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() && !TM.Options.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 (!TM.Options.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 (!TM.Options.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 (!TM.Options.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 (!TM.Options.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 (TM.Options.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 // FIXME: It might make sense to define the representative register class as the
960 // nearest super-register that has a non-null superset. For example, DPR_VFP2 is
961 // a super-register of SPR, and DPR is a superset if DPR_VFP2. Consequently,
962 // SPR's representative would be DPR_VFP2. This should work well if register
963 // pressure tracking were modified such that a register use would increment the
964 // pressure of the register class's representative and all of it's super
965 // classes' representatives transitively. We have not implemented this because
966 // of the difficulty prior to coalescing of modeling operand register classes
967 // due to the common occurrence of cross class copies and subregister insertions
969 std::pair<const TargetRegisterClass *, uint8_t>
970 ARMTargetLowering::findRepresentativeClass(const TargetRegisterInfo *TRI,
972 const TargetRegisterClass *RRC = nullptr;
974 switch (VT.SimpleTy) {
976 return TargetLowering::findRepresentativeClass(TRI, VT);
977 // Use DPR as representative register class for all floating point
978 // and vector types. Since there are 32 SPR registers and 32 DPR registers so
979 // the cost is 1 for both f32 and f64.
980 case MVT::f32: case MVT::f64: case MVT::v8i8: case MVT::v4i16:
981 case MVT::v2i32: case MVT::v1i64: case MVT::v2f32:
982 RRC = &ARM::DPRRegClass;
983 // When NEON is used for SP, only half of the register file is available
984 // because operations that define both SP and DP results will be constrained
985 // to the VFP2 class (D0-D15). We currently model this constraint prior to
986 // coalescing by double-counting the SP regs. See the FIXME above.
987 if (Subtarget->useNEONForSinglePrecisionFP())
990 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
991 case MVT::v4f32: case MVT::v2f64:
992 RRC = &ARM::DPRRegClass;
996 RRC = &ARM::DPRRegClass;
1000 RRC = &ARM::DPRRegClass;
1004 return std::make_pair(RRC, Cost);
1007 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
1009 default: return nullptr;
1010 case ARMISD::Wrapper: return "ARMISD::Wrapper";
1011 case ARMISD::WrapperPIC: return "ARMISD::WrapperPIC";
1012 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
1013 case ARMISD::CALL: return "ARMISD::CALL";
1014 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
1015 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
1016 case ARMISD::tCALL: return "ARMISD::tCALL";
1017 case ARMISD::BRCOND: return "ARMISD::BRCOND";
1018 case ARMISD::BR_JT: return "ARMISD::BR_JT";
1019 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
1020 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
1021 case ARMISD::INTRET_FLAG: return "ARMISD::INTRET_FLAG";
1022 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
1023 case ARMISD::CMP: return "ARMISD::CMP";
1024 case ARMISD::CMN: return "ARMISD::CMN";
1025 case ARMISD::CMPZ: return "ARMISD::CMPZ";
1026 case ARMISD::CMPFP: return "ARMISD::CMPFP";
1027 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
1028 case ARMISD::BCC_i64: return "ARMISD::BCC_i64";
1029 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
1031 case ARMISD::CMOV: return "ARMISD::CMOV";
1033 case ARMISD::RBIT: return "ARMISD::RBIT";
1035 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
1036 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
1037 case ARMISD::RRX: return "ARMISD::RRX";
1039 case ARMISD::ADDC: return "ARMISD::ADDC";
1040 case ARMISD::ADDE: return "ARMISD::ADDE";
1041 case ARMISD::SUBC: return "ARMISD::SUBC";
1042 case ARMISD::SUBE: return "ARMISD::SUBE";
1044 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
1045 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
1047 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
1048 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
1050 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN";
1052 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
1054 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
1056 case ARMISD::MEMBARRIER_MCR: return "ARMISD::MEMBARRIER_MCR";
1058 case ARMISD::PRELOAD: return "ARMISD::PRELOAD";
1060 case ARMISD::WIN__CHKSTK: return "ARMISD:::WIN__CHKSTK";
1062 case ARMISD::VCEQ: return "ARMISD::VCEQ";
1063 case ARMISD::VCEQZ: return "ARMISD::VCEQZ";
1064 case ARMISD::VCGE: return "ARMISD::VCGE";
1065 case ARMISD::VCGEZ: return "ARMISD::VCGEZ";
1066 case ARMISD::VCLEZ: return "ARMISD::VCLEZ";
1067 case ARMISD::VCGEU: return "ARMISD::VCGEU";
1068 case ARMISD::VCGT: return "ARMISD::VCGT";
1069 case ARMISD::VCGTZ: return "ARMISD::VCGTZ";
1070 case ARMISD::VCLTZ: return "ARMISD::VCLTZ";
1071 case ARMISD::VCGTU: return "ARMISD::VCGTU";
1072 case ARMISD::VTST: return "ARMISD::VTST";
1074 case ARMISD::VSHL: return "ARMISD::VSHL";
1075 case ARMISD::VSHRs: return "ARMISD::VSHRs";
1076 case ARMISD::VSHRu: return "ARMISD::VSHRu";
1077 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
1078 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
1079 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
1080 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
1081 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
1082 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
1083 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
1084 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
1085 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
1086 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
1087 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
1088 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
1089 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
1090 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
1091 case ARMISD::VMOVIMM: return "ARMISD::VMOVIMM";
1092 case ARMISD::VMVNIMM: return "ARMISD::VMVNIMM";
1093 case ARMISD::VMOVFPIMM: return "ARMISD::VMOVFPIMM";
1094 case ARMISD::VDUP: return "ARMISD::VDUP";
1095 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
1096 case ARMISD::VEXT: return "ARMISD::VEXT";
1097 case ARMISD::VREV64: return "ARMISD::VREV64";
1098 case ARMISD::VREV32: return "ARMISD::VREV32";
1099 case ARMISD::VREV16: return "ARMISD::VREV16";
1100 case ARMISD::VZIP: return "ARMISD::VZIP";
1101 case ARMISD::VUZP: return "ARMISD::VUZP";
1102 case ARMISD::VTRN: return "ARMISD::VTRN";
1103 case ARMISD::VTBL1: return "ARMISD::VTBL1";
1104 case ARMISD::VTBL2: return "ARMISD::VTBL2";
1105 case ARMISD::VMULLs: return "ARMISD::VMULLs";
1106 case ARMISD::VMULLu: return "ARMISD::VMULLu";
1107 case ARMISD::UMLAL: return "ARMISD::UMLAL";
1108 case ARMISD::SMLAL: return "ARMISD::SMLAL";
1109 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
1110 case ARMISD::FMAX: return "ARMISD::FMAX";
1111 case ARMISD::FMIN: return "ARMISD::FMIN";
1112 case ARMISD::VMAXNM: return "ARMISD::VMAX";
1113 case ARMISD::VMINNM: return "ARMISD::VMIN";
1114 case ARMISD::BFI: return "ARMISD::BFI";
1115 case ARMISD::VORRIMM: return "ARMISD::VORRIMM";
1116 case ARMISD::VBICIMM: return "ARMISD::VBICIMM";
1117 case ARMISD::VBSL: return "ARMISD::VBSL";
1118 case ARMISD::VLD2DUP: return "ARMISD::VLD2DUP";
1119 case ARMISD::VLD3DUP: return "ARMISD::VLD3DUP";
1120 case ARMISD::VLD4DUP: return "ARMISD::VLD4DUP";
1121 case ARMISD::VLD1_UPD: return "ARMISD::VLD1_UPD";
1122 case ARMISD::VLD2_UPD: return "ARMISD::VLD2_UPD";
1123 case ARMISD::VLD3_UPD: return "ARMISD::VLD3_UPD";
1124 case ARMISD::VLD4_UPD: return "ARMISD::VLD4_UPD";
1125 case ARMISD::VLD2LN_UPD: return "ARMISD::VLD2LN_UPD";
1126 case ARMISD::VLD3LN_UPD: return "ARMISD::VLD3LN_UPD";
1127 case ARMISD::VLD4LN_UPD: return "ARMISD::VLD4LN_UPD";
1128 case ARMISD::VLD2DUP_UPD: return "ARMISD::VLD2DUP_UPD";
1129 case ARMISD::VLD3DUP_UPD: return "ARMISD::VLD3DUP_UPD";
1130 case ARMISD::VLD4DUP_UPD: return "ARMISD::VLD4DUP_UPD";
1131 case ARMISD::VST1_UPD: return "ARMISD::VST1_UPD";
1132 case ARMISD::VST2_UPD: return "ARMISD::VST2_UPD";
1133 case ARMISD::VST3_UPD: return "ARMISD::VST3_UPD";
1134 case ARMISD::VST4_UPD: return "ARMISD::VST4_UPD";
1135 case ARMISD::VST2LN_UPD: return "ARMISD::VST2LN_UPD";
1136 case ARMISD::VST3LN_UPD: return "ARMISD::VST3LN_UPD";
1137 case ARMISD::VST4LN_UPD: return "ARMISD::VST4LN_UPD";
1141 EVT ARMTargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
1142 if (!VT.isVector()) return getPointerTy();
1143 return VT.changeVectorElementTypeToInteger();
1146 /// getRegClassFor - Return the register class that should be used for the
1147 /// specified value type.
1148 const TargetRegisterClass *ARMTargetLowering::getRegClassFor(MVT VT) const {
1149 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
1150 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
1151 // load / store 4 to 8 consecutive D registers.
1152 if (Subtarget->hasNEON()) {
1153 if (VT == MVT::v4i64)
1154 return &ARM::QQPRRegClass;
1155 if (VT == MVT::v8i64)
1156 return &ARM::QQQQPRRegClass;
1158 return TargetLowering::getRegClassFor(VT);
1161 // memcpy, and other memory intrinsics, typically tries to use LDM/STM if the
1162 // source/dest is aligned and the copy size is large enough. We therefore want
1163 // to align such objects passed to memory intrinsics.
1164 bool ARMTargetLowering::shouldAlignPointerArgs(CallInst *CI, unsigned &MinSize,
1165 unsigned &PrefAlign) const {
1166 if (!isa<MemIntrinsic>(CI))
1169 // On ARM11 onwards (excluding M class) 8-byte aligned LDM is typically 1
1170 // cycle faster than 4-byte aligned LDM.
1171 PrefAlign = (Subtarget->hasV6Ops() && !Subtarget->isMClass() ? 8 : 4);
1175 // Create a fast isel object.
1177 ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
1178 const TargetLibraryInfo *libInfo) const {
1179 return ARM::createFastISel(funcInfo, libInfo);
1182 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
1183 unsigned NumVals = N->getNumValues();
1185 return Sched::RegPressure;
1187 for (unsigned i = 0; i != NumVals; ++i) {
1188 EVT VT = N->getValueType(i);
1189 if (VT == MVT::Glue || VT == MVT::Other)
1191 if (VT.isFloatingPoint() || VT.isVector())
1195 if (!N->isMachineOpcode())
1196 return Sched::RegPressure;
1198 // Load are scheduled for latency even if there instruction itinerary
1199 // is not available.
1200 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
1201 const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
1203 if (MCID.getNumDefs() == 0)
1204 return Sched::RegPressure;
1205 if (!Itins->isEmpty() &&
1206 Itins->getOperandCycle(MCID.getSchedClass(), 0) > 2)
1209 return Sched::RegPressure;
1212 //===----------------------------------------------------------------------===//
1214 //===----------------------------------------------------------------------===//
1216 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
1217 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
1219 default: llvm_unreachable("Unknown condition code!");
1220 case ISD::SETNE: return ARMCC::NE;
1221 case ISD::SETEQ: return ARMCC::EQ;
1222 case ISD::SETGT: return ARMCC::GT;
1223 case ISD::SETGE: return ARMCC::GE;
1224 case ISD::SETLT: return ARMCC::LT;
1225 case ISD::SETLE: return ARMCC::LE;
1226 case ISD::SETUGT: return ARMCC::HI;
1227 case ISD::SETUGE: return ARMCC::HS;
1228 case ISD::SETULT: return ARMCC::LO;
1229 case ISD::SETULE: return ARMCC::LS;
1233 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
1234 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
1235 ARMCC::CondCodes &CondCode2) {
1236 CondCode2 = ARMCC::AL;
1238 default: llvm_unreachable("Unknown FP condition!");
1240 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
1242 case ISD::SETOGT: CondCode = ARMCC::GT; break;
1244 case ISD::SETOGE: CondCode = ARMCC::GE; break;
1245 case ISD::SETOLT: CondCode = ARMCC::MI; break;
1246 case ISD::SETOLE: CondCode = ARMCC::LS; break;
1247 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
1248 case ISD::SETO: CondCode = ARMCC::VC; break;
1249 case ISD::SETUO: CondCode = ARMCC::VS; break;
1250 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
1251 case ISD::SETUGT: CondCode = ARMCC::HI; break;
1252 case ISD::SETUGE: CondCode = ARMCC::PL; break;
1254 case ISD::SETULT: CondCode = ARMCC::LT; break;
1256 case ISD::SETULE: CondCode = ARMCC::LE; break;
1258 case ISD::SETUNE: CondCode = ARMCC::NE; break;
1262 //===----------------------------------------------------------------------===//
1263 // Calling Convention Implementation
1264 //===----------------------------------------------------------------------===//
1266 #include "ARMGenCallingConv.inc"
1268 /// getEffectiveCallingConv - Get the effective calling convention, taking into
1269 /// account presence of floating point hardware and calling convention
1270 /// limitations, such as support for variadic functions.
1272 ARMTargetLowering::getEffectiveCallingConv(CallingConv::ID CC,
1273 bool isVarArg) const {
1276 llvm_unreachable("Unsupported calling convention");
1277 case CallingConv::ARM_AAPCS:
1278 case CallingConv::ARM_APCS:
1279 case CallingConv::GHC:
1281 case CallingConv::ARM_AAPCS_VFP:
1282 return isVarArg ? CallingConv::ARM_AAPCS : CallingConv::ARM_AAPCS_VFP;
1283 case CallingConv::C:
1284 if (!Subtarget->isAAPCS_ABI())
1285 return CallingConv::ARM_APCS;
1286 else if (Subtarget->hasVFP2() && !Subtarget->isThumb1Only() &&
1287 getTargetMachine().Options.FloatABIType == FloatABI::Hard &&
1289 return CallingConv::ARM_AAPCS_VFP;
1291 return CallingConv::ARM_AAPCS;
1292 case CallingConv::Fast:
1293 if (!Subtarget->isAAPCS_ABI()) {
1294 if (Subtarget->hasVFP2() && !Subtarget->isThumb1Only() && !isVarArg)
1295 return CallingConv::Fast;
1296 return CallingConv::ARM_APCS;
1297 } else if (Subtarget->hasVFP2() && !Subtarget->isThumb1Only() && !isVarArg)
1298 return CallingConv::ARM_AAPCS_VFP;
1300 return CallingConv::ARM_AAPCS;
1304 /// CCAssignFnForNode - Selects the correct CCAssignFn for the given
1305 /// CallingConvention.
1306 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
1308 bool isVarArg) const {
1309 switch (getEffectiveCallingConv(CC, isVarArg)) {
1311 llvm_unreachable("Unsupported calling convention");
1312 case CallingConv::ARM_APCS:
1313 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1314 case CallingConv::ARM_AAPCS:
1315 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1316 case CallingConv::ARM_AAPCS_VFP:
1317 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1318 case CallingConv::Fast:
1319 return (Return ? RetFastCC_ARM_APCS : FastCC_ARM_APCS);
1320 case CallingConv::GHC:
1321 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS_GHC);
1325 /// LowerCallResult - Lower the result values of a call into the
1326 /// appropriate copies out of appropriate physical registers.
1328 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1329 CallingConv::ID CallConv, bool isVarArg,
1330 const SmallVectorImpl<ISD::InputArg> &Ins,
1331 SDLoc dl, SelectionDAG &DAG,
1332 SmallVectorImpl<SDValue> &InVals,
1333 bool isThisReturn, SDValue ThisVal) const {
1335 // Assign locations to each value returned by this call.
1336 SmallVector<CCValAssign, 16> RVLocs;
1337 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
1338 *DAG.getContext(), Call);
1339 CCInfo.AnalyzeCallResult(Ins,
1340 CCAssignFnForNode(CallConv, /* Return*/ true,
1343 // Copy all of the result registers out of their specified physreg.
1344 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1345 CCValAssign VA = RVLocs[i];
1347 // Pass 'this' value directly from the argument to return value, to avoid
1348 // reg unit interference
1349 if (i == 0 && isThisReturn) {
1350 assert(!VA.needsCustom() && VA.getLocVT() == MVT::i32 &&
1351 "unexpected return calling convention register assignment");
1352 InVals.push_back(ThisVal);
1357 if (VA.needsCustom()) {
1358 // Handle f64 or half of a v2f64.
1359 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1361 Chain = Lo.getValue(1);
1362 InFlag = Lo.getValue(2);
1363 VA = RVLocs[++i]; // skip ahead to next loc
1364 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1366 Chain = Hi.getValue(1);
1367 InFlag = Hi.getValue(2);
1368 if (!Subtarget->isLittle())
1370 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1372 if (VA.getLocVT() == MVT::v2f64) {
1373 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1374 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1375 DAG.getConstant(0, MVT::i32));
1377 VA = RVLocs[++i]; // skip ahead to next loc
1378 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1379 Chain = Lo.getValue(1);
1380 InFlag = Lo.getValue(2);
1381 VA = RVLocs[++i]; // skip ahead to next loc
1382 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1383 Chain = Hi.getValue(1);
1384 InFlag = Hi.getValue(2);
1385 if (!Subtarget->isLittle())
1387 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1388 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1389 DAG.getConstant(1, MVT::i32));
1392 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
1394 Chain = Val.getValue(1);
1395 InFlag = Val.getValue(2);
1398 switch (VA.getLocInfo()) {
1399 default: llvm_unreachable("Unknown loc info!");
1400 case CCValAssign::Full: break;
1401 case CCValAssign::BCvt:
1402 Val = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), Val);
1406 InVals.push_back(Val);
1412 /// LowerMemOpCallTo - Store the argument to the stack.
1414 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
1415 SDValue StackPtr, SDValue Arg,
1416 SDLoc dl, SelectionDAG &DAG,
1417 const CCValAssign &VA,
1418 ISD::ArgFlagsTy Flags) const {
1419 unsigned LocMemOffset = VA.getLocMemOffset();
1420 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1421 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1422 return DAG.getStore(Chain, dl, Arg, PtrOff,
1423 MachinePointerInfo::getStack(LocMemOffset),
1427 void ARMTargetLowering::PassF64ArgInRegs(SDLoc dl, SelectionDAG &DAG,
1428 SDValue Chain, SDValue &Arg,
1429 RegsToPassVector &RegsToPass,
1430 CCValAssign &VA, CCValAssign &NextVA,
1432 SmallVectorImpl<SDValue> &MemOpChains,
1433 ISD::ArgFlagsTy Flags) const {
1435 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1436 DAG.getVTList(MVT::i32, MVT::i32), Arg);
1437 unsigned id = Subtarget->isLittle() ? 0 : 1;
1438 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd.getValue(id)));
1440 if (NextVA.isRegLoc())
1441 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1-id)));
1443 assert(NextVA.isMemLoc());
1444 if (!StackPtr.getNode())
1445 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1447 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1-id),
1453 /// LowerCall - Lowering a call into a callseq_start <-
1454 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
1457 ARMTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
1458 SmallVectorImpl<SDValue> &InVals) const {
1459 SelectionDAG &DAG = CLI.DAG;
1461 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
1462 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
1463 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
1464 SDValue Chain = CLI.Chain;
1465 SDValue Callee = CLI.Callee;
1466 bool &isTailCall = CLI.IsTailCall;
1467 CallingConv::ID CallConv = CLI.CallConv;
1468 bool doesNotRet = CLI.DoesNotReturn;
1469 bool isVarArg = CLI.IsVarArg;
1471 MachineFunction &MF = DAG.getMachineFunction();
1472 bool isStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1473 bool isThisReturn = false;
1474 bool isSibCall = false;
1476 // Disable tail calls if they're not supported.
1477 if (!Subtarget->supportsTailCall() || MF.getTarget().Options.DisableTailCalls)
1481 // Check if it's really possible to do a tail call.
1482 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1483 isVarArg, isStructRet, MF.getFunction()->hasStructRetAttr(),
1484 Outs, OutVals, Ins, DAG);
1485 if (!isTailCall && CLI.CS && CLI.CS->isMustTailCall())
1486 report_fatal_error("failed to perform tail call elimination on a call "
1487 "site marked musttail");
1488 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1489 // detected sibcalls.
1496 // Analyze operands of the call, assigning locations to each operand.
1497 SmallVector<CCValAssign, 16> ArgLocs;
1498 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs,
1499 *DAG.getContext(), Call);
1500 CCInfo.AnalyzeCallOperands(Outs,
1501 CCAssignFnForNode(CallConv, /* Return*/ false,
1504 // Get a count of how many bytes are to be pushed on the stack.
1505 unsigned NumBytes = CCInfo.getNextStackOffset();
1507 // For tail calls, memory operands are available in our caller's stack.
1511 // Adjust the stack pointer for the new arguments...
1512 // These operations are automatically eliminated by the prolog/epilog pass
1514 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true),
1517 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1519 RegsToPassVector RegsToPass;
1520 SmallVector<SDValue, 8> MemOpChains;
1522 // Walk the register/memloc assignments, inserting copies/loads. In the case
1523 // of tail call optimization, arguments are handled later.
1524 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1526 ++i, ++realArgIdx) {
1527 CCValAssign &VA = ArgLocs[i];
1528 SDValue Arg = OutVals[realArgIdx];
1529 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1530 bool isByVal = Flags.isByVal();
1532 // Promote the value if needed.
1533 switch (VA.getLocInfo()) {
1534 default: llvm_unreachable("Unknown loc info!");
1535 case CCValAssign::Full: break;
1536 case CCValAssign::SExt:
1537 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1539 case CCValAssign::ZExt:
1540 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1542 case CCValAssign::AExt:
1543 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1545 case CCValAssign::BCvt:
1546 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1550 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1551 if (VA.needsCustom()) {
1552 if (VA.getLocVT() == MVT::v2f64) {
1553 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1554 DAG.getConstant(0, MVT::i32));
1555 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1556 DAG.getConstant(1, MVT::i32));
1558 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1559 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1561 VA = ArgLocs[++i]; // skip ahead to next loc
1562 if (VA.isRegLoc()) {
1563 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1564 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1566 assert(VA.isMemLoc());
1568 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1569 dl, DAG, VA, Flags));
1572 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1573 StackPtr, MemOpChains, Flags);
1575 } else if (VA.isRegLoc()) {
1576 if (realArgIdx == 0 && Flags.isReturned() && Outs[0].VT == MVT::i32) {
1577 assert(VA.getLocVT() == MVT::i32 &&
1578 "unexpected calling convention register assignment");
1579 assert(!Ins.empty() && Ins[0].VT == MVT::i32 &&
1580 "unexpected use of 'returned'");
1581 isThisReturn = true;
1583 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1584 } else if (isByVal) {
1585 assert(VA.isMemLoc());
1586 unsigned offset = 0;
1588 // True if this byval aggregate will be split between registers
1590 unsigned ByValArgsCount = CCInfo.getInRegsParamsCount();
1591 unsigned CurByValIdx = CCInfo.getInRegsParamsProcessed();
1593 if (CurByValIdx < ByValArgsCount) {
1595 unsigned RegBegin, RegEnd;
1596 CCInfo.getInRegsParamInfo(CurByValIdx, RegBegin, RegEnd);
1598 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1600 for (i = 0, j = RegBegin; j < RegEnd; i++, j++) {
1601 SDValue Const = DAG.getConstant(4*i, MVT::i32);
1602 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
1603 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
1604 MachinePointerInfo(),
1605 false, false, false,
1606 DAG.InferPtrAlignment(AddArg));
1607 MemOpChains.push_back(Load.getValue(1));
1608 RegsToPass.push_back(std::make_pair(j, Load));
1611 // If parameter size outsides register area, "offset" value
1612 // helps us to calculate stack slot for remained part properly.
1613 offset = RegEnd - RegBegin;
1615 CCInfo.nextInRegsParam();
1618 if (Flags.getByValSize() > 4*offset) {
1619 unsigned LocMemOffset = VA.getLocMemOffset();
1620 SDValue StkPtrOff = DAG.getIntPtrConstant(LocMemOffset);
1621 SDValue Dst = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr,
1623 SDValue SrcOffset = DAG.getIntPtrConstant(4*offset);
1624 SDValue Src = DAG.getNode(ISD::ADD, dl, getPointerTy(), Arg, SrcOffset);
1625 SDValue SizeNode = DAG.getConstant(Flags.getByValSize() - 4*offset,
1627 SDValue AlignNode = DAG.getConstant(Flags.getByValAlign(), MVT::i32);
1629 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
1630 SDValue Ops[] = { Chain, Dst, Src, SizeNode, AlignNode};
1631 MemOpChains.push_back(DAG.getNode(ARMISD::COPY_STRUCT_BYVAL, dl, VTs,
1634 } else if (!isSibCall) {
1635 assert(VA.isMemLoc());
1637 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1638 dl, DAG, VA, Flags));
1642 if (!MemOpChains.empty())
1643 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains);
1645 // Build a sequence of copy-to-reg nodes chained together with token chain
1646 // and flag operands which copy the outgoing args into the appropriate regs.
1648 // Tail call byval lowering might overwrite argument registers so in case of
1649 // tail call optimization the copies to registers are lowered later.
1651 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1652 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1653 RegsToPass[i].second, InFlag);
1654 InFlag = Chain.getValue(1);
1657 // For tail calls lower the arguments to the 'real' stack slot.
1659 // Force all the incoming stack arguments to be loaded from the stack
1660 // before any new outgoing arguments are stored to the stack, because the
1661 // outgoing stack slots may alias the incoming argument stack slots, and
1662 // the alias isn't otherwise explicit. This is slightly more conservative
1663 // than necessary, because it means that each store effectively depends
1664 // on every argument instead of just those arguments it would clobber.
1666 // Do not flag preceding copytoreg stuff together with the following stuff.
1668 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1669 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1670 RegsToPass[i].second, InFlag);
1671 InFlag = Chain.getValue(1);
1676 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1677 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1678 // node so that legalize doesn't hack it.
1679 bool isDirect = false;
1680 bool isARMFunc = false;
1681 bool isLocalARMFunc = false;
1682 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1684 if (EnableARMLongCalls) {
1685 assert((Subtarget->isTargetWindows() ||
1686 getTargetMachine().getRelocationModel() == Reloc::Static) &&
1687 "long-calls with non-static relocation model!");
1688 // Handle a global address or an external symbol. If it's not one of
1689 // those, the target's already in a register, so we don't need to do
1691 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1692 const GlobalValue *GV = G->getGlobal();
1693 // Create a constant pool entry for the callee address
1694 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1695 ARMConstantPoolValue *CPV =
1696 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 0);
1698 // Get the address of the callee into a register
1699 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1700 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1701 Callee = DAG.getLoad(getPointerTy(), dl,
1702 DAG.getEntryNode(), CPAddr,
1703 MachinePointerInfo::getConstantPool(),
1704 false, false, false, 0);
1705 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1706 const char *Sym = S->getSymbol();
1708 // Create a constant pool entry for the callee address
1709 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1710 ARMConstantPoolValue *CPV =
1711 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1712 ARMPCLabelIndex, 0);
1713 // Get the address of the callee into a register
1714 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1715 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1716 Callee = DAG.getLoad(getPointerTy(), dl,
1717 DAG.getEntryNode(), CPAddr,
1718 MachinePointerInfo::getConstantPool(),
1719 false, false, false, 0);
1721 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1722 const GlobalValue *GV = G->getGlobal();
1724 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1725 bool isStub = (isExt && Subtarget->isTargetMachO()) &&
1726 getTargetMachine().getRelocationModel() != Reloc::Static;
1727 isARMFunc = !Subtarget->isThumb() || (isStub && !Subtarget->isMClass());
1728 // ARM call to a local ARM function is predicable.
1729 isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking);
1730 // tBX takes a register source operand.
1731 if (isStub && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1732 assert(Subtarget->isTargetMachO() && "WrapperPIC use on non-MachO?");
1733 Callee = DAG.getNode(ARMISD::WrapperPIC, dl, getPointerTy(),
1734 DAG.getTargetGlobalAddress(GV, dl, getPointerTy(),
1735 0, ARMII::MO_NONLAZY));
1736 Callee = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Callee,
1737 MachinePointerInfo::getGOT(), false, false, true, 0);
1738 } else if (Subtarget->isTargetCOFF()) {
1739 assert(Subtarget->isTargetWindows() &&
1740 "Windows is the only supported COFF target");
1741 unsigned TargetFlags = GV->hasDLLImportStorageClass()
1742 ? ARMII::MO_DLLIMPORT
1743 : ARMII::MO_NO_FLAG;
1744 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), /*Offset=*/0,
1746 if (GV->hasDLLImportStorageClass())
1747 Callee = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
1748 DAG.getNode(ARMISD::Wrapper, dl, getPointerTy(),
1749 Callee), MachinePointerInfo::getGOT(),
1750 false, false, false, 0);
1752 // On ELF targets for PIC code, direct calls should go through the PLT
1753 unsigned OpFlags = 0;
1754 if (Subtarget->isTargetELF() &&
1755 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1756 OpFlags = ARMII::MO_PLT;
1757 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
1759 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1761 bool isStub = Subtarget->isTargetMachO() &&
1762 getTargetMachine().getRelocationModel() != Reloc::Static;
1763 isARMFunc = !Subtarget->isThumb() || (isStub && !Subtarget->isMClass());
1764 // tBX takes a register source operand.
1765 const char *Sym = S->getSymbol();
1766 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1767 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1768 ARMConstantPoolValue *CPV =
1769 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1770 ARMPCLabelIndex, 4);
1771 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1772 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1773 Callee = DAG.getLoad(getPointerTy(), dl,
1774 DAG.getEntryNode(), CPAddr,
1775 MachinePointerInfo::getConstantPool(),
1776 false, false, false, 0);
1777 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1778 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1779 getPointerTy(), Callee, PICLabel);
1781 unsigned OpFlags = 0;
1782 // On ELF targets for PIC code, direct calls should go through the PLT
1783 if (Subtarget->isTargetELF() &&
1784 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1785 OpFlags = ARMII::MO_PLT;
1786 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlags);
1790 // FIXME: handle tail calls differently.
1792 bool HasMinSizeAttr = MF.getFunction()->hasFnAttribute(Attribute::MinSize);
1793 if (Subtarget->isThumb()) {
1794 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1795 CallOpc = ARMISD::CALL_NOLINK;
1797 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1799 if (!isDirect && !Subtarget->hasV5TOps())
1800 CallOpc = ARMISD::CALL_NOLINK;
1801 else if (doesNotRet && isDirect && Subtarget->hasRAS() &&
1802 // Emit regular call when code size is the priority
1804 // "mov lr, pc; b _foo" to avoid confusing the RSP
1805 CallOpc = ARMISD::CALL_NOLINK;
1807 CallOpc = isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL;
1810 std::vector<SDValue> Ops;
1811 Ops.push_back(Chain);
1812 Ops.push_back(Callee);
1814 // Add argument registers to the end of the list so that they are known live
1816 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1817 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1818 RegsToPass[i].second.getValueType()));
1820 // Add a register mask operand representing the call-preserved registers.
1822 const uint32_t *Mask;
1823 const ARMBaseRegisterInfo *ARI = Subtarget->getRegisterInfo();
1825 // For 'this' returns, use the R0-preserving mask if applicable
1826 Mask = ARI->getThisReturnPreservedMask(MF, CallConv);
1828 // Set isThisReturn to false if the calling convention is not one that
1829 // allows 'returned' to be modeled in this way, so LowerCallResult does
1830 // not try to pass 'this' straight through
1831 isThisReturn = false;
1832 Mask = ARI->getCallPreservedMask(MF, CallConv);
1835 Mask = ARI->getCallPreservedMask(MF, CallConv);
1837 assert(Mask && "Missing call preserved mask for calling convention");
1838 Ops.push_back(DAG.getRegisterMask(Mask));
1841 if (InFlag.getNode())
1842 Ops.push_back(InFlag);
1844 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
1846 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, Ops);
1848 // Returns a chain and a flag for retval copy to use.
1849 Chain = DAG.getNode(CallOpc, dl, NodeTys, Ops);
1850 InFlag = Chain.getValue(1);
1852 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1853 DAG.getIntPtrConstant(0, true), InFlag, dl);
1855 InFlag = Chain.getValue(1);
1857 // Handle result values, copying them out of physregs into vregs that we
1859 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins, dl, DAG,
1860 InVals, isThisReturn,
1861 isThisReturn ? OutVals[0] : SDValue());
1864 /// HandleByVal - Every parameter *after* a byval parameter is passed
1865 /// on the stack. Remember the next parameter register to allocate,
1866 /// and then confiscate the rest of the parameter registers to insure
1868 void ARMTargetLowering::HandleByVal(CCState *State, unsigned &Size,
1869 unsigned Align) const {
1870 assert((State->getCallOrPrologue() == Prologue ||
1871 State->getCallOrPrologue() == Call) &&
1872 "unhandled ParmContext");
1874 // Byval (as with any stack) slots are always at least 4 byte aligned.
1875 Align = std::max(Align, 4U);
1877 unsigned Reg = State->AllocateReg(GPRArgRegs);
1881 unsigned AlignInRegs = Align / 4;
1882 unsigned Waste = (ARM::R4 - Reg) % AlignInRegs;
1883 for (unsigned i = 0; i < Waste; ++i)
1884 Reg = State->AllocateReg(GPRArgRegs);
1889 unsigned Excess = 4 * (ARM::R4 - Reg);
1891 // Special case when NSAA != SP and parameter size greater than size of
1892 // all remained GPR regs. In that case we can't split parameter, we must
1893 // send it to stack. We also must set NCRN to R4, so waste all
1894 // remained registers.
1895 const unsigned NSAAOffset = State->getNextStackOffset();
1896 if (NSAAOffset != 0 && Size > Excess) {
1897 while (State->AllocateReg(GPRArgRegs))
1902 // First register for byval parameter is the first register that wasn't
1903 // allocated before this method call, so it would be "reg".
1904 // If parameter is small enough to be saved in range [reg, r4), then
1905 // the end (first after last) register would be reg + param-size-in-regs,
1906 // else parameter would be splitted between registers and stack,
1907 // end register would be r4 in this case.
1908 unsigned ByValRegBegin = Reg;
1909 unsigned ByValRegEnd = std::min<unsigned>(Reg + Size / 4, ARM::R4);
1910 State->addInRegsParamInfo(ByValRegBegin, ByValRegEnd);
1911 // Note, first register is allocated in the beginning of function already,
1912 // allocate remained amount of registers we need.
1913 for (unsigned i = Reg + 1; i != ByValRegEnd; ++i)
1914 State->AllocateReg(GPRArgRegs);
1915 // A byval parameter that is split between registers and memory needs its
1916 // size truncated here.
1917 // In the case where the entire structure fits in registers, we set the
1918 // size in memory to zero.
1919 Size = std::max<int>(Size - Excess, 0);
1923 /// MatchingStackOffset - Return true if the given stack call argument is
1924 /// already available in the same position (relatively) of the caller's
1925 /// incoming argument stack.
1927 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1928 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1929 const TargetInstrInfo *TII) {
1930 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1932 if (Arg.getOpcode() == ISD::CopyFromReg) {
1933 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1934 if (!TargetRegisterInfo::isVirtualRegister(VR))
1936 MachineInstr *Def = MRI->getVRegDef(VR);
1939 if (!Flags.isByVal()) {
1940 if (!TII->isLoadFromStackSlot(Def, FI))
1945 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
1946 if (Flags.isByVal())
1947 // ByVal argument is passed in as a pointer but it's now being
1948 // dereferenced. e.g.
1949 // define @foo(%struct.X* %A) {
1950 // tail call @bar(%struct.X* byval %A)
1953 SDValue Ptr = Ld->getBasePtr();
1954 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
1957 FI = FINode->getIndex();
1961 assert(FI != INT_MAX);
1962 if (!MFI->isFixedObjectIndex(FI))
1964 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
1967 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
1968 /// for tail call optimization. Targets which want to do tail call
1969 /// optimization should implement this function.
1971 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1972 CallingConv::ID CalleeCC,
1974 bool isCalleeStructRet,
1975 bool isCallerStructRet,
1976 const SmallVectorImpl<ISD::OutputArg> &Outs,
1977 const SmallVectorImpl<SDValue> &OutVals,
1978 const SmallVectorImpl<ISD::InputArg> &Ins,
1979 SelectionDAG& DAG) const {
1980 const Function *CallerF = DAG.getMachineFunction().getFunction();
1981 CallingConv::ID CallerCC = CallerF->getCallingConv();
1982 bool CCMatch = CallerCC == CalleeCC;
1984 // Look for obvious safe cases to perform tail call optimization that do not
1985 // require ABI changes. This is what gcc calls sibcall.
1987 // Do not sibcall optimize vararg calls unless the call site is not passing
1989 if (isVarArg && !Outs.empty())
1992 // Exception-handling functions need a special set of instructions to indicate
1993 // a return to the hardware. Tail-calling another function would probably
1995 if (CallerF->hasFnAttribute("interrupt"))
1998 // Also avoid sibcall optimization if either caller or callee uses struct
1999 // return semantics.
2000 if (isCalleeStructRet || isCallerStructRet)
2003 // FIXME: Completely disable sibcall for Thumb1 since ThumbRegisterInfo::
2004 // emitEpilogue is not ready for them. Thumb tail calls also use t2B, as
2005 // the Thumb1 16-bit unconditional branch doesn't have sufficient relocation
2006 // support in the assembler and linker to be used. This would need to be
2007 // fixed to fully support tail calls in Thumb1.
2009 // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take
2010 // LR. This means if we need to reload LR, it takes an extra instructions,
2011 // which outweighs the value of the tail call; but here we don't know yet
2012 // whether LR is going to be used. Probably the right approach is to
2013 // generate the tail call here and turn it back into CALL/RET in
2014 // emitEpilogue if LR is used.
2016 // Thumb1 PIC calls to external symbols use BX, so they can be tail calls,
2017 // but we need to make sure there are enough registers; the only valid
2018 // registers are the 4 used for parameters. We don't currently do this
2020 if (Subtarget->isThumb1Only())
2023 // Externally-defined functions with weak linkage should not be
2024 // tail-called on ARM when the OS does not support dynamic
2025 // pre-emption of symbols, as the AAELF spec requires normal calls
2026 // to undefined weak functions to be replaced with a NOP or jump to the
2027 // next instruction. The behaviour of branch instructions in this
2028 // situation (as used for tail calls) is implementation-defined, so we
2029 // cannot rely on the linker replacing the tail call with a return.
2030 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2031 const GlobalValue *GV = G->getGlobal();
2032 const Triple TT(getTargetMachine().getTargetTriple());
2033 if (GV->hasExternalWeakLinkage() &&
2034 (!TT.isOSWindows() || TT.isOSBinFormatELF() || TT.isOSBinFormatMachO()))
2038 // If the calling conventions do not match, then we'd better make sure the
2039 // results are returned in the same way as what the caller expects.
2041 SmallVector<CCValAssign, 16> RVLocs1;
2042 ARMCCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(), RVLocs1,
2043 *DAG.getContext(), Call);
2044 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
2046 SmallVector<CCValAssign, 16> RVLocs2;
2047 ARMCCState CCInfo2(CallerCC, false, DAG.getMachineFunction(), RVLocs2,
2048 *DAG.getContext(), Call);
2049 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
2051 if (RVLocs1.size() != RVLocs2.size())
2053 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
2054 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
2056 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
2058 if (RVLocs1[i].isRegLoc()) {
2059 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
2062 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
2068 // If Caller's vararg or byval argument has been split between registers and
2069 // stack, do not perform tail call, since part of the argument is in caller's
2071 const ARMFunctionInfo *AFI_Caller = DAG.getMachineFunction().
2072 getInfo<ARMFunctionInfo>();
2073 if (AFI_Caller->getArgRegsSaveSize())
2076 // If the callee takes no arguments then go on to check the results of the
2078 if (!Outs.empty()) {
2079 // Check if stack adjustment is needed. For now, do not do this if any
2080 // argument is passed on the stack.
2081 SmallVector<CCValAssign, 16> ArgLocs;
2082 ARMCCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
2083 *DAG.getContext(), Call);
2084 CCInfo.AnalyzeCallOperands(Outs,
2085 CCAssignFnForNode(CalleeCC, false, isVarArg));
2086 if (CCInfo.getNextStackOffset()) {
2087 MachineFunction &MF = DAG.getMachineFunction();
2089 // Check if the arguments are already laid out in the right way as
2090 // the caller's fixed stack objects.
2091 MachineFrameInfo *MFI = MF.getFrameInfo();
2092 const MachineRegisterInfo *MRI = &MF.getRegInfo();
2093 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
2094 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
2096 ++i, ++realArgIdx) {
2097 CCValAssign &VA = ArgLocs[i];
2098 EVT RegVT = VA.getLocVT();
2099 SDValue Arg = OutVals[realArgIdx];
2100 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
2101 if (VA.getLocInfo() == CCValAssign::Indirect)
2103 if (VA.needsCustom()) {
2104 // f64 and vector types are split into multiple registers or
2105 // register/stack-slot combinations. The types will not match
2106 // the registers; give up on memory f64 refs until we figure
2107 // out what to do about this.
2110 if (!ArgLocs[++i].isRegLoc())
2112 if (RegVT == MVT::v2f64) {
2113 if (!ArgLocs[++i].isRegLoc())
2115 if (!ArgLocs[++i].isRegLoc())
2118 } else if (!VA.isRegLoc()) {
2119 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
2131 ARMTargetLowering::CanLowerReturn(CallingConv::ID CallConv,
2132 MachineFunction &MF, bool isVarArg,
2133 const SmallVectorImpl<ISD::OutputArg> &Outs,
2134 LLVMContext &Context) const {
2135 SmallVector<CCValAssign, 16> RVLocs;
2136 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, Context);
2137 return CCInfo.CheckReturn(Outs, CCAssignFnForNode(CallConv, /*Return=*/true,
2141 static SDValue LowerInterruptReturn(SmallVectorImpl<SDValue> &RetOps,
2142 SDLoc DL, SelectionDAG &DAG) {
2143 const MachineFunction &MF = DAG.getMachineFunction();
2144 const Function *F = MF.getFunction();
2146 StringRef IntKind = F->getFnAttribute("interrupt").getValueAsString();
2148 // See ARM ARM v7 B1.8.3. On exception entry LR is set to a possibly offset
2149 // version of the "preferred return address". These offsets affect the return
2150 // instruction if this is a return from PL1 without hypervisor extensions.
2151 // IRQ/FIQ: +4 "subs pc, lr, #4"
2152 // SWI: 0 "subs pc, lr, #0"
2153 // ABORT: +4 "subs pc, lr, #4"
2154 // UNDEF: +4/+2 "subs pc, lr, #0"
2155 // UNDEF varies depending on where the exception came from ARM or Thumb
2156 // mode. Alongside GCC, we throw our hands up in disgust and pretend it's 0.
2159 if (IntKind == "" || IntKind == "IRQ" || IntKind == "FIQ" ||
2162 else if (IntKind == "SWI" || IntKind == "UNDEF")
2165 report_fatal_error("Unsupported interrupt attribute. If present, value "
2166 "must be one of: IRQ, FIQ, SWI, ABORT or UNDEF");
2168 RetOps.insert(RetOps.begin() + 1, DAG.getConstant(LROffset, MVT::i32, false));
2170 return DAG.getNode(ARMISD::INTRET_FLAG, DL, MVT::Other, RetOps);
2174 ARMTargetLowering::LowerReturn(SDValue Chain,
2175 CallingConv::ID CallConv, bool isVarArg,
2176 const SmallVectorImpl<ISD::OutputArg> &Outs,
2177 const SmallVectorImpl<SDValue> &OutVals,
2178 SDLoc dl, SelectionDAG &DAG) const {
2180 // CCValAssign - represent the assignment of the return value to a location.
2181 SmallVector<CCValAssign, 16> RVLocs;
2183 // CCState - Info about the registers and stack slots.
2184 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
2185 *DAG.getContext(), Call);
2187 // Analyze outgoing return values.
2188 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
2192 SmallVector<SDValue, 4> RetOps;
2193 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
2194 bool isLittleEndian = Subtarget->isLittle();
2196 MachineFunction &MF = DAG.getMachineFunction();
2197 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2198 AFI->setReturnRegsCount(RVLocs.size());
2200 // Copy the result values into the output registers.
2201 for (unsigned i = 0, realRVLocIdx = 0;
2203 ++i, ++realRVLocIdx) {
2204 CCValAssign &VA = RVLocs[i];
2205 assert(VA.isRegLoc() && "Can only return in registers!");
2207 SDValue Arg = OutVals[realRVLocIdx];
2209 switch (VA.getLocInfo()) {
2210 default: llvm_unreachable("Unknown loc info!");
2211 case CCValAssign::Full: break;
2212 case CCValAssign::BCvt:
2213 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
2217 if (VA.needsCustom()) {
2218 if (VA.getLocVT() == MVT::v2f64) {
2219 // Extract the first half and return it in two registers.
