of
9791 // the operands is known to be nonzero.
9792 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
9793 !DAG.getTarget().Options.UnsafeFPMath &&
9794 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9796 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
9805 // If LHS is NaN, an ordered comparison will be false and the result will
9806 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
9807 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
9808 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
9809 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
9811 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
9812 // will return +0, so vmax can only be used for unsafe math or if one of
9813 // the operands is known to be nonzero.
9814 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
9815 !DAG.getTarget().Options.UnsafeFPMath &&
9816 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9818 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
9824 return DAG.getNode(Opcode, SDLoc(N), N->getValueType(0), LHS, RHS);
9827 /// PerformCMOVCombine - Target-specific DAG combining for ARMISD::CMOV.
9829 ARMTargetLowering::PerformCMOVCombine(SDNode *N, SelectionDAG &DAG) const {
9830 SDValue Cmp = N->getOperand(4);
9831 if (Cmp.getOpcode() != ARMISD::CMPZ)
9832 // Only looking at EQ and NE cases.
9835 EVT VT = N->getValueType(0);
9837 SDValue LHS = Cmp.getOperand(0);
9838 SDValue RHS = Cmp.getOperand(1);
9839 SDValue FalseVal = N->getOperand(0);
9840 SDValue TrueVal = N->getOperand(1);
9841 SDValue ARMcc = N->getOperand(2);
9842 ARMCC::CondCodes CC =
9843 (ARMCC::CondCodes)cast<ConstantSDNode>(ARMcc)->getZExtValue();
9861 /// FIXME: Turn this into a target neutral optimization?
9863 if (CC == ARMCC::NE && FalseVal == RHS && FalseVal != LHS) {
9864 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, TrueVal, ARMcc,
9865 N->getOperand(3), Cmp);
9866 } else if (CC == ARMCC::EQ && TrueVal == RHS) {
9868 SDValue NewCmp = getARMCmp(LHS, RHS, ISD::SETNE, ARMcc, DAG, dl);
9869 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, FalseVal, ARMcc,
9870 N->getOperand(3), NewCmp);
9873 if (Res.getNode()) {
9874 APInt KnownZero, KnownOne;
9875 DAG.computeKnownBits(SDValue(N,0), KnownZero, KnownOne);
9876 // Capture demanded bits information that would be otherwise lost.
9877 if (KnownZero == 0xfffffffe)
9878 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
9879 DAG.getValueType(MVT::i1));
9880 else if (KnownZero == 0xffffff00)
9881 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
9882 DAG.getValueType(MVT::i8));
9883 else if (KnownZero == 0xffff0000)
9884 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
9885 DAG.getValueType(MVT::i16));
9891 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
9892 DAGCombinerInfo &DCI) const {
9893 switch (N->getOpcode()) {
9895 case ISD::ADDC: return PerformADDCCombine(N, DCI, Subtarget);
9896 case ISD::ADD: return PerformADDCombine(N, DCI, Subtarget);
9897 case ISD::SUB: return PerformSUBCombine(N, DCI);
9898 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
9899 case ISD::OR: return PerformORCombine(N, DCI, Subtarget);
9900 case ISD::XOR: return PerformXORCombine(N, DCI, Subtarget);
9901 case ISD::AND: return PerformANDCombine(N, DCI, Subtarget);
9902 case ARMISD::BFI: return PerformBFICombine(N, DCI);
9903 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI, Subtarget);
9904 case ARMISD::VMOVDRR: return PerformVMOVDRRCombine(N, DCI.DAG);
9905 case ISD::STORE: return PerformSTORECombine(N, DCI);
9906 case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DCI, Subtarget);
9907 case ISD::INSERT_VECTOR_ELT: return PerformInsertEltCombine(N, DCI);
9908 case ISD::VECTOR_SHUFFLE: return PerformVECTOR_SHUFFLECombine(N, DCI.DAG);
9909 case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI);
9910 case ISD::FP_TO_SINT:
9911 case ISD::FP_TO_UINT: return PerformVCVTCombine(N, DCI, Subtarget);
9912 case ISD::FDIV: return PerformVDIVCombine(N, DCI, Subtarget);
9913 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
9916 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
9917 case ISD::SIGN_EXTEND:
9918 case ISD::ZERO_EXTEND:
9919 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
9920 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
9921 case ARMISD::CMOV: return PerformCMOVCombine(N, DCI.DAG);
9922 case ISD::LOAD: return PerformLOADCombine(N, DCI);
9923 case ARMISD::VLD2DUP:
9924 case ARMISD::VLD3DUP:
9925 case ARMISD::VLD4DUP:
9926 return PerformVLDCombine(N, DCI);
9927 case ARMISD::BUILD_VECTOR:
9928 return PerformARMBUILD_VECTORCombine(N, DCI);
9929 case ISD::INTRINSIC_VOID:
9930 case ISD::INTRINSIC_W_CHAIN:
9931 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
9932 case Intrinsic::arm_neon_vld1:
9933 case Intrinsic::arm_neon_vld2:
9934 case Intrinsic::arm_neon_vld3:
9935 case Intrinsic::arm_neon_vld4:
9936 case Intrinsic::arm_neon_vld2lane:
9937 case Intrinsic::arm_neon_vld3lane:
9938 case Intrinsic::arm_neon_vld4lane:
9939 case Intrinsic::arm_neon_vst1:
9940 case Intrinsic::arm_neon_vst2:
9941 case Intrinsic::arm_neon_vst3:
9942 case Intrinsic::arm_neon_vst4:
9943 case Intrinsic::arm_neon_vst2lane:
9944 case Intrinsic::arm_neon_vst3lane:
9945 case Intrinsic::arm_neon_vst4lane:
9946 return PerformVLDCombine(N, DCI);
9954 bool ARMTargetLowering::isDesirableToTransformToIntegerOp(unsigned Opc,
9956 return (VT == MVT::f32) && (Opc == ISD::LOAD || Opc == ISD::STORE);
9959 bool ARMTargetLowering::allowsMisalignedMemoryAccesses(EVT VT,
9963 // The AllowsUnaliged flag models the SCTLR.A setting in ARM cpus
9964 bool AllowsUnaligned = Subtarget->allowsUnalignedMem();
9966 switch (VT.getSimpleVT().SimpleTy) {
9972 // Unaligned access can use (for example) LRDB, LRDH, LDR
9973 if (AllowsUnaligned) {
9975 *Fast = Subtarget->hasV7Ops();
9982 // For any little-endian targets with neon, we can support unaligned ld/st
9983 // of D and Q (e.g. {D0,D1}) registers by using vld1.i8/vst1.i8.
