1 //===-- ARMISelLowering.cpp - ARM DAG Lowering Implementation -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the interfaces that ARM uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
15 #include "ARMISelLowering.h"
16 #include "ARMCallingConv.h"
17 #include "ARMConstantPoolValue.h"
18 #include "ARMMachineFunctionInfo.h"
19 #include "ARMPerfectShuffle.h"
20 #include "ARMSubtarget.h"
21 #include "ARMTargetMachine.h"
22 #include "ARMTargetObjectFile.h"
23 #include "MCTargetDesc/ARMAddressingModes.h"
24 #include "llvm/ADT/Statistic.h"
25 #include "llvm/ADT/StringExtras.h"
26 #include "llvm/ADT/StringSwitch.h"
27 #include "llvm/CodeGen/CallingConvLower.h"
28 #include "llvm/CodeGen/IntrinsicLowering.h"
29 #include "llvm/CodeGen/MachineBasicBlock.h"
30 #include "llvm/CodeGen/MachineFrameInfo.h"
31 #include "llvm/CodeGen/MachineFunction.h"
32 #include "llvm/CodeGen/MachineInstrBuilder.h"
33 #include "llvm/CodeGen/MachineJumpTableInfo.h"
34 #include "llvm/CodeGen/MachineModuleInfo.h"
35 #include "llvm/CodeGen/MachineRegisterInfo.h"
36 #include "llvm/CodeGen/SelectionDAG.h"
37 #include "llvm/IR/CallingConv.h"
38 #include "llvm/IR/Constants.h"
39 #include "llvm/IR/Function.h"
40 #include "llvm/IR/GlobalValue.h"
41 #include "llvm/IR/IRBuilder.h"
42 #include "llvm/IR/Instruction.h"
43 #include "llvm/IR/Instructions.h"
44 #include "llvm/IR/IntrinsicInst.h"
45 #include "llvm/IR/Intrinsics.h"
46 #include "llvm/IR/Type.h"
47 #include "llvm/MC/MCSectionMachO.h"
48 #include "llvm/Support/CommandLine.h"
49 #include "llvm/Support/Debug.h"
50 #include "llvm/Support/ErrorHandling.h"
51 #include "llvm/Support/MathExtras.h"
52 #include "llvm/Support/raw_ostream.h"
53 #include "llvm/Target/TargetOptions.h"
57 #define DEBUG_TYPE "arm-isel"
59 STATISTIC(NumTailCalls, "Number of tail calls");
60 STATISTIC(NumMovwMovt, "Number of GAs materialized with movw + movt");
61 STATISTIC(NumLoopByVals, "Number of loops generated for byval arguments");
64 ARMInterworking("arm-interworking", cl::Hidden,
65 cl::desc("Enable / disable ARM interworking (for debugging only)"),
69 class ARMCCState : public CCState {
71 ARMCCState(CallingConv::ID CC, bool isVarArg, MachineFunction &MF,
72 SmallVectorImpl<CCValAssign> &locs, LLVMContext &C,
74 : CCState(CC, isVarArg, MF, locs, C) {
75 assert(((PC == Call) || (PC == Prologue)) &&
76 "ARMCCState users must specify whether their context is call"
77 "or prologue generation.");
83 // The APCS parameter registers.
84 static const MCPhysReg GPRArgRegs[] = {
85 ARM::R0, ARM::R1, ARM::R2, ARM::R3
88 void ARMTargetLowering::addTypeForNEON(MVT VT, MVT PromotedLdStVT,
89 MVT PromotedBitwiseVT) {
90 if (VT != PromotedLdStVT) {
91 setOperationAction(ISD::LOAD, VT, Promote);
92 AddPromotedToType (ISD::LOAD, VT, PromotedLdStVT);
94 setOperationAction(ISD::STORE, VT, Promote);
95 AddPromotedToType (ISD::STORE, VT, PromotedLdStVT);
98 MVT ElemTy = VT.getVectorElementType();
99 if (ElemTy != MVT::i64 && ElemTy != MVT::f64)
100 setOperationAction(ISD::SETCC, VT, Custom);
101 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
102 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
103 if (ElemTy == MVT::i32) {
104 setOperationAction(ISD::SINT_TO_FP, VT, Custom);
105 setOperationAction(ISD::UINT_TO_FP, VT, Custom);
106 setOperationAction(ISD::FP_TO_SINT, VT, Custom);
107 setOperationAction(ISD::FP_TO_UINT, VT, Custom);
109 setOperationAction(ISD::SINT_TO_FP, VT, Expand);
110 setOperationAction(ISD::UINT_TO_FP, VT, Expand);
111 setOperationAction(ISD::FP_TO_SINT, VT, Expand);
112 setOperationAction(ISD::FP_TO_UINT, VT, Expand);
114 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
115 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
116 setOperationAction(ISD::CONCAT_VECTORS, VT, Legal);
117 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Legal);
118 setOperationAction(ISD::SELECT, VT, Expand);
119 setOperationAction(ISD::SELECT_CC, VT, Expand);
120 setOperationAction(ISD::VSELECT, VT, Expand);
121 setOperationAction(ISD::SIGN_EXTEND_INREG, VT, Expand);
122 if (VT.isInteger()) {
123 setOperationAction(ISD::SHL, VT, Custom);
124 setOperationAction(ISD::SRA, VT, Custom);
125 setOperationAction(ISD::SRL, VT, Custom);
128 // Promote all bit-wise operations.
129 if (VT.isInteger() && VT != PromotedBitwiseVT) {
130 setOperationAction(ISD::AND, VT, Promote);
131 AddPromotedToType (ISD::AND, VT, PromotedBitwiseVT);
132 setOperationAction(ISD::OR, VT, Promote);
133 AddPromotedToType (ISD::OR, VT, PromotedBitwiseVT);
134 setOperationAction(ISD::XOR, VT, Promote);
135 AddPromotedToType (ISD::XOR, VT, PromotedBitwiseVT);
138 // Neon does not support vector divide/remainder operations.
139 setOperationAction(ISD::SDIV, VT, Expand);
140 setOperationAction(ISD::UDIV, VT, Expand);
141 setOperationAction(ISD::FDIV, VT, Expand);
142 setOperationAction(ISD::SREM, VT, Expand);
143 setOperationAction(ISD::UREM, VT, Expand);
144 setOperationAction(ISD::FREM, VT, Expand);
147 void ARMTargetLowering::addDRTypeForNEON(MVT VT) {
148 addRegisterClass(VT, &ARM::DPRRegClass);
149 addTypeForNEON(VT, MVT::f64, MVT::v2i32);
152 void ARMTargetLowering::addQRTypeForNEON(MVT VT) {
153 addRegisterClass(VT, &ARM::DPairRegClass);
154 addTypeForNEON(VT, MVT::v2f64, MVT::v4i32);
157 ARMTargetLowering::ARMTargetLowering(const TargetMachine &TM,
158 const ARMSubtarget &STI)
159 : TargetLowering(TM), Subtarget(&STI) {
160 RegInfo = Subtarget->getRegisterInfo();
161 Itins = Subtarget->getInstrItineraryData();
163 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
165 if (Subtarget->isTargetMachO()) {
166 // Uses VFP for Thumb libfuncs if available.
167 if (Subtarget->isThumb() && Subtarget->hasVFP2() &&
168 Subtarget->hasARMOps() && !Subtarget->useSoftFloat()) {
169 // Single-precision floating-point arithmetic.
170 setLibcallName(RTLIB::ADD_F32, "__addsf3vfp");
171 setLibcallName(RTLIB::SUB_F32, "__subsf3vfp");
172 setLibcallName(RTLIB::MUL_F32, "__mulsf3vfp");
173 setLibcallName(RTLIB::DIV_F32, "__divsf3vfp");
175 // Double-precision floating-point arithmetic.
176 setLibcallName(RTLIB::ADD_F64, "__adddf3vfp");
177 setLibcallName(RTLIB::SUB_F64, "__subdf3vfp");
178 setLibcallName(RTLIB::MUL_F64, "__muldf3vfp");
179 setLibcallName(RTLIB::DIV_F64, "__divdf3vfp");
181 // Single-precision comparisons.
182 setLibcallName(RTLIB::OEQ_F32, "__eqsf2vfp");
183 setLibcallName(RTLIB::UNE_F32, "__nesf2vfp");
184 setLibcallName(RTLIB::OLT_F32, "__ltsf2vfp");
185 setLibcallName(RTLIB::OLE_F32, "__lesf2vfp");
186 setLibcallName(RTLIB::OGE_F32, "__gesf2vfp");
187 setLibcallName(RTLIB::OGT_F32, "__gtsf2vfp");
188 setLibcallName(RTLIB::UO_F32, "__unordsf2vfp");
189 setLibcallName(RTLIB::O_F32, "__unordsf2vfp");
191 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
192 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETNE);
193 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
194 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
195 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
196 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
197 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
198 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
200 // Double-precision comparisons.
201 setLibcallName(RTLIB::OEQ_F64, "__eqdf2vfp");
202 setLibcallName(RTLIB::UNE_F64, "__nedf2vfp");
203 setLibcallName(RTLIB::OLT_F64, "__ltdf2vfp");
204 setLibcallName(RTLIB::OLE_F64, "__ledf2vfp");
205 setLibcallName(RTLIB::OGE_F64, "__gedf2vfp");
206 setLibcallName(RTLIB::OGT_F64, "__gtdf2vfp");
207 setLibcallName(RTLIB::UO_F64, "__unorddf2vfp");
208 setLibcallName(RTLIB::O_F64, "__unorddf2vfp");
210 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
211 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETNE);
212 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
213 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
214 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
215 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
216 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
217 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
219 // Floating-point to integer conversions.
220 // i64 conversions are done via library routines even when generating VFP
221 // instructions, so use the same ones.
222 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__fixdfsivfp");
223 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__fixunsdfsivfp");
224 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__fixsfsivfp");
225 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__fixunssfsivfp");
227 // Conversions between floating types.
228 setLibcallName(RTLIB::FPROUND_F64_F32, "__truncdfsf2vfp");
229 setLibcallName(RTLIB::FPEXT_F32_F64, "__extendsfdf2vfp");
231 // Integer to floating-point conversions.
232 // i64 conversions are done via library routines even when generating VFP
233 // instructions, so use the same ones.
234 // FIXME: There appears to be some naming inconsistency in ARM libgcc:
235 // e.g., __floatunsidf vs. __floatunssidfvfp.
236 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__floatsidfvfp");
237 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__floatunssidfvfp");
238 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__floatsisfvfp");
239 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__floatunssisfvfp");
243 // These libcalls are not available in 32-bit.
244 setLibcallName(RTLIB::SHL_I128, nullptr);
245 setLibcallName(RTLIB::SRL_I128, nullptr);
246 setLibcallName(RTLIB::SRA_I128, nullptr);
248 if (Subtarget->isAAPCS_ABI() && !Subtarget->isTargetMachO() &&
249 !Subtarget->isTargetWindows()) {
250 static const struct {
251 const RTLIB::Libcall Op;
252 const char * const Name;
253 const CallingConv::ID CC;
254 const ISD::CondCode Cond;
256 // Double-precision floating-point arithmetic helper functions
257 // RTABI chapter 4.1.2, Table 2
258 { RTLIB::ADD_F64, "__aeabi_dadd", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
259 { RTLIB::DIV_F64, "__aeabi_ddiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
260 { RTLIB::MUL_F64, "__aeabi_dmul", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
261 { RTLIB::SUB_F64, "__aeabi_dsub", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
263 // Double-precision floating-point comparison helper functions
264 // RTABI chapter 4.1.2, Table 3
265 { RTLIB::OEQ_F64, "__aeabi_dcmpeq", CallingConv::ARM_AAPCS, ISD::SETNE },
266 { RTLIB::UNE_F64, "__aeabi_dcmpeq", CallingConv::ARM_AAPCS, ISD::SETEQ },
267 { RTLIB::OLT_F64, "__aeabi_dcmplt", CallingConv::ARM_AAPCS, ISD::SETNE },
268 { RTLIB::OLE_F64, "__aeabi_dcmple", CallingConv::ARM_AAPCS, ISD::SETNE },
269 { RTLIB::OGE_F64, "__aeabi_dcmpge", CallingConv::ARM_AAPCS, ISD::SETNE },
270 { RTLIB::OGT_F64, "__aeabi_dcmpgt", CallingConv::ARM_AAPCS, ISD::SETNE },
271 { RTLIB::UO_F64, "__aeabi_dcmpun", CallingConv::ARM_AAPCS, ISD::SETNE },
272 { RTLIB::O_F64, "__aeabi_dcmpun", CallingConv::ARM_AAPCS, ISD::SETEQ },
274 // Single-precision floating-point arithmetic helper functions
275 // RTABI chapter 4.1.2, Table 4
276 { RTLIB::ADD_F32, "__aeabi_fadd", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
277 { RTLIB::DIV_F32, "__aeabi_fdiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
278 { RTLIB::MUL_F32, "__aeabi_fmul", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
279 { RTLIB::SUB_F32, "__aeabi_fsub", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
281 // Single-precision floating-point comparison helper functions
282 // RTABI chapter 4.1.2, Table 5
283 { RTLIB::OEQ_F32, "__aeabi_fcmpeq", CallingConv::ARM_AAPCS, ISD::SETNE },
284 { RTLIB::UNE_F32, "__aeabi_fcmpeq", CallingConv::ARM_AAPCS, ISD::SETEQ },
285 { RTLIB::OLT_F32, "__aeabi_fcmplt", CallingConv::ARM_AAPCS, ISD::SETNE },
286 { RTLIB::OLE_F32, "__aeabi_fcmple", CallingConv::ARM_AAPCS, ISD::SETNE },
287 { RTLIB::OGE_F32, "__aeabi_fcmpge", CallingConv::ARM_AAPCS, ISD::SETNE },
288 { RTLIB::OGT_F32, "__aeabi_fcmpgt", CallingConv::ARM_AAPCS, ISD::SETNE },
289 { RTLIB::UO_F32, "__aeabi_fcmpun", CallingConv::ARM_AAPCS, ISD::SETNE },
290 { RTLIB::O_F32, "__aeabi_fcmpun", CallingConv::ARM_AAPCS, ISD::SETEQ },
292 // Floating-point to integer conversions.
293 // RTABI chapter 4.1.2, Table 6
294 { RTLIB::FPTOSINT_F64_I32, "__aeabi_d2iz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
295 { RTLIB::FPTOUINT_F64_I32, "__aeabi_d2uiz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
296 { RTLIB::FPTOSINT_F64_I64, "__aeabi_d2lz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
297 { RTLIB::FPTOUINT_F64_I64, "__aeabi_d2ulz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
298 { RTLIB::FPTOSINT_F32_I32, "__aeabi_f2iz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
299 { RTLIB::FPTOUINT_F32_I32, "__aeabi_f2uiz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
300 { RTLIB::FPTOSINT_F32_I64, "__aeabi_f2lz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
301 { RTLIB::FPTOUINT_F32_I64, "__aeabi_f2ulz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
303 // Conversions between floating types.
304 // RTABI chapter 4.1.2, Table 7
305 { RTLIB::FPROUND_F64_F32, "__aeabi_d2f", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
306 { RTLIB::FPROUND_F64_F16, "__aeabi_d2h", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
307 { RTLIB::FPEXT_F32_F64, "__aeabi_f2d", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
309 // Integer to floating-point conversions.
310 // RTABI chapter 4.1.2, Table 8
311 { RTLIB::SINTTOFP_I32_F64, "__aeabi_i2d", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
312 { RTLIB::UINTTOFP_I32_F64, "__aeabi_ui2d", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
313 { RTLIB::SINTTOFP_I64_F64, "__aeabi_l2d", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
314 { RTLIB::UINTTOFP_I64_F64, "__aeabi_ul2d", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
315 { RTLIB::SINTTOFP_I32_F32, "__aeabi_i2f", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
316 { RTLIB::UINTTOFP_I32_F32, "__aeabi_ui2f", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
317 { RTLIB::SINTTOFP_I64_F32, "__aeabi_l2f", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
318 { RTLIB::UINTTOFP_I64_F32, "__aeabi_ul2f", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
320 // Long long helper functions
321 // RTABI chapter 4.2, Table 9
322 { RTLIB::MUL_I64, "__aeabi_lmul", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
323 { RTLIB::SHL_I64, "__aeabi_llsl", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
324 { RTLIB::SRL_I64, "__aeabi_llsr", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
325 { RTLIB::SRA_I64, "__aeabi_lasr", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
327 // Integer division functions
328 // RTABI chapter 4.3.1
329 { RTLIB::SDIV_I8, "__aeabi_idiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
330 { RTLIB::SDIV_I16, "__aeabi_idiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
331 { RTLIB::SDIV_I32, "__aeabi_idiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
332 { RTLIB::SDIV_I64, "__aeabi_ldivmod", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
333 { RTLIB::UDIV_I8, "__aeabi_uidiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
334 { RTLIB::UDIV_I16, "__aeabi_uidiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
335 { RTLIB::UDIV_I32, "__aeabi_uidiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
336 { RTLIB::UDIV_I64, "__aeabi_uldivmod", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
339 // RTABI chapter 4.3.4
340 { RTLIB::MEMCPY, "__aeabi_memcpy", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
341 { RTLIB::MEMMOVE, "__aeabi_memmove", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
342 { RTLIB::MEMSET, "__aeabi_memset", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
345 for (const auto &LC : LibraryCalls) {
346 setLibcallName(LC.Op, LC.Name);
347 setLibcallCallingConv(LC.Op, LC.CC);
348 if (LC.Cond != ISD::SETCC_INVALID)
349 setCmpLibcallCC(LC.Op, LC.Cond);
353 if (Subtarget->isTargetWindows()) {
354 static const struct {
355 const RTLIB::Libcall Op;
356 const char * const Name;
357 const CallingConv::ID CC;
359 { RTLIB::FPTOSINT_F32_I64, "__stoi64", CallingConv::ARM_AAPCS_VFP },
360 { RTLIB::FPTOSINT_F64_I64, "__dtoi64", CallingConv::ARM_AAPCS_VFP },
361 { RTLIB::FPTOUINT_F32_I64, "__stou64", CallingConv::ARM_AAPCS_VFP },
362 { RTLIB::FPTOUINT_F64_I64, "__dtou64", CallingConv::ARM_AAPCS_VFP },
363 { RTLIB::SINTTOFP_I64_F32, "__i64tos", CallingConv::ARM_AAPCS_VFP },
364 { RTLIB::SINTTOFP_I64_F64, "__i64tod", CallingConv::ARM_AAPCS_VFP },
365 { RTLIB::UINTTOFP_I64_F32, "__u64tos", CallingConv::ARM_AAPCS_VFP },
366 { RTLIB::UINTTOFP_I64_F64, "__u64tod", CallingConv::ARM_AAPCS_VFP },
369 for (const auto &LC : LibraryCalls) {
370 setLibcallName(LC.Op, LC.Name);
371 setLibcallCallingConv(LC.Op, LC.CC);
375 // Use divmod compiler-rt calls for iOS 5.0 and later.
376 if (Subtarget->getTargetTriple().isiOS() &&
377 !Subtarget->getTargetTriple().isOSVersionLT(5, 0)) {
378 setLibcallName(RTLIB::SDIVREM_I32, "__divmodsi4");
379 setLibcallName(RTLIB::UDIVREM_I32, "__udivmodsi4");
382 // The half <-> float conversion functions are always soft-float, but are
383 // needed for some targets which use a hard-float calling convention by
385 if (Subtarget->isAAPCS_ABI()) {
386 setLibcallCallingConv(RTLIB::FPROUND_F32_F16, CallingConv::ARM_AAPCS);
387 setLibcallCallingConv(RTLIB::FPROUND_F64_F16, CallingConv::ARM_AAPCS);
388 setLibcallCallingConv(RTLIB::FPEXT_F16_F32, CallingConv::ARM_AAPCS);
390 setLibcallCallingConv(RTLIB::FPROUND_F32_F16, CallingConv::ARM_APCS);
391 setLibcallCallingConv(RTLIB::FPROUND_F64_F16, CallingConv::ARM_APCS);
392 setLibcallCallingConv(RTLIB::FPEXT_F16_F32, CallingConv::ARM_APCS);
395 if (Subtarget->isThumb1Only())
396 addRegisterClass(MVT::i32, &ARM::tGPRRegClass);
398 addRegisterClass(MVT::i32, &ARM::GPRRegClass);
399 if (!Subtarget->useSoftFloat() && Subtarget->hasVFP2() &&
400 !Subtarget->isThumb1Only()) {
401 addRegisterClass(MVT::f32, &ARM::SPRRegClass);
402 addRegisterClass(MVT::f64, &ARM::DPRRegClass);
405 for (MVT VT : MVT::vector_valuetypes()) {
406 for (MVT InnerVT : MVT::vector_valuetypes()) {
407 setTruncStoreAction(VT, InnerVT, Expand);
408 setLoadExtAction(ISD::SEXTLOAD, VT, InnerVT, Expand);
409 setLoadExtAction(ISD::ZEXTLOAD, VT, InnerVT, Expand);
410 setLoadExtAction(ISD::EXTLOAD, VT, InnerVT, Expand);
413 setOperationAction(ISD::MULHS, VT, Expand);
414 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
415 setOperationAction(ISD::MULHU, VT, Expand);
416 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
418 setOperationAction(ISD::BSWAP, VT, Expand);
421 setOperationAction(ISD::ConstantFP, MVT::f32, Custom);
422 setOperationAction(ISD::ConstantFP, MVT::f64, Custom);
424 setOperationAction(ISD::READ_REGISTER, MVT::i64, Custom);
425 setOperationAction(ISD::WRITE_REGISTER, MVT::i64, Custom);
427 if (Subtarget->hasNEON()) {
428 addDRTypeForNEON(MVT::v2f32);
429 addDRTypeForNEON(MVT::v8i8);
430 addDRTypeForNEON(MVT::v4i16);
431 addDRTypeForNEON(MVT::v2i32);
432 addDRTypeForNEON(MVT::v1i64);
434 addQRTypeForNEON(MVT::v4f32);
435 addQRTypeForNEON(MVT::v2f64);
436 addQRTypeForNEON(MVT::v16i8);
437 addQRTypeForNEON(MVT::v8i16);
438 addQRTypeForNEON(MVT::v4i32);
439 addQRTypeForNEON(MVT::v2i64);
441 // v2f64 is legal so that QR subregs can be extracted as f64 elements, but
442 // neither Neon nor VFP support any arithmetic operations on it.
443 // The same with v4f32. But keep in mind that vadd, vsub, vmul are natively
444 // supported for v4f32.
445 setOperationAction(ISD::FADD, MVT::v2f64, Expand);
446 setOperationAction(ISD::FSUB, MVT::v2f64, Expand);
447 setOperationAction(ISD::FMUL, MVT::v2f64, Expand);
448 // FIXME: Code duplication: FDIV and FREM are expanded always, see
449 // ARMTargetLowering::addTypeForNEON method for details.
450 setOperationAction(ISD::FDIV, MVT::v2f64, Expand);
451 setOperationAction(ISD::FREM, MVT::v2f64, Expand);
452 // FIXME: Create unittest.
453 // In another words, find a way when "copysign" appears in DAG with vector
455 setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand);
456 // FIXME: Code duplication: SETCC has custom operation action, see
457 // ARMTargetLowering::addTypeForNEON method for details.
458 setOperationAction(ISD::SETCC, MVT::v2f64, Expand);
459 // FIXME: Create unittest for FNEG and for FABS.
460 setOperationAction(ISD::FNEG, MVT::v2f64, Expand);
461 setOperationAction(ISD::FABS, MVT::v2f64, Expand);
462 setOperationAction(ISD::FSQRT, MVT::v2f64, Expand);
463 setOperationAction(ISD::FSIN, MVT::v2f64, Expand);
464 setOperationAction(ISD::FCOS, MVT::v2f64, Expand);
465 setOperationAction(ISD::FPOWI, MVT::v2f64, Expand);
466 setOperationAction(ISD::FPOW, MVT::v2f64, Expand);
467 setOperationAction(ISD::FLOG, MVT::v2f64, Expand);
468 setOperationAction(ISD::FLOG2, MVT::v2f64, Expand);
469 setOperationAction(ISD::FLOG10, MVT::v2f64, Expand);
470 setOperationAction(ISD::FEXP, MVT::v2f64, Expand);
471 setOperationAction(ISD::FEXP2, MVT::v2f64, Expand);
472 // FIXME: Create unittest for FCEIL, FTRUNC, FRINT, FNEARBYINT, FFLOOR.
473 setOperationAction(ISD::FCEIL, MVT::v2f64, Expand);
474 setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand);
475 setOperationAction(ISD::FRINT, MVT::v2f64, Expand);
476 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand);
477 setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand);
478 setOperationAction(ISD::FMA, MVT::v2f64, Expand);
480 setOperationAction(ISD::FSQRT, MVT::v4f32, Expand);
481 setOperationAction(ISD::FSIN, MVT::v4f32, Expand);
482 setOperationAction(ISD::FCOS, MVT::v4f32, Expand);
483 setOperationAction(ISD::FPOWI, MVT::v4f32, Expand);
484 setOperationAction(ISD::FPOW, MVT::v4f32, Expand);
485 setOperationAction(ISD::FLOG, MVT::v4f32, Expand);
486 setOperationAction(ISD::FLOG2, MVT::v4f32, Expand);
487 setOperationAction(ISD::FLOG10, MVT::v4f32, Expand);
488 setOperationAction(ISD::FEXP, MVT::v4f32, Expand);
489 setOperationAction(ISD::FEXP2, MVT::v4f32, Expand);
490 setOperationAction(ISD::FCEIL, MVT::v4f32, Expand);
491 setOperationAction(ISD::FTRUNC, MVT::v4f32, Expand);
492 setOperationAction(ISD::FRINT, MVT::v4f32, Expand);
493 setOperationAction(ISD::FNEARBYINT, MVT::v4f32, Expand);
494 setOperationAction(ISD::FFLOOR, MVT::v4f32, Expand);
496 // Mark v2f32 intrinsics.
497 setOperationAction(ISD::FSQRT, MVT::v2f32, Expand);
498 setOperationAction(ISD::FSIN, MVT::v2f32, Expand);
499 setOperationAction(ISD::FCOS, MVT::v2f32, Expand);
500 setOperationAction(ISD::FPOWI, MVT::v2f32, Expand);
501 setOperationAction(ISD::FPOW, MVT::v2f32, Expand);
502 setOperationAction(ISD::FLOG, MVT::v2f32, Expand);
503 setOperationAction(ISD::FLOG2, MVT::v2f32, Expand);
504 setOperationAction(ISD::FLOG10, MVT::v2f32, Expand);
505 setOperationAction(ISD::FEXP, MVT::v2f32, Expand);
506 setOperationAction(ISD::FEXP2, MVT::v2f32, Expand);
507 setOperationAction(ISD::FCEIL, MVT::v2f32, Expand);
508 setOperationAction(ISD::FTRUNC, MVT::v2f32, Expand);
509 setOperationAction(ISD::FRINT, MVT::v2f32, Expand);
510 setOperationAction(ISD::FNEARBYINT, MVT::v2f32, Expand);
511 setOperationAction(ISD::FFLOOR, MVT::v2f32, Expand);
513 // Neon does not support some operations on v1i64 and v2i64 types.
514 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
515 // Custom handling for some quad-vector types to detect VMULL.
516 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
517 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
518 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
519 // Custom handling for some vector types to avoid expensive expansions
520 setOperationAction(ISD::SDIV, MVT::v4i16, Custom);
521 setOperationAction(ISD::SDIV, MVT::v8i8, Custom);
522 setOperationAction(ISD::UDIV, MVT::v4i16, Custom);
523 setOperationAction(ISD::UDIV, MVT::v8i8, Custom);
524 setOperationAction(ISD::SETCC, MVT::v1i64, Expand);
525 setOperationAction(ISD::SETCC, MVT::v2i64, Expand);
526 // Neon does not have single instruction SINT_TO_FP and UINT_TO_FP with
527 // a destination type that is wider than the source, and nor does
528 // it have a FP_TO_[SU]INT instruction with a narrower destination than
530 setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Custom);
531 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
532 setOperationAction(ISD::FP_TO_UINT, MVT::v4i16, Custom);
533 setOperationAction(ISD::FP_TO_SINT, MVT::v4i16, Custom);
535 setOperationAction(ISD::FP_ROUND, MVT::v2f32, Expand);
536 setOperationAction(ISD::FP_EXTEND, MVT::v2f64, Expand);
538 // NEON does not have single instruction CTPOP for vectors with element
539 // types wider than 8-bits. However, custom lowering can leverage the
540 // v8i8/v16i8 vcnt instruction.
541 setOperationAction(ISD::CTPOP, MVT::v2i32, Custom);
542 setOperationAction(ISD::CTPOP, MVT::v4i32, Custom);
543 setOperationAction(ISD::CTPOP, MVT::v4i16, Custom);
544 setOperationAction(ISD::CTPOP, MVT::v8i16, Custom);
546 // NEON only has FMA instructions as of VFP4.
547 if (!Subtarget->hasVFP4()) {
548 setOperationAction(ISD::FMA, MVT::v2f32, Expand);
549 setOperationAction(ISD::FMA, MVT::v4f32, Expand);
552 setTargetDAGCombine(ISD::INTRINSIC_VOID);
553 setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN);
554 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
555 setTargetDAGCombine(ISD::SHL);
556 setTargetDAGCombine(ISD::SRL);
557 setTargetDAGCombine(ISD::SRA);
558 setTargetDAGCombine(ISD::SIGN_EXTEND);
559 setTargetDAGCombine(ISD::ZERO_EXTEND);
560 setTargetDAGCombine(ISD::ANY_EXTEND);
561 setTargetDAGCombine(ISD::SELECT_CC);
562 setTargetDAGCombine(ISD::BUILD_VECTOR);
563 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
564 setTargetDAGCombine(ISD::INSERT_VECTOR_ELT);
565 setTargetDAGCombine(ISD::STORE);
566 setTargetDAGCombine(ISD::FP_TO_SINT);
567 setTargetDAGCombine(ISD::FP_TO_UINT);
568 setTargetDAGCombine(ISD::FDIV);
569 setTargetDAGCombine(ISD::LOAD);
571 // It is legal to extload from v4i8 to v4i16 or v4i32.
