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::FPROUND_F64_F16, "__aeabi_d2h", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
316 { RTLIB::FPEXT_F32_F64, "__aeabi_f2d", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
318 // Integer to floating-point conversions.
319 // RTABI chapter 4.1.2, Table 8
320 { RTLIB::SINTTOFP_I32_F64, "__aeabi_i2d", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
321 { RTLIB::UINTTOFP_I32_F64, "__aeabi_ui2d", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
322 { RTLIB::SINTTOFP_I64_F64, "__aeabi_l2d", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
323 { RTLIB::UINTTOFP_I64_F64, "__aeabi_ul2d", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
324 { RTLIB::SINTTOFP_I32_F32, "__aeabi_i2f", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
325 { RTLIB::UINTTOFP_I32_F32, "__aeabi_ui2f", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
326 { RTLIB::SINTTOFP_I64_F32, "__aeabi_l2f", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
327 { RTLIB::UINTTOFP_I64_F32, "__aeabi_ul2f", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
329 // Long long helper functions
330 // RTABI chapter 4.2, Table 9
331 { RTLIB::MUL_I64, "__aeabi_lmul", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
332 { RTLIB::SHL_I64, "__aeabi_llsl", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
333 { RTLIB::SRL_I64, "__aeabi_llsr", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
334 { RTLIB::SRA_I64, "__aeabi_lasr", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
336 // Integer division functions
337 // RTABI chapter 4.3.1
338 { RTLIB::SDIV_I8, "__aeabi_idiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
339 { RTLIB::SDIV_I16, "__aeabi_idiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
340 { RTLIB::SDIV_I32, "__aeabi_idiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
341 { RTLIB::SDIV_I64, "__aeabi_ldivmod", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
342 { RTLIB::UDIV_I8, "__aeabi_uidiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
343 { RTLIB::UDIV_I16, "__aeabi_uidiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
344 { RTLIB::UDIV_I32, "__aeabi_uidiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
345 { RTLIB::UDIV_I64, "__aeabi_uldivmod", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
348 // RTABI chapter 4.3.4
349 { RTLIB::MEMCPY, "__aeabi_memcpy", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
350 { RTLIB::MEMMOVE, "__aeabi_memmove", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
351 { RTLIB::MEMSET, "__aeabi_memset", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
354 for (const auto &LC : LibraryCalls) {
355 setLibcallName(LC.Op, LC.Name);
356 setLibcallCallingConv(LC.Op, LC.CC);
357 if (LC.Cond != ISD::SETCC_INVALID)
358 setCmpLibcallCC(LC.Op, LC.Cond);
362 if (Subtarget->isTargetWindows()) {
363 static const struct {
364 const RTLIB::Libcall Op;
365 const char * const Name;
366 const CallingConv::ID CC;
368 { RTLIB::FPTOSINT_F32_I64, "__stoi64", CallingConv::ARM_AAPCS_VFP },
369 { RTLIB::FPTOSINT_F64_I64, "__dtoi64", CallingConv::ARM_AAPCS_VFP },
370 { RTLIB::FPTOUINT_F32_I64, "__stou64", CallingConv::ARM_AAPCS_VFP },
371 { RTLIB::FPTOUINT_F64_I64, "__dtou64", CallingConv::ARM_AAPCS_VFP },
372 { RTLIB::SINTTOFP_I64_F32, "__i64tos", CallingConv::ARM_AAPCS_VFP },
373 { RTLIB::SINTTOFP_I64_F64, "__i64tod", CallingConv::ARM_AAPCS_VFP },
374 { RTLIB::UINTTOFP_I64_F32, "__u64tos", CallingConv::ARM_AAPCS_VFP },
375 { RTLIB::UINTTOFP_I64_F64, "__u64tod", CallingConv::ARM_AAPCS_VFP },
378 for (const auto &LC : LibraryCalls) {
379 setLibcallName(LC.Op, LC.Name);
380 setLibcallCallingConv(LC.Op, LC.CC);
384 // Use divmod compiler-rt calls for iOS 5.0 and later.
385 if (Subtarget->getTargetTriple().isiOS() &&
386 !Subtarget->getTargetTriple().isOSVersionLT(5, 0)) {
387 setLibcallName(RTLIB::SDIVREM_I32, "__divmodsi4");
388 setLibcallName(RTLIB::UDIVREM_I32, "__udivmodsi4");
391 if (Subtarget->isThumb1Only())
392 addRegisterClass(MVT::i32, &ARM::tGPRRegClass);
394 addRegisterClass(MVT::i32, &ARM::GPRRegClass);
395 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
396 !Subtarget->isThumb1Only()) {
397 addRegisterClass(MVT::f32, &ARM::SPRRegClass);
398 if (!Subtarget->isFPOnlySP())
399 addRegisterClass(MVT::f64, &ARM::DPRRegClass);
402 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
403 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
404 for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
405 InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
406 setTruncStoreAction((MVT::SimpleValueType)VT,
407 (MVT::SimpleValueType)InnerVT, Expand);
408 setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT, Expand);
409 setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT, Expand);
410 setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT, Expand);
412 setOperationAction(ISD::MULHS, (MVT::SimpleValueType)VT, Expand);
413 setOperationAction(ISD::SMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
414 setOperationAction(ISD::MULHU, (MVT::SimpleValueType)VT, Expand);
415 setOperationAction(ISD::UMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
417 setOperationAction(ISD::BSWAP, (MVT::SimpleValueType)VT, Expand);
420 setOperationAction(ISD::ConstantFP, MVT::f32, Custom);
421 setOperationAction(ISD::ConstantFP, MVT::f64, Custom);
423 if (Subtarget->hasNEON()) {
424 addDRTypeForNEON(MVT::v2f32);
425 addDRTypeForNEON(MVT::v8i8);
426 addDRTypeForNEON(MVT::v4i16);
427 addDRTypeForNEON(MVT::v2i32);
428 addDRTypeForNEON(MVT::v1i64);
430 addQRTypeForNEON(MVT::v4f32);
431 addQRTypeForNEON(MVT::v2f64);
432 addQRTypeForNEON(MVT::v16i8);
433 addQRTypeForNEON(MVT::v8i16);
434 addQRTypeForNEON(MVT::v4i32);
435 addQRTypeForNEON(MVT::v2i64);
437 // v2f64 is legal so that QR subregs can be extracted as f64 elements, but
438 // neither Neon nor VFP support any arithmetic operations on it.
439 // The same with v4f32. But keep in mind that vadd, vsub, vmul are natively
440 // supported for v4f32.
441 setOperationAction(ISD::FADD, MVT::v2f64, Expand);
442 setOperationAction(ISD::FSUB, MVT::v2f64, Expand);
443 setOperationAction(ISD::FMUL, MVT::v2f64, Expand);
444 // FIXME: Code duplication: FDIV and FREM are expanded always, see
445 // ARMTargetLowering::addTypeForNEON method for details.
446 setOperationAction(ISD::FDIV, MVT::v2f64, Expand);
447 setOperationAction(ISD::FREM, MVT::v2f64, Expand);
448 // FIXME: Create unittest.
449 // In another words, find a way when "copysign" appears in DAG with vector
451 setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand);
452 // FIXME: Code duplication: SETCC has custom operation action, see
453 // ARMTargetLowering::addTypeForNEON method for details.
454 setOperationAction(ISD::SETCC, MVT::v2f64, Expand);
455 // FIXME: Create unittest for FNEG and for FABS.
456 setOperationAction(ISD::FNEG, MVT::v2f64, Expand);
457 setOperationAction(ISD::FABS, MVT::v2f64, Expand);
458 setOperationAction(ISD::FSQRT, MVT::v2f64, Expand);
459 setOperationAction(ISD::FSIN, MVT::v2f64, Expand);
460 setOperationAction(ISD::FCOS, MVT::v2f64, Expand);
461 setOperationAction(ISD::FPOWI, MVT::v2f64, Expand);
462 setOperationAction(ISD::FPOW, MVT::v2f64, Expand);
463 setOperationAction(ISD::FLOG, MVT::v2f64, Expand);
464 setOperationAction(ISD::FLOG2, MVT::v2f64, Expand);
465 setOperationAction(ISD::FLOG10, MVT::v2f64, Expand);
466 setOperationAction(ISD::FEXP, MVT::v2f64, Expand);
467 setOperationAction(ISD::FEXP2, MVT::v2f64, Expand);
468 // FIXME: Create unittest for FCEIL, FTRUNC, FRINT, FNEARBYINT, FFLOOR.
469 setOperationAction(ISD::FCEIL, MVT::v2f64, Expand);
470 setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand);
471 setOperationAction(ISD::FRINT, MVT::v2f64, Expand);
472 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand);
473 setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand);
474 setOperationAction(ISD::FMA, MVT::v2f64, Expand);
476 setOperationAction(ISD::FSQRT, MVT::v4f32, Expand);
477 setOperationAction(ISD::FSIN, MVT::v4f32, Expand);
478 setOperationAction(ISD::FCOS, MVT::v4f32, Expand);
479 setOperationAction(ISD::FPOWI, MVT::v4f32, Expand);
480 setOperationAction(ISD::FPOW, MVT::v4f32, Expand);
481 setOperationAction(ISD::FLOG, MVT::v4f32, Expand);
482 setOperationAction(ISD::FLOG2, MVT::v4f32, Expand);
483 setOperationAction(ISD::FLOG10, MVT::v4f32, Expand);
484 setOperationAction(ISD::FEXP, MVT::v4f32, Expand);
485 setOperationAction(ISD::FEXP2, MVT::v4f32, Expand);
486 setOperationAction(ISD::FCEIL, MVT::v4f32, Expand);
487 setOperationAction(ISD::FTRUNC, MVT::v4f32, Expand);
488 setOperationAction(ISD::FRINT, MVT::v4f32, Expand);
489 setOperationAction(ISD::FNEARBYINT, MVT::v4f32, Expand);
490 setOperationAction(ISD::FFLOOR, MVT::v4f32, Expand);
492 // Mark v2f32 intrinsics.
493 setOperationAction(ISD::FSQRT, MVT::v2f32, Expand);
494 setOperationAction(ISD::FSIN, MVT::v2f32, Expand);
495 setOperationAction(ISD::FCOS, MVT::v2f32, Expand);
496 setOperationAction(ISD::FPOWI, MVT::v2f32, Expand);
497 setOperationAction(ISD::FPOW, MVT::v2f32, Expand);
498 setOperationAction(ISD::FLOG, MVT::v2f32, Expand);
499 setOperationAction(ISD::FLOG2, MVT::v2f32, Expand);
500 setOperationAction(ISD::FLOG10, MVT::v2f32, Expand);
501 setOperationAction(ISD::FEXP, MVT::v2f32, Expand);
502 setOperationAction(ISD::FEXP2, MVT::v2f32, Expand);
503 setOperationAction(ISD::FCEIL, MVT::v2f32, Expand);
504 setOperationAction(ISD::FTRUNC, MVT::v2f32, Expand);
505 setOperationAction(ISD::FRINT, MVT::v2f32, Expand);
506 setOperationAction(ISD::FNEARBYINT, MVT::v2f32, Expand);
507 setOperationAction(ISD::FFLOOR, MVT::v2f32, Expand);
509 // Neon does not support some operations on v1i64 and v2i64 types.
510 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
511 // Custom handling for some quad-vector types to detect VMULL.
512 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
513 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
514 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
515 // Custom handling for some vector types to avoid expensive expansions
516 setOperationAction(ISD::SDIV, MVT::v4i16, Custom);
517 setOperationAction(ISD::SDIV, MVT::v8i8, Custom);
518 setOperationAction(ISD::UDIV, MVT::v4i16, Custom);
519 setOperationAction(ISD::UDIV, MVT::v8i8, Custom);
520 setOperationAction(ISD::SETCC, MVT::v1i64, Expand);
521 setOperationAction(ISD::SETCC, MVT::v2i64, Expand);
522 // Neon does not have single instruction SINT_TO_FP and UINT_TO_FP with
523 // a destination type that is wider than the source, and nor does
524 // it have a FP_TO_[SU]INT instruction with a narrower destination than
526 setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Custom);
527 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
528 setOperationAction(ISD::FP_TO_UINT, MVT::v4i16, Custom);
529 setOperationAction(ISD::FP_TO_SINT, MVT::v4i16, Custom);
531 setOperationAction(ISD::FP_ROUND, MVT::v2f32, Expand);
532 setOperationAction(ISD::FP_EXTEND, MVT::v2f64, Expand);
534 // NEON does not have single instruction CTPOP for vectors with element
535 // types wider than 8-bits. However, custom lowering can leverage the
536 // v8i8/v16i8 vcnt instruction.
537 setOperationAction(ISD::CTPOP, MVT::v2i32, Custom);
538 setOperationAction(ISD::CTPOP, MVT::v4i32, Custom);
539 setOperationAction(ISD::CTPOP, MVT::v4i16, Custom);
540 setOperationAction(ISD::CTPOP, MVT::v8i16, Custom);
542 // NEON only has FMA instructions as of VFP4.
543 if (!Subtarget->hasVFP4()) {
544 setOperationAction(ISD::FMA, MVT::v2f32, Expand);
545 setOperationAction(ISD::FMA, MVT::v4f32, Expand);
548 setTargetDAGCombine(ISD::INTRINSIC_VOID);
549 setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN);
550 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
551 setTargetDAGCombine(ISD::SHL);
552 setTargetDAGCombine(ISD::SRL);
553 setTargetDAGCombine(ISD::SRA);
554 setTargetDAGCombine(ISD::SIGN_EXTEND);
555 setTargetDAGCombine(ISD::ZERO_EXTEND);
556 setTargetDAGCombine(ISD::ANY_EXTEND);
557 setTargetDAGCombine(ISD::SELECT_CC);
558 setTargetDAGCombine(ISD::BUILD_VECTOR);
559 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
560 setTargetDAGCombine(ISD::INSERT_VECTOR_ELT);
561 setTargetDAGCombine(ISD::STORE);
562 setTargetDAGCombine(ISD::FP_TO_SINT);
563 setTargetDAGCombine(ISD::FP_TO_UINT);
564 setTargetDAGCombine(ISD::FDIV);
566 // It is legal to extload from v4i8 to v4i16 or v4i32.
567 MVT Tys[6] = {MVT::v8i8, MVT::v4i8, MVT::v2i8,
568 MVT::v4i16, MVT::v2i16,
570 for (unsigned i = 0; i < 6; ++i) {
571 setLoadExtAction(ISD::EXTLOAD, Tys[i], Legal);
572 setLoadExtAction(ISD::ZEXTLOAD, Tys[i], Legal);
573 setLoadExtAction(ISD::SEXTLOAD, Tys[i], Legal);
577 // ARM and Thumb2 support UMLAL/SMLAL.
578 if (!Subtarget->isThumb1Only())
579 setTargetDAGCombine(ISD::ADDC);
582 computeRegisterProperties();
584 // ARM does not have floating-point extending loads.
585 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
586 setLoadExtAction(ISD::EXTLOAD, MVT::f16, Expand);
588 // ... or truncating stores
589 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
590 setTruncStoreAction(MVT::f32, MVT::f16, Expand);
591 setTruncStoreAction(MVT::f64, MVT::f16, Expand);
593 // ARM does not have i1 sign extending load.
594 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
596 // ARM supports all 4 flavors of integer indexed load / store.
597 if (!Subtarget->isThumb1Only()) {
598 for (unsigned im = (unsigned)ISD::PRE_INC;
599 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
600 setIndexedLoadAction(im, MVT::i1, Legal);
601 setIndexedLoadAction(im, MVT::i8, Legal);
602 setIndexedLoadAction(im, MVT::i16, Legal);
603 setIndexedLoadAction(im, MVT::i32, Legal);
604 setIndexedStoreAction(im, MVT::i1, Legal);
605 setIndexedStoreAction(im, MVT::i8, Legal);
606 setIndexedStoreAction(im, MVT::i16, Legal);
607 setIndexedStoreAction(im, MVT::i32, Legal);
611 setOperationAction(ISD::SADDO, MVT::i32, Custom);
612 setOperationAction(ISD::UADDO, MVT::i32, Custom);
613 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
614 setOperationAction(ISD::USUBO, MVT::i32, Custom);
616 // i64 operation support.
617 setOperationAction(ISD::MUL, MVT::i64, Expand);
618 setOperationAction(ISD::MULHU, MVT::i32, Expand);
619 if (Subtarget->isThumb1Only()) {
620 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
621 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
623 if (Subtarget->isThumb1Only() || !Subtarget->hasV6Ops()
624 || (Subtarget->isThumb2() && !Subtarget->hasThumb2DSP()))
625 setOperationAction(ISD::MULHS, MVT::i32, Expand);
627 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
628 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
629 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
630 setOperationAction(ISD::SRL, MVT::i64, Custom);
631 setOperationAction(ISD::SRA, MVT::i64, Custom);
633 if (!Subtarget->isThumb1Only()) {
634 // FIXME: We should do this for Thumb1 as well.
635 setOperationAction(ISD::ADDC, MVT::i32, Custom);
636 setOperationAction(ISD::ADDE, MVT::i32, Custom);
637 setOperationAction(ISD::SUBC, MVT::i32, Custom);
638 setOperationAction(ISD::SUBE, MVT::i32, Custom);
641 // ARM does not have ROTL.
642 setOperationAction(ISD::ROTL, MVT::i32, Expand);
643 setOperationAction(ISD::CTTZ, MVT::i32, Custom);
644 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
645 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
646 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
648 // These just redirect to CTTZ and CTLZ on ARM.
649 setOperationAction(ISD::CTTZ_ZERO_UNDEF , MVT::i32 , Expand);
650 setOperationAction(ISD::CTLZ_ZERO_UNDEF , MVT::i32 , Expand);
652 setOperationAction(ISD::READCYCLECOUNTER, MVT::i64, Custom);
654 // Only ARMv6 has BSWAP.
655 if (!Subtarget->hasV6Ops())
656 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
658 if (!(Subtarget->hasDivide() && Subtarget->isThumb2()) &&
659 !(Subtarget->hasDivideInARMMode() && !Subtarget->isThumb())) {
660 // These are expanded into libcalls if the cpu doesn't have HW divider.
661 setOperationAction(ISD::SDIV, MVT::i32, Expand);
662 setOperationAction(ISD::UDIV, MVT::i32, Expand);
665 // FIXME: Also set divmod for SREM on EABI
666 setOperationAction(ISD::SREM, MVT::i32, Expand);
667 setOperationAction(ISD::UREM, MVT::i32, Expand);
668 // Register based DivRem for AEABI (RTABI 4.2)
669 if (Subtarget->isTargetAEABI()) {
670 setLibcallName(RTLIB::SDIVREM_I8, "__aeabi_idivmod");
671 setLibcallName(RTLIB::SDIVREM_I16, "__aeabi_idivmod");
672 setLibcallName(RTLIB::SDIVREM_I32, "__aeabi_idivmod");
673 setLibcallName(RTLIB::SDIVREM_I64, "__aeabi_ldivmod");
674 setLibcallName(RTLIB::UDIVREM_I8, "__aeabi_uidivmod");
675 setLibcallName(RTLIB::UDIVREM_I16, "__aeabi_uidivmod");
676 setLibcallName(RTLIB::UDIVREM_I32, "__aeabi_uidivmod");
677 setLibcallName(RTLIB::UDIVREM_I64, "__aeabi_uldivmod");
679 setLibcallCallingConv(RTLIB::SDIVREM_I8, CallingConv::ARM_AAPCS);
680 setLibcallCallingConv(RTLIB::SDIVREM_I16, CallingConv::ARM_AAPCS);
681 setLibcallCallingConv(RTLIB::SDIVREM_I32, CallingConv::ARM_AAPCS);
682 setLibcallCallingConv(RTLIB::SDIVREM_I64, CallingConv::ARM_AAPCS);
683 setLibcallCallingConv(RTLIB::UDIVREM_I8, CallingConv::ARM_AAPCS);
684 setLibcallCallingConv(RTLIB::UDIVREM_I16, CallingConv::ARM_AAPCS);
685 setLibcallCallingConv(RTLIB::UDIVREM_I32, CallingConv::ARM_AAPCS);
686 setLibcallCallingConv(RTLIB::UDIVREM_I64, CallingConv::ARM_AAPCS);
688 setOperationAction(ISD::SDIVREM, MVT::i32, Custom);
689 setOperationAction(ISD::UDIVREM, MVT::i32, Custom);
691 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
692 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
695 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
696 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
697 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
698 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
699 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
701 setOperationAction(ISD::TRAP, MVT::Other, Legal);
703 // Use the default implementation.
704 setOperationAction(ISD::VASTART, MVT::Other, Custom);
705 setOperationAction(ISD::VAARG, MVT::Other, Expand);
706 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
707 setOperationAction(ISD::VAEND, MVT::Other, Expand);
708 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
709 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
711 if (!Subtarget->isTargetMachO()) {
712 // Non-MachO platforms may return values in these registers via the
713 // personality function.
714 setExceptionPointerRegister(ARM::R0);
715 setExceptionSelectorRegister(ARM::R1);
718 if (Subtarget->getTargetTriple().isWindowsItaniumEnvironment())
719 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
721 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
723 // ARMv6 Thumb1 (except for CPUs that support dmb / dsb) and earlier use
724 // the default expansion.
725 if (Subtarget->hasAnyDataBarrier() && !Subtarget->isThumb1Only()) {
726 // ATOMIC_FENCE needs custom lowering; the others should have been expanded
727 // to ldrex/strex loops already.
728 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Custom);
730 // On v8, we have particularly efficient implementations of atomic fences
731 // if they can be combined with nearby atomic loads and stores.
732 if (!Subtarget->hasV8Ops()) {
733 // Automatically insert fences (dmb ist) around ATOMIC_SWAP etc.
734 setInsertFencesForAtomic(true);
737 // If there's anything we can use as a barrier, go through custom lowering
739 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other,
740 Subtarget->hasAnyDataBarrier() ? Custom : Expand);
742 // Set them all for expansion, which will force libcalls.
743 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Expand);
744 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Expand);
745 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Expand);
746 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Expand);
747 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Expand);
748 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Expand);
749 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Expand);
750 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Expand);
751 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i32, Expand);
752 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i32, Expand);
753 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i32, Expand);
754 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i32, Expand);
755 // Mark ATOMIC_LOAD and ATOMIC_STORE custom so we can handle the
756 // Unordered/Monotonic case.
757 setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Custom);
758 setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Custom);
761 setOperationAction(ISD::PREFETCH, MVT::Other, Custom);
763 // Requires SXTB/SXTH, available on v6 and up in both ARM and Thumb modes.
764 if (!Subtarget->hasV6Ops()) {
765 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
766 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
768 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
770 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
771 !Subtarget->isThumb1Only()) {
772 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
773 // iff target supports vfp2.
774 setOperationAction(ISD::BITCAST, MVT::i64, Custom);
775 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
778 // We want to custom lower some of our intrinsics.
779 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
780 if (Subtarget->isTargetDarwin()) {
781 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
782 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
783 setLibcallName(RTLIB::UNWIND_RESUME, "_Unwind_SjLj_Resume");
786 setOperationAction(ISD::SETCC, MVT::i32, Expand);
787 setOperationAction(ISD::SETCC, MVT::f32, Expand);
788 setOperationAction(ISD::SETCC, MVT::f64, Expand);
789 setOperationAction(ISD::SELECT, MVT::i32, Custom);
790 setOperationAction(ISD::SELECT, MVT::f32, Custom);
791 setOperationAction(ISD::SELECT, MVT::f64, Custom);
792 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
793 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
794 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
796 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
797 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
798 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
799 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
800 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
802 // We don't support sin/cos/fmod/copysign/pow
803 setOperationAction(ISD::FSIN, MVT::f64, Expand);
804 setOperationAction(ISD::FSIN, MVT::f32, Expand);
805 setOperationAction(ISD::FCOS, MVT::f32, Expand);
806 setOperationAction(ISD::FCOS, MVT::f64, Expand);
807 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
808 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
809 setOperationAction(ISD::FREM, MVT::f64, Expand);
810 setOperationAction(ISD::FREM, MVT::f32, Expand);
811 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
812 !Subtarget->isThumb1Only()) {
813 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
814 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
816 setOperationAction(ISD::FPOW, MVT::f64, Expand);
817 setOperationAction(ISD::FPOW, MVT::f32, Expand);
819 if (!Subtarget->hasVFP4()) {
820 setOperationAction(ISD::FMA, MVT::f64, Expand);
821 setOperationAction(ISD::FMA, MVT::f32, Expand);
824 // Various VFP goodness
825 if (!TM.Options.UseSoftFloat && !Subtarget->isThumb1Only()) {
826 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
827 if (Subtarget->hasVFP2()) {
828 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
829 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
830 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
831 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
834 // v8 adds f64 <-> f16 conversion. Before that it should be expanded.
835 if (!Subtarget->hasV8Ops()) {
836 setOperationAction(ISD::FP16_TO_FP, MVT::f64, Expand);
837 setOperationAction(ISD::FP_TO_FP16, MVT::f64, Expand);
840 // fp16 is a special v7 extension that adds f16 <-> f32 conversions.
841 if (!Subtarget->hasFP16()) {
842 setOperationAction(ISD::FP16_TO_FP, MVT::f32, Expand);
843 setOperationAction(ISD::FP_TO_FP16, MVT::f32, Expand);
847 // Combine sin / cos into one node or libcall if possible.
848 if (Subtarget->hasSinCos()) {
849 setLibcallName(RTLIB::SINCOS_F32, "sincosf");
850 setLibcallName(RTLIB::SINCOS_F64, "sincos");
851 if (Subtarget->getTargetTriple().getOS() == Triple::IOS) {
852 // For iOS, we don't want to the normal expansion of a libcall to
853 // sincos. We want to issue a libcall to __sincos_stret.
854 setOperationAction(ISD::FSINCOS, MVT::f64, Custom);
855 setOperationAction(ISD::FSINCOS, MVT::f32, Custom);
859 // We have target-specific dag combine patterns for the following nodes:
860 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
861 setTargetDAGCombine(ISD::ADD);
862 setTargetDAGCombine(ISD::SUB);
863 setTargetDAGCombine(ISD::MUL);
864 setTargetDAGCombine(ISD::AND);
865 setTargetDAGCombine(ISD::OR);
866 setTargetDAGCombine(ISD::XOR);
868 if (Subtarget->hasV6Ops())
869 setTargetDAGCombine(ISD::SRL);
871 setStackPointerRegisterToSaveRestore(ARM::SP);
873 if (TM.Options.UseSoftFloat || Subtarget->isThumb1Only() ||
874 !Subtarget->hasVFP2())
875 setSchedulingPreference(Sched::RegPressure);
877 setSchedulingPreference(Sched::Hybrid);
879 //// temporary - rewrite interface to use type
880 MaxStoresPerMemset = 8;
881 MaxStoresPerMemsetOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
882 MaxStoresPerMemcpy = 4; // For @llvm.memcpy -> sequence of stores
883 MaxStoresPerMemcpyOptSize = Subtarget->isTargetDarwin() ? 4 : 2;
884 MaxStoresPerMemmove = 4; // For @llvm.memmove -> sequence of stores
885 MaxStoresPerMemmoveOptSize = Subtarget->isTargetDarwin() ? 4 : 2;
887 // On ARM arguments smaller than 4 bytes are extended, so all arguments
888 // are at least 4 bytes aligned.
889 setMinStackArgumentAlignment(4);
891 // Prefer likely predicted branches to selects on out-of-order cores.
892 PredictableSelectIsExpensive = Subtarget->isLikeA9();
894 setMinFunctionAlignment(Subtarget->isThumb() ? 1 : 2);
897 // FIXME: It might make sense to define the representative register class as the
898 // nearest super-register that has a non-null superset. For example, DPR_VFP2 is
899 // a super-register of SPR, and DPR is a superset if DPR_VFP2. Consequently,
900 // SPR's representative would be DPR_VFP2. This should work well if register
901 // pressure tracking were modified such that a register use would increment the
902 // pressure of the register class's representative and all of it's super
903 // classes' representatives transitively. We have not implemented this because
904 // of the difficulty prior to coalescing of modeling operand register classes
905 // due to the common occurrence of cross class copies and subregister insertions
907 std::pair<const TargetRegisterClass*, uint8_t>
908 ARMTargetLowering::findRepresentativeClass(MVT VT) const{
909 const TargetRegisterClass *RRC = nullptr;
911 switch (VT.SimpleTy) {
913 return TargetLowering::findRepresentativeClass(VT);
914 // Use DPR as representative register class for all floating point
915 // and vector types. Since there are 32 SPR registers and 32 DPR registers so
916 // the cost is 1 for both f32 and f64.
917 case MVT::f32: case MVT::f64: case MVT::v8i8: case MVT::v4i16:
918 case MVT::v2i32: case MVT::v1i64: case MVT::v2f32:
919 RRC = &ARM::DPRRegClass;
920 // When NEON is used for SP, only half of the register file is available
921 // because operations that define both SP and DP results will be constrained
922 // to the VFP2 class (D0-D15). We currently model this constraint prior to
923 // coalescing by double-counting the SP regs. See the FIXME above.
924 if (Subtarget->useNEONForSinglePrecisionFP())
927 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
928 case MVT::v4f32: case MVT::v2f64:
929 RRC = &ARM::DPRRegClass;
933 RRC = &ARM::DPRRegClass;
937 RRC = &ARM::DPRRegClass;
941 return std::make_pair(RRC, Cost);
944 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
946 default: return nullptr;
947 case ARMISD::Wrapper: return "ARMISD::Wrapper";
948 case ARMISD::WrapperPIC: return "ARMISD::WrapperPIC";
949 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
950 case ARMISD::CALL: return "ARMISD::CALL";
951 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
952 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
953 case ARMISD::tCALL: return "ARMISD::tCALL";
954 case ARMISD::BRCOND: return "ARMISD::BRCOND";
955 case ARMISD::BR_JT: return "ARMISD::BR_JT";
956 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
957 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
958 case ARMISD::INTRET_FLAG: return "ARMISD::INTRET_FLAG";
959 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
960 case ARMISD::CMP: return "ARMISD::CMP";
961 case ARMISD::CMN: return "ARMISD::CMN";
962 case ARMISD::CMPZ: return "ARMISD::CMPZ";
963 case ARMISD::CMPFP: return "ARMISD::CMPFP";
964 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
965 case ARMISD::BCC_i64: return "ARMISD::BCC_i64";
966 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
968 case ARMISD::CMOV: return "ARMISD::CMOV";
970 case ARMISD::RBIT: return "ARMISD::RBIT";
972 case ARMISD::FTOSI: return "ARMISD::FTOSI";
973 case ARMISD::FTOUI: return "ARMISD::FTOUI";
974 case ARMISD::SITOF: return "ARMISD::SITOF";
975 case ARMISD::UITOF: return "ARMISD::UITOF";
977 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
978 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
979 case ARMISD::RRX: return "ARMISD::RRX";
981 case ARMISD::ADDC: return "ARMISD::ADDC";
982 case ARMISD::ADDE: return "ARMISD::ADDE";
983 case ARMISD::SUBC: return "ARMISD::SUBC";
984 case ARMISD::SUBE: return "ARMISD::SUBE";
986 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
987 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
989 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
990 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
992 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN";
994 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
996 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
998 case ARMISD::MEMBARRIER_MCR: return "ARMISD::MEMBARRIER_MCR";
1000 case ARMISD::PRELOAD: return "ARMISD::PRELOAD";
1002 case ARMISD::WIN__CHKSTK: return "ARMISD:::WIN__CHKSTK";
1004 case ARMISD::VCEQ: return "ARMISD::VCEQ";
1005 case ARMISD::VCEQZ: return "ARMISD::VCEQZ";
1006 case ARMISD::VCGE: return "ARMISD::VCGE";
1007 case ARMISD::VCGEZ: return "ARMISD::VCGEZ";
1008 case ARMISD::VCLEZ: return "ARMISD::VCLEZ";
1009 case ARMISD::VCGEU: return "ARMISD::VCGEU";
1010 case ARMISD::VCGT: return "ARMISD::VCGT";
1011 case ARMISD::VCGTZ: return "ARMISD::VCGTZ";
1012 case ARMISD::VCLTZ: return "ARMISD::VCLTZ";
1013 case ARMISD::VCGTU: return "ARMISD::VCGTU";
1014 case ARMISD::VTST: return "ARMISD::VTST";
1016 case ARMISD::VSHL: return "ARMISD::VSHL";
1017 case ARMISD::VSHRs: return "ARMISD::VSHRs";
1018 case ARMISD::VSHRu: return "ARMISD::VSHRu";
1019 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
1020 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
1021 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
1022 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
1023 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
1024 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
1025 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
1026 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
1027 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
1028 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
1029 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
1030 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
1031 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
1032 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
1033 case ARMISD::VMOVIMM: return "ARMISD::VMOVIMM";
1034 case ARMISD::VMVNIMM: return "ARMISD::VMVNIMM";
1035 case ARMISD::VMOVFPIMM: return "ARMISD::VMOVFPIMM";
1036 case ARMISD::VDUP: return "ARMISD::VDUP";
1037 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
1038 case ARMISD::VEXT: return "ARMISD::VEXT";
1039 case ARMISD::VREV64: return "ARMISD::VREV64";
1040 case ARMISD::VREV32: return "ARMISD::VREV32";
1041 case ARMISD::VREV16: return "ARMISD::VREV16";
1042 case ARMISD::VZIP: return "ARMISD::VZIP";
1043 case ARMISD::VUZP: return "ARMISD::VUZP";
1044 case ARMISD::VTRN: return "ARMISD::VTRN";
1045 case ARMISD::VTBL1: return "ARMISD::VTBL1";
1046 case ARMISD::VTBL2: return "ARMISD::VTBL2";
1047 case ARMISD::VMULLs: return "ARMISD::VMULLs";
1048 case ARMISD::VMULLu: return "ARMISD::VMULLu";
1049 case ARMISD::UMLAL: return "ARMISD::UMLAL";
1050 case ARMISD::SMLAL: return "ARMISD::SMLAL";
1051 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
1052 case ARMISD::FMAX: return "ARMISD::FMAX";
1053 case ARMISD::FMIN: return "ARMISD::FMIN";
1054 case ARMISD::VMAXNM: return "ARMISD::VMAX";
1055 case ARMISD::VMINNM: return "ARMISD::VMIN";
1056 case ARMISD::BFI: return "ARMISD::BFI";
1057 case ARMISD::VORRIMM: return "ARMISD::VORRIMM";
1058 case ARMISD::VBICIMM: return "ARMISD::VBICIMM";
1059 case ARMISD::VBSL: return "ARMISD::VBSL";
1060 case ARMISD::VLD2DUP: return "ARMISD::VLD2DUP";
1061 case ARMISD::VLD3DUP: return "ARMISD::VLD3DUP";
1062 case ARMISD::VLD4DUP: return "ARMISD::VLD4DUP";
1063 case ARMISD::VLD1_UPD: return "ARMISD::VLD1_UPD";
1064 case ARMISD::VLD2_UPD: return "ARMISD::VLD2_UPD";
1065 case ARMISD::VLD3_UPD: return "ARMISD::VLD3_UPD";
1066 case ARMISD::VLD4_UPD: return "ARMISD::VLD4_UPD";
1067 case ARMISD::VLD2LN_UPD: return "ARMISD::VLD2LN_UPD";
1068 case ARMISD::VLD3LN_UPD: return "ARMISD::VLD3LN_UPD";
1069 case ARMISD::VLD4LN_UPD: return "ARMISD::VLD4LN_UPD";
1070 case ARMISD::VLD2DUP_UPD: return "ARMISD::VLD2DUP_UPD";
1071 case ARMISD::VLD3DUP_UPD: return "ARMISD::VLD3DUP_UPD";
1072 case ARMISD::VLD4DUP_UPD: return "ARMISD::VLD4DUP_UPD";
1073 case ARMISD::VST1_UPD: return "ARMISD::VST1_UPD";
1074 case ARMISD::VST2_UPD: return "ARMISD::VST2_UPD";
1075 case ARMISD::VST3_UPD: return "ARMISD::VST3_UPD";
1076 case ARMISD::VST4_UPD: return "ARMISD::VST4_UPD";
1077 case ARMISD::VST2LN_UPD: return "ARMISD::VST2LN_UPD";
1078 case ARMISD::VST3LN_UPD: return "ARMISD::VST3LN_UPD";
1079 case ARMISD::VST4LN_UPD: return "ARMISD::VST4LN_UPD";
1083 EVT ARMTargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
1084 if (!VT.isVector()) return getPointerTy();
1085 return VT.changeVectorElementTypeToInteger();
1088 /// getRegClassFor - Return the register class that should be used for the
1089 /// specified value type.
1090 const TargetRegisterClass *ARMTargetLowering::getRegClassFor(MVT VT) const {
1091 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
1092 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
1093 // load / store 4 to 8 consecutive D registers.
1094 if (Subtarget->hasNEON()) {
1095 if (VT == MVT::v4i64)
1096 return &ARM::QQPRRegClass;
1097 if (VT == MVT::v8i64)
1098 return &ARM::QQQQPRRegClass;
1100 return TargetLowering::getRegClassFor(VT);
1103 // Create a fast isel object.
1105 ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
1106 const TargetLibraryInfo *libInfo) const {
1107 return ARM::createFastISel(funcInfo, libInfo);
1110 /// getMaximalGlobalOffset - Returns the maximal possible offset which can
1111 /// be used for loads / stores from the global.
1112 unsigned ARMTargetLowering::getMaximalGlobalOffset() const {
1113 return (Subtarget->isThumb1Only() ? 127 : 4095);
1116 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
1117 unsigned NumVals = N->getNumValues();
1119 return Sched::RegPressure;
1121 for (unsigned i = 0; i != NumVals; ++i) {
1122 EVT VT = N->getValueType(i);
1123 if (VT == MVT::Glue || VT == MVT::Other)
1125 if (VT.isFloatingPoint() || VT.isVector())
1129 if (!N->isMachineOpcode())
1130 return Sched::RegPressure;
1132 // Load are scheduled for latency even if there instruction itinerary
1133 // is not available.
1134 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1135 const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
1137 if (MCID.getNumDefs() == 0)
1138 return Sched::RegPressure;
1139 if (!Itins->isEmpty() &&
1140 Itins->getOperandCycle(MCID.getSchedClass(), 0) > 2)
1143 return Sched::RegPressure;
1146 //===----------------------------------------------------------------------===//
1148 //===----------------------------------------------------------------------===//
1150 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
1151 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
1153 default: llvm_unreachable("Unknown condition code!");
1154 case ISD::SETNE: return ARMCC::NE;
1155 case ISD::SETEQ: return ARMCC::EQ;
1156 case ISD::SETGT: return ARMCC::GT;
1157 case ISD::SETGE: return ARMCC::GE;
1158 case ISD::SETLT: return ARMCC::LT;
1159 case ISD::SETLE: return ARMCC::LE;
1160 case ISD::SETUGT: return ARMCC::HI;
1161 case ISD::SETUGE: return ARMCC::HS;
1162 case ISD::SETULT: return ARMCC::LO;
1163 case ISD::SETULE: return ARMCC::LS;
1167 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
1168 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
1169 ARMCC::CondCodes &CondCode2) {
1170 CondCode2 = ARMCC::AL;
1172 default: llvm_unreachable("Unknown FP condition!");
1174 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
1176 case ISD::SETOGT: CondCode = ARMCC::GT; break;
1178 case ISD::SETOGE: CondCode = ARMCC::GE; break;
1179 case ISD::SETOLT: CondCode = ARMCC::MI; break;
1180 case ISD::SETOLE: CondCode = ARMCC::LS; break;
1181 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
1182 case ISD::SETO: CondCode = ARMCC::VC; break;
1183 case ISD::SETUO: CondCode = ARMCC::VS; break;
1184 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
1185 case ISD::SETUGT: CondCode = ARMCC::HI; break;
1186 case ISD::SETUGE: CondCode = ARMCC::PL; break;
1188 case ISD::SETULT: CondCode = ARMCC::LT; break;
1190 case ISD::SETULE: CondCode = ARMCC::LE; break;
1192 case ISD::SETUNE: CondCode = ARMCC::NE; break;
1196 //===----------------------------------------------------------------------===//
1197 // Calling Convention Implementation
1198 //===----------------------------------------------------------------------===//
1200 #include "ARMGenCallingConv.inc"
1202 /// getEffectiveCallingConv - Get the effective calling convention, taking into
1203 /// account presence of floating point hardware and calling convention
1204 /// limitations, such as support for variadic functions.
