1 //===-- ARMISelLowering.cpp - ARM DAG Lowering Implementation -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the interfaces that ARM uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
15 #include "ARMISelLowering.h"
16 #include "ARMCallingConv.h"
17 #include "ARMConstantPoolValue.h"
18 #include "ARMMachineFunctionInfo.h"
19 #include "ARMPerfectShuffle.h"
20 #include "ARMSubtarget.h"
21 #include "ARMTargetMachine.h"
22 #include "ARMTargetObjectFile.h"
23 #include "MCTargetDesc/ARMAddressingModes.h"
24 #include "llvm/ADT/Statistic.h"
25 #include "llvm/ADT/StringExtras.h"
26 #include "llvm/CodeGen/CallingConvLower.h"
27 #include "llvm/CodeGen/IntrinsicLowering.h"
28 #include "llvm/CodeGen/MachineBasicBlock.h"
29 #include "llvm/CodeGen/MachineFrameInfo.h"
30 #include "llvm/CodeGen/MachineFunction.h"
31 #include "llvm/CodeGen/MachineInstrBuilder.h"
32 #include "llvm/CodeGen/MachineModuleInfo.h"
33 #include "llvm/CodeGen/MachineRegisterInfo.h"
34 #include "llvm/CodeGen/SelectionDAG.h"
35 #include "llvm/IR/CallingConv.h"
36 #include "llvm/IR/Constants.h"
37 #include "llvm/IR/Function.h"
38 #include "llvm/IR/GlobalValue.h"
39 #include "llvm/IR/IRBuilder.h"
40 #include "llvm/IR/Instruction.h"
41 #include "llvm/IR/Instructions.h"
42 #include "llvm/IR/Intrinsics.h"
43 #include "llvm/IR/Type.h"
44 #include "llvm/MC/MCSectionMachO.h"
45 #include "llvm/Support/CommandLine.h"
46 #include "llvm/Support/Debug.h"
47 #include "llvm/Support/ErrorHandling.h"
48 #include "llvm/Support/MathExtras.h"
49 #include "llvm/Target/TargetOptions.h"
53 #define DEBUG_TYPE "arm-isel"
55 STATISTIC(NumTailCalls, "Number of tail calls");
56 STATISTIC(NumMovwMovt, "Number of GAs materialized with movw + movt");
57 STATISTIC(NumLoopByVals, "Number of loops generated for byval arguments");
60 EnableARMLongCalls("arm-long-calls", cl::Hidden,
61 cl::desc("Generate calls via indirect call instructions"),
65 ARMInterworking("arm-interworking", cl::Hidden,
66 cl::desc("Enable / disable ARM interworking (for debugging only)"),
70 class ARMCCState : public CCState {
72 ARMCCState(CallingConv::ID CC, bool isVarArg, MachineFunction &MF,
73 const TargetMachine &TM, SmallVectorImpl<CCValAssign> &locs,
74 LLVMContext &C, ParmContext PC)
75 : CCState(CC, isVarArg, MF, TM, locs, C) {
76 assert(((PC == Call) || (PC == Prologue)) &&
77 "ARMCCState users must specify whether their context is call"
78 "or prologue generation.");
84 // The APCS parameter registers.
85 static const MCPhysReg GPRArgRegs[] = {
86 ARM::R0, ARM::R1, ARM::R2, ARM::R3
89 void ARMTargetLowering::addTypeForNEON(MVT VT, MVT PromotedLdStVT,
90 MVT PromotedBitwiseVT) {
91 if (VT != PromotedLdStVT) {
92 setOperationAction(ISD::LOAD, VT, Promote);
93 AddPromotedToType (ISD::LOAD, VT, PromotedLdStVT);
95 setOperationAction(ISD::STORE, VT, Promote);
96 AddPromotedToType (ISD::STORE, VT, PromotedLdStVT);
99 MVT ElemTy = VT.getVectorElementType();
100 if (ElemTy != MVT::i64 && ElemTy != MVT::f64)
101 setOperationAction(ISD::SETCC, VT, Custom);
102 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
103 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
104 if (ElemTy == MVT::i32) {
105 setOperationAction(ISD::SINT_TO_FP, VT, Custom);
106 setOperationAction(ISD::UINT_TO_FP, VT, Custom);
107 setOperationAction(ISD::FP_TO_SINT, VT, Custom);
108 setOperationAction(ISD::FP_TO_UINT, VT, Custom);
110 setOperationAction(ISD::SINT_TO_FP, VT, Expand);
111 setOperationAction(ISD::UINT_TO_FP, VT, Expand);
112 setOperationAction(ISD::FP_TO_SINT, VT, Expand);
113 setOperationAction(ISD::FP_TO_UINT, VT, Expand);
115 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
116 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
117 setOperationAction(ISD::CONCAT_VECTORS, VT, Legal);
118 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Legal);
119 setOperationAction(ISD::SELECT, VT, Expand);
120 setOperationAction(ISD::SELECT_CC, VT, Expand);
121 setOperationAction(ISD::VSELECT, VT, Expand);
122 setOperationAction(ISD::SIGN_EXTEND_INREG, VT, Expand);
123 if (VT.isInteger()) {
124 setOperationAction(ISD::SHL, VT, Custom);
125 setOperationAction(ISD::SRA, VT, Custom);
126 setOperationAction(ISD::SRL, VT, Custom);
129 // Promote all bit-wise operations.
130 if (VT.isInteger() && VT != PromotedBitwiseVT) {
131 setOperationAction(ISD::AND, VT, Promote);
132 AddPromotedToType (ISD::AND, VT, PromotedBitwiseVT);
133 setOperationAction(ISD::OR, VT, Promote);
134 AddPromotedToType (ISD::OR, VT, PromotedBitwiseVT);
135 setOperationAction(ISD::XOR, VT, Promote);
136 AddPromotedToType (ISD::XOR, VT, PromotedBitwiseVT);
139 // Neon does not support vector divide/remainder operations.
140 setOperationAction(ISD::SDIV, VT, Expand);
141 setOperationAction(ISD::UDIV, VT, Expand);
142 setOperationAction(ISD::FDIV, VT, Expand);
143 setOperationAction(ISD::SREM, VT, Expand);
144 setOperationAction(ISD::UREM, VT, Expand);
145 setOperationAction(ISD::FREM, VT, Expand);
148 void ARMTargetLowering::addDRTypeForNEON(MVT VT) {
149 addRegisterClass(VT, &ARM::DPRRegClass);
150 addTypeForNEON(VT, MVT::f64, MVT::v2i32);
153 void ARMTargetLowering::addQRTypeForNEON(MVT VT) {
154 addRegisterClass(VT, &ARM::DPairRegClass);
155 addTypeForNEON(VT, MVT::v2f64, MVT::v4i32);
158 static TargetLoweringObjectFile *createTLOF(const Triple &TT) {
159 if (TT.isOSBinFormatMachO())
160 return new TargetLoweringObjectFileMachO();
161 if (TT.isOSWindows())
162 return new TargetLoweringObjectFileCOFF();
163 return new ARMElfTargetObjectFile();
166 ARMTargetLowering::ARMTargetLowering(TargetMachine &TM)
167 : TargetLowering(TM, createTLOF(Triple(TM.getTargetTriple()))) {
168 Subtarget = &TM.getSubtarget<ARMSubtarget>();
169 RegInfo = TM.getRegisterInfo();
170 Itins = TM.getInstrItineraryData();
172 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
174 if (Subtarget->isTargetMachO()) {
175 // Uses VFP for Thumb libfuncs if available.
176 if (Subtarget->isThumb() && Subtarget->hasVFP2() &&
177 Subtarget->hasARMOps() && !TM.Options.UseSoftFloat) {
178 // Single-precision floating-point arithmetic.
179 setLibcallName(RTLIB::ADD_F32, "__addsf3vfp");
180 setLibcallName(RTLIB::SUB_F32, "__subsf3vfp");
181 setLibcallName(RTLIB::MUL_F32, "__mulsf3vfp");
182 setLibcallName(RTLIB::DIV_F32, "__divsf3vfp");
184 // Double-precision floating-point arithmetic.
185 setLibcallName(RTLIB::ADD_F64, "__adddf3vfp");
186 setLibcallName(RTLIB::SUB_F64, "__subdf3vfp");
187 setLibcallName(RTLIB::MUL_F64, "__muldf3vfp");
188 setLibcallName(RTLIB::DIV_F64, "__divdf3vfp");
190 // Single-precision comparisons.
191 setLibcallName(RTLIB::OEQ_F32, "__eqsf2vfp");
192 setLibcallName(RTLIB::UNE_F32, "__nesf2vfp");
193 setLibcallName(RTLIB::OLT_F32, "__ltsf2vfp");
194 setLibcallName(RTLIB::OLE_F32, "__lesf2vfp");
195 setLibcallName(RTLIB::OGE_F32, "__gesf2vfp");
196 setLibcallName(RTLIB::OGT_F32, "__gtsf2vfp");
197 setLibcallName(RTLIB::UO_F32, "__unordsf2vfp");
198 setLibcallName(RTLIB::O_F32, "__unordsf2vfp");
200 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
201 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETNE);
202 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
203 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
204 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
205 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
206 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
207 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
209 // Double-precision comparisons.
210 setLibcallName(RTLIB::OEQ_F64, "__eqdf2vfp");
211 setLibcallName(RTLIB::UNE_F64, "__nedf2vfp");
212 setLibcallName(RTLIB::OLT_F64, "__ltdf2vfp");
213 setLibcallName(RTLIB::OLE_F64, "__ledf2vfp");
214 setLibcallName(RTLIB::OGE_F64, "__gedf2vfp");
215 setLibcallName(RTLIB::OGT_F64, "__gtdf2vfp");
216 setLibcallName(RTLIB::UO_F64, "__unorddf2vfp");
217 setLibcallName(RTLIB::O_F64, "__unorddf2vfp");
219 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
220 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETNE);
221 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
222 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
223 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
224 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
225 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
226 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
228 // Floating-point to integer conversions.
229 // i64 conversions are done via library routines even when generating VFP
230 // instructions, so use the same ones.
231 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__fixdfsivfp");
232 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__fixunsdfsivfp");
233 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__fixsfsivfp");
234 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__fixunssfsivfp");
236 // Conversions between floating types.
237 setLibcallName(RTLIB::FPROUND_F64_F32, "__truncdfsf2vfp");
238 setLibcallName(RTLIB::FPEXT_F32_F64, "__extendsfdf2vfp");
240 // Integer to floating-point conversions.
241 // i64 conversions are done via library routines even when generating VFP
242 // instructions, so use the same ones.
243 // FIXME: There appears to be some naming inconsistency in ARM libgcc:
244 // e.g., __floatunsidf vs. __floatunssidfvfp.
245 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__floatsidfvfp");
246 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__floatunssidfvfp");
247 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__floatsisfvfp");
248 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__floatunssisfvfp");
252 // These libcalls are not available in 32-bit.
253 setLibcallName(RTLIB::SHL_I128, nullptr);
254 setLibcallName(RTLIB::SRL_I128, nullptr);
255 setLibcallName(RTLIB::SRA_I128, nullptr);
257 if (Subtarget->isAAPCS_ABI() && !Subtarget->isTargetMachO() &&
258 !Subtarget->isTargetWindows()) {
259 static const struct {
260 const RTLIB::Libcall Op;
261 const char * const Name;
262 const CallingConv::ID CC;
263 const ISD::CondCode Cond;
265 // Double-precision floating-point arithmetic helper functions
266 // RTABI chapter 4.1.2, Table 2
267 { RTLIB::ADD_F64, "__aeabi_dadd", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
268 { RTLIB::DIV_F64, "__aeabi_ddiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
269 { RTLIB::MUL_F64, "__aeabi_dmul", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
270 { RTLIB::SUB_F64, "__aeabi_dsub", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
272 // Double-precision floating-point comparison helper functions
273 // RTABI chapter 4.1.2, Table 3
274 { RTLIB::OEQ_F64, "__aeabi_dcmpeq", CallingConv::ARM_AAPCS, ISD::SETNE },
275 { RTLIB::UNE_F64, "__aeabi_dcmpeq", CallingConv::ARM_AAPCS, ISD::SETEQ },
276 { RTLIB::OLT_F64, "__aeabi_dcmplt", CallingConv::ARM_AAPCS, ISD::SETNE },
277 { RTLIB::OLE_F64, "__aeabi_dcmple", CallingConv::ARM_AAPCS, ISD::SETNE },
278 { RTLIB::OGE_F64, "__aeabi_dcmpge", CallingConv::ARM_AAPCS, ISD::SETNE },
279 { RTLIB::OGT_F64, "__aeabi_dcmpgt", CallingConv::ARM_AAPCS, ISD::SETNE },
280 { RTLIB::UO_F64, "__aeabi_dcmpun", CallingConv::ARM_AAPCS, ISD::SETNE },
281 { RTLIB::O_F64, "__aeabi_dcmpun", CallingConv::ARM_AAPCS, ISD::SETEQ },
283 // Single-precision floating-point arithmetic helper functions
284 // RTABI chapter 4.1.2, Table 4
285 { RTLIB::ADD_F32, "__aeabi_fadd", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
286 { RTLIB::DIV_F32, "__aeabi_fdiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
287 { RTLIB::MUL_F32, "__aeabi_fmul", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
288 { RTLIB::SUB_F32, "__aeabi_fsub", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
290 // Single-precision floating-point comparison helper functions
291 // RTABI chapter 4.1.2, Table 5
292 { RTLIB::OEQ_F32, "__aeabi_fcmpeq", CallingConv::ARM_AAPCS, ISD::SETNE },
293 { RTLIB::UNE_F32, "__aeabi_fcmpeq", CallingConv::ARM_AAPCS, ISD::SETEQ },
294 { RTLIB::OLT_F32, "__aeabi_fcmplt", CallingConv::ARM_AAPCS, ISD::SETNE },
295 { RTLIB::OLE_F32, "__aeabi_fcmple", CallingConv::ARM_AAPCS, ISD::SETNE },
296 { RTLIB::OGE_F32, "__aeabi_fcmpge", CallingConv::ARM_AAPCS, ISD::SETNE },
297 { RTLIB::OGT_F32, "__aeabi_fcmpgt", CallingConv::ARM_AAPCS, ISD::SETNE },
298 { RTLIB::UO_F32, "__aeabi_fcmpun", CallingConv::ARM_AAPCS, ISD::SETNE },
299 { RTLIB::O_F32, "__aeabi_fcmpun", CallingConv::ARM_AAPCS, ISD::SETEQ },
301 // Floating-point to integer conversions.
302 // RTABI chapter 4.1.2, Table 6
303 { RTLIB::FPTOSINT_F64_I32, "__aeabi_d2iz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
304 { RTLIB::FPTOUINT_F64_I32, "__aeabi_d2uiz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
305 { RTLIB::FPTOSINT_F64_I64, "__aeabi_d2lz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
306 { RTLIB::FPTOUINT_F64_I64, "__aeabi_d2ulz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
307 { RTLIB::FPTOSINT_F32_I32, "__aeabi_f2iz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
308 { RTLIB::FPTOUINT_F32_I32, "__aeabi_f2uiz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
309 { RTLIB::FPTOSINT_F32_I64, "__aeabi_f2lz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
310 { RTLIB::FPTOUINT_F32_I64, "__aeabi_f2ulz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
312 // Conversions between floating types.
313 // RTABI chapter 4.1.2, Table 7
314 { RTLIB::FPROUND_F64_F32, "__aeabi_d2f", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
315 { RTLIB::FPEXT_F32_F64, "__aeabi_f2d", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
317 // Integer to floating-point conversions.
318 // RTABI chapter 4.1.2, Table 8
319 { RTLIB::SINTTOFP_I32_F64, "__aeabi_i2d", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
320 { RTLIB::UINTTOFP_I32_F64, "__aeabi_ui2d", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
321 { RTLIB::SINTTOFP_I64_F64, "__aeabi_l2d", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
322 { RTLIB::UINTTOFP_I64_F64, "__aeabi_ul2d", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
323 { RTLIB::SINTTOFP_I32_F32, "__aeabi_i2f", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
324 { RTLIB::UINTTOFP_I32_F32, "__aeabi_ui2f", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
325 { RTLIB::SINTTOFP_I64_F32, "__aeabi_l2f", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
326 { RTLIB::UINTTOFP_I64_F32, "__aeabi_ul2f", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
328 // Long long helper functions
329 // RTABI chapter 4.2, Table 9
330 { RTLIB::MUL_I64, "__aeabi_lmul", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
331 { RTLIB::SHL_I64, "__aeabi_llsl", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
332 { RTLIB::SRL_I64, "__aeabi_llsr", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
333 { RTLIB::SRA_I64, "__aeabi_lasr", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
335 // Integer division functions
336 // RTABI chapter 4.3.1
337 { RTLIB::SDIV_I8, "__aeabi_idiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
338 { RTLIB::SDIV_I16, "__aeabi_idiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
339 { RTLIB::SDIV_I32, "__aeabi_idiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
340 { RTLIB::SDIV_I64, "__aeabi_ldivmod", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
341 { RTLIB::UDIV_I8, "__aeabi_uidiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
342 { RTLIB::UDIV_I16, "__aeabi_uidiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
343 { RTLIB::UDIV_I32, "__aeabi_uidiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
344 { RTLIB::UDIV_I64, "__aeabi_uldivmod", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
347 // RTABI chapter 4.3.4
348 { RTLIB::MEMCPY, "__aeabi_memcpy", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
349 { RTLIB::MEMMOVE, "__aeabi_memmove", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
350 { RTLIB::MEMSET, "__aeabi_memset", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
353 for (const auto &LC : LibraryCalls) {
354 setLibcallName(LC.Op, LC.Name);
355 setLibcallCallingConv(LC.Op, LC.CC);
356 if (LC.Cond != ISD::SETCC_INVALID)
357 setCmpLibcallCC(LC.Op, LC.Cond);
361 if (Subtarget->isTargetWindows()) {
362 static const struct {
363 const RTLIB::Libcall Op;
364 const char * const Name;
365 const CallingConv::ID CC;
367 { RTLIB::FPTOSINT_F32_I64, "__stoi64", CallingConv::ARM_AAPCS_VFP },
368 { RTLIB::FPTOSINT_F64_I64, "__dtoi64", CallingConv::ARM_AAPCS_VFP },
369 { RTLIB::FPTOUINT_F32_I64, "__stou64", CallingConv::ARM_AAPCS_VFP },
370 { RTLIB::FPTOUINT_F64_I64, "__dtou64", CallingConv::ARM_AAPCS_VFP },
371 { RTLIB::SINTTOFP_I64_F32, "__i64tos", CallingConv::ARM_AAPCS_VFP },
372 { RTLIB::SINTTOFP_I64_F64, "__i64tod", CallingConv::ARM_AAPCS_VFP },
373 { RTLIB::UINTTOFP_I64_F32, "__u64tos", CallingConv::ARM_AAPCS_VFP },
374 { RTLIB::UINTTOFP_I64_F64, "__u64tod", CallingConv::ARM_AAPCS_VFP },
377 for (const auto &LC : LibraryCalls) {
378 setLibcallName(LC.Op, LC.Name);
379 setLibcallCallingConv(LC.Op, LC.CC);
383 // Use divmod compiler-rt calls for iOS 5.0 and later.
384 if (Subtarget->getTargetTriple().isiOS() &&
385 !Subtarget->getTargetTriple().isOSVersionLT(5, 0)) {
386 setLibcallName(RTLIB::SDIVREM_I32, "__divmodsi4");
387 setLibcallName(RTLIB::UDIVREM_I32, "__udivmodsi4");
390 if (Subtarget->isThumb1Only())
391 addRegisterClass(MVT::i32, &ARM::tGPRRegClass);
393 addRegisterClass(MVT::i32, &ARM::GPRRegClass);
394 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
395 !Subtarget->isThumb1Only()) {
396 addRegisterClass(MVT::f32, &ARM::SPRRegClass);
397 if (!Subtarget->isFPOnlySP())
398 addRegisterClass(MVT::f64, &ARM::DPRRegClass);
400 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
403 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
404 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
405 for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
406 InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
407 setTruncStoreAction((MVT::SimpleValueType)VT,
408 (MVT::SimpleValueType)InnerVT, Expand);
409 setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT, Expand);
410 setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT, Expand);
411 setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT, Expand);
413 setOperationAction(ISD::MULHS, (MVT::SimpleValueType)VT, Expand);
414 setOperationAction(ISD::SMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
415 setOperationAction(ISD::MULHU, (MVT::SimpleValueType)VT, Expand);
416 setOperationAction(ISD::UMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
418 setOperationAction(ISD::BSWAP, (MVT::SimpleValueType)VT, Expand);
421 setOperationAction(ISD::ConstantFP, MVT::f32, Custom);
422 setOperationAction(ISD::ConstantFP, MVT::f64, Custom);
424 if (Subtarget->hasNEON()) {
425 addDRTypeForNEON(MVT::v2f32);
426 addDRTypeForNEON(MVT::v8i8);
427 addDRTypeForNEON(MVT::v4i16);
428 addDRTypeForNEON(MVT::v2i32);
429 addDRTypeForNEON(MVT::v1i64);
431 addQRTypeForNEON(MVT::v4f32);
432 addQRTypeForNEON(MVT::v2f64);
433 addQRTypeForNEON(MVT::v16i8);
434 addQRTypeForNEON(MVT::v8i16);
435 addQRTypeForNEON(MVT::v4i32);
436 addQRTypeForNEON(MVT::v2i64);
438 // v2f64 is legal so that QR subregs can be extracted as f64 elements, but
439 // neither Neon nor VFP support any arithmetic operations on it.
440 // The same with v4f32. But keep in mind that vadd, vsub, vmul are natively
441 // supported for v4f32.
442 setOperationAction(ISD::FADD, MVT::v2f64, Expand);
443 setOperationAction(ISD::FSUB, MVT::v2f64, Expand);
444 setOperationAction(ISD::FMUL, MVT::v2f64, Expand);
445 // FIXME: Code duplication: FDIV and FREM are expanded always, see
446 // ARMTargetLowering::addTypeForNEON method for details.
447 setOperationAction(ISD::FDIV, MVT::v2f64, Expand);
448 setOperationAction(ISD::FREM, MVT::v2f64, Expand);
449 // FIXME: Create unittest.
450 // In another words, find a way when "copysign" appears in DAG with vector
452 setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand);
453 // FIXME: Code duplication: SETCC has custom operation action, see
454 // ARMTargetLowering::addTypeForNEON method for details.
455 setOperationAction(ISD::SETCC, MVT::v2f64, Expand);
456 // FIXME: Create unittest for FNEG and for FABS.
457 setOperationAction(ISD::FNEG, MVT::v2f64, Expand);
458 setOperationAction(ISD::FABS, MVT::v2f64, Expand);
459 setOperationAction(ISD::FSQRT, MVT::v2f64, Expand);
460 setOperationAction(ISD::FSIN, MVT::v2f64, Expand);
461 setOperationAction(ISD::FCOS, MVT::v2f64, Expand);
462 setOperationAction(ISD::FPOWI, MVT::v2f64, Expand);
463 setOperationAction(ISD::FPOW, MVT::v2f64, Expand);
464 setOperationAction(ISD::FLOG, MVT::v2f64, Expand);
465 setOperationAction(ISD::FLOG2, MVT::v2f64, Expand);
466 setOperationAction(ISD::FLOG10, MVT::v2f64, Expand);
467 setOperationAction(ISD::FEXP, MVT::v2f64, Expand);
468 setOperationAction(ISD::FEXP2, MVT::v2f64, Expand);
469 // FIXME: Create unittest for FCEIL, FTRUNC, FRINT, FNEARBYINT, FFLOOR.
470 setOperationAction(ISD::FCEIL, MVT::v2f64, Expand);
471 setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand);
472 setOperationAction(ISD::FRINT, MVT::v2f64, Expand);
473 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand);
474 setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand);
475 setOperationAction(ISD::FMA, MVT::v2f64, Expand);
477 setOperationAction(ISD::FSQRT, MVT::v4f32, Expand);
478 setOperationAction(ISD::FSIN, MVT::v4f32, Expand);
479 setOperationAction(ISD::FCOS, MVT::v4f32, Expand);
480 setOperationAction(ISD::FPOWI, MVT::v4f32, Expand);
481 setOperationAction(ISD::FPOW, MVT::v4f32, Expand);
482 setOperationAction(ISD::FLOG, MVT::v4f32, Expand);
483 setOperationAction(ISD::FLOG2, MVT::v4f32, Expand);
484 setOperationAction(ISD::FLOG10, MVT::v4f32, Expand);
485 setOperationAction(ISD::FEXP, MVT::v4f32, Expand);
486 setOperationAction(ISD::FEXP2, MVT::v4f32, Expand);
487 setOperationAction(ISD::FCEIL, MVT::v4f32, Expand);
488 setOperationAction(ISD::FTRUNC, MVT::v4f32, Expand);
489 setOperationAction(ISD::FRINT, MVT::v4f32, Expand);
490 setOperationAction(ISD::FNEARBYINT, MVT::v4f32, Expand);
491 setOperationAction(ISD::FFLOOR, MVT::v4f32, Expand);
493 // Mark v2f32 intrinsics.
494 setOperationAction(ISD::FSQRT, MVT::v2f32, Expand);
495 setOperationAction(ISD::FSIN, MVT::v2f32, Expand);
496 setOperationAction(ISD::FCOS, MVT::v2f32, Expand);
497 setOperationAction(ISD::FPOWI, MVT::v2f32, Expand);
498 setOperationAction(ISD::FPOW, MVT::v2f32, Expand);
499 setOperationAction(ISD::FLOG, MVT::v2f32, Expand);
500 setOperationAction(ISD::FLOG2, MVT::v2f32, Expand);
501 setOperationAction(ISD::FLOG10, MVT::v2f32, Expand);
502 setOperationAction(ISD::FEXP, MVT::v2f32, Expand);
503 setOperationAction(ISD::FEXP2, MVT::v2f32, Expand);
504 setOperationAction(ISD::FCEIL, MVT::v2f32, Expand);
505 setOperationAction(ISD::FTRUNC, MVT::v2f32, Expand);
506 setOperationAction(ISD::FRINT, MVT::v2f32, Expand);
507 setOperationAction(ISD::FNEARBYINT, MVT::v2f32, Expand);
508 setOperationAction(ISD::FFLOOR, MVT::v2f32, Expand);
510 // Neon does not support some operations on v1i64 and v2i64 types.
511 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
512 // Custom handling for some quad-vector types to detect VMULL.
513 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
514 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
515 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
516 // Custom handling for some vector types to avoid expensive expansions
517 setOperationAction(ISD::SDIV, MVT::v4i16, Custom);
518 setOperationAction(ISD::SDIV, MVT::v8i8, Custom);
519 setOperationAction(ISD::UDIV, MVT::v4i16, Custom);
520 setOperationAction(ISD::UDIV, MVT::v8i8, Custom);
521 setOperationAction(ISD::SETCC, MVT::v1i64, Expand);
522 setOperationAction(ISD::SETCC, MVT::v2i64, Expand);
523 // Neon does not have single instruction SINT_TO_FP and UINT_TO_FP with
524 // a destination type that is wider than the source, and nor does
525 // it have a FP_TO_[SU]INT instruction with a narrower destination than
527 setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Custom);
528 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
529 setOperationAction(ISD::FP_TO_UINT, MVT::v4i16, Custom);
530 setOperationAction(ISD::FP_TO_SINT, MVT::v4i16, Custom);
532 setOperationAction(ISD::FP_ROUND, MVT::v2f32, Expand);
533 setOperationAction(ISD::FP_EXTEND, MVT::v2f64, Expand);
535 // NEON does not have single instruction CTPOP for vectors with element
536 // types wider than 8-bits. However, custom lowering can leverage the
537 // v8i8/v16i8 vcnt instruction.
538 setOperationAction(ISD::CTPOP, MVT::v2i32, Custom);
539 setOperationAction(ISD::CTPOP, MVT::v4i32, Custom);
540 setOperationAction(ISD::CTPOP, MVT::v4i16, Custom);
541 setOperationAction(ISD::CTPOP, MVT::v8i16, Custom);
543 // NEON only has FMA instructions as of VFP4.
544 if (!Subtarget->hasVFP4()) {
545 setOperationAction(ISD::FMA, MVT::v2f32, Expand);
546 setOperationAction(ISD::FMA, MVT::v4f32, Expand);
549 setTargetDAGCombine(ISD::INTRINSIC_VOID);
550 setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN);
551 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
552 setTargetDAGCombine(ISD::SHL);
553 setTargetDAGCombine(ISD::SRL);
554 setTargetDAGCombine(ISD::SRA);
555 setTargetDAGCombine(ISD::SIGN_EXTEND);
556 setTargetDAGCombine(ISD::ZERO_EXTEND);
557 setTargetDAGCombine(ISD::ANY_EXTEND);
558 setTargetDAGCombine(ISD::SELECT_CC);
559 setTargetDAGCombine(ISD::BUILD_VECTOR);
560 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
561 setTargetDAGCombine(ISD::INSERT_VECTOR_ELT);
562 setTargetDAGCombine(ISD::STORE);
563 setTargetDAGCombine(ISD::FP_TO_SINT);
564 setTargetDAGCombine(ISD::FP_TO_UINT);
565 setTargetDAGCombine(ISD::FDIV);
567 // It is legal to extload from v4i8 to v4i16 or v4i32.
568 MVT Tys[6] = {MVT::v8i8, MVT::v4i8, MVT::v2i8,
569 MVT::v4i16, MVT::v2i16,
571 for (unsigned i = 0; i < 6; ++i) {
572 setLoadExtAction(ISD::EXTLOAD, Tys[i], Legal);
573 setLoadExtAction(ISD::ZEXTLOAD, Tys[i], Legal);
574 setLoadExtAction(ISD::SEXTLOAD, Tys[i], Legal);
578 // ARM and Thumb2 support UMLAL/SMLAL.
579 if (!Subtarget->isThumb1Only())
580 setTargetDAGCombine(ISD::ADDC);
583 computeRegisterProperties();
585 // ARM does not have f32 extending load.
586 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
588 // ARM does not have i1 sign extending load.
589 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
591 // ARM supports all 4 flavors of integer indexed load / store.
592 if (!Subtarget->isThumb1Only()) {
593 for (unsigned im = (unsigned)ISD::PRE_INC;
594 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
595 setIndexedLoadAction(im, MVT::i1, Legal);
596 setIndexedLoadAction(im, MVT::i8, Legal);
597 setIndexedLoadAction(im, MVT::i16, Legal);
598 setIndexedLoadAction(im, MVT::i32, Legal);
599 setIndexedStoreAction(im, MVT::i1, Legal);
600 setIndexedStoreAction(im, MVT::i8, Legal);
601 setIndexedStoreAction(im, MVT::i16, Legal);
602 setIndexedStoreAction(im, MVT::i32, Legal);
606 setOperationAction(ISD::SADDO, MVT::i32, Custom);
607 setOperationAction(ISD::UADDO, MVT::i32, Custom);
608 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
609 setOperationAction(ISD::USUBO, MVT::i32, Custom);
611 // i64 operation support.
612 setOperationAction(ISD::MUL, MVT::i64, Expand);
613 setOperationAction(ISD::MULHU, MVT::i32, Expand);
614 if (Subtarget->isThumb1Only()) {
615 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
616 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
618 if (Subtarget->isThumb1Only() || !Subtarget->hasV6Ops()
619 || (Subtarget->isThumb2() && !Subtarget->hasThumb2DSP()))
620 setOperationAction(ISD::MULHS, MVT::i32, Expand);
622 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
623 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
624 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
625 setOperationAction(ISD::SRL, MVT::i64, Custom);
626 setOperationAction(ISD::SRA, MVT::i64, Custom);
628 if (!Subtarget->isThumb1Only()) {
629 // FIXME: We should do this for Thumb1 as well.
630 setOperationAction(ISD::ADDC, MVT::i32, Custom);
631 setOperationAction(ISD::ADDE, MVT::i32, Custom);
632 setOperationAction(ISD::SUBC, MVT::i32, Custom);
633 setOperationAction(ISD::SUBE, MVT::i32, Custom);
636 // ARM does not have ROTL.
637 setOperationAction(ISD::ROTL, MVT::i32, Expand);
638 setOperationAction(ISD::CTTZ, MVT::i32, Custom);
639 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
640 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
641 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
643 // These just redirect to CTTZ and CTLZ on ARM.
644 setOperationAction(ISD::CTTZ_ZERO_UNDEF , MVT::i32 , Expand);
645 setOperationAction(ISD::CTLZ_ZERO_UNDEF , MVT::i32 , Expand);
647 setOperationAction(ISD::READCYCLECOUNTER, MVT::i64, Custom);
649 // Only ARMv6 has BSWAP.
650 if (!Subtarget->hasV6Ops())
651 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
653 if (!(Subtarget->hasDivide() && Subtarget->isThumb2()) &&
654 !(Subtarget->hasDivideInARMMode() && !Subtarget->isThumb())) {
655 // These are expanded into libcalls if the cpu doesn't have HW divider.
656 setOperationAction(ISD::SDIV, MVT::i32, Expand);
657 setOperationAction(ISD::UDIV, MVT::i32, Expand);
660 // FIXME: Also set divmod for SREM on EABI
661 setOperationAction(ISD::SREM, MVT::i32, Expand);
662 setOperationAction(ISD::UREM, MVT::i32, Expand);
663 // Register based DivRem for AEABI (RTABI 4.2)
664 if (Subtarget->isTargetAEABI()) {
665 setLibcallName(RTLIB::SDIVREM_I8, "__aeabi_idivmod");
666 setLibcallName(RTLIB::SDIVREM_I16, "__aeabi_idivmod");
667 setLibcallName(RTLIB::SDIVREM_I32, "__aeabi_idivmod");
668 setLibcallName(RTLIB::SDIVREM_I64, "__aeabi_ldivmod");
669 setLibcallName(RTLIB::UDIVREM_I8, "__aeabi_uidivmod");
670 setLibcallName(RTLIB::UDIVREM_I16, "__aeabi_uidivmod");
671 setLibcallName(RTLIB::UDIVREM_I32, "__aeabi_uidivmod");
672 setLibcallName(RTLIB::UDIVREM_I64, "__aeabi_uldivmod");
674 setLibcallCallingConv(RTLIB::SDIVREM_I8, CallingConv::ARM_AAPCS);
675 setLibcallCallingConv(RTLIB::SDIVREM_I16, CallingConv::ARM_AAPCS);
676 setLibcallCallingConv(RTLIB::SDIVREM_I32, CallingConv::ARM_AAPCS);
677 setLibcallCallingConv(RTLIB::SDIVREM_I64, CallingConv::ARM_AAPCS);
678 setLibcallCallingConv(RTLIB::UDIVREM_I8, CallingConv::ARM_AAPCS);
679 setLibcallCallingConv(RTLIB::UDIVREM_I16, CallingConv::ARM_AAPCS);
680 setLibcallCallingConv(RTLIB::UDIVREM_I32, CallingConv::ARM_AAPCS);
681 setLibcallCallingConv(RTLIB::UDIVREM_I64, CallingConv::ARM_AAPCS);
683 setOperationAction(ISD::SDIVREM, MVT::i32, Custom);
684 setOperationAction(ISD::UDIVREM, MVT::i32, Custom);
686 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
687 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
690 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
691 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
692 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
693 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
694 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
696 setOperationAction(ISD::TRAP, MVT::Other, Legal);
698 // Use the default implementation.
699 setOperationAction(ISD::VASTART, MVT::Other, Custom);
700 setOperationAction(ISD::VAARG, MVT::Other, Expand);
701 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
702 setOperationAction(ISD::VAEND, MVT::Other, Expand);
703 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
704 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
706 if (!Subtarget->isTargetMachO()) {
707 // Non-MachO platforms may return values in these registers via the
708 // personality function.
709 setExceptionPointerRegister(ARM::R0);
710 setExceptionSelectorRegister(ARM::R1);
713 if (Subtarget->getTargetTriple().isWindowsItaniumEnvironment())
714 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
716 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
718 // ARMv6 Thumb1 (except for CPUs that support dmb / dsb) and earlier use
719 // the default expansion.
720 if (Subtarget->hasAnyDataBarrier() && !Subtarget->isThumb1Only()) {
721 // ATOMIC_FENCE needs custom lowering; the others should have been expanded
722 // to ldrex/strex loops already.
723 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Custom);
725 // On v8, we have particularly efficient implementations of atomic fences
726 // if they can be combined with nearby atomic loads and stores.
727 if (!Subtarget->hasV8Ops()) {
728 // Automatically insert fences (dmb ist) around ATOMIC_SWAP etc.
729 setInsertFencesForAtomic(true);
732 // If there's anything we can use as a barrier, go through custom lowering
734 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other,
735 Subtarget->hasAnyDataBarrier() ? Custom : Expand);
737 // Set them all for expansion, which will force libcalls.
738 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Expand);
739 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Expand);
740 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Expand);
741 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Expand);
742 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Expand);
743 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Expand);
744 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Expand);
745 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Expand);
746 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i32, Expand);
747 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i32, Expand);
748 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i32, Expand);
749 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i32, Expand);
750 // Mark ATOMIC_LOAD and ATOMIC_STORE custom so we can handle the
751 // Unordered/Monotonic case.
752 setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Custom);
753 setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Custom);
756 setOperationAction(ISD::PREFETCH, MVT::Other, Custom);
758 // Requires SXTB/SXTH, available on v6 and up in both ARM and Thumb modes.
759 if (!Subtarget->hasV6Ops()) {
760 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
761 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
763 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
765 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
766 !Subtarget->isThumb1Only()) {
767 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
768 // iff target supports vfp2.
769 setOperationAction(ISD::BITCAST, MVT::i64, Custom);
770 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
773 // We want to custom lower some of our intrinsics.
774 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
775 if (Subtarget->isTargetDarwin()) {
776 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
777 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
778 setLibcallName(RTLIB::UNWIND_RESUME, "_Unwind_SjLj_Resume");
781 setOperationAction(ISD::SETCC, MVT::i32, Expand);
782 setOperationAction(ISD::SETCC, MVT::f32, Expand);
783 setOperationAction(ISD::SETCC, MVT::f64, Expand);
784 setOperationAction(ISD::SELECT, MVT::i32, Custom);
785 setOperationAction(ISD::SELECT, MVT::f32, Custom);
786 setOperationAction(ISD::SELECT, MVT::f64, Custom);
787 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
788 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
789 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
791 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
792 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
793 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
794 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
795 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
797 // We don't support sin/cos/fmod/copysign/pow
798 setOperationAction(ISD::FSIN, MVT::f64, Expand);
799 setOperationAction(ISD::FSIN, MVT::f32, Expand);
800 setOperationAction(ISD::FCOS, MVT::f32, Expand);
801 setOperationAction(ISD::FCOS, MVT::f64, Expand);
802 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
803 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
804 setOperationAction(ISD::FREM, MVT::f64, Expand);
805 setOperationAction(ISD::FREM, MVT::f32, Expand);
806 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
807 !Subtarget->isThumb1Only()) {
808 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
809 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
811 setOperationAction(ISD::FPOW, MVT::f64, Expand);
812 setOperationAction(ISD::FPOW, MVT::f32, Expand);
814 if (!Subtarget->hasVFP4()) {
815 setOperationAction(ISD::FMA, MVT::f64, Expand);
816 setOperationAction(ISD::FMA, MVT::f32, Expand);
819 // Various VFP goodness
820 if (!TM.Options.UseSoftFloat && !Subtarget->isThumb1Only()) {
821 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
822 if (Subtarget->hasVFP2()) {
823 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
824 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
825 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
826 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
829 // v8 adds f64 <-> f16 conversion. Before that it should be expanded.
830 if (!Subtarget->hasV8Ops()) {
831 setOperationAction(ISD::FP16_TO_FP, MVT::f64, Expand);
832 setOperationAction(ISD::FP_TO_FP16, MVT::f64, Expand);
835 // fp16 is a special v7 extension that adds f16 <-> f32 conversions.
836 if (!Subtarget->hasFP16()) {
837 setOperationAction(ISD::FP16_TO_FP, MVT::f32, Expand);
838 setOperationAction(ISD::FP_TO_FP16, MVT::f32, Expand);
842 // Combine sin / cos into one node or libcall if possible.
843 if (Subtarget->hasSinCos()) {
844 setLibcallName(RTLIB::SINCOS_F32, "sincosf");
845 setLibcallName(RTLIB::SINCOS_F64, "sincos");
846 if (Subtarget->getTargetTriple().getOS() == Triple::IOS) {
847 // For iOS, we don't want to the normal expansion of a libcall to
848 // sincos. We want to issue a libcall to __sincos_stret.
849 setOperationAction(ISD::FSINCOS, MVT::f64, Custom);
850 setOperationAction(ISD::FSINCOS, MVT::f32, Custom);
854 // We have target-specific dag combine patterns for the following nodes:
855 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
856 setTargetDAGCombine(ISD::ADD);
857 setTargetDAGCombine(ISD::SUB);
858 setTargetDAGCombine(ISD::MUL);
859 setTargetDAGCombine(ISD::AND);
860 setTargetDAGCombine(ISD::OR);
861 setTargetDAGCombine(ISD::XOR);
863 if (Subtarget->hasV6Ops())
864 setTargetDAGCombine(ISD::SRL);
866 setStackPointerRegisterToSaveRestore(ARM::SP);
868 if (TM.Options.UseSoftFloat || Subtarget->isThumb1Only() ||
869 !Subtarget->hasVFP2())
870 setSchedulingPreference(Sched::RegPressure);
872 setSchedulingPreference(Sched::Hybrid);
874 //// temporary - rewrite interface to use type
875 MaxStoresPerMemset = 8;
876 MaxStoresPerMemsetOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
877 MaxStoresPerMemcpy = 4; // For @llvm.memcpy -> sequence of stores
878 MaxStoresPerMemcpyOptSize = Subtarget->isTargetDarwin() ? 4 : 2;
879 MaxStoresPerMemmove = 4; // For @llvm.memmove -> sequence of stores
880 MaxStoresPerMemmoveOptSize = Subtarget->isTargetDarwin() ? 4 : 2;
882 // On ARM arguments smaller than 4 bytes are extended, so all arguments
883 // are at least 4 bytes aligned.
884 setMinStackArgumentAlignment(4);
886 // Prefer likely predicted branches to selects on out-of-order cores.
887 PredictableSelectIsExpensive = Subtarget->isLikeA9();
889 setMinFunctionAlignment(Subtarget->isThumb() ? 1 : 2);
892 // FIXME: It might make sense to define the representative register class as the
893 // nearest super-register that has a non-null superset. For example, DPR_VFP2 is
894 // a super-register of SPR, and DPR is a superset if DPR_VFP2. Consequently,
895 // SPR's representative would be DPR_VFP2. This should work well if register
896 // pressure tracking were modified such that a register use would increment the
897 // pressure of the register class's representative and all of it's super
898 // classes' representatives transitively. We have not implemented this because
899 // of the difficulty prior to coalescing of modeling operand register classes
900 // due to the common occurrence of cross class copies and subregister insertions
902 std::pair<const TargetRegisterClass*, uint8_t>
903 ARMTargetLowering::findRepresentativeClass(MVT VT) const{
904 const TargetRegisterClass *RRC = nullptr;
906 switch (VT.SimpleTy) {
908 return TargetLowering::findRepresentativeClass(VT);
909 // Use DPR as representative register class for all floating point
910 // and vector types. Since there are 32 SPR registers and 32 DPR registers so
911 // the cost is 1 for both f32 and f64.
912 case MVT::f32: case MVT::f64: case MVT::v8i8: case MVT::v4i16:
913 case MVT::v2i32: case MVT::v1i64: case MVT::v2f32:
914 RRC = &ARM::DPRRegClass;
915 // When NEON is used for SP, only half of the register file is available
916 // because operations that define both SP and DP results will be constrained
917 // to the VFP2 class (D0-D15). We currently model this constraint prior to
918 // coalescing by double-counting the SP regs. See the FIXME above.
919 if (Subtarget->useNEONForSinglePrecisionFP())
922 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
923 case MVT::v4f32: case MVT::v2f64:
924 RRC = &ARM::DPRRegClass;
928 RRC = &ARM::DPRRegClass;
932 RRC = &ARM::DPRRegClass;
936 return std::make_pair(RRC, Cost);
939 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
941 default: return nullptr;
942 case ARMISD::Wrapper: return "ARMISD::Wrapper";
943 case ARMISD::WrapperPIC: return "ARMISD::WrapperPIC";
944 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
945 case ARMISD::CALL: return "ARMISD::CALL";
946 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
947 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
948 case ARMISD::tCALL: return "ARMISD::tCALL";
949 case ARMISD::BRCOND: return "ARMISD::BRCOND";
950 case ARMISD::BR_JT: return "ARMISD::BR_JT";
951 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
952 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
953 case ARMISD::INTRET_FLAG: return "ARMISD::INTRET_FLAG";
954 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
955 case ARMISD::CMP: return "ARMISD::CMP";
956 case ARMISD::CMN: return "ARMISD::CMN";
957 case ARMISD::CMPZ: return "ARMISD::CMPZ";
958 case ARMISD::CMPFP: return "ARMISD::CMPFP";
959 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
960 case ARMISD::BCC_i64: return "ARMISD::BCC_i64";
961 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
963 case ARMISD::CMOV: return "ARMISD::CMOV";
965 case ARMISD::RBIT: return "ARMISD::RBIT";
967 case ARMISD::FTOSI: return "ARMISD::FTOSI";
968 case ARMISD::FTOUI: return "ARMISD::FTOUI";
969 case ARMISD::SITOF: return "ARMISD::SITOF";
970 case ARMISD::UITOF: return "ARMISD::UITOF";
972 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
973 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
974 case ARMISD::RRX: return "ARMISD::RRX";
976 case ARMISD::ADDC: return "ARMISD::ADDC";
977 case ARMISD::ADDE: return "ARMISD::ADDE";
978 case ARMISD::SUBC: return "ARMISD::SUBC";
979 case ARMISD::SUBE: return "ARMISD::SUBE";
981 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
982 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
984 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
985 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
987 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN";
989 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
991 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
993 case ARMISD::MEMBARRIER_MCR: return "ARMISD::MEMBARRIER_MCR";
995 case ARMISD::PRELOAD: return "ARMISD::PRELOAD";
997 case ARMISD::WIN__CHKSTK: return "ARMISD:::WIN__CHKSTK";
999 case ARMISD::VCEQ: return "ARMISD::VCEQ";
1000 case ARMISD::VCEQZ: return "ARMISD::VCEQZ";
1001 case ARMISD::VCGE: return "ARMISD::VCGE";
1002 case ARMISD::VCGEZ: return "ARMISD::VCGEZ";
1003 case ARMISD::VCLEZ: return "ARMISD::VCLEZ";
1004 case ARMISD::VCGEU: return "ARMISD::VCGEU";
1005 case ARMISD::VCGT: return "ARMISD::VCGT";
1006 case ARMISD::VCGTZ: return "ARMISD::VCGTZ";
1007 case ARMISD::VCLTZ: return "ARMISD::VCLTZ";
1008 case ARMISD::VCGTU: return "ARMISD::VCGTU";
1009 case ARMISD::VTST: return "ARMISD::VTST";
1011 case ARMISD::VSHL: return "ARMISD::VSHL";
1012 case ARMISD::VSHRs: return "ARMISD::VSHRs";
1013 case ARMISD::VSHRu: return "ARMISD::VSHRu";
1014 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
1015 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
1016 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
1017 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
1018 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
1019 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
1020 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
1021 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
1022 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
1023 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
1024 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
1025 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
1026 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
1027 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
1028 case ARMISD::VMOVIMM: return "ARMISD::VMOVIMM";
1029 case ARMISD::VMVNIMM: return "ARMISD::VMVNIMM";
1030 case ARMISD::VMOVFPIMM: return "ARMISD::VMOVFPIMM";
1031 case ARMISD::VDUP: return "ARMISD::VDUP";
1032 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
1033 case ARMISD::VEXT: return "ARMISD::VEXT";
1034 case ARMISD::VREV64: return "ARMISD::VREV64";
1035 case ARMISD::VREV32: return "ARMISD::VREV32";
1036 case ARMISD::VREV16: return "ARMISD::VREV16";
1037 case ARMISD::VZIP: return "ARMISD::VZIP";
1038 case ARMISD::VUZP: return "ARMISD::VUZP";
1039 case ARMISD::VTRN: return "ARMISD::VTRN";
1040 case ARMISD::VTBL1: return "ARMISD::VTBL1";
1041 case ARMISD::VTBL2: return "ARMISD::VTBL2";
1042 case ARMISD::VMULLs: return "ARMISD::VMULLs";
1043 case ARMISD::VMULLu: return "ARMISD::VMULLu";
1044 case ARMISD::UMLAL: return "ARMISD::UMLAL";
1045 case ARMISD::SMLAL: return "ARMISD::SMLAL";
1046 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
1047 case ARMISD::FMAX: return "ARMISD::FMAX";
1048 case ARMISD::FMIN: return "ARMISD::FMIN";
1049 case ARMISD::VMAXNM: return "ARMISD::VMAX";
1050 case ARMISD::VMINNM: return "ARMISD::VMIN";
1051 case ARMISD::BFI: return "ARMISD::BFI";
1052 case ARMISD::VORRIMM: return "ARMISD::VORRIMM";
1053 case ARMISD::VBICIMM: return "ARMISD::VBICIMM";
1054 case ARMISD::VBSL: return "ARMISD::VBSL";
1055 case ARMISD::VLD2DUP: return "ARMISD::VLD2DUP";
1056 case ARMISD::VLD3DUP: return "ARMISD::VLD3DUP";
1057 case ARMISD::VLD4DUP: return "ARMISD::VLD4DUP";
1058 case ARMISD::VLD1_UPD: return "ARMISD::VLD1_UPD";
1059 case ARMISD::VLD2_UPD: return "ARMISD::VLD2_UPD";
1060 case ARMISD::VLD3_UPD: return "ARMISD::VLD3_UPD";
1061 case ARMISD::VLD4_UPD: return "ARMISD::VLD4_UPD";
1062 case ARMISD::VLD2LN_UPD: return "ARMISD::VLD2LN_UPD";
1063 case ARMISD::VLD3LN_UPD: return "ARMISD::VLD3LN_UPD";
1064 case ARMISD::VLD4LN_UPD: return "ARMISD::VLD4LN_UPD";
1065 case ARMISD::VLD2DUP_UPD: return "ARMISD::VLD2DUP_UPD";
1066 case ARMISD::VLD3DUP_UPD: return "ARMISD::VLD3DUP_UPD";
1067 case ARMISD::VLD4DUP_UPD: return "ARMISD::VLD4DUP_UPD";
1068 case ARMISD::VST1_UPD: return "ARMISD::VST1_UPD";
1069 case ARMISD::VST2_UPD: return "ARMISD::VST2_UPD";
1070 case ARMISD::VST3_UPD: return "ARMISD::VST3_UPD";
1071 case ARMISD::VST4_UPD: return "ARMISD::VST4_UPD";
1072 case ARMISD::VST2LN_UPD: return "ARMISD::VST2LN_UPD";
1073 case ARMISD::VST3LN_UPD: return "ARMISD::VST3LN_UPD";
1074 case ARMISD::VST4LN_UPD: return "ARMISD::VST4LN_UPD";
1078 EVT ARMTargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
1079 if (!VT.isVector()) return getPointerTy();
1080 return VT.changeVectorElementTypeToInteger();
1083 /// getRegClassFor - Return the register class that should be used for the
1084 /// specified value type.