2220 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
2221 DAG.getConstant(0, MVT::i32));
2222 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
2223 DAG.getVTList(MVT::i32, MVT::i32), Half);
2225 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2226 HalfGPRs.getValue(isLittleEndian ? 0 : 1),
2228 Flag = Chain.getValue(1);
2229 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2230 VA = RVLocs[++i]; // skip ahead to next loc
2231 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2232 HalfGPRs.getValue(isLittleEndian ? 1 : 0),
2234 Flag = Chain.getValue(1);
2235 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2236 VA = RVLocs[++i]; // skip ahead to next loc
2238 // Extract the 2nd half and fall through to handle it as an f64 value.
2239 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
2240 DAG.getConstant(1, MVT::i32));
2242 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
2244 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
2245 DAG.getVTList(MVT::i32, MVT::i32), Arg);
2246 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2247 fmrrd.getValue(isLittleEndian ? 0 : 1),
2249 Flag = Chain.getValue(1);
2250 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2251 VA = RVLocs[++i]; // skip ahead to next loc
2252 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2253 fmrrd.getValue(isLittleEndian ? 1 : 0),
2256 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
2258 // Guarantee that all emitted copies are
2259 // stuck together, avoiding something bad.
2260 Flag = Chain.getValue(1);
2261 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2264 // Update chain and glue.
2267 RetOps.push_back(Flag);
2269 // CPUs which aren't M-class use a special sequence to return from
2270 // exceptions (roughly, any instruction setting pc and cpsr simultaneously,
2271 // though we use "subs pc, lr, #N").
2273 // M-class CPUs actually use a normal return sequence with a special
2274 // (hardware-provided) value in LR, so the normal code path works.
2275 if (DAG.getMachineFunction().getFunction()->hasFnAttribute("interrupt") &&
2276 !Subtarget->isMClass()) {
2277 if (Subtarget->isThumb1Only())
2278 report_fatal_error("interrupt attribute is not supported in Thumb1");
2279 return LowerInterruptReturn(RetOps, dl, DAG);
2282 return DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, RetOps);
2285 bool ARMTargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
2286 if (N->getNumValues() != 1)
2288 if (!N->hasNUsesOfValue(1, 0))
2291 SDValue TCChain = Chain;
2292 SDNode *Copy = *N->use_begin();
2293 if (Copy->getOpcode() == ISD::CopyToReg) {
2294 // If the copy has a glue operand, we conservatively assume it isn't safe to
2295 // perform a tail call.
2296 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
2298 TCChain = Copy->getOperand(0);
2299 } else if (Copy->getOpcode() == ARMISD::VMOVRRD) {
2300 SDNode *VMov = Copy;
2301 // f64 returned in a pair of GPRs.
2302 SmallPtrSet<SDNode*, 2> Copies;
2303 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
2305 if (UI->getOpcode() != ISD::CopyToReg)
2309 if (Copies.size() > 2)
2312 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
2314 SDValue UseChain = UI->getOperand(0);
2315 if (Copies.count(UseChain.getNode()))
2319 // We are at the top of this chain.
2320 // If the copy has a glue operand, we conservatively assume it
2321 // isn't safe to perform a tail call.
2322 if (UI->getOperand(UI->getNumOperands()-1).getValueType() == MVT::Glue)
2328 } else if (Copy->getOpcode() == ISD::BITCAST) {
2329 // f32 returned in a single GPR.
2330 if (!Copy->hasOneUse())
2332 Copy = *Copy->use_begin();
2333 if (Copy->getOpcode() != ISD::CopyToReg || !Copy->hasNUsesOfValue(1, 0))
2335 // If the copy has a glue operand, we conservatively assume it isn't safe to
2336 // perform a tail call.
2337 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
2339 TCChain = Copy->getOperand(0);
2344 bool HasRet = false;
2345 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
2347 if (UI->getOpcode() != ARMISD::RET_FLAG &&
2348 UI->getOpcode() != ARMISD::INTRET_FLAG)
2360 bool ARMTargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
2361 if (!Subtarget->supportsTailCall())
2364 if (!CI->isTailCall() || getTargetMachine().Options.DisableTailCalls)
2367 return !Subtarget->isThumb1Only();
2370 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
2371 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
2372 // one of the above mentioned nodes. It has to be wrapped because otherwise
2373 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
2374 // be used to form addressing mode. These wrapped nodes will be selected
2376 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
2377 EVT PtrVT = Op.getValueType();
2378 // FIXME there is no actual debug info here
2380 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
2382 if (CP->isMachineConstantPoolEntry())
2383 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
2384 CP->getAlignment());
2386 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
2387 CP->getAlignment());
2388 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
2391 unsigned ARMTargetLowering::getJumpTableEncoding() const {
2392 return MachineJumpTableInfo::EK_Inline;
2395 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
2396 SelectionDAG &DAG) const {
2397 MachineFunction &MF = DAG.getMachineFunction();
2398 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2399 unsigned ARMPCLabelIndex = 0;
2401 EVT PtrVT = getPointerTy();
2402 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
2403 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2405 if (RelocM == Reloc::Static) {
2406 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
2408 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2409 ARMPCLabelIndex = AFI->createPICLabelUId();
2410 ARMConstantPoolValue *CPV =
2411 ARMConstantPoolConstant::Create(BA, ARMPCLabelIndex,
2412 ARMCP::CPBlockAddress, PCAdj);
2413 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2415 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
2416 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
2417 MachinePointerInfo::getConstantPool(),
2418 false, false, false, 0);
2419 if (RelocM == Reloc::Static)
2421 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2422 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
2425 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
2427 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
2428 SelectionDAG &DAG) const {
2430 EVT PtrVT = getPointerTy();
2431 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2432 MachineFunction &MF = DAG.getMachineFunction();
2433 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2434 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2435 ARMConstantPoolValue *CPV =
2436 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2437 ARMCP::CPValue, PCAdj, ARMCP::TLSGD, true);
2438 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2439 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
2440 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
2441 MachinePointerInfo::getConstantPool(),
2442 false, false, false, 0);
2443 SDValue Chain = Argument.getValue(1);
2445 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2446 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
2448 // call __tls_get_addr.
2451 Entry.Node = Argument;
2452 Entry.Ty = (Type *) Type::getInt32Ty(*DAG.getContext());
2453 Args.push_back(Entry);
2455 // FIXME: is there useful debug info available here?
2456 TargetLowering::CallLoweringInfo CLI(DAG);
2457 CLI.setDebugLoc(dl).setChain(Chain)
2458 .setCallee(CallingConv::C, Type::getInt32Ty(*DAG.getContext()),
2459 DAG.getExternalSymbol("__tls_get_addr", PtrVT), std::move(Args),
2462 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
2463 return CallResult.first;
2466 // Lower ISD::GlobalTLSAddress using the "initial exec" or
2467 // "local exec" model.
2469 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
2471 TLSModel::Model model) const {
2472 const GlobalValue *GV = GA->getGlobal();
2475 SDValue Chain = DAG.getEntryNode();
2476 EVT PtrVT = getPointerTy();
2477 // Get the Thread Pointer
2478 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2480 if (model == TLSModel::InitialExec) {
2481 MachineFunction &MF = DAG.getMachineFunction();
2482 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2483 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2484 // Initial exec model.
2485 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2486 ARMConstantPoolValue *CPV =
2487 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2488 ARMCP::CPValue, PCAdj, ARMCP::GOTTPOFF,
2490 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2491 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2492 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2493 MachinePointerInfo::getConstantPool(),
2494 false, false, false, 0);
2495 Chain = Offset.getValue(1);
2497 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2498 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
2500 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2501 MachinePointerInfo::getConstantPool(),
2502 false, false, false, 0);
2505 assert(model == TLSModel::LocalExec);
2506 ARMConstantPoolValue *CPV =
2507 ARMConstantPoolConstant::Create(GV, ARMCP::TPOFF);
2508 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2509 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2510 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2511 MachinePointerInfo::getConstantPool(),
2512 false, false, false, 0);
2515 // The address of the thread local variable is the add of the thread
2516 // pointer with the offset of the variable.
2517 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
2521 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
2522 // TODO: implement the "local dynamic" model
2523 assert(Subtarget->isTargetELF() &&
2524 "TLS not implemented for non-ELF targets");
2525 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
2527 TLSModel::Model model = getTargetMachine().getTLSModel(GA->getGlobal());
2530 case TLSModel::GeneralDynamic:
2531 case TLSModel::LocalDynamic:
2532 return LowerToTLSGeneralDynamicModel(GA, DAG);
2533 case TLSModel::InitialExec:
2534 case TLSModel::LocalExec:
2535 return LowerToTLSExecModels(GA, DAG, model);
2537 llvm_unreachable("bogus TLS model");
2540 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
2541 SelectionDAG &DAG) const {
2542 EVT PtrVT = getPointerTy();
2544 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2545 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2546 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
2547 ARMConstantPoolValue *CPV =
2548 ARMConstantPoolConstant::Create(GV,
2549 UseGOTOFF ? ARMCP::GOTOFF : ARMCP::GOT);
2550 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2551 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2552 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
2554 MachinePointerInfo::getConstantPool(),
2555 false, false, false, 0);
2556 SDValue Chain = Result.getValue(1);
2557 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
2558 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
2560 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
2561 MachinePointerInfo::getGOT(),
2562 false, false, false, 0);
2566 // If we have T2 ops, we can materialize the address directly via movt/movw
2567 // pair. This is always cheaper.
2568 if (Subtarget->useMovt(DAG.getMachineFunction())) {
2570 // FIXME: Once remat is capable of dealing with instructions with register
2571 // operands, expand this into two nodes.
2572 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2573 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2575 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2576 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2577 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2578 MachinePointerInfo::getConstantPool(),
2579 false, false, false, 0);
2583 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
2584 SelectionDAG &DAG) const {
2585 EVT PtrVT = getPointerTy();
2587 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2588 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2590 if (Subtarget->useMovt(DAG.getMachineFunction()))
2593 // FIXME: Once remat is capable of dealing with instructions with register
2594 // operands, expand this into multiple nodes
2596 RelocM == Reloc::PIC_ ? ARMISD::WrapperPIC : ARMISD::Wrapper;
2598 SDValue G = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, ARMII::MO_NONLAZY);
2599 SDValue Result = DAG.getNode(Wrapper, dl, PtrVT, G);
2601 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2602 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
2603 MachinePointerInfo::getGOT(), false, false, false, 0);
2607 SDValue ARMTargetLowering::LowerGlobalAddressWindows(SDValue Op,
2608 SelectionDAG &DAG) const {
2609 assert(Subtarget->isTargetWindows() && "non-Windows COFF is not supported");
2610 assert(Subtarget->useMovt(DAG.getMachineFunction()) &&
2611 "Windows on ARM expects to use movw/movt");
2613 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2614 const ARMII::TOF TargetFlags =
2615 (GV->hasDLLImportStorageClass() ? ARMII::MO_DLLIMPORT : ARMII::MO_NO_FLAG);
2616 EVT PtrVT = getPointerTy();
2622 // FIXME: Once remat is capable of dealing with instructions with register
2623 // operands, expand this into two nodes.
2624 Result = DAG.getNode(ARMISD::Wrapper, DL, PtrVT,
2625 DAG.getTargetGlobalAddress(GV, DL, PtrVT, /*Offset=*/0,
2627 if (GV->hasDLLImportStorageClass())
2628 Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), Result,
2629 MachinePointerInfo::getGOT(), false, false, false, 0);
2633 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
2634 SelectionDAG &DAG) const {
2635 assert(Subtarget->isTargetELF() &&
2636 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
2637 MachineFunction &MF = DAG.getMachineFunction();
2638 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2639 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2640 EVT PtrVT = getPointerTy();
2642 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2643 ARMConstantPoolValue *CPV =
2644 ARMConstantPoolSymbol::Create(*DAG.getContext(), "_GLOBAL_OFFSET_TABLE_",
2645 ARMPCLabelIndex, PCAdj);
2646 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2647 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2648 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2649 MachinePointerInfo::getConstantPool(),
2650 false, false, false, 0);
2651 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2652 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2656 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
2658 SDValue Val = DAG.getConstant(0, MVT::i32);
2659 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl,
2660 DAG.getVTList(MVT::i32, MVT::Other), Op.getOperand(0),
2661 Op.getOperand(1), Val);
2665 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
2667 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
2668 Op.getOperand(1), DAG.getConstant(0, MVT::i32));
2672 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
2673 const ARMSubtarget *Subtarget) const {
2674 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2677 default: return SDValue(); // Don't custom lower most intrinsics.
2678 case Intrinsic::arm_rbit: {
2679 assert(Op.getOperand(1).getValueType() == MVT::i32 &&
2680 "RBIT intrinsic must have i32 type!");
2681 return DAG.getNode(ARMISD::RBIT, dl, MVT::i32, Op.getOperand(1));
2683 case Intrinsic::arm_thread_pointer: {
2684 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2685 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2687 case Intrinsic::eh_sjlj_lsda: {
2688 MachineFunction &MF = DAG.getMachineFunction();
2689 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2690 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2691 EVT PtrVT = getPointerTy();
2692 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2694 unsigned PCAdj = (RelocM != Reloc::PIC_)
2695 ? 0 : (Subtarget->isThumb() ? 4 : 8);
2696 ARMConstantPoolValue *CPV =
2697 ARMConstantPoolConstant::Create(MF.getFunction(), ARMPCLabelIndex,
2698 ARMCP::CPLSDA, PCAdj);
2699 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2700 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2702 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2703 MachinePointerInfo::getConstantPool(),
2704 false, false, false, 0);
2706 if (RelocM == Reloc::PIC_) {
2707 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2708 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2712 case Intrinsic::arm_neon_vmulls:
2713 case Intrinsic::arm_neon_vmullu: {
2714 unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vmulls)
2715 ? ARMISD::VMULLs : ARMISD::VMULLu;
2716 return DAG.getNode(NewOpc, SDLoc(Op), Op.getValueType(),
2717 Op.getOperand(1), Op.getOperand(2));
2722 static SDValue LowerATOMIC_FENCE(SDValue Op, SelectionDAG &DAG,
2723 const ARMSubtarget *Subtarget) {
2724 // FIXME: handle "fence singlethread" more efficiently.
2726 if (!Subtarget->hasDataBarrier()) {
2727 // Some ARMv6 cpus can support data barriers with an mcr instruction.
2728 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2730 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2731 "Unexpected ISD::ATOMIC_FENCE encountered. Should be libcall!");
2732 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2733 DAG.getConstant(0, MVT::i32));
2736 ConstantSDNode *OrdN = cast<ConstantSDNode>(Op.getOperand(1));
2737 AtomicOrdering Ord = static_cast<AtomicOrdering>(OrdN->getZExtValue());
2738 ARM_MB::MemBOpt Domain = ARM_MB::ISH;
2739 if (Subtarget->isMClass()) {
2740 // Only a full system barrier exists in the M-class architectures.
2741 Domain = ARM_MB::SY;
2742 } else if (Subtarget->isSwift() && Ord == Release) {
2743 // Swift happens to implement ISHST barriers in a way that's compatible with
2744 // Release semantics but weaker than ISH so we'd be fools not to use
2745 // it. Beware: other processors probably don't!
2746 Domain = ARM_MB::ISHST;
2749 return DAG.getNode(ISD::INTRINSIC_VOID, dl, MVT::Other, Op.getOperand(0),
2750 DAG.getConstant(Intrinsic::arm_dmb, MVT::i32),
2751 DAG.getConstant(Domain, MVT::i32));
2754 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG,
2755 const ARMSubtarget *Subtarget) {
2756 // ARM pre v5TE and Thumb1 does not have preload instructions.
2757 if (!(Subtarget->isThumb2() ||
2758 (!Subtarget->isThumb1Only() && Subtarget->hasV5TEOps())))
2759 // Just preserve the chain.
2760 return Op.getOperand(0);
2763 unsigned isRead = ~cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() & 1;
2765 (!Subtarget->hasV7Ops() || !Subtarget->hasMPExtension()))
2766 // ARMv7 with MP extension has PLDW.
2767 return Op.getOperand(0);
2769 unsigned isData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
2770 if (Subtarget->isThumb()) {
2772 isRead = ~isRead & 1;
2773 isData = ~isData & 1;
2776 return DAG.getNode(ARMISD::PRELOAD, dl, MVT::Other, Op.getOperand(0),
2777 Op.getOperand(1), DAG.getConstant(isRead, MVT::i32),
2778 DAG.getConstant(isData, MVT::i32));
2781 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
2782 MachineFunction &MF = DAG.getMachineFunction();
2783 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
2785 // vastart just stores the address of the VarArgsFrameIndex slot into the
2786 // memory location argument.
2788 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2789 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2790 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
2791 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
2792 MachinePointerInfo(SV), false, false, 0);
2796 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
2797 SDValue &Root, SelectionDAG &DAG,
2799 MachineFunction &MF = DAG.getMachineFunction();
2800 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2802 const TargetRegisterClass *RC;
2803 if (AFI->isThumb1OnlyFunction())
2804 RC = &ARM::tGPRRegClass;
2806 RC = &ARM::GPRRegClass;
2808 // Transform the arguments stored in physical registers into virtual ones.
2809 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2810 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2813 if (NextVA.isMemLoc()) {
2814 MachineFrameInfo *MFI = MF.getFrameInfo();
2815 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true);
2817 // Create load node to retrieve arguments from the stack.
2818 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2819 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
2820 MachinePointerInfo::getFixedStack(FI),
2821 false, false, false, 0);
2823 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
2824 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2826 if (!Subtarget->isLittle())
2827 std::swap (ArgValue, ArgValue2);
2828 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
2831 // The remaining GPRs hold either the beginning of variable-argument
2832 // data, or the beginning of an aggregate passed by value (usually
2833 // byval). Either way, we allocate stack slots adjacent to the data
2834 // provided by our caller, and store the unallocated registers there.
2835 // If this is a variadic function, the va_list pointer will begin with
2836 // these values; otherwise, this reassembles a (byval) structure that
2837 // was split between registers and memory.
2838 // Return: The frame index registers were stored into.
2840 ARMTargetLowering::StoreByValRegs(CCState &CCInfo, SelectionDAG &DAG,
2841 SDLoc dl, SDValue &Chain,
2842 const Value *OrigArg,
2843 unsigned InRegsParamRecordIdx,
2845 unsigned ArgSize) const {
2846 // Currently, two use-cases possible:
2847 // Case #1. Non-var-args function, and we meet first byval parameter.
2848 // Setup first unallocated register as first byval register;
2849 // eat all remained registers
2850 // (these two actions are performed by HandleByVal method).
2851 // Then, here, we initialize stack frame with
2852 // "store-reg" instructions.
2853 // Case #2. Var-args function, that doesn't contain byval parameters.
2854 // The same: eat all remained unallocated registers,
2855 // initialize stack frame.
2857 MachineFunction &MF = DAG.getMachineFunction();
2858 MachineFrameInfo *MFI = MF.getFrameInfo();
2859 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2860 unsigned RBegin, REnd;
2861 if (InRegsParamRecordIdx < CCInfo.getInRegsParamsCount()) {
2862 CCInfo.getInRegsParamInfo(InRegsParamRecordIdx, RBegin, REnd);
2864 unsigned RBeginIdx = CCInfo.getFirstUnallocated(GPRArgRegs);
2865 RBegin = RBeginIdx == 4 ? (unsigned)ARM::R4 : GPRArgRegs[RBeginIdx];
2870 ArgOffset = -4 * (ARM::R4 - RBegin);
2872 int FrameIndex = MFI->CreateFixedObject(ArgSize, ArgOffset, false);
2873 SDValue FIN = DAG.getFrameIndex(FrameIndex, getPointerTy());
2875 SmallVector<SDValue, 4> MemOps;
2876 const TargetRegisterClass *RC =
2877 AFI->isThumb1OnlyFunction() ? &ARM::tGPRRegClass : &ARM::GPRRegClass;
2879 for (unsigned Reg = RBegin, i = 0; Reg < REnd; ++Reg, ++i) {
2880 unsigned VReg = MF.addLiveIn(Reg, RC);
2881 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2883 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2884 MachinePointerInfo(OrigArg, 4 * i), false, false, 0);
2885 MemOps.push_back(Store);
2886 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
2887 DAG.getConstant(4, getPointerTy()));
2890 if (!MemOps.empty())
2891 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOps);
2895 // Setup stack frame, the va_list pointer will start from.
2897 ARMTargetLowering::VarArgStyleRegisters(CCState &CCInfo, SelectionDAG &DAG,
2898 SDLoc dl, SDValue &Chain,
2900 unsigned TotalArgRegsSaveSize,
2901 bool ForceMutable) const {
2902 MachineFunction &MF = DAG.getMachineFunction();
2903 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2905 // Try to store any remaining integer argument regs
2906 // to their spots on the stack so that they may be loaded by deferencing
2907 // the result of va_next.
2908 // If there is no regs to be stored, just point address after last
2909 // argument passed via stack.
2910 int FrameIndex = StoreByValRegs(CCInfo, DAG, dl, Chain, nullptr,
2911 CCInfo.getInRegsParamsCount(),
2912 CCInfo.getNextStackOffset(), 4);
2913 AFI->setVarArgsFrameIndex(FrameIndex);
2917 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
2918 CallingConv::ID CallConv, bool isVarArg,
2919 const SmallVectorImpl<ISD::InputArg>
2921 SDLoc dl, SelectionDAG &DAG,
2922 SmallVectorImpl<SDValue> &InVals)
2924 MachineFunction &MF = DAG.getMachineFunction();
2925 MachineFrameInfo *MFI = MF.getFrameInfo();
2927 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2929 // Assign locations to all of the incoming arguments.
2930 SmallVector<CCValAssign, 16> ArgLocs;
2931 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs,
2932 *DAG.getContext(), Prologue);
2933 CCInfo.AnalyzeFormalArguments(Ins,
2934 CCAssignFnForNode(CallConv, /* Return*/ false,
2937 SmallVector<SDValue, 16> ArgValues;
2939 Function::const_arg_iterator CurOrigArg = MF.getFunction()->arg_begin();
2940 unsigned CurArgIdx = 0;
2942 // Initially ArgRegsSaveSize is zero.
2943 // Then we increase this value each time we meet byval parameter.
2944 // We also increase this value in case of varargs function.
2945 AFI->setArgRegsSaveSize(0);
2947 // Calculate the amount of stack space that we need to allocate to store
2948 // byval and variadic arguments that are passed in registers.
2949 // We need to know this before we allocate the first byval or variadic
2950 // argument, as they will be allocated a stack slot below the CFA (Canonical
2951 // Frame Address, the stack pointer at entry to the function).
2952 unsigned ArgRegBegin = ARM::R4;
2953 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2954 if (CCInfo.getInRegsParamsProcessed() >= CCInfo.getInRegsParamsCount())
2957 CCValAssign &VA = ArgLocs[i];
2958 unsigned Index = VA.getValNo();
2959 ISD::ArgFlagsTy Flags = Ins[Index].Flags;
2960 if (!Flags.isByVal())
2963 assert(VA.isMemLoc() && "unexpected byval pointer in reg");
2964 unsigned RBegin, REnd;
2965 CCInfo.getInRegsParamInfo(CCInfo.getInRegsParamsProcessed(), RBegin, REnd);
2966 ArgRegBegin = std::min(ArgRegBegin, RBegin);
2968 CCInfo.nextInRegsParam();
2970 CCInfo.rewindByValRegsInfo();
2972 int lastInsIndex = -1;
2973 if (isVarArg && MFI->hasVAStart()) {
2974 unsigned RegIdx = CCInfo.getFirstUnallocated(GPRArgRegs);
2975 if (RegIdx != array_lengthof(GPRArgRegs))
2976 ArgRegBegin = std::min(ArgRegBegin, (unsigned)GPRArgRegs[RegIdx]);
2979 unsigned TotalArgRegsSaveSize = 4 * (ARM::R4 - ArgRegBegin);
2980 AFI->setArgRegsSaveSize(TotalArgRegsSaveSize);
2982 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2983 CCValAssign &VA = ArgLocs[i];
2984 if (Ins[VA.getValNo()].isOrigArg()) {
2985 std::advance(CurOrigArg,
2986 Ins[VA.getValNo()].getOrigArgIndex() - CurArgIdx);
2987 CurArgIdx = Ins[VA.getValNo()].getOrigArgIndex();
2989 // Arguments stored in registers.
2990 if (VA.isRegLoc()) {
2991 EVT RegVT = VA.getLocVT();
2993 if (VA.needsCustom()) {
2994 // f64 and vector types are split up into multiple registers or
2995 // combinations of registers and stack slots.
2996 if (VA.getLocVT() == MVT::v2f64) {
2997 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
2999 VA = ArgLocs[++i]; // skip ahead to next loc
3001 if (VA.isMemLoc()) {
3002 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true);
3003 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
3004 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
3005 MachinePointerInfo::getFixedStack(FI),
3006 false, false, false, 0);
3008 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
3011 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
3012 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
3013 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
3014 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
3015 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
3017 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
3020 const TargetRegisterClass *RC;
3022 if (RegVT == MVT::f32)
3023 RC = &ARM::SPRRegClass;
3024 else if (RegVT == MVT::f64)
3025 RC = &ARM::DPRRegClass;
3026 else if (RegVT == MVT::v2f64)
3027 RC = &ARM::QPRRegClass;
3028 else if (RegVT == MVT::i32)
3029 RC = AFI->isThumb1OnlyFunction() ? &ARM::tGPRRegClass
3030 : &ARM::GPRRegClass;
3032 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
3034 // Transform the arguments in physical registers into virtual ones.
3035 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
3036 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
3039 // If this is an 8 or 16-bit value, it is really passed promoted
3040 // to 32 bits. Insert an assert[sz]ext to capture this, then
3041 // truncate to the right size.
3042 switch (VA.getLocInfo()) {
3043 default: llvm_unreachable("Unknown loc info!");
3044 case CCValAssign::Full: break;
3045 case CCValAssign::BCvt:
3046 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
3048 case CCValAssign::SExt:
3049 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
3050 DAG.getValueType(VA.getValVT()));
3051 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
3053 case CCValAssign::ZExt:
3054 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
3055 DAG.getValueType(VA.getValVT()));
3056 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
3060 InVals.push_back(ArgValue);
3062 } else { // VA.isRegLoc()
3065 assert(VA.isMemLoc());
3066 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
3068 int index = VA.getValNo();
3070 // Some Ins[] entries become multiple ArgLoc[] entries.
3071 // Process them only once.
3072 if (index != lastInsIndex)
3074 ISD::ArgFlagsTy Flags = Ins[index].Flags;
3075 // FIXME: For now, all byval parameter objects are marked mutable.
3076 // This can be changed with more analysis.
3077 // In case of tail call optimization mark all arguments mutable.
3078 // Since they could be overwritten by lowering of arguments in case of
3080 if (Flags.isByVal()) {
3081 assert(Ins[index].isOrigArg() &&
3082 "Byval arguments cannot be implicit");
3083 unsigned CurByValIndex = CCInfo.getInRegsParamsProcessed();
3085 int FrameIndex = StoreByValRegs(CCInfo, DAG, dl, Chain, CurOrigArg,
3086 CurByValIndex, VA.getLocMemOffset(),
3087 Flags.getByValSize());
3088 InVals.push_back(DAG.getFrameIndex(FrameIndex, getPointerTy()));
3089 CCInfo.nextInRegsParam();
3091 unsigned FIOffset = VA.getLocMemOffset();
3092 int FI = MFI->CreateFixedObject(VA.getLocVT().getSizeInBits()/8,
3095 // Create load nodes to retrieve arguments from the stack.
3096 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
3097 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
3098 MachinePointerInfo::getFixedStack(FI),
3099 false, false, false, 0));
3101 lastInsIndex = index;
3107 if (isVarArg && MFI->hasVAStart())
3108 VarArgStyleRegisters(CCInfo, DAG, dl, Chain,
3109 CCInfo.getNextStackOffset(),
3110 TotalArgRegsSaveSize);
3112 AFI->setArgumentStackSize(CCInfo.getNextStackOffset());
3117 /// isFloatingPointZero - Return true if this is +0.0.
3118 static bool isFloatingPointZero(SDValue Op) {
3119 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
3120 return CFP->getValueAPF().isPosZero();
3121 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
3122 // Maybe this has already been legalized into the constant pool?
3123 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
3124 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
3125 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
3126 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
3127 return CFP->getValueAPF().isPosZero();
3129 } else if (Op->getOpcode() == ISD::BITCAST &&
3130 Op->getValueType(0) == MVT::f64) {
3131 // Handle (ISD::BITCAST (ARMISD::VMOVIMM (ISD::TargetConstant 0)) MVT::f64)
3132 // created by LowerConstantFP().
3133 SDValue BitcastOp = Op->getOperand(0);
3134 if (BitcastOp->getOpcode() == ARMISD::VMOVIMM) {
3135 SDValue MoveOp = BitcastOp->getOperand(0);
3136 if (MoveOp->getOpcode() == ISD::TargetConstant &&
3137 cast<ConstantSDNode>(MoveOp)->getZExtValue() == 0) {
3145 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
3146 /// the given operands.
3148 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
3149 SDValue &ARMcc, SelectionDAG &DAG,
3151 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
3152 unsigned C = RHSC->getZExtValue();
3153 if (!isLegalICmpImmediate(C)) {
3154 // Constant does not fit, try adjusting it by one?
3159 if (C != 0x80000000 && isLegalICmpImmediate(C-1)) {
3160 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
3161 RHS = DAG.getConstant(C-1, MVT::i32);
3166 if (C != 0 && isLegalICmpImmediate(C-1)) {
3167 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
3168 RHS = DAG.getConstant(C-1, MVT::i32);
3173 if (C != 0x7fffffff && isLegalICmpImmediate(C+1)) {
3174 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
3175 RHS = DAG.getConstant(C+1, MVT::i32);
3180 if (C != 0xffffffff && isLegalICmpImmediate(C+1)) {
3181 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
3182 RHS = DAG.getConstant(C+1, MVT::i32);
3189 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3190 ARMISD::NodeType CompareType;
3193 CompareType = ARMISD::CMP;
3198 CompareType = ARMISD::CMPZ;
3201 ARMcc = DAG.getConstant(CondCode, MVT::i32);
3202 return DAG.getNode(CompareType, dl, MVT::Glue, LHS, RHS);
3205 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
3207 ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
3209 assert(!Subtarget->isFPOnlySP() || RHS.getValueType() != MVT::f64);
3211 if (!isFloatingPointZero(RHS))
3212 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Glue, LHS, RHS);
3214 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Glue, LHS);
3215 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Glue, Cmp);
3218 /// duplicateCmp - Glue values can have only one use, so this function
3219 /// duplicates a comparison node.
3221 ARMTargetLowering::duplicateCmp(SDValue Cmp, SelectionDAG &DAG) const {
3222 unsigned Opc = Cmp.getOpcode();
3224 if (Opc == ARMISD::CMP || Opc == ARMISD::CMPZ)
3225 return DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
3227 assert(Opc == ARMISD::FMSTAT && "unexpected comparison operation");
3228 Cmp = Cmp.getOperand(0);
3229 Opc = Cmp.getOpcode();
3230 if (Opc == ARMISD::CMPFP)
3231 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
3233 assert(Opc == ARMISD::CMPFPw0 && "unexpected operand of FMSTAT");
3234 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0));
3236 return DAG.getNode(ARMISD::FMSTAT, DL, MVT::Glue, Cmp);
3239 std::pair<SDValue, SDValue>
3240 ARMTargetLowering::getARMXALUOOp(SDValue Op, SelectionDAG &DAG,
3241 SDValue &ARMcc) const {
3242 assert(Op.getValueType() == MVT::i32 && "Unsupported value type");
3244 SDValue Value, OverflowCmp;
3245 SDValue LHS = Op.getOperand(0);
3246 SDValue RHS = Op.getOperand(1);
3249 // FIXME: We are currently always generating CMPs because we don't support
3250 // generating CMN through the backend. This is not as good as the natural
3251 // CMP case because it causes a register dependency and cannot be folded
3254 switch (Op.getOpcode()) {
3256 llvm_unreachable("Unknown overflow instruction!");
3258 ARMcc = DAG.getConstant(ARMCC::VC, MVT::i32);
3259 Value = DAG.getNode(ISD::ADD, SDLoc(Op), Op.getValueType(), LHS, RHS);
3260 OverflowCmp = DAG.getNode(ARMISD::CMP, SDLoc(Op), MVT::Glue, Value, LHS);
3263 ARMcc = DAG.getConstant(ARMCC::HS, MVT::i32);
3264 Value = DAG.getNode(ISD::ADD, SDLoc(Op), Op.getValueType(), LHS, RHS);
3265 OverflowCmp = DAG.getNode(ARMISD::CMP, SDLoc(Op), MVT::Glue, Value, LHS);
3268 ARMcc = DAG.getConstant(ARMCC::VC, MVT::i32);
3269 Value = DAG.getNode(ISD::SUB, SDLoc(Op), Op.getValueType(), LHS, RHS);
3270 OverflowCmp = DAG.getNode(ARMISD::CMP, SDLoc(Op), MVT::Glue, LHS, RHS);
3273 ARMcc = DAG.getConstant(ARMCC::HS, MVT::i32);
3274 Value = DAG.getNode(ISD::SUB, SDLoc(Op), Op.getValueType(), LHS, RHS);
3275 OverflowCmp = DAG.getNode(ARMISD::CMP, SDLoc(Op), MVT::Glue, LHS, RHS);
3279 return std::make_pair(Value, OverflowCmp);
3284 ARMTargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) const {
3285 // Let legalize expand this if it isn't a legal type yet.
3286 if (!DAG.getTargetLoweringInfo().isTypeLegal(Op.getValueType()))
3289 SDValue Value, OverflowCmp;
3291 std::tie(Value, OverflowCmp) = getARMXALUOOp(Op, DAG, ARMcc);
3292 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3293 // We use 0 and 1 as false and true values.
3294 SDValue TVal = DAG.getConstant(1, MVT::i32);
3295 SDValue FVal = DAG.getConstant(0, MVT::i32);
3296 EVT VT = Op.getValueType();
3298 SDValue Overflow = DAG.getNode(ARMISD::CMOV, SDLoc(Op), VT, TVal, FVal,
3299 ARMcc, CCR, OverflowCmp);
3301 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
3302 return DAG.getNode(ISD::MERGE_VALUES, SDLoc(Op), VTs, Value, Overflow);
3306 SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
3307 SDValue Cond = Op.getOperand(0);
3308 SDValue SelectTrue = Op.getOperand(1);
3309 SDValue SelectFalse = Op.getOperand(2);
3311 unsigned Opc = Cond.getOpcode();
3313 if (Cond.getResNo() == 1 &&
3314 (Opc == ISD::SADDO || Opc == ISD::UADDO || Opc == ISD::SSUBO ||
3315 Opc == ISD::USUBO)) {
3316 if (!DAG.getTargetLoweringInfo().isTypeLegal(Cond->getValueType(0)))
3319 SDValue Value, OverflowCmp;
3321 std::tie(Value, OverflowCmp) = getARMXALUOOp(Cond, DAG, ARMcc);
3322 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3323 EVT VT = Op.getValueType();
3325 return getCMOV(SDLoc(Op), VT, SelectTrue, SelectFalse, ARMcc, CCR,
3331 // (select (cmov 1, 0, cond), t, f) -> (cmov t, f, cond)
3332 // (select (cmov 0, 1, cond), t, f) -> (cmov f, t, cond)
3334 if (Cond.getOpcode() == ARMISD::CMOV && Cond.hasOneUse()) {
3335 const ConstantSDNode *CMOVTrue =
3336 dyn_cast<ConstantSDNode>(Cond.getOperand(0));
3337 const ConstantSDNode *CMOVFalse =
3338 dyn_cast<ConstantSDNode>(Cond.getOperand(1));
3340 if (CMOVTrue && CMOVFalse) {
3341 unsigned CMOVTrueVal = CMOVTrue->getZExtValue();
3342 unsigned CMOVFalseVal = CMOVFalse->getZExtValue();
3346 if (CMOVTrueVal == 1 && CMOVFalseVal == 0) {
3348 False = SelectFalse;
3349 } else if (CMOVTrueVal == 0 && CMOVFalseVal == 1) {
3354 if (True.getNode() && False.getNode()) {
3355 EVT VT = Op.getValueType();
3356 SDValue ARMcc = Cond.getOperand(2);
3357 SDValue CCR = Cond.getOperand(3);
3358 SDValue Cmp = duplicateCmp(Cond.getOperand(4), DAG);
3359 assert(True.getValueType() == VT);
3360 return getCMOV(dl, VT, True, False, ARMcc, CCR, Cmp, DAG);
3365 // ARM's BooleanContents value is UndefinedBooleanContent. Mask out the
3366 // undefined bits before doing a full-word comparison with zero.
3367 Cond = DAG.getNode(ISD::AND, dl, Cond.getValueType(), Cond,
3368 DAG.getConstant(1, Cond.getValueType()));
3370 return DAG.getSelectCC(dl, Cond,
3371 DAG.getConstant(0, Cond.getValueType()),
3372 SelectTrue, SelectFalse, ISD::SETNE);
3375 static ISD::CondCode getInverseCCForVSEL(ISD::CondCode CC) {
3376 if (CC == ISD::SETNE)
3378 return ISD::getSetCCInverse(CC, true);
3381 static void checkVSELConstraints(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
3382 bool &swpCmpOps, bool &swpVselOps) {
3383 // Start by selecting the GE condition code for opcodes that return true for
3385 if (CC == ISD::SETUGE || CC == ISD::SETOGE || CC == ISD::SETOLE ||
3387 CondCode = ARMCC::GE;
3389 // and GT for opcodes that return false for 'equality'.
3390 else if (CC == ISD::SETUGT || CC == ISD::SETOGT || CC == ISD::SETOLT ||
3392 CondCode = ARMCC::GT;
3394 // Since we are constrained to GE/GT, if the opcode contains 'less', we need
3395 // to swap the compare operands.
3396 if (CC == ISD::SETOLE || CC == ISD::SETULE || CC == ISD::SETOLT ||
3400 // Both GT and GE are ordered comparisons, and return false for 'unordered'.
3401 // If we have an unordered opcode, we need to swap the operands to the VSEL
3402 // instruction (effectively negating the condition).
3404 // This also has the effect of swapping which one of 'less' or 'greater'
3405 // returns true, so we also swap the compare operands. It also switches
3406 // whether we return true for 'equality', so we compensate by picking the
3407 // opposite condition code to our original choice.
3408 if (CC == ISD::SETULE || CC == ISD::SETULT || CC == ISD::SETUGE ||
3409 CC == ISD::SETUGT) {
3410 swpCmpOps = !swpCmpOps;
3411 swpVselOps = !swpVselOps;
3412 CondCode = CondCode == ARMCC::GT ? ARMCC::GE : ARMCC::GT;
3415 // 'ordered' is 'anything but unordered', so use the VS condition code and
3416 // swap the VSEL operands.
3417 if (CC == ISD::SETO) {
3418 CondCode = ARMCC::VS;
3422 // 'unordered or not equal' is 'anything but equal', so use the EQ condition
3423 // code and swap the VSEL operands.
3424 if (CC == ISD::SETUNE) {
3425 CondCode = ARMCC::EQ;
3430 SDValue ARMTargetLowering::getCMOV(SDLoc dl, EVT VT, SDValue FalseVal,
3431 SDValue TrueVal, SDValue ARMcc, SDValue CCR,
3432 SDValue Cmp, SelectionDAG &DAG) const {
3433 if (Subtarget->isFPOnlySP() && VT == MVT::f64) {
3434 FalseVal = DAG.getNode(ARMISD::VMOVRRD, dl,
3435 DAG.getVTList(MVT::i32, MVT::i32), FalseVal);
3436 TrueVal = DAG.getNode(ARMISD::VMOVRRD, dl,
3437 DAG.getVTList(MVT::i32, MVT::i32), TrueVal);
3439 SDValue TrueLow = TrueVal.getValue(0);
3440 SDValue TrueHigh = TrueVal.getValue(1);
3441 SDValue FalseLow = FalseVal.getValue(0);
3442 SDValue FalseHigh = FalseVal.getValue(1);
3444 SDValue Low = DAG.getNode(ARMISD::CMOV, dl, MVT::i32, FalseLow, TrueLow,
3446 SDValue High = DAG.getNode(ARMISD::CMOV, dl, MVT::i32, FalseHigh, TrueHigh,
3447 ARMcc, CCR, duplicateCmp(Cmp, DAG));
3449 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Low, High);
3451 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,
3456 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
3457 EVT VT = Op.getValueType();
3458 SDValue LHS = Op.getOperand(0);
3459 SDValue RHS = Op.getOperand(1);
3460 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
3461 SDValue TrueVal = Op.getOperand(2);
3462 SDValue FalseVal = Op.getOperand(3);
3465 if (Subtarget->isFPOnlySP() && LHS.getValueType() == MVT::f64) {
3466 DAG.getTargetLoweringInfo().softenSetCCOperands(DAG, MVT::f64, LHS, RHS, CC,
3469 // If softenSetCCOperands only returned one value, we should compare it to
3471 if (!RHS.getNode()) {
3472 RHS = DAG.getConstant(0, LHS.getValueType());
3477 if (LHS.getValueType() == MVT::i32) {
3478 // Try to generate VSEL on ARMv8.
3479 // The VSEL instruction can't use all the usual ARM condition
3480 // codes: it only has two bits to select the condition code, so it's
3481 // constrained to use only GE, GT, VS and EQ.
3483 // To implement all the various ISD::SETXXX opcodes, we sometimes need to
3484 // swap the operands of the previous compare instruction (effectively
3485 // inverting the compare condition, swapping 'less' and 'greater') and
3486 // sometimes need to swap the operands to the VSEL (which inverts the
3487 // condition in the sense of firing whenever the previous condition didn't)
3488 if (Subtarget->hasFPARMv8() && (TrueVal.getValueType() == MVT::f32 ||
3489 TrueVal.getValueType() == MVT::f64)) {
3490 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3491 if (CondCode == ARMCC::LT || CondCode == ARMCC::LE ||
3492 CondCode == ARMCC::VC || CondCode == ARMCC::NE) {
3493 CC = getInverseCCForVSEL(CC);
3494 std::swap(TrueVal, FalseVal);
3499 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3500 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3501 return getCMOV(dl, VT, FalseVal, TrueVal, ARMcc, CCR, Cmp, DAG);
3504 ARMCC::CondCodes CondCode, CondCode2;
3505 FPCCToARMCC(CC, CondCode, CondCode2);
3507 // Try to generate VMAXNM/VMINNM on ARMv8.
3508 if (Subtarget->hasFPARMv8() && (TrueVal.getValueType() == MVT::f32 ||
3509 TrueVal.getValueType() == MVT::f64)) {
3510 // We can use VMAXNM/VMINNM for a compare followed by a select with the
3511 // same operands, as follows:
3512 // c = fcmp [?gt, ?ge, ?lt, ?le] a, b
3514 // In NoNaNsFPMath the CC will have been changed from, e.g., 'ogt' to 'gt'.
3515 // We only do this transformation in UnsafeFPMath and for no-NaNs
3516 // comparisons, because signed zeros and NaNs are handled differently than
3517 // the original code sequence.
3518 // FIXME: There are more cases that can be transformed even with NaNs,
3519 // signed zeroes and safe math. E.g. in the following, the result will be
3520 // FalseVal if a is a NaN or -0./0. and that's what vmaxnm will give, too.
3521 // c = fcmp ogt, a, 0. ; select c, a, 0. => vmaxnm a, 0.
3522 // FIXME: There is similar code that allows some extensions in
3523 // AArch64TargetLowering::LowerSELECT_CC that should be shared with this
3525 if (getTargetMachine().Options.UnsafeFPMath) {
3526 if (LHS == TrueVal && RHS == FalseVal) {
3527 if (CC == ISD::SETGT || CC == ISD::SETGE)
3528 return DAG.getNode(ARMISD::VMAXNM, dl, VT, TrueVal, FalseVal);
3529 if (CC == ISD::SETLT || CC == ISD::SETLE)
3530 return DAG.getNode(ARMISD::VMINNM, dl, VT, TrueVal, FalseVal);
3531 } else if (LHS == FalseVal && RHS == TrueVal) {
3532 if (CC == ISD::SETLT || CC == ISD::SETLE)
3533 return DAG.getNode(ARMISD::VMAXNM, dl, VT, TrueVal, FalseVal);
3534 if (CC == ISD::SETGT || CC == ISD::SETGE)
3535 return DAG.getNode(ARMISD::VMINNM, dl, VT, TrueVal, FalseVal);
3539 bool swpCmpOps = false;
3540 bool swpVselOps = false;
3541 checkVSELConstraints(CC, CondCode, swpCmpOps, swpVselOps);
3543 if (CondCode == ARMCC::GT || CondCode == ARMCC::GE ||
3544 CondCode == ARMCC::VS || CondCode == ARMCC::EQ) {
3546 std::swap(LHS, RHS);
3548 std::swap(TrueVal, FalseVal);
3552 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
3553 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
3554 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3555 SDValue Result = getCMOV(dl, VT, FalseVal, TrueVal, ARMcc, CCR, Cmp, DAG);
3556 if (CondCode2 != ARMCC::AL) {
3557 SDValue ARMcc2 = DAG.getConstant(CondCode2, MVT::i32);
3558 // FIXME: Needs another CMP because flag can have but one use.