9984 // A big-endian target may also explicitly support unaligned accesses
9985 if (Subtarget->hasNEON() && (AllowsUnaligned || isLittleEndian())) {
9995 static bool memOpAlign(unsigned DstAlign, unsigned SrcAlign,
9996 unsigned AlignCheck) {
9997 return ((SrcAlign == 0 || SrcAlign % AlignCheck == 0) &&
9998 (DstAlign == 0 || DstAlign % AlignCheck == 0));
10001 EVT ARMTargetLowering::getOptimalMemOpType(uint64_t Size,
10002 unsigned DstAlign, unsigned SrcAlign,
10003 bool IsMemset, bool ZeroMemset,
10005 MachineFunction &MF) const {
10006 const Function *F = MF.getFunction();
10008 // See if we can use NEON instructions for this...
10009 if ((!IsMemset || ZeroMemset) && Subtarget->hasNEON() &&
10010 !F->hasFnAttribute(Attribute::NoImplicitFloat)) {
10013 (memOpAlign(SrcAlign, DstAlign, 16) ||
10014 (allowsMisalignedMemoryAccesses(MVT::v2f64, 0, 1, &Fast) && Fast))) {
10016 } else if (Size >= 8 &&
10017 (memOpAlign(SrcAlign, DstAlign, 8) ||
10018 (allowsMisalignedMemoryAccesses(MVT::f64, 0, 1, &Fast) &&
10024 // Lowering to i32/i16 if the size permits.
10027 else if (Size >= 2)
10030 // Let the target-independent logic figure it out.
10034 bool ARMTargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
10035 if (Val.getOpcode() != ISD::LOAD)
10038 EVT VT1 = Val.getValueType();
10039 if (!VT1.isSimple() || !VT1.isInteger() ||
10040 !VT2.isSimple() || !VT2.isInteger())
10043 switch (VT1.getSimpleVT().SimpleTy) {
10048 // 8-bit and 16-bit loads implicitly zero-extend to 32-bits.
10055 bool ARMTargetLowering::isVectorLoadExtDesirable(SDValue ExtVal) const {
10056 EVT VT = ExtVal.getValueType();
10058 if (!isTypeLegal(VT))
10061 // Don't create a loadext if we can fold the extension into a wide/long
10063 // If there's more than one user instruction, the loadext is desirable no
10064 // matter what. There can be two uses by the same instruction.
10065 if (ExtVal->use_empty() ||
10066 !ExtVal->use_begin()->isOnlyUserOf(ExtVal.getNode()))
10069 SDNode *U = *ExtVal->use_begin();
10070 if ((U->getOpcode() == ISD::ADD || U->getOpcode() == ISD::SUB ||
10071 U->getOpcode() == ISD::SHL || U->getOpcode() == ARMISD::VSHL))
10077 bool ARMTargetLowering::allowTruncateForTailCall(Type *Ty1, Type *Ty2) const {
10078 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
10081 if (!isTypeLegal(EVT::getEVT(Ty1)))
10084 assert(Ty1->getPrimitiveSizeInBits() <= 64 && "i128 is probably not a noop");
10086 // Assuming the caller doesn't have a zeroext or signext return parameter,
10087 // truncation all the way down to i1 is valid.
10092 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
10096 unsigned Scale = 1;
10097 switch (VT.getSimpleVT().SimpleTy) {
10098 default: return false;
10113 if ((V & (Scale - 1)) != 0)
10116 return V == (V & ((1LL << 5) - 1));
10119 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
10120 const ARMSubtarget *Subtarget) {
10121 bool isNeg = false;
10127 switch (VT.getSimpleVT().SimpleTy) {
10128 default: return false;
10133 // + imm12 or - imm8
10135 return V == (V & ((1LL << 8) - 1));
10136 return V == (V & ((1LL << 12) - 1));
10139 // Same as ARM mode. FIXME: NEON?
10140 if (!Subtarget->hasVFP2())
10145 return V == (V & ((1LL << 8) - 1));
10149 /// isLegalAddressImmediate - Return true if the integer value can be used
10150 /// as the offset of the target addressing mode for load / store of the
10152 static bool isLegalAddressImmediate(int64_t V, EVT VT,
10153 const ARMSubtarget *Subtarget) {
10157 if (!VT.isSimple())
10160 if (Subtarget->isThumb1Only())
10161 return isLegalT1AddressImmediate(V, VT);
10162 else if (Subtarget->isThumb2())
10163 return isLegalT2AddressImmediate(V, VT, Subtarget);
10168 switch (VT.getSimpleVT().SimpleTy) {
10169 default: return false;
10174 return V == (V & ((1LL << 12) - 1));
10177 return V == (V & ((1LL << 8) - 1));
10180 if (!Subtarget->hasVFP2()) // FIXME: NEON?
10185 return V == (V & ((1LL << 8) - 1));
10189 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
10191 int Scale = AM.Scale;
10195 switch (VT.getSimpleVT().SimpleTy) {
10196 default: return false;
10204 Scale = Scale & ~1;
10205 return Scale == 2 || Scale == 4 || Scale == 8;
10208 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
10212 // Note, we allow "void" uses (basically, uses that aren't loads or
10213 // stores), because arm allows folding a scale into many arithmetic
10214 // operations. This should be made more precise and revisited later.
10216 // Allow r << imm, but the imm has to be a multiple of two.
10217 if (Scale & 1) return false;
10218 return isPowerOf2_32(Scale);
10222 /// isLegalAddressingMode - Return true if the addressing mode represented
10223 /// by AM is legal for this target, for a load/store of the specified type.
10224 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
10226 EVT VT = getValueType(Ty, true);
10227 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
10230 // Can never fold addr of global into load/store.
10234 switch (AM.Scale) {
10235 case 0: // no scale reg, must be "r+i" or "r", or "i".
10238 if (Subtarget->isThumb1Only())
10242 // ARM doesn't support any R+R*scale+imm addr modes.
10246 if (!VT.isSimple())
10249 if (Subtarget->isThumb2())
10250 return isLegalT2ScaledAddressingMode(AM, VT);
10252 int Scale = AM.Scale;
10253 switch (VT.getSimpleVT().SimpleTy) {
10254 default: return false;
10258 if (Scale < 0) Scale = -Scale;
10262 return isPowerOf2_32(Scale & ~1);
10266 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
10271 // Note, we allow "void" uses (basically, uses that aren't loads or
10272 // stores), because arm allows folding a scale into many arithmetic
10273 // operations. This should be made more precise and revisited later.