1085 const TargetRegisterClass *ARMTargetLowering::getRegClassFor(MVT VT) const {
1086 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
1087 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
1088 // load / store 4 to 8 consecutive D registers.
1089 if (Subtarget->hasNEON()) {
1090 if (VT == MVT::v4i64)
1091 return &ARM::QQPRRegClass;
1092 if (VT == MVT::v8i64)
1093 return &ARM::QQQQPRRegClass;
1095 return TargetLowering::getRegClassFor(VT);
1098 // Create a fast isel object.
1100 ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
1101 const TargetLibraryInfo *libInfo) const {
1102 return ARM::createFastISel(funcInfo, libInfo);
1105 /// getMaximalGlobalOffset - Returns the maximal possible offset which can
1106 /// be used for loads / stores from the global.
1107 unsigned ARMTargetLowering::getMaximalGlobalOffset() const {
1108 return (Subtarget->isThumb1Only() ? 127 : 4095);
1111 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
1112 unsigned NumVals = N->getNumValues();
1114 return Sched::RegPressure;
1116 for (unsigned i = 0; i != NumVals; ++i) {
1117 EVT VT = N->getValueType(i);
1118 if (VT == MVT::Glue || VT == MVT::Other)
1120 if (VT.isFloatingPoint() || VT.isVector())
1124 if (!N->isMachineOpcode())
1125 return Sched::RegPressure;
1127 // Load are scheduled for latency even if there instruction itinerary
1128 // is not available.
1129 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1130 const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
1132 if (MCID.getNumDefs() == 0)
1133 return Sched::RegPressure;
1134 if (!Itins->isEmpty() &&
1135 Itins->getOperandCycle(MCID.getSchedClass(), 0) > 2)
1138 return Sched::RegPressure;
1141 //===----------------------------------------------------------------------===//
1143 //===----------------------------------------------------------------------===//
1145 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
1146 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
1148 default: llvm_unreachable("Unknown condition code!");
1149 case ISD::SETNE: return ARMCC::NE;
1150 case ISD::SETEQ: return ARMCC::EQ;
1151 case ISD::SETGT: return ARMCC::GT;
1152 case ISD::SETGE: return ARMCC::GE;
1153 case ISD::SETLT: return ARMCC::LT;
1154 case ISD::SETLE: return ARMCC::LE;
1155 case ISD::SETUGT: return ARMCC::HI;
1156 case ISD::SETUGE: return ARMCC::HS;
1157 case ISD::SETULT: return ARMCC::LO;
1158 case ISD::SETULE: return ARMCC::LS;
1162 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
1163 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
1164 ARMCC::CondCodes &CondCode2) {
1165 CondCode2 = ARMCC::AL;
1167 default: llvm_unreachable("Unknown FP condition!");
1169 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
1171 case ISD::SETOGT: CondCode = ARMCC::GT; break;
1173 case ISD::SETOGE: CondCode = ARMCC::GE; break;
1174 case ISD::SETOLT: CondCode = ARMCC::MI; break;
1175 case ISD::SETOLE: CondCode = ARMCC::LS; break;
1176 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
1177 case ISD::SETO: CondCode = ARMCC::VC; break;
1178 case ISD::SETUO: CondCode = ARMCC::VS; break;
1179 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
1180 case ISD::SETUGT: CondCode = ARMCC::HI; break;
1181 case ISD::SETUGE: CondCode = ARMCC::PL; break;
1183 case ISD::SETULT: CondCode = ARMCC::LT; break;
1185 case ISD::SETULE: CondCode = ARMCC::LE; break;
1187 case ISD::SETUNE: CondCode = ARMCC::NE; break;
1191 //===----------------------------------------------------------------------===//
1192 // Calling Convention Implementation
1193 //===----------------------------------------------------------------------===//
1195 #include "ARMGenCallingConv.inc"
1197 /// getEffectiveCallingConv - Get the effective calling convention, taking into
1198 /// account presence of floating point hardware and calling convention
1199 /// limitations, such as support for variadic functions.
1201 ARMTargetLowering::getEffectiveCallingConv(CallingConv::ID CC,
1202 bool isVarArg) const {
1205 llvm_unreachable("Unsupported calling convention");
1206 case CallingConv::ARM_AAPCS:
1207 case CallingConv::ARM_APCS:
1208 case CallingConv::GHC:
1210 case CallingConv::ARM_AAPCS_VFP:
1211 return isVarArg ? CallingConv::ARM_AAPCS : CallingConv::ARM_AAPCS_VFP;
1212 case CallingConv::C:
1213 if (!Subtarget->isAAPCS_ABI())
1214 return CallingConv::ARM_APCS;
1215 else if (Subtarget->hasVFP2() && !Subtarget->isThumb1Only() &&
1216 getTargetMachine().Options.FloatABIType == FloatABI::Hard &&
1218 return CallingConv::ARM_AAPCS_VFP;
1220 return CallingConv::ARM_AAPCS;
1221 case CallingConv::Fast:
1222 if (!Subtarget->isAAPCS_ABI()) {
1223 if (Subtarget->hasVFP2() && !Subtarget->isThumb1Only() && !isVarArg)
1224 return CallingConv::Fast;
1225 return CallingConv::ARM_APCS;
1226 } else if (Subtarget->hasVFP2() && !Subtarget->isThumb1Only() && !isVarArg)
1227 return CallingConv::ARM_AAPCS_VFP;
1229 return CallingConv::ARM_AAPCS;
1233 /// CCAssignFnForNode - Selects the correct CCAssignFn for the given
1234 /// CallingConvention.
1235 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
1237 bool isVarArg) const {
1238 switch (getEffectiveCallingConv(CC, isVarArg)) {
1240 llvm_unreachable("Unsupported calling convention");
1241 case CallingConv::ARM_APCS:
1242 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1243 case CallingConv::ARM_AAPCS:
1244 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1245 case CallingConv::ARM_AAPCS_VFP:
1246 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1247 case CallingConv::Fast:
1248 return (Return ? RetFastCC_ARM_APCS : FastCC_ARM_APCS);
1249 case CallingConv::GHC:
1250 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS_GHC);
1254 /// LowerCallResult - Lower the result values of a call into the
1255 /// appropriate copies out of appropriate physical registers.
1257 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1258 CallingConv::ID CallConv, bool isVarArg,
1259 const SmallVectorImpl<ISD::InputArg> &Ins,
1260 SDLoc dl, SelectionDAG &DAG,
1261 SmallVectorImpl<SDValue> &InVals,
1262 bool isThisReturn, SDValue ThisVal) const {
1264 // Assign locations to each value returned by this call.
1265 SmallVector<CCValAssign, 16> RVLocs;
1266 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1267 getTargetMachine(), RVLocs, *DAG.getContext(), Call);
1268 CCInfo.AnalyzeCallResult(Ins,
1269 CCAssignFnForNode(CallConv, /* Return*/ true,
1272 // Copy all of the result registers out of their specified physreg.
1273 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1274 CCValAssign VA = RVLocs[i];
1276 // Pass 'this' value directly from the argument to return value, to avoid
1277 // reg unit interference
1278 if (i == 0 && isThisReturn) {
1279 assert(!VA.needsCustom() && VA.getLocVT() == MVT::i32 &&
1280 "unexpected return calling convention register assignment");
1281 InVals.push_back(ThisVal);
1286 if (VA.needsCustom()) {
1287 // Handle f64 or half of a v2f64.
1288 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1290 Chain = Lo.getValue(1);
1291 InFlag = Lo.getValue(2);
1292 VA = RVLocs[++i]; // skip ahead to next loc
1293 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1295 Chain = Hi.getValue(1);
1296 InFlag = Hi.getValue(2);
1297 if (!Subtarget->isLittle())
1299 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1301 if (VA.getLocVT() == MVT::v2f64) {
1302 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1303 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1304 DAG.getConstant(0, MVT::i32));
1306 VA = RVLocs[++i]; // skip ahead to next loc
1307 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1308 Chain = Lo.getValue(1);
1309 InFlag = Lo.getValue(2);
1310 VA = RVLocs[++i]; // skip ahead to next loc
1311 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1312 Chain = Hi.getValue(1);
1313 InFlag = Hi.getValue(2);
1314 if (!Subtarget->isLittle())
1316 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1317 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1318 DAG.getConstant(1, MVT::i32));
1321 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
1323 Chain = Val.getValue(1);
1324 InFlag = Val.getValue(2);
1327 switch (VA.getLocInfo()) {
1328 default: llvm_unreachable("Unknown loc info!");
1329 case CCValAssign::Full: break;
1330 case CCValAssign::BCvt:
1331 Val = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), Val);
1335 InVals.push_back(Val);
1341 /// LowerMemOpCallTo - Store the argument to the stack.
1343 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
1344 SDValue StackPtr, SDValue Arg,
1345 SDLoc dl, SelectionDAG &DAG,
1346 const CCValAssign &VA,
1347 ISD::ArgFlagsTy Flags) const {
1348 unsigned LocMemOffset = VA.getLocMemOffset();
1349 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1350 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1351 return DAG.getStore(Chain, dl, Arg, PtrOff,
1352 MachinePointerInfo::getStack(LocMemOffset),
1356 void ARMTargetLowering::PassF64ArgInRegs(SDLoc dl, SelectionDAG &DAG,
1357 SDValue Chain, SDValue &Arg,
1358 RegsToPassVector &RegsToPass,
1359 CCValAssign &VA, CCValAssign &NextVA,
1361 SmallVectorImpl<SDValue> &MemOpChains,
1362 ISD::ArgFlagsTy Flags) const {
1364 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1365 DAG.getVTList(MVT::i32, MVT::i32), Arg);
1366 unsigned id = Subtarget->isLittle() ? 0 : 1;
1367 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd.getValue(id)));
1369 if (NextVA.isRegLoc())
1370 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1-id)));
1372 assert(NextVA.isMemLoc());
1373 if (!StackPtr.getNode())
1374 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1376 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1-id),
1382 /// LowerCall - Lowering a call into a callseq_start <-
1383 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
1386 ARMTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
1387 SmallVectorImpl<SDValue> &InVals) const {
1388 SelectionDAG &DAG = CLI.DAG;
1390 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
1391 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
1392 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
1393 SDValue Chain = CLI.Chain;
1394 SDValue Callee = CLI.Callee;
1395 bool &isTailCall = CLI.IsTailCall;
1396 CallingConv::ID CallConv = CLI.CallConv;
1397 bool doesNotRet = CLI.DoesNotReturn;
1398 bool isVarArg = CLI.IsVarArg;
1400 MachineFunction &MF = DAG.getMachineFunction();
1401 bool isStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1402 bool isThisReturn = false;
1403 bool isSibCall = false;
1405 // Disable tail calls if they're not supported.
1406 if (!Subtarget->supportsTailCall() || MF.getTarget().Options.DisableTailCalls)
1410 // Check if it's really possible to do a tail call.
1411 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1412 isVarArg, isStructRet, MF.getFunction()->hasStructRetAttr(),
1413 Outs, OutVals, Ins, DAG);
1414 if (!isTailCall && CLI.CS && CLI.CS->isMustTailCall())
1415 report_fatal_error("failed to perform tail call elimination on a call "
1416 "site marked musttail");
1417 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1418 // detected sibcalls.
1425 // Analyze operands of the call, assigning locations to each operand.
1426 SmallVector<CCValAssign, 16> ArgLocs;
1427 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1428 getTargetMachine(), ArgLocs, *DAG.getContext(), Call);
1429 CCInfo.AnalyzeCallOperands(Outs,
1430 CCAssignFnForNode(CallConv, /* Return*/ false,
1433 // Get a count of how many bytes are to be pushed on the stack.
1434 unsigned NumBytes = CCInfo.getNextStackOffset();
1436 // For tail calls, memory operands are available in our caller's stack.
1440 // Adjust the stack pointer for the new arguments...
1441 // These operations are automatically eliminated by the prolog/epilog pass
1443 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true),
1446 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1448 RegsToPassVector RegsToPass;
1449 SmallVector<SDValue, 8> MemOpChains;
1451 // Walk the register/memloc assignments, inserting copies/loads. In the case
1452 // of tail call optimization, arguments are handled later.
1453 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1455 ++i, ++realArgIdx) {
1456 CCValAssign &VA = ArgLocs[i];
1457 SDValue Arg = OutVals[realArgIdx];
1458 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1459 bool isByVal = Flags.isByVal();
1461 // Promote the value if needed.
1462 switch (VA.getLocInfo()) {
1463 default: llvm_unreachable("Unknown loc info!");
1464 case CCValAssign::Full: break;
1465 case CCValAssign::SExt:
1466 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1468 case CCValAssign::ZExt:
1469 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1471 case CCValAssign::AExt:
1472 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1474 case CCValAssign::BCvt:
1475 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1479 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1480 if (VA.needsCustom()) {
1481 if (VA.getLocVT() == MVT::v2f64) {
1482 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1483 DAG.getConstant(0, MVT::i32));
1484 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1485 DAG.getConstant(1, MVT::i32));
1487 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1488 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1490 VA = ArgLocs[++i]; // skip ahead to next loc
1491 if (VA.isRegLoc()) {
1492 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1493 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1495 assert(VA.isMemLoc());
1497 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1498 dl, DAG, VA, Flags));
1501 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1502 StackPtr, MemOpChains, Flags);
1504 } else if (VA.isRegLoc()) {
1505 if (realArgIdx == 0 && Flags.isReturned() && Outs[0].VT == MVT::i32) {
1506 assert(VA.getLocVT() == MVT::i32 &&
1507 "unexpected calling convention register assignment");
1508 assert(!Ins.empty() && Ins[0].VT == MVT::i32 &&
1509 "unexpected use of 'returned'");
1510 isThisReturn = true;
1512 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1513 } else if (isByVal) {
1514 assert(VA.isMemLoc());
1515 unsigned offset = 0;
1517 // True if this byval aggregate will be split between registers
1519 unsigned ByValArgsCount = CCInfo.getInRegsParamsCount();
1520 unsigned CurByValIdx = CCInfo.getInRegsParamsProceed();
1522 if (CurByValIdx < ByValArgsCount) {
1524 unsigned RegBegin, RegEnd;
1525 CCInfo.getInRegsParamInfo(CurByValIdx, RegBegin, RegEnd);
1527 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1529 for (i = 0, j = RegBegin; j < RegEnd; i++, j++) {
1530 SDValue Const = DAG.getConstant(4*i, MVT::i32);
1531 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
1532 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
1533 MachinePointerInfo(),
1534 false, false, false,
1535 DAG.InferPtrAlignment(AddArg));
1536 MemOpChains.push_back(Load.getValue(1));
1537 RegsToPass.push_back(std::make_pair(j, Load));
1540 // If parameter size outsides register area, "offset" value
1541 // helps us to calculate stack slot for remained part properly.
1542 offset = RegEnd - RegBegin;
1544 CCInfo.nextInRegsParam();
1547 if (Flags.getByValSize() > 4*offset) {
1548 unsigned LocMemOffset = VA.getLocMemOffset();
1549 SDValue StkPtrOff = DAG.getIntPtrConstant(LocMemOffset);
1550 SDValue Dst = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr,
1552 SDValue SrcOffset = DAG.getIntPtrConstant(4*offset);
1553 SDValue Src = DAG.getNode(ISD::ADD, dl, getPointerTy(), Arg, SrcOffset);
1554 SDValue SizeNode = DAG.getConstant(Flags.getByValSize() - 4*offset,
1556 SDValue AlignNode = DAG.getConstant(Flags.getByValAlign(), MVT::i32);
1558 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
1559 SDValue Ops[] = { Chain, Dst, Src, SizeNode, AlignNode};
1560 MemOpChains.push_back(DAG.getNode(ARMISD::COPY_STRUCT_BYVAL, dl, VTs,
1563 } else if (!isSibCall) {
1564 assert(VA.isMemLoc());
1566 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1567 dl, DAG, VA, Flags));
1571 if (!MemOpChains.empty())
1572 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains);
1574 // Build a sequence of copy-to-reg nodes chained together with token chain
1575 // and flag operands which copy the outgoing args into the appropriate regs.
1577 // Tail call byval lowering might overwrite argument registers so in case of
1578 // tail call optimization the copies to registers are lowered later.
1580 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1581 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1582 RegsToPass[i].second, InFlag);
1583 InFlag = Chain.getValue(1);
1586 // For tail calls lower the arguments to the 'real' stack slot.
1588 // Force all the incoming stack arguments to be loaded from the stack
1589 // before any new outgoing arguments are stored to the stack, because the
1590 // outgoing stack slots may alias the incoming argument stack slots, and
1591 // the alias isn't otherwise explicit. This is slightly more conservative
1592 // than necessary, because it means that each store effectively depends
1593 // on every argument instead of just those arguments it would clobber.
1595 // Do not flag preceding copytoreg stuff together with the following stuff.
1597 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1598 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1599 RegsToPass[i].second, InFlag);
1600 InFlag = Chain.getValue(1);
1605 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1606 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1607 // node so that legalize doesn't hack it.
1608 bool isDirect = false;
1609 bool isARMFunc = false;
1610 bool isLocalARMFunc = false;
1611 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1613 if (EnableARMLongCalls) {
1614 assert((Subtarget->isTargetWindows() ||
1615 getTargetMachine().getRelocationModel() == Reloc::Static) &&
1616 "long-calls with non-static relocation model!");
1617 // Handle a global address or an external symbol. If it's not one of
1618 // those, the target's already in a register, so we don't need to do
1620 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1621 const GlobalValue *GV = G->getGlobal();
1622 // Create a constant pool entry for the callee address
1623 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1624 ARMConstantPoolValue *CPV =
1625 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 0);
1627 // Get the address of the callee into a register
1628 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1629 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1630 Callee = DAG.getLoad(getPointerTy(), dl,
1631 DAG.getEntryNode(), CPAddr,
1632 MachinePointerInfo::getConstantPool(),
1633 false, false, false, 0);
1634 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1635 const char *Sym = S->getSymbol();
1637 // Create a constant pool entry for the callee address
1638 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1639 ARMConstantPoolValue *CPV =
1640 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1641 ARMPCLabelIndex, 0);
1642 // Get the address of the callee into a register
1643 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1644 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1645 Callee = DAG.getLoad(getPointerTy(), dl,
1646 DAG.getEntryNode(), CPAddr,
1647 MachinePointerInfo::getConstantPool(),
1648 false, false, false, 0);
1650 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1651 const GlobalValue *GV = G->getGlobal();
1653 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1654 bool isStub = (isExt && Subtarget->isTargetMachO()) &&
1655 getTargetMachine().getRelocationModel() != Reloc::Static;
1656 isARMFunc = !Subtarget->isThumb() || isStub;
1657 // ARM call to a local ARM function is predicable.
1658 isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking);
1659 // tBX takes a register source operand.
1660 if (isStub && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1661 assert(Subtarget->isTargetMachO() && "WrapperPIC use on non-MachO?");
1662 Callee = DAG.getNode(ARMISD::WrapperPIC, dl, getPointerTy(),
1663 DAG.getTargetGlobalAddress(GV, dl, getPointerTy()));
1664 } else if (Subtarget->isTargetCOFF()) {
1665 assert(Subtarget->isTargetWindows() &&
1666 "Windows is the only supported COFF target");
1667 unsigned TargetFlags = GV->hasDLLImportStorageClass()
1668 ? ARMII::MO_DLLIMPORT
1669 : ARMII::MO_NO_FLAG;
1670 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), /*Offset=*/0,
1672 if (GV->hasDLLImportStorageClass())
1673 Callee = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
1674 DAG.getNode(ARMISD::Wrapper, dl, getPointerTy(),
1675 Callee), MachinePointerInfo::getGOT(),
1676 false, false, false, 0);
1678 // On ELF targets for PIC code, direct calls should go through the PLT
1679 unsigned OpFlags = 0;
1680 if (Subtarget->isTargetELF() &&
1681 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1682 OpFlags = ARMII::MO_PLT;
1683 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
1685 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1687 bool isStub = Subtarget->isTargetMachO() &&
1688 getTargetMachine().getRelocationModel() != Reloc::Static;
1689 isARMFunc = !Subtarget->isThumb() || isStub;
1690 // tBX takes a register source operand.
1691 const char *Sym = S->getSymbol();
1692 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1693 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1694 ARMConstantPoolValue *CPV =
1695 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1696 ARMPCLabelIndex, 4);
1697 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1698 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1699 Callee = DAG.getLoad(getPointerTy(), dl,
1700 DAG.getEntryNode(), CPAddr,
1701 MachinePointerInfo::getConstantPool(),
1702 false, false, false, 0);
1703 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1704 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1705 getPointerTy(), Callee, PICLabel);
1707 unsigned OpFlags = 0;
1708 // On ELF targets for PIC code, direct calls should go through the PLT
1709 if (Subtarget->isTargetELF() &&
1710 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1711 OpFlags = ARMII::MO_PLT;
1712 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlags);
1716 // FIXME: handle tail calls differently.
1718 bool HasMinSizeAttr = MF.getFunction()->getAttributes().hasAttribute(
1719 AttributeSet::FunctionIndex, Attribute::MinSize);
1720 if (Subtarget->isThumb()) {
1721 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1722 CallOpc = ARMISD::CALL_NOLINK;
1724 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1726 if (!isDirect && !Subtarget->hasV5TOps())
1727 CallOpc = ARMISD::CALL_NOLINK;
1728 else if (doesNotRet && isDirect && Subtarget->hasRAS() &&
1729 // Emit regular call when code size is the priority
1731 // "mov lr, pc; b _foo" to avoid confusing the RSP
1732 CallOpc = ARMISD::CALL_NOLINK;
1734 CallOpc = isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL;
1737 std::vector<SDValue> Ops;
1738 Ops.push_back(Chain);
1739 Ops.push_back(Callee);
1741 // Add argument registers to the end of the list so that they are known live
1743 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1744 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1745 RegsToPass[i].second.getValueType()));
1747 // Add a register mask operand representing the call-preserved registers.
1749 const uint32_t *Mask;
1750 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
1751 const ARMBaseRegisterInfo *ARI = static_cast<const ARMBaseRegisterInfo*>(TRI);
1753 // For 'this' returns, use the R0-preserving mask if applicable
1754 Mask = ARI->getThisReturnPreservedMask(CallConv);
1756 // Set isThisReturn to false if the calling convention is not one that
1757 // allows 'returned' to be modeled in this way, so LowerCallResult does
1758 // not try to pass 'this' straight through
1759 isThisReturn = false;
1760 Mask = ARI->getCallPreservedMask(CallConv);
1763 Mask = ARI->getCallPreservedMask(CallConv);
1765 assert(Mask && "Missing call preserved mask for calling convention");
1766 Ops.push_back(DAG.getRegisterMask(Mask));
1769 if (InFlag.getNode())
1770 Ops.push_back(InFlag);
1772 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
1774 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, Ops);
1776 // Returns a chain and a flag for retval copy to use.
1777 Chain = DAG.getNode(CallOpc, dl, NodeTys, Ops);
1778 InFlag = Chain.getValue(1);
1780 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1781 DAG.getIntPtrConstant(0, true), InFlag, dl);
1783 InFlag = Chain.getValue(1);
1785 // Handle result values, copying them out of physregs into vregs that we
1787 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins, dl, DAG,
1788 InVals, isThisReturn,
1789 isThisReturn ? OutVals[0] : SDValue());
1792 /// HandleByVal - Every parameter *after* a byval parameter is passed
1793 /// on the stack. Remember the next parameter register to allocate,
1794 /// and then confiscate the rest of the parameter registers to insure
1797 ARMTargetLowering::HandleByVal(
1798 CCState *State, unsigned &size, unsigned Align) const {
1799 unsigned reg = State->AllocateReg(GPRArgRegs, 4);
1800 assert((State->getCallOrPrologue() == Prologue ||
1801 State->getCallOrPrologue() == Call) &&
1802 "unhandled ParmContext");
1804 if ((ARM::R0 <= reg) && (reg <= ARM::R3)) {
1805 if (Subtarget->isAAPCS_ABI() && Align > 4) {
1806 unsigned AlignInRegs = Align / 4;
1807 unsigned Waste = (ARM::R4 - reg) % AlignInRegs;
1808 for (unsigned i = 0; i < Waste; ++i)
1809 reg = State->AllocateReg(GPRArgRegs, 4);
1812 unsigned excess = 4 * (ARM::R4 - reg);
1814 // Special case when NSAA != SP and parameter size greater than size of
1815 // all remained GPR regs. In that case we can't split parameter, we must
1816 // send it to stack. We also must set NCRN to R4, so waste all
1817 // remained registers.
1818 const unsigned NSAAOffset = State->getNextStackOffset();
1819 if (Subtarget->isAAPCS_ABI() && NSAAOffset != 0 && size > excess) {
1820 while (State->AllocateReg(GPRArgRegs, 4))
1825 // First register for byval parameter is the first register that wasn't
1826 // allocated before this method call, so it would be "reg".
1827 // If parameter is small enough to be saved in range [reg, r4), then
1828 // the end (first after last) register would be reg + param-size-in-regs,
1829 // else parameter would be splitted between registers and stack,
1830 // end register would be r4 in this case.
1831 unsigned ByValRegBegin = reg;
1832 unsigned ByValRegEnd = (size < excess) ? reg + size/4 : (unsigned)ARM::R4;
1833 State->addInRegsParamInfo(ByValRegBegin, ByValRegEnd);
1834 // Note, first register is allocated in the beginning of function already,
1835 // allocate remained amount of registers we need.
1836 for (unsigned i = reg+1; i != ByValRegEnd; ++i)
1837 State->AllocateReg(GPRArgRegs, 4);
1838 // A byval parameter that is split between registers and memory needs its
1839 // size truncated here.
1840 // In the case where the entire structure fits in registers, we set the
1841 // size in memory to zero.
1850 /// MatchingStackOffset - Return true if the given stack call argument is
1851 /// already available in the same position (relatively) of the caller's
1852 /// incoming argument stack.
1854 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1855 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1856 const TargetInstrInfo *TII) {
1857 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1859 if (Arg.getOpcode() == ISD::CopyFromReg) {
1860 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1861 if (!TargetRegisterInfo::isVirtualRegister(VR))
1863 MachineInstr *Def = MRI->getVRegDef(VR);
1866 if (!Flags.isByVal()) {
1867 if (!TII->isLoadFromStackSlot(Def, FI))
1872 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
1873 if (Flags.isByVal())
1874 // ByVal argument is passed in as a pointer but it's now being
1875 // dereferenced. e.g.
1876 // define @foo(%struct.X* %A) {
1877 // tail call @bar(%struct.X* byval %A)
1880 SDValue Ptr = Ld->getBasePtr();
1881 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
1884 FI = FINode->getIndex();
1888 assert(FI != INT_MAX);
1889 if (!MFI->isFixedObjectIndex(FI))
1891 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
1894 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
1895 /// for tail call optimization. Targets which want to do tail call
1896 /// optimization should implement this function.
1898 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1899 CallingConv::ID CalleeCC,
1901 bool isCalleeStructRet,
1902 bool isCallerStructRet,
1903 const SmallVectorImpl<ISD::OutputArg> &Outs,
1904 const SmallVectorImpl<SDValue> &OutVals,
1905 const SmallVectorImpl<ISD::InputArg> &Ins,
1906 SelectionDAG& DAG) const {
1907 const Function *CallerF = DAG.getMachineFunction().getFunction();
1908 CallingConv::ID CallerCC = CallerF->getCallingConv();
1909 bool CCMatch = CallerCC == CalleeCC;
1911 // Look for obvious safe cases to perform tail call optimization that do not
1912 // require ABI changes. This is what gcc calls sibcall.
1914 // Do not sibcall optimize vararg calls unless the call site is not passing
1916 if (isVarArg && !Outs.empty())
1919 // Exception-handling functions need a special set of instructions to indicate
1920 // a return to the hardware. Tail-calling another function would probably
1922 if (CallerF->hasFnAttribute("interrupt"))
1925 // Also avoid sibcall optimization if either caller or callee uses struct
1926 // return semantics.
1927 if (isCalleeStructRet || isCallerStructRet)
1930 // FIXME: Completely disable sibcall for Thumb1 since Thumb1RegisterInfo::
1931 // emitEpilogue is not ready for them. Thumb tail calls also use t2B, as
1932 // the Thumb1 16-bit unconditional branch doesn't have sufficient relocation
1933 // support in the assembler and linker to be used. This would need to be
1934 // fixed to fully support tail calls in Thumb1.
1936 // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take
1937 // LR. This means if we need to reload LR, it takes an extra instructions,
1938 // which outweighs the value of the tail call; but here we don't know yet
1939 // whether LR is going to be used. Probably the right approach is to
1940 // generate the tail call here and turn it back into CALL/RET in
1941 // emitEpilogue if LR is used.
1943 // Thumb1 PIC calls to external symbols use BX, so they can be tail calls,
1944 // but we need to make sure there are enough registers; the only valid
1945 // registers are the 4 used for parameters. We don't currently do this
1947 if (Subtarget->isThumb1Only())
1950 // If the calling conventions do not match, then we'd better make sure the
1951 // results are returned in the same way as what the caller expects.
1953 SmallVector<CCValAssign, 16> RVLocs1;
1954 ARMCCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(),
1955 getTargetMachine(), RVLocs1, *DAG.getContext(), Call);
1956 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
1958 SmallVector<CCValAssign, 16> RVLocs2;
1959 ARMCCState CCInfo2(CallerCC, false, DAG.getMachineFunction(),
1960 getTargetMachine(), RVLocs2, *DAG.getContext(), Call);
1961 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
1963 if (RVLocs1.size() != RVLocs2.size())
1965 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
1966 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
1968 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
1970 if (RVLocs1[i].isRegLoc()) {
1971 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
1974 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
1980 // If Caller's vararg or byval argument has been split between registers and
1981 // stack, do not perform tail call, since part of the argument is in caller's
1983 const ARMFunctionInfo *AFI_Caller = DAG.getMachineFunction().
1984 getInfo<ARMFunctionInfo>();
1985 if (AFI_Caller->getArgRegsSaveSize())
1988 // If the callee takes no arguments then go on to check the results of the
1990 if (!Outs.empty()) {
1991 // Check if stack adjustment is needed. For now, do not do this if any
1992 // argument is passed on the stack.
1993 SmallVector<CCValAssign, 16> ArgLocs;
1994 ARMCCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(),
1995 getTargetMachine(), ArgLocs, *DAG.getContext(), Call);
1996 CCInfo.AnalyzeCallOperands(Outs,
1997 CCAssignFnForNode(CalleeCC, false, isVarArg));
1998 if (CCInfo.getNextStackOffset()) {
1999 MachineFunction &MF = DAG.getMachineFunction();
2001 // Check if the arguments are already laid out in the right way as
2002 // the caller's fixed stack objects.
2003 MachineFrameInfo *MFI = MF.getFrameInfo();
2004 const MachineRegisterInfo *MRI = &MF.getRegInfo();
2005 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
2006 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
2008 ++i, ++realArgIdx) {
2009 CCValAssign &VA = ArgLocs[i];
2010 EVT RegVT = VA.getLocVT();
2011 SDValue Arg = OutVals[realArgIdx];
2012 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
2013 if (VA.getLocInfo() == CCValAssign::Indirect)
2015 if (VA.needsCustom()) {
2016 // f64 and vector types are split into multiple registers or
2017 // register/stack-slot combinations. The types will not match
2018 // the registers; give up on memory f64 refs until we figure
2019 // out what to do about this.
2022 if (!ArgLocs[++i].isRegLoc())
2024 if (RegVT == MVT::v2f64) {
2025 if (!ArgLocs[++i].isRegLoc())
2027 if (!ArgLocs[++i].isRegLoc())
2030 } else if (!VA.isRegLoc()) {
2031 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
2043 ARMTargetLowering::CanLowerReturn(CallingConv::ID CallConv,
2044 MachineFunction &MF, bool isVarArg,
2045 const SmallVectorImpl<ISD::OutputArg> &Outs,
2046 LLVMContext &Context) const {
2047 SmallVector<CCValAssign, 16> RVLocs;
2048 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(), RVLocs, Context);
2049 return CCInfo.CheckReturn(Outs, CCAssignFnForNode(CallConv, /*Return=*/true,
2053 static SDValue LowerInterruptReturn(SmallVectorImpl<SDValue> &RetOps,
2054 SDLoc DL, SelectionDAG &DAG) {
2055 const MachineFunction &MF = DAG.getMachineFunction();
2056 const Function *F = MF.getFunction();
2058 StringRef IntKind = F->getFnAttribute("interrupt").getValueAsString();
2060 // See ARM ARM v7 B1.8.3. On exception entry LR is set to a possibly offset
2061 // version of the "preferred return address". These offsets affect the return
2062 // instruction if this is a return from PL1 without hypervisor extensions.
2063 // IRQ/FIQ: +4 "subs pc, lr, #4"
2064 // SWI: 0 "subs pc, lr, #0"
2065 // ABORT: +4 "subs pc, lr, #4"
2066 // UNDEF: +4/+2 "subs pc, lr, #0"
2067 // UNDEF varies depending on where the exception came from ARM or Thumb
2068 // mode. Alongside GCC, we throw our hands up in disgust and pretend it's 0.
2071 if (IntKind == "" || IntKind == "IRQ" || IntKind == "FIQ" ||
2074 else if (IntKind == "SWI" || IntKind == "UNDEF")
2077 report_fatal_error("Unsupported interrupt attribute. If present, value "
2078 "must be one of: IRQ, FIQ, SWI, ABORT or UNDEF");
2080 RetOps.insert(RetOps.begin() + 1, DAG.getConstant(LROffset, MVT::i32, false));
2082 return DAG.getNode(ARMISD::INTRET_FLAG, DL, MVT::Other, RetOps);
2086 ARMTargetLowering::LowerReturn(SDValue Chain,
2087 CallingConv::ID CallConv, bool isVarArg,
2088 const SmallVectorImpl<ISD::OutputArg> &Outs,
2089 const SmallVectorImpl<SDValue> &OutVals,
2090 SDLoc dl, SelectionDAG &DAG) const {
2092 // CCValAssign - represent the assignment of the return value to a location.
2093 SmallVector<CCValAssign, 16> RVLocs;
2095 // CCState - Info about the registers and stack slots.
2096 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2097 getTargetMachine(), RVLocs, *DAG.getContext(), Call);
2099 // Analyze outgoing return values.
2100 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
2104 SmallVector<SDValue, 4> RetOps;
2105 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
2106 bool isLittleEndian = Subtarget->isLittle();
2108 // Copy the result values into the output registers.
2109 for (unsigned i = 0, realRVLocIdx = 0;
2111 ++i, ++realRVLocIdx) {
2112 CCValAssign &VA = RVLocs[i];
2113 assert(VA.isRegLoc() && "Can only return in registers!");
2115 SDValue Arg = OutVals[realRVLocIdx];
2117 switch (VA.getLocInfo()) {
2118 default: llvm_unreachable("Unknown loc info!");
2119 case CCValAssign::Full: break;
2120 case CCValAssign::BCvt:
2121 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
2125 if (VA.needsCustom()) {
2126 if (VA.getLocVT() == MVT::v2f64) {
2127 // Extract the first half and return it in two registers.
2128 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
2129 DAG.getConstant(0, MVT::i32));
2130 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
2131 DAG.getVTList(MVT::i32, MVT::i32), Half);
2133 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2134 HalfGPRs.getValue(isLittleEndian ? 0 : 1),
2136 Flag = Chain.getValue(1);
2137 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2138 VA = RVLocs[++i]; // skip ahead to next loc
2139 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2140 HalfGPRs.getValue(isLittleEndian ? 1 : 0),
2142 Flag = Chain.getValue(1);
2143 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2144 VA = RVLocs[++i]; // skip ahead to next loc
2146 // Extract the 2nd half and fall through to handle it as an f64 value.
2147 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
2148 DAG.getConstant(1, MVT::i32));
2150 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
2152 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
2153 DAG.getVTList(MVT::i32, MVT::i32), Arg);
2154 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2155 fmrrd.getValue(isLittleEndian ? 0 : 1),
2157 Flag = Chain.getValue(1);
2158 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2159 VA = RVLocs[++i]; // skip ahead to next loc
2160 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2161 fmrrd.getValue(isLittleEndian ? 1 : 0),
2164 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
2166 // Guarantee that all emitted copies are
2167 // stuck together, avoiding something bad.
2168 Flag = Chain.getValue(1);
2169 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2172 // Update chain and glue.
2175 RetOps.push_back(Flag);
2177 // CPUs which aren't M-class use a special sequence to return from
2178 // exceptions (roughly, any instruction setting pc and cpsr simultaneously,
2179 // though we use "subs pc, lr, #N").
2181 // M-class CPUs actually use a normal return sequence with a special
2182 // (hardware-provided) value in LR, so the normal code path works.
2183 if (DAG.getMachineFunction().getFunction()->hasFnAttribute("interrupt") &&
2184 !Subtarget->isMClass()) {
2185 if (Subtarget->isThumb1Only())
2186 report_fatal_error("interrupt attribute is not supported in Thumb1");
2187 return LowerInterruptReturn(RetOps, dl, DAG);
2190 return DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, RetOps);
2193 bool ARMTargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
2194 if (N->getNumValues() != 1)
2196 if (!N->hasNUsesOfValue(1, 0))
2199 SDValue TCChain = Chain;
2200 SDNode *Copy = *N->use_begin();
2201 if (Copy->getOpcode() == ISD::CopyToReg) {
2202 // If the copy has a glue operand, we conservatively assume it isn't safe to
2203 // perform a tail call.
2204 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
2206 TCChain = Copy->getOperand(0);
2207 } else if (Copy->getOpcode() == ARMISD::VMOVRRD) {
2208 SDNode *VMov = Copy;
2209 // f64 returned in a pair of GPRs.
2210 SmallPtrSet<SDNode*, 2> Copies;
2211 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
2213 if (UI->getOpcode() != ISD::CopyToReg)
2217 if (Copies.size() > 2)
2220 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
2222 SDValue UseChain = UI->getOperand(0);
2223 if (Copies.count(UseChain.getNode()))
2230 } else if (Copy->getOpcode() == ISD::BITCAST) {
2231 // f32 returned in a single GPR.
2232 if (!Copy->hasOneUse())
2234 Copy = *Copy->use_begin();
2235 if (Copy->getOpcode() != ISD::CopyToReg || !Copy->hasNUsesOfValue(1, 0))
2237 TCChain = Copy->getOperand(0);
2242 bool HasRet = false;
2243 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
2245 if (UI->getOpcode() != ARMISD::RET_FLAG &&
2246 UI->getOpcode() != ARMISD::INTRET_FLAG)
2258 bool ARMTargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
2259 if (!Subtarget->supportsTailCall())
2262 if (!CI->isTailCall() || getTargetMachine().Options.DisableTailCalls)
2265 return !Subtarget->isThumb1Only();
2268 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
2269 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
2270 // one of the above mentioned nodes. It has to be wrapped because otherwise
2271 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
2272 // be used to form addressing mode. These wrapped nodes will be selected
2274 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
2275 EVT PtrVT = Op.getValueType();
2276 // FIXME there is no actual debug info here
2278 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
2280 if (CP->isMachineConstantPoolEntry())
2281 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
2282 CP->getAlignment());
2284 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
2285 CP->getAlignment());
2286 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
2289 unsigned ARMTargetLowering::getJumpTableEncoding() const {
2290 return MachineJumpTableInfo::EK_Inline;
2293 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
2294 SelectionDAG &DAG) const {
2295 MachineFunction &MF = DAG.getMachineFunction();
2296 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2297 unsigned ARMPCLabelIndex = 0;
2299 EVT PtrVT = getPointerTy();
2300 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
2301 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2303 if (RelocM == Reloc::Static) {
2304 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
2306 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2307 ARMPCLabelIndex = AFI->createPICLabelUId();
2308 ARMConstantPoolValue *CPV =
2309 ARMConstantPoolConstant::Create(BA, ARMPCLabelIndex,
2310 ARMCP::CPBlockAddress, PCAdj);
2311 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2313 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
2314 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
2315 MachinePointerInfo::getConstantPool(),
2316 false, false, false, 0);
2317 if (RelocM == Reloc::Static)
2319 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2320 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
2323 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
2325 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
2326 SelectionDAG &DAG) const {
2328 EVT PtrVT = getPointerTy();
2329 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2330 MachineFunction &MF = DAG.getMachineFunction();
2331 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2332 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2333 ARMConstantPoolValue *CPV =
2334 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2335 ARMCP::CPValue, PCAdj, ARMCP::TLSGD, true);
2336 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2337 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
2338 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
2339 MachinePointerInfo::getConstantPool(),
2340 false, false, false, 0);
2341 SDValue Chain = Argument.getValue(1);
2343 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2344 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
2346 // call __tls_get_addr.
2349 Entry.Node = Argument;
2350 Entry.Ty = (Type *) Type::getInt32Ty(*DAG.getContext());
2351 Args.push_back(Entry);
2353 // FIXME: is there useful debug info available here?
2354 TargetLowering::CallLoweringInfo CLI(DAG);
2355 CLI.setDebugLoc(dl).setChain(Chain)
2356 .setCallee(CallingConv::C, Type::getInt32Ty(*DAG.getContext()),
2357 DAG.getExternalSymbol("__tls_get_addr", PtrVT), std::move(Args),
2360 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
2361 return CallResult.first;
2364 // Lower ISD::GlobalTLSAddress using the "initial exec" or
2365 // "local exec" model.
2367 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
2369 TLSModel::Model model) const {
2370 const GlobalValue *GV = GA->getGlobal();
2373 SDValue Chain = DAG.getEntryNode();
2374 EVT PtrVT = getPointerTy();
2375 // Get the Thread Pointer
2376 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2378 if (model == TLSModel::InitialExec) {
2379 MachineFunction &MF = DAG.getMachineFunction();
2380 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2381 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2382 // Initial exec model.
2383 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2384 ARMConstantPoolValue *CPV =
2385 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2386 ARMCP::CPValue, PCAdj, ARMCP::GOTTPOFF,
2388 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2389 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2390 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2391 MachinePointerInfo::getConstantPool(),
2392 false, false, false, 0);
2393 Chain = Offset.getValue(1);
2395 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2396 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
2398 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2399 MachinePointerInfo::getConstantPool(),
2400 false, false, false, 0);
2403 assert(model == TLSModel::LocalExec);
2404 ARMConstantPoolValue *CPV =
2405 ARMConstantPoolConstant::Create(GV, ARMCP::TPOFF);
2406 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2407 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2408 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2409 MachinePointerInfo::getConstantPool(),
2410 false, false, false, 0);
2413 // The address of the thread local variable is the add of the thread
2414 // pointer with the offset of the variable.
2415 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
2419 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
2420 // TODO: implement the "local dynamic" model
2421 assert(Subtarget->isTargetELF() &&
2422 "TLS not implemented for non-ELF targets");
2423 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
2425 TLSModel::Model model = getTargetMachine().getTLSModel(GA->getGlobal());
2428 case TLSModel::GeneralDynamic:
2429 case TLSModel::LocalDynamic:
2430 return LowerToTLSGeneralDynamicModel(GA, DAG);
2431 case TLSModel::InitialExec:
2432 case TLSModel::LocalExec:
2433 return LowerToTLSExecModels(GA, DAG, model);
2435 llvm_unreachable("bogus TLS model");
2438 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
2439 SelectionDAG &DAG) const {
2440 EVT PtrVT = getPointerTy();
2442 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2443 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2444 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
2445 ARMConstantPoolValue *CPV =
2446 ARMConstantPoolConstant::Create(GV,
2447 UseGOTOFF ? ARMCP::GOTOFF : ARMCP::GOT);
2448 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2449 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2450 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
2452 MachinePointerInfo::getConstantPool(),
2453 false, false, false, 0);
2454 SDValue Chain = Result.getValue(1);
2455 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
2456 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
2458 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
2459 MachinePointerInfo::getGOT(),
2460 false, false, false, 0);
2464 // If we have T2 ops, we can materialize the address directly via movt/movw
2465 // pair. This is always cheaper.
2466 if (Subtarget->useMovt(DAG.getMachineFunction())) {
2468 // FIXME: Once remat is capable of dealing with instructions with register
2469 // operands, expand this into two nodes.
2470 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2471 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2473 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2474 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2475 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2476 MachinePointerInfo::getConstantPool(),
2477 false, false, false, 0);
2481 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
2482 SelectionDAG &DAG) const {
2483 EVT PtrVT = getPointerTy();
2485 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2486 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2488 if (Subtarget->useMovt(DAG.getMachineFunction()))
2491 // FIXME: Once remat is capable of dealing with instructions with register
2492 // operands, expand this into multiple nodes
2494 RelocM == Reloc::PIC_ ? ARMISD::WrapperPIC : ARMISD::Wrapper;
2496 SDValue G = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, ARMII::MO_NONLAZY);
2497 SDValue Result = DAG.getNode(Wrapper, dl, PtrVT, G);
2499 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2500 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
2501 MachinePointerInfo::getGOT(), false, false, false, 0);
2505 SDValue ARMTargetLowering::LowerGlobalAddressWindows(SDValue Op,
2506 SelectionDAG &DAG) const {
2507 assert(Subtarget->isTargetWindows() && "non-Windows COFF is not supported");
2508 assert(Subtarget->useMovt(DAG.getMachineFunction()) &&
2509 "Windows on ARM expects to use movw/movt");
2511 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2512 const ARMII::TOF TargetFlags =
2513 (GV->hasDLLImportStorageClass() ? ARMII::MO_DLLIMPORT : ARMII::MO_NO_FLAG);
2514 EVT PtrVT = getPointerTy();
2520 // FIXME: Once remat is capable of dealing with instructions with register
2521 // operands, expand this into two nodes.