1206 ARMTargetLowering::getEffectiveCallingConv(CallingConv::ID CC,
1207 bool isVarArg) const {
1210 llvm_unreachable("Unsupported calling convention");
1211 case CallingConv::ARM_AAPCS:
1212 case CallingConv::ARM_APCS:
1213 case CallingConv::GHC:
1215 case CallingConv::ARM_AAPCS_VFP:
1216 return isVarArg ? CallingConv::ARM_AAPCS : CallingConv::ARM_AAPCS_VFP;
1217 case CallingConv::C:
1218 if (!Subtarget->isAAPCS_ABI())
1219 return CallingConv::ARM_APCS;
1220 else if (Subtarget->hasVFP2() && !Subtarget->isThumb1Only() &&
1221 getTargetMachine().Options.FloatABIType == FloatABI::Hard &&
1223 return CallingConv::ARM_AAPCS_VFP;
1225 return CallingConv::ARM_AAPCS;
1226 case CallingConv::Fast:
1227 if (!Subtarget->isAAPCS_ABI()) {
1228 if (Subtarget->hasVFP2() && !Subtarget->isThumb1Only() && !isVarArg)
1229 return CallingConv::Fast;
1230 return CallingConv::ARM_APCS;
1231 } else if (Subtarget->hasVFP2() && !Subtarget->isThumb1Only() && !isVarArg)
1232 return CallingConv::ARM_AAPCS_VFP;
1234 return CallingConv::ARM_AAPCS;
1238 /// CCAssignFnForNode - Selects the correct CCAssignFn for the given
1239 /// CallingConvention.
1240 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
1242 bool isVarArg) const {
1243 switch (getEffectiveCallingConv(CC, isVarArg)) {
1245 llvm_unreachable("Unsupported calling convention");
1246 case CallingConv::ARM_APCS:
1247 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1248 case CallingConv::ARM_AAPCS:
1249 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1250 case CallingConv::ARM_AAPCS_VFP:
1251 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1252 case CallingConv::Fast:
1253 return (Return ? RetFastCC_ARM_APCS : FastCC_ARM_APCS);
1254 case CallingConv::GHC:
1255 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS_GHC);
1259 /// LowerCallResult - Lower the result values of a call into the
1260 /// appropriate copies out of appropriate physical registers.
1262 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1263 CallingConv::ID CallConv, bool isVarArg,
1264 const SmallVectorImpl<ISD::InputArg> &Ins,
1265 SDLoc dl, SelectionDAG &DAG,
1266 SmallVectorImpl<SDValue> &InVals,
1267 bool isThisReturn, SDValue ThisVal) const {
1269 // Assign locations to each value returned by this call.
1270 SmallVector<CCValAssign, 16> RVLocs;
1271 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1272 getTargetMachine(), RVLocs, *DAG.getContext(), Call);
1273 CCInfo.AnalyzeCallResult(Ins,
1274 CCAssignFnForNode(CallConv, /* Return*/ true,
1277 // Copy all of the result registers out of their specified physreg.
1278 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1279 CCValAssign VA = RVLocs[i];
1281 // Pass 'this' value directly from the argument to return value, to avoid
1282 // reg unit interference
1283 if (i == 0 && isThisReturn) {
1284 assert(!VA.needsCustom() && VA.getLocVT() == MVT::i32 &&
1285 "unexpected return calling convention register assignment");
1286 InVals.push_back(ThisVal);
1291 if (VA.needsCustom()) {
1292 // Handle f64 or half of a v2f64.
1293 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1295 Chain = Lo.getValue(1);
1296 InFlag = Lo.getValue(2);
1297 VA = RVLocs[++i]; // skip ahead to next loc
1298 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1300 Chain = Hi.getValue(1);
1301 InFlag = Hi.getValue(2);
1302 if (!Subtarget->isLittle())
1304 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1306 if (VA.getLocVT() == MVT::v2f64) {
1307 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1308 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1309 DAG.getConstant(0, MVT::i32));
1311 VA = RVLocs[++i]; // skip ahead to next loc
1312 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1313 Chain = Lo.getValue(1);
1314 InFlag = Lo.getValue(2);
1315 VA = RVLocs[++i]; // skip ahead to next loc
1316 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1317 Chain = Hi.getValue(1);
1318 InFlag = Hi.getValue(2);
1319 if (!Subtarget->isLittle())
1321 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1322 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1323 DAG.getConstant(1, MVT::i32));
1326 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
1328 Chain = Val.getValue(1);
1329 InFlag = Val.getValue(2);
1332 switch (VA.getLocInfo()) {
1333 default: llvm_unreachable("Unknown loc info!");
1334 case CCValAssign::Full: break;
1335 case CCValAssign::BCvt:
1336 Val = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), Val);
1340 InVals.push_back(Val);
1346 /// LowerMemOpCallTo - Store the argument to the stack.
1348 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
1349 SDValue StackPtr, SDValue Arg,
1350 SDLoc dl, SelectionDAG &DAG,
1351 const CCValAssign &VA,
1352 ISD::ArgFlagsTy Flags) const {
1353 unsigned LocMemOffset = VA.getLocMemOffset();
1354 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1355 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1356 return DAG.getStore(Chain, dl, Arg, PtrOff,
1357 MachinePointerInfo::getStack(LocMemOffset),
1361 void ARMTargetLowering::PassF64ArgInRegs(SDLoc dl, SelectionDAG &DAG,
1362 SDValue Chain, SDValue &Arg,
1363 RegsToPassVector &RegsToPass,
1364 CCValAssign &VA, CCValAssign &NextVA,
1366 SmallVectorImpl<SDValue> &MemOpChains,
1367 ISD::ArgFlagsTy Flags) const {
1369 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1370 DAG.getVTList(MVT::i32, MVT::i32), Arg);
1371 unsigned id = Subtarget->isLittle() ? 0 : 1;
1372 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd.getValue(id)));
1374 if (NextVA.isRegLoc())
1375 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1-id)));
1377 assert(NextVA.isMemLoc());
1378 if (!StackPtr.getNode())
1379 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1381 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1-id),
1387 /// LowerCall - Lowering a call into a callseq_start <-
1388 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
1391 ARMTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
1392 SmallVectorImpl<SDValue> &InVals) const {
1393 SelectionDAG &DAG = CLI.DAG;
1395 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
1396 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
1397 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
1398 SDValue Chain = CLI.Chain;
1399 SDValue Callee = CLI.Callee;
1400 bool &isTailCall = CLI.IsTailCall;
1401 CallingConv::ID CallConv = CLI.CallConv;
1402 bool doesNotRet = CLI.DoesNotReturn;
1403 bool isVarArg = CLI.IsVarArg;
1405 MachineFunction &MF = DAG.getMachineFunction();
1406 bool isStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1407 bool isThisReturn = false;
1408 bool isSibCall = false;
1410 // Disable tail calls if they're not supported.
1411 if (!Subtarget->supportsTailCall() || MF.getTarget().Options.DisableTailCalls)
1415 // Check if it's really possible to do a tail call.
1416 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1417 isVarArg, isStructRet, MF.getFunction()->hasStructRetAttr(),
1418 Outs, OutVals, Ins, DAG);
1419 if (!isTailCall && CLI.CS && CLI.CS->isMustTailCall())
1420 report_fatal_error("failed to perform tail call elimination on a call "
1421 "site marked musttail");
1422 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1423 // detected sibcalls.
1430 // Analyze operands of the call, assigning locations to each operand.
1431 SmallVector<CCValAssign, 16> ArgLocs;
1432 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1433 getTargetMachine(), ArgLocs, *DAG.getContext(), Call);
1434 CCInfo.AnalyzeCallOperands(Outs,
1435 CCAssignFnForNode(CallConv, /* Return*/ false,
1438 // Get a count of how many bytes are to be pushed on the stack.
1439 unsigned NumBytes = CCInfo.getNextStackOffset();
1441 // For tail calls, memory operands are available in our caller's stack.
1445 // Adjust the stack pointer for the new arguments...
1446 // These operations are automatically eliminated by the prolog/epilog pass
1448 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true),
1451 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1453 RegsToPassVector RegsToPass;
1454 SmallVector<SDValue, 8> MemOpChains;
1456 // Walk the register/memloc assignments, inserting copies/loads. In the case
1457 // of tail call optimization, arguments are handled later.
1458 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1460 ++i, ++realArgIdx) {
1461 CCValAssign &VA = ArgLocs[i];
1462 SDValue Arg = OutVals[realArgIdx];
1463 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1464 bool isByVal = Flags.isByVal();
1466 // Promote the value if needed.
1467 switch (VA.getLocInfo()) {
1468 default: llvm_unreachable("Unknown loc info!");
1469 case CCValAssign::Full: break;
1470 case CCValAssign::SExt:
1471 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1473 case CCValAssign::ZExt:
1474 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1476 case CCValAssign::AExt:
1477 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1479 case CCValAssign::BCvt:
1480 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1484 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1485 if (VA.needsCustom()) {
1486 if (VA.getLocVT() == MVT::v2f64) {
1487 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1488 DAG.getConstant(0, MVT::i32));
1489 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1490 DAG.getConstant(1, MVT::i32));
1492 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1493 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1495 VA = ArgLocs[++i]; // skip ahead to next loc
1496 if (VA.isRegLoc()) {
1497 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1498 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1500 assert(VA.isMemLoc());
1502 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1503 dl, DAG, VA, Flags));
1506 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1507 StackPtr, MemOpChains, Flags);
1509 } else if (VA.isRegLoc()) {
1510 if (realArgIdx == 0 && Flags.isReturned() && Outs[0].VT == MVT::i32) {
1511 assert(VA.getLocVT() == MVT::i32 &&
1512 "unexpected calling convention register assignment");
1513 assert(!Ins.empty() && Ins[0].VT == MVT::i32 &&
1514 "unexpected use of 'returned'");
1515 isThisReturn = true;
1517 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1518 } else if (isByVal) {
1519 assert(VA.isMemLoc());
1520 unsigned offset = 0;
1522 // True if this byval aggregate will be split between registers
1524 unsigned ByValArgsCount = CCInfo.getInRegsParamsCount();
1525 unsigned CurByValIdx = CCInfo.getInRegsParamsProceed();
1527 if (CurByValIdx < ByValArgsCount) {
1529 unsigned RegBegin, RegEnd;
1530 CCInfo.getInRegsParamInfo(CurByValIdx, RegBegin, RegEnd);
1532 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1534 for (i = 0, j = RegBegin; j < RegEnd; i++, j++) {
1535 SDValue Const = DAG.getConstant(4*i, MVT::i32);
1536 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
1537 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
1538 MachinePointerInfo(),
1539 false, false, false,
1540 DAG.InferPtrAlignment(AddArg));
1541 MemOpChains.push_back(Load.getValue(1));
1542 RegsToPass.push_back(std::make_pair(j, Load));
1545 // If parameter size outsides register area, "offset" value
1546 // helps us to calculate stack slot for remained part properly.
1547 offset = RegEnd - RegBegin;
1549 CCInfo.nextInRegsParam();
1552 if (Flags.getByValSize() > 4*offset) {
1553 unsigned LocMemOffset = VA.getLocMemOffset();
1554 SDValue StkPtrOff = DAG.getIntPtrConstant(LocMemOffset);
1555 SDValue Dst = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr,
1557 SDValue SrcOffset = DAG.getIntPtrConstant(4*offset);
1558 SDValue Src = DAG.getNode(ISD::ADD, dl, getPointerTy(), Arg, SrcOffset);
1559 SDValue SizeNode = DAG.getConstant(Flags.getByValSize() - 4*offset,
1561 SDValue AlignNode = DAG.getConstant(Flags.getByValAlign(), MVT::i32);
1563 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
1564 SDValue Ops[] = { Chain, Dst, Src, SizeNode, AlignNode};
1565 MemOpChains.push_back(DAG.getNode(ARMISD::COPY_STRUCT_BYVAL, dl, VTs,
1568 } else if (!isSibCall) {
1569 assert(VA.isMemLoc());
1571 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1572 dl, DAG, VA, Flags));
1576 if (!MemOpChains.empty())
1577 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains);
1579 // Build a sequence of copy-to-reg nodes chained together with token chain
1580 // and flag operands which copy the outgoing args into the appropriate regs.
1582 // Tail call byval lowering might overwrite argument registers so in case of
1583 // tail call optimization the copies to registers are lowered later.
1585 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1586 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1587 RegsToPass[i].second, InFlag);
1588 InFlag = Chain.getValue(1);
1591 // For tail calls lower the arguments to the 'real' stack slot.
1593 // Force all the incoming stack arguments to be loaded from the stack
1594 // before any new outgoing arguments are stored to the stack, because the
1595 // outgoing stack slots may alias the incoming argument stack slots, and
1596 // the alias isn't otherwise explicit. This is slightly more conservative
1597 // than necessary, because it means that each store effectively depends
1598 // on every argument instead of just those arguments it would clobber.
1600 // Do not flag preceding copytoreg stuff together with the following stuff.
1602 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1603 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1604 RegsToPass[i].second, InFlag);
1605 InFlag = Chain.getValue(1);
1610 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1611 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1612 // node so that legalize doesn't hack it.
1613 bool isDirect = false;
1614 bool isARMFunc = false;
1615 bool isLocalARMFunc = false;
1616 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1618 if (EnableARMLongCalls) {
1619 assert((Subtarget->isTargetWindows() ||
1620 getTargetMachine().getRelocationModel() == Reloc::Static) &&
1621 "long-calls with non-static relocation model!");
1622 // Handle a global address or an external symbol. If it's not one of
1623 // those, the target's already in a register, so we don't need to do
1625 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1626 const GlobalValue *GV = G->getGlobal();
1627 // Create a constant pool entry for the callee address
1628 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1629 ARMConstantPoolValue *CPV =
1630 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 0);
1632 // Get the address of the callee into a register
1633 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1634 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1635 Callee = DAG.getLoad(getPointerTy(), dl,
1636 DAG.getEntryNode(), CPAddr,
1637 MachinePointerInfo::getConstantPool(),
1638 false, false, false, 0);
1639 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1640 const char *Sym = S->getSymbol();
1642 // Create a constant pool entry for the callee address
1643 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1644 ARMConstantPoolValue *CPV =
1645 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1646 ARMPCLabelIndex, 0);
1647 // Get the address of the callee into a register
1648 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1649 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1650 Callee = DAG.getLoad(getPointerTy(), dl,
1651 DAG.getEntryNode(), CPAddr,
1652 MachinePointerInfo::getConstantPool(),
1653 false, false, false, 0);
1655 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1656 const GlobalValue *GV = G->getGlobal();
1658 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1659 bool isStub = (isExt && Subtarget->isTargetMachO()) &&
1660 getTargetMachine().getRelocationModel() != Reloc::Static;
1661 isARMFunc = !Subtarget->isThumb() || isStub;
1662 // ARM call to a local ARM function is predicable.
1663 isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking);
1664 // tBX takes a register source operand.
1665 if (isStub && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1666 assert(Subtarget->isTargetMachO() && "WrapperPIC use on non-MachO?");
1667 Callee = DAG.getNode(ARMISD::WrapperPIC, dl, getPointerTy(),
1668 DAG.getTargetGlobalAddress(GV, dl, getPointerTy()));
1669 } else if (Subtarget->isTargetCOFF()) {
1670 assert(Subtarget->isTargetWindows() &&
1671 "Windows is the only supported COFF target");
1672 unsigned TargetFlags = GV->hasDLLImportStorageClass()
1673 ? ARMII::MO_DLLIMPORT
1674 : ARMII::MO_NO_FLAG;
1675 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), /*Offset=*/0,
1677 if (GV->hasDLLImportStorageClass())
1678 Callee = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
1679 DAG.getNode(ARMISD::Wrapper, dl, getPointerTy(),
1680 Callee), MachinePointerInfo::getGOT(),
1681 false, false, false, 0);
1683 // On ELF targets for PIC code, direct calls should go through the PLT
1684 unsigned OpFlags = 0;
1685 if (Subtarget->isTargetELF() &&
1686 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1687 OpFlags = ARMII::MO_PLT;
1688 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
1690 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1692 bool isStub = Subtarget->isTargetMachO() &&
1693 getTargetMachine().getRelocationModel() != Reloc::Static;
1694 isARMFunc = !Subtarget->isThumb() || isStub;
1695 // tBX takes a register source operand.
1696 const char *Sym = S->getSymbol();
1697 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1698 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1699 ARMConstantPoolValue *CPV =
1700 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1701 ARMPCLabelIndex, 4);
1702 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1703 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1704 Callee = DAG.getLoad(getPointerTy(), dl,
1705 DAG.getEntryNode(), CPAddr,
1706 MachinePointerInfo::getConstantPool(),
1707 false, false, false, 0);
1708 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1709 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1710 getPointerTy(), Callee, PICLabel);
1712 unsigned OpFlags = 0;
1713 // On ELF targets for PIC code, direct calls should go through the PLT
1714 if (Subtarget->isTargetELF() &&
1715 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1716 OpFlags = ARMII::MO_PLT;
1717 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlags);
1721 // FIXME: handle tail calls differently.
1723 bool HasMinSizeAttr = MF.getFunction()->getAttributes().hasAttribute(
1724 AttributeSet::FunctionIndex, Attribute::MinSize);
1725 if (Subtarget->isThumb()) {
1726 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1727 CallOpc = ARMISD::CALL_NOLINK;
1729 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1731 if (!isDirect && !Subtarget->hasV5TOps())
1732 CallOpc = ARMISD::CALL_NOLINK;
1733 else if (doesNotRet && isDirect && Subtarget->hasRAS() &&
1734 // Emit regular call when code size is the priority
1736 // "mov lr, pc; b _foo" to avoid confusing the RSP
1737 CallOpc = ARMISD::CALL_NOLINK;
1739 CallOpc = isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL;
1742 std::vector<SDValue> Ops;
1743 Ops.push_back(Chain);
1744 Ops.push_back(Callee);
1746 // Add argument registers to the end of the list so that they are known live
1748 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1749 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1750 RegsToPass[i].second.getValueType()));
1752 // Add a register mask operand representing the call-preserved registers.
1754 const uint32_t *Mask;
1755 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
1756 const ARMBaseRegisterInfo *ARI = static_cast<const ARMBaseRegisterInfo*>(TRI);
1758 // For 'this' returns, use the R0-preserving mask if applicable
1759 Mask = ARI->getThisReturnPreservedMask(CallConv);
1761 // Set isThisReturn to false if the calling convention is not one that
1762 // allows 'returned' to be modeled in this way, so LowerCallResult does
1763 // not try to pass 'this' straight through
1764 isThisReturn = false;
1765 Mask = ARI->getCallPreservedMask(CallConv);
1768 Mask = ARI->getCallPreservedMask(CallConv);
1770 assert(Mask && "Missing call preserved mask for calling convention");
1771 Ops.push_back(DAG.getRegisterMask(Mask));
1774 if (InFlag.getNode())
1775 Ops.push_back(InFlag);
1777 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
1779 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, Ops);
1781 // Returns a chain and a flag for retval copy to use.
1782 Chain = DAG.getNode(CallOpc, dl, NodeTys, Ops);
1783 InFlag = Chain.getValue(1);
1785 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1786 DAG.getIntPtrConstant(0, true), InFlag, dl);
1788 InFlag = Chain.getValue(1);
1790 // Handle result values, copying them out of physregs into vregs that we
1792 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins, dl, DAG,
1793 InVals, isThisReturn,
1794 isThisReturn ? OutVals[0] : SDValue());
1797 /// HandleByVal - Every parameter *after* a byval parameter is passed
1798 /// on the stack. Remember the next parameter register to allocate,
1799 /// and then confiscate the rest of the parameter registers to insure
1802 ARMTargetLowering::HandleByVal(
1803 CCState *State, unsigned &size, unsigned Align) const {
1804 unsigned reg = State->AllocateReg(GPRArgRegs, 4);
1805 assert((State->getCallOrPrologue() == Prologue ||
1806 State->getCallOrPrologue() == Call) &&
1807 "unhandled ParmContext");
1809 if ((ARM::R0 <= reg) && (reg <= ARM::R3)) {
1810 if (Subtarget->isAAPCS_ABI() && Align > 4) {
1811 unsigned AlignInRegs = Align / 4;
1812 unsigned Waste = (ARM::R4 - reg) % AlignInRegs;
1813 for (unsigned i = 0; i < Waste; ++i)
1814 reg = State->AllocateReg(GPRArgRegs, 4);
1817 unsigned excess = 4 * (ARM::R4 - reg);
1819 // Special case when NSAA != SP and parameter size greater than size of
1820 // all remained GPR regs. In that case we can't split parameter, we must
1821 // send it to stack. We also must set NCRN to R4, so waste all
1822 // remained registers.
1823 const unsigned NSAAOffset = State->getNextStackOffset();
1824 if (Subtarget->isAAPCS_ABI() && NSAAOffset != 0 && size > excess) {
1825 while (State->AllocateReg(GPRArgRegs, 4))
1830 // First register for byval parameter is the first register that wasn't
1831 // allocated before this method call, so it would be "reg".
1832 // If parameter is small enough to be saved in range [reg, r4), then
1833 // the end (first after last) register would be reg + param-size-in-regs,
1834 // else parameter would be splitted between registers and stack,
1835 // end register would be r4 in this case.
1836 unsigned ByValRegBegin = reg;
1837 unsigned ByValRegEnd = (size < excess) ? reg + size/4 : (unsigned)ARM::R4;
1838 State->addInRegsParamInfo(ByValRegBegin, ByValRegEnd);
1839 // Note, first register is allocated in the beginning of function already,
1840 // allocate remained amount of registers we need.
1841 for (unsigned i = reg+1; i != ByValRegEnd; ++i)
1842 State->AllocateReg(GPRArgRegs, 4);
1843 // A byval parameter that is split between registers and memory needs its
1844 // size truncated here.
1845 // In the case where the entire structure fits in registers, we set the
1846 // size in memory to zero.
1855 /// MatchingStackOffset - Return true if the given stack call argument is
1856 /// already available in the same position (relatively) of the caller's
1857 /// incoming argument stack.
1859 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1860 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1861 const TargetInstrInfo *TII) {
1862 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1864 if (Arg.getOpcode() == ISD::CopyFromReg) {
1865 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1866 if (!TargetRegisterInfo::isVirtualRegister(VR))
1868 MachineInstr *Def = MRI->getVRegDef(VR);
1871 if (!Flags.isByVal()) {
1872 if (!TII->isLoadFromStackSlot(Def, FI))
1877 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
1878 if (Flags.isByVal())
1879 // ByVal argument is passed in as a pointer but it's now being
1880 // dereferenced. e.g.
1881 // define @foo(%struct.X* %A) {
1882 // tail call @bar(%struct.X* byval %A)
1885 SDValue Ptr = Ld->getBasePtr();
1886 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
1889 FI = FINode->getIndex();
1893 assert(FI != INT_MAX);
1894 if (!MFI->isFixedObjectIndex(FI))
1896 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
1899 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
1900 /// for tail call optimization. Targets which want to do tail call
1901 /// optimization should implement this function.
1903 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1904 CallingConv::ID CalleeCC,
1906 bool isCalleeStructRet,
1907 bool isCallerStructRet,
1908 const SmallVectorImpl<ISD::OutputArg> &Outs,
1909 const SmallVectorImpl<SDValue> &OutVals,
1910 const SmallVectorImpl<ISD::InputArg> &Ins,
1911 SelectionDAG& DAG) const {
1912 const Function *CallerF = DAG.getMachineFunction().getFunction();
1913 CallingConv::ID CallerCC = CallerF->getCallingConv();
1914 bool CCMatch = CallerCC == CalleeCC;
1916 // Look for obvious safe cases to perform tail call optimization that do not
1917 // require ABI changes. This is what gcc calls sibcall.
1919 // Do not sibcall optimize vararg calls unless the call site is not passing
1921 if (isVarArg && !Outs.empty())
1924 // Exception-handling functions need a special set of instructions to indicate
1925 // a return to the hardware. Tail-calling another function would probably
1927 if (CallerF->hasFnAttribute("interrupt"))
1930 // Also avoid sibcall optimization if either caller or callee uses struct
1931 // return semantics.
1932 if (isCalleeStructRet || isCallerStructRet)
1935 // FIXME: Completely disable sibcall for Thumb1 since Thumb1RegisterInfo::
1936 // emitEpilogue is not ready for them. Thumb tail calls also use t2B, as
1937 // the Thumb1 16-bit unconditional branch doesn't have sufficient relocation
1938 // support in the assembler and linker to be used. This would need to be
1939 // fixed to fully support tail calls in Thumb1.
1941 // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take
1942 // LR. This means if we need to reload LR, it takes an extra instructions,
1943 // which outweighs the value of the tail call; but here we don't know yet
1944 // whether LR is going to be used. Probably the right approach is to
1945 // generate the tail call here and turn it back into CALL/RET in
1946 // emitEpilogue if LR is used.
1948 // Thumb1 PIC calls to external symbols use BX, so they can be tail calls,
1949 // but we need to make sure there are enough registers; the only valid
1950 // registers are the 4 used for parameters. We don't currently do this
1952 if (Subtarget->isThumb1Only())
1955 // If the calling conventions do not match, then we'd better make sure the
1956 // results are returned in the same way as what the caller expects.
1958 SmallVector<CCValAssign, 16> RVLocs1;
1959 ARMCCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(),
1960 getTargetMachine(), RVLocs1, *DAG.getContext(), Call);
1961 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
1963 SmallVector<CCValAssign, 16> RVLocs2;
1964 ARMCCState CCInfo2(CallerCC, false, DAG.getMachineFunction(),
1965 getTargetMachine(), RVLocs2, *DAG.getContext(), Call);
1966 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
1968 if (RVLocs1.size() != RVLocs2.size())
1970 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
1971 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
1973 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
1975 if (RVLocs1[i].isRegLoc()) {
1976 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
1979 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
1985 // If Caller's vararg or byval argument has been split between registers and
1986 // stack, do not perform tail call, since part of the argument is in caller's
1988 const ARMFunctionInfo *AFI_Caller = DAG.getMachineFunction().
1989 getInfo<ARMFunctionInfo>();
1990 if (AFI_Caller->getArgRegsSaveSize())
1993 // If the callee takes no arguments then go on to check the results of the
1995 if (!Outs.empty()) {
1996 // Check if stack adjustment is needed. For now, do not do this if any
1997 // argument is passed on the stack.
1998 SmallVector<CCValAssign, 16> ArgLocs;
1999 ARMCCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(),
2000 getTargetMachine(), ArgLocs, *DAG.getContext(), Call);
2001 CCInfo.AnalyzeCallOperands(Outs,
2002 CCAssignFnForNode(CalleeCC, false, isVarArg));
2003 if (CCInfo.getNextStackOffset()) {
2004 MachineFunction &MF = DAG.getMachineFunction();
2006 // Check if the arguments are already laid out in the right way as
2007 // the caller's fixed stack objects.
2008 MachineFrameInfo *MFI = MF.getFrameInfo();
2009 const MachineRegisterInfo *MRI = &MF.getRegInfo();
2010 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
2011 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
2013 ++i, ++realArgIdx) {
2014 CCValAssign &VA = ArgLocs[i];
2015 EVT RegVT = VA.getLocVT();
2016 SDValue Arg = OutVals[realArgIdx];
2017 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
2018 if (VA.getLocInfo() == CCValAssign::Indirect)
2020 if (VA.needsCustom()) {
2021 // f64 and vector types are split into multiple registers or
2022 // register/stack-slot combinations. The types will not match
2023 // the registers; give up on memory f64 refs until we figure
2024 // out what to do about this.
2027 if (!ArgLocs[++i].isRegLoc())
2029 if (RegVT == MVT::v2f64) {
2030 if (!ArgLocs[++i].isRegLoc())
2032 if (!ArgLocs[++i].isRegLoc())
2035 } else if (!VA.isRegLoc()) {
2036 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
2048 ARMTargetLowering::CanLowerReturn(CallingConv::ID CallConv,
2049 MachineFunction &MF, bool isVarArg,
2050 const SmallVectorImpl<ISD::OutputArg> &Outs,
2051 LLVMContext &Context) const {
2052 SmallVector<CCValAssign, 16> RVLocs;
2053 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(), RVLocs, Context);
2054 return CCInfo.CheckReturn(Outs, CCAssignFnForNode(CallConv, /*Return=*/true,
2058 static SDValue LowerInterruptReturn(SmallVectorImpl<SDValue> &RetOps,
2059 SDLoc DL, SelectionDAG &DAG) {
2060 const MachineFunction &MF = DAG.getMachineFunction();
2061 const Function *F = MF.getFunction();
2063 StringRef IntKind = F->getFnAttribute("interrupt").getValueAsString();
2065 // See ARM ARM v7 B1.8.3. On exception entry LR is set to a possibly offset
2066 // version of the "preferred return address". These offsets affect the return
2067 // instruction if this is a return from PL1 without hypervisor extensions.
2068 // IRQ/FIQ: +4 "subs pc, lr, #4"
2069 // SWI: 0 "subs pc, lr, #0"
2070 // ABORT: +4 "subs pc, lr, #4"
2071 // UNDEF: +4/+2 "subs pc, lr, #0"
2072 // UNDEF varies depending on where the exception came from ARM or Thumb
2073 // mode. Alongside GCC, we throw our hands up in disgust and pretend it's 0.
2076 if (IntKind == "" || IntKind == "IRQ" || IntKind == "FIQ" ||
2079 else if (IntKind == "SWI" || IntKind == "UNDEF")
2082 report_fatal_error("Unsupported interrupt attribute. If present, value "
2083 "must be one of: IRQ, FIQ, SWI, ABORT or UNDEF");
2085 RetOps.insert(RetOps.begin() + 1, DAG.getConstant(LROffset, MVT::i32, false));
2087 return DAG.getNode(ARMISD::INTRET_FLAG, DL, MVT::Other, RetOps);
2091 ARMTargetLowering::LowerReturn(SDValue Chain,
2092 CallingConv::ID CallConv, bool isVarArg,
2093 const SmallVectorImpl<ISD::OutputArg> &Outs,
2094 const SmallVectorImpl<SDValue> &OutVals,
2095 SDLoc dl, SelectionDAG &DAG) const {
2097 // CCValAssign - represent the assignment of the return value to a location.
2098 SmallVector<CCValAssign, 16> RVLocs;
2100 // CCState - Info about the registers and stack slots.
2101 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2102 getTargetMachine(), RVLocs, *DAG.getContext(), Call);
2104 // Analyze outgoing return values.
2105 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
2109 SmallVector<SDValue, 4> RetOps;
2110 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
2111 bool isLittleEndian = Subtarget->isLittle();
2113 // Copy the result values into the output registers.
2114 for (unsigned i = 0, realRVLocIdx = 0;
2116 ++i, ++realRVLocIdx) {
2117 CCValAssign &VA = RVLocs[i];
2118 assert(VA.isRegLoc() && "Can only return in registers!");
2120 SDValue Arg = OutVals[realRVLocIdx];
2122 switch (VA.getLocInfo()) {
2123 default: llvm_unreachable("Unknown loc info!");
2124 case CCValAssign::Full: break;
2125 case CCValAssign::BCvt:
2126 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
2130 if (VA.needsCustom()) {
2131 if (VA.getLocVT() == MVT::v2f64) {
2132 // Extract the first half and return it in two registers.
2133 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
2134 DAG.getConstant(0, MVT::i32));
2135 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
2136 DAG.getVTList(MVT::i32, MVT::i32), Half);
2138 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2139 HalfGPRs.getValue(isLittleEndian ? 0 : 1),
2141 Flag = Chain.getValue(1);
2142 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2143 VA = RVLocs[++i]; // skip ahead to next loc
2144 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2145 HalfGPRs.getValue(isLittleEndian ? 1 : 0),
2147 Flag = Chain.getValue(1);
2148 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2149 VA = RVLocs[++i]; // skip ahead to next loc
2151 // Extract the 2nd half and fall through to handle it as an f64 value.
2152 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
2153 DAG.getConstant(1, MVT::i32));
2155 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
2157 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
2158 DAG.getVTList(MVT::i32, MVT::i32), Arg);
2159 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2160 fmrrd.getValue(isLittleEndian ? 0 : 1),
2162 Flag = Chain.getValue(1);
2163 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2164 VA = RVLocs[++i]; // skip ahead to next loc
2165 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2166 fmrrd.getValue(isLittleEndian ? 1 : 0),
2169 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
2171 // Guarantee that all emitted copies are
2172 // stuck together, avoiding something bad.
2173 Flag = Chain.getValue(1);
2174 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2177 // Update chain and glue.
2180 RetOps.push_back(Flag);
2182 // CPUs which aren't M-class use a special sequence to return from
2183 // exceptions (roughly, any instruction setting pc and cpsr simultaneously,
2184 // though we use "subs pc, lr, #N").
2186 // M-class CPUs actually use a normal return sequence with a special
2187 // (hardware-provided) value in LR, so the normal code path works.
2188 if (DAG.getMachineFunction().getFunction()->hasFnAttribute("interrupt") &&
2189 !Subtarget->isMClass()) {
2190 if (Subtarget->isThumb1Only())
2191 report_fatal_error("interrupt attribute is not supported in Thumb1");
2192 return LowerInterruptReturn(RetOps, dl, DAG);
2195 return DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, RetOps);
2198 bool ARMTargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
2199 if (N->getNumValues() != 1)
2201 if (!N->hasNUsesOfValue(1, 0))
2204 SDValue TCChain = Chain;
2205 SDNode *Copy = *N->use_begin();
2206 if (Copy->getOpcode() == ISD::CopyToReg) {
2207 // If the copy has a glue operand, we conservatively assume it isn't safe to
2208 // perform a tail call.
2209 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
2211 TCChain = Copy->getOperand(0);
2212 } else if (Copy->getOpcode() == ARMISD::VMOVRRD) {
2213 SDNode *VMov = Copy;
2214 // f64 returned in a pair of GPRs.
2215 SmallPtrSet<SDNode*, 2> Copies;
2216 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
2218 if (UI->getOpcode() != ISD::CopyToReg)
2222 if (Copies.size() > 2)
2225 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
2227 SDValue UseChain = UI->getOperand(0);
2228 if (Copies.count(UseChain.getNode()))
2235 } else if (Copy->getOpcode() == ISD::BITCAST) {
2236 // f32 returned in a single GPR.
2237 if (!Copy->hasOneUse())
2239 Copy = *Copy->use_begin();
2240 if (Copy->getOpcode() != ISD::CopyToReg || !Copy->hasNUsesOfValue(1, 0))
2242 TCChain = Copy->getOperand(0);
2247 bool HasRet = false;
2248 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
2250 if (UI->getOpcode() != ARMISD::RET_FLAG &&
2251 UI->getOpcode() != ARMISD::INTRET_FLAG)
2263 bool ARMTargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
2264 if (!Subtarget->supportsTailCall())
2267 if (!CI->isTailCall() || getTargetMachine().Options.DisableTailCalls)
2270 return !Subtarget->isThumb1Only();
2273 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
2274 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
2275 // one of the above mentioned nodes. It has to be wrapped because otherwise
2276 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
2277 // be used to form addressing mode. These wrapped nodes will be selected
2279 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
2280 EVT PtrVT = Op.getValueType();
2281 // FIXME there is no actual debug info here
2283 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
2285 if (CP->isMachineConstantPoolEntry())
2286 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
2287 CP->getAlignment());
2289 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
2290 CP->getAlignment());
2291 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
2294 unsigned ARMTargetLowering::getJumpTableEncoding() const {
2295 return MachineJumpTableInfo::EK_Inline;
2298 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
2299 SelectionDAG &DAG) const {
2300 MachineFunction &MF = DAG.getMachineFunction();
2301 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2302 unsigned ARMPCLabelIndex = 0;
2304 EVT PtrVT = getPointerTy();
2305 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
2306 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2308 if (RelocM == Reloc::Static) {
2309 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
2311 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2312 ARMPCLabelIndex = AFI->createPICLabelUId();
2313 ARMConstantPoolValue *CPV =
2314 ARMConstantPoolConstant::Create(BA, ARMPCLabelIndex,
2315 ARMCP::CPBlockAddress, PCAdj);
2316 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2318 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
2319 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
2320 MachinePointerInfo::getConstantPool(),
2321 false, false, false, 0);
2322 if (RelocM == Reloc::Static)
2324 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2325 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
2328 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
2330 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
2331 SelectionDAG &DAG) const {
2333 EVT PtrVT = getPointerTy();
2334 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2335 MachineFunction &MF = DAG.getMachineFunction();
2336 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2337 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2338 ARMConstantPoolValue *CPV =
2339 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2340 ARMCP::CPValue, PCAdj, ARMCP::TLSGD, true);
2341 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2342 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
2343 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
2344 MachinePointerInfo::getConstantPool(),
2345 false, false, false, 0);
2346 SDValue Chain = Argument.getValue(1);
2348 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2349 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
2351 // call __tls_get_addr.
2354 Entry.Node = Argument;
2355 Entry.Ty = (Type *) Type::getInt32Ty(*DAG.getContext());
2356 Args.push_back(Entry);
2358 // FIXME: is there useful debug info available here?
2359 TargetLowering::CallLoweringInfo CLI(DAG);
2360 CLI.setDebugLoc(dl).setChain(Chain)
2361 .setCallee(CallingConv::C, Type::getInt32Ty(*DAG.getContext()),
2362 DAG.getExternalSymbol("__tls_get_addr", PtrVT), std::move(Args),
2365 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
2366 return CallResult.first;
2369 // Lower ISD::GlobalTLSAddress using the "initial exec" or
2370 // "local exec" model.
2372 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
2374 TLSModel::Model model) const {
2375 const GlobalValue *GV = GA->getGlobal();
2378 SDValue Chain = DAG.getEntryNode();
2379 EVT PtrVT = getPointerTy();
2380 // Get the Thread Pointer
2381 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2383 if (model == TLSModel::InitialExec) {
2384 MachineFunction &MF = DAG.getMachineFunction();
2385 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2386 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2387 // Initial exec model.
2388 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2389 ARMConstantPoolValue *CPV =
2390 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2391 ARMCP::CPValue, PCAdj, ARMCP::GOTTPOFF,
2393 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2394 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2395 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2396 MachinePointerInfo::getConstantPool(),
2397 false, false, false, 0);
2398 Chain = Offset.getValue(1);
2400 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2401 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
2403 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2404 MachinePointerInfo::getConstantPool(),
2405 false, false, false, 0);
2408 assert(model == TLSModel::LocalExec);
2409 ARMConstantPoolValue *CPV =
2410 ARMConstantPoolConstant::Create(GV, ARMCP::TPOFF);
2411 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2412 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2413 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2414 MachinePointerInfo::getConstantPool(),
2415 false, false, false, 0);
2418 // The address of the thread local variable is the add of the thread
2419 // pointer with the offset of the variable.
2420 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
2424 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
2425 // TODO: implement the "local dynamic" model
2426 assert(Subtarget->isTargetELF() &&
2427 "TLS not implemented for non-ELF targets");
2428 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
2430 TLSModel::Model model = getTargetMachine().getTLSModel(GA->getGlobal());
2433 case TLSModel::GeneralDynamic:
2434 case TLSModel::LocalDynamic:
2435 return LowerToTLSGeneralDynamicModel(GA, DAG);
2436 case TLSModel::InitialExec:
2437 case TLSModel::LocalExec:
2438 return LowerToTLSExecModels(GA, DAG, model);
2440 llvm_unreachable("bogus TLS model");
2443 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
2444 SelectionDAG &DAG) const {
2445 EVT PtrVT = getPointerTy();
2447 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2448 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2449 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
2450 ARMConstantPoolValue *CPV =
2451 ARMConstantPoolConstant::Create(GV,
2452 UseGOTOFF ? ARMCP::GOTOFF : ARMCP::GOT);
2453 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2454 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2455 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
2457 MachinePointerInfo::getConstantPool(),
2458 false, false, false, 0);
2459 SDValue Chain = Result.getValue(1);
2460 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
2461 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
2463 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
2464 MachinePointerInfo::getGOT(),
2465 false, false, false, 0);
2469 // If we have T2 ops, we can materialize the address directly via movt/movw
2470 // pair. This is always cheaper.
2471 if (Subtarget->useMovt(DAG.getMachineFunction())) {
2473 // FIXME: Once remat is capable of dealing with instructions with register
2474 // operands, expand this into two nodes.
2475 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2476 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2478 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2479 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2480 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2481 MachinePointerInfo::getConstantPool(),
2482 false, false, false, 0);
2486 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
2487 SelectionDAG &DAG) const {
2488 EVT PtrVT = getPointerTy();
2490 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2491 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2493 if (Subtarget->useMovt(DAG.getMachineFunction()))
2496 // FIXME: Once remat is capable of dealing with instructions with register
2497 // operands, expand this into multiple nodes
2499 RelocM == Reloc::PIC_ ? ARMISD::WrapperPIC : ARMISD::Wrapper;
2501 SDValue G = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, ARMII::MO_NONLAZY);
2502 SDValue Result = DAG.getNode(Wrapper, dl, PtrVT, G);
2504 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2505 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
2506 MachinePointerInfo::getGOT(), false, false, false, 0);
2510 SDValue ARMTargetLowering::LowerGlobalAddressWindows(SDValue Op,
2511 SelectionDAG &DAG) const {
2512 assert(Subtarget->isTargetWindows() && "non-Windows COFF is not supported");
2513 assert(Subtarget->useMovt(DAG.getMachineFunction()) &&
2514 "Windows on ARM expects to use movw/movt");
2516 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2517 const ARMII::TOF TargetFlags =
2518 (GV->hasDLLImportStorageClass() ? ARMII::MO_DLLIMPORT : ARMII::MO_NO_FLAG);
2519 EVT PtrVT = getPointerTy();
2525 // FIXME: Once remat is capable of dealing with instructions with register
2526 // operands, expand this into two nodes.