3559 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
3560 Result = getCMOV(dl, VT, Result, TrueVal, ARMcc2, CCR, Cmp2, DAG);
3565 /// canChangeToInt - Given the fp compare operand, return true if it is suitable
3566 /// to morph to an integer compare sequence.
3567 static bool canChangeToInt(SDValue Op, bool &SeenZero,
3568 const ARMSubtarget *Subtarget) {
3569 SDNode *N = Op.getNode();
3570 if (!N->hasOneUse())
3571 // Otherwise it requires moving the value from fp to integer registers.
3573 if (!N->getNumValues())
3575 EVT VT = Op.getValueType();
3576 if (VT != MVT::f32 && !Subtarget->isFPBrccSlow())
3577 // f32 case is generally profitable. f64 case only makes sense when vcmpe +
3578 // vmrs are very slow, e.g. cortex-a8.
3581 if (isFloatingPointZero(Op)) {
3585 return ISD::isNormalLoad(N);
3588 static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) {
3589 if (isFloatingPointZero(Op))
3590 return DAG.getConstant(0, MVT::i32);
3592 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op))
3593 return DAG.getLoad(MVT::i32, SDLoc(Op),
3594 Ld->getChain(), Ld->getBasePtr(), Ld->getPointerInfo(),
3595 Ld->isVolatile(), Ld->isNonTemporal(),
3596 Ld->isInvariant(), Ld->getAlignment());
3598 llvm_unreachable("Unknown VFP cmp argument!");
3601 static void expandf64Toi32(SDValue Op, SelectionDAG &DAG,
3602 SDValue &RetVal1, SDValue &RetVal2) {
3603 if (isFloatingPointZero(Op)) {
3604 RetVal1 = DAG.getConstant(0, MVT::i32);
3605 RetVal2 = DAG.getConstant(0, MVT::i32);
3609 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) {
3610 SDValue Ptr = Ld->getBasePtr();
3611 RetVal1 = DAG.getLoad(MVT::i32, SDLoc(Op),
3612 Ld->getChain(), Ptr,
3613 Ld->getPointerInfo(),
3614 Ld->isVolatile(), Ld->isNonTemporal(),
3615 Ld->isInvariant(), Ld->getAlignment());
3617 EVT PtrType = Ptr.getValueType();
3618 unsigned NewAlign = MinAlign(Ld->getAlignment(), 4);
3619 SDValue NewPtr = DAG.getNode(ISD::ADD, SDLoc(Op),
3620 PtrType, Ptr, DAG.getConstant(4, PtrType));
3621 RetVal2 = DAG.getLoad(MVT::i32, SDLoc(Op),
3622 Ld->getChain(), NewPtr,
3623 Ld->getPointerInfo().getWithOffset(4),
3624 Ld->isVolatile(), Ld->isNonTemporal(),
3625 Ld->isInvariant(), NewAlign);
3629 llvm_unreachable("Unknown VFP cmp argument!");
3632 /// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some
3633 /// f32 and even f64 comparisons to integer ones.
3635 ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const {
3636 SDValue Chain = Op.getOperand(0);
3637 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3638 SDValue LHS = Op.getOperand(2);
3639 SDValue RHS = Op.getOperand(3);
3640 SDValue Dest = Op.getOperand(4);
3643 bool LHSSeenZero = false;
3644 bool LHSOk = canChangeToInt(LHS, LHSSeenZero, Subtarget);
3645 bool RHSSeenZero = false;
3646 bool RHSOk = canChangeToInt(RHS, RHSSeenZero, Subtarget);
3647 if (LHSOk && RHSOk && (LHSSeenZero || RHSSeenZero)) {
3648 // If unsafe fp math optimization is enabled and there are no other uses of
3649 // the CMP operands, and the condition code is EQ or NE, we can optimize it
3650 // to an integer comparison.
3651 if (CC == ISD::SETOEQ)
3653 else if (CC == ISD::SETUNE)
3656 SDValue Mask = DAG.getConstant(0x7fffffff, MVT::i32);
3658 if (LHS.getValueType() == MVT::f32) {
3659 LHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3660 bitcastf32Toi32(LHS, DAG), Mask);
3661 RHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3662 bitcastf32Toi32(RHS, DAG), Mask);
3663 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3664 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3665 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3666 Chain, Dest, ARMcc, CCR, Cmp);
3671 expandf64Toi32(LHS, DAG, LHS1, LHS2);
3672 expandf64Toi32(RHS, DAG, RHS1, RHS2);
3673 LHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, LHS2, Mask);
3674 RHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, RHS2, Mask);
3675 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3676 ARMcc = DAG.getConstant(CondCode, MVT::i32);
3677 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3678 SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest };
3679 return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops);
3685 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
3686 SDValue Chain = Op.getOperand(0);
3687 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3688 SDValue LHS = Op.getOperand(2);
3689 SDValue RHS = Op.getOperand(3);
3690 SDValue Dest = Op.getOperand(4);
3693 if (Subtarget->isFPOnlySP() && LHS.getValueType() == MVT::f64) {
3694 DAG.getTargetLoweringInfo().softenSetCCOperands(DAG, MVT::f64, LHS, RHS, CC,
3697 // If softenSetCCOperands only returned one value, we should compare it to
3699 if (!RHS.getNode()) {
3700 RHS = DAG.getConstant(0, LHS.getValueType());
3705 if (LHS.getValueType() == MVT::i32) {
3707 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3708 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3709 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3710 Chain, Dest, ARMcc, CCR, Cmp);
3713 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3715 if (getTargetMachine().Options.UnsafeFPMath &&
3716 (CC == ISD::SETEQ || CC == ISD::SETOEQ ||
3717 CC == ISD::SETNE || CC == ISD::SETUNE)) {
3718 SDValue Result = OptimizeVFPBrcond(Op, DAG);
3719 if (Result.getNode())
3723 ARMCC::CondCodes CondCode, CondCode2;
3724 FPCCToARMCC(CC, CondCode, CondCode2);
3726 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
3727 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
3728 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3729 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3730 SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp };
3731 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops);
3732 if (CondCode2 != ARMCC::AL) {
3733 ARMcc = DAG.getConstant(CondCode2, MVT::i32);
3734 SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) };
3735 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops);
3740 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
3741 SDValue Chain = Op.getOperand(0);
3742 SDValue Table = Op.getOperand(1);
3743 SDValue Index = Op.getOperand(2);
3746 EVT PTy = getPointerTy();
3747 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
3748 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
3749 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
3750 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
3751 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
3752 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
3753 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
3754 if (Subtarget->isThumb2()) {
3755 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
3756 // which does another jump to the destination. This also makes it easier
3757 // to translate it to TBB / TBH later.
3758 // FIXME: This might not work if the function is extremely large.
3759 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
3760 Addr, Op.getOperand(2), JTI, UId);
3762 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
3763 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
3764 MachinePointerInfo::getJumpTable(),
3765 false, false, false, 0);
3766 Chain = Addr.getValue(1);
3767 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
3768 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3770 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
3771 MachinePointerInfo::getJumpTable(),
3772 false, false, false, 0);
3773 Chain = Addr.getValue(1);
3774 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3778 static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3779 EVT VT = Op.getValueType();
3782 if (Op.getValueType().getVectorElementType() == MVT::i32) {
3783 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::f32)
3785 return DAG.UnrollVectorOp(Op.getNode());
3788 assert(Op.getOperand(0).getValueType() == MVT::v4f32 &&
3789 "Invalid type for custom lowering!");
3790 if (VT != MVT::v4i16)
3791 return DAG.UnrollVectorOp(Op.getNode());
3793 Op = DAG.getNode(Op.getOpcode(), dl, MVT::v4i32, Op.getOperand(0));
3794 return DAG.getNode(ISD::TRUNCATE, dl, VT, Op);
3797 SDValue ARMTargetLowering::LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) const {
3798 EVT VT = Op.getValueType();
3800 return LowerVectorFP_TO_INT(Op, DAG);
3801 if (Subtarget->isFPOnlySP() && Op.getOperand(0).getValueType() == MVT::f64) {
3803 if (Op.getOpcode() == ISD::FP_TO_SINT)
3804 LC = RTLIB::getFPTOSINT(Op.getOperand(0).getValueType(),
3807 LC = RTLIB::getFPTOUINT(Op.getOperand(0).getValueType(),
3809 return makeLibCall(DAG, LC, Op.getValueType(), &Op.getOperand(0), 1,
3810 /*isSigned*/ false, SDLoc(Op)).first;
3816 static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3817 EVT VT = Op.getValueType();
3820 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::i32) {
3821 if (VT.getVectorElementType() == MVT::f32)
3823 return DAG.UnrollVectorOp(Op.getNode());
3826 assert(Op.getOperand(0).getValueType() == MVT::v4i16 &&
3827 "Invalid type for custom lowering!");
3828 if (VT != MVT::v4f32)
3829 return DAG.UnrollVectorOp(Op.getNode());
3833 switch (Op.getOpcode()) {
3834 default: llvm_unreachable("Invalid opcode!");
3835 case ISD::SINT_TO_FP:
3836 CastOpc = ISD::SIGN_EXTEND;
3837 Opc = ISD::SINT_TO_FP;
3839 case ISD::UINT_TO_FP:
3840 CastOpc = ISD::ZERO_EXTEND;
3841 Opc = ISD::UINT_TO_FP;
3845 Op = DAG.getNode(CastOpc, dl, MVT::v4i32, Op.getOperand(0));
3846 return DAG.getNode(Opc, dl, VT, Op);
3849 SDValue ARMTargetLowering::LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) const {
3850 EVT VT = Op.getValueType();
3852 return LowerVectorINT_TO_FP(Op, DAG);
3853 if (Subtarget->isFPOnlySP() && Op.getValueType() == MVT::f64) {
3855 if (Op.getOpcode() == ISD::SINT_TO_FP)
3856 LC = RTLIB::getSINTTOFP(Op.getOperand(0).getValueType(),
3859 LC = RTLIB::getUINTTOFP(Op.getOperand(0).getValueType(),
3861 return makeLibCall(DAG, LC, Op.getValueType(), &Op.getOperand(0), 1,
3862 /*isSigned*/ false, SDLoc(Op)).first;
3868 SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
3869 // Implement fcopysign with a fabs and a conditional fneg.
3870 SDValue Tmp0 = Op.getOperand(0);
3871 SDValue Tmp1 = Op.getOperand(1);
3873 EVT VT = Op.getValueType();
3874 EVT SrcVT = Tmp1.getValueType();
3875 bool InGPR = Tmp0.getOpcode() == ISD::BITCAST ||
3876 Tmp0.getOpcode() == ARMISD::VMOVDRR;
3877 bool UseNEON = !InGPR && Subtarget->hasNEON();
3880 // Use VBSL to copy the sign bit.
3881 unsigned EncodedVal = ARM_AM::createNEONModImm(0x6, 0x80);
3882 SDValue Mask = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v2i32,
3883 DAG.getTargetConstant(EncodedVal, MVT::i32));
3884 EVT OpVT = (VT == MVT::f32) ? MVT::v2i32 : MVT::v1i64;
3886 Mask = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3887 DAG.getNode(ISD::BITCAST, dl, OpVT, Mask),
3888 DAG.getConstant(32, MVT::i32));
3889 else /*if (VT == MVT::f32)*/
3890 Tmp0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp0);
3891 if (SrcVT == MVT::f32) {
3892 Tmp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp1);
3894 Tmp1 = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3895 DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1),
3896 DAG.getConstant(32, MVT::i32));
3897 } else if (VT == MVT::f32)
3898 Tmp1 = DAG.getNode(ARMISD::VSHRu, dl, MVT::v1i64,
3899 DAG.getNode(ISD::BITCAST, dl, MVT::v1i64, Tmp1),
3900 DAG.getConstant(32, MVT::i32));
3901 Tmp0 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp0);
3902 Tmp1 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1);
3904 SDValue AllOnes = DAG.getTargetConstant(ARM_AM::createNEONModImm(0xe, 0xff),
3906 AllOnes = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v8i8, AllOnes);
3907 SDValue MaskNot = DAG.getNode(ISD::XOR, dl, OpVT, Mask,
3908 DAG.getNode(ISD::BITCAST, dl, OpVT, AllOnes));
3910 SDValue Res = DAG.getNode(ISD::OR, dl, OpVT,
3911 DAG.getNode(ISD::AND, dl, OpVT, Tmp1, Mask),
3912 DAG.getNode(ISD::AND, dl, OpVT, Tmp0, MaskNot));
3913 if (VT == MVT::f32) {
3914 Res = DAG.getNode(ISD::BITCAST, dl, MVT::v2f32, Res);
3915 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, Res,
3916 DAG.getConstant(0, MVT::i32));
3918 Res = DAG.getNode(ISD::BITCAST, dl, MVT::f64, Res);
3924 // Bitcast operand 1 to i32.
3925 if (SrcVT == MVT::f64)
3926 Tmp1 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
3928 Tmp1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp1);
3930 // Or in the signbit with integer operations.
3931 SDValue Mask1 = DAG.getConstant(0x80000000, MVT::i32);
3932 SDValue Mask2 = DAG.getConstant(0x7fffffff, MVT::i32);
3933 Tmp1 = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp1, Mask1);
3934 if (VT == MVT::f32) {
3935 Tmp0 = DAG.getNode(ISD::AND, dl, MVT::i32,
3936 DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp0), Mask2);
3937 return DAG.getNode(ISD::BITCAST, dl, MVT::f32,
3938 DAG.getNode(ISD::OR, dl, MVT::i32, Tmp0, Tmp1));
3941 // f64: Or the high part with signbit and then combine two parts.
3942 Tmp0 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
3944 SDValue Lo = Tmp0.getValue(0);
3945 SDValue Hi = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp0.getValue(1), Mask2);
3946 Hi = DAG.getNode(ISD::OR, dl, MVT::i32, Hi, Tmp1);
3947 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
3950 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
3951 MachineFunction &MF = DAG.getMachineFunction();
3952 MachineFrameInfo *MFI = MF.getFrameInfo();
3953 MFI->setReturnAddressIsTaken(true);
3955 if (verifyReturnAddressArgumentIsConstant(Op, DAG))
3958 EVT VT = Op.getValueType();
3960 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3962 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
3963 SDValue Offset = DAG.getConstant(4, MVT::i32);
3964 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
3965 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
3966 MachinePointerInfo(), false, false, false, 0);
3969 // Return LR, which contains the return address. Mark it an implicit live-in.
3970 unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32));
3971 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
3974 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
3975 const ARMBaseRegisterInfo &ARI =
3976 *static_cast<const ARMBaseRegisterInfo*>(RegInfo);
3977 MachineFunction &MF = DAG.getMachineFunction();
3978 MachineFrameInfo *MFI = MF.getFrameInfo();
3979 MFI->setFrameAddressIsTaken(true);
3981 EVT VT = Op.getValueType();
3982 SDLoc dl(Op); // FIXME probably not meaningful
3983 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3984 unsigned FrameReg = ARI.getFrameRegister(MF);
3985 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
3987 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
3988 MachinePointerInfo(),
3989 false, false, false, 0);
3993 // FIXME? Maybe this could be a TableGen attribute on some registers and
3994 // this table could be generated automatically from RegInfo.
3995 unsigned ARMTargetLowering::getRegisterByName(const char* RegName,
3997 unsigned Reg = StringSwitch<unsigned>(RegName)
3998 .Case("sp", ARM::SP)
4002 report_fatal_error("Invalid register name global variable");
4005 /// ExpandBITCAST - If the target supports VFP, this function is called to
4006 /// expand a bit convert where either the source or destination type is i64 to
4007 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
4008 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
4009 /// vectors), since the legalizer won't know what to do with that.
4010 static SDValue ExpandBITCAST(SDNode *N, SelectionDAG &DAG) {
4011 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4013 SDValue Op = N->getOperand(0);
4015 // This function is only supposed to be called for i64 types, either as the
4016 // source or destination of the bit convert.
4017 EVT SrcVT = Op.getValueType();
4018 EVT DstVT = N->getValueType(0);
4019 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
4020 "ExpandBITCAST called for non-i64 type");
4022 // Turn i64->f64 into VMOVDRR.
4023 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
4024 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
4025 DAG.getConstant(0, MVT::i32));
4026 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
4027 DAG.getConstant(1, MVT::i32));
4028 return DAG.getNode(ISD::BITCAST, dl, DstVT,
4029 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
4032 // Turn f64->i64 into VMOVRRD.
4033 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
4035 if (TLI.isBigEndian() && SrcVT.isVector() &&
4036 SrcVT.getVectorNumElements() > 1)
4037 Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
4038 DAG.getVTList(MVT::i32, MVT::i32),
4039 DAG.getNode(ARMISD::VREV64, dl, SrcVT, Op));
4041 Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
4042 DAG.getVTList(MVT::i32, MVT::i32), Op);
4043 // Merge the pieces into a single i64 value.
4044 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
4050 /// getZeroVector - Returns a vector of specified type with all zero elements.
4051 /// Zero vectors are used to represent vector negation and in those cases
4052 /// will be implemented with the NEON VNEG instruction. However, VNEG does
4053 /// not support i64 elements, so sometimes the zero vectors will need to be
4054 /// explicitly constructed. Regardless, use a canonical VMOV to create the
4056 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, SDLoc dl) {
4057 assert(VT.isVector() && "Expected a vector type");
4058 // The canonical modified immediate encoding of a zero vector is....0!
4059 SDValue EncodedVal = DAG.getTargetConstant(0, MVT::i32);
4060 EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
4061 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal);
4062 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4065 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
4066 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
4067 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
4068 SelectionDAG &DAG) const {
4069 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4070 EVT VT = Op.getValueType();
4071 unsigned VTBits = VT.getSizeInBits();
4073 SDValue ShOpLo = Op.getOperand(0);
4074 SDValue ShOpHi = Op.getOperand(1);
4075 SDValue ShAmt = Op.getOperand(2);
4077 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
4079 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
4081 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
4082 DAG.getConstant(VTBits, MVT::i32), ShAmt);
4083 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
4084 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
4085 DAG.getConstant(VTBits, MVT::i32));
4086 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
4087 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
4088 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
4090 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
4091 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
4093 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
4094 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc,
4097 SDValue Ops[2] = { Lo, Hi };
4098 return DAG.getMergeValues(Ops, dl);
4101 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
4102 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
4103 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
4104 SelectionDAG &DAG) const {
4105 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4106 EVT VT = Op.getValueType();
4107 unsigned VTBits = VT.getSizeInBits();
4109 SDValue ShOpLo = Op.getOperand(0);
4110 SDValue ShOpHi = Op.getOperand(1);
4111 SDValue ShAmt = Op.getOperand(2);
4114 assert(Op.getOpcode() == ISD::SHL_PARTS);
4115 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
4116 DAG.getConstant(VTBits, MVT::i32), ShAmt);
4117 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
4118 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
4119 DAG.getConstant(VTBits, MVT::i32));
4120 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
4121 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
4123 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
4124 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
4125 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
4127 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
4128 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc,
4131 SDValue Ops[2] = { Lo, Hi };
4132 return DAG.getMergeValues(Ops, dl);
4135 SDValue ARMTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
4136 SelectionDAG &DAG) const {
4137 // The rounding mode is in bits 23:22 of the FPSCR.
4138 // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0
4139 // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3)
4140 // so that the shift + and get folded into a bitfield extract.
4142 SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::i32,
4143 DAG.getConstant(Intrinsic::arm_get_fpscr,
4145 SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPSCR,
4146 DAG.getConstant(1U << 22, MVT::i32));
4147 SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds,
4148 DAG.getConstant(22, MVT::i32));
4149 return DAG.getNode(ISD::AND, dl, MVT::i32, RMODE,
4150 DAG.getConstant(3, MVT::i32));
4153 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
4154 const ARMSubtarget *ST) {
4155 EVT VT = N->getValueType(0);
4158 if (!ST->hasV6T2Ops())
4161 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
4162 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
4165 /// getCTPOP16BitCounts - Returns a v8i8/v16i8 vector containing the bit-count
4166 /// for each 16-bit element from operand, repeated. The basic idea is to
4167 /// leverage vcnt to get the 8-bit counts, gather and add the results.
4169 /// Trace for v4i16:
4170 /// input = [v0 v1 v2 v3 ] (vi 16-bit element)
4171 /// cast: N0 = [w0 w1 w2 w3 w4 w5 w6 w7] (v0 = [w0 w1], wi 8-bit element)
4172 /// vcnt: N1 = [b0 b1 b2 b3 b4 b5 b6 b7] (bi = bit-count of 8-bit element wi)
4173 /// vrev: N2 = [b1 b0 b3 b2 b5 b4 b7 b6]
4174 /// [b0 b1 b2 b3 b4 b5 b6 b7]
4175 /// +[b1 b0 b3 b2 b5 b4 b7 b6]
4176 /// N3=N1+N2 = [k0 k0 k1 k1 k2 k2 k3 k3] (k0 = b0+b1 = bit-count of 16-bit v0,
4177 /// vuzp: = [k0 k1 k2 k3 k0 k1 k2 k3] each ki is 8-bits)
4178 static SDValue getCTPOP16BitCounts(SDNode *N, SelectionDAG &DAG) {
4179 EVT VT = N->getValueType(0);
4182 EVT VT8Bit = VT.is64BitVector() ? MVT::v8i8 : MVT::v16i8;
4183 SDValue N0 = DAG.getNode(ISD::BITCAST, DL, VT8Bit, N->getOperand(0));
4184 SDValue N1 = DAG.getNode(ISD::CTPOP, DL, VT8Bit, N0);
4185 SDValue N2 = DAG.getNode(ARMISD::VREV16, DL, VT8Bit, N1);
4186 SDValue N3 = DAG.getNode(ISD::ADD, DL, VT8Bit, N1, N2);
4187 return DAG.getNode(ARMISD::VUZP, DL, VT8Bit, N3, N3);
4190 /// lowerCTPOP16BitElements - Returns a v4i16/v8i16 vector containing the
4191 /// bit-count for each 16-bit element from the operand. We need slightly
4192 /// different sequencing for v4i16 and v8i16 to stay within NEON's available
4193 /// 64/128-bit registers.
4195 /// Trace for v4i16:
4196 /// input = [v0 v1 v2 v3 ] (vi 16-bit element)
4197 /// v8i8: BitCounts = [k0 k1 k2 k3 k0 k1 k2 k3 ] (ki is the bit-count of vi)
4198 /// v8i16:Extended = [k0 k1 k2 k3 k0 k1 k2 k3 ]
4199 /// v4i16:Extracted = [k0 k1 k2 k3 ]
4200 static SDValue lowerCTPOP16BitElements(SDNode *N, SelectionDAG &DAG) {
4201 EVT VT = N->getValueType(0);
4204 SDValue BitCounts = getCTPOP16BitCounts(N, DAG);
4205 if (VT.is64BitVector()) {
4206 SDValue Extended = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v8i16, BitCounts);
4207 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i16, Extended,
4208 DAG.getIntPtrConstant(0));
4210 SDValue Extracted = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v8i8,
4211 BitCounts, DAG.getIntPtrConstant(0));
4212 return DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v8i16, Extracted);
4216 /// lowerCTPOP32BitElements - Returns a v2i32/v4i32 vector containing the
4217 /// bit-count for each 32-bit element from the operand. The idea here is
4218 /// to split the vector into 16-bit elements, leverage the 16-bit count
4219 /// routine, and then combine the results.
4221 /// Trace for v2i32 (v4i32 similar with Extracted/Extended exchanged):
4222 /// input = [v0 v1 ] (vi: 32-bit elements)
4223 /// Bitcast = [w0 w1 w2 w3 ] (wi: 16-bit elements, v0 = [w0 w1])
4224 /// Counts16 = [k0 k1 k2 k3 ] (ki: 16-bit elements, bit-count of wi)
4225 /// vrev: N0 = [k1 k0 k3 k2 ]
4227 /// N1 =+[k1 k0 k3 k2 ]
4229 /// N2 =+[k1 k3 k0 k2 ]
4231 /// Extended =+[k1 k3 k0 k2 ]
4233 /// Extracted=+[k1 k3 ]
4235 static SDValue lowerCTPOP32BitElements(SDNode *N, SelectionDAG &DAG) {
4236 EVT VT = N->getValueType(0);
4239 EVT VT16Bit = VT.is64BitVector() ? MVT::v4i16 : MVT::v8i16;
4241 SDValue Bitcast = DAG.getNode(ISD::BITCAST, DL, VT16Bit, N->getOperand(0));
4242 SDValue Counts16 = lowerCTPOP16BitElements(Bitcast.getNode(), DAG);
4243 SDValue N0 = DAG.getNode(ARMISD::VREV32, DL, VT16Bit, Counts16);
4244 SDValue N1 = DAG.getNode(ISD::ADD, DL, VT16Bit, Counts16, N0);
4245 SDValue N2 = DAG.getNode(ARMISD::VUZP, DL, VT16Bit, N1, N1);
4247 if (VT.is64BitVector()) {
4248 SDValue Extended = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v4i32, N2);
4249 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i32, Extended,
4250 DAG.getIntPtrConstant(0));
4252 SDValue Extracted = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i16, N2,
4253 DAG.getIntPtrConstant(0));
4254 return DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v4i32, Extracted);
4258 static SDValue LowerCTPOP(SDNode *N, SelectionDAG &DAG,
4259 const ARMSubtarget *ST) {
4260 EVT VT = N->getValueType(0);
4262 assert(ST->hasNEON() && "Custom ctpop lowering requires NEON.");
4263 assert((VT == MVT::v2i32 || VT == MVT::v4i32 ||
4264 VT == MVT::v4i16 || VT == MVT::v8i16) &&
4265 "Unexpected type for custom ctpop lowering");
4267 if (VT.getVectorElementType() == MVT::i32)
4268 return lowerCTPOP32BitElements(N, DAG);
4270 return lowerCTPOP16BitElements(N, DAG);
4273 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
4274 const ARMSubtarget *ST) {
4275 EVT VT = N->getValueType(0);
4281 // Lower vector shifts on NEON to use VSHL.
4282 assert(ST->hasNEON() && "unexpected vector shift");
4284 // Left shifts translate directly to the vshiftu intrinsic.
4285 if (N->getOpcode() == ISD::SHL)
4286 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
4287 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
4288 N->getOperand(0), N->getOperand(1));
4290 assert((N->getOpcode() == ISD::SRA ||
4291 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
4293 // NEON uses the same intrinsics for both left and right shifts. For
4294 // right shifts, the shift amounts are negative, so negate the vector of
4296 EVT ShiftVT = N->getOperand(1).getValueType();
4297 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
4298 getZeroVector(ShiftVT, DAG, dl),
4300 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
4301 Intrinsic::arm_neon_vshifts :
4302 Intrinsic::arm_neon_vshiftu);
4303 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
4304 DAG.getConstant(vshiftInt, MVT::i32),
4305 N->getOperand(0), NegatedCount);
4308 static SDValue Expand64BitShift(SDNode *N, SelectionDAG &DAG,
4309 const ARMSubtarget *ST) {
4310 EVT VT = N->getValueType(0);
4313 // We can get here for a node like i32 = ISD::SHL i32, i64
4317 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
4318 "Unknown shift to lower!");
4320 // We only lower SRA, SRL of 1 here, all others use generic lowering.
4321 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
4322 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
4325 // If we are in thumb mode, we don't have RRX.
4326 if (ST->isThumb1Only()) return SDValue();
4328 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
4329 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
4330 DAG.getConstant(0, MVT::i32));
4331 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
4332 DAG.getConstant(1, MVT::i32));
4334 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
4335 // captures the result into a carry flag.
4336 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
4337 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Glue), Hi);
4339 // The low part is an ARMISD::RRX operand, which shifts the carry in.
4340 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
4342 // Merge the pieces into a single i64 value.
4343 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
4346 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
4347 SDValue TmpOp0, TmpOp1;
4348 bool Invert = false;
4352 SDValue Op0 = Op.getOperand(0);
4353 SDValue Op1 = Op.getOperand(1);
4354 SDValue CC = Op.getOperand(2);
4355 EVT CmpVT = Op0.getValueType().changeVectorElementTypeToInteger();
4356 EVT VT = Op.getValueType();
4357 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
4360 if (Op1.getValueType().isFloatingPoint()) {
4361 switch (SetCCOpcode) {
4362 default: llvm_unreachable("Illegal FP comparison");
4364 case ISD::SETNE: Invert = true; // Fallthrough
4366 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
4368 case ISD::SETLT: Swap = true; // Fallthrough
4370 case ISD::SETGT: Opc = ARMISD::VCGT; break;
4372 case ISD::SETLE: Swap = true; // Fallthrough
4374 case ISD::SETGE: Opc = ARMISD::VCGE; break;
4375 case ISD::SETUGE: Swap = true; // Fallthrough
4376 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
4377 case ISD::SETUGT: Swap = true; // Fallthrough
4378 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
4379 case ISD::SETUEQ: Invert = true; // Fallthrough
4381 // Expand this to (OLT | OGT).
4385 Op0 = DAG.getNode(ARMISD::VCGT, dl, CmpVT, TmpOp1, TmpOp0);
4386 Op1 = DAG.getNode(ARMISD::VCGT, dl, CmpVT, TmpOp0, TmpOp1);
4388 case ISD::SETUO: Invert = true; // Fallthrough
4390 // Expand this to (OLT | OGE).
4394 Op0 = DAG.getNode(ARMISD::VCGT, dl, CmpVT, TmpOp1, TmpOp0);
4395 Op1 = DAG.getNode(ARMISD::VCGE, dl, CmpVT, TmpOp0, TmpOp1);
4399 // Integer comparisons.
4400 switch (SetCCOpcode) {
4401 default: llvm_unreachable("Illegal integer comparison");
4402 case ISD::SETNE: Invert = true;
4403 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
4404 case ISD::SETLT: Swap = true;
4405 case ISD::SETGT: Opc = ARMISD::VCGT; break;
4406 case ISD::SETLE: Swap = true;
4407 case ISD::SETGE: Opc = ARMISD::VCGE; break;
4408 case ISD::SETULT: Swap = true;
4409 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
4410 case ISD::SETULE: Swap = true;
4411 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
4414 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
4415 if (Opc == ARMISD::VCEQ) {
4418 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
4420 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
4423 // Ignore bitconvert.
4424 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BITCAST)
4425 AndOp = AndOp.getOperand(0);
4427 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
4429 Op0 = DAG.getNode(ISD::BITCAST, dl, CmpVT, AndOp.getOperand(0));
4430 Op1 = DAG.getNode(ISD::BITCAST, dl, CmpVT, AndOp.getOperand(1));
4437 std::swap(Op0, Op1);
4439 // If one of the operands is a constant vector zero, attempt to fold the
4440 // comparison to a specialized compare-against-zero form.
4442 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
4444 else if (ISD::isBuildVectorAllZeros(Op0.getNode())) {
4445 if (Opc == ARMISD::VCGE)
4446 Opc = ARMISD::VCLEZ;
4447 else if (Opc == ARMISD::VCGT)
4448 Opc = ARMISD::VCLTZ;
4453 if (SingleOp.getNode()) {
4456 Result = DAG.getNode(ARMISD::VCEQZ, dl, CmpVT, SingleOp); break;
4458 Result = DAG.getNode(ARMISD::VCGEZ, dl, CmpVT, SingleOp); break;
4460 Result = DAG.getNode(ARMISD::VCLEZ, dl, CmpVT, SingleOp); break;
4462 Result = DAG.getNode(ARMISD::VCGTZ, dl, CmpVT, SingleOp); break;
4464 Result = DAG.getNode(ARMISD::VCLTZ, dl, CmpVT, SingleOp); break;
4466 Result = DAG.getNode(Opc, dl, CmpVT, Op0, Op1);
4469 Result = DAG.getNode(Opc, dl, CmpVT, Op0, Op1);
4472 Result = DAG.getSExtOrTrunc(Result, dl, VT);
4475 Result = DAG.getNOT(dl, Result, VT);
4480 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
4481 /// valid vector constant for a NEON instruction with a "modified immediate"
4482 /// operand (e.g., VMOV). If so, return the encoded value.
4483 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
4484 unsigned SplatBitSize, SelectionDAG &DAG,
4485 EVT &VT, bool is128Bits, NEONModImmType type) {
4486 unsigned OpCmode, Imm;
4488 // SplatBitSize is set to the smallest size that splats the vector, so a
4489 // zero vector will always have SplatBitSize == 8. However, NEON modified
4490 // immediate instructions others than VMOV do not support the 8-bit encoding
4491 // of a zero vector, and the default encoding of zero is supposed to be the
4496 switch (SplatBitSize) {
4498 if (type != VMOVModImm)
4500 // Any 1-byte value is OK. Op=0, Cmode=1110.
4501 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
4504 VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
4508 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
4509 VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
4510 if ((SplatBits & ~0xff) == 0) {
4511 // Value = 0x00nn: Op=x, Cmode=100x.
4516 if ((SplatBits & ~0xff00) == 0) {
4517 // Value = 0xnn00: Op=x, Cmode=101x.
4519 Imm = SplatBits >> 8;
4525 // NEON's 32-bit VMOV supports splat values where:
4526 // * only one byte is nonzero, or
4527 // * the least significant byte is 0xff and the second byte is nonzero, or
4528 // * the least significant 2 bytes are 0xff and the third is nonzero.
4529 VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
4530 if ((SplatBits & ~0xff) == 0) {
4531 // Value = 0x000000nn: Op=x, Cmode=000x.
4536 if ((SplatBits & ~0xff00) == 0) {
4537 // Value = 0x0000nn00: Op=x, Cmode=001x.
4539 Imm = SplatBits >> 8;
4542 if ((SplatBits & ~0xff0000) == 0) {
4543 // Value = 0x00nn0000: Op=x, Cmode=010x.
4545 Imm = SplatBits >> 16;
4548 if ((SplatBits & ~0xff000000) == 0) {
4549 // Value = 0xnn000000: Op=x, Cmode=011x.
4551 Imm = SplatBits >> 24;
4555 // cmode == 0b1100 and cmode == 0b1101 are not supported for VORR or VBIC
4556 if (type == OtherModImm) return SDValue();
4558 if ((SplatBits & ~0xffff) == 0 &&
4559 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
4560 // Value = 0x0000nnff: Op=x, Cmode=1100.
4562 Imm = SplatBits >> 8;
4566 if ((SplatBits & ~0xffffff) == 0 &&
4567 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
4568 // Value = 0x00nnffff: Op=x, Cmode=1101.
4570 Imm = SplatBits >> 16;
4574 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
4575 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
4576 // VMOV.I32. A (very) minor optimization would be to replicate the value
4577 // and fall through here to test for a valid 64-bit splat. But, then the
4578 // caller would also need to check and handle the change in size.
4582 if (type != VMOVModImm)
4584 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
4585 uint64_t BitMask = 0xff;
4587 unsigned ImmMask = 1;
4589 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
4590 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
4593 } else if ((SplatBits & BitMask) != 0) {
4600 if (DAG.getTargetLoweringInfo().isBigEndian())
4601 // swap higher and lower 32 bit word
4602 Imm = ((Imm & 0xf) << 4) | ((Imm & 0xf0) >> 4);
4604 // Op=1, Cmode=1110.
4606 VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
4611 llvm_unreachable("unexpected size for isNEONModifiedImm");
4614 unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm);
4615 return DAG.getTargetConstant(EncodedVal, MVT::i32);
4618 SDValue ARMTargetLowering::LowerConstantFP(SDValue Op, SelectionDAG &DAG,
4619 const ARMSubtarget *ST) const {
4623 bool IsDouble = Op.getValueType() == MVT::f64;
4624 ConstantFPSDNode *CFP = cast<ConstantFPSDNode>(Op);
4626 // Use the default (constant pool) lowering for double constants when we have
4628 if (IsDouble && Subtarget->isFPOnlySP())
4631 // Try splatting with a VMOV.f32...
4632 APFloat FPVal = CFP->getValueAPF();
4633 int ImmVal = IsDouble ? ARM_AM::getFP64Imm(FPVal) : ARM_AM::getFP32Imm(FPVal);
4636 if (IsDouble || !ST->useNEONForSinglePrecisionFP()) {
4637 // We have code in place to select a valid ConstantFP already, no need to
4642 // It's a float and we are trying to use NEON operations where
4643 // possible. Lower it to a splat followed by an extract.
4645 SDValue NewVal = DAG.getTargetConstant(ImmVal, MVT::i32);
4646 SDValue VecConstant = DAG.getNode(ARMISD::VMOVFPIMM, DL, MVT::v2f32,
4648 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecConstant,
4649 DAG.getConstant(0, MVT::i32));
4652 // The rest of our options are NEON only, make sure that's allowed before
4654 if (!ST->hasNEON() || (!IsDouble && !ST->useNEONForSinglePrecisionFP()))
4658 uint64_t iVal = FPVal.bitcastToAPInt().getZExtValue();
4660 // It wouldn't really be worth bothering for doubles except for one very
4661 // important value, which does happen to match: 0.0. So make sure we don't do
4663 if (IsDouble && (iVal & 0xffffffff) != (iVal >> 32))
4666 // Try a VMOV.i32 (FIXME: i8, i16, or i64 could work too).
4667 SDValue NewVal = isNEONModifiedImm(iVal & 0xffffffffU, 0, 32, DAG, VMovVT,
4669 if (NewVal != SDValue()) {
4671 SDValue VecConstant = DAG.getNode(ARMISD::VMOVIMM, DL, VMovVT,
4674 return DAG.getNode(ISD::BITCAST, DL, MVT::f64, VecConstant);
4676 // It's a float: cast and extract a vector element.
4677 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
4679 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
4680 DAG.getConstant(0, MVT::i32));
4683 // Finally, try a VMVN.i32
4684 NewVal = isNEONModifiedImm(~iVal & 0xffffffffU, 0, 32, DAG, VMovVT,
4686 if (NewVal != SDValue()) {
4688 SDValue VecConstant = DAG.getNode(ARMISD::VMVNIMM, DL, VMovVT, NewVal);
4691 return DAG.getNode(ISD::BITCAST, DL, MVT::f64, VecConstant);
4693 // It's a float: cast and extract a vector element.
4694 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
4696 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
4697 DAG.getConstant(0, MVT::i32));
4703 // check if an VEXT instruction can handle the shuffle mask when the
4704 // vector sources of the shuffle are the same.
4705 static bool isSingletonVEXTMask(ArrayRef<int> M, EVT VT, unsigned &Imm) {
4706 unsigned NumElts = VT.getVectorNumElements();
4708 // Assume that the first shuffle index is not UNDEF. Fail if it is.
4714 // If this is a VEXT shuffle, the immediate value is the index of the first
4715 // element. The other shuffle indices must be the successive elements after
4717 unsigned ExpectedElt = Imm;
4718 for (unsigned i = 1; i < NumElts; ++i) {
4719 // Increment the expected index. If it wraps around, just follow it
4720 // back to index zero and keep going.
4722 if (ExpectedElt == NumElts)
4725 if (M[i] < 0) continue; // ignore UNDEF indices
4726 if (ExpectedElt != static_cast<unsigned>(M[i]))
4734 static bool isVEXTMask(ArrayRef<int> M, EVT VT,
4735 bool &ReverseVEXT, unsigned &Imm) {
4736 unsigned NumElts = VT.getVectorNumElements();
4737 ReverseVEXT = false;
4739 // Assume that the first shuffle index is not UNDEF. Fail if it is.
4745 // If this is a VEXT shuffle, the immediate value is the index of the first
4746 // element. The other shuffle indices must be the successive elements after
4748 unsigned ExpectedElt = Imm;
4749 for (unsigned i = 1; i < NumElts; ++i) {
4750 // Increment the expected index. If it wraps around, it may still be
4751 // a VEXT but the source vectors must be swapped.
4753 if (ExpectedElt == NumElts * 2) {
4758 if (M[i] < 0) continue; // ignore UNDEF indices
4759 if (ExpectedElt != static_cast<unsigned>(M[i]))
4763 // Adjust the index value if the source operands will be swapped.
4770 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
4771 /// instruction with the specified blocksize. (The order of the elements
4772 /// within each block of the vector is reversed.)
4773 static bool isVREVMask(ArrayRef<int> M, EVT VT, unsigned BlockSize) {
4774 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
4775 "Only possible block sizes for VREV are: 16, 32, 64");
4777 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4781 unsigned NumElts = VT.getVectorNumElements();
4782 unsigned BlockElts = M[0] + 1;
4783 // If the first shuffle index is UNDEF, be optimistic.
4785 BlockElts = BlockSize / EltSz;
4787 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
4790 for (unsigned i = 0; i < NumElts; ++i) {
4791 if (M[i] < 0) continue; // ignore UNDEF indices
4792 if ((unsigned) M[i] != (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
4799 static bool isVTBLMask(ArrayRef<int> M, EVT VT) {
4800 // We can handle <8 x i8> vector shuffles. If the index in the mask is out of
4801 // range, then 0 is placed into the resulting vector. So pretty much any mask
4802 // of 8 elements can work here.
4803 return VT == MVT::v8i8 && M.size() == 8;
4806 static bool isVTRNMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4807 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4811 unsigned NumElts = VT.getVectorNumElements();
4812 WhichResult = (M[0] == 0 ? 0 : 1);
4813 for (unsigned i = 0; i < NumElts; i += 2) {
4814 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
4815 (M[i+1] >= 0 && (unsigned) M[i+1] != i + NumElts + WhichResult))
4821 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
4822 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4823 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
4824 static bool isVTRN_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4825 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4829 unsigned NumElts = VT.getVectorNumElements();
4830 WhichResult = (M[0] == 0 ? 0 : 1);
4831 for (unsigned i = 0; i < NumElts; i += 2) {
4832 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
4833 (M[i+1] >= 0 && (unsigned) M[i+1] != i + WhichResult))
4839 static bool isVUZPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4840 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4844 unsigned NumElts = VT.getVectorNumElements();
4845 WhichResult = (M[0] == 0 ? 0 : 1);
4846 for (unsigned i = 0; i != NumElts; ++i) {
4847 if (M[i] < 0) continue; // ignore UNDEF indices
4848 if ((unsigned) M[i] != 2 * i + WhichResult)
4852 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4853 if (VT.is64BitVector() && EltSz == 32)
4859 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
4860 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4861 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
4862 static bool isVUZP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4863 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4867 unsigned Half = VT.getVectorNumElements() / 2;
4868 WhichResult = (M[0] == 0 ? 0 : 1);
4869 for (unsigned j = 0; j != 2; ++j) {
4870 unsigned Idx = WhichResult;
4871 for (unsigned i = 0; i != Half; ++i) {
4872 int MIdx = M[i + j * Half];
4873 if (MIdx >= 0 && (unsigned) MIdx != Idx)
4879 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4880 if (VT.is64BitVector() && EltSz == 32)
4886 static bool isVZIPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4887 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4891 unsigned NumElts = VT.getVectorNumElements();
4892 WhichResult = (M[0] == 0 ? 0 : 1);
4893 unsigned Idx = WhichResult * NumElts / 2;
4894 for (unsigned i = 0; i != NumElts; i += 2) {
4895 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
4896 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx + NumElts))
4901 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4902 if (VT.is64BitVector() && EltSz == 32)
4908 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
4909 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4910 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
4911 static bool isVZIP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4912 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4916 unsigned NumElts = VT.getVectorNumElements();
4917 WhichResult = (M[0] == 0 ? 0 : 1);
4918 unsigned Idx = WhichResult * NumElts / 2;
4919 for (unsigned i = 0; i != NumElts; i += 2) {
4920 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
4921 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx))
4926 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4927 if (VT.is64BitVector() && EltSz == 32)
4933 /// \return true if this is a reverse operation on an vector.
4934 static bool isReverseMask(ArrayRef<int> M, EVT VT) {
4935 unsigned NumElts = VT.getVectorNumElements();
4936 // Make sure the mask has the right size.
4937 if (NumElts != M.size())
4940 // Look for <15, ..., 3, -1, 1, 0>.
4941 for (unsigned i = 0; i != NumElts; ++i)
4942 if (M[i] >= 0 && M[i] != (int) (NumElts - 1 - i))
4948 // If N is an integer constant that can be moved into a register in one
4949 // instruction, return an SDValue of such a constant (will become a MOV
4950 // instruction). Otherwise return null.