2522 Result = DAG.getNode(ARMISD::Wrapper, DL, PtrVT,
2523 DAG.getTargetGlobalAddress(GV, DL, PtrVT, /*Offset=*/0,
2525 if (GV->hasDLLImportStorageClass())
2526 Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), Result,
2527 MachinePointerInfo::getGOT(), false, false, false, 0);
2531 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
2532 SelectionDAG &DAG) const {
2533 assert(Subtarget->isTargetELF() &&
2534 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
2535 MachineFunction &MF = DAG.getMachineFunction();
2536 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2537 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2538 EVT PtrVT = getPointerTy();
2540 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2541 ARMConstantPoolValue *CPV =
2542 ARMConstantPoolSymbol::Create(*DAG.getContext(), "_GLOBAL_OFFSET_TABLE_",
2543 ARMPCLabelIndex, PCAdj);
2544 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2545 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2546 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2547 MachinePointerInfo::getConstantPool(),
2548 false, false, false, 0);
2549 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2550 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2554 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
2556 SDValue Val = DAG.getConstant(0, MVT::i32);
2557 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl,
2558 DAG.getVTList(MVT::i32, MVT::Other), Op.getOperand(0),
2559 Op.getOperand(1), Val);
2563 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
2565 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
2566 Op.getOperand(1), DAG.getConstant(0, MVT::i32));
2570 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
2571 const ARMSubtarget *Subtarget) const {
2572 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2575 default: return SDValue(); // Don't custom lower most intrinsics.
2576 case Intrinsic::arm_rbit: {
2577 assert(Op.getOperand(0).getValueType() == MVT::i32 &&
2578 "RBIT intrinsic must have i32 type!");
2579 return DAG.getNode(ARMISD::RBIT, dl, MVT::i32, Op.getOperand(0));
2581 case Intrinsic::arm_thread_pointer: {
2582 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2583 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2585 case Intrinsic::eh_sjlj_lsda: {
2586 MachineFunction &MF = DAG.getMachineFunction();
2587 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2588 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2589 EVT PtrVT = getPointerTy();
2590 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2592 unsigned PCAdj = (RelocM != Reloc::PIC_)
2593 ? 0 : (Subtarget->isThumb() ? 4 : 8);
2594 ARMConstantPoolValue *CPV =
2595 ARMConstantPoolConstant::Create(MF.getFunction(), ARMPCLabelIndex,
2596 ARMCP::CPLSDA, PCAdj);
2597 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2598 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2600 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2601 MachinePointerInfo::getConstantPool(),
2602 false, false, false, 0);
2604 if (RelocM == Reloc::PIC_) {
2605 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2606 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2610 case Intrinsic::arm_neon_vmulls:
2611 case Intrinsic::arm_neon_vmullu: {
2612 unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vmulls)
2613 ? ARMISD::VMULLs : ARMISD::VMULLu;
2614 return DAG.getNode(NewOpc, SDLoc(Op), Op.getValueType(),
2615 Op.getOperand(1), Op.getOperand(2));
2620 static SDValue LowerATOMIC_FENCE(SDValue Op, SelectionDAG &DAG,
2621 const ARMSubtarget *Subtarget) {
2622 // FIXME: handle "fence singlethread" more efficiently.
2624 if (!Subtarget->hasDataBarrier()) {
2625 // Some ARMv6 cpus can support data barriers with an mcr instruction.
2626 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2628 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2629 "Unexpected ISD::ATOMIC_FENCE encountered. Should be libcall!");
2630 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2631 DAG.getConstant(0, MVT::i32));
2634 ConstantSDNode *OrdN = cast<ConstantSDNode>(Op.getOperand(1));
2635 AtomicOrdering Ord = static_cast<AtomicOrdering>(OrdN->getZExtValue());
2636 unsigned Domain = ARM_MB::ISH;
2637 if (Subtarget->isMClass()) {
2638 // Only a full system barrier exists in the M-class architectures.
2639 Domain = ARM_MB::SY;
2640 } else if (Subtarget->isSwift() && Ord == Release) {
2641 // Swift happens to implement ISHST barriers in a way that's compatible with
2642 // Release semantics but weaker than ISH so we'd be fools not to use
2643 // it. Beware: other processors probably don't!
2644 Domain = ARM_MB::ISHST;
2647 return DAG.getNode(ISD::INTRINSIC_VOID, dl, MVT::Other, Op.getOperand(0),
2648 DAG.getConstant(Intrinsic::arm_dmb, MVT::i32),
2649 DAG.getConstant(Domain, MVT::i32));
2652 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG,
2653 const ARMSubtarget *Subtarget) {
2654 // ARM pre v5TE and Thumb1 does not have preload instructions.
2655 if (!(Subtarget->isThumb2() ||
2656 (!Subtarget->isThumb1Only() && Subtarget->hasV5TEOps())))
2657 // Just preserve the chain.
2658 return Op.getOperand(0);
2661 unsigned isRead = ~cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() & 1;
2663 (!Subtarget->hasV7Ops() || !Subtarget->hasMPExtension()))
2664 // ARMv7 with MP extension has PLDW.
2665 return Op.getOperand(0);
2667 unsigned isData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
2668 if (Subtarget->isThumb()) {
2670 isRead = ~isRead & 1;
2671 isData = ~isData & 1;
2674 return DAG.getNode(ARMISD::PRELOAD, dl, MVT::Other, Op.getOperand(0),
2675 Op.getOperand(1), DAG.getConstant(isRead, MVT::i32),
2676 DAG.getConstant(isData, MVT::i32));
2679 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
2680 MachineFunction &MF = DAG.getMachineFunction();
2681 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
2683 // vastart just stores the address of the VarArgsFrameIndex slot into the
2684 // memory location argument.
2686 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2687 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2688 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
2689 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
2690 MachinePointerInfo(SV), false, false, 0);
2694 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
2695 SDValue &Root, SelectionDAG &DAG,
2697 MachineFunction &MF = DAG.getMachineFunction();
2698 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2700 const TargetRegisterClass *RC;
2701 if (AFI->isThumb1OnlyFunction())
2702 RC = &ARM::tGPRRegClass;
2704 RC = &ARM::GPRRegClass;
2706 // Transform the arguments stored in physical registers into virtual ones.
2707 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2708 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2711 if (NextVA.isMemLoc()) {
2712 MachineFrameInfo *MFI = MF.getFrameInfo();
2713 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true);
2715 // Create load node to retrieve arguments from the stack.
2716 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2717 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
2718 MachinePointerInfo::getFixedStack(FI),
2719 false, false, false, 0);
2721 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
2722 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2724 if (!Subtarget->isLittle())
2725 std::swap (ArgValue, ArgValue2);
2726 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
2730 ARMTargetLowering::computeRegArea(CCState &CCInfo, MachineFunction &MF,
2731 unsigned InRegsParamRecordIdx,
2733 unsigned &ArgRegsSize,
2734 unsigned &ArgRegsSaveSize)
2737 if (InRegsParamRecordIdx < CCInfo.getInRegsParamsCount()) {
2738 unsigned RBegin, REnd;
2739 CCInfo.getInRegsParamInfo(InRegsParamRecordIdx, RBegin, REnd);
2740 NumGPRs = REnd - RBegin;
2742 unsigned int firstUnalloced;
2743 firstUnalloced = CCInfo.getFirstUnallocated(GPRArgRegs,
2744 sizeof(GPRArgRegs) /
2745 sizeof(GPRArgRegs[0]));
2746 NumGPRs = (firstUnalloced <= 3) ? (4 - firstUnalloced) : 0;
2749 unsigned Align = MF.getTarget().getFrameLowering()->getStackAlignment();
2750 ArgRegsSize = NumGPRs * 4;
2752 // If parameter is split between stack and GPRs...
2753 if (NumGPRs && Align > 4 &&
2754 (ArgRegsSize < ArgSize ||
2755 InRegsParamRecordIdx >= CCInfo.getInRegsParamsCount())) {
2756 // Add padding for part of param recovered from GPRs. For example,
2757 // if Align == 8, its last byte must be at address K*8 - 1.
2758 // We need to do it, since remained (stack) part of parameter has
2759 // stack alignment, and we need to "attach" "GPRs head" without gaps
2762 // |---- 8 bytes block ----| |---- 8 bytes block ----| |---- 8 bytes...
2763 // [ [padding] [GPRs head] ] [ Tail passed via stack ....
2765 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2767 OffsetToAlignment(ArgRegsSize + AFI->getArgRegsSaveSize(), Align);
2768 ArgRegsSaveSize = ArgRegsSize + Padding;
2770 // We don't need to extend regs save size for byval parameters if they
2771 // are passed via GPRs only.
2772 ArgRegsSaveSize = ArgRegsSize;
2775 // The remaining GPRs hold either the beginning of variable-argument
2776 // data, or the beginning of an aggregate passed by value (usually
2777 // byval). Either way, we allocate stack slots adjacent to the data
2778 // provided by our caller, and store the unallocated registers there.
2779 // If this is a variadic function, the va_list pointer will begin with
2780 // these values; otherwise, this reassembles a (byval) structure that
2781 // was split between registers and memory.
2782 // Return: The frame index registers were stored into.
2784 ARMTargetLowering::StoreByValRegs(CCState &CCInfo, SelectionDAG &DAG,
2785 SDLoc dl, SDValue &Chain,
2786 const Value *OrigArg,
2787 unsigned InRegsParamRecordIdx,
2788 unsigned OffsetFromOrigArg,
2792 unsigned ByValStoreOffset,
2793 unsigned TotalArgRegsSaveSize) const {
2795 // Currently, two use-cases possible:
2796 // Case #1. Non-var-args function, and we meet first byval parameter.
2797 // Setup first unallocated register as first byval register;
2798 // eat all remained registers
2799 // (these two actions are performed by HandleByVal method).
2800 // Then, here, we initialize stack frame with
2801 // "store-reg" instructions.
2802 // Case #2. Var-args function, that doesn't contain byval parameters.
2803 // The same: eat all remained unallocated registers,
2804 // initialize stack frame.
2806 MachineFunction &MF = DAG.getMachineFunction();
2807 MachineFrameInfo *MFI = MF.getFrameInfo();
2808 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2809 unsigned firstRegToSaveIndex, lastRegToSaveIndex;
2810 unsigned RBegin, REnd;
2811 if (InRegsParamRecordIdx < CCInfo.getInRegsParamsCount()) {
2812 CCInfo.getInRegsParamInfo(InRegsParamRecordIdx, RBegin, REnd);
2813 firstRegToSaveIndex = RBegin - ARM::R0;
2814 lastRegToSaveIndex = REnd - ARM::R0;
2816 firstRegToSaveIndex = CCInfo.getFirstUnallocated
2817 (GPRArgRegs, array_lengthof(GPRArgRegs));
2818 lastRegToSaveIndex = 4;
2821 unsigned ArgRegsSize, ArgRegsSaveSize;
2822 computeRegArea(CCInfo, MF, InRegsParamRecordIdx, ArgSize,
2823 ArgRegsSize, ArgRegsSaveSize);
2825 // Store any by-val regs to their spots on the stack so that they may be
2826 // loaded by deferencing the result of formal parameter pointer or va_next.
2827 // Note: once stack area for byval/varargs registers
2828 // was initialized, it can't be initialized again.
2829 if (ArgRegsSaveSize) {
2830 unsigned Padding = ArgRegsSaveSize - ArgRegsSize;
2833 assert(AFI->getStoredByValParamsPadding() == 0 &&
2834 "The only parameter may be padded.");
2835 AFI->setStoredByValParamsPadding(Padding);
2838 int FrameIndex = MFI->CreateFixedObject(ArgRegsSaveSize,
2841 (int64_t)TotalArgRegsSaveSize,
2843 SDValue FIN = DAG.getFrameIndex(FrameIndex, getPointerTy());
2845 MFI->CreateFixedObject(Padding,
2846 ArgOffset + ByValStoreOffset -
2847 (int64_t)ArgRegsSaveSize,
2851 SmallVector<SDValue, 4> MemOps;
2852 for (unsigned i = 0; firstRegToSaveIndex < lastRegToSaveIndex;
2853 ++firstRegToSaveIndex, ++i) {
2854 const TargetRegisterClass *RC;
2855 if (AFI->isThumb1OnlyFunction())
2856 RC = &ARM::tGPRRegClass;
2858 RC = &ARM::GPRRegClass;
2860 unsigned VReg = MF.addLiveIn(GPRArgRegs[firstRegToSaveIndex], RC);
2861 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2863 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2864 MachinePointerInfo(OrigArg, OffsetFromOrigArg + 4*i),
2866 MemOps.push_back(Store);
2867 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
2868 DAG.getConstant(4, getPointerTy()));
2871 AFI->setArgRegsSaveSize(ArgRegsSaveSize + AFI->getArgRegsSaveSize());
2873 if (!MemOps.empty())
2874 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOps);
2878 // We cannot allocate a zero-byte object for the first variadic argument,
2879 // so just make up a size.
2882 // This will point to the next argument passed via stack.
2883 return MFI->CreateFixedObject(
2884 ArgSize, ArgOffset, !ForceMutable);
2888 // Setup stack frame, the va_list pointer will start from.
2890 ARMTargetLowering::VarArgStyleRegisters(CCState &CCInfo, SelectionDAG &DAG,
2891 SDLoc dl, SDValue &Chain,
2893 unsigned TotalArgRegsSaveSize,
2894 bool ForceMutable) const {
2895 MachineFunction &MF = DAG.getMachineFunction();
2896 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2898 // Try to store any remaining integer argument regs
2899 // to their spots on the stack so that they may be loaded by deferencing
2900 // the result of va_next.
2901 // If there is no regs to be stored, just point address after last
2902 // argument passed via stack.
2904 StoreByValRegs(CCInfo, DAG, dl, Chain, nullptr,
2905 CCInfo.getInRegsParamsCount(), 0, ArgOffset, 0, ForceMutable,
2906 0, TotalArgRegsSaveSize);
2908 AFI->setVarArgsFrameIndex(FrameIndex);
2912 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
2913 CallingConv::ID CallConv, bool isVarArg,
2914 const SmallVectorImpl<ISD::InputArg>
2916 SDLoc dl, SelectionDAG &DAG,
2917 SmallVectorImpl<SDValue> &InVals)
2919 MachineFunction &MF = DAG.getMachineFunction();
2920 MachineFrameInfo *MFI = MF.getFrameInfo();
2922 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2924 // Assign locations to all of the incoming arguments.
2925 SmallVector<CCValAssign, 16> ArgLocs;
2926 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2927 getTargetMachine(), ArgLocs, *DAG.getContext(), Prologue);
2928 CCInfo.AnalyzeFormalArguments(Ins,
2929 CCAssignFnForNode(CallConv, /* Return*/ false,
2932 SmallVector<SDValue, 16> ArgValues;
2933 int lastInsIndex = -1;
2935 Function::const_arg_iterator CurOrigArg = MF.getFunction()->arg_begin();
2936 unsigned CurArgIdx = 0;
2938 // Initially ArgRegsSaveSize is zero.
2939 // Then we increase this value each time we meet byval parameter.
2940 // We also increase this value in case of varargs function.
2941 AFI->setArgRegsSaveSize(0);
2943 unsigned ByValStoreOffset = 0;
2944 unsigned TotalArgRegsSaveSize = 0;
2945 unsigned ArgRegsSaveSizeMaxAlign = 4;
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 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2953 CCValAssign &VA = ArgLocs[i];
2954 if (VA.isMemLoc()) {
2955 int index = VA.getValNo();
2956 if (index != lastInsIndex) {
2957 ISD::ArgFlagsTy Flags = Ins[index].Flags;
2958 if (Flags.isByVal()) {
2959 unsigned ExtraArgRegsSize;
2960 unsigned ExtraArgRegsSaveSize;
2961 computeRegArea(CCInfo, MF, CCInfo.getInRegsParamsProceed(),
2962 Flags.getByValSize(),
2963 ExtraArgRegsSize, ExtraArgRegsSaveSize);
2965 TotalArgRegsSaveSize += ExtraArgRegsSaveSize;
2966 if (Flags.getByValAlign() > ArgRegsSaveSizeMaxAlign)
2967 ArgRegsSaveSizeMaxAlign = Flags.getByValAlign();
2968 CCInfo.nextInRegsParam();
2970 lastInsIndex = index;
2974 CCInfo.rewindByValRegsInfo();
2977 unsigned ExtraArgRegsSize;
2978 unsigned ExtraArgRegsSaveSize;
2979 computeRegArea(CCInfo, MF, CCInfo.getInRegsParamsCount(), 0,
2980 ExtraArgRegsSize, ExtraArgRegsSaveSize);
2981 TotalArgRegsSaveSize += ExtraArgRegsSaveSize;
2983 // If the arg regs save area contains N-byte aligned values, the
2984 // bottom of it must be at least N-byte aligned.
2985 TotalArgRegsSaveSize = RoundUpToAlignment(TotalArgRegsSaveSize, ArgRegsSaveSizeMaxAlign);
2986 TotalArgRegsSaveSize = std::min(TotalArgRegsSaveSize, 16U);
2988 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2989 CCValAssign &VA = ArgLocs[i];
2990 std::advance(CurOrigArg, Ins[VA.getValNo()].OrigArgIndex - CurArgIdx);
2991 CurArgIdx = Ins[VA.getValNo()].OrigArgIndex;
2992 // Arguments stored in registers.
2993 if (VA.isRegLoc()) {
2994 EVT RegVT = VA.getLocVT();
2996 if (VA.needsCustom()) {
2997 // f64 and vector types are split up into multiple registers or
2998 // combinations of registers and stack slots.
2999 if (VA.getLocVT() == MVT::v2f64) {
3000 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
3002 VA = ArgLocs[++i]; // skip ahead to next loc
3004 if (VA.isMemLoc()) {
3005 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true);
3006 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
3007 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
3008 MachinePointerInfo::getFixedStack(FI),
3009 false, false, false, 0);
3011 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
3014 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
3015 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
3016 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
3017 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
3018 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
3020 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
3023 const TargetRegisterClass *RC;
3025 if (RegVT == MVT::f32)
3026 RC = &ARM::SPRRegClass;
3027 else if (RegVT == MVT::f64)
3028 RC = &ARM::DPRRegClass;
3029 else if (RegVT == MVT::v2f64)
3030 RC = &ARM::QPRRegClass;
3031 else if (RegVT == MVT::i32)
3032 RC = AFI->isThumb1OnlyFunction() ?
3033 (const TargetRegisterClass*)&ARM::tGPRRegClass :
3034 (const TargetRegisterClass*)&ARM::GPRRegClass;
3036 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
3038 // Transform the arguments in physical registers into virtual ones.
3039 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
3040 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
3043 // If this is an 8 or 16-bit value, it is really passed promoted
3044 // to 32 bits. Insert an assert[sz]ext to capture this, then
3045 // truncate to the right size.
3046 switch (VA.getLocInfo()) {
3047 default: llvm_unreachable("Unknown loc info!");
3048 case CCValAssign::Full: break;
3049 case CCValAssign::BCvt:
3050 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
3052 case CCValAssign::SExt:
3053 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
3054 DAG.getValueType(VA.getValVT()));
3055 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
3057 case CCValAssign::ZExt:
3058 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
3059 DAG.getValueType(VA.getValVT()));
3060 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
3064 InVals.push_back(ArgValue);
3066 } else { // VA.isRegLoc()
3069 assert(VA.isMemLoc());
3070 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
3072 int index = ArgLocs[i].getValNo();
3074 // Some Ins[] entries become multiple ArgLoc[] entries.
3075 // Process them only once.
3076 if (index != lastInsIndex)
3078 ISD::ArgFlagsTy Flags = Ins[index].Flags;
3079 // FIXME: For now, all byval parameter objects are marked mutable.
3080 // This can be changed with more analysis.
3081 // In case of tail call optimization mark all arguments mutable.
3082 // Since they could be overwritten by lowering of arguments in case of
3084 if (Flags.isByVal()) {
3085 unsigned CurByValIndex = CCInfo.getInRegsParamsProceed();
3087 ByValStoreOffset = RoundUpToAlignment(ByValStoreOffset, Flags.getByValAlign());
3088 int FrameIndex = StoreByValRegs(
3089 CCInfo, DAG, dl, Chain, CurOrigArg,
3091 Ins[VA.getValNo()].PartOffset,
3092 VA.getLocMemOffset(),
3093 Flags.getByValSize(),
3094 true /*force mutable frames*/,
3096 TotalArgRegsSaveSize);
3097 ByValStoreOffset += Flags.getByValSize();
3098 ByValStoreOffset = std::min(ByValStoreOffset, 16U);
3099 InVals.push_back(DAG.getFrameIndex(FrameIndex, getPointerTy()));
3100 CCInfo.nextInRegsParam();
3102 unsigned FIOffset = VA.getLocMemOffset();
3103 int FI = MFI->CreateFixedObject(VA.getLocVT().getSizeInBits()/8,
3106 // Create load nodes to retrieve arguments from the stack.
3107 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
3108 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
3109 MachinePointerInfo::getFixedStack(FI),
3110 false, false, false, 0));
3112 lastInsIndex = index;
3119 VarArgStyleRegisters(CCInfo, DAG, dl, Chain,
3120 CCInfo.getNextStackOffset(),
3121 TotalArgRegsSaveSize);
3123 AFI->setArgumentStackSize(CCInfo.getNextStackOffset());
3128 /// isFloatingPointZero - Return true if this is +0.0.
3129 static bool isFloatingPointZero(SDValue Op) {
3130 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
3131 return CFP->getValueAPF().isPosZero();
3132 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
3133 // Maybe this has already been legalized into the constant pool?
3134 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
3135 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
3136 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
3137 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
3138 return CFP->getValueAPF().isPosZero();
3144 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
3145 /// the given operands.
3147 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
3148 SDValue &ARMcc, SelectionDAG &DAG,
3150 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
3151 unsigned C = RHSC->getZExtValue();
3152 if (!isLegalICmpImmediate(C)) {
3153 // Constant does not fit, try adjusting it by one?
3158 if (C != 0x80000000 && isLegalICmpImmediate(C-1)) {
3159 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
3160 RHS = DAG.getConstant(C-1, MVT::i32);
3165 if (C != 0 && isLegalICmpImmediate(C-1)) {
3166 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
3167 RHS = DAG.getConstant(C-1, MVT::i32);
3172 if (C != 0x7fffffff && isLegalICmpImmediate(C+1)) {
3173 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
3174 RHS = DAG.getConstant(C+1, MVT::i32);
3179 if (C != 0xffffffff && isLegalICmpImmediate(C+1)) {
3180 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
3181 RHS = DAG.getConstant(C+1, MVT::i32);
3188 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3189 ARMISD::NodeType CompareType;
3192 CompareType = ARMISD::CMP;
3197 CompareType = ARMISD::CMPZ;
3200 ARMcc = DAG.getConstant(CondCode, MVT::i32);
3201 return DAG.getNode(CompareType, dl, MVT::Glue, LHS, RHS);
3204 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
3206 ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
3209 if (!isFloatingPointZero(RHS))
3210 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Glue, LHS, RHS);
3212 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Glue, LHS);
3213 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Glue, Cmp);
3216 /// duplicateCmp - Glue values can have only one use, so this function
3217 /// duplicates a comparison node.
3219 ARMTargetLowering::duplicateCmp(SDValue Cmp, SelectionDAG &DAG) const {
3220 unsigned Opc = Cmp.getOpcode();
3222 if (Opc == ARMISD::CMP || Opc == ARMISD::CMPZ)
3223 return DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
3225 assert(Opc == ARMISD::FMSTAT && "unexpected comparison operation");
3226 Cmp = Cmp.getOperand(0);
3227 Opc = Cmp.getOpcode();
3228 if (Opc == ARMISD::CMPFP)
3229 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
3231 assert(Opc == ARMISD::CMPFPw0 && "unexpected operand of FMSTAT");
3232 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0));
3234 return DAG.getNode(ARMISD::FMSTAT, DL, MVT::Glue, Cmp);
3237 std::pair<SDValue, SDValue>
3238 ARMTargetLowering::getARMXALUOOp(SDValue Op, SelectionDAG &DAG,
3239 SDValue &ARMcc) const {
3240 assert(Op.getValueType() == MVT::i32 && "Unsupported value type");
3242 SDValue Value, OverflowCmp;
3243 SDValue LHS = Op.getOperand(0);
3244 SDValue RHS = Op.getOperand(1);
3247 // FIXME: We are currently always generating CMPs because we don't support
3248 // generating CMN through the backend. This is not as good as the natural
3249 // CMP case because it causes a register dependency and cannot be folded
3252 switch (Op.getOpcode()) {
3254 llvm_unreachable("Unknown overflow instruction!");
3256 ARMcc = DAG.getConstant(ARMCC::VC, MVT::i32);
3257 Value = DAG.getNode(ISD::ADD, SDLoc(Op), Op.getValueType(), LHS, RHS);
3258 OverflowCmp = DAG.getNode(ARMISD::CMP, SDLoc(Op), MVT::Glue, Value, LHS);
3261 ARMcc = DAG.getConstant(ARMCC::HS, MVT::i32);
3262 Value = DAG.getNode(ISD::ADD, SDLoc(Op), Op.getValueType(), LHS, RHS);
3263 OverflowCmp = DAG.getNode(ARMISD::CMP, SDLoc(Op), MVT::Glue, Value, LHS);
3266 ARMcc = DAG.getConstant(ARMCC::VC, MVT::i32);
3267 Value = DAG.getNode(ISD::SUB, SDLoc(Op), Op.getValueType(), LHS, RHS);
3268 OverflowCmp = DAG.getNode(ARMISD::CMP, SDLoc(Op), MVT::Glue, LHS, RHS);
3271 ARMcc = DAG.getConstant(ARMCC::HS, MVT::i32);
3272 Value = DAG.getNode(ISD::SUB, SDLoc(Op), Op.getValueType(), LHS, RHS);
3273 OverflowCmp = DAG.getNode(ARMISD::CMP, SDLoc(Op), MVT::Glue, LHS, RHS);
3277 return std::make_pair(Value, OverflowCmp);
3282 ARMTargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) const {
3283 // Let legalize expand this if it isn't a legal type yet.
3284 if (!DAG.getTargetLoweringInfo().isTypeLegal(Op.getValueType()))
3287 SDValue Value, OverflowCmp;
3289 std::tie(Value, OverflowCmp) = getARMXALUOOp(Op, DAG, ARMcc);
3290 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3291 // We use 0 and 1 as false and true values.
3292 SDValue TVal = DAG.getConstant(1, MVT::i32);
3293 SDValue FVal = DAG.getConstant(0, MVT::i32);
3294 EVT VT = Op.getValueType();
3296 SDValue Overflow = DAG.getNode(ARMISD::CMOV, SDLoc(Op), VT, TVal, FVal,
3297 ARMcc, CCR, OverflowCmp);
3299 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
3300 return DAG.getNode(ISD::MERGE_VALUES, SDLoc(Op), VTs, Value, Overflow);
3304 SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
3305 SDValue Cond = Op.getOperand(0);
3306 SDValue SelectTrue = Op.getOperand(1);
3307 SDValue SelectFalse = Op.getOperand(2);
3309 unsigned Opc = Cond.getOpcode();
3311 if (Cond.getResNo() == 1 &&
3312 (Opc == ISD::SADDO || Opc == ISD::UADDO || Opc == ISD::SSUBO ||
3313 Opc == ISD::USUBO)) {
3314 if (!DAG.getTargetLoweringInfo().isTypeLegal(Cond->getValueType(0)))
3317 SDValue Value, OverflowCmp;
3319 std::tie(Value, OverflowCmp) = getARMXALUOOp(Cond, DAG, ARMcc);
3320 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3321 EVT VT = Op.getValueType();
3323 return DAG.getNode(ARMISD::CMOV, SDLoc(Op), VT, SelectTrue, SelectFalse,
3324 ARMcc, CCR, OverflowCmp);
3330 // (select (cmov 1, 0, cond), t, f) -> (cmov t, f, cond)
3331 // (select (cmov 0, 1, cond), t, f) -> (cmov f, t, cond)
3333 if (Cond.getOpcode() == ARMISD::CMOV && Cond.hasOneUse()) {
3334 const ConstantSDNode *CMOVTrue =
3335 dyn_cast<ConstantSDNode>(Cond.getOperand(0));
3336 const ConstantSDNode *CMOVFalse =
3337 dyn_cast<ConstantSDNode>(Cond.getOperand(1));
3339 if (CMOVTrue && CMOVFalse) {
3340 unsigned CMOVTrueVal = CMOVTrue->getZExtValue();
3341 unsigned CMOVFalseVal = CMOVFalse->getZExtValue();
3345 if (CMOVTrueVal == 1 && CMOVFalseVal == 0) {
3347 False = SelectFalse;
3348 } else if (CMOVTrueVal == 0 && CMOVFalseVal == 1) {
3353 if (True.getNode() && False.getNode()) {
3354 EVT VT = Op.getValueType();
3355 SDValue ARMcc = Cond.getOperand(2);
3356 SDValue CCR = Cond.getOperand(3);
3357 SDValue Cmp = duplicateCmp(Cond.getOperand(4), DAG);
3358 assert(True.getValueType() == VT);
3359 return DAG.getNode(ARMISD::CMOV, dl, VT, True, False, ARMcc, CCR, Cmp);
3364 // ARM's BooleanContents value is UndefinedBooleanContent. Mask out the
3365 // undefined bits before doing a full-word comparison with zero.
3366 Cond = DAG.getNode(ISD::AND, dl, Cond.getValueType(), Cond,
3367 DAG.getConstant(1, Cond.getValueType()));
3369 return DAG.getSelectCC(dl, Cond,
3370 DAG.getConstant(0, Cond.getValueType()),
3371 SelectTrue, SelectFalse, ISD::SETNE);
3374 static ISD::CondCode getInverseCCForVSEL(ISD::CondCode CC) {
3375 if (CC == ISD::SETNE)
3377 return ISD::getSetCCInverse(CC, true);
3380 static void checkVSELConstraints(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
3381 bool &swpCmpOps, bool &swpVselOps) {
3382 // Start by selecting the GE condition code for opcodes that return true for
3384 if (CC == ISD::SETUGE || CC == ISD::SETOGE || CC == ISD::SETOLE ||
3386 CondCode = ARMCC::GE;
3388 // and GT for opcodes that return false for 'equality'.
3389 else if (CC == ISD::SETUGT || CC == ISD::SETOGT || CC == ISD::SETOLT ||
3391 CondCode = ARMCC::GT;
3393 // Since we are constrained to GE/GT, if the opcode contains 'less', we need
3394 // to swap the compare operands.
3395 if (CC == ISD::SETOLE || CC == ISD::SETULE || CC == ISD::SETOLT ||
3399 // Both GT and GE are ordered comparisons, and return false for 'unordered'.
3400 // If we have an unordered opcode, we need to swap the operands to the VSEL
3401 // instruction (effectively negating the condition).
3403 // This also has the effect of swapping which one of 'less' or 'greater'
3404 // returns true, so we also swap the compare operands. It also switches
3405 // whether we return true for 'equality', so we compensate by picking the
3406 // opposite condition code to our original choice.
3407 if (CC == ISD::SETULE || CC == ISD::SETULT || CC == ISD::SETUGE ||
3408 CC == ISD::SETUGT) {
3409 swpCmpOps = !swpCmpOps;
3410 swpVselOps = !swpVselOps;
3411 CondCode = CondCode == ARMCC::GT ? ARMCC::GE : ARMCC::GT;
3414 // 'ordered' is 'anything but unordered', so use the VS condition code and
3415 // swap the VSEL operands.
3416 if (CC == ISD::SETO) {
3417 CondCode = ARMCC::VS;
3421 // 'unordered or not equal' is 'anything but equal', so use the EQ condition
3422 // code and swap the VSEL operands.
3423 if (CC == ISD::SETUNE) {
3424 CondCode = ARMCC::EQ;
3429 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
3430 EVT VT = Op.getValueType();
3431 SDValue LHS = Op.getOperand(0);
3432 SDValue RHS = Op.getOperand(1);
3433 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
3434 SDValue TrueVal = Op.getOperand(2);
3435 SDValue FalseVal = Op.getOperand(3);
3438 if (LHS.getValueType() == MVT::i32) {
3439 // Try to generate VSEL on ARMv8.
3440 // The VSEL instruction can't use all the usual ARM condition
3441 // codes: it only has two bits to select the condition code, so it's
3442 // constrained to use only GE, GT, VS and EQ.
3444 // To implement all the various ISD::SETXXX opcodes, we sometimes need to
3445 // swap the operands of the previous compare instruction (effectively
3446 // inverting the compare condition, swapping 'less' and 'greater') and
3447 // sometimes need to swap the operands to the VSEL (which inverts the
3448 // condition in the sense of firing whenever the previous condition didn't)
3449 if (getSubtarget()->hasFPARMv8() && (TrueVal.getValueType() == MVT::f32 ||
3450 TrueVal.getValueType() == MVT::f64)) {
3451 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3452 if (CondCode == ARMCC::LT || CondCode == ARMCC::LE ||
3453 CondCode == ARMCC::VC || CondCode == ARMCC::NE) {
3454 CC = getInverseCCForVSEL(CC);
3455 std::swap(TrueVal, FalseVal);
3460 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3461 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3462 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,
3466 ARMCC::CondCodes CondCode, CondCode2;
3467 FPCCToARMCC(CC, CondCode, CondCode2);
3469 // Try to generate VSEL on ARMv8.
3470 if (getSubtarget()->hasFPARMv8() && (TrueVal.getValueType() == MVT::f32 ||
3471 TrueVal.getValueType() == MVT::f64)) {
3472 // We can select VMAXNM/VMINNM from a compare followed by a select with the
3473 // same operands, as follows:
3474 // c = fcmp [ogt, olt, ugt, ult] a, b
3476 // We only do this in unsafe-fp-math, because signed zeros and NaNs are
3477 // handled differently than the original code sequence.
3478 if (getTargetMachine().Options.UnsafeFPMath && LHS == TrueVal &&
3480 if (CC == ISD::SETOGT || CC == ISD::SETUGT)
3481 return DAG.getNode(ARMISD::VMAXNM, dl, VT, TrueVal, FalseVal);
3482 if (CC == ISD::SETOLT || CC == ISD::SETULT)
3483 return DAG.getNode(ARMISD::VMINNM, dl, VT, TrueVal, FalseVal);
3486 bool swpCmpOps = false;
3487 bool swpVselOps = false;
3488 checkVSELConstraints(CC, CondCode, swpCmpOps, swpVselOps);
3490 if (CondCode == ARMCC::GT || CondCode == ARMCC::GE ||
3491 CondCode == ARMCC::VS || CondCode == ARMCC::EQ) {
3493 std::swap(LHS, RHS);
3495 std::swap(TrueVal, FalseVal);
3499 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
3500 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
3501 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3502 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
3504 if (CondCode2 != ARMCC::AL) {
3505 SDValue ARMcc2 = DAG.getConstant(CondCode2, MVT::i32);
3506 // FIXME: Needs another CMP because flag can have but one use.
3507 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
3508 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
3509 Result, TrueVal, ARMcc2, CCR, Cmp2);
3514 /// canChangeToInt - Given the fp compare operand, return true if it is suitable
3515 /// to morph to an integer compare sequence.
3516 static bool canChangeToInt(SDValue Op, bool &SeenZero,
3517 const ARMSubtarget *Subtarget) {
3518 SDNode *N = Op.getNode();
3519 if (!N->hasOneUse())
3520 // Otherwise it requires moving the value from fp to integer registers.
3522 if (!N->getNumValues())
3524 EVT VT = Op.getValueType();
3525 if (VT != MVT::f32 && !Subtarget->isFPBrccSlow())
3526 // f32 case is generally profitable. f64 case only makes sense when vcmpe +
3527 // vmrs are very slow, e.g. cortex-a8.
3530 if (isFloatingPointZero(Op)) {
3534 return ISD::isNormalLoad(N);
3537 static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) {
3538 if (isFloatingPointZero(Op))
3539 return DAG.getConstant(0, MVT::i32);
3541 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op))
3542 return DAG.getLoad(MVT::i32, SDLoc(Op),
3543 Ld->getChain(), Ld->getBasePtr(), Ld->getPointerInfo(),
3544 Ld->isVolatile(), Ld->isNonTemporal(),
3545 Ld->isInvariant(), Ld->getAlignment());
3547 llvm_unreachable("Unknown VFP cmp argument!");
3550 static void expandf64Toi32(SDValue Op, SelectionDAG &DAG,
3551 SDValue &RetVal1, SDValue &RetVal2) {
3552 if (isFloatingPointZero(Op)) {
3553 RetVal1 = DAG.getConstant(0, MVT::i32);
3554 RetVal2 = DAG.getConstant(0, MVT::i32);
3558 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) {
3559 SDValue Ptr = Ld->getBasePtr();
3560 RetVal1 = DAG.getLoad(MVT::i32, SDLoc(Op),
3561 Ld->getChain(), Ptr,
3562 Ld->getPointerInfo(),
3563 Ld->isVolatile(), Ld->isNonTemporal(),
3564 Ld->isInvariant(), Ld->getAlignment());
3566 EVT PtrType = Ptr.getValueType();
3567 unsigned NewAlign = MinAlign(Ld->getAlignment(), 4);
3568 SDValue NewPtr = DAG.getNode(ISD::ADD, SDLoc(Op),
3569 PtrType, Ptr, DAG.getConstant(4, PtrType));
3570 RetVal2 = DAG.getLoad(MVT::i32, SDLoc(Op),
3571 Ld->getChain(), NewPtr,
3572 Ld->getPointerInfo().getWithOffset(4),
3573 Ld->isVolatile(), Ld->isNonTemporal(),
3574 Ld->isInvariant(), NewAlign);
3578 llvm_unreachable("Unknown VFP cmp argument!");
3581 /// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some
3582 /// f32 and even f64 comparisons to integer ones.
3584 ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const {
3585 SDValue Chain = Op.getOperand(0);
3586 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3587 SDValue LHS = Op.getOperand(2);
3588 SDValue RHS = Op.getOperand(3);
3589 SDValue Dest = Op.getOperand(4);
3592 bool LHSSeenZero = false;
3593 bool LHSOk = canChangeToInt(LHS, LHSSeenZero, Subtarget);
3594 bool RHSSeenZero = false;
3595 bool RHSOk = canChangeToInt(RHS, RHSSeenZero, Subtarget);
3596 if (LHSOk && RHSOk && (LHSSeenZero || RHSSeenZero)) {
3597 // If unsafe fp math optimization is enabled and there are no other uses of
3598 // the CMP operands, and the condition code is EQ or NE, we can optimize it
3599 // to an integer comparison.
3600 if (CC == ISD::SETOEQ)
3602 else if (CC == ISD::SETUNE)
3605 SDValue Mask = DAG.getConstant(0x7fffffff, MVT::i32);
3607 if (LHS.getValueType() == MVT::f32) {
3608 LHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3609 bitcastf32Toi32(LHS, DAG), Mask);
3610 RHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3611 bitcastf32Toi32(RHS, DAG), Mask);
3612 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3613 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3614 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3615 Chain, Dest, ARMcc, CCR, Cmp);
3620 expandf64Toi32(LHS, DAG, LHS1, LHS2);
3621 expandf64Toi32(RHS, DAG, RHS1, RHS2);
3622 LHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, LHS2, Mask);
3623 RHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, RHS2, Mask);
3624 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3625 ARMcc = DAG.getConstant(CondCode, MVT::i32);
3626 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3627 SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest };
3628 return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops);
3634 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
3635 SDValue Chain = Op.getOperand(0);
3636 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3637 SDValue LHS = Op.getOperand(2);
3638 SDValue RHS = Op.getOperand(3);
3639 SDValue Dest = Op.getOperand(4);
3642 if (LHS.getValueType() == MVT::i32) {
3644 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3645 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3646 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3647 Chain, Dest, ARMcc, CCR, Cmp);
3650 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3652 if (getTargetMachine().Options.UnsafeFPMath &&
3653 (CC == ISD::SETEQ || CC == ISD::SETOEQ ||
3654 CC == ISD::SETNE || CC == ISD::SETUNE)) {
3655 SDValue Result = OptimizeVFPBrcond(Op, DAG);
3656 if (Result.getNode())
3660 ARMCC::CondCodes CondCode, CondCode2;
3661 FPCCToARMCC(CC, CondCode, CondCode2);
3663 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
3664 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
3665 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3666 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3667 SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp };
3668 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops);
3669 if (CondCode2 != ARMCC::AL) {
3670 ARMcc = DAG.getConstant(CondCode2, MVT::i32);
3671 SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) };
3672 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops);
3677 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
3678 SDValue Chain = Op.getOperand(0);
3679 SDValue Table = Op.getOperand(1);
3680 SDValue Index = Op.getOperand(2);
3683 EVT PTy = getPointerTy();
3684 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
3685 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
3686 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
3687 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
3688 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
3689 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
3690 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
3691 if (Subtarget->isThumb2()) {
3692 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
3693 // which does another jump to the destination. This also makes it easier
3694 // to translate it to TBB / TBH later.
3695 // FIXME: This might not work if the function is extremely large.
3696 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
3697 Addr, Op.getOperand(2), JTI, UId);
3699 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
3700 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
3701 MachinePointerInfo::getJumpTable(),
3702 false, false, false, 0);
3703 Chain = Addr.getValue(1);
3704 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
3705 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3707 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
3708 MachinePointerInfo::getJumpTable(),
3709 false, false, false, 0);
3710 Chain = Addr.getValue(1);
3711 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3715 static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3716 EVT VT = Op.getValueType();
3719 if (Op.getValueType().getVectorElementType() == MVT::i32) {
3720 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::f32)
3722 return DAG.UnrollVectorOp(Op.getNode());
3725 assert(Op.getOperand(0).getValueType() == MVT::v4f32 &&
3726 "Invalid type for custom lowering!");
3727 if (VT != MVT::v4i16)
3728 return DAG.UnrollVectorOp(Op.getNode());
3730 Op = DAG.getNode(Op.getOpcode(), dl, MVT::v4i32, Op.getOperand(0));
3731 return DAG.getNode(ISD::TRUNCATE, dl, VT, Op);
3734 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3735 EVT VT = Op.getValueType();
3737 return LowerVectorFP_TO_INT(Op, DAG);
3742 switch (Op.getOpcode()) {
3743 default: llvm_unreachable("Invalid opcode!");
3744 case ISD::FP_TO_SINT:
3745 Opc = ARMISD::FTOSI;
3747 case ISD::FP_TO_UINT:
3748 Opc = ARMISD::FTOUI;
3751 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
3752 return DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
3755 static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3756 EVT VT = Op.getValueType();
3759 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::i32) {
3760 if (VT.getVectorElementType() == MVT::f32)
3762 return DAG.UnrollVectorOp(Op.getNode());
3765 assert(Op.getOperand(0).getValueType() == MVT::v4i16 &&
3766 "Invalid type for custom lowering!");
3767 if (VT != MVT::v4f32)
3768 return DAG.UnrollVectorOp(Op.getNode());
3772 switch (Op.getOpcode()) {
3773 default: llvm_unreachable("Invalid opcode!");
3774 case ISD::SINT_TO_FP:
3775 CastOpc = ISD::SIGN_EXTEND;
3776 Opc = ISD::SINT_TO_FP;
3778 case ISD::UINT_TO_FP:
3779 CastOpc = ISD::ZERO_EXTEND;
3780 Opc = ISD::UINT_TO_FP;
3784 Op = DAG.getNode(CastOpc, dl, MVT::v4i32, Op.getOperand(0));
3785 return DAG.getNode(Opc, dl, VT, Op);
3788 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3789 EVT VT = Op.getValueType();
3791 return LowerVectorINT_TO_FP(Op, DAG);
3796 switch (Op.getOpcode()) {
3797 default: llvm_unreachable("Invalid opcode!");
3798 case ISD::SINT_TO_FP:
3799 Opc = ARMISD::SITOF;
3801 case ISD::UINT_TO_FP:
3802 Opc = ARMISD::UITOF;
3806 Op = DAG.getNode(ISD::BITCAST, dl, MVT::f32, Op.getOperand(0));
3807 return DAG.getNode(Opc, dl, VT, Op);
3810 SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
3811 // Implement fcopysign with a fabs and a conditional fneg.
3812 SDValue Tmp0 = Op.getOperand(0);
3813 SDValue Tmp1 = Op.getOperand(1);
3815 EVT VT = Op.getValueType();
3816 EVT SrcVT = Tmp1.getValueType();
3817 bool InGPR = Tmp0.getOpcode() == ISD::BITCAST ||
3818 Tmp0.getOpcode() == ARMISD::VMOVDRR;
3819 bool UseNEON = !InGPR && Subtarget->hasNEON();
3822 // Use VBSL to copy the sign bit.
3823 unsigned EncodedVal = ARM_AM::createNEONModImm(0x6, 0x80);
3824 SDValue Mask = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v2i32,
3825 DAG.getTargetConstant(EncodedVal, MVT::i32));
3826 EVT OpVT = (VT == MVT::f32) ? MVT::v2i32 : MVT::v1i64;
3828 Mask = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3829 DAG.getNode(ISD::BITCAST, dl, OpVT, Mask),
3830 DAG.getConstant(32, MVT::i32));
3831 else /*if (VT == MVT::f32)*/
3832 Tmp0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp0);
3833 if (SrcVT == MVT::f32) {
3834 Tmp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp1);
3836 Tmp1 = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3837 DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1),
3838 DAG.getConstant(32, MVT::i32));
3839 } else if (VT == MVT::f32)
3840 Tmp1 = DAG.getNode(ARMISD::VSHRu, dl, MVT::v1i64,
3841 DAG.getNode(ISD::BITCAST, dl, MVT::v1i64, Tmp1),
3842 DAG.getConstant(32, MVT::i32));
3843 Tmp0 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp0);
3844 Tmp1 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1);
3846 SDValue AllOnes = DAG.getTargetConstant(ARM_AM::createNEONModImm(0xe, 0xff),
3848 AllOnes = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v8i8, AllOnes);
3849 SDValue MaskNot = DAG.getNode(ISD::XOR, dl, OpVT, Mask,
3850 DAG.getNode(ISD::BITCAST, dl, OpVT, AllOnes));
3852 SDValue Res = DAG.getNode(ISD::OR, dl, OpVT,
3853 DAG.getNode(ISD::AND, dl, OpVT, Tmp1, Mask),
3854 DAG.getNode(ISD::AND, dl, OpVT, Tmp0, MaskNot));
3855 if (VT == MVT::f32) {
3856 Res = DAG.getNode(ISD::BITCAST, dl, MVT::v2f32, Res);
3857 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, Res,
3858 DAG.getConstant(0, MVT::i32));
3860 Res = DAG.getNode(ISD::BITCAST, dl, MVT::f64, Res);
3866 // Bitcast operand 1 to i32.
3867 if (SrcVT == MVT::f64)
3868 Tmp1 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
3870 Tmp1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp1);
3872 // Or in the signbit with integer operations.
3873 SDValue Mask1 = DAG.getConstant(0x80000000, MVT::i32);
3874 SDValue Mask2 = DAG.getConstant(0x7fffffff, MVT::i32);
3875 Tmp1 = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp1, Mask1);
3876 if (VT == MVT::f32) {
3877 Tmp0 = DAG.getNode(ISD::AND, dl, MVT::i32,
3878 DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp0), Mask2);
3879 return DAG.getNode(ISD::BITCAST, dl, MVT::f32,
3880 DAG.getNode(ISD::OR, dl, MVT::i32, Tmp0, Tmp1));
3883 // f64: Or the high part with signbit and then combine two parts.
3884 Tmp0 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
3886 SDValue Lo = Tmp0.getValue(0);
3887 SDValue Hi = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp0.getValue(1), Mask2);
3888 Hi = DAG.getNode(ISD::OR, dl, MVT::i32, Hi, Tmp1);
3889 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
3892 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
3893 MachineFunction &MF = DAG.getMachineFunction();
3894 MachineFrameInfo *MFI = MF.getFrameInfo();
3895 MFI->setReturnAddressIsTaken(true);
3897 if (verifyReturnAddressArgumentIsConstant(Op, DAG))
3900 EVT VT = Op.getValueType();
3902 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3904 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
3905 SDValue Offset = DAG.getConstant(4, MVT::i32);
3906 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
3907 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
3908 MachinePointerInfo(), false, false, false, 0);
3911 // Return LR, which contains the return address. Mark it an implicit live-in.