10275 // Allow r << imm, but the imm has to be a multiple of two.
10276 if (Scale & 1) return false;
10277 return isPowerOf2_32(Scale);
10283 /// isLegalICmpImmediate - Return true if the specified immediate is legal
10284 /// icmp immediate, that is the target has icmp instructions which can compare
10285 /// a register against the immediate without having to materialize the
10286 /// immediate into a register.
10287 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
10288 // Thumb2 and ARM modes can use cmn for negative immediates.
10289 if (!Subtarget->isThumb())
10290 return ARM_AM::getSOImmVal(std::abs(Imm)) != -1;
10291 if (Subtarget->isThumb2())
10292 return ARM_AM::getT2SOImmVal(std::abs(Imm)) != -1;
10293 // Thumb1 doesn't have cmn, and only 8-bit immediates.
10294 return Imm >= 0 && Imm <= 255;
10297 /// isLegalAddImmediate - Return true if the specified immediate is a legal add
10298 /// *or sub* immediate, that is the target has add or sub instructions which can
10299 /// add a register with the immediate without having to materialize the
10300 /// immediate into a register.
10301 bool ARMTargetLowering::isLegalAddImmediate(int64_t Imm) const {
10302 // Same encoding for add/sub, just flip the sign.
10303 int64_t AbsImm = std::abs(Imm);
10304 if (!Subtarget->isThumb())
10305 return ARM_AM::getSOImmVal(AbsImm) != -1;
10306 if (Subtarget->isThumb2())
10307 return ARM_AM::getT2SOImmVal(AbsImm) != -1;
10308 // Thumb1 only has 8-bit unsigned immediate.
10309 return AbsImm >= 0 && AbsImm <= 255;
10312 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
10313 bool isSEXTLoad, SDValue &Base,
10314 SDValue &Offset, bool &isInc,
10315 SelectionDAG &DAG) {
10316 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
10319 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
10320 // AddressingMode 3
10321 Base = Ptr->getOperand(0);
10322 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
10323 int RHSC = (int)RHS->getZExtValue();
10324 if (RHSC < 0 && RHSC > -256) {
10325 assert(Ptr->getOpcode() == ISD::ADD);
10327 Offset = DAG.getConstant(-RHSC, SDLoc(Ptr), RHS->getValueType(0));
10331 isInc = (Ptr->getOpcode() == ISD::ADD);
10332 Offset = Ptr->getOperand(1);
10334 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
10335 // AddressingMode 2
10336 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
10337 int RHSC = (int)RHS->getZExtValue();
10338 if (RHSC < 0 && RHSC > -0x1000) {
10339 assert(Ptr->getOpcode() == ISD::ADD);
10341 Offset = DAG.getConstant(-RHSC, SDLoc(Ptr), RHS->getValueType(0));
10342 Base = Ptr->getOperand(0);
10347 if (Ptr->getOpcode() == ISD::ADD) {
10349 ARM_AM::ShiftOpc ShOpcVal=
10350 ARM_AM::getShiftOpcForNode(Ptr->getOperand(0).getOpcode());
10351 if (ShOpcVal != ARM_AM::no_shift) {
10352 Base = Ptr->getOperand(1);
10353 Offset = Ptr->getOperand(0);
10355 Base = Ptr->getOperand(0);
10356 Offset = Ptr->getOperand(1);
10361 isInc = (Ptr->getOpcode() == ISD::ADD);
10362 Base = Ptr->getOperand(0);
10363 Offset = Ptr->getOperand(1);
10367 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
10371 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
10372 bool isSEXTLoad, SDValue &Base,
10373 SDValue &Offset, bool &isInc,
10374 SelectionDAG &DAG) {
10375 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
10378 Base = Ptr->getOperand(0);
10379 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
10380 int RHSC = (int)RHS->getZExtValue();
10381 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
10382 assert(Ptr->getOpcode() == ISD::ADD);
10384 Offset = DAG.getConstant(-RHSC, SDLoc(Ptr), RHS->getValueType(0));
10386 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
10387 isInc = Ptr->getOpcode() == ISD::ADD;
10388 Offset = DAG.getConstant(RHSC, SDLoc(Ptr), RHS->getValueType(0));
10396 /// getPreIndexedAddressParts - returns true by value, base pointer and
10397 /// offset pointer and addressing mode by reference if the node's address
10398 /// can be legally represented as pre-indexed load / store address.
10400 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
10402 ISD::MemIndexedMode &AM,
10403 SelectionDAG &DAG) const {
10404 if (Subtarget->isThumb1Only())
10409 bool isSEXTLoad = false;
10410 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
10411 Ptr = LD->getBasePtr();
10412 VT = LD->getMemoryVT();
10413 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
10414 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
10415 Ptr = ST->getBasePtr();
10416 VT = ST->getMemoryVT();
10421 bool isLegal = false;
10422 if (Subtarget->isThumb2())
10423 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
10424 Offset, isInc, DAG);
10426 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
10427 Offset, isInc, DAG);
10431 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
10435 /// getPostIndexedAddressParts - returns true by value, base pointer and
10436 /// offset pointer and addressing mode by reference if this node can be
10437 /// combined with a load / store to form a post-indexed load / store.
10438 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
10441 ISD::MemIndexedMode &AM,
10442 SelectionDAG &DAG) const {
10443 if (Subtarget->isThumb1Only())
10448 bool isSEXTLoad = false;
10449 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
10450 VT = LD->getMemoryVT();
10451 Ptr = LD->getBasePtr();
10452 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
10453 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
10454 VT = ST->getMemoryVT();
10455 Ptr = ST->getBasePtr();
10460 bool isLegal = false;
10461 if (Subtarget->isThumb2())
10462 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
10465 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
10471 // Swap base ptr and offset to catch more post-index load / store when
10472 // it's legal. In Thumb2 mode, offset must be an immediate.
10473 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
10474 !Subtarget->isThumb2())
10475 std::swap(Base, Offset);
10477 // Post-indexed load / store update the base pointer.
10482 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
10486 void ARMTargetLowering::computeKnownBitsForTargetNode(const SDValue Op,
10489 const SelectionDAG &DAG,
10490 unsigned Depth) const {
10491 unsigned BitWidth = KnownOne.getBitWidth();
10492 KnownZero = KnownOne = APInt(BitWidth, 0);
10493 switch (Op.getOpcode()) {
10499 // These nodes' second result is a boolean
10500 if (Op.getResNo() == 0)
10502 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - 1);
10504 case ARMISD::CMOV: {
10505 // Bits are known zero/one if known on the LHS and RHS.