4951 static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG,
4952 const ARMSubtarget *ST, SDLoc dl) {
4954 if (!isa<ConstantSDNode>(N))
4956 Val = cast<ConstantSDNode>(N)->getZExtValue();
4958 if (ST->isThumb1Only()) {
4959 if (Val <= 255 || ~Val <= 255)
4960 return DAG.getConstant(Val, MVT::i32);
4962 if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1)
4963 return DAG.getConstant(Val, MVT::i32);
4968 // If this is a case we can't handle, return null and let the default
4969 // expansion code take care of it.
4970 SDValue ARMTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
4971 const ARMSubtarget *ST) const {
4972 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
4974 EVT VT = Op.getValueType();
4976 APInt SplatBits, SplatUndef;
4977 unsigned SplatBitSize;
4979 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
4980 if (SplatBitSize <= 64) {
4981 // Check if an immediate VMOV works.
4983 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
4984 SplatUndef.getZExtValue(), SplatBitSize,
4985 DAG, VmovVT, VT.is128BitVector(),
4987 if (Val.getNode()) {
4988 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
4989 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4992 // Try an immediate VMVN.
4993 uint64_t NegatedImm = (~SplatBits).getZExtValue();
4994 Val = isNEONModifiedImm(NegatedImm,
4995 SplatUndef.getZExtValue(), SplatBitSize,
4996 DAG, VmovVT, VT.is128BitVector(),
4998 if (Val.getNode()) {
4999 SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
5000 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
5003 // Use vmov.f32 to materialize other v2f32 and v4f32 splats.
5004 if ((VT == MVT::v2f32 || VT == MVT::v4f32) && SplatBitSize == 32) {
5005 int ImmVal = ARM_AM::getFP32Imm(SplatBits);
5007 SDValue Val = DAG.getTargetConstant(ImmVal, MVT::i32);
5008 return DAG.getNode(ARMISD::VMOVFPIMM, dl, VT, Val);
5014 // Scan through the operands to see if only one value is used.
5016 // As an optimisation, even if more than one value is used it may be more
5017 // profitable to splat with one value then change some lanes.
5019 // Heuristically we decide to do this if the vector has a "dominant" value,
5020 // defined as splatted to more than half of the lanes.
5021 unsigned NumElts = VT.getVectorNumElements();
5022 bool isOnlyLowElement = true;
5023 bool usesOnlyOneValue = true;
5024 bool hasDominantValue = false;
5025 bool isConstant = true;
5027 // Map of the number of times a particular SDValue appears in the
5029 DenseMap<SDValue, unsigned> ValueCounts;
5031 for (unsigned i = 0; i < NumElts; ++i) {
5032 SDValue V = Op.getOperand(i);
5033 if (V.getOpcode() == ISD::UNDEF)
5036 isOnlyLowElement = false;
5037 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
5040 ValueCounts.insert(std::make_pair(V, 0));
5041 unsigned &Count = ValueCounts[V];
5043 // Is this value dominant? (takes up more than half of the lanes)
5044 if (++Count > (NumElts / 2)) {
5045 hasDominantValue = true;
5049 if (ValueCounts.size() != 1)
5050 usesOnlyOneValue = false;
5051 if (!Value.getNode() && ValueCounts.size() > 0)
5052 Value = ValueCounts.begin()->first;
5054 if (ValueCounts.size() == 0)
5055 return DAG.getUNDEF(VT);
5057 // Loads are better lowered with insert_vector_elt/ARMISD::BUILD_VECTOR.
5058 // Keep going if we are hitting this case.
5059 if (isOnlyLowElement && !ISD::isNormalLoad(Value.getNode()))
5060 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
5062 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5064 // Use VDUP for non-constant splats. For f32 constant splats, reduce to
5065 // i32 and try again.
5066 if (hasDominantValue && EltSize <= 32) {
5070 // If we are VDUPing a value that comes directly from a vector, that will
5071 // cause an unnecessary move to and from a GPR, where instead we could
5072 // just use VDUPLANE. We can only do this if the lane being extracted
5073 // is at a constant index, as the VDUP from lane instructions only have
5074 // constant-index forms.
5075 if (Value->getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5076 isa<ConstantSDNode>(Value->getOperand(1))) {
5077 // We need to create a new undef vector to use for the VDUPLANE if the
5078 // size of the vector from which we get the value is different than the
5079 // size of the vector that we need to create. We will insert the element
5080 // such that the register coalescer will remove unnecessary copies.
5081 if (VT != Value->getOperand(0).getValueType()) {
5082 ConstantSDNode *constIndex;
5083 constIndex = dyn_cast<ConstantSDNode>(Value->getOperand(1));
5084 assert(constIndex && "The index is not a constant!");
5085 unsigned index = constIndex->getAPIntValue().getLimitedValue() %
5086 VT.getVectorNumElements();
5087 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
5088 DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DAG.getUNDEF(VT),
5089 Value, DAG.getConstant(index, MVT::i32)),
5090 DAG.getConstant(index, MVT::i32));
5092 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
5093 Value->getOperand(0), Value->getOperand(1));
5095 N = DAG.getNode(ARMISD::VDUP, dl, VT, Value);
5097 if (!usesOnlyOneValue) {
5098 // The dominant value was splatted as 'N', but we now have to insert
5099 // all differing elements.
5100 for (unsigned I = 0; I < NumElts; ++I) {
5101 if (Op.getOperand(I) == Value)
5103 SmallVector<SDValue, 3> Ops;
5105 Ops.push_back(Op.getOperand(I));
5106 Ops.push_back(DAG.getConstant(I, MVT::i32));
5107 N = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Ops);
5112 if (VT.getVectorElementType().isFloatingPoint()) {
5113 SmallVector<SDValue, 8> Ops;
5114 for (unsigned i = 0; i < NumElts; ++i)
5115 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, MVT::i32,
5117 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
5118 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, Ops);
5119 Val = LowerBUILD_VECTOR(Val, DAG, ST);
5121 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
5123 if (usesOnlyOneValue) {
5124 SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
5125 if (isConstant && Val.getNode())
5126 return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
5130 // If all elements are constants and the case above didn't get hit, fall back
5131 // to the default expansion, which will generate a load from the constant
5136 // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
5138 SDValue shuffle = ReconstructShuffle(Op, DAG);
5139 if (shuffle != SDValue())
5143 // Vectors with 32- or 64-bit elements can be built by directly assigning
5144 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
5145 // will be legalized.
5146 if (EltSize >= 32) {
5147 // Do the expansion with floating-point types, since that is what the VFP
5148 // registers are defined to use, and since i64 is not legal.
5149 EVT EltVT = EVT::getFloatingPointVT(EltSize);
5150 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
5151 SmallVector<SDValue, 8> Ops;
5152 for (unsigned i = 0; i < NumElts; ++i)
5153 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, EltVT, Op.getOperand(i)));
5154 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, Ops);
5155 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
5158 // If all else fails, just use a sequence of INSERT_VECTOR_ELT when we
5159 // know the default expansion would otherwise fall back on something even
5160 // worse. For a vector with one or two non-undef values, that's
5161 // scalar_to_vector for the elements followed by a shuffle (provided the
5162 // shuffle is valid for the target) and materialization element by element
5163 // on the stack followed by a load for everything else.
5164 if (!isConstant && !usesOnlyOneValue) {
5165 SDValue Vec = DAG.getUNDEF(VT);
5166 for (unsigned i = 0 ; i < NumElts; ++i) {
5167 SDValue V = Op.getOperand(i);
5168 if (V.getOpcode() == ISD::UNDEF)
5170 SDValue LaneIdx = DAG.getConstant(i, MVT::i32);
5171 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Vec, V, LaneIdx);
5179 // Gather data to see if the operation can be modelled as a
5180 // shuffle in combination with VEXTs.
5181 SDValue ARMTargetLowering::ReconstructShuffle(SDValue Op,
5182 SelectionDAG &DAG) const {
5184 EVT VT = Op.getValueType();
5185 unsigned NumElts = VT.getVectorNumElements();
5187 SmallVector<SDValue, 2> SourceVecs;
5188 SmallVector<unsigned, 2> MinElts;
5189 SmallVector<unsigned, 2> MaxElts;
5191 for (unsigned i = 0; i < NumElts; ++i) {
5192 SDValue V = Op.getOperand(i);
5193 if (V.getOpcode() == ISD::UNDEF)
5195 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
5196 // A shuffle can only come from building a vector from various
5197 // elements of other vectors.
5199 } else if (V.getOperand(0).getValueType().getVectorElementType() !=
5200 VT.getVectorElementType()) {
5201 // This code doesn't know how to handle shuffles where the vector
5202 // element types do not match (this happens because type legalization
5203 // promotes the return type of EXTRACT_VECTOR_ELT).
5204 // FIXME: It might be appropriate to extend this code to handle
5205 // mismatched types.
5209 // Record this extraction against the appropriate vector if possible...
5210 SDValue SourceVec = V.getOperand(0);
5211 // If the element number isn't a constant, we can't effectively
5212 // analyze what's going on.
5213 if (!isa<ConstantSDNode>(V.getOperand(1)))
5215 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
5216 bool FoundSource = false;
5217 for (unsigned j = 0; j < SourceVecs.size(); ++j) {
5218 if (SourceVecs[j] == SourceVec) {
5219 if (MinElts[j] > EltNo)
5221 if (MaxElts[j] < EltNo)
5228 // Or record a new source if not...
5230 SourceVecs.push_back(SourceVec);
5231 MinElts.push_back(EltNo);
5232 MaxElts.push_back(EltNo);
5236 // Currently only do something sane when at most two source vectors
5238 if (SourceVecs.size() > 2)
5241 SDValue ShuffleSrcs[2] = {DAG.getUNDEF(VT), DAG.getUNDEF(VT) };
5242 int VEXTOffsets[2] = {0, 0};
5244 // This loop extracts the usage patterns of the source vectors
5245 // and prepares appropriate SDValues for a shuffle if possible.
5246 for (unsigned i = 0; i < SourceVecs.size(); ++i) {
5247 if (SourceVecs[i].getValueType() == VT) {
5248 // No VEXT necessary
5249 ShuffleSrcs[i] = SourceVecs[i];
5252 } else if (SourceVecs[i].getValueType().getVectorNumElements() < NumElts) {
5253 // It probably isn't worth padding out a smaller vector just to
5254 // break it down again in a shuffle.
5258 // Since only 64-bit and 128-bit vectors are legal on ARM and
5259 // we've eliminated the other cases...
5260 assert(SourceVecs[i].getValueType().getVectorNumElements() == 2*NumElts &&
5261 "unexpected vector sizes in ReconstructShuffle");
5263 if (MaxElts[i] - MinElts[i] >= NumElts) {
5264 // Span too large for a VEXT to cope
5268 if (MinElts[i] >= NumElts) {
5269 // The extraction can just take the second half
5270 VEXTOffsets[i] = NumElts;
5271 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5273 DAG.getIntPtrConstant(NumElts));
5274 } else if (MaxElts[i] < NumElts) {
5275 // The extraction can just take the first half
5277 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5279 DAG.getIntPtrConstant(0));
5281 // An actual VEXT is needed
5282 VEXTOffsets[i] = MinElts[i];
5283 SDValue VEXTSrc1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5285 DAG.getIntPtrConstant(0));
5286 SDValue VEXTSrc2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5288 DAG.getIntPtrConstant(NumElts));
5289 ShuffleSrcs[i] = DAG.getNode(ARMISD::VEXT, dl, VT, VEXTSrc1, VEXTSrc2,
5290 DAG.getConstant(VEXTOffsets[i], MVT::i32));
5294 SmallVector<int, 8> Mask;
5296 for (unsigned i = 0; i < NumElts; ++i) {
5297 SDValue Entry = Op.getOperand(i);
5298 if (Entry.getOpcode() == ISD::UNDEF) {
5303 SDValue ExtractVec = Entry.getOperand(0);
5304 int ExtractElt = cast<ConstantSDNode>(Op.getOperand(i)
5305 .getOperand(1))->getSExtValue();
5306 if (ExtractVec == SourceVecs[0]) {
5307 Mask.push_back(ExtractElt - VEXTOffsets[0]);
5309 Mask.push_back(ExtractElt + NumElts - VEXTOffsets[1]);
5313 // Final check before we try to produce nonsense...
5314 if (isShuffleMaskLegal(Mask, VT))
5315 return DAG.getVectorShuffle(VT, dl, ShuffleSrcs[0], ShuffleSrcs[1],
5321 /// isShuffleMaskLegal - Targets can use this to indicate that they only
5322 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
5323 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
5324 /// are assumed to be legal.
5326 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
5328 if (VT.getVectorNumElements() == 4 &&
5329 (VT.is128BitVector() || VT.is64BitVector())) {
5330 unsigned PFIndexes[4];
5331 for (unsigned i = 0; i != 4; ++i) {
5335 PFIndexes[i] = M[i];
5338 // Compute the index in the perfect shuffle table.
5339 unsigned PFTableIndex =
5340 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
5341 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
5342 unsigned Cost = (PFEntry >> 30);
5349 unsigned Imm, WhichResult;
5351 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5352 return (EltSize >= 32 ||
5353 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
5354 isVREVMask(M, VT, 64) ||
5355 isVREVMask(M, VT, 32) ||
5356 isVREVMask(M, VT, 16) ||
5357 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
5358 isVTBLMask(M, VT) ||
5359 isVTRNMask(M, VT, WhichResult) ||
5360 isVUZPMask(M, VT, WhichResult) ||
5361 isVZIPMask(M, VT, WhichResult) ||
5362 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
5363 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
5364 isVZIP_v_undef_Mask(M, VT, WhichResult) ||
5365 ((VT == MVT::v8i16 || VT == MVT::v16i8) && isReverseMask(M, VT)));
5368 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
5369 /// the specified operations to build the shuffle.
5370 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
5371 SDValue RHS, SelectionDAG &DAG,
5373 unsigned OpNum = (PFEntry >> 26) & 0x0F;
5374 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
5375 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
5378 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
5387 OP_VUZPL, // VUZP, left result
5388 OP_VUZPR, // VUZP, right result
5389 OP_VZIPL, // VZIP, left result
5390 OP_VZIPR, // VZIP, right result
5391 OP_VTRNL, // VTRN, left result
5392 OP_VTRNR // VTRN, right result
5395 if (OpNum == OP_COPY) {
5396 if (LHSID == (1*9+2)*9+3) return LHS;
5397 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
5401 SDValue OpLHS, OpRHS;
5402 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
5403 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
5404 EVT VT = OpLHS.getValueType();
5407 default: llvm_unreachable("Unknown shuffle opcode!");
5409 // VREV divides the vector in half and swaps within the half.
5410 if (VT.getVectorElementType() == MVT::i32 ||
5411 VT.getVectorElementType() == MVT::f32)
5412 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
5413 // vrev <4 x i16> -> VREV32
5414 if (VT.getVectorElementType() == MVT::i16)
5415 return DAG.getNode(ARMISD::VREV32, dl, VT, OpLHS);
5416 // vrev <4 x i8> -> VREV16
5417 assert(VT.getVectorElementType() == MVT::i8);
5418 return DAG.getNode(ARMISD::VREV16, dl, VT, OpLHS);
5423 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
5424 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
5428 return DAG.getNode(ARMISD::VEXT, dl, VT,
5430 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
5433 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
5434 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
5437 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
5438 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
5441 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
5442 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
5446 static SDValue LowerVECTOR_SHUFFLEv8i8(SDValue Op,
5447 ArrayRef<int> ShuffleMask,
5448 SelectionDAG &DAG) {
5449 // Check to see if we can use the VTBL instruction.
5450 SDValue V1 = Op.getOperand(0);
5451 SDValue V2 = Op.getOperand(1);
5454 SmallVector<SDValue, 8> VTBLMask;
5455 for (ArrayRef<int>::iterator
5456 I = ShuffleMask.begin(), E = ShuffleMask.end(); I != E; ++I)
5457 VTBLMask.push_back(DAG.getConstant(*I, MVT::i32));
5459 if (V2.getNode()->getOpcode() == ISD::UNDEF)
5460 return DAG.getNode(ARMISD::VTBL1, DL, MVT::v8i8, V1,
5461 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8, VTBLMask));
5463 return DAG.getNode(ARMISD::VTBL2, DL, MVT::v8i8, V1, V2,
5464 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8, VTBLMask));
5467 static SDValue LowerReverse_VECTOR_SHUFFLEv16i8_v8i16(SDValue Op,
5468 SelectionDAG &DAG) {
5470 SDValue OpLHS = Op.getOperand(0);
5471 EVT VT = OpLHS.getValueType();
5473 assert((VT == MVT::v8i16 || VT == MVT::v16i8) &&
5474 "Expect an v8i16/v16i8 type");
5475 OpLHS = DAG.getNode(ARMISD::VREV64, DL, VT, OpLHS);
5476 // For a v16i8 type: After the VREV, we have got <8, ...15, 8, ..., 0>. Now,
5477 // extract the first 8 bytes into the top double word and the last 8 bytes
5478 // into the bottom double word. The v8i16 case is similar.
5479 unsigned ExtractNum = (VT == MVT::v16i8) ? 8 : 4;
5480 return DAG.getNode(ARMISD::VEXT, DL, VT, OpLHS, OpLHS,
5481 DAG.getConstant(ExtractNum, MVT::i32));
5484 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
5485 SDValue V1 = Op.getOperand(0);
5486 SDValue V2 = Op.getOperand(1);
5488 EVT VT = Op.getValueType();
5489 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
5491 // Convert shuffles that are directly supported on NEON to target-specific
5492 // DAG nodes, instead of keeping them as shuffles and matching them again
5493 // during code selection. This is more efficient and avoids the possibility
5494 // of inconsistencies between legalization and selection.
5495 // FIXME: floating-point vectors should be canonicalized to integer vectors
5496 // of the same time so that they get CSEd properly.
5497 ArrayRef<int> ShuffleMask = SVN->getMask();
5499 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5500 if (EltSize <= 32) {
5501 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
5502 int Lane = SVN->getSplatIndex();
5503 // If this is undef splat, generate it via "just" vdup, if possible.
5504 if (Lane == -1) Lane = 0;
5506 // Test if V1 is a SCALAR_TO_VECTOR.
5507 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
5508 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
5510 // Test if V1 is a BUILD_VECTOR which is equivalent to a SCALAR_TO_VECTOR
5511 // (and probably will turn into a SCALAR_TO_VECTOR once legalization
5513 if (Lane == 0 && V1.getOpcode() == ISD::BUILD_VECTOR &&
5514 !isa<ConstantSDNode>(V1.getOperand(0))) {
5515 bool IsScalarToVector = true;
5516 for (unsigned i = 1, e = V1.getNumOperands(); i != e; ++i)
5517 if (V1.getOperand(i).getOpcode() != ISD::UNDEF) {
5518 IsScalarToVector = false;
5521 if (IsScalarToVector)
5522 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
5524 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
5525 DAG.getConstant(Lane, MVT::i32));
5530 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
5533 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
5534 DAG.getConstant(Imm, MVT::i32));
5537 if (isVREVMask(ShuffleMask, VT, 64))
5538 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
5539 if (isVREVMask(ShuffleMask, VT, 32))
5540 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
5541 if (isVREVMask(ShuffleMask, VT, 16))
5542 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
5544 if (V2->getOpcode() == ISD::UNDEF &&
5545 isSingletonVEXTMask(ShuffleMask, VT, Imm)) {
5546 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V1,
5547 DAG.getConstant(Imm, MVT::i32));
5550 // Check for Neon shuffles that modify both input vectors in place.
5551 // If both results are used, i.e., if there are two shuffles with the same
5552 // source operands and with masks corresponding to both results of one of
5553 // these operations, DAG memoization will ensure that a single node is
5554 // used for both shuffles.
5555 unsigned WhichResult;
5556 if (isVTRNMask(ShuffleMask, VT, WhichResult))
5557 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
5558 V1, V2).getValue(WhichResult);
5559 if (isVUZPMask(ShuffleMask, VT, WhichResult))
5560 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
5561 V1, V2).getValue(WhichResult);
5562 if (isVZIPMask(ShuffleMask, VT, WhichResult))
5563 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
5564 V1, V2).getValue(WhichResult);
5566 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
5567 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
5568 V1, V1).getValue(WhichResult);
5569 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
5570 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
5571 V1, V1).getValue(WhichResult);
5572 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
5573 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
5574 V1, V1).getValue(WhichResult);
5577 // If the shuffle is not directly supported and it has 4 elements, use
5578 // the PerfectShuffle-generated table to synthesize it from other shuffles.
5579 unsigned NumElts = VT.getVectorNumElements();
5581 unsigned PFIndexes[4];
5582 for (unsigned i = 0; i != 4; ++i) {
5583 if (ShuffleMask[i] < 0)
5586 PFIndexes[i] = ShuffleMask[i];
5589 // Compute the index in the perfect shuffle table.
5590 unsigned PFTableIndex =
5591 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
5592 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
5593 unsigned Cost = (PFEntry >> 30);
5596 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
5599 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
5600 if (EltSize >= 32) {
5601 // Do the expansion with floating-point types, since that is what the VFP
5602 // registers are defined to use, and since i64 is not legal.
5603 EVT EltVT = EVT::getFloatingPointVT(EltSize);
5604 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
5605 V1 = DAG.getNode(ISD::BITCAST, dl, VecVT, V1);
5606 V2 = DAG.getNode(ISD::BITCAST, dl, VecVT, V2);
5607 SmallVector<SDValue, 8> Ops;
5608 for (unsigned i = 0; i < NumElts; ++i) {
5609 if (ShuffleMask[i] < 0)
5610 Ops.push_back(DAG.getUNDEF(EltVT));
5612 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
5613 ShuffleMask[i] < (int)NumElts ? V1 : V2,
5614 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
5617 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, Ops);
5618 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
5621 if ((VT == MVT::v8i16 || VT == MVT::v16i8) && isReverseMask(ShuffleMask, VT))
5622 return LowerReverse_VECTOR_SHUFFLEv16i8_v8i16(Op, DAG);
5624 if (VT == MVT::v8i8) {
5625 SDValue NewOp = LowerVECTOR_SHUFFLEv8i8(Op, ShuffleMask, DAG);
5626 if (NewOp.getNode())
5633 static SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
5634 // INSERT_VECTOR_ELT is legal only for immediate indexes.
5635 SDValue Lane = Op.getOperand(2);
5636 if (!isa<ConstantSDNode>(Lane))
5642 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
5643 // EXTRACT_VECTOR_ELT is legal only for immediate indexes.
5644 SDValue Lane = Op.getOperand(1);
5645 if (!isa<ConstantSDNode>(Lane))
5648 SDValue Vec = Op.getOperand(0);
5649 if (Op.getValueType() == MVT::i32 &&
5650 Vec.getValueType().getVectorElementType().getSizeInBits() < 32) {
5652 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
5658 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
5659 // The only time a CONCAT_VECTORS operation can have legal types is when
5660 // two 64-bit vectors are concatenated to a 128-bit vector.
5661 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
5662 "unexpected CONCAT_VECTORS");
5664 SDValue Val = DAG.getUNDEF(MVT::v2f64);
5665 SDValue Op0 = Op.getOperand(0);
5666 SDValue Op1 = Op.getOperand(1);
5667 if (Op0.getOpcode() != ISD::UNDEF)
5668 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
5669 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op0),
5670 DAG.getIntPtrConstant(0));
5671 if (Op1.getOpcode() != ISD::UNDEF)
5672 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
5673 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op1),
5674 DAG.getIntPtrConstant(1));
5675 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Val);
5678 /// isExtendedBUILD_VECTOR - Check if N is a constant BUILD_VECTOR where each
5679 /// element has been zero/sign-extended, depending on the isSigned parameter,
5680 /// from an integer type half its size.
5681 static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG,
5683 // A v2i64 BUILD_VECTOR will have been legalized to a BITCAST from v4i32.
5684 EVT VT = N->getValueType(0);
5685 if (VT == MVT::v2i64 && N->getOpcode() == ISD::BITCAST) {
5686 SDNode *BVN = N->getOperand(0).getNode();
5687 if (BVN->getValueType(0) != MVT::v4i32 ||
5688 BVN->getOpcode() != ISD::BUILD_VECTOR)
5690 unsigned LoElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
5691 unsigned HiElt = 1 - LoElt;
5692 ConstantSDNode *Lo0 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt));
5693 ConstantSDNode *Hi0 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt));
5694 ConstantSDNode *Lo1 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt+2));
5695 ConstantSDNode *Hi1 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt+2));
5696 if (!Lo0 || !Hi0 || !Lo1 || !Hi1)
5699 if (Hi0->getSExtValue() == Lo0->getSExtValue() >> 32 &&
5700 Hi1->getSExtValue() == Lo1->getSExtValue() >> 32)
5703 if (Hi0->isNullValue() && Hi1->isNullValue())
5709 if (N->getOpcode() != ISD::BUILD_VECTOR)
5712 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
5713 SDNode *Elt = N->getOperand(i).getNode();
5714 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
5715 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5716 unsigned HalfSize = EltSize / 2;
5718 if (!isIntN(HalfSize, C->getSExtValue()))
5721 if (!isUIntN(HalfSize, C->getZExtValue()))
5732 /// isSignExtended - Check if a node is a vector value that is sign-extended
5733 /// or a constant BUILD_VECTOR with sign-extended elements.
5734 static bool isSignExtended(SDNode *N, SelectionDAG &DAG) {
5735 if (N->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N))
5737 if (isExtendedBUILD_VECTOR(N, DAG, true))
5742 /// isZeroExtended - Check if a node is a vector value that is zero-extended
5743 /// or a constant BUILD_VECTOR with zero-extended elements.
5744 static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) {
5745 if (N->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N))
5747 if (isExtendedBUILD_VECTOR(N, DAG, false))
5752 static EVT getExtensionTo64Bits(const EVT &OrigVT) {
5753 if (OrigVT.getSizeInBits() >= 64)
5756 assert(OrigVT.isSimple() && "Expecting a simple value type");
5758 MVT::SimpleValueType OrigSimpleTy = OrigVT.getSimpleVT().SimpleTy;
5759 switch (OrigSimpleTy) {
5760 default: llvm_unreachable("Unexpected Vector Type");
5769 /// AddRequiredExtensionForVMULL - Add a sign/zero extension to extend the total
5770 /// value size to 64 bits. We need a 64-bit D register as an operand to VMULL.
5771 /// We insert the required extension here to get the vector to fill a D register.
5772 static SDValue AddRequiredExtensionForVMULL(SDValue N, SelectionDAG &DAG,
5775 unsigned ExtOpcode) {
5776 // The vector originally had a size of OrigTy. It was then extended to ExtTy.
5777 // We expect the ExtTy to be 128-bits total. If the OrigTy is less than
5778 // 64-bits we need to insert a new extension so that it will be 64-bits.
5779 assert(ExtTy.is128BitVector() && "Unexpected extension size");
5780 if (OrigTy.getSizeInBits() >= 64)
5783 // Must extend size to at least 64 bits to be used as an operand for VMULL.
5784 EVT NewVT = getExtensionTo64Bits(OrigTy);
5786 return DAG.getNode(ExtOpcode, SDLoc(N), NewVT, N);
5789 /// SkipLoadExtensionForVMULL - return a load of the original vector size that
5790 /// does not do any sign/zero extension. If the original vector is less
5791 /// than 64 bits, an appropriate extension will be added after the load to
5792 /// reach a total size of 64 bits. We have to add the extension separately
5793 /// because ARM does not have a sign/zero extending load for vectors.
5794 static SDValue SkipLoadExtensionForVMULL(LoadSDNode *LD, SelectionDAG& DAG) {
5795 EVT ExtendedTy = getExtensionTo64Bits(LD->getMemoryVT());
5797 // The load already has the right type.
5798 if (ExtendedTy == LD->getMemoryVT())
5799 return DAG.getLoad(LD->getMemoryVT(), SDLoc(LD), LD->getChain(),
5800 LD->getBasePtr(), LD->getPointerInfo(), LD->isVolatile(),
5801 LD->isNonTemporal(), LD->isInvariant(),
5802 LD->getAlignment());
5804 // We need to create a zextload/sextload. We cannot just create a load
5805 // followed by a zext/zext node because LowerMUL is also run during normal
5806 // operation legalization where we can't create illegal types.
5807 return DAG.getExtLoad(LD->getExtensionType(), SDLoc(LD), ExtendedTy,
5808 LD->getChain(), LD->getBasePtr(), LD->getPointerInfo(),
5809 LD->getMemoryVT(), LD->isVolatile(), LD->isInvariant(),
5810 LD->isNonTemporal(), LD->getAlignment());
5813 /// SkipExtensionForVMULL - For a node that is a SIGN_EXTEND, ZERO_EXTEND,
5814 /// extending load, or BUILD_VECTOR with extended elements, return the
5815 /// unextended value. The unextended vector should be 64 bits so that it can
5816 /// be used as an operand to a VMULL instruction. If the original vector size
5817 /// before extension is less than 64 bits we add a an extension to resize
5818 /// the vector to 64 bits.
5819 static SDValue SkipExtensionForVMULL(SDNode *N, SelectionDAG &DAG) {
5820 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
5821 return AddRequiredExtensionForVMULL(N->getOperand(0), DAG,
5822 N->getOperand(0)->getValueType(0),
5826 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
5827 return SkipLoadExtensionForVMULL(LD, DAG);
5829 // Otherwise, the value must be a BUILD_VECTOR. For v2i64, it will
5830 // have been legalized as a BITCAST from v4i32.
5831 if (N->getOpcode() == ISD::BITCAST) {
5832 SDNode *BVN = N->getOperand(0).getNode();
5833 assert(BVN->getOpcode() == ISD::BUILD_VECTOR &&
5834 BVN->getValueType(0) == MVT::v4i32 && "expected v4i32 BUILD_VECTOR");
5835 unsigned LowElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
5836 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), MVT::v2i32,
5837 BVN->getOperand(LowElt), BVN->getOperand(LowElt+2));
5839 // Construct a new BUILD_VECTOR with elements truncated to half the size.
5840 assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR");
5841 EVT VT = N->getValueType(0);
5842 unsigned EltSize = VT.getVectorElementType().getSizeInBits() / 2;
5843 unsigned NumElts = VT.getVectorNumElements();
5844 MVT TruncVT = MVT::getIntegerVT(EltSize);
5845 SmallVector<SDValue, 8> Ops;
5846 for (unsigned i = 0; i != NumElts; ++i) {
5847 ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i));
5848 const APInt &CInt = C->getAPIntValue();
5849 // Element types smaller than 32 bits are not legal, so use i32 elements.
5850 // The values are implicitly truncated so sext vs. zext doesn't matter.
5851 Ops.push_back(DAG.getConstant(CInt.zextOrTrunc(32), MVT::i32));
5853 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N),
5854 MVT::getVectorVT(TruncVT, NumElts), Ops);
5857 static bool isAddSubSExt(SDNode *N, SelectionDAG &DAG) {
5858 unsigned Opcode = N->getOpcode();
5859 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
5860 SDNode *N0 = N->getOperand(0).getNode();
5861 SDNode *N1 = N->getOperand(1).getNode();
5862 return N0->hasOneUse() && N1->hasOneUse() &&
5863 isSignExtended(N0, DAG) && isSignExtended(N1, DAG);
5868 static bool isAddSubZExt(SDNode *N, SelectionDAG &DAG) {
5869 unsigned Opcode = N->getOpcode();
5870 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
5871 SDNode *N0 = N->getOperand(0).getNode();
5872 SDNode *N1 = N->getOperand(1).getNode();
5873 return N0->hasOneUse() && N1->hasOneUse() &&
5874 isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG);
5879 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
5880 // Multiplications are only custom-lowered for 128-bit vectors so that
5881 // VMULL can be detected. Otherwise v2i64 multiplications are not legal.
5882 EVT VT = Op.getValueType();
5883 assert(VT.is128BitVector() && VT.isInteger() &&
5884 "unexpected type for custom-lowering ISD::MUL");
5885 SDNode *N0 = Op.getOperand(0).getNode();
5886 SDNode *N1 = Op.getOperand(1).getNode();
5887 unsigned NewOpc = 0;
5889 bool isN0SExt = isSignExtended(N0, DAG);
5890 bool isN1SExt = isSignExtended(N1, DAG);
5891 if (isN0SExt && isN1SExt)
5892 NewOpc = ARMISD::VMULLs;
5894 bool isN0ZExt = isZeroExtended(N0, DAG);
5895 bool isN1ZExt = isZeroExtended(N1, DAG);
5896 if (isN0ZExt && isN1ZExt)
5897 NewOpc = ARMISD::VMULLu;
5898 else if (isN1SExt || isN1ZExt) {
5899 // Look for (s/zext A + s/zext B) * (s/zext C). We want to turn these
5900 // into (s/zext A * s/zext C) + (s/zext B * s/zext C)
5901 if (isN1SExt && isAddSubSExt(N0, DAG)) {
5902 NewOpc = ARMISD::VMULLs;
5904 } else if (isN1ZExt && isAddSubZExt(N0, DAG)) {
5905 NewOpc = ARMISD::VMULLu;
5907 } else if (isN0ZExt && isAddSubZExt(N1, DAG)) {
5909 NewOpc = ARMISD::VMULLu;
5915 if (VT == MVT::v2i64)
5916 // Fall through to expand this. It is not legal.
5919 // Other vector multiplications are legal.
5924 // Legalize to a VMULL instruction.
5927 SDValue Op1 = SkipExtensionForVMULL(N1, DAG);
5929 Op0 = SkipExtensionForVMULL(N0, DAG);
5930 assert(Op0.getValueType().is64BitVector() &&
5931 Op1.getValueType().is64BitVector() &&
5932 "unexpected types for extended operands to VMULL");
5933 return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
5936 // Optimizing (zext A + zext B) * C, to (VMULL A, C) + (VMULL B, C) during
5937 // isel lowering to take advantage of no-stall back to back vmul + vmla.
5944 SDValue N00 = SkipExtensionForVMULL(N0->getOperand(0).getNode(), DAG);
5945 SDValue N01 = SkipExtensionForVMULL(N0->getOperand(1).getNode(), DAG);
5946 EVT Op1VT = Op1.getValueType();
5947 return DAG.getNode(N0->getOpcode(), DL, VT,
5948 DAG.getNode(NewOpc, DL, VT,
5949 DAG.getNode(ISD::BITCAST, DL, Op1VT, N00), Op1),
5950 DAG.getNode(NewOpc, DL, VT,
5951 DAG.getNode(ISD::BITCAST, DL, Op1VT, N01), Op1));
5955 LowerSDIV_v4i8(SDValue X, SDValue Y, SDLoc dl, SelectionDAG &DAG) {
5957 // float4 xf = vcvt_f32_s32(vmovl_s16(a.lo));
5958 // float4 yf = vcvt_f32_s32(vmovl_s16(b.lo));
5959 X = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, X);
5960 Y = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, Y);
5961 X = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, X);
5962 Y = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, Y);
5963 // Get reciprocal estimate.
5964 // float4 recip = vrecpeq_f32(yf);
5965 Y = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5966 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), Y);
5967 // Because char has a smaller range than uchar, we can actually get away
5968 // without any newton steps. This requires that we use a weird bias
5969 // of 0xb000, however (again, this has been exhaustively tested).
5970 // float4 result = as_float4(as_int4(xf*recip) + 0xb000);
5971 X = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, X, Y);
5972 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, X);
5973 Y = DAG.getConstant(0xb000, MVT::i32);
5974 Y = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Y, Y, Y, Y);
5975 X = DAG.getNode(ISD::ADD, dl, MVT::v4i32, X, Y);
5976 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, X);
5977 // Convert back to short.
5978 X = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, X);
5979 X = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, X);
5984 LowerSDIV_v4i16(SDValue N0, SDValue N1, SDLoc dl, SelectionDAG &DAG) {
5986 // Convert to float.
5987 // float4 yf = vcvt_f32_s32(vmovl_s16(y));
5988 // float4 xf = vcvt_f32_s32(vmovl_s16(x));
5989 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N0);
5990 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N1);
5991 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
5992 N1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
5994 // Use reciprocal estimate and one refinement step.
5995 // float4 recip = vrecpeq_f32(yf);
5996 // recip *= vrecpsq_f32(yf, recip);
5997 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5998 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), N1);
5999 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6000 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
6002 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
6003 // Because short has a smaller range than ushort, we can actually get away
6004 // with only a single newton step. This requires that we use a weird bias
6005 // of 89, however (again, this has been exhaustively tested).
6006 // float4 result = as_float4(as_int4(xf*recip) + 0x89);
6007 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
6008 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
6009 N1 = DAG.getConstant(0x89, MVT::i32);
6010 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
6011 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
6012 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
6013 // Convert back to integer and return.
6014 // return vmovn_s32(vcvt_s32_f32(result));
6015 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
6016 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
6020 static SDValue LowerSDIV(SDValue Op, SelectionDAG &DAG) {
6021 EVT VT = Op.getValueType();
6022 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
6023 "unexpected type for custom-lowering ISD::SDIV");
6026 SDValue N0 = Op.getOperand(0);
6027 SDValue N1 = Op.getOperand(1);
6030 if (VT == MVT::v8i8) {
6031 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N0);
6032 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N1);
6034 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
6035 DAG.getIntPtrConstant(4));
6036 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
6037 DAG.getIntPtrConstant(4));
6038 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
6039 DAG.getIntPtrConstant(0));
6040 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
6041 DAG.getIntPtrConstant(0));
6043 N0 = LowerSDIV_v4i8(N0, N1, dl, DAG); // v4i16
6044 N2 = LowerSDIV_v4i8(N2, N3, dl, DAG); // v4i16
6046 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
6047 N0 = LowerCONCAT_VECTORS(N0, DAG);
6049 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v8i8, N0);
6052 return LowerSDIV_v4i16(N0, N1, dl, DAG);
6055 static SDValue LowerUDIV(SDValue Op, SelectionDAG &DAG) {
6056 EVT VT = Op.getValueType();
6057 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
6058 "unexpected type for custom-lowering ISD::UDIV");
6061 SDValue N0 = Op.getOperand(0);
6062 SDValue N1 = Op.getOperand(1);
6065 if (VT == MVT::v8i8) {
6066 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N0);
6067 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N1);
6069 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
6070 DAG.getIntPtrConstant(4));
6071 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
6072 DAG.getIntPtrConstant(4));
6073 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
6074 DAG.getIntPtrConstant(0));
6075 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
6076 DAG.getIntPtrConstant(0));
6078 N0 = LowerSDIV_v4i16(N0, N1, dl, DAG); // v4i16
6079 N2 = LowerSDIV_v4i16(N2, N3, dl, DAG); // v4i16
6081 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
6082 N0 = LowerCONCAT_VECTORS(N0, DAG);
6084 N0 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v8i8,
6085 DAG.getConstant(Intrinsic::arm_neon_vqmovnsu, MVT::i32),
6090 // v4i16 sdiv ... Convert to float.
6091 // float4 yf = vcvt_f32_s32(vmovl_u16(y));
6092 // float4 xf = vcvt_f32_s32(vmovl_u16(x));
6093 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N0);
6094 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N1);
6095 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
6096 SDValue BN1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
6098 // Use reciprocal estimate and two refinement steps.
6099 // float4 recip = vrecpeq_f32(yf);
6100 // recip *= vrecpsq_f32(yf, recip);
6101 // recip *= vrecpsq_f32(yf, recip);
6102 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6103 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), BN1);
6104 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6105 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
6107 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
6108 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6109 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
6111 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
6112 // Simply multiplying by the reciprocal estimate can leave us a few ulps
6113 // too low, so we add 2 ulps (exhaustive testing shows that this is enough,
6114 // and that it will never cause us to return an answer too large).
6115 // float4 result = as_float4(as_int4(xf*recip) + 2);
6116 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
6117 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
6118 N1 = DAG.getConstant(2, MVT::i32);
6119 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
6120 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
6121 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
6122 // Convert back to integer and return.
6123 // return vmovn_u32(vcvt_s32_f32(result));
6124 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
6125 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
6129 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
6130 EVT VT = Op.getNode()->getValueType(0);
6131 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
6134 bool ExtraOp = false;
6135 switch (Op.getOpcode()) {
6136 default: llvm_unreachable("Invalid code");
6137 case ISD::ADDC: Opc = ARMISD::ADDC; break;
6138 case ISD::ADDE: Opc = ARMISD::ADDE; ExtraOp = true; break;
6139 case ISD::SUBC: Opc = ARMISD::SUBC; break;
6140 case ISD::SUBE: Opc = ARMISD::SUBE; ExtraOp = true; break;
6144 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
6146 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
6147 Op.getOperand(1), Op.getOperand(2));
6150 SDValue ARMTargetLowering::LowerFSINCOS(SDValue Op, SelectionDAG &DAG) const {
6151 assert(Subtarget->isTargetDarwin());
6153 // For iOS, we want to call an alternative entry point: __sincos_stret,
6154 // return values are passed via sret.
6156 SDValue Arg = Op.getOperand(0);
6157 EVT ArgVT = Arg.getValueType();
6158 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
6160 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
6161 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6163 // Pair of floats / doubles used to pass the result.
6164 StructType *RetTy = StructType::get(ArgTy, ArgTy, nullptr);
6166 // Create stack object for sret.
6167 const uint64_t ByteSize = TLI.getDataLayout()->getTypeAllocSize(RetTy);
6168 const unsigned StackAlign = TLI.getDataLayout()->getPrefTypeAlignment(RetTy);
6169 int FrameIdx = FrameInfo->CreateStackObject(ByteSize, StackAlign, false);
6170 SDValue SRet = DAG.getFrameIndex(FrameIdx, TLI.getPointerTy());
6176 Entry.Ty = RetTy->getPointerTo();
6177 Entry.isSExt = false;
6178 Entry.isZExt = false;
6179 Entry.isSRet = true;
6180 Args.push_back(Entry);
6184 Entry.isSExt = false;
6185 Entry.isZExt = false;
6186 Args.push_back(Entry);
6188 const char *LibcallName = (ArgVT == MVT::f64)
6189 ? "__sincos_stret" : "__sincosf_stret";
6190 SDValue Callee = DAG.getExternalSymbol(LibcallName, getPointerTy());
6192 TargetLowering::CallLoweringInfo CLI(DAG);
6193 CLI.setDebugLoc(dl).setChain(DAG.getEntryNode())
6194 .setCallee(CallingConv::C, Type::getVoidTy(*DAG.getContext()), Callee,
6196 .setDiscardResult();
6198 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
6200 SDValue LoadSin = DAG.getLoad(ArgVT, dl, CallResult.second, SRet,
6201 MachinePointerInfo(), false, false, false, 0);
6203 // Address of cos field.
6204 SDValue Add = DAG.getNode(ISD::ADD, dl, getPointerTy(), SRet,
6205 DAG.getIntPtrConstant(ArgVT.getStoreSize()));
6206 SDValue LoadCos = DAG.getLoad(ArgVT, dl, LoadSin.getValue(1), Add,
6207 MachinePointerInfo(), false, false, false, 0);
6209 SDVTList Tys = DAG.getVTList(ArgVT, ArgVT);
6210 return DAG.getNode(ISD::MERGE_VALUES, dl, Tys,
6211 LoadSin.getValue(0), LoadCos.getValue(0));
6214 static SDValue LowerAtomicLoadStore(SDValue Op, SelectionDAG &DAG) {
6215 // Monotonic load/store is legal for all targets
6216 if (cast<AtomicSDNode>(Op)->getOrdering() <= Monotonic)
6219 // Acquire/Release load/store is not legal for targets without a
6220 // dmb or equivalent available.
6224 static void ReplaceREADCYCLECOUNTER(SDNode *N,
6225 SmallVectorImpl<SDValue> &Results,
6227 const ARMSubtarget *Subtarget) {
6229 SDValue Cycles32, OutChain;
6231 if (Subtarget->hasPerfMon()) {
6232 // Under Power Management extensions, the cycle-count is:
6233 // mrc p15, #0, <Rt>, c9, c13, #0
6234 SDValue Ops[] = { N->getOperand(0), // Chain
6235 DAG.getConstant(Intrinsic::arm_mrc, MVT::i32),
6236 DAG.getConstant(15, MVT::i32),
6237 DAG.getConstant(0, MVT::i32),
6238 DAG.getConstant(9, MVT::i32),
6239 DAG.getConstant(13, MVT::i32),
6240 DAG.getConstant(0, MVT::i32)
6243 Cycles32 = DAG.getNode(ISD::INTRINSIC_W_CHAIN, DL,
6244 DAG.getVTList(MVT::i32, MVT::Other), Ops);
6245 OutChain = Cycles32.getValue(1);
6247 // Intrinsic is defined to return 0 on unsupported platforms. Technically
6248 // there are older ARM CPUs that have implementation-specific ways of
6249 // obtaining this information (FIXME!).