3912 unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32));
3913 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
3916 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
3917 const ARMBaseRegisterInfo &ARI =
3918 *static_cast<const ARMBaseRegisterInfo*>(RegInfo);
3919 MachineFunction &MF = DAG.getMachineFunction();
3920 MachineFrameInfo *MFI = MF.getFrameInfo();
3921 MFI->setFrameAddressIsTaken(true);
3923 EVT VT = Op.getValueType();
3924 SDLoc dl(Op); // FIXME probably not meaningful
3925 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3926 unsigned FrameReg = ARI.getFrameRegister(MF);
3927 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
3929 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
3930 MachinePointerInfo(),
3931 false, false, false, 0);
3935 // FIXME? Maybe this could be a TableGen attribute on some registers and
3936 // this table could be generated automatically from RegInfo.
3937 unsigned ARMTargetLowering::getRegisterByName(const char* RegName,
3939 unsigned Reg = StringSwitch<unsigned>(RegName)
3940 .Case("sp", ARM::SP)
3944 report_fatal_error("Invalid register name global variable");
3947 /// ExpandBITCAST - If the target supports VFP, this function is called to
3948 /// expand a bit convert where either the source or destination type is i64 to
3949 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
3950 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
3951 /// vectors), since the legalizer won't know what to do with that.
3952 static SDValue ExpandBITCAST(SDNode *N, SelectionDAG &DAG) {
3953 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3955 SDValue Op = N->getOperand(0);
3957 // This function is only supposed to be called for i64 types, either as the
3958 // source or destination of the bit convert.
3959 EVT SrcVT = Op.getValueType();
3960 EVT DstVT = N->getValueType(0);
3961 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
3962 "ExpandBITCAST called for non-i64 type");
3964 // Turn i64->f64 into VMOVDRR.
3965 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
3966 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
3967 DAG.getConstant(0, MVT::i32));
3968 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
3969 DAG.getConstant(1, MVT::i32));
3970 return DAG.getNode(ISD::BITCAST, dl, DstVT,
3971 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
3974 // Turn f64->i64 into VMOVRRD.
3975 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
3977 if (TLI.isBigEndian() && SrcVT.isVector() &&
3978 SrcVT.getVectorNumElements() > 1)
3979 Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
3980 DAG.getVTList(MVT::i32, MVT::i32),
3981 DAG.getNode(ARMISD::VREV64, dl, SrcVT, Op));
3983 Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
3984 DAG.getVTList(MVT::i32, MVT::i32), Op);
3985 // Merge the pieces into a single i64 value.
3986 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
3992 /// getZeroVector - Returns a vector of specified type with all zero elements.
3993 /// Zero vectors are used to represent vector negation and in those cases
3994 /// will be implemented with the NEON VNEG instruction. However, VNEG does
3995 /// not support i64 elements, so sometimes the zero vectors will need to be
3996 /// explicitly constructed. Regardless, use a canonical VMOV to create the
3998 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, SDLoc dl) {
3999 assert(VT.isVector() && "Expected a vector type");
4000 // The canonical modified immediate encoding of a zero vector is....0!
4001 SDValue EncodedVal = DAG.getTargetConstant(0, MVT::i32);
4002 EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
4003 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal);
4004 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4007 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
4008 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
4009 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
4010 SelectionDAG &DAG) const {
4011 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4012 EVT VT = Op.getValueType();
4013 unsigned VTBits = VT.getSizeInBits();
4015 SDValue ShOpLo = Op.getOperand(0);
4016 SDValue ShOpHi = Op.getOperand(1);
4017 SDValue ShAmt = Op.getOperand(2);
4019 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
4021 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
4023 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
4024 DAG.getConstant(VTBits, MVT::i32), ShAmt);
4025 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
4026 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
4027 DAG.getConstant(VTBits, MVT::i32));
4028 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
4029 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
4030 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
4032 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
4033 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
4035 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
4036 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc,
4039 SDValue Ops[2] = { Lo, Hi };
4040 return DAG.getMergeValues(Ops, dl);
4043 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
4044 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
4045 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
4046 SelectionDAG &DAG) const {
4047 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4048 EVT VT = Op.getValueType();
4049 unsigned VTBits = VT.getSizeInBits();
4051 SDValue ShOpLo = Op.getOperand(0);
4052 SDValue ShOpHi = Op.getOperand(1);
4053 SDValue ShAmt = Op.getOperand(2);
4056 assert(Op.getOpcode() == ISD::SHL_PARTS);
4057 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
4058 DAG.getConstant(VTBits, MVT::i32), ShAmt);
4059 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
4060 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
4061 DAG.getConstant(VTBits, MVT::i32));
4062 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
4063 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
4065 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
4066 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
4067 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
4069 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
4070 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc,
4073 SDValue Ops[2] = { Lo, Hi };
4074 return DAG.getMergeValues(Ops, dl);
4077 SDValue ARMTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
4078 SelectionDAG &DAG) const {
4079 // The rounding mode is in bits 23:22 of the FPSCR.
4080 // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0
4081 // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3)
4082 // so that the shift + and get folded into a bitfield extract.
4084 SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::i32,
4085 DAG.getConstant(Intrinsic::arm_get_fpscr,
4087 SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPSCR,
4088 DAG.getConstant(1U << 22, MVT::i32));
4089 SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds,
4090 DAG.getConstant(22, MVT::i32));
4091 return DAG.getNode(ISD::AND, dl, MVT::i32, RMODE,
4092 DAG.getConstant(3, MVT::i32));
4095 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
4096 const ARMSubtarget *ST) {
4097 EVT VT = N->getValueType(0);
4100 if (!ST->hasV6T2Ops())
4103 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
4104 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
4107 /// getCTPOP16BitCounts - Returns a v8i8/v16i8 vector containing the bit-count
4108 /// for each 16-bit element from operand, repeated. The basic idea is to
4109 /// leverage vcnt to get the 8-bit counts, gather and add the results.
4111 /// Trace for v4i16:
4112 /// input = [v0 v1 v2 v3 ] (vi 16-bit element)
4113 /// cast: N0 = [w0 w1 w2 w3 w4 w5 w6 w7] (v0 = [w0 w1], wi 8-bit element)
4114 /// vcnt: N1 = [b0 b1 b2 b3 b4 b5 b6 b7] (bi = bit-count of 8-bit element wi)
4115 /// vrev: N2 = [b1 b0 b3 b2 b5 b4 b7 b6]
4116 /// [b0 b1 b2 b3 b4 b5 b6 b7]
4117 /// +[b1 b0 b3 b2 b5 b4 b7 b6]
4118 /// N3=N1+N2 = [k0 k0 k1 k1 k2 k2 k3 k3] (k0 = b0+b1 = bit-count of 16-bit v0,
4119 /// vuzp: = [k0 k1 k2 k3 k0 k1 k2 k3] each ki is 8-bits)
4120 static SDValue getCTPOP16BitCounts(SDNode *N, SelectionDAG &DAG) {
4121 EVT VT = N->getValueType(0);
4124 EVT VT8Bit = VT.is64BitVector() ? MVT::v8i8 : MVT::v16i8;
4125 SDValue N0 = DAG.getNode(ISD::BITCAST, DL, VT8Bit, N->getOperand(0));
4126 SDValue N1 = DAG.getNode(ISD::CTPOP, DL, VT8Bit, N0);
4127 SDValue N2 = DAG.getNode(ARMISD::VREV16, DL, VT8Bit, N1);
4128 SDValue N3 = DAG.getNode(ISD::ADD, DL, VT8Bit, N1, N2);
4129 return DAG.getNode(ARMISD::VUZP, DL, VT8Bit, N3, N3);
4132 /// lowerCTPOP16BitElements - Returns a v4i16/v8i16 vector containing the
4133 /// bit-count for each 16-bit element from the operand. We need slightly
4134 /// different sequencing for v4i16 and v8i16 to stay within NEON's available
4135 /// 64/128-bit registers.
4137 /// Trace for v4i16:
4138 /// input = [v0 v1 v2 v3 ] (vi 16-bit element)
4139 /// v8i8: BitCounts = [k0 k1 k2 k3 k0 k1 k2 k3 ] (ki is the bit-count of vi)
4140 /// v8i16:Extended = [k0 k1 k2 k3 k0 k1 k2 k3 ]
4141 /// v4i16:Extracted = [k0 k1 k2 k3 ]
4142 static SDValue lowerCTPOP16BitElements(SDNode *N, SelectionDAG &DAG) {
4143 EVT VT = N->getValueType(0);
4146 SDValue BitCounts = getCTPOP16BitCounts(N, DAG);
4147 if (VT.is64BitVector()) {
4148 SDValue Extended = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v8i16, BitCounts);
4149 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i16, Extended,
4150 DAG.getIntPtrConstant(0));
4152 SDValue Extracted = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v8i8,
4153 BitCounts, DAG.getIntPtrConstant(0));
4154 return DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v8i16, Extracted);
4158 /// lowerCTPOP32BitElements - Returns a v2i32/v4i32 vector containing the
4159 /// bit-count for each 32-bit element from the operand. The idea here is
4160 /// to split the vector into 16-bit elements, leverage the 16-bit count
4161 /// routine, and then combine the results.
4163 /// Trace for v2i32 (v4i32 similar with Extracted/Extended exchanged):
4164 /// input = [v0 v1 ] (vi: 32-bit elements)
4165 /// Bitcast = [w0 w1 w2 w3 ] (wi: 16-bit elements, v0 = [w0 w1])
4166 /// Counts16 = [k0 k1 k2 k3 ] (ki: 16-bit elements, bit-count of wi)
4167 /// vrev: N0 = [k1 k0 k3 k2 ]
4169 /// N1 =+[k1 k0 k3 k2 ]
4171 /// N2 =+[k1 k3 k0 k2 ]
4173 /// Extended =+[k1 k3 k0 k2 ]
4175 /// Extracted=+[k1 k3 ]
4177 static SDValue lowerCTPOP32BitElements(SDNode *N, SelectionDAG &DAG) {
4178 EVT VT = N->getValueType(0);
4181 EVT VT16Bit = VT.is64BitVector() ? MVT::v4i16 : MVT::v8i16;
4183 SDValue Bitcast = DAG.getNode(ISD::BITCAST, DL, VT16Bit, N->getOperand(0));
4184 SDValue Counts16 = lowerCTPOP16BitElements(Bitcast.getNode(), DAG);
4185 SDValue N0 = DAG.getNode(ARMISD::VREV32, DL, VT16Bit, Counts16);
4186 SDValue N1 = DAG.getNode(ISD::ADD, DL, VT16Bit, Counts16, N0);
4187 SDValue N2 = DAG.getNode(ARMISD::VUZP, DL, VT16Bit, N1, N1);
4189 if (VT.is64BitVector()) {
4190 SDValue Extended = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v4i32, N2);
4191 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i32, Extended,
4192 DAG.getIntPtrConstant(0));
4194 SDValue Extracted = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i16, N2,
4195 DAG.getIntPtrConstant(0));
4196 return DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v4i32, Extracted);
4200 static SDValue LowerCTPOP(SDNode *N, SelectionDAG &DAG,
4201 const ARMSubtarget *ST) {
4202 EVT VT = N->getValueType(0);
4204 assert(ST->hasNEON() && "Custom ctpop lowering requires NEON.");
4205 assert((VT == MVT::v2i32 || VT == MVT::v4i32 ||
4206 VT == MVT::v4i16 || VT == MVT::v8i16) &&
4207 "Unexpected type for custom ctpop lowering");
4209 if (VT.getVectorElementType() == MVT::i32)
4210 return lowerCTPOP32BitElements(N, DAG);
4212 return lowerCTPOP16BitElements(N, DAG);
4215 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
4216 const ARMSubtarget *ST) {
4217 EVT VT = N->getValueType(0);
4223 // Lower vector shifts on NEON to use VSHL.
4224 assert(ST->hasNEON() && "unexpected vector shift");
4226 // Left shifts translate directly to the vshiftu intrinsic.
4227 if (N->getOpcode() == ISD::SHL)
4228 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
4229 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
4230 N->getOperand(0), N->getOperand(1));
4232 assert((N->getOpcode() == ISD::SRA ||
4233 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
4235 // NEON uses the same intrinsics for both left and right shifts. For
4236 // right shifts, the shift amounts are negative, so negate the vector of
4238 EVT ShiftVT = N->getOperand(1).getValueType();
4239 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
4240 getZeroVector(ShiftVT, DAG, dl),
4242 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
4243 Intrinsic::arm_neon_vshifts :
4244 Intrinsic::arm_neon_vshiftu);
4245 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
4246 DAG.getConstant(vshiftInt, MVT::i32),
4247 N->getOperand(0), NegatedCount);
4250 static SDValue Expand64BitShift(SDNode *N, SelectionDAG &DAG,
4251 const ARMSubtarget *ST) {
4252 EVT VT = N->getValueType(0);
4255 // We can get here for a node like i32 = ISD::SHL i32, i64
4259 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
4260 "Unknown shift to lower!");
4262 // We only lower SRA, SRL of 1 here, all others use generic lowering.
4263 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
4264 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
4267 // If we are in thumb mode, we don't have RRX.
4268 if (ST->isThumb1Only()) return SDValue();
4270 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
4271 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
4272 DAG.getConstant(0, MVT::i32));
4273 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
4274 DAG.getConstant(1, MVT::i32));
4276 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
4277 // captures the result into a carry flag.
4278 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
4279 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Glue), Hi);
4281 // The low part is an ARMISD::RRX operand, which shifts the carry in.
4282 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
4284 // Merge the pieces into a single i64 value.
4285 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
4288 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
4289 SDValue TmpOp0, TmpOp1;
4290 bool Invert = false;
4294 SDValue Op0 = Op.getOperand(0);
4295 SDValue Op1 = Op.getOperand(1);
4296 SDValue CC = Op.getOperand(2);
4297 EVT VT = Op.getValueType();
4298 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
4301 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
4302 switch (SetCCOpcode) {
4303 default: llvm_unreachable("Illegal FP comparison");
4305 case ISD::SETNE: Invert = true; // Fallthrough
4307 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
4309 case ISD::SETLT: Swap = true; // Fallthrough
4311 case ISD::SETGT: Opc = ARMISD::VCGT; break;
4313 case ISD::SETLE: Swap = true; // Fallthrough
4315 case ISD::SETGE: Opc = ARMISD::VCGE; break;
4316 case ISD::SETUGE: Swap = true; // Fallthrough
4317 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
4318 case ISD::SETUGT: Swap = true; // Fallthrough
4319 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
4320 case ISD::SETUEQ: Invert = true; // Fallthrough
4322 // Expand this to (OLT | OGT).
4326 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
4327 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
4329 case ISD::SETUO: Invert = true; // Fallthrough
4331 // Expand this to (OLT | OGE).
4335 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
4336 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
4340 // Integer comparisons.
4341 switch (SetCCOpcode) {
4342 default: llvm_unreachable("Illegal integer comparison");
4343 case ISD::SETNE: Invert = true;
4344 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
4345 case ISD::SETLT: Swap = true;
4346 case ISD::SETGT: Opc = ARMISD::VCGT; break;
4347 case ISD::SETLE: Swap = true;
4348 case ISD::SETGE: Opc = ARMISD::VCGE; break;
4349 case ISD::SETULT: Swap = true;
4350 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
4351 case ISD::SETULE: Swap = true;
4352 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
4355 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
4356 if (Opc == ARMISD::VCEQ) {
4359 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
4361 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
4364 // Ignore bitconvert.
4365 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BITCAST)
4366 AndOp = AndOp.getOperand(0);
4368 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
4370 Op0 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(0));
4371 Op1 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(1));
4378 std::swap(Op0, Op1);
4380 // If one of the operands is a constant vector zero, attempt to fold the
4381 // comparison to a specialized compare-against-zero form.
4383 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
4385 else if (ISD::isBuildVectorAllZeros(Op0.getNode())) {
4386 if (Opc == ARMISD::VCGE)
4387 Opc = ARMISD::VCLEZ;
4388 else if (Opc == ARMISD::VCGT)
4389 Opc = ARMISD::VCLTZ;
4394 if (SingleOp.getNode()) {
4397 Result = DAG.getNode(ARMISD::VCEQZ, dl, VT, SingleOp); break;
4399 Result = DAG.getNode(ARMISD::VCGEZ, dl, VT, SingleOp); break;
4401 Result = DAG.getNode(ARMISD::VCLEZ, dl, VT, SingleOp); break;
4403 Result = DAG.getNode(ARMISD::VCGTZ, dl, VT, SingleOp); break;
4405 Result = DAG.getNode(ARMISD::VCLTZ, dl, VT, SingleOp); break;
4407 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
4410 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
4414 Result = DAG.getNOT(dl, Result, VT);
4419 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
4420 /// valid vector constant for a NEON instruction with a "modified immediate"
4421 /// operand (e.g., VMOV). If so, return the encoded value.
4422 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
4423 unsigned SplatBitSize, SelectionDAG &DAG,
4424 EVT &VT, bool is128Bits, NEONModImmType type) {
4425 unsigned OpCmode, Imm;
4427 // SplatBitSize is set to the smallest size that splats the vector, so a
4428 // zero vector will always have SplatBitSize == 8. However, NEON modified
4429 // immediate instructions others than VMOV do not support the 8-bit encoding
4430 // of a zero vector, and the default encoding of zero is supposed to be the
4435 switch (SplatBitSize) {
4437 if (type != VMOVModImm)
4439 // Any 1-byte value is OK. Op=0, Cmode=1110.
4440 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
4443 VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
4447 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
4448 VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
4449 if ((SplatBits & ~0xff) == 0) {
4450 // Value = 0x00nn: Op=x, Cmode=100x.
4455 if ((SplatBits & ~0xff00) == 0) {
4456 // Value = 0xnn00: Op=x, Cmode=101x.
4458 Imm = SplatBits >> 8;
4464 // NEON's 32-bit VMOV supports splat values where:
4465 // * only one byte is nonzero, or
4466 // * the least significant byte is 0xff and the second byte is nonzero, or
4467 // * the least significant 2 bytes are 0xff and the third is nonzero.
4468 VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
4469 if ((SplatBits & ~0xff) == 0) {
4470 // Value = 0x000000nn: Op=x, Cmode=000x.
4475 if ((SplatBits & ~0xff00) == 0) {
4476 // Value = 0x0000nn00: Op=x, Cmode=001x.
4478 Imm = SplatBits >> 8;
4481 if ((SplatBits & ~0xff0000) == 0) {
4482 // Value = 0x00nn0000: Op=x, Cmode=010x.
4484 Imm = SplatBits >> 16;
4487 if ((SplatBits & ~0xff000000) == 0) {
4488 // Value = 0xnn000000: Op=x, Cmode=011x.
4490 Imm = SplatBits >> 24;
4494 // cmode == 0b1100 and cmode == 0b1101 are not supported for VORR or VBIC
4495 if (type == OtherModImm) return SDValue();
4497 if ((SplatBits & ~0xffff) == 0 &&
4498 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
4499 // Value = 0x0000nnff: Op=x, Cmode=1100.
4501 Imm = SplatBits >> 8;
4505 if ((SplatBits & ~0xffffff) == 0 &&
4506 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
4507 // Value = 0x00nnffff: Op=x, Cmode=1101.
4509 Imm = SplatBits >> 16;
4513 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
4514 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
4515 // VMOV.I32. A (very) minor optimization would be to replicate the value
4516 // and fall through here to test for a valid 64-bit splat. But, then the
4517 // caller would also need to check and handle the change in size.
4521 if (type != VMOVModImm)
4523 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
4524 uint64_t BitMask = 0xff;
4526 unsigned ImmMask = 1;
4528 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
4529 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
4532 } else if ((SplatBits & BitMask) != 0) {
4539 if (DAG.getTargetLoweringInfo().isBigEndian())
4540 // swap higher and lower 32 bit word
4541 Imm = ((Imm & 0xf) << 4) | ((Imm & 0xf0) >> 4);
4543 // Op=1, Cmode=1110.
4545 VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
4550 llvm_unreachable("unexpected size for isNEONModifiedImm");
4553 unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm);
4554 return DAG.getTargetConstant(EncodedVal, MVT::i32);
4557 SDValue ARMTargetLowering::LowerConstantFP(SDValue Op, SelectionDAG &DAG,
4558 const ARMSubtarget *ST) const {
4562 bool IsDouble = Op.getValueType() == MVT::f64;
4563 ConstantFPSDNode *CFP = cast<ConstantFPSDNode>(Op);
4565 // Try splatting with a VMOV.f32...
4566 APFloat FPVal = CFP->getValueAPF();
4567 int ImmVal = IsDouble ? ARM_AM::getFP64Imm(FPVal) : ARM_AM::getFP32Imm(FPVal);
4570 if (IsDouble || !ST->useNEONForSinglePrecisionFP()) {
4571 // We have code in place to select a valid ConstantFP already, no need to
4576 // It's a float and we are trying to use NEON operations where
4577 // possible. Lower it to a splat followed by an extract.
4579 SDValue NewVal = DAG.getTargetConstant(ImmVal, MVT::i32);
4580 SDValue VecConstant = DAG.getNode(ARMISD::VMOVFPIMM, DL, MVT::v2f32,
4582 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecConstant,
4583 DAG.getConstant(0, MVT::i32));
4586 // The rest of our options are NEON only, make sure that's allowed before
4588 if (!ST->hasNEON() || (!IsDouble && !ST->useNEONForSinglePrecisionFP()))
4592 uint64_t iVal = FPVal.bitcastToAPInt().getZExtValue();
4594 // It wouldn't really be worth bothering for doubles except for one very
4595 // important value, which does happen to match: 0.0. So make sure we don't do
4597 if (IsDouble && (iVal & 0xffffffff) != (iVal >> 32))
4600 // Try a VMOV.i32 (FIXME: i8, i16, or i64 could work too).
4601 SDValue NewVal = isNEONModifiedImm(iVal & 0xffffffffU, 0, 32, DAG, VMovVT,
4603 if (NewVal != SDValue()) {
4605 SDValue VecConstant = DAG.getNode(ARMISD::VMOVIMM, DL, VMovVT,
4608 return DAG.getNode(ISD::BITCAST, DL, MVT::f64, VecConstant);
4610 // It's a float: cast and extract a vector element.
4611 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
4613 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
4614 DAG.getConstant(0, MVT::i32));
4617 // Finally, try a VMVN.i32
4618 NewVal = isNEONModifiedImm(~iVal & 0xffffffffU, 0, 32, DAG, VMovVT,
4620 if (NewVal != SDValue()) {
4622 SDValue VecConstant = DAG.getNode(ARMISD::VMVNIMM, DL, VMovVT, NewVal);
4625 return DAG.getNode(ISD::BITCAST, DL, MVT::f64, VecConstant);
4627 // It's a float: cast and extract a vector element.
4628 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
4630 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
4631 DAG.getConstant(0, MVT::i32));
4637 // check if an VEXT instruction can handle the shuffle mask when the
4638 // vector sources of the shuffle are the same.
4639 static bool isSingletonVEXTMask(ArrayRef<int> M, EVT VT, unsigned &Imm) {
4640 unsigned NumElts = VT.getVectorNumElements();
4642 // Assume that the first shuffle index is not UNDEF. Fail if it is.
4648 // If this is a VEXT shuffle, the immediate value is the index of the first
4649 // element. The other shuffle indices must be the successive elements after
4651 unsigned ExpectedElt = Imm;
4652 for (unsigned i = 1; i < NumElts; ++i) {
4653 // Increment the expected index. If it wraps around, just follow it
4654 // back to index zero and keep going.
4656 if (ExpectedElt == NumElts)
4659 if (M[i] < 0) continue; // ignore UNDEF indices
4660 if (ExpectedElt != static_cast<unsigned>(M[i]))
4668 static bool isVEXTMask(ArrayRef<int> M, EVT VT,
4669 bool &ReverseVEXT, unsigned &Imm) {
4670 unsigned NumElts = VT.getVectorNumElements();
4671 ReverseVEXT = false;
4673 // Assume that the first shuffle index is not UNDEF. Fail if it is.
4679 // If this is a VEXT shuffle, the immediate value is the index of the first
4680 // element. The other shuffle indices must be the successive elements after
4682 unsigned ExpectedElt = Imm;
4683 for (unsigned i = 1; i < NumElts; ++i) {
4684 // Increment the expected index. If it wraps around, it may still be
4685 // a VEXT but the source vectors must be swapped.
4687 if (ExpectedElt == NumElts * 2) {
4692 if (M[i] < 0) continue; // ignore UNDEF indices
4693 if (ExpectedElt != static_cast<unsigned>(M[i]))
4697 // Adjust the index value if the source operands will be swapped.
4704 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
4705 /// instruction with the specified blocksize. (The order of the elements
4706 /// within each block of the vector is reversed.)
4707 static bool isVREVMask(ArrayRef<int> M, EVT VT, unsigned BlockSize) {
4708 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
4709 "Only possible block sizes for VREV are: 16, 32, 64");
4711 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4715 unsigned NumElts = VT.getVectorNumElements();
4716 unsigned BlockElts = M[0] + 1;
4717 // If the first shuffle index is UNDEF, be optimistic.
4719 BlockElts = BlockSize / EltSz;
4721 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
4724 for (unsigned i = 0; i < NumElts; ++i) {
4725 if (M[i] < 0) continue; // ignore UNDEF indices
4726 if ((unsigned) M[i] != (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
4733 static bool isVTBLMask(ArrayRef<int> M, EVT VT) {
4734 // We can handle <8 x i8> vector shuffles. If the index in the mask is out of
4735 // range, then 0 is placed into the resulting vector. So pretty much any mask
4736 // of 8 elements can work here.
4737 return VT == MVT::v8i8 && M.size() == 8;
4740 static bool isVTRNMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4741 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4745 unsigned NumElts = VT.getVectorNumElements();
4746 WhichResult = (M[0] == 0 ? 0 : 1);
4747 for (unsigned i = 0; i < NumElts; i += 2) {
4748 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
4749 (M[i+1] >= 0 && (unsigned) M[i+1] != i + NumElts + WhichResult))
4755 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
4756 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4757 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
4758 static bool isVTRN_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4759 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4763 unsigned NumElts = VT.getVectorNumElements();
4764 WhichResult = (M[0] == 0 ? 0 : 1);
4765 for (unsigned i = 0; i < NumElts; i += 2) {
4766 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
4767 (M[i+1] >= 0 && (unsigned) M[i+1] != i + WhichResult))
4773 static bool isVUZPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4774 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4778 unsigned NumElts = VT.getVectorNumElements();
4779 WhichResult = (M[0] == 0 ? 0 : 1);
4780 for (unsigned i = 0; i != NumElts; ++i) {
4781 if (M[i] < 0) continue; // ignore UNDEF indices
4782 if ((unsigned) M[i] != 2 * i + WhichResult)
4786 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4787 if (VT.is64BitVector() && EltSz == 32)
4793 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
4794 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4795 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
4796 static bool isVUZP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4797 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4801 unsigned Half = VT.getVectorNumElements() / 2;
4802 WhichResult = (M[0] == 0 ? 0 : 1);
4803 for (unsigned j = 0; j != 2; ++j) {
4804 unsigned Idx = WhichResult;
4805 for (unsigned i = 0; i != Half; ++i) {
4806 int MIdx = M[i + j * Half];
4807 if (MIdx >= 0 && (unsigned) MIdx != Idx)
4813 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4814 if (VT.is64BitVector() && EltSz == 32)
4820 static bool isVZIPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4821 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4825 unsigned NumElts = VT.getVectorNumElements();
4826 WhichResult = (M[0] == 0 ? 0 : 1);
4827 unsigned Idx = WhichResult * NumElts / 2;
4828 for (unsigned i = 0; i != NumElts; i += 2) {
4829 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
4830 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx + NumElts))
4835 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4836 if (VT.is64BitVector() && EltSz == 32)
4842 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
4843 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4844 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
4845 static bool isVZIP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4846 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4850 unsigned NumElts = VT.getVectorNumElements();
4851 WhichResult = (M[0] == 0 ? 0 : 1);
4852 unsigned Idx = WhichResult * NumElts / 2;
4853 for (unsigned i = 0; i != NumElts; i += 2) {
4854 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
4855 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx))
4860 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4861 if (VT.is64BitVector() && EltSz == 32)
4867 /// \return true if this is a reverse operation on an vector.
4868 static bool isReverseMask(ArrayRef<int> M, EVT VT) {
4869 unsigned NumElts = VT.getVectorNumElements();
4870 // Make sure the mask has the right size.
4871 if (NumElts != M.size())
4874 // Look for <15, ..., 3, -1, 1, 0>.
4875 for (unsigned i = 0; i != NumElts; ++i)
4876 if (M[i] >= 0 && M[i] != (int) (NumElts - 1 - i))
4882 // If N is an integer constant that can be moved into a register in one
4883 // instruction, return an SDValue of such a constant (will become a MOV
4884 // instruction). Otherwise return null.
4885 static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG,
4886 const ARMSubtarget *ST, SDLoc dl) {
4888 if (!isa<ConstantSDNode>(N))
4890 Val = cast<ConstantSDNode>(N)->getZExtValue();
4892 if (ST->isThumb1Only()) {
4893 if (Val <= 255 || ~Val <= 255)
4894 return DAG.getConstant(Val, MVT::i32);
4896 if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1)
4897 return DAG.getConstant(Val, MVT::i32);
4902 // If this is a case we can't handle, return null and let the default
4903 // expansion code take care of it.
4904 SDValue ARMTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
4905 const ARMSubtarget *ST) const {
4906 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
4908 EVT VT = Op.getValueType();
4910 APInt SplatBits, SplatUndef;
4911 unsigned SplatBitSize;
4913 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
4914 if (SplatBitSize <= 64) {
4915 // Check if an immediate VMOV works.
4917 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
4918 SplatUndef.getZExtValue(), SplatBitSize,
4919 DAG, VmovVT, VT.is128BitVector(),
4921 if (Val.getNode()) {
4922 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
4923 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4926 // Try an immediate VMVN.
4927 uint64_t NegatedImm = (~SplatBits).getZExtValue();
4928 Val = isNEONModifiedImm(NegatedImm,
4929 SplatUndef.getZExtValue(), SplatBitSize,
4930 DAG, VmovVT, VT.is128BitVector(),
4932 if (Val.getNode()) {
4933 SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
4934 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4937 // Use vmov.f32 to materialize other v2f32 and v4f32 splats.
4938 if ((VT == MVT::v2f32 || VT == MVT::v4f32) && SplatBitSize == 32) {
4939 int ImmVal = ARM_AM::getFP32Imm(SplatBits);
4941 SDValue Val = DAG.getTargetConstant(ImmVal, MVT::i32);
4942 return DAG.getNode(ARMISD::VMOVFPIMM, dl, VT, Val);
4948 // Scan through the operands to see if only one value is used.
4950 // As an optimisation, even if more than one value is used it may be more
4951 // profitable to splat with one value then change some lanes.
4953 // Heuristically we decide to do this if the vector has a "dominant" value,
4954 // defined as splatted to more than half of the lanes.
4955 unsigned NumElts = VT.getVectorNumElements();
4956 bool isOnlyLowElement = true;
4957 bool usesOnlyOneValue = true;
4958 bool hasDominantValue = false;
4959 bool isConstant = true;
4961 // Map of the number of times a particular SDValue appears in the
4963 DenseMap<SDValue, unsigned> ValueCounts;
4965 for (unsigned i = 0; i < NumElts; ++i) {
4966 SDValue V = Op.getOperand(i);
4967 if (V.getOpcode() == ISD::UNDEF)
4970 isOnlyLowElement = false;
4971 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
4974 ValueCounts.insert(std::make_pair(V, 0));
4975 unsigned &Count = ValueCounts[V];
4977 // Is this value dominant? (takes up more than half of the lanes)
4978 if (++Count > (NumElts / 2)) {
4979 hasDominantValue = true;
4983 if (ValueCounts.size() != 1)
4984 usesOnlyOneValue = false;
4985 if (!Value.getNode() && ValueCounts.size() > 0)
4986 Value = ValueCounts.begin()->first;
4988 if (ValueCounts.size() == 0)
4989 return DAG.getUNDEF(VT);
4991 // Loads are better lowered with insert_vector_elt/ARMISD::BUILD_VECTOR.
4992 // Keep going if we are hitting this case.
4993 if (isOnlyLowElement && !ISD::isNormalLoad(Value.getNode()))
4994 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
4996 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4998 // Use VDUP for non-constant splats. For f32 constant splats, reduce to
4999 // i32 and try again.
5000 if (hasDominantValue && EltSize <= 32) {
5004 // If we are VDUPing a value that comes directly from a vector, that will
5005 // cause an unnecessary move to and from a GPR, where instead we could
5006 // just use VDUPLANE. We can only do this if the lane being extracted
5007 // is at a constant index, as the VDUP from lane instructions only have
5008 // constant-index forms.
5009 if (Value->getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5010 isa<ConstantSDNode>(Value->getOperand(1))) {
5011 // We need to create a new undef vector to use for the VDUPLANE if the
5012 // size of the vector from which we get the value is different than the
5013 // size of the vector that we need to create. We will insert the element
5014 // such that the register coalescer will remove unnecessary copies.
5015 if (VT != Value->getOperand(0).getValueType()) {
5016 ConstantSDNode *constIndex;
5017 constIndex = dyn_cast<ConstantSDNode>(Value->getOperand(1));
5018 assert(constIndex && "The index is not a constant!");
5019 unsigned index = constIndex->getAPIntValue().getLimitedValue() %
5020 VT.getVectorNumElements();
5021 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
5022 DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DAG.getUNDEF(VT),
5023 Value, DAG.getConstant(index, MVT::i32)),
5024 DAG.getConstant(index, MVT::i32));
5026 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
5027 Value->getOperand(0), Value->getOperand(1));
5029 N = DAG.getNode(ARMISD::VDUP, dl, VT, Value);
5031 if (!usesOnlyOneValue) {
5032 // The dominant value was splatted as 'N', but we now have to insert
5033 // all differing elements.
5034 for (unsigned I = 0; I < NumElts; ++I) {
5035 if (Op.getOperand(I) == Value)
5037 SmallVector<SDValue, 3> Ops;
5039 Ops.push_back(Op.getOperand(I));
5040 Ops.push_back(DAG.getConstant(I, MVT::i32));
5041 N = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Ops);
5046 if (VT.getVectorElementType().isFloatingPoint()) {
5047 SmallVector<SDValue, 8> Ops;
5048 for (unsigned i = 0; i < NumElts; ++i)
5049 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, MVT::i32,
5051 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
5052 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, Ops);
5053 Val = LowerBUILD_VECTOR(Val, DAG, ST);
5055 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
5057 if (usesOnlyOneValue) {
5058 SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
5059 if (isConstant && Val.getNode())
5060 return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
5064 // If all elements are constants and the case above didn't get hit, fall back
5065 // to the default expansion, which will generate a load from the constant
5070 // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
5072 SDValue shuffle = ReconstructShuffle(Op, DAG);
5073 if (shuffle != SDValue())
5077 // Vectors with 32- or 64-bit elements can be built by directly assigning
5078 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
5079 // will be legalized.
5080 if (EltSize >= 32) {
5081 // Do the expansion with floating-point types, since that is what the VFP
5082 // registers are defined to use, and since i64 is not legal.
5083 EVT EltVT = EVT::getFloatingPointVT(EltSize);
5084 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
5085 SmallVector<SDValue, 8> Ops;
5086 for (unsigned i = 0; i < NumElts; ++i)
5087 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, EltVT, Op.getOperand(i)));
5088 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, Ops);
5089 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
5092 // If all else fails, just use a sequence of INSERT_VECTOR_ELT when we
5093 // know the default expansion would otherwise fall back on something even
5094 // worse. For a vector with one or two non-undef values, that's
5095 // scalar_to_vector for the elements followed by a shuffle (provided the
5096 // shuffle is valid for the target) and materialization element by element
5097 // on the stack followed by a load for everything else.
5098 if (!isConstant && !usesOnlyOneValue) {
5099 SDValue Vec = DAG.getUNDEF(VT);
5100 for (unsigned i = 0 ; i < NumElts; ++i) {
5101 SDValue V = Op.getOperand(i);
5102 if (V.getOpcode() == ISD::UNDEF)
5104 SDValue LaneIdx = DAG.getConstant(i, MVT::i32);
5105 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Vec, V, LaneIdx);
5113 // Gather data to see if the operation can be modelled as a
5114 // shuffle in combination with VEXTs.
5115 SDValue ARMTargetLowering::ReconstructShuffle(SDValue Op,
5116 SelectionDAG &DAG) const {
5118 EVT VT = Op.getValueType();
5119 unsigned NumElts = VT.getVectorNumElements();
5121 SmallVector<SDValue, 2> SourceVecs;
5122 SmallVector<unsigned, 2> MinElts;
5123 SmallVector<unsigned, 2> MaxElts;
5125 for (unsigned i = 0; i < NumElts; ++i) {
5126 SDValue V = Op.getOperand(i);
5127 if (V.getOpcode() == ISD::UNDEF)
5129 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
5130 // A shuffle can only come from building a vector from various
5131 // elements of other vectors.
5133 } else if (V.getOperand(0).getValueType().getVectorElementType() !=
5134 VT.getVectorElementType()) {
5135 // This code doesn't know how to handle shuffles where the vector
5136 // element types do not match (this happens because type legalization
5137 // promotes the return type of EXTRACT_VECTOR_ELT).
5138 // FIXME: It might be appropriate to extend this code to handle
5139 // mismatched types.
5143 // Record this extraction against the appropriate vector if possible...
5144 SDValue SourceVec = V.getOperand(0);
5145 // If the element number isn't a constant, we can't effectively
5146 // analyze what's going on.
5147 if (!isa<ConstantSDNode>(V.getOperand(1)))
5149 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
5150 bool FoundSource = false;
5151 for (unsigned j = 0; j < SourceVecs.size(); ++j) {
5152 if (SourceVecs[j] == SourceVec) {
5153 if (MinElts[j] > EltNo)
5155 if (MaxElts[j] < EltNo)
5162 // Or record a new source if not...
5164 SourceVecs.push_back(SourceVec);
5165 MinElts.push_back(EltNo);
5166 MaxElts.push_back(EltNo);
5170 // Currently only do something sane when at most two source vectors
5172 if (SourceVecs.size() > 2)
5175 SDValue ShuffleSrcs[2] = {DAG.getUNDEF(VT), DAG.getUNDEF(VT) };
5176 int VEXTOffsets[2] = {0, 0};
5178 // This loop extracts the usage patterns of the source vectors
5179 // and prepares appropriate SDValues for a shuffle if possible.
5180 for (unsigned i = 0; i < SourceVecs.size(); ++i) {
5181 if (SourceVecs[i].getValueType() == VT) {
5182 // No VEXT necessary
5183 ShuffleSrcs[i] = SourceVecs[i];
5186 } else if (SourceVecs[i].getValueType().getVectorNumElements() < NumElts) {
5187 // It probably isn't worth padding out a smaller vector just to
5188 // break it down again in a shuffle.
5192 // Since only 64-bit and 128-bit vectors are legal on ARM and
5193 // we've eliminated the other cases...
5194 assert(SourceVecs[i].getValueType().getVectorNumElements() == 2*NumElts &&
5195 "unexpected vector sizes in ReconstructShuffle");
5197 if (MaxElts[i] - MinElts[i] >= NumElts) {
5198 // Span too large for a VEXT to cope
5202 if (MinElts[i] >= NumElts) {
5203 // The extraction can just take the second half
5204 VEXTOffsets[i] = NumElts;
5205 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5207 DAG.getIntPtrConstant(NumElts));
5208 } else if (MaxElts[i] < NumElts) {
5209 // The extraction can just take the first half
5211 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5213 DAG.getIntPtrConstant(0));
5215 // An actual VEXT is needed
5216 VEXTOffsets[i] = MinElts[i];
5217 SDValue VEXTSrc1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5219 DAG.getIntPtrConstant(0));
5220 SDValue VEXTSrc2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5222 DAG.getIntPtrConstant(NumElts));
5223 ShuffleSrcs[i] = DAG.getNode(ARMISD::VEXT, dl, VT, VEXTSrc1, VEXTSrc2,
5224 DAG.getConstant(VEXTOffsets[i], MVT::i32));
5228 SmallVector<int, 8> Mask;
5230 for (unsigned i = 0; i < NumElts; ++i) {
5231 SDValue Entry = Op.getOperand(i);
5232 if (Entry.getOpcode() == ISD::UNDEF) {
5237 SDValue ExtractVec = Entry.getOperand(0);
5238 int ExtractElt = cast<ConstantSDNode>(Op.getOperand(i)
5239 .getOperand(1))->getSExtValue();
5240 if (ExtractVec == SourceVecs[0]) {
5241 Mask.push_back(ExtractElt - VEXTOffsets[0]);
5243 Mask.push_back(ExtractElt + NumElts - VEXTOffsets[1]);
5247 // Final check before we try to produce nonsense...
5248 if (isShuffleMaskLegal(Mask, VT))
5249 return DAG.getVectorShuffle(VT, dl, ShuffleSrcs[0], ShuffleSrcs[1],
5255 /// isShuffleMaskLegal - Targets can use this to indicate that they only
5256 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
5257 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
5258 /// are assumed to be legal.
5260 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
5262 if (VT.getVectorNumElements() == 4 &&
5263 (VT.is128BitVector() || VT.is64BitVector())) {
5264 unsigned PFIndexes[4];
5265 for (unsigned i = 0; i != 4; ++i) {
5269 PFIndexes[i] = M[i];
5272 // Compute the index in the perfect shuffle table.
5273 unsigned PFTableIndex =
5274 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
5275 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
5276 unsigned Cost = (PFEntry >> 30);
5283 unsigned Imm, WhichResult;
5285 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5286 return (EltSize >= 32 ||
5287 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
5288 isVREVMask(M, VT, 64) ||
5289 isVREVMask(M, VT, 32) ||
5290 isVREVMask(M, VT, 16) ||
5291 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
5292 isVTBLMask(M, VT) ||
5293 isVTRNMask(M, VT, WhichResult) ||
5294 isVUZPMask(M, VT, WhichResult) ||
5295 isVZIPMask(M, VT, WhichResult) ||
5296 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
5297 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
5298 isVZIP_v_undef_Mask(M, VT, WhichResult) ||
5299 ((VT == MVT::v8i16 || VT == MVT::v16i8) && isReverseMask(M, VT)));
5302 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
5303 /// the specified operations to build the shuffle.
5304 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
5305 SDValue RHS, SelectionDAG &DAG,
5307 unsigned OpNum = (PFEntry >> 26) & 0x0F;
5308 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
5309 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
5312 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
5321 OP_VUZPL, // VUZP, left result
5322 OP_VUZPR, // VUZP, right result
5323 OP_VZIPL, // VZIP, left result
5324 OP_VZIPR, // VZIP, right result
5325 OP_VTRNL, // VTRN, left result
5326 OP_VTRNR // VTRN, right result
5329 if (OpNum == OP_COPY) {
5330 if (LHSID == (1*9+2)*9+3) return LHS;
5331 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
5335 SDValue OpLHS, OpRHS;
5336 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
5337 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
5338 EVT VT = OpLHS.getValueType();
5341 default: llvm_unreachable("Unknown shuffle opcode!");
5343 // VREV divides the vector in half and swaps within the half.
5344 if (VT.getVectorElementType() == MVT::i32 ||
5345 VT.getVectorElementType() == MVT::f32)
5346 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
5347 // vrev <4 x i16> -> VREV32
5348 if (VT.getVectorElementType() == MVT::i16)
5349 return DAG.getNode(ARMISD::VREV32, dl, VT, OpLHS);
5350 // vrev <4 x i8> -> VREV16
5351 assert(VT.getVectorElementType() == MVT::i8);
5352 return DAG.getNode(ARMISD::VREV16, dl, VT, OpLHS);
5357 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
5358 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
5362 return DAG.getNode(ARMISD::VEXT, dl, VT,
5364 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
5367 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
5368 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
5371 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
5372 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
5375 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
5376 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
5380 static SDValue LowerVECTOR_SHUFFLEv8i8(SDValue Op,
5381 ArrayRef<int> ShuffleMask,
5382 SelectionDAG &DAG) {
5383 // Check to see if we can use the VTBL instruction.
5384 SDValue V1 = Op.getOperand(0);
5385 SDValue V2 = Op.getOperand(1);
5388 SmallVector<SDValue, 8> VTBLMask;
5389 for (ArrayRef<int>::iterator
5390 I = ShuffleMask.begin(), E = ShuffleMask.end(); I != E; ++I)
5391 VTBLMask.push_back(DAG.getConstant(*I, MVT::i32));
5393 if (V2.getNode()->getOpcode() == ISD::UNDEF)
5394 return DAG.getNode(ARMISD::VTBL1, DL, MVT::v8i8, V1,
5395 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8, VTBLMask));
5397 return DAG.getNode(ARMISD::VTBL2, DL, MVT::v8i8, V1, V2,
5398 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8, VTBLMask));
5401 static SDValue LowerReverse_VECTOR_SHUFFLEv16i8_v8i16(SDValue Op,
5402 SelectionDAG &DAG) {
5404 SDValue OpLHS = Op.getOperand(0);
5405 EVT VT = OpLHS.getValueType();
5407 assert((VT == MVT::v8i16 || VT == MVT::v16i8) &&
5408 "Expect an v8i16/v16i8 type");
5409 OpLHS = DAG.getNode(ARMISD::VREV64, DL, VT, OpLHS);
5410 // For a v16i8 type: After the VREV, we have got <8, ...15, 8, ..., 0>. Now,
5411 // extract the first 8 bytes into the top double word and the last 8 bytes
5412 // into the bottom double word. The v8i16 case is similar.
5413 unsigned ExtractNum = (VT == MVT::v16i8) ? 8 : 4;
5414 return DAG.getNode(ARMISD::VEXT, DL, VT, OpLHS, OpLHS,
5415 DAG.getConstant(ExtractNum, MVT::i32));
5418 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
5419 SDValue V1 = Op.getOperand(0);
5420 SDValue V2 = Op.getOperand(1);
5422 EVT VT = Op.getValueType();
5423 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
5425 // Convert shuffles that are directly supported on NEON to target-specific
5426 // DAG nodes, instead of keeping them as shuffles and matching them again
5427 // during code selection. This is more efficient and avoids the possibility
5428 // of inconsistencies between legalization and selection.
5429 // FIXME: floating-point vectors should be canonicalized to integer vectors
5430 // of the same time so that they get CSEd properly.
5431 ArrayRef<int> ShuffleMask = SVN->getMask();
5433 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5434 if (EltSize <= 32) {
5435 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
5436 int Lane = SVN->getSplatIndex();
5437 // If this is undef splat, generate it via "just" vdup, if possible.
5438 if (Lane == -1) Lane = 0;
5440 // Test if V1 is a SCALAR_TO_VECTOR.
5441 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
5442 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
5444 // Test if V1 is a BUILD_VECTOR which is equivalent to a SCALAR_TO_VECTOR
5445 // (and probably will turn into a SCALAR_TO_VECTOR once legalization
5447 if (Lane == 0 && V1.getOpcode() == ISD::BUILD_VECTOR &&
5448 !isa<ConstantSDNode>(V1.getOperand(0))) {
5449 bool IsScalarToVector = true;
5450 for (unsigned i = 1, e = V1.getNumOperands(); i != e; ++i)
5451 if (V1.getOperand(i).getOpcode() != ISD::UNDEF) {
5452 IsScalarToVector = false;
5455 if (IsScalarToVector)
5456 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
5458 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
5459 DAG.getConstant(Lane, MVT::i32));
5464 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
5467 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
5468 DAG.getConstant(Imm, MVT::i32));
5471 if (isVREVMask(ShuffleMask, VT, 64))
5472 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
5473 if (isVREVMask(ShuffleMask, VT, 32))
5474 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
5475 if (isVREVMask(ShuffleMask, VT, 16))
5476 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
5478 if (V2->getOpcode() == ISD::UNDEF &&
5479 isSingletonVEXTMask(ShuffleMask, VT, Imm)) {
5480 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V1,
5481 DAG.getConstant(Imm, MVT::i32));
5484 // Check for Neon shuffles that modify both input vectors in place.