10506 DAG.computeKnownBits(Op.getOperand(0), KnownZero, KnownOne, Depth+1);
10507 if (KnownZero == 0 && KnownOne == 0) return;
10509 APInt KnownZeroRHS, KnownOneRHS;
10510 DAG.computeKnownBits(Op.getOperand(1), KnownZeroRHS, KnownOneRHS, Depth+1);
10511 KnownZero &= KnownZeroRHS;
10512 KnownOne &= KnownOneRHS;
10515 case ISD::INTRINSIC_W_CHAIN: {
10516 ConstantSDNode *CN = cast<ConstantSDNode>(Op->getOperand(1));
10517 Intrinsic::ID IntID = static_cast<Intrinsic::ID>(CN->getZExtValue());
10520 case Intrinsic::arm_ldaex:
10521 case Intrinsic::arm_ldrex: {
10522 EVT VT = cast<MemIntrinsicSDNode>(Op)->getMemoryVT();
10523 unsigned MemBits = VT.getScalarType().getSizeInBits();
10524 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - MemBits);
10532 //===----------------------------------------------------------------------===//
10533 // ARM Inline Assembly Support
10534 //===----------------------------------------------------------------------===//
10536 bool ARMTargetLowering::ExpandInlineAsm(CallInst *CI) const {
10537 // Looking for "rev" which is V6+.
10538 if (!Subtarget->hasV6Ops())
10541 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
10542 std::string AsmStr = IA->getAsmString();
10543 SmallVector<StringRef, 4> AsmPieces;
10544 SplitString(AsmStr, AsmPieces, ";\n");
10546 switch (AsmPieces.size()) {
10547 default: return false;
10549 AsmStr = AsmPieces[0];
10551 SplitString(AsmStr, AsmPieces, " \t,");
10554 if (AsmPieces.size() == 3 &&
10555 AsmPieces[0] == "rev" && AsmPieces[1] == "$0" && AsmPieces[2] == "$1" &&
10556 IA->getConstraintString().compare(0, 4, "=l,l") == 0) {
10557 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
10558 if (Ty && Ty->getBitWidth() == 32)
10559 return IntrinsicLowering::LowerToByteSwap(CI);
10567 /// getConstraintType - Given a constraint letter, return the type of
10568 /// constraint it is for this target.
10569 ARMTargetLowering::ConstraintType
10570 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
10571 if (Constraint.size() == 1) {
10572 switch (Constraint[0]) {
10574 case 'l': return C_RegisterClass;
10575 case 'w': return C_RegisterClass;
10576 case 'h': return C_RegisterClass;
10577 case 'x': return C_RegisterClass;
10578 case 't': return C_RegisterClass;
10579 case 'j': return C_Other; // Constant for movw.
10580 // An address with a single base register. Due to the way we
10581 // currently handle addresses it is the same as an 'r' memory constraint.
10582 case 'Q': return C_Memory;
10584 } else if (Constraint.size() == 2) {
10585 switch (Constraint[0]) {
10587 // All 'U+' constraints are addresses.
10588 case 'U': return C_Memory;
10591 return TargetLowering::getConstraintType(Constraint);
10594 /// Examine constraint type and operand type and determine a weight value.
10595 /// This object must already have been set up with the operand type
10596 /// and the current alternative constraint selected.
10597 TargetLowering::ConstraintWeight
10598 ARMTargetLowering::getSingleConstraintMatchWeight(
10599 AsmOperandInfo &info, const char *constraint) const {
10600 ConstraintWeight weight = CW_Invalid;
10601 Value *CallOperandVal = info.CallOperandVal;
10602 // If we don't have a value, we can't do a match,
10603 // but allow it at the lowest weight.
10604 if (!CallOperandVal)
10606 Type *type = CallOperandVal->getType();
10607 // Look at the constraint type.
10608 switch (*constraint) {
10610 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
10613 if (type->isIntegerTy()) {
10614 if (Subtarget->isThumb())
10615 weight = CW_SpecificReg;
10617 weight = CW_Register;
10621 if (type->isFloatingPointTy())
10622 weight = CW_Register;
10628 typedef std::pair<unsigned, const TargetRegisterClass*> RCPair;
10630 ARMTargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
10631 const std::string &Constraint,
10633 if (Constraint.size() == 1) {
10634 // GCC ARM Constraint Letters
10635 switch (Constraint[0]) {
10636 case 'l': // Low regs or general regs.
10637 if (Subtarget->isThumb())
10638 return RCPair(0U, &ARM::tGPRRegClass);
10639 return RCPair(0U, &ARM::GPRRegClass);
10640 case 'h': // High regs or no regs.
10641 if (Subtarget->isThumb())
10642 return RCPair(0U, &ARM::hGPRRegClass);
10645 if (Subtarget->isThumb1Only())
10646 return RCPair(0U, &ARM::tGPRRegClass);
10647 return RCPair(0U, &ARM::GPRRegClass);
10649 if (VT == MVT::Other)
10651 if (VT == MVT::f32)
10652 return RCPair(0U, &ARM::SPRRegClass);
10653 if (VT.getSizeInBits() == 64)
10654 return RCPair(0U, &ARM::DPRRegClass);
10655 if (VT.getSizeInBits() == 128)
10656 return RCPair(0U, &ARM::QPRRegClass);
10659 if (VT == MVT::Other)
10661 if (VT == MVT::f32)
10662 return RCPair(0U, &ARM::SPR_8RegClass);
10663 if (VT.getSizeInBits() == 64)
10664 return RCPair(0U, &ARM::DPR_8RegClass);
10665 if (VT.getSizeInBits() == 128)
10666 return RCPair(0U, &ARM::QPR_8RegClass);
10669 if (VT == MVT::f32)
10670 return RCPair(0U, &ARM::SPRRegClass);
10674 if (StringRef("{cc}").equals_lower(Constraint))
10675 return std::make_pair(unsigned(ARM::CPSR), &ARM::CCRRegClass);
10677 return TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
10680 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
10681 /// vector. If it is invalid, don't add anything to Ops.
10682 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
10683 std::string &Constraint,
10684 std::vector<SDValue>&Ops,
10685 SelectionDAG &DAG) const {
10688 // Currently only support length 1 constraints.
