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);
401 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
402 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
403 for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
404 InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
405 setTruncStoreAction((MVT::SimpleValueType)VT,
406 (MVT::SimpleValueType)InnerVT, Expand);
407 setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT, Expand);
408 setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT, Expand);
409 setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT, Expand);
411 setOperationAction(ISD::MULHS, (MVT::SimpleValueType)VT, Expand);
412 setOperationAction(ISD::SMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
413 setOperationAction(ISD::MULHU, (MVT::SimpleValueType)VT, Expand);
414 setOperationAction(ISD::UMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
416 setOperationAction(ISD::BSWAP, (MVT::SimpleValueType)VT, Expand);
419 setOperationAction(ISD::ConstantFP, MVT::f32, Custom);
420 setOperationAction(ISD::ConstantFP, MVT::f64, Custom);
422 if (Subtarget->hasNEON()) {
423 addDRTypeForNEON(MVT::v2f32);
424 addDRTypeForNEON(MVT::v8i8);
425 addDRTypeForNEON(MVT::v4i16);
426 addDRTypeForNEON(MVT::v2i32);
427 addDRTypeForNEON(MVT::v1i64);
429 addQRTypeForNEON(MVT::v4f32);
430 addQRTypeForNEON(MVT::v2f64);
431 addQRTypeForNEON(MVT::v16i8);
432 addQRTypeForNEON(MVT::v8i16);
433 addQRTypeForNEON(MVT::v4i32);
434 addQRTypeForNEON(MVT::v2i64);
436 // v2f64 is legal so that QR subregs can be extracted as f64 elements, but
437 // neither Neon nor VFP support any arithmetic operations on it.
438 // The same with v4f32. But keep in mind that vadd, vsub, vmul are natively
439 // supported for v4f32.
440 setOperationAction(ISD::FADD, MVT::v2f64, Expand);
441 setOperationAction(ISD::FSUB, MVT::v2f64, Expand);
442 setOperationAction(ISD::FMUL, MVT::v2f64, Expand);
443 // FIXME: Code duplication: FDIV and FREM are expanded always, see
444 // ARMTargetLowering::addTypeForNEON method for details.
445 setOperationAction(ISD::FDIV, MVT::v2f64, Expand);
446 setOperationAction(ISD::FREM, MVT::v2f64, Expand);
447 // FIXME: Create unittest.
448 // In another words, find a way when "copysign" appears in DAG with vector
450 setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand);
451 // FIXME: Code duplication: SETCC has custom operation action, see
452 // ARMTargetLowering::addTypeForNEON method for details.
453 setOperationAction(ISD::SETCC, MVT::v2f64, Expand);
454 // FIXME: Create unittest for FNEG and for FABS.
455 setOperationAction(ISD::FNEG, MVT::v2f64, Expand);
456 setOperationAction(ISD::FABS, MVT::v2f64, Expand);
457 setOperationAction(ISD::FSQRT, MVT::v2f64, Expand);
458 setOperationAction(ISD::FSIN, MVT::v2f64, Expand);
459 setOperationAction(ISD::FCOS, MVT::v2f64, Expand);
460 setOperationAction(ISD::FPOWI, MVT::v2f64, Expand);
461 setOperationAction(ISD::FPOW, MVT::v2f64, Expand);
462 setOperationAction(ISD::FLOG, MVT::v2f64, Expand);
463 setOperationAction(ISD::FLOG2, MVT::v2f64, Expand);
464 setOperationAction(ISD::FLOG10, MVT::v2f64, Expand);
465 setOperationAction(ISD::FEXP, MVT::v2f64, Expand);
466 setOperationAction(ISD::FEXP2, MVT::v2f64, Expand);
467 // FIXME: Create unittest for FCEIL, FTRUNC, FRINT, FNEARBYINT, FFLOOR.
468 setOperationAction(ISD::FCEIL, MVT::v2f64, Expand);
469 setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand);
470 setOperationAction(ISD::FRINT, MVT::v2f64, Expand);
471 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand);
472 setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand);
473 setOperationAction(ISD::FMA, MVT::v2f64, Expand);
475 setOperationAction(ISD::FSQRT, MVT::v4f32, Expand);
476 setOperationAction(ISD::FSIN, MVT::v4f32, Expand);
477 setOperationAction(ISD::FCOS, MVT::v4f32, Expand);
478 setOperationAction(ISD::FPOWI, MVT::v4f32, Expand);
479 setOperationAction(ISD::FPOW, MVT::v4f32, Expand);
480 setOperationAction(ISD::FLOG, MVT::v4f32, Expand);
481 setOperationAction(ISD::FLOG2, MVT::v4f32, Expand);
482 setOperationAction(ISD::FLOG10, MVT::v4f32, Expand);
483 setOperationAction(ISD::FEXP, MVT::v4f32, Expand);
484 setOperationAction(ISD::FEXP2, MVT::v4f32, Expand);
485 setOperationAction(ISD::FCEIL, MVT::v4f32, Expand);
486 setOperationAction(ISD::FTRUNC, MVT::v4f32, Expand);
487 setOperationAction(ISD::FRINT, MVT::v4f32, Expand);
488 setOperationAction(ISD::FNEARBYINT, MVT::v4f32, Expand);
489 setOperationAction(ISD::FFLOOR, MVT::v4f32, Expand);
491 // Mark v2f32 intrinsics.
492 setOperationAction(ISD::FSQRT, MVT::v2f32, Expand);
493 setOperationAction(ISD::FSIN, MVT::v2f32, Expand);
494 setOperationAction(ISD::FCOS, MVT::v2f32, Expand);
495 setOperationAction(ISD::FPOWI, MVT::v2f32, Expand);
496 setOperationAction(ISD::FPOW, MVT::v2f32, Expand);
497 setOperationAction(ISD::FLOG, MVT::v2f32, Expand);
498 setOperationAction(ISD::FLOG2, MVT::v2f32, Expand);
499 setOperationAction(ISD::FLOG10, MVT::v2f32, Expand);
500 setOperationAction(ISD::FEXP, MVT::v2f32, Expand);
501 setOperationAction(ISD::FEXP2, MVT::v2f32, Expand);
502 setOperationAction(ISD::FCEIL, MVT::v2f32, Expand);
503 setOperationAction(ISD::FTRUNC, MVT::v2f32, Expand);
504 setOperationAction(ISD::FRINT, MVT::v2f32, Expand);
505 setOperationAction(ISD::FNEARBYINT, MVT::v2f32, Expand);
506 setOperationAction(ISD::FFLOOR, MVT::v2f32, Expand);
508 // Neon does not support some operations on v1i64 and v2i64 types.
509 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
510 // Custom handling for some quad-vector types to detect VMULL.
511 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
512 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
513 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
514 // Custom handling for some vector types to avoid expensive expansions
515 setOperationAction(ISD::SDIV, MVT::v4i16, Custom);
516 setOperationAction(ISD::SDIV, MVT::v8i8, Custom);
517 setOperationAction(ISD::UDIV, MVT::v4i16, Custom);
518 setOperationAction(ISD::UDIV, MVT::v8i8, Custom);
519 setOperationAction(ISD::SETCC, MVT::v1i64, Expand);
520 setOperationAction(ISD::SETCC, MVT::v2i64, Expand);
521 // Neon does not have single instruction SINT_TO_FP and UINT_TO_FP with
522 // a destination type that is wider than the source, and nor does
523 // it have a FP_TO_[SU]INT instruction with a narrower destination than
525 setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Custom);
526 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
527 setOperationAction(ISD::FP_TO_UINT, MVT::v4i16, Custom);
528 setOperationAction(ISD::FP_TO_SINT, MVT::v4i16, Custom);
530 setOperationAction(ISD::FP_ROUND, MVT::v2f32, Expand);
531 setOperationAction(ISD::FP_EXTEND, MVT::v2f64, Expand);
533 // NEON does not have single instruction CTPOP for vectors with element
534 // types wider than 8-bits. However, custom lowering can leverage the
535 // v8i8/v16i8 vcnt instruction.
536 setOperationAction(ISD::CTPOP, MVT::v2i32, Custom);
537 setOperationAction(ISD::CTPOP, MVT::v4i32, Custom);
538 setOperationAction(ISD::CTPOP, MVT::v4i16, Custom);
539 setOperationAction(ISD::CTPOP, MVT::v8i16, Custom);
541 // NEON only has FMA instructions as of VFP4.
542 if (!Subtarget->hasVFP4()) {
543 setOperationAction(ISD::FMA, MVT::v2f32, Expand);
544 setOperationAction(ISD::FMA, MVT::v4f32, Expand);
547 setTargetDAGCombine(ISD::INTRINSIC_VOID);
548 setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN);
549 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
550 setTargetDAGCombine(ISD::SHL);
551 setTargetDAGCombine(ISD::SRL);
552 setTargetDAGCombine(ISD::SRA);
553 setTargetDAGCombine(ISD::SIGN_EXTEND);
554 setTargetDAGCombine(ISD::ZERO_EXTEND);
555 setTargetDAGCombine(ISD::ANY_EXTEND);
556 setTargetDAGCombine(ISD::SELECT_CC);
557 setTargetDAGCombine(ISD::BUILD_VECTOR);
558 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
559 setTargetDAGCombine(ISD::INSERT_VECTOR_ELT);
560 setTargetDAGCombine(ISD::STORE);
561 setTargetDAGCombine(ISD::FP_TO_SINT);
562 setTargetDAGCombine(ISD::FP_TO_UINT);
563 setTargetDAGCombine(ISD::FDIV);
565 // It is legal to extload from v4i8 to v4i16 or v4i32.
566 MVT Tys[6] = {MVT::v8i8, MVT::v4i8, MVT::v2i8,
567 MVT::v4i16, MVT::v2i16,
569 for (unsigned i = 0; i < 6; ++i) {
570 setLoadExtAction(ISD::EXTLOAD, Tys[i], Legal);
571 setLoadExtAction(ISD::ZEXTLOAD, Tys[i], Legal);
572 setLoadExtAction(ISD::SEXTLOAD, Tys[i], Legal);
576 // ARM and Thumb2 support UMLAL/SMLAL.
577 if (!Subtarget->isThumb1Only())
578 setTargetDAGCombine(ISD::ADDC);
581 computeRegisterProperties();
583 // ARM does not have floating-point extending loads.
584 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
585 setLoadExtAction(ISD::EXTLOAD, MVT::f16, Expand);
587 // ... or truncating stores
588 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
589 setTruncStoreAction(MVT::f32, MVT::f16, Expand);
590 setTruncStoreAction(MVT::f64, MVT::f16, Expand);
592 // ARM does not have i1 sign extending load.
593 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
595 // ARM supports all 4 flavors of integer indexed load / store.
596 if (!Subtarget->isThumb1Only()) {
597 for (unsigned im = (unsigned)ISD::PRE_INC;
598 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
599 setIndexedLoadAction(im, MVT::i1, Legal);
600 setIndexedLoadAction(im, MVT::i8, Legal);
601 setIndexedLoadAction(im, MVT::i16, Legal);
602 setIndexedLoadAction(im, MVT::i32, Legal);
603 setIndexedStoreAction(im, MVT::i1, Legal);
604 setIndexedStoreAction(im, MVT::i8, Legal);
605 setIndexedStoreAction(im, MVT::i16, Legal);
606 setIndexedStoreAction(im, MVT::i32, Legal);
610 setOperationAction(ISD::SADDO, MVT::i32, Custom);
611 setOperationAction(ISD::UADDO, MVT::i32, Custom);
612 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
613 setOperationAction(ISD::USUBO, MVT::i32, Custom);
615 // i64 operation support.
616 setOperationAction(ISD::MUL, MVT::i64, Expand);
617 setOperationAction(ISD::MULHU, MVT::i32, Expand);
618 if (Subtarget->isThumb1Only()) {
619 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
620 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
622 if (Subtarget->isThumb1Only() || !Subtarget->hasV6Ops()
623 || (Subtarget->isThumb2() && !Subtarget->hasThumb2DSP()))
624 setOperationAction(ISD::MULHS, MVT::i32, Expand);
626 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
627 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
628 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
629 setOperationAction(ISD::SRL, MVT::i64, Custom);
630 setOperationAction(ISD::SRA, MVT::i64, Custom);
632 if (!Subtarget->isThumb1Only()) {
633 // FIXME: We should do this for Thumb1 as well.
634 setOperationAction(ISD::ADDC, MVT::i32, Custom);
635 setOperationAction(ISD::ADDE, MVT::i32, Custom);
636 setOperationAction(ISD::SUBC, MVT::i32, Custom);
637 setOperationAction(ISD::SUBE, MVT::i32, Custom);
640 // ARM does not have ROTL.
641 setOperationAction(ISD::ROTL, MVT::i32, Expand);
642 setOperationAction(ISD::CTTZ, MVT::i32, Custom);
643 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
644 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
645 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
647 // These just redirect to CTTZ and CTLZ on ARM.
648 setOperationAction(ISD::CTTZ_ZERO_UNDEF , MVT::i32 , Expand);
649 setOperationAction(ISD::CTLZ_ZERO_UNDEF , MVT::i32 , Expand);
651 setOperationAction(ISD::READCYCLECOUNTER, MVT::i64, Custom);
653 // Only ARMv6 has BSWAP.
654 if (!Subtarget->hasV6Ops())
655 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
657 if (!(Subtarget->hasDivide() && Subtarget->isThumb2()) &&
658 !(Subtarget->hasDivideInARMMode() && !Subtarget->isThumb())) {
659 // These are expanded into libcalls if the cpu doesn't have HW divider.
660 setOperationAction(ISD::SDIV, MVT::i32, Expand);
661 setOperationAction(ISD::UDIV, MVT::i32, Expand);
664 // FIXME: Also set divmod for SREM on EABI
665 setOperationAction(ISD::SREM, MVT::i32, Expand);
666 setOperationAction(ISD::UREM, MVT::i32, Expand);
667 // Register based DivRem for AEABI (RTABI 4.2)
668 if (Subtarget->isTargetAEABI()) {
669 setLibcallName(RTLIB::SDIVREM_I8, "__aeabi_idivmod");
670 setLibcallName(RTLIB::SDIVREM_I16, "__aeabi_idivmod");
671 setLibcallName(RTLIB::SDIVREM_I32, "__aeabi_idivmod");
672 setLibcallName(RTLIB::SDIVREM_I64, "__aeabi_ldivmod");
673 setLibcallName(RTLIB::UDIVREM_I8, "__aeabi_uidivmod");
674 setLibcallName(RTLIB::UDIVREM_I16, "__aeabi_uidivmod");
675 setLibcallName(RTLIB::UDIVREM_I32, "__aeabi_uidivmod");
676 setLibcallName(RTLIB::UDIVREM_I64, "__aeabi_uldivmod");
678 setLibcallCallingConv(RTLIB::SDIVREM_I8, CallingConv::ARM_AAPCS);
679 setLibcallCallingConv(RTLIB::SDIVREM_I16, CallingConv::ARM_AAPCS);
680 setLibcallCallingConv(RTLIB::SDIVREM_I32, CallingConv::ARM_AAPCS);
681 setLibcallCallingConv(RTLIB::SDIVREM_I64, CallingConv::ARM_AAPCS);
682 setLibcallCallingConv(RTLIB::UDIVREM_I8, CallingConv::ARM_AAPCS);
683 setLibcallCallingConv(RTLIB::UDIVREM_I16, CallingConv::ARM_AAPCS);
684 setLibcallCallingConv(RTLIB::UDIVREM_I32, CallingConv::ARM_AAPCS);
685 setLibcallCallingConv(RTLIB::UDIVREM_I64, CallingConv::ARM_AAPCS);
687 setOperationAction(ISD::SDIVREM, MVT::i32, Custom);
688 setOperationAction(ISD::UDIVREM, MVT::i32, Custom);
690 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
691 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
694 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
695 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
696 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
697 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
698 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
700 setOperationAction(ISD::TRAP, MVT::Other, Legal);
702 // Use the default implementation.
703 setOperationAction(ISD::VASTART, MVT::Other, Custom);
704 setOperationAction(ISD::VAARG, MVT::Other, Expand);
705 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
706 setOperationAction(ISD::VAEND, MVT::Other, Expand);
707 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
708 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
710 if (!Subtarget->isTargetMachO()) {
711 // Non-MachO platforms may return values in these registers via the
712 // personality function.
713 setExceptionPointerRegister(ARM::R0);
714 setExceptionSelectorRegister(ARM::R1);
717 if (Subtarget->getTargetTriple().isWindowsItaniumEnvironment())
718 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
720 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
722 // ARMv6 Thumb1 (except for CPUs that support dmb / dsb) and earlier use
723 // the default expansion.
724 if (Subtarget->hasAnyDataBarrier() && !Subtarget->isThumb1Only()) {
725 // ATOMIC_FENCE needs custom lowering; the others should have been expanded
726 // to ldrex/strex loops already.
727 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Custom);
729 // On v8, we have particularly efficient implementations of atomic fences
730 // if they can be combined with nearby atomic loads and stores.
731 if (!Subtarget->hasV8Ops()) {
732 // Automatically insert fences (dmb ist) around ATOMIC_SWAP etc.
733 setInsertFencesForAtomic(true);
736 // If there's anything we can use as a barrier, go through custom lowering
738 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other,
739 Subtarget->hasAnyDataBarrier() ? Custom : Expand);
741 // Set them all for expansion, which will force libcalls.
742 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Expand);
743 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Expand);
744 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Expand);
745 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Expand);
746 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Expand);
747 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Expand);
748 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Expand);
749 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Expand);
750 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i32, Expand);
751 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i32, Expand);
752 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i32, Expand);
753 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i32, Expand);
754 // Mark ATOMIC_LOAD and ATOMIC_STORE custom so we can handle the
755 // Unordered/Monotonic case.
756 setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Custom);
757 setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Custom);
760 setOperationAction(ISD::PREFETCH, MVT::Other, Custom);
762 // Requires SXTB/SXTH, available on v6 and up in both ARM and Thumb modes.
763 if (!Subtarget->hasV6Ops()) {
764 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
765 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
767 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
769 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
770 !Subtarget->isThumb1Only()) {
771 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
772 // iff target supports vfp2.
773 setOperationAction(ISD::BITCAST, MVT::i64, Custom);
774 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
777 // We want to custom lower some of our intrinsics.
778 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
779 if (Subtarget->isTargetDarwin()) {
780 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
781 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
782 setLibcallName(RTLIB::UNWIND_RESUME, "_Unwind_SjLj_Resume");
785 setOperationAction(ISD::SETCC, MVT::i32, Expand);
786 setOperationAction(ISD::SETCC, MVT::f32, Expand);
787 setOperationAction(ISD::SETCC, MVT::f64, Expand);
788 setOperationAction(ISD::SELECT, MVT::i32, Custom);
789 setOperationAction(ISD::SELECT, MVT::f32, Custom);
790 setOperationAction(ISD::SELECT, MVT::f64, Custom);
791 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
792 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
793 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
795 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
796 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
797 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
798 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
799 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
801 // We don't support sin/cos/fmod/copysign/pow
802 setOperationAction(ISD::FSIN, MVT::f64, Expand);
803 setOperationAction(ISD::FSIN, MVT::f32, Expand);
804 setOperationAction(ISD::FCOS, MVT::f32, Expand);
805 setOperationAction(ISD::FCOS, MVT::f64, Expand);
806 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
807 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
808 setOperationAction(ISD::FREM, MVT::f64, Expand);
809 setOperationAction(ISD::FREM, MVT::f32, Expand);
810 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
811 !Subtarget->isThumb1Only()) {
812 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
813 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
815 setOperationAction(ISD::FPOW, MVT::f64, Expand);
816 setOperationAction(ISD::FPOW, MVT::f32, Expand);
818 if (!Subtarget->hasVFP4()) {
819 setOperationAction(ISD::FMA, MVT::f64, Expand);
820 setOperationAction(ISD::FMA, MVT::f32, Expand);
823 // Various VFP goodness
824 if (!TM.Options.UseSoftFloat && !Subtarget->isThumb1Only()) {
825 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
826 if (Subtarget->hasVFP2()) {
827 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
828 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
829 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
830 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
833 // v8 adds f64 <-> f16 conversion. Before that it should be expanded.
834 if (!Subtarget->hasV8Ops()) {
835 setOperationAction(ISD::FP16_TO_FP, MVT::f64, Expand);
836 setOperationAction(ISD::FP_TO_FP16, MVT::f64, Expand);
839 // fp16 is a special v7 extension that adds f16 <-> f32 conversions.
840 if (!Subtarget->hasFP16()) {
841 setOperationAction(ISD::FP16_TO_FP, MVT::f32, Expand);
842 setOperationAction(ISD::FP_TO_FP16, MVT::f32, Expand);
846 // Combine sin / cos into one node or libcall if possible.
847 if (Subtarget->hasSinCos()) {
848 setLibcallName(RTLIB::SINCOS_F32, "sincosf");
849 setLibcallName(RTLIB::SINCOS_F64, "sincos");
850 if (Subtarget->getTargetTriple().getOS() == Triple::IOS) {
851 // For iOS, we don't want to the normal expansion of a libcall to
852 // sincos. We want to issue a libcall to __sincos_stret.
853 setOperationAction(ISD::FSINCOS, MVT::f64, Custom);
854 setOperationAction(ISD::FSINCOS, MVT::f32, Custom);
858 // We have target-specific dag combine patterns for the following nodes:
859 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
860 setTargetDAGCombine(ISD::ADD);
861 setTargetDAGCombine(ISD::SUB);
862 setTargetDAGCombine(ISD::MUL);
863 setTargetDAGCombine(ISD::AND);
864 setTargetDAGCombine(ISD::OR);
865 setTargetDAGCombine(ISD::XOR);
867 if (Subtarget->hasV6Ops())
868 setTargetDAGCombine(ISD::SRL);
870 setStackPointerRegisterToSaveRestore(ARM::SP);
872 if (TM.Options.UseSoftFloat || Subtarget->isThumb1Only() ||
873 !Subtarget->hasVFP2())
874 setSchedulingPreference(Sched::RegPressure);
876 setSchedulingPreference(Sched::Hybrid);
878 //// temporary - rewrite interface to use type
879 MaxStoresPerMemset = 8;
880 MaxStoresPerMemsetOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
881 MaxStoresPerMemcpy = 4; // For @llvm.memcpy -> sequence of stores
882 MaxStoresPerMemcpyOptSize = Subtarget->isTargetDarwin() ? 4 : 2;
883 MaxStoresPerMemmove = 4; // For @llvm.memmove -> sequence of stores
884 MaxStoresPerMemmoveOptSize = Subtarget->isTargetDarwin() ? 4 : 2;
886 // On ARM arguments smaller than 4 bytes are extended, so all arguments
887 // are at least 4 bytes aligned.
888 setMinStackArgumentAlignment(4);
890 // Prefer likely predicted branches to selects on out-of-order cores.
891 PredictableSelectIsExpensive = Subtarget->isLikeA9();
893 setMinFunctionAlignment(Subtarget->isThumb() ? 1 : 2);
896 // FIXME: It might make sense to define the representative register class as the
897 // nearest super-register that has a non-null superset. For example, DPR_VFP2 is
898 // a super-register of SPR, and DPR is a superset if DPR_VFP2. Consequently,
899 // SPR's representative would be DPR_VFP2. This should work well if register
900 // pressure tracking were modified such that a register use would increment the
901 // pressure of the register class's representative and all of it's super
902 // classes' representatives transitively. We have not implemented this because
903 // of the difficulty prior to coalescing of modeling operand register classes
904 // due to the common occurrence of cross class copies and subregister insertions
906 std::pair<const TargetRegisterClass*, uint8_t>
907 ARMTargetLowering::findRepresentativeClass(MVT VT) const{
908 const TargetRegisterClass *RRC = nullptr;
910 switch (VT.SimpleTy) {
912 return TargetLowering::findRepresentativeClass(VT);
913 // Use DPR as representative register class for all floating point
914 // and vector types. Since there are 32 SPR registers and 32 DPR registers so
915 // the cost is 1 for both f32 and f64.
916 case MVT::f32: case MVT::f64: case MVT::v8i8: case MVT::v4i16:
917 case MVT::v2i32: case MVT::v1i64: case MVT::v2f32:
918 RRC = &ARM::DPRRegClass;
919 // When NEON is used for SP, only half of the register file is available
920 // because operations that define both SP and DP results will be constrained
921 // to the VFP2 class (D0-D15). We currently model this constraint prior to
922 // coalescing by double-counting the SP regs. See the FIXME above.
923 if (Subtarget->useNEONForSinglePrecisionFP())
926 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
927 case MVT::v4f32: case MVT::v2f64:
928 RRC = &ARM::DPRRegClass;
932 RRC = &ARM::DPRRegClass;
936 RRC = &ARM::DPRRegClass;
940 return std::make_pair(RRC, Cost);
943 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
945 default: return nullptr;
946 case ARMISD::Wrapper: return "ARMISD::Wrapper";
947 case ARMISD::WrapperPIC: return "ARMISD::WrapperPIC";
948 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
949 case ARMISD::CALL: return "ARMISD::CALL";
950 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
951 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
952 case ARMISD::tCALL: return "ARMISD::tCALL";
953 case ARMISD::BRCOND: return "ARMISD::BRCOND";
954 case ARMISD::BR_JT: return "ARMISD::BR_JT";
955 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
956 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
957 case ARMISD::INTRET_FLAG: return "ARMISD::INTRET_FLAG";
958 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
959 case ARMISD::CMP: return "ARMISD::CMP";
960 case ARMISD::CMN: return "ARMISD::CMN";
961 case ARMISD::CMPZ: return "ARMISD::CMPZ";
962 case ARMISD::CMPFP: return "ARMISD::CMPFP";
963 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
964 case ARMISD::BCC_i64: return "ARMISD::BCC_i64";
965 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
967 case ARMISD::CMOV: return "ARMISD::CMOV";
969 case ARMISD::RBIT: return "ARMISD::RBIT";
971 case ARMISD::FTOSI: return "ARMISD::FTOSI";
972 case ARMISD::FTOUI: return "ARMISD::FTOUI";
973 case ARMISD::SITOF: return "ARMISD::SITOF";
974 case ARMISD::UITOF: return "ARMISD::UITOF";
976 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
977 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
978 case ARMISD::RRX: return "ARMISD::RRX";
980 case ARMISD::ADDC: return "ARMISD::ADDC";
981 case ARMISD::ADDE: return "ARMISD::ADDE";
982 case ARMISD::SUBC: return "ARMISD::SUBC";
983 case ARMISD::SUBE: return "ARMISD::SUBE";
985 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
986 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
988 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
989 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
991 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN";
993 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
995 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
997 case ARMISD::MEMBARRIER_MCR: return "ARMISD::MEMBARRIER_MCR";
999 case ARMISD::PRELOAD: return "ARMISD::PRELOAD";
1001 case ARMISD::WIN__CHKSTK: return "ARMISD:::WIN__CHKSTK";
1003 case ARMISD::VCEQ: return "ARMISD::VCEQ";
1004 case ARMISD::VCEQZ: return "ARMISD::VCEQZ";
1005 case ARMISD::VCGE: return "ARMISD::VCGE";
1006 case ARMISD::VCGEZ: return "ARMISD::VCGEZ";
1007 case ARMISD::VCLEZ: return "ARMISD::VCLEZ";
1008 case ARMISD::VCGEU: return "ARMISD::VCGEU";
1009 case ARMISD::VCGT: return "ARMISD::VCGT";
1010 case ARMISD::VCGTZ: return "ARMISD::VCGTZ";
1011 case ARMISD::VCLTZ: return "ARMISD::VCLTZ";
1012 case ARMISD::VCGTU: return "ARMISD::VCGTU";
1013 case ARMISD::VTST: return "ARMISD::VTST";
1015 case ARMISD::VSHL: return "ARMISD::VSHL";
1016 case ARMISD::VSHRs: return "ARMISD::VSHRs";
1017 case ARMISD::VSHRu: return "ARMISD::VSHRu";
1018 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
1019 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
1020 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
1021 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
1022 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
1023 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
1024 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
1025 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
1026 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
1027 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
1028 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
1029 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
1030 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
1031 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
1032 case ARMISD::VMOVIMM: return "ARMISD::VMOVIMM";
1033 case ARMISD::VMVNIMM: return "ARMISD::VMVNIMM";
1034 case ARMISD::VMOVFPIMM: return "ARMISD::VMOVFPIMM";
1035 case ARMISD::VDUP: return "ARMISD::VDUP";
1036 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
1037 case ARMISD::VEXT: return "ARMISD::VEXT";
1038 case ARMISD::VREV64: return "ARMISD::VREV64";
1039 case ARMISD::VREV32: return "ARMISD::VREV32";
1040 case ARMISD::VREV16: return "ARMISD::VREV16";
1041 case ARMISD::VZIP: return "ARMISD::VZIP";
1042 case ARMISD::VUZP: return "ARMISD::VUZP";
1043 case ARMISD::VTRN: return "ARMISD::VTRN";
1044 case ARMISD::VTBL1: return "ARMISD::VTBL1";
1045 case ARMISD::VTBL2: return "ARMISD::VTBL2";
1046 case ARMISD::VMULLs: return "ARMISD::VMULLs";
1047 case ARMISD::VMULLu: return "ARMISD::VMULLu";
1048 case ARMISD::UMLAL: return "ARMISD::UMLAL";
1049 case ARMISD::SMLAL: return "ARMISD::SMLAL";
1050 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
1051 case ARMISD::FMAX: return "ARMISD::FMAX";
1052 case ARMISD::FMIN: return "ARMISD::FMIN";
1053 case ARMISD::VMAXNM: return "ARMISD::VMAX";
1054 case ARMISD::VMINNM: return "ARMISD::VMIN";
1055 case ARMISD::BFI: return "ARMISD::BFI";
1056 case ARMISD::VORRIMM: return "ARMISD::VORRIMM";
1057 case ARMISD::VBICIMM: return "ARMISD::VBICIMM";
1058 case ARMISD::VBSL: return "ARMISD::VBSL";
1059 case ARMISD::VLD2DUP: return "ARMISD::VLD2DUP";
1060 case ARMISD::VLD3DUP: return "ARMISD::VLD3DUP";
1061 case ARMISD::VLD4DUP: return "ARMISD::VLD4DUP";
1062 case ARMISD::VLD1_UPD: return "ARMISD::VLD1_UPD";
1063 case ARMISD::VLD2_UPD: return "ARMISD::VLD2_UPD";
1064 case ARMISD::VLD3_UPD: return "ARMISD::VLD3_UPD";
1065 case ARMISD::VLD4_UPD: return "ARMISD::VLD4_UPD";
1066 case ARMISD::VLD2LN_UPD: return "ARMISD::VLD2LN_UPD";
1067 case ARMISD::VLD3LN_UPD: return "ARMISD::VLD3LN_UPD";
1068 case ARMISD::VLD4LN_UPD: return "ARMISD::VLD4LN_UPD";
1069 case ARMISD::VLD2DUP_UPD: return "ARMISD::VLD2DUP_UPD";
1070 case ARMISD::VLD3DUP_UPD: return "ARMISD::VLD3DUP_UPD";
1071 case ARMISD::VLD4DUP_UPD: return "ARMISD::VLD4DUP_UPD";
1072 case ARMISD::VST1_UPD: return "ARMISD::VST1_UPD";
1073 case ARMISD::VST2_UPD: return "ARMISD::VST2_UPD";
1074 case ARMISD::VST3_UPD: return "ARMISD::VST3_UPD";
1075 case ARMISD::VST4_UPD: return "ARMISD::VST4_UPD";
1076 case ARMISD::VST2LN_UPD: return "ARMISD::VST2LN_UPD";
1077 case ARMISD::VST3LN_UPD: return "ARMISD::VST3LN_UPD";
1078 case ARMISD::VST4LN_UPD: return "ARMISD::VST4LN_UPD";
1082 EVT ARMTargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
1083 if (!VT.isVector()) return getPointerTy();
1084 return VT.changeVectorElementTypeToInteger();
1087 /// getRegClassFor - Return the register class that should be used for the
1088 /// specified value type.
1089 const TargetRegisterClass *ARMTargetLowering::getRegClassFor(MVT VT) const {
1090 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
1091 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
1092 // load / store 4 to 8 consecutive D registers.
1093 if (Subtarget->hasNEON()) {
1094 if (VT == MVT::v4i64)
1095 return &ARM::QQPRRegClass;
1096 if (VT == MVT::v8i64)
1097 return &ARM::QQQQPRRegClass;
1099 return TargetLowering::getRegClassFor(VT);
1102 // Create a fast isel object.
1104 ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
1105 const TargetLibraryInfo *libInfo) const {
1106 return ARM::createFastISel(funcInfo, libInfo);
1109 /// getMaximalGlobalOffset - Returns the maximal possible offset which can
1110 /// be used for loads / stores from the global.
1111 unsigned ARMTargetLowering::getMaximalGlobalOffset() const {
1112 return (Subtarget->isThumb1Only() ? 127 : 4095);
1115 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
1116 unsigned NumVals = N->getNumValues();
1118 return Sched::RegPressure;
1120 for (unsigned i = 0; i != NumVals; ++i) {
1121 EVT VT = N->getValueType(i);
1122 if (VT == MVT::Glue || VT == MVT::Other)
1124 if (VT.isFloatingPoint() || VT.isVector())
1128 if (!N->isMachineOpcode())
1129 return Sched::RegPressure;
1131 // Load are scheduled for latency even if there instruction itinerary
1132 // is not available.
1133 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1134 const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
1136 if (MCID.getNumDefs() == 0)
1137 return Sched::RegPressure;
1138 if (!Itins->isEmpty() &&
1139 Itins->getOperandCycle(MCID.getSchedClass(), 0) > 2)
1142 return Sched::RegPressure;
1145 //===----------------------------------------------------------------------===//
1147 //===----------------------------------------------------------------------===//
1149 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
1150 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
1152 default: llvm_unreachable("Unknown condition code!");
1153 case ISD::SETNE: return ARMCC::NE;
1154 case ISD::SETEQ: return ARMCC::EQ;
1155 case ISD::SETGT: return ARMCC::GT;
1156 case ISD::SETGE: return ARMCC::GE;
1157 case ISD::SETLT: return ARMCC::LT;
1158 case ISD::SETLE: return ARMCC::LE;
1159 case ISD::SETUGT: return ARMCC::HI;
1160 case ISD::SETUGE: return ARMCC::HS;
1161 case ISD::SETULT: return ARMCC::LO;
1162 case ISD::SETULE: return ARMCC::LS;
1166 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
1167 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
1168 ARMCC::CondCodes &CondCode2) {
1169 CondCode2 = ARMCC::AL;
1171 default: llvm_unreachable("Unknown FP condition!");
1173 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
1175 case ISD::SETOGT: CondCode = ARMCC::GT; break;
1177 case ISD::SETOGE: CondCode = ARMCC::GE; break;
1178 case ISD::SETOLT: CondCode = ARMCC::MI; break;
1179 case ISD::SETOLE: CondCode = ARMCC::LS; break;
1180 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
1181 case ISD::SETO: CondCode = ARMCC::VC; break;
1182 case ISD::SETUO: CondCode = ARMCC::VS; break;
1183 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
1184 case ISD::SETUGT: CondCode = ARMCC::HI; break;
1185 case ISD::SETUGE: CondCode = ARMCC::PL; break;
1187 case ISD::SETULT: CondCode = ARMCC::LT; break;
1189 case ISD::SETULE: CondCode = ARMCC::LE; break;
1191 case ISD::SETUNE: CondCode = ARMCC::NE; break;
1195 //===----------------------------------------------------------------------===//
1196 // Calling Convention Implementation
1197 //===----------------------------------------------------------------------===//
1199 #include "ARMGenCallingConv.inc"
1201 /// getEffectiveCallingConv - Get the effective calling convention, taking into
1202 /// account presence of floating point hardware and calling convention
1203 /// limitations, such as support for variadic functions.
1205 ARMTargetLowering::getEffectiveCallingConv(CallingConv::ID CC,
1206 bool isVarArg) const {
1209 llvm_unreachable("Unsupported calling convention");
1210 case CallingConv::ARM_AAPCS:
1211 case CallingConv::ARM_APCS:
1212 case CallingConv::GHC:
1214 case CallingConv::ARM_AAPCS_VFP:
1215 return isVarArg ? CallingConv::ARM_AAPCS : CallingConv::ARM_AAPCS_VFP;
1216 case CallingConv::C:
1217 if (!Subtarget->isAAPCS_ABI())
1218 return CallingConv::ARM_APCS;
1219 else if (Subtarget->hasVFP2() && !Subtarget->isThumb1Only() &&
1220 getTargetMachine().Options.FloatABIType == FloatABI::Hard &&
1222 return CallingConv::ARM_AAPCS_VFP;
1224 return CallingConv::ARM_AAPCS;
1225 case CallingConv::Fast:
1226 if (!Subtarget->isAAPCS_ABI()) {
1227 if (Subtarget->hasVFP2() && !Subtarget->isThumb1Only() && !isVarArg)
1228 return CallingConv::Fast;
1229 return CallingConv::ARM_APCS;
1230 } else if (Subtarget->hasVFP2() && !Subtarget->isThumb1Only() && !isVarArg)
1231 return CallingConv::ARM_AAPCS_VFP;
1233 return CallingConv::ARM_AAPCS;
1237 /// CCAssignFnForNode - Selects the correct CCAssignFn for the given
1238 /// CallingConvention.
1239 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
1241 bool isVarArg) const {
1242 switch (getEffectiveCallingConv(CC, isVarArg)) {
1244 llvm_unreachable("Unsupported calling convention");
1245 case CallingConv::ARM_APCS:
1246 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1247 case CallingConv::ARM_AAPCS:
1248 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1249 case CallingConv::ARM_AAPCS_VFP:
1250 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1251 case CallingConv::Fast:
1252 return (Return ? RetFastCC_ARM_APCS : FastCC_ARM_APCS);
1253 case CallingConv::GHC:
1254 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS_GHC);
1258 /// LowerCallResult - Lower the result values of a call into the
1259 /// appropriate copies out of appropriate physical registers.
1261 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1262 CallingConv::ID CallConv, bool isVarArg,
1263 const SmallVectorImpl<ISD::InputArg> &Ins,
1264 SDLoc dl, SelectionDAG &DAG,
1265 SmallVectorImpl<SDValue> &InVals,
1266 bool isThisReturn, SDValue ThisVal) const {
1268 // Assign locations to each value returned by this call.
1269 SmallVector<CCValAssign, 16> RVLocs;
1270 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1271 getTargetMachine(), RVLocs, *DAG.getContext(), Call);
1272 CCInfo.AnalyzeCallResult(Ins,
1273 CCAssignFnForNode(CallConv, /* Return*/ true,
1276 // Copy all of the result registers out of their specified physreg.
1277 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1278 CCValAssign VA = RVLocs[i];
1280 // Pass 'this' value directly from the argument to return value, to avoid
1281 // reg unit interference
1282 if (i == 0 && isThisReturn) {
1283 assert(!VA.needsCustom() && VA.getLocVT() == MVT::i32 &&
1284 "unexpected return calling convention register assignment");
1285 InVals.push_back(ThisVal);
1290 if (VA.needsCustom()) {
1291 // Handle f64 or half of a v2f64.
1292 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1294 Chain = Lo.getValue(1);
1295 InFlag = Lo.getValue(2);
1296 VA = RVLocs[++i]; // skip ahead to next loc
1297 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1299 Chain = Hi.getValue(1);
1300 InFlag = Hi.getValue(2);
1301 if (!Subtarget->isLittle())
1303 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1305 if (VA.getLocVT() == MVT::v2f64) {
1306 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1307 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1308 DAG.getConstant(0, MVT::i32));
1310 VA = RVLocs[++i]; // skip ahead to next loc
1311 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1312 Chain = Lo.getValue(1);
1313 InFlag = Lo.getValue(2);
1314 VA = RVLocs[++i]; // skip ahead to next loc
1315 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1316 Chain = Hi.getValue(1);
1317 InFlag = Hi.getValue(2);
1318 if (!Subtarget->isLittle())
1320 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1321 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1322 DAG.getConstant(1, MVT::i32));
1325 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
1327 Chain = Val.getValue(1);
1328 InFlag = Val.getValue(2);
1331 switch (VA.getLocInfo()) {
1332 default: llvm_unreachable("Unknown loc info!");
1333 case CCValAssign::Full: break;
1334 case CCValAssign::BCvt:
1335 Val = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), Val);
1339 InVals.push_back(Val);
1345 /// LowerMemOpCallTo - Store the argument to the stack.
1347 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
1348 SDValue StackPtr, SDValue Arg,
1349 SDLoc dl, SelectionDAG &DAG,
1350 const CCValAssign &VA,
1351 ISD::ArgFlagsTy Flags) const {
1352 unsigned LocMemOffset = VA.getLocMemOffset();
1353 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1354 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1355 return DAG.getStore(Chain, dl, Arg, PtrOff,
1356 MachinePointerInfo::getStack(LocMemOffset),
1360 void ARMTargetLowering::PassF64ArgInRegs(SDLoc dl, SelectionDAG &DAG,
1361 SDValue Chain, SDValue &Arg,
1362 RegsToPassVector &RegsToPass,
1363 CCValAssign &VA, CCValAssign &NextVA,
1365 SmallVectorImpl<SDValue> &MemOpChains,
1366 ISD::ArgFlagsTy Flags) const {
1368 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1369 DAG.getVTList(MVT::i32, MVT::i32), Arg);
1370 unsigned id = Subtarget->isLittle() ? 0 : 1;
1371 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd.getValue(id)));
1373 if (NextVA.isRegLoc())
1374 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1-id)));
1376 assert(NextVA.isMemLoc());
1377 if (!StackPtr.getNode())
1378 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1380 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1-id),
1386 /// LowerCall - Lowering a call into a callseq_start <-
1387 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
1390 ARMTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
1391 SmallVectorImpl<SDValue> &InVals) const {
1392 SelectionDAG &DAG = CLI.DAG;
1394 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
1395 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
1396 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
1397 SDValue Chain = CLI.Chain;
1398 SDValue Callee = CLI.Callee;
1399 bool &isTailCall = CLI.IsTailCall;
1400 CallingConv::ID CallConv = CLI.CallConv;
1401 bool doesNotRet = CLI.DoesNotReturn;
1402 bool isVarArg = CLI.IsVarArg;
1404 MachineFunction &MF = DAG.getMachineFunction();
1405 bool isStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1406 bool isThisReturn = false;
1407 bool isSibCall = false;
1409 // Disable tail calls if they're not supported.
1410 if (!Subtarget->supportsTailCall() || MF.getTarget().Options.DisableTailCalls)
1414 // Check if it's really possible to do a tail call.
1415 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1416 isVarArg, isStructRet, MF.getFunction()->hasStructRetAttr(),
1417 Outs, OutVals, Ins, DAG);
1418 if (!isTailCall && CLI.CS && CLI.CS->isMustTailCall())
1419 report_fatal_error("failed to perform tail call elimination on a call "
1420 "site marked musttail");
1421 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1422 // detected sibcalls.
1429 // Analyze operands of the call, assigning locations to each operand.
1430 SmallVector<CCValAssign, 16> ArgLocs;
1431 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1432 getTargetMachine(), ArgLocs, *DAG.getContext(), Call);
1433 CCInfo.AnalyzeCallOperands(Outs,
1434 CCAssignFnForNode(CallConv, /* Return*/ false,
1437 // Get a count of how many bytes are to be pushed on the stack.
1438 unsigned NumBytes = CCInfo.getNextStackOffset();
1440 // For tail calls, memory operands are available in our caller's stack.
1444 // Adjust the stack pointer for the new arguments...
1445 // These operations are automatically eliminated by the prolog/epilog pass
1447 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true),
1450 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1452 RegsToPassVector RegsToPass;
1453 SmallVector<SDValue, 8> MemOpChains;
1455 // Walk the register/memloc assignments, inserting copies/loads. In the case
1456 // of tail call optimization, arguments are handled later.
1457 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1459 ++i, ++realArgIdx) {
1460 CCValAssign &VA = ArgLocs[i];
1461 SDValue Arg = OutVals[realArgIdx];
1462 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1463 bool isByVal = Flags.isByVal();
1465 // Promote the value if needed.
1466 switch (VA.getLocInfo()) {
1467 default: llvm_unreachable("Unknown loc info!");
1468 case CCValAssign::Full: break;
1469 case CCValAssign::SExt:
1470 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1472 case CCValAssign::ZExt:
1473 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1475 case CCValAssign::AExt:
1476 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1478 case CCValAssign::BCvt:
1479 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1483 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1484 if (VA.needsCustom()) {
1485 if (VA.getLocVT() == MVT::v2f64) {
1486 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1487 DAG.getConstant(0, MVT::i32));
1488 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1489 DAG.getConstant(1, MVT::i32));
1491 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1492 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1494 VA = ArgLocs[++i]; // skip ahead to next loc
1495 if (VA.isRegLoc()) {
1496 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1497 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1499 assert(VA.isMemLoc());
1501 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1502 dl, DAG, VA, Flags));
1505 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1506 StackPtr, MemOpChains, Flags);
1508 } else if (VA.isRegLoc()) {
1509 if (realArgIdx == 0 && Flags.isReturned() && Outs[0].VT == MVT::i32) {
1510 assert(VA.getLocVT() == MVT::i32 &&
1511 "unexpected calling convention register assignment");
1512 assert(!Ins.empty() && Ins[0].VT == MVT::i32 &&
1513 "unexpected use of 'returned'");
1514 isThisReturn = true;
1516 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1517 } else if (isByVal) {
1518 assert(VA.isMemLoc());
1519 unsigned offset = 0;
1521 // True if this byval aggregate will be split between registers
1523 unsigned ByValArgsCount = CCInfo.getInRegsParamsCount();
1524 unsigned CurByValIdx = CCInfo.getInRegsParamsProceed();
1526 if (CurByValIdx < ByValArgsCount) {
1528 unsigned RegBegin, RegEnd;
1529 CCInfo.getInRegsParamInfo(CurByValIdx, RegBegin, RegEnd);
1531 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1533 for (i = 0, j = RegBegin; j < RegEnd; i++, j++) {
1534 SDValue Const = DAG.getConstant(4*i, MVT::i32);
1535 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
1536 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
1537 MachinePointerInfo(),
1538 false, false, false,
1539 DAG.InferPtrAlignment(AddArg));
1540 MemOpChains.push_back(Load.getValue(1));
1541 RegsToPass.push_back(std::make_pair(j, Load));
1544 // If parameter size outsides register area, "offset" value
1545 // helps us to calculate stack slot for remained part properly.