5485 // If both results are used, i.e., if there are two shuffles with the same
5486 // source operands and with masks corresponding to both results of one of
5487 // these operations, DAG memoization will ensure that a single node is
5488 // used for both shuffles.
5489 unsigned WhichResult;
5490 if (isVTRNMask(ShuffleMask, VT, WhichResult))
5491 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
5492 V1, V2).getValue(WhichResult);
5493 if (isVUZPMask(ShuffleMask, VT, WhichResult))
5494 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
5495 V1, V2).getValue(WhichResult);
5496 if (isVZIPMask(ShuffleMask, VT, WhichResult))
5497 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
5498 V1, V2).getValue(WhichResult);
5500 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
5501 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
5502 V1, V1).getValue(WhichResult);
5503 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
5504 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
5505 V1, V1).getValue(WhichResult);
5506 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
5507 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
5508 V1, V1).getValue(WhichResult);
5511 // If the shuffle is not directly supported and it has 4 elements, use
5512 // the PerfectShuffle-generated table to synthesize it from other shuffles.
5513 unsigned NumElts = VT.getVectorNumElements();
5515 unsigned PFIndexes[4];
5516 for (unsigned i = 0; i != 4; ++i) {
5517 if (ShuffleMask[i] < 0)
5520 PFIndexes[i] = ShuffleMask[i];
5523 // Compute the index in the perfect shuffle table.
5524 unsigned PFTableIndex =
5525 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
5526 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
5527 unsigned Cost = (PFEntry >> 30);
5530 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
5533 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
5534 if (EltSize >= 32) {
5535 // Do the expansion with floating-point types, since that is what the VFP
5536 // registers are defined to use, and since i64 is not legal.
5537 EVT EltVT = EVT::getFloatingPointVT(EltSize);
5538 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
5539 V1 = DAG.getNode(ISD::BITCAST, dl, VecVT, V1);
5540 V2 = DAG.getNode(ISD::BITCAST, dl, VecVT, V2);
5541 SmallVector<SDValue, 8> Ops;
5542 for (unsigned i = 0; i < NumElts; ++i) {
5543 if (ShuffleMask[i] < 0)
5544 Ops.push_back(DAG.getUNDEF(EltVT));
5546 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
5547 ShuffleMask[i] < (int)NumElts ? V1 : V2,
5548 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
5551 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, Ops);
5552 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
5555 if ((VT == MVT::v8i16 || VT == MVT::v16i8) && isReverseMask(ShuffleMask, VT))
5556 return LowerReverse_VECTOR_SHUFFLEv16i8_v8i16(Op, DAG);
5558 if (VT == MVT::v8i8) {
5559 SDValue NewOp = LowerVECTOR_SHUFFLEv8i8(Op, ShuffleMask, DAG);
5560 if (NewOp.getNode())
5567 static SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
5568 // INSERT_VECTOR_ELT is legal only for immediate indexes.
5569 SDValue Lane = Op.getOperand(2);
5570 if (!isa<ConstantSDNode>(Lane))
5576 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
5577 // EXTRACT_VECTOR_ELT is legal only for immediate indexes.
5578 SDValue Lane = Op.getOperand(1);
5579 if (!isa<ConstantSDNode>(Lane))
5582 SDValue Vec = Op.getOperand(0);
5583 if (Op.getValueType() == MVT::i32 &&
5584 Vec.getValueType().getVectorElementType().getSizeInBits() < 32) {
5586 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
5592 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
5593 // The only time a CONCAT_VECTORS operation can have legal types is when
5594 // two 64-bit vectors are concatenated to a 128-bit vector.
5595 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
5596 "unexpected CONCAT_VECTORS");
5598 SDValue Val = DAG.getUNDEF(MVT::v2f64);
5599 SDValue Op0 = Op.getOperand(0);
5600 SDValue Op1 = Op.getOperand(1);
5601 if (Op0.getOpcode() != ISD::UNDEF)
5602 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
5603 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op0),
5604 DAG.getIntPtrConstant(0));
5605 if (Op1.getOpcode() != ISD::UNDEF)
5606 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
5607 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op1),
5608 DAG.getIntPtrConstant(1));
5609 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Val);
5612 /// isExtendedBUILD_VECTOR - Check if N is a constant BUILD_VECTOR where each
5613 /// element has been zero/sign-extended, depending on the isSigned parameter,
5614 /// from an integer type half its size.
5615 static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG,
5617 // A v2i64 BUILD_VECTOR will have been legalized to a BITCAST from v4i32.
5618 EVT VT = N->getValueType(0);
5619 if (VT == MVT::v2i64 && N->getOpcode() == ISD::BITCAST) {
5620 SDNode *BVN = N->getOperand(0).getNode();
5621 if (BVN->getValueType(0) != MVT::v4i32 ||
5622 BVN->getOpcode() != ISD::BUILD_VECTOR)
5624 unsigned LoElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
5625 unsigned HiElt = 1 - LoElt;
5626 ConstantSDNode *Lo0 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt));
5627 ConstantSDNode *Hi0 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt));
5628 ConstantSDNode *Lo1 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt+2));
5629 ConstantSDNode *Hi1 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt+2));
5630 if (!Lo0 || !Hi0 || !Lo1 || !Hi1)
5633 if (Hi0->getSExtValue() == Lo0->getSExtValue() >> 32 &&
5634 Hi1->getSExtValue() == Lo1->getSExtValue() >> 32)
5637 if (Hi0->isNullValue() && Hi1->isNullValue())
5643 if (N->getOpcode() != ISD::BUILD_VECTOR)
5646 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
5647 SDNode *Elt = N->getOperand(i).getNode();
5648 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
5649 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5650 unsigned HalfSize = EltSize / 2;
5652 if (!isIntN(HalfSize, C->getSExtValue()))
5655 if (!isUIntN(HalfSize, C->getZExtValue()))
5666 /// isSignExtended - Check if a node is a vector value that is sign-extended
5667 /// or a constant BUILD_VECTOR with sign-extended elements.
5668 static bool isSignExtended(SDNode *N, SelectionDAG &DAG) {
5669 if (N->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N))
5671 if (isExtendedBUILD_VECTOR(N, DAG, true))
5676 /// isZeroExtended - Check if a node is a vector value that is zero-extended
5677 /// or a constant BUILD_VECTOR with zero-extended elements.
5678 static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) {
5679 if (N->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N))
5681 if (isExtendedBUILD_VECTOR(N, DAG, false))
5686 static EVT getExtensionTo64Bits(const EVT &OrigVT) {
5687 if (OrigVT.getSizeInBits() >= 64)
5690 assert(OrigVT.isSimple() && "Expecting a simple value type");
5692 MVT::SimpleValueType OrigSimpleTy = OrigVT.getSimpleVT().SimpleTy;
5693 switch (OrigSimpleTy) {
5694 default: llvm_unreachable("Unexpected Vector Type");
5703 /// AddRequiredExtensionForVMULL - Add a sign/zero extension to extend the total
5704 /// value size to 64 bits. We need a 64-bit D register as an operand to VMULL.
5705 /// We insert the required extension here to get the vector to fill a D register.
5706 static SDValue AddRequiredExtensionForVMULL(SDValue N, SelectionDAG &DAG,
5709 unsigned ExtOpcode) {
5710 // The vector originally had a size of OrigTy. It was then extended to ExtTy.
5711 // We expect the ExtTy to be 128-bits total. If the OrigTy is less than
5712 // 64-bits we need to insert a new extension so that it will be 64-bits.
5713 assert(ExtTy.is128BitVector() && "Unexpected extension size");
5714 if (OrigTy.getSizeInBits() >= 64)
5717 // Must extend size to at least 64 bits to be used as an operand for VMULL.
5718 EVT NewVT = getExtensionTo64Bits(OrigTy);
5720 return DAG.getNode(ExtOpcode, SDLoc(N), NewVT, N);
5723 /// SkipLoadExtensionForVMULL - return a load of the original vector size that
5724 /// does not do any sign/zero extension. If the original vector is less
5725 /// than 64 bits, an appropriate extension will be added after the load to
5726 /// reach a total size of 64 bits. We have to add the extension separately
5727 /// because ARM does not have a sign/zero extending load for vectors.
5728 static SDValue SkipLoadExtensionForVMULL(LoadSDNode *LD, SelectionDAG& DAG) {
5729 EVT ExtendedTy = getExtensionTo64Bits(LD->getMemoryVT());
5731 // The load already has the right type.
5732 if (ExtendedTy == LD->getMemoryVT())
5733 return DAG.getLoad(LD->getMemoryVT(), SDLoc(LD), LD->getChain(),
5734 LD->getBasePtr(), LD->getPointerInfo(), LD->isVolatile(),
5735 LD->isNonTemporal(), LD->isInvariant(),
5736 LD->getAlignment());
5738 // We need to create a zextload/sextload. We cannot just create a load
5739 // followed by a zext/zext node because LowerMUL is also run during normal
5740 // operation legalization where we can't create illegal types.
5741 return DAG.getExtLoad(LD->getExtensionType(), SDLoc(LD), ExtendedTy,
5742 LD->getChain(), LD->getBasePtr(), LD->getPointerInfo(),
5743 LD->getMemoryVT(), LD->isVolatile(),
5744 LD->isNonTemporal(), LD->getAlignment());
5747 /// SkipExtensionForVMULL - For a node that is a SIGN_EXTEND, ZERO_EXTEND,
5748 /// extending load, or BUILD_VECTOR with extended elements, return the
5749 /// unextended value. The unextended vector should be 64 bits so that it can
5750 /// be used as an operand to a VMULL instruction. If the original vector size
5751 /// before extension is less than 64 bits we add a an extension to resize
5752 /// the vector to 64 bits.
5753 static SDValue SkipExtensionForVMULL(SDNode *N, SelectionDAG &DAG) {
5754 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
5755 return AddRequiredExtensionForVMULL(N->getOperand(0), DAG,
5756 N->getOperand(0)->getValueType(0),
5760 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
5761 return SkipLoadExtensionForVMULL(LD, DAG);
5763 // Otherwise, the value must be a BUILD_VECTOR. For v2i64, it will
5764 // have been legalized as a BITCAST from v4i32.
5765 if (N->getOpcode() == ISD::BITCAST) {
5766 SDNode *BVN = N->getOperand(0).getNode();
5767 assert(BVN->getOpcode() == ISD::BUILD_VECTOR &&
5768 BVN->getValueType(0) == MVT::v4i32 && "expected v4i32 BUILD_VECTOR");
5769 unsigned LowElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
5770 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), MVT::v2i32,
5771 BVN->getOperand(LowElt), BVN->getOperand(LowElt+2));
5773 // Construct a new BUILD_VECTOR with elements truncated to half the size.
5774 assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR");
5775 EVT VT = N->getValueType(0);
5776 unsigned EltSize = VT.getVectorElementType().getSizeInBits() / 2;
5777 unsigned NumElts = VT.getVectorNumElements();
5778 MVT TruncVT = MVT::getIntegerVT(EltSize);
5779 SmallVector<SDValue, 8> Ops;
5780 for (unsigned i = 0; i != NumElts; ++i) {
5781 ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i));
5782 const APInt &CInt = C->getAPIntValue();
5783 // Element types smaller than 32 bits are not legal, so use i32 elements.
5784 // The values are implicitly truncated so sext vs. zext doesn't matter.
5785 Ops.push_back(DAG.getConstant(CInt.zextOrTrunc(32), MVT::i32));
5787 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N),
5788 MVT::getVectorVT(TruncVT, NumElts), Ops);
5791 static bool isAddSubSExt(SDNode *N, SelectionDAG &DAG) {
5792 unsigned Opcode = N->getOpcode();
5793 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
5794 SDNode *N0 = N->getOperand(0).getNode();
5795 SDNode *N1 = N->getOperand(1).getNode();
5796 return N0->hasOneUse() && N1->hasOneUse() &&
5797 isSignExtended(N0, DAG) && isSignExtended(N1, DAG);
5802 static bool isAddSubZExt(SDNode *N, SelectionDAG &DAG) {
5803 unsigned Opcode = N->getOpcode();
5804 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
5805 SDNode *N0 = N->getOperand(0).getNode();
5806 SDNode *N1 = N->getOperand(1).getNode();
5807 return N0->hasOneUse() && N1->hasOneUse() &&
5808 isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG);
5813 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
5814 // Multiplications are only custom-lowered for 128-bit vectors so that
5815 // VMULL can be detected. Otherwise v2i64 multiplications are not legal.
5816 EVT VT = Op.getValueType();
5817 assert(VT.is128BitVector() && VT.isInteger() &&
5818 "unexpected type for custom-lowering ISD::MUL");
5819 SDNode *N0 = Op.getOperand(0).getNode();
5820 SDNode *N1 = Op.getOperand(1).getNode();
5821 unsigned NewOpc = 0;
5823 bool isN0SExt = isSignExtended(N0, DAG);
5824 bool isN1SExt = isSignExtended(N1, DAG);
5825 if (isN0SExt && isN1SExt)
5826 NewOpc = ARMISD::VMULLs;
5828 bool isN0ZExt = isZeroExtended(N0, DAG);
5829 bool isN1ZExt = isZeroExtended(N1, DAG);
5830 if (isN0ZExt && isN1ZExt)
5831 NewOpc = ARMISD::VMULLu;
5832 else if (isN1SExt || isN1ZExt) {
5833 // Look for (s/zext A + s/zext B) * (s/zext C). We want to turn these
5834 // into (s/zext A * s/zext C) + (s/zext B * s/zext C)
5835 if (isN1SExt && isAddSubSExt(N0, DAG)) {
5836 NewOpc = ARMISD::VMULLs;
5838 } else if (isN1ZExt && isAddSubZExt(N0, DAG)) {
5839 NewOpc = ARMISD::VMULLu;
5841 } else if (isN0ZExt && isAddSubZExt(N1, DAG)) {
5843 NewOpc = ARMISD::VMULLu;
5849 if (VT == MVT::v2i64)
5850 // Fall through to expand this. It is not legal.
5853 // Other vector multiplications are legal.
5858 // Legalize to a VMULL instruction.
5861 SDValue Op1 = SkipExtensionForVMULL(N1, DAG);
5863 Op0 = SkipExtensionForVMULL(N0, DAG);
5864 assert(Op0.getValueType().is64BitVector() &&
5865 Op1.getValueType().is64BitVector() &&
5866 "unexpected types for extended operands to VMULL");
5867 return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
5870 // Optimizing (zext A + zext B) * C, to (VMULL A, C) + (VMULL B, C) during
5871 // isel lowering to take advantage of no-stall back to back vmul + vmla.
5878 SDValue N00 = SkipExtensionForVMULL(N0->getOperand(0).getNode(), DAG);
5879 SDValue N01 = SkipExtensionForVMULL(N0->getOperand(1).getNode(), DAG);
5880 EVT Op1VT = Op1.getValueType();
5881 return DAG.getNode(N0->getOpcode(), DL, VT,
5882 DAG.getNode(NewOpc, DL, VT,
5883 DAG.getNode(ISD::BITCAST, DL, Op1VT, N00), Op1),
5884 DAG.getNode(NewOpc, DL, VT,
5885 DAG.getNode(ISD::BITCAST, DL, Op1VT, N01), Op1));
5889 LowerSDIV_v4i8(SDValue X, SDValue Y, SDLoc dl, SelectionDAG &DAG) {
5891 // float4 xf = vcvt_f32_s32(vmovl_s16(a.lo));
5892 // float4 yf = vcvt_f32_s32(vmovl_s16(b.lo));
5893 X = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, X);
5894 Y = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, Y);
5895 X = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, X);
5896 Y = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, Y);
5897 // Get reciprocal estimate.
5898 // float4 recip = vrecpeq_f32(yf);
5899 Y = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5900 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), Y);
5901 // Because char has a smaller range than uchar, we can actually get away
5902 // without any newton steps. This requires that we use a weird bias
5903 // of 0xb000, however (again, this has been exhaustively tested).
5904 // float4 result = as_float4(as_int4(xf*recip) + 0xb000);
5905 X = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, X, Y);
5906 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, X);
5907 Y = DAG.getConstant(0xb000, MVT::i32);
5908 Y = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Y, Y, Y, Y);
5909 X = DAG.getNode(ISD::ADD, dl, MVT::v4i32, X, Y);
5910 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, X);
5911 // Convert back to short.
5912 X = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, X);
5913 X = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, X);
5918 LowerSDIV_v4i16(SDValue N0, SDValue N1, SDLoc dl, SelectionDAG &DAG) {
5920 // Convert to float.
5921 // float4 yf = vcvt_f32_s32(vmovl_s16(y));
5922 // float4 xf = vcvt_f32_s32(vmovl_s16(x));
5923 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N0);
5924 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N1);
5925 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
5926 N1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
5928 // Use reciprocal estimate and one refinement step.
5929 // float4 recip = vrecpeq_f32(yf);
5930 // recip *= vrecpsq_f32(yf, recip);
5931 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5932 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), N1);
5933 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5934 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
5936 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
5937 // Because short has a smaller range than ushort, we can actually get away
5938 // with only a single newton step. This requires that we use a weird bias
5939 // of 89, however (again, this has been exhaustively tested).
5940 // float4 result = as_float4(as_int4(xf*recip) + 0x89);
5941 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
5942 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
5943 N1 = DAG.getConstant(0x89, MVT::i32);
5944 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
5945 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
5946 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
5947 // Convert back to integer and return.
5948 // return vmovn_s32(vcvt_s32_f32(result));
5949 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
5950 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
5954 static SDValue LowerSDIV(SDValue Op, SelectionDAG &DAG) {
5955 EVT VT = Op.getValueType();
5956 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
5957 "unexpected type for custom-lowering ISD::SDIV");
5960 SDValue N0 = Op.getOperand(0);
5961 SDValue N1 = Op.getOperand(1);
5964 if (VT == MVT::v8i8) {
5965 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N0);
5966 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N1);
5968 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5969 DAG.getIntPtrConstant(4));
5970 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5971 DAG.getIntPtrConstant(4));
5972 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5973 DAG.getIntPtrConstant(0));
5974 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5975 DAG.getIntPtrConstant(0));
5977 N0 = LowerSDIV_v4i8(N0, N1, dl, DAG); // v4i16
5978 N2 = LowerSDIV_v4i8(N2, N3, dl, DAG); // v4i16
5980 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
5981 N0 = LowerCONCAT_VECTORS(N0, DAG);
5983 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v8i8, N0);
5986 return LowerSDIV_v4i16(N0, N1, dl, DAG);
5989 static SDValue LowerUDIV(SDValue Op, SelectionDAG &DAG) {
5990 EVT VT = Op.getValueType();
5991 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
5992 "unexpected type for custom-lowering ISD::UDIV");
5995 SDValue N0 = Op.getOperand(0);
5996 SDValue N1 = Op.getOperand(1);
5999 if (VT == MVT::v8i8) {
6000 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N0);
6001 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N1);
6003 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
6004 DAG.getIntPtrConstant(4));
6005 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
6006 DAG.getIntPtrConstant(4));
6007 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
6008 DAG.getIntPtrConstant(0));
6009 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
6010 DAG.getIntPtrConstant(0));
6012 N0 = LowerSDIV_v4i16(N0, N1, dl, DAG); // v4i16
6013 N2 = LowerSDIV_v4i16(N2, N3, dl, DAG); // v4i16
6015 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
6016 N0 = LowerCONCAT_VECTORS(N0, DAG);
6018 N0 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v8i8,
6019 DAG.getConstant(Intrinsic::arm_neon_vqmovnsu, MVT::i32),
6024 // v4i16 sdiv ... Convert to float.
6025 // float4 yf = vcvt_f32_s32(vmovl_u16(y));
6026 // float4 xf = vcvt_f32_s32(vmovl_u16(x));
6027 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N0);
6028 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N1);
6029 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
6030 SDValue BN1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
6032 // Use reciprocal estimate and two refinement steps.
6033 // float4 recip = vrecpeq_f32(yf);
6034 // recip *= vrecpsq_f32(yf, recip);
6035 // recip *= vrecpsq_f32(yf, recip);
6036 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6037 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), BN1);
6038 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6039 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
6041 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
6042 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6043 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
6045 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
6046 // Simply multiplying by the reciprocal estimate can leave us a few ulps
6047 // too low, so we add 2 ulps (exhaustive testing shows that this is enough,
6048 // and that it will never cause us to return an answer too large).
6049 // float4 result = as_float4(as_int4(xf*recip) + 2);
6050 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
6051 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
6052 N1 = DAG.getConstant(2, MVT::i32);
6053 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
6054 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
6055 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
6056 // Convert back to integer and return.
6057 // return vmovn_u32(vcvt_s32_f32(result));
6058 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
6059 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
6063 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
6064 EVT VT = Op.getNode()->getValueType(0);
6065 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
6068 bool ExtraOp = false;
6069 switch (Op.getOpcode()) {
6070 default: llvm_unreachable("Invalid code");
6071 case ISD::ADDC: Opc = ARMISD::ADDC; break;
6072 case ISD::ADDE: Opc = ARMISD::ADDE; ExtraOp = true; break;
6073 case ISD::SUBC: Opc = ARMISD::SUBC; break;
6074 case ISD::SUBE: Opc = ARMISD::SUBE; ExtraOp = true; break;
6078 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
6080 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
6081 Op.getOperand(1), Op.getOperand(2));
6084 SDValue ARMTargetLowering::LowerFSINCOS(SDValue Op, SelectionDAG &DAG) const {
6085 assert(Subtarget->isTargetDarwin());
6087 // For iOS, we want to call an alternative entry point: __sincos_stret,
6088 // return values are passed via sret.
6090 SDValue Arg = Op.getOperand(0);
6091 EVT ArgVT = Arg.getValueType();
6092 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
6094 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
6095 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6097 // Pair of floats / doubles used to pass the result.
6098 StructType *RetTy = StructType::get(ArgTy, ArgTy, NULL);
6100 // Create stack object for sret.
6101 const uint64_t ByteSize = TLI.getDataLayout()->getTypeAllocSize(RetTy);
6102 const unsigned StackAlign = TLI.getDataLayout()->getPrefTypeAlignment(RetTy);
6103 int FrameIdx = FrameInfo->CreateStackObject(ByteSize, StackAlign, false);
6104 SDValue SRet = DAG.getFrameIndex(FrameIdx, TLI.getPointerTy());
6110 Entry.Ty = RetTy->getPointerTo();
6111 Entry.isSExt = false;
6112 Entry.isZExt = false;
6113 Entry.isSRet = true;
6114 Args.push_back(Entry);
6118 Entry.isSExt = false;
6119 Entry.isZExt = false;
6120 Args.push_back(Entry);
6122 const char *LibcallName = (ArgVT == MVT::f64)
6123 ? "__sincos_stret" : "__sincosf_stret";
6124 SDValue Callee = DAG.getExternalSymbol(LibcallName, getPointerTy());
6126 TargetLowering::CallLoweringInfo CLI(DAG);
6127 CLI.setDebugLoc(dl).setChain(DAG.getEntryNode())
6128 .setCallee(CallingConv::C, Type::getVoidTy(*DAG.getContext()), Callee,
6130 .setDiscardResult();
6132 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
6134 SDValue LoadSin = DAG.getLoad(ArgVT, dl, CallResult.second, SRet,
6135 MachinePointerInfo(), false, false, false, 0);
6137 // Address of cos field.
6138 SDValue Add = DAG.getNode(ISD::ADD, dl, getPointerTy(), SRet,
6139 DAG.getIntPtrConstant(ArgVT.getStoreSize()));
6140 SDValue LoadCos = DAG.getLoad(ArgVT, dl, LoadSin.getValue(1), Add,
6141 MachinePointerInfo(), false, false, false, 0);
6143 SDVTList Tys = DAG.getVTList(ArgVT, ArgVT);
6144 return DAG.getNode(ISD::MERGE_VALUES, dl, Tys,
6145 LoadSin.getValue(0), LoadCos.getValue(0));
6148 static SDValue LowerAtomicLoadStore(SDValue Op, SelectionDAG &DAG) {
6149 // Monotonic load/store is legal for all targets
6150 if (cast<AtomicSDNode>(Op)->getOrdering() <= Monotonic)
6153 // Acquire/Release load/store is not legal for targets without a
6154 // dmb or equivalent available.
6158 static void ReplaceREADCYCLECOUNTER(SDNode *N,
6159 SmallVectorImpl<SDValue> &Results,
6161 const ARMSubtarget *Subtarget) {
6163 SDValue Cycles32, OutChain;
6165 if (Subtarget->hasPerfMon()) {
6166 // Under Power Management extensions, the cycle-count is:
6167 // mrc p15, #0, <Rt>, c9, c13, #0
6168 SDValue Ops[] = { N->getOperand(0), // Chain
6169 DAG.getConstant(Intrinsic::arm_mrc, MVT::i32),
6170 DAG.getConstant(15, MVT::i32),
6171 DAG.getConstant(0, MVT::i32),
6172 DAG.getConstant(9, MVT::i32),
6173 DAG.getConstant(13, MVT::i32),
6174 DAG.getConstant(0, MVT::i32)
6177 Cycles32 = DAG.getNode(ISD::INTRINSIC_W_CHAIN, DL,
6178 DAG.getVTList(MVT::i32, MVT::Other), Ops);
6179 OutChain = Cycles32.getValue(1);
6181 // Intrinsic is defined to return 0 on unsupported platforms. Technically
6182 // there are older ARM CPUs that have implementation-specific ways of
6183 // obtaining this information (FIXME!).
6184 Cycles32 = DAG.getConstant(0, MVT::i32);
6185 OutChain = DAG.getEntryNode();
6189 SDValue Cycles64 = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64,
6190 Cycles32, DAG.getConstant(0, MVT::i32));
6191 Results.push_back(Cycles64);
6192 Results.push_back(OutChain);
6195 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
6196 switch (Op.getOpcode()) {
6197 default: llvm_unreachable("Don't know how to custom lower this!");
6198 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
6199 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
6200 case ISD::GlobalAddress:
6201 switch (Subtarget->getTargetTriple().getObjectFormat()) {
6202 default: llvm_unreachable("unknown object format");
6204 return LowerGlobalAddressWindows(Op, DAG);
6206 return LowerGlobalAddressELF(Op, DAG);
6208 return LowerGlobalAddressDarwin(Op, DAG);
6210 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
6211 case ISD::SELECT: return LowerSELECT(Op, DAG);
6212 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
6213 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
6214 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
6215 case ISD::VASTART: return LowerVASTART(Op, DAG);
6216 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, DAG, Subtarget);
6217 case ISD::PREFETCH: return LowerPREFETCH(Op, DAG, Subtarget);
6218 case ISD::SINT_TO_FP:
6219 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
6220 case ISD::FP_TO_SINT:
6221 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
6222 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
6223 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
6224 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
6225 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
6226 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
6227 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
6228 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
6230 case ISD::BITCAST: return ExpandBITCAST(Op.getNode(), DAG);
6233 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
6234 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
6235 case ISD::SRL_PARTS:
6236 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
6237 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
6238 case ISD::CTPOP: return LowerCTPOP(Op.getNode(), DAG, Subtarget);
6239 case ISD::SETCC: return LowerVSETCC(Op, DAG);
6240 case ISD::ConstantFP: return LowerConstantFP(Op, DAG, Subtarget);
6241 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG, Subtarget);
6242 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
6243 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
6244 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
6245 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
6246 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
6247 case ISD::MUL: return LowerMUL(Op, DAG);
6248 case ISD::SDIV: return LowerSDIV(Op, DAG);
6249 case ISD::UDIV: return LowerUDIV(Op, DAG);
6253 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
6258 return LowerXALUO(Op, DAG);
6259 case ISD::ATOMIC_LOAD:
6260 case ISD::ATOMIC_STORE: return LowerAtomicLoadStore(Op, DAG);
6261 case ISD::FSINCOS: return LowerFSINCOS(Op, DAG);
6263 case ISD::UDIVREM: return LowerDivRem(Op, DAG);
6264 case ISD::DYNAMIC_STACKALLOC:
6265 if (Subtarget->getTargetTriple().isWindowsItaniumEnvironment())
6266 return LowerDYNAMIC_STACKALLOC(Op, DAG);
6267 llvm_unreachable("Don't know how to custom lower this!");
6271 /// ReplaceNodeResults - Replace the results of node with an illegal result
6272 /// type with new values built out of custom code.
6273 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
6274 SmallVectorImpl<SDValue>&Results,
6275 SelectionDAG &DAG) const {
6277 switch (N->getOpcode()) {
6279 llvm_unreachable("Don't know how to custom expand this!");
6281 Res = ExpandBITCAST(N, DAG);
6285 Res = Expand64BitShift(N, DAG, Subtarget);
6287 case ISD::READCYCLECOUNTER:
6288 ReplaceREADCYCLECOUNTER(N, Results, DAG, Subtarget);
6292 Results.push_back(Res);
6295 //===----------------------------------------------------------------------===//
6296 // ARM Scheduler Hooks
6297 //===----------------------------------------------------------------------===//
6299 /// SetupEntryBlockForSjLj - Insert code into the entry block that creates and
6300 /// registers the function context.
6301 void ARMTargetLowering::
6302 SetupEntryBlockForSjLj(MachineInstr *MI, MachineBasicBlock *MBB,
6303 MachineBasicBlock *DispatchBB, int FI) const {
6304 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6305 DebugLoc dl = MI->getDebugLoc();
6306 MachineFunction *MF = MBB->getParent();
6307 MachineRegisterInfo *MRI = &MF->getRegInfo();
6308 MachineConstantPool *MCP = MF->getConstantPool();
6309 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
6310 const Function *F = MF->getFunction();
6312 bool isThumb = Subtarget->isThumb();
6313 bool isThumb2 = Subtarget->isThumb2();
6315 unsigned PCLabelId = AFI->createPICLabelUId();
6316 unsigned PCAdj = (isThumb || isThumb2) ? 4 : 8;
6317 ARMConstantPoolValue *CPV =
6318 ARMConstantPoolMBB::Create(F->getContext(), DispatchBB, PCLabelId, PCAdj);
6319 unsigned CPI = MCP->getConstantPoolIndex(CPV, 4);
6321 const TargetRegisterClass *TRC = isThumb ?
6322 (const TargetRegisterClass*)&ARM::tGPRRegClass :
6323 (const TargetRegisterClass*)&ARM::GPRRegClass;
6325 // Grab constant pool and fixed stack memory operands.
6326 MachineMemOperand *CPMMO =
6327 MF->getMachineMemOperand(MachinePointerInfo::getConstantPool(),
6328 MachineMemOperand::MOLoad, 4, 4);
6330 MachineMemOperand *FIMMOSt =
6331 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
6332 MachineMemOperand::MOStore, 4, 4);
6334 // Load the address of the dispatch MBB into the jump buffer.
6336 // Incoming value: jbuf
6337 // ldr.n r5, LCPI1_1
6340 // str r5, [$jbuf, #+4] ; &jbuf[1]
6341 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6342 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2LDRpci), NewVReg1)
6343 .addConstantPoolIndex(CPI)
6344 .addMemOperand(CPMMO));
6345 // Set the low bit because of thumb mode.
6346 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6348 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2ORRri), NewVReg2)
6349 .addReg(NewVReg1, RegState::Kill)
6351 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6352 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg3)
6353 .addReg(NewVReg2, RegState::Kill)
6355 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2STRi12))
6356 .addReg(NewVReg3, RegState::Kill)
6358 .addImm(36) // &jbuf[1] :: pc
6359 .addMemOperand(FIMMOSt));
6360 } else if (isThumb) {
6361 // Incoming value: jbuf
6362 // ldr.n r1, LCPI1_4
6366 // add r2, $jbuf, #+4 ; &jbuf[1]
6368 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6369 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tLDRpci), NewVReg1)
6370 .addConstantPoolIndex(CPI)
6371 .addMemOperand(CPMMO));
6372 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6373 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg2)
6374 .addReg(NewVReg1, RegState::Kill)
6376 // Set the low bit because of thumb mode.
6377 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6378 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tMOVi8), NewVReg3)
6379 .addReg(ARM::CPSR, RegState::Define)
6381 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6382 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tORR), NewVReg4)
6383 .addReg(ARM::CPSR, RegState::Define)
6384 .addReg(NewVReg2, RegState::Kill)
6385 .addReg(NewVReg3, RegState::Kill));
6386 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6387 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tADDrSPi), NewVReg5)
6389 .addImm(36)); // &jbuf[1] :: pc
6390 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tSTRi))
6391 .addReg(NewVReg4, RegState::Kill)
6392 .addReg(NewVReg5, RegState::Kill)
6394 .addMemOperand(FIMMOSt));
6396 // Incoming value: jbuf
6399 // str r1, [$jbuf, #+4] ; &jbuf[1]
6400 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6401 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::LDRi12), NewVReg1)
6402 .addConstantPoolIndex(CPI)
6404 .addMemOperand(CPMMO));
6405 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6406 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::PICADD), NewVReg2)
6407 .addReg(NewVReg1, RegState::Kill)
6408 .addImm(PCLabelId));
6409 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::STRi12))
6410 .addReg(NewVReg2, RegState::Kill)
6412 .addImm(36) // &jbuf[1] :: pc
6413 .addMemOperand(FIMMOSt));
6417 MachineBasicBlock *ARMTargetLowering::
6418 EmitSjLjDispatchBlock(MachineInstr *MI, MachineBasicBlock *MBB) const {
6419 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6420 DebugLoc dl = MI->getDebugLoc();
6421 MachineFunction *MF = MBB->getParent();
6422 MachineRegisterInfo *MRI = &MF->getRegInfo();
6423 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
6424 MachineFrameInfo *MFI = MF->getFrameInfo();
6425 int FI = MFI->getFunctionContextIndex();
6427 const TargetRegisterClass *TRC = Subtarget->isThumb() ?
6428 (const TargetRegisterClass*)&ARM::tGPRRegClass :
6429 (const TargetRegisterClass*)&ARM::GPRnopcRegClass;
6431 // Get a mapping of the call site numbers to all of the landing pads they're
6433 DenseMap<unsigned, SmallVector<MachineBasicBlock*, 2> > CallSiteNumToLPad;
6434 unsigned MaxCSNum = 0;
6435 MachineModuleInfo &MMI = MF->getMMI();
6436 for (MachineFunction::iterator BB = MF->begin(), E = MF->end(); BB != E;
6438 if (!BB->isLandingPad()) continue;
6440 // FIXME: We should assert that the EH_LABEL is the first MI in the landing
6442 for (MachineBasicBlock::iterator
6443 II = BB->begin(), IE = BB->end(); II != IE; ++II) {
6444 if (!II->isEHLabel()) continue;
6446 MCSymbol *Sym = II->getOperand(0).getMCSymbol();
6447 if (!MMI.hasCallSiteLandingPad(Sym)) continue;
6449 SmallVectorImpl<unsigned> &CallSiteIdxs = MMI.getCallSiteLandingPad(Sym);
6450 for (SmallVectorImpl<unsigned>::iterator
6451 CSI = CallSiteIdxs.begin(), CSE = CallSiteIdxs.end();
6452 CSI != CSE; ++CSI) {
6453 CallSiteNumToLPad[*CSI].push_back(BB);
6454 MaxCSNum = std::max(MaxCSNum, *CSI);
6460 // Get an ordered list of the machine basic blocks for the jump table.
6461 std::vector<MachineBasicBlock*> LPadList;
6462 SmallPtrSet<MachineBasicBlock*, 64> InvokeBBs;
6463 LPadList.reserve(CallSiteNumToLPad.size());
6464 for (unsigned I = 1; I <= MaxCSNum; ++I) {
6465 SmallVectorImpl<MachineBasicBlock*> &MBBList = CallSiteNumToLPad[I];
6466 for (SmallVectorImpl<MachineBasicBlock*>::iterator
6467 II = MBBList.begin(), IE = MBBList.end(); II != IE; ++II) {
6468 LPadList.push_back(*II);
6469 InvokeBBs.insert((*II)->pred_begin(), (*II)->pred_end());
6473 assert(!LPadList.empty() &&
6474 "No landing pad destinations for the dispatch jump table!");
6476 // Create the jump table and associated information.
6477 MachineJumpTableInfo *JTI =
6478 MF->getOrCreateJumpTableInfo(MachineJumpTableInfo::EK_Inline);
6479 unsigned MJTI = JTI->createJumpTableIndex(LPadList);
6480 unsigned UId = AFI->createJumpTableUId();
6481 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
6483 // Create the MBBs for the dispatch code.
6485 // Shove the dispatch's address into the return slot in the function context.
6486 MachineBasicBlock *DispatchBB = MF->CreateMachineBasicBlock();
6487 DispatchBB->setIsLandingPad();
6489 MachineBasicBlock *TrapBB = MF->CreateMachineBasicBlock();
6490 unsigned trap_opcode;
6491 if (Subtarget->isThumb())
6492 trap_opcode = ARM::tTRAP;
6494 trap_opcode = Subtarget->useNaClTrap() ? ARM::TRAPNaCl : ARM::TRAP;
6496 BuildMI(TrapBB, dl, TII->get(trap_opcode));
6497 DispatchBB->addSuccessor(TrapBB);
6499 MachineBasicBlock *DispContBB = MF->CreateMachineBasicBlock();
6500 DispatchBB->addSuccessor(DispContBB);
6503 MF->insert(MF->end(), DispatchBB);
6504 MF->insert(MF->end(), DispContBB);
6505 MF->insert(MF->end(), TrapBB);
6507 // Insert code into the entry block that creates and registers the function
6509 SetupEntryBlockForSjLj(MI, MBB, DispatchBB, FI);
6511 MachineMemOperand *FIMMOLd =
6512 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
6513 MachineMemOperand::MOLoad |
6514 MachineMemOperand::MOVolatile, 4, 4);
6516 MachineInstrBuilder MIB;
6517 MIB = BuildMI(DispatchBB, dl, TII->get(ARM::Int_eh_sjlj_dispatchsetup));
6519 const ARMBaseInstrInfo *AII = static_cast<const ARMBaseInstrInfo*>(TII);
6520 const ARMBaseRegisterInfo &RI = AII->getRegisterInfo();
6522 // Add a register mask with no preserved registers. This results in all
6523 // registers being marked as clobbered.
6524 MIB.addRegMask(RI.getNoPreservedMask());
6526 unsigned NumLPads = LPadList.size();
6527 if (Subtarget->isThumb2()) {
6528 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6529 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2LDRi12), NewVReg1)
6532 .addMemOperand(FIMMOLd));
6534 if (NumLPads < 256) {
6535 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPri))
6537 .addImm(LPadList.size()));
6539 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6540 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVi16), VReg1)
6541 .addImm(NumLPads & 0xFFFF));
6543 unsigned VReg2 = VReg1;
6544 if ((NumLPads & 0xFFFF0000) != 0) {
6545 VReg2 = MRI->createVirtualRegister(TRC);
6546 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVTi16), VReg2)
6548 .addImm(NumLPads >> 16));
6551 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPrr))
6556 BuildMI(DispatchBB, dl, TII->get(ARM::t2Bcc))
6561 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6562 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::t2LEApcrelJT),NewVReg3)
6563 .addJumpTableIndex(MJTI)
6566 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6569 BuildMI(DispContBB, dl, TII->get(ARM::t2ADDrs), NewVReg4)
6570 .addReg(NewVReg3, RegState::Kill)
6572 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
6574 BuildMI(DispContBB, dl, TII->get(ARM::t2BR_JT))
6575 .addReg(NewVReg4, RegState::Kill)
6577 .addJumpTableIndex(MJTI)
6579 } else if (Subtarget->isThumb()) {
6580 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6581 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRspi), NewVReg1)
6584 .addMemOperand(FIMMOLd));
6586 if (NumLPads < 256) {
6587 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPi8))
6591 MachineConstantPool *ConstantPool = MF->getConstantPool();
6592 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
6593 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
6595 // MachineConstantPool wants an explicit alignment.
6596 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
6598 Align = getDataLayout()->getTypeAllocSize(C->getType());
6599 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6601 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6602 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRpci))
6603 .addReg(VReg1, RegState::Define)
6604 .addConstantPoolIndex(Idx));
6605 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPr))
6610 BuildMI(DispatchBB, dl, TII->get(ARM::tBcc))
6615 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6616 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLSLri), NewVReg2)
6617 .addReg(ARM::CPSR, RegState::Define)
6621 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6622 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLEApcrelJT), NewVReg3)
6623 .addJumpTableIndex(MJTI)
6626 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6627 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg4)
6628 .addReg(ARM::CPSR, RegState::Define)
6629 .addReg(NewVReg2, RegState::Kill)
6632 MachineMemOperand *JTMMOLd =
6633 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
6634 MachineMemOperand::MOLoad, 4, 4);
6636 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6637 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLDRi), NewVReg5)
6638 .addReg(NewVReg4, RegState::Kill)
6640 .addMemOperand(JTMMOLd));
6642 unsigned NewVReg6 = NewVReg5;
6643 if (RelocM == Reloc::PIC_) {
6644 NewVReg6 = MRI->createVirtualRegister(TRC);
6645 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg6)
6646 .addReg(ARM::CPSR, RegState::Define)
6647 .addReg(NewVReg5, RegState::Kill)
6651 BuildMI(DispContBB, dl, TII->get(ARM::tBR_JTr))
6652 .addReg(NewVReg6, RegState::Kill)
6653 .addJumpTableIndex(MJTI)
6656 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6657 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRi12), NewVReg1)
6660 .addMemOperand(FIMMOLd));
6662 if (NumLPads < 256) {
6663 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPri))
6666 } else if (Subtarget->hasV6T2Ops() && isUInt<16>(NumLPads)) {
6667 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6668 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVi16), VReg1)
6669 .addImm(NumLPads & 0xFFFF));
6671 unsigned VReg2 = VReg1;
6672 if ((NumLPads & 0xFFFF0000) != 0) {
6673 VReg2 = MRI->createVirtualRegister(TRC);
6674 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVTi16), VReg2)
6676 .addImm(NumLPads >> 16));
6679 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
6683 MachineConstantPool *ConstantPool = MF->getConstantPool();
6684 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
6685 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
6687 // MachineConstantPool wants an explicit alignment.
6688 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
6690 Align = getDataLayout()->getTypeAllocSize(C->getType());
6691 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6693 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6694 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRcp))
6695 .addReg(VReg1, RegState::Define)
6696 .addConstantPoolIndex(Idx)
6698 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
6700 .addReg(VReg1, RegState::Kill));
6703 BuildMI(DispatchBB, dl, TII->get(ARM::Bcc))
6708 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6710 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::MOVsi), NewVReg3)
6712 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
6713 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6714 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::LEApcrelJT), NewVReg4)
6715 .addJumpTableIndex(MJTI)
6718 MachineMemOperand *JTMMOLd =
6719 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
6720 MachineMemOperand::MOLoad, 4, 4);
6721 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6723 BuildMI(DispContBB, dl, TII->get(ARM::LDRrs), NewVReg5)
6724 .addReg(NewVReg3, RegState::Kill)
6727 .addMemOperand(JTMMOLd));
6729 if (RelocM == Reloc::PIC_) {
6730 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTadd))
6731 .addReg(NewVReg5, RegState::Kill)
6733 .addJumpTableIndex(MJTI)
6736 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTr))
6737 .addReg(NewVReg5, RegState::Kill)
6738 .addJumpTableIndex(MJTI)
6743 // Add the jump table entries as successors to the MBB.
6744 SmallPtrSet<MachineBasicBlock*, 8> SeenMBBs;
6745 for (std::vector<MachineBasicBlock*>::iterator
6746 I = LPadList.begin(), E = LPadList.end(); I != E; ++I) {
6747 MachineBasicBlock *CurMBB = *I;
6748 if (SeenMBBs.insert(CurMBB))
6749 DispContBB->addSuccessor(CurMBB);
6752 // N.B. the order the invoke BBs are processed in doesn't matter here.
6753 const MCPhysReg *SavedRegs = RI.getCalleeSavedRegs(MF);
6754 SmallVector<MachineBasicBlock*, 64> MBBLPads;
6755 for (SmallPtrSet<MachineBasicBlock*, 64>::iterator
6756 I = InvokeBBs.begin(), E = InvokeBBs.end(); I != E; ++I) {
6757 MachineBasicBlock *BB = *I;
6759 // Remove the landing pad successor from the invoke block and replace it
6760 // with the new dispatch block.
6761 SmallVector<MachineBasicBlock*, 4> Successors(BB->succ_begin(),
6763 while (!Successors.empty()) {
6764 MachineBasicBlock *SMBB = Successors.pop_back_val();
6765 if (SMBB->isLandingPad()) {
6766 BB->removeSuccessor(SMBB);
6767 MBBLPads.push_back(SMBB);
6771 BB->addSuccessor(DispatchBB);
6773 // Find the invoke call and mark all of the callee-saved registers as
6774 // 'implicit defined' so that they're spilled. This prevents code from
6775 // moving instructions to before the EH block, where they will never be
6777 for (MachineBasicBlock::reverse_iterator
6778 II = BB->rbegin(), IE = BB->rend(); II != IE; ++II) {
6779 if (!II->isCall()) continue;
6781 DenseMap<unsigned, bool> DefRegs;
6782 for (MachineInstr::mop_iterator
6783 OI = II->operands_begin(), OE = II->operands_end();
6785 if (!OI->isReg()) continue;
6786 DefRegs[OI->getReg()] = true;
6789 MachineInstrBuilder MIB(*MF, &*II);
6791 for (unsigned i = 0; SavedRegs[i] != 0; ++i) {
6792 unsigned Reg = SavedRegs[i];
6793 if (Subtarget->isThumb2() &&
6794 !ARM::tGPRRegClass.contains(Reg) &&
6795 !ARM::hGPRRegClass.contains(Reg))
6797 if (Subtarget->isThumb1Only() && !ARM::tGPRRegClass.contains(Reg))
6799 if (!Subtarget->isThumb() && !ARM::GPRRegClass.contains(Reg))
6802 MIB.addReg(Reg, RegState::ImplicitDefine | RegState::Dead);
6809 // Mark all former landing pads as non-landing pads. The dispatch is the only
6811 for (SmallVectorImpl<MachineBasicBlock*>::iterator
6812 I = MBBLPads.begin(), E = MBBLPads.end(); I != E; ++I)
6813 (*I)->setIsLandingPad(false);
6815 // The instruction is gone now.
6816 MI->eraseFromParent();
6822 MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) {
6823 for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
6824 E = MBB->succ_end(); I != E; ++I)
6827 llvm_unreachable("Expecting a BB with two successors!");
6830 /// Return the load opcode for a given load size. If load size >= 8,
6831 /// neon opcode will be returned.
6832 static unsigned getLdOpcode(unsigned LdSize, bool IsThumb1, bool IsThumb2) {
6834 return LdSize == 16 ? ARM::VLD1q32wb_fixed
6835 : LdSize == 8 ? ARM::VLD1d32wb_fixed : 0;
6837 return LdSize == 4 ? ARM::tLDRi
6838 : LdSize == 2 ? ARM::tLDRHi
6839 : LdSize == 1 ? ARM::tLDRBi : 0;
6841 return LdSize == 4 ? ARM::t2LDR_POST
6842 : LdSize == 2 ? ARM::t2LDRH_POST
6843 : LdSize == 1 ? ARM::t2LDRB_POST : 0;
6844 return LdSize == 4 ? ARM::LDR_POST_IMM
6845 : LdSize == 2 ? ARM::LDRH_POST
6846 : LdSize == 1 ? ARM::LDRB_POST_IMM : 0;
6849 /// Return the store opcode for a given store size. If store size >= 8,
6850 /// neon opcode will be returned.
6851 static unsigned getStOpcode(unsigned StSize, bool IsThumb1, bool IsThumb2) {
6853 return StSize == 16 ? ARM::VST1q32wb_fixed
6854 : StSize == 8 ? ARM::VST1d32wb_fixed : 0;
6856 return StSize == 4 ? ARM::tSTRi
6857 : StSize == 2 ? ARM::tSTRHi
6858 : StSize == 1 ? ARM::tSTRBi : 0;
6860 return StSize == 4 ? ARM::t2STR_POST
6861 : StSize == 2 ? ARM::t2STRH_POST
6862 : StSize == 1 ? ARM::t2STRB_POST : 0;
6863 return StSize == 4 ? ARM::STR_POST_IMM
6864 : StSize == 2 ? ARM::STRH_POST
6865 : StSize == 1 ? ARM::STRB_POST_IMM : 0;
6868 /// Emit a post-increment load operation with given size. The instructions
6869 /// will be added to BB at Pos.