10689 if (Constraint.length() != 1) return;
10691 char ConstraintLetter = Constraint[0];
10692 switch (ConstraintLetter) {
10695 case 'I': case 'J': case 'K': case 'L':
10696 case 'M': case 'N': case 'O':
10697 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
10701 int64_t CVal64 = C->getSExtValue();
10702 int CVal = (int) CVal64;
10703 // None of these constraints allow values larger than 32 bits. Check
10704 // that the value fits in an int.
10705 if (CVal != CVal64)
10708 switch (ConstraintLetter) {
10710 // Constant suitable for movw, must be between 0 and
10712 if (Subtarget->hasV6T2Ops())
10713 if (CVal >= 0 && CVal <= 65535)
10717 if (Subtarget->isThumb1Only()) {
10718 // This must be a constant between 0 and 255, for ADD
10720 if (CVal >= 0 && CVal <= 255)
10722 } else if (Subtarget->isThumb2()) {
10723 // A constant that can be used as an immediate value in a
10724 // data-processing instruction.
10725 if (ARM_AM::getT2SOImmVal(CVal) != -1)
10728 // A constant that can be used as an immediate value in a
10729 // data-processing instruction.
10730 if (ARM_AM::getSOImmVal(CVal) != -1)
10736 if (Subtarget->isThumb()) { // FIXME thumb2
10737 // This must be a constant between -255 and -1, for negated ADD
10738 // immediates. This can be used in GCC with an "n" modifier that
10739 // prints the negated value, for use with SUB instructions. It is
10740 // not useful otherwise but is implemented for compatibility.
10741 if (CVal >= -255 && CVal <= -1)
10744 // This must be a constant between -4095 and 4095. It is not clear
10745 // what this constraint is intended for. Implemented for
10746 // compatibility with GCC.
10747 if (CVal >= -4095 && CVal <= 4095)
10753 if (Subtarget->isThumb1Only()) {
10754 // A 32-bit value where only one byte has a nonzero value. Exclude
10755 // zero to match GCC. This constraint is used by GCC internally for
10756 // constants that can be loaded with a move/shift combination.
10757 // It is not useful otherwise but is implemented for compatibility.
10758 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
10760 } else if (Subtarget->isThumb2()) {
10761 // A constant whose bitwise inverse can be used as an immediate
10762 // value in a data-processing instruction. This can be used in GCC
10763 // with a "B" modifier that prints the inverted value, for use with
10764 // BIC and MVN instructions. It is not useful otherwise but is
10765 // implemented for compatibility.
10766 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
10769 // A constant whose bitwise inverse can be used as an immediate
10770 // value in a data-processing instruction. This can be used in GCC
10771 // with a "B" modifier that prints the inverted value, for use with
10772 // BIC and MVN instructions. It is not useful otherwise but is
10773 // implemented for compatibility.
10774 if (ARM_AM::getSOImmVal(~CVal) != -1)
10780 if (Subtarget->isThumb1Only()) {
10781 // This must be a constant between -7 and 7,
10782 // for 3-operand ADD/SUB immediate instructions.
10783 if (CVal >= -7 && CVal < 7)
10785 } else if (Subtarget->isThumb2()) {
10786 // A constant whose negation can be used as an immediate value in a
10787 // data-processing instruction. This can be used in GCC with an "n"
10788 // modifier that prints the negated value, for use with SUB
10789 // instructions. It is not useful otherwise but is implemented for
10791 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
10794 // A constant whose negation can be used as an immediate value in a
10795 // data-processing instruction. This can be used in GCC with an "n"
10796 // modifier that prints the negated value, for use with SUB
10797 // instructions. It is not useful otherwise but is implemented for
10799 if (ARM_AM::getSOImmVal(-CVal) != -1)
10805 if (Subtarget->isThumb()) { // FIXME thumb2
10806 // This must be a multiple of 4 between 0 and 1020, for
10807 // ADD sp + immediate.
10808 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
10811 // A power of two or a constant between 0 and 32. This is used in
10812 // GCC for the shift amount on shifted register operands, but it is
10813 // useful in general for any shift amounts.
10814 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
10820 if (Subtarget->isThumb()) { // FIXME thumb2
10821 // This must be a constant between 0 and 31, for shift amounts.
10822 if (CVal >= 0 && CVal <= 31)
10828 if (Subtarget->isThumb()) { // FIXME thumb2
10829 // This must be a multiple of 4 between -508 and 508, for
10830 // ADD/SUB sp = sp + immediate.
10831 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
10836 Result = DAG.getTargetConstant(CVal, SDLoc(Op), Op.getValueType());
10840 if (Result.getNode()) {
10841 Ops.push_back(Result);
10844 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
10847 SDValue ARMTargetLowering::LowerDivRem(SDValue Op, SelectionDAG &DAG) const {
10848 assert(Subtarget->isTargetAEABI() && "Register-based DivRem lowering only");
10849 unsigned Opcode = Op->getOpcode();
10850 assert((Opcode == ISD::SDIVREM || Opcode == ISD::UDIVREM) &&
10851 "Invalid opcode for Div/Rem lowering");
10852 bool isSigned = (Opcode == ISD::SDIVREM);
10853 EVT VT = Op->getValueType(0);
10854 Type *Ty = VT.getTypeForEVT(*DAG.getContext());
10857 switch (VT.getSimpleVT().SimpleTy) {
10858 default: llvm_unreachable("Unexpected request for libcall!");
10859 case MVT::i8: LC = isSigned ? RTLIB::SDIVREM_I8 : RTLIB::UDIVREM_I8; break;
10860 case MVT::i16: LC = isSigned ? RTLIB::SDIVREM_I16 : RTLIB::UDIVREM_I16; break;
10861 case MVT::i32: LC = isSigned ? RTLIB::SDIVREM_I32 : RTLIB::UDIVREM_I32; break;
10862 case MVT::i64: LC = isSigned ? RTLIB::SDIVREM_I64 : RTLIB::UDIVREM_I64; break;
10865 SDValue InChain = DAG.getEntryNode();
10867 TargetLowering::ArgListTy Args;
10868 TargetLowering::ArgListEntry Entry;
10869 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i) {
10870 EVT ArgVT = Op->getOperand(i).getValueType();
10871 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
10872 Entry.Node = Op->getOperand(i);
10874 Entry.isSExt = isSigned;
10875 Entry.isZExt = !isSigned;
10876 Args.push_back(Entry);
10879 SDValue Callee = DAG.getExternalSymbol(getLibcallName(LC),
10882 Type *RetTy = (Type*)StructType::get(Ty, Ty, nullptr);
10885 TargetLowering::CallLoweringInfo CLI(DAG);