2527 Result = DAG.getNode(ARMISD::Wrapper, DL, PtrVT,
2528 DAG.getTargetGlobalAddress(GV, DL, PtrVT, /*Offset=*/0,
2530 if (GV->hasDLLImportStorageClass())
2531 Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), Result,
2532 MachinePointerInfo::getGOT(), false, false, false, 0);
2536 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
2537 SelectionDAG &DAG) const {
2538 assert(Subtarget->isTargetELF() &&
2539 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
2540 MachineFunction &MF = DAG.getMachineFunction();
2541 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2542 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2543 EVT PtrVT = getPointerTy();
2545 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2546 ARMConstantPoolValue *CPV =
2547 ARMConstantPoolSymbol::Create(*DAG.getContext(), "_GLOBAL_OFFSET_TABLE_",
2548 ARMPCLabelIndex, PCAdj);
2549 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2550 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2551 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2552 MachinePointerInfo::getConstantPool(),
2553 false, false, false, 0);
2554 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2555 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2559 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
2561 SDValue Val = DAG.getConstant(0, MVT::i32);
2562 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl,
2563 DAG.getVTList(MVT::i32, MVT::Other), Op.getOperand(0),
2564 Op.getOperand(1), Val);
2568 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
2570 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
2571 Op.getOperand(1), DAG.getConstant(0, MVT::i32));
2575 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
2576 const ARMSubtarget *Subtarget) const {
2577 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2580 default: return SDValue(); // Don't custom lower most intrinsics.
2581 case Intrinsic::arm_rbit: {
2582 assert(Op.getOperand(0).getValueType() == MVT::i32 &&
2583 "RBIT intrinsic must have i32 type!");
2584 return DAG.getNode(ARMISD::RBIT, dl, MVT::i32, Op.getOperand(0));
2586 case Intrinsic::arm_thread_pointer: {
2587 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2588 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2590 case Intrinsic::eh_sjlj_lsda: {
2591 MachineFunction &MF = DAG.getMachineFunction();
2592 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2593 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2594 EVT PtrVT = getPointerTy();
2595 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2597 unsigned PCAdj = (RelocM != Reloc::PIC_)
2598 ? 0 : (Subtarget->isThumb() ? 4 : 8);
2599 ARMConstantPoolValue *CPV =
2600 ARMConstantPoolConstant::Create(MF.getFunction(), ARMPCLabelIndex,
2601 ARMCP::CPLSDA, PCAdj);
2602 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2603 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2605 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2606 MachinePointerInfo::getConstantPool(),
2607 false, false, false, 0);
2609 if (RelocM == Reloc::PIC_) {
2610 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2611 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2615 case Intrinsic::arm_neon_vmulls:
2616 case Intrinsic::arm_neon_vmullu: {
2617 unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vmulls)
2618 ? ARMISD::VMULLs : ARMISD::VMULLu;
2619 return DAG.getNode(NewOpc, SDLoc(Op), Op.getValueType(),
2620 Op.getOperand(1), Op.getOperand(2));
2625 static SDValue LowerATOMIC_FENCE(SDValue Op, SelectionDAG &DAG,
2626 const ARMSubtarget *Subtarget) {
2627 // FIXME: handle "fence singlethread" more efficiently.
2629 if (!Subtarget->hasDataBarrier()) {
2630 // Some ARMv6 cpus can support data barriers with an mcr instruction.
2631 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2633 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2634 "Unexpected ISD::ATOMIC_FENCE encountered. Should be libcall!");
2635 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2636 DAG.getConstant(0, MVT::i32));
2639 ConstantSDNode *OrdN = cast<ConstantSDNode>(Op.getOperand(1));
2640 AtomicOrdering Ord = static_cast<AtomicOrdering>(OrdN->getZExtValue());
2641 unsigned Domain = ARM_MB::ISH;
2642 if (Subtarget->isMClass()) {
2643 // Only a full system barrier exists in the M-class architectures.
2644 Domain = ARM_MB::SY;
2645 } else if (Subtarget->isSwift() && Ord == Release) {
2646 // Swift happens to implement ISHST barriers in a way that's compatible with
2647 // Release semantics but weaker than ISH so we'd be fools not to use
2648 // it. Beware: other processors probably don't!
2649 Domain = ARM_MB::ISHST;
2652 return DAG.getNode(ISD::INTRINSIC_VOID, dl, MVT::Other, Op.getOperand(0),
2653 DAG.getConstant(Intrinsic::arm_dmb, MVT::i32),
2654 DAG.getConstant(Domain, MVT::i32));
2657 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG,
2658 const ARMSubtarget *Subtarget) {
2659 // ARM pre v5TE and Thumb1 does not have preload instructions.
2660 if (!(Subtarget->isThumb2() ||
2661 (!Subtarget->isThumb1Only() && Subtarget->hasV5TEOps())))
2662 // Just preserve the chain.
2663 return Op.getOperand(0);
2666 unsigned isRead = ~cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() & 1;
2668 (!Subtarget->hasV7Ops() || !Subtarget->hasMPExtension()))
2669 // ARMv7 with MP extension has PLDW.
2670 return Op.getOperand(0);
2672 unsigned isData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
2673 if (Subtarget->isThumb()) {
2675 isRead = ~isRead & 1;
2676 isData = ~isData & 1;
2679 return DAG.getNode(ARMISD::PRELOAD, dl, MVT::Other, Op.getOperand(0),
2680 Op.getOperand(1), DAG.getConstant(isRead, MVT::i32),
2681 DAG.getConstant(isData, MVT::i32));
2684 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
2685 MachineFunction &MF = DAG.getMachineFunction();
2686 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
2688 // vastart just stores the address of the VarArgsFrameIndex slot into the
2689 // memory location argument.
2691 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2692 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2693 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
2694 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
2695 MachinePointerInfo(SV), false, false, 0);
2699 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
2700 SDValue &Root, SelectionDAG &DAG,
2702 MachineFunction &MF = DAG.getMachineFunction();
2703 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2705 const TargetRegisterClass *RC;
2706 if (AFI->isThumb1OnlyFunction())
2707 RC = &ARM::tGPRRegClass;
2709 RC = &ARM::GPRRegClass;
2711 // Transform the arguments stored in physical registers into virtual ones.
2712 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2713 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2716 if (NextVA.isMemLoc()) {
2717 MachineFrameInfo *MFI = MF.getFrameInfo();
2718 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true);
2720 // Create load node to retrieve arguments from the stack.
2721 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2722 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
2723 MachinePointerInfo::getFixedStack(FI),
2724 false, false, false, 0);
2726 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
2727 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2729 if (!Subtarget->isLittle())
2730 std::swap (ArgValue, ArgValue2);
2731 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
2735 ARMTargetLowering::computeRegArea(CCState &CCInfo, MachineFunction &MF,
2736 unsigned InRegsParamRecordIdx,
2738 unsigned &ArgRegsSize,
2739 unsigned &ArgRegsSaveSize)
2742 if (InRegsParamRecordIdx < CCInfo.getInRegsParamsCount()) {
2743 unsigned RBegin, REnd;
2744 CCInfo.getInRegsParamInfo(InRegsParamRecordIdx, RBegin, REnd);
2745 NumGPRs = REnd - RBegin;
2747 unsigned int firstUnalloced;
2748 firstUnalloced = CCInfo.getFirstUnallocated(GPRArgRegs,
2749 sizeof(GPRArgRegs) /
2750 sizeof(GPRArgRegs[0]));
2751 NumGPRs = (firstUnalloced <= 3) ? (4 - firstUnalloced) : 0;
2754 unsigned Align = MF.getTarget().getFrameLowering()->getStackAlignment();
2755 ArgRegsSize = NumGPRs * 4;
2757 // If parameter is split between stack and GPRs...
2758 if (NumGPRs && Align > 4 &&
2759 (ArgRegsSize < ArgSize ||
2760 InRegsParamRecordIdx >= CCInfo.getInRegsParamsCount())) {
2761 // Add padding for part of param recovered from GPRs. For example,
2762 // if Align == 8, its last byte must be at address K*8 - 1.
2763 // We need to do it, since remained (stack) part of parameter has
2764 // stack alignment, and we need to "attach" "GPRs head" without gaps
2767 // |---- 8 bytes block ----| |---- 8 bytes block ----| |---- 8 bytes...
2768 // [ [padding] [GPRs head] ] [ Tail passed via stack ....
2770 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2772 OffsetToAlignment(ArgRegsSize + AFI->getArgRegsSaveSize(), Align);
2773 ArgRegsSaveSize = ArgRegsSize + Padding;
2775 // We don't need to extend regs save size for byval parameters if they
2776 // are passed via GPRs only.
2777 ArgRegsSaveSize = ArgRegsSize;
2780 // The remaining GPRs hold either the beginning of variable-argument
2781 // data, or the beginning of an aggregate passed by value (usually
2782 // byval). Either way, we allocate stack slots adjacent to the data
2783 // provided by our caller, and store the unallocated registers there.
2784 // If this is a variadic function, the va_list pointer will begin with
2785 // these values; otherwise, this reassembles a (byval) structure that
2786 // was split between registers and memory.
2787 // Return: The frame index registers were stored into.
2789 ARMTargetLowering::StoreByValRegs(CCState &CCInfo, SelectionDAG &DAG,
2790 SDLoc dl, SDValue &Chain,
2791 const Value *OrigArg,
2792 unsigned InRegsParamRecordIdx,
2793 unsigned OffsetFromOrigArg,
2797 unsigned ByValStoreOffset,
2798 unsigned TotalArgRegsSaveSize) const {
2800 // Currently, two use-cases possible:
2801 // Case #1. Non-var-args function, and we meet first byval parameter.
2802 // Setup first unallocated register as first byval register;
2803 // eat all remained registers
2804 // (these two actions are performed by HandleByVal method).
2805 // Then, here, we initialize stack frame with
2806 // "store-reg" instructions.
2807 // Case #2. Var-args function, that doesn't contain byval parameters.
2808 // The same: eat all remained unallocated registers,
2809 // initialize stack frame.
2811 MachineFunction &MF = DAG.getMachineFunction();
2812 MachineFrameInfo *MFI = MF.getFrameInfo();
2813 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2814 unsigned firstRegToSaveIndex, lastRegToSaveIndex;
2815 unsigned RBegin, REnd;
2816 if (InRegsParamRecordIdx < CCInfo.getInRegsParamsCount()) {
2817 CCInfo.getInRegsParamInfo(InRegsParamRecordIdx, RBegin, REnd);
2818 firstRegToSaveIndex = RBegin - ARM::R0;
2819 lastRegToSaveIndex = REnd - ARM::R0;
2821 firstRegToSaveIndex = CCInfo.getFirstUnallocated
2822 (GPRArgRegs, array_lengthof(GPRArgRegs));
2823 lastRegToSaveIndex = 4;
2826 unsigned ArgRegsSize, ArgRegsSaveSize;
2827 computeRegArea(CCInfo, MF, InRegsParamRecordIdx, ArgSize,
2828 ArgRegsSize, ArgRegsSaveSize);
2830 // Store any by-val regs to their spots on the stack so that they may be
2831 // loaded by deferencing the result of formal parameter pointer or va_next.
2832 // Note: once stack area for byval/varargs registers
2833 // was initialized, it can't be initialized again.
2834 if (ArgRegsSaveSize) {
2835 unsigned Padding = ArgRegsSaveSize - ArgRegsSize;
2838 assert(AFI->getStoredByValParamsPadding() == 0 &&
2839 "The only parameter may be padded.");
2840 AFI->setStoredByValParamsPadding(Padding);
2843 int FrameIndex = MFI->CreateFixedObject(ArgRegsSaveSize,
2846 (int64_t)TotalArgRegsSaveSize,
2848 SDValue FIN = DAG.getFrameIndex(FrameIndex, getPointerTy());
2850 MFI->CreateFixedObject(Padding,
2851 ArgOffset + ByValStoreOffset -
2852 (int64_t)ArgRegsSaveSize,
2856 SmallVector<SDValue, 4> MemOps;
2857 for (unsigned i = 0; firstRegToSaveIndex < lastRegToSaveIndex;
2858 ++firstRegToSaveIndex, ++i) {
2859 const TargetRegisterClass *RC;
2860 if (AFI->isThumb1OnlyFunction())
2861 RC = &ARM::tGPRRegClass;
2863 RC = &ARM::GPRRegClass;
2865 unsigned VReg = MF.addLiveIn(GPRArgRegs[firstRegToSaveIndex], RC);
2866 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2868 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2869 MachinePointerInfo(OrigArg, OffsetFromOrigArg + 4*i),
2871 MemOps.push_back(Store);
2872 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
2873 DAG.getConstant(4, getPointerTy()));
2876 AFI->setArgRegsSaveSize(ArgRegsSaveSize + AFI->getArgRegsSaveSize());
2878 if (!MemOps.empty())
2879 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOps);
2883 // We cannot allocate a zero-byte object for the first variadic argument,
2884 // so just make up a size.
2887 // This will point to the next argument passed via stack.
2888 return MFI->CreateFixedObject(
2889 ArgSize, ArgOffset, !ForceMutable);
2893 // Setup stack frame, the va_list pointer will start from.
2895 ARMTargetLowering::VarArgStyleRegisters(CCState &CCInfo, SelectionDAG &DAG,
2896 SDLoc dl, SDValue &Chain,
2898 unsigned TotalArgRegsSaveSize,
2899 bool ForceMutable) const {
2900 MachineFunction &MF = DAG.getMachineFunction();
2901 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2903 // Try to store any remaining integer argument regs
2904 // to their spots on the stack so that they may be loaded by deferencing
2905 // the result of va_next.
2906 // If there is no regs to be stored, just point address after last
2907 // argument passed via stack.
2909 StoreByValRegs(CCInfo, DAG, dl, Chain, nullptr,
2910 CCInfo.getInRegsParamsCount(), 0, ArgOffset, 0, ForceMutable,
2911 0, TotalArgRegsSaveSize);
2913 AFI->setVarArgsFrameIndex(FrameIndex);
2917 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
2918 CallingConv::ID CallConv, bool isVarArg,
2919 const SmallVectorImpl<ISD::InputArg>
2921 SDLoc dl, SelectionDAG &DAG,
2922 SmallVectorImpl<SDValue> &InVals)
2924 MachineFunction &MF = DAG.getMachineFunction();
2925 MachineFrameInfo *MFI = MF.getFrameInfo();
2927 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2929 // Assign locations to all of the incoming arguments.
2930 SmallVector<CCValAssign, 16> ArgLocs;
2931 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2932 getTargetMachine(), ArgLocs, *DAG.getContext(), Prologue);
2933 CCInfo.AnalyzeFormalArguments(Ins,
2934 CCAssignFnForNode(CallConv, /* Return*/ false,
2937 SmallVector<SDValue, 16> ArgValues;
2938 int lastInsIndex = -1;
2940 Function::const_arg_iterator CurOrigArg = MF.getFunction()->arg_begin();
2941 unsigned CurArgIdx = 0;
2943 // Initially ArgRegsSaveSize is zero.
2944 // Then we increase this value each time we meet byval parameter.
2945 // We also increase this value in case of varargs function.
2946 AFI->setArgRegsSaveSize(0);
2948 unsigned ByValStoreOffset = 0;
2949 unsigned TotalArgRegsSaveSize = 0;
2950 unsigned ArgRegsSaveSizeMaxAlign = 4;
2952 // Calculate the amount of stack space that we need to allocate to store
2953 // byval and variadic arguments that are passed in registers.
2954 // We need to know this before we allocate the first byval or variadic
2955 // argument, as they will be allocated a stack slot below the CFA (Canonical
2956 // Frame Address, the stack pointer at entry to the function).
2957 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2958 CCValAssign &VA = ArgLocs[i];
2959 if (VA.isMemLoc()) {
2960 int index = VA.getValNo();
2961 if (index != lastInsIndex) {
2962 ISD::ArgFlagsTy Flags = Ins[index].Flags;
2963 if (Flags.isByVal()) {
2964 unsigned ExtraArgRegsSize;
2965 unsigned ExtraArgRegsSaveSize;
2966 computeRegArea(CCInfo, MF, CCInfo.getInRegsParamsProceed(),
2967 Flags.getByValSize(),
2968 ExtraArgRegsSize, ExtraArgRegsSaveSize);
2970 TotalArgRegsSaveSize += ExtraArgRegsSaveSize;
2971 if (Flags.getByValAlign() > ArgRegsSaveSizeMaxAlign)
2972 ArgRegsSaveSizeMaxAlign = Flags.getByValAlign();
2973 CCInfo.nextInRegsParam();
2975 lastInsIndex = index;
2979 CCInfo.rewindByValRegsInfo();
2982 unsigned ExtraArgRegsSize;
2983 unsigned ExtraArgRegsSaveSize;
2984 computeRegArea(CCInfo, MF, CCInfo.getInRegsParamsCount(), 0,
2985 ExtraArgRegsSize, ExtraArgRegsSaveSize);
2986 TotalArgRegsSaveSize += ExtraArgRegsSaveSize;
2988 // If the arg regs save area contains N-byte aligned values, the
2989 // bottom of it must be at least N-byte aligned.
2990 TotalArgRegsSaveSize = RoundUpToAlignment(TotalArgRegsSaveSize, ArgRegsSaveSizeMaxAlign);
2991 TotalArgRegsSaveSize = std::min(TotalArgRegsSaveSize, 16U);
2993 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2994 CCValAssign &VA = ArgLocs[i];
2995 std::advance(CurOrigArg, Ins[VA.getValNo()].OrigArgIndex - CurArgIdx);
2996 CurArgIdx = Ins[VA.getValNo()].OrigArgIndex;
2997 // Arguments stored in registers.
2998 if (VA.isRegLoc()) {
2999 EVT RegVT = VA.getLocVT();
3001 if (VA.needsCustom()) {
3002 // f64 and vector types are split up into multiple registers or
3003 // combinations of registers and stack slots.
3004 if (VA.getLocVT() == MVT::v2f64) {
3005 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
3007 VA = ArgLocs[++i]; // skip ahead to next loc
3009 if (VA.isMemLoc()) {
3010 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true);
3011 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
3012 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
3013 MachinePointerInfo::getFixedStack(FI),
3014 false, false, false, 0);
3016 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
3019 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
3020 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
3021 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
3022 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
3023 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
3025 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
3028 const TargetRegisterClass *RC;
3030 if (RegVT == MVT::f32)
3031 RC = &ARM::SPRRegClass;
3032 else if (RegVT == MVT::f64)
3033 RC = &ARM::DPRRegClass;
3034 else if (RegVT == MVT::v2f64)
3035 RC = &ARM::QPRRegClass;
3036 else if (RegVT == MVT::i32)
3037 RC = AFI->isThumb1OnlyFunction() ?
3038 (const TargetRegisterClass*)&ARM::tGPRRegClass :
3039 (const TargetRegisterClass*)&ARM::GPRRegClass;
3041 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
3043 // Transform the arguments in physical registers into virtual ones.
3044 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
3045 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
3048 // If this is an 8 or 16-bit value, it is really passed promoted
3049 // to 32 bits. Insert an assert[sz]ext to capture this, then
3050 // truncate to the right size.
3051 switch (VA.getLocInfo()) {
3052 default: llvm_unreachable("Unknown loc info!");
3053 case CCValAssign::Full: break;
3054 case CCValAssign::BCvt:
3055 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
3057 case CCValAssign::SExt:
3058 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
3059 DAG.getValueType(VA.getValVT()));
3060 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
3062 case CCValAssign::ZExt:
3063 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
3064 DAG.getValueType(VA.getValVT()));
3065 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
3069 InVals.push_back(ArgValue);
3071 } else { // VA.isRegLoc()
3074 assert(VA.isMemLoc());
3075 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
3077 int index = ArgLocs[i].getValNo();
3079 // Some Ins[] entries become multiple ArgLoc[] entries.
3080 // Process them only once.
3081 if (index != lastInsIndex)
3083 ISD::ArgFlagsTy Flags = Ins[index].Flags;
3084 // FIXME: For now, all byval parameter objects are marked mutable.
3085 // This can be changed with more analysis.
3086 // In case of tail call optimization mark all arguments mutable.
3087 // Since they could be overwritten by lowering of arguments in case of
3089 if (Flags.isByVal()) {
3090 unsigned CurByValIndex = CCInfo.getInRegsParamsProceed();
3092 ByValStoreOffset = RoundUpToAlignment(ByValStoreOffset, Flags.getByValAlign());
3093 int FrameIndex = StoreByValRegs(
3094 CCInfo, DAG, dl, Chain, CurOrigArg,
3096 Ins[VA.getValNo()].PartOffset,
3097 VA.getLocMemOffset(),
3098 Flags.getByValSize(),
3099 true /*force mutable frames*/,
3101 TotalArgRegsSaveSize);
3102 ByValStoreOffset += Flags.getByValSize();
3103 ByValStoreOffset = std::min(ByValStoreOffset, 16U);
3104 InVals.push_back(DAG.getFrameIndex(FrameIndex, getPointerTy()));
3105 CCInfo.nextInRegsParam();
3107 unsigned FIOffset = VA.getLocMemOffset();
3108 int FI = MFI->CreateFixedObject(VA.getLocVT().getSizeInBits()/8,
3111 // Create load nodes to retrieve arguments from the stack.
3112 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
3113 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
3114 MachinePointerInfo::getFixedStack(FI),
3115 false, false, false, 0));
3117 lastInsIndex = index;
3124 VarArgStyleRegisters(CCInfo, DAG, dl, Chain,
3125 CCInfo.getNextStackOffset(),
3126 TotalArgRegsSaveSize);
3128 AFI->setArgumentStackSize(CCInfo.getNextStackOffset());
3133 /// isFloatingPointZero - Return true if this is +0.0.
3134 static bool isFloatingPointZero(SDValue Op) {
3135 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
3136 return CFP->getValueAPF().isPosZero();
3137 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
3138 // Maybe this has already been legalized into the constant pool?
3139 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
3140 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
3141 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
3142 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
3143 return CFP->getValueAPF().isPosZero();
3149 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
3150 /// the given operands.
3152 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
3153 SDValue &ARMcc, SelectionDAG &DAG,
3155 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
3156 unsigned C = RHSC->getZExtValue();
3157 if (!isLegalICmpImmediate(C)) {
3158 // Constant does not fit, try adjusting it by one?
3163 if (C != 0x80000000 && isLegalICmpImmediate(C-1)) {
3164 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
3165 RHS = DAG.getConstant(C-1, MVT::i32);
3170 if (C != 0 && isLegalICmpImmediate(C-1)) {
3171 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
3172 RHS = DAG.getConstant(C-1, MVT::i32);
3177 if (C != 0x7fffffff && isLegalICmpImmediate(C+1)) {
3178 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
3179 RHS = DAG.getConstant(C+1, MVT::i32);
3184 if (C != 0xffffffff && isLegalICmpImmediate(C+1)) {
3185 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
3186 RHS = DAG.getConstant(C+1, MVT::i32);
3193 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3194 ARMISD::NodeType CompareType;
3197 CompareType = ARMISD::CMP;
3202 CompareType = ARMISD::CMPZ;
3205 ARMcc = DAG.getConstant(CondCode, MVT::i32);
3206 return DAG.getNode(CompareType, dl, MVT::Glue, LHS, RHS);
3209 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
3211 ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
3214 if (!isFloatingPointZero(RHS))
3215 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Glue, LHS, RHS);
3217 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Glue, LHS);
3218 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Glue, Cmp);
3221 /// duplicateCmp - Glue values can have only one use, so this function
3222 /// duplicates a comparison node.
3224 ARMTargetLowering::duplicateCmp(SDValue Cmp, SelectionDAG &DAG) const {
3225 unsigned Opc = Cmp.getOpcode();
3227 if (Opc == ARMISD::CMP || Opc == ARMISD::CMPZ)
3228 return DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
3230 assert(Opc == ARMISD::FMSTAT && "unexpected comparison operation");
3231 Cmp = Cmp.getOperand(0);
3232 Opc = Cmp.getOpcode();
3233 if (Opc == ARMISD::CMPFP)
3234 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
3236 assert(Opc == ARMISD::CMPFPw0 && "unexpected operand of FMSTAT");
3237 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0));
3239 return DAG.getNode(ARMISD::FMSTAT, DL, MVT::Glue, Cmp);
3242 std::pair<SDValue, SDValue>
3243 ARMTargetLowering::getARMXALUOOp(SDValue Op, SelectionDAG &DAG,
3244 SDValue &ARMcc) const {
3245 assert(Op.getValueType() == MVT::i32 && "Unsupported value type");
3247 SDValue Value, OverflowCmp;
3248 SDValue LHS = Op.getOperand(0);
3249 SDValue RHS = Op.getOperand(1);
3252 // FIXME: We are currently always generating CMPs because we don't support
3253 // generating CMN through the backend. This is not as good as the natural
3254 // CMP case because it causes a register dependency and cannot be folded
3257 switch (Op.getOpcode()) {
3259 llvm_unreachable("Unknown overflow instruction!");
3261 ARMcc = DAG.getConstant(ARMCC::VC, MVT::i32);
3262 Value = DAG.getNode(ISD::ADD, SDLoc(Op), Op.getValueType(), LHS, RHS);
3263 OverflowCmp = DAG.getNode(ARMISD::CMP, SDLoc(Op), MVT::Glue, Value, LHS);
3266 ARMcc = DAG.getConstant(ARMCC::HS, MVT::i32);
3267 Value = DAG.getNode(ISD::ADD, SDLoc(Op), Op.getValueType(), LHS, RHS);
3268 OverflowCmp = DAG.getNode(ARMISD::CMP, SDLoc(Op), MVT::Glue, Value, LHS);
3271 ARMcc = DAG.getConstant(ARMCC::VC, MVT::i32);
3272 Value = DAG.getNode(ISD::SUB, SDLoc(Op), Op.getValueType(), LHS, RHS);
3273 OverflowCmp = DAG.getNode(ARMISD::CMP, SDLoc(Op), MVT::Glue, LHS, RHS);
3276 ARMcc = DAG.getConstant(ARMCC::HS, MVT::i32);
3277 Value = DAG.getNode(ISD::SUB, SDLoc(Op), Op.getValueType(), LHS, RHS);
3278 OverflowCmp = DAG.getNode(ARMISD::CMP, SDLoc(Op), MVT::Glue, LHS, RHS);
3282 return std::make_pair(Value, OverflowCmp);
3287 ARMTargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) const {
3288 // Let legalize expand this if it isn't a legal type yet.
3289 if (!DAG.getTargetLoweringInfo().isTypeLegal(Op.getValueType()))
3292 SDValue Value, OverflowCmp;
3294 std::tie(Value, OverflowCmp) = getARMXALUOOp(Op, DAG, ARMcc);
3295 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3296 // We use 0 and 1 as false and true values.
3297 SDValue TVal = DAG.getConstant(1, MVT::i32);
3298 SDValue FVal = DAG.getConstant(0, MVT::i32);
3299 EVT VT = Op.getValueType();
3301 SDValue Overflow = DAG.getNode(ARMISD::CMOV, SDLoc(Op), VT, TVal, FVal,
3302 ARMcc, CCR, OverflowCmp);
3304 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
3305 return DAG.getNode(ISD::MERGE_VALUES, SDLoc(Op), VTs, Value, Overflow);
3309 SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
3310 SDValue Cond = Op.getOperand(0);
3311 SDValue SelectTrue = Op.getOperand(1);
3312 SDValue SelectFalse = Op.getOperand(2);
3314 unsigned Opc = Cond.getOpcode();
3316 if (Cond.getResNo() == 1 &&
3317 (Opc == ISD::SADDO || Opc == ISD::UADDO || Opc == ISD::SSUBO ||
3318 Opc == ISD::USUBO)) {
3319 if (!DAG.getTargetLoweringInfo().isTypeLegal(Cond->getValueType(0)))
3322 SDValue Value, OverflowCmp;
3324 std::tie(Value, OverflowCmp) = getARMXALUOOp(Cond, DAG, ARMcc);
3325 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3326 EVT VT = Op.getValueType();
3328 return DAG.getNode(ARMISD::CMOV, SDLoc(Op), VT, SelectTrue, SelectFalse,
3329 ARMcc, CCR, OverflowCmp);
3335 // (select (cmov 1, 0, cond), t, f) -> (cmov t, f, cond)
3336 // (select (cmov 0, 1, cond), t, f) -> (cmov f, t, cond)
3338 if (Cond.getOpcode() == ARMISD::CMOV && Cond.hasOneUse()) {
3339 const ConstantSDNode *CMOVTrue =
3340 dyn_cast<ConstantSDNode>(Cond.getOperand(0));
3341 const ConstantSDNode *CMOVFalse =
3342 dyn_cast<ConstantSDNode>(Cond.getOperand(1));
3344 if (CMOVTrue && CMOVFalse) {
3345 unsigned CMOVTrueVal = CMOVTrue->getZExtValue();
3346 unsigned CMOVFalseVal = CMOVFalse->getZExtValue();
3350 if (CMOVTrueVal == 1 && CMOVFalseVal == 0) {
3352 False = SelectFalse;
3353 } else if (CMOVTrueVal == 0 && CMOVFalseVal == 1) {
3358 if (True.getNode() && False.getNode()) {
3359 EVT VT = Op.getValueType();
3360 SDValue ARMcc = Cond.getOperand(2);
3361 SDValue CCR = Cond.getOperand(3);
3362 SDValue Cmp = duplicateCmp(Cond.getOperand(4), DAG);
3363 assert(True.getValueType() == VT);
3364 return DAG.getNode(ARMISD::CMOV, dl, VT, True, False, ARMcc, CCR, Cmp);
3369 // ARM's BooleanContents value is UndefinedBooleanContent. Mask out the
3370 // undefined bits before doing a full-word comparison with zero.
3371 Cond = DAG.getNode(ISD::AND, dl, Cond.getValueType(), Cond,
3372 DAG.getConstant(1, Cond.getValueType()));
3374 return DAG.getSelectCC(dl, Cond,
3375 DAG.getConstant(0, Cond.getValueType()),
3376 SelectTrue, SelectFalse, ISD::SETNE);
3379 static ISD::CondCode getInverseCCForVSEL(ISD::CondCode CC) {
3380 if (CC == ISD::SETNE)
3382 return ISD::getSetCCInverse(CC, true);
3385 static void checkVSELConstraints(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
3386 bool &swpCmpOps, bool &swpVselOps) {
3387 // Start by selecting the GE condition code for opcodes that return true for
3389 if (CC == ISD::SETUGE || CC == ISD::SETOGE || CC == ISD::SETOLE ||
3391 CondCode = ARMCC::GE;
3393 // and GT for opcodes that return false for 'equality'.
3394 else if (CC == ISD::SETUGT || CC == ISD::SETOGT || CC == ISD::SETOLT ||
3396 CondCode = ARMCC::GT;
3398 // Since we are constrained to GE/GT, if the opcode contains 'less', we need
3399 // to swap the compare operands.
3400 if (CC == ISD::SETOLE || CC == ISD::SETULE || CC == ISD::SETOLT ||
3404 // Both GT and GE are ordered comparisons, and return false for 'unordered'.
3405 // If we have an unordered opcode, we need to swap the operands to the VSEL
3406 // instruction (effectively negating the condition).
3408 // This also has the effect of swapping which one of 'less' or 'greater'
3409 // returns true, so we also swap the compare operands. It also switches
3410 // whether we return true for 'equality', so we compensate by picking the
3411 // opposite condition code to our original choice.
3412 if (CC == ISD::SETULE || CC == ISD::SETULT || CC == ISD::SETUGE ||
3413 CC == ISD::SETUGT) {
3414 swpCmpOps = !swpCmpOps;
3415 swpVselOps = !swpVselOps;
3416 CondCode = CondCode == ARMCC::GT ? ARMCC::GE : ARMCC::GT;
3419 // 'ordered' is 'anything but unordered', so use the VS condition code and
3420 // swap the VSEL operands.
3421 if (CC == ISD::SETO) {
3422 CondCode = ARMCC::VS;
3426 // 'unordered or not equal' is 'anything but equal', so use the EQ condition
3427 // code and swap the VSEL operands.
3428 if (CC == ISD::SETUNE) {
3429 CondCode = ARMCC::EQ;
3434 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
3435 EVT VT = Op.getValueType();
3436 SDValue LHS = Op.getOperand(0);
3437 SDValue RHS = Op.getOperand(1);
3438 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
3439 SDValue TrueVal = Op.getOperand(2);
3440 SDValue FalseVal = Op.getOperand(3);
3443 if (LHS.getValueType() == MVT::i32) {
3444 // Try to generate VSEL on ARMv8.
3445 // The VSEL instruction can't use all the usual ARM condition
3446 // codes: it only has two bits to select the condition code, so it's
3447 // constrained to use only GE, GT, VS and EQ.
3449 // To implement all the various ISD::SETXXX opcodes, we sometimes need to
3450 // swap the operands of the previous compare instruction (effectively
3451 // inverting the compare condition, swapping 'less' and 'greater') and
3452 // sometimes need to swap the operands to the VSEL (which inverts the
3453 // condition in the sense of firing whenever the previous condition didn't)
3454 if (getSubtarget()->hasFPARMv8() && (TrueVal.getValueType() == MVT::f32 ||
3455 TrueVal.getValueType() == MVT::f64)) {
3456 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3457 if (CondCode == ARMCC::LT || CondCode == ARMCC::LE ||
3458 CondCode == ARMCC::VC || CondCode == ARMCC::NE) {
3459 CC = getInverseCCForVSEL(CC);
3460 std::swap(TrueVal, FalseVal);
3465 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3466 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3467 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,
3471 ARMCC::CondCodes CondCode, CondCode2;
3472 FPCCToARMCC(CC, CondCode, CondCode2);
3474 // Try to generate VSEL on ARMv8.
3475 if (getSubtarget()->hasFPARMv8() && (TrueVal.getValueType() == MVT::f32 ||
3476 TrueVal.getValueType() == MVT::f64)) {
3477 // We can select VMAXNM/VMINNM from a compare followed by a select with the
3478 // same operands, as follows:
3479 // c = fcmp [ogt, olt, ugt, ult] a, b
3481 // We only do this in unsafe-fp-math, because signed zeros and NaNs are
3482 // handled differently than the original code sequence.
3483 if (getTargetMachine().Options.UnsafeFPMath && LHS == TrueVal &&
3485 if (CC == ISD::SETOGT || CC == ISD::SETUGT)
3486 return DAG.getNode(ARMISD::VMAXNM, dl, VT, TrueVal, FalseVal);
3487 if (CC == ISD::SETOLT || CC == ISD::SETULT)
3488 return DAG.getNode(ARMISD::VMINNM, dl, VT, TrueVal, FalseVal);
3491 bool swpCmpOps = false;
3492 bool swpVselOps = false;
3493 checkVSELConstraints(CC, CondCode, swpCmpOps, swpVselOps);
3495 if (CondCode == ARMCC::GT || CondCode == ARMCC::GE ||
3496 CondCode == ARMCC::VS || CondCode == ARMCC::EQ) {
3498 std::swap(LHS, RHS);
3500 std::swap(TrueVal, FalseVal);
3504 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
3505 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
3506 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3507 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
3509 if (CondCode2 != ARMCC::AL) {
3510 SDValue ARMcc2 = DAG.getConstant(CondCode2, MVT::i32);
3511 // FIXME: Needs another CMP because flag can have but one use.
3512 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
3513 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
3514 Result, TrueVal, ARMcc2, CCR, Cmp2);
3519 /// canChangeToInt - Given the fp compare operand, return true if it is suitable
3520 /// to morph to an integer compare sequence.
3521 static bool canChangeToInt(SDValue Op, bool &SeenZero,
3522 const ARMSubtarget *Subtarget) {
3523 SDNode *N = Op.getNode();
3524 if (!N->hasOneUse())
3525 // Otherwise it requires moving the value from fp to integer registers.
3527 if (!N->getNumValues())
3529 EVT VT = Op.getValueType();
3530 if (VT != MVT::f32 && !Subtarget->isFPBrccSlow())
3531 // f32 case is generally profitable. f64 case only makes sense when vcmpe +
3532 // vmrs are very slow, e.g. cortex-a8.
3535 if (isFloatingPointZero(Op)) {
3539 return ISD::isNormalLoad(N);
3542 static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) {
3543 if (isFloatingPointZero(Op))
3544 return DAG.getConstant(0, MVT::i32);
3546 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op))
3547 return DAG.getLoad(MVT::i32, SDLoc(Op),
3548 Ld->getChain(), Ld->getBasePtr(), Ld->getPointerInfo(),
3549 Ld->isVolatile(), Ld->isNonTemporal(),
3550 Ld->isInvariant(), Ld->getAlignment());
3552 llvm_unreachable("Unknown VFP cmp argument!");
3555 static void expandf64Toi32(SDValue Op, SelectionDAG &DAG,
3556 SDValue &RetVal1, SDValue &RetVal2) {
3557 if (isFloatingPointZero(Op)) {
3558 RetVal1 = DAG.getConstant(0, MVT::i32);
3559 RetVal2 = DAG.getConstant(0, MVT::i32);
3563 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) {
3564 SDValue Ptr = Ld->getBasePtr();
3565 RetVal1 = DAG.getLoad(MVT::i32, SDLoc(Op),
3566 Ld->getChain(), Ptr,
3567 Ld->getPointerInfo(),
3568 Ld->isVolatile(), Ld->isNonTemporal(),
3569 Ld->isInvariant(), Ld->getAlignment());
3571 EVT PtrType = Ptr.getValueType();
3572 unsigned NewAlign = MinAlign(Ld->getAlignment(), 4);
3573 SDValue NewPtr = DAG.getNode(ISD::ADD, SDLoc(Op),
3574 PtrType, Ptr, DAG.getConstant(4, PtrType));
3575 RetVal2 = DAG.getLoad(MVT::i32, SDLoc(Op),
3576 Ld->getChain(), NewPtr,
3577 Ld->getPointerInfo().getWithOffset(4),
3578 Ld->isVolatile(), Ld->isNonTemporal(),
3579 Ld->isInvariant(), NewAlign);
3583 llvm_unreachable("Unknown VFP cmp argument!");
3586 /// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some
3587 /// f32 and even f64 comparisons to integer ones.
3589 ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const {
3590 SDValue Chain = Op.getOperand(0);
3591 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3592 SDValue LHS = Op.getOperand(2);
3593 SDValue RHS = Op.getOperand(3);
3594 SDValue Dest = Op.getOperand(4);
3597 bool LHSSeenZero = false;
3598 bool LHSOk = canChangeToInt(LHS, LHSSeenZero, Subtarget);
3599 bool RHSSeenZero = false;
3600 bool RHSOk = canChangeToInt(RHS, RHSSeenZero, Subtarget);
3601 if (LHSOk && RHSOk && (LHSSeenZero || RHSSeenZero)) {
3602 // If unsafe fp math optimization is enabled and there are no other uses of
3603 // the CMP operands, and the condition code is EQ or NE, we can optimize it
3604 // to an integer comparison.
3605 if (CC == ISD::SETOEQ)
3607 else if (CC == ISD::SETUNE)
3610 SDValue Mask = DAG.getConstant(0x7fffffff, MVT::i32);
3612 if (LHS.getValueType() == MVT::f32) {
3613 LHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3614 bitcastf32Toi32(LHS, DAG), Mask);
3615 RHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3616 bitcastf32Toi32(RHS, DAG), Mask);
3617 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3618 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3619 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3620 Chain, Dest, ARMcc, CCR, Cmp);
3625 expandf64Toi32(LHS, DAG, LHS1, LHS2);
3626 expandf64Toi32(RHS, DAG, RHS1, RHS2);
3627 LHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, LHS2, Mask);
3628 RHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, RHS2, Mask);
3629 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3630 ARMcc = DAG.getConstant(CondCode, MVT::i32);
3631 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3632 SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest };
3633 return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops);
3639 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
3640 SDValue Chain = Op.getOperand(0);
3641 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3642 SDValue LHS = Op.getOperand(2);
3643 SDValue RHS = Op.getOperand(3);
3644 SDValue Dest = Op.getOperand(4);
3647 if (LHS.getValueType() == MVT::i32) {
3649 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3650 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3651 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3652 Chain, Dest, ARMcc, CCR, Cmp);
3655 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3657 if (getTargetMachine().Options.UnsafeFPMath &&
3658 (CC == ISD::SETEQ || CC == ISD::SETOEQ ||
3659 CC == ISD::SETNE || CC == ISD::SETUNE)) {
3660 SDValue Result = OptimizeVFPBrcond(Op, DAG);
3661 if (Result.getNode())
3665 ARMCC::CondCodes CondCode, CondCode2;
3666 FPCCToARMCC(CC, CondCode, CondCode2);
3668 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
3669 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
3670 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3671 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3672 SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp };
3673 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops);
3674 if (CondCode2 != ARMCC::AL) {
3675 ARMcc = DAG.getConstant(CondCode2, MVT::i32);
3676 SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) };
3677 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops);
3682 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
3683 SDValue Chain = Op.getOperand(0);
3684 SDValue Table = Op.getOperand(1);
3685 SDValue Index = Op.getOperand(2);
3688 EVT PTy = getPointerTy();
3689 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
3690 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
3691 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
3692 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
3693 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
3694 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
3695 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
3696 if (Subtarget->isThumb2()) {
3697 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
3698 // which does another jump to the destination. This also makes it easier
3699 // to translate it to TBB / TBH later.