6870 static void emitPostLd(MachineBasicBlock *BB, MachineInstr *Pos,
6871 const TargetInstrInfo *TII, DebugLoc dl,
6872 unsigned LdSize, unsigned Data, unsigned AddrIn,
6873 unsigned AddrOut, bool IsThumb1, bool IsThumb2) {
6874 unsigned LdOpc = getLdOpcode(LdSize, IsThumb1, IsThumb2);
6875 assert(LdOpc != 0 && "Should have a load opcode");
6877 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
6878 .addReg(AddrOut, RegState::Define).addReg(AddrIn)
6880 } else if (IsThumb1) {
6881 // load + update AddrIn
6882 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
6883 .addReg(AddrIn).addImm(0));
6884 MachineInstrBuilder MIB =
6885 BuildMI(*BB, Pos, dl, TII->get(ARM::tADDi8), AddrOut);
6886 MIB = AddDefaultT1CC(MIB);
6887 MIB.addReg(AddrIn).addImm(LdSize);
6888 AddDefaultPred(MIB);
6889 } else if (IsThumb2) {
6890 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
6891 .addReg(AddrOut, RegState::Define).addReg(AddrIn)
6894 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
6895 .addReg(AddrOut, RegState::Define).addReg(AddrIn)
6896 .addReg(0).addImm(LdSize));
6900 /// Emit a post-increment store operation with given size. The instructions
6901 /// will be added to BB at Pos.
6902 static void emitPostSt(MachineBasicBlock *BB, MachineInstr *Pos,
6903 const TargetInstrInfo *TII, DebugLoc dl,
6904 unsigned StSize, unsigned Data, unsigned AddrIn,
6905 unsigned AddrOut, bool IsThumb1, bool IsThumb2) {
6906 unsigned StOpc = getStOpcode(StSize, IsThumb1, IsThumb2);
6907 assert(StOpc != 0 && "Should have a store opcode");
6909 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc), AddrOut)
6910 .addReg(AddrIn).addImm(0).addReg(Data));
6911 } else if (IsThumb1) {
6912 // store + update AddrIn
6913 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc)).addReg(Data)
6914 .addReg(AddrIn).addImm(0));
6915 MachineInstrBuilder MIB =
6916 BuildMI(*BB, Pos, dl, TII->get(ARM::tADDi8), AddrOut);
6917 MIB = AddDefaultT1CC(MIB);
6918 MIB.addReg(AddrIn).addImm(StSize);
6919 AddDefaultPred(MIB);
6920 } else if (IsThumb2) {
6921 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc), AddrOut)
6922 .addReg(Data).addReg(AddrIn).addImm(StSize));
6924 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc), AddrOut)
6925 .addReg(Data).addReg(AddrIn).addReg(0)
6931 ARMTargetLowering::EmitStructByval(MachineInstr *MI,
6932 MachineBasicBlock *BB) const {
6933 // This pseudo instruction has 3 operands: dst, src, size
6934 // We expand it to a loop if size > Subtarget->getMaxInlineSizeThreshold().
6935 // Otherwise, we will generate unrolled scalar copies.
6936 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6937 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6938 MachineFunction::iterator It = BB;
6941 unsigned dest = MI->getOperand(0).getReg();
6942 unsigned src = MI->getOperand(1).getReg();
6943 unsigned SizeVal = MI->getOperand(2).getImm();
6944 unsigned Align = MI->getOperand(3).getImm();
6945 DebugLoc dl = MI->getDebugLoc();
6947 MachineFunction *MF = BB->getParent();
6948 MachineRegisterInfo &MRI = MF->getRegInfo();
6949 unsigned UnitSize = 0;
6950 const TargetRegisterClass *TRC = nullptr;
6951 const TargetRegisterClass *VecTRC = nullptr;
6953 bool IsThumb1 = Subtarget->isThumb1Only();
6954 bool IsThumb2 = Subtarget->isThumb2();
6958 } else if (Align & 2) {
6961 // Check whether we can use NEON instructions.
6962 if (!MF->getFunction()->getAttributes().
6963 hasAttribute(AttributeSet::FunctionIndex,
6964 Attribute::NoImplicitFloat) &&
6965 Subtarget->hasNEON()) {
6966 if ((Align % 16 == 0) && SizeVal >= 16)
6968 else if ((Align % 8 == 0) && SizeVal >= 8)
6971 // Can't use NEON instructions.
6976 // Select the correct opcode and register class for unit size load/store
6977 bool IsNeon = UnitSize >= 8;
6978 TRC = (IsThumb1 || IsThumb2) ? (const TargetRegisterClass *)&ARM::tGPRRegClass
6979 : (const TargetRegisterClass *)&ARM::GPRRegClass;
6981 VecTRC = UnitSize == 16
6982 ? (const TargetRegisterClass *)&ARM::DPairRegClass
6984 ? (const TargetRegisterClass *)&ARM::DPRRegClass
6987 unsigned BytesLeft = SizeVal % UnitSize;
6988 unsigned LoopSize = SizeVal - BytesLeft;
6990 if (SizeVal <= Subtarget->getMaxInlineSizeThreshold()) {
6991 // Use LDR and STR to copy.
6992 // [scratch, srcOut] = LDR_POST(srcIn, UnitSize)
6993 // [destOut] = STR_POST(scratch, destIn, UnitSize)
6994 unsigned srcIn = src;
6995 unsigned destIn = dest;
6996 for (unsigned i = 0; i < LoopSize; i+=UnitSize) {
6997 unsigned srcOut = MRI.createVirtualRegister(TRC);
6998 unsigned destOut = MRI.createVirtualRegister(TRC);
6999 unsigned scratch = MRI.createVirtualRegister(IsNeon ? VecTRC : TRC);
7000 emitPostLd(BB, MI, TII, dl, UnitSize, scratch, srcIn, srcOut,
7001 IsThumb1, IsThumb2);
7002 emitPostSt(BB, MI, TII, dl, UnitSize, scratch, destIn, destOut,
7003 IsThumb1, IsThumb2);
7008 // Handle the leftover bytes with LDRB and STRB.
7009 // [scratch, srcOut] = LDRB_POST(srcIn, 1)
7010 // [destOut] = STRB_POST(scratch, destIn, 1)
7011 for (unsigned i = 0; i < BytesLeft; i++) {
7012 unsigned srcOut = MRI.createVirtualRegister(TRC);
7013 unsigned destOut = MRI.createVirtualRegister(TRC);
7014 unsigned scratch = MRI.createVirtualRegister(TRC);
7015 emitPostLd(BB, MI, TII, dl, 1, scratch, srcIn, srcOut,
7016 IsThumb1, IsThumb2);
7017 emitPostSt(BB, MI, TII, dl, 1, scratch, destIn, destOut,
7018 IsThumb1, IsThumb2);
7022 MI->eraseFromParent(); // The instruction is gone now.
7026 // Expand the pseudo op to a loop.
7029 // movw varEnd, # --> with thumb2
7031 // ldrcp varEnd, idx --> without thumb2
7032 // fallthrough --> loopMBB
7034 // PHI varPhi, varEnd, varLoop
7035 // PHI srcPhi, src, srcLoop
7036 // PHI destPhi, dst, destLoop
7037 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
7038 // [destLoop] = STR_POST(scratch, destPhi, UnitSize)
7039 // subs varLoop, varPhi, #UnitSize
7041 // fallthrough --> exitMBB
7043 // epilogue to handle left-over bytes
7044 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
7045 // [destOut] = STRB_POST(scratch, destLoop, 1)
7046 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
7047 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
7048 MF->insert(It, loopMBB);
7049 MF->insert(It, exitMBB);
7051 // Transfer the remainder of BB and its successor edges to exitMBB.
7052 exitMBB->splice(exitMBB->begin(), BB,
7053 std::next(MachineBasicBlock::iterator(MI)), BB->end());
7054 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
7056 // Load an immediate to varEnd.
7057 unsigned varEnd = MRI.createVirtualRegister(TRC);
7059 unsigned Vtmp = varEnd;
7060 if ((LoopSize & 0xFFFF0000) != 0)
7061 Vtmp = MRI.createVirtualRegister(TRC);
7062 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2MOVi16), Vtmp)
7063 .addImm(LoopSize & 0xFFFF));
7065 if ((LoopSize & 0xFFFF0000) != 0)
7066 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2MOVTi16), varEnd)
7067 .addReg(Vtmp).addImm(LoopSize >> 16));
7069 MachineConstantPool *ConstantPool = MF->getConstantPool();
7070 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
7071 const Constant *C = ConstantInt::get(Int32Ty, LoopSize);
7073 // MachineConstantPool wants an explicit alignment.
7074 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
7076 Align = getDataLayout()->getTypeAllocSize(C->getType());
7077 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
7080 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(ARM::tLDRpci)).addReg(
7081 varEnd, RegState::Define).addConstantPoolIndex(Idx));
7083 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(ARM::LDRcp)).addReg(
7084 varEnd, RegState::Define).addConstantPoolIndex(Idx).addImm(0));
7086 BB->addSuccessor(loopMBB);
7088 // Generate the loop body:
7089 // varPhi = PHI(varLoop, varEnd)
7090 // srcPhi = PHI(srcLoop, src)
7091 // destPhi = PHI(destLoop, dst)
7092 MachineBasicBlock *entryBB = BB;
7094 unsigned varLoop = MRI.createVirtualRegister(TRC);
7095 unsigned varPhi = MRI.createVirtualRegister(TRC);
7096 unsigned srcLoop = MRI.createVirtualRegister(TRC);
7097 unsigned srcPhi = MRI.createVirtualRegister(TRC);
7098 unsigned destLoop = MRI.createVirtualRegister(TRC);
7099 unsigned destPhi = MRI.createVirtualRegister(TRC);
7101 BuildMI(*BB, BB->begin(), dl, TII->get(ARM::PHI), varPhi)
7102 .addReg(varLoop).addMBB(loopMBB)
7103 .addReg(varEnd).addMBB(entryBB);
7104 BuildMI(BB, dl, TII->get(ARM::PHI), srcPhi)
7105 .addReg(srcLoop).addMBB(loopMBB)
7106 .addReg(src).addMBB(entryBB);
7107 BuildMI(BB, dl, TII->get(ARM::PHI), destPhi)
7108 .addReg(destLoop).addMBB(loopMBB)
7109 .addReg(dest).addMBB(entryBB);
7111 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
7112 // [destLoop] = STR_POST(scratch, destPhi, UnitSiz)
7113 unsigned scratch = MRI.createVirtualRegister(IsNeon ? VecTRC : TRC);
7114 emitPostLd(BB, BB->end(), TII, dl, UnitSize, scratch, srcPhi, srcLoop,
7115 IsThumb1, IsThumb2);
7116 emitPostSt(BB, BB->end(), TII, dl, UnitSize, scratch, destPhi, destLoop,
7117 IsThumb1, IsThumb2);
7119 // Decrement loop variable by UnitSize.
7121 MachineInstrBuilder MIB =
7122 BuildMI(*BB, BB->end(), dl, TII->get(ARM::tSUBi8), varLoop);
7123 MIB = AddDefaultT1CC(MIB);
7124 MIB.addReg(varPhi).addImm(UnitSize);
7125 AddDefaultPred(MIB);
7127 MachineInstrBuilder MIB =
7128 BuildMI(*BB, BB->end(), dl,
7129 TII->get(IsThumb2 ? ARM::t2SUBri : ARM::SUBri), varLoop);
7130 AddDefaultCC(AddDefaultPred(MIB.addReg(varPhi).addImm(UnitSize)));
7131 MIB->getOperand(5).setReg(ARM::CPSR);
7132 MIB->getOperand(5).setIsDef(true);
7134 BuildMI(*BB, BB->end(), dl,
7135 TII->get(IsThumb1 ? ARM::tBcc : IsThumb2 ? ARM::t2Bcc : ARM::Bcc))
7136 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
7138 // loopMBB can loop back to loopMBB or fall through to exitMBB.
7139 BB->addSuccessor(loopMBB);
7140 BB->addSuccessor(exitMBB);
7142 // Add epilogue to handle BytesLeft.
7144 MachineInstr *StartOfExit = exitMBB->begin();
7146 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
7147 // [destOut] = STRB_POST(scratch, destLoop, 1)
7148 unsigned srcIn = srcLoop;
7149 unsigned destIn = destLoop;
7150 for (unsigned i = 0; i < BytesLeft; i++) {
7151 unsigned srcOut = MRI.createVirtualRegister(TRC);
7152 unsigned destOut = MRI.createVirtualRegister(TRC);
7153 unsigned scratch = MRI.createVirtualRegister(TRC);
7154 emitPostLd(BB, StartOfExit, TII, dl, 1, scratch, srcIn, srcOut,
7155 IsThumb1, IsThumb2);
7156 emitPostSt(BB, StartOfExit, TII, dl, 1, scratch, destIn, destOut,
7157 IsThumb1, IsThumb2);
7162 MI->eraseFromParent(); // The instruction is gone now.
7167 ARMTargetLowering::EmitLowered__chkstk(MachineInstr *MI,
7168 MachineBasicBlock *MBB) const {
7169 const TargetMachine &TM = getTargetMachine();
7170 const TargetInstrInfo &TII = *TM.getInstrInfo();
7171 DebugLoc DL = MI->getDebugLoc();
7173 assert(Subtarget->isTargetWindows() &&
7174 "__chkstk is only supported on Windows");
7175 assert(Subtarget->isThumb2() && "Windows on ARM requires Thumb-2 mode");
7177 // __chkstk takes the number of words to allocate on the stack in R4, and
7178 // returns the stack adjustment in number of bytes in R4. This will not
7179 // clober any other registers (other than the obvious lr).
7181 // Although, technically, IP should be considered a register which may be
7182 // clobbered, the call itself will not touch it. Windows on ARM is a pure
7183 // thumb-2 environment, so there is no interworking required. As a result, we
7184 // do not expect a veneer to be emitted by the linker, clobbering IP.
7186 // Each module receives its own copy of __chkstk, so no import thunk is
7187 // required, again, ensuring that IP is not clobbered.
7189 // Finally, although some linkers may theoretically provide a trampoline for
7190 // out of range calls (which is quite common due to a 32M range limitation of
7191 // branches for Thumb), we can generate the long-call version via
7192 // -mcmodel=large, alleviating the need for the trampoline which may clobber
7195 switch (TM.getCodeModel()) {
7196 case CodeModel::Small:
7197 case CodeModel::Medium:
7198 case CodeModel::Default:
7199 case CodeModel::Kernel:
7200 BuildMI(*MBB, MI, DL, TII.get(ARM::tBL))
7201 .addImm((unsigned)ARMCC::AL).addReg(0)
7202 .addExternalSymbol("__chkstk")
7203 .addReg(ARM::R4, RegState::Implicit | RegState::Kill)
7204 .addReg(ARM::R4, RegState::Implicit | RegState::Define)
7205 .addReg(ARM::R12, RegState::Implicit | RegState::Define | RegState::Dead);
7207 case CodeModel::Large:
7208 case CodeModel::JITDefault: {
7209 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
7210 unsigned Reg = MRI.createVirtualRegister(&ARM::rGPRRegClass);
7212 BuildMI(*MBB, MI, DL, TII.get(ARM::t2MOVi32imm), Reg)
7213 .addExternalSymbol("__chkstk");
7214 BuildMI(*MBB, MI, DL, TII.get(ARM::tBLXr))
7215 .addImm((unsigned)ARMCC::AL).addReg(0)
7216 .addReg(Reg, RegState::Kill)
7217 .addReg(ARM::R4, RegState::Implicit | RegState::Kill)
7218 .addReg(ARM::R4, RegState::Implicit | RegState::Define)
7219 .addReg(ARM::R12, RegState::Implicit | RegState::Define | RegState::Dead);
7224 AddDefaultCC(AddDefaultPred(BuildMI(*MBB, MI, DL, TII.get(ARM::t2SUBrr),
7226 .addReg(ARM::SP, RegState::Define)
7227 .addReg(ARM::R4, RegState::Kill)));
7229 MI->eraseFromParent();
7234 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
7235 MachineBasicBlock *BB) const {
7236 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7237 DebugLoc dl = MI->getDebugLoc();
7238 bool isThumb2 = Subtarget->isThumb2();
7239 switch (MI->getOpcode()) {
7242 llvm_unreachable("Unexpected instr type to insert");
7244 // The Thumb2 pre-indexed stores have the same MI operands, they just
7245 // define them differently in the .td files from the isel patterns, so
7246 // they need pseudos.
7247 case ARM::t2STR_preidx:
7248 MI->setDesc(TII->get(ARM::t2STR_PRE));
7250 case ARM::t2STRB_preidx:
7251 MI->setDesc(TII->get(ARM::t2STRB_PRE));
7253 case ARM::t2STRH_preidx:
7254 MI->setDesc(TII->get(ARM::t2STRH_PRE));
7257 case ARM::STRi_preidx:
7258 case ARM::STRBi_preidx: {
7259 unsigned NewOpc = MI->getOpcode() == ARM::STRi_preidx ?
7260 ARM::STR_PRE_IMM : ARM::STRB_PRE_IMM;
7261 // Decode the offset.
7262 unsigned Offset = MI->getOperand(4).getImm();
7263 bool isSub = ARM_AM::getAM2Op(Offset) == ARM_AM::sub;
7264 Offset = ARM_AM::getAM2Offset(Offset);
7268 MachineMemOperand *MMO = *MI->memoperands_begin();
7269 BuildMI(*BB, MI, dl, TII->get(NewOpc))
7270 .addOperand(MI->getOperand(0)) // Rn_wb
7271 .addOperand(MI->getOperand(1)) // Rt
7272 .addOperand(MI->getOperand(2)) // Rn
7273 .addImm(Offset) // offset (skip GPR==zero_reg)
7274 .addOperand(MI->getOperand(5)) // pred
7275 .addOperand(MI->getOperand(6))
7276 .addMemOperand(MMO);
7277 MI->eraseFromParent();
7280 case ARM::STRr_preidx:
7281 case ARM::STRBr_preidx:
7282 case ARM::STRH_preidx: {
7284 switch (MI->getOpcode()) {
7285 default: llvm_unreachable("unexpected opcode!");
7286 case ARM::STRr_preidx: NewOpc = ARM::STR_PRE_REG; break;
7287 case ARM::STRBr_preidx: NewOpc = ARM::STRB_PRE_REG; break;
7288 case ARM::STRH_preidx: NewOpc = ARM::STRH_PRE; break;
7290 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(NewOpc));
7291 for (unsigned i = 0; i < MI->getNumOperands(); ++i)
7292 MIB.addOperand(MI->getOperand(i));
7293 MI->eraseFromParent();
7297 case ARM::tMOVCCr_pseudo: {
7298 // To "insert" a SELECT_CC instruction, we actually have to insert the
7299 // diamond control-flow pattern. The incoming instruction knows the
7300 // destination vreg to set, the condition code register to branch on, the
7301 // true/false values to select between, and a branch opcode to use.
7302 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7303 MachineFunction::iterator It = BB;
7309 // cmpTY ccX, r1, r2
7311 // fallthrough --> copy0MBB
7312 MachineBasicBlock *thisMBB = BB;
7313 MachineFunction *F = BB->getParent();
7314 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
7315 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
7316 F->insert(It, copy0MBB);
7317 F->insert(It, sinkMBB);
7319 // Transfer the remainder of BB and its successor edges to sinkMBB.
7320 sinkMBB->splice(sinkMBB->begin(), BB,
7321 std::next(MachineBasicBlock::iterator(MI)), BB->end());
7322 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
7324 BB->addSuccessor(copy0MBB);
7325 BB->addSuccessor(sinkMBB);
7327 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
7328 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
7331 // %FalseValue = ...
7332 // # fallthrough to sinkMBB
7335 // Update machine-CFG edges
7336 BB->addSuccessor(sinkMBB);
7339 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
7342 BuildMI(*BB, BB->begin(), dl,
7343 TII->get(ARM::PHI), MI->getOperand(0).getReg())
7344 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
7345 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
7347 MI->eraseFromParent(); // The pseudo instruction is gone now.
7352 case ARM::BCCZi64: {
7353 // If there is an unconditional branch to the other successor, remove it.
7354 BB->erase(std::next(MachineBasicBlock::iterator(MI)), BB->end());
7356 // Compare both parts that make up the double comparison separately for
7358 bool RHSisZero = MI->getOpcode() == ARM::BCCZi64;
7360 unsigned LHS1 = MI->getOperand(1).getReg();
7361 unsigned LHS2 = MI->getOperand(2).getReg();
7363 AddDefaultPred(BuildMI(BB, dl,
7364 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7365 .addReg(LHS1).addImm(0));
7366 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7367 .addReg(LHS2).addImm(0)
7368 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
7370 unsigned RHS1 = MI->getOperand(3).getReg();
7371 unsigned RHS2 = MI->getOperand(4).getReg();
7372 AddDefaultPred(BuildMI(BB, dl,
7373 TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
7374 .addReg(LHS1).addReg(RHS1));
7375 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
7376 .addReg(LHS2).addReg(RHS2)
7377 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
7380 MachineBasicBlock *destMBB = MI->getOperand(RHSisZero ? 3 : 5).getMBB();
7381 MachineBasicBlock *exitMBB = OtherSucc(BB, destMBB);
7382 if (MI->getOperand(0).getImm() == ARMCC::NE)
7383 std::swap(destMBB, exitMBB);
7385 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
7386 .addMBB(destMBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
7388 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2B)).addMBB(exitMBB));
7390 BuildMI(BB, dl, TII->get(ARM::B)) .addMBB(exitMBB);
7392 MI->eraseFromParent(); // The pseudo instruction is gone now.
7396 case ARM::Int_eh_sjlj_setjmp:
7397 case ARM::Int_eh_sjlj_setjmp_nofp:
7398 case ARM::tInt_eh_sjlj_setjmp:
7399 case ARM::t2Int_eh_sjlj_setjmp:
7400 case ARM::t2Int_eh_sjlj_setjmp_nofp:
7401 EmitSjLjDispatchBlock(MI, BB);
7406 // To insert an ABS instruction, we have to insert the
7407 // diamond control-flow pattern. The incoming instruction knows the
7408 // source vreg to test against 0, the destination vreg to set,
7409 // the condition code register to branch on, the
7410 // true/false values to select between, and a branch opcode to use.
7415 // BCC (branch to SinkBB if V0 >= 0)
7416 // RSBBB: V3 = RSBri V2, 0 (compute ABS if V2 < 0)
7417 // SinkBB: V1 = PHI(V2, V3)
7418 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7419 MachineFunction::iterator BBI = BB;
7421 MachineFunction *Fn = BB->getParent();
7422 MachineBasicBlock *RSBBB = Fn->CreateMachineBasicBlock(LLVM_BB);
7423 MachineBasicBlock *SinkBB = Fn->CreateMachineBasicBlock(LLVM_BB);
7424 Fn->insert(BBI, RSBBB);
7425 Fn->insert(BBI, SinkBB);
7427 unsigned int ABSSrcReg = MI->getOperand(1).getReg();
7428 unsigned int ABSDstReg = MI->getOperand(0).getReg();
7429 bool isThumb2 = Subtarget->isThumb2();
7430 MachineRegisterInfo &MRI = Fn->getRegInfo();
7431 // In Thumb mode S must not be specified if source register is the SP or
7432 // PC and if destination register is the SP, so restrict register class
7433 unsigned NewRsbDstReg = MRI.createVirtualRegister(isThumb2 ?
7434 (const TargetRegisterClass*)&ARM::rGPRRegClass :
7435 (const TargetRegisterClass*)&ARM::GPRRegClass);
7437 // Transfer the remainder of BB and its successor edges to sinkMBB.
7438 SinkBB->splice(SinkBB->begin(), BB,
7439 std::next(MachineBasicBlock::iterator(MI)), BB->end());
7440 SinkBB->transferSuccessorsAndUpdatePHIs(BB);
7442 BB->addSuccessor(RSBBB);
7443 BB->addSuccessor(SinkBB);
7445 // fall through to SinkMBB
7446 RSBBB->addSuccessor(SinkBB);
7448 // insert a cmp at the end of BB
7449 AddDefaultPred(BuildMI(BB, dl,
7450 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7451 .addReg(ABSSrcReg).addImm(0));
7453 // insert a bcc with opposite CC to ARMCC::MI at the end of BB
7455 TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)).addMBB(SinkBB)
7456 .addImm(ARMCC::getOppositeCondition(ARMCC::MI)).addReg(ARM::CPSR);
7458 // insert rsbri in RSBBB
7459 // Note: BCC and rsbri will be converted into predicated rsbmi
7460 // by if-conversion pass
7461 BuildMI(*RSBBB, RSBBB->begin(), dl,
7462 TII->get(isThumb2 ? ARM::t2RSBri : ARM::RSBri), NewRsbDstReg)
7463 .addReg(ABSSrcReg, RegState::Kill)
7464 .addImm(0).addImm((unsigned)ARMCC::AL).addReg(0).addReg(0);
7466 // insert PHI in SinkBB,
7467 // reuse ABSDstReg to not change uses of ABS instruction
7468 BuildMI(*SinkBB, SinkBB->begin(), dl,
7469 TII->get(ARM::PHI), ABSDstReg)
7470 .addReg(NewRsbDstReg).addMBB(RSBBB)
7471 .addReg(ABSSrcReg).addMBB(BB);
7473 // remove ABS instruction
7474 MI->eraseFromParent();
7476 // return last added BB
7479 case ARM::COPY_STRUCT_BYVAL_I32:
7481 return EmitStructByval(MI, BB);
7482 case ARM::WIN__CHKSTK:
7483 return EmitLowered__chkstk(MI, BB);
7487 void ARMTargetLowering::AdjustInstrPostInstrSelection(MachineInstr *MI,
7488 SDNode *Node) const {
7489 if (!MI->hasPostISelHook()) {
7490 assert(!convertAddSubFlagsOpcode(MI->getOpcode()) &&
7491 "Pseudo flag-setting opcodes must be marked with 'hasPostISelHook'");
7495 const MCInstrDesc *MCID = &MI->getDesc();
7496 // Adjust potentially 's' setting instructions after isel, i.e. ADC, SBC, RSB,
7497 // RSC. Coming out of isel, they have an implicit CPSR def, but the optional
7498 // operand is still set to noreg. If needed, set the optional operand's
7499 // register to CPSR, and remove the redundant implicit def.
7501 // e.g. ADCS (..., CPSR<imp-def>) -> ADC (... opt:CPSR<def>).
7503 // Rename pseudo opcodes.
7504 unsigned NewOpc = convertAddSubFlagsOpcode(MI->getOpcode());
7506 const ARMBaseInstrInfo *TII =
7507 static_cast<const ARMBaseInstrInfo*>(getTargetMachine().getInstrInfo());
7508 MCID = &TII->get(NewOpc);
7510 assert(MCID->getNumOperands() == MI->getDesc().getNumOperands() + 1 &&
7511 "converted opcode should be the same except for cc_out");
7515 // Add the optional cc_out operand
7516 MI->addOperand(MachineOperand::CreateReg(0, /*isDef=*/true));
7518 unsigned ccOutIdx = MCID->getNumOperands() - 1;
7520 // Any ARM instruction that sets the 's' bit should specify an optional
7521 // "cc_out" operand in the last operand position.
7522 if (!MI->hasOptionalDef() || !MCID->OpInfo[ccOutIdx].isOptionalDef()) {
7523 assert(!NewOpc && "Optional cc_out operand required");
7526 // Look for an implicit def of CPSR added by MachineInstr ctor. Remove it
7527 // since we already have an optional CPSR def.
7528 bool definesCPSR = false;
7529 bool deadCPSR = false;
7530 for (unsigned i = MCID->getNumOperands(), e = MI->getNumOperands();
7532 const MachineOperand &MO = MI->getOperand(i);
7533 if (MO.isReg() && MO.isDef() && MO.getReg() == ARM::CPSR) {
7537 MI->RemoveOperand(i);
7542 assert(!NewOpc && "Optional cc_out operand required");
7545 assert(deadCPSR == !Node->hasAnyUseOfValue(1) && "inconsistent dead flag");
7547 assert(!MI->getOperand(ccOutIdx).getReg() &&
7548 "expect uninitialized optional cc_out operand");
7552 // If this instruction was defined with an optional CPSR def and its dag node
7553 // had a live implicit CPSR def, then activate the optional CPSR def.
7554 MachineOperand &MO = MI->getOperand(ccOutIdx);
7555 MO.setReg(ARM::CPSR);
7559 //===----------------------------------------------------------------------===//
7560 // ARM Optimization Hooks
7561 //===----------------------------------------------------------------------===//
7563 // Helper function that checks if N is a null or all ones constant.
7564 static inline bool isZeroOrAllOnes(SDValue N, bool AllOnes) {
7565 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N);
7568 return AllOnes ? C->isAllOnesValue() : C->isNullValue();
7571 // Return true if N is conditionally 0 or all ones.
7572 // Detects these expressions where cc is an i1 value:
7574 // (select cc 0, y) [AllOnes=0]
7575 // (select cc y, 0) [AllOnes=0]
7576 // (zext cc) [AllOnes=0]
7577 // (sext cc) [AllOnes=0/1]
7578 // (select cc -1, y) [AllOnes=1]
7579 // (select cc y, -1) [AllOnes=1]
7581 // Invert is set when N is the null/all ones constant when CC is false.
7582 // OtherOp is set to the alternative value of N.
7583 static bool isConditionalZeroOrAllOnes(SDNode *N, bool AllOnes,
7584 SDValue &CC, bool &Invert,
7586 SelectionDAG &DAG) {
7587 switch (N->getOpcode()) {
7588 default: return false;
7590 CC = N->getOperand(0);
7591 SDValue N1 = N->getOperand(1);
7592 SDValue N2 = N->getOperand(2);
7593 if (isZeroOrAllOnes(N1, AllOnes)) {
7598 if (isZeroOrAllOnes(N2, AllOnes)) {
7605 case ISD::ZERO_EXTEND:
7606 // (zext cc) can never be the all ones value.
7610 case ISD::SIGN_EXTEND: {
7611 EVT VT = N->getValueType(0);
7612 CC = N->getOperand(0);
7613 if (CC.getValueType() != MVT::i1)
7617 // When looking for an AllOnes constant, N is an sext, and the 'other'
7619 OtherOp = DAG.getConstant(0, VT);
7620 else if (N->getOpcode() == ISD::ZERO_EXTEND)
7621 // When looking for a 0 constant, N can be zext or sext.
7622 OtherOp = DAG.getConstant(1, VT);
7624 OtherOp = DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), VT);
7630 // Combine a constant select operand into its use:
7632 // (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
7633 // (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
7634 // (and (select cc, -1, c), x) -> (select cc, x, (and, x, c)) [AllOnes=1]
7635 // (or (select cc, 0, c), x) -> (select cc, x, (or, x, c))
7636 // (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c))
7638 // The transform is rejected if the select doesn't have a constant operand that
7639 // is null, or all ones when AllOnes is set.
7641 // Also recognize sext/zext from i1:
7643 // (add (zext cc), x) -> (select cc (add x, 1), x)
7644 // (add (sext cc), x) -> (select cc (add x, -1), x)
7646 // These transformations eventually create predicated instructions.
7648 // @param N The node to transform.
7649 // @param Slct The N operand that is a select.
7650 // @param OtherOp The other N operand (x above).
7651 // @param DCI Context.
7652 // @param AllOnes Require the select constant to be all ones instead of null.
7653 // @returns The new node, or SDValue() on failure.
7655 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
7656 TargetLowering::DAGCombinerInfo &DCI,
7657 bool AllOnes = false) {
7658 SelectionDAG &DAG = DCI.DAG;
7659 EVT VT = N->getValueType(0);
7660 SDValue NonConstantVal;
7663 if (!isConditionalZeroOrAllOnes(Slct.getNode(), AllOnes, CCOp, SwapSelectOps,
7664 NonConstantVal, DAG))
7667 // Slct is now know to be the desired identity constant when CC is true.
7668 SDValue TrueVal = OtherOp;
7669 SDValue FalseVal = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
7670 OtherOp, NonConstantVal);
7671 // Unless SwapSelectOps says CC should be false.
7673 std::swap(TrueVal, FalseVal);
7675 return DAG.getNode(ISD::SELECT, SDLoc(N), VT,
7676 CCOp, TrueVal, FalseVal);
7679 // Attempt combineSelectAndUse on each operand of a commutative operator N.
7681 SDValue combineSelectAndUseCommutative(SDNode *N, bool AllOnes,
7682 TargetLowering::DAGCombinerInfo &DCI) {
7683 SDValue N0 = N->getOperand(0);
7684 SDValue N1 = N->getOperand(1);
7685 if (N0.getNode()->hasOneUse()) {
7686 SDValue Result = combineSelectAndUse(N, N0, N1, DCI, AllOnes);
7687 if (Result.getNode())
7690 if (N1.getNode()->hasOneUse()) {
7691 SDValue Result = combineSelectAndUse(N, N1, N0, DCI, AllOnes);
7692 if (Result.getNode())
7698 // AddCombineToVPADDL- For pair-wise add on neon, use the vpaddl instruction
7699 // (only after legalization).
7700 static SDValue AddCombineToVPADDL(SDNode *N, SDValue N0, SDValue N1,
7701 TargetLowering::DAGCombinerInfo &DCI,
7702 const ARMSubtarget *Subtarget) {
7704 // Only perform optimization if after legalize, and if NEON is available. We
7705 // also expected both operands to be BUILD_VECTORs.
7706 if (DCI.isBeforeLegalize() || !Subtarget->hasNEON()
7707 || N0.getOpcode() != ISD::BUILD_VECTOR
7708 || N1.getOpcode() != ISD::BUILD_VECTOR)
7711 // Check output type since VPADDL operand elements can only be 8, 16, or 32.
7712 EVT VT = N->getValueType(0);
7713 if (!VT.isInteger() || VT.getVectorElementType() == MVT::i64)
7716 // Check that the vector operands are of the right form.
7717 // N0 and N1 are BUILD_VECTOR nodes with N number of EXTRACT_VECTOR
7718 // operands, where N is the size of the formed vector.
7719 // Each EXTRACT_VECTOR should have the same input vector and odd or even
7720 // index such that we have a pair wise add pattern.
7722 // Grab the vector that all EXTRACT_VECTOR nodes should be referencing.
7723 if (N0->getOperand(0)->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
7725 SDValue Vec = N0->getOperand(0)->getOperand(0);
7726 SDNode *V = Vec.getNode();
7727 unsigned nextIndex = 0;
7729 // For each operands to the ADD which are BUILD_VECTORs,
7730 // check to see if each of their operands are an EXTRACT_VECTOR with
7731 // the same vector and appropriate index.
7732 for (unsigned i = 0, e = N0->getNumOperands(); i != e; ++i) {
7733 if (N0->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT
7734 && N1->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
7736 SDValue ExtVec0 = N0->getOperand(i);
7737 SDValue ExtVec1 = N1->getOperand(i);
7739 // First operand is the vector, verify its the same.
7740 if (V != ExtVec0->getOperand(0).getNode() ||
7741 V != ExtVec1->getOperand(0).getNode())
7744 // Second is the constant, verify its correct.
7745 ConstantSDNode *C0 = dyn_cast<ConstantSDNode>(ExtVec0->getOperand(1));
7746 ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(ExtVec1->getOperand(1));
7748 // For the constant, we want to see all the even or all the odd.
7749 if (!C0 || !C1 || C0->getZExtValue() != nextIndex
7750 || C1->getZExtValue() != nextIndex+1)
7759 // Create VPADDL node.
7760 SelectionDAG &DAG = DCI.DAG;
7761 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7763 // Build operand list.
7764 SmallVector<SDValue, 8> Ops;
7765 Ops.push_back(DAG.getConstant(Intrinsic::arm_neon_vpaddls,
7766 TLI.getPointerTy()));
7768 // Input is the vector.
7771 // Get widened type and narrowed type.
7773 unsigned numElem = VT.getVectorNumElements();
7775 EVT inputLaneType = Vec.getValueType().getVectorElementType();
7776 switch (inputLaneType.getSimpleVT().SimpleTy) {
7777 case MVT::i8: widenType = MVT::getVectorVT(MVT::i16, numElem); break;
7778 case MVT::i16: widenType = MVT::getVectorVT(MVT::i32, numElem); break;
7779 case MVT::i32: widenType = MVT::getVectorVT(MVT::i64, numElem); break;
7781 llvm_unreachable("Invalid vector element type for padd optimization.");
7784 SDValue tmp = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N), widenType, Ops);
7785 unsigned ExtOp = VT.bitsGT(tmp.getValueType()) ? ISD::ANY_EXTEND : ISD::TRUNCATE;
7786 return DAG.getNode(ExtOp, SDLoc(N), VT, tmp);
7789 static SDValue findMUL_LOHI(SDValue V) {
7790 if (V->getOpcode() == ISD::UMUL_LOHI ||
7791 V->getOpcode() == ISD::SMUL_LOHI)
7796 static SDValue AddCombineTo64bitMLAL(SDNode *AddcNode,
7797 TargetLowering::DAGCombinerInfo &DCI,
7798 const ARMSubtarget *Subtarget) {
7800 if (Subtarget->isThumb1Only()) return SDValue();
7802 // Only perform the checks after legalize when the pattern is available.
7803 if (DCI.isBeforeLegalize()) return SDValue();
7805 // Look for multiply add opportunities.
7806 // The pattern is a ISD::UMUL_LOHI followed by two add nodes, where
7807 // each add nodes consumes a value from ISD::UMUL_LOHI and there is
7808 // a glue link from the first add to the second add.
7809 // If we find this pattern, we can replace the U/SMUL_LOHI, ADDC, and ADDE by
7810 // a S/UMLAL instruction.
7813 // \ / \ [no multiline comment]
7819 assert(AddcNode->getOpcode() == ISD::ADDC && "Expect an ADDC");
7820 SDValue AddcOp0 = AddcNode->getOperand(0);
7821 SDValue AddcOp1 = AddcNode->getOperand(1);
7823 // Check if the two operands are from the same mul_lohi node.
7824 if (AddcOp0.getNode() == AddcOp1.getNode())
7827 assert(AddcNode->getNumValues() == 2 &&
7828 AddcNode->getValueType(0) == MVT::i32 &&
7829 "Expect ADDC with two result values. First: i32");
7831 // Check that we have a glued ADDC node.
7832 if (AddcNode->getValueType(1) != MVT::Glue)
7835 // Check that the ADDC adds the low result of the S/UMUL_LOHI.
7836 if (AddcOp0->getOpcode() != ISD::UMUL_LOHI &&
7837 AddcOp0->getOpcode() != ISD::SMUL_LOHI &&
7838 AddcOp1->getOpcode() != ISD::UMUL_LOHI &&
7839 AddcOp1->getOpcode() != ISD::SMUL_LOHI)
7842 // Look for the glued ADDE.
7843 SDNode* AddeNode = AddcNode->getGluedUser();
7847 // Make sure it is really an ADDE.
7848 if (AddeNode->getOpcode() != ISD::ADDE)
7851 assert(AddeNode->getNumOperands() == 3 &&
7852 AddeNode->getOperand(2).getValueType() == MVT::Glue &&
7853 "ADDE node has the wrong inputs");
7855 // Check for the triangle shape.
7856 SDValue AddeOp0 = AddeNode->getOperand(0);
7857 SDValue AddeOp1 = AddeNode->getOperand(1);
7859 // Make sure that the ADDE operands are not coming from the same node.
7860 if (AddeOp0.getNode() == AddeOp1.getNode())
7863 // Find the MUL_LOHI node walking up ADDE's operands.
7864 bool IsLeftOperandMUL = false;
7865 SDValue MULOp = findMUL_LOHI(AddeOp0);
7866 if (MULOp == SDValue())
7867 MULOp = findMUL_LOHI(AddeOp1);
7869 IsLeftOperandMUL = true;
7870 if (MULOp == SDValue())
7873 // Figure out the right opcode.
7874 unsigned Opc = MULOp->getOpcode();
7875 unsigned FinalOpc = (Opc == ISD::SMUL_LOHI) ? ARMISD::SMLAL : ARMISD::UMLAL;
7877 // Figure out the high and low input values to the MLAL node.
7878 SDValue* HiMul = &MULOp;
7879 SDValue* HiAdd = nullptr;
7880 SDValue* LoMul = nullptr;
7881 SDValue* LowAdd = nullptr;
7883 if (IsLeftOperandMUL)
7889 if (AddcOp0->getOpcode() == Opc) {
7893 if (AddcOp1->getOpcode() == Opc) {
7901 if (LoMul->getNode() != HiMul->getNode())
7904 // Create the merged node.
7905 SelectionDAG &DAG = DCI.DAG;
7907 // Build operand list.
7908 SmallVector<SDValue, 8> Ops;
7909 Ops.push_back(LoMul->getOperand(0));
7910 Ops.push_back(LoMul->getOperand(1));
7911 Ops.push_back(*LowAdd);
7912 Ops.push_back(*HiAdd);
7914 SDValue MLALNode = DAG.getNode(FinalOpc, SDLoc(AddcNode),
7915 DAG.getVTList(MVT::i32, MVT::i32), Ops);
7917 // Replace the ADDs' nodes uses by the MLA node's values.
7918 SDValue HiMLALResult(MLALNode.getNode(), 1);
7919 DAG.ReplaceAllUsesOfValueWith(SDValue(AddeNode, 0), HiMLALResult);
7921 SDValue LoMLALResult(MLALNode.getNode(), 0);
7922 DAG.ReplaceAllUsesOfValueWith(SDValue(AddcNode, 0), LoMLALResult);
7924 // Return original node to notify the driver to stop replacing.
7925 SDValue resNode(AddcNode, 0);
7929 /// PerformADDCCombine - Target-specific dag combine transform from
7930 /// ISD::ADDC, ISD::ADDE, and ISD::MUL_LOHI to MLAL.
7931 static SDValue PerformADDCCombine(SDNode *N,
7932 TargetLowering::DAGCombinerInfo &DCI,
7933 const ARMSubtarget *Subtarget) {
7935 return AddCombineTo64bitMLAL(N, DCI, Subtarget);
7939 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
7940 /// operands N0 and N1. This is a helper for PerformADDCombine that is
7941 /// called with the default operands, and if that fails, with commuted
7943 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
7944 TargetLowering::DAGCombinerInfo &DCI,
7945 const ARMSubtarget *Subtarget){
7947 // Attempt to create vpaddl for this add.
7948 SDValue Result = AddCombineToVPADDL(N, N0, N1, DCI, Subtarget);
7949 if (Result.getNode())
7952 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
7953 if (N0.getNode()->hasOneUse()) {
7954 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
7955 if (Result.getNode()) return Result;
7960 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
7962 static SDValue PerformADDCombine(SDNode *N,
7963 TargetLowering::DAGCombinerInfo &DCI,
7964 const ARMSubtarget *Subtarget) {
7965 SDValue N0 = N->getOperand(0);
7966 SDValue N1 = N->getOperand(1);
7968 // First try with the default operand order.
7969 SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI, Subtarget);
7970 if (Result.getNode())
7973 // If that didn't work, try again with the operands commuted.
7974 return PerformADDCombineWithOperands(N, N1, N0, DCI, Subtarget);
7977 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
7979 static SDValue PerformSUBCombine(SDNode *N,
7980 TargetLowering::DAGCombinerInfo &DCI) {
7981 SDValue N0 = N->getOperand(0);
7982 SDValue N1 = N->getOperand(1);
7984 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
7985 if (N1.getNode()->hasOneUse()) {
7986 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
7987 if (Result.getNode()) return Result;
7993 /// PerformVMULCombine
7994 /// Distribute (A + B) * C to (A * C) + (B * C) to take advantage of the
7995 /// special multiplier accumulator forwarding.
8001 // However, for (A + B) * (A + B),
8008 static SDValue PerformVMULCombine(SDNode *N,
8009 TargetLowering::DAGCombinerInfo &DCI,
8010 const ARMSubtarget *Subtarget) {
8011 if (!Subtarget->hasVMLxForwarding())
8014 SelectionDAG &DAG = DCI.DAG;
8015 SDValue N0 = N->getOperand(0);
8016 SDValue N1 = N->getOperand(1);
8017 unsigned Opcode = N0.getOpcode();
8018 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
8019 Opcode != ISD::FADD && Opcode != ISD::FSUB) {
8020 Opcode = N1.getOpcode();
8021 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
8022 Opcode != ISD::FADD && Opcode != ISD::FSUB)
8030 EVT VT = N->getValueType(0);
8032 SDValue N00 = N0->getOperand(0);
8033 SDValue N01 = N0->getOperand(1);
8034 return DAG.getNode(Opcode, DL, VT,
8035 DAG.getNode(ISD::MUL, DL, VT, N00, N1),
8036 DAG.getNode(ISD::MUL, DL, VT, N01, N1));
8039 static SDValue PerformMULCombine(SDNode *N,
8040 TargetLowering::DAGCombinerInfo &DCI,
8041 const ARMSubtarget *Subtarget) {
8042 SelectionDAG &DAG = DCI.DAG;
8044 if (Subtarget->isThumb1Only())
8047 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
8050 EVT VT = N->getValueType(0);
8051 if (VT.is64BitVector() || VT.is128BitVector())
8052 return PerformVMULCombine(N, DCI, Subtarget);
8056 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
8060 int64_t MulAmt = C->getSExtValue();
8061 unsigned ShiftAmt = countTrailingZeros<uint64_t>(MulAmt);
8063 ShiftAmt = ShiftAmt & (32 - 1);
8064 SDValue V = N->getOperand(0);
8068 MulAmt >>= ShiftAmt;
8071 if (isPowerOf2_32(MulAmt - 1)) {
8072 // (mul x, 2^N + 1) => (add (shl x, N), x)
8073 Res = DAG.getNode(ISD::ADD, DL, VT,
8075 DAG.getNode(ISD::SHL, DL, VT,
8077 DAG.getConstant(Log2_32(MulAmt - 1),
8079 } else if (isPowerOf2_32(MulAmt + 1)) {
8080 // (mul x, 2^N - 1) => (sub (shl x, N), x)
8081 Res = DAG.getNode(ISD::SUB, DL, VT,
8082 DAG.getNode(ISD::SHL, DL, VT,
8084 DAG.getConstant(Log2_32(MulAmt + 1),
8090 uint64_t MulAmtAbs = -MulAmt;
8091 if (isPowerOf2_32(MulAmtAbs + 1)) {
8092 // (mul x, -(2^N - 1)) => (sub x, (shl x, N))
8093 Res = DAG.getNode(ISD::SUB, DL, VT,
8095 DAG.getNode(ISD::SHL, DL, VT,
8097 DAG.getConstant(Log2_32(MulAmtAbs + 1),
8099 } else if (isPowerOf2_32(MulAmtAbs - 1)) {
8100 // (mul x, -(2^N + 1)) => - (add (shl x, N), x)
8101 Res = DAG.getNode(ISD::ADD, DL, VT,
8103 DAG.getNode(ISD::SHL, DL, VT,
8105 DAG.getConstant(Log2_32(MulAmtAbs-1),
8107 Res = DAG.getNode(ISD::SUB, DL, VT,
8108 DAG.getConstant(0, MVT::i32),Res);
8115 Res = DAG.getNode(ISD::SHL, DL, VT,
8116 Res, DAG.getConstant(ShiftAmt, MVT::i32));
8118 // Do not add new nodes to DAG combiner worklist.