1546 offset = RegEnd - RegBegin;
1548 CCInfo.nextInRegsParam();
1551 if (Flags.getByValSize() > 4*offset) {
1552 unsigned LocMemOffset = VA.getLocMemOffset();
1553 SDValue StkPtrOff = DAG.getIntPtrConstant(LocMemOffset);
1554 SDValue Dst = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr,
1556 SDValue SrcOffset = DAG.getIntPtrConstant(4*offset);
1557 SDValue Src = DAG.getNode(ISD::ADD, dl, getPointerTy(), Arg, SrcOffset);
1558 SDValue SizeNode = DAG.getConstant(Flags.getByValSize() - 4*offset,
1560 SDValue AlignNode = DAG.getConstant(Flags.getByValAlign(), MVT::i32);
1562 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
1563 SDValue Ops[] = { Chain, Dst, Src, SizeNode, AlignNode};
1564 MemOpChains.push_back(DAG.getNode(ARMISD::COPY_STRUCT_BYVAL, dl, VTs,
1567 } else if (!isSibCall) {
1568 assert(VA.isMemLoc());
1570 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1571 dl, DAG, VA, Flags));
1575 if (!MemOpChains.empty())
1576 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains);
1578 // Build a sequence of copy-to-reg nodes chained together with token chain
1579 // and flag operands which copy the outgoing args into the appropriate regs.
1581 // Tail call byval lowering might overwrite argument registers so in case of
1582 // tail call optimization the copies to registers are lowered later.
1584 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1585 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1586 RegsToPass[i].second, InFlag);
1587 InFlag = Chain.getValue(1);
1590 // For tail calls lower the arguments to the 'real' stack slot.
1592 // Force all the incoming stack arguments to be loaded from the stack
1593 // before any new outgoing arguments are stored to the stack, because the
1594 // outgoing stack slots may alias the incoming argument stack slots, and
1595 // the alias isn't otherwise explicit. This is slightly more conservative
1596 // than necessary, because it means that each store effectively depends
1597 // on every argument instead of just those arguments it would clobber.
1599 // Do not flag preceding copytoreg stuff together with the following stuff.
1601 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1602 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1603 RegsToPass[i].second, InFlag);
1604 InFlag = Chain.getValue(1);
1609 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1610 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1611 // node so that legalize doesn't hack it.
1612 bool isDirect = false;
1613 bool isARMFunc = false;
1614 bool isLocalARMFunc = false;
1615 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1617 if (EnableARMLongCalls) {
1618 assert((Subtarget->isTargetWindows() ||
1619 getTargetMachine().getRelocationModel() == Reloc::Static) &&
1620 "long-calls with non-static relocation model!");
1621 // Handle a global address or an external symbol. If it's not one of
1622 // those, the target's already in a register, so we don't need to do
1624 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1625 const GlobalValue *GV = G->getGlobal();
1626 // Create a constant pool entry for the callee address
1627 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1628 ARMConstantPoolValue *CPV =
1629 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 0);
1631 // Get the address of the callee into a register
1632 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1633 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1634 Callee = DAG.getLoad(getPointerTy(), dl,
1635 DAG.getEntryNode(), CPAddr,
1636 MachinePointerInfo::getConstantPool(),
1637 false, false, false, 0);
1638 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1639 const char *Sym = S->getSymbol();
1641 // Create a constant pool entry for the callee address
1642 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1643 ARMConstantPoolValue *CPV =
1644 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1645 ARMPCLabelIndex, 0);
1646 // Get the address of the callee into a register
1647 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1648 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1649 Callee = DAG.getLoad(getPointerTy(), dl,
1650 DAG.getEntryNode(), CPAddr,
1651 MachinePointerInfo::getConstantPool(),
1652 false, false, false, 0);
1654 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1655 const GlobalValue *GV = G->getGlobal();
1657 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1658 bool isStub = (isExt && Subtarget->isTargetMachO()) &&
1659 getTargetMachine().getRelocationModel() != Reloc::Static;
1660 isARMFunc = !Subtarget->isThumb() || isStub;
1661 // ARM call to a local ARM function is predicable.
1662 isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking);
1663 // tBX takes a register source operand.
1664 if (isStub && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1665 assert(Subtarget->isTargetMachO() && "WrapperPIC use on non-MachO?");
1666 Callee = DAG.getNode(ARMISD::WrapperPIC, dl, getPointerTy(),
1667 DAG.getTargetGlobalAddress(GV, dl, getPointerTy()));
1668 } else if (Subtarget->isTargetCOFF()) {
1669 assert(Subtarget->isTargetWindows() &&
1670 "Windows is the only supported COFF target");
1671 unsigned TargetFlags = GV->hasDLLImportStorageClass()
1672 ? ARMII::MO_DLLIMPORT
1673 : ARMII::MO_NO_FLAG;
1674 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), /*Offset=*/0,
1676 if (GV->hasDLLImportStorageClass())
1677 Callee = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
1678 DAG.getNode(ARMISD::Wrapper, dl, getPointerTy(),
1679 Callee), MachinePointerInfo::getGOT(),
1680 false, false, false, 0);
1682 // On ELF targets for PIC code, direct calls should go through the PLT
1683 unsigned OpFlags = 0;
1684 if (Subtarget->isTargetELF() &&
1685 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1686 OpFlags = ARMII::MO_PLT;
1687 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
1689 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1691 bool isStub = Subtarget->isTargetMachO() &&
1692 getTargetMachine().getRelocationModel() != Reloc::Static;
1693 isARMFunc = !Subtarget->isThumb() || isStub;
1694 // tBX takes a register source operand.
1695 const char *Sym = S->getSymbol();
1696 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1697 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1698 ARMConstantPoolValue *CPV =
1699 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1700 ARMPCLabelIndex, 4);
1701 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1702 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1703 Callee = DAG.getLoad(getPointerTy(), dl,
1704 DAG.getEntryNode(), CPAddr,
1705 MachinePointerInfo::getConstantPool(),
1706 false, false, false, 0);
1707 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1708 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1709 getPointerTy(), Callee, PICLabel);
1711 unsigned OpFlags = 0;
1712 // On ELF targets for PIC code, direct calls should go through the PLT
1713 if (Subtarget->isTargetELF() &&
1714 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1715 OpFlags = ARMII::MO_PLT;
1716 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlags);
1720 // FIXME: handle tail calls differently.
1722 bool HasMinSizeAttr = MF.getFunction()->getAttributes().hasAttribute(
1723 AttributeSet::FunctionIndex, Attribute::MinSize);
1724 if (Subtarget->isThumb()) {
1725 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1726 CallOpc = ARMISD::CALL_NOLINK;
1728 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1730 if (!isDirect && !Subtarget->hasV5TOps())
1731 CallOpc = ARMISD::CALL_NOLINK;
1732 else if (doesNotRet && isDirect && Subtarget->hasRAS() &&
1733 // Emit regular call when code size is the priority
1735 // "mov lr, pc; b _foo" to avoid confusing the RSP
1736 CallOpc = ARMISD::CALL_NOLINK;
1738 CallOpc = isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL;
1741 std::vector<SDValue> Ops;
1742 Ops.push_back(Chain);
1743 Ops.push_back(Callee);
1745 // Add argument registers to the end of the list so that they are known live
1747 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1748 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1749 RegsToPass[i].second.getValueType()));
1751 // Add a register mask operand representing the call-preserved registers.
1753 const uint32_t *Mask;
1754 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
1755 const ARMBaseRegisterInfo *ARI = static_cast<const ARMBaseRegisterInfo*>(TRI);
1757 // For 'this' returns, use the R0-preserving mask if applicable
1758 Mask = ARI->getThisReturnPreservedMask(CallConv);
1760 // Set isThisReturn to false if the calling convention is not one that
1761 // allows 'returned' to be modeled in this way, so LowerCallResult does
1762 // not try to pass 'this' straight through
1763 isThisReturn = false;
1764 Mask = ARI->getCallPreservedMask(CallConv);
1767 Mask = ARI->getCallPreservedMask(CallConv);
1769 assert(Mask && "Missing call preserved mask for calling convention");
1770 Ops.push_back(DAG.getRegisterMask(Mask));
1773 if (InFlag.getNode())
1774 Ops.push_back(InFlag);
1776 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
1778 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, Ops);
1780 // Returns a chain and a flag for retval copy to use.
1781 Chain = DAG.getNode(CallOpc, dl, NodeTys, Ops);
1782 InFlag = Chain.getValue(1);
1784 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1785 DAG.getIntPtrConstant(0, true), InFlag, dl);
1787 InFlag = Chain.getValue(1);
1789 // Handle result values, copying them out of physregs into vregs that we
1791 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins, dl, DAG,
1792 InVals, isThisReturn,
1793 isThisReturn ? OutVals[0] : SDValue());
1796 /// HandleByVal - Every parameter *after* a byval parameter is passed
1797 /// on the stack. Remember the next parameter register to allocate,
1798 /// and then confiscate the rest of the parameter registers to insure
1801 ARMTargetLowering::HandleByVal(
1802 CCState *State, unsigned &size, unsigned Align) const {
1803 unsigned reg = State->AllocateReg(GPRArgRegs, 4);
1804 assert((State->getCallOrPrologue() == Prologue ||
1805 State->getCallOrPrologue() == Call) &&
1806 "unhandled ParmContext");
1808 if ((ARM::R0 <= reg) && (reg <= ARM::R3)) {
1809 if (Subtarget->isAAPCS_ABI() && Align > 4) {
1810 unsigned AlignInRegs = Align / 4;
1811 unsigned Waste = (ARM::R4 - reg) % AlignInRegs;
1812 for (unsigned i = 0; i < Waste; ++i)
1813 reg = State->AllocateReg(GPRArgRegs, 4);
1816 unsigned excess = 4 * (ARM::R4 - reg);
1818 // Special case when NSAA != SP and parameter size greater than size of
1819 // all remained GPR regs. In that case we can't split parameter, we must
1820 // send it to stack. We also must set NCRN to R4, so waste all
1821 // remained registers.
1822 const unsigned NSAAOffset = State->getNextStackOffset();
1823 if (Subtarget->isAAPCS_ABI() && NSAAOffset != 0 && size > excess) {
1824 while (State->AllocateReg(GPRArgRegs, 4))
1829 // First register for byval parameter is the first register that wasn't
1830 // allocated before this method call, so it would be "reg".
1831 // If parameter is small enough to be saved in range [reg, r4), then
1832 // the end (first after last) register would be reg + param-size-in-regs,
1833 // else parameter would be splitted between registers and stack,
1834 // end register would be r4 in this case.
1835 unsigned ByValRegBegin = reg;
1836 unsigned ByValRegEnd = (size < excess) ? reg + size/4 : (unsigned)ARM::R4;
1837 State->addInRegsParamInfo(ByValRegBegin, ByValRegEnd);
1838 // Note, first register is allocated in the beginning of function already,
1839 // allocate remained amount of registers we need.
1840 for (unsigned i = reg+1; i != ByValRegEnd; ++i)
1841 State->AllocateReg(GPRArgRegs, 4);
1842 // A byval parameter that is split between registers and memory needs its
1843 // size truncated here.
1844 // In the case where the entire structure fits in registers, we set the
1845 // size in memory to zero.
1854 /// MatchingStackOffset - Return true if the given stack call argument is
1855 /// already available in the same position (relatively) of the caller's
1856 /// incoming argument stack.
1858 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1859 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1860 const TargetInstrInfo *TII) {
1861 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1863 if (Arg.getOpcode() == ISD::CopyFromReg) {
1864 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1865 if (!TargetRegisterInfo::isVirtualRegister(VR))
1867 MachineInstr *Def = MRI->getVRegDef(VR);
1870 if (!Flags.isByVal()) {
1871 if (!TII->isLoadFromStackSlot(Def, FI))
1876 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
1877 if (Flags.isByVal())
1878 // ByVal argument is passed in as a pointer but it's now being
1879 // dereferenced. e.g.
1880 // define @foo(%struct.X* %A) {
1881 // tail call @bar(%struct.X* byval %A)
1884 SDValue Ptr = Ld->getBasePtr();
1885 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
1888 FI = FINode->getIndex();
1892 assert(FI != INT_MAX);
1893 if (!MFI->isFixedObjectIndex(FI))
1895 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
1898 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
1899 /// for tail call optimization. Targets which want to do tail call
1900 /// optimization should implement this function.
1902 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1903 CallingConv::ID CalleeCC,
1905 bool isCalleeStructRet,
1906 bool isCallerStructRet,
1907 const SmallVectorImpl<ISD::OutputArg> &Outs,
1908 const SmallVectorImpl<SDValue> &OutVals,
1909 const SmallVectorImpl<ISD::InputArg> &Ins,
1910 SelectionDAG& DAG) const {
1911 const Function *CallerF = DAG.getMachineFunction().getFunction();
1912 CallingConv::ID CallerCC = CallerF->getCallingConv();
1913 bool CCMatch = CallerCC == CalleeCC;
1915 // Look for obvious safe cases to perform tail call optimization that do not
1916 // require ABI changes. This is what gcc calls sibcall.
1918 // Do not sibcall optimize vararg calls unless the call site is not passing
1920 if (isVarArg && !Outs.empty())
1923 // Exception-handling functions need a special set of instructions to indicate
1924 // a return to the hardware. Tail-calling another function would probably
1926 if (CallerF->hasFnAttribute("interrupt"))
1929 // Also avoid sibcall optimization if either caller or callee uses struct
1930 // return semantics.
1931 if (isCalleeStructRet || isCallerStructRet)
1934 // FIXME: Completely disable sibcall for Thumb1 since Thumb1RegisterInfo::
1935 // emitEpilogue is not ready for them. Thumb tail calls also use t2B, as
1936 // the Thumb1 16-bit unconditional branch doesn't have sufficient relocation
1937 // support in the assembler and linker to be used. This would need to be
1938 // fixed to fully support tail calls in Thumb1.
1940 // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take
1941 // LR. This means if we need to reload LR, it takes an extra instructions,
1942 // which outweighs the value of the tail call; but here we don't know yet
1943 // whether LR is going to be used. Probably the right approach is to
1944 // generate the tail call here and turn it back into CALL/RET in
1945 // emitEpilogue if LR is used.
1947 // Thumb1 PIC calls to external symbols use BX, so they can be tail calls,
1948 // but we need to make sure there are enough registers; the only valid
1949 // registers are the 4 used for parameters. We don't currently do this
1951 if (Subtarget->isThumb1Only())
1954 // If the calling conventions do not match, then we'd better make sure the
1955 // results are returned in the same way as what the caller expects.
1957 SmallVector<CCValAssign, 16> RVLocs1;
1958 ARMCCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(),
1959 getTargetMachine(), RVLocs1, *DAG.getContext(), Call);
1960 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
1962 SmallVector<CCValAssign, 16> RVLocs2;
1963 ARMCCState CCInfo2(CallerCC, false, DAG.getMachineFunction(),
1964 getTargetMachine(), RVLocs2, *DAG.getContext(), Call);
1965 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
1967 if (RVLocs1.size() != RVLocs2.size())
1969 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
1970 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
1972 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
1974 if (RVLocs1[i].isRegLoc()) {
1975 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
1978 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
1984 // If Caller's vararg or byval argument has been split between registers and
1985 // stack, do not perform tail call, since part of the argument is in caller's
1987 const ARMFunctionInfo *AFI_Caller = DAG.getMachineFunction().
1988 getInfo<ARMFunctionInfo>();
1989 if (AFI_Caller->getArgRegsSaveSize())
1992 // If the callee takes no arguments then go on to check the results of the
1994 if (!Outs.empty()) {
1995 // Check if stack adjustment is needed. For now, do not do this if any
1996 // argument is passed on the stack.
1997 SmallVector<CCValAssign, 16> ArgLocs;
1998 ARMCCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(),
1999 getTargetMachine(), ArgLocs, *DAG.getContext(), Call);
2000 CCInfo.AnalyzeCallOperands(Outs,
2001 CCAssignFnForNode(CalleeCC, false, isVarArg));
2002 if (CCInfo.getNextStackOffset()) {
2003 MachineFunction &MF = DAG.getMachineFunction();
2005 // Check if the arguments are already laid out in the right way as
2006 // the caller's fixed stack objects.
2007 MachineFrameInfo *MFI = MF.getFrameInfo();
2008 const MachineRegisterInfo *MRI = &MF.getRegInfo();
2009 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
2010 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
2012 ++i, ++realArgIdx) {
2013 CCValAssign &VA = ArgLocs[i];
2014 EVT RegVT = VA.getLocVT();
2015 SDValue Arg = OutVals[realArgIdx];
2016 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
2017 if (VA.getLocInfo() == CCValAssign::Indirect)
2019 if (VA.needsCustom()) {
2020 // f64 and vector types are split into multiple registers or
2021 // register/stack-slot combinations. The types will not match
2022 // the registers; give up on memory f64 refs until we figure
2023 // out what to do about this.
2026 if (!ArgLocs[++i].isRegLoc())
2028 if (RegVT == MVT::v2f64) {
2029 if (!ArgLocs[++i].isRegLoc())
2031 if (!ArgLocs[++i].isRegLoc())
2034 } else if (!VA.isRegLoc()) {
2035 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
2047 ARMTargetLowering::CanLowerReturn(CallingConv::ID CallConv,
2048 MachineFunction &MF, bool isVarArg,
2049 const SmallVectorImpl<ISD::OutputArg> &Outs,
2050 LLVMContext &Context) const {
2051 SmallVector<CCValAssign, 16> RVLocs;
2052 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(), RVLocs, Context);
2053 return CCInfo.CheckReturn(Outs, CCAssignFnForNode(CallConv, /*Return=*/true,
2057 static SDValue LowerInterruptReturn(SmallVectorImpl<SDValue> &RetOps,
2058 SDLoc DL, SelectionDAG &DAG) {
2059 const MachineFunction &MF = DAG.getMachineFunction();
2060 const Function *F = MF.getFunction();
2062 StringRef IntKind = F->getFnAttribute("interrupt").getValueAsString();
2064 // See ARM ARM v7 B1.8.3. On exception entry LR is set to a possibly offset
2065 // version of the "preferred return address". These offsets affect the return
2066 // instruction if this is a return from PL1 without hypervisor extensions.
2067 // IRQ/FIQ: +4 "subs pc, lr, #4"
2068 // SWI: 0 "subs pc, lr, #0"
2069 // ABORT: +4 "subs pc, lr, #4"
2070 // UNDEF: +4/+2 "subs pc, lr, #0"
2071 // UNDEF varies depending on where the exception came from ARM or Thumb
2072 // mode. Alongside GCC, we throw our hands up in disgust and pretend it's 0.
2075 if (IntKind == "" || IntKind == "IRQ" || IntKind == "FIQ" ||
2078 else if (IntKind == "SWI" || IntKind == "UNDEF")
2081 report_fatal_error("Unsupported interrupt attribute. If present, value "
2082 "must be one of: IRQ, FIQ, SWI, ABORT or UNDEF");
2084 RetOps.insert(RetOps.begin() + 1, DAG.getConstant(LROffset, MVT::i32, false));
2086 return DAG.getNode(ARMISD::INTRET_FLAG, DL, MVT::Other, RetOps);
2090 ARMTargetLowering::LowerReturn(SDValue Chain,
2091 CallingConv::ID CallConv, bool isVarArg,
2092 const SmallVectorImpl<ISD::OutputArg> &Outs,
2093 const SmallVectorImpl<SDValue> &OutVals,
2094 SDLoc dl, SelectionDAG &DAG) const {
2096 // CCValAssign - represent the assignment of the return value to a location.
2097 SmallVector<CCValAssign, 16> RVLocs;
2099 // CCState - Info about the registers and stack slots.
2100 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2101 getTargetMachine(), RVLocs, *DAG.getContext(), Call);
2103 // Analyze outgoing return values.
2104 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
2108 SmallVector<SDValue, 4> RetOps;
2109 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
2110 bool isLittleEndian = Subtarget->isLittle();
2112 // Copy the result values into the output registers.
2113 for (unsigned i = 0, realRVLocIdx = 0;
2115 ++i, ++realRVLocIdx) {
2116 CCValAssign &VA = RVLocs[i];
2117 assert(VA.isRegLoc() && "Can only return in registers!");
2119 SDValue Arg = OutVals[realRVLocIdx];
2121 switch (VA.getLocInfo()) {
2122 default: llvm_unreachable("Unknown loc info!");
2123 case CCValAssign::Full: break;
2124 case CCValAssign::BCvt:
2125 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
2129 if (VA.needsCustom()) {
2130 if (VA.getLocVT() == MVT::v2f64) {
2131 // Extract the first half and return it in two registers.
2132 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
2133 DAG.getConstant(0, MVT::i32));
2134 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
2135 DAG.getVTList(MVT::i32, MVT::i32), Half);
2137 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2138 HalfGPRs.getValue(isLittleEndian ? 0 : 1),
2140 Flag = Chain.getValue(1);
2141 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2142 VA = RVLocs[++i]; // skip ahead to next loc
2143 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2144 HalfGPRs.getValue(isLittleEndian ? 1 : 0),
2146 Flag = Chain.getValue(1);
2147 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2148 VA = RVLocs[++i]; // skip ahead to next loc
2150 // Extract the 2nd half and fall through to handle it as an f64 value.
2151 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
2152 DAG.getConstant(1, MVT::i32));
2154 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
2156 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
2157 DAG.getVTList(MVT::i32, MVT::i32), Arg);
2158 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2159 fmrrd.getValue(isLittleEndian ? 0 : 1),
2161 Flag = Chain.getValue(1);
2162 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2163 VA = RVLocs[++i]; // skip ahead to next loc
2164 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2165 fmrrd.getValue(isLittleEndian ? 1 : 0),
2168 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
2170 // Guarantee that all emitted copies are
2171 // stuck together, avoiding something bad.
2172 Flag = Chain.getValue(1);
2173 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2176 // Update chain and glue.
2179 RetOps.push_back(Flag);
2181 // CPUs which aren't M-class use a special sequence to return from
2182 // exceptions (roughly, any instruction setting pc and cpsr simultaneously,
2183 // though we use "subs pc, lr, #N").
2185 // M-class CPUs actually use a normal return sequence with a special
2186 // (hardware-provided) value in LR, so the normal code path works.
2187 if (DAG.getMachineFunction().getFunction()->hasFnAttribute("interrupt") &&
2188 !Subtarget->isMClass()) {
2189 if (Subtarget->isThumb1Only())
2190 report_fatal_error("interrupt attribute is not supported in Thumb1");
2191 return LowerInterruptReturn(RetOps, dl, DAG);
2194 return DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, RetOps);
2197 bool ARMTargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
2198 if (N->getNumValues() != 1)
2200 if (!N->hasNUsesOfValue(1, 0))
2203 SDValue TCChain = Chain;
2204 SDNode *Copy = *N->use_begin();
2205 if (Copy->getOpcode() == ISD::CopyToReg) {
2206 // If the copy has a glue operand, we conservatively assume it isn't safe to
2207 // perform a tail call.
2208 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
2210 TCChain = Copy->getOperand(0);
2211 } else if (Copy->getOpcode() == ARMISD::VMOVRRD) {
2212 SDNode *VMov = Copy;
2213 // f64 returned in a pair of GPRs.
2214 SmallPtrSet<SDNode*, 2> Copies;
2215 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
2217 if (UI->getOpcode() != ISD::CopyToReg)
2221 if (Copies.size() > 2)
2224 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
2226 SDValue UseChain = UI->getOperand(0);
2227 if (Copies.count(UseChain.getNode()))
2234 } else if (Copy->getOpcode() == ISD::BITCAST) {
2235 // f32 returned in a single GPR.
2236 if (!Copy->hasOneUse())
2238 Copy = *Copy->use_begin();
2239 if (Copy->getOpcode() != ISD::CopyToReg || !Copy->hasNUsesOfValue(1, 0))
2241 TCChain = Copy->getOperand(0);
2246 bool HasRet = false;
2247 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
2249 if (UI->getOpcode() != ARMISD::RET_FLAG &&
2250 UI->getOpcode() != ARMISD::INTRET_FLAG)
2262 bool ARMTargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
2263 if (!Subtarget->supportsTailCall())
2266 if (!CI->isTailCall() || getTargetMachine().Options.DisableTailCalls)
2269 return !Subtarget->isThumb1Only();
2272 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
2273 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
2274 // one of the above mentioned nodes. It has to be wrapped because otherwise
2275 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
2276 // be used to form addressing mode. These wrapped nodes will be selected
2278 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
2279 EVT PtrVT = Op.getValueType();
2280 // FIXME there is no actual debug info here
2282 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
2284 if (CP->isMachineConstantPoolEntry())
2285 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
2286 CP->getAlignment());
2288 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
2289 CP->getAlignment());
2290 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
2293 unsigned ARMTargetLowering::getJumpTableEncoding() const {
2294 return MachineJumpTableInfo::EK_Inline;
2297 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
2298 SelectionDAG &DAG) const {
2299 MachineFunction &MF = DAG.getMachineFunction();
2300 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2301 unsigned ARMPCLabelIndex = 0;
2303 EVT PtrVT = getPointerTy();
2304 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
2305 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2307 if (RelocM == Reloc::Static) {
2308 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
2310 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2311 ARMPCLabelIndex = AFI->createPICLabelUId();
2312 ARMConstantPoolValue *CPV =
2313 ARMConstantPoolConstant::Create(BA, ARMPCLabelIndex,
2314 ARMCP::CPBlockAddress, PCAdj);
2315 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2317 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
2318 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
2319 MachinePointerInfo::getConstantPool(),
2320 false, false, false, 0);
2321 if (RelocM == Reloc::Static)
2323 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2324 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
2327 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
2329 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
2330 SelectionDAG &DAG) const {
2332 EVT PtrVT = getPointerTy();
2333 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2334 MachineFunction &MF = DAG.getMachineFunction();
2335 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2336 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2337 ARMConstantPoolValue *CPV =
2338 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2339 ARMCP::CPValue, PCAdj, ARMCP::TLSGD, true);
2340 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2341 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
2342 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
2343 MachinePointerInfo::getConstantPool(),
2344 false, false, false, 0);
2345 SDValue Chain = Argument.getValue(1);
2347 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2348 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
2350 // call __tls_get_addr.
2353 Entry.Node = Argument;
2354 Entry.Ty = (Type *) Type::getInt32Ty(*DAG.getContext());
2355 Args.push_back(Entry);
2357 // FIXME: is there useful debug info available here?
2358 TargetLowering::CallLoweringInfo CLI(DAG);
2359 CLI.setDebugLoc(dl).setChain(Chain)
2360 .setCallee(CallingConv::C, Type::getInt32Ty(*DAG.getContext()),
2361 DAG.getExternalSymbol("__tls_get_addr", PtrVT), std::move(Args),
2364 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
2365 return CallResult.first;
2368 // Lower ISD::GlobalTLSAddress using the "initial exec" or
2369 // "local exec" model.
2371 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
2373 TLSModel::Model model) const {
2374 const GlobalValue *GV = GA->getGlobal();
2377 SDValue Chain = DAG.getEntryNode();
2378 EVT PtrVT = getPointerTy();
2379 // Get the Thread Pointer
2380 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2382 if (model == TLSModel::InitialExec) {
2383 MachineFunction &MF = DAG.getMachineFunction();
2384 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2385 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2386 // Initial exec model.
2387 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2388 ARMConstantPoolValue *CPV =
2389 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2390 ARMCP::CPValue, PCAdj, ARMCP::GOTTPOFF,
2392 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2393 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2394 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2395 MachinePointerInfo::getConstantPool(),
2396 false, false, false, 0);
2397 Chain = Offset.getValue(1);
2399 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2400 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
2402 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2403 MachinePointerInfo::getConstantPool(),
2404 false, false, false, 0);
2407 assert(model == TLSModel::LocalExec);
2408 ARMConstantPoolValue *CPV =
2409 ARMConstantPoolConstant::Create(GV, ARMCP::TPOFF);
2410 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2411 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2412 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2413 MachinePointerInfo::getConstantPool(),
2414 false, false, false, 0);
2417 // The address of the thread local variable is the add of the thread
2418 // pointer with the offset of the variable.
2419 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
2423 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
2424 // TODO: implement the "local dynamic" model
2425 assert(Subtarget->isTargetELF() &&
2426 "TLS not implemented for non-ELF targets");
2427 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
2429 TLSModel::Model model = getTargetMachine().getTLSModel(GA->getGlobal());
2432 case TLSModel::GeneralDynamic:
2433 case TLSModel::LocalDynamic:
2434 return LowerToTLSGeneralDynamicModel(GA, DAG);
2435 case TLSModel::InitialExec:
2436 case TLSModel::LocalExec:
2437 return LowerToTLSExecModels(GA, DAG, model);
2439 llvm_unreachable("bogus TLS model");
2442 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
2443 SelectionDAG &DAG) const {
2444 EVT PtrVT = getPointerTy();
2446 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2447 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2448 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
2449 ARMConstantPoolValue *CPV =
2450 ARMConstantPoolConstant::Create(GV,
2451 UseGOTOFF ? ARMCP::GOTOFF : ARMCP::GOT);
2452 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2453 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2454 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
2456 MachinePointerInfo::getConstantPool(),
2457 false, false, false, 0);
2458 SDValue Chain = Result.getValue(1);
2459 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
2460 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
2462 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
2463 MachinePointerInfo::getGOT(),
2464 false, false, false, 0);
2468 // If we have T2 ops, we can materialize the address directly via movt/movw
2469 // pair. This is always cheaper.
2470 if (Subtarget->useMovt(DAG.getMachineFunction())) {
2472 // FIXME: Once remat is capable of dealing with instructions with register
2473 // operands, expand this into two nodes.
2474 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2475 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2477 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2478 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2479 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2480 MachinePointerInfo::getConstantPool(),
2481 false, false, false, 0);
2485 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
2486 SelectionDAG &DAG) const {
2487 EVT PtrVT = getPointerTy();
2489 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2490 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2492 if (Subtarget->useMovt(DAG.getMachineFunction()))
2495 // FIXME: Once remat is capable of dealing with instructions with register
2496 // operands, expand this into multiple nodes
2498 RelocM == Reloc::PIC_ ? ARMISD::WrapperPIC : ARMISD::Wrapper;
2500 SDValue G = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, ARMII::MO_NONLAZY);
2501 SDValue Result = DAG.getNode(Wrapper, dl, PtrVT, G);
2503 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2504 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
2505 MachinePointerInfo::getGOT(), false, false, false, 0);
2509 SDValue ARMTargetLowering::LowerGlobalAddressWindows(SDValue Op,
2510 SelectionDAG &DAG) const {
2511 assert(Subtarget->isTargetWindows() && "non-Windows COFF is not supported");
2512 assert(Subtarget->useMovt(DAG.getMachineFunction()) &&
2513 "Windows on ARM expects to use movw/movt");
2515 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2516 const ARMII::TOF TargetFlags =
2517 (GV->hasDLLImportStorageClass() ? ARMII::MO_DLLIMPORT : ARMII::MO_NO_FLAG);
2518 EVT PtrVT = getPointerTy();
2524 // FIXME: Once remat is capable of dealing with instructions with register
2525 // operands, expand this into two nodes.
2526 Result = DAG.getNode(ARMISD::Wrapper, DL, PtrVT,
2527 DAG.getTargetGlobalAddress(GV, DL, PtrVT, /*Offset=*/0,
2529 if (GV->hasDLLImportStorageClass())
2530 Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), Result,
2531 MachinePointerInfo::getGOT(), false, false, false, 0);
2535 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
2536 SelectionDAG &DAG) const {
2537 assert(Subtarget->isTargetELF() &&
2538 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
2539 MachineFunction &MF = DAG.getMachineFunction();
2540 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2541 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2542 EVT PtrVT = getPointerTy();
2544 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2545 ARMConstantPoolValue *CPV =
2546 ARMConstantPoolSymbol::Create(*DAG.getContext(), "_GLOBAL_OFFSET_TABLE_",
2547 ARMPCLabelIndex, PCAdj);
2548 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2549 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2550 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2551 MachinePointerInfo::getConstantPool(),
2552 false, false, false, 0);
2553 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2554 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2558 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
2560 SDValue Val = DAG.getConstant(0, MVT::i32);
2561 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl,
2562 DAG.getVTList(MVT::i32, MVT::Other), Op.getOperand(0),
2563 Op.getOperand(1), Val);
2567 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
2569 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
2570 Op.getOperand(1), DAG.getConstant(0, MVT::i32));
2574 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
2575 const ARMSubtarget *Subtarget) const {
2576 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2579 default: return SDValue(); // Don't custom lower most intrinsics.
2580 case Intrinsic::arm_rbit: {
2581 assert(Op.getOperand(0).getValueType() == MVT::i32 &&
2582 "RBIT intrinsic must have i32 type!");
2583 return DAG.getNode(ARMISD::RBIT, dl, MVT::i32, Op.getOperand(0));
2585 case Intrinsic::arm_thread_pointer: {
2586 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2587 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2589 case Intrinsic::eh_sjlj_lsda: {
2590 MachineFunction &MF = DAG.getMachineFunction();
2591 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2592 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2593 EVT PtrVT = getPointerTy();
2594 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2596 unsigned PCAdj = (RelocM != Reloc::PIC_)
2597 ? 0 : (Subtarget->isThumb() ? 4 : 8);
2598 ARMConstantPoolValue *CPV =
2599 ARMConstantPoolConstant::Create(MF.getFunction(), ARMPCLabelIndex,
2600 ARMCP::CPLSDA, PCAdj);
2601 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2602 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2604 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2605 MachinePointerInfo::getConstantPool(),
2606 false, false, false, 0);
2608 if (RelocM == Reloc::PIC_) {
2609 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2610 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2614 case Intrinsic::arm_neon_vmulls:
2615 case Intrinsic::arm_neon_vmullu: {
2616 unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vmulls)
2617 ? ARMISD::VMULLs : ARMISD::VMULLu;
2618 return DAG.getNode(NewOpc, SDLoc(Op), Op.getValueType(),
2619 Op.getOperand(1), Op.getOperand(2));
2624 static SDValue LowerATOMIC_FENCE(SDValue Op, SelectionDAG &DAG,
2625 const ARMSubtarget *Subtarget) {
2626 // FIXME: handle "fence singlethread" more efficiently.
2628 if (!Subtarget->hasDataBarrier()) {
2629 // Some ARMv6 cpus can support data barriers with an mcr instruction.
2630 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2632 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2633 "Unexpected ISD::ATOMIC_FENCE encountered. Should be libcall!");
2634 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2635 DAG.getConstant(0, MVT::i32));
2638 ConstantSDNode *OrdN = cast<ConstantSDNode>(Op.getOperand(1));
2639 AtomicOrdering Ord = static_cast<AtomicOrdering>(OrdN->getZExtValue());
2640 unsigned Domain = ARM_MB::ISH;
2641 if (Subtarget->isMClass()) {
2642 // Only a full system barrier exists in the M-class architectures.
2643 Domain = ARM_MB::SY;
2644 } else if (Subtarget->isSwift() && Ord == Release) {
2645 // Swift happens to implement ISHST barriers in a way that's compatible with
2646 // Release semantics but weaker than ISH so we'd be fools not to use
2647 // it. Beware: other processors probably don't!
2648 Domain = ARM_MB::ISHST;
2651 return DAG.getNode(ISD::INTRINSIC_VOID, dl, MVT::Other, Op.getOperand(0),
2652 DAG.getConstant(Intrinsic::arm_dmb, MVT::i32),
2653 DAG.getConstant(Domain, MVT::i32));
2656 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG,
2657 const ARMSubtarget *Subtarget) {
2658 // ARM pre v5TE and Thumb1 does not have preload instructions.
2659 if (!(Subtarget->isThumb2() ||
2660 (!Subtarget->isThumb1Only() && Subtarget->hasV5TEOps())))
2661 // Just preserve the chain.
2662 return Op.getOperand(0);
2665 unsigned isRead = ~cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() & 1;
2667 (!Subtarget->hasV7Ops() || !Subtarget->hasMPExtension()))
2668 // ARMv7 with MP extension has PLDW.
2669 return Op.getOperand(0);
2671 unsigned isData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
2672 if (Subtarget->isThumb()) {
2674 isRead = ~isRead & 1;
2675 isData = ~isData & 1;
2678 return DAG.getNode(ARMISD::PRELOAD, dl, MVT::Other, Op.getOperand(0),
2679 Op.getOperand(1), DAG.getConstant(isRead, MVT::i32),
2680 DAG.getConstant(isData, MVT::i32));
2683 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
2684 MachineFunction &MF = DAG.getMachineFunction();
2685 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
2687 // vastart just stores the address of the VarArgsFrameIndex slot into the
2688 // memory location argument.
2690 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2691 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2692 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
2693 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
2694 MachinePointerInfo(SV), false, false, 0);
2698 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
2699 SDValue &Root, SelectionDAG &DAG,
2701 MachineFunction &MF = DAG.getMachineFunction();
2702 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2704 const TargetRegisterClass *RC;
2705 if (AFI->isThumb1OnlyFunction())
2706 RC = &ARM::tGPRRegClass;
2708 RC = &ARM::GPRRegClass;
2710 // Transform the arguments stored in physical registers into virtual ones.
2711 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2712 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2715 if (NextVA.isMemLoc()) {
2716 MachineFrameInfo *MFI = MF.getFrameInfo();
2717 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true);
2719 // Create load node to retrieve arguments from the stack.
2720 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2721 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
2722 MachinePointerInfo::getFixedStack(FI),
2723 false, false, false, 0);
2725 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
2726 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2728 if (!Subtarget->isLittle())
2729 std::swap (ArgValue, ArgValue2);
2730 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
2734 ARMTargetLowering::computeRegArea(CCState &CCInfo, MachineFunction &MF,
2735 unsigned InRegsParamRecordIdx,
2737 unsigned &ArgRegsSize,
2738 unsigned &ArgRegsSaveSize)
2741 if (InRegsParamRecordIdx < CCInfo.getInRegsParamsCount()) {
2742 unsigned RBegin, REnd;
2743 CCInfo.getInRegsParamInfo(InRegsParamRecordIdx, RBegin, REnd);
2744 NumGPRs = REnd - RBegin;
2746 unsigned int firstUnalloced;
2747 firstUnalloced = CCInfo.getFirstUnallocated(GPRArgRegs,
2748 sizeof(GPRArgRegs) /
2749 sizeof(GPRArgRegs[0]));
2750 NumGPRs = (firstUnalloced <= 3) ? (4 - firstUnalloced) : 0;
2753 unsigned Align = MF.getTarget().getFrameLowering()->getStackAlignment();
2754 ArgRegsSize = NumGPRs * 4;
2756 // If parameter is split between stack and GPRs...
2757 if (NumGPRs && Align > 4 &&
2758 (ArgRegsSize < ArgSize ||
2759 InRegsParamRecordIdx >= CCInfo.getInRegsParamsCount())) {
2760 // Add padding for part of param recovered from GPRs. For example,
2761 // if Align == 8, its last byte must be at address K*8 - 1.
2762 // We need to do it, since remained (stack) part of parameter has
2763 // stack alignment, and we need to "attach" "GPRs head" without gaps
2766 // |---- 8 bytes block ----| |---- 8 bytes block ----| |---- 8 bytes...
2767 // [ [padding] [GPRs head] ] [ Tail passed via stack ....
2769 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2771 OffsetToAlignment(ArgRegsSize + AFI->getArgRegsSaveSize(), Align);
2772 ArgRegsSaveSize = ArgRegsSize + Padding;
2774 // We don't need to extend regs save size for byval parameters if they
2775 // are passed via GPRs only.
2776 ArgRegsSaveSize = ArgRegsSize;
2779 // The remaining GPRs hold either the beginning of variable-argument
2780 // data, or the beginning of an aggregate passed by value (usually
2781 // byval). Either way, we allocate stack slots adjacent to the data
2782 // provided by our caller, and store the unallocated registers there.
2783 // If this is a variadic function, the va_list pointer will begin with
2784 // these values; otherwise, this reassembles a (byval) structure that
2785 // was split between registers and memory.
2786 // Return: The frame index registers were stored into.
2788 ARMTargetLowering::StoreByValRegs(CCState &CCInfo, SelectionDAG &DAG,
2789 SDLoc dl, SDValue &Chain,
2790 const Value *OrigArg,
2791 unsigned InRegsParamRecordIdx,
2792 unsigned OffsetFromOrigArg,
2796 unsigned ByValStoreOffset,
2797 unsigned TotalArgRegsSaveSize) const {
2799 // Currently, two use-cases possible:
2800 // Case #1. Non-var-args function, and we meet first byval parameter.
2801 // Setup first unallocated register as first byval register;
2802 // eat all remained registers
2803 // (these two actions are performed by HandleByVal method).
2804 // Then, here, we initialize stack frame with
2805 // "store-reg" instructions.
2806 // Case #2. Var-args function, that doesn't contain byval parameters.
2807 // The same: eat all remained unallocated registers,
2808 // initialize stack frame.
2810 MachineFunction &MF = DAG.getMachineFunction();
2811 MachineFrameInfo *MFI = MF.getFrameInfo();
2812 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2813 unsigned firstRegToSaveIndex, lastRegToSaveIndex;
2814 unsigned RBegin, REnd;
2815 if (InRegsParamRecordIdx < CCInfo.getInRegsParamsCount()) {
2816 CCInfo.getInRegsParamInfo(InRegsParamRecordIdx, RBegin, REnd);
2817 firstRegToSaveIndex = RBegin - ARM::R0;
2818 lastRegToSaveIndex = REnd - ARM::R0;
2820 firstRegToSaveIndex = CCInfo.getFirstUnallocated
2821 (GPRArgRegs, array_lengthof(GPRArgRegs));
2822 lastRegToSaveIndex = 4;
2825 unsigned ArgRegsSize, ArgRegsSaveSize;
2826 computeRegArea(CCInfo, MF, InRegsParamRecordIdx, ArgSize,
2827 ArgRegsSize, ArgRegsSaveSize);
2829 // Store any by-val regs to their spots on the stack so that they may be
2830 // loaded by deferencing the result of formal parameter pointer or va_next.
2831 // Note: once stack area for byval/varargs registers
2832 // was initialized, it can't be initialized again.
2833 if (ArgRegsSaveSize) {
2834 unsigned Padding = ArgRegsSaveSize - ArgRegsSize;
2837 assert(AFI->getStoredByValParamsPadding() == 0 &&
2838 "The only parameter may be padded.");
2839 AFI->setStoredByValParamsPadding(Padding);
2842 int FrameIndex = MFI->CreateFixedObject(ArgRegsSaveSize,
2845 (int64_t)TotalArgRegsSaveSize,
2847 SDValue FIN = DAG.getFrameIndex(FrameIndex, getPointerTy());
2849 MFI->CreateFixedObject(Padding,
2850 ArgOffset + ByValStoreOffset -
2851 (int64_t)ArgRegsSaveSize,
2855 SmallVector<SDValue, 4> MemOps;
2856 for (unsigned i = 0; firstRegToSaveIndex < lastRegToSaveIndex;
2857 ++firstRegToSaveIndex, ++i) {
2858 const TargetRegisterClass *RC;
2859 if (AFI->isThumb1OnlyFunction())
2860 RC = &ARM::tGPRRegClass;
2862 RC = &ARM::GPRRegClass;
2864 unsigned VReg = MF.addLiveIn(GPRArgRegs[firstRegToSaveIndex], RC);
2865 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2867 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2868 MachinePointerInfo(OrigArg, OffsetFromOrigArg + 4*i),
2870 MemOps.push_back(Store);
2871 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
2872 DAG.getConstant(4, getPointerTy()));
2875 AFI->setArgRegsSaveSize(ArgRegsSaveSize + AFI->getArgRegsSaveSize());
2877 if (!MemOps.empty())
2878 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOps);
2882 // We cannot allocate a zero-byte object for the first variadic argument,
2883 // so just make up a size.
2886 // This will point to the next argument passed via stack.
2887 return MFI->CreateFixedObject(
2888 ArgSize, ArgOffset, !ForceMutable);
2892 // Setup stack frame, the va_list pointer will start from.
2894 ARMTargetLowering::VarArgStyleRegisters(CCState &CCInfo, SelectionDAG &DAG,
2895 SDLoc dl, SDValue &Chain,
2897 unsigned TotalArgRegsSaveSize,
2898 bool ForceMutable) const {
2899 MachineFunction &MF = DAG.getMachineFunction();
2900 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2902 // Try to store any remaining integer argument regs
2903 // to their spots on the stack so that they may be loaded by deferencing
2904 // the result of va_next.
2905 // If there is no regs to be stored, just point address after last
2906 // argument passed via stack.
2908 StoreByValRegs(CCInfo, DAG, dl, Chain, nullptr,
2909 CCInfo.getInRegsParamsCount(), 0, ArgOffset, 0, ForceMutable,
2910 0, TotalArgRegsSaveSize);
2912 AFI->setVarArgsFrameIndex(FrameIndex);
2916 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
2917 CallingConv::ID CallConv, bool isVarArg,
2918 const SmallVectorImpl<ISD::InputArg>
2920 SDLoc dl, SelectionDAG &DAG,
2921 SmallVectorImpl<SDValue> &InVals)
2923 MachineFunction &MF = DAG.getMachineFunction();
2924 MachineFrameInfo *MFI = MF.getFrameInfo();
2926 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2928 // Assign locations to all of the incoming arguments.