3700 // FIXME: This might not work if the function is extremely large.
3701 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
3702 Addr, Op.getOperand(2), JTI, UId);
3704 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
3705 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
3706 MachinePointerInfo::getJumpTable(),
3707 false, false, false, 0);
3708 Chain = Addr.getValue(1);
3709 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
3710 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3712 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
3713 MachinePointerInfo::getJumpTable(),
3714 false, false, false, 0);
3715 Chain = Addr.getValue(1);
3716 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3720 static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3721 EVT VT = Op.getValueType();
3724 if (Op.getValueType().getVectorElementType() == MVT::i32) {
3725 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::f32)
3727 return DAG.UnrollVectorOp(Op.getNode());
3730 assert(Op.getOperand(0).getValueType() == MVT::v4f32 &&
3731 "Invalid type for custom lowering!");
3732 if (VT != MVT::v4i16)
3733 return DAG.UnrollVectorOp(Op.getNode());
3735 Op = DAG.getNode(Op.getOpcode(), dl, MVT::v4i32, Op.getOperand(0));
3736 return DAG.getNode(ISD::TRUNCATE, dl, VT, Op);
3739 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3740 EVT VT = Op.getValueType();
3742 return LowerVectorFP_TO_INT(Op, DAG);
3747 switch (Op.getOpcode()) {
3748 default: llvm_unreachable("Invalid opcode!");
3749 case ISD::FP_TO_SINT:
3750 Opc = ARMISD::FTOSI;
3752 case ISD::FP_TO_UINT:
3753 Opc = ARMISD::FTOUI;
3756 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
3757 return DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
3760 static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3761 EVT VT = Op.getValueType();
3764 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::i32) {
3765 if (VT.getVectorElementType() == MVT::f32)
3767 return DAG.UnrollVectorOp(Op.getNode());
3770 assert(Op.getOperand(0).getValueType() == MVT::v4i16 &&
3771 "Invalid type for custom lowering!");
3772 if (VT != MVT::v4f32)
3773 return DAG.UnrollVectorOp(Op.getNode());
3777 switch (Op.getOpcode()) {
3778 default: llvm_unreachable("Invalid opcode!");
3779 case ISD::SINT_TO_FP:
3780 CastOpc = ISD::SIGN_EXTEND;
3781 Opc = ISD::SINT_TO_FP;
3783 case ISD::UINT_TO_FP:
3784 CastOpc = ISD::ZERO_EXTEND;
3785 Opc = ISD::UINT_TO_FP;
3789 Op = DAG.getNode(CastOpc, dl, MVT::v4i32, Op.getOperand(0));
3790 return DAG.getNode(Opc, dl, VT, Op);
3793 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3794 EVT VT = Op.getValueType();
3796 return LowerVectorINT_TO_FP(Op, DAG);
3801 switch (Op.getOpcode()) {
3802 default: llvm_unreachable("Invalid opcode!");
3803 case ISD::SINT_TO_FP:
3804 Opc = ARMISD::SITOF;
3806 case ISD::UINT_TO_FP:
3807 Opc = ARMISD::UITOF;
3811 Op = DAG.getNode(ISD::BITCAST, dl, MVT::f32, Op.getOperand(0));
3812 return DAG.getNode(Opc, dl, VT, Op);
3815 SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
3816 // Implement fcopysign with a fabs and a conditional fneg.
3817 SDValue Tmp0 = Op.getOperand(0);
3818 SDValue Tmp1 = Op.getOperand(1);
3820 EVT VT = Op.getValueType();
3821 EVT SrcVT = Tmp1.getValueType();
3822 bool InGPR = Tmp0.getOpcode() == ISD::BITCAST ||
3823 Tmp0.getOpcode() == ARMISD::VMOVDRR;
3824 bool UseNEON = !InGPR && Subtarget->hasNEON();
3827 // Use VBSL to copy the sign bit.
3828 unsigned EncodedVal = ARM_AM::createNEONModImm(0x6, 0x80);
3829 SDValue Mask = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v2i32,
3830 DAG.getTargetConstant(EncodedVal, MVT::i32));
3831 EVT OpVT = (VT == MVT::f32) ? MVT::v2i32 : MVT::v1i64;
3833 Mask = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3834 DAG.getNode(ISD::BITCAST, dl, OpVT, Mask),
3835 DAG.getConstant(32, MVT::i32));
3836 else /*if (VT == MVT::f32)*/
3837 Tmp0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp0);
3838 if (SrcVT == MVT::f32) {
3839 Tmp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp1);
3841 Tmp1 = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3842 DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1),
3843 DAG.getConstant(32, MVT::i32));
3844 } else if (VT == MVT::f32)
3845 Tmp1 = DAG.getNode(ARMISD::VSHRu, dl, MVT::v1i64,
3846 DAG.getNode(ISD::BITCAST, dl, MVT::v1i64, Tmp1),
3847 DAG.getConstant(32, MVT::i32));
3848 Tmp0 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp0);
3849 Tmp1 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1);
3851 SDValue AllOnes = DAG.getTargetConstant(ARM_AM::createNEONModImm(0xe, 0xff),
3853 AllOnes = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v8i8, AllOnes);
3854 SDValue MaskNot = DAG.getNode(ISD::XOR, dl, OpVT, Mask,
3855 DAG.getNode(ISD::BITCAST, dl, OpVT, AllOnes));
3857 SDValue Res = DAG.getNode(ISD::OR, dl, OpVT,
3858 DAG.getNode(ISD::AND, dl, OpVT, Tmp1, Mask),
3859 DAG.getNode(ISD::AND, dl, OpVT, Tmp0, MaskNot));
3860 if (VT == MVT::f32) {
3861 Res = DAG.getNode(ISD::BITCAST, dl, MVT::v2f32, Res);
3862 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, Res,
3863 DAG.getConstant(0, MVT::i32));
3865 Res = DAG.getNode(ISD::BITCAST, dl, MVT::f64, Res);
3871 // Bitcast operand 1 to i32.
3872 if (SrcVT == MVT::f64)
3873 Tmp1 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
3875 Tmp1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp1);
3877 // Or in the signbit with integer operations.
3878 SDValue Mask1 = DAG.getConstant(0x80000000, MVT::i32);
3879 SDValue Mask2 = DAG.getConstant(0x7fffffff, MVT::i32);
3880 Tmp1 = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp1, Mask1);
3881 if (VT == MVT::f32) {
3882 Tmp0 = DAG.getNode(ISD::AND, dl, MVT::i32,
3883 DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp0), Mask2);
3884 return DAG.getNode(ISD::BITCAST, dl, MVT::f32,
3885 DAG.getNode(ISD::OR, dl, MVT::i32, Tmp0, Tmp1));
3888 // f64: Or the high part with signbit and then combine two parts.
3889 Tmp0 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
3891 SDValue Lo = Tmp0.getValue(0);
3892 SDValue Hi = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp0.getValue(1), Mask2);
3893 Hi = DAG.getNode(ISD::OR, dl, MVT::i32, Hi, Tmp1);
3894 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
3897 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
3898 MachineFunction &MF = DAG.getMachineFunction();
3899 MachineFrameInfo *MFI = MF.getFrameInfo();
3900 MFI->setReturnAddressIsTaken(true);
3902 if (verifyReturnAddressArgumentIsConstant(Op, DAG))
3905 EVT VT = Op.getValueType();
3907 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3909 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
3910 SDValue Offset = DAG.getConstant(4, MVT::i32);
3911 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
3912 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
3913 MachinePointerInfo(), false, false, false, 0);
3916 // Return LR, which contains the return address. Mark it an implicit live-in.
3917 unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32));
3918 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
3921 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
3922 const ARMBaseRegisterInfo &ARI =
3923 *static_cast<const ARMBaseRegisterInfo*>(RegInfo);
3924 MachineFunction &MF = DAG.getMachineFunction();
3925 MachineFrameInfo *MFI = MF.getFrameInfo();
3926 MFI->setFrameAddressIsTaken(true);
3928 EVT VT = Op.getValueType();
3929 SDLoc dl(Op); // FIXME probably not meaningful
3930 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3931 unsigned FrameReg = ARI.getFrameRegister(MF);
3932 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
3934 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
3935 MachinePointerInfo(),
3936 false, false, false, 0);
3940 // FIXME? Maybe this could be a TableGen attribute on some registers and
3941 // this table could be generated automatically from RegInfo.
3942 unsigned ARMTargetLowering::getRegisterByName(const char* RegName,
3944 unsigned Reg = StringSwitch<unsigned>(RegName)
3945 .Case("sp", ARM::SP)
3949 report_fatal_error("Invalid register name global variable");
3952 /// ExpandBITCAST - If the target supports VFP, this function is called to
3953 /// expand a bit convert where either the source or destination type is i64 to
3954 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
3955 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
3956 /// vectors), since the legalizer won't know what to do with that.
3957 static SDValue ExpandBITCAST(SDNode *N, SelectionDAG &DAG) {
3958 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3960 SDValue Op = N->getOperand(0);
3962 // This function is only supposed to be called for i64 types, either as the
3963 // source or destination of the bit convert.
3964 EVT SrcVT = Op.getValueType();
3965 EVT DstVT = N->getValueType(0);
3966 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
3967 "ExpandBITCAST called for non-i64 type");
3969 // Turn i64->f64 into VMOVDRR.
3970 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
3971 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
3972 DAG.getConstant(0, MVT::i32));
3973 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
3974 DAG.getConstant(1, MVT::i32));
3975 return DAG.getNode(ISD::BITCAST, dl, DstVT,
3976 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
3979 // Turn f64->i64 into VMOVRRD.
3980 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
3982 if (TLI.isBigEndian() && SrcVT.isVector() &&
3983 SrcVT.getVectorNumElements() > 1)
3984 Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
3985 DAG.getVTList(MVT::i32, MVT::i32),
3986 DAG.getNode(ARMISD::VREV64, dl, SrcVT, Op));
3988 Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
3989 DAG.getVTList(MVT::i32, MVT::i32), Op);
3990 // Merge the pieces into a single i64 value.
3991 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
3997 /// getZeroVector - Returns a vector of specified type with all zero elements.
3998 /// Zero vectors are used to represent vector negation and in those cases
3999 /// will be implemented with the NEON VNEG instruction. However, VNEG does
4000 /// not support i64 elements, so sometimes the zero vectors will need to be
4001 /// explicitly constructed. Regardless, use a canonical VMOV to create the
4003 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, SDLoc dl) {
4004 assert(VT.isVector() && "Expected a vector type");
4005 // The canonical modified immediate encoding of a zero vector is....0!
4006 SDValue EncodedVal = DAG.getTargetConstant(0, MVT::i32);
4007 EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
4008 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal);
4009 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4012 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
4013 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
4014 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
4015 SelectionDAG &DAG) const {
4016 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4017 EVT VT = Op.getValueType();
4018 unsigned VTBits = VT.getSizeInBits();
4020 SDValue ShOpLo = Op.getOperand(0);
4021 SDValue ShOpHi = Op.getOperand(1);
4022 SDValue ShAmt = Op.getOperand(2);
4024 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
4026 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
4028 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
4029 DAG.getConstant(VTBits, MVT::i32), ShAmt);
4030 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
4031 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
4032 DAG.getConstant(VTBits, MVT::i32));
4033 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
4034 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
4035 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
4037 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
4038 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
4040 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
4041 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc,
4044 SDValue Ops[2] = { Lo, Hi };
4045 return DAG.getMergeValues(Ops, dl);
4048 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
4049 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
4050 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
4051 SelectionDAG &DAG) const {
4052 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4053 EVT VT = Op.getValueType();
4054 unsigned VTBits = VT.getSizeInBits();
4056 SDValue ShOpLo = Op.getOperand(0);
4057 SDValue ShOpHi = Op.getOperand(1);
4058 SDValue ShAmt = Op.getOperand(2);
4061 assert(Op.getOpcode() == ISD::SHL_PARTS);
4062 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
4063 DAG.getConstant(VTBits, MVT::i32), ShAmt);
4064 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
4065 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
4066 DAG.getConstant(VTBits, MVT::i32));
4067 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
4068 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
4070 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
4071 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
4072 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
4074 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
4075 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc,
4078 SDValue Ops[2] = { Lo, Hi };
4079 return DAG.getMergeValues(Ops, dl);
4082 SDValue ARMTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
4083 SelectionDAG &DAG) const {
4084 // The rounding mode is in bits 23:22 of the FPSCR.
4085 // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0
4086 // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3)
4087 // so that the shift + and get folded into a bitfield extract.
4089 SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::i32,
4090 DAG.getConstant(Intrinsic::arm_get_fpscr,
4092 SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPSCR,
4093 DAG.getConstant(1U << 22, MVT::i32));
4094 SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds,
4095 DAG.getConstant(22, MVT::i32));
4096 return DAG.getNode(ISD::AND, dl, MVT::i32, RMODE,
4097 DAG.getConstant(3, MVT::i32));
4100 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
4101 const ARMSubtarget *ST) {
4102 EVT VT = N->getValueType(0);
4105 if (!ST->hasV6T2Ops())
4108 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
4109 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
4112 /// getCTPOP16BitCounts - Returns a v8i8/v16i8 vector containing the bit-count
4113 /// for each 16-bit element from operand, repeated. The basic idea is to
4114 /// leverage vcnt to get the 8-bit counts, gather and add the results.
4116 /// Trace for v4i16:
4117 /// input = [v0 v1 v2 v3 ] (vi 16-bit element)
4118 /// cast: N0 = [w0 w1 w2 w3 w4 w5 w6 w7] (v0 = [w0 w1], wi 8-bit element)
4119 /// vcnt: N1 = [b0 b1 b2 b3 b4 b5 b6 b7] (bi = bit-count of 8-bit element wi)
4120 /// vrev: N2 = [b1 b0 b3 b2 b5 b4 b7 b6]
4121 /// [b0 b1 b2 b3 b4 b5 b6 b7]
4122 /// +[b1 b0 b3 b2 b5 b4 b7 b6]
4123 /// N3=N1+N2 = [k0 k0 k1 k1 k2 k2 k3 k3] (k0 = b0+b1 = bit-count of 16-bit v0,
4124 /// vuzp: = [k0 k1 k2 k3 k0 k1 k2 k3] each ki is 8-bits)
4125 static SDValue getCTPOP16BitCounts(SDNode *N, SelectionDAG &DAG) {
4126 EVT VT = N->getValueType(0);
4129 EVT VT8Bit = VT.is64BitVector() ? MVT::v8i8 : MVT::v16i8;
4130 SDValue N0 = DAG.getNode(ISD::BITCAST, DL, VT8Bit, N->getOperand(0));
4131 SDValue N1 = DAG.getNode(ISD::CTPOP, DL, VT8Bit, N0);
4132 SDValue N2 = DAG.getNode(ARMISD::VREV16, DL, VT8Bit, N1);
4133 SDValue N3 = DAG.getNode(ISD::ADD, DL, VT8Bit, N1, N2);
4134 return DAG.getNode(ARMISD::VUZP, DL, VT8Bit, N3, N3);
4137 /// lowerCTPOP16BitElements - Returns a v4i16/v8i16 vector containing the
4138 /// bit-count for each 16-bit element from the operand. We need slightly
4139 /// different sequencing for v4i16 and v8i16 to stay within NEON's available
4140 /// 64/128-bit registers.
4142 /// Trace for v4i16:
4143 /// input = [v0 v1 v2 v3 ] (vi 16-bit element)
4144 /// v8i8: BitCounts = [k0 k1 k2 k3 k0 k1 k2 k3 ] (ki is the bit-count of vi)
4145 /// v8i16:Extended = [k0 k1 k2 k3 k0 k1 k2 k3 ]
4146 /// v4i16:Extracted = [k0 k1 k2 k3 ]
4147 static SDValue lowerCTPOP16BitElements(SDNode *N, SelectionDAG &DAG) {
4148 EVT VT = N->getValueType(0);
4151 SDValue BitCounts = getCTPOP16BitCounts(N, DAG);
4152 if (VT.is64BitVector()) {
4153 SDValue Extended = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v8i16, BitCounts);
4154 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i16, Extended,
4155 DAG.getIntPtrConstant(0));
4157 SDValue Extracted = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v8i8,
4158 BitCounts, DAG.getIntPtrConstant(0));
4159 return DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v8i16, Extracted);
4163 /// lowerCTPOP32BitElements - Returns a v2i32/v4i32 vector containing the
4164 /// bit-count for each 32-bit element from the operand. The idea here is
4165 /// to split the vector into 16-bit elements, leverage the 16-bit count
4166 /// routine, and then combine the results.
4168 /// Trace for v2i32 (v4i32 similar with Extracted/Extended exchanged):
4169 /// input = [v0 v1 ] (vi: 32-bit elements)
4170 /// Bitcast = [w0 w1 w2 w3 ] (wi: 16-bit elements, v0 = [w0 w1])
4171 /// Counts16 = [k0 k1 k2 k3 ] (ki: 16-bit elements, bit-count of wi)
4172 /// vrev: N0 = [k1 k0 k3 k2 ]
4174 /// N1 =+[k1 k0 k3 k2 ]
4176 /// N2 =+[k1 k3 k0 k2 ]
4178 /// Extended =+[k1 k3 k0 k2 ]
4180 /// Extracted=+[k1 k3 ]
4182 static SDValue lowerCTPOP32BitElements(SDNode *N, SelectionDAG &DAG) {
4183 EVT VT = N->getValueType(0);
4186 EVT VT16Bit = VT.is64BitVector() ? MVT::v4i16 : MVT::v8i16;
4188 SDValue Bitcast = DAG.getNode(ISD::BITCAST, DL, VT16Bit, N->getOperand(0));
4189 SDValue Counts16 = lowerCTPOP16BitElements(Bitcast.getNode(), DAG);
4190 SDValue N0 = DAG.getNode(ARMISD::VREV32, DL, VT16Bit, Counts16);
4191 SDValue N1 = DAG.getNode(ISD::ADD, DL, VT16Bit, Counts16, N0);
4192 SDValue N2 = DAG.getNode(ARMISD::VUZP, DL, VT16Bit, N1, N1);
4194 if (VT.is64BitVector()) {
4195 SDValue Extended = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v4i32, N2);
4196 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i32, Extended,
4197 DAG.getIntPtrConstant(0));
4199 SDValue Extracted = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i16, N2,
4200 DAG.getIntPtrConstant(0));
4201 return DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v4i32, Extracted);
4205 static SDValue LowerCTPOP(SDNode *N, SelectionDAG &DAG,
4206 const ARMSubtarget *ST) {
4207 EVT VT = N->getValueType(0);
4209 assert(ST->hasNEON() && "Custom ctpop lowering requires NEON.");
4210 assert((VT == MVT::v2i32 || VT == MVT::v4i32 ||
4211 VT == MVT::v4i16 || VT == MVT::v8i16) &&
4212 "Unexpected type for custom ctpop lowering");
4214 if (VT.getVectorElementType() == MVT::i32)
4215 return lowerCTPOP32BitElements(N, DAG);
4217 return lowerCTPOP16BitElements(N, DAG);
4220 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
4221 const ARMSubtarget *ST) {
4222 EVT VT = N->getValueType(0);
4228 // Lower vector shifts on NEON to use VSHL.
4229 assert(ST->hasNEON() && "unexpected vector shift");
4231 // Left shifts translate directly to the vshiftu intrinsic.
4232 if (N->getOpcode() == ISD::SHL)
4233 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
4234 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
4235 N->getOperand(0), N->getOperand(1));
4237 assert((N->getOpcode() == ISD::SRA ||
4238 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
4240 // NEON uses the same intrinsics for both left and right shifts. For
4241 // right shifts, the shift amounts are negative, so negate the vector of
4243 EVT ShiftVT = N->getOperand(1).getValueType();
4244 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
4245 getZeroVector(ShiftVT, DAG, dl),
4247 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
4248 Intrinsic::arm_neon_vshifts :
4249 Intrinsic::arm_neon_vshiftu);
4250 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
4251 DAG.getConstant(vshiftInt, MVT::i32),
4252 N->getOperand(0), NegatedCount);
4255 static SDValue Expand64BitShift(SDNode *N, SelectionDAG &DAG,
4256 const ARMSubtarget *ST) {
4257 EVT VT = N->getValueType(0);
4260 // We can get here for a node like i32 = ISD::SHL i32, i64
4264 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
4265 "Unknown shift to lower!");
4267 // We only lower SRA, SRL of 1 here, all others use generic lowering.
4268 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
4269 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
4272 // If we are in thumb mode, we don't have RRX.
4273 if (ST->isThumb1Only()) return SDValue();
4275 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
4276 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
4277 DAG.getConstant(0, MVT::i32));
4278 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
4279 DAG.getConstant(1, MVT::i32));
4281 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
4282 // captures the result into a carry flag.
4283 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
4284 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Glue), Hi);
4286 // The low part is an ARMISD::RRX operand, which shifts the carry in.
4287 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
4289 // Merge the pieces into a single i64 value.
4290 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
4293 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
4294 SDValue TmpOp0, TmpOp1;
4295 bool Invert = false;
4299 SDValue Op0 = Op.getOperand(0);
4300 SDValue Op1 = Op.getOperand(1);
4301 SDValue CC = Op.getOperand(2);
4302 EVT VT = Op.getValueType();
4303 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
4306 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
4307 switch (SetCCOpcode) {
4308 default: llvm_unreachable("Illegal FP comparison");
4310 case ISD::SETNE: Invert = true; // Fallthrough
4312 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
4314 case ISD::SETLT: Swap = true; // Fallthrough
4316 case ISD::SETGT: Opc = ARMISD::VCGT; break;
4318 case ISD::SETLE: Swap = true; // Fallthrough
4320 case ISD::SETGE: Opc = ARMISD::VCGE; break;
4321 case ISD::SETUGE: Swap = true; // Fallthrough
4322 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
4323 case ISD::SETUGT: Swap = true; // Fallthrough
4324 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
4325 case ISD::SETUEQ: Invert = true; // Fallthrough
4327 // Expand this to (OLT | OGT).
4331 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
4332 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
4334 case ISD::SETUO: Invert = true; // Fallthrough
4336 // Expand this to (OLT | OGE).
4340 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
4341 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
4345 // Integer comparisons.
4346 switch (SetCCOpcode) {
4347 default: llvm_unreachable("Illegal integer comparison");
4348 case ISD::SETNE: Invert = true;
4349 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
4350 case ISD::SETLT: Swap = true;
4351 case ISD::SETGT: Opc = ARMISD::VCGT; break;
4352 case ISD::SETLE: Swap = true;
4353 case ISD::SETGE: Opc = ARMISD::VCGE; break;
4354 case ISD::SETULT: Swap = true;
4355 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
4356 case ISD::SETULE: Swap = true;
4357 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
4360 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
4361 if (Opc == ARMISD::VCEQ) {
4364 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
4366 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
4369 // Ignore bitconvert.
4370 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BITCAST)
4371 AndOp = AndOp.getOperand(0);
4373 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
4375 Op0 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(0));
4376 Op1 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(1));
4383 std::swap(Op0, Op1);
4385 // If one of the operands is a constant vector zero, attempt to fold the
4386 // comparison to a specialized compare-against-zero form.
4388 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
4390 else if (ISD::isBuildVectorAllZeros(Op0.getNode())) {
4391 if (Opc == ARMISD::VCGE)
4392 Opc = ARMISD::VCLEZ;
4393 else if (Opc == ARMISD::VCGT)
4394 Opc = ARMISD::VCLTZ;
4399 if (SingleOp.getNode()) {
4402 Result = DAG.getNode(ARMISD::VCEQZ, dl, VT, SingleOp); break;
4404 Result = DAG.getNode(ARMISD::VCGEZ, dl, VT, SingleOp); break;
4406 Result = DAG.getNode(ARMISD::VCLEZ, dl, VT, SingleOp); break;
4408 Result = DAG.getNode(ARMISD::VCGTZ, dl, VT, SingleOp); break;
4410 Result = DAG.getNode(ARMISD::VCLTZ, dl, VT, SingleOp); break;
4412 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
4415 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
4419 Result = DAG.getNOT(dl, Result, VT);
4424 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
4425 /// valid vector constant for a NEON instruction with a "modified immediate"
4426 /// operand (e.g., VMOV). If so, return the encoded value.
4427 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
4428 unsigned SplatBitSize, SelectionDAG &DAG,
4429 EVT &VT, bool is128Bits, NEONModImmType type) {
4430 unsigned OpCmode, Imm;
4432 // SplatBitSize is set to the smallest size that splats the vector, so a
4433 // zero vector will always have SplatBitSize == 8. However, NEON modified
4434 // immediate instructions others than VMOV do not support the 8-bit encoding
4435 // of a zero vector, and the default encoding of zero is supposed to be the
4440 switch (SplatBitSize) {
4442 if (type != VMOVModImm)
4444 // Any 1-byte value is OK. Op=0, Cmode=1110.
4445 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
4448 VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
4452 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
4453 VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
4454 if ((SplatBits & ~0xff) == 0) {
4455 // Value = 0x00nn: Op=x, Cmode=100x.
4460 if ((SplatBits & ~0xff00) == 0) {
4461 // Value = 0xnn00: Op=x, Cmode=101x.
4463 Imm = SplatBits >> 8;
4469 // NEON's 32-bit VMOV supports splat values where:
4470 // * only one byte is nonzero, or
4471 // * the least significant byte is 0xff and the second byte is nonzero, or
4472 // * the least significant 2 bytes are 0xff and the third is nonzero.
4473 VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
4474 if ((SplatBits & ~0xff) == 0) {
4475 // Value = 0x000000nn: Op=x, Cmode=000x.
4480 if ((SplatBits & ~0xff00) == 0) {
4481 // Value = 0x0000nn00: Op=x, Cmode=001x.
4483 Imm = SplatBits >> 8;
4486 if ((SplatBits & ~0xff0000) == 0) {
4487 // Value = 0x00nn0000: Op=x, Cmode=010x.
4489 Imm = SplatBits >> 16;
4492 if ((SplatBits & ~0xff000000) == 0) {
4493 // Value = 0xnn000000: Op=x, Cmode=011x.
4495 Imm = SplatBits >> 24;
4499 // cmode == 0b1100 and cmode == 0b1101 are not supported for VORR or VBIC
4500 if (type == OtherModImm) return SDValue();
4502 if ((SplatBits & ~0xffff) == 0 &&
4503 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
4504 // Value = 0x0000nnff: Op=x, Cmode=1100.
4506 Imm = SplatBits >> 8;
4510 if ((SplatBits & ~0xffffff) == 0 &&
4511 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
4512 // Value = 0x00nnffff: Op=x, Cmode=1101.
4514 Imm = SplatBits >> 16;
4518 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
4519 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
4520 // VMOV.I32. A (very) minor optimization would be to replicate the value
4521 // and fall through here to test for a valid 64-bit splat. But, then the
4522 // caller would also need to check and handle the change in size.
4526 if (type != VMOVModImm)
4528 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
4529 uint64_t BitMask = 0xff;
4531 unsigned ImmMask = 1;
4533 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
4534 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
4537 } else if ((SplatBits & BitMask) != 0) {
4544 if (DAG.getTargetLoweringInfo().isBigEndian())
4545 // swap higher and lower 32 bit word
4546 Imm = ((Imm & 0xf) << 4) | ((Imm & 0xf0) >> 4);
4548 // Op=1, Cmode=1110.
4550 VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
4555 llvm_unreachable("unexpected size for isNEONModifiedImm");
4558 unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm);
4559 return DAG.getTargetConstant(EncodedVal, MVT::i32);
4562 SDValue ARMTargetLowering::LowerConstantFP(SDValue Op, SelectionDAG &DAG,
4563 const ARMSubtarget *ST) const {
4567 bool IsDouble = Op.getValueType() == MVT::f64;
4568 ConstantFPSDNode *CFP = cast<ConstantFPSDNode>(Op);
4570 // Try splatting with a VMOV.f32...
4571 APFloat FPVal = CFP->getValueAPF();
4572 int ImmVal = IsDouble ? ARM_AM::getFP64Imm(FPVal) : ARM_AM::getFP32Imm(FPVal);
4575 if (IsDouble || !ST->useNEONForSinglePrecisionFP()) {
4576 // We have code in place to select a valid ConstantFP already, no need to
4581 // It's a float and we are trying to use NEON operations where
4582 // possible. Lower it to a splat followed by an extract.
4584 SDValue NewVal = DAG.getTargetConstant(ImmVal, MVT::i32);
4585 SDValue VecConstant = DAG.getNode(ARMISD::VMOVFPIMM, DL, MVT::v2f32,
4587 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecConstant,
4588 DAG.getConstant(0, MVT::i32));
4591 // The rest of our options are NEON only, make sure that's allowed before
4593 if (!ST->hasNEON() || (!IsDouble && !ST->useNEONForSinglePrecisionFP()))
4597 uint64_t iVal = FPVal.bitcastToAPInt().getZExtValue();
4599 // It wouldn't really be worth bothering for doubles except for one very
4600 // important value, which does happen to match: 0.0. So make sure we don't do
4602 if (IsDouble && (iVal & 0xffffffff) != (iVal >> 32))
4605 // Try a VMOV.i32 (FIXME: i8, i16, or i64 could work too).
4606 SDValue NewVal = isNEONModifiedImm(iVal & 0xffffffffU, 0, 32, DAG, VMovVT,
4608 if (NewVal != SDValue()) {
4610 SDValue VecConstant = DAG.getNode(ARMISD::VMOVIMM, DL, VMovVT,
4613 return DAG.getNode(ISD::BITCAST, DL, MVT::f64, VecConstant);
4615 // It's a float: cast and extract a vector element.
4616 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
4618 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
4619 DAG.getConstant(0, MVT::i32));
4622 // Finally, try a VMVN.i32
4623 NewVal = isNEONModifiedImm(~iVal & 0xffffffffU, 0, 32, DAG, VMovVT,
4625 if (NewVal != SDValue()) {
4627 SDValue VecConstant = DAG.getNode(ARMISD::VMVNIMM, DL, VMovVT, NewVal);
4630 return DAG.getNode(ISD::BITCAST, DL, MVT::f64, VecConstant);
4632 // It's a float: cast and extract a vector element.
4633 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
4635 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
4636 DAG.getConstant(0, MVT::i32));
4642 // check if an VEXT instruction can handle the shuffle mask when the
4643 // vector sources of the shuffle are the same.
4644 static bool isSingletonVEXTMask(ArrayRef<int> M, EVT VT, unsigned &Imm) {
4645 unsigned NumElts = VT.getVectorNumElements();
4647 // Assume that the first shuffle index is not UNDEF. Fail if it is.
4653 // If this is a VEXT shuffle, the immediate value is the index of the first
4654 // element. The other shuffle indices must be the successive elements after
4656 unsigned ExpectedElt = Imm;
4657 for (unsigned i = 1; i < NumElts; ++i) {
4658 // Increment the expected index. If it wraps around, just follow it
4659 // back to index zero and keep going.
4661 if (ExpectedElt == NumElts)
4664 if (M[i] < 0) continue; // ignore UNDEF indices
4665 if (ExpectedElt != static_cast<unsigned>(M[i]))
4673 static bool isVEXTMask(ArrayRef<int> M, EVT VT,
4674 bool &ReverseVEXT, unsigned &Imm) {
4675 unsigned NumElts = VT.getVectorNumElements();
4676 ReverseVEXT = false;
4678 // Assume that the first shuffle index is not UNDEF. Fail if it is.
4684 // If this is a VEXT shuffle, the immediate value is the index of the first
4685 // element. The other shuffle indices must be the successive elements after
4687 unsigned ExpectedElt = Imm;
4688 for (unsigned i = 1; i < NumElts; ++i) {
4689 // Increment the expected index. If it wraps around, it may still be
4690 // a VEXT but the source vectors must be swapped.
4692 if (ExpectedElt == NumElts * 2) {
4697 if (M[i] < 0) continue; // ignore UNDEF indices
4698 if (ExpectedElt != static_cast<unsigned>(M[i]))
4702 // Adjust the index value if the source operands will be swapped.
4709 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
4710 /// instruction with the specified blocksize. (The order of the elements
4711 /// within each block of the vector is reversed.)
4712 static bool isVREVMask(ArrayRef<int> M, EVT VT, unsigned BlockSize) {
4713 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
4714 "Only possible block sizes for VREV are: 16, 32, 64");
4716 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4720 unsigned NumElts = VT.getVectorNumElements();
4721 unsigned BlockElts = M[0] + 1;
4722 // If the first shuffle index is UNDEF, be optimistic.
4724 BlockElts = BlockSize / EltSz;
4726 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
4729 for (unsigned i = 0; i < NumElts; ++i) {
4730 if (M[i] < 0) continue; // ignore UNDEF indices
4731 if ((unsigned) M[i] != (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
4738 static bool isVTBLMask(ArrayRef<int> M, EVT VT) {
4739 // We can handle <8 x i8> vector shuffles. If the index in the mask is out of
4740 // range, then 0 is placed into the resulting vector. So pretty much any mask
4741 // of 8 elements can work here.
4742 return VT == MVT::v8i8 && M.size() == 8;
4745 static bool isVTRNMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4746 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4750 unsigned NumElts = VT.getVectorNumElements();
4751 WhichResult = (M[0] == 0 ? 0 : 1);
4752 for (unsigned i = 0; i < NumElts; i += 2) {
4753 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
4754 (M[i+1] >= 0 && (unsigned) M[i+1] != i + NumElts + WhichResult))
4760 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
4761 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4762 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
4763 static bool isVTRN_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4764 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4768 unsigned NumElts = VT.getVectorNumElements();
4769 WhichResult = (M[0] == 0 ? 0 : 1);
4770 for (unsigned i = 0; i < NumElts; i += 2) {
4771 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
4772 (M[i+1] >= 0 && (unsigned) M[i+1] != i + WhichResult))
4778 static bool isVUZPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4779 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4783 unsigned NumElts = VT.getVectorNumElements();
4784 WhichResult = (M[0] == 0 ? 0 : 1);
4785 for (unsigned i = 0; i != NumElts; ++i) {
4786 if (M[i] < 0) continue; // ignore UNDEF indices
4787 if ((unsigned) M[i] != 2 * i + WhichResult)
4791 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4792 if (VT.is64BitVector() && EltSz == 32)
4798 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
4799 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4800 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
4801 static bool isVUZP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4802 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4806 unsigned Half = VT.getVectorNumElements() / 2;
4807 WhichResult = (M[0] == 0 ? 0 : 1);
4808 for (unsigned j = 0; j != 2; ++j) {
4809 unsigned Idx = WhichResult;
4810 for (unsigned i = 0; i != Half; ++i) {
4811 int MIdx = M[i + j * Half];
4812 if (MIdx >= 0 && (unsigned) MIdx != Idx)
4818 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4819 if (VT.is64BitVector() && EltSz == 32)
4825 static bool isVZIPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4826 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4830 unsigned NumElts = VT.getVectorNumElements();
4831 WhichResult = (M[0] == 0 ? 0 : 1);
4832 unsigned Idx = WhichResult * NumElts / 2;
4833 for (unsigned i = 0; i != NumElts; i += 2) {
4834 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
4835 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx + NumElts))
4840 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4841 if (VT.is64BitVector() && EltSz == 32)
4847 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
4848 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4849 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
4850 static bool isVZIP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4851 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4855 unsigned NumElts = VT.getVectorNumElements();
4856 WhichResult = (M[0] == 0 ? 0 : 1);
4857 unsigned Idx = WhichResult * NumElts / 2;
4858 for (unsigned i = 0; i != NumElts; i += 2) {
4859 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
4860 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx))
4865 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4866 if (VT.is64BitVector() && EltSz == 32)
4872 /// \return true if this is a reverse operation on an vector.
4873 static bool isReverseMask(ArrayRef<int> M, EVT VT) {
4874 unsigned NumElts = VT.getVectorNumElements();
4875 // Make sure the mask has the right size.
4876 if (NumElts != M.size())
4879 // Look for <15, ..., 3, -1, 1, 0>.
4880 for (unsigned i = 0; i != NumElts; ++i)
4881 if (M[i] >= 0 && M[i] != (int) (NumElts - 1 - i))
4887 // If N is an integer constant that can be moved into a register in one
4888 // instruction, return an SDValue of such a constant (will become a MOV
4889 // instruction). Otherwise return null.
4890 static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG,
4891 const ARMSubtarget *ST, SDLoc dl) {
4893 if (!isa<ConstantSDNode>(N))
4895 Val = cast<ConstantSDNode>(N)->getZExtValue();
4897 if (ST->isThumb1Only()) {
4898 if (Val <= 255 || ~Val <= 255)
4899 return DAG.getConstant(Val, MVT::i32);
4901 if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1)
4902 return DAG.getConstant(Val, MVT::i32);
4907 // If this is a case we can't handle, return null and let the default
4908 // expansion code take care of it.
4909 SDValue ARMTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
4910 const ARMSubtarget *ST) const {
4911 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
4913 EVT VT = Op.getValueType();
4915 APInt SplatBits, SplatUndef;
4916 unsigned SplatBitSize;
4918 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
4919 if (SplatBitSize <= 64) {
4920 // Check if an immediate VMOV works.
4922 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
4923 SplatUndef.getZExtValue(), SplatBitSize,
4924 DAG, VmovVT, VT.is128BitVector(),
4926 if (Val.getNode()) {
4927 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
4928 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4931 // Try an immediate VMVN.
4932 uint64_t NegatedImm = (~SplatBits).getZExtValue();
4933 Val = isNEONModifiedImm(NegatedImm,
4934 SplatUndef.getZExtValue(), SplatBitSize,
4935 DAG, VmovVT, VT.is128BitVector(),
4937 if (Val.getNode()) {
4938 SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
4939 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4942 // Use vmov.f32 to materialize other v2f32 and v4f32 splats.
4943 if ((VT == MVT::v2f32 || VT == MVT::v4f32) && SplatBitSize == 32) {
4944 int ImmVal = ARM_AM::getFP32Imm(SplatBits);
4946 SDValue Val = DAG.getTargetConstant(ImmVal, MVT::i32);
4947 return DAG.getNode(ARMISD::VMOVFPIMM, dl, VT, Val);
4953 // Scan through the operands to see if only one value is used.
4955 // As an optimisation, even if more than one value is used it may be more
4956 // profitable to splat with one value then change some lanes.
4958 // Heuristically we decide to do this if the vector has a "dominant" value,
4959 // defined as splatted to more than half of the lanes.
4960 unsigned NumElts = VT.getVectorNumElements();
4961 bool isOnlyLowElement = true;
4962 bool usesOnlyOneValue = true;
4963 bool hasDominantValue = false;
4964 bool isConstant = true;
4966 // Map of the number of times a particular SDValue appears in the
4968 DenseMap<SDValue, unsigned> ValueCounts;
4970 for (unsigned i = 0; i < NumElts; ++i) {
4971 SDValue V = Op.getOperand(i);
4972 if (V.getOpcode() == ISD::UNDEF)
4975 isOnlyLowElement = false;
4976 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
4979 ValueCounts.insert(std::make_pair(V, 0));
4980 unsigned &Count = ValueCounts[V];
4982 // Is this value dominant? (takes up more than half of the lanes)
4983 if (++Count > (NumElts / 2)) {
4984 hasDominantValue = true;
4988 if (ValueCounts.size() != 1)
4989 usesOnlyOneValue = false;
4990 if (!Value.getNode() && ValueCounts.size() > 0)
4991 Value = ValueCounts.begin()->first;
4993 if (ValueCounts.size() == 0)
4994 return DAG.getUNDEF(VT);
4996 // Loads are better lowered with insert_vector_elt/ARMISD::BUILD_VECTOR.
4997 // Keep going if we are hitting this case.
4998 if (isOnlyLowElement && !ISD::isNormalLoad(Value.getNode()))
4999 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
5001 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5003 // Use VDUP for non-constant splats. For f32 constant splats, reduce to
5004 // i32 and try again.
5005 if (hasDominantValue && EltSize <= 32) {
5009 // If we are VDUPing a value that comes directly from a vector, that will
5010 // cause an unnecessary move to and from a GPR, where instead we could
5011 // just use VDUPLANE. We can only do this if the lane being extracted
5012 // is at a constant index, as the VDUP from lane instructions only have
5013 // constant-index forms.
5014 if (Value->getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5015 isa<ConstantSDNode>(Value->getOperand(1))) {
5016 // We need to create a new undef vector to use for the VDUPLANE if the
5017 // size of the vector from which we get the value is different than the
5018 // size of the vector that we need to create. We will insert the element
5019 // such that the register coalescer will remove unnecessary copies.