572 for (MVT Ty : {MVT::v8i8, MVT::v4i8, MVT::v2i8, MVT::v4i16, MVT::v2i16,
574 for (MVT VT : MVT::integer_vector_valuetypes()) {
575 setLoadExtAction(ISD::EXTLOAD, VT, Ty, Legal);
576 setLoadExtAction(ISD::ZEXTLOAD, VT, Ty, Legal);
577 setLoadExtAction(ISD::SEXTLOAD, VT, Ty, Legal);
582 // ARM and Thumb2 support UMLAL/SMLAL.
583 if (!Subtarget->isThumb1Only())
584 setTargetDAGCombine(ISD::ADDC);
586 if (Subtarget->isFPOnlySP()) {
587 // When targetting a floating-point unit with only single-precision
588 // operations, f64 is legal for the few double-precision instructions which
589 // are present However, no double-precision operations other than moves,
590 // loads and stores are provided by the hardware.
591 setOperationAction(ISD::FADD, MVT::f64, Expand);
592 setOperationAction(ISD::FSUB, MVT::f64, Expand);
593 setOperationAction(ISD::FMUL, MVT::f64, Expand);
594 setOperationAction(ISD::FMA, MVT::f64, Expand);
595 setOperationAction(ISD::FDIV, MVT::f64, Expand);
596 setOperationAction(ISD::FREM, MVT::f64, Expand);
597 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
598 setOperationAction(ISD::FGETSIGN, MVT::f64, Expand);
599 setOperationAction(ISD::FNEG, MVT::f64, Expand);
600 setOperationAction(ISD::FABS, MVT::f64, Expand);
601 setOperationAction(ISD::FSQRT, MVT::f64, Expand);
602 setOperationAction(ISD::FSIN, MVT::f64, Expand);
603 setOperationAction(ISD::FCOS, MVT::f64, Expand);
604 setOperationAction(ISD::FPOWI, MVT::f64, Expand);
605 setOperationAction(ISD::FPOW, MVT::f64, Expand);
606 setOperationAction(ISD::FLOG, MVT::f64, Expand);
607 setOperationAction(ISD::FLOG2, MVT::f64, Expand);
608 setOperationAction(ISD::FLOG10, MVT::f64, Expand);
609 setOperationAction(ISD::FEXP, MVT::f64, Expand);
610 setOperationAction(ISD::FEXP2, MVT::f64, Expand);
611 setOperationAction(ISD::FCEIL, MVT::f64, Expand);
612 setOperationAction(ISD::FTRUNC, MVT::f64, Expand);
613 setOperationAction(ISD::FRINT, MVT::f64, Expand);
614 setOperationAction(ISD::FNEARBYINT, MVT::f64, Expand);
615 setOperationAction(ISD::FFLOOR, MVT::f64, Expand);
616 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
617 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
618 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
619 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
620 setOperationAction(ISD::FP_TO_SINT, MVT::f64, Custom);
621 setOperationAction(ISD::FP_TO_UINT, MVT::f64, Custom);
622 setOperationAction(ISD::FP_ROUND, MVT::f32, Custom);
623 setOperationAction(ISD::FP_EXTEND, MVT::f64, Custom);
626 computeRegisterProperties(Subtarget->getRegisterInfo());
628 // ARM does not have floating-point extending loads.
629 for (MVT VT : MVT::fp_valuetypes()) {
630 setLoadExtAction(ISD::EXTLOAD, VT, MVT::f32, Expand);
631 setLoadExtAction(ISD::EXTLOAD, VT, MVT::f16, Expand);
634 // ... or truncating stores
635 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
636 setTruncStoreAction(MVT::f32, MVT::f16, Expand);
637 setTruncStoreAction(MVT::f64, MVT::f16, Expand);
639 // ARM does not have i1 sign extending load.
640 for (MVT VT : MVT::integer_valuetypes())
641 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Promote);
643 // ARM supports all 4 flavors of integer indexed load / store.
644 if (!Subtarget->isThumb1Only()) {
645 for (unsigned im = (unsigned)ISD::PRE_INC;
646 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
647 setIndexedLoadAction(im, MVT::i1, Legal);
648 setIndexedLoadAction(im, MVT::i8, Legal);
649 setIndexedLoadAction(im, MVT::i16, Legal);
650 setIndexedLoadAction(im, MVT::i32, Legal);
651 setIndexedStoreAction(im, MVT::i1, Legal);
652 setIndexedStoreAction(im, MVT::i8, Legal);
653 setIndexedStoreAction(im, MVT::i16, Legal);
654 setIndexedStoreAction(im, MVT::i32, Legal);
658 setOperationAction(ISD::SADDO, MVT::i32, Custom);
659 setOperationAction(ISD::UADDO, MVT::i32, Custom);
660 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
661 setOperationAction(ISD::USUBO, MVT::i32, Custom);
663 // i64 operation support.
664 setOperationAction(ISD::MUL, MVT::i64, Expand);
665 setOperationAction(ISD::MULHU, MVT::i32, Expand);
666 if (Subtarget->isThumb1Only()) {
667 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
668 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
670 if (Subtarget->isThumb1Only() || !Subtarget->hasV6Ops()
671 || (Subtarget->isThumb2() && !Subtarget->hasThumb2DSP()))
672 setOperationAction(ISD::MULHS, MVT::i32, Expand);
674 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
675 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
676 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
677 setOperationAction(ISD::SRL, MVT::i64, Custom);
678 setOperationAction(ISD::SRA, MVT::i64, Custom);
680 if (!Subtarget->isThumb1Only()) {
681 // FIXME: We should do this for Thumb1 as well.
682 setOperationAction(ISD::ADDC, MVT::i32, Custom);
683 setOperationAction(ISD::ADDE, MVT::i32, Custom);
684 setOperationAction(ISD::SUBC, MVT::i32, Custom);
685 setOperationAction(ISD::SUBE, MVT::i32, Custom);
688 // ARM does not have ROTL.
689 setOperationAction(ISD::ROTL, MVT::i32, Expand);
690 setOperationAction(ISD::CTTZ, MVT::i32, Custom);
691 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
692 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
693 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
695 // These just redirect to CTTZ and CTLZ on ARM.
696 setOperationAction(ISD::CTTZ_ZERO_UNDEF , MVT::i32 , Expand);
697 setOperationAction(ISD::CTLZ_ZERO_UNDEF , MVT::i32 , Expand);
699 setOperationAction(ISD::READCYCLECOUNTER, MVT::i64, Custom);
701 // Only ARMv6 has BSWAP.
702 if (!Subtarget->hasV6Ops())
703 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
705 if (!(Subtarget->hasDivide() && Subtarget->isThumb2()) &&
706 !(Subtarget->hasDivideInARMMode() && !Subtarget->isThumb())) {
707 // These are expanded into libcalls if the cpu doesn't have HW divider.
708 setOperationAction(ISD::SDIV, MVT::i32, Expand);
709 setOperationAction(ISD::UDIV, MVT::i32, Expand);
712 // FIXME: Also set divmod for SREM on EABI
713 setOperationAction(ISD::SREM, MVT::i32, Expand);
714 setOperationAction(ISD::UREM, MVT::i32, Expand);
715 // Register based DivRem for AEABI (RTABI 4.2)
716 if (Subtarget->isTargetAEABI()) {
717 setLibcallName(RTLIB::SDIVREM_I8, "__aeabi_idivmod");
718 setLibcallName(RTLIB::SDIVREM_I16, "__aeabi_idivmod");
719 setLibcallName(RTLIB::SDIVREM_I32, "__aeabi_idivmod");
720 setLibcallName(RTLIB::SDIVREM_I64, "__aeabi_ldivmod");
721 setLibcallName(RTLIB::UDIVREM_I8, "__aeabi_uidivmod");
722 setLibcallName(RTLIB::UDIVREM_I16, "__aeabi_uidivmod");
723 setLibcallName(RTLIB::UDIVREM_I32, "__aeabi_uidivmod");
724 setLibcallName(RTLIB::UDIVREM_I64, "__aeabi_uldivmod");
726 setLibcallCallingConv(RTLIB::SDIVREM_I8, CallingConv::ARM_AAPCS);
727 setLibcallCallingConv(RTLIB::SDIVREM_I16, CallingConv::ARM_AAPCS);
728 setLibcallCallingConv(RTLIB::SDIVREM_I32, CallingConv::ARM_AAPCS);
729 setLibcallCallingConv(RTLIB::SDIVREM_I64, CallingConv::ARM_AAPCS);
730 setLibcallCallingConv(RTLIB::UDIVREM_I8, CallingConv::ARM_AAPCS);
731 setLibcallCallingConv(RTLIB::UDIVREM_I16, CallingConv::ARM_AAPCS);
732 setLibcallCallingConv(RTLIB::UDIVREM_I32, CallingConv::ARM_AAPCS);
733 setLibcallCallingConv(RTLIB::UDIVREM_I64, CallingConv::ARM_AAPCS);
735 setOperationAction(ISD::SDIVREM, MVT::i32, Custom);
736 setOperationAction(ISD::UDIVREM, MVT::i32, Custom);
738 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
739 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
742 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
743 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
744 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
745 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
746 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
748 setOperationAction(ISD::TRAP, MVT::Other, Legal);
750 // Use the default implementation.
751 setOperationAction(ISD::VASTART, MVT::Other, Custom);
752 setOperationAction(ISD::VAARG, MVT::Other, Expand);
753 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
754 setOperationAction(ISD::VAEND, MVT::Other, Expand);
755 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
756 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
758 if (!Subtarget->isTargetMachO()) {
759 // Non-MachO platforms may return values in these registers via the
760 // personality function.
761 setExceptionPointerRegister(ARM::R0);
762 setExceptionSelectorRegister(ARM::R1);
765 if (Subtarget->getTargetTriple().isWindowsItaniumEnvironment())
766 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
768 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
770 // ARMv6 Thumb1 (except for CPUs that support dmb / dsb) and earlier use
771 // the default expansion. If we are targeting a single threaded system,
772 // then set them all for expand so we can lower them later into their
774 if (TM.Options.ThreadModel == ThreadModel::Single)
775 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Expand);
776 else if (Subtarget->hasAnyDataBarrier() && !Subtarget->isThumb1Only()) {
777 // ATOMIC_FENCE needs custom lowering; the others should have been expanded
778 // to ldrex/strex loops already.
779 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Custom);
781 // On v8, we have particularly efficient implementations of atomic fences
782 // if they can be combined with nearby atomic loads and stores.
783 if (!Subtarget->hasV8Ops()) {
784 // Automatically insert fences (dmb ish) around ATOMIC_SWAP etc.
785 setInsertFencesForAtomic(true);
788 // If there's anything we can use as a barrier, go through custom lowering
790 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other,
791 Subtarget->hasAnyDataBarrier() ? Custom : Expand);
793 // Set them all for expansion, which will force libcalls.
794 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Expand);
795 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Expand);
796 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Expand);
797 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Expand);
798 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Expand);
799 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Expand);
800 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Expand);
801 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Expand);
802 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i32, Expand);
803 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i32, Expand);
804 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i32, Expand);
805 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i32, Expand);
806 // Mark ATOMIC_LOAD and ATOMIC_STORE custom so we can handle the
807 // Unordered/Monotonic case.
808 setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Custom);
809 setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Custom);
812 setOperationAction(ISD::PREFETCH, MVT::Other, Custom);
814 // Requires SXTB/SXTH, available on v6 and up in both ARM and Thumb modes.
815 if (!Subtarget->hasV6Ops()) {
816 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
817 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
819 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
821 if (!Subtarget->useSoftFloat() && Subtarget->hasVFP2() &&
822 !Subtarget->isThumb1Only()) {
823 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
824 // iff target supports vfp2.
825 setOperationAction(ISD::BITCAST, MVT::i64, Custom);
826 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
829 // We want to custom lower some of our intrinsics.
830 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
831 if (Subtarget->isTargetDarwin()) {
832 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
833 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
834 setLibcallName(RTLIB::UNWIND_RESUME, "_Unwind_SjLj_Resume");
837 setOperationAction(ISD::SETCC, MVT::i32, Expand);
838 setOperationAction(ISD::SETCC, MVT::f32, Expand);
839 setOperationAction(ISD::SETCC, MVT::f64, Expand);
840 setOperationAction(ISD::SELECT, MVT::i32, Custom);
841 setOperationAction(ISD::SELECT, MVT::f32, Custom);
842 setOperationAction(ISD::SELECT, MVT::f64, Custom);
843 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
844 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
845 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
847 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
848 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
849 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
850 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
851 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
853 // We don't support sin/cos/fmod/copysign/pow
854 setOperationAction(ISD::FSIN, MVT::f64, Expand);
855 setOperationAction(ISD::FSIN, MVT::f32, Expand);
856 setOperationAction(ISD::FCOS, MVT::f32, Expand);
857 setOperationAction(ISD::FCOS, MVT::f64, Expand);
858 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
859 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
860 setOperationAction(ISD::FREM, MVT::f64, Expand);
861 setOperationAction(ISD::FREM, MVT::f32, Expand);
862 if (!Subtarget->useSoftFloat() && Subtarget->hasVFP2() &&
863 !Subtarget->isThumb1Only()) {
864 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
865 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
867 setOperationAction(ISD::FPOW, MVT::f64, Expand);
868 setOperationAction(ISD::FPOW, MVT::f32, Expand);
870 if (!Subtarget->hasVFP4()) {
871 setOperationAction(ISD::FMA, MVT::f64, Expand);
872 setOperationAction(ISD::FMA, MVT::f32, Expand);
875 // Various VFP goodness
876 if (!Subtarget->useSoftFloat() && !Subtarget->isThumb1Only()) {
877 // FP-ARMv8 adds f64 <-> f16 conversion. Before that it should be expanded.
878 if (!Subtarget->hasFPARMv8() || Subtarget->isFPOnlySP()) {
879 setOperationAction(ISD::FP16_TO_FP, MVT::f64, Expand);
880 setOperationAction(ISD::FP_TO_FP16, MVT::f64, Expand);
883 // fp16 is a special v7 extension that adds f16 <-> f32 conversions.
884 if (!Subtarget->hasFP16()) {
885 setOperationAction(ISD::FP16_TO_FP, MVT::f32, Expand);
886 setOperationAction(ISD::FP_TO_FP16, MVT::f32, Expand);
890 // Combine sin / cos into one node or libcall if possible.
891 if (Subtarget->hasSinCos()) {
892 setLibcallName(RTLIB::SINCOS_F32, "sincosf");
893 setLibcallName(RTLIB::SINCOS_F64, "sincos");
894 if (Subtarget->getTargetTriple().isiOS()) {
895 // For iOS, we don't want to the normal expansion of a libcall to
896 // sincos. We want to issue a libcall to __sincos_stret.
897 setOperationAction(ISD::FSINCOS, MVT::f64, Custom);
898 setOperationAction(ISD::FSINCOS, MVT::f32, Custom);
902 // FP-ARMv8 implements a lot of rounding-like FP operations.
903 if (Subtarget->hasFPARMv8()) {
904 setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
905 setOperationAction(ISD::FCEIL, MVT::f32, Legal);
906 setOperationAction(ISD::FROUND, MVT::f32, Legal);
907 setOperationAction(ISD::FTRUNC, MVT::f32, Legal);
908 setOperationAction(ISD::FNEARBYINT, MVT::f32, Legal);
909 setOperationAction(ISD::FRINT, MVT::f32, Legal);
910 if (!Subtarget->isFPOnlySP()) {
911 setOperationAction(ISD::FFLOOR, MVT::f64, Legal);
912 setOperationAction(ISD::FCEIL, MVT::f64, Legal);
913 setOperationAction(ISD::FROUND, MVT::f64, Legal);
914 setOperationAction(ISD::FTRUNC, MVT::f64, Legal);
915 setOperationAction(ISD::FNEARBYINT, MVT::f64, Legal);
916 setOperationAction(ISD::FRINT, MVT::f64, Legal);
919 // We have target-specific dag combine patterns for the following nodes:
920 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
921 setTargetDAGCombine(ISD::ADD);
922 setTargetDAGCombine(ISD::SUB);
923 setTargetDAGCombine(ISD::MUL);
924 setTargetDAGCombine(ISD::AND);
925 setTargetDAGCombine(ISD::OR);
926 setTargetDAGCombine(ISD::XOR);
928 if (Subtarget->hasV6Ops())
929 setTargetDAGCombine(ISD::SRL);
931 setStackPointerRegisterToSaveRestore(ARM::SP);
933 if (Subtarget->useSoftFloat() || Subtarget->isThumb1Only() ||
934 !Subtarget->hasVFP2())
935 setSchedulingPreference(Sched::RegPressure);
937 setSchedulingPreference(Sched::Hybrid);
939 //// temporary - rewrite interface to use type
940 MaxStoresPerMemset = 8;
941 MaxStoresPerMemsetOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
942 MaxStoresPerMemcpy = 4; // For @llvm.memcpy -> sequence of stores
943 MaxStoresPerMemcpyOptSize = Subtarget->isTargetDarwin() ? 4 : 2;
944 MaxStoresPerMemmove = 4; // For @llvm.memmove -> sequence of stores
945 MaxStoresPerMemmoveOptSize = Subtarget->isTargetDarwin() ? 4 : 2;
947 // On ARM arguments smaller than 4 bytes are extended, so all arguments
948 // are at least 4 bytes aligned.
949 setMinStackArgumentAlignment(4);
951 // Prefer likely predicted branches to selects on out-of-order cores.
952 PredictableSelectIsExpensive = Subtarget->isLikeA9();
954 setMinFunctionAlignment(Subtarget->isThumb() ? 1 : 2);
957 bool ARMTargetLowering::useSoftFloat() const {
958 return Subtarget->useSoftFloat();
961 // FIXME: It might make sense to define the representative register class as the
962 // nearest super-register that has a non-null superset. For example, DPR_VFP2 is
963 // a super-register of SPR, and DPR is a superset if DPR_VFP2. Consequently,
964 // SPR's representative would be DPR_VFP2. This should work well if register
965 // pressure tracking were modified such that a register use would increment the
966 // pressure of the register class's representative and all of it's super
967 // classes' representatives transitively. We have not implemented this because
968 // of the difficulty prior to coalescing of modeling operand register classes
969 // due to the common occurrence of cross class copies and subregister insertions
971 std::pair<const TargetRegisterClass *, uint8_t>
972 ARMTargetLowering::findRepresentativeClass(const TargetRegisterInfo *TRI,
974 const TargetRegisterClass *RRC = nullptr;
976 switch (VT.SimpleTy) {
978 return TargetLowering::findRepresentativeClass(TRI, VT);
979 // Use DPR as representative register class for all floating point
980 // and vector types. Since there are 32 SPR registers and 32 DPR registers so
981 // the cost is 1 for both f32 and f64.
982 case MVT::f32: case MVT::f64: case MVT::v8i8: case MVT::v4i16:
983 case MVT::v2i32: case MVT::v1i64: case MVT::v2f32:
984 RRC = &ARM::DPRRegClass;
985 // When NEON is used for SP, only half of the register file is available
986 // because operations that define both SP and DP results will be constrained
987 // to the VFP2 class (D0-D15). We currently model this constraint prior to
988 // coalescing by double-counting the SP regs. See the FIXME above.
989 if (Subtarget->useNEONForSinglePrecisionFP())
992 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
993 case MVT::v4f32: case MVT::v2f64:
994 RRC = &ARM::DPRRegClass;
998 RRC = &ARM::DPRRegClass;
1002 RRC = &ARM::DPRRegClass;
1006 return std::make_pair(RRC, Cost);
1009 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
1010 switch ((ARMISD::NodeType)Opcode) {
1011 case ARMISD::FIRST_NUMBER: break;
1012 case ARMISD::Wrapper: return "ARMISD::Wrapper";
1013 case ARMISD::WrapperPIC: return "ARMISD::WrapperPIC";
1014 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
1015 case ARMISD::COPY_STRUCT_BYVAL: return "ARMISD::COPY_STRUCT_BYVAL";
1016 case ARMISD::CALL: return "ARMISD::CALL";
1017 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
1018 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
1019 case ARMISD::tCALL: return "ARMISD::tCALL";
1020 case ARMISD::BRCOND: return "ARMISD::BRCOND";
1021 case ARMISD::BR_JT: return "ARMISD::BR_JT";
1022 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
1023 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
1024 case ARMISD::INTRET_FLAG: return "ARMISD::INTRET_FLAG";
1025 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
1026 case ARMISD::CMP: return "ARMISD::CMP";
1027 case ARMISD::CMN: return "ARMISD::CMN";
1028 case ARMISD::CMPZ: return "ARMISD::CMPZ";
1029 case ARMISD::CMPFP: return "ARMISD::CMPFP";
1030 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
1031 case ARMISD::BCC_i64: return "ARMISD::BCC_i64";
1032 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
1034 case ARMISD::CMOV: return "ARMISD::CMOV";
1036 case ARMISD::RBIT: return "ARMISD::RBIT";
1038 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
1039 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
1040 case ARMISD::RRX: return "ARMISD::RRX";
1042 case ARMISD::ADDC: return "ARMISD::ADDC";
1043 case ARMISD::ADDE: return "ARMISD::ADDE";
1044 case ARMISD::SUBC: return "ARMISD::SUBC";
1045 case ARMISD::SUBE: return "ARMISD::SUBE";
1047 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
1048 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
1050 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
1051 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
1053 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN";
1055 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
1057 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
1059 case ARMISD::MEMBARRIER_MCR: return "ARMISD::MEMBARRIER_MCR";
1061 case ARMISD::PRELOAD: return "ARMISD::PRELOAD";
1063 case ARMISD::WIN__CHKSTK: return "ARMISD:::WIN__CHKSTK";
1065 case ARMISD::VCEQ: return "ARMISD::VCEQ";
1066 case ARMISD::VCEQZ: return "ARMISD::VCEQZ";
1067 case ARMISD::VCGE: return "ARMISD::VCGE";
1068 case ARMISD::VCGEZ: return "ARMISD::VCGEZ";
1069 case ARMISD::VCLEZ: return "ARMISD::VCLEZ";
1070 case ARMISD::VCGEU: return "ARMISD::VCGEU";
1071 case ARMISD::VCGT: return "ARMISD::VCGT";
1072 case ARMISD::VCGTZ: return "ARMISD::VCGTZ";
1073 case ARMISD::VCLTZ: return "ARMISD::VCLTZ";
1074 case ARMISD::VCGTU: return "ARMISD::VCGTU";
1075 case ARMISD::VTST: return "ARMISD::VTST";
1077 case ARMISD::VSHL: return "ARMISD::VSHL";
1078 case ARMISD::VSHRs: return "ARMISD::VSHRs";
1079 case ARMISD::VSHRu: return "ARMISD::VSHRu";
1080 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
1081 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
1082 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
1083 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
1084 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
1085 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
1086 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
1087 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
1088 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
1089 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
1090 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
1091 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
1092 case ARMISD::VSLI: return "ARMISD::VSLI";
1093 case ARMISD::VSRI: return "ARMISD::VSRI";
1094 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
1095 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
1096 case ARMISD::VMOVIMM: return "ARMISD::VMOVIMM";
1097 case ARMISD::VMVNIMM: return "ARMISD::VMVNIMM";
1098 case ARMISD::VMOVFPIMM: return "ARMISD::VMOVFPIMM";
1099 case ARMISD::VDUP: return "ARMISD::VDUP";
1100 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
1101 case ARMISD::VEXT: return "ARMISD::VEXT";
1102 case ARMISD::VREV64: return "ARMISD::VREV64";
1103 case ARMISD::VREV32: return "ARMISD::VREV32";
1104 case ARMISD::VREV16: return "ARMISD::VREV16";
1105 case ARMISD::VZIP: return "ARMISD::VZIP";
1106 case ARMISD::VUZP: return "ARMISD::VUZP";
1107 case ARMISD::VTRN: return "ARMISD::VTRN";
1108 case ARMISD::VTBL1: return "ARMISD::VTBL1";
1109 case ARMISD::VTBL2: return "ARMISD::VTBL2";
1110 case ARMISD::VMULLs: return "ARMISD::VMULLs";
1111 case ARMISD::VMULLu: return "ARMISD::VMULLu";
1112 case ARMISD::UMLAL: return "ARMISD::UMLAL";
1113 case ARMISD::SMLAL: return "ARMISD::SMLAL";
1114 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
1115 case ARMISD::FMAX: return "ARMISD::FMAX";
1116 case ARMISD::FMIN: return "ARMISD::FMIN";
1117 case ARMISD::VMAXNM: return "ARMISD::VMAX";
1118 case ARMISD::VMINNM: return "ARMISD::VMIN";
1119 case ARMISD::BFI: return "ARMISD::BFI";
1120 case ARMISD::VORRIMM: return "ARMISD::VORRIMM";
1121 case ARMISD::VBICIMM: return "ARMISD::VBICIMM";
1122 case ARMISD::VBSL: return "ARMISD::VBSL";
1123 case ARMISD::VLD2DUP: return "ARMISD::VLD2DUP";
1124 case ARMISD::VLD3DUP: return "ARMISD::VLD3DUP";
1125 case ARMISD::VLD4DUP: return "ARMISD::VLD4DUP";
1126 case ARMISD::VLD1_UPD: return "ARMISD::VLD1_UPD";
1127 case ARMISD::VLD2_UPD: return "ARMISD::VLD2_UPD";
1128 case ARMISD::VLD3_UPD: return "ARMISD::VLD3_UPD";
1129 case ARMISD::VLD4_UPD: return "ARMISD::VLD4_UPD";
1130 case ARMISD::VLD2LN_UPD: return "ARMISD::VLD2LN_UPD";
1131 case ARMISD::VLD3LN_UPD: return "ARMISD::VLD3LN_UPD";
1132 case ARMISD::VLD4LN_UPD: return "ARMISD::VLD4LN_UPD";
1133 case ARMISD::VLD2DUP_UPD: return "ARMISD::VLD2DUP_UPD";
1134 case ARMISD::VLD3DUP_UPD: return "ARMISD::VLD3DUP_UPD";
1135 case ARMISD::VLD4DUP_UPD: return "ARMISD::VLD4DUP_UPD";
1136 case ARMISD::VST1_UPD: return "ARMISD::VST1_UPD";
1137 case ARMISD::VST2_UPD: return "ARMISD::VST2_UPD";
1138 case ARMISD::VST3_UPD: return "ARMISD::VST3_UPD";
1139 case ARMISD::VST4_UPD: return "ARMISD::VST4_UPD";
1140 case ARMISD::VST2LN_UPD: return "ARMISD::VST2LN_UPD";
1141 case ARMISD::VST3LN_UPD: return "ARMISD::VST3LN_UPD";
1142 case ARMISD::VST4LN_UPD: return "ARMISD::VST4LN_UPD";
1147 EVT ARMTargetLowering::getSetCCResultType(const DataLayout &DL, LLVMContext &,
1150 return getPointerTy(DL);
1151 return VT.changeVectorElementTypeToInteger();
1154 /// getRegClassFor - Return the register class that should be used for the
1155 /// specified value type.
1156 const TargetRegisterClass *ARMTargetLowering::getRegClassFor(MVT VT) const {
1157 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
1158 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
1159 // load / store 4 to 8 consecutive D registers.
1160 if (Subtarget->hasNEON()) {
1161 if (VT == MVT::v4i64)
1162 return &ARM::QQPRRegClass;
1163 if (VT == MVT::v8i64)
1164 return &ARM::QQQQPRRegClass;
1166 return TargetLowering::getRegClassFor(VT);
1169 // memcpy, and other memory intrinsics, typically tries to use LDM/STM if the
1170 // source/dest is aligned and the copy size is large enough. We therefore want
1171 // to align such objects passed to memory intrinsics.
1172 bool ARMTargetLowering::shouldAlignPointerArgs(CallInst *CI, unsigned &MinSize,
1173 unsigned &PrefAlign) const {
1174 if (!isa<MemIntrinsic>(CI))
1177 // On ARM11 onwards (excluding M class) 8-byte aligned LDM is typically 1
1178 // cycle faster than 4-byte aligned LDM.
1179 PrefAlign = (Subtarget->hasV6Ops() && !Subtarget->isMClass() ? 8 : 4);
1183 // Create a fast isel object.
1185 ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
1186 const TargetLibraryInfo *libInfo) const {
1187 return ARM::createFastISel(funcInfo, libInfo);
1190 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
1191 unsigned NumVals = N->getNumValues();
1193 return Sched::RegPressure;
1195 for (unsigned i = 0; i != NumVals; ++i) {
1196 EVT VT = N->getValueType(i);
1197 if (VT == MVT::Glue || VT == MVT::Other)
1199 if (VT.isFloatingPoint() || VT.isVector())
1203 if (!N->isMachineOpcode())
1204 return Sched::RegPressure;
1206 // Load are scheduled for latency even if there instruction itinerary
1207 // is not available.
1208 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
1209 const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
1211 if (MCID.getNumDefs() == 0)
1212 return Sched::RegPressure;
1213 if (!Itins->isEmpty() &&
1214 Itins->getOperandCycle(MCID.getSchedClass(), 0) > 2)
1217 return Sched::RegPressure;
1220 //===----------------------------------------------------------------------===//
1222 //===----------------------------------------------------------------------===//
1224 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
1225 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
1227 default: llvm_unreachable("Unknown condition code!");
1228 case ISD::SETNE: return ARMCC::NE;
1229 case ISD::SETEQ: return ARMCC::EQ;
1230 case ISD::SETGT: return ARMCC::GT;
1231 case ISD::SETGE: return ARMCC::GE;
1232 case ISD::SETLT: return ARMCC::LT;
1233 case ISD::SETLE: return ARMCC::LE;
1234 case ISD::SETUGT: return ARMCC::HI;
1235 case ISD::SETUGE: return ARMCC::HS;
1236 case ISD::SETULT: return ARMCC::LO;
1237 case ISD::SETULE: return ARMCC::LS;
1241 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
1242 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
1243 ARMCC::CondCodes &CondCode2) {
1244 CondCode2 = ARMCC::AL;
1246 default: llvm_unreachable("Unknown FP condition!");
1248 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
1250 case ISD::SETOGT: CondCode = ARMCC::GT; break;
1252 case ISD::SETOGE: CondCode = ARMCC::GE; break;
1253 case ISD::SETOLT: CondCode = ARMCC::MI; break;
1254 case ISD::SETOLE: CondCode = ARMCC::LS; break;
1255 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
1256 case ISD::SETO: CondCode = ARMCC::VC; break;
1257 case ISD::SETUO: CondCode = ARMCC::VS; break;
1258 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
1259 case ISD::SETUGT: CondCode = ARMCC::HI; break;
1260 case ISD::SETUGE: CondCode = ARMCC::PL; break;
1262 case ISD::SETULT: CondCode = ARMCC::LT; break;
1264 case ISD::SETULE: CondCode = ARMCC::LE; break;
1266 case ISD::SETUNE: CondCode = ARMCC::NE; break;
1270 //===----------------------------------------------------------------------===//
1271 // Calling Convention Implementation
1272 //===----------------------------------------------------------------------===//
1274 #include "ARMGenCallingConv.inc"
1276 /// getEffectiveCallingConv - Get the effective calling convention, taking into
1277 /// account presence of floating point hardware and calling convention
1278 /// limitations, such as support for variadic functions.