2929 SmallVector<CCValAssign, 16> ArgLocs;
2930 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2931 getTargetMachine(), ArgLocs, *DAG.getContext(), Prologue);
2932 CCInfo.AnalyzeFormalArguments(Ins,
2933 CCAssignFnForNode(CallConv, /* Return*/ false,
2936 SmallVector<SDValue, 16> ArgValues;
2937 int lastInsIndex = -1;
2939 Function::const_arg_iterator CurOrigArg = MF.getFunction()->arg_begin();
2940 unsigned CurArgIdx = 0;
2942 // Initially ArgRegsSaveSize is zero.
2943 // Then we increase this value each time we meet byval parameter.
2944 // We also increase this value in case of varargs function.
2945 AFI->setArgRegsSaveSize(0);
2947 unsigned ByValStoreOffset = 0;
2948 unsigned TotalArgRegsSaveSize = 0;
2949 unsigned ArgRegsSaveSizeMaxAlign = 4;
2951 // Calculate the amount of stack space that we need to allocate to store
2952 // byval and variadic arguments that are passed in registers.
2953 // We need to know this before we allocate the first byval or variadic
2954 // argument, as they will be allocated a stack slot below the CFA (Canonical
2955 // Frame Address, the stack pointer at entry to the function).
2956 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2957 CCValAssign &VA = ArgLocs[i];
2958 if (VA.isMemLoc()) {
2959 int index = VA.getValNo();
2960 if (index != lastInsIndex) {
2961 ISD::ArgFlagsTy Flags = Ins[index].Flags;
2962 if (Flags.isByVal()) {
2963 unsigned ExtraArgRegsSize;
2964 unsigned ExtraArgRegsSaveSize;
2965 computeRegArea(CCInfo, MF, CCInfo.getInRegsParamsProceed(),
2966 Flags.getByValSize(),
2967 ExtraArgRegsSize, ExtraArgRegsSaveSize);
2969 TotalArgRegsSaveSize += ExtraArgRegsSaveSize;
2970 if (Flags.getByValAlign() > ArgRegsSaveSizeMaxAlign)
2971 ArgRegsSaveSizeMaxAlign = Flags.getByValAlign();
2972 CCInfo.nextInRegsParam();
2974 lastInsIndex = index;
2978 CCInfo.rewindByValRegsInfo();
2981 unsigned ExtraArgRegsSize;
2982 unsigned ExtraArgRegsSaveSize;
2983 computeRegArea(CCInfo, MF, CCInfo.getInRegsParamsCount(), 0,
2984 ExtraArgRegsSize, ExtraArgRegsSaveSize);
2985 TotalArgRegsSaveSize += ExtraArgRegsSaveSize;
2987 // If the arg regs save area contains N-byte aligned values, the
2988 // bottom of it must be at least N-byte aligned.
2989 TotalArgRegsSaveSize = RoundUpToAlignment(TotalArgRegsSaveSize, ArgRegsSaveSizeMaxAlign);
2990 TotalArgRegsSaveSize = std::min(TotalArgRegsSaveSize, 16U);
2992 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2993 CCValAssign &VA = ArgLocs[i];
2994 std::advance(CurOrigArg, Ins[VA.getValNo()].OrigArgIndex - CurArgIdx);
2995 CurArgIdx = Ins[VA.getValNo()].OrigArgIndex;
2996 // Arguments stored in registers.
2997 if (VA.isRegLoc()) {
2998 EVT RegVT = VA.getLocVT();
3000 if (VA.needsCustom()) {
3001 // f64 and vector types are split up into multiple registers or
3002 // combinations of registers and stack slots.
3003 if (VA.getLocVT() == MVT::v2f64) {
3004 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
3006 VA = ArgLocs[++i]; // skip ahead to next loc
3008 if (VA.isMemLoc()) {
3009 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true);
3010 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
3011 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
3012 MachinePointerInfo::getFixedStack(FI),
3013 false, false, false, 0);
3015 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
3018 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
3019 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
3020 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
3021 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
3022 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
3024 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
3027 const TargetRegisterClass *RC;
3029 if (RegVT == MVT::f32)
3030 RC = &ARM::SPRRegClass;
3031 else if (RegVT == MVT::f64)
3032 RC = &ARM::DPRRegClass;
3033 else if (RegVT == MVT::v2f64)
3034 RC = &ARM::QPRRegClass;
3035 else if (RegVT == MVT::i32)
3036 RC = AFI->isThumb1OnlyFunction() ?
3037 (const TargetRegisterClass*)&ARM::tGPRRegClass :
3038 (const TargetRegisterClass*)&ARM::GPRRegClass;
3040 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
3042 // Transform the arguments in physical registers into virtual ones.
3043 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
3044 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
3047 // If this is an 8 or 16-bit value, it is really passed promoted
3048 // to 32 bits. Insert an assert[sz]ext to capture this, then
3049 // truncate to the right size.
3050 switch (VA.getLocInfo()) {
3051 default: llvm_unreachable("Unknown loc info!");
3052 case CCValAssign::Full: break;
3053 case CCValAssign::BCvt:
3054 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
3056 case CCValAssign::SExt:
3057 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
3058 DAG.getValueType(VA.getValVT()));
3059 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
3061 case CCValAssign::ZExt:
3062 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
3063 DAG.getValueType(VA.getValVT()));
3064 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
3068 InVals.push_back(ArgValue);
3070 } else { // VA.isRegLoc()
3073 assert(VA.isMemLoc());
3074 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
3076 int index = ArgLocs[i].getValNo();
3078 // Some Ins[] entries become multiple ArgLoc[] entries.
3079 // Process them only once.
3080 if (index != lastInsIndex)
3082 ISD::ArgFlagsTy Flags = Ins[index].Flags;
3083 // FIXME: For now, all byval parameter objects are marked mutable.
3084 // This can be changed with more analysis.
3085 // In case of tail call optimization mark all arguments mutable.
3086 // Since they could be overwritten by lowering of arguments in case of
3088 if (Flags.isByVal()) {
3089 unsigned CurByValIndex = CCInfo.getInRegsParamsProceed();
3091 ByValStoreOffset = RoundUpToAlignment(ByValStoreOffset, Flags.getByValAlign());
3092 int FrameIndex = StoreByValRegs(
3093 CCInfo, DAG, dl, Chain, CurOrigArg,
3095 Ins[VA.getValNo()].PartOffset,
3096 VA.getLocMemOffset(),
3097 Flags.getByValSize(),
3098 true /*force mutable frames*/,
3100 TotalArgRegsSaveSize);
3101 ByValStoreOffset += Flags.getByValSize();
3102 ByValStoreOffset = std::min(ByValStoreOffset, 16U);
3103 InVals.push_back(DAG.getFrameIndex(FrameIndex, getPointerTy()));
3104 CCInfo.nextInRegsParam();
3106 unsigned FIOffset = VA.getLocMemOffset();
3107 int FI = MFI->CreateFixedObject(VA.getLocVT().getSizeInBits()/8,
3110 // Create load nodes to retrieve arguments from the stack.
3111 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
3112 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
3113 MachinePointerInfo::getFixedStack(FI),
3114 false, false, false, 0));
3116 lastInsIndex = index;
3123 VarArgStyleRegisters(CCInfo, DAG, dl, Chain,
3124 CCInfo.getNextStackOffset(),
3125 TotalArgRegsSaveSize);
3127 AFI->setArgumentStackSize(CCInfo.getNextStackOffset());
3132 /// isFloatingPointZero - Return true if this is +0.0.
3133 static bool isFloatingPointZero(SDValue Op) {
3134 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
3135 return CFP->getValueAPF().isPosZero();
3136 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
3137 // Maybe this has already been legalized into the constant pool?
3138 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
3139 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
3140 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
3141 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
3142 return CFP->getValueAPF().isPosZero();
3148 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
3149 /// the given operands.
3151 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
3152 SDValue &ARMcc, SelectionDAG &DAG,
3154 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
3155 unsigned C = RHSC->getZExtValue();
3156 if (!isLegalICmpImmediate(C)) {
3157 // Constant does not fit, try adjusting it by one?
3162 if (C != 0x80000000 && isLegalICmpImmediate(C-1)) {
3163 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
3164 RHS = DAG.getConstant(C-1, MVT::i32);
3169 if (C != 0 && isLegalICmpImmediate(C-1)) {
3170 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
3171 RHS = DAG.getConstant(C-1, MVT::i32);
3176 if (C != 0x7fffffff && isLegalICmpImmediate(C+1)) {
3177 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
3178 RHS = DAG.getConstant(C+1, MVT::i32);
3183 if (C != 0xffffffff && isLegalICmpImmediate(C+1)) {
3184 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
3185 RHS = DAG.getConstant(C+1, MVT::i32);
3192 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3193 ARMISD::NodeType CompareType;
3196 CompareType = ARMISD::CMP;
3201 CompareType = ARMISD::CMPZ;
3204 ARMcc = DAG.getConstant(CondCode, MVT::i32);
3205 return DAG.getNode(CompareType, dl, MVT::Glue, LHS, RHS);
3208 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
3210 ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
3213 if (!isFloatingPointZero(RHS))
3214 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Glue, LHS, RHS);
3216 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Glue, LHS);
3217 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Glue, Cmp);
3220 /// duplicateCmp - Glue values can have only one use, so this function
3221 /// duplicates a comparison node.
3223 ARMTargetLowering::duplicateCmp(SDValue Cmp, SelectionDAG &DAG) const {
3224 unsigned Opc = Cmp.getOpcode();
3226 if (Opc == ARMISD::CMP || Opc == ARMISD::CMPZ)
3227 return DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
3229 assert(Opc == ARMISD::FMSTAT && "unexpected comparison operation");
3230 Cmp = Cmp.getOperand(0);
3231 Opc = Cmp.getOpcode();
3232 if (Opc == ARMISD::CMPFP)
3233 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
3235 assert(Opc == ARMISD::CMPFPw0 && "unexpected operand of FMSTAT");
3236 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0));
3238 return DAG.getNode(ARMISD::FMSTAT, DL, MVT::Glue, Cmp);
3241 std::pair<SDValue, SDValue>
3242 ARMTargetLowering::getARMXALUOOp(SDValue Op, SelectionDAG &DAG,
3243 SDValue &ARMcc) const {
3244 assert(Op.getValueType() == MVT::i32 && "Unsupported value type");
3246 SDValue Value, OverflowCmp;
3247 SDValue LHS = Op.getOperand(0);
3248 SDValue RHS = Op.getOperand(1);
3251 // FIXME: We are currently always generating CMPs because we don't support
3252 // generating CMN through the backend. This is not as good as the natural
3253 // CMP case because it causes a register dependency and cannot be folded
3256 switch (Op.getOpcode()) {
3258 llvm_unreachable("Unknown overflow instruction!");
3260 ARMcc = DAG.getConstant(ARMCC::VC, MVT::i32);
3261 Value = DAG.getNode(ISD::ADD, SDLoc(Op), Op.getValueType(), LHS, RHS);
3262 OverflowCmp = DAG.getNode(ARMISD::CMP, SDLoc(Op), MVT::Glue, Value, LHS);
3265 ARMcc = DAG.getConstant(ARMCC::HS, MVT::i32);
3266 Value = DAG.getNode(ISD::ADD, SDLoc(Op), Op.getValueType(), LHS, RHS);
3267 OverflowCmp = DAG.getNode(ARMISD::CMP, SDLoc(Op), MVT::Glue, Value, LHS);
3270 ARMcc = DAG.getConstant(ARMCC::VC, MVT::i32);
3271 Value = DAG.getNode(ISD::SUB, SDLoc(Op), Op.getValueType(), LHS, RHS);
3272 OverflowCmp = DAG.getNode(ARMISD::CMP, SDLoc(Op), MVT::Glue, LHS, RHS);
3275 ARMcc = DAG.getConstant(ARMCC::HS, MVT::i32);
3276 Value = DAG.getNode(ISD::SUB, SDLoc(Op), Op.getValueType(), LHS, RHS);
3277 OverflowCmp = DAG.getNode(ARMISD::CMP, SDLoc(Op), MVT::Glue, LHS, RHS);
3281 return std::make_pair(Value, OverflowCmp);
3286 ARMTargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) const {
3287 // Let legalize expand this if it isn't a legal type yet.
3288 if (!DAG.getTargetLoweringInfo().isTypeLegal(Op.getValueType()))
3291 SDValue Value, OverflowCmp;
3293 std::tie(Value, OverflowCmp) = getARMXALUOOp(Op, DAG, ARMcc);
3294 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3295 // We use 0 and 1 as false and true values.
3296 SDValue TVal = DAG.getConstant(1, MVT::i32);
3297 SDValue FVal = DAG.getConstant(0, MVT::i32);
3298 EVT VT = Op.getValueType();
3300 SDValue Overflow = DAG.getNode(ARMISD::CMOV, SDLoc(Op), VT, TVal, FVal,
3301 ARMcc, CCR, OverflowCmp);
3303 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
3304 return DAG.getNode(ISD::MERGE_VALUES, SDLoc(Op), VTs, Value, Overflow);
3308 SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
3309 SDValue Cond = Op.getOperand(0);
3310 SDValue SelectTrue = Op.getOperand(1);
3311 SDValue SelectFalse = Op.getOperand(2);
3313 unsigned Opc = Cond.getOpcode();
3315 if (Cond.getResNo() == 1 &&
3316 (Opc == ISD::SADDO || Opc == ISD::UADDO || Opc == ISD::SSUBO ||
3317 Opc == ISD::USUBO)) {
3318 if (!DAG.getTargetLoweringInfo().isTypeLegal(Cond->getValueType(0)))
3321 SDValue Value, OverflowCmp;
3323 std::tie(Value, OverflowCmp) = getARMXALUOOp(Cond, DAG, ARMcc);
3324 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3325 EVT VT = Op.getValueType();
3327 return DAG.getNode(ARMISD::CMOV, SDLoc(Op), VT, SelectTrue, SelectFalse,
3328 ARMcc, CCR, OverflowCmp);
3334 // (select (cmov 1, 0, cond), t, f) -> (cmov t, f, cond)
3335 // (select (cmov 0, 1, cond), t, f) -> (cmov f, t, cond)
3337 if (Cond.getOpcode() == ARMISD::CMOV && Cond.hasOneUse()) {
3338 const ConstantSDNode *CMOVTrue =
3339 dyn_cast<ConstantSDNode>(Cond.getOperand(0));
3340 const ConstantSDNode *CMOVFalse =
3341 dyn_cast<ConstantSDNode>(Cond.getOperand(1));
3343 if (CMOVTrue && CMOVFalse) {
3344 unsigned CMOVTrueVal = CMOVTrue->getZExtValue();
3345 unsigned CMOVFalseVal = CMOVFalse->getZExtValue();
3349 if (CMOVTrueVal == 1 && CMOVFalseVal == 0) {
3351 False = SelectFalse;
3352 } else if (CMOVTrueVal == 0 && CMOVFalseVal == 1) {
3357 if (True.getNode() && False.getNode()) {
3358 EVT VT = Op.getValueType();
3359 SDValue ARMcc = Cond.getOperand(2);
3360 SDValue CCR = Cond.getOperand(3);
3361 SDValue Cmp = duplicateCmp(Cond.getOperand(4), DAG);
3362 assert(True.getValueType() == VT);
3363 return DAG.getNode(ARMISD::CMOV, dl, VT, True, False, ARMcc, CCR, Cmp);
3368 // ARM's BooleanContents value is UndefinedBooleanContent. Mask out the
3369 // undefined bits before doing a full-word comparison with zero.
3370 Cond = DAG.getNode(ISD::AND, dl, Cond.getValueType(), Cond,
3371 DAG.getConstant(1, Cond.getValueType()));
3373 return DAG.getSelectCC(dl, Cond,
3374 DAG.getConstant(0, Cond.getValueType()),
3375 SelectTrue, SelectFalse, ISD::SETNE);
3378 static ISD::CondCode getInverseCCForVSEL(ISD::CondCode CC) {
3379 if (CC == ISD::SETNE)
3381 return ISD::getSetCCInverse(CC, true);
3384 static void checkVSELConstraints(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
3385 bool &swpCmpOps, bool &swpVselOps) {
3386 // Start by selecting the GE condition code for opcodes that return true for
3388 if (CC == ISD::SETUGE || CC == ISD::SETOGE || CC == ISD::SETOLE ||
3390 CondCode = ARMCC::GE;
3392 // and GT for opcodes that return false for 'equality'.
3393 else if (CC == ISD::SETUGT || CC == ISD::SETOGT || CC == ISD::SETOLT ||
3395 CondCode = ARMCC::GT;
3397 // Since we are constrained to GE/GT, if the opcode contains 'less', we need
3398 // to swap the compare operands.
3399 if (CC == ISD::SETOLE || CC == ISD::SETULE || CC == ISD::SETOLT ||
3403 // Both GT and GE are ordered comparisons, and return false for 'unordered'.
3404 // If we have an unordered opcode, we need to swap the operands to the VSEL
3405 // instruction (effectively negating the condition).
3407 // This also has the effect of swapping which one of 'less' or 'greater'
3408 // returns true, so we also swap the compare operands. It also switches
3409 // whether we return true for 'equality', so we compensate by picking the
3410 // opposite condition code to our original choice.
3411 if (CC == ISD::SETULE || CC == ISD::SETULT || CC == ISD::SETUGE ||
3412 CC == ISD::SETUGT) {
3413 swpCmpOps = !swpCmpOps;
3414 swpVselOps = !swpVselOps;
3415 CondCode = CondCode == ARMCC::GT ? ARMCC::GE : ARMCC::GT;
3418 // 'ordered' is 'anything but unordered', so use the VS condition code and
3419 // swap the VSEL operands.
3420 if (CC == ISD::SETO) {
3421 CondCode = ARMCC::VS;
3425 // 'unordered or not equal' is 'anything but equal', so use the EQ condition
3426 // code and swap the VSEL operands.
3427 if (CC == ISD::SETUNE) {
3428 CondCode = ARMCC::EQ;
3433 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
3434 EVT VT = Op.getValueType();
3435 SDValue LHS = Op.getOperand(0);
3436 SDValue RHS = Op.getOperand(1);
3437 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
3438 SDValue TrueVal = Op.getOperand(2);
3439 SDValue FalseVal = Op.getOperand(3);
3442 if (LHS.getValueType() == MVT::i32) {
3443 // Try to generate VSEL on ARMv8.
3444 // The VSEL instruction can't use all the usual ARM condition
3445 // codes: it only has two bits to select the condition code, so it's
3446 // constrained to use only GE, GT, VS and EQ.
3448 // To implement all the various ISD::SETXXX opcodes, we sometimes need to
3449 // swap the operands of the previous compare instruction (effectively
3450 // inverting the compare condition, swapping 'less' and 'greater') and
3451 // sometimes need to swap the operands to the VSEL (which inverts the
3452 // condition in the sense of firing whenever the previous condition didn't)
3453 if (getSubtarget()->hasFPARMv8() && (TrueVal.getValueType() == MVT::f32 ||
3454 TrueVal.getValueType() == MVT::f64)) {
3455 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3456 if (CondCode == ARMCC::LT || CondCode == ARMCC::LE ||
3457 CondCode == ARMCC::VC || CondCode == ARMCC::NE) {
3458 CC = getInverseCCForVSEL(CC);
3459 std::swap(TrueVal, FalseVal);
3464 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3465 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3466 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,
3470 ARMCC::CondCodes CondCode, CondCode2;
3471 FPCCToARMCC(CC, CondCode, CondCode2);
3473 // Try to generate VSEL on ARMv8.
3474 if (getSubtarget()->hasFPARMv8() && (TrueVal.getValueType() == MVT::f32 ||
3475 TrueVal.getValueType() == MVT::f64)) {
3476 // We can select VMAXNM/VMINNM from a compare followed by a select with the
3477 // same operands, as follows:
3478 // c = fcmp [ogt, olt, ugt, ult] a, b
3480 // We only do this in unsafe-fp-math, because signed zeros and NaNs are
3481 // handled differently than the original code sequence.
3482 if (getTargetMachine().Options.UnsafeFPMath && LHS == TrueVal &&
3484 if (CC == ISD::SETOGT || CC == ISD::SETUGT)
3485 return DAG.getNode(ARMISD::VMAXNM, dl, VT, TrueVal, FalseVal);
3486 if (CC == ISD::SETOLT || CC == ISD::SETULT)
3487 return DAG.getNode(ARMISD::VMINNM, dl, VT, TrueVal, FalseVal);
3490 bool swpCmpOps = false;
3491 bool swpVselOps = false;
3492 checkVSELConstraints(CC, CondCode, swpCmpOps, swpVselOps);
3494 if (CondCode == ARMCC::GT || CondCode == ARMCC::GE ||
3495 CondCode == ARMCC::VS || CondCode == ARMCC::EQ) {
3497 std::swap(LHS, RHS);
3499 std::swap(TrueVal, FalseVal);
3503 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
3504 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
3505 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3506 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
3508 if (CondCode2 != ARMCC::AL) {
3509 SDValue ARMcc2 = DAG.getConstant(CondCode2, MVT::i32);
3510 // FIXME: Needs another CMP because flag can have but one use.
3511 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
3512 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
3513 Result, TrueVal, ARMcc2, CCR, Cmp2);
3518 /// canChangeToInt - Given the fp compare operand, return true if it is suitable
3519 /// to morph to an integer compare sequence.
3520 static bool canChangeToInt(SDValue Op, bool &SeenZero,
3521 const ARMSubtarget *Subtarget) {
3522 SDNode *N = Op.getNode();
3523 if (!N->hasOneUse())
3524 // Otherwise it requires moving the value from fp to integer registers.
3526 if (!N->getNumValues())
3528 EVT VT = Op.getValueType();
3529 if (VT != MVT::f32 && !Subtarget->isFPBrccSlow())
3530 // f32 case is generally profitable. f64 case only makes sense when vcmpe +
3531 // vmrs are very slow, e.g. cortex-a8.
3534 if (isFloatingPointZero(Op)) {
3538 return ISD::isNormalLoad(N);
3541 static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) {
3542 if (isFloatingPointZero(Op))
3543 return DAG.getConstant(0, MVT::i32);
3545 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op))
3546 return DAG.getLoad(MVT::i32, SDLoc(Op),
3547 Ld->getChain(), Ld->getBasePtr(), Ld->getPointerInfo(),
3548 Ld->isVolatile(), Ld->isNonTemporal(),
3549 Ld->isInvariant(), Ld->getAlignment());
3551 llvm_unreachable("Unknown VFP cmp argument!");
3554 static void expandf64Toi32(SDValue Op, SelectionDAG &DAG,
3555 SDValue &RetVal1, SDValue &RetVal2) {
3556 if (isFloatingPointZero(Op)) {
3557 RetVal1 = DAG.getConstant(0, MVT::i32);
3558 RetVal2 = DAG.getConstant(0, MVT::i32);
3562 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) {
3563 SDValue Ptr = Ld->getBasePtr();
3564 RetVal1 = DAG.getLoad(MVT::i32, SDLoc(Op),
3565 Ld->getChain(), Ptr,
3566 Ld->getPointerInfo(),
3567 Ld->isVolatile(), Ld->isNonTemporal(),
3568 Ld->isInvariant(), Ld->getAlignment());
3570 EVT PtrType = Ptr.getValueType();
3571 unsigned NewAlign = MinAlign(Ld->getAlignment(), 4);
3572 SDValue NewPtr = DAG.getNode(ISD::ADD, SDLoc(Op),
3573 PtrType, Ptr, DAG.getConstant(4, PtrType));
3574 RetVal2 = DAG.getLoad(MVT::i32, SDLoc(Op),
3575 Ld->getChain(), NewPtr,
3576 Ld->getPointerInfo().getWithOffset(4),
3577 Ld->isVolatile(), Ld->isNonTemporal(),
3578 Ld->isInvariant(), NewAlign);
3582 llvm_unreachable("Unknown VFP cmp argument!");
3585 /// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some
3586 /// f32 and even f64 comparisons to integer ones.
3588 ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const {
3589 SDValue Chain = Op.getOperand(0);
3590 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3591 SDValue LHS = Op.getOperand(2);
3592 SDValue RHS = Op.getOperand(3);
3593 SDValue Dest = Op.getOperand(4);
3596 bool LHSSeenZero = false;
3597 bool LHSOk = canChangeToInt(LHS, LHSSeenZero, Subtarget);
3598 bool RHSSeenZero = false;
3599 bool RHSOk = canChangeToInt(RHS, RHSSeenZero, Subtarget);
3600 if (LHSOk && RHSOk && (LHSSeenZero || RHSSeenZero)) {
3601 // If unsafe fp math optimization is enabled and there are no other uses of
3602 // the CMP operands, and the condition code is EQ or NE, we can optimize it
3603 // to an integer comparison.
3604 if (CC == ISD::SETOEQ)
3606 else if (CC == ISD::SETUNE)
3609 SDValue Mask = DAG.getConstant(0x7fffffff, MVT::i32);
3611 if (LHS.getValueType() == MVT::f32) {
3612 LHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3613 bitcastf32Toi32(LHS, DAG), Mask);
3614 RHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3615 bitcastf32Toi32(RHS, DAG), Mask);
3616 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3617 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3618 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3619 Chain, Dest, ARMcc, CCR, Cmp);
3624 expandf64Toi32(LHS, DAG, LHS1, LHS2);
3625 expandf64Toi32(RHS, DAG, RHS1, RHS2);
3626 LHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, LHS2, Mask);
3627 RHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, RHS2, Mask);
3628 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3629 ARMcc = DAG.getConstant(CondCode, MVT::i32);
3630 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3631 SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest };
3632 return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops);
3638 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
3639 SDValue Chain = Op.getOperand(0);
3640 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3641 SDValue LHS = Op.getOperand(2);
3642 SDValue RHS = Op.getOperand(3);
3643 SDValue Dest = Op.getOperand(4);
3646 if (LHS.getValueType() == MVT::i32) {
3648 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3649 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3650 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3651 Chain, Dest, ARMcc, CCR, Cmp);
3654 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3656 if (getTargetMachine().Options.UnsafeFPMath &&
3657 (CC == ISD::SETEQ || CC == ISD::SETOEQ ||
3658 CC == ISD::SETNE || CC == ISD::SETUNE)) {
3659 SDValue Result = OptimizeVFPBrcond(Op, DAG);
3660 if (Result.getNode())
3664 ARMCC::CondCodes CondCode, CondCode2;
3665 FPCCToARMCC(CC, CondCode, CondCode2);
3667 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
3668 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
3669 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3670 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3671 SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp };
3672 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops);
3673 if (CondCode2 != ARMCC::AL) {
3674 ARMcc = DAG.getConstant(CondCode2, MVT::i32);
3675 SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) };
3676 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops);
3681 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
3682 SDValue Chain = Op.getOperand(0);
3683 SDValue Table = Op.getOperand(1);
3684 SDValue Index = Op.getOperand(2);
3687 EVT PTy = getPointerTy();
3688 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
3689 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
3690 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
3691 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
3692 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
3693 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
3694 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
3695 if (Subtarget->isThumb2()) {
3696 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
3697 // which does another jump to the destination. This also makes it easier
3698 // to translate it to TBB / TBH later.
3699 // FIXME: This might not work if the function is extremely large.
3700 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
3701 Addr, Op.getOperand(2), JTI, UId);
3703 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
3704 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
3705 MachinePointerInfo::getJumpTable(),
3706 false, false, false, 0);
3707 Chain = Addr.getValue(1);
3708 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
3709 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3711 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
3712 MachinePointerInfo::getJumpTable(),
3713 false, false, false, 0);
3714 Chain = Addr.getValue(1);
3715 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3719 static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3720 EVT VT = Op.getValueType();
3723 if (Op.getValueType().getVectorElementType() == MVT::i32) {
3724 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::f32)
3726 return DAG.UnrollVectorOp(Op.getNode());
3729 assert(Op.getOperand(0).getValueType() == MVT::v4f32 &&
3730 "Invalid type for custom lowering!");
3731 if (VT != MVT::v4i16)
3732 return DAG.UnrollVectorOp(Op.getNode());
3734 Op = DAG.getNode(Op.getOpcode(), dl, MVT::v4i32, Op.getOperand(0));
3735 return DAG.getNode(ISD::TRUNCATE, dl, VT, Op);
3738 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3739 EVT VT = Op.getValueType();
3741 return LowerVectorFP_TO_INT(Op, DAG);
3746 switch (Op.getOpcode()) {
3747 default: llvm_unreachable("Invalid opcode!");
3748 case ISD::FP_TO_SINT:
3749 Opc = ARMISD::FTOSI;
3751 case ISD::FP_TO_UINT:
3752 Opc = ARMISD::FTOUI;
3755 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
3756 return DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
3759 static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3760 EVT VT = Op.getValueType();
3763 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::i32) {
3764 if (VT.getVectorElementType() == MVT::f32)
3766 return DAG.UnrollVectorOp(Op.getNode());
3769 assert(Op.getOperand(0).getValueType() == MVT::v4i16 &&
3770 "Invalid type for custom lowering!");
3771 if (VT != MVT::v4f32)
3772 return DAG.UnrollVectorOp(Op.getNode());
3776 switch (Op.getOpcode()) {
3777 default: llvm_unreachable("Invalid opcode!");
3778 case ISD::SINT_TO_FP:
3779 CastOpc = ISD::SIGN_EXTEND;
3780 Opc = ISD::SINT_TO_FP;
3782 case ISD::UINT_TO_FP:
3783 CastOpc = ISD::ZERO_EXTEND;
3784 Opc = ISD::UINT_TO_FP;
3788 Op = DAG.getNode(CastOpc, dl, MVT::v4i32, Op.getOperand(0));
3789 return DAG.getNode(Opc, dl, VT, Op);
3792 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3793 EVT VT = Op.getValueType();
3795 return LowerVectorINT_TO_FP(Op, DAG);
3800 switch (Op.getOpcode()) {
3801 default: llvm_unreachable("Invalid opcode!");
3802 case ISD::SINT_TO_FP:
3803 Opc = ARMISD::SITOF;
3805 case ISD::UINT_TO_FP:
3806 Opc = ARMISD::UITOF;
3810 Op = DAG.getNode(ISD::BITCAST, dl, MVT::f32, Op.getOperand(0));
3811 return DAG.getNode(Opc, dl, VT, Op);
3814 SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
3815 // Implement fcopysign with a fabs and a conditional fneg.
3816 SDValue Tmp0 = Op.getOperand(0);
3817 SDValue Tmp1 = Op.getOperand(1);
3819 EVT VT = Op.getValueType();
3820 EVT SrcVT = Tmp1.getValueType();
3821 bool InGPR = Tmp0.getOpcode() == ISD::BITCAST ||
3822 Tmp0.getOpcode() == ARMISD::VMOVDRR;
3823 bool UseNEON = !InGPR && Subtarget->hasNEON();
3826 // Use VBSL to copy the sign bit.
3827 unsigned EncodedVal = ARM_AM::createNEONModImm(0x6, 0x80);
3828 SDValue Mask = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v2i32,
3829 DAG.getTargetConstant(EncodedVal, MVT::i32));
3830 EVT OpVT = (VT == MVT::f32) ? MVT::v2i32 : MVT::v1i64;
3832 Mask = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3833 DAG.getNode(ISD::BITCAST, dl, OpVT, Mask),
3834 DAG.getConstant(32, MVT::i32));
3835 else /*if (VT == MVT::f32)*/
3836 Tmp0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp0);
3837 if (SrcVT == MVT::f32) {
3838 Tmp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp1);
3840 Tmp1 = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3841 DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1),
3842 DAG.getConstant(32, MVT::i32));
3843 } else if (VT == MVT::f32)
3844 Tmp1 = DAG.getNode(ARMISD::VSHRu, dl, MVT::v1i64,
3845 DAG.getNode(ISD::BITCAST, dl, MVT::v1i64, Tmp1),
3846 DAG.getConstant(32, MVT::i32));
3847 Tmp0 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp0);
3848 Tmp1 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1);
3850 SDValue AllOnes = DAG.getTargetConstant(ARM_AM::createNEONModImm(0xe, 0xff),
3852 AllOnes = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v8i8, AllOnes);
3853 SDValue MaskNot = DAG.getNode(ISD::XOR, dl, OpVT, Mask,
3854 DAG.getNode(ISD::BITCAST, dl, OpVT, AllOnes));
3856 SDValue Res = DAG.getNode(ISD::OR, dl, OpVT,
3857 DAG.getNode(ISD::AND, dl, OpVT, Tmp1, Mask),
3858 DAG.getNode(ISD::AND, dl, OpVT, Tmp0, MaskNot));
3859 if (VT == MVT::f32) {
3860 Res = DAG.getNode(ISD::BITCAST, dl, MVT::v2f32, Res);
3861 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, Res,
3862 DAG.getConstant(0, MVT::i32));
3864 Res = DAG.getNode(ISD::BITCAST, dl, MVT::f64, Res);
3870 // Bitcast operand 1 to i32.
3871 if (SrcVT == MVT::f64)
3872 Tmp1 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
3874 Tmp1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp1);
3876 // Or in the signbit with integer operations.
3877 SDValue Mask1 = DAG.getConstant(0x80000000, MVT::i32);
3878 SDValue Mask2 = DAG.getConstant(0x7fffffff, MVT::i32);
3879 Tmp1 = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp1, Mask1);
3880 if (VT == MVT::f32) {
3881 Tmp0 = DAG.getNode(ISD::AND, dl, MVT::i32,
3882 DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp0), Mask2);
3883 return DAG.getNode(ISD::BITCAST, dl, MVT::f32,
3884 DAG.getNode(ISD::OR, dl, MVT::i32, Tmp0, Tmp1));
3887 // f64: Or the high part with signbit and then combine two parts.
3888 Tmp0 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
3890 SDValue Lo = Tmp0.getValue(0);
3891 SDValue Hi = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp0.getValue(1), Mask2);
3892 Hi = DAG.getNode(ISD::OR, dl, MVT::i32, Hi, Tmp1);
3893 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
3896 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
3897 MachineFunction &MF = DAG.getMachineFunction();
3898 MachineFrameInfo *MFI = MF.getFrameInfo();
3899 MFI->setReturnAddressIsTaken(true);
3901 if (verifyReturnAddressArgumentIsConstant(Op, DAG))
3904 EVT VT = Op.getValueType();
3906 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3908 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
3909 SDValue Offset = DAG.getConstant(4, MVT::i32);
3910 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
3911 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
3912 MachinePointerInfo(), false, false, false, 0);
3915 // Return LR, which contains the return address. Mark it an implicit live-in.
3916 unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32));
3917 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
3920 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
3921 const ARMBaseRegisterInfo &ARI =
3922 *static_cast<const ARMBaseRegisterInfo*>(RegInfo);
3923 MachineFunction &MF = DAG.getMachineFunction();
3924 MachineFrameInfo *MFI = MF.getFrameInfo();
3925 MFI->setFrameAddressIsTaken(true);
3927 EVT VT = Op.getValueType();
3928 SDLoc dl(Op); // FIXME probably not meaningful
3929 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3930 unsigned FrameReg = ARI.getFrameRegister(MF);
3931 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
3933 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
3934 MachinePointerInfo(),
3935 false, false, false, 0);
3939 // FIXME? Maybe this could be a TableGen attribute on some registers and
3940 // this table could be generated automatically from RegInfo.
3941 unsigned ARMTargetLowering::getRegisterByName(const char* RegName,
3943 unsigned Reg = StringSwitch<unsigned>(RegName)
3944 .Case("sp", ARM::SP)
3948 report_fatal_error("Invalid register name global variable");
3951 /// ExpandBITCAST - If the target supports VFP, this function is called to
3952 /// expand a bit convert where either the source or destination type is i64 to
3953 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
3954 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
3955 /// vectors), since the legalizer won't know what to do with that.
3956 static SDValue ExpandBITCAST(SDNode *N, SelectionDAG &DAG) {
3957 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3959 SDValue Op = N->getOperand(0);
3961 // This function is only supposed to be called for i64 types, either as the
3962 // source or destination of the bit convert.
3963 EVT SrcVT = Op.getValueType();
3964 EVT DstVT = N->getValueType(0);
3965 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
3966 "ExpandBITCAST called for non-i64 type");
3968 // Turn i64->f64 into VMOVDRR.
3969 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
3970 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
3971 DAG.getConstant(0, MVT::i32));
3972 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
3973 DAG.getConstant(1, MVT::i32));
3974 return DAG.getNode(ISD::BITCAST, dl, DstVT,
3975 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
3978 // Turn f64->i64 into VMOVRRD.
3979 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
3981 if (TLI.isBigEndian() && SrcVT.isVector() &&
3982 SrcVT.getVectorNumElements() > 1)
3983 Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
3984 DAG.getVTList(MVT::i32, MVT::i32),
3985 DAG.getNode(ARMISD::VREV64, dl, SrcVT, Op));
3987 Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
3988 DAG.getVTList(MVT::i32, MVT::i32), Op);
3989 // Merge the pieces into a single i64 value.
3990 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
3996 /// getZeroVector - Returns a vector of specified type with all zero elements.
3997 /// Zero vectors are used to represent vector negation and in those cases
3998 /// will be implemented with the NEON VNEG instruction. However, VNEG does
3999 /// not support i64 elements, so sometimes the zero vectors will need to be
4000 /// explicitly constructed. Regardless, use a canonical VMOV to create the
4002 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, SDLoc dl) {
4003 assert(VT.isVector() && "Expected a vector type");
4004 // The canonical modified immediate encoding of a zero vector is....0!
4005 SDValue EncodedVal = DAG.getTargetConstant(0, MVT::i32);
4006 EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
4007 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal);
4008 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4011 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
4012 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
4013 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
4014 SelectionDAG &DAG) const {
4015 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4016 EVT VT = Op.getValueType();
4017 unsigned VTBits = VT.getSizeInBits();
4019 SDValue ShOpLo = Op.getOperand(0);
4020 SDValue ShOpHi = Op.getOperand(1);
4021 SDValue ShAmt = Op.getOperand(2);
4023 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
4025 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
4027 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
4028 DAG.getConstant(VTBits, MVT::i32), ShAmt);
4029 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
4030 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
4031 DAG.getConstant(VTBits, MVT::i32));
4032 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
4033 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
4034 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
4036 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
4037 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
4039 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
4040 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc,
4043 SDValue Ops[2] = { Lo, Hi };
4044 return DAG.getMergeValues(Ops, dl);
4047 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
4048 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
4049 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
4050 SelectionDAG &DAG) const {
4051 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4052 EVT VT = Op.getValueType();
4053 unsigned VTBits = VT.getSizeInBits();
4055 SDValue ShOpLo = Op.getOperand(0);
4056 SDValue ShOpHi = Op.getOperand(1);
4057 SDValue ShAmt = Op.getOperand(2);
4060 assert(Op.getOpcode() == ISD::SHL_PARTS);
4061 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
4062 DAG.getConstant(VTBits, MVT::i32), ShAmt);
4063 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
4064 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
4065 DAG.getConstant(VTBits, MVT::i32));
4066 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
4067 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
4069 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
4070 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
4071 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
4073 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
4074 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc,
4077 SDValue Ops[2] = { Lo, Hi };
4078 return DAG.getMergeValues(Ops, dl);
4081 SDValue ARMTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
4082 SelectionDAG &DAG) const {
4083 // The rounding mode is in bits 23:22 of the FPSCR.
4084 // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0
4085 // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3)
4086 // so that the shift + and get folded into a bitfield extract.
4088 SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::i32,
4089 DAG.getConstant(Intrinsic::arm_get_fpscr,
4091 SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPSCR,
4092 DAG.getConstant(1U << 22, MVT::i32));
4093 SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds,
4094 DAG.getConstant(22, MVT::i32));
4095 return DAG.getNode(ISD::AND, dl, MVT::i32, RMODE,
4096 DAG.getConstant(3, MVT::i32));
4099 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
4100 const ARMSubtarget *ST) {
4101 EVT VT = N->getValueType(0);
4104 if (!ST->hasV6T2Ops())
4107 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
4108 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
4111 /// getCTPOP16BitCounts - Returns a v8i8/v16i8 vector containing the bit-count
4112 /// for each 16-bit element from operand, repeated. The basic idea is to
4113 /// leverage vcnt to get the 8-bit counts, gather and add the results.
4115 /// Trace for v4i16:
4116 /// input = [v0 v1 v2 v3 ] (vi 16-bit element)
4117 /// cast: N0 = [w0 w1 w2 w3 w4 w5 w6 w7] (v0 = [w0 w1], wi 8-bit element)
4118 /// vcnt: N1 = [b0 b1 b2 b3 b4 b5 b6 b7] (bi = bit-count of 8-bit element wi)
4119 /// vrev: N2 = [b1 b0 b3 b2 b5 b4 b7 b6]
4120 /// [b0 b1 b2 b3 b4 b5 b6 b7]
4121 /// +[b1 b0 b3 b2 b5 b4 b7 b6]
4122 /// N3=N1+N2 = [k0 k0 k1 k1 k2 k2 k3 k3] (k0 = b0+b1 = bit-count of 16-bit v0,
4123 /// vuzp: = [k0 k1 k2 k3 k0 k1 k2 k3] each ki is 8-bits)
4124 static SDValue getCTPOP16BitCounts(SDNode *N, SelectionDAG &DAG) {
4125 EVT VT = N->getValueType(0);
4128 EVT VT8Bit = VT.is64BitVector() ? MVT::v8i8 : MVT::v16i8;
4129 SDValue N0 = DAG.getNode(ISD::BITCAST, DL, VT8Bit, N->getOperand(0));
4130 SDValue N1 = DAG.getNode(ISD::CTPOP, DL, VT8Bit, N0);
4131 SDValue N2 = DAG.getNode(ARMISD::VREV16, DL, VT8Bit, N1);
4132 SDValue N3 = DAG.getNode(ISD::ADD, DL, VT8Bit, N1, N2);
4133 return DAG.getNode(ARMISD::VUZP, DL, VT8Bit, N3, N3);
4136 /// lowerCTPOP16BitElements - Returns a v4i16/v8i16 vector containing the
4137 /// bit-count for each 16-bit element from the operand. We need slightly
4138 /// different sequencing for v4i16 and v8i16 to stay within NEON's available
4139 /// 64/128-bit registers.
4141 /// Trace for v4i16:
4142 /// input = [v0 v1 v2 v3 ] (vi 16-bit element)
4143 /// v8i8: BitCounts = [k0 k1 k2 k3 k0 k1 k2 k3 ] (ki is the bit-count of vi)
4144 /// v8i16:Extended = [k0 k1 k2 k3 k0 k1 k2 k3 ]
4145 /// v4i16:Extracted = [k0 k1 k2 k3 ]
4146 static SDValue lowerCTPOP16BitElements(SDNode *N, SelectionDAG &DAG) {
4147 EVT VT = N->getValueType(0);
4150 SDValue BitCounts = getCTPOP16BitCounts(N, DAG);
4151 if (VT.is64BitVector()) {
4152 SDValue Extended = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v8i16, BitCounts);
4153 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i16, Extended,
4154 DAG.getIntPtrConstant(0));
4156 SDValue Extracted = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v8i8,
4157 BitCounts, DAG.getIntPtrConstant(0));
4158 return DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v8i16, Extracted);
4162 /// lowerCTPOP32BitElements - Returns a v2i32/v4i32 vector containing the
4163 /// bit-count for each 32-bit element from the operand. The idea here is
4164 /// to split the vector into 16-bit elements, leverage the 16-bit count
4165 /// routine, and then combine the results.
4167 /// Trace for v2i32 (v4i32 similar with Extracted/Extended exchanged):
4168 /// input = [v0 v1 ] (vi: 32-bit elements)
4169 /// Bitcast = [w0 w1 w2 w3 ] (wi: 16-bit elements, v0 = [w0 w1])
4170 /// Counts16 = [k0 k1 k2 k3 ] (ki: 16-bit elements, bit-count of wi)
4171 /// vrev: N0 = [k1 k0 k3 k2 ]
4173 /// N1 =+[k1 k0 k3 k2 ]
4175 /// N2 =+[k1 k3 k0 k2 ]
4177 /// Extended =+[k1 k3 k0 k2 ]
4179 /// Extracted=+[k1 k3 ]
4181 static SDValue lowerCTPOP32BitElements(SDNode *N, SelectionDAG &DAG) {
4182 EVT VT = N->getValueType(0);
4185 EVT VT16Bit = VT.is64BitVector() ? MVT::v4i16 : MVT::v8i16;
4187 SDValue Bitcast = DAG.getNode(ISD::BITCAST, DL, VT16Bit, N->getOperand(0));
4188 SDValue Counts16 = lowerCTPOP16BitElements(Bitcast.getNode(), DAG);
4189 SDValue N0 = DAG.getNode(ARMISD::VREV32, DL, VT16Bit, Counts16);
4190 SDValue N1 = DAG.getNode(ISD::ADD, DL, VT16Bit, Counts16, N0);
4191 SDValue N2 = DAG.getNode(ARMISD::VUZP, DL, VT16Bit, N1, N1);
4193 if (VT.is64BitVector()) {
4194 SDValue Extended = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v4i32, N2);
4195 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i32, Extended,
4196 DAG.getIntPtrConstant(0));
4198 SDValue Extracted = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i16, N2,
4199 DAG.getIntPtrConstant(0));
4200 return DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v4i32, Extracted);
4204 static SDValue LowerCTPOP(SDNode *N, SelectionDAG &DAG,
4205 const ARMSubtarget *ST) {
4206 EVT VT = N->getValueType(0);
4208 assert(ST->hasNEON() && "Custom ctpop lowering requires NEON.");
4209 assert((VT == MVT::v2i32 || VT == MVT::v4i32 ||
4210 VT == MVT::v4i16 || VT == MVT::v8i16) &&
4211 "Unexpected type for custom ctpop lowering");
4213 if (VT.getVectorElementType() == MVT::i32)
4214 return lowerCTPOP32BitElements(N, DAG);
4216 return lowerCTPOP16BitElements(N, DAG);
4219 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
4220 const ARMSubtarget *ST) {
4221 EVT VT = N->getValueType(0);
4227 // Lower vector shifts on NEON to use VSHL.
4228 assert(ST->hasNEON() && "unexpected vector shift");
4230 // Left shifts translate directly to the vshiftu intrinsic.