8119 DCI.CombineTo(N, Res, false);
8123 static SDValue PerformANDCombine(SDNode *N,
8124 TargetLowering::DAGCombinerInfo &DCI,
8125 const ARMSubtarget *Subtarget) {
8127 // Attempt to use immediate-form VBIC
8128 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
8130 EVT VT = N->getValueType(0);
8131 SelectionDAG &DAG = DCI.DAG;
8133 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8136 APInt SplatBits, SplatUndef;
8137 unsigned SplatBitSize;
8140 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
8141 if (SplatBitSize <= 64) {
8143 SDValue Val = isNEONModifiedImm((~SplatBits).getZExtValue(),
8144 SplatUndef.getZExtValue(), SplatBitSize,
8145 DAG, VbicVT, VT.is128BitVector(),
8147 if (Val.getNode()) {
8149 DAG.getNode(ISD::BITCAST, dl, VbicVT, N->getOperand(0));
8150 SDValue Vbic = DAG.getNode(ARMISD::VBICIMM, dl, VbicVT, Input, Val);
8151 return DAG.getNode(ISD::BITCAST, dl, VT, Vbic);
8156 if (!Subtarget->isThumb1Only()) {
8157 // fold (and (select cc, -1, c), x) -> (select cc, x, (and, x, c))
8158 SDValue Result = combineSelectAndUseCommutative(N, true, DCI);
8159 if (Result.getNode())
8166 /// PerformORCombine - Target-specific dag combine xforms for ISD::OR
8167 static SDValue PerformORCombine(SDNode *N,
8168 TargetLowering::DAGCombinerInfo &DCI,
8169 const ARMSubtarget *Subtarget) {
8170 // Attempt to use immediate-form VORR
8171 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
8173 EVT VT = N->getValueType(0);
8174 SelectionDAG &DAG = DCI.DAG;
8176 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8179 APInt SplatBits, SplatUndef;
8180 unsigned SplatBitSize;
8182 if (BVN && Subtarget->hasNEON() &&
8183 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
8184 if (SplatBitSize <= 64) {
8186 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
8187 SplatUndef.getZExtValue(), SplatBitSize,
8188 DAG, VorrVT, VT.is128BitVector(),
8190 if (Val.getNode()) {
8192 DAG.getNode(ISD::BITCAST, dl, VorrVT, N->getOperand(0));
8193 SDValue Vorr = DAG.getNode(ARMISD::VORRIMM, dl, VorrVT, Input, Val);
8194 return DAG.getNode(ISD::BITCAST, dl, VT, Vorr);
8199 if (!Subtarget->isThumb1Only()) {
8200 // fold (or (select cc, 0, c), x) -> (select cc, x, (or, x, c))
8201 SDValue Result = combineSelectAndUseCommutative(N, false, DCI);
8202 if (Result.getNode())
8206 // The code below optimizes (or (and X, Y), Z).
8207 // The AND operand needs to have a single user to make these optimizations
8209 SDValue N0 = N->getOperand(0);
8210 if (N0.getOpcode() != ISD::AND || !N0.hasOneUse())
8212 SDValue N1 = N->getOperand(1);
8214 // (or (and B, A), (and C, ~A)) => (VBSL A, B, C) when A is a constant.
8215 if (Subtarget->hasNEON() && N1.getOpcode() == ISD::AND && VT.isVector() &&
8216 DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
8218 unsigned SplatBitSize;
8221 APInt SplatBits0, SplatBits1;
8222 BuildVectorSDNode *BVN0 = dyn_cast<BuildVectorSDNode>(N0->getOperand(1));
8223 BuildVectorSDNode *BVN1 = dyn_cast<BuildVectorSDNode>(N1->getOperand(1));
8224 // Ensure that the second operand of both ands are constants
8225 if (BVN0 && BVN0->isConstantSplat(SplatBits0, SplatUndef, SplatBitSize,
8226 HasAnyUndefs) && !HasAnyUndefs) {
8227 if (BVN1 && BVN1->isConstantSplat(SplatBits1, SplatUndef, SplatBitSize,
8228 HasAnyUndefs) && !HasAnyUndefs) {
8229 // Ensure that the bit width of the constants are the same and that
8230 // the splat arguments are logical inverses as per the pattern we
8231 // are trying to simplify.
8232 if (SplatBits0.getBitWidth() == SplatBits1.getBitWidth() &&
8233 SplatBits0 == ~SplatBits1) {
8234 // Canonicalize the vector type to make instruction selection
8236 EVT CanonicalVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
8237 SDValue Result = DAG.getNode(ARMISD::VBSL, dl, CanonicalVT,
8241 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
8247 // Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when
8250 // BFI is only available on V6T2+
8251 if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops())
8255 // 1) or (and A, mask), val => ARMbfi A, val, mask
8256 // iff (val & mask) == val
8258 // 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
8259 // 2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2)
8260 // && mask == ~mask2
8261 // 2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2)
8262 // && ~mask == mask2
8263 // (i.e., copy a bitfield value into another bitfield of the same width)
8268 SDValue N00 = N0.getOperand(0);
8270 // The value and the mask need to be constants so we can verify this is
8271 // actually a bitfield set. If the mask is 0xffff, we can do better
8272 // via a movt instruction, so don't use BFI in that case.
8273 SDValue MaskOp = N0.getOperand(1);
8274 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(MaskOp);
8277 unsigned Mask = MaskC->getZExtValue();
8281 // Case (1): or (and A, mask), val => ARMbfi A, val, mask
8282 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
8284 unsigned Val = N1C->getZExtValue();
8285 if ((Val & ~Mask) != Val)
8288 if (ARM::isBitFieldInvertedMask(Mask)) {
8289 Val >>= countTrailingZeros(~Mask);
8291 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00,
8292 DAG.getConstant(Val, MVT::i32),
8293 DAG.getConstant(Mask, MVT::i32));
8295 // Do not add new nodes to DAG combiner worklist.
8296 DCI.CombineTo(N, Res, false);
8299 } else if (N1.getOpcode() == ISD::AND) {
8300 // case (2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
8301 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
8304 unsigned Mask2 = N11C->getZExtValue();
8306 // Mask and ~Mask2 (or reverse) must be equivalent for the BFI pattern
8308 if (ARM::isBitFieldInvertedMask(Mask) &&
8310 // The pack halfword instruction works better for masks that fit it,
8311 // so use that when it's available.
8312 if (Subtarget->hasT2ExtractPack() &&
8313 (Mask == 0xffff || Mask == 0xffff0000))
8316 unsigned amt = countTrailingZeros(Mask2);
8317 Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0),
8318 DAG.getConstant(amt, MVT::i32));
8319 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00, Res,
8320 DAG.getConstant(Mask, MVT::i32));
8321 // Do not add new nodes to DAG combiner worklist.
8322 DCI.CombineTo(N, Res, false);
8324 } else if (ARM::isBitFieldInvertedMask(~Mask) &&
8326 // The pack halfword instruction works better for masks that fit it,
8327 // so use that when it's available.
8328 if (Subtarget->hasT2ExtractPack() &&
8329 (Mask2 == 0xffff || Mask2 == 0xffff0000))
8332 unsigned lsb = countTrailingZeros(Mask);
8333 Res = DAG.getNode(ISD::SRL, DL, VT, N00,
8334 DAG.getConstant(lsb, MVT::i32));
8335 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res,
8336 DAG.getConstant(Mask2, MVT::i32));
8337 // Do not add new nodes to DAG combiner worklist.
8338 DCI.CombineTo(N, Res, false);
8343 if (DAG.MaskedValueIsZero(N1, MaskC->getAPIntValue()) &&
8344 N00.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N00.getOperand(1)) &&
8345 ARM::isBitFieldInvertedMask(~Mask)) {
8346 // Case (3): or (and (shl A, #shamt), mask), B => ARMbfi B, A, ~mask
8347 // where lsb(mask) == #shamt and masked bits of B are known zero.
8348 SDValue ShAmt = N00.getOperand(1);
8349 unsigned ShAmtC = cast<ConstantSDNode>(ShAmt)->getZExtValue();
8350 unsigned LSB = countTrailingZeros(Mask);
8354 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1, N00.getOperand(0),
8355 DAG.getConstant(~Mask, MVT::i32));
8357 // Do not add new nodes to DAG combiner worklist.
8358 DCI.CombineTo(N, Res, false);
8364 static SDValue PerformXORCombine(SDNode *N,
8365 TargetLowering::DAGCombinerInfo &DCI,
8366 const ARMSubtarget *Subtarget) {
8367 EVT VT = N->getValueType(0);
8368 SelectionDAG &DAG = DCI.DAG;
8370 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8373 if (!Subtarget->isThumb1Only()) {
8374 // fold (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c))
8375 SDValue Result = combineSelectAndUseCommutative(N, false, DCI);
8376 if (Result.getNode())
8383 /// PerformBFICombine - (bfi A, (and B, Mask1), Mask2) -> (bfi A, B, Mask2) iff
8384 /// the bits being cleared by the AND are not demanded by the BFI.
8385 static SDValue PerformBFICombine(SDNode *N,
8386 TargetLowering::DAGCombinerInfo &DCI) {
8387 SDValue N1 = N->getOperand(1);
8388 if (N1.getOpcode() == ISD::AND) {
8389 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
8392 unsigned InvMask = cast<ConstantSDNode>(N->getOperand(2))->getZExtValue();
8393 unsigned LSB = countTrailingZeros(~InvMask);
8394 unsigned Width = (32 - countLeadingZeros(~InvMask)) - LSB;
8395 unsigned Mask = (1 << Width)-1;
8396 unsigned Mask2 = N11C->getZExtValue();
8397 if ((Mask & (~Mask2)) == 0)
8398 return DCI.DAG.getNode(ARMISD::BFI, SDLoc(N), N->getValueType(0),
8399 N->getOperand(0), N1.getOperand(0),
8405 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
8406 /// ARMISD::VMOVRRD.
8407 static SDValue PerformVMOVRRDCombine(SDNode *N,
8408 TargetLowering::DAGCombinerInfo &DCI) {
8409 // vmovrrd(vmovdrr x, y) -> x,y
8410 SDValue InDouble = N->getOperand(0);
8411 if (InDouble.getOpcode() == ARMISD::VMOVDRR)
8412 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
8414 // vmovrrd(load f64) -> (load i32), (load i32)
8415 SDNode *InNode = InDouble.getNode();
8416 if (ISD::isNormalLoad(InNode) && InNode->hasOneUse() &&
8417 InNode->getValueType(0) == MVT::f64 &&
8418 InNode->getOperand(1).getOpcode() == ISD::FrameIndex &&
8419 !cast<LoadSDNode>(InNode)->isVolatile()) {
8420 // TODO: Should this be done for non-FrameIndex operands?
8421 LoadSDNode *LD = cast<LoadSDNode>(InNode);
8423 SelectionDAG &DAG = DCI.DAG;
8425 SDValue BasePtr = LD->getBasePtr();
8426 SDValue NewLD1 = DAG.getLoad(MVT::i32, DL, LD->getChain(), BasePtr,
8427 LD->getPointerInfo(), LD->isVolatile(),
8428 LD->isNonTemporal(), LD->isInvariant(),
8429 LD->getAlignment());
8431 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
8432 DAG.getConstant(4, MVT::i32));
8433 SDValue NewLD2 = DAG.getLoad(MVT::i32, DL, NewLD1.getValue(1), OffsetPtr,
8434 LD->getPointerInfo(), LD->isVolatile(),
8435 LD->isNonTemporal(), LD->isInvariant(),
8436 std::min(4U, LD->getAlignment() / 2));
8438 DAG.ReplaceAllUsesOfValueWith(SDValue(LD, 1), NewLD2.getValue(1));
8439 if (DCI.DAG.getTargetLoweringInfo().isBigEndian())
8440 std::swap (NewLD1, NewLD2);
8441 SDValue Result = DCI.CombineTo(N, NewLD1, NewLD2);
8442 DCI.RemoveFromWorklist(LD);
8450 /// PerformVMOVDRRCombine - Target-specific dag combine xforms for
8451 /// ARMISD::VMOVDRR. This is also used for BUILD_VECTORs with 2 operands.
8452 static SDValue PerformVMOVDRRCombine(SDNode *N, SelectionDAG &DAG) {
8453 // N=vmovrrd(X); vmovdrr(N:0, N:1) -> bit_convert(X)
8454 SDValue Op0 = N->getOperand(0);
8455 SDValue Op1 = N->getOperand(1);
8456 if (Op0.getOpcode() == ISD::BITCAST)
8457 Op0 = Op0.getOperand(0);
8458 if (Op1.getOpcode() == ISD::BITCAST)
8459 Op1 = Op1.getOperand(0);
8460 if (Op0.getOpcode() == ARMISD::VMOVRRD &&
8461 Op0.getNode() == Op1.getNode() &&
8462 Op0.getResNo() == 0 && Op1.getResNo() == 1)
8463 return DAG.getNode(ISD::BITCAST, SDLoc(N),
8464 N->getValueType(0), Op0.getOperand(0));
8468 /// PerformSTORECombine - Target-specific dag combine xforms for
8470 static SDValue PerformSTORECombine(SDNode *N,
8471 TargetLowering::DAGCombinerInfo &DCI) {
8472 StoreSDNode *St = cast<StoreSDNode>(N);
8473 if (St->isVolatile())
8476 // Optimize trunc store (of multiple scalars) to shuffle and store. First,
8477 // pack all of the elements in one place. Next, store to memory in fewer
8479 SDValue StVal = St->getValue();
8480 EVT VT = StVal.getValueType();
8481 if (St->isTruncatingStore() && VT.isVector()) {
8482 SelectionDAG &DAG = DCI.DAG;
8483 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8484 EVT StVT = St->getMemoryVT();
8485 unsigned NumElems = VT.getVectorNumElements();
8486 assert(StVT != VT && "Cannot truncate to the same type");
8487 unsigned FromEltSz = VT.getVectorElementType().getSizeInBits();
8488 unsigned ToEltSz = StVT.getVectorElementType().getSizeInBits();
8490 // From, To sizes and ElemCount must be pow of two
8491 if (!isPowerOf2_32(NumElems * FromEltSz * ToEltSz)) return SDValue();
8493 // We are going to use the original vector elt for storing.
8494 // Accumulated smaller vector elements must be a multiple of the store size.
8495 if (0 != (NumElems * FromEltSz) % ToEltSz) return SDValue();
8497 unsigned SizeRatio = FromEltSz / ToEltSz;
8498 assert(SizeRatio * NumElems * ToEltSz == VT.getSizeInBits());
8500 // Create a type on which we perform the shuffle.
8501 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(), StVT.getScalarType(),
8502 NumElems*SizeRatio);
8503 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
8506 SDValue WideVec = DAG.getNode(ISD::BITCAST, DL, WideVecVT, StVal);
8507 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
8508 for (unsigned i = 0; i < NumElems; ++i)
8509 ShuffleVec[i] = TLI.isBigEndian() ? (i+1) * SizeRatio - 1 : i * SizeRatio;
8511 // Can't shuffle using an illegal type.
8512 if (!TLI.isTypeLegal(WideVecVT)) return SDValue();
8514 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, DL, WideVec,
8515 DAG.getUNDEF(WideVec.getValueType()),
8517 // At this point all of the data is stored at the bottom of the
8518 // register. We now need to save it to mem.
8520 // Find the largest store unit
8521 MVT StoreType = MVT::i8;
8522 for (unsigned tp = MVT::FIRST_INTEGER_VALUETYPE;
8523 tp < MVT::LAST_INTEGER_VALUETYPE; ++tp) {
8524 MVT Tp = (MVT::SimpleValueType)tp;
8525 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToEltSz)
8528 // Didn't find a legal store type.
8529 if (!TLI.isTypeLegal(StoreType))
8532 // Bitcast the original vector into a vector of store-size units
8533 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
8534 StoreType, VT.getSizeInBits()/EVT(StoreType).getSizeInBits());
8535 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
8536 SDValue ShuffWide = DAG.getNode(ISD::BITCAST, DL, StoreVecVT, Shuff);
8537 SmallVector<SDValue, 8> Chains;
8538 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits()/8,
8539 TLI.getPointerTy());
8540 SDValue BasePtr = St->getBasePtr();
8542 // Perform one or more big stores into memory.
8543 unsigned E = (ToEltSz*NumElems)/StoreType.getSizeInBits();
8544 for (unsigned I = 0; I < E; I++) {
8545 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
8546 StoreType, ShuffWide,
8547 DAG.getIntPtrConstant(I));
8548 SDValue Ch = DAG.getStore(St->getChain(), DL, SubVec, BasePtr,
8549 St->getPointerInfo(), St->isVolatile(),
8550 St->isNonTemporal(), St->getAlignment());
8551 BasePtr = DAG.getNode(ISD::ADD, DL, BasePtr.getValueType(), BasePtr,
8553 Chains.push_back(Ch);
8555 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
8558 if (!ISD::isNormalStore(St))
8561 // Split a store of a VMOVDRR into two integer stores to avoid mixing NEON and
8562 // ARM stores of arguments in the same cache line.
8563 if (StVal.getNode()->getOpcode() == ARMISD::VMOVDRR &&
8564 StVal.getNode()->hasOneUse()) {
8565 SelectionDAG &DAG = DCI.DAG;
8566 bool isBigEndian = DAG.getTargetLoweringInfo().isBigEndian();
8568 SDValue BasePtr = St->getBasePtr();
8569 SDValue NewST1 = DAG.getStore(St->getChain(), DL,
8570 StVal.getNode()->getOperand(isBigEndian ? 1 : 0 ),
8571 BasePtr, St->getPointerInfo(), St->isVolatile(),
8572 St->isNonTemporal(), St->getAlignment());
8574 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
8575 DAG.getConstant(4, MVT::i32));
8576 return DAG.getStore(NewST1.getValue(0), DL,
8577 StVal.getNode()->getOperand(isBigEndian ? 0 : 1),
8578 OffsetPtr, St->getPointerInfo(), St->isVolatile(),
8579 St->isNonTemporal(),
8580 std::min(4U, St->getAlignment() / 2));
8583 if (StVal.getValueType() != MVT::i64 ||
8584 StVal.getNode()->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
8587 // Bitcast an i64 store extracted from a vector to f64.
8588 // Otherwise, the i64 value will be legalized to a pair of i32 values.
8589 SelectionDAG &DAG = DCI.DAG;
8591 SDValue IntVec = StVal.getOperand(0);
8592 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
8593 IntVec.getValueType().getVectorNumElements());
8594 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, IntVec);
8595 SDValue ExtElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
8596 Vec, StVal.getOperand(1));
8598 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ExtElt);
8599 // Make the DAGCombiner fold the bitcasts.
8600 DCI.AddToWorklist(Vec.getNode());
8601 DCI.AddToWorklist(ExtElt.getNode());
8602 DCI.AddToWorklist(V.getNode());
8603 return DAG.getStore(St->getChain(), dl, V, St->getBasePtr(),
8604 St->getPointerInfo(), St->isVolatile(),
8605 St->isNonTemporal(), St->getAlignment(),
8609 /// hasNormalLoadOperand - Check if any of the operands of a BUILD_VECTOR node
8610 /// are normal, non-volatile loads. If so, it is profitable to bitcast an
8611 /// i64 vector to have f64 elements, since the value can then be loaded
8612 /// directly into a VFP register.
8613 static bool hasNormalLoadOperand(SDNode *N) {
8614 unsigned NumElts = N->getValueType(0).getVectorNumElements();
8615 for (unsigned i = 0; i < NumElts; ++i) {
8616 SDNode *Elt = N->getOperand(i).getNode();
8617 if (ISD::isNormalLoad(Elt) && !cast<LoadSDNode>(Elt)->isVolatile())
8623 /// PerformBUILD_VECTORCombine - Target-specific dag combine xforms for
8624 /// ISD::BUILD_VECTOR.
8625 static SDValue PerformBUILD_VECTORCombine(SDNode *N,
8626 TargetLowering::DAGCombinerInfo &DCI){
8627 // build_vector(N=ARMISD::VMOVRRD(X), N:1) -> bit_convert(X):
8628 // VMOVRRD is introduced when legalizing i64 types. It forces the i64 value
8629 // into a pair of GPRs, which is fine when the value is used as a scalar,
8630 // but if the i64 value is converted to a vector, we need to undo the VMOVRRD.
8631 SelectionDAG &DAG = DCI.DAG;
8632 if (N->getNumOperands() == 2) {
8633 SDValue RV = PerformVMOVDRRCombine(N, DAG);
8638 // Load i64 elements as f64 values so that type legalization does not split
8639 // them up into i32 values.
8640 EVT VT = N->getValueType(0);
8641 if (VT.getVectorElementType() != MVT::i64 || !hasNormalLoadOperand(N))
8644 SmallVector<SDValue, 8> Ops;
8645 unsigned NumElts = VT.getVectorNumElements();
8646 for (unsigned i = 0; i < NumElts; ++i) {
8647 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(i));
8649 // Make the DAGCombiner fold the bitcast.
8650 DCI.AddToWorklist(V.getNode());
8652 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, NumElts);
8653 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, FloatVT, Ops);
8654 return DAG.getNode(ISD::BITCAST, dl, VT, BV);
8657 /// \brief Target-specific dag combine xforms for ARMISD::BUILD_VECTOR.
8659 PerformARMBUILD_VECTORCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
8660 // ARMISD::BUILD_VECTOR is introduced when legalizing ISD::BUILD_VECTOR.
8661 // At that time, we may have inserted bitcasts from integer to float.
8662 // If these bitcasts have survived DAGCombine, change the lowering of this
8663 // BUILD_VECTOR in something more vector friendly, i.e., that does not
8664 // force to use floating point types.
8666 // Make sure we can change the type of the vector.
8667 // This is possible iff:
8668 // 1. The vector is only used in a bitcast to a integer type. I.e.,
8669 // 1.1. Vector is used only once.
8670 // 1.2. Use is a bit convert to an integer type.
8671 // 2. The size of its operands are 32-bits (64-bits are not legal).
8672 EVT VT = N->getValueType(0);
8673 EVT EltVT = VT.getVectorElementType();
8675 // Check 1.1. and 2.
8676 if (EltVT.getSizeInBits() != 32 || !N->hasOneUse())
8679 // By construction, the input type must be float.
8680 assert(EltVT == MVT::f32 && "Unexpected type!");
8683 SDNode *Use = *N->use_begin();
8684 if (Use->getOpcode() != ISD::BITCAST ||
8685 Use->getValueType(0).isFloatingPoint())
8688 // Check profitability.
8689 // Model is, if more than half of the relevant operands are bitcast from
8690 // i32, turn the build_vector into a sequence of insert_vector_elt.
8691 // Relevant operands are everything that is not statically
8692 // (i.e., at compile time) bitcasted.
8693 unsigned NumOfBitCastedElts = 0;
8694 unsigned NumElts = VT.getVectorNumElements();
8695 unsigned NumOfRelevantElts = NumElts;
8696 for (unsigned Idx = 0; Idx < NumElts; ++Idx) {
8697 SDValue Elt = N->getOperand(Idx);
8698 if (Elt->getOpcode() == ISD::BITCAST) {
8699 // Assume only bit cast to i32 will go away.
8700 if (Elt->getOperand(0).getValueType() == MVT::i32)
8701 ++NumOfBitCastedElts;
8702 } else if (Elt.getOpcode() == ISD::UNDEF || isa<ConstantSDNode>(Elt))
8703 // Constants are statically casted, thus do not count them as
8704 // relevant operands.
8705 --NumOfRelevantElts;
8708 // Check if more than half of the elements require a non-free bitcast.
8709 if (NumOfBitCastedElts <= NumOfRelevantElts / 2)
8712 SelectionDAG &DAG = DCI.DAG;
8713 // Create the new vector type.
8714 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
8715 // Check if the type is legal.
8716 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8717 if (!TLI.isTypeLegal(VecVT))
8721 // ARMISD::BUILD_VECTOR E1, E2, ..., EN.
8722 // => BITCAST INSERT_VECTOR_ELT
8723 // (INSERT_VECTOR_ELT (...), (BITCAST EN-1), N-1),
8725 SDValue Vec = DAG.getUNDEF(VecVT);
8727 for (unsigned Idx = 0 ; Idx < NumElts; ++Idx) {
8728 SDValue V = N->getOperand(Idx);
8729 if (V.getOpcode() == ISD::UNDEF)
8731 if (V.getOpcode() == ISD::BITCAST &&
8732 V->getOperand(0).getValueType() == MVT::i32)
8733 // Fold obvious case.
8734 V = V.getOperand(0);
8736 V = DAG.getNode(ISD::BITCAST, SDLoc(V), MVT::i32, V);
8737 // Make the DAGCombiner fold the bitcasts.
8738 DCI.AddToWorklist(V.getNode());
8740 SDValue LaneIdx = DAG.getConstant(Idx, MVT::i32);
8741 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VecVT, Vec, V, LaneIdx);
8743 Vec = DAG.getNode(ISD::BITCAST, dl, VT, Vec);
8744 // Make the DAGCombiner fold the bitcasts.
8745 DCI.AddToWorklist(Vec.getNode());
8749 /// PerformInsertEltCombine - Target-specific dag combine xforms for
8750 /// ISD::INSERT_VECTOR_ELT.
8751 static SDValue PerformInsertEltCombine(SDNode *N,
8752 TargetLowering::DAGCombinerInfo &DCI) {
8753 // Bitcast an i64 load inserted into a vector to f64.
8754 // Otherwise, the i64 value will be legalized to a pair of i32 values.
8755 EVT VT = N->getValueType(0);
8756 SDNode *Elt = N->getOperand(1).getNode();
8757 if (VT.getVectorElementType() != MVT::i64 ||
8758 !ISD::isNormalLoad(Elt) || cast<LoadSDNode>(Elt)->isVolatile())
8761 SelectionDAG &DAG = DCI.DAG;
8763 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
8764 VT.getVectorNumElements());
8765 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, N->getOperand(0));
8766 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(1));
8767 // Make the DAGCombiner fold the bitcasts.
8768 DCI.AddToWorklist(Vec.getNode());
8769 DCI.AddToWorklist(V.getNode());
8770 SDValue InsElt = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, FloatVT,
8771 Vec, V, N->getOperand(2));
8772 return DAG.getNode(ISD::BITCAST, dl, VT, InsElt);
8775 /// PerformVECTOR_SHUFFLECombine - Target-specific dag combine xforms for
8776 /// ISD::VECTOR_SHUFFLE.
8777 static SDValue PerformVECTOR_SHUFFLECombine(SDNode *N, SelectionDAG &DAG) {
8778 // The LLVM shufflevector instruction does not require the shuffle mask
8779 // length to match the operand vector length, but ISD::VECTOR_SHUFFLE does
8780 // have that requirement. When translating to ISD::VECTOR_SHUFFLE, if the
8781 // operands do not match the mask length, they are extended by concatenating
8782 // them with undef vectors. That is probably the right thing for other
8783 // targets, but for NEON it is better to concatenate two double-register
8784 // size vector operands into a single quad-register size vector. Do that
8785 // transformation here:
8786 // shuffle(concat(v1, undef), concat(v2, undef)) ->
8787 // shuffle(concat(v1, v2), undef)
8788 SDValue Op0 = N->getOperand(0);
8789 SDValue Op1 = N->getOperand(1);
8790 if (Op0.getOpcode() != ISD::CONCAT_VECTORS ||
8791 Op1.getOpcode() != ISD::CONCAT_VECTORS ||
8792 Op0.getNumOperands() != 2 ||
8793 Op1.getNumOperands() != 2)
8795 SDValue Concat0Op1 = Op0.getOperand(1);
8796 SDValue Concat1Op1 = Op1.getOperand(1);
8797 if (Concat0Op1.getOpcode() != ISD::UNDEF ||
8798 Concat1Op1.getOpcode() != ISD::UNDEF)
8800 // Skip the transformation if any of the types are illegal.
8801 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8802 EVT VT = N->getValueType(0);
8803 if (!TLI.isTypeLegal(VT) ||
8804 !TLI.isTypeLegal(Concat0Op1.getValueType()) ||
8805 !TLI.isTypeLegal(Concat1Op1.getValueType()))
8808 SDValue NewConcat = DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT,
8809 Op0.getOperand(0), Op1.getOperand(0));
8810 // Translate the shuffle mask.
8811 SmallVector<int, 16> NewMask;
8812 unsigned NumElts = VT.getVectorNumElements();
8813 unsigned HalfElts = NumElts/2;
8814 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
8815 for (unsigned n = 0; n < NumElts; ++n) {
8816 int MaskElt = SVN->getMaskElt(n);
8818 if (MaskElt < (int)HalfElts)
8820 else if (MaskElt >= (int)NumElts && MaskElt < (int)(NumElts + HalfElts))
8821 NewElt = HalfElts + MaskElt - NumElts;
8822 NewMask.push_back(NewElt);
8824 return DAG.getVectorShuffle(VT, SDLoc(N), NewConcat,
8825 DAG.getUNDEF(VT), NewMask.data());
8828 /// CombineBaseUpdate - Target-specific DAG combine function for VLDDUP and
8829 /// NEON load/store intrinsics to merge base address updates.
8830 static SDValue CombineBaseUpdate(SDNode *N,
8831 TargetLowering::DAGCombinerInfo &DCI) {
8832 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
8835 SelectionDAG &DAG = DCI.DAG;
8836 bool isIntrinsic = (N->getOpcode() == ISD::INTRINSIC_VOID ||
8837 N->getOpcode() == ISD::INTRINSIC_W_CHAIN);
8838 unsigned AddrOpIdx = (isIntrinsic ? 2 : 1);
8839 SDValue Addr = N->getOperand(AddrOpIdx);
8841 // Search for a use of the address operand that is an increment.
8842 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(),
8843 UE = Addr.getNode()->use_end(); UI != UE; ++UI) {
8845 if (User->getOpcode() != ISD::ADD ||
8846 UI.getUse().getResNo() != Addr.getResNo())
8849 // Check that the add is independent of the load/store. Otherwise, folding
8850 // it would create a cycle.
8851 if (User->isPredecessorOf(N) || N->isPredecessorOf(User))
8854 // Find the new opcode for the updating load/store.
8856 bool isLaneOp = false;
8857 unsigned NewOpc = 0;
8858 unsigned NumVecs = 0;
8860 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
8862 default: llvm_unreachable("unexpected intrinsic for Neon base update");
8863 case Intrinsic::arm_neon_vld1: NewOpc = ARMISD::VLD1_UPD;
8865 case Intrinsic::arm_neon_vld2: NewOpc = ARMISD::VLD2_UPD;
8867 case Intrinsic::arm_neon_vld3: NewOpc = ARMISD::VLD3_UPD;
8869 case Intrinsic::arm_neon_vld4: NewOpc = ARMISD::VLD4_UPD;
8871 case Intrinsic::arm_neon_vld2lane: NewOpc = ARMISD::VLD2LN_UPD;
8872 NumVecs = 2; isLaneOp = true; break;
8873 case Intrinsic::arm_neon_vld3lane: NewOpc = ARMISD::VLD3LN_UPD;
8874 NumVecs = 3; isLaneOp = true; break;
8875 case Intrinsic::arm_neon_vld4lane: NewOpc = ARMISD::VLD4LN_UPD;
8876 NumVecs = 4; isLaneOp = true; break;
8877 case Intrinsic::arm_neon_vst1: NewOpc = ARMISD::VST1_UPD;
8878 NumVecs = 1; isLoad = false; break;
8879 case Intrinsic::arm_neon_vst2: NewOpc = ARMISD::VST2_UPD;
8880 NumVecs = 2; isLoad = false; break;
8881 case Intrinsic::arm_neon_vst3: NewOpc = ARMISD::VST3_UPD;
8882 NumVecs = 3; isLoad = false; break;
8883 case Intrinsic::arm_neon_vst4: NewOpc = ARMISD::VST4_UPD;
8884 NumVecs = 4; isLoad = false; break;
8885 case Intrinsic::arm_neon_vst2lane: NewOpc = ARMISD::VST2LN_UPD;
8886 NumVecs = 2; isLoad = false; isLaneOp = true; break;
8887 case Intrinsic::arm_neon_vst3lane: NewOpc = ARMISD::VST3LN_UPD;
8888 NumVecs = 3; isLoad = false; isLaneOp = true; break;
8889 case Intrinsic::arm_neon_vst4lane: NewOpc = ARMISD::VST4LN_UPD;
8890 NumVecs = 4; isLoad = false; isLaneOp = true; break;
8894 switch (N->getOpcode()) {
8895 default: llvm_unreachable("unexpected opcode for Neon base update");
8896 case ARMISD::VLD2DUP: NewOpc = ARMISD::VLD2DUP_UPD; NumVecs = 2; break;
8897 case ARMISD::VLD3DUP: NewOpc = ARMISD::VLD3DUP_UPD; NumVecs = 3; break;
8898 case ARMISD::VLD4DUP: NewOpc = ARMISD::VLD4DUP_UPD; NumVecs = 4; break;
8902 // Find the size of memory referenced by the load/store.
8905 VecTy = N->getValueType(0);
8907 VecTy = N->getOperand(AddrOpIdx+1).getValueType();
8908 unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8;
8910 NumBytes /= VecTy.getVectorNumElements();
8912 // If the increment is a constant, it must match the memory ref size.
8913 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
8914 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
8915 uint64_t IncVal = CInc->getZExtValue();
8916 if (IncVal != NumBytes)
8918 } else if (NumBytes >= 3 * 16) {
8919 // VLD3/4 and VST3/4 for 128-bit vectors are implemented with two
8920 // separate instructions that make it harder to use a non-constant update.
8924 // Create the new updating load/store node.
8926 unsigned NumResultVecs = (isLoad ? NumVecs : 0);
8928 for (n = 0; n < NumResultVecs; ++n)
8930 Tys[n++] = MVT::i32;
8931 Tys[n] = MVT::Other;
8932 SDVTList SDTys = DAG.getVTList(ArrayRef<EVT>(Tys, NumResultVecs+2));
8933 SmallVector<SDValue, 8> Ops;
8934 Ops.push_back(N->getOperand(0)); // incoming chain
8935 Ops.push_back(N->getOperand(AddrOpIdx));
8937 for (unsigned i = AddrOpIdx + 1; i < N->getNumOperands(); ++i) {
8938 Ops.push_back(N->getOperand(i));
8940 MemIntrinsicSDNode *MemInt = cast<MemIntrinsicSDNode>(N);
8941 SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, SDLoc(N), SDTys,
8942 Ops, MemInt->getMemoryVT(),
8943 MemInt->getMemOperand());
8946 std::vector<SDValue> NewResults;
8947 for (unsigned i = 0; i < NumResultVecs; ++i) {
8948 NewResults.push_back(SDValue(UpdN.getNode(), i));
8950 NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs+1)); // chain
8951 DCI.CombineTo(N, NewResults);
8952 DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs));
8959 /// CombineVLDDUP - For a VDUPLANE node N, check if its source operand is a
8960 /// vldN-lane (N > 1) intrinsic, and if all the other uses of that intrinsic
8961 /// are also VDUPLANEs. If so, combine them to a vldN-dup operation and
8963 static bool CombineVLDDUP(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
8964 SelectionDAG &DAG = DCI.DAG;
8965 EVT VT = N->getValueType(0);
8966 // vldN-dup instructions only support 64-bit vectors for N > 1.
8967 if (!VT.is64BitVector())
8970 // Check if the VDUPLANE operand is a vldN-dup intrinsic.
8971 SDNode *VLD = N->getOperand(0).getNode();
8972 if (VLD->getOpcode() != ISD::INTRINSIC_W_CHAIN)
8974 unsigned NumVecs = 0;
8975 unsigned NewOpc = 0;
8976 unsigned IntNo = cast<ConstantSDNode>(VLD->getOperand(1))->getZExtValue();
8977 if (IntNo == Intrinsic::arm_neon_vld2lane) {
8979 NewOpc = ARMISD::VLD2DUP;
8980 } else if (IntNo == Intrinsic::arm_neon_vld3lane) {
8982 NewOpc = ARMISD::VLD3DUP;
8983 } else if (IntNo == Intrinsic::arm_neon_vld4lane) {
8985 NewOpc = ARMISD::VLD4DUP;
8990 // First check that all the vldN-lane uses are VDUPLANEs and that the lane
8991 // numbers match the load.
8992 unsigned VLDLaneNo =
8993 cast<ConstantSDNode>(VLD->getOperand(NumVecs+3))->getZExtValue();
8994 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
8996 // Ignore uses of the chain result.
8997 if (UI.getUse().getResNo() == NumVecs)
9000 if (User->getOpcode() != ARMISD::VDUPLANE ||
9001 VLDLaneNo != cast<ConstantSDNode>(User->getOperand(1))->getZExtValue())
9005 // Create the vldN-dup node.
9008 for (n = 0; n < NumVecs; ++n)
9010 Tys[n] = MVT::Other;
9011 SDVTList SDTys = DAG.getVTList(ArrayRef<EVT>(Tys, NumVecs+1));
9012 SDValue Ops[] = { VLD->getOperand(0), VLD->getOperand(2) };
9013 MemIntrinsicSDNode *VLDMemInt = cast<MemIntrinsicSDNode>(VLD);
9014 SDValue VLDDup = DAG.getMemIntrinsicNode(NewOpc, SDLoc(VLD), SDTys,
9015 Ops, VLDMemInt->getMemoryVT(),
9016 VLDMemInt->getMemOperand());
9019 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
9021 unsigned ResNo = UI.getUse().getResNo();
9022 // Ignore uses of the chain result.
9023 if (ResNo == NumVecs)
9026 DCI.CombineTo(User, SDValue(VLDDup.getNode(), ResNo));
9029 // Now the vldN-lane intrinsic is dead except for its chain result.
9030 // Update uses of the chain.
9031 std::vector<SDValue> VLDDupResults;
9032 for (unsigned n = 0; n < NumVecs; ++n)
9033 VLDDupResults.push_back(SDValue(VLDDup.getNode(), n));
9034 VLDDupResults.push_back(SDValue(VLDDup.getNode(), NumVecs));
9035 DCI.CombineTo(VLD, VLDDupResults);
9040 /// PerformVDUPLANECombine - Target-specific dag combine xforms for
9041 /// ARMISD::VDUPLANE.
9042 static SDValue PerformVDUPLANECombine(SDNode *N,
9043 TargetLowering::DAGCombinerInfo &DCI) {
9044 SDValue Op = N->getOperand(0);
9046 // If the source is a vldN-lane (N > 1) intrinsic, and all the other uses
9047 // of that intrinsic are also VDUPLANEs, combine them to a vldN-dup operation.
9048 if (CombineVLDDUP(N, DCI))
9049 return SDValue(N, 0);
9051 // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is
9052 // redundant. Ignore bit_converts for now; element sizes are checked below.
9053 while (Op.getOpcode() == ISD::BITCAST)
9054 Op = Op.getOperand(0);
9055 if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM)
9058 // Make sure the VMOV element size is not bigger than the VDUPLANE elements.
9059 unsigned EltSize = Op.getValueType().getVectorElementType().getSizeInBits();
9060 // The canonical VMOV for a zero vector uses a 32-bit element size.
9061 unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
9063 if (ARM_AM::decodeNEONModImm(Imm, EltBits) == 0)
9065 EVT VT = N->getValueType(0);
9066 if (EltSize > VT.getVectorElementType().getSizeInBits())
9069 return DCI.DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Op);
9072 // isConstVecPow2 - Return true if each vector element is a power of 2, all
9073 // elements are the same constant, C, and Log2(C) ranges from 1 to 32.
9074 static bool isConstVecPow2(SDValue ConstVec, bool isSigned, uint64_t &C)
9078 for (unsigned I = 0, E = ConstVec.getValueType().getVectorNumElements();
9080 ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(ConstVec.getOperand(I));
9085 APFloat APF = C->getValueAPF();
9086 if (APF.convertToInteger(&cN, 64, isSigned, APFloat::rmTowardZero, &isExact)
9087 != APFloat::opOK || !isExact)
9090 c0 = (I == 0) ? cN : c0;
9091 if (!isPowerOf2_64(cN) || c0 != cN || Log2_64(c0) < 1 || Log2_64(c0) > 32)
9098 /// PerformVCVTCombine - VCVT (floating-point to fixed-point, Advanced SIMD)
9099 /// can replace combinations of VMUL and VCVT (floating-point to integer)
9100 /// when the VMUL has a constant operand that is a power of 2.
9102 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
9103 /// vmul.f32 d16, d17, d16
9104 /// vcvt.s32.f32 d16, d16
9106 /// vcvt.s32.f32 d16, d16, #3
9107 static SDValue PerformVCVTCombine(SDNode *N,
9108 TargetLowering::DAGCombinerInfo &DCI,
9109 const ARMSubtarget *Subtarget) {
9110 SelectionDAG &DAG = DCI.DAG;
9111 SDValue Op = N->getOperand(0);
9113 if (!Subtarget->hasNEON() || !Op.getValueType().isVector() ||
9114 Op.getOpcode() != ISD::FMUL)
9118 SDValue N0 = Op->getOperand(0);
9119 SDValue ConstVec = Op->getOperand(1);
9120 bool isSigned = N->getOpcode() == ISD::FP_TO_SINT;
9122 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
9123 !isConstVecPow2(ConstVec, isSigned, C))
9126 MVT FloatTy = Op.getSimpleValueType().getVectorElementType();
9127 MVT IntTy = N->getSimpleValueType(0).getVectorElementType();
9128 if (FloatTy.getSizeInBits() != 32 || IntTy.getSizeInBits() > 32) {
9129 // These instructions only exist converting from f32 to i32. We can handle
9130 // smaller integers by generating an extra truncate, but larger ones would
9135 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfp2fxs :
9136 Intrinsic::arm_neon_vcvtfp2fxu;
9137 unsigned NumLanes = Op.getValueType().getVectorNumElements();
9138 SDValue FixConv = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N),
9139 NumLanes == 2 ? MVT::v2i32 : MVT::v4i32,
9140 DAG.getConstant(IntrinsicOpcode, MVT::i32), N0,
9141 DAG.getConstant(Log2_64(C), MVT::i32));
9143 if (IntTy.getSizeInBits() < FloatTy.getSizeInBits())
9144 FixConv = DAG.getNode(ISD::TRUNCATE, SDLoc(N), N->getValueType(0), FixConv);
9149 /// PerformVDIVCombine - VCVT (fixed-point to floating-point, Advanced SIMD)
9150 /// can replace combinations of VCVT (integer to floating-point) and VDIV
9151 /// when the VDIV has a constant operand that is a power of 2.
9153 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
9154 /// vcvt.f32.s32 d16, d16
9155 /// vdiv.f32 d16, d17, d16
9157 /// vcvt.f32.s32 d16, d16, #3
9158 static SDValue PerformVDIVCombine(SDNode *N,
9159 TargetLowering::DAGCombinerInfo &DCI,
9160 const ARMSubtarget *Subtarget) {
9161 SelectionDAG &DAG = DCI.DAG;
9162 SDValue Op = N->getOperand(0);
9163 unsigned OpOpcode = Op.getNode()->getOpcode();
9165 if (!Subtarget->hasNEON() || !N->getValueType(0).isVector() ||
9166 (OpOpcode != ISD::SINT_TO_FP && OpOpcode != ISD::UINT_TO_FP))
9170 SDValue ConstVec = N->getOperand(1);
9171 bool isSigned = OpOpcode == ISD::SINT_TO_FP;
9173 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
9174 !isConstVecPow2(ConstVec, isSigned, C))
9177 MVT FloatTy = N->getSimpleValueType(0).getVectorElementType();
9178 MVT IntTy = Op.getOperand(0).getSimpleValueType().getVectorElementType();
9179 if (FloatTy.getSizeInBits() != 32 || IntTy.getSizeInBits() > 32) {
9180 // These instructions only exist converting from i32 to f32. We can handle
9181 // smaller integers by generating an extra extend, but larger ones would
9186 SDValue ConvInput = Op.getOperand(0);
9187 unsigned NumLanes = Op.getValueType().getVectorNumElements();
9188 if (IntTy.getSizeInBits() < FloatTy.getSizeInBits())
9189 ConvInput = DAG.getNode(isSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND,
9190 SDLoc(N), NumLanes == 2 ? MVT::v2i32 : MVT::v4i32,
9193 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfxs2fp :
9194 Intrinsic::arm_neon_vcvtfxu2fp;
9195 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N),
9197 DAG.getConstant(IntrinsicOpcode, MVT::i32),
9198 ConvInput, DAG.getConstant(Log2_64(C), MVT::i32));
9201 /// Getvshiftimm - Check if this is a valid build_vector for the immediate
9202 /// operand of a vector shift operation, where all the elements of the
9203 /// build_vector must have the same constant integer value.
9204 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
9205 // Ignore bit_converts.
9206 while (Op.getOpcode() == ISD::BITCAST)
9207 Op = Op.getOperand(0);
9208 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
9209 APInt SplatBits, SplatUndef;
9210 unsigned SplatBitSize;
9212 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
9213 HasAnyUndefs, ElementBits) ||
9214 SplatBitSize > ElementBits)
9216 Cnt = SplatBits.getSExtValue();
9220 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
9221 /// operand of a vector shift left operation. That value must be in the range:
9222 /// 0 <= Value < ElementBits for a left shift; or
9223 /// 0 <= Value <= ElementBits for a long left shift.
9224 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
9225 assert(VT.isVector() && "vector shift count is not a vector type");
9226 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
9227 if (! getVShiftImm(Op, ElementBits, Cnt))
9229 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
9232 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
9233 /// operand of a vector shift right operation. For a shift opcode, the value
9234 /// is positive, but for an intrinsic the value count must be negative. The
9235 /// absolute value must be in the range:
9236 /// 1 <= |Value| <= ElementBits for a right shift; or
9237 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
9238 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
9240 assert(VT.isVector() && "vector shift count is not a vector type");
9241 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
9242 if (! getVShiftImm(Op, ElementBits, Cnt))
9246 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
9249 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
9250 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
9251 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
9254 // Don't do anything for most intrinsics.
9257 // Vector shifts: check for immediate versions and lower them.
9258 // Note: This is done during DAG combining instead of DAG legalizing because
9259 // the build_vectors for 64-bit vector element shift counts are generally
9260 // not legal, and it is hard to see their values after they get legalized to
9261 // loads from a constant pool.
9262 case Intrinsic::arm_neon_vshifts:
9263 case Intrinsic::arm_neon_vshiftu:
9264 case Intrinsic::arm_neon_vrshifts:
9265 case Intrinsic::arm_neon_vrshiftu:
9266 case Intrinsic::arm_neon_vrshiftn:
9267 case Intrinsic::arm_neon_vqshifts:
9268 case Intrinsic::arm_neon_vqshiftu:
9269 case Intrinsic::arm_neon_vqshiftsu:
9270 case Intrinsic::arm_neon_vqshiftns:
9271 case Intrinsic::arm_neon_vqshiftnu:
9272 case Intrinsic::arm_neon_vqshiftnsu:
9273 case Intrinsic::arm_neon_vqrshiftns:
9274 case Intrinsic::arm_neon_vqrshiftnu:
9275 case Intrinsic::arm_neon_vqrshiftnsu: {
9276 EVT VT = N->getOperand(1).getValueType();
9278 unsigned VShiftOpc = 0;
9281 case Intrinsic::arm_neon_vshifts:
9282 case Intrinsic::arm_neon_vshiftu:
9283 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
9284 VShiftOpc = ARMISD::VSHL;
9287 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
9288 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
9289 ARMISD::VSHRs : ARMISD::VSHRu);
9294 case Intrinsic::arm_neon_vrshifts:
9295 case Intrinsic::arm_neon_vrshiftu:
9296 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
9300 case Intrinsic::arm_neon_vqshifts:
9301 case Intrinsic::arm_neon_vqshiftu:
9302 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
9306 case Intrinsic::arm_neon_vqshiftsu:
9307 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
9309 llvm_unreachable("invalid shift count for vqshlu intrinsic");
9311 case Intrinsic::arm_neon_vrshiftn:
9312 case Intrinsic::arm_neon_vqshiftns:
9313 case Intrinsic::arm_neon_vqshiftnu:
9314 case Intrinsic::arm_neon_vqshiftnsu:
9315 case Intrinsic::arm_neon_vqrshiftns:
9316 case Intrinsic::arm_neon_vqrshiftnu:
9317 case Intrinsic::arm_neon_vqrshiftnsu:
9318 // Narrowing shifts require an immediate right shift.
9319 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
9321 llvm_unreachable("invalid shift count for narrowing vector shift "
9325 llvm_unreachable("unhandled vector shift");
9329 case Intrinsic::arm_neon_vshifts:
9330 case Intrinsic::arm_neon_vshiftu:
9331 // Opcode already set above.