5020 if (VT != Value->getOperand(0).getValueType()) {
5021 ConstantSDNode *constIndex;
5022 constIndex = dyn_cast<ConstantSDNode>(Value->getOperand(1));
5023 assert(constIndex && "The index is not a constant!");
5024 unsigned index = constIndex->getAPIntValue().getLimitedValue() %
5025 VT.getVectorNumElements();
5026 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
5027 DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DAG.getUNDEF(VT),
5028 Value, DAG.getConstant(index, MVT::i32)),
5029 DAG.getConstant(index, MVT::i32));
5031 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
5032 Value->getOperand(0), Value->getOperand(1));
5034 N = DAG.getNode(ARMISD::VDUP, dl, VT, Value);
5036 if (!usesOnlyOneValue) {
5037 // The dominant value was splatted as 'N', but we now have to insert
5038 // all differing elements.
5039 for (unsigned I = 0; I < NumElts; ++I) {
5040 if (Op.getOperand(I) == Value)
5042 SmallVector<SDValue, 3> Ops;
5044 Ops.push_back(Op.getOperand(I));
5045 Ops.push_back(DAG.getConstant(I, MVT::i32));
5046 N = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Ops);
5051 if (VT.getVectorElementType().isFloatingPoint()) {
5052 SmallVector<SDValue, 8> Ops;
5053 for (unsigned i = 0; i < NumElts; ++i)
5054 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, MVT::i32,
5056 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
5057 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, Ops);
5058 Val = LowerBUILD_VECTOR(Val, DAG, ST);
5060 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
5062 if (usesOnlyOneValue) {
5063 SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
5064 if (isConstant && Val.getNode())
5065 return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
5069 // If all elements are constants and the case above didn't get hit, fall back
5070 // to the default expansion, which will generate a load from the constant
5075 // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
5077 SDValue shuffle = ReconstructShuffle(Op, DAG);
5078 if (shuffle != SDValue())
5082 // Vectors with 32- or 64-bit elements can be built by directly assigning
5083 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
5084 // will be legalized.
5085 if (EltSize >= 32) {
5086 // Do the expansion with floating-point types, since that is what the VFP
5087 // registers are defined to use, and since i64 is not legal.
5088 EVT EltVT = EVT::getFloatingPointVT(EltSize);
5089 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
5090 SmallVector<SDValue, 8> Ops;
5091 for (unsigned i = 0; i < NumElts; ++i)
5092 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, EltVT, Op.getOperand(i)));
5093 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, Ops);
5094 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
5097 // If all else fails, just use a sequence of INSERT_VECTOR_ELT when we
5098 // know the default expansion would otherwise fall back on something even
5099 // worse. For a vector with one or two non-undef values, that's
5100 // scalar_to_vector for the elements followed by a shuffle (provided the
5101 // shuffle is valid for the target) and materialization element by element
5102 // on the stack followed by a load for everything else.
5103 if (!isConstant && !usesOnlyOneValue) {
5104 SDValue Vec = DAG.getUNDEF(VT);
5105 for (unsigned i = 0 ; i < NumElts; ++i) {
5106 SDValue V = Op.getOperand(i);
5107 if (V.getOpcode() == ISD::UNDEF)
5109 SDValue LaneIdx = DAG.getConstant(i, MVT::i32);
5110 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Vec, V, LaneIdx);
5118 // Gather data to see if the operation can be modelled as a
5119 // shuffle in combination with VEXTs.
5120 SDValue ARMTargetLowering::ReconstructShuffle(SDValue Op,
5121 SelectionDAG &DAG) const {
5123 EVT VT = Op.getValueType();
5124 unsigned NumElts = VT.getVectorNumElements();
5126 SmallVector<SDValue, 2> SourceVecs;
5127 SmallVector<unsigned, 2> MinElts;
5128 SmallVector<unsigned, 2> MaxElts;
5130 for (unsigned i = 0; i < NumElts; ++i) {
5131 SDValue V = Op.getOperand(i);
5132 if (V.getOpcode() == ISD::UNDEF)
5134 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
5135 // A shuffle can only come from building a vector from various
5136 // elements of other vectors.
5138 } else if (V.getOperand(0).getValueType().getVectorElementType() !=
5139 VT.getVectorElementType()) {
5140 // This code doesn't know how to handle shuffles where the vector
5141 // element types do not match (this happens because type legalization
5142 // promotes the return type of EXTRACT_VECTOR_ELT).
5143 // FIXME: It might be appropriate to extend this code to handle
5144 // mismatched types.
5148 // Record this extraction against the appropriate vector if possible...
5149 SDValue SourceVec = V.getOperand(0);
5150 // If the element number isn't a constant, we can't effectively
5151 // analyze what's going on.
5152 if (!isa<ConstantSDNode>(V.getOperand(1)))
5154 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
5155 bool FoundSource = false;
5156 for (unsigned j = 0; j < SourceVecs.size(); ++j) {
5157 if (SourceVecs[j] == SourceVec) {
5158 if (MinElts[j] > EltNo)
5160 if (MaxElts[j] < EltNo)
5167 // Or record a new source if not...
5169 SourceVecs.push_back(SourceVec);
5170 MinElts.push_back(EltNo);
5171 MaxElts.push_back(EltNo);
5175 // Currently only do something sane when at most two source vectors
5177 if (SourceVecs.size() > 2)
5180 SDValue ShuffleSrcs[2] = {DAG.getUNDEF(VT), DAG.getUNDEF(VT) };
5181 int VEXTOffsets[2] = {0, 0};
5183 // This loop extracts the usage patterns of the source vectors
5184 // and prepares appropriate SDValues for a shuffle if possible.
5185 for (unsigned i = 0; i < SourceVecs.size(); ++i) {
5186 if (SourceVecs[i].getValueType() == VT) {
5187 // No VEXT necessary
5188 ShuffleSrcs[i] = SourceVecs[i];
5191 } else if (SourceVecs[i].getValueType().getVectorNumElements() < NumElts) {
5192 // It probably isn't worth padding out a smaller vector just to
5193 // break it down again in a shuffle.
5197 // Since only 64-bit and 128-bit vectors are legal on ARM and
5198 // we've eliminated the other cases...
5199 assert(SourceVecs[i].getValueType().getVectorNumElements() == 2*NumElts &&
5200 "unexpected vector sizes in ReconstructShuffle");
5202 if (MaxElts[i] - MinElts[i] >= NumElts) {
5203 // Span too large for a VEXT to cope
5207 if (MinElts[i] >= NumElts) {
5208 // The extraction can just take the second half
5209 VEXTOffsets[i] = NumElts;
5210 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5212 DAG.getIntPtrConstant(NumElts));
5213 } else if (MaxElts[i] < NumElts) {
5214 // The extraction can just take the first half
5216 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5218 DAG.getIntPtrConstant(0));
5220 // An actual VEXT is needed
5221 VEXTOffsets[i] = MinElts[i];
5222 SDValue VEXTSrc1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5224 DAG.getIntPtrConstant(0));
5225 SDValue VEXTSrc2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5227 DAG.getIntPtrConstant(NumElts));
5228 ShuffleSrcs[i] = DAG.getNode(ARMISD::VEXT, dl, VT, VEXTSrc1, VEXTSrc2,
5229 DAG.getConstant(VEXTOffsets[i], MVT::i32));
5233 SmallVector<int, 8> Mask;
5235 for (unsigned i = 0; i < NumElts; ++i) {
5236 SDValue Entry = Op.getOperand(i);
5237 if (Entry.getOpcode() == ISD::UNDEF) {
5242 SDValue ExtractVec = Entry.getOperand(0);
5243 int ExtractElt = cast<ConstantSDNode>(Op.getOperand(i)
5244 .getOperand(1))->getSExtValue();
5245 if (ExtractVec == SourceVecs[0]) {
5246 Mask.push_back(ExtractElt - VEXTOffsets[0]);
5248 Mask.push_back(ExtractElt + NumElts - VEXTOffsets[1]);
5252 // Final check before we try to produce nonsense...
5253 if (isShuffleMaskLegal(Mask, VT))
5254 return DAG.getVectorShuffle(VT, dl, ShuffleSrcs[0], ShuffleSrcs[1],
5260 /// isShuffleMaskLegal - Targets can use this to indicate that they only
5261 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
5262 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
5263 /// are assumed to be legal.
5265 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
5267 if (VT.getVectorNumElements() == 4 &&
5268 (VT.is128BitVector() || VT.is64BitVector())) {
5269 unsigned PFIndexes[4];
5270 for (unsigned i = 0; i != 4; ++i) {
5274 PFIndexes[i] = M[i];
5277 // Compute the index in the perfect shuffle table.
5278 unsigned PFTableIndex =
5279 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
5280 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
5281 unsigned Cost = (PFEntry >> 30);
5288 unsigned Imm, WhichResult;
5290 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5291 return (EltSize >= 32 ||
5292 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
5293 isVREVMask(M, VT, 64) ||
5294 isVREVMask(M, VT, 32) ||
5295 isVREVMask(M, VT, 16) ||
5296 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
5297 isVTBLMask(M, VT) ||
5298 isVTRNMask(M, VT, WhichResult) ||
5299 isVUZPMask(M, VT, WhichResult) ||
5300 isVZIPMask(M, VT, WhichResult) ||
5301 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
5302 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
5303 isVZIP_v_undef_Mask(M, VT, WhichResult) ||
5304 ((VT == MVT::v8i16 || VT == MVT::v16i8) && isReverseMask(M, VT)));
5307 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
5308 /// the specified operations to build the shuffle.
5309 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
5310 SDValue RHS, SelectionDAG &DAG,
5312 unsigned OpNum = (PFEntry >> 26) & 0x0F;
5313 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
5314 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
5317 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
5326 OP_VUZPL, // VUZP, left result
5327 OP_VUZPR, // VUZP, right result
5328 OP_VZIPL, // VZIP, left result
5329 OP_VZIPR, // VZIP, right result
5330 OP_VTRNL, // VTRN, left result
5331 OP_VTRNR // VTRN, right result
5334 if (OpNum == OP_COPY) {
5335 if (LHSID == (1*9+2)*9+3) return LHS;
5336 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
5340 SDValue OpLHS, OpRHS;
5341 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
5342 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
5343 EVT VT = OpLHS.getValueType();
5346 default: llvm_unreachable("Unknown shuffle opcode!");
5348 // VREV divides the vector in half and swaps within the half.
5349 if (VT.getVectorElementType() == MVT::i32 ||
5350 VT.getVectorElementType() == MVT::f32)
5351 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
5352 // vrev <4 x i16> -> VREV32
5353 if (VT.getVectorElementType() == MVT::i16)
5354 return DAG.getNode(ARMISD::VREV32, dl, VT, OpLHS);
5355 // vrev <4 x i8> -> VREV16
5356 assert(VT.getVectorElementType() == MVT::i8);
5357 return DAG.getNode(ARMISD::VREV16, dl, VT, OpLHS);
5362 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
5363 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
5367 return DAG.getNode(ARMISD::VEXT, dl, VT,
5369 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
5372 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
5373 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
5376 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
5377 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
5380 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
5381 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
5385 static SDValue LowerVECTOR_SHUFFLEv8i8(SDValue Op,
5386 ArrayRef<int> ShuffleMask,
5387 SelectionDAG &DAG) {
5388 // Check to see if we can use the VTBL instruction.
5389 SDValue V1 = Op.getOperand(0);
5390 SDValue V2 = Op.getOperand(1);
5393 SmallVector<SDValue, 8> VTBLMask;
5394 for (ArrayRef<int>::iterator
5395 I = ShuffleMask.begin(), E = ShuffleMask.end(); I != E; ++I)
5396 VTBLMask.push_back(DAG.getConstant(*I, MVT::i32));
5398 if (V2.getNode()->getOpcode() == ISD::UNDEF)
5399 return DAG.getNode(ARMISD::VTBL1, DL, MVT::v8i8, V1,
5400 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8, VTBLMask));
5402 return DAG.getNode(ARMISD::VTBL2, DL, MVT::v8i8, V1, V2,
5403 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8, VTBLMask));
5406 static SDValue LowerReverse_VECTOR_SHUFFLEv16i8_v8i16(SDValue Op,
5407 SelectionDAG &DAG) {
5409 SDValue OpLHS = Op.getOperand(0);
5410 EVT VT = OpLHS.getValueType();
5412 assert((VT == MVT::v8i16 || VT == MVT::v16i8) &&
5413 "Expect an v8i16/v16i8 type");
5414 OpLHS = DAG.getNode(ARMISD::VREV64, DL, VT, OpLHS);
5415 // For a v16i8 type: After the VREV, we have got <8, ...15, 8, ..., 0>. Now,
5416 // extract the first 8 bytes into the top double word and the last 8 bytes
5417 // into the bottom double word. The v8i16 case is similar.
5418 unsigned ExtractNum = (VT == MVT::v16i8) ? 8 : 4;
5419 return DAG.getNode(ARMISD::VEXT, DL, VT, OpLHS, OpLHS,
5420 DAG.getConstant(ExtractNum, MVT::i32));
5423 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
5424 SDValue V1 = Op.getOperand(0);
5425 SDValue V2 = Op.getOperand(1);
5427 EVT VT = Op.getValueType();
5428 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
5430 // Convert shuffles that are directly supported on NEON to target-specific
5431 // DAG nodes, instead of keeping them as shuffles and matching them again
5432 // during code selection. This is more efficient and avoids the possibility
5433 // of inconsistencies between legalization and selection.
5434 // FIXME: floating-point vectors should be canonicalized to integer vectors
5435 // of the same time so that they get CSEd properly.
5436 ArrayRef<int> ShuffleMask = SVN->getMask();
5438 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5439 if (EltSize <= 32) {
5440 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
5441 int Lane = SVN->getSplatIndex();
5442 // If this is undef splat, generate it via "just" vdup, if possible.
5443 if (Lane == -1) Lane = 0;
5445 // Test if V1 is a SCALAR_TO_VECTOR.
5446 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
5447 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
5449 // Test if V1 is a BUILD_VECTOR which is equivalent to a SCALAR_TO_VECTOR
5450 // (and probably will turn into a SCALAR_TO_VECTOR once legalization
5452 if (Lane == 0 && V1.getOpcode() == ISD::BUILD_VECTOR &&
5453 !isa<ConstantSDNode>(V1.getOperand(0))) {
5454 bool IsScalarToVector = true;
5455 for (unsigned i = 1, e = V1.getNumOperands(); i != e; ++i)
5456 if (V1.getOperand(i).getOpcode() != ISD::UNDEF) {
5457 IsScalarToVector = false;
5460 if (IsScalarToVector)
5461 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
5463 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
5464 DAG.getConstant(Lane, MVT::i32));
5469 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
5472 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
5473 DAG.getConstant(Imm, MVT::i32));
5476 if (isVREVMask(ShuffleMask, VT, 64))
5477 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
5478 if (isVREVMask(ShuffleMask, VT, 32))
5479 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
5480 if (isVREVMask(ShuffleMask, VT, 16))
5481 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
5483 if (V2->getOpcode() == ISD::UNDEF &&
5484 isSingletonVEXTMask(ShuffleMask, VT, Imm)) {
5485 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V1,
5486 DAG.getConstant(Imm, MVT::i32));
5489 // Check for Neon shuffles that modify both input vectors in place.
5490 // If both results are used, i.e., if there are two shuffles with the same
5491 // source operands and with masks corresponding to both results of one of
5492 // these operations, DAG memoization will ensure that a single node is
5493 // used for both shuffles.
5494 unsigned WhichResult;
5495 if (isVTRNMask(ShuffleMask, VT, WhichResult))
5496 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
5497 V1, V2).getValue(WhichResult);
5498 if (isVUZPMask(ShuffleMask, VT, WhichResult))
5499 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
5500 V1, V2).getValue(WhichResult);
5501 if (isVZIPMask(ShuffleMask, VT, WhichResult))
5502 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
5503 V1, V2).getValue(WhichResult);
5505 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
5506 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
5507 V1, V1).getValue(WhichResult);
5508 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
5509 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
5510 V1, V1).getValue(WhichResult);
5511 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
5512 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
5513 V1, V1).getValue(WhichResult);
5516 // If the shuffle is not directly supported and it has 4 elements, use
5517 // the PerfectShuffle-generated table to synthesize it from other shuffles.
5518 unsigned NumElts = VT.getVectorNumElements();
5520 unsigned PFIndexes[4];
5521 for (unsigned i = 0; i != 4; ++i) {
5522 if (ShuffleMask[i] < 0)
5525 PFIndexes[i] = ShuffleMask[i];
5528 // Compute the index in the perfect shuffle table.
5529 unsigned PFTableIndex =
5530 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
5531 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
5532 unsigned Cost = (PFEntry >> 30);
5535 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
5538 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
5539 if (EltSize >= 32) {
5540 // Do the expansion with floating-point types, since that is what the VFP
5541 // registers are defined to use, and since i64 is not legal.
5542 EVT EltVT = EVT::getFloatingPointVT(EltSize);
5543 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
5544 V1 = DAG.getNode(ISD::BITCAST, dl, VecVT, V1);
5545 V2 = DAG.getNode(ISD::BITCAST, dl, VecVT, V2);
5546 SmallVector<SDValue, 8> Ops;
5547 for (unsigned i = 0; i < NumElts; ++i) {
5548 if (ShuffleMask[i] < 0)
5549 Ops.push_back(DAG.getUNDEF(EltVT));
5551 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
5552 ShuffleMask[i] < (int)NumElts ? V1 : V2,
5553 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
5556 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, Ops);
5557 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
5560 if ((VT == MVT::v8i16 || VT == MVT::v16i8) && isReverseMask(ShuffleMask, VT))
5561 return LowerReverse_VECTOR_SHUFFLEv16i8_v8i16(Op, DAG);
5563 if (VT == MVT::v8i8) {
5564 SDValue NewOp = LowerVECTOR_SHUFFLEv8i8(Op, ShuffleMask, DAG);
5565 if (NewOp.getNode())
5572 static SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
5573 // INSERT_VECTOR_ELT is legal only for immediate indexes.
5574 SDValue Lane = Op.getOperand(2);
5575 if (!isa<ConstantSDNode>(Lane))
5581 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
5582 // EXTRACT_VECTOR_ELT is legal only for immediate indexes.
5583 SDValue Lane = Op.getOperand(1);
5584 if (!isa<ConstantSDNode>(Lane))
5587 SDValue Vec = Op.getOperand(0);
5588 if (Op.getValueType() == MVT::i32 &&
5589 Vec.getValueType().getVectorElementType().getSizeInBits() < 32) {
5591 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
5597 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
5598 // The only time a CONCAT_VECTORS operation can have legal types is when
5599 // two 64-bit vectors are concatenated to a 128-bit vector.
5600 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
5601 "unexpected CONCAT_VECTORS");
5603 SDValue Val = DAG.getUNDEF(MVT::v2f64);
5604 SDValue Op0 = Op.getOperand(0);
5605 SDValue Op1 = Op.getOperand(1);
5606 if (Op0.getOpcode() != ISD::UNDEF)
5607 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
5608 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op0),
5609 DAG.getIntPtrConstant(0));
5610 if (Op1.getOpcode() != ISD::UNDEF)
5611 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
5612 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op1),
5613 DAG.getIntPtrConstant(1));
5614 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Val);
5617 /// isExtendedBUILD_VECTOR - Check if N is a constant BUILD_VECTOR where each
5618 /// element has been zero/sign-extended, depending on the isSigned parameter,
5619 /// from an integer type half its size.
5620 static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG,
5622 // A v2i64 BUILD_VECTOR will have been legalized to a BITCAST from v4i32.
5623 EVT VT = N->getValueType(0);
5624 if (VT == MVT::v2i64 && N->getOpcode() == ISD::BITCAST) {
5625 SDNode *BVN = N->getOperand(0).getNode();
5626 if (BVN->getValueType(0) != MVT::v4i32 ||
5627 BVN->getOpcode() != ISD::BUILD_VECTOR)
5629 unsigned LoElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
5630 unsigned HiElt = 1 - LoElt;
5631 ConstantSDNode *Lo0 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt));
5632 ConstantSDNode *Hi0 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt));
5633 ConstantSDNode *Lo1 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt+2));
5634 ConstantSDNode *Hi1 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt+2));
5635 if (!Lo0 || !Hi0 || !Lo1 || !Hi1)
5638 if (Hi0->getSExtValue() == Lo0->getSExtValue() >> 32 &&
5639 Hi1->getSExtValue() == Lo1->getSExtValue() >> 32)
5642 if (Hi0->isNullValue() && Hi1->isNullValue())
5648 if (N->getOpcode() != ISD::BUILD_VECTOR)
5651 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
5652 SDNode *Elt = N->getOperand(i).getNode();
5653 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
5654 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5655 unsigned HalfSize = EltSize / 2;
5657 if (!isIntN(HalfSize, C->getSExtValue()))
5660 if (!isUIntN(HalfSize, C->getZExtValue()))
5671 /// isSignExtended - Check if a node is a vector value that is sign-extended
5672 /// or a constant BUILD_VECTOR with sign-extended elements.
5673 static bool isSignExtended(SDNode *N, SelectionDAG &DAG) {
5674 if (N->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N))
5676 if (isExtendedBUILD_VECTOR(N, DAG, true))
5681 /// isZeroExtended - Check if a node is a vector value that is zero-extended
5682 /// or a constant BUILD_VECTOR with zero-extended elements.
5683 static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) {
5684 if (N->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N))
5686 if (isExtendedBUILD_VECTOR(N, DAG, false))
5691 static EVT getExtensionTo64Bits(const EVT &OrigVT) {
5692 if (OrigVT.getSizeInBits() >= 64)
5695 assert(OrigVT.isSimple() && "Expecting a simple value type");
5697 MVT::SimpleValueType OrigSimpleTy = OrigVT.getSimpleVT().SimpleTy;
5698 switch (OrigSimpleTy) {
5699 default: llvm_unreachable("Unexpected Vector Type");
5708 /// AddRequiredExtensionForVMULL - Add a sign/zero extension to extend the total
5709 /// value size to 64 bits. We need a 64-bit D register as an operand to VMULL.
5710 /// We insert the required extension here to get the vector to fill a D register.
5711 static SDValue AddRequiredExtensionForVMULL(SDValue N, SelectionDAG &DAG,
5714 unsigned ExtOpcode) {
5715 // The vector originally had a size of OrigTy. It was then extended to ExtTy.
5716 // We expect the ExtTy to be 128-bits total. If the OrigTy is less than
5717 // 64-bits we need to insert a new extension so that it will be 64-bits.
5718 assert(ExtTy.is128BitVector() && "Unexpected extension size");
5719 if (OrigTy.getSizeInBits() >= 64)
5722 // Must extend size to at least 64 bits to be used as an operand for VMULL.
5723 EVT NewVT = getExtensionTo64Bits(OrigTy);
5725 return DAG.getNode(ExtOpcode, SDLoc(N), NewVT, N);
5728 /// SkipLoadExtensionForVMULL - return a load of the original vector size that
5729 /// does not do any sign/zero extension. If the original vector is less
5730 /// than 64 bits, an appropriate extension will be added after the load to
5731 /// reach a total size of 64 bits. We have to add the extension separately
5732 /// because ARM does not have a sign/zero extending load for vectors.
5733 static SDValue SkipLoadExtensionForVMULL(LoadSDNode *LD, SelectionDAG& DAG) {
5734 EVT ExtendedTy = getExtensionTo64Bits(LD->getMemoryVT());
5736 // The load already has the right type.
5737 if (ExtendedTy == LD->getMemoryVT())
5738 return DAG.getLoad(LD->getMemoryVT(), SDLoc(LD), LD->getChain(),
5739 LD->getBasePtr(), LD->getPointerInfo(), LD->isVolatile(),
5740 LD->isNonTemporal(), LD->isInvariant(),
5741 LD->getAlignment());
5743 // We need to create a zextload/sextload. We cannot just create a load
5744 // followed by a zext/zext node because LowerMUL is also run during normal
5745 // operation legalization where we can't create illegal types.
5746 return DAG.getExtLoad(LD->getExtensionType(), SDLoc(LD), ExtendedTy,
5747 LD->getChain(), LD->getBasePtr(), LD->getPointerInfo(),
5748 LD->getMemoryVT(), LD->isVolatile(), LD->isInvariant(),
5749 LD->isNonTemporal(), LD->getAlignment());
5752 /// SkipExtensionForVMULL - For a node that is a SIGN_EXTEND, ZERO_EXTEND,
5753 /// extending load, or BUILD_VECTOR with extended elements, return the
5754 /// unextended value. The unextended vector should be 64 bits so that it can
5755 /// be used as an operand to a VMULL instruction. If the original vector size
5756 /// before extension is less than 64 bits we add a an extension to resize
5757 /// the vector to 64 bits.
5758 static SDValue SkipExtensionForVMULL(SDNode *N, SelectionDAG &DAG) {
5759 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
5760 return AddRequiredExtensionForVMULL(N->getOperand(0), DAG,
5761 N->getOperand(0)->getValueType(0),
5765 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
5766 return SkipLoadExtensionForVMULL(LD, DAG);
5768 // Otherwise, the value must be a BUILD_VECTOR. For v2i64, it will
5769 // have been legalized as a BITCAST from v4i32.
5770 if (N->getOpcode() == ISD::BITCAST) {
5771 SDNode *BVN = N->getOperand(0).getNode();
5772 assert(BVN->getOpcode() == ISD::BUILD_VECTOR &&
5773 BVN->getValueType(0) == MVT::v4i32 && "expected v4i32 BUILD_VECTOR");
5774 unsigned LowElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
5775 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), MVT::v2i32,
5776 BVN->getOperand(LowElt), BVN->getOperand(LowElt+2));
5778 // Construct a new BUILD_VECTOR with elements truncated to half the size.
5779 assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR");
5780 EVT VT = N->getValueType(0);
5781 unsigned EltSize = VT.getVectorElementType().getSizeInBits() / 2;
5782 unsigned NumElts = VT.getVectorNumElements();
5783 MVT TruncVT = MVT::getIntegerVT(EltSize);
5784 SmallVector<SDValue, 8> Ops;
5785 for (unsigned i = 0; i != NumElts; ++i) {
5786 ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i));
5787 const APInt &CInt = C->getAPIntValue();
5788 // Element types smaller than 32 bits are not legal, so use i32 elements.
5789 // The values are implicitly truncated so sext vs. zext doesn't matter.
5790 Ops.push_back(DAG.getConstant(CInt.zextOrTrunc(32), MVT::i32));
5792 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N),
5793 MVT::getVectorVT(TruncVT, NumElts), Ops);
5796 static bool isAddSubSExt(SDNode *N, SelectionDAG &DAG) {
5797 unsigned Opcode = N->getOpcode();
5798 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
5799 SDNode *N0 = N->getOperand(0).getNode();
5800 SDNode *N1 = N->getOperand(1).getNode();
5801 return N0->hasOneUse() && N1->hasOneUse() &&
5802 isSignExtended(N0, DAG) && isSignExtended(N1, DAG);
5807 static bool isAddSubZExt(SDNode *N, SelectionDAG &DAG) {
5808 unsigned Opcode = N->getOpcode();
5809 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
5810 SDNode *N0 = N->getOperand(0).getNode();
5811 SDNode *N1 = N->getOperand(1).getNode();
5812 return N0->hasOneUse() && N1->hasOneUse() &&
5813 isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG);
5818 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
5819 // Multiplications are only custom-lowered for 128-bit vectors so that
5820 // VMULL can be detected. Otherwise v2i64 multiplications are not legal.
5821 EVT VT = Op.getValueType();
5822 assert(VT.is128BitVector() && VT.isInteger() &&
5823 "unexpected type for custom-lowering ISD::MUL");
5824 SDNode *N0 = Op.getOperand(0).getNode();
5825 SDNode *N1 = Op.getOperand(1).getNode();
5826 unsigned NewOpc = 0;
5828 bool isN0SExt = isSignExtended(N0, DAG);
5829 bool isN1SExt = isSignExtended(N1, DAG);
5830 if (isN0SExt && isN1SExt)
5831 NewOpc = ARMISD::VMULLs;
5833 bool isN0ZExt = isZeroExtended(N0, DAG);
5834 bool isN1ZExt = isZeroExtended(N1, DAG);
5835 if (isN0ZExt && isN1ZExt)
5836 NewOpc = ARMISD::VMULLu;
5837 else if (isN1SExt || isN1ZExt) {
5838 // Look for (s/zext A + s/zext B) * (s/zext C). We want to turn these
5839 // into (s/zext A * s/zext C) + (s/zext B * s/zext C)
5840 if (isN1SExt && isAddSubSExt(N0, DAG)) {
5841 NewOpc = ARMISD::VMULLs;
5843 } else if (isN1ZExt && isAddSubZExt(N0, DAG)) {
5844 NewOpc = ARMISD::VMULLu;
5846 } else if (isN0ZExt && isAddSubZExt(N1, DAG)) {
5848 NewOpc = ARMISD::VMULLu;
5854 if (VT == MVT::v2i64)
5855 // Fall through to expand this. It is not legal.
5858 // Other vector multiplications are legal.
5863 // Legalize to a VMULL instruction.
5866 SDValue Op1 = SkipExtensionForVMULL(N1, DAG);
5868 Op0 = SkipExtensionForVMULL(N0, DAG);
5869 assert(Op0.getValueType().is64BitVector() &&
5870 Op1.getValueType().is64BitVector() &&
5871 "unexpected types for extended operands to VMULL");
5872 return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
5875 // Optimizing (zext A + zext B) * C, to (VMULL A, C) + (VMULL B, C) during
5876 // isel lowering to take advantage of no-stall back to back vmul + vmla.
5883 SDValue N00 = SkipExtensionForVMULL(N0->getOperand(0).getNode(), DAG);
5884 SDValue N01 = SkipExtensionForVMULL(N0->getOperand(1).getNode(), DAG);
5885 EVT Op1VT = Op1.getValueType();
5886 return DAG.getNode(N0->getOpcode(), DL, VT,
5887 DAG.getNode(NewOpc, DL, VT,
5888 DAG.getNode(ISD::BITCAST, DL, Op1VT, N00), Op1),
5889 DAG.getNode(NewOpc, DL, VT,
5890 DAG.getNode(ISD::BITCAST, DL, Op1VT, N01), Op1));
5894 LowerSDIV_v4i8(SDValue X, SDValue Y, SDLoc dl, SelectionDAG &DAG) {
5896 // float4 xf = vcvt_f32_s32(vmovl_s16(a.lo));
5897 // float4 yf = vcvt_f32_s32(vmovl_s16(b.lo));
5898 X = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, X);
5899 Y = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, Y);
5900 X = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, X);
5901 Y = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, Y);
5902 // Get reciprocal estimate.
5903 // float4 recip = vrecpeq_f32(yf);
5904 Y = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5905 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), Y);
5906 // Because char has a smaller range than uchar, we can actually get away
5907 // without any newton steps. This requires that we use a weird bias
5908 // of 0xb000, however (again, this has been exhaustively tested).
5909 // float4 result = as_float4(as_int4(xf*recip) + 0xb000);
5910 X = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, X, Y);
5911 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, X);
5912 Y = DAG.getConstant(0xb000, MVT::i32);
5913 Y = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Y, Y, Y, Y);
5914 X = DAG.getNode(ISD::ADD, dl, MVT::v4i32, X, Y);
5915 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, X);
5916 // Convert back to short.
5917 X = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, X);
5918 X = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, X);
5923 LowerSDIV_v4i16(SDValue N0, SDValue N1, SDLoc dl, SelectionDAG &DAG) {
5925 // Convert to float.
5926 // float4 yf = vcvt_f32_s32(vmovl_s16(y));
5927 // float4 xf = vcvt_f32_s32(vmovl_s16(x));
5928 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N0);
5929 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N1);
5930 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
5931 N1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
5933 // Use reciprocal estimate and one refinement step.
5934 // float4 recip = vrecpeq_f32(yf);
5935 // recip *= vrecpsq_f32(yf, recip);
5936 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5937 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), N1);
5938 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5939 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
5941 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
5942 // Because short has a smaller range than ushort, we can actually get away
5943 // with only a single newton step. This requires that we use a weird bias
5944 // of 89, however (again, this has been exhaustively tested).
5945 // float4 result = as_float4(as_int4(xf*recip) + 0x89);
5946 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
5947 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
5948 N1 = DAG.getConstant(0x89, MVT::i32);
5949 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
5950 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
5951 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
5952 // Convert back to integer and return.
5953 // return vmovn_s32(vcvt_s32_f32(result));
5954 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
5955 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
5959 static SDValue LowerSDIV(SDValue Op, SelectionDAG &DAG) {
5960 EVT VT = Op.getValueType();
5961 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
5962 "unexpected type for custom-lowering ISD::SDIV");
5965 SDValue N0 = Op.getOperand(0);
5966 SDValue N1 = Op.getOperand(1);
5969 if (VT == MVT::v8i8) {
5970 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N0);
5971 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N1);
5973 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5974 DAG.getIntPtrConstant(4));
5975 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5976 DAG.getIntPtrConstant(4));
5977 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
5978 DAG.getIntPtrConstant(0));
5979 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
5980 DAG.getIntPtrConstant(0));
5982 N0 = LowerSDIV_v4i8(N0, N1, dl, DAG); // v4i16
5983 N2 = LowerSDIV_v4i8(N2, N3, dl, DAG); // v4i16
5985 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
5986 N0 = LowerCONCAT_VECTORS(N0, DAG);
5988 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v8i8, N0);
5991 return LowerSDIV_v4i16(N0, N1, dl, DAG);
5994 static SDValue LowerUDIV(SDValue Op, SelectionDAG &DAG) {
5995 EVT VT = Op.getValueType();
5996 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
5997 "unexpected type for custom-lowering ISD::UDIV");
6000 SDValue N0 = Op.getOperand(0);
6001 SDValue N1 = Op.getOperand(1);
6004 if (VT == MVT::v8i8) {
6005 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N0);
6006 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N1);
6008 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
6009 DAG.getIntPtrConstant(4));
6010 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
6011 DAG.getIntPtrConstant(4));
6012 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
6013 DAG.getIntPtrConstant(0));
6014 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
6015 DAG.getIntPtrConstant(0));
6017 N0 = LowerSDIV_v4i16(N0, N1, dl, DAG); // v4i16
6018 N2 = LowerSDIV_v4i16(N2, N3, dl, DAG); // v4i16
6020 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
6021 N0 = LowerCONCAT_VECTORS(N0, DAG);
6023 N0 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v8i8,
6024 DAG.getConstant(Intrinsic::arm_neon_vqmovnsu, MVT::i32),
6029 // v4i16 sdiv ... Convert to float.
6030 // float4 yf = vcvt_f32_s32(vmovl_u16(y));
6031 // float4 xf = vcvt_f32_s32(vmovl_u16(x));
6032 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N0);
6033 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N1);
6034 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
6035 SDValue BN1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
6037 // Use reciprocal estimate and two refinement steps.
6038 // float4 recip = vrecpeq_f32(yf);
6039 // recip *= vrecpsq_f32(yf, recip);
6040 // recip *= vrecpsq_f32(yf, recip);
6041 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6042 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), BN1);
6043 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6044 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
6046 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
6047 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6048 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
6050 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
6051 // Simply multiplying by the reciprocal estimate can leave us a few ulps
6052 // too low, so we add 2 ulps (exhaustive testing shows that this is enough,
6053 // and that it will never cause us to return an answer too large).
6054 // float4 result = as_float4(as_int4(xf*recip) + 2);
6055 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
6056 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
6057 N1 = DAG.getConstant(2, MVT::i32);
6058 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
6059 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
6060 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
6061 // Convert back to integer and return.
6062 // return vmovn_u32(vcvt_s32_f32(result));
6063 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
6064 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
6068 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
6069 EVT VT = Op.getNode()->getValueType(0);
6070 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
6073 bool ExtraOp = false;
6074 switch (Op.getOpcode()) {
6075 default: llvm_unreachable("Invalid code");
6076 case ISD::ADDC: Opc = ARMISD::ADDC; break;
6077 case ISD::ADDE: Opc = ARMISD::ADDE; ExtraOp = true; break;
6078 case ISD::SUBC: Opc = ARMISD::SUBC; break;
6079 case ISD::SUBE: Opc = ARMISD::SUBE; ExtraOp = true; break;
6083 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
6085 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
6086 Op.getOperand(1), Op.getOperand(2));
6089 SDValue ARMTargetLowering::LowerFSINCOS(SDValue Op, SelectionDAG &DAG) const {
6090 assert(Subtarget->isTargetDarwin());
6092 // For iOS, we want to call an alternative entry point: __sincos_stret,
6093 // return values are passed via sret.
6095 SDValue Arg = Op.getOperand(0);
6096 EVT ArgVT = Arg.getValueType();
6097 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
6099 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
6100 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6102 // Pair of floats / doubles used to pass the result.
6103 StructType *RetTy = StructType::get(ArgTy, ArgTy, NULL);
6105 // Create stack object for sret.
6106 const uint64_t ByteSize = TLI.getDataLayout()->getTypeAllocSize(RetTy);
6107 const unsigned StackAlign = TLI.getDataLayout()->getPrefTypeAlignment(RetTy);
6108 int FrameIdx = FrameInfo->CreateStackObject(ByteSize, StackAlign, false);
6109 SDValue SRet = DAG.getFrameIndex(FrameIdx, TLI.getPointerTy());
6115 Entry.Ty = RetTy->getPointerTo();
6116 Entry.isSExt = false;
6117 Entry.isZExt = false;
6118 Entry.isSRet = true;
6119 Args.push_back(Entry);
6123 Entry.isSExt = false;
6124 Entry.isZExt = false;
6125 Args.push_back(Entry);
6127 const char *LibcallName = (ArgVT == MVT::f64)
6128 ? "__sincos_stret" : "__sincosf_stret";
6129 SDValue Callee = DAG.getExternalSymbol(LibcallName, getPointerTy());
6131 TargetLowering::CallLoweringInfo CLI(DAG);
6132 CLI.setDebugLoc(dl).setChain(DAG.getEntryNode())
6133 .setCallee(CallingConv::C, Type::getVoidTy(*DAG.getContext()), Callee,
6135 .setDiscardResult();
6137 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
6139 SDValue LoadSin = DAG.getLoad(ArgVT, dl, CallResult.second, SRet,
6140 MachinePointerInfo(), false, false, false, 0);
6142 // Address of cos field.
6143 SDValue Add = DAG.getNode(ISD::ADD, dl, getPointerTy(), SRet,
6144 DAG.getIntPtrConstant(ArgVT.getStoreSize()));
6145 SDValue LoadCos = DAG.getLoad(ArgVT, dl, LoadSin.getValue(1), Add,
6146 MachinePointerInfo(), false, false, false, 0);
6148 SDVTList Tys = DAG.getVTList(ArgVT, ArgVT);
6149 return DAG.getNode(ISD::MERGE_VALUES, dl, Tys,
6150 LoadSin.getValue(0), LoadCos.getValue(0));
6153 static SDValue LowerAtomicLoadStore(SDValue Op, SelectionDAG &DAG) {
6154 // Monotonic load/store is legal for all targets
6155 if (cast<AtomicSDNode>(Op)->getOrdering() <= Monotonic)
6158 // Acquire/Release load/store is not legal for targets without a
6159 // dmb or equivalent available.
6163 static void ReplaceREADCYCLECOUNTER(SDNode *N,
6164 SmallVectorImpl<SDValue> &Results,
6166 const ARMSubtarget *Subtarget) {
6168 SDValue Cycles32, OutChain;
6170 if (Subtarget->hasPerfMon()) {
6171 // Under Power Management extensions, the cycle-count is:
6172 // mrc p15, #0, <Rt>, c9, c13, #0
6173 SDValue Ops[] = { N->getOperand(0), // Chain
6174 DAG.getConstant(Intrinsic::arm_mrc, MVT::i32),
6175 DAG.getConstant(15, MVT::i32),
6176 DAG.getConstant(0, MVT::i32),
6177 DAG.getConstant(9, MVT::i32),
6178 DAG.getConstant(13, MVT::i32),
6179 DAG.getConstant(0, MVT::i32)
6182 Cycles32 = DAG.getNode(ISD::INTRINSIC_W_CHAIN, DL,
6183 DAG.getVTList(MVT::i32, MVT::Other), Ops);
6184 OutChain = Cycles32.getValue(1);
6186 // Intrinsic is defined to return 0 on unsupported platforms. Technically
6187 // there are older ARM CPUs that have implementation-specific ways of
6188 // obtaining this information (FIXME!).