4231 if (N->getOpcode() == ISD::SHL)
4232 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
4233 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
4234 N->getOperand(0), N->getOperand(1));
4236 assert((N->getOpcode() == ISD::SRA ||
4237 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
4239 // NEON uses the same intrinsics for both left and right shifts. For
4240 // right shifts, the shift amounts are negative, so negate the vector of
4242 EVT ShiftVT = N->getOperand(1).getValueType();
4243 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
4244 getZeroVector(ShiftVT, DAG, dl),
4246 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
4247 Intrinsic::arm_neon_vshifts :
4248 Intrinsic::arm_neon_vshiftu);
4249 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
4250 DAG.getConstant(vshiftInt, MVT::i32),
4251 N->getOperand(0), NegatedCount);
4254 static SDValue Expand64BitShift(SDNode *N, SelectionDAG &DAG,
4255 const ARMSubtarget *ST) {
4256 EVT VT = N->getValueType(0);
4259 // We can get here for a node like i32 = ISD::SHL i32, i64
4263 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
4264 "Unknown shift to lower!");
4266 // We only lower SRA, SRL of 1 here, all others use generic lowering.
4267 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
4268 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
4271 // If we are in thumb mode, we don't have RRX.
4272 if (ST->isThumb1Only()) return SDValue();
4274 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
4275 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
4276 DAG.getConstant(0, MVT::i32));
4277 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
4278 DAG.getConstant(1, MVT::i32));
4280 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
4281 // captures the result into a carry flag.
4282 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
4283 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Glue), Hi);
4285 // The low part is an ARMISD::RRX operand, which shifts the carry in.
4286 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
4288 // Merge the pieces into a single i64 value.
4289 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
4292 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
4293 SDValue TmpOp0, TmpOp1;
4294 bool Invert = false;
4298 SDValue Op0 = Op.getOperand(0);
4299 SDValue Op1 = Op.getOperand(1);
4300 SDValue CC = Op.getOperand(2);
4301 EVT VT = Op.getValueType();
4302 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
4305 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
4306 switch (SetCCOpcode) {
4307 default: llvm_unreachable("Illegal FP comparison");
4309 case ISD::SETNE: Invert = true; // Fallthrough
4311 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
4313 case ISD::SETLT: Swap = true; // Fallthrough
4315 case ISD::SETGT: Opc = ARMISD::VCGT; break;
4317 case ISD::SETLE: Swap = true; // Fallthrough
4319 case ISD::SETGE: Opc = ARMISD::VCGE; break;
4320 case ISD::SETUGE: Swap = true; // Fallthrough
4321 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
4322 case ISD::SETUGT: Swap = true; // Fallthrough
4323 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
4324 case ISD::SETUEQ: Invert = true; // Fallthrough
4326 // Expand this to (OLT | OGT).
4330 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
4331 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
4333 case ISD::SETUO: Invert = true; // Fallthrough
4335 // Expand this to (OLT | OGE).
4339 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
4340 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
4344 // Integer comparisons.
4345 switch (SetCCOpcode) {
4346 default: llvm_unreachable("Illegal integer comparison");
4347 case ISD::SETNE: Invert = true;
4348 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
4349 case ISD::SETLT: Swap = true;
4350 case ISD::SETGT: Opc = ARMISD::VCGT; break;
4351 case ISD::SETLE: Swap = true;
4352 case ISD::SETGE: Opc = ARMISD::VCGE; break;
4353 case ISD::SETULT: Swap = true;
4354 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
4355 case ISD::SETULE: Swap = true;
4356 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
4359 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
4360 if (Opc == ARMISD::VCEQ) {
4363 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
4365 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
4368 // Ignore bitconvert.
4369 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BITCAST)
4370 AndOp = AndOp.getOperand(0);
4372 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
4374 Op0 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(0));
4375 Op1 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(1));
4382 std::swap(Op0, Op1);
4384 // If one of the operands is a constant vector zero, attempt to fold the
4385 // comparison to a specialized compare-against-zero form.
4387 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
4389 else if (ISD::isBuildVectorAllZeros(Op0.getNode())) {
4390 if (Opc == ARMISD::VCGE)
4391 Opc = ARMISD::VCLEZ;
4392 else if (Opc == ARMISD::VCGT)
4393 Opc = ARMISD::VCLTZ;
4398 if (SingleOp.getNode()) {
4401 Result = DAG.getNode(ARMISD::VCEQZ, dl, VT, SingleOp); break;
4403 Result = DAG.getNode(ARMISD::VCGEZ, dl, VT, SingleOp); break;
4405 Result = DAG.getNode(ARMISD::VCLEZ, dl, VT, SingleOp); break;
4407 Result = DAG.getNode(ARMISD::VCGTZ, dl, VT, SingleOp); break;
4409 Result = DAG.getNode(ARMISD::VCLTZ, dl, VT, SingleOp); break;
4411 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
4414 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
4418 Result = DAG.getNOT(dl, Result, VT);
4423 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
4424 /// valid vector constant for a NEON instruction with a "modified immediate"
4425 /// operand (e.g., VMOV). If so, return the encoded value.
4426 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
4427 unsigned SplatBitSize, SelectionDAG &DAG,
4428 EVT &VT, bool is128Bits, NEONModImmType type) {
4429 unsigned OpCmode, Imm;
4431 // SplatBitSize is set to the smallest size that splats the vector, so a
4432 // zero vector will always have SplatBitSize == 8. However, NEON modified
4433 // immediate instructions others than VMOV do not support the 8-bit encoding
4434 // of a zero vector, and the default encoding of zero is supposed to be the
4439 switch (SplatBitSize) {
4441 if (type != VMOVModImm)
4443 // Any 1-byte value is OK. Op=0, Cmode=1110.
4444 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
4447 VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
4451 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
4452 VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
4453 if ((SplatBits & ~0xff) == 0) {
4454 // Value = 0x00nn: Op=x, Cmode=100x.
4459 if ((SplatBits & ~0xff00) == 0) {
4460 // Value = 0xnn00: Op=x, Cmode=101x.
4462 Imm = SplatBits >> 8;
4468 // NEON's 32-bit VMOV supports splat values where:
4469 // * only one byte is nonzero, or
4470 // * the least significant byte is 0xff and the second byte is nonzero, or
4471 // * the least significant 2 bytes are 0xff and the third is nonzero.
4472 VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
4473 if ((SplatBits & ~0xff) == 0) {
4474 // Value = 0x000000nn: Op=x, Cmode=000x.
4479 if ((SplatBits & ~0xff00) == 0) {
4480 // Value = 0x0000nn00: Op=x, Cmode=001x.
4482 Imm = SplatBits >> 8;
4485 if ((SplatBits & ~0xff0000) == 0) {
4486 // Value = 0x00nn0000: Op=x, Cmode=010x.
4488 Imm = SplatBits >> 16;
4491 if ((SplatBits & ~0xff000000) == 0) {
4492 // Value = 0xnn000000: Op=x, Cmode=011x.
4494 Imm = SplatBits >> 24;
4498 // cmode == 0b1100 and cmode == 0b1101 are not supported for VORR or VBIC
4499 if (type == OtherModImm) return SDValue();
4501 if ((SplatBits & ~0xffff) == 0 &&
4502 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
4503 // Value = 0x0000nnff: Op=x, Cmode=1100.
4505 Imm = SplatBits >> 8;
4509 if ((SplatBits & ~0xffffff) == 0 &&
4510 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
4511 // Value = 0x00nnffff: Op=x, Cmode=1101.
4513 Imm = SplatBits >> 16;
4517 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
4518 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
4519 // VMOV.I32. A (very) minor optimization would be to replicate the value
4520 // and fall through here to test for a valid 64-bit splat. But, then the
4521 // caller would also need to check and handle the change in size.
4525 if (type != VMOVModImm)
4527 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
4528 uint64_t BitMask = 0xff;
4530 unsigned ImmMask = 1;
4532 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
4533 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
4536 } else if ((SplatBits & BitMask) != 0) {
4543 if (DAG.getTargetLoweringInfo().isBigEndian())
4544 // swap higher and lower 32 bit word
4545 Imm = ((Imm & 0xf) << 4) | ((Imm & 0xf0) >> 4);
4547 // Op=1, Cmode=1110.
4549 VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
4554 llvm_unreachable("unexpected size for isNEONModifiedImm");
4557 unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm);
4558 return DAG.getTargetConstant(EncodedVal, MVT::i32);
4561 SDValue ARMTargetLowering::LowerConstantFP(SDValue Op, SelectionDAG &DAG,
4562 const ARMSubtarget *ST) const {
4566 bool IsDouble = Op.getValueType() == MVT::f64;
4567 ConstantFPSDNode *CFP = cast<ConstantFPSDNode>(Op);
4569 // Try splatting with a VMOV.f32...
4570 APFloat FPVal = CFP->getValueAPF();
4571 int ImmVal = IsDouble ? ARM_AM::getFP64Imm(FPVal) : ARM_AM::getFP32Imm(FPVal);
4574 if (IsDouble || !ST->useNEONForSinglePrecisionFP()) {
4575 // We have code in place to select a valid ConstantFP already, no need to
4580 // It's a float and we are trying to use NEON operations where
4581 // possible. Lower it to a splat followed by an extract.
4583 SDValue NewVal = DAG.getTargetConstant(ImmVal, MVT::i32);
4584 SDValue VecConstant = DAG.getNode(ARMISD::VMOVFPIMM, DL, MVT::v2f32,
4586 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecConstant,
4587 DAG.getConstant(0, MVT::i32));
4590 // The rest of our options are NEON only, make sure that's allowed before
4592 if (!ST->hasNEON() || (!IsDouble && !ST->useNEONForSinglePrecisionFP()))
4596 uint64_t iVal = FPVal.bitcastToAPInt().getZExtValue();
4598 // It wouldn't really be worth bothering for doubles except for one very
4599 // important value, which does happen to match: 0.0. So make sure we don't do
4601 if (IsDouble && (iVal & 0xffffffff) != (iVal >> 32))
4604 // Try a VMOV.i32 (FIXME: i8, i16, or i64 could work too).
4605 SDValue NewVal = isNEONModifiedImm(iVal & 0xffffffffU, 0, 32, DAG, VMovVT,
4607 if (NewVal != SDValue()) {
4609 SDValue VecConstant = DAG.getNode(ARMISD::VMOVIMM, DL, VMovVT,
4612 return DAG.getNode(ISD::BITCAST, DL, MVT::f64, VecConstant);
4614 // It's a float: cast and extract a vector element.
4615 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
4617 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
4618 DAG.getConstant(0, MVT::i32));
4621 // Finally, try a VMVN.i32
4622 NewVal = isNEONModifiedImm(~iVal & 0xffffffffU, 0, 32, DAG, VMovVT,
4624 if (NewVal != SDValue()) {
4626 SDValue VecConstant = DAG.getNode(ARMISD::VMVNIMM, DL, VMovVT, NewVal);
4629 return DAG.getNode(ISD::BITCAST, DL, MVT::f64, VecConstant);
4631 // It's a float: cast and extract a vector element.
4632 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
4634 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
4635 DAG.getConstant(0, MVT::i32));
4641 // check if an VEXT instruction can handle the shuffle mask when the
4642 // vector sources of the shuffle are the same.
4643 static bool isSingletonVEXTMask(ArrayRef<int> M, EVT VT, unsigned &Imm) {
4644 unsigned NumElts = VT.getVectorNumElements();
4646 // Assume that the first shuffle index is not UNDEF. Fail if it is.
4652 // If this is a VEXT shuffle, the immediate value is the index of the first
4653 // element. The other shuffle indices must be the successive elements after
4655 unsigned ExpectedElt = Imm;
4656 for (unsigned i = 1; i < NumElts; ++i) {
4657 // Increment the expected index. If it wraps around, just follow it
4658 // back to index zero and keep going.
4660 if (ExpectedElt == NumElts)
4663 if (M[i] < 0) continue; // ignore UNDEF indices
4664 if (ExpectedElt != static_cast<unsigned>(M[i]))
4672 static bool isVEXTMask(ArrayRef<int> M, EVT VT,
4673 bool &ReverseVEXT, unsigned &Imm) {
4674 unsigned NumElts = VT.getVectorNumElements();
4675 ReverseVEXT = false;
4677 // Assume that the first shuffle index is not UNDEF. Fail if it is.
4683 // If this is a VEXT shuffle, the immediate value is the index of the first
4684 // element. The other shuffle indices must be the successive elements after
4686 unsigned ExpectedElt = Imm;
4687 for (unsigned i = 1; i < NumElts; ++i) {
4688 // Increment the expected index. If it wraps around, it may still be
4689 // a VEXT but the source vectors must be swapped.
4691 if (ExpectedElt == NumElts * 2) {
4696 if (M[i] < 0) continue; // ignore UNDEF indices
4697 if (ExpectedElt != static_cast<unsigned>(M[i]))
4701 // Adjust the index value if the source operands will be swapped.
4708 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
4709 /// instruction with the specified blocksize. (The order of the elements
4710 /// within each block of the vector is reversed.)
4711 static bool isVREVMask(ArrayRef<int> M, EVT VT, unsigned BlockSize) {
4712 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
4713 "Only possible block sizes for VREV are: 16, 32, 64");
4715 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4719 unsigned NumElts = VT.getVectorNumElements();
4720 unsigned BlockElts = M[0] + 1;
4721 // If the first shuffle index is UNDEF, be optimistic.
4723 BlockElts = BlockSize / EltSz;
4725 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
4728 for (unsigned i = 0; i < NumElts; ++i) {
4729 if (M[i] < 0) continue; // ignore UNDEF indices
4730 if ((unsigned) M[i] != (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
4737 static bool isVTBLMask(ArrayRef<int> M, EVT VT) {
4738 // We can handle <8 x i8> vector shuffles. If the index in the mask is out of
4739 // range, then 0 is placed into the resulting vector. So pretty much any mask
4740 // of 8 elements can work here.
4741 return VT == MVT::v8i8 && M.size() == 8;
4744 static bool isVTRNMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4745 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4749 unsigned NumElts = VT.getVectorNumElements();
4750 WhichResult = (M[0] == 0 ? 0 : 1);
4751 for (unsigned i = 0; i < NumElts; i += 2) {
4752 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
4753 (M[i+1] >= 0 && (unsigned) M[i+1] != i + NumElts + WhichResult))
4759 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
4760 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4761 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
4762 static bool isVTRN_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4763 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4767 unsigned NumElts = VT.getVectorNumElements();
4768 WhichResult = (M[0] == 0 ? 0 : 1);
4769 for (unsigned i = 0; i < NumElts; i += 2) {
4770 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
4771 (M[i+1] >= 0 && (unsigned) M[i+1] != i + WhichResult))
4777 static bool isVUZPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4778 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4782 unsigned NumElts = VT.getVectorNumElements();
4783 WhichResult = (M[0] == 0 ? 0 : 1);
4784 for (unsigned i = 0; i != NumElts; ++i) {
4785 if (M[i] < 0) continue; // ignore UNDEF indices
4786 if ((unsigned) M[i] != 2 * i + WhichResult)
4790 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4791 if (VT.is64BitVector() && EltSz == 32)
4797 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
4798 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4799 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
4800 static bool isVUZP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4801 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4805 unsigned Half = VT.getVectorNumElements() / 2;
4806 WhichResult = (M[0] == 0 ? 0 : 1);
4807 for (unsigned j = 0; j != 2; ++j) {
4808 unsigned Idx = WhichResult;
4809 for (unsigned i = 0; i != Half; ++i) {
4810 int MIdx = M[i + j * Half];
4811 if (MIdx >= 0 && (unsigned) MIdx != Idx)
4817 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4818 if (VT.is64BitVector() && EltSz == 32)
4824 static bool isVZIPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4825 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4829 unsigned NumElts = VT.getVectorNumElements();
4830 WhichResult = (M[0] == 0 ? 0 : 1);
4831 unsigned Idx = WhichResult * NumElts / 2;
4832 for (unsigned i = 0; i != NumElts; i += 2) {
4833 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
4834 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx + NumElts))
4839 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4840 if (VT.is64BitVector() && EltSz == 32)
4846 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
4847 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4848 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
4849 static bool isVZIP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4850 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4854 unsigned NumElts = VT.getVectorNumElements();
4855 WhichResult = (M[0] == 0 ? 0 : 1);
4856 unsigned Idx = WhichResult * NumElts / 2;
4857 for (unsigned i = 0; i != NumElts; i += 2) {
4858 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
4859 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx))
4864 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4865 if (VT.is64BitVector() && EltSz == 32)
4871 /// \return true if this is a reverse operation on an vector.
4872 static bool isReverseMask(ArrayRef<int> M, EVT VT) {
4873 unsigned NumElts = VT.getVectorNumElements();
4874 // Make sure the mask has the right size.
4875 if (NumElts != M.size())
4878 // Look for <15, ..., 3, -1, 1, 0>.
4879 for (unsigned i = 0; i != NumElts; ++i)
4880 if (M[i] >= 0 && M[i] != (int) (NumElts - 1 - i))
4886 // If N is an integer constant that can be moved into a register in one
4887 // instruction, return an SDValue of such a constant (will become a MOV
4888 // instruction). Otherwise return null.
4889 static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG,
4890 const ARMSubtarget *ST, SDLoc dl) {
4892 if (!isa<ConstantSDNode>(N))
4894 Val = cast<ConstantSDNode>(N)->getZExtValue();
4896 if (ST->isThumb1Only()) {
4897 if (Val <= 255 || ~Val <= 255)
4898 return DAG.getConstant(Val, MVT::i32);
4900 if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1)
4901 return DAG.getConstant(Val, MVT::i32);
4906 // If this is a case we can't handle, return null and let the default
4907 // expansion code take care of it.
4908 SDValue ARMTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
4909 const ARMSubtarget *ST) const {
4910 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
4912 EVT VT = Op.getValueType();
4914 APInt SplatBits, SplatUndef;
4915 unsigned SplatBitSize;
4917 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
4918 if (SplatBitSize <= 64) {
4919 // Check if an immediate VMOV works.
4921 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
4922 SplatUndef.getZExtValue(), SplatBitSize,
4923 DAG, VmovVT, VT.is128BitVector(),
4925 if (Val.getNode()) {
4926 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
4927 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4930 // Try an immediate VMVN.
4931 uint64_t NegatedImm = (~SplatBits).getZExtValue();
4932 Val = isNEONModifiedImm(NegatedImm,
4933 SplatUndef.getZExtValue(), SplatBitSize,
4934 DAG, VmovVT, VT.is128BitVector(),
4936 if (Val.getNode()) {
4937 SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
4938 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4941 // Use vmov.f32 to materialize other v2f32 and v4f32 splats.
4942 if ((VT == MVT::v2f32 || VT == MVT::v4f32) && SplatBitSize == 32) {
4943 int ImmVal = ARM_AM::getFP32Imm(SplatBits);
4945 SDValue Val = DAG.getTargetConstant(ImmVal, MVT::i32);
4946 return DAG.getNode(ARMISD::VMOVFPIMM, dl, VT, Val);
4952 // Scan through the operands to see if only one value is used.
4954 // As an optimisation, even if more than one value is used it may be more
4955 // profitable to splat with one value then change some lanes.
4957 // Heuristically we decide to do this if the vector has a "dominant" value,
4958 // defined as splatted to more than half of the lanes.
4959 unsigned NumElts = VT.getVectorNumElements();
4960 bool isOnlyLowElement = true;
4961 bool usesOnlyOneValue = true;
4962 bool hasDominantValue = false;
4963 bool isConstant = true;
4965 // Map of the number of times a particular SDValue appears in the
4967 DenseMap<SDValue, unsigned> ValueCounts;
4969 for (unsigned i = 0; i < NumElts; ++i) {
4970 SDValue V = Op.getOperand(i);
4971 if (V.getOpcode() == ISD::UNDEF)
4974 isOnlyLowElement = false;
4975 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
4978 ValueCounts.insert(std::make_pair(V, 0));
4979 unsigned &Count = ValueCounts[V];
4981 // Is this value dominant? (takes up more than half of the lanes)
4982 if (++Count > (NumElts / 2)) {
4983 hasDominantValue = true;
4987 if (ValueCounts.size() != 1)
4988 usesOnlyOneValue = false;
4989 if (!Value.getNode() && ValueCounts.size() > 0)
4990 Value = ValueCounts.begin()->first;
4992 if (ValueCounts.size() == 0)
4993 return DAG.getUNDEF(VT);
4995 // Loads are better lowered with insert_vector_elt/ARMISD::BUILD_VECTOR.
4996 // Keep going if we are hitting this case.
4997 if (isOnlyLowElement && !ISD::isNormalLoad(Value.getNode()))
4998 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
5000 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5002 // Use VDUP for non-constant splats. For f32 constant splats, reduce to
5003 // i32 and try again.
5004 if (hasDominantValue && EltSize <= 32) {
5008 // If we are VDUPing a value that comes directly from a vector, that will
5009 // cause an unnecessary move to and from a GPR, where instead we could
5010 // just use VDUPLANE. We can only do this if the lane being extracted
5011 // is at a constant index, as the VDUP from lane instructions only have
5012 // constant-index forms.
5013 if (Value->getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5014 isa<ConstantSDNode>(Value->getOperand(1))) {
5015 // We need to create a new undef vector to use for the VDUPLANE if the
5016 // size of the vector from which we get the value is different than the
5017 // size of the vector that we need to create. We will insert the element
5018 // such that the register coalescer will remove unnecessary copies.
5019 if (VT != Value->getOperand(0).getValueType()) {
5020 ConstantSDNode *constIndex;
5021 constIndex = dyn_cast<ConstantSDNode>(Value->getOperand(1));
5022 assert(constIndex && "The index is not a constant!");
5023 unsigned index = constIndex->getAPIntValue().getLimitedValue() %
5024 VT.getVectorNumElements();
5025 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
5026 DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DAG.getUNDEF(VT),
5027 Value, DAG.getConstant(index, MVT::i32)),
5028 DAG.getConstant(index, MVT::i32));
5030 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
5031 Value->getOperand(0), Value->getOperand(1));
5033 N = DAG.getNode(ARMISD::VDUP, dl, VT, Value);
5035 if (!usesOnlyOneValue) {
5036 // The dominant value was splatted as 'N', but we now have to insert
5037 // all differing elements.
5038 for (unsigned I = 0; I < NumElts; ++I) {
5039 if (Op.getOperand(I) == Value)
5041 SmallVector<SDValue, 3> Ops;
5043 Ops.push_back(Op.getOperand(I));
5044 Ops.push_back(DAG.getConstant(I, MVT::i32));
5045 N = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Ops);
5050 if (VT.getVectorElementType().isFloatingPoint()) {
5051 SmallVector<SDValue, 8> Ops;
5052 for (unsigned i = 0; i < NumElts; ++i)
5053 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, MVT::i32,
5055 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
5056 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, Ops);
5057 Val = LowerBUILD_VECTOR(Val, DAG, ST);
5059 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
5061 if (usesOnlyOneValue) {
5062 SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
5063 if (isConstant && Val.getNode())
5064 return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
5068 // If all elements are constants and the case above didn't get hit, fall back
5069 // to the default expansion, which will generate a load from the constant
5074 // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
5076 SDValue shuffle = ReconstructShuffle(Op, DAG);
5077 if (shuffle != SDValue())
5081 // Vectors with 32- or 64-bit elements can be built by directly assigning
5082 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
5083 // will be legalized.
5084 if (EltSize >= 32) {
5085 // Do the expansion with floating-point types, since that is what the VFP
5086 // registers are defined to use, and since i64 is not legal.
5087 EVT EltVT = EVT::getFloatingPointVT(EltSize);
5088 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
5089 SmallVector<SDValue, 8> Ops;
5090 for (unsigned i = 0; i < NumElts; ++i)
5091 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, EltVT, Op.getOperand(i)));
5092 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, Ops);
5093 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
5096 // If all else fails, just use a sequence of INSERT_VECTOR_ELT when we
5097 // know the default expansion would otherwise fall back on something even
5098 // worse. For a vector with one or two non-undef values, that's
5099 // scalar_to_vector for the elements followed by a shuffle (provided the
5100 // shuffle is valid for the target) and materialization element by element
5101 // on the stack followed by a load for everything else.
5102 if (!isConstant && !usesOnlyOneValue) {
5103 SDValue Vec = DAG.getUNDEF(VT);
5104 for (unsigned i = 0 ; i < NumElts; ++i) {
5105 SDValue V = Op.getOperand(i);
5106 if (V.getOpcode() == ISD::UNDEF)
5108 SDValue LaneIdx = DAG.getConstant(i, MVT::i32);
5109 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Vec, V, LaneIdx);
5117 // Gather data to see if the operation can be modelled as a
5118 // shuffle in combination with VEXTs.
5119 SDValue ARMTargetLowering::ReconstructShuffle(SDValue Op,
5120 SelectionDAG &DAG) const {
5122 EVT VT = Op.getValueType();
5123 unsigned NumElts = VT.getVectorNumElements();
5125 SmallVector<SDValue, 2> SourceVecs;
5126 SmallVector<unsigned, 2> MinElts;
5127 SmallVector<unsigned, 2> MaxElts;
5129 for (unsigned i = 0; i < NumElts; ++i) {
5130 SDValue V = Op.getOperand(i);
5131 if (V.getOpcode() == ISD::UNDEF)
5133 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
5134 // A shuffle can only come from building a vector from various
5135 // elements of other vectors.
5137 } else if (V.getOperand(0).getValueType().getVectorElementType() !=
5138 VT.getVectorElementType()) {
5139 // This code doesn't know how to handle shuffles where the vector
5140 // element types do not match (this happens because type legalization
5141 // promotes the return type of EXTRACT_VECTOR_ELT).
5142 // FIXME: It might be appropriate to extend this code to handle
5143 // mismatched types.
5147 // Record this extraction against the appropriate vector if possible...
5148 SDValue SourceVec = V.getOperand(0);
5149 // If the element number isn't a constant, we can't effectively
5150 // analyze what's going on.
5151 if (!isa<ConstantSDNode>(V.getOperand(1)))
5153 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
5154 bool FoundSource = false;
5155 for (unsigned j = 0; j < SourceVecs.size(); ++j) {
5156 if (SourceVecs[j] == SourceVec) {
5157 if (MinElts[j] > EltNo)
5159 if (MaxElts[j] < EltNo)
5166 // Or record a new source if not...
5168 SourceVecs.push_back(SourceVec);
5169 MinElts.push_back(EltNo);
5170 MaxElts.push_back(EltNo);
5174 // Currently only do something sane when at most two source vectors
5176 if (SourceVecs.size() > 2)
5179 SDValue ShuffleSrcs[2] = {DAG.getUNDEF(VT), DAG.getUNDEF(VT) };
5180 int VEXTOffsets[2] = {0, 0};
5182 // This loop extracts the usage patterns of the source vectors
5183 // and prepares appropriate SDValues for a shuffle if possible.
5184 for (unsigned i = 0; i < SourceVecs.size(); ++i) {
5185 if (SourceVecs[i].getValueType() == VT) {
5186 // No VEXT necessary
5187 ShuffleSrcs[i] = SourceVecs[i];
5190 } else if (SourceVecs[i].getValueType().getVectorNumElements() < NumElts) {
5191 // It probably isn't worth padding out a smaller vector just to
5192 // break it down again in a shuffle.
5196 // Since only 64-bit and 128-bit vectors are legal on ARM and
5197 // we've eliminated the other cases...
5198 assert(SourceVecs[i].getValueType().getVectorNumElements() == 2*NumElts &&
5199 "unexpected vector sizes in ReconstructShuffle");
5201 if (MaxElts[i] - MinElts[i] >= NumElts) {
5202 // Span too large for a VEXT to cope
5206 if (MinElts[i] >= NumElts) {
5207 // The extraction can just take the second half
5208 VEXTOffsets[i] = NumElts;
5209 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5211 DAG.getIntPtrConstant(NumElts));
5212 } else if (MaxElts[i] < NumElts) {
5213 // The extraction can just take the first half
5215 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5217 DAG.getIntPtrConstant(0));
5219 // An actual VEXT is needed
5220 VEXTOffsets[i] = MinElts[i];
5221 SDValue VEXTSrc1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5223 DAG.getIntPtrConstant(0));
5224 SDValue VEXTSrc2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5226 DAG.getIntPtrConstant(NumElts));
5227 ShuffleSrcs[i] = DAG.getNode(ARMISD::VEXT, dl, VT, VEXTSrc1, VEXTSrc2,
5228 DAG.getConstant(VEXTOffsets[i], MVT::i32));
5232 SmallVector<int, 8> Mask;
5234 for (unsigned i = 0; i < NumElts; ++i) {
5235 SDValue Entry = Op.getOperand(i);
5236 if (Entry.getOpcode() == ISD::UNDEF) {
5241 SDValue ExtractVec = Entry.getOperand(0);
5242 int ExtractElt = cast<ConstantSDNode>(Op.getOperand(i)
5243 .getOperand(1))->getSExtValue();
5244 if (ExtractVec == SourceVecs[0]) {
5245 Mask.push_back(ExtractElt - VEXTOffsets[0]);
5247 Mask.push_back(ExtractElt + NumElts - VEXTOffsets[1]);
5251 // Final check before we try to produce nonsense...
5252 if (isShuffleMaskLegal(Mask, VT))
5253 return DAG.getVectorShuffle(VT, dl, ShuffleSrcs[0], ShuffleSrcs[1],
5259 /// isShuffleMaskLegal - Targets can use this to indicate that they only
5260 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
5261 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
5262 /// are assumed to be legal.
5264 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
5266 if (VT.getVectorNumElements() == 4 &&
5267 (VT.is128BitVector() || VT.is64BitVector())) {
5268 unsigned PFIndexes[4];
5269 for (unsigned i = 0; i != 4; ++i) {
5273 PFIndexes[i] = M[i];
5276 // Compute the index in the perfect shuffle table.
5277 unsigned PFTableIndex =
5278 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
5279 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
5280 unsigned Cost = (PFEntry >> 30);
5287 unsigned Imm, WhichResult;
5289 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5290 return (EltSize >= 32 ||
5291 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
5292 isVREVMask(M, VT, 64) ||
5293 isVREVMask(M, VT, 32) ||
5294 isVREVMask(M, VT, 16) ||
5295 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
5296 isVTBLMask(M, VT) ||
5297 isVTRNMask(M, VT, WhichResult) ||
5298 isVUZPMask(M, VT, WhichResult) ||
5299 isVZIPMask(M, VT, WhichResult) ||
5300 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
5301 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
5302 isVZIP_v_undef_Mask(M, VT, WhichResult) ||
5303 ((VT == MVT::v8i16 || VT == MVT::v16i8) && isReverseMask(M, VT)));
5306 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
5307 /// the specified operations to build the shuffle.
5308 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
5309 SDValue RHS, SelectionDAG &DAG,
5311 unsigned OpNum = (PFEntry >> 26) & 0x0F;
5312 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
5313 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
5316 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
5325 OP_VUZPL, // VUZP, left result
5326 OP_VUZPR, // VUZP, right result
5327 OP_VZIPL, // VZIP, left result
5328 OP_VZIPR, // VZIP, right result
5329 OP_VTRNL, // VTRN, left result
5330 OP_VTRNR // VTRN, right result
5333 if (OpNum == OP_COPY) {
5334 if (LHSID == (1*9+2)*9+3) return LHS;
5335 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
5339 SDValue OpLHS, OpRHS;
5340 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
5341 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
5342 EVT VT = OpLHS.getValueType();
5345 default: llvm_unreachable("Unknown shuffle opcode!");
5347 // VREV divides the vector in half and swaps within the half.
5348 if (VT.getVectorElementType() == MVT::i32 ||
5349 VT.getVectorElementType() == MVT::f32)
5350 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
5351 // vrev <4 x i16> -> VREV32
5352 if (VT.getVectorElementType() == MVT::i16)
5353 return DAG.getNode(ARMISD::VREV32, dl, VT, OpLHS);
5354 // vrev <4 x i8> -> VREV16
5355 assert(VT.getVectorElementType() == MVT::i8);
5356 return DAG.getNode(ARMISD::VREV16, dl, VT, OpLHS);
5361 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
5362 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
5366 return DAG.getNode(ARMISD::VEXT, dl, VT,
5368 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
5371 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
5372 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
5375 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
5376 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
5379 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
5380 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
5384 static SDValue LowerVECTOR_SHUFFLEv8i8(SDValue Op,
5385 ArrayRef<int> ShuffleMask,
5386 SelectionDAG &DAG) {
5387 // Check to see if we can use the VTBL instruction.
5388 SDValue V1 = Op.getOperand(0);
5389 SDValue V2 = Op.getOperand(1);
5392 SmallVector<SDValue, 8> VTBLMask;
5393 for (ArrayRef<int>::iterator
5394 I = ShuffleMask.begin(), E = ShuffleMask.end(); I != E; ++I)
5395 VTBLMask.push_back(DAG.getConstant(*I, MVT::i32));
5397 if (V2.getNode()->getOpcode() == ISD::UNDEF)
5398 return DAG.getNode(ARMISD::VTBL1, DL, MVT::v8i8, V1,
5399 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8, VTBLMask));
5401 return DAG.getNode(ARMISD::VTBL2, DL, MVT::v8i8, V1, V2,
5402 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8, VTBLMask));
5405 static SDValue LowerReverse_VECTOR_SHUFFLEv16i8_v8i16(SDValue Op,
5406 SelectionDAG &DAG) {
5408 SDValue OpLHS = Op.getOperand(0);
5409 EVT VT = OpLHS.getValueType();
5411 assert((VT == MVT::v8i16 || VT == MVT::v16i8) &&
5412 "Expect an v8i16/v16i8 type");
5413 OpLHS = DAG.getNode(ARMISD::VREV64, DL, VT, OpLHS);
5414 // For a v16i8 type: After the VREV, we have got <8, ...15, 8, ..., 0>. Now,
5415 // extract the first 8 bytes into the top double word and the last 8 bytes
5416 // into the bottom double word. The v8i16 case is similar.
5417 unsigned ExtractNum = (VT == MVT::v16i8) ? 8 : 4;
5418 return DAG.getNode(ARMISD::VEXT, DL, VT, OpLHS, OpLHS,
5419 DAG.getConstant(ExtractNum, MVT::i32));
5422 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
5423 SDValue V1 = Op.getOperand(0);
5424 SDValue V2 = Op.getOperand(1);
5426 EVT VT = Op.getValueType();
5427 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
5429 // Convert shuffles that are directly supported on NEON to target-specific
5430 // DAG nodes, instead of keeping them as shuffles and matching them again
5431 // during code selection. This is more efficient and avoids the possibility
5432 // of inconsistencies between legalization and selection.
5433 // FIXME: floating-point vectors should be canonicalized to integer vectors
5434 // of the same time so that they get CSEd properly.
5435 ArrayRef<int> ShuffleMask = SVN->getMask();
5437 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5438 if (EltSize <= 32) {
5439 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
5440 int Lane = SVN->getSplatIndex();
5441 // If this is undef splat, generate it via "just" vdup, if possible.
5442 if (Lane == -1) Lane = 0;
5444 // Test if V1 is a SCALAR_TO_VECTOR.
5445 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
5446 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
5448 // Test if V1 is a BUILD_VECTOR which is equivalent to a SCALAR_TO_VECTOR
5449 // (and probably will turn into a SCALAR_TO_VECTOR once legalization
5451 if (Lane == 0 && V1.getOpcode() == ISD::BUILD_VECTOR &&
5452 !isa<ConstantSDNode>(V1.getOperand(0))) {
5453 bool IsScalarToVector = true;
5454 for (unsigned i = 1, e = V1.getNumOperands(); i != e; ++i)
5455 if (V1.getOperand(i).getOpcode() != ISD::UNDEF) {
5456 IsScalarToVector = false;
5459 if (IsScalarToVector)
5460 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
5462 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
5463 DAG.getConstant(Lane, MVT::i32));
5468 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
5471 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
5472 DAG.getConstant(Imm, MVT::i32));
5475 if (isVREVMask(ShuffleMask, VT, 64))
5476 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
5477 if (isVREVMask(ShuffleMask, VT, 32))
5478 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
5479 if (isVREVMask(ShuffleMask, VT, 16))
5480 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
5482 if (V2->getOpcode() == ISD::UNDEF &&
5483 isSingletonVEXTMask(ShuffleMask, VT, Imm)) {
5484 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V1,
5485 DAG.getConstant(Imm, MVT::i32));
5488 // Check for Neon shuffles that modify both input vectors in place.
5489 // If both results are used, i.e., if there are two shuffles with the same
5490 // source operands and with masks corresponding to both results of one of
5491 // these operations, DAG memoization will ensure that a single node is
5492 // used for both shuffles.
5493 unsigned WhichResult;
5494 if (isVTRNMask(ShuffleMask, VT, WhichResult))
5495 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
5496 V1, V2).getValue(WhichResult);
5497 if (isVUZPMask(ShuffleMask, VT, WhichResult))
5498 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
5499 V1, V2).getValue(WhichResult);
5500 if (isVZIPMask(ShuffleMask, VT, WhichResult))
5501 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
5502 V1, V2).getValue(WhichResult);
5504 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
5505 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
5506 V1, V1).getValue(WhichResult);
5507 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
5508 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
5509 V1, V1).getValue(WhichResult);
5510 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
5511 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
5512 V1, V1).getValue(WhichResult);
5515 // If the shuffle is not directly supported and it has 4 elements, use
5516 // the PerfectShuffle-generated table to synthesize it from other shuffles.
5517 unsigned NumElts = VT.getVectorNumElements();
5519 unsigned PFIndexes[4];
5520 for (unsigned i = 0; i != 4; ++i) {
5521 if (ShuffleMask[i] < 0)
5524 PFIndexes[i] = ShuffleMask[i];
5527 // Compute the index in the perfect shuffle table.
5528 unsigned PFTableIndex =
5529 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
5530 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
5531 unsigned Cost = (PFEntry >> 30);
5534 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
5537 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
5538 if (EltSize >= 32) {
5539 // Do the expansion with floating-point types, since that is what the VFP
5540 // registers are defined to use, and since i64 is not legal.
5541 EVT EltVT = EVT::getFloatingPointVT(EltSize);
5542 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
5543 V1 = DAG.getNode(ISD::BITCAST, dl, VecVT, V1);
5544 V2 = DAG.getNode(ISD::BITCAST, dl, VecVT, V2);
5545 SmallVector<SDValue, 8> Ops;
5546 for (unsigned i = 0; i < NumElts; ++i) {
5547 if (ShuffleMask[i] < 0)
5548 Ops.push_back(DAG.getUNDEF(EltVT));
5550 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
5551 ShuffleMask[i] < (int)NumElts ? V1 : V2,
5552 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
5555 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, Ops);
5556 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
5559 if ((VT == MVT::v8i16 || VT == MVT::v16i8) && isReverseMask(ShuffleMask, VT))
5560 return LowerReverse_VECTOR_SHUFFLEv16i8_v8i16(Op, DAG);
5562 if (VT == MVT::v8i8) {
5563 SDValue NewOp = LowerVECTOR_SHUFFLEv8i8(Op, ShuffleMask, DAG);
5564 if (NewOp.getNode())
5571 static SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
5572 // INSERT_VECTOR_ELT is legal only for immediate indexes.
5573 SDValue Lane = Op.getOperand(2);
5574 if (!isa<ConstantSDNode>(Lane))
5580 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
5581 // EXTRACT_VECTOR_ELT is legal only for immediate indexes.
5582 SDValue Lane = Op.getOperand(1);
5583 if (!isa<ConstantSDNode>(Lane))
5586 SDValue Vec = Op.getOperand(0);
5587 if (Op.getValueType() == MVT::i32 &&
5588 Vec.getValueType().getVectorElementType().getSizeInBits() < 32) {
5590 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
5596 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
5597 // The only time a CONCAT_VECTORS operation can have legal types is when
5598 // two 64-bit vectors are concatenated to a 128-bit vector.
5599 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
5600 "unexpected CONCAT_VECTORS");
5602 SDValue Val = DAG.getUNDEF(MVT::v2f64);
5603 SDValue Op0 = Op.getOperand(0);
5604 SDValue Op1 = Op.getOperand(1);
5605 if (Op0.getOpcode() != ISD::UNDEF)
5606 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
5607 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op0),
5608 DAG.getIntPtrConstant(0));
5609 if (Op1.getOpcode() != ISD::UNDEF)
5610 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
5611 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op1),
5612 DAG.getIntPtrConstant(1));
5613 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Val);
5616 /// isExtendedBUILD_VECTOR - Check if N is a constant BUILD_VECTOR where each
5617 /// element has been zero/sign-extended, depending on the isSigned parameter,
5618 /// from an integer type half its size.
5619 static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG,
5621 // A v2i64 BUILD_VECTOR will have been legalized to a BITCAST from v4i32.
5622 EVT VT = N->getValueType(0);
5623 if (VT == MVT::v2i64 && N->getOpcode() == ISD::BITCAST) {
5624 SDNode *BVN = N->getOperand(0).getNode();
5625 if (BVN->getValueType(0) != MVT::v4i32 ||
5626 BVN->getOpcode() != ISD::BUILD_VECTOR)
5628 unsigned LoElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
5629 unsigned HiElt = 1 - LoElt;
5630 ConstantSDNode *Lo0 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt));
5631 ConstantSDNode *Hi0 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt));
5632 ConstantSDNode *Lo1 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt+2));
5633 ConstantSDNode *Hi1 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt+2));
5634 if (!Lo0 || !Hi0 || !Lo1 || !Hi1)
5637 if (Hi0->getSExtValue() == Lo0->getSExtValue() >> 32 &&
5638 Hi1->getSExtValue() == Lo1->getSExtValue() >> 32)
5641 if (Hi0->isNullValue() && Hi1->isNullValue())
5647 if (N->getOpcode() != ISD::BUILD_VECTOR)
5650 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
5651 SDNode *Elt = N->getOperand(i).getNode();
5652 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
5653 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5654 unsigned HalfSize = EltSize / 2;
5656 if (!isIntN(HalfSize, C->getSExtValue()))
5659 if (!isUIntN(HalfSize, C->getZExtValue()))
5670 /// isSignExtended - Check if a node is a vector value that is sign-extended
5671 /// or a constant BUILD_VECTOR with sign-extended elements.
5672 static bool isSignExtended(SDNode *N, SelectionDAG &DAG) {
5673 if (N->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N))
5675 if (isExtendedBUILD_VECTOR(N, DAG, true))
5680 /// isZeroExtended - Check if a node is a vector value that is zero-extended
5681 /// or a constant BUILD_VECTOR with zero-extended elements.
5682 static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) {
5683 if (N->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N))
5685 if (isExtendedBUILD_VECTOR(N, DAG, false))
5690 static EVT getExtensionTo64Bits(const EVT &OrigVT) {
5691 if (OrigVT.getSizeInBits() >= 64)
5694 assert(OrigVT.isSimple() && "Expecting a simple value type");
5696 MVT::SimpleValueType OrigSimpleTy = OrigVT.getSimpleVT().SimpleTy;
5697 switch (OrigSimpleTy) {
5698 default: llvm_unreachable("Unexpected Vector Type");
5707 /// AddRequiredExtensionForVMULL - Add a sign/zero extension to extend the total
5708 /// value size to 64 bits. We need a 64-bit D register as an operand to VMULL.
5709 /// We insert the required extension here to get the vector to fill a D register.
5710 static SDValue AddRequiredExtensionForVMULL(SDValue N, SelectionDAG &DAG,
5713 unsigned ExtOpcode) {
5714 // The vector originally had a size of OrigTy. It was then extended to ExtTy.
5715 // We expect the ExtTy to be 128-bits total. If the OrigTy is less than
5716 // 64-bits we need to insert a new extension so that it will be 64-bits.
5717 assert(ExtTy.is128BitVector() && "Unexpected extension size");
5718 if (OrigTy.getSizeInBits() >= 64)
5721 // Must extend size to at least 64 bits to be used as an operand for VMULL.
5722 EVT NewVT = getExtensionTo64Bits(OrigTy);
5724 return DAG.getNode(ExtOpcode, SDLoc(N), NewVT, N);
5727 /// SkipLoadExtensionForVMULL - return a load of the original vector size that
5728 /// does not do any sign/zero extension. If the original vector is less
5729 /// than 64 bits, an appropriate extension will be added after the load to
5730 /// reach a total size of 64 bits. We have to add the extension separately
5731 /// because ARM does not have a sign/zero extending load for vectors.
5732 static SDValue SkipLoadExtensionForVMULL(LoadSDNode *LD, SelectionDAG& DAG) {
5733 EVT ExtendedTy = getExtensionTo64Bits(LD->getMemoryVT());
5735 // The load already has the right type.
5736 if (ExtendedTy == LD->getMemoryVT())
5737 return DAG.getLoad(LD->getMemoryVT(), SDLoc(LD), LD->getChain(),
5738 LD->getBasePtr(), LD->getPointerInfo(), LD->isVolatile(),
5739 LD->isNonTemporal(), LD->isInvariant(),
5740 LD->getAlignment());
5742 // We need to create a zextload/sextload. We cannot just create a load
5743 // followed by a zext/zext node because LowerMUL is also run during normal
5744 // operation legalization where we can't create illegal types.
5745 return DAG.getExtLoad(LD->getExtensionType(), SDLoc(LD), ExtendedTy,
5746 LD->getChain(), LD->getBasePtr(), LD->getPointerInfo(),
5747 LD->getMemoryVT(), LD->isVolatile(),
5748 LD->isNonTemporal(), LD->getAlignment());
5751 /// SkipExtensionForVMULL - For a node that is a SIGN_EXTEND, ZERO_EXTEND,
5752 /// extending load, or BUILD_VECTOR with extended elements, return the
5753 /// unextended value. The unextended vector should be 64 bits so that it can
5754 /// be used as an operand to a VMULL instruction. If the original vector size
5755 /// before extension is less than 64 bits we add a an extension to resize
5756 /// the vector to 64 bits.