6250 Cycles32 = DAG.getConstant(0, MVT::i32);
6251 OutChain = DAG.getEntryNode();
6255 SDValue Cycles64 = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64,
6256 Cycles32, DAG.getConstant(0, MVT::i32));
6257 Results.push_back(Cycles64);
6258 Results.push_back(OutChain);
6261 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
6262 switch (Op.getOpcode()) {
6263 default: llvm_unreachable("Don't know how to custom lower this!");
6264 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
6265 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
6266 case ISD::GlobalAddress:
6267 switch (Subtarget->getTargetTriple().getObjectFormat()) {
6268 default: llvm_unreachable("unknown object format");
6270 return LowerGlobalAddressWindows(Op, DAG);
6272 return LowerGlobalAddressELF(Op, DAG);
6274 return LowerGlobalAddressDarwin(Op, DAG);
6276 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
6277 case ISD::SELECT: return LowerSELECT(Op, DAG);
6278 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
6279 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
6280 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
6281 case ISD::VASTART: return LowerVASTART(Op, DAG);
6282 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, DAG, Subtarget);
6283 case ISD::PREFETCH: return LowerPREFETCH(Op, DAG, Subtarget);
6284 case ISD::SINT_TO_FP:
6285 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
6286 case ISD::FP_TO_SINT:
6287 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
6288 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
6289 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
6290 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
6291 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
6292 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
6293 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
6294 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
6296 case ISD::BITCAST: return ExpandBITCAST(Op.getNode(), DAG);
6299 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
6300 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
6301 case ISD::SRL_PARTS:
6302 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
6303 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
6304 case ISD::CTPOP: return LowerCTPOP(Op.getNode(), DAG, Subtarget);
6305 case ISD::SETCC: return LowerVSETCC(Op, DAG);
6306 case ISD::ConstantFP: return LowerConstantFP(Op, DAG, Subtarget);
6307 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG, Subtarget);
6308 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
6309 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
6310 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
6311 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
6312 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
6313 case ISD::MUL: return LowerMUL(Op, DAG);
6314 case ISD::SDIV: return LowerSDIV(Op, DAG);
6315 case ISD::UDIV: return LowerUDIV(Op, DAG);
6319 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
6324 return LowerXALUO(Op, DAG);
6325 case ISD::ATOMIC_LOAD:
6326 case ISD::ATOMIC_STORE: return LowerAtomicLoadStore(Op, DAG);
6327 case ISD::FSINCOS: return LowerFSINCOS(Op, DAG);
6329 case ISD::UDIVREM: return LowerDivRem(Op, DAG);
6330 case ISD::DYNAMIC_STACKALLOC:
6331 if (Subtarget->getTargetTriple().isWindowsItaniumEnvironment())
6332 return LowerDYNAMIC_STACKALLOC(Op, DAG);
6333 llvm_unreachable("Don't know how to custom lower this!");
6334 case ISD::FP_ROUND: return LowerFP_ROUND(Op, DAG);
6335 case ISD::FP_EXTEND: return LowerFP_EXTEND(Op, DAG);
6339 /// ReplaceNodeResults - Replace the results of node with an illegal result
6340 /// type with new values built out of custom code.
6341 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
6342 SmallVectorImpl<SDValue>&Results,
6343 SelectionDAG &DAG) const {
6345 switch (N->getOpcode()) {
6347 llvm_unreachable("Don't know how to custom expand this!");
6349 Res = ExpandBITCAST(N, DAG);
6353 Res = Expand64BitShift(N, DAG, Subtarget);
6355 case ISD::READCYCLECOUNTER:
6356 ReplaceREADCYCLECOUNTER(N, Results, DAG, Subtarget);
6360 Results.push_back(Res);
6363 //===----------------------------------------------------------------------===//
6364 // ARM Scheduler Hooks
6365 //===----------------------------------------------------------------------===//
6367 /// SetupEntryBlockForSjLj - Insert code into the entry block that creates and
6368 /// registers the function context.
6369 void ARMTargetLowering::
6370 SetupEntryBlockForSjLj(MachineInstr *MI, MachineBasicBlock *MBB,
6371 MachineBasicBlock *DispatchBB, int FI) const {
6372 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
6373 DebugLoc dl = MI->getDebugLoc();
6374 MachineFunction *MF = MBB->getParent();
6375 MachineRegisterInfo *MRI = &MF->getRegInfo();
6376 MachineConstantPool *MCP = MF->getConstantPool();
6377 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
6378 const Function *F = MF->getFunction();
6380 bool isThumb = Subtarget->isThumb();
6381 bool isThumb2 = Subtarget->isThumb2();
6383 unsigned PCLabelId = AFI->createPICLabelUId();
6384 unsigned PCAdj = (isThumb || isThumb2) ? 4 : 8;
6385 ARMConstantPoolValue *CPV =
6386 ARMConstantPoolMBB::Create(F->getContext(), DispatchBB, PCLabelId, PCAdj);
6387 unsigned CPI = MCP->getConstantPoolIndex(CPV, 4);
6389 const TargetRegisterClass *TRC = isThumb ? &ARM::tGPRRegClass
6390 : &ARM::GPRRegClass;
6392 // Grab constant pool and fixed stack memory operands.
6393 MachineMemOperand *CPMMO =
6394 MF->getMachineMemOperand(MachinePointerInfo::getConstantPool(),
6395 MachineMemOperand::MOLoad, 4, 4);
6397 MachineMemOperand *FIMMOSt =
6398 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
6399 MachineMemOperand::MOStore, 4, 4);
6401 // Load the address of the dispatch MBB into the jump buffer.
6403 // Incoming value: jbuf
6404 // ldr.n r5, LCPI1_1
6407 // str r5, [$jbuf, #+4] ; &jbuf[1]
6408 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6409 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2LDRpci), NewVReg1)
6410 .addConstantPoolIndex(CPI)
6411 .addMemOperand(CPMMO));
6412 // Set the low bit because of thumb mode.
6413 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6415 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2ORRri), NewVReg2)
6416 .addReg(NewVReg1, RegState::Kill)
6418 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6419 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg3)
6420 .addReg(NewVReg2, RegState::Kill)
6422 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2STRi12))
6423 .addReg(NewVReg3, RegState::Kill)
6425 .addImm(36) // &jbuf[1] :: pc
6426 .addMemOperand(FIMMOSt));
6427 } else if (isThumb) {
6428 // Incoming value: jbuf
6429 // ldr.n r1, LCPI1_4
6433 // add r2, $jbuf, #+4 ; &jbuf[1]
6435 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6436 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tLDRpci), NewVReg1)
6437 .addConstantPoolIndex(CPI)
6438 .addMemOperand(CPMMO));
6439 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6440 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg2)
6441 .addReg(NewVReg1, RegState::Kill)
6443 // Set the low bit because of thumb mode.
6444 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6445 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tMOVi8), NewVReg3)
6446 .addReg(ARM::CPSR, RegState::Define)
6448 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6449 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tORR), NewVReg4)
6450 .addReg(ARM::CPSR, RegState::Define)
6451 .addReg(NewVReg2, RegState::Kill)
6452 .addReg(NewVReg3, RegState::Kill));
6453 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6454 BuildMI(*MBB, MI, dl, TII->get(ARM::tADDframe), NewVReg5)
6456 .addImm(36); // &jbuf[1] :: pc
6457 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tSTRi))
6458 .addReg(NewVReg4, RegState::Kill)
6459 .addReg(NewVReg5, RegState::Kill)
6461 .addMemOperand(FIMMOSt));
6463 // Incoming value: jbuf
6466 // str r1, [$jbuf, #+4] ; &jbuf[1]
6467 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6468 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::LDRi12), NewVReg1)
6469 .addConstantPoolIndex(CPI)
6471 .addMemOperand(CPMMO));
6472 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6473 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::PICADD), NewVReg2)
6474 .addReg(NewVReg1, RegState::Kill)
6475 .addImm(PCLabelId));
6476 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::STRi12))
6477 .addReg(NewVReg2, RegState::Kill)
6479 .addImm(36) // &jbuf[1] :: pc
6480 .addMemOperand(FIMMOSt));
6484 MachineBasicBlock *ARMTargetLowering::
6485 EmitSjLjDispatchBlock(MachineInstr *MI, MachineBasicBlock *MBB) const {
6486 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
6487 DebugLoc dl = MI->getDebugLoc();
6488 MachineFunction *MF = MBB->getParent();
6489 MachineRegisterInfo *MRI = &MF->getRegInfo();
6490 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
6491 MachineFrameInfo *MFI = MF->getFrameInfo();
6492 int FI = MFI->getFunctionContextIndex();
6494 const TargetRegisterClass *TRC = Subtarget->isThumb() ? &ARM::tGPRRegClass
6495 : &ARM::GPRnopcRegClass;
6497 // Get a mapping of the call site numbers to all of the landing pads they're
6499 DenseMap<unsigned, SmallVector<MachineBasicBlock*, 2> > CallSiteNumToLPad;
6500 unsigned MaxCSNum = 0;
6501 MachineModuleInfo &MMI = MF->getMMI();
6502 for (MachineFunction::iterator BB = MF->begin(), E = MF->end(); BB != E;
6504 if (!BB->isLandingPad()) continue;
6506 // FIXME: We should assert that the EH_LABEL is the first MI in the landing
6508 for (MachineBasicBlock::iterator
6509 II = BB->begin(), IE = BB->end(); II != IE; ++II) {
6510 if (!II->isEHLabel()) continue;
6512 MCSymbol *Sym = II->getOperand(0).getMCSymbol();
6513 if (!MMI.hasCallSiteLandingPad(Sym)) continue;
6515 SmallVectorImpl<unsigned> &CallSiteIdxs = MMI.getCallSiteLandingPad(Sym);
6516 for (SmallVectorImpl<unsigned>::iterator
6517 CSI = CallSiteIdxs.begin(), CSE = CallSiteIdxs.end();
6518 CSI != CSE; ++CSI) {
6519 CallSiteNumToLPad[*CSI].push_back(BB);
6520 MaxCSNum = std::max(MaxCSNum, *CSI);
6526 // Get an ordered list of the machine basic blocks for the jump table.
6527 std::vector<MachineBasicBlock*> LPadList;
6528 SmallPtrSet<MachineBasicBlock*, 64> InvokeBBs;
6529 LPadList.reserve(CallSiteNumToLPad.size());
6530 for (unsigned I = 1; I <= MaxCSNum; ++I) {
6531 SmallVectorImpl<MachineBasicBlock*> &MBBList = CallSiteNumToLPad[I];
6532 for (SmallVectorImpl<MachineBasicBlock*>::iterator
6533 II = MBBList.begin(), IE = MBBList.end(); II != IE; ++II) {
6534 LPadList.push_back(*II);
6535 InvokeBBs.insert((*II)->pred_begin(), (*II)->pred_end());
6539 assert(!LPadList.empty() &&
6540 "No landing pad destinations for the dispatch jump table!");
6542 // Create the jump table and associated information.
6543 MachineJumpTableInfo *JTI =
6544 MF->getOrCreateJumpTableInfo(MachineJumpTableInfo::EK_Inline);
6545 unsigned MJTI = JTI->createJumpTableIndex(LPadList);
6546 unsigned UId = AFI->createJumpTableUId();
6547 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
6549 // Create the MBBs for the dispatch code.
6551 // Shove the dispatch's address into the return slot in the function context.
6552 MachineBasicBlock *DispatchBB = MF->CreateMachineBasicBlock();
6553 DispatchBB->setIsLandingPad();
6555 MachineBasicBlock *TrapBB = MF->CreateMachineBasicBlock();
6556 unsigned trap_opcode;
6557 if (Subtarget->isThumb())
6558 trap_opcode = ARM::tTRAP;
6560 trap_opcode = Subtarget->useNaClTrap() ? ARM::TRAPNaCl : ARM::TRAP;
6562 BuildMI(TrapBB, dl, TII->get(trap_opcode));
6563 DispatchBB->addSuccessor(TrapBB);
6565 MachineBasicBlock *DispContBB = MF->CreateMachineBasicBlock();
6566 DispatchBB->addSuccessor(DispContBB);
6569 MF->insert(MF->end(), DispatchBB);
6570 MF->insert(MF->end(), DispContBB);
6571 MF->insert(MF->end(), TrapBB);
6573 // Insert code into the entry block that creates and registers the function
6575 SetupEntryBlockForSjLj(MI, MBB, DispatchBB, FI);
6577 MachineMemOperand *FIMMOLd =
6578 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
6579 MachineMemOperand::MOLoad |
6580 MachineMemOperand::MOVolatile, 4, 4);
6582 MachineInstrBuilder MIB;
6583 MIB = BuildMI(DispatchBB, dl, TII->get(ARM::Int_eh_sjlj_dispatchsetup));
6585 const ARMBaseInstrInfo *AII = static_cast<const ARMBaseInstrInfo*>(TII);
6586 const ARMBaseRegisterInfo &RI = AII->getRegisterInfo();
6588 // Add a register mask with no preserved registers. This results in all
6589 // registers being marked as clobbered.
6590 MIB.addRegMask(RI.getNoPreservedMask());
6592 unsigned NumLPads = LPadList.size();
6593 if (Subtarget->isThumb2()) {
6594 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6595 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2LDRi12), NewVReg1)
6598 .addMemOperand(FIMMOLd));
6600 if (NumLPads < 256) {
6601 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPri))
6603 .addImm(LPadList.size()));
6605 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6606 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVi16), VReg1)
6607 .addImm(NumLPads & 0xFFFF));
6609 unsigned VReg2 = VReg1;
6610 if ((NumLPads & 0xFFFF0000) != 0) {
6611 VReg2 = MRI->createVirtualRegister(TRC);
6612 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVTi16), VReg2)
6614 .addImm(NumLPads >> 16));
6617 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPrr))
6622 BuildMI(DispatchBB, dl, TII->get(ARM::t2Bcc))
6627 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6628 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::t2LEApcrelJT),NewVReg3)
6629 .addJumpTableIndex(MJTI)
6632 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6635 BuildMI(DispContBB, dl, TII->get(ARM::t2ADDrs), NewVReg4)
6636 .addReg(NewVReg3, RegState::Kill)
6638 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
6640 BuildMI(DispContBB, dl, TII->get(ARM::t2BR_JT))
6641 .addReg(NewVReg4, RegState::Kill)
6643 .addJumpTableIndex(MJTI)
6645 } else if (Subtarget->isThumb()) {
6646 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6647 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRspi), NewVReg1)
6650 .addMemOperand(FIMMOLd));
6652 if (NumLPads < 256) {
6653 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPi8))
6657 MachineConstantPool *ConstantPool = MF->getConstantPool();
6658 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
6659 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
6661 // MachineConstantPool wants an explicit alignment.
6662 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
6664 Align = getDataLayout()->getTypeAllocSize(C->getType());
6665 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6667 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6668 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRpci))
6669 .addReg(VReg1, RegState::Define)
6670 .addConstantPoolIndex(Idx));
6671 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPr))
6676 BuildMI(DispatchBB, dl, TII->get(ARM::tBcc))
6681 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6682 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLSLri), NewVReg2)
6683 .addReg(ARM::CPSR, RegState::Define)
6687 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6688 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLEApcrelJT), NewVReg3)
6689 .addJumpTableIndex(MJTI)
6692 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6693 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg4)
6694 .addReg(ARM::CPSR, RegState::Define)
6695 .addReg(NewVReg2, RegState::Kill)
6698 MachineMemOperand *JTMMOLd =
6699 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
6700 MachineMemOperand::MOLoad, 4, 4);
6702 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6703 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLDRi), NewVReg5)
6704 .addReg(NewVReg4, RegState::Kill)
6706 .addMemOperand(JTMMOLd));
6708 unsigned NewVReg6 = NewVReg5;
6709 if (RelocM == Reloc::PIC_) {
6710 NewVReg6 = MRI->createVirtualRegister(TRC);
6711 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg6)
6712 .addReg(ARM::CPSR, RegState::Define)
6713 .addReg(NewVReg5, RegState::Kill)
6717 BuildMI(DispContBB, dl, TII->get(ARM::tBR_JTr))
6718 .addReg(NewVReg6, RegState::Kill)
6719 .addJumpTableIndex(MJTI)
6722 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6723 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRi12), NewVReg1)
6726 .addMemOperand(FIMMOLd));
6728 if (NumLPads < 256) {
6729 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPri))
6732 } else if (Subtarget->hasV6T2Ops() && isUInt<16>(NumLPads)) {
6733 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6734 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVi16), VReg1)
6735 .addImm(NumLPads & 0xFFFF));
6737 unsigned VReg2 = VReg1;
6738 if ((NumLPads & 0xFFFF0000) != 0) {
6739 VReg2 = MRI->createVirtualRegister(TRC);
6740 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVTi16), VReg2)
6742 .addImm(NumLPads >> 16));
6745 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
6749 MachineConstantPool *ConstantPool = MF->getConstantPool();
6750 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
6751 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
6753 // MachineConstantPool wants an explicit alignment.
6754 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
6756 Align = getDataLayout()->getTypeAllocSize(C->getType());
6757 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6759 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6760 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRcp))
6761 .addReg(VReg1, RegState::Define)
6762 .addConstantPoolIndex(Idx)
6764 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
6766 .addReg(VReg1, RegState::Kill));
6769 BuildMI(DispatchBB, dl, TII->get(ARM::Bcc))
6774 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6776 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::MOVsi), NewVReg3)
6778 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
6779 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6780 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::LEApcrelJT), NewVReg4)
6781 .addJumpTableIndex(MJTI)
6784 MachineMemOperand *JTMMOLd =
6785 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
6786 MachineMemOperand::MOLoad, 4, 4);
6787 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6789 BuildMI(DispContBB, dl, TII->get(ARM::LDRrs), NewVReg5)
6790 .addReg(NewVReg3, RegState::Kill)
6793 .addMemOperand(JTMMOLd));
6795 if (RelocM == Reloc::PIC_) {
6796 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTadd))
6797 .addReg(NewVReg5, RegState::Kill)
6799 .addJumpTableIndex(MJTI)
6802 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTr))
6803 .addReg(NewVReg5, RegState::Kill)
6804 .addJumpTableIndex(MJTI)
6809 // Add the jump table entries as successors to the MBB.
6810 SmallPtrSet<MachineBasicBlock*, 8> SeenMBBs;
6811 for (std::vector<MachineBasicBlock*>::iterator
6812 I = LPadList.begin(), E = LPadList.end(); I != E; ++I) {
6813 MachineBasicBlock *CurMBB = *I;
6814 if (SeenMBBs.insert(CurMBB).second)
6815 DispContBB->addSuccessor(CurMBB);
6818 // N.B. the order the invoke BBs are processed in doesn't matter here.
6819 const MCPhysReg *SavedRegs = RI.getCalleeSavedRegs(MF);
6820 SmallVector<MachineBasicBlock*, 64> MBBLPads;
6821 for (MachineBasicBlock *BB : InvokeBBs) {
6823 // Remove the landing pad successor from the invoke block and replace it
6824 // with the new dispatch block.
6825 SmallVector<MachineBasicBlock*, 4> Successors(BB->succ_begin(),
6827 while (!Successors.empty()) {
6828 MachineBasicBlock *SMBB = Successors.pop_back_val();
6829 if (SMBB->isLandingPad()) {
6830 BB->removeSuccessor(SMBB);
6831 MBBLPads.push_back(SMBB);
6835 BB->addSuccessor(DispatchBB);
6837 // Find the invoke call and mark all of the callee-saved registers as
6838 // 'implicit defined' so that they're spilled. This prevents code from
6839 // moving instructions to before the EH block, where they will never be
6841 for (MachineBasicBlock::reverse_iterator
6842 II = BB->rbegin(), IE = BB->rend(); II != IE; ++II) {
6843 if (!II->isCall()) continue;
6845 DenseMap<unsigned, bool> DefRegs;
6846 for (MachineInstr::mop_iterator
6847 OI = II->operands_begin(), OE = II->operands_end();
6849 if (!OI->isReg()) continue;
6850 DefRegs[OI->getReg()] = true;
6853 MachineInstrBuilder MIB(*MF, &*II);
6855 for (unsigned i = 0; SavedRegs[i] != 0; ++i) {
6856 unsigned Reg = SavedRegs[i];
6857 if (Subtarget->isThumb2() &&
6858 !ARM::tGPRRegClass.contains(Reg) &&
6859 !ARM::hGPRRegClass.contains(Reg))
6861 if (Subtarget->isThumb1Only() && !ARM::tGPRRegClass.contains(Reg))
6863 if (!Subtarget->isThumb() && !ARM::GPRRegClass.contains(Reg))
6866 MIB.addReg(Reg, RegState::ImplicitDefine | RegState::Dead);
6873 // Mark all former landing pads as non-landing pads. The dispatch is the only
6875 for (SmallVectorImpl<MachineBasicBlock*>::iterator
6876 I = MBBLPads.begin(), E = MBBLPads.end(); I != E; ++I)
6877 (*I)->setIsLandingPad(false);
6879 // The instruction is gone now.
6880 MI->eraseFromParent();
6886 MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) {
6887 for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
6888 E = MBB->succ_end(); I != E; ++I)
6891 llvm_unreachable("Expecting a BB with two successors!");
6894 /// Return the load opcode for a given load size. If load size >= 8,
6895 /// neon opcode will be returned.
6896 static unsigned getLdOpcode(unsigned LdSize, bool IsThumb1, bool IsThumb2) {
6898 return LdSize == 16 ? ARM::VLD1q32wb_fixed
6899 : LdSize == 8 ? ARM::VLD1d32wb_fixed : 0;
6901 return LdSize == 4 ? ARM::tLDRi
6902 : LdSize == 2 ? ARM::tLDRHi
6903 : LdSize == 1 ? ARM::tLDRBi : 0;
6905 return LdSize == 4 ? ARM::t2LDR_POST
6906 : LdSize == 2 ? ARM::t2LDRH_POST
6907 : LdSize == 1 ? ARM::t2LDRB_POST : 0;
6908 return LdSize == 4 ? ARM::LDR_POST_IMM
6909 : LdSize == 2 ? ARM::LDRH_POST
6910 : LdSize == 1 ? ARM::LDRB_POST_IMM : 0;
6913 /// Return the store opcode for a given store size. If store size >= 8,
6914 /// neon opcode will be returned.
6915 static unsigned getStOpcode(unsigned StSize, bool IsThumb1, bool IsThumb2) {
6917 return StSize == 16 ? ARM::VST1q32wb_fixed
6918 : StSize == 8 ? ARM::VST1d32wb_fixed : 0;
6920 return StSize == 4 ? ARM::tSTRi
6921 : StSize == 2 ? ARM::tSTRHi
6922 : StSize == 1 ? ARM::tSTRBi : 0;
6924 return StSize == 4 ? ARM::t2STR_POST
6925 : StSize == 2 ? ARM::t2STRH_POST
6926 : StSize == 1 ? ARM::t2STRB_POST : 0;
6927 return StSize == 4 ? ARM::STR_POST_IMM
6928 : StSize == 2 ? ARM::STRH_POST
6929 : StSize == 1 ? ARM::STRB_POST_IMM : 0;
6932 /// Emit a post-increment load operation with given size. The instructions
6933 /// will be added to BB at Pos.
6934 static void emitPostLd(MachineBasicBlock *BB, MachineInstr *Pos,
6935 const TargetInstrInfo *TII, DebugLoc dl,
6936 unsigned LdSize, unsigned Data, unsigned AddrIn,
6937 unsigned AddrOut, bool IsThumb1, bool IsThumb2) {
6938 unsigned LdOpc = getLdOpcode(LdSize, IsThumb1, IsThumb2);
6939 assert(LdOpc != 0 && "Should have a load opcode");
6941 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
6942 .addReg(AddrOut, RegState::Define).addReg(AddrIn)
6944 } else if (IsThumb1) {
6945 // load + update AddrIn
6946 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
6947 .addReg(AddrIn).addImm(0));
6948 MachineInstrBuilder MIB =
6949 BuildMI(*BB, Pos, dl, TII->get(ARM::tADDi8), AddrOut);
6950 MIB = AddDefaultT1CC(MIB);
6951 MIB.addReg(AddrIn).addImm(LdSize);
6952 AddDefaultPred(MIB);
6953 } else if (IsThumb2) {
6954 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
6955 .addReg(AddrOut, RegState::Define).addReg(AddrIn)
6958 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
6959 .addReg(AddrOut, RegState::Define).addReg(AddrIn)
6960 .addReg(0).addImm(LdSize));
6964 /// Emit a post-increment store operation with given size. The instructions
6965 /// will be added to BB at Pos.
6966 static void emitPostSt(MachineBasicBlock *BB, MachineInstr *Pos,
6967 const TargetInstrInfo *TII, DebugLoc dl,
6968 unsigned StSize, unsigned Data, unsigned AddrIn,
6969 unsigned AddrOut, bool IsThumb1, bool IsThumb2) {
6970 unsigned StOpc = getStOpcode(StSize, IsThumb1, IsThumb2);
6971 assert(StOpc != 0 && "Should have a store opcode");
6973 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc), AddrOut)
6974 .addReg(AddrIn).addImm(0).addReg(Data));
6975 } else if (IsThumb1) {
6976 // store + update AddrIn
6977 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc)).addReg(Data)
6978 .addReg(AddrIn).addImm(0));
6979 MachineInstrBuilder MIB =
6980 BuildMI(*BB, Pos, dl, TII->get(ARM::tADDi8), AddrOut);
6981 MIB = AddDefaultT1CC(MIB);
6982 MIB.addReg(AddrIn).addImm(StSize);
6983 AddDefaultPred(MIB);
6984 } else if (IsThumb2) {
6985 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc), AddrOut)
6986 .addReg(Data).addReg(AddrIn).addImm(StSize));
6988 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc), AddrOut)
6989 .addReg(Data).addReg(AddrIn).addReg(0)
6995 ARMTargetLowering::EmitStructByval(MachineInstr *MI,
6996 MachineBasicBlock *BB) const {
6997 // This pseudo instruction has 3 operands: dst, src, size
6998 // We expand it to a loop if size > Subtarget->getMaxInlineSizeThreshold().
6999 // Otherwise, we will generate unrolled scalar copies.
7000 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
7001 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7002 MachineFunction::iterator It = BB;
7005 unsigned dest = MI->getOperand(0).getReg();
7006 unsigned src = MI->getOperand(1).getReg();
7007 unsigned SizeVal = MI->getOperand(2).getImm();
7008 unsigned Align = MI->getOperand(3).getImm();
7009 DebugLoc dl = MI->getDebugLoc();
7011 MachineFunction *MF = BB->getParent();
7012 MachineRegisterInfo &MRI = MF->getRegInfo();
7013 unsigned UnitSize = 0;
7014 const TargetRegisterClass *TRC = nullptr;
7015 const TargetRegisterClass *VecTRC = nullptr;
7017 bool IsThumb1 = Subtarget->isThumb1Only();
7018 bool IsThumb2 = Subtarget->isThumb2();
7022 } else if (Align & 2) {
7025 // Check whether we can use NEON instructions.
7026 if (!MF->getFunction()->hasFnAttribute(Attribute::NoImplicitFloat) &&
7027 Subtarget->hasNEON()) {
7028 if ((Align % 16 == 0) && SizeVal >= 16)
7030 else if ((Align % 8 == 0) && SizeVal >= 8)
7033 // Can't use NEON instructions.
7038 // Select the correct opcode and register class for unit size load/store
7039 bool IsNeon = UnitSize >= 8;
7040 TRC = (IsThumb1 || IsThumb2) ? &ARM::tGPRRegClass : &ARM::GPRRegClass;
7042 VecTRC = UnitSize == 16 ? &ARM::DPairRegClass
7043 : UnitSize == 8 ? &ARM::DPRRegClass
7046 unsigned BytesLeft = SizeVal % UnitSize;
7047 unsigned LoopSize = SizeVal - BytesLeft;
7049 if (SizeVal <= Subtarget->getMaxInlineSizeThreshold()) {
7050 // Use LDR and STR to copy.
7051 // [scratch, srcOut] = LDR_POST(srcIn, UnitSize)
7052 // [destOut] = STR_POST(scratch, destIn, UnitSize)
7053 unsigned srcIn = src;
7054 unsigned destIn = dest;
7055 for (unsigned i = 0; i < LoopSize; i+=UnitSize) {
7056 unsigned srcOut = MRI.createVirtualRegister(TRC);
7057 unsigned destOut = MRI.createVirtualRegister(TRC);
7058 unsigned scratch = MRI.createVirtualRegister(IsNeon ? VecTRC : TRC);
7059 emitPostLd(BB, MI, TII, dl, UnitSize, scratch, srcIn, srcOut,
7060 IsThumb1, IsThumb2);
7061 emitPostSt(BB, MI, TII, dl, UnitSize, scratch, destIn, destOut,
7062 IsThumb1, IsThumb2);
7067 // Handle the leftover bytes with LDRB and STRB.
7068 // [scratch, srcOut] = LDRB_POST(srcIn, 1)
7069 // [destOut] = STRB_POST(scratch, destIn, 1)
7070 for (unsigned i = 0; i < BytesLeft; i++) {
7071 unsigned srcOut = MRI.createVirtualRegister(TRC);
7072 unsigned destOut = MRI.createVirtualRegister(TRC);
7073 unsigned scratch = MRI.createVirtualRegister(TRC);
7074 emitPostLd(BB, MI, TII, dl, 1, scratch, srcIn, srcOut,
7075 IsThumb1, IsThumb2);
7076 emitPostSt(BB, MI, TII, dl, 1, scratch, destIn, destOut,
7077 IsThumb1, IsThumb2);
7081 MI->eraseFromParent(); // The instruction is gone now.
7085 // Expand the pseudo op to a loop.
7088 // movw varEnd, # --> with thumb2
7090 // ldrcp varEnd, idx --> without thumb2
7091 // fallthrough --> loopMBB
7093 // PHI varPhi, varEnd, varLoop
7094 // PHI srcPhi, src, srcLoop
7095 // PHI destPhi, dst, destLoop
7096 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
7097 // [destLoop] = STR_POST(scratch, destPhi, UnitSize)
7098 // subs varLoop, varPhi, #UnitSize
7100 // fallthrough --> exitMBB
7102 // epilogue to handle left-over bytes
7103 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
7104 // [destOut] = STRB_POST(scratch, destLoop, 1)
7105 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
7106 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
7107 MF->insert(It, loopMBB);
7108 MF->insert(It, exitMBB);
7110 // Transfer the remainder of BB and its successor edges to exitMBB.
7111 exitMBB->splice(exitMBB->begin(), BB,
7112 std::next(MachineBasicBlock::iterator(MI)), BB->end());
7113 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
7115 // Load an immediate to varEnd.
7116 unsigned varEnd = MRI.createVirtualRegister(TRC);
7117 if (Subtarget->useMovt(*MF)) {
7118 unsigned Vtmp = varEnd;
7119 if ((LoopSize & 0xFFFF0000) != 0)
7120 Vtmp = MRI.createVirtualRegister(TRC);
7121 AddDefaultPred(BuildMI(BB, dl,
7122 TII->get(IsThumb2 ? ARM::t2MOVi16 : ARM::MOVi16),
7123 Vtmp).addImm(LoopSize & 0xFFFF));
7125 if ((LoopSize & 0xFFFF0000) != 0)
7126 AddDefaultPred(BuildMI(BB, dl,
7127 TII->get(IsThumb2 ? ARM::t2MOVTi16 : ARM::MOVTi16),
7130 .addImm(LoopSize >> 16));
7132 MachineConstantPool *ConstantPool = MF->getConstantPool();
7133 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
7134 const Constant *C = ConstantInt::get(Int32Ty, LoopSize);
7136 // MachineConstantPool wants an explicit alignment.
7137 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
7139 Align = getDataLayout()->getTypeAllocSize(C->getType());
7140 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
7143 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(ARM::tLDRpci)).addReg(
7144 varEnd, RegState::Define).addConstantPoolIndex(Idx));
7146 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(ARM::LDRcp)).addReg(
7147 varEnd, RegState::Define).addConstantPoolIndex(Idx).addImm(0));
7149 BB->addSuccessor(loopMBB);
7151 // Generate the loop body:
7152 // varPhi = PHI(varLoop, varEnd)
7153 // srcPhi = PHI(srcLoop, src)
7154 // destPhi = PHI(destLoop, dst)
7155 MachineBasicBlock *entryBB = BB;
7157 unsigned varLoop = MRI.createVirtualRegister(TRC);
7158 unsigned varPhi = MRI.createVirtualRegister(TRC);
7159 unsigned srcLoop = MRI.createVirtualRegister(TRC);
7160 unsigned srcPhi = MRI.createVirtualRegister(TRC);
7161 unsigned destLoop = MRI.createVirtualRegister(TRC);
7162 unsigned destPhi = MRI.createVirtualRegister(TRC);
7164 BuildMI(*BB, BB->begin(), dl, TII->get(ARM::PHI), varPhi)
7165 .addReg(varLoop).addMBB(loopMBB)
7166 .addReg(varEnd).addMBB(entryBB);
7167 BuildMI(BB, dl, TII->get(ARM::PHI), srcPhi)
7168 .addReg(srcLoop).addMBB(loopMBB)
7169 .addReg(src).addMBB(entryBB);
7170 BuildMI(BB, dl, TII->get(ARM::PHI), destPhi)
7171 .addReg(destLoop).addMBB(loopMBB)
7172 .addReg(dest).addMBB(entryBB);
7174 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
7175 // [destLoop] = STR_POST(scratch, destPhi, UnitSiz)
7176 unsigned scratch = MRI.createVirtualRegister(IsNeon ? VecTRC : TRC);
7177 emitPostLd(BB, BB->end(), TII, dl, UnitSize, scratch, srcPhi, srcLoop,
7178 IsThumb1, IsThumb2);
7179 emitPostSt(BB, BB->end(), TII, dl, UnitSize, scratch, destPhi, destLoop,
7180 IsThumb1, IsThumb2);
7182 // Decrement loop variable by UnitSize.
7184 MachineInstrBuilder MIB =
7185 BuildMI(*BB, BB->end(), dl, TII->get(ARM::tSUBi8), varLoop);
7186 MIB = AddDefaultT1CC(MIB);
7187 MIB.addReg(varPhi).addImm(UnitSize);
7188 AddDefaultPred(MIB);
7190 MachineInstrBuilder MIB =
7191 BuildMI(*BB, BB->end(), dl,
7192 TII->get(IsThumb2 ? ARM::t2SUBri : ARM::SUBri), varLoop);
7193 AddDefaultCC(AddDefaultPred(MIB.addReg(varPhi).addImm(UnitSize)));
7194 MIB->getOperand(5).setReg(ARM::CPSR);
7195 MIB->getOperand(5).setIsDef(true);
7197 BuildMI(*BB, BB->end(), dl,
7198 TII->get(IsThumb1 ? ARM::tBcc : IsThumb2 ? ARM::t2Bcc : ARM::Bcc))
7199 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
7201 // loopMBB can loop back to loopMBB or fall through to exitMBB.
7202 BB->addSuccessor(loopMBB);
7203 BB->addSuccessor(exitMBB);
7205 // Add epilogue to handle BytesLeft.
7207 MachineInstr *StartOfExit = exitMBB->begin();
7209 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
7210 // [destOut] = STRB_POST(scratch, destLoop, 1)
7211 unsigned srcIn = srcLoop;
7212 unsigned destIn = destLoop;
7213 for (unsigned i = 0; i < BytesLeft; i++) {
7214 unsigned srcOut = MRI.createVirtualRegister(TRC);
7215 unsigned destOut = MRI.createVirtualRegister(TRC);
7216 unsigned scratch = MRI.createVirtualRegister(TRC);
7217 emitPostLd(BB, StartOfExit, TII, dl, 1, scratch, srcIn, srcOut,
7218 IsThumb1, IsThumb2);
7219 emitPostSt(BB, StartOfExit, TII, dl, 1, scratch, destIn, destOut,
7220 IsThumb1, IsThumb2);
7225 MI->eraseFromParent(); // The instruction is gone now.
7230 ARMTargetLowering::EmitLowered__chkstk(MachineInstr *MI,
7231 MachineBasicBlock *MBB) const {
7232 const TargetMachine &TM = getTargetMachine();
7233 const TargetInstrInfo &TII = *Subtarget->getInstrInfo();
7234 DebugLoc DL = MI->getDebugLoc();
7236 assert(Subtarget->isTargetWindows() &&
7237 "__chkstk is only supported on Windows");
7238 assert(Subtarget->isThumb2() && "Windows on ARM requires Thumb-2 mode");
7240 // __chkstk takes the number of words to allocate on the stack in R4, and
7241 // returns the stack adjustment in number of bytes in R4. This will not
7242 // clober any other registers (other than the obvious lr).
7244 // Although, technically, IP should be considered a register which may be
7245 // clobbered, the call itself will not touch it. Windows on ARM is a pure
7246 // thumb-2 environment, so there is no interworking required. As a result, we
7247 // do not expect a veneer to be emitted by the linker, clobbering IP.
7249 // Each module receives its own copy of __chkstk, so no import thunk is
7250 // required, again, ensuring that IP is not clobbered.
7252 // Finally, although some linkers may theoretically provide a trampoline for
7253 // out of range calls (which is quite common due to a 32M range limitation of
7254 // branches for Thumb), we can generate the long-call version via
7255 // -mcmodel=large, alleviating the need for the trampoline which may clobber
7258 switch (TM.getCodeModel()) {
7259 case CodeModel::Small:
7260 case CodeModel::Medium:
7261 case CodeModel::Default:
7262 case CodeModel::Kernel:
7263 BuildMI(*MBB, MI, DL, TII.get(ARM::tBL))
7264 .addImm((unsigned)ARMCC::AL).addReg(0)
7265 .addExternalSymbol("__chkstk")
7266 .addReg(ARM::R4, RegState::Implicit | RegState::Kill)
7267 .addReg(ARM::R4, RegState::Implicit | RegState::Define)
7268 .addReg(ARM::R12, RegState::Implicit | RegState::Define | RegState::Dead);
7270 case CodeModel::Large:
7271 case CodeModel::JITDefault: {
7272 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
7273 unsigned Reg = MRI.createVirtualRegister(&ARM::rGPRRegClass);
7275 BuildMI(*MBB, MI, DL, TII.get(ARM::t2MOVi32imm), Reg)
7276 .addExternalSymbol("__chkstk");
7277 BuildMI(*MBB, MI, DL, TII.get(ARM::tBLXr))
7278 .addImm((unsigned)ARMCC::AL).addReg(0)
7279 .addReg(Reg, RegState::Kill)
7280 .addReg(ARM::R4, RegState::Implicit | RegState::Kill)
7281 .addReg(ARM::R4, RegState::Implicit | RegState::Define)
7282 .addReg(ARM::R12, RegState::Implicit | RegState::Define | RegState::Dead);
7287 AddDefaultCC(AddDefaultPred(BuildMI(*MBB, MI, DL, TII.get(ARM::t2SUBrr),
7289 .addReg(ARM::SP).addReg(ARM::R4)));
7291 MI->eraseFromParent();
7296 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
7297 MachineBasicBlock *BB) const {
7298 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
7299 DebugLoc dl = MI->getDebugLoc();
7300 bool isThumb2 = Subtarget->isThumb2();
7301 switch (MI->getOpcode()) {
7304 llvm_unreachable("Unexpected instr type to insert");
7306 // The Thumb2 pre-indexed stores have the same MI operands, they just
7307 // define them differently in the .td files from the isel patterns, so
7308 // they need pseudos.
7309 case ARM::t2STR_preidx:
7310 MI->setDesc(TII->get(ARM::t2STR_PRE));
7312 case ARM::t2STRB_preidx:
7313 MI->setDesc(TII->get(ARM::t2STRB_PRE));
7315 case ARM::t2STRH_preidx:
7316 MI->setDesc(TII->get(ARM::t2STRH_PRE));
7319 case ARM::STRi_preidx:
7320 case ARM::STRBi_preidx: {
7321 unsigned NewOpc = MI->getOpcode() == ARM::STRi_preidx ?
7322 ARM::STR_PRE_IMM : ARM::STRB_PRE_IMM;
7323 // Decode the offset.
7324 unsigned Offset = MI->getOperand(4).getImm();
7325 bool isSub = ARM_AM::getAM2Op(Offset) == ARM_AM::sub;
7326 Offset = ARM_AM::getAM2Offset(Offset);
7330 MachineMemOperand *MMO = *MI->memoperands_begin();
7331 BuildMI(*BB, MI, dl, TII->get(NewOpc))
7332 .addOperand(MI->getOperand(0)) // Rn_wb
7333 .addOperand(MI->getOperand(1)) // Rt
7334 .addOperand(MI->getOperand(2)) // Rn
7335 .addImm(Offset) // offset (skip GPR==zero_reg)
7336 .addOperand(MI->getOperand(5)) // pred
7337 .addOperand(MI->getOperand(6))
7338 .addMemOperand(MMO);
7339 MI->eraseFromParent();
7342 case ARM::STRr_preidx:
7343 case ARM::STRBr_preidx:
7344 case ARM::STRH_preidx: {
7346 switch (MI->getOpcode()) {
7347 default: llvm_unreachable("unexpected opcode!");
7348 case ARM::STRr_preidx: NewOpc = ARM::STR_PRE_REG; break;
7349 case ARM::STRBr_preidx: NewOpc = ARM::STRB_PRE_REG; break;
7350 case ARM::STRH_preidx: NewOpc = ARM::STRH_PRE; break;
7352 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(NewOpc));
7353 for (unsigned i = 0; i < MI->getNumOperands(); ++i)
7354 MIB.addOperand(MI->getOperand(i));
7355 MI->eraseFromParent();
7359 case ARM::tMOVCCr_pseudo: {
7360 // To "insert" a SELECT_CC instruction, we actually have to insert the
7361 // diamond control-flow pattern. The incoming instruction knows the
7362 // destination vreg to set, the condition code register to branch on, the
7363 // true/false values to select between, and a branch opcode to use.
7364 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7365 MachineFunction::iterator It = BB;
7371 // cmpTY ccX, r1, r2
7373 // fallthrough --> copy0MBB
7374 MachineBasicBlock *thisMBB = BB;
7375 MachineFunction *F = BB->getParent();
7376 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
7377 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
7378 F->insert(It, copy0MBB);
7379 F->insert(It, sinkMBB);
7381 // Transfer the remainder of BB and its successor edges to sinkMBB.
7382 sinkMBB->splice(sinkMBB->begin(), BB,
7383 std::next(MachineBasicBlock::iterator(MI)), BB->end());
7384 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
7386 BB->addSuccessor(copy0MBB);
7387 BB->addSuccessor(sinkMBB);
7389 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
7390 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
7393 // %FalseValue = ...
7394 // # fallthrough to sinkMBB
7397 // Update machine-CFG edges
7398 BB->addSuccessor(sinkMBB);
7401 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
7404 BuildMI(*BB, BB->begin(), dl,
7405 TII->get(ARM::PHI), MI->getOperand(0).getReg())
7406 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
7407 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
7409 MI->eraseFromParent(); // The pseudo instruction is gone now.
7414 case ARM::BCCZi64: {
7415 // If there is an unconditional branch to the other successor, remove it.
7416 BB->erase(std::next(MachineBasicBlock::iterator(MI)), BB->end());
7418 // Compare both parts that make up the double comparison separately for
7420 bool RHSisZero = MI->getOpcode() == ARM::BCCZi64;
7422 unsigned LHS1 = MI->getOperand(1).getReg();
7423 unsigned LHS2 = MI->getOperand(2).getReg();
7425 AddDefaultPred(BuildMI(BB, dl,
7426 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7427 .addReg(LHS1).addImm(0));
7428 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7429 .addReg(LHS2).addImm(0)
7430 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
7432 unsigned RHS1 = MI->getOperand(3).getReg();
7433 unsigned RHS2 = MI->getOperand(4).getReg();
7434 AddDefaultPred(BuildMI(BB, dl,
7435 TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
7436 .addReg(LHS1).addReg(RHS1));
7437 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
7438 .addReg(LHS2).addReg(RHS2)
7439 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
7442 MachineBasicBlock *destMBB = MI->getOperand(RHSisZero ? 3 : 5).getMBB();
7443 MachineBasicBlock *exitMBB = OtherSucc(BB, destMBB);
7444 if (MI->getOperand(0).getImm() == ARMCC::NE)
7445 std::swap(destMBB, exitMBB);
7447 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
7448 .addMBB(destMBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
7450 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2B)).addMBB(exitMBB));
7452 BuildMI(BB, dl, TII->get(ARM::B)) .addMBB(exitMBB);
7454 MI->eraseFromParent(); // The pseudo instruction is gone now.
7458 case ARM::Int_eh_sjlj_setjmp:
7459 case ARM::Int_eh_sjlj_setjmp_nofp:
7460 case ARM::tInt_eh_sjlj_setjmp:
7461 case ARM::t2Int_eh_sjlj_setjmp:
7462 case ARM::t2Int_eh_sjlj_setjmp_nofp:
7463 EmitSjLjDispatchBlock(MI, BB);
7468 // To insert an ABS instruction, we have to insert the
7469 // diamond control-flow pattern. The incoming instruction knows the
7470 // source vreg to test against 0, the destination vreg to set,
7471 // the condition code register to branch on, the
7472 // true/false values to select between, and a branch opcode to use.