10886 CLI.setDebugLoc(dl).setChain(InChain)
10887 .setCallee(getLibcallCallingConv(LC), RetTy, Callee, std::move(Args), 0)
10888 .setInRegister().setSExtResult(isSigned).setZExtResult(!isSigned);
10890 std::pair<SDValue, SDValue> CallInfo = LowerCallTo(CLI);
10891 return CallInfo.first;
10895 ARMTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const {
10896 assert(Subtarget->isTargetWindows() && "unsupported target platform");
10900 SDValue Chain = Op.getOperand(0);
10901 SDValue Size = Op.getOperand(1);
10903 SDValue Words = DAG.getNode(ISD::SRL, DL, MVT::i32, Size,
10904 DAG.getConstant(2, DL, MVT::i32));
10907 Chain = DAG.getCopyToReg(Chain, DL, ARM::R4, Words, Flag);
10908 Flag = Chain.getValue(1);
10910 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
10911 Chain = DAG.getNode(ARMISD::WIN__CHKSTK, DL, NodeTys, Chain, Flag);
10913 SDValue NewSP = DAG.getCopyFromReg(Chain, DL, ARM::SP, MVT::i32);
10914 Chain = NewSP.getValue(1);
10916 SDValue Ops[2] = { NewSP, Chain };
10917 return DAG.getMergeValues(Ops, DL);
10920 SDValue ARMTargetLowering::LowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) const {
10921 assert(Op.getValueType() == MVT::f64 && Subtarget->isFPOnlySP() &&
10922 "Unexpected type for custom-lowering FP_EXTEND");
10925 LC = RTLIB::getFPEXT(Op.getOperand(0).getValueType(), Op.getValueType());
10927 SDValue SrcVal = Op.getOperand(0);
10928 return makeLibCall(DAG, LC, Op.getValueType(), &SrcVal, 1,
10929 /*isSigned*/ false, SDLoc(Op)).first;
10932 SDValue ARMTargetLowering::LowerFP_ROUND(SDValue Op, SelectionDAG &DAG) const {
10933 assert(Op.getOperand(0).getValueType() == MVT::f64 &&
10934 Subtarget->isFPOnlySP() &&
10935 "Unexpected type for custom-lowering FP_ROUND");
10938 LC = RTLIB::getFPROUND(Op.getOperand(0).getValueType(), Op.getValueType());
10940 SDValue SrcVal = Op.getOperand(0);
10941 return makeLibCall(DAG, LC, Op.getValueType(), &SrcVal, 1,
10942 /*isSigned*/ false, SDLoc(Op)).first;
10946 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
10947 // The ARM target isn't yet aware of offsets.
10951 bool ARM::isBitFieldInvertedMask(unsigned v) {
10952 if (v == 0xffffffff)
10955 // there can be 1's on either or both "outsides", all the "inside"
10956 // bits must be 0's
10957 return isShiftedMask_32(~v);
10960 /// isFPImmLegal - Returns true if the target can instruction select the
10961 /// specified FP immediate natively. If false, the legalizer will
10962 /// materialize the FP immediate as a load from a constant pool.
10963 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
10964 if (!Subtarget->hasVFP3())
10966 if (VT == MVT::f32)
10967 return ARM_AM::getFP32Imm(Imm) != -1;
10968 if (VT == MVT::f64 && !Subtarget->isFPOnlySP())
10969 return ARM_AM::getFP64Imm(Imm) != -1;
10973 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
10974 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
10975 /// specified in the intrinsic calls.
10976 bool ARMTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
10978 unsigned Intrinsic) const {
10979 switch (Intrinsic) {
10980 case Intrinsic::arm_neon_vld1:
10981 case Intrinsic::arm_neon_vld2:
10982 case Intrinsic::arm_neon_vld3:
10983 case Intrinsic::arm_neon_vld4:
10984 case Intrinsic::arm_neon_vld2lane:
10985 case Intrinsic::arm_neon_vld3lane:
10986 case Intrinsic::arm_neon_vld4lane: {
10987 Info.opc = ISD::INTRINSIC_W_CHAIN;
10988 // Conservatively set memVT to the entire set of vectors loaded.
10989 uint64_t NumElts = getDataLayout()->getTypeAllocSize(I.getType()) / 8;
10990 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
10991 Info.ptrVal = I.getArgOperand(0);
10993 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
10994 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
10995 Info.vol = false; // volatile loads with NEON intrinsics not supported
10996 Info.readMem = true;
10997 Info.writeMem = false;
11000 case Intrinsic::arm_neon_vst1:
11001 case Intrinsic::arm_neon_vst2:
11002 case Intrinsic::arm_neon_vst3:
11003 case Intrinsic::arm_neon_vst4:
11004 case Intrinsic::arm_neon_vst2lane:
11005 case Intrinsic::arm_neon_vst3lane:
11006 case Intrinsic::arm_neon_vst4lane: {
11007 Info.opc = ISD::INTRINSIC_VOID;
11008 // Conservatively set memVT to the entire set of vectors stored.
11009 unsigned NumElts = 0;
11010 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
11011 Type *ArgTy = I.getArgOperand(ArgI)->getType();
11012 if (!ArgTy->isVectorTy())
11014 NumElts += getDataLayout()->getTypeAllocSize(ArgTy) / 8;
11016 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
11017 Info.ptrVal = I.getArgOperand(0);
11019 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
11020 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
11021 Info.vol = false; // volatile stores with NEON intrinsics not supported
11022 Info.readMem = false;
11023 Info.writeMem = true;
11026 case Intrinsic::arm_ldaex:
11027 case Intrinsic::arm_ldrex: {
11028 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(0)->getType());
11029 Info.opc = ISD::INTRINSIC_W_CHAIN;
11030 Info.memVT = MVT::getVT(PtrTy->getElementType());
11031 Info.ptrVal = I.getArgOperand(0);
11033 Info.align = getDataLayout()->getABITypeAlignment(PtrTy->getElementType());
11035 Info.readMem = true;
11036 Info.writeMem = false;
11039 case Intrinsic::arm_stlex:
11040 case Intrinsic::arm_strex: {
11041 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(1)->getType());
11042 Info.opc = ISD::INTRINSIC_W_CHAIN;
11043 Info.memVT = MVT::getVT(PtrTy->getElementType());
11044 Info.ptrVal = I.getArgOperand(1);
11046 Info.align = getDataLayout()->getABITypeAlignment(PtrTy->getElementType());
11048 Info.readMem = false;
11049 Info.writeMem = true;
11052 case Intrinsic::arm_stlexd:
11053 case Intrinsic::arm_strexd: {
11054 Info.opc = ISD::INTRINSIC_W_CHAIN;
11055 Info.memVT = MVT::i64;
11056 Info.ptrVal = I.getArgOperand(2);
11060 Info.readMem = false;
11061 Info.writeMem = true;
11064 case Intrinsic::arm_ldaexd:
11065 case Intrinsic::arm_ldrexd: {
11066 Info.opc = ISD::INTRINSIC_W_CHAIN;
11067 Info.memVT = MVT::i64;
11068 Info.ptrVal = I.getArgOperand(0);
11072 Info.readMem = true;
11073 Info.writeMem = false;
11083 /// \brief Returns true if it is beneficial to convert a load of a constant
11084 /// to just the constant itself.