1280 ARMTargetLowering::getEffectiveCallingConv(CallingConv::ID CC,
1281 bool isVarArg) const {
1284 llvm_unreachable("Unsupported calling convention");
1285 case CallingConv::ARM_AAPCS:
1286 case CallingConv::ARM_APCS:
1287 case CallingConv::GHC:
1289 case CallingConv::ARM_AAPCS_VFP:
1290 return isVarArg ? CallingConv::ARM_AAPCS : CallingConv::ARM_AAPCS_VFP;
1291 case CallingConv::C:
1292 if (!Subtarget->isAAPCS_ABI())
1293 return CallingConv::ARM_APCS;
1294 else if (Subtarget->hasVFP2() && !Subtarget->isThumb1Only() &&
1295 getTargetMachine().Options.FloatABIType == FloatABI::Hard &&
1297 return CallingConv::ARM_AAPCS_VFP;
1299 return CallingConv::ARM_AAPCS;
1300 case CallingConv::Fast:
1301 if (!Subtarget->isAAPCS_ABI()) {
1302 if (Subtarget->hasVFP2() && !Subtarget->isThumb1Only() && !isVarArg)
1303 return CallingConv::Fast;
1304 return CallingConv::ARM_APCS;
1305 } else if (Subtarget->hasVFP2() && !Subtarget->isThumb1Only() && !isVarArg)
1306 return CallingConv::ARM_AAPCS_VFP;
1308 return CallingConv::ARM_AAPCS;
1312 /// CCAssignFnForNode - Selects the correct CCAssignFn for the given
1313 /// CallingConvention.
1314 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
1316 bool isVarArg) const {
1317 switch (getEffectiveCallingConv(CC, isVarArg)) {
1319 llvm_unreachable("Unsupported calling convention");
1320 case CallingConv::ARM_APCS:
1321 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1322 case CallingConv::ARM_AAPCS:
1323 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1324 case CallingConv::ARM_AAPCS_VFP:
1325 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1326 case CallingConv::Fast:
1327 return (Return ? RetFastCC_ARM_APCS : FastCC_ARM_APCS);
1328 case CallingConv::GHC:
1329 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS_GHC);
1333 /// LowerCallResult - Lower the result values of a call into the
1334 /// appropriate copies out of appropriate physical registers.
1336 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1337 CallingConv::ID CallConv, bool isVarArg,
1338 const SmallVectorImpl<ISD::InputArg> &Ins,
1339 SDLoc dl, SelectionDAG &DAG,
1340 SmallVectorImpl<SDValue> &InVals,
1341 bool isThisReturn, SDValue ThisVal) const {
1343 // Assign locations to each value returned by this call.
1344 SmallVector<CCValAssign, 16> RVLocs;
1345 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
1346 *DAG.getContext(), Call);
1347 CCInfo.AnalyzeCallResult(Ins,
1348 CCAssignFnForNode(CallConv, /* Return*/ true,
1351 // Copy all of the result registers out of their specified physreg.
1352 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1353 CCValAssign VA = RVLocs[i];
1355 // Pass 'this' value directly from the argument to return value, to avoid
1356 // reg unit interference
1357 if (i == 0 && isThisReturn) {
1358 assert(!VA.needsCustom() && VA.getLocVT() == MVT::i32 &&
1359 "unexpected return calling convention register assignment");
1360 InVals.push_back(ThisVal);
1365 if (VA.needsCustom()) {
1366 // Handle f64 or half of a v2f64.
1367 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1369 Chain = Lo.getValue(1);
1370 InFlag = Lo.getValue(2);
1371 VA = RVLocs[++i]; // skip ahead to next loc
1372 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1374 Chain = Hi.getValue(1);
1375 InFlag = Hi.getValue(2);
1376 if (!Subtarget->isLittle())
1378 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1380 if (VA.getLocVT() == MVT::v2f64) {
1381 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1382 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1383 DAG.getConstant(0, dl, MVT::i32));
1385 VA = RVLocs[++i]; // skip ahead to next loc
1386 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1387 Chain = Lo.getValue(1);
1388 InFlag = Lo.getValue(2);
1389 VA = RVLocs[++i]; // skip ahead to next loc
1390 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1391 Chain = Hi.getValue(1);
1392 InFlag = Hi.getValue(2);
1393 if (!Subtarget->isLittle())
1395 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1396 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1397 DAG.getConstant(1, dl, MVT::i32));
1400 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
1402 Chain = Val.getValue(1);
1403 InFlag = Val.getValue(2);
1406 switch (VA.getLocInfo()) {
1407 default: llvm_unreachable("Unknown loc info!");
1408 case CCValAssign::Full: break;
1409 case CCValAssign::BCvt:
1410 Val = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), Val);
1414 InVals.push_back(Val);
1420 /// LowerMemOpCallTo - Store the argument to the stack.
1422 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
1423 SDValue StackPtr, SDValue Arg,
1424 SDLoc dl, SelectionDAG &DAG,
1425 const CCValAssign &VA,
1426 ISD::ArgFlagsTy Flags) const {
1427 unsigned LocMemOffset = VA.getLocMemOffset();
1428 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset, dl);
1429 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(DAG.getDataLayout()),
1431 return DAG.getStore(Chain, dl, Arg, PtrOff,
1432 MachinePointerInfo::getStack(LocMemOffset),
1436 void ARMTargetLowering::PassF64ArgInRegs(SDLoc dl, SelectionDAG &DAG,
1437 SDValue Chain, SDValue &Arg,
1438 RegsToPassVector &RegsToPass,
1439 CCValAssign &VA, CCValAssign &NextVA,
1441 SmallVectorImpl<SDValue> &MemOpChains,
1442 ISD::ArgFlagsTy Flags) const {
1444 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1445 DAG.getVTList(MVT::i32, MVT::i32), Arg);
1446 unsigned id = Subtarget->isLittle() ? 0 : 1;
1447 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd.getValue(id)));
1449 if (NextVA.isRegLoc())
1450 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1-id)));
1452 assert(NextVA.isMemLoc());
1453 if (!StackPtr.getNode())
1454 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP,
1455 getPointerTy(DAG.getDataLayout()));
1457 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1-id),
1463 /// LowerCall - Lowering a call into a callseq_start <-
1464 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
1467 ARMTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
1468 SmallVectorImpl<SDValue> &InVals) const {
1469 SelectionDAG &DAG = CLI.DAG;
1471 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
1472 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
1473 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
1474 SDValue Chain = CLI.Chain;
1475 SDValue Callee = CLI.Callee;
1476 bool &isTailCall = CLI.IsTailCall;
1477 CallingConv::ID CallConv = CLI.CallConv;
1478 bool doesNotRet = CLI.DoesNotReturn;
1479 bool isVarArg = CLI.IsVarArg;
1481 MachineFunction &MF = DAG.getMachineFunction();
1482 bool isStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1483 bool isThisReturn = false;
1484 bool isSibCall = false;
1485 auto Attr = MF.getFunction()->getFnAttribute("disable-tail-calls");
1487 // Disable tail calls if they're not supported.
1488 if (!Subtarget->supportsTailCall() || Attr.getValueAsString() == "true")
1492 // Check if it's really possible to do a tail call.
1493 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1494 isVarArg, isStructRet, MF.getFunction()->hasStructRetAttr(),
1495 Outs, OutVals, Ins, DAG);
1496 if (!isTailCall && CLI.CS && CLI.CS->isMustTailCall())
1497 report_fatal_error("failed to perform tail call elimination on a call "
1498 "site marked musttail");
1499 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1500 // detected sibcalls.
1507 // Analyze operands of the call, assigning locations to each operand.
1508 SmallVector<CCValAssign, 16> ArgLocs;
1509 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs,
1510 *DAG.getContext(), Call);
1511 CCInfo.AnalyzeCallOperands(Outs,
1512 CCAssignFnForNode(CallConv, /* Return*/ false,
1515 // Get a count of how many bytes are to be pushed on the stack.
1516 unsigned NumBytes = CCInfo.getNextStackOffset();
1518 // For tail calls, memory operands are available in our caller's stack.
1522 // Adjust the stack pointer for the new arguments...
1523 // These operations are automatically eliminated by the prolog/epilog pass
1525 Chain = DAG.getCALLSEQ_START(Chain,
1526 DAG.getIntPtrConstant(NumBytes, dl, true), dl);
1529 DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy(DAG.getDataLayout()));
1531 RegsToPassVector RegsToPass;
1532 SmallVector<SDValue, 8> MemOpChains;
1534 // Walk the register/memloc assignments, inserting copies/loads. In the case
1535 // of tail call optimization, arguments are handled later.
1536 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1538 ++i, ++realArgIdx) {
1539 CCValAssign &VA = ArgLocs[i];
1540 SDValue Arg = OutVals[realArgIdx];
1541 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1542 bool isByVal = Flags.isByVal();
1544 // Promote the value if needed.
1545 switch (VA.getLocInfo()) {
1546 default: llvm_unreachable("Unknown loc info!");
1547 case CCValAssign::Full: break;
1548 case CCValAssign::SExt:
1549 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1551 case CCValAssign::ZExt:
1552 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1554 case CCValAssign::AExt:
1555 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1557 case CCValAssign::BCvt:
1558 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1562 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1563 if (VA.needsCustom()) {
1564 if (VA.getLocVT() == MVT::v2f64) {
1565 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1566 DAG.getConstant(0, dl, MVT::i32));
1567 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1568 DAG.getConstant(1, dl, MVT::i32));
1570 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1571 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1573 VA = ArgLocs[++i]; // skip ahead to next loc
1574 if (VA.isRegLoc()) {
1575 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1576 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1578 assert(VA.isMemLoc());
1580 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1581 dl, DAG, VA, Flags));
1584 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1585 StackPtr, MemOpChains, Flags);
1587 } else if (VA.isRegLoc()) {
1588 if (realArgIdx == 0 && Flags.isReturned() && Outs[0].VT == MVT::i32) {
1589 assert(VA.getLocVT() == MVT::i32 &&
1590 "unexpected calling convention register assignment");
1591 assert(!Ins.empty() && Ins[0].VT == MVT::i32 &&
1592 "unexpected use of 'returned'");
1593 isThisReturn = true;
1595 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1596 } else if (isByVal) {
1597 assert(VA.isMemLoc());
1598 unsigned offset = 0;
1600 // True if this byval aggregate will be split between registers
1602 unsigned ByValArgsCount = CCInfo.getInRegsParamsCount();
1603 unsigned CurByValIdx = CCInfo.getInRegsParamsProcessed();
1605 if (CurByValIdx < ByValArgsCount) {
1607 unsigned RegBegin, RegEnd;
1608 CCInfo.getInRegsParamInfo(CurByValIdx, RegBegin, RegEnd);
1611 DAG.getTargetLoweringInfo().getPointerTy(DAG.getDataLayout());
1613 for (i = 0, j = RegBegin; j < RegEnd; i++, j++) {
1614 SDValue Const = DAG.getConstant(4*i, dl, MVT::i32);
1615 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
1616 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
1617 MachinePointerInfo(),
1618 false, false, false,
1619 DAG.InferPtrAlignment(AddArg));
1620 MemOpChains.push_back(Load.getValue(1));
1621 RegsToPass.push_back(std::make_pair(j, Load));
1624 // If parameter size outsides register area, "offset" value
1625 // helps us to calculate stack slot for remained part properly.
1626 offset = RegEnd - RegBegin;
1628 CCInfo.nextInRegsParam();
1631 if (Flags.getByValSize() > 4*offset) {
1632 auto PtrVT = getPointerTy(DAG.getDataLayout());
1633 unsigned LocMemOffset = VA.getLocMemOffset();
1634 SDValue StkPtrOff = DAG.getIntPtrConstant(LocMemOffset, dl);
1635 SDValue Dst = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, StkPtrOff);
1636 SDValue SrcOffset = DAG.getIntPtrConstant(4*offset, dl);
1637 SDValue Src = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, SrcOffset);
1638 SDValue SizeNode = DAG.getConstant(Flags.getByValSize() - 4*offset, dl,
1640 SDValue AlignNode = DAG.getConstant(Flags.getByValAlign(), dl,
1643 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
1644 SDValue Ops[] = { Chain, Dst, Src, SizeNode, AlignNode};
1645 MemOpChains.push_back(DAG.getNode(ARMISD::COPY_STRUCT_BYVAL, dl, VTs,
1648 } else if (!isSibCall) {
1649 assert(VA.isMemLoc());
1651 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1652 dl, DAG, VA, Flags));
1656 if (!MemOpChains.empty())
1657 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains);
1659 // Build a sequence of copy-to-reg nodes chained together with token chain
1660 // and flag operands which copy the outgoing args into the appropriate regs.
1662 // Tail call byval lowering might overwrite argument registers so in case of
1663 // tail call optimization the copies to registers are lowered later.
1665 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1666 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1667 RegsToPass[i].second, InFlag);
1668 InFlag = Chain.getValue(1);
1671 // For tail calls lower the arguments to the 'real' stack slot.
1673 // Force all the incoming stack arguments to be loaded from the stack
1674 // before any new outgoing arguments are stored to the stack, because the
1675 // outgoing stack slots may alias the incoming argument stack slots, and
1676 // the alias isn't otherwise explicit. This is slightly more conservative
1677 // than necessary, because it means that each store effectively depends
1678 // on every argument instead of just those arguments it would clobber.
1680 // Do not flag preceding copytoreg stuff together with the following stuff.
1682 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1683 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1684 RegsToPass[i].second, InFlag);
1685 InFlag = Chain.getValue(1);
1690 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1691 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1692 // node so that legalize doesn't hack it.
1693 bool isDirect = false;
1694 bool isARMFunc = false;
1695 bool isLocalARMFunc = false;
1696 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1697 auto PtrVt = getPointerTy(DAG.getDataLayout());
1699 if (Subtarget->genLongCalls()) {
1700 assert((Subtarget->isTargetWindows() ||
1701 getTargetMachine().getRelocationModel() == Reloc::Static) &&
1702 "long-calls with non-static relocation model!");
1703 // Handle a global address or an external symbol. If it's not one of
1704 // those, the target's already in a register, so we don't need to do
1706 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1707 const GlobalValue *GV = G->getGlobal();
1708 // Create a constant pool entry for the callee address
1709 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1710 ARMConstantPoolValue *CPV =
1711 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 0);
1713 // Get the address of the callee into a register
1714 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVt, 4);
1715 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1716 Callee = DAG.getLoad(PtrVt, dl, DAG.getEntryNode(), CPAddr,
1717 MachinePointerInfo::getConstantPool(), false, false,
1719 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1720 const char *Sym = S->getSymbol();
1722 // Create a constant pool entry for the callee address
1723 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1724 ARMConstantPoolValue *CPV =
1725 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1726 ARMPCLabelIndex, 0);
1727 // Get the address of the callee into a register
1728 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVt, 4);
1729 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1730 Callee = DAG.getLoad(PtrVt, dl, DAG.getEntryNode(), CPAddr,
1731 MachinePointerInfo::getConstantPool(), false, false,
1734 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1735 const GlobalValue *GV = G->getGlobal();
1737 bool isDef = GV->isStrongDefinitionForLinker();
1738 bool isStub = (!isDef && Subtarget->isTargetMachO()) &&
1739 getTargetMachine().getRelocationModel() != Reloc::Static;
1740 isARMFunc = !Subtarget->isThumb() || (isStub && !Subtarget->isMClass());
1741 // ARM call to a local ARM function is predicable.
1742 isLocalARMFunc = !Subtarget->isThumb() && (isDef || !ARMInterworking);
1743 // tBX takes a register source operand.
1744 if (isStub && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1745 assert(Subtarget->isTargetMachO() && "WrapperPIC use on non-MachO?");
1746 Callee = DAG.getNode(
1747 ARMISD::WrapperPIC, dl, PtrVt,
1748 DAG.getTargetGlobalAddress(GV, dl, PtrVt, 0, ARMII::MO_NONLAZY));
1749 Callee = DAG.getLoad(PtrVt, dl, DAG.getEntryNode(), Callee,
1750 MachinePointerInfo::getGOT(), false, false, true, 0);
1751 } else if (Subtarget->isTargetCOFF()) {
1752 assert(Subtarget->isTargetWindows() &&
1753 "Windows is the only supported COFF target");
1754 unsigned TargetFlags = GV->hasDLLImportStorageClass()
1755 ? ARMII::MO_DLLIMPORT
1756 : ARMII::MO_NO_FLAG;
1758 DAG.getTargetGlobalAddress(GV, dl, PtrVt, /*Offset=*/0, TargetFlags);
1759 if (GV->hasDLLImportStorageClass())
1761 DAG.getLoad(PtrVt, dl, DAG.getEntryNode(),
1762 DAG.getNode(ARMISD::Wrapper, dl, PtrVt, Callee),
1763 MachinePointerInfo::getGOT(), false, false, false, 0);
1765 // On ELF targets for PIC code, direct calls should go through the PLT
1766 unsigned OpFlags = 0;
1767 if (Subtarget->isTargetELF() &&
1768 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1769 OpFlags = ARMII::MO_PLT;
1770 Callee = DAG.getTargetGlobalAddress(GV, dl, PtrVt, 0, OpFlags);
1772 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1774 bool isStub = Subtarget->isTargetMachO() &&
1775 getTargetMachine().getRelocationModel() != Reloc::Static;
1776 isARMFunc = !Subtarget->isThumb() || (isStub && !Subtarget->isMClass());
1777 // tBX takes a register source operand.
1778 const char *Sym = S->getSymbol();
1779 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1780 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1781 ARMConstantPoolValue *CPV =
1782 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1783 ARMPCLabelIndex, 4);
1784 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVt, 4);
1785 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1786 Callee = DAG.getLoad(PtrVt, dl, DAG.getEntryNode(), CPAddr,
1787 MachinePointerInfo::getConstantPool(), false, false,
1789 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, dl, MVT::i32);
1790 Callee = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVt, Callee, PICLabel);
1792 unsigned OpFlags = 0;
1793 // On ELF targets for PIC code, direct calls should go through the PLT
1794 if (Subtarget->isTargetELF() &&
1795 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1796 OpFlags = ARMII::MO_PLT;
1797 Callee = DAG.getTargetExternalSymbol(Sym, PtrVt, OpFlags);
1801 // FIXME: handle tail calls differently.
1803 bool HasMinSizeAttr = MF.getFunction()->hasFnAttribute(Attribute::MinSize);
1804 if (Subtarget->isThumb()) {
1805 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1806 CallOpc = ARMISD::CALL_NOLINK;
1808 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1810 if (!isDirect && !Subtarget->hasV5TOps())
1811 CallOpc = ARMISD::CALL_NOLINK;
1812 else if (doesNotRet && isDirect && Subtarget->hasRAS() &&
1813 // Emit regular call when code size is the priority
1815 // "mov lr, pc; b _foo" to avoid confusing the RSP
1816 CallOpc = ARMISD::CALL_NOLINK;
1818 CallOpc = isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL;
1821 std::vector<SDValue> Ops;
1822 Ops.push_back(Chain);
1823 Ops.push_back(Callee);
1825 // Add argument registers to the end of the list so that they are known live
1827 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1828 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1829 RegsToPass[i].second.getValueType()));
1831 // Add a register mask operand representing the call-preserved registers.
1833 const uint32_t *Mask;
1834 const ARMBaseRegisterInfo *ARI = Subtarget->getRegisterInfo();
1836 // For 'this' returns, use the R0-preserving mask if applicable
1837 Mask = ARI->getThisReturnPreservedMask(MF, CallConv);
1839 // Set isThisReturn to false if the calling convention is not one that
1840 // allows 'returned' to be modeled in this way, so LowerCallResult does
1841 // not try to pass 'this' straight through
1842 isThisReturn = false;
1843 Mask = ARI->getCallPreservedMask(MF, CallConv);
1846 Mask = ARI->getCallPreservedMask(MF, CallConv);
1848 assert(Mask && "Missing call preserved mask for calling convention");
1849 Ops.push_back(DAG.getRegisterMask(Mask));
1852 if (InFlag.getNode())
1853 Ops.push_back(InFlag);
1855 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
1857 MF.getFrameInfo()->setHasTailCall();
1858 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, Ops);
1861 // Returns a chain and a flag for retval copy to use.
1862 Chain = DAG.getNode(CallOpc, dl, NodeTys, Ops);
1863 InFlag = Chain.getValue(1);
1865 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, dl, true),
1866 DAG.getIntPtrConstant(0, dl, true), InFlag, dl);
1868 InFlag = Chain.getValue(1);
1870 // Handle result values, copying them out of physregs into vregs that we
1872 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins, dl, DAG,
1873 InVals, isThisReturn,
1874 isThisReturn ? OutVals[0] : SDValue());
1877 /// HandleByVal - Every parameter *after* a byval parameter is passed
1878 /// on the stack. Remember the next parameter register to allocate,
1879 /// and then confiscate the rest of the parameter registers to insure
1881 void ARMTargetLowering::HandleByVal(CCState *State, unsigned &Size,
1882 unsigned Align) const {
1883 assert((State->getCallOrPrologue() == Prologue ||
1884 State->getCallOrPrologue() == Call) &&
1885 "unhandled ParmContext");
1887 // Byval (as with any stack) slots are always at least 4 byte aligned.
1888 Align = std::max(Align, 4U);
1890 unsigned Reg = State->AllocateReg(GPRArgRegs);
1894 unsigned AlignInRegs = Align / 4;
1895 unsigned Waste = (ARM::R4 - Reg) % AlignInRegs;
1896 for (unsigned i = 0; i < Waste; ++i)
1897 Reg = State->AllocateReg(GPRArgRegs);
1902 unsigned Excess = 4 * (ARM::R4 - Reg);
1904 // Special case when NSAA != SP and parameter size greater than size of
1905 // all remained GPR regs. In that case we can't split parameter, we must
1906 // send it to stack. We also must set NCRN to R4, so waste all
1907 // remained registers.
1908 const unsigned NSAAOffset = State->getNextStackOffset();
1909 if (NSAAOffset != 0 && Size > Excess) {
1910 while (State->AllocateReg(GPRArgRegs))
1915 // First register for byval parameter is the first register that wasn't
1916 // allocated before this method call, so it would be "reg".
1917 // If parameter is small enough to be saved in range [reg, r4), then
1918 // the end (first after last) register would be reg + param-size-in-regs,
1919 // else parameter would be splitted between registers and stack,
1920 // end register would be r4 in this case.
1921 unsigned ByValRegBegin = Reg;
1922 unsigned ByValRegEnd = std::min<unsigned>(Reg + Size / 4, ARM::R4);
1923 State->addInRegsParamInfo(ByValRegBegin, ByValRegEnd);
1924 // Note, first register is allocated in the beginning of function already,
1925 // allocate remained amount of registers we need.
1926 for (unsigned i = Reg + 1; i != ByValRegEnd; ++i)
1927 State->AllocateReg(GPRArgRegs);
1928 // A byval parameter that is split between registers and memory needs its
1929 // size truncated here.
1930 // In the case where the entire structure fits in registers, we set the
1931 // size in memory to zero.
1932 Size = std::max<int>(Size - Excess, 0);
1935 /// MatchingStackOffset - Return true if the given stack call argument is
1936 /// already available in the same position (relatively) of the caller's
1937 /// incoming argument stack.
1939 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1940 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1941 const TargetInstrInfo *TII) {
1942 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1944 if (Arg.getOpcode() == ISD::CopyFromReg) {
1945 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1946 if (!TargetRegisterInfo::isVirtualRegister(VR))
1948 MachineInstr *Def = MRI->getVRegDef(VR);
1951 if (!Flags.isByVal()) {
1952 if (!TII->isLoadFromStackSlot(Def, FI))
1957 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
1958 if (Flags.isByVal())
1959 // ByVal argument is passed in as a pointer but it's now being
1960 // dereferenced. e.g.
1961 // define @foo(%struct.X* %A) {
1962 // tail call @bar(%struct.X* byval %A)
1965 SDValue Ptr = Ld->getBasePtr();
1966 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
1969 FI = FINode->getIndex();
1973 assert(FI != INT_MAX);
1974 if (!MFI->isFixedObjectIndex(FI))
1976 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
1979 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
1980 /// for tail call optimization. Targets which want to do tail call
1981 /// optimization should implement this function.
1983 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1984 CallingConv::ID CalleeCC,
1986 bool isCalleeStructRet,
1987 bool isCallerStructRet,
1988 const SmallVectorImpl<ISD::OutputArg> &Outs,
1989 const SmallVectorImpl<SDValue> &OutVals,
1990 const SmallVectorImpl<ISD::InputArg> &Ins,
1991 SelectionDAG& DAG) const {
1992 const Function *CallerF = DAG.getMachineFunction().getFunction();
1993 CallingConv::ID CallerCC = CallerF->getCallingConv();
1994 bool CCMatch = CallerCC == CalleeCC;
1996 // Look for obvious safe cases to perform tail call optimization that do not
1997 // require ABI changes. This is what gcc calls sibcall.
1999 // Do not sibcall optimize vararg calls unless the call site is not passing
2001 if (isVarArg && !Outs.empty())
2004 // Exception-handling functions need a special set of instructions to indicate
2005 // a return to the hardware. Tail-calling another function would probably
2007 if (CallerF->hasFnAttribute("interrupt"))
2010 // Also avoid sibcall optimization if either caller or callee uses struct
2011 // return semantics.
2012 if (isCalleeStructRet || isCallerStructRet)
2015 // FIXME: Completely disable sibcall for Thumb1 since ThumbRegisterInfo::
2016 // emitEpilogue is not ready for them. Thumb tail calls also use t2B, as
2017 // the Thumb1 16-bit unconditional branch doesn't have sufficient relocation
2018 // support in the assembler and linker to be used. This would need to be
2019 // fixed to fully support tail calls in Thumb1.
2021 // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take
2022 // LR. This means if we need to reload LR, it takes an extra instructions,
2023 // which outweighs the value of the tail call; but here we don't know yet
2024 // whether LR is going to be used. Probably the right approach is to
2025 // generate the tail call here and turn it back into CALL/RET in
2026 // emitEpilogue if LR is used.
2028 // Thumb1 PIC calls to external symbols use BX, so they can be tail calls,
2029 // but we need to make sure there are enough registers; the only valid
2030 // registers are the 4 used for parameters. We don't currently do this
2032 if (Subtarget->isThumb1Only())
2035 // Externally-defined functions with weak linkage should not be
2036 // tail-called on ARM when the OS does not support dynamic
2037 // pre-emption of symbols, as the AAELF spec requires normal calls
2038 // to undefined weak functions to be replaced with a NOP or jump to the
2039 // next instruction. The behaviour of branch instructions in this
2040 // situation (as used for tail calls) is implementation-defined, so we
2041 // cannot rely on the linker replacing the tail call with a return.
2042 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2043 const GlobalValue *GV = G->getGlobal();
2044 const Triple &TT = getTargetMachine().getTargetTriple();
2045 if (GV->hasExternalWeakLinkage() &&
2046 (!TT.isOSWindows() || TT.isOSBinFormatELF() || TT.isOSBinFormatMachO()))
2050 // If the calling conventions do not match, then we'd better make sure the
2051 // results are returned in the same way as what the caller expects.
2053 SmallVector<CCValAssign, 16> RVLocs1;
2054 ARMCCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(), RVLocs1,
2055 *DAG.getContext(), Call);
2056 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
2058 SmallVector<CCValAssign, 16> RVLocs2;
2059 ARMCCState CCInfo2(CallerCC, false, DAG.getMachineFunction(), RVLocs2,
2060 *DAG.getContext(), Call);
2061 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
2063 if (RVLocs1.size() != RVLocs2.size())
2065 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
2066 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
2068 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
2070 if (RVLocs1[i].isRegLoc()) {
2071 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
2074 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
2080 // If Caller's vararg or byval argument has been split between registers and
2081 // stack, do not perform tail call, since part of the argument is in caller's
2083 const ARMFunctionInfo *AFI_Caller = DAG.getMachineFunction().
2084 getInfo<ARMFunctionInfo>();
2085 if (AFI_Caller->getArgRegsSaveSize())
2088 // If the callee takes no arguments then go on to check the results of the
2090 if (!Outs.empty()) {
2091 // Check if stack adjustment is needed. For now, do not do this if any
2092 // argument is passed on the stack.
2093 SmallVector<CCValAssign, 16> ArgLocs;
2094 ARMCCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
2095 *DAG.getContext(), Call);
2096 CCInfo.AnalyzeCallOperands(Outs,
2097 CCAssignFnForNode(CalleeCC, false, isVarArg));
2098 if (CCInfo.getNextStackOffset()) {
2099 MachineFunction &MF = DAG.getMachineFunction();
2101 // Check if the arguments are already laid out in the right way as
2102 // the caller's fixed stack objects.
2103 MachineFrameInfo *MFI = MF.getFrameInfo();
2104 const MachineRegisterInfo *MRI = &MF.getRegInfo();
2105 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
2106 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
2108 ++i, ++realArgIdx) {
2109 CCValAssign &VA = ArgLocs[i];
2110 EVT RegVT = VA.getLocVT();
2111 SDValue Arg = OutVals[realArgIdx];
2112 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
2113 if (VA.getLocInfo() == CCValAssign::Indirect)
2115 if (VA.needsCustom()) {
2116 // f64 and vector types are split into multiple registers or
2117 // register/stack-slot combinations. The types will not match
2118 // the registers; give up on memory f64 refs until we figure
2119 // out what to do about this.