7477 // BCC (branch to SinkBB if V0 >= 0)
7478 // RSBBB: V3 = RSBri V2, 0 (compute ABS if V2 < 0)
7479 // SinkBB: V1 = PHI(V2, V3)
7480 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7481 MachineFunction::iterator BBI = BB;
7483 MachineFunction *Fn = BB->getParent();
7484 MachineBasicBlock *RSBBB = Fn->CreateMachineBasicBlock(LLVM_BB);
7485 MachineBasicBlock *SinkBB = Fn->CreateMachineBasicBlock(LLVM_BB);
7486 Fn->insert(BBI, RSBBB);
7487 Fn->insert(BBI, SinkBB);
7489 unsigned int ABSSrcReg = MI->getOperand(1).getReg();
7490 unsigned int ABSDstReg = MI->getOperand(0).getReg();
7491 bool isThumb2 = Subtarget->isThumb2();
7492 MachineRegisterInfo &MRI = Fn->getRegInfo();
7493 // In Thumb mode S must not be specified if source register is the SP or
7494 // PC and if destination register is the SP, so restrict register class
7495 unsigned NewRsbDstReg =
7496 MRI.createVirtualRegister(isThumb2 ? &ARM::rGPRRegClass : &ARM::GPRRegClass);
7498 // Transfer the remainder of BB and its successor edges to sinkMBB.
7499 SinkBB->splice(SinkBB->begin(), BB,
7500 std::next(MachineBasicBlock::iterator(MI)), BB->end());
7501 SinkBB->transferSuccessorsAndUpdatePHIs(BB);
7503 BB->addSuccessor(RSBBB);
7504 BB->addSuccessor(SinkBB);
7506 // fall through to SinkMBB
7507 RSBBB->addSuccessor(SinkBB);
7509 // insert a cmp at the end of BB
7510 AddDefaultPred(BuildMI(BB, dl,
7511 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7512 .addReg(ABSSrcReg).addImm(0));
7514 // insert a bcc with opposite CC to ARMCC::MI at the end of BB
7516 TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)).addMBB(SinkBB)
7517 .addImm(ARMCC::getOppositeCondition(ARMCC::MI)).addReg(ARM::CPSR);
7519 // insert rsbri in RSBBB
7520 // Note: BCC and rsbri will be converted into predicated rsbmi
7521 // by if-conversion pass
7522 BuildMI(*RSBBB, RSBBB->begin(), dl,
7523 TII->get(isThumb2 ? ARM::t2RSBri : ARM::RSBri), NewRsbDstReg)
7524 .addReg(ABSSrcReg, RegState::Kill)
7525 .addImm(0).addImm((unsigned)ARMCC::AL).addReg(0).addReg(0);
7527 // insert PHI in SinkBB,
7528 // reuse ABSDstReg to not change uses of ABS instruction
7529 BuildMI(*SinkBB, SinkBB->begin(), dl,
7530 TII->get(ARM::PHI), ABSDstReg)
7531 .addReg(NewRsbDstReg).addMBB(RSBBB)
7532 .addReg(ABSSrcReg).addMBB(BB);
7534 // remove ABS instruction
7535 MI->eraseFromParent();
7537 // return last added BB
7540 case ARM::COPY_STRUCT_BYVAL_I32:
7542 return EmitStructByval(MI, BB);
7543 case ARM::WIN__CHKSTK:
7544 return EmitLowered__chkstk(MI, BB);
7548 void ARMTargetLowering::AdjustInstrPostInstrSelection(MachineInstr *MI,
7549 SDNode *Node) const {
7550 const MCInstrDesc *MCID = &MI->getDesc();
7551 // Adjust potentially 's' setting instructions after isel, i.e. ADC, SBC, RSB,
7552 // RSC. Coming out of isel, they have an implicit CPSR def, but the optional
7553 // operand is still set to noreg. If needed, set the optional operand's
7554 // register to CPSR, and remove the redundant implicit def.
7556 // e.g. ADCS (..., CPSR<imp-def>) -> ADC (... opt:CPSR<def>).
7558 // Rename pseudo opcodes.
7559 unsigned NewOpc = convertAddSubFlagsOpcode(MI->getOpcode());
7561 const ARMBaseInstrInfo *TII = Subtarget->getInstrInfo();
7562 MCID = &TII->get(NewOpc);
7564 assert(MCID->getNumOperands() == MI->getDesc().getNumOperands() + 1 &&
7565 "converted opcode should be the same except for cc_out");
7569 // Add the optional cc_out operand
7570 MI->addOperand(MachineOperand::CreateReg(0, /*isDef=*/true));
7572 unsigned ccOutIdx = MCID->getNumOperands() - 1;
7574 // Any ARM instruction that sets the 's' bit should specify an optional
7575 // "cc_out" operand in the last operand position.
7576 if (!MI->hasOptionalDef() || !MCID->OpInfo[ccOutIdx].isOptionalDef()) {
7577 assert(!NewOpc && "Optional cc_out operand required");
7580 // Look for an implicit def of CPSR added by MachineInstr ctor. Remove it
7581 // since we already have an optional CPSR def.
7582 bool definesCPSR = false;
7583 bool deadCPSR = false;
7584 for (unsigned i = MCID->getNumOperands(), e = MI->getNumOperands();
7586 const MachineOperand &MO = MI->getOperand(i);
7587 if (MO.isReg() && MO.isDef() && MO.getReg() == ARM::CPSR) {
7591 MI->RemoveOperand(i);
7596 assert(!NewOpc && "Optional cc_out operand required");
7599 assert(deadCPSR == !Node->hasAnyUseOfValue(1) && "inconsistent dead flag");
7601 assert(!MI->getOperand(ccOutIdx).getReg() &&
7602 "expect uninitialized optional cc_out operand");
7606 // If this instruction was defined with an optional CPSR def and its dag node
7607 // had a live implicit CPSR def, then activate the optional CPSR def.
7608 MachineOperand &MO = MI->getOperand(ccOutIdx);
7609 MO.setReg(ARM::CPSR);
7613 //===----------------------------------------------------------------------===//
7614 // ARM Optimization Hooks
7615 //===----------------------------------------------------------------------===//
7617 // Helper function that checks if N is a null or all ones constant.
7618 static inline bool isZeroOrAllOnes(SDValue N, bool AllOnes) {
7619 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N);
7622 return AllOnes ? C->isAllOnesValue() : C->isNullValue();
7625 // Return true if N is conditionally 0 or all ones.
7626 // Detects these expressions where cc is an i1 value:
7628 // (select cc 0, y) [AllOnes=0]
7629 // (select cc y, 0) [AllOnes=0]
7630 // (zext cc) [AllOnes=0]
7631 // (sext cc) [AllOnes=0/1]
7632 // (select cc -1, y) [AllOnes=1]
7633 // (select cc y, -1) [AllOnes=1]
7635 // Invert is set when N is the null/all ones constant when CC is false.
7636 // OtherOp is set to the alternative value of N.
7637 static bool isConditionalZeroOrAllOnes(SDNode *N, bool AllOnes,
7638 SDValue &CC, bool &Invert,
7640 SelectionDAG &DAG) {
7641 switch (N->getOpcode()) {
7642 default: return false;
7644 CC = N->getOperand(0);
7645 SDValue N1 = N->getOperand(1);
7646 SDValue N2 = N->getOperand(2);
7647 if (isZeroOrAllOnes(N1, AllOnes)) {
7652 if (isZeroOrAllOnes(N2, AllOnes)) {
7659 case ISD::ZERO_EXTEND:
7660 // (zext cc) can never be the all ones value.
7664 case ISD::SIGN_EXTEND: {
7665 EVT VT = N->getValueType(0);
7666 CC = N->getOperand(0);
7667 if (CC.getValueType() != MVT::i1)
7671 // When looking for an AllOnes constant, N is an sext, and the 'other'
7673 OtherOp = DAG.getConstant(0, VT);
7674 else if (N->getOpcode() == ISD::ZERO_EXTEND)
7675 // When looking for a 0 constant, N can be zext or sext.
7676 OtherOp = DAG.getConstant(1, VT);
7678 OtherOp = DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), VT);
7684 // Combine a constant select operand into its use:
7686 // (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
7687 // (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
7688 // (and (select cc, -1, c), x) -> (select cc, x, (and, x, c)) [AllOnes=1]
7689 // (or (select cc, 0, c), x) -> (select cc, x, (or, x, c))
7690 // (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c))
7692 // The transform is rejected if the select doesn't have a constant operand that
7693 // is null, or all ones when AllOnes is set.
7695 // Also recognize sext/zext from i1:
7697 // (add (zext cc), x) -> (select cc (add x, 1), x)
7698 // (add (sext cc), x) -> (select cc (add x, -1), x)
7700 // These transformations eventually create predicated instructions.
7702 // @param N The node to transform.
7703 // @param Slct The N operand that is a select.
7704 // @param OtherOp The other N operand (x above).
7705 // @param DCI Context.
7706 // @param AllOnes Require the select constant to be all ones instead of null.
7707 // @returns The new node, or SDValue() on failure.
7709 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
7710 TargetLowering::DAGCombinerInfo &DCI,
7711 bool AllOnes = false) {
7712 SelectionDAG &DAG = DCI.DAG;
7713 EVT VT = N->getValueType(0);
7714 SDValue NonConstantVal;
7717 if (!isConditionalZeroOrAllOnes(Slct.getNode(), AllOnes, CCOp, SwapSelectOps,
7718 NonConstantVal, DAG))
7721 // Slct is now know to be the desired identity constant when CC is true.
7722 SDValue TrueVal = OtherOp;
7723 SDValue FalseVal = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
7724 OtherOp, NonConstantVal);
7725 // Unless SwapSelectOps says CC should be false.
7727 std::swap(TrueVal, FalseVal);
7729 return DAG.getNode(ISD::SELECT, SDLoc(N), VT,
7730 CCOp, TrueVal, FalseVal);
7733 // Attempt combineSelectAndUse on each operand of a commutative operator N.
7735 SDValue combineSelectAndUseCommutative(SDNode *N, bool AllOnes,
7736 TargetLowering::DAGCombinerInfo &DCI) {
7737 SDValue N0 = N->getOperand(0);
7738 SDValue N1 = N->getOperand(1);
7739 if (N0.getNode()->hasOneUse()) {
7740 SDValue Result = combineSelectAndUse(N, N0, N1, DCI, AllOnes);
7741 if (Result.getNode())
7744 if (N1.getNode()->hasOneUse()) {
7745 SDValue Result = combineSelectAndUse(N, N1, N0, DCI, AllOnes);
7746 if (Result.getNode())
7752 // AddCombineToVPADDL- For pair-wise add on neon, use the vpaddl instruction
7753 // (only after legalization).
7754 static SDValue AddCombineToVPADDL(SDNode *N, SDValue N0, SDValue N1,
7755 TargetLowering::DAGCombinerInfo &DCI,
7756 const ARMSubtarget *Subtarget) {
7758 // Only perform optimization if after legalize, and if NEON is available. We
7759 // also expected both operands to be BUILD_VECTORs.
7760 if (DCI.isBeforeLegalize() || !Subtarget->hasNEON()
7761 || N0.getOpcode() != ISD::BUILD_VECTOR
7762 || N1.getOpcode() != ISD::BUILD_VECTOR)
7765 // Check output type since VPADDL operand elements can only be 8, 16, or 32.
7766 EVT VT = N->getValueType(0);
7767 if (!VT.isInteger() || VT.getVectorElementType() == MVT::i64)
7770 // Check that the vector operands are of the right form.
7771 // N0 and N1 are BUILD_VECTOR nodes with N number of EXTRACT_VECTOR
7772 // operands, where N is the size of the formed vector.
7773 // Each EXTRACT_VECTOR should have the same input vector and odd or even
7774 // index such that we have a pair wise add pattern.
7776 // Grab the vector that all EXTRACT_VECTOR nodes should be referencing.
7777 if (N0->getOperand(0)->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
7779 SDValue Vec = N0->getOperand(0)->getOperand(0);
7780 SDNode *V = Vec.getNode();
7781 unsigned nextIndex = 0;
7783 // For each operands to the ADD which are BUILD_VECTORs,
7784 // check to see if each of their operands are an EXTRACT_VECTOR with
7785 // the same vector and appropriate index.
7786 for (unsigned i = 0, e = N0->getNumOperands(); i != e; ++i) {
7787 if (N0->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT
7788 && N1->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
7790 SDValue ExtVec0 = N0->getOperand(i);
7791 SDValue ExtVec1 = N1->getOperand(i);
7793 // First operand is the vector, verify its the same.
7794 if (V != ExtVec0->getOperand(0).getNode() ||
7795 V != ExtVec1->getOperand(0).getNode())
7798 // Second is the constant, verify its correct.
7799 ConstantSDNode *C0 = dyn_cast<ConstantSDNode>(ExtVec0->getOperand(1));
7800 ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(ExtVec1->getOperand(1));
7802 // For the constant, we want to see all the even or all the odd.
7803 if (!C0 || !C1 || C0->getZExtValue() != nextIndex
7804 || C1->getZExtValue() != nextIndex+1)
7813 // Create VPADDL node.
7814 SelectionDAG &DAG = DCI.DAG;
7815 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7817 // Build operand list.
7818 SmallVector<SDValue, 8> Ops;
7819 Ops.push_back(DAG.getConstant(Intrinsic::arm_neon_vpaddls,
7820 TLI.getPointerTy()));
7822 // Input is the vector.
7825 // Get widened type and narrowed type.
7827 unsigned numElem = VT.getVectorNumElements();
7829 EVT inputLaneType = Vec.getValueType().getVectorElementType();
7830 switch (inputLaneType.getSimpleVT().SimpleTy) {
7831 case MVT::i8: widenType = MVT::getVectorVT(MVT::i16, numElem); break;
7832 case MVT::i16: widenType = MVT::getVectorVT(MVT::i32, numElem); break;
7833 case MVT::i32: widenType = MVT::getVectorVT(MVT::i64, numElem); break;
7835 llvm_unreachable("Invalid vector element type for padd optimization.");
7838 SDValue tmp = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N), widenType, Ops);
7839 unsigned ExtOp = VT.bitsGT(tmp.getValueType()) ? ISD::ANY_EXTEND : ISD::TRUNCATE;
7840 return DAG.getNode(ExtOp, SDLoc(N), VT, tmp);
7843 static SDValue findMUL_LOHI(SDValue V) {
7844 if (V->getOpcode() == ISD::UMUL_LOHI ||
7845 V->getOpcode() == ISD::SMUL_LOHI)
7850 static SDValue AddCombineTo64bitMLAL(SDNode *AddcNode,
7851 TargetLowering::DAGCombinerInfo &DCI,
7852 const ARMSubtarget *Subtarget) {
7854 if (Subtarget->isThumb1Only()) return SDValue();
7856 // Only perform the checks after legalize when the pattern is available.
7857 if (DCI.isBeforeLegalize()) return SDValue();
7859 // Look for multiply add opportunities.
7860 // The pattern is a ISD::UMUL_LOHI followed by two add nodes, where
7861 // each add nodes consumes a value from ISD::UMUL_LOHI and there is
7862 // a glue link from the first add to the second add.
7863 // If we find this pattern, we can replace the U/SMUL_LOHI, ADDC, and ADDE by
7864 // a S/UMLAL instruction.
7867 // \ / \ [no multiline comment]
7873 assert(AddcNode->getOpcode() == ISD::ADDC && "Expect an ADDC");
7874 SDValue AddcOp0 = AddcNode->getOperand(0);
7875 SDValue AddcOp1 = AddcNode->getOperand(1);
7877 // Check if the two operands are from the same mul_lohi node.
7878 if (AddcOp0.getNode() == AddcOp1.getNode())
7881 assert(AddcNode->getNumValues() == 2 &&
7882 AddcNode->getValueType(0) == MVT::i32 &&
7883 "Expect ADDC with two result values. First: i32");
7885 // Check that we have a glued ADDC node.
7886 if (AddcNode->getValueType(1) != MVT::Glue)
7889 // Check that the ADDC adds the low result of the S/UMUL_LOHI.
7890 if (AddcOp0->getOpcode() != ISD::UMUL_LOHI &&
7891 AddcOp0->getOpcode() != ISD::SMUL_LOHI &&
7892 AddcOp1->getOpcode() != ISD::UMUL_LOHI &&
7893 AddcOp1->getOpcode() != ISD::SMUL_LOHI)
7896 // Look for the glued ADDE.
7897 SDNode* AddeNode = AddcNode->getGluedUser();
7901 // Make sure it is really an ADDE.
7902 if (AddeNode->getOpcode() != ISD::ADDE)
7905 assert(AddeNode->getNumOperands() == 3 &&
7906 AddeNode->getOperand(2).getValueType() == MVT::Glue &&
7907 "ADDE node has the wrong inputs");
7909 // Check for the triangle shape.
7910 SDValue AddeOp0 = AddeNode->getOperand(0);
7911 SDValue AddeOp1 = AddeNode->getOperand(1);
7913 // Make sure that the ADDE operands are not coming from the same node.
7914 if (AddeOp0.getNode() == AddeOp1.getNode())
7917 // Find the MUL_LOHI node walking up ADDE's operands.
7918 bool IsLeftOperandMUL = false;
7919 SDValue MULOp = findMUL_LOHI(AddeOp0);
7920 if (MULOp == SDValue())
7921 MULOp = findMUL_LOHI(AddeOp1);
7923 IsLeftOperandMUL = true;
7924 if (MULOp == SDValue())
7927 // Figure out the right opcode.
7928 unsigned Opc = MULOp->getOpcode();
7929 unsigned FinalOpc = (Opc == ISD::SMUL_LOHI) ? ARMISD::SMLAL : ARMISD::UMLAL;
7931 // Figure out the high and low input values to the MLAL node.
7932 SDValue* HiAdd = nullptr;
7933 SDValue* LoMul = nullptr;
7934 SDValue* LowAdd = nullptr;
7936 // Ensure that ADDE is from high result of ISD::SMUL_LOHI.
7937 if ((AddeOp0 != MULOp.getValue(1)) && (AddeOp1 != MULOp.getValue(1)))
7940 if (IsLeftOperandMUL)
7946 // Ensure that LoMul and LowAdd are taken from correct ISD::SMUL_LOHI node
7947 // whose low result is fed to the ADDC we are checking.
7949 if (AddcOp0 == MULOp.getValue(0)) {
7953 if (AddcOp1 == MULOp.getValue(0)) {
7961 // Create the merged node.
7962 SelectionDAG &DAG = DCI.DAG;
7964 // Build operand list.
7965 SmallVector<SDValue, 8> Ops;
7966 Ops.push_back(LoMul->getOperand(0));
7967 Ops.push_back(LoMul->getOperand(1));
7968 Ops.push_back(*LowAdd);
7969 Ops.push_back(*HiAdd);
7971 SDValue MLALNode = DAG.getNode(FinalOpc, SDLoc(AddcNode),
7972 DAG.getVTList(MVT::i32, MVT::i32), Ops);
7974 // Replace the ADDs' nodes uses by the MLA node's values.
7975 SDValue HiMLALResult(MLALNode.getNode(), 1);
7976 DAG.ReplaceAllUsesOfValueWith(SDValue(AddeNode, 0), HiMLALResult);
7978 SDValue LoMLALResult(MLALNode.getNode(), 0);
7979 DAG.ReplaceAllUsesOfValueWith(SDValue(AddcNode, 0), LoMLALResult);
7981 // Return original node to notify the driver to stop replacing.
7982 SDValue resNode(AddcNode, 0);
7986 /// PerformADDCCombine - Target-specific dag combine transform from
7987 /// ISD::ADDC, ISD::ADDE, and ISD::MUL_LOHI to MLAL.
7988 static SDValue PerformADDCCombine(SDNode *N,
7989 TargetLowering::DAGCombinerInfo &DCI,
7990 const ARMSubtarget *Subtarget) {
7992 return AddCombineTo64bitMLAL(N, DCI, Subtarget);
7996 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
7997 /// operands N0 and N1. This is a helper for PerformADDCombine that is
7998 /// called with the default operands, and if that fails, with commuted
8000 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
8001 TargetLowering::DAGCombinerInfo &DCI,
8002 const ARMSubtarget *Subtarget){
8004 // Attempt to create vpaddl for this add.
8005 SDValue Result = AddCombineToVPADDL(N, N0, N1, DCI, Subtarget);
8006 if (Result.getNode())
8009 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
8010 if (N0.getNode()->hasOneUse()) {
8011 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
8012 if (Result.getNode()) return Result;
8017 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
8019 static SDValue PerformADDCombine(SDNode *N,
8020 TargetLowering::DAGCombinerInfo &DCI,
8021 const ARMSubtarget *Subtarget) {
8022 SDValue N0 = N->getOperand(0);
8023 SDValue N1 = N->getOperand(1);
8025 // First try with the default operand order.
8026 SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI, Subtarget);
8027 if (Result.getNode())
8030 // If that didn't work, try again with the operands commuted.
8031 return PerformADDCombineWithOperands(N, N1, N0, DCI, Subtarget);
8034 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
8036 static SDValue PerformSUBCombine(SDNode *N,
8037 TargetLowering::DAGCombinerInfo &DCI) {
8038 SDValue N0 = N->getOperand(0);
8039 SDValue N1 = N->getOperand(1);
8041 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
8042 if (N1.getNode()->hasOneUse()) {
8043 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
8044 if (Result.getNode()) return Result;
8050 /// PerformVMULCombine
8051 /// Distribute (A + B) * C to (A * C) + (B * C) to take advantage of the
8052 /// special multiplier accumulator forwarding.
8058 // However, for (A + B) * (A + B),
8065 static SDValue PerformVMULCombine(SDNode *N,
8066 TargetLowering::DAGCombinerInfo &DCI,
8067 const ARMSubtarget *Subtarget) {
8068 if (!Subtarget->hasVMLxForwarding())
8071 SelectionDAG &DAG = DCI.DAG;
8072 SDValue N0 = N->getOperand(0);
8073 SDValue N1 = N->getOperand(1);
8074 unsigned Opcode = N0.getOpcode();
8075 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
8076 Opcode != ISD::FADD && Opcode != ISD::FSUB) {
8077 Opcode = N1.getOpcode();
8078 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
8079 Opcode != ISD::FADD && Opcode != ISD::FSUB)
8087 EVT VT = N->getValueType(0);
8089 SDValue N00 = N0->getOperand(0);
8090 SDValue N01 = N0->getOperand(1);
8091 return DAG.getNode(Opcode, DL, VT,
8092 DAG.getNode(ISD::MUL, DL, VT, N00, N1),
8093 DAG.getNode(ISD::MUL, DL, VT, N01, N1));
8096 static SDValue PerformMULCombine(SDNode *N,
8097 TargetLowering::DAGCombinerInfo &DCI,
8098 const ARMSubtarget *Subtarget) {
8099 SelectionDAG &DAG = DCI.DAG;
8101 if (Subtarget->isThumb1Only())
8104 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
8107 EVT VT = N->getValueType(0);
8108 if (VT.is64BitVector() || VT.is128BitVector())
8109 return PerformVMULCombine(N, DCI, Subtarget);
8113 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
8117 int64_t MulAmt = C->getSExtValue();
8118 unsigned ShiftAmt = countTrailingZeros<uint64_t>(MulAmt);
8120 ShiftAmt = ShiftAmt & (32 - 1);
8121 SDValue V = N->getOperand(0);
8125 MulAmt >>= ShiftAmt;
8128 if (isPowerOf2_32(MulAmt - 1)) {
8129 // (mul x, 2^N + 1) => (add (shl x, N), x)
8130 Res = DAG.getNode(ISD::ADD, DL, VT,
8132 DAG.getNode(ISD::SHL, DL, VT,
8134 DAG.getConstant(Log2_32(MulAmt - 1),
8136 } else if (isPowerOf2_32(MulAmt + 1)) {
8137 // (mul x, 2^N - 1) => (sub (shl x, N), x)
8138 Res = DAG.getNode(ISD::SUB, DL, VT,
8139 DAG.getNode(ISD::SHL, DL, VT,
8141 DAG.getConstant(Log2_32(MulAmt + 1),
8147 uint64_t MulAmtAbs = -MulAmt;
8148 if (isPowerOf2_32(MulAmtAbs + 1)) {
8149 // (mul x, -(2^N - 1)) => (sub x, (shl x, N))
8150 Res = DAG.getNode(ISD::SUB, DL, VT,
8152 DAG.getNode(ISD::SHL, DL, VT,
8154 DAG.getConstant(Log2_32(MulAmtAbs + 1),
8156 } else if (isPowerOf2_32(MulAmtAbs - 1)) {
8157 // (mul x, -(2^N + 1)) => - (add (shl x, N), x)
8158 Res = DAG.getNode(ISD::ADD, DL, VT,
8160 DAG.getNode(ISD::SHL, DL, VT,
8162 DAG.getConstant(Log2_32(MulAmtAbs-1),
8164 Res = DAG.getNode(ISD::SUB, DL, VT,
8165 DAG.getConstant(0, MVT::i32),Res);
8172 Res = DAG.getNode(ISD::SHL, DL, VT,
8173 Res, DAG.getConstant(ShiftAmt, MVT::i32));
8175 // Do not add new nodes to DAG combiner worklist.
8176 DCI.CombineTo(N, Res, false);
8180 static SDValue PerformANDCombine(SDNode *N,
8181 TargetLowering::DAGCombinerInfo &DCI,
8182 const ARMSubtarget *Subtarget) {
8184 // Attempt to use immediate-form VBIC
8185 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
8187 EVT VT = N->getValueType(0);
8188 SelectionDAG &DAG = DCI.DAG;
8190 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8193 APInt SplatBits, SplatUndef;
8194 unsigned SplatBitSize;
8197 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
8198 if (SplatBitSize <= 64) {
8200 SDValue Val = isNEONModifiedImm((~SplatBits).getZExtValue(),
8201 SplatUndef.getZExtValue(), SplatBitSize,
8202 DAG, VbicVT, VT.is128BitVector(),
8204 if (Val.getNode()) {
8206 DAG.getNode(ISD::BITCAST, dl, VbicVT, N->getOperand(0));
8207 SDValue Vbic = DAG.getNode(ARMISD::VBICIMM, dl, VbicVT, Input, Val);
8208 return DAG.getNode(ISD::BITCAST, dl, VT, Vbic);
8213 if (!Subtarget->isThumb1Only()) {
8214 // fold (and (select cc, -1, c), x) -> (select cc, x, (and, x, c))
8215 SDValue Result = combineSelectAndUseCommutative(N, true, DCI);
8216 if (Result.getNode())
8223 /// PerformORCombine - Target-specific dag combine xforms for ISD::OR
8224 static SDValue PerformORCombine(SDNode *N,
8225 TargetLowering::DAGCombinerInfo &DCI,
8226 const ARMSubtarget *Subtarget) {
8227 // Attempt to use immediate-form VORR
8228 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
8230 EVT VT = N->getValueType(0);
8231 SelectionDAG &DAG = DCI.DAG;
8233 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8236 APInt SplatBits, SplatUndef;
8237 unsigned SplatBitSize;
8239 if (BVN && Subtarget->hasNEON() &&
8240 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
8241 if (SplatBitSize <= 64) {
8243 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
8244 SplatUndef.getZExtValue(), SplatBitSize,
8245 DAG, VorrVT, VT.is128BitVector(),
8247 if (Val.getNode()) {
8249 DAG.getNode(ISD::BITCAST, dl, VorrVT, N->getOperand(0));
8250 SDValue Vorr = DAG.getNode(ARMISD::VORRIMM, dl, VorrVT, Input, Val);
8251 return DAG.getNode(ISD::BITCAST, dl, VT, Vorr);
8256 if (!Subtarget->isThumb1Only()) {
8257 // fold (or (select cc, 0, c), x) -> (select cc, x, (or, x, c))
8258 SDValue Result = combineSelectAndUseCommutative(N, false, DCI);
8259 if (Result.getNode())
8263 // The code below optimizes (or (and X, Y), Z).
8264 // The AND operand needs to have a single user to make these optimizations
8266 SDValue N0 = N->getOperand(0);
8267 if (N0.getOpcode() != ISD::AND || !N0.hasOneUse())
8269 SDValue N1 = N->getOperand(1);
8271 // (or (and B, A), (and C, ~A)) => (VBSL A, B, C) when A is a constant.
8272 if (Subtarget->hasNEON() && N1.getOpcode() == ISD::AND && VT.isVector() &&
8273 DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
8275 unsigned SplatBitSize;
8278 APInt SplatBits0, SplatBits1;
8279 BuildVectorSDNode *BVN0 = dyn_cast<BuildVectorSDNode>(N0->getOperand(1));
8280 BuildVectorSDNode *BVN1 = dyn_cast<BuildVectorSDNode>(N1->getOperand(1));
8281 // Ensure that the second operand of both ands are constants
8282 if (BVN0 && BVN0->isConstantSplat(SplatBits0, SplatUndef, SplatBitSize,
8283 HasAnyUndefs) && !HasAnyUndefs) {
8284 if (BVN1 && BVN1->isConstantSplat(SplatBits1, SplatUndef, SplatBitSize,
8285 HasAnyUndefs) && !HasAnyUndefs) {
8286 // Ensure that the bit width of the constants are the same and that
8287 // the splat arguments are logical inverses as per the pattern we
8288 // are trying to simplify.
8289 if (SplatBits0.getBitWidth() == SplatBits1.getBitWidth() &&
8290 SplatBits0 == ~SplatBits1) {
8291 // Canonicalize the vector type to make instruction selection
8293 EVT CanonicalVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
8294 SDValue Result = DAG.getNode(ARMISD::VBSL, dl, CanonicalVT,
8298 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
8304 // Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when
8307 // BFI is only available on V6T2+
8308 if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops())
8312 // 1) or (and A, mask), val => ARMbfi A, val, mask
8313 // iff (val & mask) == val
8315 // 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
8316 // 2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2)
8317 // && mask == ~mask2
8318 // 2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2)
8319 // && ~mask == mask2
8320 // (i.e., copy a bitfield value into another bitfield of the same width)
8325 SDValue N00 = N0.getOperand(0);
8327 // The value and the mask need to be constants so we can verify this is
8328 // actually a bitfield set. If the mask is 0xffff, we can do better
8329 // via a movt instruction, so don't use BFI in that case.
8330 SDValue MaskOp = N0.getOperand(1);
8331 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(MaskOp);
8334 unsigned Mask = MaskC->getZExtValue();
8338 // Case (1): or (and A, mask), val => ARMbfi A, val, mask
8339 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
8341 unsigned Val = N1C->getZExtValue();
8342 if ((Val & ~Mask) != Val)
8345 if (ARM::isBitFieldInvertedMask(Mask)) {
8346 Val >>= countTrailingZeros(~Mask);
8348 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00,
8349 DAG.getConstant(Val, MVT::i32),
8350 DAG.getConstant(Mask, MVT::i32));
8352 // Do not add new nodes to DAG combiner worklist.
8353 DCI.CombineTo(N, Res, false);
8356 } else if (N1.getOpcode() == ISD::AND) {
8357 // case (2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
8358 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
8361 unsigned Mask2 = N11C->getZExtValue();
8363 // Mask and ~Mask2 (or reverse) must be equivalent for the BFI pattern
8365 if (ARM::isBitFieldInvertedMask(Mask) &&
8367 // The pack halfword instruction works better for masks that fit it,
8368 // so use that when it's available.
8369 if (Subtarget->hasT2ExtractPack() &&
8370 (Mask == 0xffff || Mask == 0xffff0000))
8373 unsigned amt = countTrailingZeros(Mask2);
8374 Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0),
8375 DAG.getConstant(amt, MVT::i32));
8376 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00, Res,
8377 DAG.getConstant(Mask, MVT::i32));
8378 // Do not add new nodes to DAG combiner worklist.
8379 DCI.CombineTo(N, Res, false);
8381 } else if (ARM::isBitFieldInvertedMask(~Mask) &&
8383 // The pack halfword instruction works better for masks that fit it,
8384 // so use that when it's available.
8385 if (Subtarget->hasT2ExtractPack() &&
8386 (Mask2 == 0xffff || Mask2 == 0xffff0000))
8389 unsigned lsb = countTrailingZeros(Mask);
8390 Res = DAG.getNode(ISD::SRL, DL, VT, N00,
8391 DAG.getConstant(lsb, MVT::i32));
8392 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res,
8393 DAG.getConstant(Mask2, MVT::i32));
8394 // Do not add new nodes to DAG combiner worklist.
8395 DCI.CombineTo(N, Res, false);
8400 if (DAG.MaskedValueIsZero(N1, MaskC->getAPIntValue()) &&
8401 N00.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N00.getOperand(1)) &&
8402 ARM::isBitFieldInvertedMask(~Mask)) {
8403 // Case (3): or (and (shl A, #shamt), mask), B => ARMbfi B, A, ~mask
8404 // where lsb(mask) == #shamt and masked bits of B are known zero.
8405 SDValue ShAmt = N00.getOperand(1);
8406 unsigned ShAmtC = cast<ConstantSDNode>(ShAmt)->getZExtValue();
8407 unsigned LSB = countTrailingZeros(Mask);
8411 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1, N00.getOperand(0),
8412 DAG.getConstant(~Mask, MVT::i32));
8414 // Do not add new nodes to DAG combiner worklist.
8415 DCI.CombineTo(N, Res, false);
8421 static SDValue PerformXORCombine(SDNode *N,
8422 TargetLowering::DAGCombinerInfo &DCI,
8423 const ARMSubtarget *Subtarget) {
8424 EVT VT = N->getValueType(0);
8425 SelectionDAG &DAG = DCI.DAG;
8427 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8430 if (!Subtarget->isThumb1Only()) {
8431 // fold (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c))
8432 SDValue Result = combineSelectAndUseCommutative(N, false, DCI);
8433 if (Result.getNode())
8440 /// PerformBFICombine - (bfi A, (and B, Mask1), Mask2) -> (bfi A, B, Mask2) iff
8441 /// the bits being cleared by the AND are not demanded by the BFI.
8442 static SDValue PerformBFICombine(SDNode *N,
8443 TargetLowering::DAGCombinerInfo &DCI) {
8444 SDValue N1 = N->getOperand(1);
8445 if (N1.getOpcode() == ISD::AND) {
8446 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
8449 unsigned InvMask = cast<ConstantSDNode>(N->getOperand(2))->getZExtValue();
8450 unsigned LSB = countTrailingZeros(~InvMask);
8451 unsigned Width = (32 - countLeadingZeros(~InvMask)) - LSB;
8453 static_cast<unsigned>(std::numeric_limits<unsigned>::digits) &&
8454 "undefined behavior");
8455 unsigned Mask = (1u << Width) - 1;
8456 unsigned Mask2 = N11C->getZExtValue();
8457 if ((Mask & (~Mask2)) == 0)
8458 return DCI.DAG.getNode(ARMISD::BFI, SDLoc(N), N->getValueType(0),
8459 N->getOperand(0), N1.getOperand(0),
8465 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
8466 /// ARMISD::VMOVRRD.
8467 static SDValue PerformVMOVRRDCombine(SDNode *N,
8468 TargetLowering::DAGCombinerInfo &DCI,
8469 const ARMSubtarget *Subtarget) {
8470 // vmovrrd(vmovdrr x, y) -> x,y
8471 SDValue InDouble = N->getOperand(0);
8472 if (InDouble.getOpcode() == ARMISD::VMOVDRR && !Subtarget->isFPOnlySP())
8473 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
8475 // vmovrrd(load f64) -> (load i32), (load i32)
8476 SDNode *InNode = InDouble.getNode();
8477 if (ISD::isNormalLoad(InNode) && InNode->hasOneUse() &&
8478 InNode->getValueType(0) == MVT::f64 &&
8479 InNode->getOperand(1).getOpcode() == ISD::FrameIndex &&
8480 !cast<LoadSDNode>(InNode)->isVolatile()) {
8481 // TODO: Should this be done for non-FrameIndex operands?
8482 LoadSDNode *LD = cast<LoadSDNode>(InNode);
8484 SelectionDAG &DAG = DCI.DAG;
8486 SDValue BasePtr = LD->getBasePtr();
8487 SDValue NewLD1 = DAG.getLoad(MVT::i32, DL, LD->getChain(), BasePtr,
8488 LD->getPointerInfo(), LD->isVolatile(),
8489 LD->isNonTemporal(), LD->isInvariant(),
8490 LD->getAlignment());
8492 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
8493 DAG.getConstant(4, MVT::i32));
8494 SDValue NewLD2 = DAG.getLoad(MVT::i32, DL, NewLD1.getValue(1), OffsetPtr,
8495 LD->getPointerInfo(), LD->isVolatile(),
8496 LD->isNonTemporal(), LD->isInvariant(),
8497 std::min(4U, LD->getAlignment() / 2));
8499 DAG.ReplaceAllUsesOfValueWith(SDValue(LD, 1), NewLD2.getValue(1));
8500 if (DCI.DAG.getTargetLoweringInfo().isBigEndian())
8501 std::swap (NewLD1, NewLD2);
8502 SDValue Result = DCI.CombineTo(N, NewLD1, NewLD2);
8509 /// PerformVMOVDRRCombine - Target-specific dag combine xforms for
8510 /// ARMISD::VMOVDRR. This is also used for BUILD_VECTORs with 2 operands.
8511 static SDValue PerformVMOVDRRCombine(SDNode *N, SelectionDAG &DAG) {
8512 // N=vmovrrd(X); vmovdrr(N:0, N:1) -> bit_convert(X)
8513 SDValue Op0 = N->getOperand(0);
8514 SDValue Op1 = N->getOperand(1);
8515 if (Op0.getOpcode() == ISD::BITCAST)
8516 Op0 = Op0.getOperand(0);
8517 if (Op1.getOpcode() == ISD::BITCAST)
8518 Op1 = Op1.getOperand(0);
8519 if (Op0.getOpcode() == ARMISD::VMOVRRD &&
8520 Op0.getNode() == Op1.getNode() &&
8521 Op0.getResNo() == 0 && Op1.getResNo() == 1)
8522 return DAG.getNode(ISD::BITCAST, SDLoc(N),
8523 N->getValueType(0), Op0.getOperand(0));
8527 /// hasNormalLoadOperand - Check if any of the operands of a BUILD_VECTOR node
8528 /// are normal, non-volatile loads. If so, it is profitable to bitcast an
8529 /// i64 vector to have f64 elements, since the value can then be loaded
8530 /// directly into a VFP register.
8531 static bool hasNormalLoadOperand(SDNode *N) {
8532 unsigned NumElts = N->getValueType(0).getVectorNumElements();
8533 for (unsigned i = 0; i < NumElts; ++i) {
8534 SDNode *Elt = N->getOperand(i).getNode();
8535 if (ISD::isNormalLoad(Elt) && !cast<LoadSDNode>(Elt)->isVolatile())
8541 /// PerformBUILD_VECTORCombine - Target-specific dag combine xforms for
8542 /// ISD::BUILD_VECTOR.
8543 static SDValue PerformBUILD_VECTORCombine(SDNode *N,
8544 TargetLowering::DAGCombinerInfo &DCI,
8545 const ARMSubtarget *Subtarget) {
8546 // build_vector(N=ARMISD::VMOVRRD(X), N:1) -> bit_convert(X):
8547 // VMOVRRD is introduced when legalizing i64 types. It forces the i64 value
8548 // into a pair of GPRs, which is fine when the value is used as a scalar,
8549 // but if the i64 value is converted to a vector, we need to undo the VMOVRRD.
8550 SelectionDAG &DAG = DCI.DAG;
8551 if (N->getNumOperands() == 2) {
8552 SDValue RV = PerformVMOVDRRCombine(N, DAG);
8557 // Load i64 elements as f64 values so that type legalization does not split
8558 // them up into i32 values.
8559 EVT VT = N->getValueType(0);
8560 if (VT.getVectorElementType() != MVT::i64 || !hasNormalLoadOperand(N))
8563 SmallVector<SDValue, 8> Ops;
8564 unsigned NumElts = VT.getVectorNumElements();
8565 for (unsigned i = 0; i < NumElts; ++i) {
8566 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(i));
8568 // Make the DAGCombiner fold the bitcast.
8569 DCI.AddToWorklist(V.getNode());
8571 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, NumElts);
8572 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, FloatVT, Ops);
8573 return DAG.getNode(ISD::BITCAST, dl, VT, BV);
8576 /// \brief Target-specific dag combine xforms for ARMISD::BUILD_VECTOR.
8578 PerformARMBUILD_VECTORCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
8579 // ARMISD::BUILD_VECTOR is introduced when legalizing ISD::BUILD_VECTOR.
8580 // At that time, we may have inserted bitcasts from integer to float.
8581 // If these bitcasts have survived DAGCombine, change the lowering of this
8582 // BUILD_VECTOR in something more vector friendly, i.e., that does not
8583 // force to use floating point types.
8585 // Make sure we can change the type of the vector.
8586 // This is possible iff:
8587 // 1. The vector is only used in a bitcast to a integer type. I.e.,
8588 // 1.1. Vector is used only once.
8589 // 1.2. Use is a bit convert to an integer type.
8590 // 2. The size of its operands are 32-bits (64-bits are not legal).
8591 EVT VT = N->getValueType(0);
8592 EVT EltVT = VT.getVectorElementType();
8594 // Check 1.1. and 2.
8595 if (EltVT.getSizeInBits() != 32 || !N->hasOneUse())
8598 // By construction, the input type must be float.
8599 assert(EltVT == MVT::f32 && "Unexpected type!");
8602 SDNode *Use = *N->use_begin();
8603 if (Use->getOpcode() != ISD::BITCAST ||
8604 Use->getValueType(0).isFloatingPoint())
8607 // Check profitability.
8608 // Model is, if more than half of the relevant operands are bitcast from
8609 // i32, turn the build_vector into a sequence of insert_vector_elt.
8610 // Relevant operands are everything that is not statically
8611 // (i.e., at compile time) bitcasted.
8612 unsigned NumOfBitCastedElts = 0;
8613 unsigned NumElts = VT.getVectorNumElements();
8614 unsigned NumOfRelevantElts = NumElts;
8615 for (unsigned Idx = 0; Idx < NumElts; ++Idx) {
8616 SDValue Elt = N->getOperand(Idx);
8617 if (Elt->getOpcode() == ISD::BITCAST) {
8618 // Assume only bit cast to i32 will go away.
8619 if (Elt->getOperand(0).getValueType() == MVT::i32)
8620 ++NumOfBitCastedElts;
8621 } else if (Elt.getOpcode() == ISD::UNDEF || isa<ConstantSDNode>(Elt))
8622 // Constants are statically casted, thus do not count them as
8623 // relevant operands.
8624 --NumOfRelevantElts;
8627 // Check if more than half of the elements require a non-free bitcast.
8628 if (NumOfBitCastedElts <= NumOfRelevantElts / 2)
8631 SelectionDAG &DAG = DCI.DAG;
8632 // Create the new vector type.
8633 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
8634 // Check if the type is legal.
8635 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8636 if (!TLI.isTypeLegal(VecVT))
8640 // ARMISD::BUILD_VECTOR E1, E2, ..., EN.
8641 // => BITCAST INSERT_VECTOR_ELT
8642 // (INSERT_VECTOR_ELT (...), (BITCAST EN-1), N-1),
8644 SDValue Vec = DAG.getUNDEF(VecVT);
8646 for (unsigned Idx = 0 ; Idx < NumElts; ++Idx) {
8647 SDValue V = N->getOperand(Idx);
8648 if (V.getOpcode() == ISD::UNDEF)
8650 if (V.getOpcode() == ISD::BITCAST &&
8651 V->getOperand(0).getValueType() == MVT::i32)
8652 // Fold obvious case.
8653 V = V.getOperand(0);
8655 V = DAG.getNode(ISD::BITCAST, SDLoc(V), MVT::i32, V);
8656 // Make the DAGCombiner fold the bitcasts.
8657 DCI.AddToWorklist(V.getNode());
8659 SDValue LaneIdx = DAG.getConstant(Idx, MVT::i32);
8660 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VecVT, Vec, V, LaneIdx);
8662 Vec = DAG.getNode(ISD::BITCAST, dl, VT, Vec);
8663 // Make the DAGCombiner fold the bitcasts.
8664 DCI.AddToWorklist(Vec.getNode());
8668 /// PerformInsertEltCombine - Target-specific dag combine xforms for
8669 /// ISD::INSERT_VECTOR_ELT.
8670 static SDValue PerformInsertEltCombine(SDNode *N,
8671 TargetLowering::DAGCombinerInfo &DCI) {
8672 // Bitcast an i64 load inserted into a vector to f64.
8673 // Otherwise, the i64 value will be legalized to a pair of i32 values.
8674 EVT VT = N->getValueType(0);
8675 SDNode *Elt = N->getOperand(1).getNode();
8676 if (VT.getVectorElementType() != MVT::i64 ||
8677 !ISD::isNormalLoad(Elt) || cast<LoadSDNode>(Elt)->isVolatile())
8680 SelectionDAG &DAG = DCI.DAG;
8682 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
8683 VT.getVectorNumElements());
8684 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, N->getOperand(0));
8685 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(1));
8686 // Make the DAGCombiner fold the bitcasts.