5757 static SDValue SkipExtensionForVMULL(SDNode *N, SelectionDAG &DAG) {
5758 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
5759 return AddRequiredExtensionForVMULL(N->getOperand(0), DAG,
5760 N->getOperand(0)->getValueType(0),
5764 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
5765 return SkipLoadExtensionForVMULL(LD, DAG);
5767 // Otherwise, the value must be a BUILD_VECTOR. For v2i64, it will
5768 // have been legalized as a BITCAST from v4i32.
5769 if (N->getOpcode() == ISD::BITCAST) {
5770 SDNode *BVN = N->getOperand(0).getNode();
5771 assert(BVN->getOpcode() == ISD::BUILD_VECTOR &&
5772 BVN->getValueType(0) == MVT::v4i32 && "expected v4i32 BUILD_VECTOR");
5773 unsigned LowElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
5774 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), MVT::v2i32,
5775 BVN->getOperand(LowElt), BVN->getOperand(LowElt+2));
5777 // Construct a new BUILD_VECTOR with elements truncated to half the size.
5778 assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR");
5779 EVT VT = N->getValueType(0);
5780 unsigned EltSize = VT.getVectorElementType().getSizeInBits() / 2;
5781 unsigned NumElts = VT.getVectorNumElements();
5782 MVT TruncVT = MVT::getIntegerVT(EltSize);
5783 SmallVector<SDValue, 8> Ops;
5784 for (unsigned i = 0; i != NumElts; ++i) {
5785 ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i));
5786 const APInt &CInt = C->getAPIntValue();
5787 // Element types smaller than 32 bits are not legal, so use i32 elements.
5788 // The values are implicitly truncated so sext vs. zext doesn't matter.
5789 Ops.push_back(DAG.getConstant(CInt.zextOrTrunc(32), MVT::i32));
5791 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N),
5792 MVT::getVectorVT(TruncVT, NumElts), Ops);
5795 static bool isAddSubSExt(SDNode *N, SelectionDAG &DAG) {
5796 unsigned Opcode = N->getOpcode();
5797 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
5798 SDNode *N0 = N->getOperand(0).getNode();
5799 SDNode *N1 = N->getOperand(1).getNode();
5800 return N0->hasOneUse() && N1->hasOneUse() &&
5801 isSignExtended(N0, DAG) && isSignExtended(N1, DAG);
5806 static bool isAddSubZExt(SDNode *N, SelectionDAG &DAG) {
5807 unsigned Opcode = N->getOpcode();
5808 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
5809 SDNode *N0 = N->getOperand(0).getNode();
5810 SDNode *N1 = N->getOperand(1).getNode();
5811 return N0->hasOneUse() && N1->hasOneUse() &&
5812 isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG);
5817 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
5818 // Multiplications are only custom-lowered for 128-bit vectors so that
5819 // VMULL can be detected. Otherwise v2i64 multiplications are not legal.
5820 EVT VT = Op.getValueType();
5821 assert(VT.is128BitVector() && VT.isInteger() &&
5822 "unexpected type for custom-lowering ISD::MUL");
5823 SDNode *N0 = Op.getOperand(0).getNode();
5824 SDNode *N1 = Op.getOperand(1).getNode();
5825 unsigned NewOpc = 0;
5827 bool isN0SExt = isSignExtended(N0, DAG);
5828 bool isN1SExt = isSignExtended(N1, DAG);
5829 if (isN0SExt && isN1SExt)
5830 NewOpc = ARMISD::VMULLs;
5832 bool isN0ZExt = isZeroExtended(N0, DAG);
5833 bool isN1ZExt = isZeroExtended(N1, DAG);
5834 if (isN0ZExt && isN1ZExt)
5835 NewOpc = ARMISD::VMULLu;
5836 else if (isN1SExt || isN1ZExt) {
5837 // Look for (s/zext A + s/zext B) * (s/zext C). We want to turn these
5838 // into (s/zext A * s/zext C) + (s/zext B * s/zext C)
5839 if (isN1SExt && isAddSubSExt(N0, DAG)) {
5840 NewOpc = ARMISD::VMULLs;
5842 } else if (isN1ZExt && isAddSubZExt(N0, DAG)) {
5843 NewOpc = ARMISD::VMULLu;
5845 } else if (isN0ZExt && isAddSubZExt(N1, DAG)) {
5847 NewOpc = ARMISD::VMULLu;
5853 if (VT == MVT::v2i64)
5854 // Fall through to expand this. It is not legal.
5857 // Other vector multiplications are legal.
5862 // Legalize to a VMULL instruction.
5865 SDValue Op1 = SkipExtensionForVMULL(N1, DAG);
5867 Op0 = SkipExtensionForVMULL(N0, DAG);
5868 assert(Op0.getValueType().is64BitVector() &&
5869 Op1.getValueType().is64BitVector() &&
5870 "unexpected types for extended operands to VMULL");
5871 return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
5874 // Optimizing (zext A + zext B) * C, to (VMULL A, C) + (VMULL B, C) during
5875 // isel lowering to take advantage of no-stall back to back vmul + vmla.
5882 SDValue N00 = SkipExtensionForVMULL(N0->getOperand(0).getNode(), DAG);
5883 SDValue N01 = SkipExtensionForVMULL(N0->getOperand(1).getNode(), DAG);
5884 EVT Op1VT = Op1.getValueType();
5885 return DAG.getNode(N0->getOpcode(), DL, VT,
5886 DAG.getNode(NewOpc, DL, VT,
5887 DAG.getNode(ISD::BITCAST, DL, Op1VT, N00), Op1),
5888 DAG.getNode(NewOpc, DL, VT,
5889 DAG.getNode(ISD::BITCAST, DL, Op1VT, N01), Op1));
5893 LowerSDIV_v4i8(SDValue X, SDValue Y, SDLoc dl, SelectionDAG &DAG) {
5895 // float4 xf = vcvt_f32_s32(vmovl_s16(a.lo));
5896 // float4 yf = vcvt_f32_s32(vmovl_s16(b.lo));
5897 X = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, X);
5898 Y = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, Y);
5899 X = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, X);
5900 Y = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, Y);
5901 // Get reciprocal estimate.
5902 // float4 recip = vrecpeq_f32(yf);
5903 Y = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5904 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), Y);
5905 // Because char has a smaller range than uchar, we can actually get away
5906 // without any newton steps. This requires that we use a weird bias
5907 // of 0xb000, however (again, this has been exhaustively tested).
5908 // float4 result = as_float4(as_int4(xf*recip) + 0xb000);
5909 X = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, X, Y);
5910 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, X);
5911 Y = DAG.getConstant(0xb000, MVT::i32);
5912 Y = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Y, Y, Y, Y);
5913 X = DAG.getNode(ISD::ADD, dl, MVT::v4i32, X, Y);
5914 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, X);
5915 // Convert back to short.
5916 X = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, X);
5917 X = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, X);
5922 LowerSDIV_v4i16(SDValue N0, SDValue N1, SDLoc dl, SelectionDAG &DAG) {
5924 // Convert to float.
5925 // float4 yf = vcvt_f32_s32(vmovl_s16(y));
5926 // float4 xf = vcvt_f32_s32(vmovl_s16(x));
5927 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N0);
5928 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N1);
5929 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
5930 N1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
5932 // Use reciprocal estimate and one refinement step.
5933 // float4 recip = vrecpeq_f32(yf);
5934 // recip *= vrecpsq_f32(yf, recip);
5935 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5936 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), N1);
5937 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5938 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
5940 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
5941 // Because short has a smaller range than ushort, we can actually get away
5942 // with only a single newton step. This requires that we use a weird bias
5943 // of 89, however (again, this has been exhaustively tested).
5944 // float4 result = as_float4(as_int4(xf*recip) + 0x89);
5945 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
5946 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
5947 N1 = DAG.getConstant(0x89, MVT::i32);
5948 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
5949 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
5950 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
5951 // Convert back to integer and return.
5952 // return vmovn_s32(vcvt_s32_f32(result));
5953 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
5954 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
5958 static SDValue LowerSDIV(SDValue Op, SelectionDAG &DAG) {
5959 EVT VT = Op.getValueType();
5960 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
5961 "unexpected type for custom-lowering ISD::SDIV");
5964 SDValue N0 = Op.getOperand(0);
5965 SDValue N1 = Op.getOperand(1);
5968 if (VT == MVT::v8i8) {
5969 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N0);
5970 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N1);
5972 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5973 DAG.getIntPtrConstant(4));
5974 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5975 DAG.getIntPtrConstant(4));
5976 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5977 DAG.getIntPtrConstant(0));
5978 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5979 DAG.getIntPtrConstant(0));
5981 N0 = LowerSDIV_v4i8(N0, N1, dl, DAG); // v4i16
5982 N2 = LowerSDIV_v4i8(N2, N3, dl, DAG); // v4i16
5984 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
5985 N0 = LowerCONCAT_VECTORS(N0, DAG);
5987 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v8i8, N0);
5990 return LowerSDIV_v4i16(N0, N1, dl, DAG);
5993 static SDValue LowerUDIV(SDValue Op, SelectionDAG &DAG) {
5994 EVT VT = Op.getValueType();
5995 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
5996 "unexpected type for custom-lowering ISD::UDIV");
5999 SDValue N0 = Op.getOperand(0);
6000 SDValue N1 = Op.getOperand(1);
6003 if (VT == MVT::v8i8) {
6004 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N0);
6005 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N1);
6007 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
6008 DAG.getIntPtrConstant(4));
6009 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
6010 DAG.getIntPtrConstant(4));
6011 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
6012 DAG.getIntPtrConstant(0));
6013 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
6014 DAG.getIntPtrConstant(0));
6016 N0 = LowerSDIV_v4i16(N0, N1, dl, DAG); // v4i16
6017 N2 = LowerSDIV_v4i16(N2, N3, dl, DAG); // v4i16
6019 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
6020 N0 = LowerCONCAT_VECTORS(N0, DAG);
6022 N0 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v8i8,
6023 DAG.getConstant(Intrinsic::arm_neon_vqmovnsu, MVT::i32),
6028 // v4i16 sdiv ... Convert to float.
6029 // float4 yf = vcvt_f32_s32(vmovl_u16(y));
6030 // float4 xf = vcvt_f32_s32(vmovl_u16(x));
6031 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N0);
6032 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N1);
6033 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
6034 SDValue BN1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
6036 // Use reciprocal estimate and two refinement steps.
6037 // float4 recip = vrecpeq_f32(yf);
6038 // recip *= vrecpsq_f32(yf, recip);
6039 // recip *= vrecpsq_f32(yf, recip);
6040 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6041 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), BN1);
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 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6047 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
6049 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
6050 // Simply multiplying by the reciprocal estimate can leave us a few ulps
6051 // too low, so we add 2 ulps (exhaustive testing shows that this is enough,
6052 // and that it will never cause us to return an answer too large).
6053 // float4 result = as_float4(as_int4(xf*recip) + 2);
6054 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
6055 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
6056 N1 = DAG.getConstant(2, MVT::i32);
6057 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
6058 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
6059 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
6060 // Convert back to integer and return.
6061 // return vmovn_u32(vcvt_s32_f32(result));
6062 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
6063 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
6067 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
6068 EVT VT = Op.getNode()->getValueType(0);
6069 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
6072 bool ExtraOp = false;
6073 switch (Op.getOpcode()) {
6074 default: llvm_unreachable("Invalid code");
6075 case ISD::ADDC: Opc = ARMISD::ADDC; break;
6076 case ISD::ADDE: Opc = ARMISD::ADDE; ExtraOp = true; break;
6077 case ISD::SUBC: Opc = ARMISD::SUBC; break;
6078 case ISD::SUBE: Opc = ARMISD::SUBE; ExtraOp = true; break;
6082 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
6084 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
6085 Op.getOperand(1), Op.getOperand(2));
6088 SDValue ARMTargetLowering::LowerFSINCOS(SDValue Op, SelectionDAG &DAG) const {
6089 assert(Subtarget->isTargetDarwin());
6091 // For iOS, we want to call an alternative entry point: __sincos_stret,
6092 // return values are passed via sret.
6094 SDValue Arg = Op.getOperand(0);
6095 EVT ArgVT = Arg.getValueType();
6096 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
6098 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
6099 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6101 // Pair of floats / doubles used to pass the result.
6102 StructType *RetTy = StructType::get(ArgTy, ArgTy, NULL);
6104 // Create stack object for sret.
6105 const uint64_t ByteSize = TLI.getDataLayout()->getTypeAllocSize(RetTy);
6106 const unsigned StackAlign = TLI.getDataLayout()->getPrefTypeAlignment(RetTy);
6107 int FrameIdx = FrameInfo->CreateStackObject(ByteSize, StackAlign, false);
6108 SDValue SRet = DAG.getFrameIndex(FrameIdx, TLI.getPointerTy());
6114 Entry.Ty = RetTy->getPointerTo();
6115 Entry.isSExt = false;
6116 Entry.isZExt = false;
6117 Entry.isSRet = true;
6118 Args.push_back(Entry);
6122 Entry.isSExt = false;
6123 Entry.isZExt = false;
6124 Args.push_back(Entry);
6126 const char *LibcallName = (ArgVT == MVT::f64)
6127 ? "__sincos_stret" : "__sincosf_stret";
6128 SDValue Callee = DAG.getExternalSymbol(LibcallName, getPointerTy());
6130 TargetLowering::CallLoweringInfo CLI(DAG);
6131 CLI.setDebugLoc(dl).setChain(DAG.getEntryNode())
6132 .setCallee(CallingConv::C, Type::getVoidTy(*DAG.getContext()), Callee,
6134 .setDiscardResult();
6136 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
6138 SDValue LoadSin = DAG.getLoad(ArgVT, dl, CallResult.second, SRet,
6139 MachinePointerInfo(), false, false, false, 0);
6141 // Address of cos field.
6142 SDValue Add = DAG.getNode(ISD::ADD, dl, getPointerTy(), SRet,
6143 DAG.getIntPtrConstant(ArgVT.getStoreSize()));
6144 SDValue LoadCos = DAG.getLoad(ArgVT, dl, LoadSin.getValue(1), Add,
6145 MachinePointerInfo(), false, false, false, 0);
6147 SDVTList Tys = DAG.getVTList(ArgVT, ArgVT);
6148 return DAG.getNode(ISD::MERGE_VALUES, dl, Tys,
6149 LoadSin.getValue(0), LoadCos.getValue(0));
6152 static SDValue LowerAtomicLoadStore(SDValue Op, SelectionDAG &DAG) {
6153 // Monotonic load/store is legal for all targets
6154 if (cast<AtomicSDNode>(Op)->getOrdering() <= Monotonic)
6157 // Acquire/Release load/store is not legal for targets without a
6158 // dmb or equivalent available.
6162 static void ReplaceREADCYCLECOUNTER(SDNode *N,
6163 SmallVectorImpl<SDValue> &Results,
6165 const ARMSubtarget *Subtarget) {
6167 SDValue Cycles32, OutChain;
6169 if (Subtarget->hasPerfMon()) {
6170 // Under Power Management extensions, the cycle-count is:
6171 // mrc p15, #0, <Rt>, c9, c13, #0
6172 SDValue Ops[] = { N->getOperand(0), // Chain
6173 DAG.getConstant(Intrinsic::arm_mrc, MVT::i32),
6174 DAG.getConstant(15, MVT::i32),
6175 DAG.getConstant(0, MVT::i32),
6176 DAG.getConstant(9, MVT::i32),
6177 DAG.getConstant(13, MVT::i32),
6178 DAG.getConstant(0, MVT::i32)
6181 Cycles32 = DAG.getNode(ISD::INTRINSIC_W_CHAIN, DL,
6182 DAG.getVTList(MVT::i32, MVT::Other), Ops);
6183 OutChain = Cycles32.getValue(1);
6185 // Intrinsic is defined to return 0 on unsupported platforms. Technically
6186 // there are older ARM CPUs that have implementation-specific ways of
6187 // obtaining this information (FIXME!).
6188 Cycles32 = DAG.getConstant(0, MVT::i32);
6189 OutChain = DAG.getEntryNode();
6193 SDValue Cycles64 = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64,
6194 Cycles32, DAG.getConstant(0, MVT::i32));
6195 Results.push_back(Cycles64);
6196 Results.push_back(OutChain);
6199 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
6200 switch (Op.getOpcode()) {
6201 default: llvm_unreachable("Don't know how to custom lower this!");
6202 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
6203 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
6204 case ISD::GlobalAddress:
6205 switch (Subtarget->getTargetTriple().getObjectFormat()) {
6206 default: llvm_unreachable("unknown object format");
6208 return LowerGlobalAddressWindows(Op, DAG);
6210 return LowerGlobalAddressELF(Op, DAG);
6212 return LowerGlobalAddressDarwin(Op, DAG);
6214 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
6215 case ISD::SELECT: return LowerSELECT(Op, DAG);
6216 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
6217 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
6218 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
6219 case ISD::VASTART: return LowerVASTART(Op, DAG);
6220 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, DAG, Subtarget);
6221 case ISD::PREFETCH: return LowerPREFETCH(Op, DAG, Subtarget);
6222 case ISD::SINT_TO_FP:
6223 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
6224 case ISD::FP_TO_SINT:
6225 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
6226 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
6227 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
6228 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
6229 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
6230 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
6231 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
6232 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
6234 case ISD::BITCAST: return ExpandBITCAST(Op.getNode(), DAG);
6237 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
6238 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
6239 case ISD::SRL_PARTS:
6240 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
6241 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
6242 case ISD::CTPOP: return LowerCTPOP(Op.getNode(), DAG, Subtarget);
6243 case ISD::SETCC: return LowerVSETCC(Op, DAG);
6244 case ISD::ConstantFP: return LowerConstantFP(Op, DAG, Subtarget);
6245 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG, Subtarget);
6246 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
6247 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
6248 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
6249 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
6250 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
6251 case ISD::MUL: return LowerMUL(Op, DAG);
6252 case ISD::SDIV: return LowerSDIV(Op, DAG);
6253 case ISD::UDIV: return LowerUDIV(Op, DAG);
6257 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
6262 return LowerXALUO(Op, DAG);
6263 case ISD::ATOMIC_LOAD:
6264 case ISD::ATOMIC_STORE: return LowerAtomicLoadStore(Op, DAG);
6265 case ISD::FSINCOS: return LowerFSINCOS(Op, DAG);
6267 case ISD::UDIVREM: return LowerDivRem(Op, DAG);
6268 case ISD::DYNAMIC_STACKALLOC:
6269 if (Subtarget->getTargetTriple().isWindowsItaniumEnvironment())
6270 return LowerDYNAMIC_STACKALLOC(Op, DAG);
6271 llvm_unreachable("Don't know how to custom lower this!");
6275 /// ReplaceNodeResults - Replace the results of node with an illegal result
6276 /// type with new values built out of custom code.
6277 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
6278 SmallVectorImpl<SDValue>&Results,
6279 SelectionDAG &DAG) const {
6281 switch (N->getOpcode()) {
6283 llvm_unreachable("Don't know how to custom expand this!");
6285 Res = ExpandBITCAST(N, DAG);
6289 Res = Expand64BitShift(N, DAG, Subtarget);
6291 case ISD::READCYCLECOUNTER:
6292 ReplaceREADCYCLECOUNTER(N, Results, DAG, Subtarget);
6296 Results.push_back(Res);
6299 //===----------------------------------------------------------------------===//
6300 // ARM Scheduler Hooks
6301 //===----------------------------------------------------------------------===//
6303 /// SetupEntryBlockForSjLj - Insert code into the entry block that creates and
6304 /// registers the function context.
6305 void ARMTargetLowering::
6306 SetupEntryBlockForSjLj(MachineInstr *MI, MachineBasicBlock *MBB,
6307 MachineBasicBlock *DispatchBB, int FI) const {
6308 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6309 DebugLoc dl = MI->getDebugLoc();
6310 MachineFunction *MF = MBB->getParent();
6311 MachineRegisterInfo *MRI = &MF->getRegInfo();
6312 MachineConstantPool *MCP = MF->getConstantPool();
6313 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
6314 const Function *F = MF->getFunction();
6316 bool isThumb = Subtarget->isThumb();
6317 bool isThumb2 = Subtarget->isThumb2();
6319 unsigned PCLabelId = AFI->createPICLabelUId();
6320 unsigned PCAdj = (isThumb || isThumb2) ? 4 : 8;
6321 ARMConstantPoolValue *CPV =
6322 ARMConstantPoolMBB::Create(F->getContext(), DispatchBB, PCLabelId, PCAdj);
6323 unsigned CPI = MCP->getConstantPoolIndex(CPV, 4);
6325 const TargetRegisterClass *TRC = isThumb ?
6326 (const TargetRegisterClass*)&ARM::tGPRRegClass :
6327 (const TargetRegisterClass*)&ARM::GPRRegClass;
6329 // Grab constant pool and fixed stack memory operands.
6330 MachineMemOperand *CPMMO =
6331 MF->getMachineMemOperand(MachinePointerInfo::getConstantPool(),
6332 MachineMemOperand::MOLoad, 4, 4);
6334 MachineMemOperand *FIMMOSt =
6335 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
6336 MachineMemOperand::MOStore, 4, 4);
6338 // Load the address of the dispatch MBB into the jump buffer.
6340 // Incoming value: jbuf
6341 // ldr.n r5, LCPI1_1
6344 // str r5, [$jbuf, #+4] ; &jbuf[1]
6345 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6346 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2LDRpci), NewVReg1)
6347 .addConstantPoolIndex(CPI)
6348 .addMemOperand(CPMMO));
6349 // Set the low bit because of thumb mode.
6350 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6352 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2ORRri), NewVReg2)
6353 .addReg(NewVReg1, RegState::Kill)
6355 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6356 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg3)
6357 .addReg(NewVReg2, RegState::Kill)
6359 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2STRi12))
6360 .addReg(NewVReg3, RegState::Kill)
6362 .addImm(36) // &jbuf[1] :: pc
6363 .addMemOperand(FIMMOSt));
6364 } else if (isThumb) {
6365 // Incoming value: jbuf
6366 // ldr.n r1, LCPI1_4
6370 // add r2, $jbuf, #+4 ; &jbuf[1]
6372 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6373 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tLDRpci), NewVReg1)
6374 .addConstantPoolIndex(CPI)
6375 .addMemOperand(CPMMO));
6376 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6377 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg2)
6378 .addReg(NewVReg1, RegState::Kill)
6380 // Set the low bit because of thumb mode.
6381 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6382 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tMOVi8), NewVReg3)
6383 .addReg(ARM::CPSR, RegState::Define)
6385 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6386 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tORR), NewVReg4)
6387 .addReg(ARM::CPSR, RegState::Define)
6388 .addReg(NewVReg2, RegState::Kill)
6389 .addReg(NewVReg3, RegState::Kill));
6390 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6391 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tADDrSPi), NewVReg5)
6393 .addImm(36)); // &jbuf[1] :: pc
6394 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tSTRi))
6395 .addReg(NewVReg4, RegState::Kill)
6396 .addReg(NewVReg5, RegState::Kill)
6398 .addMemOperand(FIMMOSt));
6400 // Incoming value: jbuf
6403 // str r1, [$jbuf, #+4] ; &jbuf[1]
6404 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6405 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::LDRi12), NewVReg1)
6406 .addConstantPoolIndex(CPI)
6408 .addMemOperand(CPMMO));
6409 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6410 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::PICADD), NewVReg2)
6411 .addReg(NewVReg1, RegState::Kill)
6412 .addImm(PCLabelId));
6413 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::STRi12))
6414 .addReg(NewVReg2, RegState::Kill)
6416 .addImm(36) // &jbuf[1] :: pc
6417 .addMemOperand(FIMMOSt));
6421 MachineBasicBlock *ARMTargetLowering::
6422 EmitSjLjDispatchBlock(MachineInstr *MI, MachineBasicBlock *MBB) const {
6423 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6424 DebugLoc dl = MI->getDebugLoc();
6425 MachineFunction *MF = MBB->getParent();
6426 MachineRegisterInfo *MRI = &MF->getRegInfo();
6427 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
6428 MachineFrameInfo *MFI = MF->getFrameInfo();
6429 int FI = MFI->getFunctionContextIndex();
6431 const TargetRegisterClass *TRC = Subtarget->isThumb() ?
6432 (const TargetRegisterClass*)&ARM::tGPRRegClass :
6433 (const TargetRegisterClass*)&ARM::GPRnopcRegClass;
6435 // Get a mapping of the call site numbers to all of the landing pads they're
6437 DenseMap<unsigned, SmallVector<MachineBasicBlock*, 2> > CallSiteNumToLPad;
6438 unsigned MaxCSNum = 0;
6439 MachineModuleInfo &MMI = MF->getMMI();
6440 for (MachineFunction::iterator BB = MF->begin(), E = MF->end(); BB != E;
6442 if (!BB->isLandingPad()) continue;
6444 // FIXME: We should assert that the EH_LABEL is the first MI in the landing
6446 for (MachineBasicBlock::iterator
6447 II = BB->begin(), IE = BB->end(); II != IE; ++II) {
6448 if (!II->isEHLabel()) continue;
6450 MCSymbol *Sym = II->getOperand(0).getMCSymbol();
6451 if (!MMI.hasCallSiteLandingPad(Sym)) continue;
6453 SmallVectorImpl<unsigned> &CallSiteIdxs = MMI.getCallSiteLandingPad(Sym);
6454 for (SmallVectorImpl<unsigned>::iterator
6455 CSI = CallSiteIdxs.begin(), CSE = CallSiteIdxs.end();
6456 CSI != CSE; ++CSI) {
6457 CallSiteNumToLPad[*CSI].push_back(BB);
6458 MaxCSNum = std::max(MaxCSNum, *CSI);
6464 // Get an ordered list of the machine basic blocks for the jump table.
6465 std::vector<MachineBasicBlock*> LPadList;
6466 SmallPtrSet<MachineBasicBlock*, 64> InvokeBBs;
6467 LPadList.reserve(CallSiteNumToLPad.size());
6468 for (unsigned I = 1; I <= MaxCSNum; ++I) {
6469 SmallVectorImpl<MachineBasicBlock*> &MBBList = CallSiteNumToLPad[I];
6470 for (SmallVectorImpl<MachineBasicBlock*>::iterator
6471 II = MBBList.begin(), IE = MBBList.end(); II != IE; ++II) {
6472 LPadList.push_back(*II);
6473 InvokeBBs.insert((*II)->pred_begin(), (*II)->pred_end());
6477 assert(!LPadList.empty() &&
6478 "No landing pad destinations for the dispatch jump table!");
6480 // Create the jump table and associated information.
6481 MachineJumpTableInfo *JTI =
6482 MF->getOrCreateJumpTableInfo(MachineJumpTableInfo::EK_Inline);
6483 unsigned MJTI = JTI->createJumpTableIndex(LPadList);
6484 unsigned UId = AFI->createJumpTableUId();
6485 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
6487 // Create the MBBs for the dispatch code.
6489 // Shove the dispatch's address into the return slot in the function context.
6490 MachineBasicBlock *DispatchBB = MF->CreateMachineBasicBlock();
6491 DispatchBB->setIsLandingPad();
6493 MachineBasicBlock *TrapBB = MF->CreateMachineBasicBlock();
6494 unsigned trap_opcode;
6495 if (Subtarget->isThumb())
6496 trap_opcode = ARM::tTRAP;
6498 trap_opcode = Subtarget->useNaClTrap() ? ARM::TRAPNaCl : ARM::TRAP;
6500 BuildMI(TrapBB, dl, TII->get(trap_opcode));
6501 DispatchBB->addSuccessor(TrapBB);
6503 MachineBasicBlock *DispContBB = MF->CreateMachineBasicBlock();
6504 DispatchBB->addSuccessor(DispContBB);
6507 MF->insert(MF->end(), DispatchBB);
6508 MF->insert(MF->end(), DispContBB);
6509 MF->insert(MF->end(), TrapBB);
6511 // Insert code into the entry block that creates and registers the function
6513 SetupEntryBlockForSjLj(MI, MBB, DispatchBB, FI);
6515 MachineMemOperand *FIMMOLd =
6516 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
6517 MachineMemOperand::MOLoad |
6518 MachineMemOperand::MOVolatile, 4, 4);
6520 MachineInstrBuilder MIB;
6521 MIB = BuildMI(DispatchBB, dl, TII->get(ARM::Int_eh_sjlj_dispatchsetup));
6523 const ARMBaseInstrInfo *AII = static_cast<const ARMBaseInstrInfo*>(TII);
6524 const ARMBaseRegisterInfo &RI = AII->getRegisterInfo();
6526 // Add a register mask with no preserved registers. This results in all
6527 // registers being marked as clobbered.
6528 MIB.addRegMask(RI.getNoPreservedMask());
6530 unsigned NumLPads = LPadList.size();
6531 if (Subtarget->isThumb2()) {
6532 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6533 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2LDRi12), NewVReg1)
6536 .addMemOperand(FIMMOLd));
6538 if (NumLPads < 256) {
6539 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPri))
6541 .addImm(LPadList.size()));
6543 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6544 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVi16), VReg1)
6545 .addImm(NumLPads & 0xFFFF));
6547 unsigned VReg2 = VReg1;
6548 if ((NumLPads & 0xFFFF0000) != 0) {
6549 VReg2 = MRI->createVirtualRegister(TRC);
6550 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVTi16), VReg2)
6552 .addImm(NumLPads >> 16));
6555 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPrr))
6560 BuildMI(DispatchBB, dl, TII->get(ARM::t2Bcc))
6565 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6566 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::t2LEApcrelJT),NewVReg3)
6567 .addJumpTableIndex(MJTI)
6570 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6573 BuildMI(DispContBB, dl, TII->get(ARM::t2ADDrs), NewVReg4)
6574 .addReg(NewVReg3, RegState::Kill)
6576 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
6578 BuildMI(DispContBB, dl, TII->get(ARM::t2BR_JT))
6579 .addReg(NewVReg4, RegState::Kill)
6581 .addJumpTableIndex(MJTI)
6583 } else if (Subtarget->isThumb()) {
6584 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6585 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRspi), NewVReg1)
6588 .addMemOperand(FIMMOLd));
6590 if (NumLPads < 256) {
6591 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPi8))
6595 MachineConstantPool *ConstantPool = MF->getConstantPool();
6596 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
6597 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
6599 // MachineConstantPool wants an explicit alignment.
6600 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
6602 Align = getDataLayout()->getTypeAllocSize(C->getType());
6603 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6605 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6606 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRpci))
6607 .addReg(VReg1, RegState::Define)
6608 .addConstantPoolIndex(Idx));
6609 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPr))
6614 BuildMI(DispatchBB, dl, TII->get(ARM::tBcc))
6619 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6620 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLSLri), NewVReg2)
6621 .addReg(ARM::CPSR, RegState::Define)
6625 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6626 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLEApcrelJT), NewVReg3)
6627 .addJumpTableIndex(MJTI)
6630 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6631 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg4)
6632 .addReg(ARM::CPSR, RegState::Define)
6633 .addReg(NewVReg2, RegState::Kill)
6636 MachineMemOperand *JTMMOLd =
6637 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
6638 MachineMemOperand::MOLoad, 4, 4);
6640 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6641 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLDRi), NewVReg5)
6642 .addReg(NewVReg4, RegState::Kill)
6644 .addMemOperand(JTMMOLd));
6646 unsigned NewVReg6 = NewVReg5;
6647 if (RelocM == Reloc::PIC_) {
6648 NewVReg6 = MRI->createVirtualRegister(TRC);
6649 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg6)
6650 .addReg(ARM::CPSR, RegState::Define)
6651 .addReg(NewVReg5, RegState::Kill)
6655 BuildMI(DispContBB, dl, TII->get(ARM::tBR_JTr))
6656 .addReg(NewVReg6, RegState::Kill)
6657 .addJumpTableIndex(MJTI)
6660 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6661 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRi12), NewVReg1)
6664 .addMemOperand(FIMMOLd));
6666 if (NumLPads < 256) {
6667 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPri))
6670 } else if (Subtarget->hasV6T2Ops() && isUInt<16>(NumLPads)) {
6671 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6672 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVi16), VReg1)
6673 .addImm(NumLPads & 0xFFFF));
6675 unsigned VReg2 = VReg1;
6676 if ((NumLPads & 0xFFFF0000) != 0) {
6677 VReg2 = MRI->createVirtualRegister(TRC);
6678 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVTi16), VReg2)
6680 .addImm(NumLPads >> 16));
6683 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
6687 MachineConstantPool *ConstantPool = MF->getConstantPool();
6688 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
6689 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
6691 // MachineConstantPool wants an explicit alignment.
6692 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
6694 Align = getDataLayout()->getTypeAllocSize(C->getType());
6695 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6697 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6698 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRcp))
6699 .addReg(VReg1, RegState::Define)
6700 .addConstantPoolIndex(Idx)
6702 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
6704 .addReg(VReg1, RegState::Kill));
6707 BuildMI(DispatchBB, dl, TII->get(ARM::Bcc))
6712 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6714 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::MOVsi), NewVReg3)
6716 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
6717 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6718 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::LEApcrelJT), NewVReg4)
6719 .addJumpTableIndex(MJTI)
6722 MachineMemOperand *JTMMOLd =
6723 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
6724 MachineMemOperand::MOLoad, 4, 4);
6725 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6727 BuildMI(DispContBB, dl, TII->get(ARM::LDRrs), NewVReg5)
6728 .addReg(NewVReg3, RegState::Kill)
6731 .addMemOperand(JTMMOLd));
6733 if (RelocM == Reloc::PIC_) {
6734 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTadd))
6735 .addReg(NewVReg5, RegState::Kill)
6737 .addJumpTableIndex(MJTI)
6740 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTr))
6741 .addReg(NewVReg5, RegState::Kill)
6742 .addJumpTableIndex(MJTI)
6747 // Add the jump table entries as successors to the MBB.
6748 SmallPtrSet<MachineBasicBlock*, 8> SeenMBBs;
6749 for (std::vector<MachineBasicBlock*>::iterator
6750 I = LPadList.begin(), E = LPadList.end(); I != E; ++I) {
6751 MachineBasicBlock *CurMBB = *I;
6752 if (SeenMBBs.insert(CurMBB))
6753 DispContBB->addSuccessor(CurMBB);
6756 // N.B. the order the invoke BBs are processed in doesn't matter here.
6757 const MCPhysReg *SavedRegs = RI.getCalleeSavedRegs(MF);
6758 SmallVector<MachineBasicBlock*, 64> MBBLPads;
6759 for (SmallPtrSet<MachineBasicBlock*, 64>::iterator
6760 I = InvokeBBs.begin(), E = InvokeBBs.end(); I != E; ++I) {
6761 MachineBasicBlock *BB = *I;
6763 // Remove the landing pad successor from the invoke block and replace it
6764 // with the new dispatch block.
6765 SmallVector<MachineBasicBlock*, 4> Successors(BB->succ_begin(),
6767 while (!Successors.empty()) {
6768 MachineBasicBlock *SMBB = Successors.pop_back_val();
6769 if (SMBB->isLandingPad()) {
6770 BB->removeSuccessor(SMBB);
6771 MBBLPads.push_back(SMBB);
6775 BB->addSuccessor(DispatchBB);
6777 // Find the invoke call and mark all of the callee-saved registers as
6778 // 'implicit defined' so that they're spilled. This prevents code from
6779 // moving instructions to before the EH block, where they will never be
6781 for (MachineBasicBlock::reverse_iterator
6782 II = BB->rbegin(), IE = BB->rend(); II != IE; ++II) {
6783 if (!II->isCall()) continue;
6785 DenseMap<unsigned, bool> DefRegs;
6786 for (MachineInstr::mop_iterator
6787 OI = II->operands_begin(), OE = II->operands_end();
6789 if (!OI->isReg()) continue;
6790 DefRegs[OI->getReg()] = true;
6793 MachineInstrBuilder MIB(*MF, &*II);
6795 for (unsigned i = 0; SavedRegs[i] != 0; ++i) {
6796 unsigned Reg = SavedRegs[i];
6797 if (Subtarget->isThumb2() &&
6798 !ARM::tGPRRegClass.contains(Reg) &&
6799 !ARM::hGPRRegClass.contains(Reg))
6801 if (Subtarget->isThumb1Only() && !ARM::tGPRRegClass.contains(Reg))
6803 if (!Subtarget->isThumb() && !ARM::GPRRegClass.contains(Reg))
6806 MIB.addReg(Reg, RegState::ImplicitDefine | RegState::Dead);
6813 // Mark all former landing pads as non-landing pads. The dispatch is the only
6815 for (SmallVectorImpl<MachineBasicBlock*>::iterator
6816 I = MBBLPads.begin(), E = MBBLPads.end(); I != E; ++I)
6817 (*I)->setIsLandingPad(false);
6819 // The instruction is gone now.
6820 MI->eraseFromParent();
6826 MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) {
6827 for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
6828 E = MBB->succ_end(); I != E; ++I)
6831 llvm_unreachable("Expecting a BB with two successors!");
6834 /// Return the load opcode for a given load size. If load size >= 8,
6835 /// neon opcode will be returned.
6836 static unsigned getLdOpcode(unsigned LdSize, bool IsThumb1, bool IsThumb2) {
6838 return LdSize == 16 ? ARM::VLD1q32wb_fixed
6839 : LdSize == 8 ? ARM::VLD1d32wb_fixed : 0;
6841 return LdSize == 4 ? ARM::tLDRi
6842 : LdSize == 2 ? ARM::tLDRHi
6843 : LdSize == 1 ? ARM::tLDRBi : 0;
6845 return LdSize == 4 ? ARM::t2LDR_POST
6846 : LdSize == 2 ? ARM::t2LDRH_POST
6847 : LdSize == 1 ? ARM::t2LDRB_POST : 0;
6848 return LdSize == 4 ? ARM::LDR_POST_IMM
6849 : LdSize == 2 ? ARM::LDRH_POST
6850 : LdSize == 1 ? ARM::LDRB_POST_IMM : 0;
6853 /// Return the store opcode for a given store size. If store size >= 8,
6854 /// neon opcode will be returned.
6855 static unsigned getStOpcode(unsigned StSize, bool IsThumb1, bool IsThumb2) {
6857 return StSize == 16 ? ARM::VST1q32wb_fixed
6858 : StSize == 8 ? ARM::VST1d32wb_fixed : 0;
6860 return StSize == 4 ? ARM::tSTRi
6861 : StSize == 2 ? ARM::tSTRHi
6862 : StSize == 1 ? ARM::tSTRBi : 0;
6864 return StSize == 4 ? ARM::t2STR_POST
6865 : StSize == 2 ? ARM::t2STRH_POST
6866 : StSize == 1 ? ARM::t2STRB_POST : 0;
6867 return StSize == 4 ? ARM::STR_POST_IMM
6868 : StSize == 2 ? ARM::STRH_POST
6869 : StSize == 1 ? ARM::STRB_POST_IMM : 0;
6872 /// Emit a post-increment load operation with given size. The instructions
6873 /// will be added to BB at Pos.
6874 static void emitPostLd(MachineBasicBlock *BB, MachineInstr *Pos,
6875 const TargetInstrInfo *TII, DebugLoc dl,
6876 unsigned LdSize, unsigned Data, unsigned AddrIn,
6877 unsigned AddrOut, bool IsThumb1, bool IsThumb2) {
6878 unsigned LdOpc = getLdOpcode(LdSize, IsThumb1, IsThumb2);
6879 assert(LdOpc != 0 && "Should have a load opcode");
6881 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
6882 .addReg(AddrOut, RegState::Define).addReg(AddrIn)
6884 } else if (IsThumb1) {
6885 // load + update AddrIn
6886 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
6887 .addReg(AddrIn).addImm(0));
6888 MachineInstrBuilder MIB =
6889 BuildMI(*BB, Pos, dl, TII->get(ARM::tADDi8), AddrOut);
6890 MIB = AddDefaultT1CC(MIB);
6891 MIB.addReg(AddrIn).addImm(LdSize);
6892 AddDefaultPred(MIB);
6893 } else if (IsThumb2) {
6894 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
6895 .addReg(AddrOut, RegState::Define).addReg(AddrIn)
6898 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
6899 .addReg(AddrOut, RegState::Define).addReg(AddrIn)
6900 .addReg(0).addImm(LdSize));
6904 /// Emit a post-increment store operation with given size. The instructions
6905 /// will be added to BB at Pos.
6906 static void emitPostSt(MachineBasicBlock *BB, MachineInstr *Pos,
6907 const TargetInstrInfo *TII, DebugLoc dl,
6908 unsigned StSize, unsigned Data, unsigned AddrIn,
6909 unsigned AddrOut, bool IsThumb1, bool IsThumb2) {
6910 unsigned StOpc = getStOpcode(StSize, IsThumb1, IsThumb2);
6911 assert(StOpc != 0 && "Should have a store opcode");
6913 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc), AddrOut)
6914 .addReg(AddrIn).addImm(0).addReg(Data));
6915 } else if (IsThumb1) {
6916 // store + update AddrIn
6917 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc)).addReg(Data)
6918 .addReg(AddrIn).addImm(0));
6919 MachineInstrBuilder MIB =
6920 BuildMI(*BB, Pos, dl, TII->get(ARM::tADDi8), AddrOut);
6921 MIB = AddDefaultT1CC(MIB);
6922 MIB.addReg(AddrIn).addImm(StSize);
6923 AddDefaultPred(MIB);
6924 } else if (IsThumb2) {
6925 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc), AddrOut)
6926 .addReg(Data).addReg(AddrIn).addImm(StSize));
6928 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc), AddrOut)
6929 .addReg(Data).addReg(AddrIn).addReg(0)
6935 ARMTargetLowering::EmitStructByval(MachineInstr *MI,
6936 MachineBasicBlock *BB) const {
6937 // This pseudo instruction has 3 operands: dst, src, size
6938 // We expand it to a loop if size > Subtarget->getMaxInlineSizeThreshold().
6939 // Otherwise, we will generate unrolled scalar copies.
6940 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6941 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6942 MachineFunction::iterator It = BB;
6945 unsigned dest = MI->getOperand(0).getReg();
6946 unsigned src = MI->getOperand(1).getReg();
6947 unsigned SizeVal = MI->getOperand(2).getImm();
6948 unsigned Align = MI->getOperand(3).getImm();
6949 DebugLoc dl = MI->getDebugLoc();
6951 MachineFunction *MF = BB->getParent();
6952 MachineRegisterInfo &MRI = MF->getRegInfo();
6953 unsigned UnitSize = 0;
6954 const TargetRegisterClass *TRC = nullptr;
6955 const TargetRegisterClass *VecTRC = nullptr;
6957 bool IsThumb1 = Subtarget->isThumb1Only();
6958 bool IsThumb2 = Subtarget->isThumb2();
6962 } else if (Align & 2) {
6965 // Check whether we can use NEON instructions.
6966 if (!MF->getFunction()->getAttributes().
6967 hasAttribute(AttributeSet::FunctionIndex,
6968 Attribute::NoImplicitFloat) &&
6969 Subtarget->hasNEON()) {
6970 if ((Align % 16 == 0) && SizeVal >= 16)
6972 else if ((Align % 8 == 0) && SizeVal >= 8)
6975 // Can't use NEON instructions.
6980 // Select the correct opcode and register class for unit size load/store
6981 bool IsNeon = UnitSize >= 8;
6982 TRC = (IsThumb1 || IsThumb2) ? (const TargetRegisterClass *)&ARM::tGPRRegClass
6983 : (const TargetRegisterClass *)&ARM::GPRRegClass;
6985 VecTRC = UnitSize == 16
6986 ? (const TargetRegisterClass *)&ARM::DPairRegClass
6988 ? (const TargetRegisterClass *)&ARM::DPRRegClass
6991 unsigned BytesLeft = SizeVal % UnitSize;
6992 unsigned LoopSize = SizeVal - BytesLeft;
6994 if (SizeVal <= Subtarget->getMaxInlineSizeThreshold()) {
6995 // Use LDR and STR to copy.
6996 // [scratch, srcOut] = LDR_POST(srcIn, UnitSize)
6997 // [destOut] = STR_POST(scratch, destIn, UnitSize)
6998 unsigned srcIn = src;
6999 unsigned destIn = dest;
7000 for (unsigned i = 0; i < LoopSize; i+=UnitSize) {
7001 unsigned srcOut = MRI.createVirtualRegister(TRC);
7002 unsigned destOut = MRI.createVirtualRegister(TRC);
7003 unsigned scratch = MRI.createVirtualRegister(IsNeon ? VecTRC : TRC);
7004 emitPostLd(BB, MI, TII, dl, UnitSize, scratch, srcIn, srcOut,
7005 IsThumb1, IsThumb2);
7006 emitPostSt(BB, MI, TII, dl, UnitSize, scratch, destIn, destOut,
7007 IsThumb1, IsThumb2);
7012 // Handle the leftover bytes with LDRB and STRB.
7013 // [scratch, srcOut] = LDRB_POST(srcIn, 1)
7014 // [destOut] = STRB_POST(scratch, destIn, 1)
7015 for (unsigned i = 0; i < BytesLeft; i++) {
7016 unsigned srcOut = MRI.createVirtualRegister(TRC);
7017 unsigned destOut = MRI.createVirtualRegister(TRC);
7018 unsigned scratch = MRI.createVirtualRegister(TRC);
7019 emitPostLd(BB, MI, TII, dl, 1, scratch, srcIn, srcOut,
7020 IsThumb1, IsThumb2);
7021 emitPostSt(BB, MI, TII, dl, 1, scratch, destIn, destOut,
7022 IsThumb1, IsThumb2);
7026 MI->eraseFromParent(); // The instruction is gone now.
7030 // Expand the pseudo op to a loop.
7033 // movw varEnd, # --> with thumb2
7035 // ldrcp varEnd, idx --> without thumb2
7036 // fallthrough --> loopMBB
7038 // PHI varPhi, varEnd, varLoop
7039 // PHI srcPhi, src, srcLoop
7040 // PHI destPhi, dst, destLoop
7041 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
7042 // [destLoop] = STR_POST(scratch, destPhi, UnitSize)
7043 // subs varLoop, varPhi, #UnitSize
7045 // fallthrough --> exitMBB
7047 // epilogue to handle left-over bytes
7048 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
7049 // [destOut] = STRB_POST(scratch, destLoop, 1)
7050 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
7051 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
7052 MF->insert(It, loopMBB);
7053 MF->insert(It, exitMBB);
7055 // Transfer the remainder of BB and its successor edges to exitMBB.