2122 if (!ArgLocs[++i].isRegLoc())
2124 if (RegVT == MVT::v2f64) {
2125 if (!ArgLocs[++i].isRegLoc())
2127 if (!ArgLocs[++i].isRegLoc())
2130 } else if (!VA.isRegLoc()) {
2131 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
2143 ARMTargetLowering::CanLowerReturn(CallingConv::ID CallConv,
2144 MachineFunction &MF, bool isVarArg,
2145 const SmallVectorImpl<ISD::OutputArg> &Outs,
2146 LLVMContext &Context) const {
2147 SmallVector<CCValAssign, 16> RVLocs;
2148 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, Context);
2149 return CCInfo.CheckReturn(Outs, CCAssignFnForNode(CallConv, /*Return=*/true,
2153 static SDValue LowerInterruptReturn(SmallVectorImpl<SDValue> &RetOps,
2154 SDLoc DL, SelectionDAG &DAG) {
2155 const MachineFunction &MF = DAG.getMachineFunction();
2156 const Function *F = MF.getFunction();
2158 StringRef IntKind = F->getFnAttribute("interrupt").getValueAsString();
2160 // See ARM ARM v7 B1.8.3. On exception entry LR is set to a possibly offset
2161 // version of the "preferred return address". These offsets affect the return
2162 // instruction if this is a return from PL1 without hypervisor extensions.
2163 // IRQ/FIQ: +4 "subs pc, lr, #4"
2164 // SWI: 0 "subs pc, lr, #0"
2165 // ABORT: +4 "subs pc, lr, #4"
2166 // UNDEF: +4/+2 "subs pc, lr, #0"
2167 // UNDEF varies depending on where the exception came from ARM or Thumb
2168 // mode. Alongside GCC, we throw our hands up in disgust and pretend it's 0.
2171 if (IntKind == "" || IntKind == "IRQ" || IntKind == "FIQ" ||
2174 else if (IntKind == "SWI" || IntKind == "UNDEF")
2177 report_fatal_error("Unsupported interrupt attribute. If present, value "
2178 "must be one of: IRQ, FIQ, SWI, ABORT or UNDEF");
2180 RetOps.insert(RetOps.begin() + 1,
2181 DAG.getConstant(LROffset, DL, MVT::i32, false));
2183 return DAG.getNode(ARMISD::INTRET_FLAG, DL, MVT::Other, RetOps);
2187 ARMTargetLowering::LowerReturn(SDValue Chain,
2188 CallingConv::ID CallConv, bool isVarArg,
2189 const SmallVectorImpl<ISD::OutputArg> &Outs,
2190 const SmallVectorImpl<SDValue> &OutVals,
2191 SDLoc dl, SelectionDAG &DAG) const {
2193 // CCValAssign - represent the assignment of the return value to a location.
2194 SmallVector<CCValAssign, 16> RVLocs;
2196 // CCState - Info about the registers and stack slots.
2197 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
2198 *DAG.getContext(), Call);
2200 // Analyze outgoing return values.
2201 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
2205 SmallVector<SDValue, 4> RetOps;
2206 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
2207 bool isLittleEndian = Subtarget->isLittle();
2209 MachineFunction &MF = DAG.getMachineFunction();
2210 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2211 AFI->setReturnRegsCount(RVLocs.size());
2213 // Copy the result values into the output registers.
2214 for (unsigned i = 0, realRVLocIdx = 0;
2216 ++i, ++realRVLocIdx) {
2217 CCValAssign &VA = RVLocs[i];
2218 assert(VA.isRegLoc() && "Can only return in registers!");
2220 SDValue Arg = OutVals[realRVLocIdx];
2222 switch (VA.getLocInfo()) {
2223 default: llvm_unreachable("Unknown loc info!");
2224 case CCValAssign::Full: break;
2225 case CCValAssign::BCvt:
2226 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
2230 if (VA.needsCustom()) {
2231 if (VA.getLocVT() == MVT::v2f64) {
2232 // Extract the first half and return it in two registers.
2233 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
2234 DAG.getConstant(0, dl, MVT::i32));
2235 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
2236 DAG.getVTList(MVT::i32, MVT::i32), Half);
2238 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2239 HalfGPRs.getValue(isLittleEndian ? 0 : 1),
2241 Flag = Chain.getValue(1);
2242 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2243 VA = RVLocs[++i]; // skip ahead to next loc
2244 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2245 HalfGPRs.getValue(isLittleEndian ? 1 : 0),
2247 Flag = Chain.getValue(1);
2248 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2249 VA = RVLocs[++i]; // skip ahead to next loc
2251 // Extract the 2nd half and fall through to handle it as an f64 value.
2252 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
2253 DAG.getConstant(1, dl, MVT::i32));
2255 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
2257 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
2258 DAG.getVTList(MVT::i32, MVT::i32), Arg);
2259 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2260 fmrrd.getValue(isLittleEndian ? 0 : 1),
2262 Flag = Chain.getValue(1);
2263 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2264 VA = RVLocs[++i]; // skip ahead to next loc
2265 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2266 fmrrd.getValue(isLittleEndian ? 1 : 0),
2269 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
2271 // Guarantee that all emitted copies are
2272 // stuck together, avoiding something bad.
2273 Flag = Chain.getValue(1);
2274 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2277 // Update chain and glue.
2280 RetOps.push_back(Flag);
2282 // CPUs which aren't M-class use a special sequence to return from
2283 // exceptions (roughly, any instruction setting pc and cpsr simultaneously,
2284 // though we use "subs pc, lr, #N").
2286 // M-class CPUs actually use a normal return sequence with a special
2287 // (hardware-provided) value in LR, so the normal code path works.
2288 if (DAG.getMachineFunction().getFunction()->hasFnAttribute("interrupt") &&
2289 !Subtarget->isMClass()) {
2290 if (Subtarget->isThumb1Only())
2291 report_fatal_error("interrupt attribute is not supported in Thumb1");
2292 return LowerInterruptReturn(RetOps, dl, DAG);
2295 return DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, RetOps);
2298 bool ARMTargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
2299 if (N->getNumValues() != 1)
2301 if (!N->hasNUsesOfValue(1, 0))
2304 SDValue TCChain = Chain;
2305 SDNode *Copy = *N->use_begin();
2306 if (Copy->getOpcode() == ISD::CopyToReg) {
2307 // If the copy has a glue operand, we conservatively assume it isn't safe to
2308 // perform a tail call.
2309 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
2311 TCChain = Copy->getOperand(0);
2312 } else if (Copy->getOpcode() == ARMISD::VMOVRRD) {
2313 SDNode *VMov = Copy;
2314 // f64 returned in a pair of GPRs.
2315 SmallPtrSet<SDNode*, 2> Copies;
2316 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
2318 if (UI->getOpcode() != ISD::CopyToReg)
2322 if (Copies.size() > 2)
2325 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
2327 SDValue UseChain = UI->getOperand(0);
2328 if (Copies.count(UseChain.getNode()))
2332 // We are at the top of this chain.
2333 // If the copy has a glue operand, we conservatively assume it
2334 // isn't safe to perform a tail call.
2335 if (UI->getOperand(UI->getNumOperands()-1).getValueType() == MVT::Glue)
2341 } else if (Copy->getOpcode() == ISD::BITCAST) {
2342 // f32 returned in a single GPR.
2343 if (!Copy->hasOneUse())
2345 Copy = *Copy->use_begin();
2346 if (Copy->getOpcode() != ISD::CopyToReg || !Copy->hasNUsesOfValue(1, 0))
2348 // If the copy has a glue operand, we conservatively assume it isn't safe to
2349 // perform a tail call.
2350 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
2352 TCChain = Copy->getOperand(0);
2357 bool HasRet = false;
2358 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
2360 if (UI->getOpcode() != ARMISD::RET_FLAG &&
2361 UI->getOpcode() != ARMISD::INTRET_FLAG)
2373 bool ARMTargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
2374 if (!Subtarget->supportsTailCall())
2378 CI->getParent()->getParent()->getFnAttribute("disable-tail-calls");
2379 if (!CI->isTailCall() || Attr.getValueAsString() == "true")
2382 return !Subtarget->isThumb1Only();
2385 // Trying to write a 64 bit value so need to split into two 32 bit values first,
2386 // and pass the lower and high parts through.
2387 static SDValue LowerWRITE_REGISTER(SDValue Op, SelectionDAG &DAG) {
2389 SDValue WriteValue = Op->getOperand(2);
2391 // This function is only supposed to be called for i64 type argument.
2392 assert(WriteValue.getValueType() == MVT::i64
2393 && "LowerWRITE_REGISTER called for non-i64 type argument.");
2395 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, WriteValue,
2396 DAG.getConstant(0, DL, MVT::i32));
2397 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, WriteValue,
2398 DAG.getConstant(1, DL, MVT::i32));
2399 SDValue Ops[] = { Op->getOperand(0), Op->getOperand(1), Lo, Hi };
2400 return DAG.getNode(ISD::WRITE_REGISTER, DL, MVT::Other, Ops);
2403 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
2404 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
2405 // one of the above mentioned nodes. It has to be wrapped because otherwise
2406 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
2407 // be used to form addressing mode. These wrapped nodes will be selected
2409 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
2410 EVT PtrVT = Op.getValueType();
2411 // FIXME there is no actual debug info here
2413 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
2415 if (CP->isMachineConstantPoolEntry())
2416 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
2417 CP->getAlignment());
2419 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
2420 CP->getAlignment());
2421 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
2424 unsigned ARMTargetLowering::getJumpTableEncoding() const {
2425 return MachineJumpTableInfo::EK_Inline;
2428 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
2429 SelectionDAG &DAG) const {
2430 MachineFunction &MF = DAG.getMachineFunction();
2431 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2432 unsigned ARMPCLabelIndex = 0;
2434 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2435 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
2436 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2438 if (RelocM == Reloc::Static) {
2439 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
2441 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2442 ARMPCLabelIndex = AFI->createPICLabelUId();
2443 ARMConstantPoolValue *CPV =
2444 ARMConstantPoolConstant::Create(BA, ARMPCLabelIndex,
2445 ARMCP::CPBlockAddress, PCAdj);
2446 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2448 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
2449 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
2450 MachinePointerInfo::getConstantPool(),
2451 false, false, false, 0);
2452 if (RelocM == Reloc::Static)
2454 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, DL, MVT::i32);
2455 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
2458 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
2460 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
2461 SelectionDAG &DAG) const {
2463 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2464 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2465 MachineFunction &MF = DAG.getMachineFunction();
2466 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2467 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2468 ARMConstantPoolValue *CPV =
2469 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2470 ARMCP::CPValue, PCAdj, ARMCP::TLSGD, true);
2471 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2472 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
2473 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
2474 MachinePointerInfo::getConstantPool(),
2475 false, false, false, 0);
2476 SDValue Chain = Argument.getValue(1);
2478 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, dl, MVT::i32);
2479 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
2481 // call __tls_get_addr.
2484 Entry.Node = Argument;
2485 Entry.Ty = (Type *) Type::getInt32Ty(*DAG.getContext());
2486 Args.push_back(Entry);
2488 // FIXME: is there useful debug info available here?
2489 TargetLowering::CallLoweringInfo CLI(DAG);
2490 CLI.setDebugLoc(dl).setChain(Chain)
2491 .setCallee(CallingConv::C, Type::getInt32Ty(*DAG.getContext()),
2492 DAG.getExternalSymbol("__tls_get_addr", PtrVT), std::move(Args),
2495 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
2496 return CallResult.first;
2499 // Lower ISD::GlobalTLSAddress using the "initial exec" or
2500 // "local exec" model.
2502 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
2504 TLSModel::Model model) const {
2505 const GlobalValue *GV = GA->getGlobal();
2508 SDValue Chain = DAG.getEntryNode();
2509 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2510 // Get the Thread Pointer
2511 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2513 if (model == TLSModel::InitialExec) {
2514 MachineFunction &MF = DAG.getMachineFunction();
2515 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2516 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2517 // Initial exec model.
2518 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2519 ARMConstantPoolValue *CPV =
2520 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2521 ARMCP::CPValue, PCAdj, ARMCP::GOTTPOFF,
2523 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2524 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2525 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2526 MachinePointerInfo::getConstantPool(),
2527 false, false, false, 0);
2528 Chain = Offset.getValue(1);
2530 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, dl, MVT::i32);
2531 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
2533 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2534 MachinePointerInfo::getConstantPool(),
2535 false, false, false, 0);
2538 assert(model == TLSModel::LocalExec);
2539 ARMConstantPoolValue *CPV =
2540 ARMConstantPoolConstant::Create(GV, ARMCP::TPOFF);
2541 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2542 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2543 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2544 MachinePointerInfo::getConstantPool(),
2545 false, false, false, 0);
2548 // The address of the thread local variable is the add of the thread
2549 // pointer with the offset of the variable.
2550 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
2554 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
2555 // TODO: implement the "local dynamic" model
2556 assert(Subtarget->isTargetELF() &&
2557 "TLS not implemented for non-ELF targets");
2558 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
2560 TLSModel::Model model = getTargetMachine().getTLSModel(GA->getGlobal());
2563 case TLSModel::GeneralDynamic:
2564 case TLSModel::LocalDynamic:
2565 return LowerToTLSGeneralDynamicModel(GA, DAG);
2566 case TLSModel::InitialExec:
2567 case TLSModel::LocalExec:
2568 return LowerToTLSExecModels(GA, DAG, model);
2570 llvm_unreachable("bogus TLS model");
2573 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
2574 SelectionDAG &DAG) const {
2575 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2577 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2578 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2579 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
2580 ARMConstantPoolValue *CPV =
2581 ARMConstantPoolConstant::Create(GV,
2582 UseGOTOFF ? ARMCP::GOTOFF : ARMCP::GOT);
2583 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2584 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2585 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
2587 MachinePointerInfo::getConstantPool(),
2588 false, false, false, 0);
2589 SDValue Chain = Result.getValue(1);
2590 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
2591 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
2593 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
2594 MachinePointerInfo::getGOT(),
2595 false, false, false, 0);
2599 // If we have T2 ops, we can materialize the address directly via movt/movw
2600 // pair. This is always cheaper.
2601 if (Subtarget->useMovt(DAG.getMachineFunction())) {
2603 // FIXME: Once remat is capable of dealing with instructions with register
2604 // operands, expand this into two nodes.
2605 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2606 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2608 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2609 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2610 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2611 MachinePointerInfo::getConstantPool(),
2612 false, false, false, 0);
2616 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
2617 SelectionDAG &DAG) const {
2618 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2620 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2621 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2623 if (Subtarget->useMovt(DAG.getMachineFunction()))
2626 // FIXME: Once remat is capable of dealing with instructions with register
2627 // operands, expand this into multiple nodes
2629 RelocM == Reloc::PIC_ ? ARMISD::WrapperPIC : ARMISD::Wrapper;
2631 SDValue G = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, ARMII::MO_NONLAZY);
2632 SDValue Result = DAG.getNode(Wrapper, dl, PtrVT, G);
2634 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2635 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
2636 MachinePointerInfo::getGOT(), false, false, false, 0);
2640 SDValue ARMTargetLowering::LowerGlobalAddressWindows(SDValue Op,
2641 SelectionDAG &DAG) const {
2642 assert(Subtarget->isTargetWindows() && "non-Windows COFF is not supported");
2643 assert(Subtarget->useMovt(DAG.getMachineFunction()) &&
2644 "Windows on ARM expects to use movw/movt");
2646 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2647 const ARMII::TOF TargetFlags =
2648 (GV->hasDLLImportStorageClass() ? ARMII::MO_DLLIMPORT : ARMII::MO_NO_FLAG);
2649 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2655 // FIXME: Once remat is capable of dealing with instructions with register
2656 // operands, expand this into two nodes.
2657 Result = DAG.getNode(ARMISD::Wrapper, DL, PtrVT,
2658 DAG.getTargetGlobalAddress(GV, DL, PtrVT, /*Offset=*/0,
2660 if (GV->hasDLLImportStorageClass())
2661 Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), Result,
2662 MachinePointerInfo::getGOT(), false, false, false, 0);
2666 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
2667 SelectionDAG &DAG) const {
2668 assert(Subtarget->isTargetELF() &&
2669 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
2670 MachineFunction &MF = DAG.getMachineFunction();
2671 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2672 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2673 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2675 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2676 ARMConstantPoolValue *CPV =
2677 ARMConstantPoolSymbol::Create(*DAG.getContext(), "_GLOBAL_OFFSET_TABLE_",
2678 ARMPCLabelIndex, PCAdj);
2679 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2680 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2681 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2682 MachinePointerInfo::getConstantPool(),
2683 false, false, false, 0);
2684 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, dl, MVT::i32);
2685 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2689 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
2691 SDValue Val = DAG.getConstant(0, dl, MVT::i32);
2692 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl,
2693 DAG.getVTList(MVT::i32, MVT::Other), Op.getOperand(0),
2694 Op.getOperand(1), Val);
2698 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
2700 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
2701 Op.getOperand(1), DAG.getConstant(0, dl, MVT::i32));
2705 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
2706 const ARMSubtarget *Subtarget) const {
2707 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2710 default: return SDValue(); // Don't custom lower most intrinsics.
2711 case Intrinsic::arm_rbit: {
2712 assert(Op.getOperand(1).getValueType() == MVT::i32 &&
2713 "RBIT intrinsic must have i32 type!");
2714 return DAG.getNode(ARMISD::RBIT, dl, MVT::i32, Op.getOperand(1));
2716 case Intrinsic::arm_thread_pointer: {
2717 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2718 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2720 case Intrinsic::eh_sjlj_lsda: {
2721 MachineFunction &MF = DAG.getMachineFunction();
2722 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2723 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2724 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2725 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2727 unsigned PCAdj = (RelocM != Reloc::PIC_)
2728 ? 0 : (Subtarget->isThumb() ? 4 : 8);
2729 ARMConstantPoolValue *CPV =
2730 ARMConstantPoolConstant::Create(MF.getFunction(), ARMPCLabelIndex,
2731 ARMCP::CPLSDA, PCAdj);
2732 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2733 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2735 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2736 MachinePointerInfo::getConstantPool(),
2737 false, false, false, 0);
2739 if (RelocM == Reloc::PIC_) {
2740 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, dl, MVT::i32);
2741 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2745 case Intrinsic::arm_neon_vmulls:
2746 case Intrinsic::arm_neon_vmullu: {
2747 unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vmulls)
2748 ? ARMISD::VMULLs : ARMISD::VMULLu;
2749 return DAG.getNode(NewOpc, SDLoc(Op), Op.getValueType(),
2750 Op.getOperand(1), Op.getOperand(2));
2755 static SDValue LowerATOMIC_FENCE(SDValue Op, SelectionDAG &DAG,
2756 const ARMSubtarget *Subtarget) {
2757 // FIXME: handle "fence singlethread" more efficiently.
2759 if (!Subtarget->hasDataBarrier()) {
2760 // Some ARMv6 cpus can support data barriers with an mcr instruction.
2761 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2763 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2764 "Unexpected ISD::ATOMIC_FENCE encountered. Should be libcall!");
2765 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2766 DAG.getConstant(0, dl, MVT::i32));
2769 ConstantSDNode *OrdN = cast<ConstantSDNode>(Op.getOperand(1));
2770 AtomicOrdering Ord = static_cast<AtomicOrdering>(OrdN->getZExtValue());
2771 ARM_MB::MemBOpt Domain = ARM_MB::ISH;
2772 if (Subtarget->isMClass()) {
2773 // Only a full system barrier exists in the M-class architectures.
2774 Domain = ARM_MB::SY;
2775 } else if (Subtarget->isSwift() && Ord == Release) {
2776 // Swift happens to implement ISHST barriers in a way that's compatible with
2777 // Release semantics but weaker than ISH so we'd be fools not to use
2778 // it. Beware: other processors probably don't!
2779 Domain = ARM_MB::ISHST;
2782 return DAG.getNode(ISD::INTRINSIC_VOID, dl, MVT::Other, Op.getOperand(0),
2783 DAG.getConstant(Intrinsic::arm_dmb, dl, MVT::i32),
2784 DAG.getConstant(Domain, dl, MVT::i32));
2787 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG,
2788 const ARMSubtarget *Subtarget) {
2789 // ARM pre v5TE and Thumb1 does not have preload instructions.
2790 if (!(Subtarget->isThumb2() ||
2791 (!Subtarget->isThumb1Only() && Subtarget->hasV5TEOps())))
2792 // Just preserve the chain.
2793 return Op.getOperand(0);
2796 unsigned isRead = ~cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() & 1;
2798 (!Subtarget->hasV7Ops() || !Subtarget->hasMPExtension()))
2799 // ARMv7 with MP extension has PLDW.
2800 return Op.getOperand(0);
2802 unsigned isData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
2803 if (Subtarget->isThumb()) {
2805 isRead = ~isRead & 1;
2806 isData = ~isData & 1;
2809 return DAG.getNode(ARMISD::PRELOAD, dl, MVT::Other, Op.getOperand(0),
2810 Op.getOperand(1), DAG.getConstant(isRead, dl, MVT::i32),
2811 DAG.getConstant(isData, dl, MVT::i32));
2814 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
2815 MachineFunction &MF = DAG.getMachineFunction();
2816 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
2818 // vastart just stores the address of the VarArgsFrameIndex slot into the
2819 // memory location argument.
2821 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy(DAG.getDataLayout());
2822 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2823 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
2824 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
2825 MachinePointerInfo(SV), false, false, 0);
2829 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
2830 SDValue &Root, SelectionDAG &DAG,
2832 MachineFunction &MF = DAG.getMachineFunction();
2833 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2835 const TargetRegisterClass *RC;
2836 if (AFI->isThumb1OnlyFunction())
2837 RC = &ARM::tGPRRegClass;
2839 RC = &ARM::GPRRegClass;
2841 // Transform the arguments stored in physical registers into virtual ones.
2842 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2843 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2846 if (NextVA.isMemLoc()) {
2847 MachineFrameInfo *MFI = MF.getFrameInfo();
2848 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true);
2850 // Create load node to retrieve arguments from the stack.
2851 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
2852 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
2853 MachinePointerInfo::getFixedStack(FI),
2854 false, false, false, 0);
2856 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
2857 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2859 if (!Subtarget->isLittle())
2860 std::swap (ArgValue, ArgValue2);
2861 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
2864 // The remaining GPRs hold either the beginning of variable-argument
2865 // data, or the beginning of an aggregate passed by value (usually
2866 // byval). Either way, we allocate stack slots adjacent to the data
2867 // provided by our caller, and store the unallocated registers there.
2868 // If this is a variadic function, the va_list pointer will begin with
2869 // these values; otherwise, this reassembles a (byval) structure that
2870 // was split between registers and memory.
2871 // Return: The frame index registers were stored into.
2873 ARMTargetLowering::StoreByValRegs(CCState &CCInfo, SelectionDAG &DAG,
2874 SDLoc dl, SDValue &Chain,
2875 const Value *OrigArg,
2876 unsigned InRegsParamRecordIdx,
2878 unsigned ArgSize) const {
2879 // Currently, two use-cases possible:
2880 // Case #1. Non-var-args function, and we meet first byval parameter.
2881 // Setup first unallocated register as first byval register;
2882 // eat all remained registers
2883 // (these two actions are performed by HandleByVal method).
2884 // Then, here, we initialize stack frame with
2885 // "store-reg" instructions.
2886 // Case #2. Var-args function, that doesn't contain byval parameters.
2887 // The same: eat all remained unallocated registers,
2888 // initialize stack frame.
2890 MachineFunction &MF = DAG.getMachineFunction();
2891 MachineFrameInfo *MFI = MF.getFrameInfo();
2892 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2893 unsigned RBegin, REnd;
2894 if (InRegsParamRecordIdx < CCInfo.getInRegsParamsCount()) {
2895 CCInfo.getInRegsParamInfo(InRegsParamRecordIdx, RBegin, REnd);
2897 unsigned RBeginIdx = CCInfo.getFirstUnallocated(GPRArgRegs);
2898 RBegin = RBeginIdx == 4 ? (unsigned)ARM::R4 : GPRArgRegs[RBeginIdx];
2903 ArgOffset = -4 * (ARM::R4 - RBegin);
2905 auto PtrVT = getPointerTy(DAG.getDataLayout());
2906 int FrameIndex = MFI->CreateFixedObject(ArgSize, ArgOffset, false);
2907 SDValue FIN = DAG.getFrameIndex(FrameIndex, PtrVT);
2909 SmallVector<SDValue, 4> MemOps;
2910 const TargetRegisterClass *RC =
2911 AFI->isThumb1OnlyFunction() ? &ARM::tGPRRegClass : &ARM::GPRRegClass;
2913 for (unsigned Reg = RBegin, i = 0; Reg < REnd; ++Reg, ++i) {
2914 unsigned VReg = MF.addLiveIn(Reg, RC);
2915 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2917 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2918 MachinePointerInfo(OrigArg, 4 * i), false, false, 0);
2919 MemOps.push_back(Store);
2920 FIN = DAG.getNode(ISD::ADD, dl, PtrVT, FIN, DAG.getConstant(4, dl, PtrVT));
2923 if (!MemOps.empty())
2924 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOps);
2928 // Setup stack frame, the va_list pointer will start from.
2930 ARMTargetLowering::VarArgStyleRegisters(CCState &CCInfo, SelectionDAG &DAG,
2931 SDLoc dl, SDValue &Chain,
2933 unsigned TotalArgRegsSaveSize,
2934 bool ForceMutable) const {
2935 MachineFunction &MF = DAG.getMachineFunction();
2936 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2938 // Try to store any remaining integer argument regs
2939 // to their spots on the stack so that they may be loaded by deferencing
2940 // the result of va_next.
2941 // If there is no regs to be stored, just point address after last
2942 // argument passed via stack.
2943 int FrameIndex = StoreByValRegs(CCInfo, DAG, dl, Chain, nullptr,
2944 CCInfo.getInRegsParamsCount(),
2945 CCInfo.getNextStackOffset(), 4);
2946 AFI->setVarArgsFrameIndex(FrameIndex);
2950 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
2951 CallingConv::ID CallConv, bool isVarArg,
2952 const SmallVectorImpl<ISD::InputArg>
2954 SDLoc dl, SelectionDAG &DAG,
2955 SmallVectorImpl<SDValue> &InVals)
2957 MachineFunction &MF = DAG.getMachineFunction();
2958 MachineFrameInfo *MFI = MF.getFrameInfo();
2960 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2962 // Assign locations to all of the incoming arguments.
2963 SmallVector<CCValAssign, 16> ArgLocs;
2964 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs,
2965 *DAG.getContext(), Prologue);
2966 CCInfo.AnalyzeFormalArguments(Ins,
2967 CCAssignFnForNode(CallConv, /* Return*/ false,
2970 SmallVector<SDValue, 16> ArgValues;
2972 Function::const_arg_iterator CurOrigArg = MF.getFunction()->arg_begin();
2973 unsigned CurArgIdx = 0;
2975 // Initially ArgRegsSaveSize is zero.
2976 // Then we increase this value each time we meet byval parameter.
2977 // We also increase this value in case of varargs function.
2978 AFI->setArgRegsSaveSize(0);
2980 // Calculate the amount of stack space that we need to allocate to store
2981 // byval and variadic arguments that are passed in registers.
2982 // We need to know this before we allocate the first byval or variadic
2983 // argument, as they will be allocated a stack slot below the CFA (Canonical
2984 // Frame Address, the stack pointer at entry to the function).
2985 unsigned ArgRegBegin = ARM::R4;
2986 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2987 if (CCInfo.getInRegsParamsProcessed() >= CCInfo.getInRegsParamsCount())
2990 CCValAssign &VA = ArgLocs[i];
2991 unsigned Index = VA.getValNo();
2992 ISD::ArgFlagsTy Flags = Ins[Index].Flags;
2993 if (!Flags.isByVal())
2996 assert(VA.isMemLoc() && "unexpected byval pointer in reg");
2997 unsigned RBegin, REnd;
2998 CCInfo.getInRegsParamInfo(CCInfo.getInRegsParamsProcessed(), RBegin, REnd);
2999 ArgRegBegin = std::min(ArgRegBegin, RBegin);
3001 CCInfo.nextInRegsParam();
3003 CCInfo.rewindByValRegsInfo();
3005 int lastInsIndex = -1;
3006 if (isVarArg && MFI->hasVAStart()) {
3007 unsigned RegIdx = CCInfo.getFirstUnallocated(GPRArgRegs);
3008 if (RegIdx != array_lengthof(GPRArgRegs))
3009 ArgRegBegin = std::min(ArgRegBegin, (unsigned)GPRArgRegs[RegIdx]);
3012 unsigned TotalArgRegsSaveSize = 4 * (ARM::R4 - ArgRegBegin);
3013 AFI->setArgRegsSaveSize(TotalArgRegsSaveSize);
3014 auto PtrVT = getPointerTy(DAG.getDataLayout());
3016 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3017 CCValAssign &VA = ArgLocs[i];
3018 if (Ins[VA.getValNo()].isOrigArg()) {
3019 std::advance(CurOrigArg,
3020 Ins[VA.getValNo()].getOrigArgIndex() - CurArgIdx);
3021 CurArgIdx = Ins[VA.getValNo()].getOrigArgIndex();
3023 // Arguments stored in registers.
3024 if (VA.isRegLoc()) {
3025 EVT RegVT = VA.getLocVT();
3027 if (VA.needsCustom()) {
3028 // f64 and vector types are split up into multiple registers or
3029 // combinations of registers and stack slots.
3030 if (VA.getLocVT() == MVT::v2f64) {
3031 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
3033 VA = ArgLocs[++i]; // skip ahead to next loc
3035 if (VA.isMemLoc()) {
3036 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true);
3037 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
3038 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
3039 MachinePointerInfo::getFixedStack(FI),
3040 false, false, false, 0);
3042 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
3045 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
3046 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
3047 ArgValue, ArgValue1,
3048 DAG.getIntPtrConstant(0, dl));
3049 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
3050 ArgValue, ArgValue2,
3051 DAG.getIntPtrConstant(1, dl));
3053 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
3056 const TargetRegisterClass *RC;
3058 if (RegVT == MVT::f32)
3059 RC = &ARM::SPRRegClass;
3060 else if (RegVT == MVT::f64)
3061 RC = &ARM::DPRRegClass;
3062 else if (RegVT == MVT::v2f64)
3063 RC = &ARM::QPRRegClass;
3064 else if (RegVT == MVT::i32)
3065 RC = AFI->isThumb1OnlyFunction() ? &ARM::tGPRRegClass
3066 : &ARM::GPRRegClass;
3068 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
3070 // Transform the arguments in physical registers into virtual ones.