8687 DCI.AddToWorklist(Vec.getNode());
8688 DCI.AddToWorklist(V.getNode());
8689 SDValue InsElt = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, FloatVT,
8690 Vec, V, N->getOperand(2));
8691 return DAG.getNode(ISD::BITCAST, dl, VT, InsElt);
8694 /// PerformVECTOR_SHUFFLECombine - Target-specific dag combine xforms for
8695 /// ISD::VECTOR_SHUFFLE.
8696 static SDValue PerformVECTOR_SHUFFLECombine(SDNode *N, SelectionDAG &DAG) {
8697 // The LLVM shufflevector instruction does not require the shuffle mask
8698 // length to match the operand vector length, but ISD::VECTOR_SHUFFLE does
8699 // have that requirement. When translating to ISD::VECTOR_SHUFFLE, if the
8700 // operands do not match the mask length, they are extended by concatenating
8701 // them with undef vectors. That is probably the right thing for other
8702 // targets, but for NEON it is better to concatenate two double-register
8703 // size vector operands into a single quad-register size vector. Do that
8704 // transformation here:
8705 // shuffle(concat(v1, undef), concat(v2, undef)) ->
8706 // shuffle(concat(v1, v2), undef)
8707 SDValue Op0 = N->getOperand(0);
8708 SDValue Op1 = N->getOperand(1);
8709 if (Op0.getOpcode() != ISD::CONCAT_VECTORS ||
8710 Op1.getOpcode() != ISD::CONCAT_VECTORS ||
8711 Op0.getNumOperands() != 2 ||
8712 Op1.getNumOperands() != 2)
8714 SDValue Concat0Op1 = Op0.getOperand(1);
8715 SDValue Concat1Op1 = Op1.getOperand(1);
8716 if (Concat0Op1.getOpcode() != ISD::UNDEF ||
8717 Concat1Op1.getOpcode() != ISD::UNDEF)
8719 // Skip the transformation if any of the types are illegal.
8720 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8721 EVT VT = N->getValueType(0);
8722 if (!TLI.isTypeLegal(VT) ||
8723 !TLI.isTypeLegal(Concat0Op1.getValueType()) ||
8724 !TLI.isTypeLegal(Concat1Op1.getValueType()))
8727 SDValue NewConcat = DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT,
8728 Op0.getOperand(0), Op1.getOperand(0));
8729 // Translate the shuffle mask.
8730 SmallVector<int, 16> NewMask;
8731 unsigned NumElts = VT.getVectorNumElements();
8732 unsigned HalfElts = NumElts/2;
8733 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
8734 for (unsigned n = 0; n < NumElts; ++n) {
8735 int MaskElt = SVN->getMaskElt(n);
8737 if (MaskElt < (int)HalfElts)
8739 else if (MaskElt >= (int)NumElts && MaskElt < (int)(NumElts + HalfElts))
8740 NewElt = HalfElts + MaskElt - NumElts;
8741 NewMask.push_back(NewElt);
8743 return DAG.getVectorShuffle(VT, SDLoc(N), NewConcat,
8744 DAG.getUNDEF(VT), NewMask.data());
8747 /// CombineBaseUpdate - Target-specific DAG combine function for VLDDUP,
8748 /// NEON load/store intrinsics, and generic vector load/stores, to merge
8749 /// base address updates.
8750 /// For generic load/stores, the memory type is assumed to be a vector.
8751 /// The caller is assumed to have checked legality.
8752 static SDValue CombineBaseUpdate(SDNode *N,
8753 TargetLowering::DAGCombinerInfo &DCI) {
8754 SelectionDAG &DAG = DCI.DAG;
8755 const bool isIntrinsic = (N->getOpcode() == ISD::INTRINSIC_VOID ||
8756 N->getOpcode() == ISD::INTRINSIC_W_CHAIN);
8757 const bool isStore = N->getOpcode() == ISD::STORE;
8758 const unsigned AddrOpIdx = ((isIntrinsic || isStore) ? 2 : 1);
8759 SDValue Addr = N->getOperand(AddrOpIdx);
8760 MemSDNode *MemN = cast<MemSDNode>(N);
8762 // Search for a use of the address operand that is an increment.
8763 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(),
8764 UE = Addr.getNode()->use_end(); UI != UE; ++UI) {
8766 if (User->getOpcode() != ISD::ADD ||
8767 UI.getUse().getResNo() != Addr.getResNo())
8770 // Check that the add is independent of the load/store. Otherwise, folding
8771 // it would create a cycle.
8772 if (User->isPredecessorOf(N) || N->isPredecessorOf(User))
8775 // Find the new opcode for the updating load/store.
8776 bool isLoadOp = true;
8777 bool isLaneOp = false;
8778 unsigned NewOpc = 0;
8779 unsigned NumVecs = 0;
8781 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
8783 default: llvm_unreachable("unexpected intrinsic for Neon base update");
8784 case Intrinsic::arm_neon_vld1: NewOpc = ARMISD::VLD1_UPD;
8786 case Intrinsic::arm_neon_vld2: NewOpc = ARMISD::VLD2_UPD;
8788 case Intrinsic::arm_neon_vld3: NewOpc = ARMISD::VLD3_UPD;
8790 case Intrinsic::arm_neon_vld4: NewOpc = ARMISD::VLD4_UPD;
8792 case Intrinsic::arm_neon_vld2lane: NewOpc = ARMISD::VLD2LN_UPD;
8793 NumVecs = 2; isLaneOp = true; break;
8794 case Intrinsic::arm_neon_vld3lane: NewOpc = ARMISD::VLD3LN_UPD;
8795 NumVecs = 3; isLaneOp = true; break;
8796 case Intrinsic::arm_neon_vld4lane: NewOpc = ARMISD::VLD4LN_UPD;
8797 NumVecs = 4; isLaneOp = true; break;
8798 case Intrinsic::arm_neon_vst1: NewOpc = ARMISD::VST1_UPD;
8799 NumVecs = 1; isLoadOp = false; break;
8800 case Intrinsic::arm_neon_vst2: NewOpc = ARMISD::VST2_UPD;
8801 NumVecs = 2; isLoadOp = false; break;
8802 case Intrinsic::arm_neon_vst3: NewOpc = ARMISD::VST3_UPD;
8803 NumVecs = 3; isLoadOp = false; break;
8804 case Intrinsic::arm_neon_vst4: NewOpc = ARMISD::VST4_UPD;
8805 NumVecs = 4; isLoadOp = false; break;
8806 case Intrinsic::arm_neon_vst2lane: NewOpc = ARMISD::VST2LN_UPD;
8807 NumVecs = 2; isLoadOp = false; isLaneOp = true; break;
8808 case Intrinsic::arm_neon_vst3lane: NewOpc = ARMISD::VST3LN_UPD;
8809 NumVecs = 3; isLoadOp = false; isLaneOp = true; break;
8810 case Intrinsic::arm_neon_vst4lane: NewOpc = ARMISD::VST4LN_UPD;
8811 NumVecs = 4; isLoadOp = false; isLaneOp = true; break;
8815 switch (N->getOpcode()) {
8816 default: llvm_unreachable("unexpected opcode for Neon base update");
8817 case ARMISD::VLD2DUP: NewOpc = ARMISD::VLD2DUP_UPD; NumVecs = 2; break;
8818 case ARMISD::VLD3DUP: NewOpc = ARMISD::VLD3DUP_UPD; NumVecs = 3; break;
8819 case ARMISD::VLD4DUP: NewOpc = ARMISD::VLD4DUP_UPD; NumVecs = 4; break;
8820 case ISD::LOAD: NewOpc = ARMISD::VLD1_UPD;
8821 NumVecs = 1; isLaneOp = false; break;
8822 case ISD::STORE: NewOpc = ARMISD::VST1_UPD;
8823 NumVecs = 1; isLaneOp = false; isLoadOp = false; break;
8827 // Find the size of memory referenced by the load/store.
8830 VecTy = N->getValueType(0);
8831 } else if (isIntrinsic) {
8832 VecTy = N->getOperand(AddrOpIdx+1).getValueType();
8834 assert(isStore && "Node has to be a load, a store, or an intrinsic!");
8835 VecTy = N->getOperand(1).getValueType();
8838 unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8;
8840 NumBytes /= VecTy.getVectorNumElements();
8842 // If the increment is a constant, it must match the memory ref size.
8843 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
8844 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
8845 uint64_t IncVal = CInc->getZExtValue();
8846 if (IncVal != NumBytes)
8848 } else if (NumBytes >= 3 * 16) {
8849 // VLD3/4 and VST3/4 for 128-bit vectors are implemented with two
8850 // separate instructions that make it harder to use a non-constant update.
8854 // OK, we found an ADD we can fold into the base update.
8855 // Now, create a _UPD node, taking care of not breaking alignment.
8857 EVT AlignedVecTy = VecTy;
8858 unsigned Alignment = MemN->getAlignment();
8860 // If this is a less-than-standard-aligned load/store, change the type to
8861 // match the standard alignment.
8862 // The alignment is overlooked when selecting _UPD variants; and it's
8863 // easier to introduce bitcasts here than fix that.
8864 // There are 3 ways to get to this base-update combine:
8865 // - intrinsics: they are assumed to be properly aligned (to the standard
8866 // alignment of the memory type), so we don't need to do anything.
8867 // - ARMISD::VLDx nodes: they are only generated from the aforementioned
8868 // intrinsics, so, likewise, there's nothing to do.
8869 // - generic load/store instructions: the alignment is specified as an
8870 // explicit operand, rather than implicitly as the standard alignment
8871 // of the memory type (like the intrisics). We need to change the
8872 // memory type to match the explicit alignment. That way, we don't
8873 // generate non-standard-aligned ARMISD::VLDx nodes.
8874 if (isa<LSBaseSDNode>(N)) {
8877 if (Alignment < VecTy.getScalarSizeInBits() / 8) {
8878 MVT EltTy = MVT::getIntegerVT(Alignment * 8);
8879 assert(NumVecs == 1 && "Unexpected multi-element generic load/store.");
8880 assert(!isLaneOp && "Unexpected generic load/store lane.");
8881 unsigned NumElts = NumBytes / (EltTy.getSizeInBits() / 8);
8882 AlignedVecTy = MVT::getVectorVT(EltTy, NumElts);
8884 // Don't set an explicit alignment on regular load/stores that we want
8885 // to transform to VLD/VST 1_UPD nodes.
8886 // This matches the behavior of regular load/stores, which only get an
8887 // explicit alignment if the MMO alignment is larger than the standard
8888 // alignment of the memory type.
8889 // Intrinsics, however, always get an explicit alignment, set to the
8890 // alignment of the MMO.
8894 // Create the new updating load/store node.
8895 // First, create an SDVTList for the new updating node's results.
8897 unsigned NumResultVecs = (isLoadOp ? NumVecs : 0);
8899 for (n = 0; n < NumResultVecs; ++n)
8900 Tys[n] = AlignedVecTy;
8901 Tys[n++] = MVT::i32;
8902 Tys[n] = MVT::Other;
8903 SDVTList SDTys = DAG.getVTList(makeArrayRef(Tys, NumResultVecs+2));
8905 // Then, gather the new node's operands.
8906 SmallVector<SDValue, 8> Ops;
8907 Ops.push_back(N->getOperand(0)); // incoming chain
8908 Ops.push_back(N->getOperand(AddrOpIdx));
8911 if (StoreSDNode *StN = dyn_cast<StoreSDNode>(N)) {
8912 // Try to match the intrinsic's signature
8913 Ops.push_back(StN->getValue());
8915 // Loads (and of course intrinsics) match the intrinsics' signature,
8916 // so just add all but the alignment operand.
8917 for (unsigned i = AddrOpIdx + 1; i < N->getNumOperands() - 1; ++i)
8918 Ops.push_back(N->getOperand(i));
8921 // For all node types, the alignment operand is always the last one.
8922 Ops.push_back(DAG.getConstant(Alignment, MVT::i32));
8924 // If this is a non-standard-aligned STORE, the penultimate operand is the
8925 // stored value. Bitcast it to the aligned type.
8926 if (AlignedVecTy != VecTy && N->getOpcode() == ISD::STORE) {
8927 SDValue &StVal = Ops[Ops.size()-2];
8928 StVal = DAG.getNode(ISD::BITCAST, SDLoc(N), AlignedVecTy, StVal);
8931 SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, SDLoc(N), SDTys,
8933 MemN->getMemOperand());
8936 SmallVector<SDValue, 5> NewResults;
8937 for (unsigned i = 0; i < NumResultVecs; ++i)
8938 NewResults.push_back(SDValue(UpdN.getNode(), i));
8940 // If this is an non-standard-aligned LOAD, the first result is the loaded
8941 // value. Bitcast it to the expected result type.
8942 if (AlignedVecTy != VecTy && N->getOpcode() == ISD::LOAD) {
8943 SDValue &LdVal = NewResults[0];
8944 LdVal = DAG.getNode(ISD::BITCAST, SDLoc(N), VecTy, LdVal);
8947 NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs+1)); // chain
8948 DCI.CombineTo(N, NewResults);
8949 DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs));
8956 static SDValue PerformVLDCombine(SDNode *N,
8957 TargetLowering::DAGCombinerInfo &DCI) {
8958 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
8961 return CombineBaseUpdate(N, DCI);
8964 /// CombineVLDDUP - For a VDUPLANE node N, check if its source operand is a
8965 /// vldN-lane (N > 1) intrinsic, and if all the other uses of that intrinsic
8966 /// are also VDUPLANEs. If so, combine them to a vldN-dup operation and
8968 static bool CombineVLDDUP(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
8969 SelectionDAG &DAG = DCI.DAG;
8970 EVT VT = N->getValueType(0);
8971 // vldN-dup instructions only support 64-bit vectors for N > 1.
8972 if (!VT.is64BitVector())
8975 // Check if the VDUPLANE operand is a vldN-dup intrinsic.
8976 SDNode *VLD = N->getOperand(0).getNode();
8977 if (VLD->getOpcode() != ISD::INTRINSIC_W_CHAIN)
8979 unsigned NumVecs = 0;
8980 unsigned NewOpc = 0;
8981 unsigned IntNo = cast<ConstantSDNode>(VLD->getOperand(1))->getZExtValue();
8982 if (IntNo == Intrinsic::arm_neon_vld2lane) {
8984 NewOpc = ARMISD::VLD2DUP;
8985 } else if (IntNo == Intrinsic::arm_neon_vld3lane) {
8987 NewOpc = ARMISD::VLD3DUP;
8988 } else if (IntNo == Intrinsic::arm_neon_vld4lane) {
8990 NewOpc = ARMISD::VLD4DUP;
8995 // First check that all the vldN-lane uses are VDUPLANEs and that the lane
8996 // numbers match the load.
8997 unsigned VLDLaneNo =
8998 cast<ConstantSDNode>(VLD->getOperand(NumVecs+3))->getZExtValue();
8999 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
9001 // Ignore uses of the chain result.
9002 if (UI.getUse().getResNo() == NumVecs)
9005 if (User->getOpcode() != ARMISD::VDUPLANE ||
9006 VLDLaneNo != cast<ConstantSDNode>(User->getOperand(1))->getZExtValue())
9010 // Create the vldN-dup node.
9013 for (n = 0; n < NumVecs; ++n)
9015 Tys[n] = MVT::Other;
9016 SDVTList SDTys = DAG.getVTList(makeArrayRef(Tys, NumVecs+1));
9017 SDValue Ops[] = { VLD->getOperand(0), VLD->getOperand(2) };
9018 MemIntrinsicSDNode *VLDMemInt = cast<MemIntrinsicSDNode>(VLD);
9019 SDValue VLDDup = DAG.getMemIntrinsicNode(NewOpc, SDLoc(VLD), SDTys,
9020 Ops, VLDMemInt->getMemoryVT(),
9021 VLDMemInt->getMemOperand());
9024 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
9026 unsigned ResNo = UI.getUse().getResNo();
9027 // Ignore uses of the chain result.
9028 if (ResNo == NumVecs)
9031 DCI.CombineTo(User, SDValue(VLDDup.getNode(), ResNo));
9034 // Now the vldN-lane intrinsic is dead except for its chain result.
9035 // Update uses of the chain.
9036 std::vector<SDValue> VLDDupResults;
9037 for (unsigned n = 0; n < NumVecs; ++n)
9038 VLDDupResults.push_back(SDValue(VLDDup.getNode(), n));
9039 VLDDupResults.push_back(SDValue(VLDDup.getNode(), NumVecs));
9040 DCI.CombineTo(VLD, VLDDupResults);
9045 /// PerformVDUPLANECombine - Target-specific dag combine xforms for
9046 /// ARMISD::VDUPLANE.
9047 static SDValue PerformVDUPLANECombine(SDNode *N,
9048 TargetLowering::DAGCombinerInfo &DCI) {
9049 SDValue Op = N->getOperand(0);
9051 // If the source is a vldN-lane (N > 1) intrinsic, and all the other uses
9052 // of that intrinsic are also VDUPLANEs, combine them to a vldN-dup operation.
9053 if (CombineVLDDUP(N, DCI))
9054 return SDValue(N, 0);
9056 // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is
9057 // redundant. Ignore bit_converts for now; element sizes are checked below.
9058 while (Op.getOpcode() == ISD::BITCAST)
9059 Op = Op.getOperand(0);
9060 if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM)
9063 // Make sure the VMOV element size is not bigger than the VDUPLANE elements.
9064 unsigned EltSize = Op.getValueType().getVectorElementType().getSizeInBits();
9065 // The canonical VMOV for a zero vector uses a 32-bit element size.
9066 unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
9068 if (ARM_AM::decodeNEONModImm(Imm, EltBits) == 0)
9070 EVT VT = N->getValueType(0);
9071 if (EltSize > VT.getVectorElementType().getSizeInBits())
9074 return DCI.DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Op);
9077 static SDValue PerformLOADCombine(SDNode *N,
9078 TargetLowering::DAGCombinerInfo &DCI) {
9079 EVT VT = N->getValueType(0);
9081 // If this is a legal vector load, try to combine it into a VLD1_UPD.
9082 if (ISD::isNormalLoad(N) && VT.isVector() &&
9083 DCI.DAG.getTargetLoweringInfo().isTypeLegal(VT))
9084 return CombineBaseUpdate(N, DCI);
9089 /// PerformSTORECombine - Target-specific dag combine xforms for
9091 static SDValue PerformSTORECombine(SDNode *N,
9092 TargetLowering::DAGCombinerInfo &DCI) {
9093 StoreSDNode *St = cast<StoreSDNode>(N);
9094 if (St->isVolatile())
9097 // Optimize trunc store (of multiple scalars) to shuffle and store. First,
9098 // pack all of the elements in one place. Next, store to memory in fewer
9100 SDValue StVal = St->getValue();
9101 EVT VT = StVal.getValueType();
9102 if (St->isTruncatingStore() && VT.isVector()) {
9103 SelectionDAG &DAG = DCI.DAG;
9104 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9105 EVT StVT = St->getMemoryVT();
9106 unsigned NumElems = VT.getVectorNumElements();
9107 assert(StVT != VT && "Cannot truncate to the same type");
9108 unsigned FromEltSz = VT.getVectorElementType().getSizeInBits();
9109 unsigned ToEltSz = StVT.getVectorElementType().getSizeInBits();
9111 // From, To sizes and ElemCount must be pow of two
9112 if (!isPowerOf2_32(NumElems * FromEltSz * ToEltSz)) return SDValue();
9114 // We are going to use the original vector elt for storing.
9115 // Accumulated smaller vector elements must be a multiple of the store size.
9116 if (0 != (NumElems * FromEltSz) % ToEltSz) return SDValue();
9118 unsigned SizeRatio = FromEltSz / ToEltSz;
9119 assert(SizeRatio * NumElems * ToEltSz == VT.getSizeInBits());
9121 // Create a type on which we perform the shuffle.
9122 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(), StVT.getScalarType(),
9123 NumElems*SizeRatio);
9124 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
9127 SDValue WideVec = DAG.getNode(ISD::BITCAST, DL, WideVecVT, StVal);
9128 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
9129 for (unsigned i = 0; i < NumElems; ++i)
9130 ShuffleVec[i] = TLI.isBigEndian() ? (i+1) * SizeRatio - 1 : i * SizeRatio;
9132 // Can't shuffle using an illegal type.
9133 if (!TLI.isTypeLegal(WideVecVT)) return SDValue();
9135 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, DL, WideVec,
9136 DAG.getUNDEF(WideVec.getValueType()),
9138 // At this point all of the data is stored at the bottom of the
9139 // register. We now need to save it to mem.
9141 // Find the largest store unit
9142 MVT StoreType = MVT::i8;
9143 for (MVT Tp : MVT::integer_valuetypes()) {
9144 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToEltSz)
9147 // Didn't find a legal store type.
9148 if (!TLI.isTypeLegal(StoreType))
9151 // Bitcast the original vector into a vector of store-size units
9152 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
9153 StoreType, VT.getSizeInBits()/EVT(StoreType).getSizeInBits());
9154 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
9155 SDValue ShuffWide = DAG.getNode(ISD::BITCAST, DL, StoreVecVT, Shuff);
9156 SmallVector<SDValue, 8> Chains;
9157 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits()/8,
9158 TLI.getPointerTy());
9159 SDValue BasePtr = St->getBasePtr();
9161 // Perform one or more big stores into memory.
9162 unsigned E = (ToEltSz*NumElems)/StoreType.getSizeInBits();
9163 for (unsigned I = 0; I < E; I++) {
9164 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
9165 StoreType, ShuffWide,
9166 DAG.getIntPtrConstant(I));
9167 SDValue Ch = DAG.getStore(St->getChain(), DL, SubVec, BasePtr,
9168 St->getPointerInfo(), St->isVolatile(),
9169 St->isNonTemporal(), St->getAlignment());
9170 BasePtr = DAG.getNode(ISD::ADD, DL, BasePtr.getValueType(), BasePtr,
9172 Chains.push_back(Ch);
9174 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
9177 if (!ISD::isNormalStore(St))
9180 // Split a store of a VMOVDRR into two integer stores to avoid mixing NEON and
9181 // ARM stores of arguments in the same cache line.
9182 if (StVal.getNode()->getOpcode() == ARMISD::VMOVDRR &&
9183 StVal.getNode()->hasOneUse()) {
9184 SelectionDAG &DAG = DCI.DAG;
9185 bool isBigEndian = DAG.getTargetLoweringInfo().isBigEndian();
9187 SDValue BasePtr = St->getBasePtr();
9188 SDValue NewST1 = DAG.getStore(St->getChain(), DL,
9189 StVal.getNode()->getOperand(isBigEndian ? 1 : 0 ),
9190 BasePtr, St->getPointerInfo(), St->isVolatile(),
9191 St->isNonTemporal(), St->getAlignment());
9193 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
9194 DAG.getConstant(4, MVT::i32));
9195 return DAG.getStore(NewST1.getValue(0), DL,
9196 StVal.getNode()->getOperand(isBigEndian ? 0 : 1),
9197 OffsetPtr, St->getPointerInfo(), St->isVolatile(),
9198 St->isNonTemporal(),
9199 std::min(4U, St->getAlignment() / 2));
9202 if (StVal.getValueType() == MVT::i64 &&
9203 StVal.getNode()->getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
9205 // Bitcast an i64 store extracted from a vector to f64.
9206 // Otherwise, the i64 value will be legalized to a pair of i32 values.
9207 SelectionDAG &DAG = DCI.DAG;
9209 SDValue IntVec = StVal.getOperand(0);
9210 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
9211 IntVec.getValueType().getVectorNumElements());
9212 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, IntVec);
9213 SDValue ExtElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
9214 Vec, StVal.getOperand(1));
9216 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ExtElt);
9217 // Make the DAGCombiner fold the bitcasts.
9218 DCI.AddToWorklist(Vec.getNode());
9219 DCI.AddToWorklist(ExtElt.getNode());
9220 DCI.AddToWorklist(V.getNode());
9221 return DAG.getStore(St->getChain(), dl, V, St->getBasePtr(),
9222 St->getPointerInfo(), St->isVolatile(),
9223 St->isNonTemporal(), St->getAlignment(),
9227 // If this is a legal vector store, try to combine it into a VST1_UPD.
9228 if (ISD::isNormalStore(N) && VT.isVector() &&
9229 DCI.DAG.getTargetLoweringInfo().isTypeLegal(VT))
9230 return CombineBaseUpdate(N, DCI);
9235 // isConstVecPow2 - Return true if each vector element is a power of 2, all
9236 // elements are the same constant, C, and Log2(C) ranges from 1 to 32.
9237 static bool isConstVecPow2(SDValue ConstVec, bool isSigned, uint64_t &C)
9241 for (unsigned I = 0, E = ConstVec.getValueType().getVectorNumElements();
9243 ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(ConstVec.getOperand(I));
9248 APFloat APF = C->getValueAPF();
9249 if (APF.convertToInteger(&cN, 64, isSigned, APFloat::rmTowardZero, &isExact)
9250 != APFloat::opOK || !isExact)
9253 c0 = (I == 0) ? cN : c0;
9254 if (!isPowerOf2_64(cN) || c0 != cN || Log2_64(c0) < 1 || Log2_64(c0) > 32)
9261 /// PerformVCVTCombine - VCVT (floating-point to fixed-point, Advanced SIMD)
9262 /// can replace combinations of VMUL and VCVT (floating-point to integer)
9263 /// when the VMUL has a constant operand that is a power of 2.
9265 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
9266 /// vmul.f32 d16, d17, d16
9267 /// vcvt.s32.f32 d16, d16
9269 /// vcvt.s32.f32 d16, d16, #3
9270 static SDValue PerformVCVTCombine(SDNode *N,
9271 TargetLowering::DAGCombinerInfo &DCI,
9272 const ARMSubtarget *Subtarget) {
9273 SelectionDAG &DAG = DCI.DAG;
9274 SDValue Op = N->getOperand(0);
9276 if (!Subtarget->hasNEON() || !Op.getValueType().isVector() ||
9277 Op.getOpcode() != ISD::FMUL)
9281 SDValue N0 = Op->getOperand(0);
9282 SDValue ConstVec = Op->getOperand(1);
9283 bool isSigned = N->getOpcode() == ISD::FP_TO_SINT;
9285 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
9286 !isConstVecPow2(ConstVec, isSigned, C))
9289 MVT FloatTy = Op.getSimpleValueType().getVectorElementType();
9290 MVT IntTy = N->getSimpleValueType(0).getVectorElementType();
9291 unsigned NumLanes = Op.getValueType().getVectorNumElements();
9292 if (FloatTy.getSizeInBits() != 32 || IntTy.getSizeInBits() > 32 ||
9294 // These instructions only exist converting from f32 to i32. We can handle
9295 // smaller integers by generating an extra truncate, but larger ones would
9296 // be lossy. We also can't handle more then 4 lanes, since these intructions
9297 // only support v2i32/v4i32 types.
9301 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfp2fxs :
9302 Intrinsic::arm_neon_vcvtfp2fxu;
9303 SDValue FixConv = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N),
9304 NumLanes == 2 ? MVT::v2i32 : MVT::v4i32,
9305 DAG.getConstant(IntrinsicOpcode, MVT::i32), N0,
9306 DAG.getConstant(Log2_64(C), MVT::i32));
9308 if (IntTy.getSizeInBits() < FloatTy.getSizeInBits())
9309 FixConv = DAG.getNode(ISD::TRUNCATE, SDLoc(N), N->getValueType(0), FixConv);
9314 /// PerformVDIVCombine - VCVT (fixed-point to floating-point, Advanced SIMD)
9315 /// can replace combinations of VCVT (integer to floating-point) and VDIV
9316 /// when the VDIV has a constant operand that is a power of 2.
9318 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
9319 /// vcvt.f32.s32 d16, d16
9320 /// vdiv.f32 d16, d17, d16
9322 /// vcvt.f32.s32 d16, d16, #3
9323 static SDValue PerformVDIVCombine(SDNode *N,
9324 TargetLowering::DAGCombinerInfo &DCI,
9325 const ARMSubtarget *Subtarget) {
9326 SelectionDAG &DAG = DCI.DAG;
9327 SDValue Op = N->getOperand(0);
9328 unsigned OpOpcode = Op.getNode()->getOpcode();
9330 if (!Subtarget->hasNEON() || !N->getValueType(0).isVector() ||
9331 (OpOpcode != ISD::SINT_TO_FP && OpOpcode != ISD::UINT_TO_FP))
9335 SDValue ConstVec = N->getOperand(1);
9336 bool isSigned = OpOpcode == ISD::SINT_TO_FP;
9338 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
9339 !isConstVecPow2(ConstVec, isSigned, C))
9342 MVT FloatTy = N->getSimpleValueType(0).getVectorElementType();
9343 MVT IntTy = Op.getOperand(0).getSimpleValueType().getVectorElementType();
9344 if (FloatTy.getSizeInBits() != 32 || IntTy.getSizeInBits() > 32) {
9345 // These instructions only exist converting from i32 to f32. We can handle
9346 // smaller integers by generating an extra extend, but larger ones would
9351 SDValue ConvInput = Op.getOperand(0);
9352 unsigned NumLanes = Op.getValueType().getVectorNumElements();
9353 if (IntTy.getSizeInBits() < FloatTy.getSizeInBits())
9354 ConvInput = DAG.getNode(isSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND,
9355 SDLoc(N), NumLanes == 2 ? MVT::v2i32 : MVT::v4i32,
9358 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfxs2fp :
9359 Intrinsic::arm_neon_vcvtfxu2fp;
9360 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N),
9362 DAG.getConstant(IntrinsicOpcode, MVT::i32),
9363 ConvInput, DAG.getConstant(Log2_64(C), MVT::i32));
9366 /// Getvshiftimm - Check if this is a valid build_vector for the immediate
9367 /// operand of a vector shift operation, where all the elements of the
9368 /// build_vector must have the same constant integer value.
9369 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
9370 // Ignore bit_converts.
9371 while (Op.getOpcode() == ISD::BITCAST)
9372 Op = Op.getOperand(0);
9373 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
9374 APInt SplatBits, SplatUndef;
9375 unsigned SplatBitSize;
9377 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
9378 HasAnyUndefs, ElementBits) ||
9379 SplatBitSize > ElementBits)
9381 Cnt = SplatBits.getSExtValue();
9385 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
9386 /// operand of a vector shift left operation. That value must be in the range:
9387 /// 0 <= Value < ElementBits for a left shift; or
9388 /// 0 <= Value <= ElementBits for a long left shift.
9389 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
9390 assert(VT.isVector() && "vector shift count is not a vector type");
9391 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
9392 if (! getVShiftImm(Op, ElementBits, Cnt))
9394 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
9397 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
9398 /// operand of a vector shift right operation. For a shift opcode, the value
9399 /// is positive, but for an intrinsic the value count must be negative. The
9400 /// absolute value must be in the range:
9401 /// 1 <= |Value| <= ElementBits for a right shift; or
9402 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
9403 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
9405 assert(VT.isVector() && "vector shift count is not a vector type");
9406 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
9407 if (! getVShiftImm(Op, ElementBits, Cnt))
9411 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
9414 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
9415 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
9416 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
9419 // Don't do anything for most intrinsics.
9422 // Vector shifts: check for immediate versions and lower them.
9423 // Note: This is done during DAG combining instead of DAG legalizing because
9424 // the build_vectors for 64-bit vector element shift counts are generally
9425 // not legal, and it is hard to see their values after they get legalized to
9426 // loads from a constant pool.
9427 case Intrinsic::arm_neon_vshifts:
9428 case Intrinsic::arm_neon_vshiftu:
9429 case Intrinsic::arm_neon_vrshifts:
9430 case Intrinsic::arm_neon_vrshiftu:
9431 case Intrinsic::arm_neon_vrshiftn:
9432 case Intrinsic::arm_neon_vqshifts:
9433 case Intrinsic::arm_neon_vqshiftu:
9434 case Intrinsic::arm_neon_vqshiftsu:
9435 case Intrinsic::arm_neon_vqshiftns:
9436 case Intrinsic::arm_neon_vqshiftnu:
9437 case Intrinsic::arm_neon_vqshiftnsu:
9438 case Intrinsic::arm_neon_vqrshiftns:
9439 case Intrinsic::arm_neon_vqrshiftnu:
9440 case Intrinsic::arm_neon_vqrshiftnsu: {
9441 EVT VT = N->getOperand(1).getValueType();
9443 unsigned VShiftOpc = 0;
9446 case Intrinsic::arm_neon_vshifts:
9447 case Intrinsic::arm_neon_vshiftu:
9448 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
9449 VShiftOpc = ARMISD::VSHL;
9452 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
9453 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
9454 ARMISD::VSHRs : ARMISD::VSHRu);
9459 case Intrinsic::arm_neon_vrshifts:
9460 case Intrinsic::arm_neon_vrshiftu:
9461 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
9465 case Intrinsic::arm_neon_vqshifts:
9466 case Intrinsic::arm_neon_vqshiftu:
9467 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
9471 case Intrinsic::arm_neon_vqshiftsu:
9472 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
9474 llvm_unreachable("invalid shift count for vqshlu intrinsic");
9476 case Intrinsic::arm_neon_vrshiftn:
9477 case Intrinsic::arm_neon_vqshiftns:
9478 case Intrinsic::arm_neon_vqshiftnu:
9479 case Intrinsic::arm_neon_vqshiftnsu:
9480 case Intrinsic::arm_neon_vqrshiftns:
9481 case Intrinsic::arm_neon_vqrshiftnu:
9482 case Intrinsic::arm_neon_vqrshiftnsu:
9483 // Narrowing shifts require an immediate right shift.
9484 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
9486 llvm_unreachable("invalid shift count for narrowing vector shift "
9490 llvm_unreachable("unhandled vector shift");
9494 case Intrinsic::arm_neon_vshifts:
9495 case Intrinsic::arm_neon_vshiftu:
9496 // Opcode already set above.
9498 case Intrinsic::arm_neon_vrshifts:
9499 VShiftOpc = ARMISD::VRSHRs; break;
9500 case Intrinsic::arm_neon_vrshiftu:
9501 VShiftOpc = ARMISD::VRSHRu; break;
9502 case Intrinsic::arm_neon_vrshiftn:
9503 VShiftOpc = ARMISD::VRSHRN; break;
9504 case Intrinsic::arm_neon_vqshifts:
9505 VShiftOpc = ARMISD::VQSHLs; break;
9506 case Intrinsic::arm_neon_vqshiftu:
9507 VShiftOpc = ARMISD::VQSHLu; break;
9508 case Intrinsic::arm_neon_vqshiftsu:
9509 VShiftOpc = ARMISD::VQSHLsu; break;
9510 case Intrinsic::arm_neon_vqshiftns:
9511 VShiftOpc = ARMISD::VQSHRNs; break;
9512 case Intrinsic::arm_neon_vqshiftnu:
9513 VShiftOpc = ARMISD::VQSHRNu; break;
9514 case Intrinsic::arm_neon_vqshiftnsu:
9515 VShiftOpc = ARMISD::VQSHRNsu; break;
9516 case Intrinsic::arm_neon_vqrshiftns:
9517 VShiftOpc = ARMISD::VQRSHRNs; break;
9518 case Intrinsic::arm_neon_vqrshiftnu:
9519 VShiftOpc = ARMISD::VQRSHRNu; break;
9520 case Intrinsic::arm_neon_vqrshiftnsu:
9521 VShiftOpc = ARMISD::VQRSHRNsu; break;
9524 return DAG.getNode(VShiftOpc, SDLoc(N), N->getValueType(0),
9525 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
9528 case Intrinsic::arm_neon_vshiftins: {
9529 EVT VT = N->getOperand(1).getValueType();
9531 unsigned VShiftOpc = 0;
9533 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
9534 VShiftOpc = ARMISD::VSLI;
9535 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
9536 VShiftOpc = ARMISD::VSRI;
9538 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
9541 return DAG.getNode(VShiftOpc, SDLoc(N), N->getValueType(0),
9542 N->getOperand(1), N->getOperand(2),
9543 DAG.getConstant(Cnt, MVT::i32));
9546 case Intrinsic::arm_neon_vqrshifts:
9547 case Intrinsic::arm_neon_vqrshiftu:
9548 // No immediate versions of these to check for.
9555 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
9556 /// lowers them. As with the vector shift intrinsics, this is done during DAG
9557 /// combining instead of DAG legalizing because the build_vectors for 64-bit
9558 /// vector element shift counts are generally not legal, and it is hard to see
9559 /// their values after they get legalized to loads from a constant pool.
9560 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
9561 const ARMSubtarget *ST) {
9562 EVT VT = N->getValueType(0);
9563 if (N->getOpcode() == ISD::SRL && VT == MVT::i32 && ST->hasV6Ops()) {
9564 // Canonicalize (srl (bswap x), 16) to (rotr (bswap x), 16) if the high
9565 // 16-bits of x is zero. This optimizes rev + lsr 16 to rev16.
9566 SDValue N1 = N->getOperand(1);
9567 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) {
9568 SDValue N0 = N->getOperand(0);
9569 if (C->getZExtValue() == 16 && N0.getOpcode() == ISD::BSWAP &&
9570 DAG.MaskedValueIsZero(N0.getOperand(0),
9571 APInt::getHighBitsSet(32, 16)))
9572 return DAG.getNode(ISD::ROTR, SDLoc(N), VT, N0, N1);
9576 // Nothing to be done for scalar shifts.
9577 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9578 if (!VT.isVector() || !TLI.isTypeLegal(VT))
9581 assert(ST->hasNEON() && "unexpected vector shift");
9584 switch (N->getOpcode()) {
9585 default: llvm_unreachable("unexpected shift opcode");
9588 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
9589 return DAG.getNode(ARMISD::VSHL, SDLoc(N), VT, N->getOperand(0),
9590 DAG.getConstant(Cnt, MVT::i32));
9595 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
9596 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
9597 ARMISD::VSHRs : ARMISD::VSHRu);
9598 return DAG.getNode(VShiftOpc, SDLoc(N), VT, N->getOperand(0),
9599 DAG.getConstant(Cnt, MVT::i32));
9605 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
9606 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
9607 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
9608 const ARMSubtarget *ST) {
9609 SDValue N0 = N->getOperand(0);
9611 // Check for sign- and zero-extensions of vector extract operations of 8-
9612 // and 16-bit vector elements. NEON supports these directly. They are
9613 // handled during DAG combining because type legalization will promote them
9614 // to 32-bit types and it is messy to recognize the operations after that.
9615 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
9616 SDValue Vec = N0.getOperand(0);
9617 SDValue Lane = N0.getOperand(1);
9618 EVT VT = N->getValueType(0);
9619 EVT EltVT = N0.getValueType();
9620 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9622 if (VT == MVT::i32 &&
9623 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
9624 TLI.isTypeLegal(Vec.getValueType()) &&
9625 isa<ConstantSDNode>(Lane)) {
9628 switch (N->getOpcode()) {
9629 default: llvm_unreachable("unexpected opcode");
9630 case ISD::SIGN_EXTEND:
9631 Opc = ARMISD::VGETLANEs;
9633 case ISD::ZERO_EXTEND:
9634 case ISD::ANY_EXTEND:
9635 Opc = ARMISD::VGETLANEu;
9638 return DAG.getNode(Opc, SDLoc(N), VT, Vec, Lane);
9645 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
9646 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
9647 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
9648 const ARMSubtarget *ST) {
9649 // If the target supports NEON, try to use vmax/vmin instructions for f32
9650 // selects like "x < y ? x : y". Unless the NoNaNsFPMath option is set,
9651 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
9652 // a NaN; only do the transformation when it matches that behavior.
9654 // For now only do this when using NEON for FP operations; if using VFP, it
9655 // is not obvious that the benefit outweighs the cost of switching to the
9657 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
9658 N->getValueType(0) != MVT::f32)
9661 SDValue CondLHS = N->getOperand(0);
9662 SDValue CondRHS = N->getOperand(1);
9663 SDValue LHS = N->getOperand(2);
9664 SDValue RHS = N->getOperand(3);
9665 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
9667 unsigned Opcode = 0;
9669 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
9670 IsReversed = false; // x CC y ? x : y
9671 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
9672 IsReversed = true ; // x CC y ? y : x
9686 // If LHS is NaN, an ordered comparison will be false and the result will
9687 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
9688 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
9689 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
9690 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
9692 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
9693 // will return -0, so vmin can only be used for unsafe math or if one of
9694 // the operands is known to be nonzero.
9695 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
9696 !DAG.getTarget().Options.UnsafeFPMath &&
9697 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9699 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
9708 // If LHS is NaN, an ordered comparison will be false and the result will
9709 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
9710 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
9711 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
9712 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
9714 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
9715 // will return +0, so vmax can only be used for unsafe math or if one of
9716 // the operands is known to be nonzero.
9717 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
9718 !DAG.getTarget().Options.UnsafeFPMath &&
9719 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9721 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
9727 return DAG.getNode(Opcode, SDLoc(N), N->getValueType(0), LHS, RHS);
9730 /// PerformCMOVCombine - Target-specific DAG combining for ARMISD::CMOV.
9732 ARMTargetLowering::PerformCMOVCombine(SDNode *N, SelectionDAG &DAG) const {
9733 SDValue Cmp = N->getOperand(4);
9734 if (Cmp.getOpcode() != ARMISD::CMPZ)
9735 // Only looking at EQ and NE cases.
9738 EVT VT = N->getValueType(0);
9740 SDValue LHS = Cmp.getOperand(0);
9741 SDValue RHS = Cmp.getOperand(1);
9742 SDValue FalseVal = N->getOperand(0);
9743 SDValue TrueVal = N->getOperand(1);
9744 SDValue ARMcc = N->getOperand(2);
9745 ARMCC::CondCodes CC =
9746 (ARMCC::CondCodes)cast<ConstantSDNode>(ARMcc)->getZExtValue();
9764 /// FIXME: Turn this into a target neutral optimization?
9766 if (CC == ARMCC::NE && FalseVal == RHS && FalseVal != LHS) {
9767 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, TrueVal, ARMcc,
9768 N->getOperand(3), Cmp);
9769 } else if (CC == ARMCC::EQ && TrueVal == RHS) {
9771 SDValue NewCmp = getARMCmp(LHS, RHS, ISD::SETNE, ARMcc, DAG, dl);
9772 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, FalseVal, ARMcc,
9773 N->getOperand(3), NewCmp);
9776 if (Res.getNode()) {
9777 APInt KnownZero, KnownOne;
9778 DAG.computeKnownBits(SDValue(N,0), KnownZero, KnownOne);
9779 // Capture demanded bits information that would be otherwise lost.
9780 if (KnownZero == 0xfffffffe)
9781 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
9782 DAG.getValueType(MVT::i1));
9783 else if (KnownZero == 0xffffff00)
9784 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
9785 DAG.getValueType(MVT::i8));
9786 else if (KnownZero == 0xffff0000)
9787 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
9788 DAG.getValueType(MVT::i16));
9794 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
9795 DAGCombinerInfo &DCI) const {
9796 switch (N->getOpcode()) {
9798 case ISD::ADDC: return PerformADDCCombine(N, DCI, Subtarget);
9799 case ISD::ADD: return PerformADDCombine(N, DCI, Subtarget);
9800 case ISD::SUB: return PerformSUBCombine(N, DCI);
9801 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
9802 case ISD::OR: return PerformORCombine(N, DCI, Subtarget);
9803 case ISD::XOR: return PerformXORCombine(N, DCI, Subtarget);
9804 case ISD::AND: return PerformANDCombine(N, DCI, Subtarget);
9805 case ARMISD::BFI: return PerformBFICombine(N, DCI);
9806 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI, Subtarget);
9807 case ARMISD::VMOVDRR: return PerformVMOVDRRCombine(N, DCI.DAG);
9808 case ISD::STORE: return PerformSTORECombine(N, DCI);
9809 case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DCI, Subtarget);
9810 case ISD::INSERT_VECTOR_ELT: return PerformInsertEltCombine(N, DCI);
9811 case ISD::VECTOR_SHUFFLE: return PerformVECTOR_SHUFFLECombine(N, DCI.DAG);
9812 case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI);
9813 case ISD::FP_TO_SINT:
9814 case ISD::FP_TO_UINT: return PerformVCVTCombine(N, DCI, Subtarget);
9815 case ISD::FDIV: return PerformVDIVCombine(N, DCI, Subtarget);
9816 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
9819 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
9820 case ISD::SIGN_EXTEND:
9821 case ISD::ZERO_EXTEND:
9822 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
9823 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
9824 case ARMISD::CMOV: return PerformCMOVCombine(N, DCI.DAG);
9825 case ISD::LOAD: return PerformLOADCombine(N, DCI);
9826 case ARMISD::VLD2DUP:
9827 case ARMISD::VLD3DUP:
9828 case ARMISD::VLD4DUP:
9829 return PerformVLDCombine(N, DCI);
9830 case ARMISD::BUILD_VECTOR:
9831 return PerformARMBUILD_VECTORCombine(N, DCI);
9832 case ISD::INTRINSIC_VOID:
9833 case ISD::INTRINSIC_W_CHAIN:
9834 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
9835 case Intrinsic::arm_neon_vld1:
9836 case Intrinsic::arm_neon_vld2:
9837 case Intrinsic::arm_neon_vld3:
9838 case Intrinsic::arm_neon_vld4:
9839 case Intrinsic::arm_neon_vld2lane:
9840 case Intrinsic::arm_neon_vld3lane:
9841 case Intrinsic::arm_neon_vld4lane:
9842 case Intrinsic::arm_neon_vst1:
9843 case Intrinsic::arm_neon_vst2:
9844 case Intrinsic::arm_neon_vst3:
9845 case Intrinsic::arm_neon_vst4:
9846 case Intrinsic::arm_neon_vst2lane:
9847 case Intrinsic::arm_neon_vst3lane:
9848 case Intrinsic::arm_neon_vst4lane:
9849 return PerformVLDCombine(N, DCI);
9857 bool ARMTargetLowering::isDesirableToTransformToIntegerOp(unsigned Opc,
9859 return (VT == MVT::f32) && (Opc == ISD::LOAD || Opc == ISD::STORE);
9862 bool ARMTargetLowering::allowsMisalignedMemoryAccesses(EVT VT,
9866 // The AllowsUnaliged flag models the SCTLR.A setting in ARM cpus
9867 bool AllowsUnaligned = Subtarget->allowsUnalignedMem();
9869 switch (VT.getSimpleVT().SimpleTy) {
9875 // Unaligned access can use (for example) LRDB, LRDH, LDR
9876 if (AllowsUnaligned) {
9878 *Fast = Subtarget->hasV7Ops();
9885 // For any little-endian targets with neon, we can support unaligned ld/st
9886 // of D and Q (e.g. {D0,D1}) registers by using vld1.i8/vst1.i8.