11085 bool ARMTargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
11087 assert(Ty->isIntegerTy());
11089 unsigned Bits = Ty->getPrimitiveSizeInBits();
11090 if (Bits == 0 || Bits > 32)
11095 bool ARMTargetLowering::hasLoadLinkedStoreConditional() const { return true; }
11097 Instruction* ARMTargetLowering::makeDMB(IRBuilder<> &Builder,
11098 ARM_MB::MemBOpt Domain) const {
11099 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
11101 // First, if the target has no DMB, see what fallback we can use.
11102 if (!Subtarget->hasDataBarrier()) {
11103 // Some ARMv6 cpus can support data barriers with an mcr instruction.
11104 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
11106 if (Subtarget->hasV6Ops() && !Subtarget->isThumb()) {
11107 Function *MCR = llvm::Intrinsic::getDeclaration(M, Intrinsic::arm_mcr);
11108 Value* args[6] = {Builder.getInt32(15), Builder.getInt32(0),
11109 Builder.getInt32(0), Builder.getInt32(7),
11110 Builder.getInt32(10), Builder.getInt32(5)};
11111 return Builder.CreateCall(MCR, args);
11113 // Instead of using barriers, atomic accesses on these subtargets use
11115 llvm_unreachable("makeDMB on a target so old that it has no barriers");
11118 Function *DMB = llvm::Intrinsic::getDeclaration(M, Intrinsic::arm_dmb);
11119 // Only a full system barrier exists in the M-class architectures.
11120 Domain = Subtarget->isMClass() ? ARM_MB::SY : Domain;
11121 Constant *CDomain = Builder.getInt32(Domain);
11122 return Builder.CreateCall(DMB, CDomain);
11126 // Based on http://www.cl.cam.ac.uk/~pes20/cpp/cpp0xmappings.html
11127 Instruction* ARMTargetLowering::emitLeadingFence(IRBuilder<> &Builder,
11128 AtomicOrdering Ord, bool IsStore,
11129 bool IsLoad) const {
11130 if (!getInsertFencesForAtomic())
11136 llvm_unreachable("Invalid fence: unordered/non-atomic");
11139 return nullptr; // Nothing to do
11140 case SequentiallyConsistent:
11142 return nullptr; // Nothing to do
11145 case AcquireRelease:
11146 if (Subtarget->isSwift())
11147 return makeDMB(Builder, ARM_MB::ISHST);
11148 // FIXME: add a comment with a link to documentation justifying this.
11150 return makeDMB(Builder, ARM_MB::ISH);
11152 llvm_unreachable("Unknown fence ordering in emitLeadingFence");
11155 Instruction* ARMTargetLowering::emitTrailingFence(IRBuilder<> &Builder,
11156 AtomicOrdering Ord, bool IsStore,
11157 bool IsLoad) const {
11158 if (!getInsertFencesForAtomic())
11164 llvm_unreachable("Invalid fence: unordered/not-atomic");
11167 return nullptr; // Nothing to do
11169 case AcquireRelease:
11170 case SequentiallyConsistent:
11171 return makeDMB(Builder, ARM_MB::ISH);
11173 llvm_unreachable("Unknown fence ordering in emitTrailingFence");
11176 // Loads and stores less than 64-bits are already atomic; ones above that
11177 // are doomed anyway, so defer to the default libcall and blame the OS when
11178 // things go wrong. Cortex M doesn't have ldrexd/strexd though, so don't emit
11179 // anything for those.
11180 bool ARMTargetLowering::shouldExpandAtomicStoreInIR(StoreInst *SI) const {
11181 unsigned Size = SI->getValueOperand()->getType()->getPrimitiveSizeInBits();
11182 return (Size == 64) && !Subtarget->isMClass();
11185 // Loads and stores less than 64-bits are already atomic; ones above that
11186 // are doomed anyway, so defer to the default libcall and blame the OS when
11187 // things go wrong. Cortex M doesn't have ldrexd/strexd though, so don't emit
11188 // anything for those.
11189 // FIXME: ldrd and strd are atomic if the CPU has LPAE (e.g. A15 has that
11190 // guarantee, see DDI0406C ARM architecture reference manual,
11191 // sections A8.8.72-74 LDRD)
11192 bool ARMTargetLowering::shouldExpandAtomicLoadInIR(LoadInst *LI) const {
11193 unsigned Size = LI->getType()->getPrimitiveSizeInBits();
11194 return (Size == 64) && !Subtarget->isMClass();
11197 // For the real atomic operations, we have ldrex/strex up to 32 bits,
11198 // and up to 64 bits on the non-M profiles
11199 TargetLoweringBase::AtomicRMWExpansionKind
11200 ARMTargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
11201 unsigned Size = AI->getType()->getPrimitiveSizeInBits();
11202 return (Size <= (Subtarget->isMClass() ? 32U : 64U))
11203 ? AtomicRMWExpansionKind::LLSC
11204 : AtomicRMWExpansionKind::None;
11207 // This has so far only been implemented for MachO.
11208 bool ARMTargetLowering::useLoadStackGuardNode() const {
11209 return Subtarget->isTargetMachO();
11212 bool ARMTargetLowering::canCombineStoreAndExtract(Type *VectorTy, Value *Idx,
11213 unsigned &Cost) const {
11214 // If we do not have NEON, vector types are not natively supported.
11215 if (!Subtarget->hasNEON())
11218 // Floating point values and vector values map to the same register file.
11219 // Therefore, althought we could do a store extract of a vector type, this is
11220 // better to leave at float as we have more freedom in the addressing mode for
11222 if (VectorTy->isFPOrFPVectorTy())
11225 // If the index is unknown at compile time, this is very expensive to lower
11226 // and it is not possible to combine the store with the extract.