7056 exitMBB->splice(exitMBB->begin(), BB,
7057 std::next(MachineBasicBlock::iterator(MI)), BB->end());
7058 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
7060 // Load an immediate to varEnd.
7061 unsigned varEnd = MRI.createVirtualRegister(TRC);
7063 unsigned Vtmp = varEnd;
7064 if ((LoopSize & 0xFFFF0000) != 0)
7065 Vtmp = MRI.createVirtualRegister(TRC);
7066 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2MOVi16), Vtmp)
7067 .addImm(LoopSize & 0xFFFF));
7069 if ((LoopSize & 0xFFFF0000) != 0)
7070 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2MOVTi16), varEnd)
7071 .addReg(Vtmp).addImm(LoopSize >> 16));
7073 MachineConstantPool *ConstantPool = MF->getConstantPool();
7074 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
7075 const Constant *C = ConstantInt::get(Int32Ty, LoopSize);
7077 // MachineConstantPool wants an explicit alignment.
7078 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
7080 Align = getDataLayout()->getTypeAllocSize(C->getType());
7081 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
7084 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(ARM::tLDRpci)).addReg(
7085 varEnd, RegState::Define).addConstantPoolIndex(Idx));
7087 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(ARM::LDRcp)).addReg(
7088 varEnd, RegState::Define).addConstantPoolIndex(Idx).addImm(0));
7090 BB->addSuccessor(loopMBB);
7092 // Generate the loop body:
7093 // varPhi = PHI(varLoop, varEnd)
7094 // srcPhi = PHI(srcLoop, src)
7095 // destPhi = PHI(destLoop, dst)
7096 MachineBasicBlock *entryBB = BB;
7098 unsigned varLoop = MRI.createVirtualRegister(TRC);
7099 unsigned varPhi = MRI.createVirtualRegister(TRC);
7100 unsigned srcLoop = MRI.createVirtualRegister(TRC);
7101 unsigned srcPhi = MRI.createVirtualRegister(TRC);
7102 unsigned destLoop = MRI.createVirtualRegister(TRC);
7103 unsigned destPhi = MRI.createVirtualRegister(TRC);
7105 BuildMI(*BB, BB->begin(), dl, TII->get(ARM::PHI), varPhi)
7106 .addReg(varLoop).addMBB(loopMBB)
7107 .addReg(varEnd).addMBB(entryBB);
7108 BuildMI(BB, dl, TII->get(ARM::PHI), srcPhi)
7109 .addReg(srcLoop).addMBB(loopMBB)
7110 .addReg(src).addMBB(entryBB);
7111 BuildMI(BB, dl, TII->get(ARM::PHI), destPhi)
7112 .addReg(destLoop).addMBB(loopMBB)
7113 .addReg(dest).addMBB(entryBB);
7115 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
7116 // [destLoop] = STR_POST(scratch, destPhi, UnitSiz)
7117 unsigned scratch = MRI.createVirtualRegister(IsNeon ? VecTRC : TRC);
7118 emitPostLd(BB, BB->end(), TII, dl, UnitSize, scratch, srcPhi, srcLoop,
7119 IsThumb1, IsThumb2);
7120 emitPostSt(BB, BB->end(), TII, dl, UnitSize, scratch, destPhi, destLoop,
7121 IsThumb1, IsThumb2);
7123 // Decrement loop variable by UnitSize.
7125 MachineInstrBuilder MIB =
7126 BuildMI(*BB, BB->end(), dl, TII->get(ARM::tSUBi8), varLoop);
7127 MIB = AddDefaultT1CC(MIB);
7128 MIB.addReg(varPhi).addImm(UnitSize);
7129 AddDefaultPred(MIB);
7131 MachineInstrBuilder MIB =
7132 BuildMI(*BB, BB->end(), dl,
7133 TII->get(IsThumb2 ? ARM::t2SUBri : ARM::SUBri), varLoop);
7134 AddDefaultCC(AddDefaultPred(MIB.addReg(varPhi).addImm(UnitSize)));
7135 MIB->getOperand(5).setReg(ARM::CPSR);
7136 MIB->getOperand(5).setIsDef(true);
7138 BuildMI(*BB, BB->end(), dl,
7139 TII->get(IsThumb1 ? ARM::tBcc : IsThumb2 ? ARM::t2Bcc : ARM::Bcc))
7140 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
7142 // loopMBB can loop back to loopMBB or fall through to exitMBB.
7143 BB->addSuccessor(loopMBB);
7144 BB->addSuccessor(exitMBB);
7146 // Add epilogue to handle BytesLeft.
7148 MachineInstr *StartOfExit = exitMBB->begin();
7150 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
7151 // [destOut] = STRB_POST(scratch, destLoop, 1)
7152 unsigned srcIn = srcLoop;
7153 unsigned destIn = destLoop;
7154 for (unsigned i = 0; i < BytesLeft; i++) {
7155 unsigned srcOut = MRI.createVirtualRegister(TRC);
7156 unsigned destOut = MRI.createVirtualRegister(TRC);
7157 unsigned scratch = MRI.createVirtualRegister(TRC);
7158 emitPostLd(BB, StartOfExit, TII, dl, 1, scratch, srcIn, srcOut,
7159 IsThumb1, IsThumb2);
7160 emitPostSt(BB, StartOfExit, TII, dl, 1, scratch, destIn, destOut,
7161 IsThumb1, IsThumb2);
7166 MI->eraseFromParent(); // The instruction is gone now.
7171 ARMTargetLowering::EmitLowered__chkstk(MachineInstr *MI,
7172 MachineBasicBlock *MBB) const {
7173 const TargetMachine &TM = getTargetMachine();
7174 const TargetInstrInfo &TII = *TM.getInstrInfo();
7175 DebugLoc DL = MI->getDebugLoc();
7177 assert(Subtarget->isTargetWindows() &&
7178 "__chkstk is only supported on Windows");
7179 assert(Subtarget->isThumb2() && "Windows on ARM requires Thumb-2 mode");
7181 // __chkstk takes the number of words to allocate on the stack in R4, and
7182 // returns the stack adjustment in number of bytes in R4. This will not
7183 // clober any other registers (other than the obvious lr).
7185 // Although, technically, IP should be considered a register which may be
7186 // clobbered, the call itself will not touch it. Windows on ARM is a pure
7187 // thumb-2 environment, so there is no interworking required. As a result, we
7188 // do not expect a veneer to be emitted by the linker, clobbering IP.
7190 // Each module receives its own copy of __chkstk, so no import thunk is
7191 // required, again, ensuring that IP is not clobbered.
7193 // Finally, although some linkers may theoretically provide a trampoline for
7194 // out of range calls (which is quite common due to a 32M range limitation of
7195 // branches for Thumb), we can generate the long-call version via
7196 // -mcmodel=large, alleviating the need for the trampoline which may clobber
7199 switch (TM.getCodeModel()) {
7200 case CodeModel::Small:
7201 case CodeModel::Medium:
7202 case CodeModel::Default:
7203 case CodeModel::Kernel:
7204 BuildMI(*MBB, MI, DL, TII.get(ARM::tBL))
7205 .addImm((unsigned)ARMCC::AL).addReg(0)
7206 .addExternalSymbol("__chkstk")
7207 .addReg(ARM::R4, RegState::Implicit | RegState::Kill)
7208 .addReg(ARM::R4, RegState::Implicit | RegState::Define)
7209 .addReg(ARM::R12, RegState::Implicit | RegState::Define | RegState::Dead);
7211 case CodeModel::Large:
7212 case CodeModel::JITDefault: {
7213 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
7214 unsigned Reg = MRI.createVirtualRegister(&ARM::rGPRRegClass);
7216 BuildMI(*MBB, MI, DL, TII.get(ARM::t2MOVi32imm), Reg)
7217 .addExternalSymbol("__chkstk");
7218 BuildMI(*MBB, MI, DL, TII.get(ARM::tBLXr))
7219 .addImm((unsigned)ARMCC::AL).addReg(0)
7220 .addReg(Reg, RegState::Kill)
7221 .addReg(ARM::R4, RegState::Implicit | RegState::Kill)
7222 .addReg(ARM::R4, RegState::Implicit | RegState::Define)
7223 .addReg(ARM::R12, RegState::Implicit | RegState::Define | RegState::Dead);
7228 AddDefaultCC(AddDefaultPred(BuildMI(*MBB, MI, DL, TII.get(ARM::t2SUBrr),
7230 .addReg(ARM::SP, RegState::Define)
7231 .addReg(ARM::R4, RegState::Kill)));
7233 MI->eraseFromParent();
7238 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
7239 MachineBasicBlock *BB) const {
7240 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7241 DebugLoc dl = MI->getDebugLoc();
7242 bool isThumb2 = Subtarget->isThumb2();
7243 switch (MI->getOpcode()) {
7246 llvm_unreachable("Unexpected instr type to insert");
7248 // The Thumb2 pre-indexed stores have the same MI operands, they just
7249 // define them differently in the .td files from the isel patterns, so
7250 // they need pseudos.
7251 case ARM::t2STR_preidx:
7252 MI->setDesc(TII->get(ARM::t2STR_PRE));
7254 case ARM::t2STRB_preidx:
7255 MI->setDesc(TII->get(ARM::t2STRB_PRE));
7257 case ARM::t2STRH_preidx:
7258 MI->setDesc(TII->get(ARM::t2STRH_PRE));
7261 case ARM::STRi_preidx:
7262 case ARM::STRBi_preidx: {
7263 unsigned NewOpc = MI->getOpcode() == ARM::STRi_preidx ?
7264 ARM::STR_PRE_IMM : ARM::STRB_PRE_IMM;
7265 // Decode the offset.
7266 unsigned Offset = MI->getOperand(4).getImm();
7267 bool isSub = ARM_AM::getAM2Op(Offset) == ARM_AM::sub;
7268 Offset = ARM_AM::getAM2Offset(Offset);
7272 MachineMemOperand *MMO = *MI->memoperands_begin();
7273 BuildMI(*BB, MI, dl, TII->get(NewOpc))
7274 .addOperand(MI->getOperand(0)) // Rn_wb
7275 .addOperand(MI->getOperand(1)) // Rt
7276 .addOperand(MI->getOperand(2)) // Rn
7277 .addImm(Offset) // offset (skip GPR==zero_reg)
7278 .addOperand(MI->getOperand(5)) // pred
7279 .addOperand(MI->getOperand(6))
7280 .addMemOperand(MMO);
7281 MI->eraseFromParent();
7284 case ARM::STRr_preidx:
7285 case ARM::STRBr_preidx:
7286 case ARM::STRH_preidx: {
7288 switch (MI->getOpcode()) {
7289 default: llvm_unreachable("unexpected opcode!");
7290 case ARM::STRr_preidx: NewOpc = ARM::STR_PRE_REG; break;
7291 case ARM::STRBr_preidx: NewOpc = ARM::STRB_PRE_REG; break;
7292 case ARM::STRH_preidx: NewOpc = ARM::STRH_PRE; break;
7294 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(NewOpc));
7295 for (unsigned i = 0; i < MI->getNumOperands(); ++i)
7296 MIB.addOperand(MI->getOperand(i));
7297 MI->eraseFromParent();
7301 case ARM::tMOVCCr_pseudo: {
7302 // To "insert" a SELECT_CC instruction, we actually have to insert the
7303 // diamond control-flow pattern. The incoming instruction knows the
7304 // destination vreg to set, the condition code register to branch on, the
7305 // true/false values to select between, and a branch opcode to use.
7306 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7307 MachineFunction::iterator It = BB;
7313 // cmpTY ccX, r1, r2
7315 // fallthrough --> copy0MBB
7316 MachineBasicBlock *thisMBB = BB;
7317 MachineFunction *F = BB->getParent();
7318 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
7319 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
7320 F->insert(It, copy0MBB);
7321 F->insert(It, sinkMBB);
7323 // Transfer the remainder of BB and its successor edges to sinkMBB.
7324 sinkMBB->splice(sinkMBB->begin(), BB,
7325 std::next(MachineBasicBlock::iterator(MI)), BB->end());
7326 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
7328 BB->addSuccessor(copy0MBB);
7329 BB->addSuccessor(sinkMBB);
7331 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
7332 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
7335 // %FalseValue = ...
7336 // # fallthrough to sinkMBB
7339 // Update machine-CFG edges
7340 BB->addSuccessor(sinkMBB);
7343 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
7346 BuildMI(*BB, BB->begin(), dl,
7347 TII->get(ARM::PHI), MI->getOperand(0).getReg())
7348 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
7349 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
7351 MI->eraseFromParent(); // The pseudo instruction is gone now.
7356 case ARM::BCCZi64: {
7357 // If there is an unconditional branch to the other successor, remove it.
7358 BB->erase(std::next(MachineBasicBlock::iterator(MI)), BB->end());
7360 // Compare both parts that make up the double comparison separately for
7362 bool RHSisZero = MI->getOpcode() == ARM::BCCZi64;
7364 unsigned LHS1 = MI->getOperand(1).getReg();
7365 unsigned LHS2 = MI->getOperand(2).getReg();
7367 AddDefaultPred(BuildMI(BB, dl,
7368 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7369 .addReg(LHS1).addImm(0));
7370 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7371 .addReg(LHS2).addImm(0)
7372 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
7374 unsigned RHS1 = MI->getOperand(3).getReg();
7375 unsigned RHS2 = MI->getOperand(4).getReg();
7376 AddDefaultPred(BuildMI(BB, dl,
7377 TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
7378 .addReg(LHS1).addReg(RHS1));
7379 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
7380 .addReg(LHS2).addReg(RHS2)
7381 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
7384 MachineBasicBlock *destMBB = MI->getOperand(RHSisZero ? 3 : 5).getMBB();
7385 MachineBasicBlock *exitMBB = OtherSucc(BB, destMBB);
7386 if (MI->getOperand(0).getImm() == ARMCC::NE)
7387 std::swap(destMBB, exitMBB);
7389 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
7390 .addMBB(destMBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
7392 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2B)).addMBB(exitMBB));
7394 BuildMI(BB, dl, TII->get(ARM::B)) .addMBB(exitMBB);
7396 MI->eraseFromParent(); // The pseudo instruction is gone now.
7400 case ARM::Int_eh_sjlj_setjmp:
7401 case ARM::Int_eh_sjlj_setjmp_nofp:
7402 case ARM::tInt_eh_sjlj_setjmp:
7403 case ARM::t2Int_eh_sjlj_setjmp:
7404 case ARM::t2Int_eh_sjlj_setjmp_nofp:
7405 EmitSjLjDispatchBlock(MI, BB);
7410 // To insert an ABS instruction, we have to insert the
7411 // diamond control-flow pattern. The incoming instruction knows the
7412 // source vreg to test against 0, the destination vreg to set,
7413 // the condition code register to branch on, the
7414 // true/false values to select between, and a branch opcode to use.
7419 // BCC (branch to SinkBB if V0 >= 0)
7420 // RSBBB: V3 = RSBri V2, 0 (compute ABS if V2 < 0)
7421 // SinkBB: V1 = PHI(V2, V3)
7422 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7423 MachineFunction::iterator BBI = BB;
7425 MachineFunction *Fn = BB->getParent();
7426 MachineBasicBlock *RSBBB = Fn->CreateMachineBasicBlock(LLVM_BB);
7427 MachineBasicBlock *SinkBB = Fn->CreateMachineBasicBlock(LLVM_BB);
7428 Fn->insert(BBI, RSBBB);
7429 Fn->insert(BBI, SinkBB);
7431 unsigned int ABSSrcReg = MI->getOperand(1).getReg();
7432 unsigned int ABSDstReg = MI->getOperand(0).getReg();
7433 bool isThumb2 = Subtarget->isThumb2();
7434 MachineRegisterInfo &MRI = Fn->getRegInfo();
7435 // In Thumb mode S must not be specified if source register is the SP or
7436 // PC and if destination register is the SP, so restrict register class
7437 unsigned NewRsbDstReg = MRI.createVirtualRegister(isThumb2 ?
7438 (const TargetRegisterClass*)&ARM::rGPRRegClass :
7439 (const TargetRegisterClass*)&ARM::GPRRegClass);
7441 // Transfer the remainder of BB and its successor edges to sinkMBB.
7442 SinkBB->splice(SinkBB->begin(), BB,
7443 std::next(MachineBasicBlock::iterator(MI)), BB->end());
7444 SinkBB->transferSuccessorsAndUpdatePHIs(BB);
7446 BB->addSuccessor(RSBBB);
7447 BB->addSuccessor(SinkBB);
7449 // fall through to SinkMBB
7450 RSBBB->addSuccessor(SinkBB);
7452 // insert a cmp at the end of BB
7453 AddDefaultPred(BuildMI(BB, dl,
7454 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7455 .addReg(ABSSrcReg).addImm(0));
7457 // insert a bcc with opposite CC to ARMCC::MI at the end of BB
7459 TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)).addMBB(SinkBB)
7460 .addImm(ARMCC::getOppositeCondition(ARMCC::MI)).addReg(ARM::CPSR);
7462 // insert rsbri in RSBBB
7463 // Note: BCC and rsbri will be converted into predicated rsbmi
7464 // by if-conversion pass
7465 BuildMI(*RSBBB, RSBBB->begin(), dl,
7466 TII->get(isThumb2 ? ARM::t2RSBri : ARM::RSBri), NewRsbDstReg)
7467 .addReg(ABSSrcReg, RegState::Kill)
7468 .addImm(0).addImm((unsigned)ARMCC::AL).addReg(0).addReg(0);
7470 // insert PHI in SinkBB,
7471 // reuse ABSDstReg to not change uses of ABS instruction
7472 BuildMI(*SinkBB, SinkBB->begin(), dl,
7473 TII->get(ARM::PHI), ABSDstReg)
7474 .addReg(NewRsbDstReg).addMBB(RSBBB)
7475 .addReg(ABSSrcReg).addMBB(BB);
7477 // remove ABS instruction
7478 MI->eraseFromParent();
7480 // return last added BB
7483 case ARM::COPY_STRUCT_BYVAL_I32:
7485 return EmitStructByval(MI, BB);
7486 case ARM::WIN__CHKSTK:
7487 return EmitLowered__chkstk(MI, BB);
7491 void ARMTargetLowering::AdjustInstrPostInstrSelection(MachineInstr *MI,
7492 SDNode *Node) const {
7493 if (!MI->hasPostISelHook()) {
7494 assert(!convertAddSubFlagsOpcode(MI->getOpcode()) &&
7495 "Pseudo flag-setting opcodes must be marked with 'hasPostISelHook'");
7499 const MCInstrDesc *MCID = &MI->getDesc();
7500 // Adjust potentially 's' setting instructions after isel, i.e. ADC, SBC, RSB,
7501 // RSC. Coming out of isel, they have an implicit CPSR def, but the optional
7502 // operand is still set to noreg. If needed, set the optional operand's
7503 // register to CPSR, and remove the redundant implicit def.
7505 // e.g. ADCS (..., CPSR<imp-def>) -> ADC (... opt:CPSR<def>).
7507 // Rename pseudo opcodes.
7508 unsigned NewOpc = convertAddSubFlagsOpcode(MI->getOpcode());
7510 const ARMBaseInstrInfo *TII =
7511 static_cast<const ARMBaseInstrInfo*>(getTargetMachine().getInstrInfo());
7512 MCID = &TII->get(NewOpc);
7514 assert(MCID->getNumOperands() == MI->getDesc().getNumOperands() + 1 &&
7515 "converted opcode should be the same except for cc_out");
7519 // Add the optional cc_out operand
7520 MI->addOperand(MachineOperand::CreateReg(0, /*isDef=*/true));
7522 unsigned ccOutIdx = MCID->getNumOperands() - 1;
7524 // Any ARM instruction that sets the 's' bit should specify an optional
7525 // "cc_out" operand in the last operand position.
7526 if (!MI->hasOptionalDef() || !MCID->OpInfo[ccOutIdx].isOptionalDef()) {
7527 assert(!NewOpc && "Optional cc_out operand required");
7530 // Look for an implicit def of CPSR added by MachineInstr ctor. Remove it
7531 // since we already have an optional CPSR def.
7532 bool definesCPSR = false;
7533 bool deadCPSR = false;
7534 for (unsigned i = MCID->getNumOperands(), e = MI->getNumOperands();
7536 const MachineOperand &MO = MI->getOperand(i);
7537 if (MO.isReg() && MO.isDef() && MO.getReg() == ARM::CPSR) {
7541 MI->RemoveOperand(i);
7546 assert(!NewOpc && "Optional cc_out operand required");
7549 assert(deadCPSR == !Node->hasAnyUseOfValue(1) && "inconsistent dead flag");
7551 assert(!MI->getOperand(ccOutIdx).getReg() &&
7552 "expect uninitialized optional cc_out operand");
7556 // If this instruction was defined with an optional CPSR def and its dag node
7557 // had a live implicit CPSR def, then activate the optional CPSR def.
7558 MachineOperand &MO = MI->getOperand(ccOutIdx);
7559 MO.setReg(ARM::CPSR);
7563 //===----------------------------------------------------------------------===//
7564 // ARM Optimization Hooks
7565 //===----------------------------------------------------------------------===//
7567 // Helper function that checks if N is a null or all ones constant.
7568 static inline bool isZeroOrAllOnes(SDValue N, bool AllOnes) {
7569 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N);
7572 return AllOnes ? C->isAllOnesValue() : C->isNullValue();
7575 // Return true if N is conditionally 0 or all ones.
7576 // Detects these expressions where cc is an i1 value:
7578 // (select cc 0, y) [AllOnes=0]
7579 // (select cc y, 0) [AllOnes=0]
7580 // (zext cc) [AllOnes=0]
7581 // (sext cc) [AllOnes=0/1]
7582 // (select cc -1, y) [AllOnes=1]
7583 // (select cc y, -1) [AllOnes=1]
7585 // Invert is set when N is the null/all ones constant when CC is false.
7586 // OtherOp is set to the alternative value of N.
7587 static bool isConditionalZeroOrAllOnes(SDNode *N, bool AllOnes,
7588 SDValue &CC, bool &Invert,
7590 SelectionDAG &DAG) {
7591 switch (N->getOpcode()) {
7592 default: return false;
7594 CC = N->getOperand(0);
7595 SDValue N1 = N->getOperand(1);
7596 SDValue N2 = N->getOperand(2);
7597 if (isZeroOrAllOnes(N1, AllOnes)) {
7602 if (isZeroOrAllOnes(N2, AllOnes)) {
7609 case ISD::ZERO_EXTEND:
7610 // (zext cc) can never be the all ones value.
7614 case ISD::SIGN_EXTEND: {
7615 EVT VT = N->getValueType(0);
7616 CC = N->getOperand(0);
7617 if (CC.getValueType() != MVT::i1)
7621 // When looking for an AllOnes constant, N is an sext, and the 'other'
7623 OtherOp = DAG.getConstant(0, VT);
7624 else if (N->getOpcode() == ISD::ZERO_EXTEND)
7625 // When looking for a 0 constant, N can be zext or sext.
7626 OtherOp = DAG.getConstant(1, VT);
7628 OtherOp = DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), VT);
7634 // Combine a constant select operand into its use:
7636 // (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
7637 // (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
7638 // (and (select cc, -1, c), x) -> (select cc, x, (and, x, c)) [AllOnes=1]
7639 // (or (select cc, 0, c), x) -> (select cc, x, (or, x, c))
7640 // (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c))
7642 // The transform is rejected if the select doesn't have a constant operand that
7643 // is null, or all ones when AllOnes is set.
7645 // Also recognize sext/zext from i1:
7647 // (add (zext cc), x) -> (select cc (add x, 1), x)
7648 // (add (sext cc), x) -> (select cc (add x, -1), x)
7650 // These transformations eventually create predicated instructions.
7652 // @param N The node to transform.
7653 // @param Slct The N operand that is a select.
7654 // @param OtherOp The other N operand (x above).
7655 // @param DCI Context.
7656 // @param AllOnes Require the select constant to be all ones instead of null.
7657 // @returns The new node, or SDValue() on failure.
7659 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
7660 TargetLowering::DAGCombinerInfo &DCI,
7661 bool AllOnes = false) {
7662 SelectionDAG &DAG = DCI.DAG;
7663 EVT VT = N->getValueType(0);
7664 SDValue NonConstantVal;
7667 if (!isConditionalZeroOrAllOnes(Slct.getNode(), AllOnes, CCOp, SwapSelectOps,
7668 NonConstantVal, DAG))
7671 // Slct is now know to be the desired identity constant when CC is true.
7672 SDValue TrueVal = OtherOp;
7673 SDValue FalseVal = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
7674 OtherOp, NonConstantVal);
7675 // Unless SwapSelectOps says CC should be false.
7677 std::swap(TrueVal, FalseVal);
7679 return DAG.getNode(ISD::SELECT, SDLoc(N), VT,
7680 CCOp, TrueVal, FalseVal);
7683 // Attempt combineSelectAndUse on each operand of a commutative operator N.
7685 SDValue combineSelectAndUseCommutative(SDNode *N, bool AllOnes,
7686 TargetLowering::DAGCombinerInfo &DCI) {
7687 SDValue N0 = N->getOperand(0);
7688 SDValue N1 = N->getOperand(1);
7689 if (N0.getNode()->hasOneUse()) {
7690 SDValue Result = combineSelectAndUse(N, N0, N1, DCI, AllOnes);
7691 if (Result.getNode())
7694 if (N1.getNode()->hasOneUse()) {
7695 SDValue Result = combineSelectAndUse(N, N1, N0, DCI, AllOnes);
7696 if (Result.getNode())
7702 // AddCombineToVPADDL- For pair-wise add on neon, use the vpaddl instruction
7703 // (only after legalization).
7704 static SDValue AddCombineToVPADDL(SDNode *N, SDValue N0, SDValue N1,
7705 TargetLowering::DAGCombinerInfo &DCI,
7706 const ARMSubtarget *Subtarget) {
7708 // Only perform optimization if after legalize, and if NEON is available. We
7709 // also expected both operands to be BUILD_VECTORs.
7710 if (DCI.isBeforeLegalize() || !Subtarget->hasNEON()
7711 || N0.getOpcode() != ISD::BUILD_VECTOR
7712 || N1.getOpcode() != ISD::BUILD_VECTOR)
7715 // Check output type since VPADDL operand elements can only be 8, 16, or 32.
7716 EVT VT = N->getValueType(0);
7717 if (!VT.isInteger() || VT.getVectorElementType() == MVT::i64)
7720 // Check that the vector operands are of the right form.
7721 // N0 and N1 are BUILD_VECTOR nodes with N number of EXTRACT_VECTOR
7722 // operands, where N is the size of the formed vector.
7723 // Each EXTRACT_VECTOR should have the same input vector and odd or even
7724 // index such that we have a pair wise add pattern.
7726 // Grab the vector that all EXTRACT_VECTOR nodes should be referencing.
7727 if (N0->getOperand(0)->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
7729 SDValue Vec = N0->getOperand(0)->getOperand(0);
7730 SDNode *V = Vec.getNode();
7731 unsigned nextIndex = 0;
7733 // For each operands to the ADD which are BUILD_VECTORs,
7734 // check to see if each of their operands are an EXTRACT_VECTOR with
7735 // the same vector and appropriate index.
7736 for (unsigned i = 0, e = N0->getNumOperands(); i != e; ++i) {
7737 if (N0->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT
7738 && N1->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
7740 SDValue ExtVec0 = N0->getOperand(i);
7741 SDValue ExtVec1 = N1->getOperand(i);
7743 // First operand is the vector, verify its the same.
7744 if (V != ExtVec0->getOperand(0).getNode() ||
7745 V != ExtVec1->getOperand(0).getNode())
7748 // Second is the constant, verify its correct.
7749 ConstantSDNode *C0 = dyn_cast<ConstantSDNode>(ExtVec0->getOperand(1));
7750 ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(ExtVec1->getOperand(1));
7752 // For the constant, we want to see all the even or all the odd.
7753 if (!C0 || !C1 || C0->getZExtValue() != nextIndex
7754 || C1->getZExtValue() != nextIndex+1)
7763 // Create VPADDL node.
7764 SelectionDAG &DAG = DCI.DAG;
7765 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7767 // Build operand list.
7768 SmallVector<SDValue, 8> Ops;
7769 Ops.push_back(DAG.getConstant(Intrinsic::arm_neon_vpaddls,
7770 TLI.getPointerTy()));
7772 // Input is the vector.
7775 // Get widened type and narrowed type.
7777 unsigned numElem = VT.getVectorNumElements();
7779 EVT inputLaneType = Vec.getValueType().getVectorElementType();
7780 switch (inputLaneType.getSimpleVT().SimpleTy) {
7781 case MVT::i8: widenType = MVT::getVectorVT(MVT::i16, numElem); break;
7782 case MVT::i16: widenType = MVT::getVectorVT(MVT::i32, numElem); break;
7783 case MVT::i32: widenType = MVT::getVectorVT(MVT::i64, numElem); break;
7785 llvm_unreachable("Invalid vector element type for padd optimization.");
7788 SDValue tmp = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N), widenType, Ops);
7789 unsigned ExtOp = VT.bitsGT(tmp.getValueType()) ? ISD::ANY_EXTEND : ISD::TRUNCATE;
7790 return DAG.getNode(ExtOp, SDLoc(N), VT, tmp);
7793 static SDValue findMUL_LOHI(SDValue V) {
7794 if (V->getOpcode() == ISD::UMUL_LOHI ||
7795 V->getOpcode() == ISD::SMUL_LOHI)
7800 static SDValue AddCombineTo64bitMLAL(SDNode *AddcNode,
7801 TargetLowering::DAGCombinerInfo &DCI,
7802 const ARMSubtarget *Subtarget) {
7804 if (Subtarget->isThumb1Only()) return SDValue();
7806 // Only perform the checks after legalize when the pattern is available.
7807 if (DCI.isBeforeLegalize()) return SDValue();
7809 // Look for multiply add opportunities.
7810 // The pattern is a ISD::UMUL_LOHI followed by two add nodes, where
7811 // each add nodes consumes a value from ISD::UMUL_LOHI and there is
7812 // a glue link from the first add to the second add.
7813 // If we find this pattern, we can replace the U/SMUL_LOHI, ADDC, and ADDE by
7814 // a S/UMLAL instruction.
7817 // \ / \ [no multiline comment]
7823 assert(AddcNode->getOpcode() == ISD::ADDC && "Expect an ADDC");
7824 SDValue AddcOp0 = AddcNode->getOperand(0);
7825 SDValue AddcOp1 = AddcNode->getOperand(1);
7827 // Check if the two operands are from the same mul_lohi node.
7828 if (AddcOp0.getNode() == AddcOp1.getNode())
7831 assert(AddcNode->getNumValues() == 2 &&
7832 AddcNode->getValueType(0) == MVT::i32 &&
7833 "Expect ADDC with two result values. First: i32");
7835 // Check that we have a glued ADDC node.
7836 if (AddcNode->getValueType(1) != MVT::Glue)
7839 // Check that the ADDC adds the low result of the S/UMUL_LOHI.
7840 if (AddcOp0->getOpcode() != ISD::UMUL_LOHI &&
7841 AddcOp0->getOpcode() != ISD::SMUL_LOHI &&
7842 AddcOp1->getOpcode() != ISD::UMUL_LOHI &&
7843 AddcOp1->getOpcode() != ISD::SMUL_LOHI)
7846 // Look for the glued ADDE.
7847 SDNode* AddeNode = AddcNode->getGluedUser();
7851 // Make sure it is really an ADDE.
7852 if (AddeNode->getOpcode() != ISD::ADDE)
7855 assert(AddeNode->getNumOperands() == 3 &&
7856 AddeNode->getOperand(2).getValueType() == MVT::Glue &&
7857 "ADDE node has the wrong inputs");
7859 // Check for the triangle shape.
7860 SDValue AddeOp0 = AddeNode->getOperand(0);
7861 SDValue AddeOp1 = AddeNode->getOperand(1);
7863 // Make sure that the ADDE operands are not coming from the same node.
7864 if (AddeOp0.getNode() == AddeOp1.getNode())
7867 // Find the MUL_LOHI node walking up ADDE's operands.
7868 bool IsLeftOperandMUL = false;
7869 SDValue MULOp = findMUL_LOHI(AddeOp0);
7870 if (MULOp == SDValue())
7871 MULOp = findMUL_LOHI(AddeOp1);
7873 IsLeftOperandMUL = true;
7874 if (MULOp == SDValue())
7877 // Figure out the right opcode.
7878 unsigned Opc = MULOp->getOpcode();
7879 unsigned FinalOpc = (Opc == ISD::SMUL_LOHI) ? ARMISD::SMLAL : ARMISD::UMLAL;
7881 // Figure out the high and low input values to the MLAL node.
7882 SDValue* HiMul = &MULOp;
7883 SDValue* HiAdd = nullptr;
7884 SDValue* LoMul = nullptr;
7885 SDValue* LowAdd = nullptr;
7887 if (IsLeftOperandMUL)
7893 if (AddcOp0->getOpcode() == Opc) {
7897 if (AddcOp1->getOpcode() == Opc) {
7905 if (LoMul->getNode() != HiMul->getNode())
7908 // Create the merged node.
7909 SelectionDAG &DAG = DCI.DAG;
7911 // Build operand list.
7912 SmallVector<SDValue, 8> Ops;
7913 Ops.push_back(LoMul->getOperand(0));
7914 Ops.push_back(LoMul->getOperand(1));
7915 Ops.push_back(*LowAdd);
7916 Ops.push_back(*HiAdd);
7918 SDValue MLALNode = DAG.getNode(FinalOpc, SDLoc(AddcNode),
7919 DAG.getVTList(MVT::i32, MVT::i32), Ops);
7921 // Replace the ADDs' nodes uses by the MLA node's values.
7922 SDValue HiMLALResult(MLALNode.getNode(), 1);
7923 DAG.ReplaceAllUsesOfValueWith(SDValue(AddeNode, 0), HiMLALResult);
7925 SDValue LoMLALResult(MLALNode.getNode(), 0);
7926 DAG.ReplaceAllUsesOfValueWith(SDValue(AddcNode, 0), LoMLALResult);
7928 // Return original node to notify the driver to stop replacing.
7929 SDValue resNode(AddcNode, 0);
7933 /// PerformADDCCombine - Target-specific dag combine transform from
7934 /// ISD::ADDC, ISD::ADDE, and ISD::MUL_LOHI to MLAL.
7935 static SDValue PerformADDCCombine(SDNode *N,
7936 TargetLowering::DAGCombinerInfo &DCI,
7937 const ARMSubtarget *Subtarget) {
7939 return AddCombineTo64bitMLAL(N, DCI, Subtarget);
7943 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
7944 /// operands N0 and N1. This is a helper for PerformADDCombine that is
7945 /// called with the default operands, and if that fails, with commuted
7947 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
7948 TargetLowering::DAGCombinerInfo &DCI,
7949 const ARMSubtarget *Subtarget){
7951 // Attempt to create vpaddl for this add.
7952 SDValue Result = AddCombineToVPADDL(N, N0, N1, DCI, Subtarget);
7953 if (Result.getNode())
7956 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
7957 if (N0.getNode()->hasOneUse()) {
7958 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
7959 if (Result.getNode()) return Result;
7964 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
7966 static SDValue PerformADDCombine(SDNode *N,
7967 TargetLowering::DAGCombinerInfo &DCI,
7968 const ARMSubtarget *Subtarget) {
7969 SDValue N0 = N->getOperand(0);
7970 SDValue N1 = N->getOperand(1);
7972 // First try with the default operand order.
7973 SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI, Subtarget);
7974 if (Result.getNode())
7977 // If that didn't work, try again with the operands commuted.
7978 return PerformADDCombineWithOperands(N, N1, N0, DCI, Subtarget);
7981 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
7983 static SDValue PerformSUBCombine(SDNode *N,
7984 TargetLowering::DAGCombinerInfo &DCI) {
7985 SDValue N0 = N->getOperand(0);
7986 SDValue N1 = N->getOperand(1);
7988 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
7989 if (N1.getNode()->hasOneUse()) {
7990 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
7991 if (Result.getNode()) return Result;
7997 /// PerformVMULCombine
7998 /// Distribute (A + B) * C to (A * C) + (B * C) to take advantage of the
7999 /// special multiplier accumulator forwarding.
8005 // However, for (A + B) * (A + B),
8012 static SDValue PerformVMULCombine(SDNode *N,
8013 TargetLowering::DAGCombinerInfo &DCI,
8014 const ARMSubtarget *Subtarget) {
8015 if (!Subtarget->hasVMLxForwarding())
8018 SelectionDAG &DAG = DCI.DAG;
8019 SDValue N0 = N->getOperand(0);
8020 SDValue N1 = N->getOperand(1);
8021 unsigned Opcode = N0.getOpcode();
8022 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
8023 Opcode != ISD::FADD && Opcode != ISD::FSUB) {
8024 Opcode = N1.getOpcode();
8025 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
8026 Opcode != ISD::FADD && Opcode != ISD::FSUB)
8034 EVT VT = N->getValueType(0);
8036 SDValue N00 = N0->getOperand(0);
8037 SDValue N01 = N0->getOperand(1);
8038 return DAG.getNode(Opcode, DL, VT,
8039 DAG.getNode(ISD::MUL, DL, VT, N00, N1),
8040 DAG.getNode(ISD::MUL, DL, VT, N01, N1));
8043 static SDValue PerformMULCombine(SDNode *N,
8044 TargetLowering::DAGCombinerInfo &DCI,
8045 const ARMSubtarget *Subtarget) {
8046 SelectionDAG &DAG = DCI.DAG;
8048 if (Subtarget->isThumb1Only())
8051 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
8054 EVT VT = N->getValueType(0);
8055 if (VT.is64BitVector() || VT.is128BitVector())
8056 return PerformVMULCombine(N, DCI, Subtarget);
8060 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
8064 int64_t MulAmt = C->getSExtValue();
8065 unsigned ShiftAmt = countTrailingZeros<uint64_t>(MulAmt);
8067 ShiftAmt = ShiftAmt & (32 - 1);
8068 SDValue V = N->getOperand(0);
8072 MulAmt >>= ShiftAmt;
8075 if (isPowerOf2_32(MulAmt - 1)) {
8076 // (mul x, 2^N + 1) => (add (shl x, N), x)
8077 Res = DAG.getNode(ISD::ADD, DL, VT,
8079 DAG.getNode(ISD::SHL, DL, VT,
8081 DAG.getConstant(Log2_32(MulAmt - 1),
8083 } else if (isPowerOf2_32(MulAmt + 1)) {
8084 // (mul x, 2^N - 1) => (sub (shl x, N), x)
8085 Res = DAG.getNode(ISD::SUB, DL, VT,
8086 DAG.getNode(ISD::SHL, DL, VT,
8088 DAG.getConstant(Log2_32(MulAmt + 1),
8094 uint64_t MulAmtAbs = -MulAmt;
8095 if (isPowerOf2_32(MulAmtAbs + 1)) {
8096 // (mul x, -(2^N - 1)) => (sub x, (shl x, N))
8097 Res = DAG.getNode(ISD::SUB, DL, VT,
8099 DAG.getNode(ISD::SHL, DL, VT,
8101 DAG.getConstant(Log2_32(MulAmtAbs + 1),
8103 } else if (isPowerOf2_32(MulAmtAbs - 1)) {
8104 // (mul x, -(2^N + 1)) => - (add (shl x, N), x)
8105 Res = DAG.getNode(ISD::ADD, DL, VT,
8107 DAG.getNode(ISD::SHL, DL, VT,
8109 DAG.getConstant(Log2_32(MulAmtAbs-1),
8111 Res = DAG.getNode(ISD::SUB, DL, VT,
8112 DAG.getConstant(0, MVT::i32),Res);
8119 Res = DAG.getNode(ISD::SHL, DL, VT,
8120 Res, DAG.getConstant(ShiftAmt, MVT::i32));
8122 // Do not add new nodes to DAG combiner worklist.
8123 DCI.CombineTo(N, Res, false);
8127 static SDValue PerformANDCombine(SDNode *N,
8128 TargetLowering::DAGCombinerInfo &DCI,
8129 const ARMSubtarget *Subtarget) {
8131 // Attempt to use immediate-form VBIC
8132 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
8134 EVT VT = N->getValueType(0);
8135 SelectionDAG &DAG = DCI.DAG;
8137 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8140 APInt SplatBits, SplatUndef;
8141 unsigned SplatBitSize;
8144 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
8145 if (SplatBitSize <= 64) {
8147 SDValue Val = isNEONModifiedImm((~SplatBits).getZExtValue(),
8148 SplatUndef.getZExtValue(), SplatBitSize,
8149 DAG, VbicVT, VT.is128BitVector(),
8151 if (Val.getNode()) {
8153 DAG.getNode(ISD::BITCAST, dl, VbicVT, N->getOperand(0));
8154 SDValue Vbic = DAG.getNode(ARMISD::VBICIMM, dl, VbicVT, Input, Val);
8155 return DAG.getNode(ISD::BITCAST, dl, VT, Vbic);
8160 if (!Subtarget->isThumb1Only()) {
8161 // fold (and (select cc, -1, c), x) -> (select cc, x, (and, x, c))
8162 SDValue Result = combineSelectAndUseCommutative(N, true, DCI);
8163 if (Result.getNode())
8170 /// PerformORCombine - Target-specific dag combine xforms for ISD::OR
8171 static SDValue PerformORCombine(SDNode *N,
8172 TargetLowering::DAGCombinerInfo &DCI,
8173 const ARMSubtarget *Subtarget) {
8174 // Attempt to use immediate-form VORR
8175 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
8177 EVT VT = N->getValueType(0);
8178 SelectionDAG &DAG = DCI.DAG;
8180 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8183 APInt SplatBits, SplatUndef;
8184 unsigned SplatBitSize;
8186 if (BVN && Subtarget->hasNEON() &&
8187 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
8188 if (SplatBitSize <= 64) {
8190 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
8191 SplatUndef.getZExtValue(), SplatBitSize,
8192 DAG, VorrVT, VT.is128BitVector(),
8194 if (Val.getNode()) {
8196 DAG.getNode(ISD::BITCAST, dl, VorrVT, N->getOperand(0));
8197 SDValue Vorr = DAG.getNode(ARMISD::VORRIMM, dl, VorrVT, Input, Val);
8198 return DAG.getNode(ISD::BITCAST, dl, VT, Vorr);
8203 if (!Subtarget->isThumb1Only()) {
8204 // fold (or (select cc, 0, c), x) -> (select cc, x, (or, x, c))
8205 SDValue Result = combineSelectAndUseCommutative(N, false, DCI);
8206 if (Result.getNode())
8210 // The code below optimizes (or (and X, Y), Z).
8211 // The AND operand needs to have a single user to make these optimizations
8213 SDValue N0 = N->getOperand(0);
8214 if (N0.getOpcode() != ISD::AND || !N0.hasOneUse())
8216 SDValue N1 = N->getOperand(1);
8218 // (or (and B, A), (and C, ~A)) => (VBSL A, B, C) when A is a constant.
8219 if (Subtarget->hasNEON() && N1.getOpcode() == ISD::AND && VT.isVector() &&
8220 DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
8222 unsigned SplatBitSize;
8225 APInt SplatBits0, SplatBits1;
8226 BuildVectorSDNode *BVN0 = dyn_cast<BuildVectorSDNode>(N0->getOperand(1));
8227 BuildVectorSDNode *BVN1 = dyn_cast<BuildVectorSDNode>(N1->getOperand(1));
8228 // Ensure that the second operand of both ands are constants
8229 if (BVN0 && BVN0->isConstantSplat(SplatBits0, SplatUndef, SplatBitSize,
8230 HasAnyUndefs) && !HasAnyUndefs) {
8231 if (BVN1 && BVN1->isConstantSplat(SplatBits1, SplatUndef, SplatBitSize,
8232 HasAnyUndefs) && !HasAnyUndefs) {
8233 // Ensure that the bit width of the constants are the same and that
8234 // the splat arguments are logical inverses as per the pattern we
8235 // are trying to simplify.
8236 if (SplatBits0.getBitWidth() == SplatBits1.getBitWidth() &&
8237 SplatBits0 == ~SplatBits1) {
8238 // Canonicalize the vector type to make instruction selection
8240 EVT CanonicalVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
8241 SDValue Result = DAG.getNode(ARMISD::VBSL, dl, CanonicalVT,
8245 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
8251 // Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when
8254 // BFI is only available on V6T2+
8255 if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops())
8259 // 1) or (and A, mask), val => ARMbfi A, val, mask
8260 // iff (val & mask) == val
8262 // 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
8263 // 2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2)
8264 // && mask == ~mask2
8265 // 2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2)
8266 // && ~mask == mask2
8267 // (i.e., copy a bitfield value into another bitfield of the same width)
8272 SDValue N00 = N0.getOperand(0);
8274 // The value and the mask need to be constants so we can verify this is
8275 // actually a bitfield set. If the mask is 0xffff, we can do better
8276 // via a movt instruction, so don't use BFI in that case.
8277 SDValue MaskOp = N0.getOperand(1);
8278 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(MaskOp);
8281 unsigned Mask = MaskC->getZExtValue();
8285 // Case (1): or (and A, mask), val => ARMbfi A, val, mask
8286 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
8288 unsigned Val = N1C->getZExtValue();
8289 if ((Val & ~Mask) != Val)
8292 if (ARM::isBitFieldInvertedMask(Mask)) {
8293 Val >>= countTrailingZeros(~Mask);
8295 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00,
8296 DAG.getConstant(Val, MVT::i32),
8297 DAG.getConstant(Mask, MVT::i32));
8299 // Do not add new nodes to DAG combiner worklist.