3071 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
3072 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
3075 // If this is an 8 or 16-bit value, it is really passed promoted
3076 // to 32 bits. Insert an assert[sz]ext to capture this, then
3077 // truncate to the right size.
3078 switch (VA.getLocInfo()) {
3079 default: llvm_unreachable("Unknown loc info!");
3080 case CCValAssign::Full: break;
3081 case CCValAssign::BCvt:
3082 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
3084 case CCValAssign::SExt:
3085 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
3086 DAG.getValueType(VA.getValVT()));
3087 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
3089 case CCValAssign::ZExt:
3090 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
3091 DAG.getValueType(VA.getValVT()));
3092 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
3096 InVals.push_back(ArgValue);
3098 } else { // VA.isRegLoc()
3101 assert(VA.isMemLoc());
3102 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
3104 int index = VA.getValNo();
3106 // Some Ins[] entries become multiple ArgLoc[] entries.
3107 // Process them only once.
3108 if (index != lastInsIndex)
3110 ISD::ArgFlagsTy Flags = Ins[index].Flags;
3111 // FIXME: For now, all byval parameter objects are marked mutable.
3112 // This can be changed with more analysis.
3113 // In case of tail call optimization mark all arguments mutable.
3114 // Since they could be overwritten by lowering of arguments in case of
3116 if (Flags.isByVal()) {
3117 assert(Ins[index].isOrigArg() &&
3118 "Byval arguments cannot be implicit");
3119 unsigned CurByValIndex = CCInfo.getInRegsParamsProcessed();
3121 int FrameIndex = StoreByValRegs(CCInfo, DAG, dl, Chain, CurOrigArg,
3122 CurByValIndex, VA.getLocMemOffset(),
3123 Flags.getByValSize());
3124 InVals.push_back(DAG.getFrameIndex(FrameIndex, PtrVT));
3125 CCInfo.nextInRegsParam();
3127 unsigned FIOffset = VA.getLocMemOffset();
3128 int FI = MFI->CreateFixedObject(VA.getLocVT().getSizeInBits()/8,
3131 // Create load nodes to retrieve arguments from the stack.
3132 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
3133 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
3134 MachinePointerInfo::getFixedStack(FI),
3135 false, false, false, 0));
3137 lastInsIndex = index;
3143 if (isVarArg && MFI->hasVAStart())
3144 VarArgStyleRegisters(CCInfo, DAG, dl, Chain,
3145 CCInfo.getNextStackOffset(),
3146 TotalArgRegsSaveSize);
3148 AFI->setArgumentStackSize(CCInfo.getNextStackOffset());
3153 /// isFloatingPointZero - Return true if this is +0.0.
3154 static bool isFloatingPointZero(SDValue Op) {
3155 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
3156 return CFP->getValueAPF().isPosZero();
3157 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
3158 // Maybe this has already been legalized into the constant pool?
3159 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
3160 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
3161 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
3162 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
3163 return CFP->getValueAPF().isPosZero();
3165 } else if (Op->getOpcode() == ISD::BITCAST &&
3166 Op->getValueType(0) == MVT::f64) {
3167 // Handle (ISD::BITCAST (ARMISD::VMOVIMM (ISD::TargetConstant 0)) MVT::f64)
3168 // created by LowerConstantFP().
3169 SDValue BitcastOp = Op->getOperand(0);
3170 if (BitcastOp->getOpcode() == ARMISD::VMOVIMM) {
3171 SDValue MoveOp = BitcastOp->getOperand(0);
3172 if (MoveOp->getOpcode() == ISD::TargetConstant &&
3173 cast<ConstantSDNode>(MoveOp)->getZExtValue() == 0) {
3181 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
3182 /// the given operands.
3184 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
3185 SDValue &ARMcc, SelectionDAG &DAG,
3187 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
3188 unsigned C = RHSC->getZExtValue();
3189 if (!isLegalICmpImmediate(C)) {
3190 // Constant does not fit, try adjusting it by one?
3195 if (C != 0x80000000 && isLegalICmpImmediate(C-1)) {
3196 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
3197 RHS = DAG.getConstant(C - 1, dl, MVT::i32);
3202 if (C != 0 && isLegalICmpImmediate(C-1)) {
3203 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
3204 RHS = DAG.getConstant(C - 1, dl, MVT::i32);
3209 if (C != 0x7fffffff && isLegalICmpImmediate(C+1)) {
3210 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
3211 RHS = DAG.getConstant(C + 1, dl, MVT::i32);
3216 if (C != 0xffffffff && isLegalICmpImmediate(C+1)) {
3217 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
3218 RHS = DAG.getConstant(C + 1, dl, MVT::i32);
3225 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3226 ARMISD::NodeType CompareType;
3229 CompareType = ARMISD::CMP;
3234 CompareType = ARMISD::CMPZ;
3237 ARMcc = DAG.getConstant(CondCode, dl, MVT::i32);
3238 return DAG.getNode(CompareType, dl, MVT::Glue, LHS, RHS);
3241 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
3243 ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
3245 assert(!Subtarget->isFPOnlySP() || RHS.getValueType() != MVT::f64);
3247 if (!isFloatingPointZero(RHS))
3248 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Glue, LHS, RHS);
3250 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Glue, LHS);
3251 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Glue, Cmp);
3254 /// duplicateCmp - Glue values can have only one use, so this function
3255 /// duplicates a comparison node.
3257 ARMTargetLowering::duplicateCmp(SDValue Cmp, SelectionDAG &DAG) const {
3258 unsigned Opc = Cmp.getOpcode();
3260 if (Opc == ARMISD::CMP || Opc == ARMISD::CMPZ)
3261 return DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
3263 assert(Opc == ARMISD::FMSTAT && "unexpected comparison operation");
3264 Cmp = Cmp.getOperand(0);
3265 Opc = Cmp.getOpcode();
3266 if (Opc == ARMISD::CMPFP)
3267 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
3269 assert(Opc == ARMISD::CMPFPw0 && "unexpected operand of FMSTAT");
3270 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0));
3272 return DAG.getNode(ARMISD::FMSTAT, DL, MVT::Glue, Cmp);
3275 std::pair<SDValue, SDValue>
3276 ARMTargetLowering::getARMXALUOOp(SDValue Op, SelectionDAG &DAG,
3277 SDValue &ARMcc) const {
3278 assert(Op.getValueType() == MVT::i32 && "Unsupported value type");
3280 SDValue Value, OverflowCmp;
3281 SDValue LHS = Op.getOperand(0);
3282 SDValue RHS = Op.getOperand(1);
3285 // FIXME: We are currently always generating CMPs because we don't support
3286 // generating CMN through the backend. This is not as good as the natural
3287 // CMP case because it causes a register dependency and cannot be folded
3290 switch (Op.getOpcode()) {
3292 llvm_unreachable("Unknown overflow instruction!");
3294 ARMcc = DAG.getConstant(ARMCC::VC, dl, MVT::i32);
3295 Value = DAG.getNode(ISD::ADD, dl, Op.getValueType(), LHS, RHS);
3296 OverflowCmp = DAG.getNode(ARMISD::CMP, dl, MVT::Glue, Value, LHS);
3299 ARMcc = DAG.getConstant(ARMCC::HS, dl, MVT::i32);
3300 Value = DAG.getNode(ISD::ADD, dl, Op.getValueType(), LHS, RHS);
3301 OverflowCmp = DAG.getNode(ARMISD::CMP, dl, MVT::Glue, Value, LHS);
3304 ARMcc = DAG.getConstant(ARMCC::VC, dl, MVT::i32);
3305 Value = DAG.getNode(ISD::SUB, dl, Op.getValueType(), LHS, RHS);
3306 OverflowCmp = DAG.getNode(ARMISD::CMP, dl, MVT::Glue, LHS, RHS);
3309 ARMcc = DAG.getConstant(ARMCC::HS, dl, MVT::i32);
3310 Value = DAG.getNode(ISD::SUB, dl, Op.getValueType(), LHS, RHS);
3311 OverflowCmp = DAG.getNode(ARMISD::CMP, dl, MVT::Glue, LHS, RHS);
3315 return std::make_pair(Value, OverflowCmp);
3320 ARMTargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) const {
3321 // Let legalize expand this if it isn't a legal type yet.
3322 if (!DAG.getTargetLoweringInfo().isTypeLegal(Op.getValueType()))
3325 SDValue Value, OverflowCmp;
3327 std::tie(Value, OverflowCmp) = getARMXALUOOp(Op, DAG, ARMcc);
3328 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3330 // We use 0 and 1 as false and true values.
3331 SDValue TVal = DAG.getConstant(1, dl, MVT::i32);
3332 SDValue FVal = DAG.getConstant(0, dl, MVT::i32);
3333 EVT VT = Op.getValueType();
3335 SDValue Overflow = DAG.getNode(ARMISD::CMOV, dl, VT, TVal, FVal,
3336 ARMcc, CCR, OverflowCmp);
3338 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
3339 return DAG.getNode(ISD::MERGE_VALUES, dl, VTs, Value, Overflow);
3343 SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
3344 SDValue Cond = Op.getOperand(0);
3345 SDValue SelectTrue = Op.getOperand(1);
3346 SDValue SelectFalse = Op.getOperand(2);
3348 unsigned Opc = Cond.getOpcode();
3350 if (Cond.getResNo() == 1 &&
3351 (Opc == ISD::SADDO || Opc == ISD::UADDO || Opc == ISD::SSUBO ||
3352 Opc == ISD::USUBO)) {
3353 if (!DAG.getTargetLoweringInfo().isTypeLegal(Cond->getValueType(0)))
3356 SDValue Value, OverflowCmp;
3358 std::tie(Value, OverflowCmp) = getARMXALUOOp(Cond, DAG, ARMcc);
3359 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3360 EVT VT = Op.getValueType();
3362 return getCMOV(dl, VT, SelectTrue, SelectFalse, ARMcc, CCR,
3368 // (select (cmov 1, 0, cond), t, f) -> (cmov t, f, cond)
3369 // (select (cmov 0, 1, cond), t, f) -> (cmov f, t, cond)
3371 if (Cond.getOpcode() == ARMISD::CMOV && Cond.hasOneUse()) {
3372 const ConstantSDNode *CMOVTrue =
3373 dyn_cast<ConstantSDNode>(Cond.getOperand(0));
3374 const ConstantSDNode *CMOVFalse =
3375 dyn_cast<ConstantSDNode>(Cond.getOperand(1));
3377 if (CMOVTrue && CMOVFalse) {
3378 unsigned CMOVTrueVal = CMOVTrue->getZExtValue();
3379 unsigned CMOVFalseVal = CMOVFalse->getZExtValue();
3383 if (CMOVTrueVal == 1 && CMOVFalseVal == 0) {
3385 False = SelectFalse;
3386 } else if (CMOVTrueVal == 0 && CMOVFalseVal == 1) {
3391 if (True.getNode() && False.getNode()) {
3392 EVT VT = Op.getValueType();
3393 SDValue ARMcc = Cond.getOperand(2);
3394 SDValue CCR = Cond.getOperand(3);
3395 SDValue Cmp = duplicateCmp(Cond.getOperand(4), DAG);
3396 assert(True.getValueType() == VT);
3397 return getCMOV(dl, VT, True, False, ARMcc, CCR, Cmp, DAG);
3402 // ARM's BooleanContents value is UndefinedBooleanContent. Mask out the
3403 // undefined bits before doing a full-word comparison with zero.
3404 Cond = DAG.getNode(ISD::AND, dl, Cond.getValueType(), Cond,
3405 DAG.getConstant(1, dl, Cond.getValueType()));
3407 return DAG.getSelectCC(dl, Cond,
3408 DAG.getConstant(0, dl, Cond.getValueType()),
3409 SelectTrue, SelectFalse, ISD::SETNE);
3412 static void checkVSELConstraints(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
3413 bool &swpCmpOps, bool &swpVselOps) {
3414 // Start by selecting the GE condition code for opcodes that return true for
3416 if (CC == ISD::SETUGE || CC == ISD::SETOGE || CC == ISD::SETOLE ||
3418 CondCode = ARMCC::GE;
3420 // and GT for opcodes that return false for 'equality'.
3421 else if (CC == ISD::SETUGT || CC == ISD::SETOGT || CC == ISD::SETOLT ||
3423 CondCode = ARMCC::GT;
3425 // Since we are constrained to GE/GT, if the opcode contains 'less', we need
3426 // to swap the compare operands.
3427 if (CC == ISD::SETOLE || CC == ISD::SETULE || CC == ISD::SETOLT ||
3431 // Both GT and GE are ordered comparisons, and return false for 'unordered'.
3432 // If we have an unordered opcode, we need to swap the operands to the VSEL
3433 // instruction (effectively negating the condition).
3435 // This also has the effect of swapping which one of 'less' or 'greater'
3436 // returns true, so we also swap the compare operands. It also switches
3437 // whether we return true for 'equality', so we compensate by picking the
3438 // opposite condition code to our original choice.
3439 if (CC == ISD::SETULE || CC == ISD::SETULT || CC == ISD::SETUGE ||
3440 CC == ISD::SETUGT) {
3441 swpCmpOps = !swpCmpOps;
3442 swpVselOps = !swpVselOps;
3443 CondCode = CondCode == ARMCC::GT ? ARMCC::GE : ARMCC::GT;
3446 // 'ordered' is 'anything but unordered', so use the VS condition code and
3447 // swap the VSEL operands.
3448 if (CC == ISD::SETO) {
3449 CondCode = ARMCC::VS;
3453 // 'unordered or not equal' is 'anything but equal', so use the EQ condition
3454 // code and swap the VSEL operands.
3455 if (CC == ISD::SETUNE) {
3456 CondCode = ARMCC::EQ;
3461 SDValue ARMTargetLowering::getCMOV(SDLoc dl, EVT VT, SDValue FalseVal,
3462 SDValue TrueVal, SDValue ARMcc, SDValue CCR,
3463 SDValue Cmp, SelectionDAG &DAG) const {
3464 if (Subtarget->isFPOnlySP() && VT == MVT::f64) {
3465 FalseVal = DAG.getNode(ARMISD::VMOVRRD, dl,
3466 DAG.getVTList(MVT::i32, MVT::i32), FalseVal);
3467 TrueVal = DAG.getNode(ARMISD::VMOVRRD, dl,
3468 DAG.getVTList(MVT::i32, MVT::i32), TrueVal);
3470 SDValue TrueLow = TrueVal.getValue(0);
3471 SDValue TrueHigh = TrueVal.getValue(1);
3472 SDValue FalseLow = FalseVal.getValue(0);
3473 SDValue FalseHigh = FalseVal.getValue(1);
3475 SDValue Low = DAG.getNode(ARMISD::CMOV, dl, MVT::i32, FalseLow, TrueLow,
3477 SDValue High = DAG.getNode(ARMISD::CMOV, dl, MVT::i32, FalseHigh, TrueHigh,
3478 ARMcc, CCR, duplicateCmp(Cmp, DAG));
3480 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Low, High);
3482 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,
3487 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
3488 EVT VT = Op.getValueType();
3489 SDValue LHS = Op.getOperand(0);
3490 SDValue RHS = Op.getOperand(1);
3491 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
3492 SDValue TrueVal = Op.getOperand(2);
3493 SDValue FalseVal = Op.getOperand(3);
3496 if (Subtarget->isFPOnlySP() && LHS.getValueType() == MVT::f64) {
3497 DAG.getTargetLoweringInfo().softenSetCCOperands(DAG, MVT::f64, LHS, RHS, CC,
3500 // If softenSetCCOperands only returned one value, we should compare it to
3502 if (!RHS.getNode()) {
3503 RHS = DAG.getConstant(0, dl, LHS.getValueType());
3508 if (LHS.getValueType() == MVT::i32) {
3509 // Try to generate VSEL on ARMv8.
3510 // The VSEL instruction can't use all the usual ARM condition
3511 // codes: it only has two bits to select the condition code, so it's
3512 // constrained to use only GE, GT, VS and EQ.
3514 // To implement all the various ISD::SETXXX opcodes, we sometimes need to
3515 // swap the operands of the previous compare instruction (effectively
3516 // inverting the compare condition, swapping 'less' and 'greater') and
3517 // sometimes need to swap the operands to the VSEL (which inverts the
3518 // condition in the sense of firing whenever the previous condition didn't)
3519 if (Subtarget->hasFPARMv8() && (TrueVal.getValueType() == MVT::f32 ||
3520 TrueVal.getValueType() == MVT::f64)) {
3521 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3522 if (CondCode == ARMCC::LT || CondCode == ARMCC::LE ||
3523 CondCode == ARMCC::VC || CondCode == ARMCC::NE) {
3524 CC = ISD::getSetCCInverse(CC, true);
3525 std::swap(TrueVal, FalseVal);
3530 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3531 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3532 return getCMOV(dl, VT, FalseVal, TrueVal, ARMcc, CCR, Cmp, DAG);
3535 ARMCC::CondCodes CondCode, CondCode2;
3536 FPCCToARMCC(CC, CondCode, CondCode2);
3538 // Try to generate VMAXNM/VMINNM on ARMv8.
3539 if (Subtarget->hasFPARMv8() && (TrueVal.getValueType() == MVT::f32 ||
3540 TrueVal.getValueType() == MVT::f64)) {
3541 // We can use VMAXNM/VMINNM for a compare followed by a select with the
3542 // same operands, as follows:
3543 // c = fcmp [?gt, ?ge, ?lt, ?le] a, b
3545 // In NoNaNsFPMath the CC will have been changed from, e.g., 'ogt' to 'gt'.
3546 bool swapSides = false;
3547 if (!getTargetMachine().Options.NoNaNsFPMath) {
3548 // transformability may depend on which way around we compare
3556 // the non-NaN should be RHS
3557 swapSides = DAG.isKnownNeverNaN(LHS) && !DAG.isKnownNeverNaN(RHS);
3563 // the non-NaN should be LHS
3564 swapSides = DAG.isKnownNeverNaN(RHS) && !DAG.isKnownNeverNaN(LHS);
3568 swapSides = swapSides || (LHS == FalseVal && RHS == TrueVal);
3570 CC = ISD::getSetCCSwappedOperands(CC);
3571 std::swap(LHS, RHS);
3573 if (LHS == TrueVal && RHS == FalseVal) {
3574 bool canTransform = true;
3575 // FIXME: FastMathFlags::noSignedZeros() doesn't appear reachable from here
3576 if (!getTargetMachine().Options.UnsafeFPMath &&
3577 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
3578 const ConstantFPSDNode *Zero;
3585 // RHS must not be -0
3586 canTransform = (Zero = dyn_cast<ConstantFPSDNode>(RHS)) &&
3587 !Zero->isNegative();
3592 // LHS must not be -0
3593 canTransform = (Zero = dyn_cast<ConstantFPSDNode>(LHS)) &&
3594 !Zero->isNegative();
3599 // RHS must not be +0
3600 canTransform = (Zero = dyn_cast<ConstantFPSDNode>(RHS)) &&
3606 // LHS must not be +0
3607 canTransform = (Zero = dyn_cast<ConstantFPSDNode>(LHS)) &&
3613 // Note: If one of the elements in a pair is a number and the other
3614 // element is NaN, the corresponding result element is the number.
3615 // This is consistent with the IEEE 754-2008 standard.
3616 // Therefore, a > b ? a : b <=> vmax(a,b), if b is constant and a is NaN
3622 if (!DAG.isKnownNeverNaN(RHS))
3624 return DAG.getNode(ARMISD::VMAXNM, dl, VT, LHS, RHS);
3627 if (!DAG.isKnownNeverNaN(LHS))
3631 return DAG.getNode(ARMISD::VMAXNM, dl, VT, LHS, RHS);
3634 if (!DAG.isKnownNeverNaN(RHS))
3636 return DAG.getNode(ARMISD::VMINNM, dl, VT, LHS, RHS);
3639 if (!DAG.isKnownNeverNaN(LHS))
3643 return DAG.getNode(ARMISD::VMINNM, dl, VT, LHS, RHS);
3648 bool swpCmpOps = false;
3649 bool swpVselOps = false;
3650 checkVSELConstraints(CC, CondCode, swpCmpOps, swpVselOps);
3652 if (CondCode == ARMCC::GT || CondCode == ARMCC::GE ||
3653 CondCode == ARMCC::VS || CondCode == ARMCC::EQ) {
3655 std::swap(LHS, RHS);
3657 std::swap(TrueVal, FalseVal);
3661 SDValue ARMcc = DAG.getConstant(CondCode, dl, MVT::i32);
3662 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
3663 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3664 SDValue Result = getCMOV(dl, VT, FalseVal, TrueVal, ARMcc, CCR, Cmp, DAG);
3665 if (CondCode2 != ARMCC::AL) {
3666 SDValue ARMcc2 = DAG.getConstant(CondCode2, dl, MVT::i32);
3667 // FIXME: Needs another CMP because flag can have but one use.
3668 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
3669 Result = getCMOV(dl, VT, Result, TrueVal, ARMcc2, CCR, Cmp2, DAG);
3674 /// canChangeToInt - Given the fp compare operand, return true if it is suitable
3675 /// to morph to an integer compare sequence.
3676 static bool canChangeToInt(SDValue Op, bool &SeenZero,
3677 const ARMSubtarget *Subtarget) {
3678 SDNode *N = Op.getNode();
3679 if (!N->hasOneUse())
3680 // Otherwise it requires moving the value from fp to integer registers.
3682 if (!N->getNumValues())
3684 EVT VT = Op.getValueType();
3685 if (VT != MVT::f32 && !Subtarget->isFPBrccSlow())
3686 // f32 case is generally profitable. f64 case only makes sense when vcmpe +
3687 // vmrs are very slow, e.g. cortex-a8.
3690 if (isFloatingPointZero(Op)) {
3694 return ISD::isNormalLoad(N);
3697 static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) {
3698 if (isFloatingPointZero(Op))
3699 return DAG.getConstant(0, SDLoc(Op), MVT::i32);
3701 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op))
3702 return DAG.getLoad(MVT::i32, SDLoc(Op),
3703 Ld->getChain(), Ld->getBasePtr(), Ld->getPointerInfo(),
3704 Ld->isVolatile(), Ld->isNonTemporal(),
3705 Ld->isInvariant(), Ld->getAlignment());
3707 llvm_unreachable("Unknown VFP cmp argument!");
3710 static void expandf64Toi32(SDValue Op, SelectionDAG &DAG,
3711 SDValue &RetVal1, SDValue &RetVal2) {
3714 if (isFloatingPointZero(Op)) {
3715 RetVal1 = DAG.getConstant(0, dl, MVT::i32);
3716 RetVal2 = DAG.getConstant(0, dl, MVT::i32);
3720 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) {
3721 SDValue Ptr = Ld->getBasePtr();
3722 RetVal1 = DAG.getLoad(MVT::i32, dl,
3723 Ld->getChain(), Ptr,
3724 Ld->getPointerInfo(),
3725 Ld->isVolatile(), Ld->isNonTemporal(),
3726 Ld->isInvariant(), Ld->getAlignment());
3728 EVT PtrType = Ptr.getValueType();
3729 unsigned NewAlign = MinAlign(Ld->getAlignment(), 4);
3730 SDValue NewPtr = DAG.getNode(ISD::ADD, dl,
3731 PtrType, Ptr, DAG.getConstant(4, dl, PtrType));
3732 RetVal2 = DAG.getLoad(MVT::i32, dl,
3733 Ld->getChain(), NewPtr,
3734 Ld->getPointerInfo().getWithOffset(4),
3735 Ld->isVolatile(), Ld->isNonTemporal(),
3736 Ld->isInvariant(), NewAlign);
3740 llvm_unreachable("Unknown VFP cmp argument!");
3743 /// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some
3744 /// f32 and even f64 comparisons to integer ones.
3746 ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const {
3747 SDValue Chain = Op.getOperand(0);
3748 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3749 SDValue LHS = Op.getOperand(2);
3750 SDValue RHS = Op.getOperand(3);
3751 SDValue Dest = Op.getOperand(4);
3754 bool LHSSeenZero = false;
3755 bool LHSOk = canChangeToInt(LHS, LHSSeenZero, Subtarget);
3756 bool RHSSeenZero = false;
3757 bool RHSOk = canChangeToInt(RHS, RHSSeenZero, Subtarget);
3758 if (LHSOk && RHSOk && (LHSSeenZero || RHSSeenZero)) {
3759 // If unsafe fp math optimization is enabled and there are no other uses of
3760 // the CMP operands, and the condition code is EQ or NE, we can optimize it
3761 // to an integer comparison.
3762 if (CC == ISD::SETOEQ)
3764 else if (CC == ISD::SETUNE)
3767 SDValue Mask = DAG.getConstant(0x7fffffff, dl, MVT::i32);
3769 if (LHS.getValueType() == MVT::f32) {
3770 LHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3771 bitcastf32Toi32(LHS, DAG), Mask);
3772 RHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3773 bitcastf32Toi32(RHS, DAG), Mask);
3774 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3775 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3776 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3777 Chain, Dest, ARMcc, CCR, Cmp);
3782 expandf64Toi32(LHS, DAG, LHS1, LHS2);
3783 expandf64Toi32(RHS, DAG, RHS1, RHS2);
3784 LHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, LHS2, Mask);
3785 RHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, RHS2, Mask);
3786 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3787 ARMcc = DAG.getConstant(CondCode, dl, MVT::i32);
3788 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3789 SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest };
3790 return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops);
3796 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
3797 SDValue Chain = Op.getOperand(0);
3798 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3799 SDValue LHS = Op.getOperand(2);
3800 SDValue RHS = Op.getOperand(3);
3801 SDValue Dest = Op.getOperand(4);
3804 if (Subtarget->isFPOnlySP() && LHS.getValueType() == MVT::f64) {
3805 DAG.getTargetLoweringInfo().softenSetCCOperands(DAG, MVT::f64, LHS, RHS, CC,
3808 // If softenSetCCOperands only returned one value, we should compare it to
3810 if (!RHS.getNode()) {
3811 RHS = DAG.getConstant(0, dl, LHS.getValueType());
3816 if (LHS.getValueType() == MVT::i32) {
3818 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3819 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3820 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3821 Chain, Dest, ARMcc, CCR, Cmp);
3824 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3826 if (getTargetMachine().Options.UnsafeFPMath &&
3827 (CC == ISD::SETEQ || CC == ISD::SETOEQ ||
3828 CC == ISD::SETNE || CC == ISD::SETUNE)) {
3829 SDValue Result = OptimizeVFPBrcond(Op, DAG);
3830 if (Result.getNode())
3834 ARMCC::CondCodes CondCode, CondCode2;
3835 FPCCToARMCC(CC, CondCode, CondCode2);
3837 SDValue ARMcc = DAG.getConstant(CondCode, dl, MVT::i32);
3838 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
3839 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3840 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3841 SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp };
3842 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops);
3843 if (CondCode2 != ARMCC::AL) {
3844 ARMcc = DAG.getConstant(CondCode2, dl, MVT::i32);
3845 SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) };
3846 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops);
3851 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
3852 SDValue Chain = Op.getOperand(0);
3853 SDValue Table = Op.getOperand(1);
3854 SDValue Index = Op.getOperand(2);
3857 EVT PTy = getPointerTy(DAG.getDataLayout());
3858 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
3859 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
3860 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI);
3861 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, dl, PTy));
3862 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
3863 if (Subtarget->isThumb2()) {
3864 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
3865 // which does another jump to the destination. This also makes it easier
3866 // to translate it to TBB / TBH later.
3867 // FIXME: This might not work if the function is extremely large.
3868 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
3869 Addr, Op.getOperand(2), JTI);
3871 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
3872 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
3873 MachinePointerInfo::getJumpTable(),
3874 false, false, false, 0);
3875 Chain = Addr.getValue(1);
3876 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
3877 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI);
3879 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
3880 MachinePointerInfo::getJumpTable(),
3881 false, false, false, 0);
3882 Chain = Addr.getValue(1);
3883 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI);
3887 static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3888 EVT VT = Op.getValueType();
3891 if (Op.getValueType().getVectorElementType() == MVT::i32) {
3892 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::f32)
3894 return DAG.UnrollVectorOp(Op.getNode());
3897 assert(Op.getOperand(0).getValueType() == MVT::v4f32 &&
3898 "Invalid type for custom lowering!");
3899 if (VT != MVT::v4i16)
3900 return DAG.UnrollVectorOp(Op.getNode());
3902 Op = DAG.getNode(Op.getOpcode(), dl, MVT::v4i32, Op.getOperand(0));
3903 return DAG.getNode(ISD::TRUNCATE, dl, VT, Op);
3906 SDValue ARMTargetLowering::LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) const {
3907 EVT VT = Op.getValueType();
3909 return LowerVectorFP_TO_INT(Op, DAG);
3910 if (Subtarget->isFPOnlySP() && Op.getOperand(0).getValueType() == MVT::f64) {
3912 if (Op.getOpcode() == ISD::FP_TO_SINT)
3913 LC = RTLIB::getFPTOSINT(Op.getOperand(0).getValueType(),
3916 LC = RTLIB::getFPTOUINT(Op.getOperand(0).getValueType(),
3918 return makeLibCall(DAG, LC, Op.getValueType(), &Op.getOperand(0), 1,
3919 /*isSigned*/ false, SDLoc(Op)).first;
3925 static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3926 EVT VT = Op.getValueType();
3929 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::i32) {
3930 if (VT.getVectorElementType() == MVT::f32)
3932 return DAG.UnrollVectorOp(Op.getNode());
3935 assert(Op.getOperand(0).getValueType() == MVT::v4i16 &&
3936 "Invalid type for custom lowering!");
3937 if (VT != MVT::v4f32)
3938 return DAG.UnrollVectorOp(Op.getNode());
3942 switch (Op.getOpcode()) {
3943 default: llvm_unreachable("Invalid opcode!");
3944 case ISD::SINT_TO_FP:
3945 CastOpc = ISD::SIGN_EXTEND;
3946 Opc = ISD::SINT_TO_FP;
3948 case ISD::UINT_TO_FP:
3949 CastOpc = ISD::ZERO_EXTEND;
3950 Opc = ISD::UINT_TO_FP;
3954 Op = DAG.getNode(CastOpc, dl, MVT::v4i32, Op.getOperand(0));
3955 return DAG.getNode(Opc, dl, VT, Op);
3958 SDValue ARMTargetLowering::LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) const {
3959 EVT VT = Op.getValueType();
3961 return LowerVectorINT_TO_FP(Op, DAG);
3962 if (Subtarget->isFPOnlySP() && Op.getValueType() == MVT::f64) {
3964 if (Op.getOpcode() == ISD::SINT_TO_FP)
3965 LC = RTLIB::getSINTTOFP(Op.getOperand(0).getValueType(),
3968 LC = RTLIB::getUINTTOFP(Op.getOperand(0).getValueType(),
3970 return makeLibCall(DAG, LC, Op.getValueType(), &Op.getOperand(0), 1,
3971 /*isSigned*/ false, SDLoc(Op)).first;
3977 SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
3978 // Implement fcopysign with a fabs and a conditional fneg.