6189 Cycles32 = DAG.getConstant(0, MVT::i32);
6190 OutChain = DAG.getEntryNode();
6194 SDValue Cycles64 = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64,
6195 Cycles32, DAG.getConstant(0, MVT::i32));
6196 Results.push_back(Cycles64);
6197 Results.push_back(OutChain);
6200 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
6201 switch (Op.getOpcode()) {
6202 default: llvm_unreachable("Don't know how to custom lower this!");
6203 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
6204 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
6205 case ISD::GlobalAddress:
6206 switch (Subtarget->getTargetTriple().getObjectFormat()) {
6207 default: llvm_unreachable("unknown object format");
6209 return LowerGlobalAddressWindows(Op, DAG);
6211 return LowerGlobalAddressELF(Op, DAG);
6213 return LowerGlobalAddressDarwin(Op, DAG);
6215 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
6216 case ISD::SELECT: return LowerSELECT(Op, DAG);
6217 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
6218 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
6219 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
6220 case ISD::VASTART: return LowerVASTART(Op, DAG);
6221 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, DAG, Subtarget);
6222 case ISD::PREFETCH: return LowerPREFETCH(Op, DAG, Subtarget);
6223 case ISD::SINT_TO_FP:
6224 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
6225 case ISD::FP_TO_SINT:
6226 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
6227 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
6228 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
6229 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
6230 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
6231 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
6232 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
6233 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
6235 case ISD::BITCAST: return ExpandBITCAST(Op.getNode(), DAG);
6238 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
6239 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
6240 case ISD::SRL_PARTS:
6241 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
6242 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
6243 case ISD::CTPOP: return LowerCTPOP(Op.getNode(), DAG, Subtarget);
6244 case ISD::SETCC: return LowerVSETCC(Op, DAG);
6245 case ISD::ConstantFP: return LowerConstantFP(Op, DAG, Subtarget);
6246 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG, Subtarget);
6247 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
6248 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
6249 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
6250 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
6251 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
6252 case ISD::MUL: return LowerMUL(Op, DAG);
6253 case ISD::SDIV: return LowerSDIV(Op, DAG);
6254 case ISD::UDIV: return LowerUDIV(Op, DAG);
6258 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
6263 return LowerXALUO(Op, DAG);
6264 case ISD::ATOMIC_LOAD:
6265 case ISD::ATOMIC_STORE: return LowerAtomicLoadStore(Op, DAG);
6266 case ISD::FSINCOS: return LowerFSINCOS(Op, DAG);
6268 case ISD::UDIVREM: return LowerDivRem(Op, DAG);
6269 case ISD::DYNAMIC_STACKALLOC:
6270 if (Subtarget->getTargetTriple().isWindowsItaniumEnvironment())
6271 return LowerDYNAMIC_STACKALLOC(Op, DAG);
6272 llvm_unreachable("Don't know how to custom lower this!");
6276 /// ReplaceNodeResults - Replace the results of node with an illegal result
6277 /// type with new values built out of custom code.
6278 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
6279 SmallVectorImpl<SDValue>&Results,
6280 SelectionDAG &DAG) const {
6282 switch (N->getOpcode()) {
6284 llvm_unreachable("Don't know how to custom expand this!");
6286 Res = ExpandBITCAST(N, DAG);
6290 Res = Expand64BitShift(N, DAG, Subtarget);
6292 case ISD::READCYCLECOUNTER:
6293 ReplaceREADCYCLECOUNTER(N, Results, DAG, Subtarget);
6297 Results.push_back(Res);
6300 //===----------------------------------------------------------------------===//
6301 // ARM Scheduler Hooks
6302 //===----------------------------------------------------------------------===//
6304 /// SetupEntryBlockForSjLj - Insert code into the entry block that creates and
6305 /// registers the function context.
6306 void ARMTargetLowering::
6307 SetupEntryBlockForSjLj(MachineInstr *MI, MachineBasicBlock *MBB,
6308 MachineBasicBlock *DispatchBB, int FI) const {
6309 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6310 DebugLoc dl = MI->getDebugLoc();
6311 MachineFunction *MF = MBB->getParent();
6312 MachineRegisterInfo *MRI = &MF->getRegInfo();
6313 MachineConstantPool *MCP = MF->getConstantPool();
6314 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
6315 const Function *F = MF->getFunction();
6317 bool isThumb = Subtarget->isThumb();
6318 bool isThumb2 = Subtarget->isThumb2();
6320 unsigned PCLabelId = AFI->createPICLabelUId();
6321 unsigned PCAdj = (isThumb || isThumb2) ? 4 : 8;
6322 ARMConstantPoolValue *CPV =
6323 ARMConstantPoolMBB::Create(F->getContext(), DispatchBB, PCLabelId, PCAdj);
6324 unsigned CPI = MCP->getConstantPoolIndex(CPV, 4);
6326 const TargetRegisterClass *TRC = isThumb ?
6327 (const TargetRegisterClass*)&ARM::tGPRRegClass :
6328 (const TargetRegisterClass*)&ARM::GPRRegClass;
6330 // Grab constant pool and fixed stack memory operands.
6331 MachineMemOperand *CPMMO =
6332 MF->getMachineMemOperand(MachinePointerInfo::getConstantPool(),
6333 MachineMemOperand::MOLoad, 4, 4);
6335 MachineMemOperand *FIMMOSt =
6336 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
6337 MachineMemOperand::MOStore, 4, 4);
6339 // Load the address of the dispatch MBB into the jump buffer.
6341 // Incoming value: jbuf
6342 // ldr.n r5, LCPI1_1
6345 // str r5, [$jbuf, #+4] ; &jbuf[1]
6346 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6347 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2LDRpci), NewVReg1)
6348 .addConstantPoolIndex(CPI)
6349 .addMemOperand(CPMMO));
6350 // Set the low bit because of thumb mode.
6351 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6353 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2ORRri), NewVReg2)
6354 .addReg(NewVReg1, RegState::Kill)
6356 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6357 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg3)
6358 .addReg(NewVReg2, RegState::Kill)
6360 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2STRi12))
6361 .addReg(NewVReg3, RegState::Kill)
6363 .addImm(36) // &jbuf[1] :: pc
6364 .addMemOperand(FIMMOSt));
6365 } else if (isThumb) {
6366 // Incoming value: jbuf
6367 // ldr.n r1, LCPI1_4
6371 // add r2, $jbuf, #+4 ; &jbuf[1]
6373 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6374 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tLDRpci), NewVReg1)
6375 .addConstantPoolIndex(CPI)
6376 .addMemOperand(CPMMO));
6377 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6378 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg2)
6379 .addReg(NewVReg1, RegState::Kill)
6381 // Set the low bit because of thumb mode.
6382 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6383 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tMOVi8), NewVReg3)
6384 .addReg(ARM::CPSR, RegState::Define)
6386 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6387 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tORR), NewVReg4)
6388 .addReg(ARM::CPSR, RegState::Define)
6389 .addReg(NewVReg2, RegState::Kill)
6390 .addReg(NewVReg3, RegState::Kill));
6391 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6392 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tADDrSPi), NewVReg5)
6394 .addImm(36)); // &jbuf[1] :: pc
6395 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tSTRi))
6396 .addReg(NewVReg4, RegState::Kill)
6397 .addReg(NewVReg5, RegState::Kill)
6399 .addMemOperand(FIMMOSt));
6401 // Incoming value: jbuf
6404 // str r1, [$jbuf, #+4] ; &jbuf[1]
6405 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6406 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::LDRi12), NewVReg1)
6407 .addConstantPoolIndex(CPI)
6409 .addMemOperand(CPMMO));
6410 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6411 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::PICADD), NewVReg2)
6412 .addReg(NewVReg1, RegState::Kill)
6413 .addImm(PCLabelId));
6414 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::STRi12))
6415 .addReg(NewVReg2, RegState::Kill)
6417 .addImm(36) // &jbuf[1] :: pc
6418 .addMemOperand(FIMMOSt));
6422 MachineBasicBlock *ARMTargetLowering::
6423 EmitSjLjDispatchBlock(MachineInstr *MI, MachineBasicBlock *MBB) const {
6424 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6425 DebugLoc dl = MI->getDebugLoc();
6426 MachineFunction *MF = MBB->getParent();
6427 MachineRegisterInfo *MRI = &MF->getRegInfo();
6428 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
6429 MachineFrameInfo *MFI = MF->getFrameInfo();
6430 int FI = MFI->getFunctionContextIndex();
6432 const TargetRegisterClass *TRC = Subtarget->isThumb() ?
6433 (const TargetRegisterClass*)&ARM::tGPRRegClass :
6434 (const TargetRegisterClass*)&ARM::GPRnopcRegClass;
6436 // Get a mapping of the call site numbers to all of the landing pads they're
6438 DenseMap<unsigned, SmallVector<MachineBasicBlock*, 2> > CallSiteNumToLPad;
6439 unsigned MaxCSNum = 0;
6440 MachineModuleInfo &MMI = MF->getMMI();
6441 for (MachineFunction::iterator BB = MF->begin(), E = MF->end(); BB != E;
6443 if (!BB->isLandingPad()) continue;
6445 // FIXME: We should assert that the EH_LABEL is the first MI in the landing
6447 for (MachineBasicBlock::iterator
6448 II = BB->begin(), IE = BB->end(); II != IE; ++II) {
6449 if (!II->isEHLabel()) continue;
6451 MCSymbol *Sym = II->getOperand(0).getMCSymbol();
6452 if (!MMI.hasCallSiteLandingPad(Sym)) continue;
6454 SmallVectorImpl<unsigned> &CallSiteIdxs = MMI.getCallSiteLandingPad(Sym);
6455 for (SmallVectorImpl<unsigned>::iterator
6456 CSI = CallSiteIdxs.begin(), CSE = CallSiteIdxs.end();
6457 CSI != CSE; ++CSI) {
6458 CallSiteNumToLPad[*CSI].push_back(BB);
6459 MaxCSNum = std::max(MaxCSNum, *CSI);
6465 // Get an ordered list of the machine basic blocks for the jump table.
6466 std::vector<MachineBasicBlock*> LPadList;
6467 SmallPtrSet<MachineBasicBlock*, 64> InvokeBBs;
6468 LPadList.reserve(CallSiteNumToLPad.size());
6469 for (unsigned I = 1; I <= MaxCSNum; ++I) {
6470 SmallVectorImpl<MachineBasicBlock*> &MBBList = CallSiteNumToLPad[I];
6471 for (SmallVectorImpl<MachineBasicBlock*>::iterator
6472 II = MBBList.begin(), IE = MBBList.end(); II != IE; ++II) {
6473 LPadList.push_back(*II);
6474 InvokeBBs.insert((*II)->pred_begin(), (*II)->pred_end());
6478 assert(!LPadList.empty() &&
6479 "No landing pad destinations for the dispatch jump table!");
6481 // Create the jump table and associated information.
6482 MachineJumpTableInfo *JTI =
6483 MF->getOrCreateJumpTableInfo(MachineJumpTableInfo::EK_Inline);
6484 unsigned MJTI = JTI->createJumpTableIndex(LPadList);
6485 unsigned UId = AFI->createJumpTableUId();
6486 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
6488 // Create the MBBs for the dispatch code.
6490 // Shove the dispatch's address into the return slot in the function context.
6491 MachineBasicBlock *DispatchBB = MF->CreateMachineBasicBlock();
6492 DispatchBB->setIsLandingPad();
6494 MachineBasicBlock *TrapBB = MF->CreateMachineBasicBlock();
6495 unsigned trap_opcode;
6496 if (Subtarget->isThumb())
6497 trap_opcode = ARM::tTRAP;
6499 trap_opcode = Subtarget->useNaClTrap() ? ARM::TRAPNaCl : ARM::TRAP;
6501 BuildMI(TrapBB, dl, TII->get(trap_opcode));
6502 DispatchBB->addSuccessor(TrapBB);
6504 MachineBasicBlock *DispContBB = MF->CreateMachineBasicBlock();
6505 DispatchBB->addSuccessor(DispContBB);
6508 MF->insert(MF->end(), DispatchBB);
6509 MF->insert(MF->end(), DispContBB);
6510 MF->insert(MF->end(), TrapBB);
6512 // Insert code into the entry block that creates and registers the function
6514 SetupEntryBlockForSjLj(MI, MBB, DispatchBB, FI);
6516 MachineMemOperand *FIMMOLd =
6517 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
6518 MachineMemOperand::MOLoad |
6519 MachineMemOperand::MOVolatile, 4, 4);
6521 MachineInstrBuilder MIB;
6522 MIB = BuildMI(DispatchBB, dl, TII->get(ARM::Int_eh_sjlj_dispatchsetup));
6524 const ARMBaseInstrInfo *AII = static_cast<const ARMBaseInstrInfo*>(TII);
6525 const ARMBaseRegisterInfo &RI = AII->getRegisterInfo();
6527 // Add a register mask with no preserved registers. This results in all
6528 // registers being marked as clobbered.
6529 MIB.addRegMask(RI.getNoPreservedMask());
6531 unsigned NumLPads = LPadList.size();
6532 if (Subtarget->isThumb2()) {
6533 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6534 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2LDRi12), NewVReg1)
6537 .addMemOperand(FIMMOLd));
6539 if (NumLPads < 256) {
6540 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPri))
6542 .addImm(LPadList.size()));
6544 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6545 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVi16), VReg1)
6546 .addImm(NumLPads & 0xFFFF));
6548 unsigned VReg2 = VReg1;
6549 if ((NumLPads & 0xFFFF0000) != 0) {
6550 VReg2 = MRI->createVirtualRegister(TRC);
6551 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVTi16), VReg2)
6553 .addImm(NumLPads >> 16));
6556 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPrr))
6561 BuildMI(DispatchBB, dl, TII->get(ARM::t2Bcc))
6566 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6567 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::t2LEApcrelJT),NewVReg3)
6568 .addJumpTableIndex(MJTI)
6571 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6574 BuildMI(DispContBB, dl, TII->get(ARM::t2ADDrs), NewVReg4)
6575 .addReg(NewVReg3, RegState::Kill)
6577 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
6579 BuildMI(DispContBB, dl, TII->get(ARM::t2BR_JT))
6580 .addReg(NewVReg4, RegState::Kill)
6582 .addJumpTableIndex(MJTI)
6584 } else if (Subtarget->isThumb()) {
6585 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6586 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRspi), NewVReg1)
6589 .addMemOperand(FIMMOLd));
6591 if (NumLPads < 256) {
6592 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPi8))
6596 MachineConstantPool *ConstantPool = MF->getConstantPool();
6597 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
6598 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
6600 // MachineConstantPool wants an explicit alignment.
6601 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
6603 Align = getDataLayout()->getTypeAllocSize(C->getType());
6604 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6606 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6607 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRpci))
6608 .addReg(VReg1, RegState::Define)
6609 .addConstantPoolIndex(Idx));
6610 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPr))
6615 BuildMI(DispatchBB, dl, TII->get(ARM::tBcc))
6620 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6621 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLSLri), NewVReg2)
6622 .addReg(ARM::CPSR, RegState::Define)
6626 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6627 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLEApcrelJT), NewVReg3)
6628 .addJumpTableIndex(MJTI)
6631 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6632 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg4)
6633 .addReg(ARM::CPSR, RegState::Define)
6634 .addReg(NewVReg2, RegState::Kill)
6637 MachineMemOperand *JTMMOLd =
6638 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
6639 MachineMemOperand::MOLoad, 4, 4);
6641 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6642 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLDRi), NewVReg5)
6643 .addReg(NewVReg4, RegState::Kill)
6645 .addMemOperand(JTMMOLd));
6647 unsigned NewVReg6 = NewVReg5;
6648 if (RelocM == Reloc::PIC_) {
6649 NewVReg6 = MRI->createVirtualRegister(TRC);
6650 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg6)
6651 .addReg(ARM::CPSR, RegState::Define)
6652 .addReg(NewVReg5, RegState::Kill)
6656 BuildMI(DispContBB, dl, TII->get(ARM::tBR_JTr))
6657 .addReg(NewVReg6, RegState::Kill)
6658 .addJumpTableIndex(MJTI)
6661 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6662 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRi12), NewVReg1)
6665 .addMemOperand(FIMMOLd));
6667 if (NumLPads < 256) {
6668 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPri))
6671 } else if (Subtarget->hasV6T2Ops() && isUInt<16>(NumLPads)) {
6672 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6673 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVi16), VReg1)
6674 .addImm(NumLPads & 0xFFFF));
6676 unsigned VReg2 = VReg1;
6677 if ((NumLPads & 0xFFFF0000) != 0) {
6678 VReg2 = MRI->createVirtualRegister(TRC);
6679 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVTi16), VReg2)
6681 .addImm(NumLPads >> 16));
6684 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
6688 MachineConstantPool *ConstantPool = MF->getConstantPool();
6689 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
6690 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
6692 // MachineConstantPool wants an explicit alignment.
6693 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
6695 Align = getDataLayout()->getTypeAllocSize(C->getType());
6696 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6698 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6699 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRcp))
6700 .addReg(VReg1, RegState::Define)
6701 .addConstantPoolIndex(Idx)
6703 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
6705 .addReg(VReg1, RegState::Kill));
6708 BuildMI(DispatchBB, dl, TII->get(ARM::Bcc))
6713 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6715 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::MOVsi), NewVReg3)
6717 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
6718 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6719 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::LEApcrelJT), NewVReg4)
6720 .addJumpTableIndex(MJTI)
6723 MachineMemOperand *JTMMOLd =
6724 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
6725 MachineMemOperand::MOLoad, 4, 4);
6726 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6728 BuildMI(DispContBB, dl, TII->get(ARM::LDRrs), NewVReg5)
6729 .addReg(NewVReg3, RegState::Kill)
6732 .addMemOperand(JTMMOLd));
6734 if (RelocM == Reloc::PIC_) {
6735 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTadd))
6736 .addReg(NewVReg5, RegState::Kill)
6738 .addJumpTableIndex(MJTI)
6741 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTr))
6742 .addReg(NewVReg5, RegState::Kill)
6743 .addJumpTableIndex(MJTI)
6748 // Add the jump table entries as successors to the MBB.
6749 SmallPtrSet<MachineBasicBlock*, 8> SeenMBBs;
6750 for (std::vector<MachineBasicBlock*>::iterator
6751 I = LPadList.begin(), E = LPadList.end(); I != E; ++I) {
6752 MachineBasicBlock *CurMBB = *I;
6753 if (SeenMBBs.insert(CurMBB))
6754 DispContBB->addSuccessor(CurMBB);
6757 // N.B. the order the invoke BBs are processed in doesn't matter here.
6758 const MCPhysReg *SavedRegs = RI.getCalleeSavedRegs(MF);
6759 SmallVector<MachineBasicBlock*, 64> MBBLPads;
6760 for (SmallPtrSet<MachineBasicBlock*, 64>::iterator
6761 I = InvokeBBs.begin(), E = InvokeBBs.end(); I != E; ++I) {
6762 MachineBasicBlock *BB = *I;
6764 // Remove the landing pad successor from the invoke block and replace it
6765 // with the new dispatch block.
6766 SmallVector<MachineBasicBlock*, 4> Successors(BB->succ_begin(),
6768 while (!Successors.empty()) {
6769 MachineBasicBlock *SMBB = Successors.pop_back_val();
6770 if (SMBB->isLandingPad()) {
6771 BB->removeSuccessor(SMBB);
6772 MBBLPads.push_back(SMBB);
6776 BB->addSuccessor(DispatchBB);
6778 // Find the invoke call and mark all of the callee-saved registers as
6779 // 'implicit defined' so that they're spilled. This prevents code from
6780 // moving instructions to before the EH block, where they will never be
6782 for (MachineBasicBlock::reverse_iterator
6783 II = BB->rbegin(), IE = BB->rend(); II != IE; ++II) {
6784 if (!II->isCall()) continue;
6786 DenseMap<unsigned, bool> DefRegs;
6787 for (MachineInstr::mop_iterator
6788 OI = II->operands_begin(), OE = II->operands_end();
6790 if (!OI->isReg()) continue;
6791 DefRegs[OI->getReg()] = true;
6794 MachineInstrBuilder MIB(*MF, &*II);
6796 for (unsigned i = 0; SavedRegs[i] != 0; ++i) {
6797 unsigned Reg = SavedRegs[i];
6798 if (Subtarget->isThumb2() &&
6799 !ARM::tGPRRegClass.contains(Reg) &&
6800 !ARM::hGPRRegClass.contains(Reg))
6802 if (Subtarget->isThumb1Only() && !ARM::tGPRRegClass.contains(Reg))
6804 if (!Subtarget->isThumb() && !ARM::GPRRegClass.contains(Reg))
6807 MIB.addReg(Reg, RegState::ImplicitDefine | RegState::Dead);
6814 // Mark all former landing pads as non-landing pads. The dispatch is the only
6816 for (SmallVectorImpl<MachineBasicBlock*>::iterator
6817 I = MBBLPads.begin(), E = MBBLPads.end(); I != E; ++I)
6818 (*I)->setIsLandingPad(false);
6820 // The instruction is gone now.
6821 MI->eraseFromParent();
6827 MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) {
6828 for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
6829 E = MBB->succ_end(); I != E; ++I)
6832 llvm_unreachable("Expecting a BB with two successors!");
6835 /// Return the load opcode for a given load size. If load size >= 8,
6836 /// neon opcode will be returned.
6837 static unsigned getLdOpcode(unsigned LdSize, bool IsThumb1, bool IsThumb2) {
6839 return LdSize == 16 ? ARM::VLD1q32wb_fixed
6840 : LdSize == 8 ? ARM::VLD1d32wb_fixed : 0;
6842 return LdSize == 4 ? ARM::tLDRi
6843 : LdSize == 2 ? ARM::tLDRHi
6844 : LdSize == 1 ? ARM::tLDRBi : 0;
6846 return LdSize == 4 ? ARM::t2LDR_POST
6847 : LdSize == 2 ? ARM::t2LDRH_POST
6848 : LdSize == 1 ? ARM::t2LDRB_POST : 0;
6849 return LdSize == 4 ? ARM::LDR_POST_IMM
6850 : LdSize == 2 ? ARM::LDRH_POST
6851 : LdSize == 1 ? ARM::LDRB_POST_IMM : 0;
6854 /// Return the store opcode for a given store size. If store size >= 8,
6855 /// neon opcode will be returned.
6856 static unsigned getStOpcode(unsigned StSize, bool IsThumb1, bool IsThumb2) {
6858 return StSize == 16 ? ARM::VST1q32wb_fixed
6859 : StSize == 8 ? ARM::VST1d32wb_fixed : 0;
6861 return StSize == 4 ? ARM::tSTRi
6862 : StSize == 2 ? ARM::tSTRHi
6863 : StSize == 1 ? ARM::tSTRBi : 0;
6865 return StSize == 4 ? ARM::t2STR_POST
6866 : StSize == 2 ? ARM::t2STRH_POST
6867 : StSize == 1 ? ARM::t2STRB_POST : 0;
6868 return StSize == 4 ? ARM::STR_POST_IMM
6869 : StSize == 2 ? ARM::STRH_POST
6870 : StSize == 1 ? ARM::STRB_POST_IMM : 0;
6873 /// Emit a post-increment load operation with given size. The instructions
6874 /// will be added to BB at Pos.
6875 static void emitPostLd(MachineBasicBlock *BB, MachineInstr *Pos,
6876 const TargetInstrInfo *TII, DebugLoc dl,
6877 unsigned LdSize, unsigned Data, unsigned AddrIn,
6878 unsigned AddrOut, bool IsThumb1, bool IsThumb2) {
6879 unsigned LdOpc = getLdOpcode(LdSize, IsThumb1, IsThumb2);
6880 assert(LdOpc != 0 && "Should have a load opcode");
6882 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
6883 .addReg(AddrOut, RegState::Define).addReg(AddrIn)
6885 } else if (IsThumb1) {
6886 // load + update AddrIn
6887 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
6888 .addReg(AddrIn).addImm(0));
6889 MachineInstrBuilder MIB =
6890 BuildMI(*BB, Pos, dl, TII->get(ARM::tADDi8), AddrOut);
6891 MIB = AddDefaultT1CC(MIB);
6892 MIB.addReg(AddrIn).addImm(LdSize);
6893 AddDefaultPred(MIB);
6894 } else if (IsThumb2) {
6895 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
6896 .addReg(AddrOut, RegState::Define).addReg(AddrIn)
6899 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
6900 .addReg(AddrOut, RegState::Define).addReg(AddrIn)
6901 .addReg(0).addImm(LdSize));
6905 /// Emit a post-increment store operation with given size. The instructions
6906 /// will be added to BB at Pos.
6907 static void emitPostSt(MachineBasicBlock *BB, MachineInstr *Pos,
6908 const TargetInstrInfo *TII, DebugLoc dl,
6909 unsigned StSize, unsigned Data, unsigned AddrIn,
6910 unsigned AddrOut, bool IsThumb1, bool IsThumb2) {
6911 unsigned StOpc = getStOpcode(StSize, IsThumb1, IsThumb2);
6912 assert(StOpc != 0 && "Should have a store opcode");
6914 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc), AddrOut)
6915 .addReg(AddrIn).addImm(0).addReg(Data));
6916 } else if (IsThumb1) {
6917 // store + update AddrIn
6918 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc)).addReg(Data)
6919 .addReg(AddrIn).addImm(0));
6920 MachineInstrBuilder MIB =
6921 BuildMI(*BB, Pos, dl, TII->get(ARM::tADDi8), AddrOut);
6922 MIB = AddDefaultT1CC(MIB);
6923 MIB.addReg(AddrIn).addImm(StSize);
6924 AddDefaultPred(MIB);
6925 } else if (IsThumb2) {
6926 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc), AddrOut)
6927 .addReg(Data).addReg(AddrIn).addImm(StSize));
6929 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc), AddrOut)
6930 .addReg(Data).addReg(AddrIn).addReg(0)
6936 ARMTargetLowering::EmitStructByval(MachineInstr *MI,
6937 MachineBasicBlock *BB) const {
6938 // This pseudo instruction has 3 operands: dst, src, size
6939 // We expand it to a loop if size > Subtarget->getMaxInlineSizeThreshold().
6940 // Otherwise, we will generate unrolled scalar copies.
6941 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6942 const BasicBlock *LLVM_BB = BB->getBasicBlock();
6943 MachineFunction::iterator It = BB;
6946 unsigned dest = MI->getOperand(0).getReg();
6947 unsigned src = MI->getOperand(1).getReg();
6948 unsigned SizeVal = MI->getOperand(2).getImm();
6949 unsigned Align = MI->getOperand(3).getImm();
6950 DebugLoc dl = MI->getDebugLoc();
6952 MachineFunction *MF = BB->getParent();
6953 MachineRegisterInfo &MRI = MF->getRegInfo();
6954 unsigned UnitSize = 0;
6955 const TargetRegisterClass *TRC = nullptr;
6956 const TargetRegisterClass *VecTRC = nullptr;
6958 bool IsThumb1 = Subtarget->isThumb1Only();
6959 bool IsThumb2 = Subtarget->isThumb2();
6963 } else if (Align & 2) {
6966 // Check whether we can use NEON instructions.
6967 if (!MF->getFunction()->getAttributes().
6968 hasAttribute(AttributeSet::FunctionIndex,
6969 Attribute::NoImplicitFloat) &&
6970 Subtarget->hasNEON()) {
6971 if ((Align % 16 == 0) && SizeVal >= 16)
6973 else if ((Align % 8 == 0) && SizeVal >= 8)
6976 // Can't use NEON instructions.
6981 // Select the correct opcode and register class for unit size load/store
6982 bool IsNeon = UnitSize >= 8;
6983 TRC = (IsThumb1 || IsThumb2) ? (const TargetRegisterClass *)&ARM::tGPRRegClass
6984 : (const TargetRegisterClass *)&ARM::GPRRegClass;
6986 VecTRC = UnitSize == 16
6987 ? (const TargetRegisterClass *)&ARM::DPairRegClass
6989 ? (const TargetRegisterClass *)&ARM::DPRRegClass
6992 unsigned BytesLeft = SizeVal % UnitSize;
6993 unsigned LoopSize = SizeVal - BytesLeft;
6995 if (SizeVal <= Subtarget->getMaxInlineSizeThreshold()) {
6996 // Use LDR and STR to copy.
6997 // [scratch, srcOut] = LDR_POST(srcIn, UnitSize)
6998 // [destOut] = STR_POST(scratch, destIn, UnitSize)
6999 unsigned srcIn = src;
7000 unsigned destIn = dest;
7001 for (unsigned i = 0; i < LoopSize; i+=UnitSize) {
7002 unsigned srcOut = MRI.createVirtualRegister(TRC);
7003 unsigned destOut = MRI.createVirtualRegister(TRC);
7004 unsigned scratch = MRI.createVirtualRegister(IsNeon ? VecTRC : TRC);
7005 emitPostLd(BB, MI, TII, dl, UnitSize, scratch, srcIn, srcOut,
7006 IsThumb1, IsThumb2);
7007 emitPostSt(BB, MI, TII, dl, UnitSize, scratch, destIn, destOut,
7008 IsThumb1, IsThumb2);
7013 // Handle the leftover bytes with LDRB and STRB.
7014 // [scratch, srcOut] = LDRB_POST(srcIn, 1)
7015 // [destOut] = STRB_POST(scratch, destIn, 1)
7016 for (unsigned i = 0; i < BytesLeft; i++) {
7017 unsigned srcOut = MRI.createVirtualRegister(TRC);
7018 unsigned destOut = MRI.createVirtualRegister(TRC);
7019 unsigned scratch = MRI.createVirtualRegister(TRC);
7020 emitPostLd(BB, MI, TII, dl, 1, scratch, srcIn, srcOut,
7021 IsThumb1, IsThumb2);
7022 emitPostSt(BB, MI, TII, dl, 1, scratch, destIn, destOut,
7023 IsThumb1, IsThumb2);
7027 MI->eraseFromParent(); // The instruction is gone now.
7031 // Expand the pseudo op to a loop.
7034 // movw varEnd, # --> with thumb2
7036 // ldrcp varEnd, idx --> without thumb2
7037 // fallthrough --> loopMBB
7039 // PHI varPhi, varEnd, varLoop
7040 // PHI srcPhi, src, srcLoop
7041 // PHI destPhi, dst, destLoop
7042 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
7043 // [destLoop] = STR_POST(scratch, destPhi, UnitSize)
7044 // subs varLoop, varPhi, #UnitSize
7046 // fallthrough --> exitMBB
7048 // epilogue to handle left-over bytes
7049 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
7050 // [destOut] = STRB_POST(scratch, destLoop, 1)
7051 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
7052 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
7053 MF->insert(It, loopMBB);
7054 MF->insert(It, exitMBB);
7056 // Transfer the remainder of BB and its successor edges to exitMBB.
7057 exitMBB->splice(exitMBB->begin(), BB,
7058 std::next(MachineBasicBlock::iterator(MI)), BB->end());
7059 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
7061 // Load an immediate to varEnd.
7062 unsigned varEnd = MRI.createVirtualRegister(TRC);
7064 unsigned Vtmp = varEnd;
7065 if ((LoopSize & 0xFFFF0000) != 0)
7066 Vtmp = MRI.createVirtualRegister(TRC);
7067 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2MOVi16), Vtmp)
7068 .addImm(LoopSize & 0xFFFF));
7070 if ((LoopSize & 0xFFFF0000) != 0)
7071 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2MOVTi16), varEnd)
7072 .addReg(Vtmp).addImm(LoopSize >> 16));
7074 MachineConstantPool *ConstantPool = MF->getConstantPool();
7075 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
7076 const Constant *C = ConstantInt::get(Int32Ty, LoopSize);
7078 // MachineConstantPool wants an explicit alignment.
7079 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
7081 Align = getDataLayout()->getTypeAllocSize(C->getType());
7082 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
7085 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(ARM::tLDRpci)).addReg(
7086 varEnd, RegState::Define).addConstantPoolIndex(Idx));
7088 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(ARM::LDRcp)).addReg(
7089 varEnd, RegState::Define).addConstantPoolIndex(Idx).addImm(0));
7091 BB->addSuccessor(loopMBB);
7093 // Generate the loop body:
7094 // varPhi = PHI(varLoop, varEnd)
7095 // srcPhi = PHI(srcLoop, src)
7096 // destPhi = PHI(destLoop, dst)
7097 MachineBasicBlock *entryBB = BB;
7099 unsigned varLoop = MRI.createVirtualRegister(TRC);
7100 unsigned varPhi = MRI.createVirtualRegister(TRC);
7101 unsigned srcLoop = MRI.createVirtualRegister(TRC);
7102 unsigned srcPhi = MRI.createVirtualRegister(TRC);
7103 unsigned destLoop = MRI.createVirtualRegister(TRC);
7104 unsigned destPhi = MRI.createVirtualRegister(TRC);
7106 BuildMI(*BB, BB->begin(), dl, TII->get(ARM::PHI), varPhi)
7107 .addReg(varLoop).addMBB(loopMBB)
7108 .addReg(varEnd).addMBB(entryBB);
7109 BuildMI(BB, dl, TII->get(ARM::PHI), srcPhi)
7110 .addReg(srcLoop).addMBB(loopMBB)
7111 .addReg(src).addMBB(entryBB);
7112 BuildMI(BB, dl, TII->get(ARM::PHI), destPhi)
7113 .addReg(destLoop).addMBB(loopMBB)
7114 .addReg(dest).addMBB(entryBB);
7116 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
7117 // [destLoop] = STR_POST(scratch, destPhi, UnitSiz)
7118 unsigned scratch = MRI.createVirtualRegister(IsNeon ? VecTRC : TRC);
7119 emitPostLd(BB, BB->end(), TII, dl, UnitSize, scratch, srcPhi, srcLoop,
7120 IsThumb1, IsThumb2);
7121 emitPostSt(BB, BB->end(), TII, dl, UnitSize, scratch, destPhi, destLoop,
7122 IsThumb1, IsThumb2);
7124 // Decrement loop variable by UnitSize.
7126 MachineInstrBuilder MIB =
7127 BuildMI(*BB, BB->end(), dl, TII->get(ARM::tSUBi8), varLoop);
7128 MIB = AddDefaultT1CC(MIB);
7129 MIB.addReg(varPhi).addImm(UnitSize);
7130 AddDefaultPred(MIB);
7132 MachineInstrBuilder MIB =
7133 BuildMI(*BB, BB->end(), dl,
7134 TII->get(IsThumb2 ? ARM::t2SUBri : ARM::SUBri), varLoop);
7135 AddDefaultCC(AddDefaultPred(MIB.addReg(varPhi).addImm(UnitSize)));
7136 MIB->getOperand(5).setReg(ARM::CPSR);
7137 MIB->getOperand(5).setIsDef(true);
7139 BuildMI(*BB, BB->end(), dl,
7140 TII->get(IsThumb1 ? ARM::tBcc : IsThumb2 ? ARM::t2Bcc : ARM::Bcc))
7141 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
7143 // loopMBB can loop back to loopMBB or fall through to exitMBB.
7144 BB->addSuccessor(loopMBB);
7145 BB->addSuccessor(exitMBB);
7147 // Add epilogue to handle BytesLeft.
7149 MachineInstr *StartOfExit = exitMBB->begin();
7151 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
7152 // [destOut] = STRB_POST(scratch, destLoop, 1)
7153 unsigned srcIn = srcLoop;
7154 unsigned destIn = destLoop;
7155 for (unsigned i = 0; i < BytesLeft; i++) {
7156 unsigned srcOut = MRI.createVirtualRegister(TRC);
7157 unsigned destOut = MRI.createVirtualRegister(TRC);
7158 unsigned scratch = MRI.createVirtualRegister(TRC);
7159 emitPostLd(BB, StartOfExit, TII, dl, 1, scratch, srcIn, srcOut,
7160 IsThumb1, IsThumb2);
7161 emitPostSt(BB, StartOfExit, TII, dl, 1, scratch, destIn, destOut,
7162 IsThumb1, IsThumb2);
7167 MI->eraseFromParent(); // The instruction is gone now.
7172 ARMTargetLowering::EmitLowered__chkstk(MachineInstr *MI,
7173 MachineBasicBlock *MBB) const {
7174 const TargetMachine &TM = getTargetMachine();
7175 const TargetInstrInfo &TII = *TM.getInstrInfo();
7176 DebugLoc DL = MI->getDebugLoc();
7178 assert(Subtarget->isTargetWindows() &&
7179 "__chkstk is only supported on Windows");
7180 assert(Subtarget->isThumb2() && "Windows on ARM requires Thumb-2 mode");
7182 // __chkstk takes the number of words to allocate on the stack in R4, and
7183 // returns the stack adjustment in number of bytes in R4. This will not
7184 // clober any other registers (other than the obvious lr).
7186 // Although, technically, IP should be considered a register which may be
7187 // clobbered, the call itself will not touch it. Windows on ARM is a pure
7188 // thumb-2 environment, so there is no interworking required. As a result, we
7189 // do not expect a veneer to be emitted by the linker, clobbering IP.
7191 // Each module receives its own copy of __chkstk, so no import thunk is
7192 // required, again, ensuring that IP is not clobbered.
7194 // Finally, although some linkers may theoretically provide a trampoline for
7195 // out of range calls (which is quite common due to a 32M range limitation of
7196 // branches for Thumb), we can generate the long-call version via
7197 // -mcmodel=large, alleviating the need for the trampoline which may clobber
7200 switch (TM.getCodeModel()) {
7201 case CodeModel::Small:
7202 case CodeModel::Medium:
7203 case CodeModel::Default:
7204 case CodeModel::Kernel:
7205 BuildMI(*MBB, MI, DL, TII.get(ARM::tBL))
7206 .addImm((unsigned)ARMCC::AL).addReg(0)
7207 .addExternalSymbol("__chkstk")
7208 .addReg(ARM::R4, RegState::Implicit | RegState::Kill)
7209 .addReg(ARM::R4, RegState::Implicit | RegState::Define)
7210 .addReg(ARM::R12, RegState::Implicit | RegState::Define | RegState::Dead);
7212 case CodeModel::Large:
7213 case CodeModel::JITDefault: {
7214 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
7215 unsigned Reg = MRI.createVirtualRegister(&ARM::rGPRRegClass);
7217 BuildMI(*MBB, MI, DL, TII.get(ARM::t2MOVi32imm), Reg)
7218 .addExternalSymbol("__chkstk");
7219 BuildMI(*MBB, MI, DL, TII.get(ARM::tBLXr))
7220 .addImm((unsigned)ARMCC::AL).addReg(0)
7221 .addReg(Reg, RegState::Kill)
7222 .addReg(ARM::R4, RegState::Implicit | RegState::Kill)
7223 .addReg(ARM::R4, RegState::Implicit | RegState::Define)
7224 .addReg(ARM::R12, RegState::Implicit | RegState::Define | RegState::Dead);
7229 AddDefaultCC(AddDefaultPred(BuildMI(*MBB, MI, DL, TII.get(ARM::t2SUBrr),
7231 .addReg(ARM::SP).addReg(ARM::R4)));
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);
8452 /// PerformVMOVDRRCombine - Target-specific dag combine xforms for
8453 /// ARMISD::VMOVDRR. This is also used for BUILD_VECTORs with 2 operands.
8454 static SDValue PerformVMOVDRRCombine(SDNode *N, SelectionDAG &DAG) {
8455 // N=vmovrrd(X); vmovdrr(N:0, N:1) -> bit_convert(X)
8456 SDValue Op0 = N->getOperand(0);
8457 SDValue Op1 = N->getOperand(1);
8458 if (Op0.getOpcode() == ISD::BITCAST)
8459 Op0 = Op0.getOperand(0);
8460 if (Op1.getOpcode() == ISD::BITCAST)
8461 Op1 = Op1.getOperand(0);
8462 if (Op0.getOpcode() == ARMISD::VMOVRRD &&
8463 Op0.getNode() == Op1.getNode() &&
8464 Op0.getResNo() == 0 && Op1.getResNo() == 1)
8465 return DAG.getNode(ISD::BITCAST, SDLoc(N),
8466 N->getValueType(0), Op0.getOperand(0));
8470 /// PerformSTORECombine - Target-specific dag combine xforms for
8472 static SDValue PerformSTORECombine(SDNode *N,
8473 TargetLowering::DAGCombinerInfo &DCI) {
8474 StoreSDNode *St = cast<StoreSDNode>(N);
8475 if (St->isVolatile())
8478 // Optimize trunc store (of multiple scalars) to shuffle and store. First,
8479 // pack all of the elements in one place. Next, store to memory in fewer
8481 SDValue StVal = St->getValue();
8482 EVT VT = StVal.getValueType();
8483 if (St->isTruncatingStore() && VT.isVector()) {
8484 SelectionDAG &DAG = DCI.DAG;
8485 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8486 EVT StVT = St->getMemoryVT();
8487 unsigned NumElems = VT.getVectorNumElements();
8488 assert(StVT != VT && "Cannot truncate to the same type");
8489 unsigned FromEltSz = VT.getVectorElementType().getSizeInBits();
8490 unsigned ToEltSz = StVT.getVectorElementType().getSizeInBits();
8492 // From, To sizes and ElemCount must be pow of two
8493 if (!isPowerOf2_32(NumElems * FromEltSz * ToEltSz)) return SDValue();
8495 // We are going to use the original vector elt for storing.
8496 // Accumulated smaller vector elements must be a multiple of the store size.
8497 if (0 != (NumElems * FromEltSz) % ToEltSz) return SDValue();
8499 unsigned SizeRatio = FromEltSz / ToEltSz;
8500 assert(SizeRatio * NumElems * ToEltSz == VT.getSizeInBits());
8502 // Create a type on which we perform the shuffle.