11227 if (!isa<ConstantInt>(Idx))
11230 assert(VectorTy->isVectorTy() && "VectorTy is not a vector type");
11231 unsigned BitWidth = cast<VectorType>(VectorTy)->getBitWidth();
11232 // We can do a store + vector extract on any vector that fits perfectly in a D
11234 if (BitWidth == 64 || BitWidth == 128) {
11241 Value *ARMTargetLowering::emitLoadLinked(IRBuilder<> &Builder, Value *Addr,
11242 AtomicOrdering Ord) const {
11243 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
11244 Type *ValTy = cast<PointerType>(Addr->getType())->getElementType();
11245 bool IsAcquire = isAtLeastAcquire(Ord);
11247 // Since i64 isn't legal and intrinsics don't get type-lowered, the ldrexd
11248 // intrinsic must return {i32, i32} and we have to recombine them into a
11249 // single i64 here.
11250 if (ValTy->getPrimitiveSizeInBits() == 64) {
11251 Intrinsic::ID Int =
11252 IsAcquire ? Intrinsic::arm_ldaexd : Intrinsic::arm_ldrexd;
11253 Function *Ldrex = llvm::Intrinsic::getDeclaration(M, Int);
11255 Addr = Builder.CreateBitCast(Addr, Type::getInt8PtrTy(M->getContext()));
11256 Value *LoHi = Builder.CreateCall(Ldrex, Addr, "lohi");
11258 Value *Lo = Builder.CreateExtractValue(LoHi, 0, "lo");
11259 Value *Hi = Builder.CreateExtractValue(LoHi, 1, "hi");
11260 if (!Subtarget->isLittle())
11261 std::swap (Lo, Hi);
11262 Lo = Builder.CreateZExt(Lo, ValTy, "lo64");
11263 Hi = Builder.CreateZExt(Hi, ValTy, "hi64");
11264 return Builder.CreateOr(
11265 Lo, Builder.CreateShl(Hi, ConstantInt::get(ValTy, 32)), "val64");
11268 Type *Tys[] = { Addr->getType() };
11269 Intrinsic::ID Int = IsAcquire ? Intrinsic::arm_ldaex : Intrinsic::arm_ldrex;
11270 Function *Ldrex = llvm::Intrinsic::getDeclaration(M, Int, Tys);
11272 return Builder.CreateTruncOrBitCast(
11273 Builder.CreateCall(Ldrex, Addr),
11274 cast<PointerType>(Addr->getType())->getElementType());
11277 Value *ARMTargetLowering::emitStoreConditional(IRBuilder<> &Builder, Value *Val,
11279 AtomicOrdering Ord) const {
11280 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
11281 bool IsRelease = isAtLeastRelease(Ord);
11283 // Since the intrinsics must have legal type, the i64 intrinsics take two
11284 // parameters: "i32, i32". We must marshal Val into the appropriate form
11285 // before the call.
11286 if (Val->getType()->getPrimitiveSizeInBits() == 64) {
11287 Intrinsic::ID Int =
11288 IsRelease ? Intrinsic::arm_stlexd : Intrinsic::arm_strexd;
11289 Function *Strex = Intrinsic::getDeclaration(M, Int);
11290 Type *Int32Ty = Type::getInt32Ty(M->getContext());
11292 Value *Lo = Builder.CreateTrunc(Val, Int32Ty, "lo");
11293 Value *Hi = Builder.CreateTrunc(Builder.CreateLShr(Val, 32), Int32Ty, "hi");
11294 if (!Subtarget->isLittle())
11295 std::swap (Lo, Hi);
11296 Addr = Builder.CreateBitCast(Addr, Type::getInt8PtrTy(M->getContext()));
11297 return Builder.CreateCall(Strex, {Lo, Hi, Addr});
11300 Intrinsic::ID Int = IsRelease ? Intrinsic::arm_stlex : Intrinsic::arm_strex;
11301 Type *Tys[] = { Addr->getType() };
11302 Function *Strex = Intrinsic::getDeclaration(M, Int, Tys);
11304 return Builder.CreateCall(
11305 Strex, {Builder.CreateZExtOrBitCast(
11306 Val, Strex->getFunctionType()->getParamType(0)),
11318 static bool isHomogeneousAggregate(Type *Ty, HABaseType &Base,
11319 uint64_t &Members) {
11320 if (const StructType *ST = dyn_cast<StructType>(Ty)) {
11321 for (unsigned i = 0; i < ST->getNumElements(); ++i) {
11322 uint64_t SubMembers = 0;
11323 if (!isHomogeneousAggregate(ST->getElementType(i), Base, SubMembers))
11325 Members += SubMembers;
11327 } else if (const ArrayType *AT = dyn_cast<ArrayType>(Ty)) {
11328 uint64_t SubMembers = 0;
11329 if (!isHomogeneousAggregate(AT->getElementType(), Base, SubMembers))
11331 Members += SubMembers * AT->getNumElements();
11332 } else if (Ty->isFloatTy()) {
11333 if (Base != HA_UNKNOWN && Base != HA_FLOAT)
11337 } else if (Ty->isDoubleTy()) {
11338 if (Base != HA_UNKNOWN && Base != HA_DOUBLE)
11342 } else if (const VectorType *VT = dyn_cast<VectorType>(Ty)) {
11349 return VT->getBitWidth() == 64;
11351 return VT->getBitWidth() == 128;
11353 switch (VT->getBitWidth()) {
11366 return (Members > 0 && Members <= 4);
11369 /// \brief Return true if a type is an AAPCS-VFP homogeneous aggregate or one of
11370 /// [N x i32] or [N x i64]. This allows front-ends to skip emitting padding when
11371 /// passing according to AAPCS rules.
11372 bool ARMTargetLowering::functionArgumentNeedsConsecutiveRegisters(
11373 Type *Ty, CallingConv::ID CallConv, bool isVarArg) const {
11374 if (getEffectiveCallingConv(CallConv, isVarArg) !=
11375 CallingConv::ARM_AAPCS_VFP)
11378 HABaseType Base = HA_UNKNOWN;
11379 uint64_t Members = 0;
11380 bool IsHA = isHomogeneousAggregate(Ty, Base, Members);
11381 DEBUG(dbgs() << "isHA: " << IsHA << " "; Ty->dump());
11383 bool IsIntArray = Ty->isArrayTy() && Ty->getArrayElementType()->isIntegerTy();
11384 return IsHA || IsIntArray;