9333 case Intrinsic::arm_neon_vrshifts:
9334 VShiftOpc = ARMISD::VRSHRs; break;
9335 case Intrinsic::arm_neon_vrshiftu:
9336 VShiftOpc = ARMISD::VRSHRu; break;
9337 case Intrinsic::arm_neon_vrshiftn:
9338 VShiftOpc = ARMISD::VRSHRN; break;
9339 case Intrinsic::arm_neon_vqshifts:
9340 VShiftOpc = ARMISD::VQSHLs; break;
9341 case Intrinsic::arm_neon_vqshiftu:
9342 VShiftOpc = ARMISD::VQSHLu; break;
9343 case Intrinsic::arm_neon_vqshiftsu:
9344 VShiftOpc = ARMISD::VQSHLsu; break;
9345 case Intrinsic::arm_neon_vqshiftns:
9346 VShiftOpc = ARMISD::VQSHRNs; break;
9347 case Intrinsic::arm_neon_vqshiftnu:
9348 VShiftOpc = ARMISD::VQSHRNu; break;
9349 case Intrinsic::arm_neon_vqshiftnsu:
9350 VShiftOpc = ARMISD::VQSHRNsu; break;
9351 case Intrinsic::arm_neon_vqrshiftns:
9352 VShiftOpc = ARMISD::VQRSHRNs; break;
9353 case Intrinsic::arm_neon_vqrshiftnu:
9354 VShiftOpc = ARMISD::VQRSHRNu; break;
9355 case Intrinsic::arm_neon_vqrshiftnsu:
9356 VShiftOpc = ARMISD::VQRSHRNsu; break;
9359 return DAG.getNode(VShiftOpc, SDLoc(N), N->getValueType(0),
9360 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
9363 case Intrinsic::arm_neon_vshiftins: {
9364 EVT VT = N->getOperand(1).getValueType();
9366 unsigned VShiftOpc = 0;
9368 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
9369 VShiftOpc = ARMISD::VSLI;
9370 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
9371 VShiftOpc = ARMISD::VSRI;
9373 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
9376 return DAG.getNode(VShiftOpc, SDLoc(N), N->getValueType(0),
9377 N->getOperand(1), N->getOperand(2),
9378 DAG.getConstant(Cnt, MVT::i32));
9381 case Intrinsic::arm_neon_vqrshifts:
9382 case Intrinsic::arm_neon_vqrshiftu:
9383 // No immediate versions of these to check for.
9390 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
9391 /// lowers them. As with the vector shift intrinsics, this is done during DAG
9392 /// combining instead of DAG legalizing because the build_vectors for 64-bit
9393 /// vector element shift counts are generally not legal, and it is hard to see
9394 /// their values after they get legalized to loads from a constant pool.
9395 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
9396 const ARMSubtarget *ST) {
9397 EVT VT = N->getValueType(0);
9398 if (N->getOpcode() == ISD::SRL && VT == MVT::i32 && ST->hasV6Ops()) {
9399 // Canonicalize (srl (bswap x), 16) to (rotr (bswap x), 16) if the high
9400 // 16-bits of x is zero. This optimizes rev + lsr 16 to rev16.
9401 SDValue N1 = N->getOperand(1);
9402 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) {
9403 SDValue N0 = N->getOperand(0);
9404 if (C->getZExtValue() == 16 && N0.getOpcode() == ISD::BSWAP &&
9405 DAG.MaskedValueIsZero(N0.getOperand(0),
9406 APInt::getHighBitsSet(32, 16)))
9407 return DAG.getNode(ISD::ROTR, SDLoc(N), VT, N0, N1);
9411 // Nothing to be done for scalar shifts.
9412 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9413 if (!VT.isVector() || !TLI.isTypeLegal(VT))
9416 assert(ST->hasNEON() && "unexpected vector shift");
9419 switch (N->getOpcode()) {
9420 default: llvm_unreachable("unexpected shift opcode");
9423 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
9424 return DAG.getNode(ARMISD::VSHL, SDLoc(N), VT, N->getOperand(0),
9425 DAG.getConstant(Cnt, MVT::i32));
9430 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
9431 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
9432 ARMISD::VSHRs : ARMISD::VSHRu);
9433 return DAG.getNode(VShiftOpc, SDLoc(N), VT, N->getOperand(0),
9434 DAG.getConstant(Cnt, MVT::i32));
9440 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
9441 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
9442 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
9443 const ARMSubtarget *ST) {
9444 SDValue N0 = N->getOperand(0);
9446 // Check for sign- and zero-extensions of vector extract operations of 8-
9447 // and 16-bit vector elements. NEON supports these directly. They are
9448 // handled during DAG combining because type legalization will promote them
9449 // to 32-bit types and it is messy to recognize the operations after that.
9450 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
9451 SDValue Vec = N0.getOperand(0);
9452 SDValue Lane = N0.getOperand(1);
9453 EVT VT = N->getValueType(0);
9454 EVT EltVT = N0.getValueType();
9455 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9457 if (VT == MVT::i32 &&
9458 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
9459 TLI.isTypeLegal(Vec.getValueType()) &&
9460 isa<ConstantSDNode>(Lane)) {
9463 switch (N->getOpcode()) {
9464 default: llvm_unreachable("unexpected opcode");
9465 case ISD::SIGN_EXTEND:
9466 Opc = ARMISD::VGETLANEs;
9468 case ISD::ZERO_EXTEND:
9469 case ISD::ANY_EXTEND:
9470 Opc = ARMISD::VGETLANEu;
9473 return DAG.getNode(Opc, SDLoc(N), VT, Vec, Lane);
9480 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
9481 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
9482 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
9483 const ARMSubtarget *ST) {
9484 // If the target supports NEON, try to use vmax/vmin instructions for f32
9485 // selects like "x < y ? x : y". Unless the NoNaNsFPMath option is set,
9486 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
9487 // a NaN; only do the transformation when it matches that behavior.
9489 // For now only do this when using NEON for FP operations; if using VFP, it
9490 // is not obvious that the benefit outweighs the cost of switching to the
9492 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
9493 N->getValueType(0) != MVT::f32)
9496 SDValue CondLHS = N->getOperand(0);
9497 SDValue CondRHS = N->getOperand(1);
9498 SDValue LHS = N->getOperand(2);
9499 SDValue RHS = N->getOperand(3);
9500 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
9502 unsigned Opcode = 0;
9504 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
9505 IsReversed = false; // x CC y ? x : y
9506 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
9507 IsReversed = true ; // x CC y ? y : x
9521 // If LHS is NaN, an ordered comparison will be false and the result will
9522 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
9523 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
9524 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
9525 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
9527 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
9528 // will return -0, so vmin can only be used for unsafe math or if one of
9529 // the operands is known to be nonzero.
9530 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
9531 !DAG.getTarget().Options.UnsafeFPMath &&
9532 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9534 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
9543 // If LHS is NaN, an ordered comparison will be false and the result will
9544 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
9545 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
9546 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
9547 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
9549 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
9550 // will return +0, so vmax can only be used for unsafe math or if one of
9551 // the operands is known to be nonzero.
9552 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
9553 !DAG.getTarget().Options.UnsafeFPMath &&
9554 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9556 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
9562 return DAG.getNode(Opcode, SDLoc(N), N->getValueType(0), LHS, RHS);
9565 /// PerformCMOVCombine - Target-specific DAG combining for ARMISD::CMOV.
9567 ARMTargetLowering::PerformCMOVCombine(SDNode *N, SelectionDAG &DAG) const {
9568 SDValue Cmp = N->getOperand(4);
9569 if (Cmp.getOpcode() != ARMISD::CMPZ)
9570 // Only looking at EQ and NE cases.
9573 EVT VT = N->getValueType(0);
9575 SDValue LHS = Cmp.getOperand(0);
9576 SDValue RHS = Cmp.getOperand(1);
9577 SDValue FalseVal = N->getOperand(0);
9578 SDValue TrueVal = N->getOperand(1);
9579 SDValue ARMcc = N->getOperand(2);
9580 ARMCC::CondCodes CC =
9581 (ARMCC::CondCodes)cast<ConstantSDNode>(ARMcc)->getZExtValue();
9599 /// FIXME: Turn this into a target neutral optimization?
9601 if (CC == ARMCC::NE && FalseVal == RHS && FalseVal != LHS) {
9602 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, TrueVal, ARMcc,
9603 N->getOperand(3), Cmp);
9604 } else if (CC == ARMCC::EQ && TrueVal == RHS) {
9606 SDValue NewCmp = getARMCmp(LHS, RHS, ISD::SETNE, ARMcc, DAG, dl);
9607 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, FalseVal, ARMcc,
9608 N->getOperand(3), NewCmp);
9611 if (Res.getNode()) {
9612 APInt KnownZero, KnownOne;
9613 DAG.computeKnownBits(SDValue(N,0), KnownZero, KnownOne);
9614 // Capture demanded bits information that would be otherwise lost.
9615 if (KnownZero == 0xfffffffe)
9616 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
9617 DAG.getValueType(MVT::i1));
9618 else if (KnownZero == 0xffffff00)
9619 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
9620 DAG.getValueType(MVT::i8));
9621 else if (KnownZero == 0xffff0000)
9622 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
9623 DAG.getValueType(MVT::i16));
9629 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
9630 DAGCombinerInfo &DCI) const {
9631 switch (N->getOpcode()) {
9633 case ISD::ADDC: return PerformADDCCombine(N, DCI, Subtarget);
9634 case ISD::ADD: return PerformADDCombine(N, DCI, Subtarget);
9635 case ISD::SUB: return PerformSUBCombine(N, DCI);
9636 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
9637 case ISD::OR: return PerformORCombine(N, DCI, Subtarget);
9638 case ISD::XOR: return PerformXORCombine(N, DCI, Subtarget);
9639 case ISD::AND: return PerformANDCombine(N, DCI, Subtarget);
9640 case ARMISD::BFI: return PerformBFICombine(N, DCI);
9641 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
9642 case ARMISD::VMOVDRR: return PerformVMOVDRRCombine(N, DCI.DAG);
9643 case ISD::STORE: return PerformSTORECombine(N, DCI);
9644 case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DCI);
9645 case ISD::INSERT_VECTOR_ELT: return PerformInsertEltCombine(N, DCI);
9646 case ISD::VECTOR_SHUFFLE: return PerformVECTOR_SHUFFLECombine(N, DCI.DAG);
9647 case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI);
9648 case ISD::FP_TO_SINT:
9649 case ISD::FP_TO_UINT: return PerformVCVTCombine(N, DCI, Subtarget);
9650 case ISD::FDIV: return PerformVDIVCombine(N, DCI, Subtarget);
9651 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
9654 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
9655 case ISD::SIGN_EXTEND:
9656 case ISD::ZERO_EXTEND:
9657 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
9658 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
9659 case ARMISD::CMOV: return PerformCMOVCombine(N, DCI.DAG);
9660 case ARMISD::VLD2DUP:
9661 case ARMISD::VLD3DUP:
9662 case ARMISD::VLD4DUP:
9663 return CombineBaseUpdate(N, DCI);
9664 case ARMISD::BUILD_VECTOR:
9665 return PerformARMBUILD_VECTORCombine(N, DCI);
9666 case ISD::INTRINSIC_VOID:
9667 case ISD::INTRINSIC_W_CHAIN:
9668 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
9669 case Intrinsic::arm_neon_vld1:
9670 case Intrinsic::arm_neon_vld2:
9671 case Intrinsic::arm_neon_vld3:
9672 case Intrinsic::arm_neon_vld4:
9673 case Intrinsic::arm_neon_vld2lane:
9674 case Intrinsic::arm_neon_vld3lane:
9675 case Intrinsic::arm_neon_vld4lane:
9676 case Intrinsic::arm_neon_vst1:
9677 case Intrinsic::arm_neon_vst2:
9678 case Intrinsic::arm_neon_vst3:
9679 case Intrinsic::arm_neon_vst4:
9680 case Intrinsic::arm_neon_vst2lane:
9681 case Intrinsic::arm_neon_vst3lane:
9682 case Intrinsic::arm_neon_vst4lane:
9683 return CombineBaseUpdate(N, DCI);
9691 bool ARMTargetLowering::isDesirableToTransformToIntegerOp(unsigned Opc,
9693 return (VT == MVT::f32) && (Opc == ISD::LOAD || Opc == ISD::STORE);
9696 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT, unsigned,
9698 // The AllowsUnaliged flag models the SCTLR.A setting in ARM cpus
9699 bool AllowsUnaligned = Subtarget->allowsUnalignedMem();
9701 switch (VT.getSimpleVT().SimpleTy) {
9707 // Unaligned access can use (for example) LRDB, LRDH, LDR
9708 if (AllowsUnaligned) {
9710 *Fast = Subtarget->hasV7Ops();
9717 // For any little-endian targets with neon, we can support unaligned ld/st
9718 // of D and Q (e.g. {D0,D1}) registers by using vld1.i8/vst1.i8.
9719 // A big-endian target may also explicitly support unaligned accesses
9720 if (Subtarget->hasNEON() && (AllowsUnaligned || isLittleEndian())) {
9730 static bool memOpAlign(unsigned DstAlign, unsigned SrcAlign,
9731 unsigned AlignCheck) {
9732 return ((SrcAlign == 0 || SrcAlign % AlignCheck == 0) &&
9733 (DstAlign == 0 || DstAlign % AlignCheck == 0));
9736 EVT ARMTargetLowering::getOptimalMemOpType(uint64_t Size,
9737 unsigned DstAlign, unsigned SrcAlign,
9738 bool IsMemset, bool ZeroMemset,
9740 MachineFunction &MF) const {
9741 const Function *F = MF.getFunction();
9743 // See if we can use NEON instructions for this...
9744 if ((!IsMemset || ZeroMemset) &&
9745 Subtarget->hasNEON() &&
9746 !F->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
9747 Attribute::NoImplicitFloat)) {
9750 (memOpAlign(SrcAlign, DstAlign, 16) ||
9751 (allowsUnalignedMemoryAccesses(MVT::v2f64, 0, &Fast) && Fast))) {
9753 } else if (Size >= 8 &&
9754 (memOpAlign(SrcAlign, DstAlign, 8) ||
9755 (allowsUnalignedMemoryAccesses(MVT::f64, 0, &Fast) && Fast))) {
9760 // Lowering to i32/i16 if the size permits.
9766 // Let the target-independent logic figure it out.
9770 bool ARMTargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
9771 if (Val.getOpcode() != ISD::LOAD)
9774 EVT VT1 = Val.getValueType();
9775 if (!VT1.isSimple() || !VT1.isInteger() ||
9776 !VT2.isSimple() || !VT2.isInteger())
9779 switch (VT1.getSimpleVT().SimpleTy) {
9784 // 8-bit and 16-bit loads implicitly zero-extend to 32-bits.
9791 bool ARMTargetLowering::allowTruncateForTailCall(Type *Ty1, Type *Ty2) const {
9792 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
9795 if (!isTypeLegal(EVT::getEVT(Ty1)))
9798 assert(Ty1->getPrimitiveSizeInBits() <= 64 && "i128 is probably not a noop");
9800 // Assuming the caller doesn't have a zeroext or signext return parameter,
9801 // truncation all the way down to i1 is valid.
9806 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
9811 switch (VT.getSimpleVT().SimpleTy) {
9812 default: return false;
9827 if ((V & (Scale - 1)) != 0)
9830 return V == (V & ((1LL << 5) - 1));
9833 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
9834 const ARMSubtarget *Subtarget) {
9841 switch (VT.getSimpleVT().SimpleTy) {
9842 default: return false;
9847 // + imm12 or - imm8
9849 return V == (V & ((1LL << 8) - 1));
9850 return V == (V & ((1LL << 12) - 1));
9853 // Same as ARM mode. FIXME: NEON?
9854 if (!Subtarget->hasVFP2())
9859 return V == (V & ((1LL << 8) - 1));
9863 /// isLegalAddressImmediate - Return true if the integer value can be used
9864 /// as the offset of the target addressing mode for load / store of the
9866 static bool isLegalAddressImmediate(int64_t V, EVT VT,
9867 const ARMSubtarget *Subtarget) {
9874 if (Subtarget->isThumb1Only())
9875 return isLegalT1AddressImmediate(V, VT);
9876 else if (Subtarget->isThumb2())
9877 return isLegalT2AddressImmediate(V, VT, Subtarget);
9882 switch (VT.getSimpleVT().SimpleTy) {
9883 default: return false;
9888 return V == (V & ((1LL << 12) - 1));
9891 return V == (V & ((1LL << 8) - 1));
9894 if (!Subtarget->hasVFP2()) // FIXME: NEON?
9899 return V == (V & ((1LL << 8) - 1));
9903 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
9905 int Scale = AM.Scale;
9909 switch (VT.getSimpleVT().SimpleTy) {
9910 default: return false;
9919 return Scale == 2 || Scale == 4 || Scale == 8;
9922 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
9926 // Note, we allow "void" uses (basically, uses that aren't loads or
9927 // stores), because arm allows folding a scale into many arithmetic
9928 // operations. This should be made more precise and revisited later.
9930 // Allow r << imm, but the imm has to be a multiple of two.
9931 if (Scale & 1) return false;
9932 return isPowerOf2_32(Scale);
9936 /// isLegalAddressingMode - Return true if the addressing mode represented
9937 /// by AM is legal for this target, for a load/store of the specified type.
9938 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
9940 EVT VT = getValueType(Ty, true);
9941 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
9944 // Can never fold addr of global into load/store.
9949 case 0: // no scale reg, must be "r+i" or "r", or "i".
9952 if (Subtarget->isThumb1Only())
9956 // ARM doesn't support any R+R*scale+imm addr modes.
9963 if (Subtarget->isThumb2())
9964 return isLegalT2ScaledAddressingMode(AM, VT);
9966 int Scale = AM.Scale;
9967 switch (VT.getSimpleVT().SimpleTy) {
9968 default: return false;
9972 if (Scale < 0) Scale = -Scale;
9976 return isPowerOf2_32(Scale & ~1);
9980 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
9985 // Note, we allow "void" uses (basically, uses that aren't loads or
9986 // stores), because arm allows folding a scale into many arithmetic
9987 // operations. This should be made more precise and revisited later.
9989 // Allow r << imm, but the imm has to be a multiple of two.
9990 if (Scale & 1) return false;
9991 return isPowerOf2_32(Scale);
9997 /// isLegalICmpImmediate - Return true if the specified immediate is legal
9998 /// icmp immediate, that is the target has icmp instructions which can compare
9999 /// a register against the immediate without having to materialize the
10000 /// immediate into a register.
10001 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
10002 // Thumb2 and ARM modes can use cmn for negative immediates.
10003 if (!Subtarget->isThumb())
10004 return ARM_AM::getSOImmVal(llvm::abs64(Imm)) != -1;
10005 if (Subtarget->isThumb2())
10006 return ARM_AM::getT2SOImmVal(llvm::abs64(Imm)) != -1;
10007 // Thumb1 doesn't have cmn, and only 8-bit immediates.
10008 return Imm >= 0 && Imm <= 255;
10011 /// isLegalAddImmediate - Return true if the specified immediate is a legal add
10012 /// *or sub* immediate, that is the target has add or sub instructions which can
10013 /// add a register with the immediate without having to materialize the
10014 /// immediate into a register.
10015 bool ARMTargetLowering::isLegalAddImmediate(int64_t Imm) const {
10016 // Same encoding for add/sub, just flip the sign.
10017 int64_t AbsImm = llvm::abs64(Imm);
10018 if (!Subtarget->isThumb())
10019 return ARM_AM::getSOImmVal(AbsImm) != -1;
10020 if (Subtarget->isThumb2())
10021 return ARM_AM::getT2SOImmVal(AbsImm) != -1;
10022 // Thumb1 only has 8-bit unsigned immediate.
10023 return AbsImm >= 0 && AbsImm <= 255;
10026 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
10027 bool isSEXTLoad, SDValue &Base,
10028 SDValue &Offset, bool &isInc,
10029 SelectionDAG &DAG) {
10030 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
10033 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
10034 // AddressingMode 3
10035 Base = Ptr->getOperand(0);
10036 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
10037 int RHSC = (int)RHS->getZExtValue();
10038 if (RHSC < 0 && RHSC > -256) {
10039 assert(Ptr->getOpcode() == ISD::ADD);
10041 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
10045 isInc = (Ptr->getOpcode() == ISD::ADD);
10046 Offset = Ptr->getOperand(1);
10048 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
10049 // AddressingMode 2
10050 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
10051 int RHSC = (int)RHS->getZExtValue();
10052 if (RHSC < 0 && RHSC > -0x1000) {
10053 assert(Ptr->getOpcode() == ISD::ADD);
10055 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
10056 Base = Ptr->getOperand(0);
10061 if (Ptr->getOpcode() == ISD::ADD) {
10063 ARM_AM::ShiftOpc ShOpcVal=
10064 ARM_AM::getShiftOpcForNode(Ptr->getOperand(0).getOpcode());
10065 if (ShOpcVal != ARM_AM::no_shift) {
10066 Base = Ptr->getOperand(1);
10067 Offset = Ptr->getOperand(0);
10069 Base = Ptr->getOperand(0);
10070 Offset = Ptr->getOperand(1);
10075 isInc = (Ptr->getOpcode() == ISD::ADD);
10076 Base = Ptr->getOperand(0);
10077 Offset = Ptr->getOperand(1);
10081 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
10085 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
10086 bool isSEXTLoad, SDValue &Base,
10087 SDValue &Offset, bool &isInc,
10088 SelectionDAG &DAG) {
10089 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
10092 Base = Ptr->getOperand(0);
10093 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
10094 int RHSC = (int)RHS->getZExtValue();
10095 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
10096 assert(Ptr->getOpcode() == ISD::ADD);
10098 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
10100 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
10101 isInc = Ptr->getOpcode() == ISD::ADD;
10102 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
10110 /// getPreIndexedAddressParts - returns true by value, base pointer and
10111 /// offset pointer and addressing mode by reference if the node's address
10112 /// can be legally represented as pre-indexed load / store address.
10114 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
10116 ISD::MemIndexedMode &AM,
10117 SelectionDAG &DAG) const {
10118 if (Subtarget->isThumb1Only())
10123 bool isSEXTLoad = false;
10124 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
10125 Ptr = LD->getBasePtr();
10126 VT = LD->getMemoryVT();
10127 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
10128 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
10129 Ptr = ST->getBasePtr();
10130 VT = ST->getMemoryVT();
10135 bool isLegal = false;
10136 if (Subtarget->isThumb2())
10137 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
10138 Offset, isInc, DAG);
10140 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
10141 Offset, isInc, DAG);
10145 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
10149 /// getPostIndexedAddressParts - returns true by value, base pointer and
10150 /// offset pointer and addressing mode by reference if this node can be
10151 /// combined with a load / store to form a post-indexed load / store.
10152 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
10155 ISD::MemIndexedMode &AM,
10156 SelectionDAG &DAG) const {
10157 if (Subtarget->isThumb1Only())
10162 bool isSEXTLoad = false;
10163 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
10164 VT = LD->getMemoryVT();
10165 Ptr = LD->getBasePtr();
10166 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
10167 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
10168 VT = ST->getMemoryVT();
10169 Ptr = ST->getBasePtr();
10174 bool isLegal = false;
10175 if (Subtarget->isThumb2())
10176 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
10179 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
10185 // Swap base ptr and offset to catch more post-index load / store when
10186 // it's legal. In Thumb2 mode, offset must be an immediate.
10187 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
10188 !Subtarget->isThumb2())
10189 std::swap(Base, Offset);
10191 // Post-indexed load / store update the base pointer.
10196 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
10200 void ARMTargetLowering::computeKnownBitsForTargetNode(const SDValue Op,
10203 const SelectionDAG &DAG,
10204 unsigned Depth) const {
10205 unsigned BitWidth = KnownOne.getBitWidth();
10206 KnownZero = KnownOne = APInt(BitWidth, 0);
10207 switch (Op.getOpcode()) {
10213 // These nodes' second result is a boolean
10214 if (Op.getResNo() == 0)
10216 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - 1);
10218 case ARMISD::CMOV: {
10219 // Bits are known zero/one if known on the LHS and RHS.
10220 DAG.computeKnownBits(Op.getOperand(0), KnownZero, KnownOne, Depth+1);
10221 if (KnownZero == 0 && KnownOne == 0) return;
10223 APInt KnownZeroRHS, KnownOneRHS;
10224 DAG.computeKnownBits(Op.getOperand(1), KnownZeroRHS, KnownOneRHS, Depth+1);
10225 KnownZero &= KnownZeroRHS;
10226 KnownOne &= KnownOneRHS;
10229 case ISD::INTRINSIC_W_CHAIN: {
10230 ConstantSDNode *CN = cast<ConstantSDNode>(Op->getOperand(1));
10231 Intrinsic::ID IntID = static_cast<Intrinsic::ID>(CN->getZExtValue());
10234 case Intrinsic::arm_ldaex:
10235 case Intrinsic::arm_ldrex: {
10236 EVT VT = cast<MemIntrinsicSDNode>(Op)->getMemoryVT();
10237 unsigned MemBits = VT.getScalarType().getSizeInBits();
10238 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - MemBits);
10246 //===----------------------------------------------------------------------===//
10247 // ARM Inline Assembly Support
10248 //===----------------------------------------------------------------------===//
10250 bool ARMTargetLowering::ExpandInlineAsm(CallInst *CI) const {
10251 // Looking for "rev" which is V6+.
10252 if (!Subtarget->hasV6Ops())
10255 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
10256 std::string AsmStr = IA->getAsmString();
10257 SmallVector<StringRef, 4> AsmPieces;
10258 SplitString(AsmStr, AsmPieces, ";\n");
10260 switch (AsmPieces.size()) {
10261 default: return false;
10263 AsmStr = AsmPieces[0];
10265 SplitString(AsmStr, AsmPieces, " \t,");
10268 if (AsmPieces.size() == 3 &&
10269 AsmPieces[0] == "rev" && AsmPieces[1] == "$0" && AsmPieces[2] == "$1" &&
10270 IA->getConstraintString().compare(0, 4, "=l,l") == 0) {
10271 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
10272 if (Ty && Ty->getBitWidth() == 32)
10273 return IntrinsicLowering::LowerToByteSwap(CI);
10281 /// getConstraintType - Given a constraint letter, return the type of
10282 /// constraint it is for this target.
10283 ARMTargetLowering::ConstraintType
10284 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
10285 if (Constraint.size() == 1) {
10286 switch (Constraint[0]) {
10288 case 'l': return C_RegisterClass;
10289 case 'w': return C_RegisterClass;
10290 case 'h': return C_RegisterClass;
10291 case 'x': return C_RegisterClass;
10292 case 't': return C_RegisterClass;
10293 case 'j': return C_Other; // Constant for movw.
10294 // An address with a single base register. Due to the way we
10295 // currently handle addresses it is the same as an 'r' memory constraint.
10296 case 'Q': return C_Memory;
10298 } else if (Constraint.size() == 2) {
10299 switch (Constraint[0]) {
10301 // All 'U+' constraints are addresses.
10302 case 'U': return C_Memory;
10305 return TargetLowering::getConstraintType(Constraint);
10308 /// Examine constraint type and operand type and determine a weight value.
10309 /// This object must already have been set up with the operand type
10310 /// and the current alternative constraint selected.
10311 TargetLowering::ConstraintWeight
10312 ARMTargetLowering::getSingleConstraintMatchWeight(
10313 AsmOperandInfo &info, const char *constraint) const {
10314 ConstraintWeight weight = CW_Invalid;
10315 Value *CallOperandVal = info.CallOperandVal;
10316 // If we don't have a value, we can't do a match,
10317 // but allow it at the lowest weight.
10318 if (!CallOperandVal)
10320 Type *type = CallOperandVal->getType();
10321 // Look at the constraint type.
10322 switch (*constraint) {
10324 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
10327 if (type->isIntegerTy()) {
10328 if (Subtarget->isThumb())
10329 weight = CW_SpecificReg;
10331 weight = CW_Register;
10335 if (type->isFloatingPointTy())
10336 weight = CW_Register;
10342 typedef std::pair<unsigned, const TargetRegisterClass*> RCPair;
10344 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
10346 if (Constraint.size() == 1) {
10347 // GCC ARM Constraint Letters
10348 switch (Constraint[0]) {
10349 case 'l': // Low regs or general regs.
10350 if (Subtarget->isThumb())
10351 return RCPair(0U, &ARM::tGPRRegClass);
10352 return RCPair(0U, &ARM::GPRRegClass);
10353 case 'h': // High regs or no regs.
10354 if (Subtarget->isThumb())
10355 return RCPair(0U, &ARM::hGPRRegClass);
10358 return RCPair(0U, &ARM::GPRRegClass);
10360 if (VT == MVT::Other)
10362 if (VT == MVT::f32)
10363 return RCPair(0U, &ARM::SPRRegClass);
10364 if (VT.getSizeInBits() == 64)
10365 return RCPair(0U, &ARM::DPRRegClass);
10366 if (VT.getSizeInBits() == 128)
10367 return RCPair(0U, &ARM::QPRRegClass);
10370 if (VT == MVT::Other)
10372 if (VT == MVT::f32)
10373 return RCPair(0U, &ARM::SPR_8RegClass);
10374 if (VT.getSizeInBits() == 64)
10375 return RCPair(0U, &ARM::DPR_8RegClass);
10376 if (VT.getSizeInBits() == 128)
10377 return RCPair(0U, &ARM::QPR_8RegClass);
10380 if (VT == MVT::f32)
10381 return RCPair(0U, &ARM::SPRRegClass);
10385 if (StringRef("{cc}").equals_lower(Constraint))
10386 return std::make_pair(unsigned(ARM::CPSR), &ARM::CCRRegClass);
10388 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
10391 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
10392 /// vector. If it is invalid, don't add anything to Ops.
10393 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
10394 std::string &Constraint,
10395 std::vector<SDValue>&Ops,
10396 SelectionDAG &DAG) const {
10399 // Currently only support length 1 constraints.
10400 if (Constraint.length() != 1) return;
10402 char ConstraintLetter = Constraint[0];
10403 switch (ConstraintLetter) {
10406 case 'I': case 'J': case 'K': case 'L':
10407 case 'M': case 'N': case 'O':
10408 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
10412 int64_t CVal64 = C->getSExtValue();
10413 int CVal = (int) CVal64;
10414 // None of these constraints allow values larger than 32 bits. Check
10415 // that the value fits in an int.
10416 if (CVal != CVal64)
10419 switch (ConstraintLetter) {
10421 // Constant suitable for movw, must be between 0 and
10423 if (Subtarget->hasV6T2Ops())
10424 if (CVal >= 0 && CVal <= 65535)
10428 if (Subtarget->isThumb1Only()) {
10429 // This must be a constant between 0 and 255, for ADD
10431 if (CVal >= 0 && CVal <= 255)
10433 } else if (Subtarget->isThumb2()) {
10434 // A constant that can be used as an immediate value in a
10435 // data-processing instruction.
10436 if (ARM_AM::getT2SOImmVal(CVal) != -1)
10439 // A constant that can be used as an immediate value in a
10440 // data-processing instruction.
10441 if (ARM_AM::getSOImmVal(CVal) != -1)
10447 if (Subtarget->isThumb()) { // FIXME thumb2
10448 // This must be a constant between -255 and -1, for negated ADD
10449 // immediates. This can be used in GCC with an "n" modifier that
10450 // prints the negated value, for use with SUB instructions. It is
10451 // not useful otherwise but is implemented for compatibility.
10452 if (CVal >= -255 && CVal <= -1)
10455 // This must be a constant between -4095 and 4095. It is not clear
10456 // what this constraint is intended for. Implemented for
10457 // compatibility with GCC.
10458 if (CVal >= -4095 && CVal <= 4095)
10464 if (Subtarget->isThumb1Only()) {
10465 // A 32-bit value where only one byte has a nonzero value. Exclude
10466 // zero to match GCC. This constraint is used by GCC internally for
10467 // constants that can be loaded with a move/shift combination.
10468 // It is not useful otherwise but is implemented for compatibility.
10469 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
10471 } else if (Subtarget->isThumb2()) {
10472 // A constant whose bitwise inverse can be used as an immediate
10473 // value in a data-processing instruction. This can be used in GCC
10474 // with a "B" modifier that prints the inverted value, for use with
10475 // BIC and MVN instructions. It is not useful otherwise but is
10476 // implemented for compatibility.
10477 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
10480 // A constant whose bitwise inverse can be used as an immediate
10481 // value in a data-processing instruction. This can be used in GCC
10482 // with a "B" modifier that prints the inverted value, for use with
10483 // BIC and MVN instructions. It is not useful otherwise but is
10484 // implemented for compatibility.
10485 if (ARM_AM::getSOImmVal(~CVal) != -1)
10491 if (Subtarget->isThumb1Only()) {
10492 // This must be a constant between -7 and 7,
10493 // for 3-operand ADD/SUB immediate instructions.
10494 if (CVal >= -7 && CVal < 7)
10496 } else if (Subtarget->isThumb2()) {
10497 // A constant whose negation can be used as an immediate value in a
10498 // data-processing instruction. This can be used in GCC with an "n"
10499 // modifier that prints the negated value, for use with SUB
10500 // instructions. It is not useful otherwise but is implemented for
10502 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
10505 // A constant whose negation can be used as an immediate value in a
10506 // data-processing instruction. This can be used in GCC with an "n"
10507 // modifier that prints the negated value, for use with SUB
10508 // instructions. It is not useful otherwise but is implemented for
10510 if (ARM_AM::getSOImmVal(-CVal) != -1)
10516 if (Subtarget->isThumb()) { // FIXME thumb2
10517 // This must be a multiple of 4 between 0 and 1020, for
10518 // ADD sp + immediate.
10519 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
10522 // A power of two or a constant between 0 and 32. This is used in
10523 // GCC for the shift amount on shifted register operands, but it is
10524 // useful in general for any shift amounts.
10525 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
10531 if (Subtarget->isThumb()) { // FIXME thumb2
10532 // This must be a constant between 0 and 31, for shift amounts.
10533 if (CVal >= 0 && CVal <= 31)
10539 if (Subtarget->isThumb()) { // FIXME thumb2
10540 // This must be a multiple of 4 between -508 and 508, for
10541 // ADD/SUB sp = sp + immediate.
10542 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
10547 Result = DAG.getTargetConstant(CVal, Op.getValueType());
10551 if (Result.getNode()) {
10552 Ops.push_back(Result);
10555 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
10558 SDValue ARMTargetLowering::LowerDivRem(SDValue Op, SelectionDAG &DAG) const {
10559 assert(Subtarget->isTargetAEABI() && "Register-based DivRem lowering only");
10560 unsigned Opcode = Op->getOpcode();
10561 assert((Opcode == ISD::SDIVREM || Opcode == ISD::UDIVREM) &&
10562 "Invalid opcode for Div/Rem lowering");
10563 bool isSigned = (Opcode == ISD::SDIVREM);
10564 EVT VT = Op->getValueType(0);
10565 Type *Ty = VT.getTypeForEVT(*DAG.getContext());
10568 switch (VT.getSimpleVT().SimpleTy) {
10569 default: llvm_unreachable("Unexpected request for libcall!");
10570 case MVT::i8: LC= isSigned ? RTLIB::SDIVREM_I8 : RTLIB::UDIVREM_I8; break;
10571 case MVT::i16: LC= isSigned ? RTLIB::SDIVREM_I16 : RTLIB::UDIVREM_I16; break;
10572 case MVT::i32: LC= isSigned ? RTLIB::SDIVREM_I32 : RTLIB::UDIVREM_I32; break;
10573 case MVT::i64: LC= isSigned ? RTLIB::SDIVREM_I64 : RTLIB::UDIVREM_I64; break;
10576 SDValue InChain = DAG.getEntryNode();
10578 TargetLowering::ArgListTy Args;
10579 TargetLowering::ArgListEntry Entry;
10580 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i) {
10581 EVT ArgVT = Op->getOperand(i).getValueType();
10582 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
10583 Entry.Node = Op->getOperand(i);
10585 Entry.isSExt = isSigned;
10586 Entry.isZExt = !isSigned;
10587 Args.push_back(Entry);
10590 SDValue Callee = DAG.getExternalSymbol(getLibcallName(LC),
10593 Type *RetTy = (Type*)StructType::get(Ty, Ty, NULL);
10596 TargetLowering::CallLoweringInfo CLI(DAG);
10597 CLI.setDebugLoc(dl).setChain(InChain)
10598 .setCallee(getLibcallCallingConv(LC), RetTy, Callee, std::move(Args), 0)
10599 .setInRegister().setSExtResult(isSigned).setZExtResult(!isSigned);
10601 std::pair<SDValue, SDValue> CallInfo = LowerCallTo(CLI);
10602 return CallInfo.first;
10606 ARMTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const {
10607 assert(Subtarget->isTargetWindows() && "unsupported target platform");
10611 SDValue Chain = Op.getOperand(0);
10612 SDValue Size = Op.getOperand(1);
10614 SDValue Words = DAG.getNode(ISD::SRL, DL, MVT::i32, Size,
10615 DAG.getConstant(2, MVT::i32));
10618 Chain = DAG.getCopyToReg(Chain, DL, ARM::R4, Words, Flag);
10619 Flag = Chain.getValue(1);
10621 SDVTList NodeTys = DAG.getVTList(MVT::i32, MVT::Glue);
10622 Chain = DAG.getNode(ARMISD::WIN__CHKSTK, DL, NodeTys, Chain, Flag);
10624 SDValue NewSP = DAG.getCopyFromReg(Chain, DL, ARM::SP, MVT::i32);
10625 Chain = NewSP.getValue(1);
10627 SDValue Ops[2] = { NewSP, Chain };
10628 return DAG.getMergeValues(Ops, DL);
10632 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
10633 // The ARM target isn't yet aware of offsets.
10637 bool ARM::isBitFieldInvertedMask(unsigned v) {
10638 if (v == 0xffffffff)
10641 // there can be 1's on either or both "outsides", all the "inside"
10642 // bits must be 0's
10643 unsigned TO = CountTrailingOnes_32(v);
10644 unsigned LO = CountLeadingOnes_32(v);
10645 v = (v >> TO) << TO;
10646 v = (v << LO) >> LO;
10650 /// isFPImmLegal - Returns true if the target can instruction select the
10651 /// specified FP immediate natively. If false, the legalizer will
10652 /// materialize the FP immediate as a load from a constant pool.
10653 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
10654 if (!Subtarget->hasVFP3())
10656 if (VT == MVT::f32)
10657 return ARM_AM::getFP32Imm(Imm) != -1;
10658 if (VT == MVT::f64)
10659 return ARM_AM::getFP64Imm(Imm) != -1;
10663 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
10664 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
10665 /// specified in the intrinsic calls.
10666 bool ARMTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
10668 unsigned Intrinsic) const {
10669 switch (Intrinsic) {
10670 case Intrinsic::arm_neon_vld1:
10671 case Intrinsic::arm_neon_vld2:
10672 case Intrinsic::arm_neon_vld3:
10673 case Intrinsic::arm_neon_vld4:
10674 case Intrinsic::arm_neon_vld2lane:
10675 case Intrinsic::arm_neon_vld3lane:
10676 case Intrinsic::arm_neon_vld4lane: {
10677 Info.opc = ISD::INTRINSIC_W_CHAIN;
10678 // Conservatively set memVT to the entire set of vectors loaded.
10679 uint64_t NumElts = getDataLayout()->getTypeAllocSize(I.getType()) / 8;
10680 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
10681 Info.ptrVal = I.getArgOperand(0);
10683 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
10684 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
10685 Info.vol = false; // volatile loads with NEON intrinsics not supported
10686 Info.readMem = true;
10687 Info.writeMem = false;
10690 case Intrinsic::arm_neon_vst1:
10691 case Intrinsic::arm_neon_vst2:
10692 case Intrinsic::arm_neon_vst3:
10693 case Intrinsic::arm_neon_vst4:
10694 case Intrinsic::arm_neon_vst2lane:
10695 case Intrinsic::arm_neon_vst3lane:
10696 case Intrinsic::arm_neon_vst4lane: {
10697 Info.opc = ISD::INTRINSIC_VOID;
10698 // Conservatively set memVT to the entire set of vectors stored.
10699 unsigned NumElts = 0;
10700 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
10701 Type *ArgTy = I.getArgOperand(ArgI)->getType();
10702 if (!ArgTy->isVectorTy())
10704 NumElts += getDataLayout()->getTypeAllocSize(ArgTy) / 8;
10706 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
10707 Info.ptrVal = I.getArgOperand(0);
10709 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
10710 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
10711 Info.vol = false; // volatile stores with NEON intrinsics not supported
10712 Info.readMem = false;
10713 Info.writeMem = true;
10716 case Intrinsic::arm_ldaex:
10717 case Intrinsic::arm_ldrex: {
10718 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(0)->getType());
10719 Info.opc = ISD::INTRINSIC_W_CHAIN;
10720 Info.memVT = MVT::getVT(PtrTy->getElementType());
10721 Info.ptrVal = I.getArgOperand(0);
10723 Info.align = getDataLayout()->getABITypeAlignment(PtrTy->getElementType());
10725 Info.readMem = true;
10726 Info.writeMem = false;
10729 case Intrinsic::arm_stlex:
10730 case Intrinsic::arm_strex: {
10731 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(1)->getType());
10732 Info.opc = ISD::INTRINSIC_W_CHAIN;
10733 Info.memVT = MVT::getVT(PtrTy->getElementType());
10734 Info.ptrVal = I.getArgOperand(1);
10736 Info.align = getDataLayout()->getABITypeAlignment(PtrTy->getElementType());
10738 Info.readMem = false;
10739 Info.writeMem = true;
10742 case Intrinsic::arm_stlexd:
10743 case Intrinsic::arm_strexd: {
10744 Info.opc = ISD::INTRINSIC_W_CHAIN;
10745 Info.memVT = MVT::i64;
10746 Info.ptrVal = I.getArgOperand(2);
10750 Info.readMem = false;
10751 Info.writeMem = true;
10754 case Intrinsic::arm_ldaexd:
10755 case Intrinsic::arm_ldrexd: {
10756 Info.opc = ISD::INTRINSIC_W_CHAIN;
10757 Info.memVT = MVT::i64;
10758 Info.ptrVal = I.getArgOperand(0);
10762 Info.readMem = true;
10763 Info.writeMem = false;
10773 /// \brief Returns true if it is beneficial to convert a load of a constant
10774 /// to just the constant itself.
10775 bool ARMTargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
10777 assert(Ty->isIntegerTy());
10779 unsigned Bits = Ty->getPrimitiveSizeInBits();
10780 if (Bits == 0 || Bits > 32)
10785 bool ARMTargetLowering::shouldExpandAtomicInIR(Instruction *Inst) const {
10786 // Loads and stores less than 64-bits are already atomic; ones above that
10787 // are doomed anyway, so defer to the default libcall and blame the OS when
10788 // things go wrong. Cortex M doesn't have ldrexd/strexd though, so don't emit
10789 // anything for those.
10790 bool IsMClass = Subtarget->isMClass();
10791 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
10792 unsigned Size = SI->getValueOperand()->getType()->getPrimitiveSizeInBits();
10793 return Size == 64 && !IsMClass;
10794 } else if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
10795 return LI->getType()->getPrimitiveSizeInBits() == 64 && !IsMClass;
10798 // For the real atomic operations, we have ldrex/strex up to 32 bits,
10799 // and up to 64 bits on the non-M profiles
10800 unsigned AtomicLimit = IsMClass ? 32 : 64;
10801 return Inst->getType()->getPrimitiveSizeInBits() <= AtomicLimit;
10804 Value *ARMTargetLowering::emitLoadLinked(IRBuilder<> &Builder, Value *Addr,
10805 AtomicOrdering Ord) const {
10806 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
10807 Type *ValTy = cast<PointerType>(Addr->getType())->getElementType();
10809 Ord == Acquire || Ord == AcquireRelease || Ord == SequentiallyConsistent;
10811 // Since i64 isn't legal and intrinsics don't get type-lowered, the ldrexd
10812 // intrinsic must return {i32, i32} and we have to recombine them into a
10813 // single i64 here.
10814 if (ValTy->getPrimitiveSizeInBits() == 64) {
10815 Intrinsic::ID Int =
10816 IsAcquire ? Intrinsic::arm_ldaexd : Intrinsic::arm_ldrexd;
10817 Function *Ldrex = llvm::Intrinsic::getDeclaration(M, Int);
10819 Addr = Builder.CreateBitCast(Addr, Type::getInt8PtrTy(M->getContext()));
10820 Value *LoHi = Builder.CreateCall(Ldrex, Addr, "lohi");
10822 Value *Lo = Builder.CreateExtractValue(LoHi, 0, "lo");
10823 Value *Hi = Builder.CreateExtractValue(LoHi, 1, "hi");
10824 if (!Subtarget->isLittle())
10825 std::swap (Lo, Hi);
10826 Lo = Builder.CreateZExt(Lo, ValTy, "lo64");
10827 Hi = Builder.CreateZExt(Hi, ValTy, "hi64");
10828 return Builder.CreateOr(
10829 Lo, Builder.CreateShl(Hi, ConstantInt::get(ValTy, 32)), "val64");
10832 Type *Tys[] = { Addr->getType() };
10833 Intrinsic::ID Int = IsAcquire ? Intrinsic::arm_ldaex : Intrinsic::arm_ldrex;
10834 Function *Ldrex = llvm::Intrinsic::getDeclaration(M, Int, Tys);
10836 return Builder.CreateTruncOrBitCast(
10837 Builder.CreateCall(Ldrex, Addr),
10838 cast<PointerType>(Addr->getType())->getElementType());
10841 Value *ARMTargetLowering::emitStoreConditional(IRBuilder<> &Builder, Value *Val,
10843 AtomicOrdering Ord) const {
10844 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
10846 Ord == Release || Ord == AcquireRelease || Ord == SequentiallyConsistent;
10848 // Since the intrinsics must have legal type, the i64 intrinsics take two
10849 // parameters: "i32, i32". We must marshal Val into the appropriate form
10850 // before the call.
10851 if (Val->getType()->getPrimitiveSizeInBits() == 64) {
10852 Intrinsic::ID Int =
10853 IsRelease ? Intrinsic::arm_stlexd : Intrinsic::arm_strexd;
10854 Function *Strex = Intrinsic::getDeclaration(M, Int);
10855 Type *Int32Ty = Type::getInt32Ty(M->getContext());
10857 Value *Lo = Builder.CreateTrunc(Val, Int32Ty, "lo");
10858 Value *Hi = Builder.CreateTrunc(Builder.CreateLShr(Val, 32), Int32Ty, "hi");
10859 if (!Subtarget->isLittle())
10860 std::swap (Lo, Hi);
10861 Addr = Builder.CreateBitCast(Addr, Type::getInt8PtrTy(M->getContext()));
10862 return Builder.CreateCall3(Strex, Lo, Hi, Addr);
10865 Intrinsic::ID Int = IsRelease ? Intrinsic::arm_stlex : Intrinsic::arm_strex;
10866 Type *Tys[] = { Addr->getType() };
10867 Function *Strex = Intrinsic::getDeclaration(M, Int, Tys);
10869 return Builder.CreateCall2(
10870 Strex, Builder.CreateZExtOrBitCast(
10871 Val, Strex->getFunctionType()->getParamType(0)),
10883 static bool isHomogeneousAggregate(Type *Ty, HABaseType &Base,
10884 uint64_t &Members) {
10885 if (const StructType *ST = dyn_cast<StructType>(Ty)) {
10886 for (unsigned i = 0; i < ST->getNumElements(); ++i) {
10887 uint64_t SubMembers = 0;
10888 if (!isHomogeneousAggregate(ST->getElementType(i), Base, SubMembers))
10890 Members += SubMembers;
10892 } else if (const ArrayType *AT = dyn_cast<ArrayType>(Ty)) {
10893 uint64_t SubMembers = 0;
10894 if (!isHomogeneousAggregate(AT->getElementType(), Base, SubMembers))
10896 Members += SubMembers * AT->getNumElements();
10897 } else if (Ty->isFloatTy()) {
10898 if (Base != HA_UNKNOWN && Base != HA_FLOAT)
10902 } else if (Ty->isDoubleTy()) {
10903 if (Base != HA_UNKNOWN && Base != HA_DOUBLE)
10907 } else if (const VectorType *VT = dyn_cast<VectorType>(Ty)) {
10914 return VT->getBitWidth() == 64;
10916 return VT->getBitWidth() == 128;
10918 switch (VT->getBitWidth()) {
10931 return (Members > 0 && Members <= 4);
10934 /// \brief Return true if a type is an AAPCS-VFP homogeneous aggregate.
10935 bool ARMTargetLowering::functionArgumentNeedsConsecutiveRegisters(
10936 Type *Ty, CallingConv::ID CallConv, bool isVarArg) const {
10937 if (getEffectiveCallingConv(CallConv, isVarArg) !=
10938 CallingConv::ARM_AAPCS_VFP)
10941 HABaseType Base = HA_UNKNOWN;
10942 uint64_t Members = 0;
10943 bool result = isHomogeneousAggregate(Ty, Base, Members);
10944 DEBUG(dbgs() << "isHA: " << result << " "; Ty->dump(); dbgs() << "\n");