8300 DCI.CombineTo(N, Res, false);
8303 } else if (N1.getOpcode() == ISD::AND) {
8304 // case (2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
8305 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
8308 unsigned Mask2 = N11C->getZExtValue();
8310 // Mask and ~Mask2 (or reverse) must be equivalent for the BFI pattern
8312 if (ARM::isBitFieldInvertedMask(Mask) &&
8314 // The pack halfword instruction works better for masks that fit it,
8315 // so use that when it's available.
8316 if (Subtarget->hasT2ExtractPack() &&
8317 (Mask == 0xffff || Mask == 0xffff0000))
8320 unsigned amt = countTrailingZeros(Mask2);
8321 Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0),
8322 DAG.getConstant(amt, MVT::i32));
8323 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00, Res,
8324 DAG.getConstant(Mask, MVT::i32));
8325 // Do not add new nodes to DAG combiner worklist.
8326 DCI.CombineTo(N, Res, false);
8328 } else if (ARM::isBitFieldInvertedMask(~Mask) &&
8330 // The pack halfword instruction works better for masks that fit it,
8331 // so use that when it's available.
8332 if (Subtarget->hasT2ExtractPack() &&
8333 (Mask2 == 0xffff || Mask2 == 0xffff0000))
8336 unsigned lsb = countTrailingZeros(Mask);
8337 Res = DAG.getNode(ISD::SRL, DL, VT, N00,
8338 DAG.getConstant(lsb, MVT::i32));
8339 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res,
8340 DAG.getConstant(Mask2, MVT::i32));
8341 // Do not add new nodes to DAG combiner worklist.
8342 DCI.CombineTo(N, Res, false);
8347 if (DAG.MaskedValueIsZero(N1, MaskC->getAPIntValue()) &&
8348 N00.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N00.getOperand(1)) &&
8349 ARM::isBitFieldInvertedMask(~Mask)) {
8350 // Case (3): or (and (shl A, #shamt), mask), B => ARMbfi B, A, ~mask
8351 // where lsb(mask) == #shamt and masked bits of B are known zero.
8352 SDValue ShAmt = N00.getOperand(1);
8353 unsigned ShAmtC = cast<ConstantSDNode>(ShAmt)->getZExtValue();
8354 unsigned LSB = countTrailingZeros(Mask);
8358 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1, N00.getOperand(0),
8359 DAG.getConstant(~Mask, MVT::i32));
8361 // Do not add new nodes to DAG combiner worklist.
8362 DCI.CombineTo(N, Res, false);
8368 static SDValue PerformXORCombine(SDNode *N,
8369 TargetLowering::DAGCombinerInfo &DCI,
8370 const ARMSubtarget *Subtarget) {
8371 EVT VT = N->getValueType(0);
8372 SelectionDAG &DAG = DCI.DAG;
8374 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8377 if (!Subtarget->isThumb1Only()) {
8378 // fold (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c))
8379 SDValue Result = combineSelectAndUseCommutative(N, false, DCI);
8380 if (Result.getNode())
8387 /// PerformBFICombine - (bfi A, (and B, Mask1), Mask2) -> (bfi A, B, Mask2) iff
8388 /// the bits being cleared by the AND are not demanded by the BFI.
8389 static SDValue PerformBFICombine(SDNode *N,
8390 TargetLowering::DAGCombinerInfo &DCI) {
8391 SDValue N1 = N->getOperand(1);
8392 if (N1.getOpcode() == ISD::AND) {
8393 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
8396 unsigned InvMask = cast<ConstantSDNode>(N->getOperand(2))->getZExtValue();
8397 unsigned LSB = countTrailingZeros(~InvMask);
8398 unsigned Width = (32 - countLeadingZeros(~InvMask)) - LSB;
8399 unsigned Mask = (1 << Width)-1;
8400 unsigned Mask2 = N11C->getZExtValue();
8401 if ((Mask & (~Mask2)) == 0)
8402 return DCI.DAG.getNode(ARMISD::BFI, SDLoc(N), N->getValueType(0),
8403 N->getOperand(0), N1.getOperand(0),
8409 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
8410 /// ARMISD::VMOVRRD.
8411 static SDValue PerformVMOVRRDCombine(SDNode *N,
8412 TargetLowering::DAGCombinerInfo &DCI) {
8413 // vmovrrd(vmovdrr x, y) -> x,y
8414 SDValue InDouble = N->getOperand(0);
8415 if (InDouble.getOpcode() == ARMISD::VMOVDRR)
8416 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
8418 // vmovrrd(load f64) -> (load i32), (load i32)
8419 SDNode *InNode = InDouble.getNode();
8420 if (ISD::isNormalLoad(InNode) && InNode->hasOneUse() &&
8421 InNode->getValueType(0) == MVT::f64 &&
8422 InNode->getOperand(1).getOpcode() == ISD::FrameIndex &&
8423 !cast<LoadSDNode>(InNode)->isVolatile()) {
8424 // TODO: Should this be done for non-FrameIndex operands?
8425 LoadSDNode *LD = cast<LoadSDNode>(InNode);
8427 SelectionDAG &DAG = DCI.DAG;
8429 SDValue BasePtr = LD->getBasePtr();
8430 SDValue NewLD1 = DAG.getLoad(MVT::i32, DL, LD->getChain(), BasePtr,
8431 LD->getPointerInfo(), LD->isVolatile(),
8432 LD->isNonTemporal(), LD->isInvariant(),
8433 LD->getAlignment());
8435 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
8436 DAG.getConstant(4, MVT::i32));
8437 SDValue NewLD2 = DAG.getLoad(MVT::i32, DL, NewLD1.getValue(1), OffsetPtr,
8438 LD->getPointerInfo(), LD->isVolatile(),
8439 LD->isNonTemporal(), LD->isInvariant(),
8440 std::min(4U, LD->getAlignment() / 2));
8442 DAG.ReplaceAllUsesOfValueWith(SDValue(LD, 1), NewLD2.getValue(1));
8443 if (DCI.DAG.getTargetLoweringInfo().isBigEndian())
8444 std::swap (NewLD1, NewLD2);
8445 SDValue Result = DCI.CombineTo(N, NewLD1, NewLD2);
8446 DCI.RemoveFromWorklist(LD);
8454 /// PerformVMOVDRRCombine - Target-specific dag combine xforms for
8455 /// ARMISD::VMOVDRR. This is also used for BUILD_VECTORs with 2 operands.
8456 static SDValue PerformVMOVDRRCombine(SDNode *N, SelectionDAG &DAG) {
8457 // N=vmovrrd(X); vmovdrr(N:0, N:1) -> bit_convert(X)
8458 SDValue Op0 = N->getOperand(0);
8459 SDValue Op1 = N->getOperand(1);
8460 if (Op0.getOpcode() == ISD::BITCAST)
8461 Op0 = Op0.getOperand(0);
8462 if (Op1.getOpcode() == ISD::BITCAST)
8463 Op1 = Op1.getOperand(0);
8464 if (Op0.getOpcode() == ARMISD::VMOVRRD &&
8465 Op0.getNode() == Op1.getNode() &&
8466 Op0.getResNo() == 0 && Op1.getResNo() == 1)
8467 return DAG.getNode(ISD::BITCAST, SDLoc(N),
8468 N->getValueType(0), Op0.getOperand(0));
8472 /// PerformSTORECombine - Target-specific dag combine xforms for
8474 static SDValue PerformSTORECombine(SDNode *N,
8475 TargetLowering::DAGCombinerInfo &DCI) {
8476 StoreSDNode *St = cast<StoreSDNode>(N);
8477 if (St->isVolatile())
8480 // Optimize trunc store (of multiple scalars) to shuffle and store. First,
8481 // pack all of the elements in one place. Next, store to memory in fewer
8483 SDValue StVal = St->getValue();
8484 EVT VT = StVal.getValueType();
8485 if (St->isTruncatingStore() && VT.isVector()) {
8486 SelectionDAG &DAG = DCI.DAG;
8487 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8488 EVT StVT = St->getMemoryVT();
8489 unsigned NumElems = VT.getVectorNumElements();
8490 assert(StVT != VT && "Cannot truncate to the same type");
8491 unsigned FromEltSz = VT.getVectorElementType().getSizeInBits();
8492 unsigned ToEltSz = StVT.getVectorElementType().getSizeInBits();
8494 // From, To sizes and ElemCount must be pow of two
8495 if (!isPowerOf2_32(NumElems * FromEltSz * ToEltSz)) return SDValue();
8497 // We are going to use the original vector elt for storing.
8498 // Accumulated smaller vector elements must be a multiple of the store size.
8499 if (0 != (NumElems * FromEltSz) % ToEltSz) return SDValue();
8501 unsigned SizeRatio = FromEltSz / ToEltSz;
8502 assert(SizeRatio * NumElems * ToEltSz == VT.getSizeInBits());
8504 // Create a type on which we perform the shuffle.
8505 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(), StVT.getScalarType(),
8506 NumElems*SizeRatio);
8507 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
8510 SDValue WideVec = DAG.getNode(ISD::BITCAST, DL, WideVecVT, StVal);
8511 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
8512 for (unsigned i = 0; i < NumElems; ++i)
8513 ShuffleVec[i] = TLI.isBigEndian() ? (i+1) * SizeRatio - 1 : i * SizeRatio;
8515 // Can't shuffle using an illegal type.
8516 if (!TLI.isTypeLegal(WideVecVT)) return SDValue();
8518 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, DL, WideVec,
8519 DAG.getUNDEF(WideVec.getValueType()),
8521 // At this point all of the data is stored at the bottom of the
8522 // register. We now need to save it to mem.
8524 // Find the largest store unit
8525 MVT StoreType = MVT::i8;
8526 for (unsigned tp = MVT::FIRST_INTEGER_VALUETYPE;
8527 tp < MVT::LAST_INTEGER_VALUETYPE; ++tp) {
8528 MVT Tp = (MVT::SimpleValueType)tp;
8529 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToEltSz)
8532 // Didn't find a legal store type.
8533 if (!TLI.isTypeLegal(StoreType))
8536 // Bitcast the original vector into a vector of store-size units
8537 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
8538 StoreType, VT.getSizeInBits()/EVT(StoreType).getSizeInBits());
8539 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
8540 SDValue ShuffWide = DAG.getNode(ISD::BITCAST, DL, StoreVecVT, Shuff);
8541 SmallVector<SDValue, 8> Chains;
8542 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits()/8,
8543 TLI.getPointerTy());
8544 SDValue BasePtr = St->getBasePtr();
8546 // Perform one or more big stores into memory.
8547 unsigned E = (ToEltSz*NumElems)/StoreType.getSizeInBits();
8548 for (unsigned I = 0; I < E; I++) {
8549 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
8550 StoreType, ShuffWide,
8551 DAG.getIntPtrConstant(I));
8552 SDValue Ch = DAG.getStore(St->getChain(), DL, SubVec, BasePtr,
8553 St->getPointerInfo(), St->isVolatile(),
8554 St->isNonTemporal(), St->getAlignment());
8555 BasePtr = DAG.getNode(ISD::ADD, DL, BasePtr.getValueType(), BasePtr,
8557 Chains.push_back(Ch);
8559 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
8562 if (!ISD::isNormalStore(St))
8565 // Split a store of a VMOVDRR into two integer stores to avoid mixing NEON and
8566 // ARM stores of arguments in the same cache line.
8567 if (StVal.getNode()->getOpcode() == ARMISD::VMOVDRR &&
8568 StVal.getNode()->hasOneUse()) {
8569 SelectionDAG &DAG = DCI.DAG;
8570 bool isBigEndian = DAG.getTargetLoweringInfo().isBigEndian();
8572 SDValue BasePtr = St->getBasePtr();
8573 SDValue NewST1 = DAG.getStore(St->getChain(), DL,
8574 StVal.getNode()->getOperand(isBigEndian ? 1 : 0 ),
8575 BasePtr, St->getPointerInfo(), St->isVolatile(),
8576 St->isNonTemporal(), St->getAlignment());
8578 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
8579 DAG.getConstant(4, MVT::i32));
8580 return DAG.getStore(NewST1.getValue(0), DL,
8581 StVal.getNode()->getOperand(isBigEndian ? 0 : 1),
8582 OffsetPtr, St->getPointerInfo(), St->isVolatile(),
8583 St->isNonTemporal(),
8584 std::min(4U, St->getAlignment() / 2));
8587 if (StVal.getValueType() != MVT::i64 ||
8588 StVal.getNode()->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
8591 // Bitcast an i64 store extracted from a vector to f64.
8592 // Otherwise, the i64 value will be legalized to a pair of i32 values.
8593 SelectionDAG &DAG = DCI.DAG;
8595 SDValue IntVec = StVal.getOperand(0);
8596 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
8597 IntVec.getValueType().getVectorNumElements());
8598 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, IntVec);
8599 SDValue ExtElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
8600 Vec, StVal.getOperand(1));
8602 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ExtElt);
8603 // Make the DAGCombiner fold the bitcasts.
8604 DCI.AddToWorklist(Vec.getNode());
8605 DCI.AddToWorklist(ExtElt.getNode());
8606 DCI.AddToWorklist(V.getNode());
8607 return DAG.getStore(St->getChain(), dl, V, St->getBasePtr(),
8608 St->getPointerInfo(), St->isVolatile(),
8609 St->isNonTemporal(), St->getAlignment(),
8613 /// hasNormalLoadOperand - Check if any of the operands of a BUILD_VECTOR node
8614 /// are normal, non-volatile loads. If so, it is profitable to bitcast an
8615 /// i64 vector to have f64 elements, since the value can then be loaded
8616 /// directly into a VFP register.
8617 static bool hasNormalLoadOperand(SDNode *N) {
8618 unsigned NumElts = N->getValueType(0).getVectorNumElements();
8619 for (unsigned i = 0; i < NumElts; ++i) {
8620 SDNode *Elt = N->getOperand(i).getNode();
8621 if (ISD::isNormalLoad(Elt) && !cast<LoadSDNode>(Elt)->isVolatile())
8627 /// PerformBUILD_VECTORCombine - Target-specific dag combine xforms for
8628 /// ISD::BUILD_VECTOR.
8629 static SDValue PerformBUILD_VECTORCombine(SDNode *N,
8630 TargetLowering::DAGCombinerInfo &DCI){
8631 // build_vector(N=ARMISD::VMOVRRD(X), N:1) -> bit_convert(X):
8632 // VMOVRRD is introduced when legalizing i64 types. It forces the i64 value
8633 // into a pair of GPRs, which is fine when the value is used as a scalar,
8634 // but if the i64 value is converted to a vector, we need to undo the VMOVRRD.
8635 SelectionDAG &DAG = DCI.DAG;
8636 if (N->getNumOperands() == 2) {
8637 SDValue RV = PerformVMOVDRRCombine(N, DAG);
8642 // Load i64 elements as f64 values so that type legalization does not split
8643 // them up into i32 values.
8644 EVT VT = N->getValueType(0);
8645 if (VT.getVectorElementType() != MVT::i64 || !hasNormalLoadOperand(N))
8648 SmallVector<SDValue, 8> Ops;
8649 unsigned NumElts = VT.getVectorNumElements();
8650 for (unsigned i = 0; i < NumElts; ++i) {
8651 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(i));
8653 // Make the DAGCombiner fold the bitcast.
8654 DCI.AddToWorklist(V.getNode());
8656 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, NumElts);
8657 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, FloatVT, Ops);
8658 return DAG.getNode(ISD::BITCAST, dl, VT, BV);
8661 /// \brief Target-specific dag combine xforms for ARMISD::BUILD_VECTOR.
8663 PerformARMBUILD_VECTORCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
8664 // ARMISD::BUILD_VECTOR is introduced when legalizing ISD::BUILD_VECTOR.
8665 // At that time, we may have inserted bitcasts from integer to float.
8666 // If these bitcasts have survived DAGCombine, change the lowering of this
8667 // BUILD_VECTOR in something more vector friendly, i.e., that does not
8668 // force to use floating point types.
8670 // Make sure we can change the type of the vector.
8671 // This is possible iff:
8672 // 1. The vector is only used in a bitcast to a integer type. I.e.,
8673 // 1.1. Vector is used only once.
8674 // 1.2. Use is a bit convert to an integer type.
8675 // 2. The size of its operands are 32-bits (64-bits are not legal).
8676 EVT VT = N->getValueType(0);
8677 EVT EltVT = VT.getVectorElementType();
8679 // Check 1.1. and 2.
8680 if (EltVT.getSizeInBits() != 32 || !N->hasOneUse())
8683 // By construction, the input type must be float.
8684 assert(EltVT == MVT::f32 && "Unexpected type!");
8687 SDNode *Use = *N->use_begin();
8688 if (Use->getOpcode() != ISD::BITCAST ||
8689 Use->getValueType(0).isFloatingPoint())
8692 // Check profitability.
8693 // Model is, if more than half of the relevant operands are bitcast from
8694 // i32, turn the build_vector into a sequence of insert_vector_elt.
8695 // Relevant operands are everything that is not statically
8696 // (i.e., at compile time) bitcasted.
8697 unsigned NumOfBitCastedElts = 0;
8698 unsigned NumElts = VT.getVectorNumElements();
8699 unsigned NumOfRelevantElts = NumElts;
8700 for (unsigned Idx = 0; Idx < NumElts; ++Idx) {
8701 SDValue Elt = N->getOperand(Idx);
8702 if (Elt->getOpcode() == ISD::BITCAST) {
8703 // Assume only bit cast to i32 will go away.
8704 if (Elt->getOperand(0).getValueType() == MVT::i32)
8705 ++NumOfBitCastedElts;
8706 } else if (Elt.getOpcode() == ISD::UNDEF || isa<ConstantSDNode>(Elt))
8707 // Constants are statically casted, thus do not count them as
8708 // relevant operands.
8709 --NumOfRelevantElts;
8712 // Check if more than half of the elements require a non-free bitcast.
8713 if (NumOfBitCastedElts <= NumOfRelevantElts / 2)
8716 SelectionDAG &DAG = DCI.DAG;
8717 // Create the new vector type.
8718 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
8719 // Check if the type is legal.
8720 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8721 if (!TLI.isTypeLegal(VecVT))
8725 // ARMISD::BUILD_VECTOR E1, E2, ..., EN.
8726 // => BITCAST INSERT_VECTOR_ELT
8727 // (INSERT_VECTOR_ELT (...), (BITCAST EN-1), N-1),
8729 SDValue Vec = DAG.getUNDEF(VecVT);
8731 for (unsigned Idx = 0 ; Idx < NumElts; ++Idx) {
8732 SDValue V = N->getOperand(Idx);
8733 if (V.getOpcode() == ISD::UNDEF)
8735 if (V.getOpcode() == ISD::BITCAST &&
8736 V->getOperand(0).getValueType() == MVT::i32)
8737 // Fold obvious case.
8738 V = V.getOperand(0);
8740 V = DAG.getNode(ISD::BITCAST, SDLoc(V), MVT::i32, V);
8741 // Make the DAGCombiner fold the bitcasts.
8742 DCI.AddToWorklist(V.getNode());
8744 SDValue LaneIdx = DAG.getConstant(Idx, MVT::i32);
8745 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VecVT, Vec, V, LaneIdx);
8747 Vec = DAG.getNode(ISD::BITCAST, dl, VT, Vec);
8748 // Make the DAGCombiner fold the bitcasts.
8749 DCI.AddToWorklist(Vec.getNode());
8753 /// PerformInsertEltCombine - Target-specific dag combine xforms for
8754 /// ISD::INSERT_VECTOR_ELT.
8755 static SDValue PerformInsertEltCombine(SDNode *N,
8756 TargetLowering::DAGCombinerInfo &DCI) {
8757 // Bitcast an i64 load inserted into a vector to f64.
8758 // Otherwise, the i64 value will be legalized to a pair of i32 values.
8759 EVT VT = N->getValueType(0);
8760 SDNode *Elt = N->getOperand(1).getNode();
8761 if (VT.getVectorElementType() != MVT::i64 ||
8762 !ISD::isNormalLoad(Elt) || cast<LoadSDNode>(Elt)->isVolatile())
8765 SelectionDAG &DAG = DCI.DAG;
8767 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
8768 VT.getVectorNumElements());
8769 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, N->getOperand(0));
8770 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(1));
8771 // Make the DAGCombiner fold the bitcasts.
8772 DCI.AddToWorklist(Vec.getNode());
8773 DCI.AddToWorklist(V.getNode());
8774 SDValue InsElt = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, FloatVT,
8775 Vec, V, N->getOperand(2));
8776 return DAG.getNode(ISD::BITCAST, dl, VT, InsElt);
8779 /// PerformVECTOR_SHUFFLECombine - Target-specific dag combine xforms for
8780 /// ISD::VECTOR_SHUFFLE.
8781 static SDValue PerformVECTOR_SHUFFLECombine(SDNode *N, SelectionDAG &DAG) {
8782 // The LLVM shufflevector instruction does not require the shuffle mask
8783 // length to match the operand vector length, but ISD::VECTOR_SHUFFLE does
8784 // have that requirement. When translating to ISD::VECTOR_SHUFFLE, if the
8785 // operands do not match the mask length, they are extended by concatenating
8786 // them with undef vectors. That is probably the right thing for other
8787 // targets, but for NEON it is better to concatenate two double-register
8788 // size vector operands into a single quad-register size vector. Do that
8789 // transformation here:
8790 // shuffle(concat(v1, undef), concat(v2, undef)) ->
8791 // shuffle(concat(v1, v2), undef)
8792 SDValue Op0 = N->getOperand(0);
8793 SDValue Op1 = N->getOperand(1);
8794 if (Op0.getOpcode() != ISD::CONCAT_VECTORS ||
8795 Op1.getOpcode() != ISD::CONCAT_VECTORS ||
8796 Op0.getNumOperands() != 2 ||
8797 Op1.getNumOperands() != 2)
8799 SDValue Concat0Op1 = Op0.getOperand(1);
8800 SDValue Concat1Op1 = Op1.getOperand(1);
8801 if (Concat0Op1.getOpcode() != ISD::UNDEF ||
8802 Concat1Op1.getOpcode() != ISD::UNDEF)
8804 // Skip the transformation if any of the types are illegal.
8805 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8806 EVT VT = N->getValueType(0);
8807 if (!TLI.isTypeLegal(VT) ||
8808 !TLI.isTypeLegal(Concat0Op1.getValueType()) ||
8809 !TLI.isTypeLegal(Concat1Op1.getValueType()))
8812 SDValue NewConcat = DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT,
8813 Op0.getOperand(0), Op1.getOperand(0));
8814 // Translate the shuffle mask.
8815 SmallVector<int, 16> NewMask;
8816 unsigned NumElts = VT.getVectorNumElements();
8817 unsigned HalfElts = NumElts/2;
8818 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
8819 for (unsigned n = 0; n < NumElts; ++n) {
8820 int MaskElt = SVN->getMaskElt(n);
8822 if (MaskElt < (int)HalfElts)
8824 else if (MaskElt >= (int)NumElts && MaskElt < (int)(NumElts + HalfElts))
8825 NewElt = HalfElts + MaskElt - NumElts;
8826 NewMask.push_back(NewElt);
8828 return DAG.getVectorShuffle(VT, SDLoc(N), NewConcat,
8829 DAG.getUNDEF(VT), NewMask.data());
8832 /// CombineBaseUpdate - Target-specific DAG combine function for VLDDUP and
8833 /// NEON load/store intrinsics to merge base address updates.
8834 static SDValue CombineBaseUpdate(SDNode *N,
8835 TargetLowering::DAGCombinerInfo &DCI) {
8836 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
8839 SelectionDAG &DAG = DCI.DAG;
8840 bool isIntrinsic = (N->getOpcode() == ISD::INTRINSIC_VOID ||
8841 N->getOpcode() == ISD::INTRINSIC_W_CHAIN);
8842 unsigned AddrOpIdx = (isIntrinsic ? 2 : 1);
8843 SDValue Addr = N->getOperand(AddrOpIdx);
8845 // Search for a use of the address operand that is an increment.
8846 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(),
8847 UE = Addr.getNode()->use_end(); UI != UE; ++UI) {
8849 if (User->getOpcode() != ISD::ADD ||
8850 UI.getUse().getResNo() != Addr.getResNo())
8853 // Check that the add is independent of the load/store. Otherwise, folding
8854 // it would create a cycle.
8855 if (User->isPredecessorOf(N) || N->isPredecessorOf(User))
8858 // Find the new opcode for the updating load/store.
8860 bool isLaneOp = false;
8861 unsigned NewOpc = 0;
8862 unsigned NumVecs = 0;
8864 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
8866 default: llvm_unreachable("unexpected intrinsic for Neon base update");
8867 case Intrinsic::arm_neon_vld1: NewOpc = ARMISD::VLD1_UPD;
8869 case Intrinsic::arm_neon_vld2: NewOpc = ARMISD::VLD2_UPD;
8871 case Intrinsic::arm_neon_vld3: NewOpc = ARMISD::VLD3_UPD;
8873 case Intrinsic::arm_neon_vld4: NewOpc = ARMISD::VLD4_UPD;
8875 case Intrinsic::arm_neon_vld2lane: NewOpc = ARMISD::VLD2LN_UPD;
8876 NumVecs = 2; isLaneOp = true; break;
8877 case Intrinsic::arm_neon_vld3lane: NewOpc = ARMISD::VLD3LN_UPD;
8878 NumVecs = 3; isLaneOp = true; break;
8879 case Intrinsic::arm_neon_vld4lane: NewOpc = ARMISD::VLD4LN_UPD;
8880 NumVecs = 4; isLaneOp = true; break;
8881 case Intrinsic::arm_neon_vst1: NewOpc = ARMISD::VST1_UPD;
8882 NumVecs = 1; isLoad = false; break;
8883 case Intrinsic::arm_neon_vst2: NewOpc = ARMISD::VST2_UPD;
8884 NumVecs = 2; isLoad = false; break;
8885 case Intrinsic::arm_neon_vst3: NewOpc = ARMISD::VST3_UPD;
8886 NumVecs = 3; isLoad = false; break;
8887 case Intrinsic::arm_neon_vst4: NewOpc = ARMISD::VST4_UPD;
8888 NumVecs = 4; isLoad = false; break;
8889 case Intrinsic::arm_neon_vst2lane: NewOpc = ARMISD::VST2LN_UPD;
8890 NumVecs = 2; isLoad = false; isLaneOp = true; break;
8891 case Intrinsic::arm_neon_vst3lane: NewOpc = ARMISD::VST3LN_UPD;
8892 NumVecs = 3; isLoad = false; isLaneOp = true; break;
8893 case Intrinsic::arm_neon_vst4lane: NewOpc = ARMISD::VST4LN_UPD;
8894 NumVecs = 4; isLoad = false; isLaneOp = true; break;
8898 switch (N->getOpcode()) {
8899 default: llvm_unreachable("unexpected opcode for Neon base update");
8900 case ARMISD::VLD2DUP: NewOpc = ARMISD::VLD2DUP_UPD; NumVecs = 2; break;
8901 case ARMISD::VLD3DUP: NewOpc = ARMISD::VLD3DUP_UPD; NumVecs = 3; break;
8902 case ARMISD::VLD4DUP: NewOpc = ARMISD::VLD4DUP_UPD; NumVecs = 4; break;
8906 // Find the size of memory referenced by the load/store.
8909 VecTy = N->getValueType(0);
8911 VecTy = N->getOperand(AddrOpIdx+1).getValueType();
8912 unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8;
8914 NumBytes /= VecTy.getVectorNumElements();
8916 // If the increment is a constant, it must match the memory ref size.
8917 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
8918 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
8919 uint64_t IncVal = CInc->getZExtValue();
8920 if (IncVal != NumBytes)
8922 } else if (NumBytes >= 3 * 16) {
8923 // VLD3/4 and VST3/4 for 128-bit vectors are implemented with two
8924 // separate instructions that make it harder to use a non-constant update.
8928 // Create the new updating load/store node.
8930 unsigned NumResultVecs = (isLoad ? NumVecs : 0);
8932 for (n = 0; n < NumResultVecs; ++n)
8934 Tys[n++] = MVT::i32;
8935 Tys[n] = MVT::Other;
8936 SDVTList SDTys = DAG.getVTList(ArrayRef<EVT>(Tys, NumResultVecs+2));
8937 SmallVector<SDValue, 8> Ops;
8938 Ops.push_back(N->getOperand(0)); // incoming chain
8939 Ops.push_back(N->getOperand(AddrOpIdx));
8941 for (unsigned i = AddrOpIdx + 1; i < N->getNumOperands(); ++i) {
8942 Ops.push_back(N->getOperand(i));
8944 MemIntrinsicSDNode *MemInt = cast<MemIntrinsicSDNode>(N);
8945 SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, SDLoc(N), SDTys,
8946 Ops, MemInt->getMemoryVT(),
8947 MemInt->getMemOperand());
8950 std::vector<SDValue> NewResults;
8951 for (unsigned i = 0; i < NumResultVecs; ++i) {
8952 NewResults.push_back(SDValue(UpdN.getNode(), i));
8954 NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs+1)); // chain
8955 DCI.CombineTo(N, NewResults);
8956 DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs));
8963 /// CombineVLDDUP - For a VDUPLANE node N, check if its source operand is a
8964 /// vldN-lane (N > 1) intrinsic, and if all the other uses of that intrinsic
8965 /// are also VDUPLANEs. If so, combine them to a vldN-dup operation and
8967 static bool CombineVLDDUP(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
8968 SelectionDAG &DAG = DCI.DAG;
8969 EVT VT = N->getValueType(0);
8970 // vldN-dup instructions only support 64-bit vectors for N > 1.
8971 if (!VT.is64BitVector())
8974 // Check if the VDUPLANE operand is a vldN-dup intrinsic.
8975 SDNode *VLD = N->getOperand(0).getNode();
8976 if (VLD->getOpcode() != ISD::INTRINSIC_W_CHAIN)
8978 unsigned NumVecs = 0;
8979 unsigned NewOpc = 0;
8980 unsigned IntNo = cast<ConstantSDNode>(VLD->getOperand(1))->getZExtValue();
8981 if (IntNo == Intrinsic::arm_neon_vld2lane) {
8983 NewOpc = ARMISD::VLD2DUP;
8984 } else if (IntNo == Intrinsic::arm_neon_vld3lane) {
8986 NewOpc = ARMISD::VLD3DUP;
8987 } else if (IntNo == Intrinsic::arm_neon_vld4lane) {
8989 NewOpc = ARMISD::VLD4DUP;
8994 // First check that all the vldN-lane uses are VDUPLANEs and that the lane
8995 // numbers match the load.
8996 unsigned VLDLaneNo =
8997 cast<ConstantSDNode>(VLD->getOperand(NumVecs+3))->getZExtValue();
8998 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
9000 // Ignore uses of the chain result.
9001 if (UI.getUse().getResNo() == NumVecs)
9004 if (User->getOpcode() != ARMISD::VDUPLANE ||
9005 VLDLaneNo != cast<ConstantSDNode>(User->getOperand(1))->getZExtValue())
9009 // Create the vldN-dup node.
9012 for (n = 0; n < NumVecs; ++n)
9014 Tys[n] = MVT::Other;
9015 SDVTList SDTys = DAG.getVTList(ArrayRef<EVT>(Tys, NumVecs+1));
9016 SDValue Ops[] = { VLD->getOperand(0), VLD->getOperand(2) };
9017 MemIntrinsicSDNode *VLDMemInt = cast<MemIntrinsicSDNode>(VLD);
9018 SDValue VLDDup = DAG.getMemIntrinsicNode(NewOpc, SDLoc(VLD), SDTys,
9019 Ops, VLDMemInt->getMemoryVT(),
9020 VLDMemInt->getMemOperand());
9023 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
9025 unsigned ResNo = UI.getUse().getResNo();
9026 // Ignore uses of the chain result.
9027 if (ResNo == NumVecs)
9030 DCI.CombineTo(User, SDValue(VLDDup.getNode(), ResNo));
9033 // Now the vldN-lane intrinsic is dead except for its chain result.
9034 // Update uses of the chain.
9035 std::vector<SDValue> VLDDupResults;
9036 for (unsigned n = 0; n < NumVecs; ++n)
9037 VLDDupResults.push_back(SDValue(VLDDup.getNode(), n));
9038 VLDDupResults.push_back(SDValue(VLDDup.getNode(), NumVecs));
9039 DCI.CombineTo(VLD, VLDDupResults);
9044 /// PerformVDUPLANECombine - Target-specific dag combine xforms for
9045 /// ARMISD::VDUPLANE.
9046 static SDValue PerformVDUPLANECombine(SDNode *N,
9047 TargetLowering::DAGCombinerInfo &DCI) {
9048 SDValue Op = N->getOperand(0);
9050 // If the source is a vldN-lane (N > 1) intrinsic, and all the other uses
9051 // of that intrinsic are also VDUPLANEs, combine them to a vldN-dup operation.
9052 if (CombineVLDDUP(N, DCI))
9053 return SDValue(N, 0);
9055 // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is
9056 // redundant. Ignore bit_converts for now; element sizes are checked below.
9057 while (Op.getOpcode() == ISD::BITCAST)
9058 Op = Op.getOperand(0);
9059 if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM)
9062 // Make sure the VMOV element size is not bigger than the VDUPLANE elements.
9063 unsigned EltSize = Op.getValueType().getVectorElementType().getSizeInBits();
9064 // The canonical VMOV for a zero vector uses a 32-bit element size.
9065 unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
9067 if (ARM_AM::decodeNEONModImm(Imm, EltBits) == 0)
9069 EVT VT = N->getValueType(0);
9070 if (EltSize > VT.getVectorElementType().getSizeInBits())
9073 return DCI.DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Op);
9076 // isConstVecPow2 - Return true if each vector element is a power of 2, all
9077 // elements are the same constant, C, and Log2(C) ranges from 1 to 32.
9078 static bool isConstVecPow2(SDValue ConstVec, bool isSigned, uint64_t &C)
9082 for (unsigned I = 0, E = ConstVec.getValueType().getVectorNumElements();
9084 ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(ConstVec.getOperand(I));
9089 APFloat APF = C->getValueAPF();
9090 if (APF.convertToInteger(&cN, 64, isSigned, APFloat::rmTowardZero, &isExact)
9091 != APFloat::opOK || !isExact)
9094 c0 = (I == 0) ? cN : c0;
9095 if (!isPowerOf2_64(cN) || c0 != cN || Log2_64(c0) < 1 || Log2_64(c0) > 32)
9102 /// PerformVCVTCombine - VCVT (floating-point to fixed-point, Advanced SIMD)
9103 /// can replace combinations of VMUL and VCVT (floating-point to integer)
9104 /// when the VMUL has a constant operand that is a power of 2.
9106 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
9107 /// vmul.f32 d16, d17, d16
9108 /// vcvt.s32.f32 d16, d16
9110 /// vcvt.s32.f32 d16, d16, #3
9111 static SDValue PerformVCVTCombine(SDNode *N,
9112 TargetLowering::DAGCombinerInfo &DCI,
9113 const ARMSubtarget *Subtarget) {
9114 SelectionDAG &DAG = DCI.DAG;
9115 SDValue Op = N->getOperand(0);
9117 if (!Subtarget->hasNEON() || !Op.getValueType().isVector() ||
9118 Op.getOpcode() != ISD::FMUL)
9122 SDValue N0 = Op->getOperand(0);
9123 SDValue ConstVec = Op->getOperand(1);
9124 bool isSigned = N->getOpcode() == ISD::FP_TO_SINT;
9126 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
9127 !isConstVecPow2(ConstVec, isSigned, C))
9130 MVT FloatTy = Op.getSimpleValueType().getVectorElementType();
9131 MVT IntTy = N->getSimpleValueType(0).getVectorElementType();
9132 if (FloatTy.getSizeInBits() != 32 || IntTy.getSizeInBits() > 32) {
9133 // These instructions only exist converting from f32 to i32. We can handle
9134 // smaller integers by generating an extra truncate, but larger ones would
9139 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfp2fxs :
9140 Intrinsic::arm_neon_vcvtfp2fxu;
9141 unsigned NumLanes = Op.getValueType().getVectorNumElements();
9142 SDValue FixConv = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N),
9143 NumLanes == 2 ? MVT::v2i32 : MVT::v4i32,
9144 DAG.getConstant(IntrinsicOpcode, MVT::i32), N0,
9145 DAG.getConstant(Log2_64(C), MVT::i32));
9147 if (IntTy.getSizeInBits() < FloatTy.getSizeInBits())
9148 FixConv = DAG.getNode(ISD::TRUNCATE, SDLoc(N), N->getValueType(0), FixConv);
9153 /// PerformVDIVCombine - VCVT (fixed-point to floating-point, Advanced SIMD)
9154 /// can replace combinations of VCVT (integer to floating-point) and VDIV
9155 /// when the VDIV has a constant operand that is a power of 2.
9157 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
9158 /// vcvt.f32.s32 d16, d16
9159 /// vdiv.f32 d16, d17, d16
9161 /// vcvt.f32.s32 d16, d16, #3
9162 static SDValue PerformVDIVCombine(SDNode *N,
9163 TargetLowering::DAGCombinerInfo &DCI,
9164 const ARMSubtarget *Subtarget) {
9165 SelectionDAG &DAG = DCI.DAG;
9166 SDValue Op = N->getOperand(0);
9167 unsigned OpOpcode = Op.getNode()->getOpcode();
9169 if (!Subtarget->hasNEON() || !N->getValueType(0).isVector() ||
9170 (OpOpcode != ISD::SINT_TO_FP && OpOpcode != ISD::UINT_TO_FP))
9174 SDValue ConstVec = N->getOperand(1);
9175 bool isSigned = OpOpcode == ISD::SINT_TO_FP;
9177 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
9178 !isConstVecPow2(ConstVec, isSigned, C))
9181 MVT FloatTy = N->getSimpleValueType(0).getVectorElementType();
9182 MVT IntTy = Op.getOperand(0).getSimpleValueType().getVectorElementType();
9183 if (FloatTy.getSizeInBits() != 32 || IntTy.getSizeInBits() > 32) {
9184 // These instructions only exist converting from i32 to f32. We can handle
9185 // smaller integers by generating an extra extend, but larger ones would
9190 SDValue ConvInput = Op.getOperand(0);
9191 unsigned NumLanes = Op.getValueType().getVectorNumElements();
9192 if (IntTy.getSizeInBits() < FloatTy.getSizeInBits())
9193 ConvInput = DAG.getNode(isSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND,
9194 SDLoc(N), NumLanes == 2 ? MVT::v2i32 : MVT::v4i32,
9197 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfxs2fp :
9198 Intrinsic::arm_neon_vcvtfxu2fp;
9199 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N),
9201 DAG.getConstant(IntrinsicOpcode, MVT::i32),
9202 ConvInput, DAG.getConstant(Log2_64(C), MVT::i32));
9205 /// Getvshiftimm - Check if this is a valid build_vector for the immediate
9206 /// operand of a vector shift operation, where all the elements of the
9207 /// build_vector must have the same constant integer value.
9208 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
9209 // Ignore bit_converts.
9210 while (Op.getOpcode() == ISD::BITCAST)
9211 Op = Op.getOperand(0);
9212 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
9213 APInt SplatBits, SplatUndef;
9214 unsigned SplatBitSize;
9216 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
9217 HasAnyUndefs, ElementBits) ||
9218 SplatBitSize > ElementBits)
9220 Cnt = SplatBits.getSExtValue();
9224 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
9225 /// operand of a vector shift left operation. That value must be in the range:
9226 /// 0 <= Value < ElementBits for a left shift; or
9227 /// 0 <= Value <= ElementBits for a long left shift.
9228 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
9229 assert(VT.isVector() && "vector shift count is not a vector type");
9230 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
9231 if (! getVShiftImm(Op, ElementBits, Cnt))
9233 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
9236 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
9237 /// operand of a vector shift right operation. For a shift opcode, the value
9238 /// is positive, but for an intrinsic the value count must be negative. The
9239 /// absolute value must be in the range:
9240 /// 1 <= |Value| <= ElementBits for a right shift; or
9241 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
9242 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
9244 assert(VT.isVector() && "vector shift count is not a vector type");
9245 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
9246 if (! getVShiftImm(Op, ElementBits, Cnt))
9250 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
9253 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
9254 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
9255 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
9258 // Don't do anything for most intrinsics.
9261 // Vector shifts: check for immediate versions and lower them.
9262 // Note: This is done during DAG combining instead of DAG legalizing because
9263 // the build_vectors for 64-bit vector element shift counts are generally
9264 // not legal, and it is hard to see their values after they get legalized to
9265 // loads from a constant pool.
9266 case Intrinsic::arm_neon_vshifts:
9267 case Intrinsic::arm_neon_vshiftu:
9268 case Intrinsic::arm_neon_vrshifts:
9269 case Intrinsic::arm_neon_vrshiftu:
9270 case Intrinsic::arm_neon_vrshiftn:
9271 case Intrinsic::arm_neon_vqshifts:
9272 case Intrinsic::arm_neon_vqshiftu:
9273 case Intrinsic::arm_neon_vqshiftsu:
9274 case Intrinsic::arm_neon_vqshiftns:
9275 case Intrinsic::arm_neon_vqshiftnu:
9276 case Intrinsic::arm_neon_vqshiftnsu:
9277 case Intrinsic::arm_neon_vqrshiftns:
9278 case Intrinsic::arm_neon_vqrshiftnu:
9279 case Intrinsic::arm_neon_vqrshiftnsu: {
9280 EVT VT = N->getOperand(1).getValueType();
9282 unsigned VShiftOpc = 0;
9285 case Intrinsic::arm_neon_vshifts:
9286 case Intrinsic::arm_neon_vshiftu:
9287 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
9288 VShiftOpc = ARMISD::VSHL;
9291 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
9292 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
9293 ARMISD::VSHRs : ARMISD::VSHRu);
9298 case Intrinsic::arm_neon_vrshifts:
9299 case Intrinsic::arm_neon_vrshiftu:
9300 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
9304 case Intrinsic::arm_neon_vqshifts:
9305 case Intrinsic::arm_neon_vqshiftu:
9306 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
9310 case Intrinsic::arm_neon_vqshiftsu:
9311 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
9313 llvm_unreachable("invalid shift count for vqshlu intrinsic");
9315 case Intrinsic::arm_neon_vrshiftn:
9316 case Intrinsic::arm_neon_vqshiftns:
9317 case Intrinsic::arm_neon_vqshiftnu:
9318 case Intrinsic::arm_neon_vqshiftnsu:
9319 case Intrinsic::arm_neon_vqrshiftns:
9320 case Intrinsic::arm_neon_vqrshiftnu:
9321 case Intrinsic::arm_neon_vqrshiftnsu:
9322 // Narrowing shifts require an immediate right shift.
9323 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
9325 llvm_unreachable("invalid shift count for narrowing vector shift "
9329 llvm_unreachable("unhandled vector shift");
9333 case Intrinsic::arm_neon_vshifts:
9334 case Intrinsic::arm_neon_vshiftu:
9335 // Opcode already set above.
9337 case Intrinsic::arm_neon_vrshifts:
9338 VShiftOpc = ARMISD::VRSHRs; break;
9339 case Intrinsic::arm_neon_vrshiftu:
9340 VShiftOpc = ARMISD::VRSHRu; break;
9341 case Intrinsic::arm_neon_vrshiftn:
9342 VShiftOpc = ARMISD::VRSHRN; break;
9343 case Intrinsic::arm_neon_vqshifts:
9344 VShiftOpc = ARMISD::VQSHLs; break;
9345 case Intrinsic::arm_neon_vqshiftu:
9346 VShiftOpc = ARMISD::VQSHLu; break;
9347 case Intrinsic::arm_neon_vqshiftsu:
9348 VShiftOpc = ARMISD::VQSHLsu; break;
9349 case Intrinsic::arm_neon_vqshiftns:
9350 VShiftOpc = ARMISD::VQSHRNs; break;
9351 case Intrinsic::arm_neon_vqshiftnu:
9352 VShiftOpc = ARMISD::VQSHRNu; break;
9353 case Intrinsic::arm_neon_vqshiftnsu:
9354 VShiftOpc = ARMISD::VQSHRNsu; break;
9355 case Intrinsic::arm_neon_vqrshiftns:
9356 VShiftOpc = ARMISD::VQRSHRNs; break;
9357 case Intrinsic::arm_neon_vqrshiftnu:
9358 VShiftOpc = ARMISD::VQRSHRNu; break;
9359 case Intrinsic::arm_neon_vqrshiftnsu:
9360 VShiftOpc = ARMISD::VQRSHRNsu; break;
9363 return DAG.getNode(VShiftOpc, SDLoc(N), N->getValueType(0),
9364 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
9367 case Intrinsic::arm_neon_vshiftins: {
9368 EVT VT = N->getOperand(1).getValueType();
9370 unsigned VShiftOpc = 0;
9372 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
9373 VShiftOpc = ARMISD::VSLI;
9374 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
9375 VShiftOpc = ARMISD::VSRI;
9377 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
9380 return DAG.getNode(VShiftOpc, SDLoc(N), N->getValueType(0),
9381 N->getOperand(1), N->getOperand(2),
9382 DAG.getConstant(Cnt, MVT::i32));
9385 case Intrinsic::arm_neon_vqrshifts:
9386 case Intrinsic::arm_neon_vqrshiftu:
9387 // No immediate versions of these to check for.
9394 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
9395 /// lowers them. As with the vector shift intrinsics, this is done during DAG
9396 /// combining instead of DAG legalizing because the build_vectors for 64-bit
9397 /// vector element shift counts are generally not legal, and it is hard to see
9398 /// their values after they get legalized to loads from a constant pool.
9399 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
9400 const ARMSubtarget *ST) {
9401 EVT VT = N->getValueType(0);
9402 if (N->getOpcode() == ISD::SRL && VT == MVT::i32 && ST->hasV6Ops()) {
9403 // Canonicalize (srl (bswap x), 16) to (rotr (bswap x), 16) if the high
9404 // 16-bits of x is zero. This optimizes rev + lsr 16 to rev16.