3979 SDValue Tmp0 = Op.getOperand(0);
3980 SDValue Tmp1 = Op.getOperand(1);
3982 EVT VT = Op.getValueType();
3983 EVT SrcVT = Tmp1.getValueType();
3984 bool InGPR = Tmp0.getOpcode() == ISD::BITCAST ||
3985 Tmp0.getOpcode() == ARMISD::VMOVDRR;
3986 bool UseNEON = !InGPR && Subtarget->hasNEON();
3989 // Use VBSL to copy the sign bit.
3990 unsigned EncodedVal = ARM_AM::createNEONModImm(0x6, 0x80);
3991 SDValue Mask = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v2i32,
3992 DAG.getTargetConstant(EncodedVal, dl, MVT::i32));
3993 EVT OpVT = (VT == MVT::f32) ? MVT::v2i32 : MVT::v1i64;
3995 Mask = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3996 DAG.getNode(ISD::BITCAST, dl, OpVT, Mask),
3997 DAG.getConstant(32, dl, MVT::i32));
3998 else /*if (VT == MVT::f32)*/
3999 Tmp0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp0);
4000 if (SrcVT == MVT::f32) {
4001 Tmp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp1);
4003 Tmp1 = DAG.getNode(ARMISD::VSHL, dl, OpVT,
4004 DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1),
4005 DAG.getConstant(32, dl, MVT::i32));
4006 } else if (VT == MVT::f32)
4007 Tmp1 = DAG.getNode(ARMISD::VSHRu, dl, MVT::v1i64,
4008 DAG.getNode(ISD::BITCAST, dl, MVT::v1i64, Tmp1),
4009 DAG.getConstant(32, dl, MVT::i32));
4010 Tmp0 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp0);
4011 Tmp1 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1);
4013 SDValue AllOnes = DAG.getTargetConstant(ARM_AM::createNEONModImm(0xe, 0xff),
4015 AllOnes = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v8i8, AllOnes);
4016 SDValue MaskNot = DAG.getNode(ISD::XOR, dl, OpVT, Mask,
4017 DAG.getNode(ISD::BITCAST, dl, OpVT, AllOnes));
4019 SDValue Res = DAG.getNode(ISD::OR, dl, OpVT,
4020 DAG.getNode(ISD::AND, dl, OpVT, Tmp1, Mask),
4021 DAG.getNode(ISD::AND, dl, OpVT, Tmp0, MaskNot));
4022 if (VT == MVT::f32) {
4023 Res = DAG.getNode(ISD::BITCAST, dl, MVT::v2f32, Res);
4024 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, Res,
4025 DAG.getConstant(0, dl, MVT::i32));
4027 Res = DAG.getNode(ISD::BITCAST, dl, MVT::f64, Res);
4033 // Bitcast operand 1 to i32.
4034 if (SrcVT == MVT::f64)
4035 Tmp1 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
4037 Tmp1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp1);
4039 // Or in the signbit with integer operations.
4040 SDValue Mask1 = DAG.getConstant(0x80000000, dl, MVT::i32);
4041 SDValue Mask2 = DAG.getConstant(0x7fffffff, dl, MVT::i32);
4042 Tmp1 = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp1, Mask1);
4043 if (VT == MVT::f32) {
4044 Tmp0 = DAG.getNode(ISD::AND, dl, MVT::i32,
4045 DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp0), Mask2);
4046 return DAG.getNode(ISD::BITCAST, dl, MVT::f32,
4047 DAG.getNode(ISD::OR, dl, MVT::i32, Tmp0, Tmp1));
4050 // f64: Or the high part with signbit and then combine two parts.
4051 Tmp0 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
4053 SDValue Lo = Tmp0.getValue(0);
4054 SDValue Hi = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp0.getValue(1), Mask2);
4055 Hi = DAG.getNode(ISD::OR, dl, MVT::i32, Hi, Tmp1);
4056 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
4059 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
4060 MachineFunction &MF = DAG.getMachineFunction();
4061 MachineFrameInfo *MFI = MF.getFrameInfo();
4062 MFI->setReturnAddressIsTaken(true);
4064 if (verifyReturnAddressArgumentIsConstant(Op, DAG))
4067 EVT VT = Op.getValueType();
4069 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
4071 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
4072 SDValue Offset = DAG.getConstant(4, dl, MVT::i32);
4073 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
4074 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
4075 MachinePointerInfo(), false, false, false, 0);
4078 // Return LR, which contains the return address. Mark it an implicit live-in.
4079 unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32));
4080 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
4083 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
4084 const ARMBaseRegisterInfo &ARI =
4085 *static_cast<const ARMBaseRegisterInfo*>(RegInfo);
4086 MachineFunction &MF = DAG.getMachineFunction();
4087 MachineFrameInfo *MFI = MF.getFrameInfo();
4088 MFI->setFrameAddressIsTaken(true);
4090 EVT VT = Op.getValueType();
4091 SDLoc dl(Op); // FIXME probably not meaningful
4092 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
4093 unsigned FrameReg = ARI.getFrameRegister(MF);
4094 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
4096 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
4097 MachinePointerInfo(),
4098 false, false, false, 0);
4102 // FIXME? Maybe this could be a TableGen attribute on some registers and
4103 // this table could be generated automatically from RegInfo.
4104 unsigned ARMTargetLowering::getRegisterByName(const char* RegName, EVT VT,
4105 SelectionDAG &DAG) const {
4106 unsigned Reg = StringSwitch<unsigned>(RegName)
4107 .Case("sp", ARM::SP)
4111 report_fatal_error(Twine("Invalid register name \""
4112 + StringRef(RegName) + "\"."));
4115 // Result is 64 bit value so split into two 32 bit values and return as a
4117 static void ExpandREAD_REGISTER(SDNode *N, SmallVectorImpl<SDValue> &Results,
4118 SelectionDAG &DAG) {
4121 // This function is only supposed to be called for i64 type destination.
4122 assert(N->getValueType(0) == MVT::i64
4123 && "ExpandREAD_REGISTER called for non-i64 type result.");
4125 SDValue Read = DAG.getNode(ISD::READ_REGISTER, DL,
4126 DAG.getVTList(MVT::i32, MVT::i32, MVT::Other),
4130 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Read.getValue(0),
4132 Results.push_back(Read.getOperand(0));
4135 /// ExpandBITCAST - If the target supports VFP, this function is called to
4136 /// expand a bit convert where either the source or destination type is i64 to
4137 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
4138 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
4139 /// vectors), since the legalizer won't know what to do with that.
4140 static SDValue ExpandBITCAST(SDNode *N, SelectionDAG &DAG) {
4141 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4143 SDValue Op = N->getOperand(0);
4145 // This function is only supposed to be called for i64 types, either as the
4146 // source or destination of the bit convert.
4147 EVT SrcVT = Op.getValueType();
4148 EVT DstVT = N->getValueType(0);
4149 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
4150 "ExpandBITCAST called for non-i64 type");
4152 // Turn i64->f64 into VMOVDRR.
4153 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
4154 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
4155 DAG.getConstant(0, dl, MVT::i32));
4156 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
4157 DAG.getConstant(1, dl, MVT::i32));
4158 return DAG.getNode(ISD::BITCAST, dl, DstVT,
4159 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
4162 // Turn f64->i64 into VMOVRRD.
4163 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
4165 if (DAG.getDataLayout().isBigEndian() && SrcVT.isVector() &&
4166 SrcVT.getVectorNumElements() > 1)
4167 Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
4168 DAG.getVTList(MVT::i32, MVT::i32),
4169 DAG.getNode(ARMISD::VREV64, dl, SrcVT, Op));
4171 Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
4172 DAG.getVTList(MVT::i32, MVT::i32), Op);
4173 // Merge the pieces into a single i64 value.
4174 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
4180 /// getZeroVector - Returns a vector of specified type with all zero elements.
4181 /// Zero vectors are used to represent vector negation and in those cases
4182 /// will be implemented with the NEON VNEG instruction. However, VNEG does
4183 /// not support i64 elements, so sometimes the zero vectors will need to be
4184 /// explicitly constructed. Regardless, use a canonical VMOV to create the
4186 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, SDLoc dl) {
4187 assert(VT.isVector() && "Expected a vector type");
4188 // The canonical modified immediate encoding of a zero vector is....0!
4189 SDValue EncodedVal = DAG.getTargetConstant(0, dl, MVT::i32);
4190 EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
4191 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal);
4192 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4195 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
4196 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
4197 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
4198 SelectionDAG &DAG) const {
4199 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4200 EVT VT = Op.getValueType();
4201 unsigned VTBits = VT.getSizeInBits();
4203 SDValue ShOpLo = Op.getOperand(0);
4204 SDValue ShOpHi = Op.getOperand(1);
4205 SDValue ShAmt = Op.getOperand(2);
4207 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
4209 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
4211 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
4212 DAG.getConstant(VTBits, dl, MVT::i32), ShAmt);
4213 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
4214 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
4215 DAG.getConstant(VTBits, dl, MVT::i32));
4216 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
4217 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
4218 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
4220 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
4221 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, dl, MVT::i32),
4222 ISD::SETGE, ARMcc, DAG, dl);
4223 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
4224 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc,
4227 SDValue Ops[2] = { Lo, Hi };
4228 return DAG.getMergeValues(Ops, dl);
4231 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
4232 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
4233 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
4234 SelectionDAG &DAG) const {
4235 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4236 EVT VT = Op.getValueType();
4237 unsigned VTBits = VT.getSizeInBits();
4239 SDValue ShOpLo = Op.getOperand(0);
4240 SDValue ShOpHi = Op.getOperand(1);
4241 SDValue ShAmt = Op.getOperand(2);
4244 assert(Op.getOpcode() == ISD::SHL_PARTS);
4245 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
4246 DAG.getConstant(VTBits, dl, MVT::i32), ShAmt);
4247 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
4248 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
4249 DAG.getConstant(VTBits, dl, MVT::i32));
4250 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
4251 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
4253 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
4254 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
4255 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, dl, MVT::i32),
4256 ISD::SETGE, ARMcc, DAG, dl);
4257 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
4258 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc,
4261 SDValue Ops[2] = { Lo, Hi };
4262 return DAG.getMergeValues(Ops, dl);
4265 SDValue ARMTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
4266 SelectionDAG &DAG) const {
4267 // The rounding mode is in bits 23:22 of the FPSCR.
4268 // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0
4269 // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3)
4270 // so that the shift + and get folded into a bitfield extract.
4272 SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::i32,
4273 DAG.getConstant(Intrinsic::arm_get_fpscr, dl,
4275 SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPSCR,
4276 DAG.getConstant(1U << 22, dl, MVT::i32));
4277 SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds,
4278 DAG.getConstant(22, dl, MVT::i32));
4279 return DAG.getNode(ISD::AND, dl, MVT::i32, RMODE,
4280 DAG.getConstant(3, dl, MVT::i32));
4283 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
4284 const ARMSubtarget *ST) {
4285 EVT VT = N->getValueType(0);
4288 if (!ST->hasV6T2Ops())
4291 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
4292 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
4295 /// getCTPOP16BitCounts - Returns a v8i8/v16i8 vector containing the bit-count
4296 /// for each 16-bit element from operand, repeated. The basic idea is to
4297 /// leverage vcnt to get the 8-bit counts, gather and add the results.
4299 /// Trace for v4i16:
4300 /// input = [v0 v1 v2 v3 ] (vi 16-bit element)
4301 /// cast: N0 = [w0 w1 w2 w3 w4 w5 w6 w7] (v0 = [w0 w1], wi 8-bit element)
4302 /// vcnt: N1 = [b0 b1 b2 b3 b4 b5 b6 b7] (bi = bit-count of 8-bit element wi)
4303 /// vrev: N2 = [b1 b0 b3 b2 b5 b4 b7 b6]
4304 /// [b0 b1 b2 b3 b4 b5 b6 b7]
4305 /// +[b1 b0 b3 b2 b5 b4 b7 b6]
4306 /// N3=N1+N2 = [k0 k0 k1 k1 k2 k2 k3 k3] (k0 = b0+b1 = bit-count of 16-bit v0,
4307 /// vuzp: = [k0 k1 k2 k3 k0 k1 k2 k3] each ki is 8-bits)
4308 static SDValue getCTPOP16BitCounts(SDNode *N, SelectionDAG &DAG) {
4309 EVT VT = N->getValueType(0);
4312 EVT VT8Bit = VT.is64BitVector() ? MVT::v8i8 : MVT::v16i8;
4313 SDValue N0 = DAG.getNode(ISD::BITCAST, DL, VT8Bit, N->getOperand(0));
4314 SDValue N1 = DAG.getNode(ISD::CTPOP, DL, VT8Bit, N0);
4315 SDValue N2 = DAG.getNode(ARMISD::VREV16, DL, VT8Bit, N1);
4316 SDValue N3 = DAG.getNode(ISD::ADD, DL, VT8Bit, N1, N2);
4317 return DAG.getNode(ARMISD::VUZP, DL, VT8Bit, N3, N3);
4320 /// lowerCTPOP16BitElements - Returns a v4i16/v8i16 vector containing the
4321 /// bit-count for each 16-bit element from the operand. We need slightly
4322 /// different sequencing for v4i16 and v8i16 to stay within NEON's available
4323 /// 64/128-bit registers.
4325 /// Trace for v4i16:
4326 /// input = [v0 v1 v2 v3 ] (vi 16-bit element)
4327 /// v8i8: BitCounts = [k0 k1 k2 k3 k0 k1 k2 k3 ] (ki is the bit-count of vi)
4328 /// v8i16:Extended = [k0 k1 k2 k3 k0 k1 k2 k3 ]
4329 /// v4i16:Extracted = [k0 k1 k2 k3 ]
4330 static SDValue lowerCTPOP16BitElements(SDNode *N, SelectionDAG &DAG) {
4331 EVT VT = N->getValueType(0);
4334 SDValue BitCounts = getCTPOP16BitCounts(N, DAG);
4335 if (VT.is64BitVector()) {
4336 SDValue Extended = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v8i16, BitCounts);
4337 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i16, Extended,
4338 DAG.getIntPtrConstant(0, DL));
4340 SDValue Extracted = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v8i8,
4341 BitCounts, DAG.getIntPtrConstant(0, DL));
4342 return DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v8i16, Extracted);
4346 /// lowerCTPOP32BitElements - Returns a v2i32/v4i32 vector containing the
4347 /// bit-count for each 32-bit element from the operand. The idea here is
4348 /// to split the vector into 16-bit elements, leverage the 16-bit count
4349 /// routine, and then combine the results.
4351 /// Trace for v2i32 (v4i32 similar with Extracted/Extended exchanged):
4352 /// input = [v0 v1 ] (vi: 32-bit elements)
4353 /// Bitcast = [w0 w1 w2 w3 ] (wi: 16-bit elements, v0 = [w0 w1])
4354 /// Counts16 = [k0 k1 k2 k3 ] (ki: 16-bit elements, bit-count of wi)
4355 /// vrev: N0 = [k1 k0 k3 k2 ]
4357 /// N1 =+[k1 k0 k3 k2 ]
4359 /// N2 =+[k1 k3 k0 k2 ]
4361 /// Extended =+[k1 k3 k0 k2 ]
4363 /// Extracted=+[k1 k3 ]
4365 static SDValue lowerCTPOP32BitElements(SDNode *N, SelectionDAG &DAG) {
4366 EVT VT = N->getValueType(0);
4369 EVT VT16Bit = VT.is64BitVector() ? MVT::v4i16 : MVT::v8i16;
4371 SDValue Bitcast = DAG.getNode(ISD::BITCAST, DL, VT16Bit, N->getOperand(0));
4372 SDValue Counts16 = lowerCTPOP16BitElements(Bitcast.getNode(), DAG);
4373 SDValue N0 = DAG.getNode(ARMISD::VREV32, DL, VT16Bit, Counts16);
4374 SDValue N1 = DAG.getNode(ISD::ADD, DL, VT16Bit, Counts16, N0);
4375 SDValue N2 = DAG.getNode(ARMISD::VUZP, DL, VT16Bit, N1, N1);
4377 if (VT.is64BitVector()) {
4378 SDValue Extended = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v4i32, N2);
4379 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i32, Extended,
4380 DAG.getIntPtrConstant(0, DL));
4382 SDValue Extracted = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i16, N2,
4383 DAG.getIntPtrConstant(0, DL));
4384 return DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v4i32, Extracted);
4388 static SDValue LowerCTPOP(SDNode *N, SelectionDAG &DAG,
4389 const ARMSubtarget *ST) {
4390 EVT VT = N->getValueType(0);
4392 assert(ST->hasNEON() && "Custom ctpop lowering requires NEON.");
4393 assert((VT == MVT::v2i32 || VT == MVT::v4i32 ||
4394 VT == MVT::v4i16 || VT == MVT::v8i16) &&
4395 "Unexpected type for custom ctpop lowering");
4397 if (VT.getVectorElementType() == MVT::i32)
4398 return lowerCTPOP32BitElements(N, DAG);
4400 return lowerCTPOP16BitElements(N, DAG);
4403 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
4404 const ARMSubtarget *ST) {
4405 EVT VT = N->getValueType(0);
4411 // Lower vector shifts on NEON to use VSHL.
4412 assert(ST->hasNEON() && "unexpected vector shift");
4414 // Left shifts translate directly to the vshiftu intrinsic.
4415 if (N->getOpcode() == ISD::SHL)
4416 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
4417 DAG.getConstant(Intrinsic::arm_neon_vshiftu, dl,
4419 N->getOperand(0), N->getOperand(1));
4421 assert((N->getOpcode() == ISD::SRA ||
4422 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
4424 // NEON uses the same intrinsics for both left and right shifts. For
4425 // right shifts, the shift amounts are negative, so negate the vector of
4427 EVT ShiftVT = N->getOperand(1).getValueType();
4428 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
4429 getZeroVector(ShiftVT, DAG, dl),
4431 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
4432 Intrinsic::arm_neon_vshifts :
4433 Intrinsic::arm_neon_vshiftu);
4434 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
4435 DAG.getConstant(vshiftInt, dl, MVT::i32),
4436 N->getOperand(0), NegatedCount);
4439 static SDValue Expand64BitShift(SDNode *N, SelectionDAG &DAG,
4440 const ARMSubtarget *ST) {
4441 EVT VT = N->getValueType(0);
4444 // We can get here for a node like i32 = ISD::SHL i32, i64
4448 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
4449 "Unknown shift to lower!");
4451 // We only lower SRA, SRL of 1 here, all others use generic lowering.
4452 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
4453 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
4456 // If we are in thumb mode, we don't have RRX.
4457 if (ST->isThumb1Only()) return SDValue();
4459 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
4460 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
4461 DAG.getConstant(0, dl, MVT::i32));
4462 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
4463 DAG.getConstant(1, dl, MVT::i32));
4465 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
4466 // captures the result into a carry flag.
4467 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
4468 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Glue), Hi);
4470 // The low part is an ARMISD::RRX operand, which shifts the carry in.
4471 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
4473 // Merge the pieces into a single i64 value.
4474 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
4477 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
4478 SDValue TmpOp0, TmpOp1;
4479 bool Invert = false;
4483 SDValue Op0 = Op.getOperand(0);
4484 SDValue Op1 = Op.getOperand(1);
4485 SDValue CC = Op.getOperand(2);
4486 EVT CmpVT = Op0.getValueType().changeVectorElementTypeToInteger();
4487 EVT VT = Op.getValueType();
4488 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
4491 if (Op1.getValueType().isFloatingPoint()) {
4492 switch (SetCCOpcode) {
4493 default: llvm_unreachable("Illegal FP comparison");
4495 case ISD::SETNE: Invert = true; // Fallthrough
4497 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
4499 case ISD::SETLT: Swap = true; // Fallthrough
4501 case ISD::SETGT: Opc = ARMISD::VCGT; break;
4503 case ISD::SETLE: Swap = true; // Fallthrough
4505 case ISD::SETGE: Opc = ARMISD::VCGE; break;
4506 case ISD::SETUGE: Swap = true; // Fallthrough
4507 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
4508 case ISD::SETUGT: Swap = true; // Fallthrough
4509 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
4510 case ISD::SETUEQ: Invert = true; // Fallthrough
4512 // Expand this to (OLT | OGT).
4516 Op0 = DAG.getNode(ARMISD::VCGT, dl, CmpVT, TmpOp1, TmpOp0);
4517 Op1 = DAG.getNode(ARMISD::VCGT, dl, CmpVT, TmpOp0, TmpOp1);
4519 case ISD::SETUO: Invert = true; // Fallthrough
4521 // Expand this to (OLT | OGE).
4525 Op0 = DAG.getNode(ARMISD::VCGT, dl, CmpVT, TmpOp1, TmpOp0);
4526 Op1 = DAG.getNode(ARMISD::VCGE, dl, CmpVT, TmpOp0, TmpOp1);
4530 // Integer comparisons.
4531 switch (SetCCOpcode) {
4532 default: llvm_unreachable("Illegal integer comparison");
4533 case ISD::SETNE: Invert = true;
4534 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
4535 case ISD::SETLT: Swap = true;
4536 case ISD::SETGT: Opc = ARMISD::VCGT; break;
4537 case ISD::SETLE: Swap = true;
4538 case ISD::SETGE: Opc = ARMISD::VCGE; break;
4539 case ISD::SETULT: Swap = true;
4540 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
4541 case ISD::SETULE: Swap = true;
4542 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
4545 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
4546 if (Opc == ARMISD::VCEQ) {
4549 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
4551 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
4554 // Ignore bitconvert.
4555 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BITCAST)
4556 AndOp = AndOp.getOperand(0);
4558 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
4560 Op0 = DAG.getNode(ISD::BITCAST, dl, CmpVT, AndOp.getOperand(0));
4561 Op1 = DAG.getNode(ISD::BITCAST, dl, CmpVT, AndOp.getOperand(1));
4568 std::swap(Op0, Op1);
4570 // If one of the operands is a constant vector zero, attempt to fold the
4571 // comparison to a specialized compare-against-zero form.
4573 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
4575 else if (ISD::isBuildVectorAllZeros(Op0.getNode())) {
4576 if (Opc == ARMISD::VCGE)
4577 Opc = ARMISD::VCLEZ;
4578 else if (Opc == ARMISD::VCGT)
4579 Opc = ARMISD::VCLTZ;
4584 if (SingleOp.getNode()) {
4587 Result = DAG.getNode(ARMISD::VCEQZ, dl, CmpVT, SingleOp); break;
4589 Result = DAG.getNode(ARMISD::VCGEZ, dl, CmpVT, SingleOp); break;
4591 Result = DAG.getNode(ARMISD::VCLEZ, dl, CmpVT, SingleOp); break;
4593 Result = DAG.getNode(ARMISD::VCGTZ, dl, CmpVT, SingleOp); break;
4595 Result = DAG.getNode(ARMISD::VCLTZ, dl, CmpVT, SingleOp); break;
4597 Result = DAG.getNode(Opc, dl, CmpVT, Op0, Op1);
4600 Result = DAG.getNode(Opc, dl, CmpVT, Op0, Op1);
4603 Result = DAG.getSExtOrTrunc(Result, dl, VT);
4606 Result = DAG.getNOT(dl, Result, VT);
4611 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
4612 /// valid vector constant for a NEON instruction with a "modified immediate"
4613 /// operand (e.g., VMOV). If so, return the encoded value.
4614 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
4615 unsigned SplatBitSize, SelectionDAG &DAG,
4616 SDLoc dl, EVT &VT, bool is128Bits,
4617 NEONModImmType type) {
4618 unsigned OpCmode, Imm;
4620 // SplatBitSize is set to the smallest size that splats the vector, so a
4621 // zero vector will always have SplatBitSize == 8. However, NEON modified
4622 // immediate instructions others than VMOV do not support the 8-bit encoding
4623 // of a zero vector, and the default encoding of zero is supposed to be the
4628 switch (SplatBitSize) {
4630 if (type != VMOVModImm)
4632 // Any 1-byte value is OK. Op=0, Cmode=1110.
4633 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
4636 VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
4640 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
4641 VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
4642 if ((SplatBits & ~0xff) == 0) {
4643 // Value = 0x00nn: Op=x, Cmode=100x.
4648 if ((SplatBits & ~0xff00) == 0) {
4649 // Value = 0xnn00: Op=x, Cmode=101x.
4651 Imm = SplatBits >> 8;
4657 // NEON's 32-bit VMOV supports splat values where:
4658 // * only one byte is nonzero, or
4659 // * the least significant byte is 0xff and the second byte is nonzero, or
4660 // * the least significant 2 bytes are 0xff and the third is nonzero.
4661 VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
4662 if ((SplatBits & ~0xff) == 0) {
4663 // Value = 0x000000nn: Op=x, Cmode=000x.
4668 if ((SplatBits & ~0xff00) == 0) {
4669 // Value = 0x0000nn00: Op=x, Cmode=001x.
4671 Imm = SplatBits >> 8;
4674 if ((SplatBits & ~0xff0000) == 0) {
4675 // Value = 0x00nn0000: Op=x, Cmode=010x.
4677 Imm = SplatBits >> 16;
4680 if ((SplatBits & ~0xff000000) == 0) {
4681 // Value = 0xnn000000: Op=x, Cmode=011x.
4683 Imm = SplatBits >> 24;
4687 // cmode == 0b1100 and cmode == 0b1101 are not supported for VORR or VBIC
4688 if (type == OtherModImm) return SDValue();
4690 if ((SplatBits & ~0xffff) == 0 &&
4691 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
4692 // Value = 0x0000nnff: Op=x, Cmode=1100.
4694 Imm = SplatBits >> 8;
4698 if ((SplatBits & ~0xffffff) == 0 &&
4699 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
4700 // Value = 0x00nnffff: Op=x, Cmode=1101.
4702 Imm = SplatBits >> 16;
4706 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
4707 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
4708 // VMOV.I32. A (very) minor optimization would be to replicate the value
4709 // and fall through here to test for a valid 64-bit splat. But, then the
4710 // caller would also need to check and handle the change in size.
4714 if (type != VMOVModImm)
4716 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
4717 uint64_t BitMask = 0xff;
4719 unsigned ImmMask = 1;
4721 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
4722 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
4725 } else if ((SplatBits & BitMask) != 0) {
4732 if (DAG.getDataLayout().isBigEndian())
4733 // swap higher and lower 32 bit word
4734 Imm = ((Imm & 0xf) << 4) | ((Imm & 0xf0) >> 4);
4736 // Op=1, Cmode=1110.
4738 VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
4743 llvm_unreachable("unexpected size for isNEONModifiedImm");
4746 unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm);
4747 return DAG.getTargetConstant(EncodedVal, dl, MVT::i32);
4750 SDValue ARMTargetLowering::LowerConstantFP(SDValue Op, SelectionDAG &DAG,
4751 const ARMSubtarget *ST) const {
4755 bool IsDouble = Op.getValueType() == MVT::f64;
4756 ConstantFPSDNode *CFP = cast<ConstantFPSDNode>(Op);
4758 // Use the default (constant pool) lowering for double constants when we have
4760 if (IsDouble && Subtarget->isFPOnlySP())
4763 // Try splatting with a VMOV.f32...
4764 APFloat FPVal = CFP->getValueAPF();
4765 int ImmVal = IsDouble ? ARM_AM::getFP64Imm(FPVal) : ARM_AM::getFP32Imm(FPVal);
4768 if (IsDouble || !ST->useNEONForSinglePrecisionFP()) {
4769 // We have code in place to select a valid ConstantFP already, no need to
4774 // It's a float and we are trying to use NEON operations where
4775 // possible. Lower it to a splat followed by an extract.
4777 SDValue NewVal = DAG.getTargetConstant(ImmVal, DL, MVT::i32);
4778 SDValue VecConstant = DAG.getNode(ARMISD::VMOVFPIMM, DL, MVT::v2f32,
4780 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecConstant,
4781 DAG.getConstant(0, DL, MVT::i32));
4784 // The rest of our options are NEON only, make sure that's allowed before
4786 if (!ST->hasNEON() || (!IsDouble && !ST->useNEONForSinglePrecisionFP()))
4790 uint64_t iVal = FPVal.bitcastToAPInt().getZExtValue();
4792 // It wouldn't really be worth bothering for doubles except for one very
4793 // important value, which does happen to match: 0.0. So make sure we don't do
4795 if (IsDouble && (iVal & 0xffffffff) != (iVal >> 32))
4798 // Try a VMOV.i32 (FIXME: i8, i16, or i64 could work too).
4799 SDValue NewVal = isNEONModifiedImm(iVal & 0xffffffffU, 0, 32, DAG, SDLoc(Op),
4800 VMovVT, false, VMOVModImm);
4801 if (NewVal != SDValue()) {
4803 SDValue VecConstant = DAG.getNode(ARMISD::VMOVIMM, DL, VMovVT,
4806 return DAG.getNode(ISD::BITCAST, DL, MVT::f64, VecConstant);
4808 // It's a float: cast and extract a vector element.
4809 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
4811 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
4812 DAG.getConstant(0, DL, MVT::i32));
4815 // Finally, try a VMVN.i32
4816 NewVal = isNEONModifiedImm(~iVal & 0xffffffffU, 0, 32, DAG, SDLoc(Op), VMovVT,
4818 if (NewVal != SDValue()) {
4820 SDValue VecConstant = DAG.getNode(ARMISD::VMVNIMM, DL, VMovVT, NewVal);
4823 return DAG.getNode(ISD::BITCAST, DL, MVT::f64, VecConstant);
4825 // It's a float: cast and extract a vector element.