9887 // A big-endian target may also explicitly support unaligned accesses
9888 if (Subtarget->hasNEON() && (AllowsUnaligned || isLittleEndian())) {
9898 static bool memOpAlign(unsigned DstAlign, unsigned SrcAlign,
9899 unsigned AlignCheck) {
9900 return ((SrcAlign == 0 || SrcAlign % AlignCheck == 0) &&
9901 (DstAlign == 0 || DstAlign % AlignCheck == 0));
9904 EVT ARMTargetLowering::getOptimalMemOpType(uint64_t Size,
9905 unsigned DstAlign, unsigned SrcAlign,
9906 bool IsMemset, bool ZeroMemset,
9908 MachineFunction &MF) const {
9909 const Function *F = MF.getFunction();
9911 // See if we can use NEON instructions for this...
9912 if ((!IsMemset || ZeroMemset) && Subtarget->hasNEON() &&
9913 !F->hasFnAttribute(Attribute::NoImplicitFloat)) {
9916 (memOpAlign(SrcAlign, DstAlign, 16) ||
9917 (allowsMisalignedMemoryAccesses(MVT::v2f64, 0, 1, &Fast) && Fast))) {
9919 } else if (Size >= 8 &&
9920 (memOpAlign(SrcAlign, DstAlign, 8) ||
9921 (allowsMisalignedMemoryAccesses(MVT::f64, 0, 1, &Fast) &&
9927 // Lowering to i32/i16 if the size permits.
9933 // Let the target-independent logic figure it out.
9937 bool ARMTargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
9938 if (Val.getOpcode() != ISD::LOAD)
9941 EVT VT1 = Val.getValueType();
9942 if (!VT1.isSimple() || !VT1.isInteger() ||
9943 !VT2.isSimple() || !VT2.isInteger())
9946 switch (VT1.getSimpleVT().SimpleTy) {
9951 // 8-bit and 16-bit loads implicitly zero-extend to 32-bits.
9958 bool ARMTargetLowering::isVectorLoadExtDesirable(SDValue ExtVal) const {
9959 EVT VT = ExtVal.getValueType();
9961 if (!isTypeLegal(VT))
9964 // Don't create a loadext if we can fold the extension into a wide/long
9966 // If there's more than one user instruction, the loadext is desirable no
9967 // matter what. There can be two uses by the same instruction.
9968 if (ExtVal->use_empty() ||
9969 !ExtVal->use_begin()->isOnlyUserOf(ExtVal.getNode()))
9972 SDNode *U = *ExtVal->use_begin();
9973 if ((U->getOpcode() == ISD::ADD || U->getOpcode() == ISD::SUB ||
9974 U->getOpcode() == ISD::SHL || U->getOpcode() == ARMISD::VSHL))
9980 bool ARMTargetLowering::allowTruncateForTailCall(Type *Ty1, Type *Ty2) const {
9981 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
9984 if (!isTypeLegal(EVT::getEVT(Ty1)))
9987 assert(Ty1->getPrimitiveSizeInBits() <= 64 && "i128 is probably not a noop");
9989 // Assuming the caller doesn't have a zeroext or signext return parameter,
9990 // truncation all the way down to i1 is valid.
9995 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
10000 switch (VT.getSimpleVT().SimpleTy) {
10001 default: return false;
10016 if ((V & (Scale - 1)) != 0)
10019 return V == (V & ((1LL << 5) - 1));
10022 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
10023 const ARMSubtarget *Subtarget) {
10024 bool isNeg = false;
10030 switch (VT.getSimpleVT().SimpleTy) {
10031 default: return false;
10036 // + imm12 or - imm8
10038 return V == (V & ((1LL << 8) - 1));
10039 return V == (V & ((1LL << 12) - 1));
10042 // Same as ARM mode. FIXME: NEON?
10043 if (!Subtarget->hasVFP2())
10048 return V == (V & ((1LL << 8) - 1));
10052 /// isLegalAddressImmediate - Return true if the integer value can be used
10053 /// as the offset of the target addressing mode for load / store of the
10055 static bool isLegalAddressImmediate(int64_t V, EVT VT,
10056 const ARMSubtarget *Subtarget) {
10060 if (!VT.isSimple())
10063 if (Subtarget->isThumb1Only())
10064 return isLegalT1AddressImmediate(V, VT);
10065 else if (Subtarget->isThumb2())
10066 return isLegalT2AddressImmediate(V, VT, Subtarget);
10071 switch (VT.getSimpleVT().SimpleTy) {
10072 default: return false;
10077 return V == (V & ((1LL << 12) - 1));
10080 return V == (V & ((1LL << 8) - 1));
10083 if (!Subtarget->hasVFP2()) // FIXME: NEON?
10088 return V == (V & ((1LL << 8) - 1));
10092 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
10094 int Scale = AM.Scale;
10098 switch (VT.getSimpleVT().SimpleTy) {
10099 default: return false;
10107 Scale = Scale & ~1;
10108 return Scale == 2 || Scale == 4 || Scale == 8;
10111 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
10115 // Note, we allow "void" uses (basically, uses that aren't loads or
10116 // stores), because arm allows folding a scale into many arithmetic
10117 // operations. This should be made more precise and revisited later.
10119 // Allow r << imm, but the imm has to be a multiple of two.
10120 if (Scale & 1) return false;
10121 return isPowerOf2_32(Scale);
10125 /// isLegalAddressingMode - Return true if the addressing mode represented
10126 /// by AM is legal for this target, for a load/store of the specified type.
10127 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
10129 EVT VT = getValueType(Ty, true);
10130 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
10133 // Can never fold addr of global into load/store.
10137 switch (AM.Scale) {
10138 case 0: // no scale reg, must be "r+i" or "r", or "i".
10141 if (Subtarget->isThumb1Only())
10145 // ARM doesn't support any R+R*scale+imm addr modes.
10149 if (!VT.isSimple())
10152 if (Subtarget->isThumb2())
10153 return isLegalT2ScaledAddressingMode(AM, VT);
10155 int Scale = AM.Scale;
10156 switch (VT.getSimpleVT().SimpleTy) {
10157 default: return false;
10161 if (Scale < 0) Scale = -Scale;
10165 return isPowerOf2_32(Scale & ~1);
10169 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
10174 // Note, we allow "void" uses (basically, uses that aren't loads or
10175 // stores), because arm allows folding a scale into many arithmetic
10176 // operations. This should be made more precise and revisited later.
10178 // Allow r << imm, but the imm has to be a multiple of two.
10179 if (Scale & 1) return false;
10180 return isPowerOf2_32(Scale);
10186 /// isLegalICmpImmediate - Return true if the specified immediate is legal
10187 /// icmp immediate, that is the target has icmp instructions which can compare
10188 /// a register against the immediate without having to materialize the
10189 /// immediate into a register.
10190 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
10191 // Thumb2 and ARM modes can use cmn for negative immediates.
10192 if (!Subtarget->isThumb())
10193 return ARM_AM::getSOImmVal(std::abs(Imm)) != -1;
10194 if (Subtarget->isThumb2())
10195 return ARM_AM::getT2SOImmVal(std::abs(Imm)) != -1;
10196 // Thumb1 doesn't have cmn, and only 8-bit immediates.
10197 return Imm >= 0 && Imm <= 255;
10200 /// isLegalAddImmediate - Return true if the specified immediate is a legal add
10201 /// *or sub* immediate, that is the target has add or sub instructions which can
10202 /// add a register with the immediate without having to materialize the
10203 /// immediate into a register.
10204 bool ARMTargetLowering::isLegalAddImmediate(int64_t Imm) const {
10205 // Same encoding for add/sub, just flip the sign.
10206 int64_t AbsImm = std::abs(Imm);
10207 if (!Subtarget->isThumb())
10208 return ARM_AM::getSOImmVal(AbsImm) != -1;
10209 if (Subtarget->isThumb2())
10210 return ARM_AM::getT2SOImmVal(AbsImm) != -1;
10211 // Thumb1 only has 8-bit unsigned immediate.
10212 return AbsImm >= 0 && AbsImm <= 255;
10215 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
10216 bool isSEXTLoad, SDValue &Base,
10217 SDValue &Offset, bool &isInc,
10218 SelectionDAG &DAG) {
10219 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
10222 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
10223 // AddressingMode 3
10224 Base = Ptr->getOperand(0);
10225 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
10226 int RHSC = (int)RHS->getZExtValue();
10227 if (RHSC < 0 && RHSC > -256) {
10228 assert(Ptr->getOpcode() == ISD::ADD);
10230 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
10234 isInc = (Ptr->getOpcode() == ISD::ADD);
10235 Offset = Ptr->getOperand(1);
10237 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
10238 // AddressingMode 2
10239 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
10240 int RHSC = (int)RHS->getZExtValue();
10241 if (RHSC < 0 && RHSC > -0x1000) {
10242 assert(Ptr->getOpcode() == ISD::ADD);
10244 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
10245 Base = Ptr->getOperand(0);
10250 if (Ptr->getOpcode() == ISD::ADD) {
10252 ARM_AM::ShiftOpc ShOpcVal=
10253 ARM_AM::getShiftOpcForNode(Ptr->getOperand(0).getOpcode());
10254 if (ShOpcVal != ARM_AM::no_shift) {
10255 Base = Ptr->getOperand(1);
10256 Offset = Ptr->getOperand(0);
10258 Base = Ptr->getOperand(0);
10259 Offset = Ptr->getOperand(1);
10264 isInc = (Ptr->getOpcode() == ISD::ADD);
10265 Base = Ptr->getOperand(0);
10266 Offset = Ptr->getOperand(1);
10270 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
10274 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
10275 bool isSEXTLoad, SDValue &Base,
10276 SDValue &Offset, bool &isInc,
10277 SelectionDAG &DAG) {
10278 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
10281 Base = Ptr->getOperand(0);
10282 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
10283 int RHSC = (int)RHS->getZExtValue();
10284 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
10285 assert(Ptr->getOpcode() == ISD::ADD);
10287 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
10289 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
10290 isInc = Ptr->getOpcode() == ISD::ADD;
10291 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
10299 /// getPreIndexedAddressParts - returns true by value, base pointer and
10300 /// offset pointer and addressing mode by reference if the node's address
10301 /// can be legally represented as pre-indexed load / store address.
10303 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
10305 ISD::MemIndexedMode &AM,
10306 SelectionDAG &DAG) const {
10307 if (Subtarget->isThumb1Only())
10312 bool isSEXTLoad = false;
10313 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
10314 Ptr = LD->getBasePtr();
10315 VT = LD->getMemoryVT();
10316 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
10317 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
10318 Ptr = ST->getBasePtr();
10319 VT = ST->getMemoryVT();
10324 bool isLegal = false;
10325 if (Subtarget->isThumb2())
10326 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
10327 Offset, isInc, DAG);
10329 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
10330 Offset, isInc, DAG);
10334 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
10338 /// getPostIndexedAddressParts - returns true by value, base pointer and
10339 /// offset pointer and addressing mode by reference if this node can be
10340 /// combined with a load / store to form a post-indexed load / store.
10341 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
10344 ISD::MemIndexedMode &AM,
10345 SelectionDAG &DAG) const {
10346 if (Subtarget->isThumb1Only())
10351 bool isSEXTLoad = false;
10352 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
10353 VT = LD->getMemoryVT();
10354 Ptr = LD->getBasePtr();
10355 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
10356 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
10357 VT = ST->getMemoryVT();
10358 Ptr = ST->getBasePtr();
10363 bool isLegal = false;
10364 if (Subtarget->isThumb2())
10365 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
10368 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
10374 // Swap base ptr and offset to catch more post-index load / store when
10375 // it's legal. In Thumb2 mode, offset must be an immediate.
10376 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
10377 !Subtarget->isThumb2())
10378 std::swap(Base, Offset);
10380 // Post-indexed load / store update the base pointer.
10385 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
10389 void ARMTargetLowering::computeKnownBitsForTargetNode(const SDValue Op,
10392 const SelectionDAG &DAG,
10393 unsigned Depth) const {
10394 unsigned BitWidth = KnownOne.getBitWidth();
10395 KnownZero = KnownOne = APInt(BitWidth, 0);
10396 switch (Op.getOpcode()) {
10402 // These nodes' second result is a boolean
10403 if (Op.getResNo() == 0)
10405 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - 1);
10407 case ARMISD::CMOV: {
10408 // Bits are known zero/one if known on the LHS and RHS.
10409 DAG.computeKnownBits(Op.getOperand(0), KnownZero, KnownOne, Depth+1);
10410 if (KnownZero == 0 && KnownOne == 0) return;
10412 APInt KnownZeroRHS, KnownOneRHS;
10413 DAG.computeKnownBits(Op.getOperand(1), KnownZeroRHS, KnownOneRHS, Depth+1);
10414 KnownZero &= KnownZeroRHS;
10415 KnownOne &= KnownOneRHS;
10418 case ISD::INTRINSIC_W_CHAIN: {
10419 ConstantSDNode *CN = cast<ConstantSDNode>(Op->getOperand(1));
10420 Intrinsic::ID IntID = static_cast<Intrinsic::ID>(CN->getZExtValue());
10423 case Intrinsic::arm_ldaex:
10424 case Intrinsic::arm_ldrex: {
10425 EVT VT = cast<MemIntrinsicSDNode>(Op)->getMemoryVT();
10426 unsigned MemBits = VT.getScalarType().getSizeInBits();
10427 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - MemBits);
10435 //===----------------------------------------------------------------------===//
10436 // ARM Inline Assembly Support
10437 //===----------------------------------------------------------------------===//
10439 bool ARMTargetLowering::ExpandInlineAsm(CallInst *CI) const {
10440 // Looking for "rev" which is V6+.
10441 if (!Subtarget->hasV6Ops())
10444 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
10445 std::string AsmStr = IA->getAsmString();
10446 SmallVector<StringRef, 4> AsmPieces;
10447 SplitString(AsmStr, AsmPieces, ";\n");
10449 switch (AsmPieces.size()) {
10450 default: return false;
10452 AsmStr = AsmPieces[0];
10454 SplitString(AsmStr, AsmPieces, " \t,");
10457 if (AsmPieces.size() == 3 &&
10458 AsmPieces[0] == "rev" && AsmPieces[1] == "$0" && AsmPieces[2] == "$1" &&
10459 IA->getConstraintString().compare(0, 4, "=l,l") == 0) {
10460 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
10461 if (Ty && Ty->getBitWidth() == 32)
10462 return IntrinsicLowering::LowerToByteSwap(CI);
10470 /// getConstraintType - Given a constraint letter, return the type of
10471 /// constraint it is for this target.
10472 ARMTargetLowering::ConstraintType
10473 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
10474 if (Constraint.size() == 1) {
10475 switch (Constraint[0]) {
10477 case 'l': return C_RegisterClass;
10478 case 'w': return C_RegisterClass;
10479 case 'h': return C_RegisterClass;
10480 case 'x': return C_RegisterClass;
10481 case 't': return C_RegisterClass;
10482 case 'j': return C_Other; // Constant for movw.
10483 // An address with a single base register. Due to the way we
10484 // currently handle addresses it is the same as an 'r' memory constraint.
10485 case 'Q': return C_Memory;
10487 } else if (Constraint.size() == 2) {
10488 switch (Constraint[0]) {
10490 // All 'U+' constraints are addresses.
10491 case 'U': return C_Memory;
10494 return TargetLowering::getConstraintType(Constraint);
10497 /// Examine constraint type and operand type and determine a weight value.
10498 /// This object must already have been set up with the operand type
10499 /// and the current alternative constraint selected.
10500 TargetLowering::ConstraintWeight
10501 ARMTargetLowering::getSingleConstraintMatchWeight(
10502 AsmOperandInfo &info, const char *constraint) const {
10503 ConstraintWeight weight = CW_Invalid;
10504 Value *CallOperandVal = info.CallOperandVal;
10505 // If we don't have a value, we can't do a match,
10506 // but allow it at the lowest weight.
10507 if (!CallOperandVal)
10509 Type *type = CallOperandVal->getType();
10510 // Look at the constraint type.
10511 switch (*constraint) {
10513 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
10516 if (type->isIntegerTy()) {
10517 if (Subtarget->isThumb())
10518 weight = CW_SpecificReg;
10520 weight = CW_Register;
10524 if (type->isFloatingPointTy())
10525 weight = CW_Register;
10531 typedef std::pair<unsigned, const TargetRegisterClass*> RCPair;
10533 ARMTargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
10534 const std::string &Constraint,
10536 if (Constraint.size() == 1) {
10537 // GCC ARM Constraint Letters
10538 switch (Constraint[0]) {
10539 case 'l': // Low regs or general regs.
10540 if (Subtarget->isThumb())
10541 return RCPair(0U, &ARM::tGPRRegClass);
10542 return RCPair(0U, &ARM::GPRRegClass);
10543 case 'h': // High regs or no regs.
10544 if (Subtarget->isThumb())
10545 return RCPair(0U, &ARM::hGPRRegClass);
10548 if (Subtarget->isThumb1Only())
10549 return RCPair(0U, &ARM::tGPRRegClass);
10550 return RCPair(0U, &ARM::GPRRegClass);
10552 if (VT == MVT::Other)
10554 if (VT == MVT::f32)
10555 return RCPair(0U, &ARM::SPRRegClass);
10556 if (VT.getSizeInBits() == 64)
10557 return RCPair(0U, &ARM::DPRRegClass);
10558 if (VT.getSizeInBits() == 128)
10559 return RCPair(0U, &ARM::QPRRegClass);
10562 if (VT == MVT::Other)
10564 if (VT == MVT::f32)
10565 return RCPair(0U, &ARM::SPR_8RegClass);
10566 if (VT.getSizeInBits() == 64)
10567 return RCPair(0U, &ARM::DPR_8RegClass);
10568 if (VT.getSizeInBits() == 128)
10569 return RCPair(0U, &ARM::QPR_8RegClass);
10572 if (VT == MVT::f32)
10573 return RCPair(0U, &ARM::SPRRegClass);
10577 if (StringRef("{cc}").equals_lower(Constraint))
10578 return std::make_pair(unsigned(ARM::CPSR), &ARM::CCRRegClass);
10580 return TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
10583 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
10584 /// vector. If it is invalid, don't add anything to Ops.
10585 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
10586 std::string &Constraint,
10587 std::vector<SDValue>&Ops,
10588 SelectionDAG &DAG) const {
10591 // Currently only support length 1 constraints.
10592 if (Constraint.length() != 1) return;
10594 char ConstraintLetter = Constraint[0];
10595 switch (ConstraintLetter) {
10598 case 'I': case 'J': case 'K': case 'L':
10599 case 'M': case 'N': case 'O':
10600 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
10604 int64_t CVal64 = C->getSExtValue();
10605 int CVal = (int) CVal64;
10606 // None of these constraints allow values larger than 32 bits. Check
10607 // that the value fits in an int.
10608 if (CVal != CVal64)
10611 switch (ConstraintLetter) {
10613 // Constant suitable for movw, must be between 0 and
10615 if (Subtarget->hasV6T2Ops())
10616 if (CVal >= 0 && CVal <= 65535)
10620 if (Subtarget->isThumb1Only()) {
10621 // This must be a constant between 0 and 255, for ADD
10623 if (CVal >= 0 && CVal <= 255)
10625 } else if (Subtarget->isThumb2()) {
10626 // A constant that can be used as an immediate value in a
10627 // data-processing instruction.
10628 if (ARM_AM::getT2SOImmVal(CVal) != -1)
10631 // A constant that can be used as an immediate value in a
10632 // data-processing instruction.
10633 if (ARM_AM::getSOImmVal(CVal) != -1)
10639 if (Subtarget->isThumb()) { // FIXME thumb2
10640 // This must be a constant between -255 and -1, for negated ADD
10641 // immediates. This can be used in GCC with an "n" modifier that
10642 // prints the negated value, for use with SUB instructions. It is
10643 // not useful otherwise but is implemented for compatibility.
10644 if (CVal >= -255 && CVal <= -1)
10647 // This must be a constant between -4095 and 4095. It is not clear
10648 // what this constraint is intended for. Implemented for
10649 // compatibility with GCC.
10650 if (CVal >= -4095 && CVal <= 4095)
10656 if (Subtarget->isThumb1Only()) {
10657 // A 32-bit value where only one byte has a nonzero value. Exclude
10658 // zero to match GCC. This constraint is used by GCC internally for
10659 // constants that can be loaded with a move/shift combination.
10660 // It is not useful otherwise but is implemented for compatibility.
10661 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
10663 } else if (Subtarget->isThumb2()) {
10664 // A constant whose bitwise inverse can be used as an immediate
10665 // value in a data-processing instruction. This can be used in GCC
10666 // with a "B" modifier that prints the inverted value, for use with
10667 // BIC and MVN instructions. It is not useful otherwise but is
10668 // implemented for compatibility.
10669 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
10672 // A constant whose bitwise inverse can be used as an immediate
10673 // value in a data-processing instruction. This can be used in GCC
10674 // with a "B" modifier that prints the inverted value, for use with
10675 // BIC and MVN instructions. It is not useful otherwise but is
10676 // implemented for compatibility.
10677 if (ARM_AM::getSOImmVal(~CVal) != -1)
10683 if (Subtarget->isThumb1Only()) {
10684 // This must be a constant between -7 and 7,
10685 // for 3-operand ADD/SUB immediate instructions.
10686 if (CVal >= -7 && CVal < 7)
10688 } else if (Subtarget->isThumb2()) {
10689 // A constant whose negation can be used as an immediate value in a
10690 // data-processing instruction. This can be used in GCC with an "n"
10691 // modifier that prints the negated value, for use with SUB
10692 // instructions. It is not useful otherwise but is implemented for
10694 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
10697 // A constant whose negation can be used as an immediate value in a
10698 // data-processing instruction. This can be used in GCC with an "n"
10699 // modifier that prints the negated value, for use with SUB
10700 // instructions. It is not useful otherwise but is implemented for
10702 if (ARM_AM::getSOImmVal(-CVal) != -1)
10708 if (Subtarget->isThumb()) { // FIXME thumb2
10709 // This must be a multiple of 4 between 0 and 1020, for
10710 // ADD sp + immediate.
10711 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
10714 // A power of two or a constant between 0 and 32. This is used in
10715 // GCC for the shift amount on shifted register operands, but it is
10716 // useful in general for any shift amounts.
10717 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
10723 if (Subtarget->isThumb()) { // FIXME thumb2
10724 // This must be a constant between 0 and 31, for shift amounts.
10725 if (CVal >= 0 && CVal <= 31)
10731 if (Subtarget->isThumb()) { // FIXME thumb2
10732 // This must be a multiple of 4 between -508 and 508, for
10733 // ADD/SUB sp = sp + immediate.
10734 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
10739 Result = DAG.getTargetConstant(CVal, Op.getValueType());
10743 if (Result.getNode()) {
10744 Ops.push_back(Result);
10747 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
10750 SDValue ARMTargetLowering::LowerDivRem(SDValue Op, SelectionDAG &DAG) const {
10751 assert(Subtarget->isTargetAEABI() && "Register-based DivRem lowering only");
10752 unsigned Opcode = Op->getOpcode();
10753 assert((Opcode == ISD::SDIVREM || Opcode == ISD::UDIVREM) &&
10754 "Invalid opcode for Div/Rem lowering");
10755 bool isSigned = (Opcode == ISD::SDIVREM);
10756 EVT VT = Op->getValueType(0);
10757 Type *Ty = VT.getTypeForEVT(*DAG.getContext());
10760 switch (VT.getSimpleVT().SimpleTy) {
10761 default: llvm_unreachable("Unexpected request for libcall!");
10762 case MVT::i8: LC = isSigned ? RTLIB::SDIVREM_I8 : RTLIB::UDIVREM_I8; break;
10763 case MVT::i16: LC = isSigned ? RTLIB::SDIVREM_I16 : RTLIB::UDIVREM_I16; break;
10764 case MVT::i32: LC = isSigned ? RTLIB::SDIVREM_I32 : RTLIB::UDIVREM_I32; break;
10765 case MVT::i64: LC = isSigned ? RTLIB::SDIVREM_I64 : RTLIB::UDIVREM_I64; break;
10768 SDValue InChain = DAG.getEntryNode();
10770 TargetLowering::ArgListTy Args;
10771 TargetLowering::ArgListEntry Entry;
10772 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i) {
10773 EVT ArgVT = Op->getOperand(i).getValueType();
10774 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
10775 Entry.Node = Op->getOperand(i);
10777 Entry.isSExt = isSigned;
10778 Entry.isZExt = !isSigned;
10779 Args.push_back(Entry);
10782 SDValue Callee = DAG.getExternalSymbol(getLibcallName(LC),
10785 Type *RetTy = (Type*)StructType::get(Ty, Ty, nullptr);
10788 TargetLowering::CallLoweringInfo CLI(DAG);
10789 CLI.setDebugLoc(dl).setChain(InChain)
10790 .setCallee(getLibcallCallingConv(LC), RetTy, Callee, std::move(Args), 0)
10791 .setInRegister().setSExtResult(isSigned).setZExtResult(!isSigned);
10793 std::pair<SDValue, SDValue> CallInfo = LowerCallTo(CLI);
10794 return CallInfo.first;
10798 ARMTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const {
10799 assert(Subtarget->isTargetWindows() && "unsupported target platform");
10803 SDValue Chain = Op.getOperand(0);
10804 SDValue Size = Op.getOperand(1);
10806 SDValue Words = DAG.getNode(ISD::SRL, DL, MVT::i32, Size,
10807 DAG.getConstant(2, MVT::i32));
10810 Chain = DAG.getCopyToReg(Chain, DL, ARM::R4, Words, Flag);
10811 Flag = Chain.getValue(1);
10813 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
10814 Chain = DAG.getNode(ARMISD::WIN__CHKSTK, DL, NodeTys, Chain, Flag);
10816 SDValue NewSP = DAG.getCopyFromReg(Chain, DL, ARM::SP, MVT::i32);
10817 Chain = NewSP.getValue(1);
10819 SDValue Ops[2] = { NewSP, Chain };
10820 return DAG.getMergeValues(Ops, DL);
10823 SDValue ARMTargetLowering::LowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) const {
10824 assert(Op.getValueType() == MVT::f64 && Subtarget->isFPOnlySP() &&
10825 "Unexpected type for custom-lowering FP_EXTEND");
10828 LC = RTLIB::getFPEXT(Op.getOperand(0).getValueType(), Op.getValueType());
10830 SDValue SrcVal = Op.getOperand(0);
10831 return makeLibCall(DAG, LC, Op.getValueType(), &SrcVal, 1,
10832 /*isSigned*/ false, SDLoc(Op)).first;
10835 SDValue ARMTargetLowering::LowerFP_ROUND(SDValue Op, SelectionDAG &DAG) const {
10836 assert(Op.getOperand(0).getValueType() == MVT::f64 &&
10837 Subtarget->isFPOnlySP() &&
10838 "Unexpected type for custom-lowering FP_ROUND");
10841 LC = RTLIB::getFPROUND(Op.getOperand(0).getValueType(), Op.getValueType());
10843 SDValue SrcVal = Op.getOperand(0);
10844 return makeLibCall(DAG, LC, Op.getValueType(), &SrcVal, 1,
10845 /*isSigned*/ false, SDLoc(Op)).first;
10849 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
10850 // The ARM target isn't yet aware of offsets.
10854 bool ARM::isBitFieldInvertedMask(unsigned v) {
10855 if (v == 0xffffffff)
10858 // there can be 1's on either or both "outsides", all the "inside"
10859 // bits must be 0's
10860 return isShiftedMask_32(~v);
10863 /// isFPImmLegal - Returns true if the target can instruction select the
10864 /// specified FP immediate natively. If false, the legalizer will
10865 /// materialize the FP immediate as a load from a constant pool.
10866 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
10867 if (!Subtarget->hasVFP3())
10869 if (VT == MVT::f32)
10870 return ARM_AM::getFP32Imm(Imm) != -1;
10871 if (VT == MVT::f64 && !Subtarget->isFPOnlySP())
10872 return ARM_AM::getFP64Imm(Imm) != -1;
10876 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
10877 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
10878 /// specified in the intrinsic calls.
10879 bool ARMTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
10881 unsigned Intrinsic) const {
10882 switch (Intrinsic) {
10883 case Intrinsic::arm_neon_vld1:
10884 case Intrinsic::arm_neon_vld2:
10885 case Intrinsic::arm_neon_vld3:
10886 case Intrinsic::arm_neon_vld4:
10887 case Intrinsic::arm_neon_vld2lane:
10888 case Intrinsic::arm_neon_vld3lane:
10889 case Intrinsic::arm_neon_vld4lane: {
10890 Info.opc = ISD::INTRINSIC_W_CHAIN;
10891 // Conservatively set memVT to the entire set of vectors loaded.
10892 uint64_t NumElts = getDataLayout()->getTypeAllocSize(I.getType()) / 8;
10893 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
10894 Info.ptrVal = I.getArgOperand(0);
10896 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
10897 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
10898 Info.vol = false; // volatile loads with NEON intrinsics not supported
10899 Info.readMem = true;
10900 Info.writeMem = false;
10903 case Intrinsic::arm_neon_vst1:
10904 case Intrinsic::arm_neon_vst2:
10905 case Intrinsic::arm_neon_vst3:
10906 case Intrinsic::arm_neon_vst4:
10907 case Intrinsic::arm_neon_vst2lane:
10908 case Intrinsic::arm_neon_vst3lane:
10909 case Intrinsic::arm_neon_vst4lane: {
10910 Info.opc = ISD::INTRINSIC_VOID;
10911 // Conservatively set memVT to the entire set of vectors stored.
10912 unsigned NumElts = 0;
10913 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
10914 Type *ArgTy = I.getArgOperand(ArgI)->getType();
10915 if (!ArgTy->isVectorTy())
10917 NumElts += getDataLayout()->getTypeAllocSize(ArgTy) / 8;
10919 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
10920 Info.ptrVal = I.getArgOperand(0);
10922 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
10923 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
10924 Info.vol = false; // volatile stores with NEON intrinsics not supported
10925 Info.readMem = false;
10926 Info.writeMem = true;
10929 case Intrinsic::arm_ldaex:
10930 case Intrinsic::arm_ldrex: {
10931 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(0)->getType());
10932 Info.opc = ISD::INTRINSIC_W_CHAIN;
10933 Info.memVT = MVT::getVT(PtrTy->getElementType());
10934 Info.ptrVal = I.getArgOperand(0);
10936 Info.align = getDataLayout()->getABITypeAlignment(PtrTy->getElementType());
10938 Info.readMem = true;
10939 Info.writeMem = false;
10942 case Intrinsic::arm_stlex:
10943 case Intrinsic::arm_strex: {
10944 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(1)->getType());
10945 Info.opc = ISD::INTRINSIC_W_CHAIN;
10946 Info.memVT = MVT::getVT(PtrTy->getElementType());
10947 Info.ptrVal = I.getArgOperand(1);
10949 Info.align = getDataLayout()->getABITypeAlignment(PtrTy->getElementType());
10951 Info.readMem = false;
10952 Info.writeMem = true;
10955 case Intrinsic::arm_stlexd:
10956 case Intrinsic::arm_strexd: {
10957 Info.opc = ISD::INTRINSIC_W_CHAIN;
10958 Info.memVT = MVT::i64;
10959 Info.ptrVal = I.getArgOperand(2);
10963 Info.readMem = false;
10964 Info.writeMem = true;
10967 case Intrinsic::arm_ldaexd:
10968 case Intrinsic::arm_ldrexd: {
10969 Info.opc = ISD::INTRINSIC_W_CHAIN;
10970 Info.memVT = MVT::i64;
10971 Info.ptrVal = I.getArgOperand(0);
10975 Info.readMem = true;
10976 Info.writeMem = false;
10986 /// \brief Returns true if it is beneficial to convert a load of a constant
10987 /// to just the constant itself.
10988 bool ARMTargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
10990 assert(Ty->isIntegerTy());
10992 unsigned Bits = Ty->getPrimitiveSizeInBits();
10993 if (Bits == 0 || Bits > 32)
10998 bool ARMTargetLowering::hasLoadLinkedStoreConditional() const { return true; }
11000 Instruction* ARMTargetLowering::makeDMB(IRBuilder<> &Builder,
11001 ARM_MB::MemBOpt Domain) const {
11002 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
11004 // First, if the target has no DMB, see what fallback we can use.
11005 if (!Subtarget->hasDataBarrier()) {
11006 // Some ARMv6 cpus can support data barriers with an mcr instruction.
11007 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
11009 if (Subtarget->hasV6Ops() && !Subtarget->isThumb()) {
11010 Function *MCR = llvm::Intrinsic::getDeclaration(M, Intrinsic::arm_mcr);
11011 Value* args[6] = {Builder.getInt32(15), Builder.getInt32(0),
11012 Builder.getInt32(0), Builder.getInt32(7),
11013 Builder.getInt32(10), Builder.getInt32(5)};
11014 return Builder.CreateCall(MCR, args);
11016 // Instead of using barriers, atomic accesses on these subtargets use
11018 llvm_unreachable("makeDMB on a target so old that it has no barriers");
11021 Function *DMB = llvm::Intrinsic::getDeclaration(M, Intrinsic::arm_dmb);
11022 // Only a full system barrier exists in the M-class architectures.
11023 Domain = Subtarget->isMClass() ? ARM_MB::SY : Domain;
11024 Constant *CDomain = Builder.getInt32(Domain);
11025 return Builder.CreateCall(DMB, CDomain);
11029 // Based on http://www.cl.cam.ac.uk/~pes20/cpp/cpp0xmappings.html
11030 Instruction* ARMTargetLowering::emitLeadingFence(IRBuilder<> &Builder,
11031 AtomicOrdering Ord, bool IsStore,
11032 bool IsLoad) const {
11033 if (!getInsertFencesForAtomic())
11039 llvm_unreachable("Invalid fence: unordered/non-atomic");
11042 return nullptr; // Nothing to do
11043 case SequentiallyConsistent:
11045 return nullptr; // Nothing to do
11048 case AcquireRelease:
11049 if (Subtarget->isSwift())
11050 return makeDMB(Builder, ARM_MB::ISHST);
11051 // FIXME: add a comment with a link to documentation justifying this.
11053 return makeDMB(Builder, ARM_MB::ISH);
11055 llvm_unreachable("Unknown fence ordering in emitLeadingFence");
11058 Instruction* ARMTargetLowering::emitTrailingFence(IRBuilder<> &Builder,
11059 AtomicOrdering Ord, bool IsStore,
11060 bool IsLoad) const {
11061 if (!getInsertFencesForAtomic())
11067 llvm_unreachable("Invalid fence: unordered/not-atomic");
11070 return nullptr; // Nothing to do
11072 case AcquireRelease:
11073 case SequentiallyConsistent:
11074 return makeDMB(Builder, ARM_MB::ISH);
11076 llvm_unreachable("Unknown fence ordering in emitTrailingFence");
11079 // Loads and stores less than 64-bits are already atomic; ones above that
11080 // are doomed anyway, so defer to the default libcall and blame the OS when
11081 // things go wrong. Cortex M doesn't have ldrexd/strexd though, so don't emit
11082 // anything for those.
11083 bool ARMTargetLowering::shouldExpandAtomicStoreInIR(StoreInst *SI) const {
11084 unsigned Size = SI->getValueOperand()->getType()->getPrimitiveSizeInBits();
11085 return (Size == 64) && !Subtarget->isMClass();
11088 // Loads and stores less than 64-bits are already atomic; ones above that
11089 // are doomed anyway, so defer to the default libcall and blame the OS when
11090 // things go wrong. Cortex M doesn't have ldrexd/strexd though, so don't emit
11091 // anything for those.
11092 // FIXME: ldrd and strd are atomic if the CPU has LPAE (e.g. A15 has that
11093 // guarantee, see DDI0406C ARM architecture reference manual,
11094 // sections A8.8.72-74 LDRD)
11095 bool ARMTargetLowering::shouldExpandAtomicLoadInIR(LoadInst *LI) const {
11096 unsigned Size = LI->getType()->getPrimitiveSizeInBits();
11097 return (Size == 64) && !Subtarget->isMClass();
11100 // For the real atomic operations, we have ldrex/strex up to 32 bits,
11101 // and up to 64 bits on the non-M profiles
11102 TargetLoweringBase::AtomicRMWExpansionKind
11103 ARMTargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
11104 unsigned Size = AI->getType()->getPrimitiveSizeInBits();
11105 return (Size <= (Subtarget->isMClass() ? 32U : 64U))
11106 ? AtomicRMWExpansionKind::LLSC
11107 : AtomicRMWExpansionKind::None;
11110 // This has so far only been implemented for MachO.
11111 bool ARMTargetLowering::useLoadStackGuardNode() const {
11112 return Subtarget->isTargetMachO();
11115 bool ARMTargetLowering::canCombineStoreAndExtract(Type *VectorTy, Value *Idx,
11116 unsigned &Cost) const {
11117 // If we do not have NEON, vector types are not natively supported.
11118 if (!Subtarget->hasNEON())
11121 // Floating point values and vector values map to the same register file.
11122 // Therefore, althought we could do a store extract of a vector type, this is
11123 // better to leave at float as we have more freedom in the addressing mode for
11125 if (VectorTy->isFPOrFPVectorTy())
11128 // If the index is unknown at compile time, this is very expensive to lower
11129 // and it is not possible to combine the store with the extract.
11130 if (!isa<ConstantInt>(Idx))
11133 assert(VectorTy->isVectorTy() && "VectorTy is not a vector type");
11134 unsigned BitWidth = cast<VectorType>(VectorTy)->getBitWidth();
11135 // We can do a store + vector extract on any vector that fits perfectly in a D
11137 if (BitWidth == 64 || BitWidth == 128) {
11144 Value *ARMTargetLowering::emitLoadLinked(IRBuilder<> &Builder, Value *Addr,
11145 AtomicOrdering Ord) const {
11146 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
11147 Type *ValTy = cast<PointerType>(Addr->getType())->getElementType();
11148 bool IsAcquire = isAtLeastAcquire(Ord);
11150 // Since i64 isn't legal and intrinsics don't get type-lowered, the ldrexd
11151 // intrinsic must return {i32, i32} and we have to recombine them into a
11152 // single i64 here.
11153 if (ValTy->getPrimitiveSizeInBits() == 64) {
11154 Intrinsic::ID Int =
11155 IsAcquire ? Intrinsic::arm_ldaexd : Intrinsic::arm_ldrexd;
11156 Function *Ldrex = llvm::Intrinsic::getDeclaration(M, Int);
11158 Addr = Builder.CreateBitCast(Addr, Type::getInt8PtrTy(M->getContext()));
11159 Value *LoHi = Builder.CreateCall(Ldrex, Addr, "lohi");
11161 Value *Lo = Builder.CreateExtractValue(LoHi, 0, "lo");
11162 Value *Hi = Builder.CreateExtractValue(LoHi, 1, "hi");
11163 if (!Subtarget->isLittle())
11164 std::swap (Lo, Hi);
11165 Lo = Builder.CreateZExt(Lo, ValTy, "lo64");
11166 Hi = Builder.CreateZExt(Hi, ValTy, "hi64");
11167 return Builder.CreateOr(
11168 Lo, Builder.CreateShl(Hi, ConstantInt::get(ValTy, 32)), "val64");
11171 Type *Tys[] = { Addr->getType() };
11172 Intrinsic::ID Int = IsAcquire ? Intrinsic::arm_ldaex : Intrinsic::arm_ldrex;
11173 Function *Ldrex = llvm::Intrinsic::getDeclaration(M, Int, Tys);
11175 return Builder.CreateTruncOrBitCast(
11176 Builder.CreateCall(Ldrex, Addr),
11177 cast<PointerType>(Addr->getType())->getElementType());
11180 Value *ARMTargetLowering::emitStoreConditional(IRBuilder<> &Builder, Value *Val,
11182 AtomicOrdering Ord) const {
11183 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
11184 bool IsRelease = isAtLeastRelease(Ord);
11186 // Since the intrinsics must have legal type, the i64 intrinsics take two
11187 // parameters: "i32, i32". We must marshal Val into the appropriate form
11188 // before the call.
11189 if (Val->getType()->getPrimitiveSizeInBits() == 64) {
11190 Intrinsic::ID Int =
11191 IsRelease ? Intrinsic::arm_stlexd : Intrinsic::arm_strexd;
11192 Function *Strex = Intrinsic::getDeclaration(M, Int);
11193 Type *Int32Ty = Type::getInt32Ty(M->getContext());
11195 Value *Lo = Builder.CreateTrunc(Val, Int32Ty, "lo");
11196 Value *Hi = Builder.CreateTrunc(Builder.CreateLShr(Val, 32), Int32Ty, "hi");
11197 if (!Subtarget->isLittle())
11198 std::swap (Lo, Hi);
11199 Addr = Builder.CreateBitCast(Addr, Type::getInt8PtrTy(M->getContext()));
11200 return Builder.CreateCall3(Strex, Lo, Hi, Addr);
11203 Intrinsic::ID Int = IsRelease ? Intrinsic::arm_stlex : Intrinsic::arm_strex;
11204 Type *Tys[] = { Addr->getType() };
11205 Function *Strex = Intrinsic::getDeclaration(M, Int, Tys);
11207 return Builder.CreateCall2(
11208 Strex, Builder.CreateZExtOrBitCast(
11209 Val, Strex->getFunctionType()->getParamType(0)),
11221 static bool isHomogeneousAggregate(Type *Ty, HABaseType &Base,
11222 uint64_t &Members) {
11223 if (const StructType *ST = dyn_cast<StructType>(Ty)) {
11224 for (unsigned i = 0; i < ST->getNumElements(); ++i) {
11225 uint64_t SubMembers = 0;
11226 if (!isHomogeneousAggregate(ST->getElementType(i), Base, SubMembers))
11228 Members += SubMembers;
11230 } else if (const ArrayType *AT = dyn_cast<ArrayType>(Ty)) {
11231 uint64_t SubMembers = 0;
11232 if (!isHomogeneousAggregate(AT->getElementType(), Base, SubMembers))
11234 Members += SubMembers * AT->getNumElements();
11235 } else if (Ty->isFloatTy()) {
11236 if (Base != HA_UNKNOWN && Base != HA_FLOAT)
11240 } else if (Ty->isDoubleTy()) {
11241 if (Base != HA_UNKNOWN && Base != HA_DOUBLE)
11245 } else if (const VectorType *VT = dyn_cast<VectorType>(Ty)) {
11252 return VT->getBitWidth() == 64;
11254 return VT->getBitWidth() == 128;
11256 switch (VT->getBitWidth()) {
11269 return (Members > 0 && Members <= 4);
11272 /// \brief Return true if a type is an AAPCS-VFP homogeneous aggregate or one of
11273 /// [N x i32] or [N x i64]. This allows front-ends to skip emitting padding when
11274 /// passing according to AAPCS rules.
11275 bool ARMTargetLowering::functionArgumentNeedsConsecutiveRegisters(
11276 Type *Ty, CallingConv::ID CallConv, bool isVarArg) const {
11277 if (getEffectiveCallingConv(CallConv, isVarArg) !=
11278 CallingConv::ARM_AAPCS_VFP)
11281 HABaseType Base = HA_UNKNOWN;
11282 uint64_t Members = 0;
11283 bool IsHA = isHomogeneousAggregate(Ty, Base, Members);
11284 DEBUG(dbgs() << "isHA: " << IsHA << " "; Ty->dump());
11286 bool IsIntArray = Ty->isArrayTy() && Ty->getArrayElementType()->isIntegerTy();
11287 return IsHA || IsIntArray;