9405 SDValue N1 = N->getOperand(1);
9406 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) {
9407 SDValue N0 = N->getOperand(0);
9408 if (C->getZExtValue() == 16 && N0.getOpcode() == ISD::BSWAP &&
9409 DAG.MaskedValueIsZero(N0.getOperand(0),
9410 APInt::getHighBitsSet(32, 16)))
9411 return DAG.getNode(ISD::ROTR, SDLoc(N), VT, N0, N1);
9415 // Nothing to be done for scalar shifts.
9416 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9417 if (!VT.isVector() || !TLI.isTypeLegal(VT))
9420 assert(ST->hasNEON() && "unexpected vector shift");
9423 switch (N->getOpcode()) {
9424 default: llvm_unreachable("unexpected shift opcode");
9427 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
9428 return DAG.getNode(ARMISD::VSHL, SDLoc(N), VT, N->getOperand(0),
9429 DAG.getConstant(Cnt, MVT::i32));
9434 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
9435 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
9436 ARMISD::VSHRs : ARMISD::VSHRu);
9437 return DAG.getNode(VShiftOpc, SDLoc(N), VT, N->getOperand(0),
9438 DAG.getConstant(Cnt, MVT::i32));
9444 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
9445 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
9446 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
9447 const ARMSubtarget *ST) {
9448 SDValue N0 = N->getOperand(0);
9450 // Check for sign- and zero-extensions of vector extract operations of 8-
9451 // and 16-bit vector elements. NEON supports these directly. They are
9452 // handled during DAG combining because type legalization will promote them
9453 // to 32-bit types and it is messy to recognize the operations after that.
9454 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
9455 SDValue Vec = N0.getOperand(0);
9456 SDValue Lane = N0.getOperand(1);
9457 EVT VT = N->getValueType(0);
9458 EVT EltVT = N0.getValueType();
9459 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9461 if (VT == MVT::i32 &&
9462 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
9463 TLI.isTypeLegal(Vec.getValueType()) &&
9464 isa<ConstantSDNode>(Lane)) {
9467 switch (N->getOpcode()) {
9468 default: llvm_unreachable("unexpected opcode");
9469 case ISD::SIGN_EXTEND:
9470 Opc = ARMISD::VGETLANEs;
9472 case ISD::ZERO_EXTEND:
9473 case ISD::ANY_EXTEND:
9474 Opc = ARMISD::VGETLANEu;
9477 return DAG.getNode(Opc, SDLoc(N), VT, Vec, Lane);
9484 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
9485 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
9486 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
9487 const ARMSubtarget *ST) {
9488 // If the target supports NEON, try to use vmax/vmin instructions for f32
9489 // selects like "x < y ? x : y". Unless the NoNaNsFPMath option is set,
9490 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
9491 // a NaN; only do the transformation when it matches that behavior.
9493 // For now only do this when using NEON for FP operations; if using VFP, it
9494 // is not obvious that the benefit outweighs the cost of switching to the
9496 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
9497 N->getValueType(0) != MVT::f32)
9500 SDValue CondLHS = N->getOperand(0);
9501 SDValue CondRHS = N->getOperand(1);
9502 SDValue LHS = N->getOperand(2);
9503 SDValue RHS = N->getOperand(3);
9504 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
9506 unsigned Opcode = 0;
9508 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
9509 IsReversed = false; // x CC y ? x : y
9510 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
9511 IsReversed = true ; // x CC y ? y : x
9525 // If LHS is NaN, an ordered comparison will be false and the result will
9526 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
9527 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
9528 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
9529 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
9531 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
9532 // will return -0, so vmin can only be used for unsafe math or if one of
9533 // the operands is known to be nonzero.
9534 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
9535 !DAG.getTarget().Options.UnsafeFPMath &&
9536 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9538 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
9547 // If LHS is NaN, an ordered comparison will be false and the result will
9548 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
9549 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
9550 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
9551 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
9553 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
9554 // will return +0, so vmax can only be used for unsafe math or if one of
9555 // the operands is known to be nonzero.
9556 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
9557 !DAG.getTarget().Options.UnsafeFPMath &&
9558 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9560 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
9566 return DAG.getNode(Opcode, SDLoc(N), N->getValueType(0), LHS, RHS);
9569 /// PerformCMOVCombine - Target-specific DAG combining for ARMISD::CMOV.
9571 ARMTargetLowering::PerformCMOVCombine(SDNode *N, SelectionDAG &DAG) const {
9572 SDValue Cmp = N->getOperand(4);
9573 if (Cmp.getOpcode() != ARMISD::CMPZ)
9574 // Only looking at EQ and NE cases.
9577 EVT VT = N->getValueType(0);
9579 SDValue LHS = Cmp.getOperand(0);
9580 SDValue RHS = Cmp.getOperand(1);
9581 SDValue FalseVal = N->getOperand(0);
9582 SDValue TrueVal = N->getOperand(1);
9583 SDValue ARMcc = N->getOperand(2);
9584 ARMCC::CondCodes CC =
9585 (ARMCC::CondCodes)cast<ConstantSDNode>(ARMcc)->getZExtValue();
9603 /// FIXME: Turn this into a target neutral optimization?
9605 if (CC == ARMCC::NE && FalseVal == RHS && FalseVal != LHS) {
9606 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, TrueVal, ARMcc,
9607 N->getOperand(3), Cmp);
9608 } else if (CC == ARMCC::EQ && TrueVal == RHS) {
9610 SDValue NewCmp = getARMCmp(LHS, RHS, ISD::SETNE, ARMcc, DAG, dl);
9611 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, FalseVal, ARMcc,
9612 N->getOperand(3), NewCmp);
9615 if (Res.getNode()) {
9616 APInt KnownZero, KnownOne;
9617 DAG.computeKnownBits(SDValue(N,0), KnownZero, KnownOne);
9618 // Capture demanded bits information that would be otherwise lost.
9619 if (KnownZero == 0xfffffffe)
9620 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
9621 DAG.getValueType(MVT::i1));
9622 else if (KnownZero == 0xffffff00)
9623 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
9624 DAG.getValueType(MVT::i8));
9625 else if (KnownZero == 0xffff0000)
9626 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
9627 DAG.getValueType(MVT::i16));
9633 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
9634 DAGCombinerInfo &DCI) const {
9635 switch (N->getOpcode()) {
9637 case ISD::ADDC: return PerformADDCCombine(N, DCI, Subtarget);
9638 case ISD::ADD: return PerformADDCombine(N, DCI, Subtarget);
9639 case ISD::SUB: return PerformSUBCombine(N, DCI);
9640 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
9641 case ISD::OR: return PerformORCombine(N, DCI, Subtarget);
9642 case ISD::XOR: return PerformXORCombine(N, DCI, Subtarget);
9643 case ISD::AND: return PerformANDCombine(N, DCI, Subtarget);
9644 case ARMISD::BFI: return PerformBFICombine(N, DCI);
9645 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
9646 case ARMISD::VMOVDRR: return PerformVMOVDRRCombine(N, DCI.DAG);
9647 case ISD::STORE: return PerformSTORECombine(N, DCI);
9648 case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DCI);
9649 case ISD::INSERT_VECTOR_ELT: return PerformInsertEltCombine(N, DCI);
9650 case ISD::VECTOR_SHUFFLE: return PerformVECTOR_SHUFFLECombine(N, DCI.DAG);
9651 case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI);
9652 case ISD::FP_TO_SINT:
9653 case ISD::FP_TO_UINT: return PerformVCVTCombine(N, DCI, Subtarget);
9654 case ISD::FDIV: return PerformVDIVCombine(N, DCI, Subtarget);
9655 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
9658 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
9659 case ISD::SIGN_EXTEND:
9660 case ISD::ZERO_EXTEND:
9661 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
9662 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
9663 case ARMISD::CMOV: return PerformCMOVCombine(N, DCI.DAG);
9664 case ARMISD::VLD2DUP:
9665 case ARMISD::VLD3DUP:
9666 case ARMISD::VLD4DUP:
9667 return CombineBaseUpdate(N, DCI);
9668 case ARMISD::BUILD_VECTOR:
9669 return PerformARMBUILD_VECTORCombine(N, DCI);
9670 case ISD::INTRINSIC_VOID:
9671 case ISD::INTRINSIC_W_CHAIN:
9672 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
9673 case Intrinsic::arm_neon_vld1:
9674 case Intrinsic::arm_neon_vld2:
9675 case Intrinsic::arm_neon_vld3:
9676 case Intrinsic::arm_neon_vld4:
9677 case Intrinsic::arm_neon_vld2lane:
9678 case Intrinsic::arm_neon_vld3lane:
9679 case Intrinsic::arm_neon_vld4lane:
9680 case Intrinsic::arm_neon_vst1:
9681 case Intrinsic::arm_neon_vst2:
9682 case Intrinsic::arm_neon_vst3:
9683 case Intrinsic::arm_neon_vst4:
9684 case Intrinsic::arm_neon_vst2lane:
9685 case Intrinsic::arm_neon_vst3lane:
9686 case Intrinsic::arm_neon_vst4lane:
9687 return CombineBaseUpdate(N, DCI);
9695 bool ARMTargetLowering::isDesirableToTransformToIntegerOp(unsigned Opc,
9697 return (VT == MVT::f32) && (Opc == ISD::LOAD || Opc == ISD::STORE);
9700 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT, unsigned,
9702 // The AllowsUnaliged flag models the SCTLR.A setting in ARM cpus
9703 bool AllowsUnaligned = Subtarget->allowsUnalignedMem();
9705 switch (VT.getSimpleVT().SimpleTy) {
9711 // Unaligned access can use (for example) LRDB, LRDH, LDR
9712 if (AllowsUnaligned) {
9714 *Fast = Subtarget->hasV7Ops();
9721 // For any little-endian targets with neon, we can support unaligned ld/st
9722 // of D and Q (e.g. {D0,D1}) registers by using vld1.i8/vst1.i8.
9723 // A big-endian target may also explicitly support unaligned accesses
9724 if (Subtarget->hasNEON() && (AllowsUnaligned || isLittleEndian())) {
9734 static bool memOpAlign(unsigned DstAlign, unsigned SrcAlign,
9735 unsigned AlignCheck) {
9736 return ((SrcAlign == 0 || SrcAlign % AlignCheck == 0) &&
9737 (DstAlign == 0 || DstAlign % AlignCheck == 0));
9740 EVT ARMTargetLowering::getOptimalMemOpType(uint64_t Size,
9741 unsigned DstAlign, unsigned SrcAlign,
9742 bool IsMemset, bool ZeroMemset,
9744 MachineFunction &MF) const {
9745 const Function *F = MF.getFunction();
9747 // See if we can use NEON instructions for this...
9748 if ((!IsMemset || ZeroMemset) &&
9749 Subtarget->hasNEON() &&
9750 !F->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
9751 Attribute::NoImplicitFloat)) {
9754 (memOpAlign(SrcAlign, DstAlign, 16) ||
9755 (allowsUnalignedMemoryAccesses(MVT::v2f64, 0, &Fast) && Fast))) {
9757 } else if (Size >= 8 &&
9758 (memOpAlign(SrcAlign, DstAlign, 8) ||
9759 (allowsUnalignedMemoryAccesses(MVT::f64, 0, &Fast) && Fast))) {
9764 // Lowering to i32/i16 if the size permits.
9770 // Let the target-independent logic figure it out.
9774 bool ARMTargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
9775 if (Val.getOpcode() != ISD::LOAD)
9778 EVT VT1 = Val.getValueType();
9779 if (!VT1.isSimple() || !VT1.isInteger() ||
9780 !VT2.isSimple() || !VT2.isInteger())
9783 switch (VT1.getSimpleVT().SimpleTy) {
9788 // 8-bit and 16-bit loads implicitly zero-extend to 32-bits.
9795 bool ARMTargetLowering::allowTruncateForTailCall(Type *Ty1, Type *Ty2) const {
9796 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
9799 if (!isTypeLegal(EVT::getEVT(Ty1)))
9802 assert(Ty1->getPrimitiveSizeInBits() <= 64 && "i128 is probably not a noop");
9804 // Assuming the caller doesn't have a zeroext or signext return parameter,
9805 // truncation all the way down to i1 is valid.
9810 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
9815 switch (VT.getSimpleVT().SimpleTy) {
9816 default: return false;
9831 if ((V & (Scale - 1)) != 0)
9834 return V == (V & ((1LL << 5) - 1));
9837 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
9838 const ARMSubtarget *Subtarget) {
9845 switch (VT.getSimpleVT().SimpleTy) {
9846 default: return false;
9851 // + imm12 or - imm8
9853 return V == (V & ((1LL << 8) - 1));
9854 return V == (V & ((1LL << 12) - 1));
9857 // Same as ARM mode. FIXME: NEON?
9858 if (!Subtarget->hasVFP2())
9863 return V == (V & ((1LL << 8) - 1));
9867 /// isLegalAddressImmediate - Return true if the integer value can be used
9868 /// as the offset of the target addressing mode for load / store of the
9870 static bool isLegalAddressImmediate(int64_t V, EVT VT,
9871 const ARMSubtarget *Subtarget) {
9878 if (Subtarget->isThumb1Only())
9879 return isLegalT1AddressImmediate(V, VT);
9880 else if (Subtarget->isThumb2())
9881 return isLegalT2AddressImmediate(V, VT, Subtarget);
9886 switch (VT.getSimpleVT().SimpleTy) {
9887 default: return false;
9892 return V == (V & ((1LL << 12) - 1));
9895 return V == (V & ((1LL << 8) - 1));
9898 if (!Subtarget->hasVFP2()) // FIXME: NEON?
9903 return V == (V & ((1LL << 8) - 1));
9907 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
9909 int Scale = AM.Scale;
9913 switch (VT.getSimpleVT().SimpleTy) {
9914 default: return false;
9923 return Scale == 2 || Scale == 4 || Scale == 8;
9926 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
9930 // Note, we allow "void" uses (basically, uses that aren't loads or
9931 // stores), because arm allows folding a scale into many arithmetic
9932 // operations. This should be made more precise and revisited later.
9934 // Allow r << imm, but the imm has to be a multiple of two.
9935 if (Scale & 1) return false;
9936 return isPowerOf2_32(Scale);
9940 /// isLegalAddressingMode - Return true if the addressing mode represented
9941 /// by AM is legal for this target, for a load/store of the specified type.
9942 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
9944 EVT VT = getValueType(Ty, true);
9945 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
9948 // Can never fold addr of global into load/store.
9953 case 0: // no scale reg, must be "r+i" or "r", or "i".
9956 if (Subtarget->isThumb1Only())
9960 // ARM doesn't support any R+R*scale+imm addr modes.
9967 if (Subtarget->isThumb2())
9968 return isLegalT2ScaledAddressingMode(AM, VT);
9970 int Scale = AM.Scale;
9971 switch (VT.getSimpleVT().SimpleTy) {
9972 default: return false;
9976 if (Scale < 0) Scale = -Scale;
9980 return isPowerOf2_32(Scale & ~1);
9984 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
9989 // Note, we allow "void" uses (basically, uses that aren't loads or
9990 // stores), because arm allows folding a scale into many arithmetic
9991 // operations. This should be made more precise and revisited later.
9993 // Allow r << imm, but the imm has to be a multiple of two.
9994 if (Scale & 1) return false;
9995 return isPowerOf2_32(Scale);
10001 /// isLegalICmpImmediate - Return true if the specified immediate is legal
10002 /// icmp immediate, that is the target has icmp instructions which can compare
10003 /// a register against the immediate without having to materialize the
10004 /// immediate into a register.
10005 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
10006 // Thumb2 and ARM modes can use cmn for negative immediates.
10007 if (!Subtarget->isThumb())
10008 return ARM_AM::getSOImmVal(llvm::abs64(Imm)) != -1;
10009 if (Subtarget->isThumb2())
10010 return ARM_AM::getT2SOImmVal(llvm::abs64(Imm)) != -1;
10011 // Thumb1 doesn't have cmn, and only 8-bit immediates.
10012 return Imm >= 0 && Imm <= 255;
10015 /// isLegalAddImmediate - Return true if the specified immediate is a legal add
10016 /// *or sub* immediate, that is the target has add or sub instructions which can
10017 /// add a register with the immediate without having to materialize the
10018 /// immediate into a register.
10019 bool ARMTargetLowering::isLegalAddImmediate(int64_t Imm) const {
10020 // Same encoding for add/sub, just flip the sign.
10021 int64_t AbsImm = llvm::abs64(Imm);
10022 if (!Subtarget->isThumb())
10023 return ARM_AM::getSOImmVal(AbsImm) != -1;
10024 if (Subtarget->isThumb2())
10025 return ARM_AM::getT2SOImmVal(AbsImm) != -1;
10026 // Thumb1 only has 8-bit unsigned immediate.
10027 return AbsImm >= 0 && AbsImm <= 255;
10030 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
10031 bool isSEXTLoad, SDValue &Base,
10032 SDValue &Offset, bool &isInc,
10033 SelectionDAG &DAG) {
10034 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
10037 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
10038 // AddressingMode 3
10039 Base = Ptr->getOperand(0);
10040 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
10041 int RHSC = (int)RHS->getZExtValue();
10042 if (RHSC < 0 && RHSC > -256) {
10043 assert(Ptr->getOpcode() == ISD::ADD);
10045 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
10049 isInc = (Ptr->getOpcode() == ISD::ADD);
10050 Offset = Ptr->getOperand(1);
10052 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
10053 // AddressingMode 2
10054 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
10055 int RHSC = (int)RHS->getZExtValue();
10056 if (RHSC < 0 && RHSC > -0x1000) {
10057 assert(Ptr->getOpcode() == ISD::ADD);
10059 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
10060 Base = Ptr->getOperand(0);
10065 if (Ptr->getOpcode() == ISD::ADD) {
10067 ARM_AM::ShiftOpc ShOpcVal=
10068 ARM_AM::getShiftOpcForNode(Ptr->getOperand(0).getOpcode());
10069 if (ShOpcVal != ARM_AM::no_shift) {
10070 Base = Ptr->getOperand(1);
10071 Offset = Ptr->getOperand(0);
10073 Base = Ptr->getOperand(0);
10074 Offset = Ptr->getOperand(1);
10079 isInc = (Ptr->getOpcode() == ISD::ADD);
10080 Base = Ptr->getOperand(0);
10081 Offset = Ptr->getOperand(1);
10085 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
10089 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
10090 bool isSEXTLoad, SDValue &Base,
10091 SDValue &Offset, bool &isInc,
10092 SelectionDAG &DAG) {
10093 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
10096 Base = Ptr->getOperand(0);
10097 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
10098 int RHSC = (int)RHS->getZExtValue();
10099 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
10100 assert(Ptr->getOpcode() == ISD::ADD);
10102 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
10104 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
10105 isInc = Ptr->getOpcode() == ISD::ADD;
10106 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
10114 /// getPreIndexedAddressParts - returns true by value, base pointer and
10115 /// offset pointer and addressing mode by reference if the node's address
10116 /// can be legally represented as pre-indexed load / store address.
10118 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
10120 ISD::MemIndexedMode &AM,
10121 SelectionDAG &DAG) const {
10122 if (Subtarget->isThumb1Only())
10127 bool isSEXTLoad = false;
10128 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
10129 Ptr = LD->getBasePtr();
10130 VT = LD->getMemoryVT();
10131 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
10132 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
10133 Ptr = ST->getBasePtr();
10134 VT = ST->getMemoryVT();
10139 bool isLegal = false;
10140 if (Subtarget->isThumb2())
10141 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
10142 Offset, isInc, DAG);
10144 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
10145 Offset, isInc, DAG);
10149 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
10153 /// getPostIndexedAddressParts - returns true by value, base pointer and
10154 /// offset pointer and addressing mode by reference if this node can be
10155 /// combined with a load / store to form a post-indexed load / store.
10156 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
10159 ISD::MemIndexedMode &AM,
10160 SelectionDAG &DAG) const {
10161 if (Subtarget->isThumb1Only())
10166 bool isSEXTLoad = false;
10167 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
10168 VT = LD->getMemoryVT();
10169 Ptr = LD->getBasePtr();
10170 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
10171 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
10172 VT = ST->getMemoryVT();
10173 Ptr = ST->getBasePtr();
10178 bool isLegal = false;
10179 if (Subtarget->isThumb2())
10180 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
10183 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
10189 // Swap base ptr and offset to catch more post-index load / store when
10190 // it's legal. In Thumb2 mode, offset must be an immediate.
10191 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
10192 !Subtarget->isThumb2())
10193 std::swap(Base, Offset);
10195 // Post-indexed load / store update the base pointer.
10200 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
10204 void ARMTargetLowering::computeKnownBitsForTargetNode(const SDValue Op,
10207 const SelectionDAG &DAG,
10208 unsigned Depth) const {
10209 unsigned BitWidth = KnownOne.getBitWidth();
10210 KnownZero = KnownOne = APInt(BitWidth, 0);
10211 switch (Op.getOpcode()) {
10217 // These nodes' second result is a boolean
10218 if (Op.getResNo() == 0)
10220 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - 1);
10222 case ARMISD::CMOV: {
10223 // Bits are known zero/one if known on the LHS and RHS.
10224 DAG.computeKnownBits(Op.getOperand(0), KnownZero, KnownOne, Depth+1);
10225 if (KnownZero == 0 && KnownOne == 0) return;
10227 APInt KnownZeroRHS, KnownOneRHS;
10228 DAG.computeKnownBits(Op.getOperand(1), KnownZeroRHS, KnownOneRHS, Depth+1);
10229 KnownZero &= KnownZeroRHS;
10230 KnownOne &= KnownOneRHS;
10233 case ISD::INTRINSIC_W_CHAIN: {
10234 ConstantSDNode *CN = cast<ConstantSDNode>(Op->getOperand(1));
10235 Intrinsic::ID IntID = static_cast<Intrinsic::ID>(CN->getZExtValue());
10238 case Intrinsic::arm_ldaex:
10239 case Intrinsic::arm_ldrex: {
10240 EVT VT = cast<MemIntrinsicSDNode>(Op)->getMemoryVT();
10241 unsigned MemBits = VT.getScalarType().getSizeInBits();
10242 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - MemBits);
10250 //===----------------------------------------------------------------------===//
10251 // ARM Inline Assembly Support
10252 //===----------------------------------------------------------------------===//
10254 bool ARMTargetLowering::ExpandInlineAsm(CallInst *CI) const {
10255 // Looking for "rev" which is V6+.
10256 if (!Subtarget->hasV6Ops())
10259 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
10260 std::string AsmStr = IA->getAsmString();
10261 SmallVector<StringRef, 4> AsmPieces;
10262 SplitString(AsmStr, AsmPieces, ";\n");
10264 switch (AsmPieces.size()) {
10265 default: return false;
10267 AsmStr = AsmPieces[0];
10269 SplitString(AsmStr, AsmPieces, " \t,");
10272 if (AsmPieces.size() == 3 &&
10273 AsmPieces[0] == "rev" && AsmPieces[1] == "$0" && AsmPieces[2] == "$1" &&
10274 IA->getConstraintString().compare(0, 4, "=l,l") == 0) {
10275 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
10276 if (Ty && Ty->getBitWidth() == 32)
10277 return IntrinsicLowering::LowerToByteSwap(CI);
10285 /// getConstraintType - Given a constraint letter, return the type of
10286 /// constraint it is for this target.
10287 ARMTargetLowering::ConstraintType
10288 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
10289 if (Constraint.size() == 1) {
10290 switch (Constraint[0]) {
10292 case 'l': return C_RegisterClass;
10293 case 'w': return C_RegisterClass;
10294 case 'h': return C_RegisterClass;
10295 case 'x': return C_RegisterClass;
10296 case 't': return C_RegisterClass;
10297 case 'j': return C_Other; // Constant for movw.
10298 // An address with a single base register. Due to the way we
10299 // currently handle addresses it is the same as an 'r' memory constraint.
10300 case 'Q': return C_Memory;
10302 } else if (Constraint.size() == 2) {
10303 switch (Constraint[0]) {
10305 // All 'U+' constraints are addresses.
10306 case 'U': return C_Memory;
10309 return TargetLowering::getConstraintType(Constraint);
10312 /// Examine constraint type and operand type and determine a weight value.
10313 /// This object must already have been set up with the operand type
10314 /// and the current alternative constraint selected.
10315 TargetLowering::ConstraintWeight
10316 ARMTargetLowering::getSingleConstraintMatchWeight(
10317 AsmOperandInfo &info, const char *constraint) const {
10318 ConstraintWeight weight = CW_Invalid;
10319 Value *CallOperandVal = info.CallOperandVal;
10320 // If we don't have a value, we can't do a match,
10321 // but allow it at the lowest weight.
10322 if (!CallOperandVal)
10324 Type *type = CallOperandVal->getType();
10325 // Look at the constraint type.
10326 switch (*constraint) {
10328 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
10331 if (type->isIntegerTy()) {
10332 if (Subtarget->isThumb())
10333 weight = CW_SpecificReg;
10335 weight = CW_Register;
10339 if (type->isFloatingPointTy())
10340 weight = CW_Register;
10346 typedef std::pair<unsigned, const TargetRegisterClass*> RCPair;
10348 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
10350 if (Constraint.size() == 1) {
10351 // GCC ARM Constraint Letters
10352 switch (Constraint[0]) {
10353 case 'l': // Low regs or general regs.
10354 if (Subtarget->isThumb())
10355 return RCPair(0U, &ARM::tGPRRegClass);
10356 return RCPair(0U, &ARM::GPRRegClass);
10357 case 'h': // High regs or no regs.
10358 if (Subtarget->isThumb())
10359 return RCPair(0U, &ARM::hGPRRegClass);
10362 return RCPair(0U, &ARM::GPRRegClass);
10364 if (VT == MVT::Other)
10366 if (VT == MVT::f32)
10367 return RCPair(0U, &ARM::SPRRegClass);
10368 if (VT.getSizeInBits() == 64)
10369 return RCPair(0U, &ARM::DPRRegClass);
10370 if (VT.getSizeInBits() == 128)
10371 return RCPair(0U, &ARM::QPRRegClass);
10374 if (VT == MVT::Other)
10376 if (VT == MVT::f32)
10377 return RCPair(0U, &ARM::SPR_8RegClass);
10378 if (VT.getSizeInBits() == 64)
10379 return RCPair(0U, &ARM::DPR_8RegClass);
10380 if (VT.getSizeInBits() == 128)
10381 return RCPair(0U, &ARM::QPR_8RegClass);
10384 if (VT == MVT::f32)
10385 return RCPair(0U, &ARM::SPRRegClass);
10389 if (StringRef("{cc}").equals_lower(Constraint))
10390 return std::make_pair(unsigned(ARM::CPSR), &ARM::CCRRegClass);
10392 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
10395 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
10396 /// vector. If it is invalid, don't add anything to Ops.
10397 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
10398 std::string &Constraint,
10399 std::vector<SDValue>&Ops,
10400 SelectionDAG &DAG) const {
10403 // Currently only support length 1 constraints.
10404 if (Constraint.length() != 1) return;
10406 char ConstraintLetter = Constraint[0];
10407 switch (ConstraintLetter) {
10410 case 'I': case 'J': case 'K': case 'L':
10411 case 'M': case 'N': case 'O':
10412 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
10416 int64_t CVal64 = C->getSExtValue();
10417 int CVal = (int) CVal64;
10418 // None of these constraints allow values larger than 32 bits. Check
10419 // that the value fits in an int.
10420 if (CVal != CVal64)
10423 switch (ConstraintLetter) {
10425 // Constant suitable for movw, must be between 0 and
10427 if (Subtarget->hasV6T2Ops())
10428 if (CVal >= 0 && CVal <= 65535)
10432 if (Subtarget->isThumb1Only()) {
10433 // This must be a constant between 0 and 255, for ADD
10435 if (CVal >= 0 && CVal <= 255)
10437 } else if (Subtarget->isThumb2()) {
10438 // A constant that can be used as an immediate value in a
10439 // data-processing instruction.
10440 if (ARM_AM::getT2SOImmVal(CVal) != -1)
10443 // A constant that can be used as an immediate value in a
10444 // data-processing instruction.
10445 if (ARM_AM::getSOImmVal(CVal) != -1)
10451 if (Subtarget->isThumb()) { // FIXME thumb2
10452 // This must be a constant between -255 and -1, for negated ADD
10453 // immediates. This can be used in GCC with an "n" modifier that
10454 // prints the negated value, for use with SUB instructions. It is
10455 // not useful otherwise but is implemented for compatibility.
10456 if (CVal >= -255 && CVal <= -1)
10459 // This must be a constant between -4095 and 4095. It is not clear
10460 // what this constraint is intended for. Implemented for
10461 // compatibility with GCC.
10462 if (CVal >= -4095 && CVal <= 4095)
10468 if (Subtarget->isThumb1Only()) {
10469 // A 32-bit value where only one byte has a nonzero value. Exclude
10470 // zero to match GCC. This constraint is used by GCC internally for
10471 // constants that can be loaded with a move/shift combination.
10472 // It is not useful otherwise but is implemented for compatibility.
10473 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
10475 } else if (Subtarget->isThumb2()) {
10476 // A constant whose bitwise inverse can be used as an immediate
10477 // value in a data-processing instruction. This can be used in GCC
10478 // with a "B" modifier that prints the inverted value, for use with
10479 // BIC and MVN instructions. It is not useful otherwise but is
10480 // implemented for compatibility.
10481 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
10484 // A constant whose bitwise inverse can be used as an immediate
10485 // value in a data-processing instruction. This can be used in GCC
10486 // with a "B" modifier that prints the inverted value, for use with
10487 // BIC and MVN instructions. It is not useful otherwise but is
10488 // implemented for compatibility.
10489 if (ARM_AM::getSOImmVal(~CVal) != -1)
10495 if (Subtarget->isThumb1Only()) {
10496 // This must be a constant between -7 and 7,
10497 // for 3-operand ADD/SUB immediate instructions.
10498 if (CVal >= -7 && CVal < 7)
10500 } else if (Subtarget->isThumb2()) {
10501 // A constant whose negation can be used as an immediate value in a
10502 // data-processing instruction. This can be used in GCC with an "n"
10503 // modifier that prints the negated value, for use with SUB
10504 // instructions. It is not useful otherwise but is implemented for
10506 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
10509 // A constant whose negation can be used as an immediate value in a
10510 // data-processing instruction. This can be used in GCC with an "n"
10511 // modifier that prints the negated value, for use with SUB
10512 // instructions. It is not useful otherwise but is implemented for
10514 if (ARM_AM::getSOImmVal(-CVal) != -1)
10520 if (Subtarget->isThumb()) { // FIXME thumb2
10521 // This must be a multiple of 4 between 0 and 1020, for
10522 // ADD sp + immediate.
10523 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
10526 // A power of two or a constant between 0 and 32. This is used in
10527 // GCC for the shift amount on shifted register operands, but it is
10528 // useful in general for any shift amounts.
10529 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
10535 if (Subtarget->isThumb()) { // FIXME thumb2
10536 // This must be a constant between 0 and 31, for shift amounts.
10537 if (CVal >= 0 && CVal <= 31)
10543 if (Subtarget->isThumb()) { // FIXME thumb2
10544 // This must be a multiple of 4 between -508 and 508, for
10545 // ADD/SUB sp = sp + immediate.
10546 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
10551 Result = DAG.getTargetConstant(CVal, Op.getValueType());
10555 if (Result.getNode()) {
10556 Ops.push_back(Result);
10559 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
10562 SDValue ARMTargetLowering::LowerDivRem(SDValue Op, SelectionDAG &DAG) const {
10563 assert(Subtarget->isTargetAEABI() && "Register-based DivRem lowering only");
10564 unsigned Opcode = Op->getOpcode();
10565 assert((Opcode == ISD::SDIVREM || Opcode == ISD::UDIVREM) &&
10566 "Invalid opcode for Div/Rem lowering");
10567 bool isSigned = (Opcode == ISD::SDIVREM);
10568 EVT VT = Op->getValueType(0);
10569 Type *Ty = VT.getTypeForEVT(*DAG.getContext());
10572 switch (VT.getSimpleVT().SimpleTy) {
10573 default: llvm_unreachable("Unexpected request for libcall!");
10574 case MVT::i8: LC= isSigned ? RTLIB::SDIVREM_I8 : RTLIB::UDIVREM_I8; break;
10575 case MVT::i16: LC= isSigned ? RTLIB::SDIVREM_I16 : RTLIB::UDIVREM_I16; break;
10576 case MVT::i32: LC= isSigned ? RTLIB::SDIVREM_I32 : RTLIB::UDIVREM_I32; break;
10577 case MVT::i64: LC= isSigned ? RTLIB::SDIVREM_I64 : RTLIB::UDIVREM_I64; break;
10580 SDValue InChain = DAG.getEntryNode();
10582 TargetLowering::ArgListTy Args;
10583 TargetLowering::ArgListEntry Entry;
10584 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i) {
10585 EVT ArgVT = Op->getOperand(i).getValueType();
10586 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
10587 Entry.Node = Op->getOperand(i);
10589 Entry.isSExt = isSigned;
10590 Entry.isZExt = !isSigned;
10591 Args.push_back(Entry);
10594 SDValue Callee = DAG.getExternalSymbol(getLibcallName(LC),
10597 Type *RetTy = (Type*)StructType::get(Ty, Ty, NULL);
10600 TargetLowering::CallLoweringInfo CLI(DAG);
10601 CLI.setDebugLoc(dl).setChain(InChain)
10602 .setCallee(getLibcallCallingConv(LC), RetTy, Callee, std::move(Args), 0)
10603 .setInRegister().setSExtResult(isSigned).setZExtResult(!isSigned);
10605 std::pair<SDValue, SDValue> CallInfo = LowerCallTo(CLI);
10606 return CallInfo.first;
10610 ARMTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const {
10611 assert(Subtarget->isTargetWindows() && "unsupported target platform");
10615 SDValue Chain = Op.getOperand(0);
10616 SDValue Size = Op.getOperand(1);
10618 SDValue Words = DAG.getNode(ISD::SRL, DL, MVT::i32, Size,
10619 DAG.getConstant(2, MVT::i32));
10622 Chain = DAG.getCopyToReg(Chain, DL, ARM::R4, Words, Flag);
10623 Flag = Chain.getValue(1);
10625 SDVTList NodeTys = DAG.getVTList(MVT::i32, MVT::Glue);
10626 Chain = DAG.getNode(ARMISD::WIN__CHKSTK, DL, NodeTys, Chain, Flag);
10628 SDValue NewSP = DAG.getCopyFromReg(Chain, DL, ARM::SP, MVT::i32);
10629 Chain = NewSP.getValue(1);
10631 SDValue Ops[2] = { NewSP, Chain };
10632 return DAG.getMergeValues(Ops, DL);
10636 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
10637 // The ARM target isn't yet aware of offsets.
10641 bool ARM::isBitFieldInvertedMask(unsigned v) {
10642 if (v == 0xffffffff)
10645 // there can be 1's on either or both "outsides", all the "inside"
10646 // bits must be 0's
10647 unsigned TO = CountTrailingOnes_32(v);
10648 unsigned LO = CountLeadingOnes_32(v);
10649 v = (v >> TO) << TO;
10650 v = (v << LO) >> LO;
10654 /// isFPImmLegal - Returns true if the target can instruction select the
10655 /// specified FP immediate natively. If false, the legalizer will
10656 /// materialize the FP immediate as a load from a constant pool.
10657 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
10658 if (!Subtarget->hasVFP3())
10660 if (VT == MVT::f32)
10661 return ARM_AM::getFP32Imm(Imm) != -1;
10662 if (VT == MVT::f64)
10663 return ARM_AM::getFP64Imm(Imm) != -1;
10667 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
10668 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
10669 /// specified in the intrinsic calls.
10670 bool ARMTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
10672 unsigned Intrinsic) const {
10673 switch (Intrinsic) {
10674 case Intrinsic::arm_neon_vld1:
10675 case Intrinsic::arm_neon_vld2:
10676 case Intrinsic::arm_neon_vld3:
10677 case Intrinsic::arm_neon_vld4:
10678 case Intrinsic::arm_neon_vld2lane:
10679 case Intrinsic::arm_neon_vld3lane:
10680 case Intrinsic::arm_neon_vld4lane: {
10681 Info.opc = ISD::INTRINSIC_W_CHAIN;
10682 // Conservatively set memVT to the entire set of vectors loaded.
10683 uint64_t NumElts = getDataLayout()->getTypeAllocSize(I.getType()) / 8;
10684 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
10685 Info.ptrVal = I.getArgOperand(0);
10687 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
10688 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
10689 Info.vol = false; // volatile loads with NEON intrinsics not supported
10690 Info.readMem = true;
10691 Info.writeMem = false;
10694 case Intrinsic::arm_neon_vst1:
10695 case Intrinsic::arm_neon_vst2:
10696 case Intrinsic::arm_neon_vst3:
10697 case Intrinsic::arm_neon_vst4:
10698 case Intrinsic::arm_neon_vst2lane:
10699 case Intrinsic::arm_neon_vst3lane:
10700 case Intrinsic::arm_neon_vst4lane: {
10701 Info.opc = ISD::INTRINSIC_VOID;
10702 // Conservatively set memVT to the entire set of vectors stored.
10703 unsigned NumElts = 0;
10704 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
10705 Type *ArgTy = I.getArgOperand(ArgI)->getType();
10706 if (!ArgTy->isVectorTy())
10708 NumElts += getDataLayout()->getTypeAllocSize(ArgTy) / 8;
10710 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
10711 Info.ptrVal = I.getArgOperand(0);
10713 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
10714 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
10715 Info.vol = false; // volatile stores with NEON intrinsics not supported
10716 Info.readMem = false;
10717 Info.writeMem = true;
10720 case Intrinsic::arm_ldaex:
10721 case Intrinsic::arm_ldrex: {
10722 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(0)->getType());
10723 Info.opc = ISD::INTRINSIC_W_CHAIN;
10724 Info.memVT = MVT::getVT(PtrTy->getElementType());
10725 Info.ptrVal = I.getArgOperand(0);
10727 Info.align = getDataLayout()->getABITypeAlignment(PtrTy->getElementType());
10729 Info.readMem = true;
10730 Info.writeMem = false;
10733 case Intrinsic::arm_stlex:
10734 case Intrinsic::arm_strex: {
10735 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(1)->getType());
10736 Info.opc = ISD::INTRINSIC_W_CHAIN;
10737 Info.memVT = MVT::getVT(PtrTy->getElementType());
10738 Info.ptrVal = I.getArgOperand(1);
10740 Info.align = getDataLayout()->getABITypeAlignment(PtrTy->getElementType());
10742 Info.readMem = false;
10743 Info.writeMem = true;
10746 case Intrinsic::arm_stlexd:
10747 case Intrinsic::arm_strexd: {
10748 Info.opc = ISD::INTRINSIC_W_CHAIN;
10749 Info.memVT = MVT::i64;
10750 Info.ptrVal = I.getArgOperand(2);
10754 Info.readMem = false;
10755 Info.writeMem = true;
10758 case Intrinsic::arm_ldaexd:
10759 case Intrinsic::arm_ldrexd: {
10760 Info.opc = ISD::INTRINSIC_W_CHAIN;
10761 Info.memVT = MVT::i64;
10762 Info.ptrVal = I.getArgOperand(0);
10766 Info.readMem = true;
10767 Info.writeMem = false;
10777 /// \brief Returns true if it is beneficial to convert a load of a constant
10778 /// to just the constant itself.
10779 bool ARMTargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
10781 assert(Ty->isIntegerTy());
10783 unsigned Bits = Ty->getPrimitiveSizeInBits();
10784 if (Bits == 0 || Bits > 32)
10789 bool ARMTargetLowering::shouldExpandAtomicInIR(Instruction *Inst) const {
10790 // Loads and stores less than 64-bits are already atomic; ones above that
10791 // are doomed anyway, so defer to the default libcall and blame the OS when
10792 // things go wrong. Cortex M doesn't have ldrexd/strexd though, so don't emit
10793 // anything for those.
10794 bool IsMClass = Subtarget->isMClass();
10795 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
10796 unsigned Size = SI->getValueOperand()->getType()->getPrimitiveSizeInBits();
10797 return Size == 64 && !IsMClass;
10798 } else if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
10799 return LI->getType()->getPrimitiveSizeInBits() == 64 && !IsMClass;
10802 // For the real atomic operations, we have ldrex/strex up to 32 bits,
10803 // and up to 64 bits on the non-M profiles
10804 unsigned AtomicLimit = IsMClass ? 32 : 64;
10805 return Inst->getType()->getPrimitiveSizeInBits() <= AtomicLimit;
10808 Value *ARMTargetLowering::emitLoadLinked(IRBuilder<> &Builder, Value *Addr,
10809 AtomicOrdering Ord) const {
10810 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
10811 Type *ValTy = cast<PointerType>(Addr->getType())->getElementType();
10813 Ord == Acquire || Ord == AcquireRelease || Ord == SequentiallyConsistent;
10815 // Since i64 isn't legal and intrinsics don't get type-lowered, the ldrexd
10816 // intrinsic must return {i32, i32} and we have to recombine them into a
10817 // single i64 here.
10818 if (ValTy->getPrimitiveSizeInBits() == 64) {
10819 Intrinsic::ID Int =
10820 IsAcquire ? Intrinsic::arm_ldaexd : Intrinsic::arm_ldrexd;
10821 Function *Ldrex = llvm::Intrinsic::getDeclaration(M, Int);
10823 Addr = Builder.CreateBitCast(Addr, Type::getInt8PtrTy(M->getContext()));
10824 Value *LoHi = Builder.CreateCall(Ldrex, Addr, "lohi");
10826 Value *Lo = Builder.CreateExtractValue(LoHi, 0, "lo");
10827 Value *Hi = Builder.CreateExtractValue(LoHi, 1, "hi");
10828 if (!Subtarget->isLittle())
10829 std::swap (Lo, Hi);
10830 Lo = Builder.CreateZExt(Lo, ValTy, "lo64");
10831 Hi = Builder.CreateZExt(Hi, ValTy, "hi64");
10832 return Builder.CreateOr(
10833 Lo, Builder.CreateShl(Hi, ConstantInt::get(ValTy, 32)), "val64");
10836 Type *Tys[] = { Addr->getType() };
10837 Intrinsic::ID Int = IsAcquire ? Intrinsic::arm_ldaex : Intrinsic::arm_ldrex;
10838 Function *Ldrex = llvm::Intrinsic::getDeclaration(M, Int, Tys);
10840 return Builder.CreateTruncOrBitCast(
10841 Builder.CreateCall(Ldrex, Addr),
10842 cast<PointerType>(Addr->getType())->getElementType());
10845 Value *ARMTargetLowering::emitStoreConditional(IRBuilder<> &Builder, Value *Val,
10847 AtomicOrdering Ord) const {
10848 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
10850 Ord == Release || Ord == AcquireRelease || Ord == SequentiallyConsistent;
10852 // Since the intrinsics must have legal type, the i64 intrinsics take two
10853 // parameters: "i32, i32". We must marshal Val into the appropriate form
10854 // before the call.
10855 if (Val->getType()->getPrimitiveSizeInBits() == 64) {
10856 Intrinsic::ID Int =
10857 IsRelease ? Intrinsic::arm_stlexd : Intrinsic::arm_strexd;
10858 Function *Strex = Intrinsic::getDeclaration(M, Int);
10859 Type *Int32Ty = Type::getInt32Ty(M->getContext());
10861 Value *Lo = Builder.CreateTrunc(Val, Int32Ty, "lo");
10862 Value *Hi = Builder.CreateTrunc(Builder.CreateLShr(Val, 32), Int32Ty, "hi");
10863 if (!Subtarget->isLittle())
10864 std::swap (Lo, Hi);
10865 Addr = Builder.CreateBitCast(Addr, Type::getInt8PtrTy(M->getContext()));
10866 return Builder.CreateCall3(Strex, Lo, Hi, Addr);
10869 Intrinsic::ID Int = IsRelease ? Intrinsic::arm_stlex : Intrinsic::arm_strex;
10870 Type *Tys[] = { Addr->getType() };
10871 Function *Strex = Intrinsic::getDeclaration(M, Int, Tys);
10873 return Builder.CreateCall2(
10874 Strex, Builder.CreateZExtOrBitCast(
10875 Val, Strex->getFunctionType()->getParamType(0)),
10887 static bool isHomogeneousAggregate(Type *Ty, HABaseType &Base,
10888 uint64_t &Members) {
10889 if (const StructType *ST = dyn_cast<StructType>(Ty)) {
10890 for (unsigned i = 0; i < ST->getNumElements(); ++i) {
10891 uint64_t SubMembers = 0;
10892 if (!isHomogeneousAggregate(ST->getElementType(i), Base, SubMembers))
10894 Members += SubMembers;
10896 } else if (const ArrayType *AT = dyn_cast<ArrayType>(Ty)) {
10897 uint64_t SubMembers = 0;
10898 if (!isHomogeneousAggregate(AT->getElementType(), Base, SubMembers))
10900 Members += SubMembers * AT->getNumElements();
10901 } else if (Ty->isFloatTy()) {
10902 if (Base != HA_UNKNOWN && Base != HA_FLOAT)
10906 } else if (Ty->isDoubleTy()) {
10907 if (Base != HA_UNKNOWN && Base != HA_DOUBLE)
10911 } else if (const VectorType *VT = dyn_cast<VectorType>(Ty)) {
10918 return VT->getBitWidth() == 64;
10920 return VT->getBitWidth() == 128;
10922 switch (VT->getBitWidth()) {
10935 return (Members > 0 && Members <= 4);
10938 /// \brief Return true if a type is an AAPCS-VFP homogeneous aggregate.
10939 bool ARMTargetLowering::functionArgumentNeedsConsecutiveRegisters(
10940 Type *Ty, CallingConv::ID CallConv, bool isVarArg) const {
10941 if (getEffectiveCallingConv(CallConv, isVarArg) !=
10942 CallingConv::ARM_AAPCS_VFP)
10945 HABaseType Base = HA_UNKNOWN;
10946 uint64_t Members = 0;
10947 bool result = isHomogeneousAggregate(Ty, Base, Members);
10948 DEBUG(dbgs() << "isHA: " << result << " "; Ty->dump(); dbgs() << "\n");