8503 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(), StVT.getScalarType(),
8504 NumElems*SizeRatio);
8505 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
8508 SDValue WideVec = DAG.getNode(ISD::BITCAST, DL, WideVecVT, StVal);
8509 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
8510 for (unsigned i = 0; i < NumElems; ++i)
8511 ShuffleVec[i] = TLI.isBigEndian() ? (i+1) * SizeRatio - 1 : i * SizeRatio;
8513 // Can't shuffle using an illegal type.
8514 if (!TLI.isTypeLegal(WideVecVT)) return SDValue();
8516 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, DL, WideVec,
8517 DAG.getUNDEF(WideVec.getValueType()),
8519 // At this point all of the data is stored at the bottom of the
8520 // register. We now need to save it to mem.
8522 // Find the largest store unit
8523 MVT StoreType = MVT::i8;
8524 for (unsigned tp = MVT::FIRST_INTEGER_VALUETYPE;
8525 tp < MVT::LAST_INTEGER_VALUETYPE; ++tp) {
8526 MVT Tp = (MVT::SimpleValueType)tp;
8527 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToEltSz)
8530 // Didn't find a legal store type.
8531 if (!TLI.isTypeLegal(StoreType))
8534 // Bitcast the original vector into a vector of store-size units
8535 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
8536 StoreType, VT.getSizeInBits()/EVT(StoreType).getSizeInBits());
8537 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
8538 SDValue ShuffWide = DAG.getNode(ISD::BITCAST, DL, StoreVecVT, Shuff);
8539 SmallVector<SDValue, 8> Chains;
8540 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits()/8,
8541 TLI.getPointerTy());
8542 SDValue BasePtr = St->getBasePtr();
8544 // Perform one or more big stores into memory.
8545 unsigned E = (ToEltSz*NumElems)/StoreType.getSizeInBits();
8546 for (unsigned I = 0; I < E; I++) {
8547 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
8548 StoreType, ShuffWide,
8549 DAG.getIntPtrConstant(I));
8550 SDValue Ch = DAG.getStore(St->getChain(), DL, SubVec, BasePtr,
8551 St->getPointerInfo(), St->isVolatile(),
8552 St->isNonTemporal(), St->getAlignment());
8553 BasePtr = DAG.getNode(ISD::ADD, DL, BasePtr.getValueType(), BasePtr,
8555 Chains.push_back(Ch);
8557 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
8560 if (!ISD::isNormalStore(St))
8563 // Split a store of a VMOVDRR into two integer stores to avoid mixing NEON and
8564 // ARM stores of arguments in the same cache line.
8565 if (StVal.getNode()->getOpcode() == ARMISD::VMOVDRR &&
8566 StVal.getNode()->hasOneUse()) {
8567 SelectionDAG &DAG = DCI.DAG;
8568 bool isBigEndian = DAG.getTargetLoweringInfo().isBigEndian();
8570 SDValue BasePtr = St->getBasePtr();
8571 SDValue NewST1 = DAG.getStore(St->getChain(), DL,
8572 StVal.getNode()->getOperand(isBigEndian ? 1 : 0 ),
8573 BasePtr, St->getPointerInfo(), St->isVolatile(),
8574 St->isNonTemporal(), St->getAlignment());
8576 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
8577 DAG.getConstant(4, MVT::i32));
8578 return DAG.getStore(NewST1.getValue(0), DL,
8579 StVal.getNode()->getOperand(isBigEndian ? 0 : 1),
8580 OffsetPtr, St->getPointerInfo(), St->isVolatile(),
8581 St->isNonTemporal(),
8582 std::min(4U, St->getAlignment() / 2));
8585 if (StVal.getValueType() != MVT::i64 ||
8586 StVal.getNode()->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
8589 // Bitcast an i64 store extracted from a vector to f64.
8590 // Otherwise, the i64 value will be legalized to a pair of i32 values.
8591 SelectionDAG &DAG = DCI.DAG;
8593 SDValue IntVec = StVal.getOperand(0);
8594 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
8595 IntVec.getValueType().getVectorNumElements());
8596 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, IntVec);
8597 SDValue ExtElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
8598 Vec, StVal.getOperand(1));
8600 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ExtElt);
8601 // Make the DAGCombiner fold the bitcasts.
8602 DCI.AddToWorklist(Vec.getNode());
8603 DCI.AddToWorklist(ExtElt.getNode());
8604 DCI.AddToWorklist(V.getNode());
8605 return DAG.getStore(St->getChain(), dl, V, St->getBasePtr(),
8606 St->getPointerInfo(), St->isVolatile(),
8607 St->isNonTemporal(), St->getAlignment(),
8611 /// hasNormalLoadOperand - Check if any of the operands of a BUILD_VECTOR node
8612 /// are normal, non-volatile loads. If so, it is profitable to bitcast an
8613 /// i64 vector to have f64 elements, since the value can then be loaded
8614 /// directly into a VFP register.
8615 static bool hasNormalLoadOperand(SDNode *N) {
8616 unsigned NumElts = N->getValueType(0).getVectorNumElements();
8617 for (unsigned i = 0; i < NumElts; ++i) {
8618 SDNode *Elt = N->getOperand(i).getNode();
8619 if (ISD::isNormalLoad(Elt) && !cast<LoadSDNode>(Elt)->isVolatile())
8625 /// PerformBUILD_VECTORCombine - Target-specific dag combine xforms for
8626 /// ISD::BUILD_VECTOR.
8627 static SDValue PerformBUILD_VECTORCombine(SDNode *N,
8628 TargetLowering::DAGCombinerInfo &DCI){
8629 // build_vector(N=ARMISD::VMOVRRD(X), N:1) -> bit_convert(X):
8630 // VMOVRRD is introduced when legalizing i64 types. It forces the i64 value
8631 // into a pair of GPRs, which is fine when the value is used as a scalar,
8632 // but if the i64 value is converted to a vector, we need to undo the VMOVRRD.
8633 SelectionDAG &DAG = DCI.DAG;
8634 if (N->getNumOperands() == 2) {
8635 SDValue RV = PerformVMOVDRRCombine(N, DAG);
8640 // Load i64 elements as f64 values so that type legalization does not split
8641 // them up into i32 values.
8642 EVT VT = N->getValueType(0);
8643 if (VT.getVectorElementType() != MVT::i64 || !hasNormalLoadOperand(N))
8646 SmallVector<SDValue, 8> Ops;
8647 unsigned NumElts = VT.getVectorNumElements();
8648 for (unsigned i = 0; i < NumElts; ++i) {
8649 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(i));
8651 // Make the DAGCombiner fold the bitcast.
8652 DCI.AddToWorklist(V.getNode());
8654 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, NumElts);
8655 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, FloatVT, Ops);
8656 return DAG.getNode(ISD::BITCAST, dl, VT, BV);
8659 /// \brief Target-specific dag combine xforms for ARMISD::BUILD_VECTOR.
8661 PerformARMBUILD_VECTORCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
8662 // ARMISD::BUILD_VECTOR is introduced when legalizing ISD::BUILD_VECTOR.
8663 // At that time, we may have inserted bitcasts from integer to float.
8664 // If these bitcasts have survived DAGCombine, change the lowering of this
8665 // BUILD_VECTOR in something more vector friendly, i.e., that does not
8666 // force to use floating point types.
8668 // Make sure we can change the type of the vector.
8669 // This is possible iff:
8670 // 1. The vector is only used in a bitcast to a integer type. I.e.,
8671 // 1.1. Vector is used only once.
8672 // 1.2. Use is a bit convert to an integer type.
8673 // 2. The size of its operands are 32-bits (64-bits are not legal).
8674 EVT VT = N->getValueType(0);
8675 EVT EltVT = VT.getVectorElementType();
8677 // Check 1.1. and 2.
8678 if (EltVT.getSizeInBits() != 32 || !N->hasOneUse())
8681 // By construction, the input type must be float.
8682 assert(EltVT == MVT::f32 && "Unexpected type!");
8685 SDNode *Use = *N->use_begin();
8686 if (Use->getOpcode() != ISD::BITCAST ||
8687 Use->getValueType(0).isFloatingPoint())
8690 // Check profitability.
8691 // Model is, if more than half of the relevant operands are bitcast from
8692 // i32, turn the build_vector into a sequence of insert_vector_elt.
8693 // Relevant operands are everything that is not statically
8694 // (i.e., at compile time) bitcasted.
8695 unsigned NumOfBitCastedElts = 0;
8696 unsigned NumElts = VT.getVectorNumElements();
8697 unsigned NumOfRelevantElts = NumElts;
8698 for (unsigned Idx = 0; Idx < NumElts; ++Idx) {
8699 SDValue Elt = N->getOperand(Idx);
8700 if (Elt->getOpcode() == ISD::BITCAST) {
8701 // Assume only bit cast to i32 will go away.
8702 if (Elt->getOperand(0).getValueType() == MVT::i32)
8703 ++NumOfBitCastedElts;
8704 } else if (Elt.getOpcode() == ISD::UNDEF || isa<ConstantSDNode>(Elt))
8705 // Constants are statically casted, thus do not count them as
8706 // relevant operands.
8707 --NumOfRelevantElts;
8710 // Check if more than half of the elements require a non-free bitcast.
8711 if (NumOfBitCastedElts <= NumOfRelevantElts / 2)
8714 SelectionDAG &DAG = DCI.DAG;
8715 // Create the new vector type.
8716 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
8717 // Check if the type is legal.
8718 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8719 if (!TLI.isTypeLegal(VecVT))
8723 // ARMISD::BUILD_VECTOR E1, E2, ..., EN.
8724 // => BITCAST INSERT_VECTOR_ELT
8725 // (INSERT_VECTOR_ELT (...), (BITCAST EN-1), N-1),
8727 SDValue Vec = DAG.getUNDEF(VecVT);
8729 for (unsigned Idx = 0 ; Idx < NumElts; ++Idx) {
8730 SDValue V = N->getOperand(Idx);
8731 if (V.getOpcode() == ISD::UNDEF)
8733 if (V.getOpcode() == ISD::BITCAST &&
8734 V->getOperand(0).getValueType() == MVT::i32)
8735 // Fold obvious case.
8736 V = V.getOperand(0);
8738 V = DAG.getNode(ISD::BITCAST, SDLoc(V), MVT::i32, V);
8739 // Make the DAGCombiner fold the bitcasts.
8740 DCI.AddToWorklist(V.getNode());
8742 SDValue LaneIdx = DAG.getConstant(Idx, MVT::i32);
8743 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VecVT, Vec, V, LaneIdx);
8745 Vec = DAG.getNode(ISD::BITCAST, dl, VT, Vec);
8746 // Make the DAGCombiner fold the bitcasts.
8747 DCI.AddToWorklist(Vec.getNode());
8751 /// PerformInsertEltCombine - Target-specific dag combine xforms for
8752 /// ISD::INSERT_VECTOR_ELT.
8753 static SDValue PerformInsertEltCombine(SDNode *N,
8754 TargetLowering::DAGCombinerInfo &DCI) {
8755 // Bitcast an i64 load inserted into a vector to f64.
8756 // Otherwise, the i64 value will be legalized to a pair of i32 values.
8757 EVT VT = N->getValueType(0);
8758 SDNode *Elt = N->getOperand(1).getNode();
8759 if (VT.getVectorElementType() != MVT::i64 ||
8760 !ISD::isNormalLoad(Elt) || cast<LoadSDNode>(Elt)->isVolatile())
8763 SelectionDAG &DAG = DCI.DAG;
8765 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
8766 VT.getVectorNumElements());
8767 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, N->getOperand(0));
8768 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(1));
8769 // Make the DAGCombiner fold the bitcasts.
8770 DCI.AddToWorklist(Vec.getNode());
8771 DCI.AddToWorklist(V.getNode());
8772 SDValue InsElt = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, FloatVT,
8773 Vec, V, N->getOperand(2));
8774 return DAG.getNode(ISD::BITCAST, dl, VT, InsElt);
8777 /// PerformVECTOR_SHUFFLECombine - Target-specific dag combine xforms for
8778 /// ISD::VECTOR_SHUFFLE.
8779 static SDValue PerformVECTOR_SHUFFLECombine(SDNode *N, SelectionDAG &DAG) {
8780 // The LLVM shufflevector instruction does not require the shuffle mask
8781 // length to match the operand vector length, but ISD::VECTOR_SHUFFLE does
8782 // have that requirement. When translating to ISD::VECTOR_SHUFFLE, if the
8783 // operands do not match the mask length, they are extended by concatenating
8784 // them with undef vectors. That is probably the right thing for other
8785 // targets, but for NEON it is better to concatenate two double-register
8786 // size vector operands into a single quad-register size vector. Do that
8787 // transformation here:
8788 // shuffle(concat(v1, undef), concat(v2, undef)) ->
8789 // shuffle(concat(v1, v2), undef)
8790 SDValue Op0 = N->getOperand(0);
8791 SDValue Op1 = N->getOperand(1);
8792 if (Op0.getOpcode() != ISD::CONCAT_VECTORS ||
8793 Op1.getOpcode() != ISD::CONCAT_VECTORS ||
8794 Op0.getNumOperands() != 2 ||
8795 Op1.getNumOperands() != 2)
8797 SDValue Concat0Op1 = Op0.getOperand(1);
8798 SDValue Concat1Op1 = Op1.getOperand(1);
8799 if (Concat0Op1.getOpcode() != ISD::UNDEF ||
8800 Concat1Op1.getOpcode() != ISD::UNDEF)
8802 // Skip the transformation if any of the types are illegal.
8803 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8804 EVT VT = N->getValueType(0);
8805 if (!TLI.isTypeLegal(VT) ||
8806 !TLI.isTypeLegal(Concat0Op1.getValueType()) ||
8807 !TLI.isTypeLegal(Concat1Op1.getValueType()))
8810 SDValue NewConcat = DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT,
8811 Op0.getOperand(0), Op1.getOperand(0));
8812 // Translate the shuffle mask.
8813 SmallVector<int, 16> NewMask;
8814 unsigned NumElts = VT.getVectorNumElements();
8815 unsigned HalfElts = NumElts/2;
8816 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
8817 for (unsigned n = 0; n < NumElts; ++n) {
8818 int MaskElt = SVN->getMaskElt(n);
8820 if (MaskElt < (int)HalfElts)
8822 else if (MaskElt >= (int)NumElts && MaskElt < (int)(NumElts + HalfElts))
8823 NewElt = HalfElts + MaskElt - NumElts;
8824 NewMask.push_back(NewElt);
8826 return DAG.getVectorShuffle(VT, SDLoc(N), NewConcat,
8827 DAG.getUNDEF(VT), NewMask.data());
8830 /// CombineBaseUpdate - Target-specific DAG combine function for VLDDUP and
8831 /// NEON load/store intrinsics to merge base address updates.
8832 static SDValue CombineBaseUpdate(SDNode *N,
8833 TargetLowering::DAGCombinerInfo &DCI) {
8834 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
8837 SelectionDAG &DAG = DCI.DAG;
8838 bool isIntrinsic = (N->getOpcode() == ISD::INTRINSIC_VOID ||
8839 N->getOpcode() == ISD::INTRINSIC_W_CHAIN);
8840 unsigned AddrOpIdx = (isIntrinsic ? 2 : 1);
8841 SDValue Addr = N->getOperand(AddrOpIdx);
8843 // Search for a use of the address operand that is an increment.
8844 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(),
8845 UE = Addr.getNode()->use_end(); UI != UE; ++UI) {
8847 if (User->getOpcode() != ISD::ADD ||
8848 UI.getUse().getResNo() != Addr.getResNo())
8851 // Check that the add is independent of the load/store. Otherwise, folding
8852 // it would create a cycle.
8853 if (User->isPredecessorOf(N) || N->isPredecessorOf(User))
8856 // Find the new opcode for the updating load/store.
8858 bool isLaneOp = false;
8859 unsigned NewOpc = 0;
8860 unsigned NumVecs = 0;
8862 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
8864 default: llvm_unreachable("unexpected intrinsic for Neon base update");
8865 case Intrinsic::arm_neon_vld1: NewOpc = ARMISD::VLD1_UPD;
8867 case Intrinsic::arm_neon_vld2: NewOpc = ARMISD::VLD2_UPD;
8869 case Intrinsic::arm_neon_vld3: NewOpc = ARMISD::VLD3_UPD;
8871 case Intrinsic::arm_neon_vld4: NewOpc = ARMISD::VLD4_UPD;
8873 case Intrinsic::arm_neon_vld2lane: NewOpc = ARMISD::VLD2LN_UPD;
8874 NumVecs = 2; isLaneOp = true; break;
8875 case Intrinsic::arm_neon_vld3lane: NewOpc = ARMISD::VLD3LN_UPD;
8876 NumVecs = 3; isLaneOp = true; break;
8877 case Intrinsic::arm_neon_vld4lane: NewOpc = ARMISD::VLD4LN_UPD;
8878 NumVecs = 4; isLaneOp = true; break;
8879 case Intrinsic::arm_neon_vst1: NewOpc = ARMISD::VST1_UPD;
8880 NumVecs = 1; isLoad = false; break;
8881 case Intrinsic::arm_neon_vst2: NewOpc = ARMISD::VST2_UPD;
8882 NumVecs = 2; isLoad = false; break;
8883 case Intrinsic::arm_neon_vst3: NewOpc = ARMISD::VST3_UPD;
8884 NumVecs = 3; isLoad = false; break;
8885 case Intrinsic::arm_neon_vst4: NewOpc = ARMISD::VST4_UPD;
8886 NumVecs = 4; isLoad = false; break;
8887 case Intrinsic::arm_neon_vst2lane: NewOpc = ARMISD::VST2LN_UPD;
8888 NumVecs = 2; isLoad = false; isLaneOp = true; break;
8889 case Intrinsic::arm_neon_vst3lane: NewOpc = ARMISD::VST3LN_UPD;
8890 NumVecs = 3; isLoad = false; isLaneOp = true; break;
8891 case Intrinsic::arm_neon_vst4lane: NewOpc = ARMISD::VST4LN_UPD;
8892 NumVecs = 4; isLoad = false; isLaneOp = true; break;
8896 switch (N->getOpcode()) {
8897 default: llvm_unreachable("unexpected opcode for Neon base update");
8898 case ARMISD::VLD2DUP: NewOpc = ARMISD::VLD2DUP_UPD; NumVecs = 2; break;
8899 case ARMISD::VLD3DUP: NewOpc = ARMISD::VLD3DUP_UPD; NumVecs = 3; break;
8900 case ARMISD::VLD4DUP: NewOpc = ARMISD::VLD4DUP_UPD; NumVecs = 4; break;
8904 // Find the size of memory referenced by the load/store.
8907 VecTy = N->getValueType(0);
8909 VecTy = N->getOperand(AddrOpIdx+1).getValueType();
8910 unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8;
8912 NumBytes /= VecTy.getVectorNumElements();
8914 // If the increment is a constant, it must match the memory ref size.
8915 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
8916 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
8917 uint64_t IncVal = CInc->getZExtValue();
8918 if (IncVal != NumBytes)
8920 } else if (NumBytes >= 3 * 16) {
8921 // VLD3/4 and VST3/4 for 128-bit vectors are implemented with two
8922 // separate instructions that make it harder to use a non-constant update.
8926 // Create the new updating load/store node.
8928 unsigned NumResultVecs = (isLoad ? NumVecs : 0);
8930 for (n = 0; n < NumResultVecs; ++n)
8932 Tys[n++] = MVT::i32;
8933 Tys[n] = MVT::Other;
8934 SDVTList SDTys = DAG.getVTList(ArrayRef<EVT>(Tys, NumResultVecs+2));
8935 SmallVector<SDValue, 8> Ops;
8936 Ops.push_back(N->getOperand(0)); // incoming chain
8937 Ops.push_back(N->getOperand(AddrOpIdx));
8939 for (unsigned i = AddrOpIdx + 1; i < N->getNumOperands(); ++i) {
8940 Ops.push_back(N->getOperand(i));
8942 MemIntrinsicSDNode *MemInt = cast<MemIntrinsicSDNode>(N);
8943 SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, SDLoc(N), SDTys,
8944 Ops, MemInt->getMemoryVT(),
8945 MemInt->getMemOperand());
8948 std::vector<SDValue> NewResults;
8949 for (unsigned i = 0; i < NumResultVecs; ++i) {
8950 NewResults.push_back(SDValue(UpdN.getNode(), i));
8952 NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs+1)); // chain
8953 DCI.CombineTo(N, NewResults);
8954 DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs));
8961 /// CombineVLDDUP - For a VDUPLANE node N, check if its source operand is a
8962 /// vldN-lane (N > 1) intrinsic, and if all the other uses of that intrinsic
8963 /// are also VDUPLANEs. If so, combine them to a vldN-dup operation and
8965 static bool CombineVLDDUP(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
8966 SelectionDAG &DAG = DCI.DAG;
8967 EVT VT = N->getValueType(0);
8968 // vldN-dup instructions only support 64-bit vectors for N > 1.
8969 if (!VT.is64BitVector())
8972 // Check if the VDUPLANE operand is a vldN-dup intrinsic.
8973 SDNode *VLD = N->getOperand(0).getNode();
8974 if (VLD->getOpcode() != ISD::INTRINSIC_W_CHAIN)
8976 unsigned NumVecs = 0;
8977 unsigned NewOpc = 0;
8978 unsigned IntNo = cast<ConstantSDNode>(VLD->getOperand(1))->getZExtValue();
8979 if (IntNo == Intrinsic::arm_neon_vld2lane) {
8981 NewOpc = ARMISD::VLD2DUP;
8982 } else if (IntNo == Intrinsic::arm_neon_vld3lane) {
8984 NewOpc = ARMISD::VLD3DUP;
8985 } else if (IntNo == Intrinsic::arm_neon_vld4lane) {
8987 NewOpc = ARMISD::VLD4DUP;
8992 // First check that all the vldN-lane uses are VDUPLANEs and that the lane
8993 // numbers match the load.
8994 unsigned VLDLaneNo =
8995 cast<ConstantSDNode>(VLD->getOperand(NumVecs+3))->getZExtValue();
8996 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
8998 // Ignore uses of the chain result.
8999 if (UI.getUse().getResNo() == NumVecs)
9002 if (User->getOpcode() != ARMISD::VDUPLANE ||
9003 VLDLaneNo != cast<ConstantSDNode>(User->getOperand(1))->getZExtValue())
9007 // Create the vldN-dup node.
9010 for (n = 0; n < NumVecs; ++n)
9012 Tys[n] = MVT::Other;
9013 SDVTList SDTys = DAG.getVTList(ArrayRef<EVT>(Tys, NumVecs+1));
9014 SDValue Ops[] = { VLD->getOperand(0), VLD->getOperand(2) };
9015 MemIntrinsicSDNode *VLDMemInt = cast<MemIntrinsicSDNode>(VLD);
9016 SDValue VLDDup = DAG.getMemIntrinsicNode(NewOpc, SDLoc(VLD), SDTys,
9017 Ops, VLDMemInt->getMemoryVT(),
9018 VLDMemInt->getMemOperand());
9021 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
9023 unsigned ResNo = UI.getUse().getResNo();
9024 // Ignore uses of the chain result.
9025 if (ResNo == NumVecs)
9028 DCI.CombineTo(User, SDValue(VLDDup.getNode(), ResNo));
9031 // Now the vldN-lane intrinsic is dead except for its chain result.
9032 // Update uses of the chain.
9033 std::vector<SDValue> VLDDupResults;
9034 for (unsigned n = 0; n < NumVecs; ++n)
9035 VLDDupResults.push_back(SDValue(VLDDup.getNode(), n));
9036 VLDDupResults.push_back(SDValue(VLDDup.getNode(), NumVecs));
9037 DCI.CombineTo(VLD, VLDDupResults);
9042 /// PerformVDUPLANECombine - Target-specific dag combine xforms for
9043 /// ARMISD::VDUPLANE.
9044 static SDValue PerformVDUPLANECombine(SDNode *N,
9045 TargetLowering::DAGCombinerInfo &DCI) {
9046 SDValue Op = N->getOperand(0);
9048 // If the source is a vldN-lane (N > 1) intrinsic, and all the other uses
9049 // of that intrinsic are also VDUPLANEs, combine them to a vldN-dup operation.
9050 if (CombineVLDDUP(N, DCI))
9051 return SDValue(N, 0);
9053 // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is
9054 // redundant. Ignore bit_converts for now; element sizes are checked below.
9055 while (Op.getOpcode() == ISD::BITCAST)
9056 Op = Op.getOperand(0);
9057 if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM)
9060 // Make sure the VMOV element size is not bigger than the VDUPLANE elements.
9061 unsigned EltSize = Op.getValueType().getVectorElementType().getSizeInBits();
9062 // The canonical VMOV for a zero vector uses a 32-bit element size.
9063 unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
9065 if (ARM_AM::decodeNEONModImm(Imm, EltBits) == 0)
9067 EVT VT = N->getValueType(0);
9068 if (EltSize > VT.getVectorElementType().getSizeInBits())
9071 return DCI.DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Op);
9074 // isConstVecPow2 - Return true if each vector element is a power of 2, all
9075 // elements are the same constant, C, and Log2(C) ranges from 1 to 32.
9076 static bool isConstVecPow2(SDValue ConstVec, bool isSigned, uint64_t &C)
9080 for (unsigned I = 0, E = ConstVec.getValueType().getVectorNumElements();
9082 ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(ConstVec.getOperand(I));
9087 APFloat APF = C->getValueAPF();
9088 if (APF.convertToInteger(&cN, 64, isSigned, APFloat::rmTowardZero, &isExact)
9089 != APFloat::opOK || !isExact)
9092 c0 = (I == 0) ? cN : c0;
9093 if (!isPowerOf2_64(cN) || c0 != cN || Log2_64(c0) < 1 || Log2_64(c0) > 32)
9100 /// PerformVCVTCombine - VCVT (floating-point to fixed-point, Advanced SIMD)
9101 /// can replace combinations of VMUL and VCVT (floating-point to integer)
9102 /// when the VMUL has a constant operand that is a power of 2.
9104 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
9105 /// vmul.f32 d16, d17, d16
9106 /// vcvt.s32.f32 d16, d16
9108 /// vcvt.s32.f32 d16, d16, #3
9109 static SDValue PerformVCVTCombine(SDNode *N,
9110 TargetLowering::DAGCombinerInfo &DCI,
9111 const ARMSubtarget *Subtarget) {
9112 SelectionDAG &DAG = DCI.DAG;
9113 SDValue Op = N->getOperand(0);
9115 if (!Subtarget->hasNEON() || !Op.getValueType().isVector() ||
9116 Op.getOpcode() != ISD::FMUL)
9120 SDValue N0 = Op->getOperand(0);
9121 SDValue ConstVec = Op->getOperand(1);
9122 bool isSigned = N->getOpcode() == ISD::FP_TO_SINT;
9124 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
9125 !isConstVecPow2(ConstVec, isSigned, C))
9128 MVT FloatTy = Op.getSimpleValueType().getVectorElementType();
9129 MVT IntTy = N->getSimpleValueType(0).getVectorElementType();
9130 if (FloatTy.getSizeInBits() != 32 || IntTy.getSizeInBits() > 32) {
9131 // These instructions only exist converting from f32 to i32. We can handle
9132 // smaller integers by generating an extra truncate, but larger ones would
9137 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfp2fxs :
9138 Intrinsic::arm_neon_vcvtfp2fxu;
9139 unsigned NumLanes = Op.getValueType().getVectorNumElements();
9140 SDValue FixConv = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N),
9141 NumLanes == 2 ? MVT::v2i32 : MVT::v4i32,
9142 DAG.getConstant(IntrinsicOpcode, MVT::i32), N0,
9143 DAG.getConstant(Log2_64(C), MVT::i32));
9145 if (IntTy.getSizeInBits() < FloatTy.getSizeInBits())
9146 FixConv = DAG.getNode(ISD::TRUNCATE, SDLoc(N), N->getValueType(0), FixConv);
9151 /// PerformVDIVCombine - VCVT (fixed-point to floating-point, Advanced SIMD)
9152 /// can replace combinations of VCVT (integer to floating-point) and VDIV
9153 /// when the VDIV has a constant operand that is a power of 2.
9155 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
9156 /// vcvt.f32.s32 d16, d16
9157 /// vdiv.f32 d16, d17, d16
9159 /// vcvt.f32.s32 d16, d16, #3
9160 static SDValue PerformVDIVCombine(SDNode *N,
9161 TargetLowering::DAGCombinerInfo &DCI,
9162 const ARMSubtarget *Subtarget) {
9163 SelectionDAG &DAG = DCI.DAG;
9164 SDValue Op = N->getOperand(0);
9165 unsigned OpOpcode = Op.getNode()->getOpcode();
9167 if (!Subtarget->hasNEON() || !N->getValueType(0).isVector() ||
9168 (OpOpcode != ISD::SINT_TO_FP && OpOpcode != ISD::UINT_TO_FP))
9172 SDValue ConstVec = N->getOperand(1);
9173 bool isSigned = OpOpcode == ISD::SINT_TO_FP;
9175 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
9176 !isConstVecPow2(ConstVec, isSigned, C))
9179 MVT FloatTy = N->getSimpleValueType(0).getVectorElementType();
9180 MVT IntTy = Op.getOperand(0).getSimpleValueType().getVectorElementType();
9181 if (FloatTy.getSizeInBits() != 32 || IntTy.getSizeInBits() > 32) {
9182 // These instructions only exist converting from i32 to f32. We can handle
9183 // smaller integers by generating an extra extend, but larger ones would
9188 SDValue ConvInput = Op.getOperand(0);
9189 unsigned NumLanes = Op.getValueType().getVectorNumElements();
9190 if (IntTy.getSizeInBits() < FloatTy.getSizeInBits())
9191 ConvInput = DAG.getNode(isSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND,
9192 SDLoc(N), NumLanes == 2 ? MVT::v2i32 : MVT::v4i32,
9195 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfxs2fp :
9196 Intrinsic::arm_neon_vcvtfxu2fp;
9197 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N),
9199 DAG.getConstant(IntrinsicOpcode, MVT::i32),
9200 ConvInput, DAG.getConstant(Log2_64(C), MVT::i32));
9203 /// Getvshiftimm - Check if this is a valid build_vector for the immediate
9204 /// operand of a vector shift operation, where all the elements of the
9205 /// build_vector must have the same constant integer value.
9206 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
9207 // Ignore bit_converts.
9208 while (Op.getOpcode() == ISD::BITCAST)
9209 Op = Op.getOperand(0);
9210 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
9211 APInt SplatBits, SplatUndef;
9212 unsigned SplatBitSize;
9214 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
9215 HasAnyUndefs, ElementBits) ||
9216 SplatBitSize > ElementBits)
9218 Cnt = SplatBits.getSExtValue();
9222 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
9223 /// operand of a vector shift left operation. That value must be in the range:
9224 /// 0 <= Value < ElementBits for a left shift; or
9225 /// 0 <= Value <= ElementBits for a long left shift.
9226 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
9227 assert(VT.isVector() && "vector shift count is not a vector type");
9228 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
9229 if (! getVShiftImm(Op, ElementBits, Cnt))
9231 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
9234 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
9235 /// operand of a vector shift right operation. For a shift opcode, the value
9236 /// is positive, but for an intrinsic the value count must be negative. The
9237 /// absolute value must be in the range:
9238 /// 1 <= |Value| <= ElementBits for a right shift; or
9239 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
9240 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
9242 assert(VT.isVector() && "vector shift count is not a vector type");
9243 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
9244 if (! getVShiftImm(Op, ElementBits, Cnt))
9248 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
9251 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
9252 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
9253 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
9256 // Don't do anything for most intrinsics.
9259 // Vector shifts: check for immediate versions and lower them.
9260 // Note: This is done during DAG combining instead of DAG legalizing because
9261 // the build_vectors for 64-bit vector element shift counts are generally
9262 // not legal, and it is hard to see their values after they get legalized to
9263 // loads from a constant pool.
9264 case Intrinsic::arm_neon_vshifts:
9265 case Intrinsic::arm_neon_vshiftu:
9266 case Intrinsic::arm_neon_vrshifts:
9267 case Intrinsic::arm_neon_vrshiftu:
9268 case Intrinsic::arm_neon_vrshiftn:
9269 case Intrinsic::arm_neon_vqshifts:
9270 case Intrinsic::arm_neon_vqshiftu:
9271 case Intrinsic::arm_neon_vqshiftsu:
9272 case Intrinsic::arm_neon_vqshiftns:
9273 case Intrinsic::arm_neon_vqshiftnu:
9274 case Intrinsic::arm_neon_vqshiftnsu:
9275 case Intrinsic::arm_neon_vqrshiftns:
9276 case Intrinsic::arm_neon_vqrshiftnu:
9277 case Intrinsic::arm_neon_vqrshiftnsu: {
9278 EVT VT = N->getOperand(1).getValueType();
9280 unsigned VShiftOpc = 0;
9283 case Intrinsic::arm_neon_vshifts:
9284 case Intrinsic::arm_neon_vshiftu:
9285 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
9286 VShiftOpc = ARMISD::VSHL;
9289 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
9290 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
9291 ARMISD::VSHRs : ARMISD::VSHRu);
9296 case Intrinsic::arm_neon_vrshifts:
9297 case Intrinsic::arm_neon_vrshiftu:
9298 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
9302 case Intrinsic::arm_neon_vqshifts:
9303 case Intrinsic::arm_neon_vqshiftu:
9304 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
9308 case Intrinsic::arm_neon_vqshiftsu:
9309 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
9311 llvm_unreachable("invalid shift count for vqshlu intrinsic");
9313 case Intrinsic::arm_neon_vrshiftn:
9314 case Intrinsic::arm_neon_vqshiftns:
9315 case Intrinsic::arm_neon_vqshiftnu:
9316 case Intrinsic::arm_neon_vqshiftnsu:
9317 case Intrinsic::arm_neon_vqrshiftns:
9318 case Intrinsic::arm_neon_vqrshiftnu:
9319 case Intrinsic::arm_neon_vqrshiftnsu:
9320 // Narrowing shifts require an immediate right shift.
9321 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
9323 llvm_unreachable("invalid shift count for narrowing vector shift "
9327 llvm_unreachable("unhandled vector shift");
9331 case Intrinsic::arm_neon_vshifts:
9332 case Intrinsic::arm_neon_vshiftu:
9333 // Opcode already set above.
9335 case Intrinsic::arm_neon_vrshifts:
9336 VShiftOpc = ARMISD::VRSHRs; break;
9337 case Intrinsic::arm_neon_vrshiftu:
9338 VShiftOpc = ARMISD::VRSHRu; break;
9339 case Intrinsic::arm_neon_vrshiftn:
9340 VShiftOpc = ARMISD::VRSHRN; break;
9341 case Intrinsic::arm_neon_vqshifts:
9342 VShiftOpc = ARMISD::VQSHLs; break;
9343 case Intrinsic::arm_neon_vqshiftu:
9344 VShiftOpc = ARMISD::VQSHLu; break;
9345 case Intrinsic::arm_neon_vqshiftsu:
9346 VShiftOpc = ARMISD::VQSHLsu; break;
9347 case Intrinsic::arm_neon_vqshiftns:
9348 VShiftOpc = ARMISD::VQSHRNs; break;
9349 case Intrinsic::arm_neon_vqshiftnu:
9350 VShiftOpc = ARMISD::VQSHRNu; break;
9351 case Intrinsic::arm_neon_vqshiftnsu:
9352 VShiftOpc = ARMISD::VQSHRNsu; break;
9353 case Intrinsic::arm_neon_vqrshiftns:
9354 VShiftOpc = ARMISD::VQRSHRNs; break;
9355 case Intrinsic::arm_neon_vqrshiftnu:
9356 VShiftOpc = ARMISD::VQRSHRNu; break;
9357 case Intrinsic::arm_neon_vqrshiftnsu:
9358 VShiftOpc = ARMISD::VQRSHRNsu; break;
9361 return DAG.getNode(VShiftOpc, SDLoc(N), N->getValueType(0),
9362 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
9365 case Intrinsic::arm_neon_vshiftins: {
9366 EVT VT = N->getOperand(1).getValueType();
9368 unsigned VShiftOpc = 0;
9370 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
9371 VShiftOpc = ARMISD::VSLI;
9372 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
9373 VShiftOpc = ARMISD::VSRI;
9375 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
9378 return DAG.getNode(VShiftOpc, SDLoc(N), N->getValueType(0),
9379 N->getOperand(1), N->getOperand(2),
9380 DAG.getConstant(Cnt, MVT::i32));
9383 case Intrinsic::arm_neon_vqrshifts:
9384 case Intrinsic::arm_neon_vqrshiftu:
9385 // No immediate versions of these to check for.
9392 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
9393 /// lowers them. As with the vector shift intrinsics, this is done during DAG
9394 /// combining instead of DAG legalizing because the build_vectors for 64-bit
9395 /// vector element shift counts are generally not legal, and it is hard to see
9396 /// their values after they get legalized to loads from a constant pool.
9397 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
9398 const ARMSubtarget *ST) {
9399 EVT VT = N->getValueType(0);
9400 if (N->getOpcode() == ISD::SRL && VT == MVT::i32 && ST->hasV6Ops()) {
9401 // Canonicalize (srl (bswap x), 16) to (rotr (bswap x), 16) if the high
9402 // 16-bits of x is zero. This optimizes rev + lsr 16 to rev16.
9403 SDValue N1 = N->getOperand(1);
9404 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) {
9405 SDValue N0 = N->getOperand(0);
9406 if (C->getZExtValue() == 16 && N0.getOpcode() == ISD::BSWAP &&
9407 DAG.MaskedValueIsZero(N0.getOperand(0),
9408 APInt::getHighBitsSet(32, 16)))
9409 return DAG.getNode(ISD::ROTR, SDLoc(N), VT, N0, N1);
9413 // Nothing to be done for scalar shifts.
9414 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9415 if (!VT.isVector() || !TLI.isTypeLegal(VT))
9418 assert(ST->hasNEON() && "unexpected vector shift");
9421 switch (N->getOpcode()) {
9422 default: llvm_unreachable("unexpected shift opcode");
9425 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
9426 return DAG.getNode(ARMISD::VSHL, SDLoc(N), VT, N->getOperand(0),
9427 DAG.getConstant(Cnt, MVT::i32));
9432 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
9433 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
9434 ARMISD::VSHRs : ARMISD::VSHRu);
9435 return DAG.getNode(VShiftOpc, SDLoc(N), VT, N->getOperand(0),
9436 DAG.getConstant(Cnt, MVT::i32));
9442 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
9443 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
9444 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
9445 const ARMSubtarget *ST) {
9446 SDValue N0 = N->getOperand(0);
9448 // Check for sign- and zero-extensions of vector extract operations of 8-
9449 // and 16-bit vector elements. NEON supports these directly. They are
9450 // handled during DAG combining because type legalization will promote them
9451 // to 32-bit types and it is messy to recognize the operations after that.
9452 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
9453 SDValue Vec = N0.getOperand(0);
9454 SDValue Lane = N0.getOperand(1);
9455 EVT VT = N->getValueType(0);
9456 EVT EltVT = N0.getValueType();
9457 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9459 if (VT == MVT::i32 &&
9460 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
9461 TLI.isTypeLegal(Vec.getValueType()) &&
9462 isa<ConstantSDNode>(Lane)) {
9465 switch (N->getOpcode()) {
9466 default: llvm_unreachable("unexpected opcode");
9467 case ISD::SIGN_EXTEND:
9468 Opc = ARMISD::VGETLANEs;
9470 case ISD::ZERO_EXTEND:
9471 case ISD::ANY_EXTEND:
9472 Opc = ARMISD::VGETLANEu;
9475 return DAG.getNode(Opc, SDLoc(N), VT, Vec, Lane);
9482 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
9483 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
9484 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
9485 const ARMSubtarget *ST) {
9486 // If the target supports NEON, try to use vmax/vmin instructions for f32
9487 // selects like "x < y ? x : y". Unless the NoNaNsFPMath option is set,
9488 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
9489 // a NaN; only do the transformation when it matches that behavior.
9491 // For now only do this when using NEON for FP operations; if using VFP, it
9492 // is not obvious that the benefit outweighs the cost of switching to the
9494 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
9495 N->getValueType(0) != MVT::f32)
9498 SDValue CondLHS = N->getOperand(0);
9499 SDValue CondRHS = N->getOperand(1);
9500 SDValue LHS = N->getOperand(2);
9501 SDValue RHS = N->getOperand(3);
9502 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
9504 unsigned Opcode = 0;
9506 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
9507 IsReversed = false; // x CC y ? x : y
9508 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
9509 IsReversed = true ; // x CC y ? y : x
9523 // If LHS is NaN, an ordered comparison will be false and the result will
9524 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
9525 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
9526 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
9527 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
9529 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
9530 // will return -0, so vmin can only be used for unsafe math or if one of
9531 // the operands is known to be nonzero.
9532 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
9533 !DAG.getTarget().Options.UnsafeFPMath &&
9534 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9536 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
9545 // If LHS is NaN, an ordered comparison will be false and the result will
9546 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
9547 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
9548 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
9549 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
9551 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
9552 // will return +0, so vmax can only be used for unsafe math or if one of
9553 // the operands is known to be nonzero.