4826 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
4828 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
4829 DAG.getConstant(0, DL, MVT::i32));
4835 // check if an VEXT instruction can handle the shuffle mask when the
4836 // vector sources of the shuffle are the same.
4837 static bool isSingletonVEXTMask(ArrayRef<int> M, EVT VT, unsigned &Imm) {
4838 unsigned NumElts = VT.getVectorNumElements();
4840 // Assume that the first shuffle index is not UNDEF. Fail if it is.
4846 // If this is a VEXT shuffle, the immediate value is the index of the first
4847 // element. The other shuffle indices must be the successive elements after
4849 unsigned ExpectedElt = Imm;
4850 for (unsigned i = 1; i < NumElts; ++i) {
4851 // Increment the expected index. If it wraps around, just follow it
4852 // back to index zero and keep going.
4854 if (ExpectedElt == NumElts)
4857 if (M[i] < 0) continue; // ignore UNDEF indices
4858 if (ExpectedElt != static_cast<unsigned>(M[i]))
4866 static bool isVEXTMask(ArrayRef<int> M, EVT VT,
4867 bool &ReverseVEXT, unsigned &Imm) {
4868 unsigned NumElts = VT.getVectorNumElements();
4869 ReverseVEXT = false;
4871 // Assume that the first shuffle index is not UNDEF. Fail if it is.
4877 // If this is a VEXT shuffle, the immediate value is the index of the first
4878 // element. The other shuffle indices must be the successive elements after
4880 unsigned ExpectedElt = Imm;
4881 for (unsigned i = 1; i < NumElts; ++i) {
4882 // Increment the expected index. If it wraps around, it may still be
4883 // a VEXT but the source vectors must be swapped.
4885 if (ExpectedElt == NumElts * 2) {
4890 if (M[i] < 0) continue; // ignore UNDEF indices
4891 if (ExpectedElt != static_cast<unsigned>(M[i]))
4895 // Adjust the index value if the source operands will be swapped.
4902 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
4903 /// instruction with the specified blocksize. (The order of the elements
4904 /// within each block of the vector is reversed.)
4905 static bool isVREVMask(ArrayRef<int> M, EVT VT, unsigned BlockSize) {
4906 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
4907 "Only possible block sizes for VREV are: 16, 32, 64");
4909 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4913 unsigned NumElts = VT.getVectorNumElements();
4914 unsigned BlockElts = M[0] + 1;
4915 // If the first shuffle index is UNDEF, be optimistic.
4917 BlockElts = BlockSize / EltSz;
4919 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
4922 for (unsigned i = 0; i < NumElts; ++i) {
4923 if (M[i] < 0) continue; // ignore UNDEF indices
4924 if ((unsigned) M[i] != (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
4931 static bool isVTBLMask(ArrayRef<int> M, EVT VT) {
4932 // We can handle <8 x i8> vector shuffles. If the index in the mask is out of
4933 // range, then 0 is placed into the resulting vector. So pretty much any mask
4934 // of 8 elements can work here.
4935 return VT == MVT::v8i8 && M.size() == 8;
4938 static bool isVTRNMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4939 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4943 unsigned NumElts = VT.getVectorNumElements();
4944 WhichResult = (M[0] == 0 ? 0 : 1);
4945 for (unsigned i = 0; i < NumElts; i += 2) {
4946 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
4947 (M[i+1] >= 0 && (unsigned) M[i+1] != i + NumElts + WhichResult))
4953 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
4954 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4955 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
4956 static bool isVTRN_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4957 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4961 unsigned NumElts = VT.getVectorNumElements();
4962 WhichResult = (M[0] == 0 ? 0 : 1);
4963 for (unsigned i = 0; i < NumElts; i += 2) {
4964 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
4965 (M[i+1] >= 0 && (unsigned) M[i+1] != i + WhichResult))
4971 static bool isVUZPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4972 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4976 unsigned NumElts = VT.getVectorNumElements();
4977 WhichResult = (M[0] == 0 ? 0 : 1);
4978 for (unsigned i = 0; i != NumElts; ++i) {
4979 if (M[i] < 0) continue; // ignore UNDEF indices
4980 if ((unsigned) M[i] != 2 * i + WhichResult)
4984 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4985 if (VT.is64BitVector() && EltSz == 32)
4991 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
4992 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4993 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
4994 static bool isVUZP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4995 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4999 unsigned Half = VT.getVectorNumElements() / 2;
5000 WhichResult = (M[0] == 0 ? 0 : 1);
5001 for (unsigned j = 0; j != 2; ++j) {
5002 unsigned Idx = WhichResult;
5003 for (unsigned i = 0; i != Half; ++i) {
5004 int MIdx = M[i + j * Half];
5005 if (MIdx >= 0 && (unsigned) MIdx != Idx)
5011 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
5012 if (VT.is64BitVector() && EltSz == 32)
5018 static bool isVZIPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
5019 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
5023 unsigned NumElts = VT.getVectorNumElements();
5024 WhichResult = (M[0] == 0 ? 0 : 1);
5025 unsigned Idx = WhichResult * NumElts / 2;
5026 for (unsigned i = 0; i != NumElts; i += 2) {
5027 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
5028 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx + NumElts))
5033 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
5034 if (VT.is64BitVector() && EltSz == 32)
5040 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
5041 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
5042 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
5043 static bool isVZIP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
5044 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
5048 unsigned NumElts = VT.getVectorNumElements();
5049 WhichResult = (M[0] == 0 ? 0 : 1);
5050 unsigned Idx = WhichResult * NumElts / 2;
5051 for (unsigned i = 0; i != NumElts; i += 2) {
5052 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
5053 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx))
5058 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
5059 if (VT.is64BitVector() && EltSz == 32)
5065 /// Check if \p ShuffleMask is a NEON two-result shuffle (VZIP, VUZP, VTRN),
5066 /// and return the corresponding ARMISD opcode if it is, or 0 if it isn't.
5067 static unsigned isNEONTwoResultShuffleMask(ArrayRef<int> ShuffleMask, EVT VT,
5068 unsigned &WhichResult,
5071 if (isVTRNMask(ShuffleMask, VT, WhichResult))
5072 return ARMISD::VTRN;
5073 if (isVUZPMask(ShuffleMask, VT, WhichResult))
5074 return ARMISD::VUZP;
5075 if (isVZIPMask(ShuffleMask, VT, WhichResult))
5076 return ARMISD::VZIP;
5079 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
5080 return ARMISD::VTRN;
5081 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
5082 return ARMISD::VUZP;
5083 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
5084 return ARMISD::VZIP;
5089 /// \return true if this is a reverse operation on an vector.
5090 static bool isReverseMask(ArrayRef<int> M, EVT VT) {
5091 unsigned NumElts = VT.getVectorNumElements();
5092 // Make sure the mask has the right size.
5093 if (NumElts != M.size())
5096 // Look for <15, ..., 3, -1, 1, 0>.
5097 for (unsigned i = 0; i != NumElts; ++i)
5098 if (M[i] >= 0 && M[i] != (int) (NumElts - 1 - i))
5104 // If N is an integer constant that can be moved into a register in one
5105 // instruction, return an SDValue of such a constant (will become a MOV
5106 // instruction). Otherwise return null.
5107 static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG,
5108 const ARMSubtarget *ST, SDLoc dl) {
5110 if (!isa<ConstantSDNode>(N))
5112 Val = cast<ConstantSDNode>(N)->getZExtValue();
5114 if (ST->isThumb1Only()) {
5115 if (Val <= 255 || ~Val <= 255)
5116 return DAG.getConstant(Val, dl, MVT::i32);
5118 if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1)
5119 return DAG.getConstant(Val, dl, MVT::i32);
5124 // If this is a case we can't handle, return null and let the default
5125 // expansion code take care of it.
5126 SDValue ARMTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
5127 const ARMSubtarget *ST) const {
5128 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
5130 EVT VT = Op.getValueType();
5132 APInt SplatBits, SplatUndef;
5133 unsigned SplatBitSize;
5135 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
5136 if (SplatBitSize <= 64) {
5137 // Check if an immediate VMOV works.
5139 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
5140 SplatUndef.getZExtValue(), SplatBitSize,
5141 DAG, dl, VmovVT, VT.is128BitVector(),
5143 if (Val.getNode()) {
5144 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
5145 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
5148 // Try an immediate VMVN.
5149 uint64_t NegatedImm = (~SplatBits).getZExtValue();
5150 Val = isNEONModifiedImm(NegatedImm,
5151 SplatUndef.getZExtValue(), SplatBitSize,
5152 DAG, dl, VmovVT, VT.is128BitVector(),
5154 if (Val.getNode()) {
5155 SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
5156 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
5159 // Use vmov.f32 to materialize other v2f32 and v4f32 splats.
5160 if ((VT == MVT::v2f32 || VT == MVT::v4f32) && SplatBitSize == 32) {
5161 int ImmVal = ARM_AM::getFP32Imm(SplatBits);
5163 SDValue Val = DAG.getTargetConstant(ImmVal, dl, MVT::i32);
5164 return DAG.getNode(ARMISD::VMOVFPIMM, dl, VT, Val);
5170 // Scan through the operands to see if only one value is used.
5172 // As an optimisation, even if more than one value is used it may be more
5173 // profitable to splat with one value then change some lanes.
5175 // Heuristically we decide to do this if the vector has a "dominant" value,
5176 // defined as splatted to more than half of the lanes.
5177 unsigned NumElts = VT.getVectorNumElements();
5178 bool isOnlyLowElement = true;
5179 bool usesOnlyOneValue = true;
5180 bool hasDominantValue = false;
5181 bool isConstant = true;
5183 // Map of the number of times a particular SDValue appears in the
5185 DenseMap<SDValue, unsigned> ValueCounts;
5187 for (unsigned i = 0; i < NumElts; ++i) {
5188 SDValue V = Op.getOperand(i);
5189 if (V.getOpcode() == ISD::UNDEF)
5192 isOnlyLowElement = false;
5193 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
5196 ValueCounts.insert(std::make_pair(V, 0));
5197 unsigned &Count = ValueCounts[V];
5199 // Is this value dominant? (takes up more than half of the lanes)
5200 if (++Count > (NumElts / 2)) {
5201 hasDominantValue = true;
5205 if (ValueCounts.size() != 1)
5206 usesOnlyOneValue = false;
5207 if (!Value.getNode() && ValueCounts.size() > 0)
5208 Value = ValueCounts.begin()->first;
5210 if (ValueCounts.size() == 0)
5211 return DAG.getUNDEF(VT);
5213 // Loads are better lowered with insert_vector_elt/ARMISD::BUILD_VECTOR.
5214 // Keep going if we are hitting this case.
5215 if (isOnlyLowElement && !ISD::isNormalLoad(Value.getNode()))
5216 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
5218 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5220 // Use VDUP for non-constant splats. For f32 constant splats, reduce to
5221 // i32 and try again.
5222 if (hasDominantValue && EltSize <= 32) {
5226 // If we are VDUPing a value that comes directly from a vector, that will
5227 // cause an unnecessary move to and from a GPR, where instead we could
5228 // just use VDUPLANE. We can only do this if the lane being extracted
5229 // is at a constant index, as the VDUP from lane instructions only have
5230 // constant-index forms.
5231 if (Value->getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5232 isa<ConstantSDNode>(Value->getOperand(1))) {
5233 // We need to create a new undef vector to use for the VDUPLANE if the
5234 // size of the vector from which we get the value is different than the
5235 // size of the vector that we need to create. We will insert the element
5236 // such that the register coalescer will remove unnecessary copies.
5237 if (VT != Value->getOperand(0).getValueType()) {
5238 ConstantSDNode *constIndex;
5239 constIndex = dyn_cast<ConstantSDNode>(Value->getOperand(1));
5240 assert(constIndex && "The index is not a constant!");
5241 unsigned index = constIndex->getAPIntValue().getLimitedValue() %
5242 VT.getVectorNumElements();
5243 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
5244 DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DAG.getUNDEF(VT),
5245 Value, DAG.getConstant(index, dl, MVT::i32)),
5246 DAG.getConstant(index, dl, MVT::i32));
5248 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
5249 Value->getOperand(0), Value->getOperand(1));
5251 N = DAG.getNode(ARMISD::VDUP, dl, VT, Value);
5253 if (!usesOnlyOneValue) {
5254 // The dominant value was splatted as 'N', but we now have to insert
5255 // all differing elements.
5256 for (unsigned I = 0; I < NumElts; ++I) {
5257 if (Op.getOperand(I) == Value)
5259 SmallVector<SDValue, 3> Ops;
5261 Ops.push_back(Op.getOperand(I));
5262 Ops.push_back(DAG.getConstant(I, dl, MVT::i32));
5263 N = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Ops);
5268 if (VT.getVectorElementType().isFloatingPoint()) {
5269 SmallVector<SDValue, 8> Ops;
5270 for (unsigned i = 0; i < NumElts; ++i)
5271 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, MVT::i32,
5273 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
5274 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, Ops);
5275 Val = LowerBUILD_VECTOR(Val, DAG, ST);
5277 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
5279 if (usesOnlyOneValue) {
5280 SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
5281 if (isConstant && Val.getNode())
5282 return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
5286 // If all elements are constants and the case above didn't get hit, fall back
5287 // to the default expansion, which will generate a load from the constant
5292 // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
5294 SDValue shuffle = ReconstructShuffle(Op, DAG);
5295 if (shuffle != SDValue())
5299 // Vectors with 32- or 64-bit elements can be built by directly assigning
5300 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
5301 // will be legalized.
5302 if (EltSize >= 32) {
5303 // Do the expansion with floating-point types, since that is what the VFP
5304 // registers are defined to use, and since i64 is not legal.
5305 EVT EltVT = EVT::getFloatingPointVT(EltSize);
5306 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
5307 SmallVector<SDValue, 8> Ops;
5308 for (unsigned i = 0; i < NumElts; ++i)
5309 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, EltVT, Op.getOperand(i)));
5310 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, Ops);
5311 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
5314 // If all else fails, just use a sequence of INSERT_VECTOR_ELT when we
5315 // know the default expansion would otherwise fall back on something even
5316 // worse. For a vector with one or two non-undef values, that's
5317 // scalar_to_vector for the elements followed by a shuffle (provided the
5318 // shuffle is valid for the target) and materialization element by element
5319 // on the stack followed by a load for everything else.
5320 if (!isConstant && !usesOnlyOneValue) {
5321 SDValue Vec = DAG.getUNDEF(VT);
5322 for (unsigned i = 0 ; i < NumElts; ++i) {
5323 SDValue V = Op.getOperand(i);
5324 if (V.getOpcode() == ISD::UNDEF)
5326 SDValue LaneIdx = DAG.getConstant(i, dl, MVT::i32);
5327 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Vec, V, LaneIdx);
5335 // Gather data to see if the operation can be modelled as a
5336 // shuffle in combination with VEXTs.
5337 SDValue ARMTargetLowering::ReconstructShuffle(SDValue Op,
5338 SelectionDAG &DAG) const {
5340 EVT VT = Op.getValueType();
5341 unsigned NumElts = VT.getVectorNumElements();
5343 SmallVector<SDValue, 2> SourceVecs;
5344 SmallVector<unsigned, 2> MinElts;
5345 SmallVector<unsigned, 2> MaxElts;
5347 for (unsigned i = 0; i < NumElts; ++i) {
5348 SDValue V = Op.getOperand(i);
5349 if (V.getOpcode() == ISD::UNDEF)
5351 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
5352 // A shuffle can only come from building a vector from various
5353 // elements of other vectors.
5355 } else if (V.getOperand(0).getValueType().getVectorElementType() !=
5356 VT.getVectorElementType()) {
5357 // This code doesn't know how to handle shuffles where the vector
5358 // element types do not match (this happens because type legalization
5359 // promotes the return type of EXTRACT_VECTOR_ELT).
5360 // FIXME: It might be appropriate to extend this code to handle
5361 // mismatched types.
5365 // Record this extraction against the appropriate vector if possible...
5366 SDValue SourceVec = V.getOperand(0);
5367 // If the element number isn't a constant, we can't effectively
5368 // analyze what's going on.
5369 if (!isa<ConstantSDNode>(V.getOperand(1)))
5371 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
5372 bool FoundSource = false;
5373 for (unsigned j = 0; j < SourceVecs.size(); ++j) {
5374 if (SourceVecs[j] == SourceVec) {
5375 if (MinElts[j] > EltNo)
5377 if (MaxElts[j] < EltNo)
5384 // Or record a new source if not...
5386 SourceVecs.push_back(SourceVec);
5387 MinElts.push_back(EltNo);
5388 MaxElts.push_back(EltNo);
5392 // Currently only do something sane when at most two source vectors
5394 if (SourceVecs.size() > 2)
5397 SDValue ShuffleSrcs[2] = {DAG.getUNDEF(VT), DAG.getUNDEF(VT) };
5398 int VEXTOffsets[2] = {0, 0};
5400 // This loop extracts the usage patterns of the source vectors
5401 // and prepares appropriate SDValues for a shuffle if possible.
5402 for (unsigned i = 0; i < SourceVecs.size(); ++i) {
5403 if (SourceVecs[i].getValueType() == VT) {
5404 // No VEXT necessary
5405 ShuffleSrcs[i] = SourceVecs[i];
5408 } else if (SourceVecs[i].getValueType().getVectorNumElements() < NumElts) {
5409 // It probably isn't worth padding out a smaller vector just to
5410 // break it down again in a shuffle.
5414 // Since only 64-bit and 128-bit vectors are legal on ARM and
5415 // we've eliminated the other cases...
5416 assert(SourceVecs[i].getValueType().getVectorNumElements() == 2*NumElts &&
5417 "unexpected vector sizes in ReconstructShuffle");
5419 if (MaxElts[i] - MinElts[i] >= NumElts) {
5420 // Span too large for a VEXT to cope
5424 if (MinElts[i] >= NumElts) {
5425 // The extraction can just take the second half
5426 VEXTOffsets[i] = NumElts;
5427 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5429 DAG.getIntPtrConstant(NumElts, dl));
5430 } else if (MaxElts[i] < NumElts) {
5431 // The extraction can just take the first half
5433 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5435 DAG.getIntPtrConstant(0, dl));
5437 // An actual VEXT is needed
5438 VEXTOffsets[i] = MinElts[i];
5439 SDValue VEXTSrc1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5441 DAG.getIntPtrConstant(0, dl));
5442 SDValue VEXTSrc2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5444 DAG.getIntPtrConstant(NumElts, dl));
5445 ShuffleSrcs[i] = DAG.getNode(ARMISD::VEXT, dl, VT, VEXTSrc1, VEXTSrc2,
5446 DAG.getConstant(VEXTOffsets[i], dl,
5451 SmallVector<int, 8> Mask;
5453 for (unsigned i = 0; i < NumElts; ++i) {
5454 SDValue Entry = Op.getOperand(i);
5455 if (Entry.getOpcode() == ISD::UNDEF) {
5460 SDValue ExtractVec = Entry.getOperand(0);
5461 int ExtractElt = cast<ConstantSDNode>(Op.getOperand(i)
5462 .getOperand(1))->getSExtValue();
5463 if (ExtractVec == SourceVecs[0]) {
5464 Mask.push_back(ExtractElt - VEXTOffsets[0]);
5466 Mask.push_back(ExtractElt + NumElts - VEXTOffsets[1]);
5470 // Final check before we try to produce nonsense...
5471 if (isShuffleMaskLegal(Mask, VT))
5472 return DAG.getVectorShuffle(VT, dl, ShuffleSrcs[0], ShuffleSrcs[1],
5478 /// isShuffleMaskLegal - Targets can use this to indicate that they only
5479 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
5480 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
5481 /// are assumed to be legal.
5483 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
5485 if (VT.getVectorNumElements() == 4 &&
5486 (VT.is128BitVector() || VT.is64BitVector())) {
5487 unsigned PFIndexes[4];
5488 for (unsigned i = 0; i != 4; ++i) {
5492 PFIndexes[i] = M[i];
5495 // Compute the index in the perfect shuffle table.
5496 unsigned PFTableIndex =
5497 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
5498 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
5499 unsigned Cost = (PFEntry >> 30);
5505 bool ReverseVEXT, isV_UNDEF;
5506 unsigned Imm, WhichResult;
5508 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5509 return (EltSize >= 32 ||
5510 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
5511 isVREVMask(M, VT, 64) ||
5512 isVREVMask(M, VT, 32) ||
5513 isVREVMask(M, VT, 16) ||
5514 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
5515 isVTBLMask(M, VT) ||
5516 isNEONTwoResultShuffleMask(M, VT, WhichResult, isV_UNDEF) ||
5517 ((VT == MVT::v8i16 || VT == MVT::v16i8) && isReverseMask(M, VT)));
5520 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
5521 /// the specified operations to build the shuffle.
5522 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
5523 SDValue RHS, SelectionDAG &DAG,
5525 unsigned OpNum = (PFEntry >> 26) & 0x0F;
5526 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
5527 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
5530 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
5539 OP_VUZPL, // VUZP, left result
5540 OP_VUZPR, // VUZP, right result
5541 OP_VZIPL, // VZIP, left result
5542 OP_VZIPR, // VZIP, right result
5543 OP_VTRNL, // VTRN, left result
5544 OP_VTRNR // VTRN, right result
5547 if (OpNum == OP_COPY) {
5548 if (LHSID == (1*9+2)*9+3) return LHS;
5549 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
5553 SDValue OpLHS, OpRHS;
5554 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
5555 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
5556 EVT VT = OpLHS.getValueType();
5559 default: llvm_unreachable("Unknown shuffle opcode!");
5561 // VREV divides the vector in half and swaps within the half.
5562 if (VT.getVectorElementType() == MVT::i32 ||
5563 VT.getVectorElementType() == MVT::f32)
5564 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
5565 // vrev <4 x i16> -> VREV32
5566 if (VT.getVectorElementType() == MVT::i16)
5567 return DAG.getNode(ARMISD::VREV32, dl, VT, OpLHS);
5568 // vrev <4 x i8> -> VREV16
5569 assert(VT.getVectorElementType() == MVT::i8);
5570 return DAG.getNode(ARMISD::VREV16, dl, VT, OpLHS);
5575 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
5576 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, dl, MVT::i32));
5580 return DAG.getNode(ARMISD::VEXT, dl, VT,
5582 DAG.getConstant(OpNum - OP_VEXT1 + 1, dl, MVT::i32));
5585 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
5586 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
5589 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
5590 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
5593 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
5594 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
5598 static SDValue LowerVECTOR_SHUFFLEv8i8(SDValue Op,
5599 ArrayRef<int> ShuffleMask,
5600 SelectionDAG &DAG) {
5601 // Check to see if we can use the VTBL instruction.
5602 SDValue V1 = Op.getOperand(0);
5603 SDValue V2 = Op.getOperand(1);
5606 SmallVector<SDValue, 8> VTBLMask;
5607 for (ArrayRef<int>::iterator
5608 I = ShuffleMask.begin(), E = ShuffleMask.end(); I != E; ++I)
5609 VTBLMask.push_back(DAG.getConstant(*I, DL, MVT::i32));
5611 if (V2.getNode()->getOpcode() == ISD::UNDEF)
5612 return DAG.getNode(ARMISD::VTBL1, DL, MVT::v8i8, V1,
5613 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8, VTBLMask));
5615 return DAG.getNode(ARMISD::VTBL2, DL, MVT::v8i8, V1, V2,
5616 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8, VTBLMask));
5619 static SDValue LowerReverse_VECTOR_SHUFFLEv16i8_v8i16(SDValue Op,
5620 SelectionDAG &DAG) {
5622 SDValue OpLHS = Op.getOperand(0);
5623 EVT VT = OpLHS.getValueType();
5625 assert((VT == MVT::v8i16 || VT == MVT::v16i8) &&
5626 "Expect an v8i16/v16i8 type");
5627 OpLHS = DAG.getNode(ARMISD::VREV64, DL, VT, OpLHS);
5628 // For a v16i8 type: After the VREV, we have got <8, ...15, 8, ..., 0>. Now,
5629 // extract the first 8 bytes into the top double word and the last 8 bytes
5630 // into the bottom double word. The v8i16 case is similar.
5631 unsigned ExtractNum = (VT == MVT::v16i8) ? 8 : 4;
5632 return DAG.getNode(ARMISD::VEXT, DL, VT, OpLHS, OpLHS,
5633 DAG.getConstant(ExtractNum, DL, MVT::i32));
5636 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
5637 SDValue V1 = Op.getOperand(0);
5638 SDValue V2 = Op.getOperand(1);
5640 EVT VT = Op.getValueType();
5641 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
5643 // Convert shuffles that are directly supported on NEON to target-specific
5644 // DAG nodes, instead of keeping them as shuffles and matching them again
5645 // during code selection. This is more efficient and avoids the possibility
5646 // of inconsistencies between legalization and selection.
5647 // FIXME: floating-point vectors should be canonicalized to integer vectors
5648 // of the same time so that they get CSEd properly.
5649 ArrayRef<int> ShuffleMask = SVN->getMask();
5651 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5652 if (EltSize <= 32) {
5653 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
5654 int Lane = SVN->getSplatIndex();
5655 // If this is undef splat, generate it via "just" vdup, if possible.
5656 if (Lane == -1) Lane = 0;
5658 // Test if V1 is a SCALAR_TO_VECTOR.
5659 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
5660 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
5662 // Test if V1 is a BUILD_VECTOR which is equivalent to a SCALAR_TO_VECTOR
5663 // (and probably will turn into a SCALAR_TO_VECTOR once legalization
5665 if (Lane == 0 && V1.getOpcode() == ISD::BUILD_VECTOR &&
5666 !isa<ConstantSDNode>(V1.getOperand(0))) {
5667 bool IsScalarToVector = true;
5668 for (unsigned i = 1, e = V1.getNumOperands(); i != e; ++i)
5669 if (V1.getOperand(i).getOpcode() != ISD::UNDEF) {
5670 IsScalarToVector = false;
5673 if (IsScalarToVector)
5674 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
5676 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
5677 DAG.getConstant(Lane, dl, MVT::i32));
5682 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
5685 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
5686 DAG.getConstant(Imm, dl, MVT::i32));
5689 if (isVREVMask(ShuffleMask, VT, 64))
5690 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
5691 if (isVREVMask(ShuffleMask, VT, 32))
5692 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
5693 if (isVREVMask(ShuffleMask, VT, 16))
5694 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
5696 if (V2->getOpcode() == ISD::UNDEF &&
5697 isSingletonVEXTMask(ShuffleMask, VT, Imm)) {
5698 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V1,
5699 DAG.getConstant(Imm, dl, MVT::i32));
5702 // Check for Neon shuffles that modify both input vectors in place.
5703 // If both results are used, i.e., if there are two shuffles with the same
5704 // source operands and with masks corresponding to both results of one of
5705 // these operations, DAG memoization will ensure that a single node is
5706 // used for both shuffles.
5707 unsigned WhichResult;
5709 if (unsigned ShuffleOpc = isNEONTwoResultShuffleMask(
5710 ShuffleMask, VT, WhichResult, isV_UNDEF)) {
5713 return DAG.getNode(ShuffleOpc, dl, DAG.getVTList(VT, VT), V1, V2)
5714 .getValue(WhichResult);
5717 // Also check for these shuffles through CONCAT_VECTORS: we canonicalize
5718 // shuffles that produce a result larger than their operands with:
5719 // shuffle(concat(v1, undef), concat(v2, undef))
5721 // shuffle(concat(v1, v2), undef)
5722 // because we can access quad vectors (see PerformVECTOR_SHUFFLECombine).
5724 // This is useful in the general case, but there are special cases where
5725 // native shuffles produce larger results: the two-result ops.
5727 // Look through the concat when lowering them:
5728 // shuffle(concat(v1, v2), undef)
5730 // concat(VZIP(v1, v2):0, :1)
5732 if (V1->getOpcode() == ISD::CONCAT_VECTORS &&
5733 V2->getOpcode() == ISD::UNDEF) {
5734 SDValue SubV1 = V1->getOperand(0);
5735 SDValue SubV2 = V1->getOperand(1);
5736 EVT SubVT = SubV1.getValueType();
5738 // We expect these to have been canonicalized to -1.
5739 assert(std::all_of(ShuffleMask.begin(), ShuffleMask.end(), [&](int i) {
5740 return i < (int)VT.getVectorNumElements();
5741 }) && "Unexpected shuffle index into UNDEF operand!");
5743 if (unsigned ShuffleOpc = isNEONTwoResultShuffleMask(
5744 ShuffleMask, SubVT, WhichResult, isV_UNDEF)) {
5747 assert((WhichResult == 0) &&
5748 "In-place shuffle of concat can only have one result!");
5749 SDValue Res = DAG.getNode(ShuffleOpc, dl, DAG.getVTList(SubVT, SubVT),
5751 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, Res.getValue(0),
5757 // If the shuffle is not directly supported and it has 4 elements, use
5758 // the PerfectShuffle-generated table to synthesize it from other shuffles.
5759 unsigned NumElts = VT.getVectorNumElements();
5761 unsigned PFIndexes[4];
5762 for (unsigned i = 0; i != 4; ++i) {
5763 if (ShuffleMask[i] < 0)
5766 PFIndexes[i] = ShuffleMask[i];
5769 // Compute the index in the perfect shuffle table.
5770 unsigned PFTableIndex =
5771 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
5772 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
5773 unsigned Cost = (PFEntry >> 30);
5776 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
5779 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
5780 if (EltSize >= 32) {
5781 // Do the expansion with floating-point types, since that is what the VFP
5782 // registers are defined to use, and since i64 is not legal.
5783 EVT EltVT = EVT::getFloatingPointVT(EltSize);
5784 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
5785 V1 = DAG.getNode(ISD::BITCAST, dl, VecVT, V1);
5786 V2 = DAG.getNode(ISD::BITCAST, dl, VecVT, V2);
5787 SmallVector<SDValue, 8> Ops;
5788 for (unsigned i = 0; i < NumElts; ++i) {
5789 if (ShuffleMask[i] < 0)
5790 Ops.push_back(DAG.getUNDEF(EltVT));
5792 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
5793 ShuffleMask[i] < (int)NumElts ? V1 : V2,
5794 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
5797 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, Ops);
5798 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
5801 if ((VT == MVT::v8i16 || VT == MVT::v16i8) && isReverseMask(ShuffleMask, VT))
5802 return LowerReverse_VECTOR_SHUFFLEv16i8_v8i16(Op, DAG);
5804 if (VT == MVT::v8i8) {
5805 SDValue NewOp = LowerVECTOR_SHUFFLEv8i8(Op, ShuffleMask, DAG);
5806 if (NewOp.getNode())
5813 static SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
5814 // INSERT_VECTOR_ELT is legal only for immediate indexes.