9554 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
9555 !DAG.getTarget().Options.UnsafeFPMath &&
9556 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9558 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
9564 return DAG.getNode(Opcode, SDLoc(N), N->getValueType(0), LHS, RHS);
9567 /// PerformCMOVCombine - Target-specific DAG combining for ARMISD::CMOV.
9569 ARMTargetLowering::PerformCMOVCombine(SDNode *N, SelectionDAG &DAG) const {
9570 SDValue Cmp = N->getOperand(4);
9571 if (Cmp.getOpcode() != ARMISD::CMPZ)
9572 // Only looking at EQ and NE cases.
9575 EVT VT = N->getValueType(0);
9577 SDValue LHS = Cmp.getOperand(0);
9578 SDValue RHS = Cmp.getOperand(1);
9579 SDValue FalseVal = N->getOperand(0);
9580 SDValue TrueVal = N->getOperand(1);
9581 SDValue ARMcc = N->getOperand(2);
9582 ARMCC::CondCodes CC =
9583 (ARMCC::CondCodes)cast<ConstantSDNode>(ARMcc)->getZExtValue();
9601 /// FIXME: Turn this into a target neutral optimization?
9603 if (CC == ARMCC::NE && FalseVal == RHS && FalseVal != LHS) {
9604 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, TrueVal, ARMcc,
9605 N->getOperand(3), Cmp);
9606 } else if (CC == ARMCC::EQ && TrueVal == RHS) {
9608 SDValue NewCmp = getARMCmp(LHS, RHS, ISD::SETNE, ARMcc, DAG, dl);
9609 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, FalseVal, ARMcc,
9610 N->getOperand(3), NewCmp);
9613 if (Res.getNode()) {
9614 APInt KnownZero, KnownOne;
9615 DAG.computeKnownBits(SDValue(N,0), KnownZero, KnownOne);
9616 // Capture demanded bits information that would be otherwise lost.
9617 if (KnownZero == 0xfffffffe)
9618 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
9619 DAG.getValueType(MVT::i1));
9620 else if (KnownZero == 0xffffff00)
9621 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
9622 DAG.getValueType(MVT::i8));
9623 else if (KnownZero == 0xffff0000)
9624 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
9625 DAG.getValueType(MVT::i16));
9631 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
9632 DAGCombinerInfo &DCI) const {
9633 switch (N->getOpcode()) {
9635 case ISD::ADDC: return PerformADDCCombine(N, DCI, Subtarget);
9636 case ISD::ADD: return PerformADDCombine(N, DCI, Subtarget);
9637 case ISD::SUB: return PerformSUBCombine(N, DCI);
9638 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
9639 case ISD::OR: return PerformORCombine(N, DCI, Subtarget);
9640 case ISD::XOR: return PerformXORCombine(N, DCI, Subtarget);
9641 case ISD::AND: return PerformANDCombine(N, DCI, Subtarget);
9642 case ARMISD::BFI: return PerformBFICombine(N, DCI);
9643 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
9644 case ARMISD::VMOVDRR: return PerformVMOVDRRCombine(N, DCI.DAG);
9645 case ISD::STORE: return PerformSTORECombine(N, DCI);
9646 case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DCI);
9647 case ISD::INSERT_VECTOR_ELT: return PerformInsertEltCombine(N, DCI);
9648 case ISD::VECTOR_SHUFFLE: return PerformVECTOR_SHUFFLECombine(N, DCI.DAG);
9649 case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI);
9650 case ISD::FP_TO_SINT:
9651 case ISD::FP_TO_UINT: return PerformVCVTCombine(N, DCI, Subtarget);
9652 case ISD::FDIV: return PerformVDIVCombine(N, DCI, Subtarget);
9653 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
9656 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
9657 case ISD::SIGN_EXTEND:
9658 case ISD::ZERO_EXTEND:
9659 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
9660 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
9661 case ARMISD::CMOV: return PerformCMOVCombine(N, DCI.DAG);
9662 case ARMISD::VLD2DUP:
9663 case ARMISD::VLD3DUP:
9664 case ARMISD::VLD4DUP:
9665 return CombineBaseUpdate(N, DCI);
9666 case ARMISD::BUILD_VECTOR:
9667 return PerformARMBUILD_VECTORCombine(N, DCI);
9668 case ISD::INTRINSIC_VOID:
9669 case ISD::INTRINSIC_W_CHAIN:
9670 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
9671 case Intrinsic::arm_neon_vld1:
9672 case Intrinsic::arm_neon_vld2:
9673 case Intrinsic::arm_neon_vld3:
9674 case Intrinsic::arm_neon_vld4:
9675 case Intrinsic::arm_neon_vld2lane:
9676 case Intrinsic::arm_neon_vld3lane:
9677 case Intrinsic::arm_neon_vld4lane:
9678 case Intrinsic::arm_neon_vst1:
9679 case Intrinsic::arm_neon_vst2:
9680 case Intrinsic::arm_neon_vst3:
9681 case Intrinsic::arm_neon_vst4:
9682 case Intrinsic::arm_neon_vst2lane:
9683 case Intrinsic::arm_neon_vst3lane:
9684 case Intrinsic::arm_neon_vst4lane:
9685 return CombineBaseUpdate(N, DCI);
9693 bool ARMTargetLowering::isDesirableToTransformToIntegerOp(unsigned Opc,
9695 return (VT == MVT::f32) && (Opc == ISD::LOAD || Opc == ISD::STORE);
9698 bool ARMTargetLowering::allowsMisalignedMemoryAccesses(EVT VT,
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 (allowsMisalignedMemoryAccesses(MVT::v2f64, 0, 1, &Fast) && Fast))) {
9757 } else if (Size >= 8 &&
9758 (memOpAlign(SrcAlign, DstAlign, 8) ||
9759 (allowsMisalignedMemoryAccesses(MVT::f64, 0, 1, &Fast) &&
9765 // Lowering to i32/i16 if the size permits.
9771 // Let the target-independent logic figure it out.
9775 bool ARMTargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
9776 if (Val.getOpcode() != ISD::LOAD)
9779 EVT VT1 = Val.getValueType();
9780 if (!VT1.isSimple() || !VT1.isInteger() ||
9781 !VT2.isSimple() || !VT2.isInteger())
9784 switch (VT1.getSimpleVT().SimpleTy) {
9789 // 8-bit and 16-bit loads implicitly zero-extend to 32-bits.
9796 bool ARMTargetLowering::allowTruncateForTailCall(Type *Ty1, Type *Ty2) const {
9797 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
9800 if (!isTypeLegal(EVT::getEVT(Ty1)))
9803 assert(Ty1->getPrimitiveSizeInBits() <= 64 && "i128 is probably not a noop");
9805 // Assuming the caller doesn't have a zeroext or signext return parameter,
9806 // truncation all the way down to i1 is valid.
9811 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
9816 switch (VT.getSimpleVT().SimpleTy) {
9817 default: return false;
9832 if ((V & (Scale - 1)) != 0)
9835 return V == (V & ((1LL << 5) - 1));
9838 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
9839 const ARMSubtarget *Subtarget) {
9846 switch (VT.getSimpleVT().SimpleTy) {
9847 default: return false;
9852 // + imm12 or - imm8
9854 return V == (V & ((1LL << 8) - 1));
9855 return V == (V & ((1LL << 12) - 1));
9858 // Same as ARM mode. FIXME: NEON?
9859 if (!Subtarget->hasVFP2())
9864 return V == (V & ((1LL << 8) - 1));
9868 /// isLegalAddressImmediate - Return true if the integer value can be used
9869 /// as the offset of the target addressing mode for load / store of the
9871 static bool isLegalAddressImmediate(int64_t V, EVT VT,
9872 const ARMSubtarget *Subtarget) {
9879 if (Subtarget->isThumb1Only())
9880 return isLegalT1AddressImmediate(V, VT);
9881 else if (Subtarget->isThumb2())
9882 return isLegalT2AddressImmediate(V, VT, Subtarget);
9887 switch (VT.getSimpleVT().SimpleTy) {
9888 default: return false;
9893 return V == (V & ((1LL << 12) - 1));
9896 return V == (V & ((1LL << 8) - 1));
9899 if (!Subtarget->hasVFP2()) // FIXME: NEON?
9904 return V == (V & ((1LL << 8) - 1));
9908 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
9910 int Scale = AM.Scale;
9914 switch (VT.getSimpleVT().SimpleTy) {
9915 default: return false;
9924 return Scale == 2 || Scale == 4 || Scale == 8;
9927 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
9931 // Note, we allow "void" uses (basically, uses that aren't loads or
9932 // stores), because arm allows folding a scale into many arithmetic
9933 // operations. This should be made more precise and revisited later.
9935 // Allow r << imm, but the imm has to be a multiple of two.
9936 if (Scale & 1) return false;
9937 return isPowerOf2_32(Scale);
9941 /// isLegalAddressingMode - Return true if the addressing mode represented
9942 /// by AM is legal for this target, for a load/store of the specified type.
9943 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
9945 EVT VT = getValueType(Ty, true);
9946 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
9949 // Can never fold addr of global into load/store.
9954 case 0: // no scale reg, must be "r+i" or "r", or "i".
9957 if (Subtarget->isThumb1Only())
9961 // ARM doesn't support any R+R*scale+imm addr modes.
9968 if (Subtarget->isThumb2())
9969 return isLegalT2ScaledAddressingMode(AM, VT);
9971 int Scale = AM.Scale;
9972 switch (VT.getSimpleVT().SimpleTy) {
9973 default: return false;
9977 if (Scale < 0) Scale = -Scale;
9981 return isPowerOf2_32(Scale & ~1);
9985 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
9990 // Note, we allow "void" uses (basically, uses that aren't loads or
9991 // stores), because arm allows folding a scale into many arithmetic
9992 // operations. This should be made more precise and revisited later.
9994 // Allow r << imm, but the imm has to be a multiple of two.
9995 if (Scale & 1) return false;
9996 return isPowerOf2_32(Scale);
10002 /// isLegalICmpImmediate - Return true if the specified immediate is legal
10003 /// icmp immediate, that is the target has icmp instructions which can compare
10004 /// a register against the immediate without having to materialize the
10005 /// immediate into a register.
10006 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
10007 // Thumb2 and ARM modes can use cmn for negative immediates.
10008 if (!Subtarget->isThumb())
10009 return ARM_AM::getSOImmVal(llvm::abs64(Imm)) != -1;
10010 if (Subtarget->isThumb2())
10011 return ARM_AM::getT2SOImmVal(llvm::abs64(Imm)) != -1;
10012 // Thumb1 doesn't have cmn, and only 8-bit immediates.
10013 return Imm >= 0 && Imm <= 255;
10016 /// isLegalAddImmediate - Return true if the specified immediate is a legal add
10017 /// *or sub* immediate, that is the target has add or sub instructions which can
10018 /// add a register with the immediate without having to materialize the
10019 /// immediate into a register.
10020 bool ARMTargetLowering::isLegalAddImmediate(int64_t Imm) const {
10021 // Same encoding for add/sub, just flip the sign.
10022 int64_t AbsImm = llvm::abs64(Imm);
10023 if (!Subtarget->isThumb())
10024 return ARM_AM::getSOImmVal(AbsImm) != -1;
10025 if (Subtarget->isThumb2())
10026 return ARM_AM::getT2SOImmVal(AbsImm) != -1;
10027 // Thumb1 only has 8-bit unsigned immediate.
10028 return AbsImm >= 0 && AbsImm <= 255;
10031 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
10032 bool isSEXTLoad, SDValue &Base,
10033 SDValue &Offset, bool &isInc,
10034 SelectionDAG &DAG) {
10035 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
10038 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
10039 // AddressingMode 3
10040 Base = Ptr->getOperand(0);
10041 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
10042 int RHSC = (int)RHS->getZExtValue();
10043 if (RHSC < 0 && RHSC > -256) {
10044 assert(Ptr->getOpcode() == ISD::ADD);
10046 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
10050 isInc = (Ptr->getOpcode() == ISD::ADD);
10051 Offset = Ptr->getOperand(1);
10053 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
10054 // AddressingMode 2
10055 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
10056 int RHSC = (int)RHS->getZExtValue();
10057 if (RHSC < 0 && RHSC > -0x1000) {
10058 assert(Ptr->getOpcode() == ISD::ADD);
10060 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
10061 Base = Ptr->getOperand(0);
10066 if (Ptr->getOpcode() == ISD::ADD) {
10068 ARM_AM::ShiftOpc ShOpcVal=
10069 ARM_AM::getShiftOpcForNode(Ptr->getOperand(0).getOpcode());
10070 if (ShOpcVal != ARM_AM::no_shift) {
10071 Base = Ptr->getOperand(1);
10072 Offset = Ptr->getOperand(0);
10074 Base = Ptr->getOperand(0);
10075 Offset = Ptr->getOperand(1);
10080 isInc = (Ptr->getOpcode() == ISD::ADD);
10081 Base = Ptr->getOperand(0);
10082 Offset = Ptr->getOperand(1);
10086 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
10090 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
10091 bool isSEXTLoad, SDValue &Base,
10092 SDValue &Offset, bool &isInc,
10093 SelectionDAG &DAG) {
10094 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
10097 Base = Ptr->getOperand(0);
10098 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
10099 int RHSC = (int)RHS->getZExtValue();
10100 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
10101 assert(Ptr->getOpcode() == ISD::ADD);
10103 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
10105 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
10106 isInc = Ptr->getOpcode() == ISD::ADD;
10107 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
10115 /// getPreIndexedAddressParts - returns true by value, base pointer and
10116 /// offset pointer and addressing mode by reference if the node's address
10117 /// can be legally represented as pre-indexed load / store address.
10119 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
10121 ISD::MemIndexedMode &AM,
10122 SelectionDAG &DAG) const {
10123 if (Subtarget->isThumb1Only())
10128 bool isSEXTLoad = false;
10129 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
10130 Ptr = LD->getBasePtr();
10131 VT = LD->getMemoryVT();
10132 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
10133 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
10134 Ptr = ST->getBasePtr();
10135 VT = ST->getMemoryVT();
10140 bool isLegal = false;
10141 if (Subtarget->isThumb2())
10142 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
10143 Offset, isInc, DAG);
10145 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
10146 Offset, isInc, DAG);
10150 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
10154 /// getPostIndexedAddressParts - returns true by value, base pointer and
10155 /// offset pointer and addressing mode by reference if this node can be
10156 /// combined with a load / store to form a post-indexed load / store.
10157 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
10160 ISD::MemIndexedMode &AM,
10161 SelectionDAG &DAG) const {
10162 if (Subtarget->isThumb1Only())
10167 bool isSEXTLoad = false;
10168 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
10169 VT = LD->getMemoryVT();
10170 Ptr = LD->getBasePtr();
10171 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
10172 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
10173 VT = ST->getMemoryVT();
10174 Ptr = ST->getBasePtr();
10179 bool isLegal = false;
10180 if (Subtarget->isThumb2())
10181 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
10184 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
10190 // Swap base ptr and offset to catch more post-index load / store when
10191 // it's legal. In Thumb2 mode, offset must be an immediate.
10192 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
10193 !Subtarget->isThumb2())
10194 std::swap(Base, Offset);
10196 // Post-indexed load / store update the base pointer.
10201 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
10205 void ARMTargetLowering::computeKnownBitsForTargetNode(const SDValue Op,
10208 const SelectionDAG &DAG,
10209 unsigned Depth) const {
10210 unsigned BitWidth = KnownOne.getBitWidth();
10211 KnownZero = KnownOne = APInt(BitWidth, 0);
10212 switch (Op.getOpcode()) {
10218 // These nodes' second result is a boolean
10219 if (Op.getResNo() == 0)
10221 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - 1);
10223 case ARMISD::CMOV: {
10224 // Bits are known zero/one if known on the LHS and RHS.
10225 DAG.computeKnownBits(Op.getOperand(0), KnownZero, KnownOne, Depth+1);
10226 if (KnownZero == 0 && KnownOne == 0) return;
10228 APInt KnownZeroRHS, KnownOneRHS;
10229 DAG.computeKnownBits(Op.getOperand(1), KnownZeroRHS, KnownOneRHS, Depth+1);
10230 KnownZero &= KnownZeroRHS;
10231 KnownOne &= KnownOneRHS;
10234 case ISD::INTRINSIC_W_CHAIN: {
10235 ConstantSDNode *CN = cast<ConstantSDNode>(Op->getOperand(1));
10236 Intrinsic::ID IntID = static_cast<Intrinsic::ID>(CN->getZExtValue());
10239 case Intrinsic::arm_ldaex:
10240 case Intrinsic::arm_ldrex: {
10241 EVT VT = cast<MemIntrinsicSDNode>(Op)->getMemoryVT();
10242 unsigned MemBits = VT.getScalarType().getSizeInBits();
10243 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - MemBits);
10251 //===----------------------------------------------------------------------===//
10252 // ARM Inline Assembly Support
10253 //===----------------------------------------------------------------------===//
10255 bool ARMTargetLowering::ExpandInlineAsm(CallInst *CI) const {
10256 // Looking for "rev" which is V6+.
10257 if (!Subtarget->hasV6Ops())
10260 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
10261 std::string AsmStr = IA->getAsmString();
10262 SmallVector<StringRef, 4> AsmPieces;
10263 SplitString(AsmStr, AsmPieces, ";\n");
10265 switch (AsmPieces.size()) {
10266 default: return false;
10268 AsmStr = AsmPieces[0];
10270 SplitString(AsmStr, AsmPieces, " \t,");
10273 if (AsmPieces.size() == 3 &&
10274 AsmPieces[0] == "rev" && AsmPieces[1] == "$0" && AsmPieces[2] == "$1" &&
10275 IA->getConstraintString().compare(0, 4, "=l,l") == 0) {
10276 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
10277 if (Ty && Ty->getBitWidth() == 32)
10278 return IntrinsicLowering::LowerToByteSwap(CI);
10286 /// getConstraintType - Given a constraint letter, return the type of
10287 /// constraint it is for this target.
10288 ARMTargetLowering::ConstraintType
10289 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
10290 if (Constraint.size() == 1) {
10291 switch (Constraint[0]) {
10293 case 'l': return C_RegisterClass;
10294 case 'w': return C_RegisterClass;
10295 case 'h': return C_RegisterClass;
10296 case 'x': return C_RegisterClass;
10297 case 't': return C_RegisterClass;
10298 case 'j': return C_Other; // Constant for movw.
10299 // An address with a single base register. Due to the way we
10300 // currently handle addresses it is the same as an 'r' memory constraint.
10301 case 'Q': return C_Memory;
10303 } else if (Constraint.size() == 2) {
10304 switch (Constraint[0]) {
10306 // All 'U+' constraints are addresses.
10307 case 'U': return C_Memory;
10310 return TargetLowering::getConstraintType(Constraint);
10313 /// Examine constraint type and operand type and determine a weight value.
10314 /// This object must already have been set up with the operand type
10315 /// and the current alternative constraint selected.
10316 TargetLowering::ConstraintWeight
10317 ARMTargetLowering::getSingleConstraintMatchWeight(
10318 AsmOperandInfo &info, const char *constraint) const {
10319 ConstraintWeight weight = CW_Invalid;
10320 Value *CallOperandVal = info.CallOperandVal;
10321 // If we don't have a value, we can't do a match,
10322 // but allow it at the lowest weight.
10323 if (!CallOperandVal)
10325 Type *type = CallOperandVal->getType();
10326 // Look at the constraint type.
10327 switch (*constraint) {
10329 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
10332 if (type->isIntegerTy()) {
10333 if (Subtarget->isThumb())
10334 weight = CW_SpecificReg;
10336 weight = CW_Register;
10340 if (type->isFloatingPointTy())
10341 weight = CW_Register;
10347 typedef std::pair<unsigned, const TargetRegisterClass*> RCPair;
10349 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
10351 if (Constraint.size() == 1) {
10352 // GCC ARM Constraint Letters
10353 switch (Constraint[0]) {
10354 case 'l': // Low regs or general regs.
10355 if (Subtarget->isThumb())
10356 return RCPair(0U, &ARM::tGPRRegClass);
10357 return RCPair(0U, &ARM::GPRRegClass);
10358 case 'h': // High regs or no regs.
10359 if (Subtarget->isThumb())
10360 return RCPair(0U, &ARM::hGPRRegClass);
10363 return RCPair(0U, &ARM::GPRRegClass);
10365 if (VT == MVT::Other)
10367 if (VT == MVT::f32)
10368 return RCPair(0U, &ARM::SPRRegClass);
10369 if (VT.getSizeInBits() == 64)
10370 return RCPair(0U, &ARM::DPRRegClass);
10371 if (VT.getSizeInBits() == 128)
10372 return RCPair(0U, &ARM::QPRRegClass);
10375 if (VT == MVT::Other)
10377 if (VT == MVT::f32)
10378 return RCPair(0U, &ARM::SPR_8RegClass);
10379 if (VT.getSizeInBits() == 64)
10380 return RCPair(0U, &ARM::DPR_8RegClass);
10381 if (VT.getSizeInBits() == 128)
10382 return RCPair(0U, &ARM::QPR_8RegClass);
10385 if (VT == MVT::f32)
10386 return RCPair(0U, &ARM::SPRRegClass);
10390 if (StringRef("{cc}").equals_lower(Constraint))
10391 return std::make_pair(unsigned(ARM::CPSR), &ARM::CCRRegClass);
10393 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
10396 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
10397 /// vector. If it is invalid, don't add anything to Ops.
10398 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
10399 std::string &Constraint,
10400 std::vector<SDValue>&Ops,
10401 SelectionDAG &DAG) const {
10404 // Currently only support length 1 constraints.
10405 if (Constraint.length() != 1) return;
10407 char ConstraintLetter = Constraint[0];
10408 switch (ConstraintLetter) {
10411 case 'I': case 'J': case 'K': case 'L':
10412 case 'M': case 'N': case 'O':
10413 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
10417 int64_t CVal64 = C->getSExtValue();
10418 int CVal = (int) CVal64;
10419 // None of these constraints allow values larger than 32 bits. Check
10420 // that the value fits in an int.
10421 if (CVal != CVal64)
10424 switch (ConstraintLetter) {
10426 // Constant suitable for movw, must be between 0 and
10428 if (Subtarget->hasV6T2Ops())
10429 if (CVal >= 0 && CVal <= 65535)
10433 if (Subtarget->isThumb1Only()) {
10434 // This must be a constant between 0 and 255, for ADD
10436 if (CVal >= 0 && CVal <= 255)
10438 } else if (Subtarget->isThumb2()) {
10439 // A constant that can be used as an immediate value in a
10440 // data-processing instruction.
10441 if (ARM_AM::getT2SOImmVal(CVal) != -1)
10444 // A constant that can be used as an immediate value in a
10445 // data-processing instruction.
10446 if (ARM_AM::getSOImmVal(CVal) != -1)
10452 if (Subtarget->isThumb()) { // FIXME thumb2
10453 // This must be a constant between -255 and -1, for negated ADD
10454 // immediates. This can be used in GCC with an "n" modifier that
10455 // prints the negated value, for use with SUB instructions. It is
10456 // not useful otherwise but is implemented for compatibility.
10457 if (CVal >= -255 && CVal <= -1)
10460 // This must be a constant between -4095 and 4095. It is not clear
10461 // what this constraint is intended for. Implemented for
10462 // compatibility with GCC.
10463 if (CVal >= -4095 && CVal <= 4095)
10469 if (Subtarget->isThumb1Only()) {
10470 // A 32-bit value where only one byte has a nonzero value. Exclude
10471 // zero to match GCC. This constraint is used by GCC internally for
10472 // constants that can be loaded with a move/shift combination.
10473 // It is not useful otherwise but is implemented for compatibility.
10474 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
10476 } else if (Subtarget->isThumb2()) {
10477 // A constant whose bitwise inverse can be used as an immediate
10478 // value in a data-processing instruction. This can be used in GCC
10479 // with a "B" modifier that prints the inverted value, for use with
10480 // BIC and MVN instructions. It is not useful otherwise but is
10481 // implemented for compatibility.
10482 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
10485 // A constant whose bitwise inverse can be used as an immediate
10486 // value in a data-processing instruction. This can be used in GCC
10487 // with a "B" modifier that prints the inverted value, for use with
10488 // BIC and MVN instructions. It is not useful otherwise but is
10489 // implemented for compatibility.
10490 if (ARM_AM::getSOImmVal(~CVal) != -1)
10496 if (Subtarget->isThumb1Only()) {
10497 // This must be a constant between -7 and 7,
10498 // for 3-operand ADD/SUB immediate instructions.
10499 if (CVal >= -7 && CVal < 7)
10501 } else if (Subtarget->isThumb2()) {
10502 // A constant whose negation can be used as an immediate value in a
10503 // data-processing instruction. This can be used in GCC with an "n"
10504 // modifier that prints the negated value, for use with SUB
10505 // instructions. It is not useful otherwise but is implemented for
10507 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
10510 // A constant whose negation can be used as an immediate value in a
10511 // data-processing instruction. This can be used in GCC with an "n"
10512 // modifier that prints the negated value, for use with SUB
10513 // instructions. It is not useful otherwise but is implemented for
10515 if (ARM_AM::getSOImmVal(-CVal) != -1)
10521 if (Subtarget->isThumb()) { // FIXME thumb2
10522 // This must be a multiple of 4 between 0 and 1020, for
10523 // ADD sp + immediate.
10524 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
10527 // A power of two or a constant between 0 and 32. This is used in
10528 // GCC for the shift amount on shifted register operands, but it is
10529 // useful in general for any shift amounts.
10530 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
10536 if (Subtarget->isThumb()) { // FIXME thumb2
10537 // This must be a constant between 0 and 31, for shift amounts.
10538 if (CVal >= 0 && CVal <= 31)
10544 if (Subtarget->isThumb()) { // FIXME thumb2
10545 // This must be a multiple of 4 between -508 and 508, for
10546 // ADD/SUB sp = sp + immediate.
10547 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
10552 Result = DAG.getTargetConstant(CVal, Op.getValueType());
10556 if (Result.getNode()) {
10557 Ops.push_back(Result);
10560 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
10563 SDValue ARMTargetLowering::LowerDivRem(SDValue Op, SelectionDAG &DAG) const {
10564 assert(Subtarget->isTargetAEABI() && "Register-based DivRem lowering only");
10565 unsigned Opcode = Op->getOpcode();
10566 assert((Opcode == ISD::SDIVREM || Opcode == ISD::UDIVREM) &&
10567 "Invalid opcode for Div/Rem lowering");
10568 bool isSigned = (Opcode == ISD::SDIVREM);
10569 EVT VT = Op->getValueType(0);
10570 Type *Ty = VT.getTypeForEVT(*DAG.getContext());
10573 switch (VT.getSimpleVT().SimpleTy) {
10574 default: llvm_unreachable("Unexpected request for libcall!");
10575 case MVT::i8: LC= isSigned ? RTLIB::SDIVREM_I8 : RTLIB::UDIVREM_I8; break;
10576 case MVT::i16: LC= isSigned ? RTLIB::SDIVREM_I16 : RTLIB::UDIVREM_I16; break;
10577 case MVT::i32: LC= isSigned ? RTLIB::SDIVREM_I32 : RTLIB::UDIVREM_I32; break;
10578 case MVT::i64: LC= isSigned ? RTLIB::SDIVREM_I64 : RTLIB::UDIVREM_I64; break;
10581 SDValue InChain = DAG.getEntryNode();
10583 TargetLowering::ArgListTy Args;
10584 TargetLowering::ArgListEntry Entry;
10585 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i) {
10586 EVT ArgVT = Op->getOperand(i).getValueType();
10587 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
10588 Entry.Node = Op->getOperand(i);
10590 Entry.isSExt = isSigned;
10591 Entry.isZExt = !isSigned;
10592 Args.push_back(Entry);
10595 SDValue Callee = DAG.getExternalSymbol(getLibcallName(LC),
10598 Type *RetTy = (Type*)StructType::get(Ty, Ty, NULL);
10601 TargetLowering::CallLoweringInfo CLI(DAG);
10602 CLI.setDebugLoc(dl).setChain(InChain)
10603 .setCallee(getLibcallCallingConv(LC), RetTy, Callee, std::move(Args), 0)
10604 .setInRegister().setSExtResult(isSigned).setZExtResult(!isSigned);
10606 std::pair<SDValue, SDValue> CallInfo = LowerCallTo(CLI);
10607 return CallInfo.first;
10611 ARMTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const {
10612 assert(Subtarget->isTargetWindows() && "unsupported target platform");
10616 SDValue Chain = Op.getOperand(0);
10617 SDValue Size = Op.getOperand(1);
10619 SDValue Words = DAG.getNode(ISD::SRL, DL, MVT::i32, Size,
10620 DAG.getConstant(2, MVT::i32));
10623 Chain = DAG.getCopyToReg(Chain, DL, ARM::R4, Words, Flag);
10624 Flag = Chain.getValue(1);
10626 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
10627 Chain = DAG.getNode(ARMISD::WIN__CHKSTK, DL, NodeTys, Chain, Flag);
10629 SDValue NewSP = DAG.getCopyFromReg(Chain, DL, ARM::SP, MVT::i32);
10630 Chain = NewSP.getValue(1);
10632 SDValue Ops[2] = { NewSP, Chain };
10633 return DAG.getMergeValues(Ops, DL);
10637 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
10638 // The ARM target isn't yet aware of offsets.
10642 bool ARM::isBitFieldInvertedMask(unsigned v) {
10643 if (v == 0xffffffff)
10646 // there can be 1's on either or both "outsides", all the "inside"
10647 // bits must be 0's
10648 unsigned TO = CountTrailingOnes_32(v);
10649 unsigned LO = CountLeadingOnes_32(v);
10650 v = (v >> TO) << TO;
10651 v = (v << LO) >> LO;
10655 /// isFPImmLegal - Returns true if the target can instruction select the
10656 /// specified FP immediate natively. If false, the legalizer will
10657 /// materialize the FP immediate as a load from a constant pool.
10658 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
10659 if (!Subtarget->hasVFP3())
10661 if (VT == MVT::f32)
10662 return ARM_AM::getFP32Imm(Imm) != -1;
10663 if (VT == MVT::f64)
10664 return ARM_AM::getFP64Imm(Imm) != -1;
10668 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
10669 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
10670 /// specified in the intrinsic calls.
10671 bool ARMTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
10673 unsigned Intrinsic) const {
10674 switch (Intrinsic) {
10675 case Intrinsic::arm_neon_vld1:
10676 case Intrinsic::arm_neon_vld2:
10677 case Intrinsic::arm_neon_vld3:
10678 case Intrinsic::arm_neon_vld4:
10679 case Intrinsic::arm_neon_vld2lane:
10680 case Intrinsic::arm_neon_vld3lane:
10681 case Intrinsic::arm_neon_vld4lane: {
10682 Info.opc = ISD::INTRINSIC_W_CHAIN;
10683 // Conservatively set memVT to the entire set of vectors loaded.
10684 uint64_t NumElts = getDataLayout()->getTypeAllocSize(I.getType()) / 8;
10685 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
10686 Info.ptrVal = I.getArgOperand(0);
10688 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
10689 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
10690 Info.vol = false; // volatile loads with NEON intrinsics not supported
10691 Info.readMem = true;
10692 Info.writeMem = false;
10695 case Intrinsic::arm_neon_vst1:
10696 case Intrinsic::arm_neon_vst2:
10697 case Intrinsic::arm_neon_vst3:
10698 case Intrinsic::arm_neon_vst4:
10699 case Intrinsic::arm_neon_vst2lane:
10700 case Intrinsic::arm_neon_vst3lane:
10701 case Intrinsic::arm_neon_vst4lane: {
10702 Info.opc = ISD::INTRINSIC_VOID;
10703 // Conservatively set memVT to the entire set of vectors stored.
10704 unsigned NumElts = 0;
10705 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
10706 Type *ArgTy = I.getArgOperand(ArgI)->getType();
10707 if (!ArgTy->isVectorTy())
10709 NumElts += getDataLayout()->getTypeAllocSize(ArgTy) / 8;
10711 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
10712 Info.ptrVal = I.getArgOperand(0);
10714 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
10715 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
10716 Info.vol = false; // volatile stores with NEON intrinsics not supported
10717 Info.readMem = false;
10718 Info.writeMem = true;
10721 case Intrinsic::arm_ldaex:
10722 case Intrinsic::arm_ldrex: {
10723 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(0)->getType());
10724 Info.opc = ISD::INTRINSIC_W_CHAIN;
10725 Info.memVT = MVT::getVT(PtrTy->getElementType());
10726 Info.ptrVal = I.getArgOperand(0);
10728 Info.align = getDataLayout()->getABITypeAlignment(PtrTy->getElementType());
10730 Info.readMem = true;
10731 Info.writeMem = false;
10734 case Intrinsic::arm_stlex:
10735 case Intrinsic::arm_strex: {
10736 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(1)->getType());
10737 Info.opc = ISD::INTRINSIC_W_CHAIN;
10738 Info.memVT = MVT::getVT(PtrTy->getElementType());
10739 Info.ptrVal = I.getArgOperand(1);
10741 Info.align = getDataLayout()->getABITypeAlignment(PtrTy->getElementType());
10743 Info.readMem = false;
10744 Info.writeMem = true;
10747 case Intrinsic::arm_stlexd:
10748 case Intrinsic::arm_strexd: {
10749 Info.opc = ISD::INTRINSIC_W_CHAIN;
10750 Info.memVT = MVT::i64;
10751 Info.ptrVal = I.getArgOperand(2);
10755 Info.readMem = false;
10756 Info.writeMem = true;
10759 case Intrinsic::arm_ldaexd:
10760 case Intrinsic::arm_ldrexd: {
10761 Info.opc = ISD::INTRINSIC_W_CHAIN;
10762 Info.memVT = MVT::i64;
10763 Info.ptrVal = I.getArgOperand(0);
10767 Info.readMem = true;
10768 Info.writeMem = false;
10778 /// \brief Returns true if it is beneficial to convert a load of a constant
10779 /// to just the constant itself.
10780 bool ARMTargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
10782 assert(Ty->isIntegerTy());
10784 unsigned Bits = Ty->getPrimitiveSizeInBits();
10785 if (Bits == 0 || Bits > 32)
10790 bool ARMTargetLowering::shouldExpandAtomicInIR(Instruction *Inst) const {
10791 // Loads and stores less than 64-bits are already atomic; ones above that
10792 // are doomed anyway, so defer to the default libcall and blame the OS when
10793 // things go wrong. Cortex M doesn't have ldrexd/strexd though, so don't emit
10794 // anything for those.
10795 bool IsMClass = Subtarget->isMClass();
10796 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
10797 unsigned Size = SI->getValueOperand()->getType()->getPrimitiveSizeInBits();
10798 return Size == 64 && !IsMClass;
10799 } else if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
10800 return LI->getType()->getPrimitiveSizeInBits() == 64 && !IsMClass;
10803 // For the real atomic operations, we have ldrex/strex up to 32 bits,
10804 // and up to 64 bits on the non-M profiles
10805 unsigned AtomicLimit = IsMClass ? 32 : 64;
10806 return Inst->getType()->getPrimitiveSizeInBits() <= AtomicLimit;
10809 // This has so far only been implemented for MachO.
10810 bool ARMTargetLowering::useLoadStackGuardNode() const {
10811 return Subtarget->getTargetTriple().getObjectFormat() == Triple::MachO;
10814 Value *ARMTargetLowering::emitLoadLinked(IRBuilder<> &Builder, Value *Addr,
10815 AtomicOrdering Ord) const {
10816 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
10817 Type *ValTy = cast<PointerType>(Addr->getType())->getElementType();
10819 Ord == Acquire || Ord == AcquireRelease || Ord == SequentiallyConsistent;
10821 // Since i64 isn't legal and intrinsics don't get type-lowered, the ldrexd
10822 // intrinsic must return {i32, i32} and we have to recombine them into a
10823 // single i64 here.
10824 if (ValTy->getPrimitiveSizeInBits() == 64) {
10825 Intrinsic::ID Int =
10826 IsAcquire ? Intrinsic::arm_ldaexd : Intrinsic::arm_ldrexd;
10827 Function *Ldrex = llvm::Intrinsic::getDeclaration(M, Int);
10829 Addr = Builder.CreateBitCast(Addr, Type::getInt8PtrTy(M->getContext()));
10830 Value *LoHi = Builder.CreateCall(Ldrex, Addr, "lohi");
10832 Value *Lo = Builder.CreateExtractValue(LoHi, 0, "lo");
10833 Value *Hi = Builder.CreateExtractValue(LoHi, 1, "hi");
10834 if (!Subtarget->isLittle())
10835 std::swap (Lo, Hi);
10836 Lo = Builder.CreateZExt(Lo, ValTy, "lo64");
10837 Hi = Builder.CreateZExt(Hi, ValTy, "hi64");
10838 return Builder.CreateOr(
10839 Lo, Builder.CreateShl(Hi, ConstantInt::get(ValTy, 32)), "val64");
10842 Type *Tys[] = { Addr->getType() };
10843 Intrinsic::ID Int = IsAcquire ? Intrinsic::arm_ldaex : Intrinsic::arm_ldrex;
10844 Function *Ldrex = llvm::Intrinsic::getDeclaration(M, Int, Tys);
10846 return Builder.CreateTruncOrBitCast(
10847 Builder.CreateCall(Ldrex, Addr),
10848 cast<PointerType>(Addr->getType())->getElementType());
10851 Value *ARMTargetLowering::emitStoreConditional(IRBuilder<> &Builder, Value *Val,
10853 AtomicOrdering Ord) const {
10854 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
10856 Ord == Release || Ord == AcquireRelease || Ord == SequentiallyConsistent;
10858 // Since the intrinsics must have legal type, the i64 intrinsics take two
10859 // parameters: "i32, i32". We must marshal Val into the appropriate form
10860 // before the call.
10861 if (Val->getType()->getPrimitiveSizeInBits() == 64) {
10862 Intrinsic::ID Int =
10863 IsRelease ? Intrinsic::arm_stlexd : Intrinsic::arm_strexd;
10864 Function *Strex = Intrinsic::getDeclaration(M, Int);
10865 Type *Int32Ty = Type::getInt32Ty(M->getContext());
10867 Value *Lo = Builder.CreateTrunc(Val, Int32Ty, "lo");
10868 Value *Hi = Builder.CreateTrunc(Builder.CreateLShr(Val, 32), Int32Ty, "hi");
10869 if (!Subtarget->isLittle())
10870 std::swap (Lo, Hi);
10871 Addr = Builder.CreateBitCast(Addr, Type::getInt8PtrTy(M->getContext()));
10872 return Builder.CreateCall3(Strex, Lo, Hi, Addr);
10875 Intrinsic::ID Int = IsRelease ? Intrinsic::arm_stlex : Intrinsic::arm_strex;
10876 Type *Tys[] = { Addr->getType() };
10877 Function *Strex = Intrinsic::getDeclaration(M, Int, Tys);
10879 return Builder.CreateCall2(
10880 Strex, Builder.CreateZExtOrBitCast(
10881 Val, Strex->getFunctionType()->getParamType(0)),
10893 static bool isHomogeneousAggregate(Type *Ty, HABaseType &Base,
10894 uint64_t &Members) {
10895 if (const StructType *ST = dyn_cast<StructType>(Ty)) {
10896 for (unsigned i = 0; i < ST->getNumElements(); ++i) {
10897 uint64_t SubMembers = 0;
10898 if (!isHomogeneousAggregate(ST->getElementType(i), Base, SubMembers))
10900 Members += SubMembers;
10902 } else if (const ArrayType *AT = dyn_cast<ArrayType>(Ty)) {
10903 uint64_t SubMembers = 0;
10904 if (!isHomogeneousAggregate(AT->getElementType(), Base, SubMembers))
10906 Members += SubMembers * AT->getNumElements();
10907 } else if (Ty->isFloatTy()) {
10908 if (Base != HA_UNKNOWN && Base != HA_FLOAT)
10912 } else if (Ty->isDoubleTy()) {
10913 if (Base != HA_UNKNOWN && Base != HA_DOUBLE)
10917 } else if (const VectorType *VT = dyn_cast<VectorType>(Ty)) {
10924 return VT->getBitWidth() == 64;
10926 return VT->getBitWidth() == 128;
10928 switch (VT->getBitWidth()) {
10941 return (Members > 0 && Members <= 4);
10944 /// \brief Return true if a type is an AAPCS-VFP homogeneous aggregate.
10945 bool ARMTargetLowering::functionArgumentNeedsConsecutiveRegisters(
10946 Type *Ty, CallingConv::ID CallConv, bool isVarArg) const {
10947 if (getEffectiveCallingConv(CallConv, isVarArg) !=
10948 CallingConv::ARM_AAPCS_VFP)
10951 HABaseType Base = HA_UNKNOWN;
10952 uint64_t Members = 0;
10953 bool result = isHomogeneousAggregate(Ty, Base, Members);
10954 DEBUG(dbgs() << "isHA: " << result << " "; Ty->dump(); dbgs() << "\n");