5815 SDValue Lane = Op.getOperand(2);
5816 if (!isa<ConstantSDNode>(Lane))
5822 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
5823 // EXTRACT_VECTOR_ELT is legal only for immediate indexes.
5824 SDValue Lane = Op.getOperand(1);
5825 if (!isa<ConstantSDNode>(Lane))
5828 SDValue Vec = Op.getOperand(0);
5829 if (Op.getValueType() == MVT::i32 &&
5830 Vec.getValueType().getVectorElementType().getSizeInBits() < 32) {
5832 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
5838 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
5839 // The only time a CONCAT_VECTORS operation can have legal types is when
5840 // two 64-bit vectors are concatenated to a 128-bit vector.
5841 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
5842 "unexpected CONCAT_VECTORS");
5844 SDValue Val = DAG.getUNDEF(MVT::v2f64);
5845 SDValue Op0 = Op.getOperand(0);
5846 SDValue Op1 = Op.getOperand(1);
5847 if (Op0.getOpcode() != ISD::UNDEF)
5848 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
5849 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op0),
5850 DAG.getIntPtrConstant(0, dl));
5851 if (Op1.getOpcode() != ISD::UNDEF)
5852 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
5853 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op1),
5854 DAG.getIntPtrConstant(1, dl));
5855 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Val);
5858 /// isExtendedBUILD_VECTOR - Check if N is a constant BUILD_VECTOR where each
5859 /// element has been zero/sign-extended, depending on the isSigned parameter,
5860 /// from an integer type half its size.
5861 static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG,
5863 // A v2i64 BUILD_VECTOR will have been legalized to a BITCAST from v4i32.
5864 EVT VT = N->getValueType(0);
5865 if (VT == MVT::v2i64 && N->getOpcode() == ISD::BITCAST) {
5866 SDNode *BVN = N->getOperand(0).getNode();
5867 if (BVN->getValueType(0) != MVT::v4i32 ||
5868 BVN->getOpcode() != ISD::BUILD_VECTOR)
5870 unsigned LoElt = DAG.getDataLayout().isBigEndian() ? 1 : 0;
5871 unsigned HiElt = 1 - LoElt;
5872 ConstantSDNode *Lo0 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt));
5873 ConstantSDNode *Hi0 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt));
5874 ConstantSDNode *Lo1 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt+2));
5875 ConstantSDNode *Hi1 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt+2));
5876 if (!Lo0 || !Hi0 || !Lo1 || !Hi1)
5879 if (Hi0->getSExtValue() == Lo0->getSExtValue() >> 32 &&
5880 Hi1->getSExtValue() == Lo1->getSExtValue() >> 32)
5883 if (Hi0->isNullValue() && Hi1->isNullValue())
5889 if (N->getOpcode() != ISD::BUILD_VECTOR)
5892 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
5893 SDNode *Elt = N->getOperand(i).getNode();
5894 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
5895 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5896 unsigned HalfSize = EltSize / 2;
5898 if (!isIntN(HalfSize, C->getSExtValue()))
5901 if (!isUIntN(HalfSize, C->getZExtValue()))
5912 /// isSignExtended - Check if a node is a vector value that is sign-extended
5913 /// or a constant BUILD_VECTOR with sign-extended elements.
5914 static bool isSignExtended(SDNode *N, SelectionDAG &DAG) {
5915 if (N->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N))
5917 if (isExtendedBUILD_VECTOR(N, DAG, true))
5922 /// isZeroExtended - Check if a node is a vector value that is zero-extended
5923 /// or a constant BUILD_VECTOR with zero-extended elements.
5924 static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) {
5925 if (N->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N))
5927 if (isExtendedBUILD_VECTOR(N, DAG, false))
5932 static EVT getExtensionTo64Bits(const EVT &OrigVT) {
5933 if (OrigVT.getSizeInBits() >= 64)
5936 assert(OrigVT.isSimple() && "Expecting a simple value type");
5938 MVT::SimpleValueType OrigSimpleTy = OrigVT.getSimpleVT().SimpleTy;
5939 switch (OrigSimpleTy) {
5940 default: llvm_unreachable("Unexpected Vector Type");
5949 /// AddRequiredExtensionForVMULL - Add a sign/zero extension to extend the total
5950 /// value size to 64 bits. We need a 64-bit D register as an operand to VMULL.
5951 /// We insert the required extension here to get the vector to fill a D register.
5952 static SDValue AddRequiredExtensionForVMULL(SDValue N, SelectionDAG &DAG,
5955 unsigned ExtOpcode) {
5956 // The vector originally had a size of OrigTy. It was then extended to ExtTy.
5957 // We expect the ExtTy to be 128-bits total. If the OrigTy is less than
5958 // 64-bits we need to insert a new extension so that it will be 64-bits.
5959 assert(ExtTy.is128BitVector() && "Unexpected extension size");
5960 if (OrigTy.getSizeInBits() >= 64)
5963 // Must extend size to at least 64 bits to be used as an operand for VMULL.
5964 EVT NewVT = getExtensionTo64Bits(OrigTy);
5966 return DAG.getNode(ExtOpcode, SDLoc(N), NewVT, N);
5969 /// SkipLoadExtensionForVMULL - return a load of the original vector size that
5970 /// does not do any sign/zero extension. If the original vector is less
5971 /// than 64 bits, an appropriate extension will be added after the load to
5972 /// reach a total size of 64 bits. We have to add the extension separately
5973 /// because ARM does not have a sign/zero extending load for vectors.
5974 static SDValue SkipLoadExtensionForVMULL(LoadSDNode *LD, SelectionDAG& DAG) {
5975 EVT ExtendedTy = getExtensionTo64Bits(LD->getMemoryVT());
5977 // The load already has the right type.
5978 if (ExtendedTy == LD->getMemoryVT())
5979 return DAG.getLoad(LD->getMemoryVT(), SDLoc(LD), LD->getChain(),
5980 LD->getBasePtr(), LD->getPointerInfo(), LD->isVolatile(),
5981 LD->isNonTemporal(), LD->isInvariant(),
5982 LD->getAlignment());
5984 // We need to create a zextload/sextload. We cannot just create a load
5985 // followed by a zext/zext node because LowerMUL is also run during normal
5986 // operation legalization where we can't create illegal types.
5987 return DAG.getExtLoad(LD->getExtensionType(), SDLoc(LD), ExtendedTy,
5988 LD->getChain(), LD->getBasePtr(), LD->getPointerInfo(),
5989 LD->getMemoryVT(), LD->isVolatile(), LD->isInvariant(),
5990 LD->isNonTemporal(), LD->getAlignment());
5993 /// SkipExtensionForVMULL - For a node that is a SIGN_EXTEND, ZERO_EXTEND,
5994 /// extending load, or BUILD_VECTOR with extended elements, return the
5995 /// unextended value. The unextended vector should be 64 bits so that it can
5996 /// be used as an operand to a VMULL instruction. If the original vector size
5997 /// before extension is less than 64 bits we add a an extension to resize
5998 /// the vector to 64 bits.
5999 static SDValue SkipExtensionForVMULL(SDNode *N, SelectionDAG &DAG) {
6000 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
6001 return AddRequiredExtensionForVMULL(N->getOperand(0), DAG,
6002 N->getOperand(0)->getValueType(0),
6006 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
6007 return SkipLoadExtensionForVMULL(LD, DAG);
6009 // Otherwise, the value must be a BUILD_VECTOR. For v2i64, it will
6010 // have been legalized as a BITCAST from v4i32.
6011 if (N->getOpcode() == ISD::BITCAST) {
6012 SDNode *BVN = N->getOperand(0).getNode();
6013 assert(BVN->getOpcode() == ISD::BUILD_VECTOR &&
6014 BVN->getValueType(0) == MVT::v4i32 && "expected v4i32 BUILD_VECTOR");
6015 unsigned LowElt = DAG.getDataLayout().isBigEndian() ? 1 : 0;
6016 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), MVT::v2i32,
6017 BVN->getOperand(LowElt), BVN->getOperand(LowElt+2));
6019 // Construct a new BUILD_VECTOR with elements truncated to half the size.
6020 assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR");
6021 EVT VT = N->getValueType(0);
6022 unsigned EltSize = VT.getVectorElementType().getSizeInBits() / 2;
6023 unsigned NumElts = VT.getVectorNumElements();
6024 MVT TruncVT = MVT::getIntegerVT(EltSize);
6025 SmallVector<SDValue, 8> Ops;
6027 for (unsigned i = 0; i != NumElts; ++i) {
6028 ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i));
6029 const APInt &CInt = C->getAPIntValue();
6030 // Element types smaller than 32 bits are not legal, so use i32 elements.
6031 // The values are implicitly truncated so sext vs. zext doesn't matter.
6032 Ops.push_back(DAG.getConstant(CInt.zextOrTrunc(32), dl, MVT::i32));
6034 return DAG.getNode(ISD::BUILD_VECTOR, dl,
6035 MVT::getVectorVT(TruncVT, NumElts), Ops);
6038 static bool isAddSubSExt(SDNode *N, SelectionDAG &DAG) {
6039 unsigned Opcode = N->getOpcode();
6040 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
6041 SDNode *N0 = N->getOperand(0).getNode();
6042 SDNode *N1 = N->getOperand(1).getNode();
6043 return N0->hasOneUse() && N1->hasOneUse() &&
6044 isSignExtended(N0, DAG) && isSignExtended(N1, DAG);
6049 static bool isAddSubZExt(SDNode *N, SelectionDAG &DAG) {
6050 unsigned Opcode = N->getOpcode();
6051 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
6052 SDNode *N0 = N->getOperand(0).getNode();
6053 SDNode *N1 = N->getOperand(1).getNode();
6054 return N0->hasOneUse() && N1->hasOneUse() &&
6055 isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG);
6060 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
6061 // Multiplications are only custom-lowered for 128-bit vectors so that
6062 // VMULL can be detected. Otherwise v2i64 multiplications are not legal.
6063 EVT VT = Op.getValueType();
6064 assert(VT.is128BitVector() && VT.isInteger() &&
6065 "unexpected type for custom-lowering ISD::MUL");
6066 SDNode *N0 = Op.getOperand(0).getNode();
6067 SDNode *N1 = Op.getOperand(1).getNode();
6068 unsigned NewOpc = 0;
6070 bool isN0SExt = isSignExtended(N0, DAG);
6071 bool isN1SExt = isSignExtended(N1, DAG);
6072 if (isN0SExt && isN1SExt)
6073 NewOpc = ARMISD::VMULLs;
6075 bool isN0ZExt = isZeroExtended(N0, DAG);
6076 bool isN1ZExt = isZeroExtended(N1, DAG);
6077 if (isN0ZExt && isN1ZExt)
6078 NewOpc = ARMISD::VMULLu;
6079 else if (isN1SExt || isN1ZExt) {
6080 // Look for (s/zext A + s/zext B) * (s/zext C). We want to turn these
6081 // into (s/zext A * s/zext C) + (s/zext B * s/zext C)
6082 if (isN1SExt && isAddSubSExt(N0, DAG)) {
6083 NewOpc = ARMISD::VMULLs;
6085 } else if (isN1ZExt && isAddSubZExt(N0, DAG)) {
6086 NewOpc = ARMISD::VMULLu;
6088 } else if (isN0ZExt && isAddSubZExt(N1, DAG)) {
6090 NewOpc = ARMISD::VMULLu;
6096 if (VT == MVT::v2i64)
6097 // Fall through to expand this. It is not legal.
6100 // Other vector multiplications are legal.
6105 // Legalize to a VMULL instruction.
6108 SDValue Op1 = SkipExtensionForVMULL(N1, DAG);
6110 Op0 = SkipExtensionForVMULL(N0, DAG);
6111 assert(Op0.getValueType().is64BitVector() &&
6112 Op1.getValueType().is64BitVector() &&
6113 "unexpected types for extended operands to VMULL");
6114 return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
6117 // Optimizing (zext A + zext B) * C, to (VMULL A, C) + (VMULL B, C) during
6118 // isel lowering to take advantage of no-stall back to back vmul + vmla.
6125 SDValue N00 = SkipExtensionForVMULL(N0->getOperand(0).getNode(), DAG);
6126 SDValue N01 = SkipExtensionForVMULL(N0->getOperand(1).getNode(), DAG);
6127 EVT Op1VT = Op1.getValueType();
6128 return DAG.getNode(N0->getOpcode(), DL, VT,
6129 DAG.getNode(NewOpc, DL, VT,
6130 DAG.getNode(ISD::BITCAST, DL, Op1VT, N00), Op1),
6131 DAG.getNode(NewOpc, DL, VT,
6132 DAG.getNode(ISD::BITCAST, DL, Op1VT, N01), Op1));
6136 LowerSDIV_v4i8(SDValue X, SDValue Y, SDLoc dl, SelectionDAG &DAG) {
6138 // float4 xf = vcvt_f32_s32(vmovl_s16(a.lo));
6139 // float4 yf = vcvt_f32_s32(vmovl_s16(b.lo));
6140 X = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, X);
6141 Y = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, Y);
6142 X = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, X);
6143 Y = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, Y);
6144 // Get reciprocal estimate.
6145 // float4 recip = vrecpeq_f32(yf);
6146 Y = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6147 DAG.getConstant(Intrinsic::arm_neon_vrecpe, dl, MVT::i32),
6149 // Because char has a smaller range than uchar, we can actually get away
6150 // without any newton steps. This requires that we use a weird bias
6151 // of 0xb000, however (again, this has been exhaustively tested).
6152 // float4 result = as_float4(as_int4(xf*recip) + 0xb000);
6153 X = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, X, Y);
6154 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, X);
6155 Y = DAG.getConstant(0xb000, dl, MVT::i32);
6156 Y = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Y, Y, Y, Y);
6157 X = DAG.getNode(ISD::ADD, dl, MVT::v4i32, X, Y);
6158 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, X);
6159 // Convert back to short.
6160 X = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, X);
6161 X = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, X);
6166 LowerSDIV_v4i16(SDValue N0, SDValue N1, SDLoc dl, SelectionDAG &DAG) {
6168 // Convert to float.
6169 // float4 yf = vcvt_f32_s32(vmovl_s16(y));
6170 // float4 xf = vcvt_f32_s32(vmovl_s16(x));
6171 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N0);
6172 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N1);
6173 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
6174 N1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
6176 // Use reciprocal estimate and one refinement step.
6177 // float4 recip = vrecpeq_f32(yf);
6178 // recip *= vrecpsq_f32(yf, recip);
6179 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6180 DAG.getConstant(Intrinsic::arm_neon_vrecpe, dl, MVT::i32),
6182 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6183 DAG.getConstant(Intrinsic::arm_neon_vrecps, dl, MVT::i32),
6185 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
6186 // Because short has a smaller range than ushort, we can actually get away
6187 // with only a single newton step. This requires that we use a weird bias
6188 // of 89, however (again, this has been exhaustively tested).
6189 // float4 result = as_float4(as_int4(xf*recip) + 0x89);
6190 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
6191 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
6192 N1 = DAG.getConstant(0x89, dl, MVT::i32);
6193 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
6194 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
6195 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
6196 // Convert back to integer and return.
6197 // return vmovn_s32(vcvt_s32_f32(result));
6198 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
6199 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
6203 static SDValue LowerSDIV(SDValue Op, SelectionDAG &DAG) {
6204 EVT VT = Op.getValueType();
6205 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
6206 "unexpected type for custom-lowering ISD::SDIV");
6209 SDValue N0 = Op.getOperand(0);
6210 SDValue N1 = Op.getOperand(1);
6213 if (VT == MVT::v8i8) {
6214 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N0);
6215 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N1);
6217 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
6218 DAG.getIntPtrConstant(4, dl));
6219 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
6220 DAG.getIntPtrConstant(4, dl));
6221 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
6222 DAG.getIntPtrConstant(0, dl));
6223 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
6224 DAG.getIntPtrConstant(0, dl));
6226 N0 = LowerSDIV_v4i8(N0, N1, dl, DAG); // v4i16
6227 N2 = LowerSDIV_v4i8(N2, N3, dl, DAG); // v4i16
6229 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
6230 N0 = LowerCONCAT_VECTORS(N0, DAG);
6232 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v8i8, N0);
6235 return LowerSDIV_v4i16(N0, N1, dl, DAG);
6238 static SDValue LowerUDIV(SDValue Op, SelectionDAG &DAG) {
6239 EVT VT = Op.getValueType();
6240 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
6241 "unexpected type for custom-lowering ISD::UDIV");
6244 SDValue N0 = Op.getOperand(0);
6245 SDValue N1 = Op.getOperand(1);
6248 if (VT == MVT::v8i8) {
6249 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N0);
6250 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N1);
6252 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
6253 DAG.getIntPtrConstant(4, dl));
6254 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
6255 DAG.getIntPtrConstant(4, dl));
6256 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
6257 DAG.getIntPtrConstant(0, dl));
6258 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
6259 DAG.getIntPtrConstant(0, dl));
6261 N0 = LowerSDIV_v4i16(N0, N1, dl, DAG); // v4i16
6262 N2 = LowerSDIV_v4i16(N2, N3, dl, DAG); // v4i16
6264 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
6265 N0 = LowerCONCAT_VECTORS(N0, DAG);
6267 N0 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v8i8,
6268 DAG.getConstant(Intrinsic::arm_neon_vqmovnsu, dl,
6274 // v4i16 sdiv ... Convert to float.
6275 // float4 yf = vcvt_f32_s32(vmovl_u16(y));
6276 // float4 xf = vcvt_f32_s32(vmovl_u16(x));
6277 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N0);
6278 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N1);
6279 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
6280 SDValue BN1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
6282 // Use reciprocal estimate and two refinement steps.
6283 // float4 recip = vrecpeq_f32(yf);
6284 // recip *= vrecpsq_f32(yf, recip);
6285 // recip *= vrecpsq_f32(yf, recip);
6286 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6287 DAG.getConstant(Intrinsic::arm_neon_vrecpe, dl, MVT::i32),
6289 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6290 DAG.getConstant(Intrinsic::arm_neon_vrecps, dl, MVT::i32),
6292 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
6293 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6294 DAG.getConstant(Intrinsic::arm_neon_vrecps, dl, MVT::i32),
6296 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
6297 // Simply multiplying by the reciprocal estimate can leave us a few ulps
6298 // too low, so we add 2 ulps (exhaustive testing shows that this is enough,
6299 // and that it will never cause us to return an answer too large).
6300 // float4 result = as_float4(as_int4(xf*recip) + 2);
6301 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
6302 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
6303 N1 = DAG.getConstant(2, dl, MVT::i32);
6304 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
6305 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
6306 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
6307 // Convert back to integer and return.
6308 // return vmovn_u32(vcvt_s32_f32(result));
6309 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
6310 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
6314 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
6315 EVT VT = Op.getNode()->getValueType(0);
6316 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
6319 bool ExtraOp = false;
6320 switch (Op.getOpcode()) {
6321 default: llvm_unreachable("Invalid code");
6322 case ISD::ADDC: Opc = ARMISD::ADDC; break;
6323 case ISD::ADDE: Opc = ARMISD::ADDE; ExtraOp = true; break;
6324 case ISD::SUBC: Opc = ARMISD::SUBC; break;
6325 case ISD::SUBE: Opc = ARMISD::SUBE; ExtraOp = true; break;
6329 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
6331 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
6332 Op.getOperand(1), Op.getOperand(2));
6335 SDValue ARMTargetLowering::LowerFSINCOS(SDValue Op, SelectionDAG &DAG) const {
6336 assert(Subtarget->isTargetDarwin());
6338 // For iOS, we want to call an alternative entry point: __sincos_stret,
6339 // return values are passed via sret.
6341 SDValue Arg = Op.getOperand(0);
6342 EVT ArgVT = Arg.getValueType();
6343 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
6344 auto PtrVT = getPointerTy(DAG.getDataLayout());
6346 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
6348 // Pair of floats / doubles used to pass the result.
6349 StructType *RetTy = StructType::get(ArgTy, ArgTy, nullptr);
6351 // Create stack object for sret.
6352 auto &DL = DAG.getDataLayout();
6353 const uint64_t ByteSize = DL.getTypeAllocSize(RetTy);
6354 const unsigned StackAlign = DL.getPrefTypeAlignment(RetTy);
6355 int FrameIdx = FrameInfo->CreateStackObject(ByteSize, StackAlign, false);
6356 SDValue SRet = DAG.getFrameIndex(FrameIdx, getPointerTy(DL));
6362 Entry.Ty = RetTy->getPointerTo();
6363 Entry.isSExt = false;
6364 Entry.isZExt = false;
6365 Entry.isSRet = true;
6366 Args.push_back(Entry);
6370 Entry.isSExt = false;
6371 Entry.isZExt = false;
6372 Args.push_back(Entry);
6374 const char *LibcallName = (ArgVT == MVT::f64)
6375 ? "__sincos_stret" : "__sincosf_stret";
6376 SDValue Callee = DAG.getExternalSymbol(LibcallName, getPointerTy(DL));
6378 TargetLowering::CallLoweringInfo CLI(DAG);
6379 CLI.setDebugLoc(dl).setChain(DAG.getEntryNode())
6380 .setCallee(CallingConv::C, Type::getVoidTy(*DAG.getContext()), Callee,
6382 .setDiscardResult();
6384 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
6386 SDValue LoadSin = DAG.getLoad(ArgVT, dl, CallResult.second, SRet,
6387 MachinePointerInfo(), false, false, false, 0);
6389 // Address of cos field.
6390 SDValue Add = DAG.getNode(ISD::ADD, dl, PtrVT, SRet,
6391 DAG.getIntPtrConstant(ArgVT.getStoreSize(), dl));
6392 SDValue LoadCos = DAG.getLoad(ArgVT, dl, LoadSin.getValue(1), Add,
6393 MachinePointerInfo(), false, false, false, 0);
6395 SDVTList Tys = DAG.getVTList(ArgVT, ArgVT);
6396 return DAG.getNode(ISD::MERGE_VALUES, dl, Tys,
6397 LoadSin.getValue(0), LoadCos.getValue(0));
6400 static SDValue LowerAtomicLoadStore(SDValue Op, SelectionDAG &DAG) {
6401 // Monotonic load/store is legal for all targets
6402 if (cast<AtomicSDNode>(Op)->getOrdering() <= Monotonic)
6405 // Acquire/Release load/store is not legal for targets without a
6406 // dmb or equivalent available.
6410 static void ReplaceREADCYCLECOUNTER(SDNode *N,
6411 SmallVectorImpl<SDValue> &Results,
6413 const ARMSubtarget *Subtarget) {
6415 SDValue Cycles32, OutChain;
6417 if (Subtarget->hasPerfMon()) {
6418 // Under Power Management extensions, the cycle-count is:
6419 // mrc p15, #0, <Rt>, c9, c13, #0
6420 SDValue Ops[] = { N->getOperand(0), // Chain
6421 DAG.getConstant(Intrinsic::arm_mrc, DL, MVT::i32),
6422 DAG.getConstant(15, DL, MVT::i32),
6423 DAG.getConstant(0, DL, MVT::i32),
6424 DAG.getConstant(9, DL, MVT::i32),
6425 DAG.getConstant(13, DL, MVT::i32),
6426 DAG.getConstant(0, DL, MVT::i32)
6429 Cycles32 = DAG.getNode(ISD::INTRINSIC_W_CHAIN, DL,
6430 DAG.getVTList(MVT::i32, MVT::Other), Ops);
6431 OutChain = Cycles32.getValue(1);
6433 // Intrinsic is defined to return 0 on unsupported platforms. Technically
6434 // there are older ARM CPUs that have implementation-specific ways of
6435 // obtaining this information (FIXME!).
6436 Cycles32 = DAG.getConstant(0, DL, MVT::i32);
6437 OutChain = DAG.getEntryNode();
6441 SDValue Cycles64 = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64,
6442 Cycles32, DAG.getConstant(0, DL, MVT::i32));
6443 Results.push_back(Cycles64);
6444 Results.push_back(OutChain);
6447 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
6448 switch (Op.getOpcode()) {
6449 default: llvm_unreachable("Don't know how to custom lower this!");
6450 case ISD::WRITE_REGISTER: return LowerWRITE_REGISTER(Op, DAG);
6451 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
6452 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
6453 case ISD::GlobalAddress:
6454 switch (Subtarget->getTargetTriple().getObjectFormat()) {
6455 default: llvm_unreachable("unknown object format");
6457 return LowerGlobalAddressWindows(Op, DAG);
6459 return LowerGlobalAddressELF(Op, DAG);
6461 return LowerGlobalAddressDarwin(Op, DAG);
6463 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
6464 case ISD::SELECT: return LowerSELECT(Op, DAG);
6465 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
6466 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
6467 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
6468 case ISD::VASTART: return LowerVASTART(Op, DAG);
6469 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, DAG, Subtarget);
6470 case ISD::PREFETCH: return LowerPREFETCH(Op, DAG, Subtarget);
6471 case ISD::SINT_TO_FP:
6472 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
6473 case ISD::FP_TO_SINT:
6474 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
6475 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
6476 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
6477 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
6478 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
6479 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
6480 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
6481 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
6483 case ISD::BITCAST: return ExpandBITCAST(Op.getNode(), DAG);
6486 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
6487 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
6488 case ISD::SRL_PARTS:
6489 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
6490 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
6491 case ISD::CTPOP: return LowerCTPOP(Op.getNode(), DAG, Subtarget);
6492 case ISD::SETCC: return LowerVSETCC(Op, DAG);
6493 case ISD::ConstantFP: return LowerConstantFP(Op, DAG, Subtarget);
6494 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG, Subtarget);
6495 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
6496 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
6497 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
6498 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
6499 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
6500 case ISD::MUL: return LowerMUL(Op, DAG);
6501 case ISD::SDIV: return LowerSDIV(Op, DAG);
6502 case ISD::UDIV: return LowerUDIV(Op, DAG);
6506 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
6511 return LowerXALUO(Op, DAG);
6512 case ISD::ATOMIC_LOAD:
6513 case ISD::ATOMIC_STORE: return LowerAtomicLoadStore(Op, DAG);
6514 case ISD::FSINCOS: return LowerFSINCOS(Op, DAG);
6516 case ISD::UDIVREM: return LowerDivRem(Op, DAG);
6517 case ISD::DYNAMIC_STACKALLOC:
6518 if (Subtarget->getTargetTriple().isWindowsItaniumEnvironment())
6519 return LowerDYNAMIC_STACKALLOC(Op, DAG);
6520 llvm_unreachable("Don't know how to custom lower this!");
6521 case ISD::FP_ROUND: return LowerFP_ROUND(Op, DAG);
6522 case ISD::FP_EXTEND: return LowerFP_EXTEND(Op, DAG);
6526 /// ReplaceNodeResults - Replace the results of node with an illegal result
6527 /// type with new values built out of custom code.
6528 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
6529 SmallVectorImpl<SDValue>&Results,
6530 SelectionDAG &DAG) const {
6532 switch (N->getOpcode()) {
6534 llvm_unreachable("Don't know how to custom expand this!");
6535 case ISD::READ_REGISTER:
6536 ExpandREAD_REGISTER(N, Results, DAG);
6539 Res = ExpandBITCAST(N, DAG);
6543 Res = Expand64BitShift(N, DAG, Subtarget);
6545 case ISD::READCYCLECOUNTER:
6546 ReplaceREADCYCLECOUNTER(N, Results, DAG, Subtarget);
6550 Results.push_back(Res);
6553 //===----------------------------------------------------------------------===//
6554 // ARM Scheduler Hooks
6555 //===----------------------------------------------------------------------===//
6557 /// SetupEntryBlockForSjLj - Insert code into the entry block that creates and
6558 /// registers the function context.
6559 void ARMTargetLowering::
6560 SetupEntryBlockForSjLj(MachineInstr *MI, MachineBasicBlock *MBB,
6561 MachineBasicBlock *DispatchBB, int FI) const {
6562 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
6563 DebugLoc dl = MI->getDebugLoc();
6564 MachineFunction *MF = MBB->getParent();
6565 MachineRegisterInfo *MRI = &MF->getRegInfo();
6566 MachineConstantPool *MCP = MF->getConstantPool();
6567 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
6568 const Function *F = MF->getFunction();
6570 bool isThumb = Subtarget->isThumb();
6571 bool isThumb2 = Subtarget->isThumb2();
6573 unsigned PCLabelId = AFI->createPICLabelUId();
6574 unsigned PCAdj = (isThumb || isThumb2) ? 4 : 8;
6575 ARMConstantPoolValue *CPV =
6576 ARMConstantPoolMBB::Create(F->getContext(), DispatchBB, PCLabelId, PCAdj);
6577 unsigned CPI = MCP->getConstantPoolIndex(CPV, 4);
6579 const TargetRegisterClass *TRC = isThumb ? &ARM::tGPRRegClass
6580 : &ARM::GPRRegClass;
6582 // Grab constant pool and fixed stack memory operands.
6583 MachineMemOperand *CPMMO =
6584 MF->getMachineMemOperand(MachinePointerInfo::getConstantPool(),
6585 MachineMemOperand::MOLoad, 4, 4);
6587 MachineMemOperand *FIMMOSt =
6588 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
6589 MachineMemOperand::MOStore, 4, 4);
6591 // Load the address of the dispatch MBB into the jump buffer.
6593 // Incoming value: jbuf
6594 // ldr.n r5, LCPI1_1
6597 // str r5, [$jbuf, #+4] ; &jbuf[1]
6598 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6599 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2LDRpci), NewVReg1)
6600 .addConstantPoolIndex(CPI)
6601 .addMemOperand(CPMMO));
6602 // Set the low bit because of thumb mode.
6603 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6605 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2ORRri), NewVReg2)
6606 .addReg(NewVReg1, RegState::Kill)
6608 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6609 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg3)
6610 .addReg(NewVReg2, RegState::Kill)
6612 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2STRi12))
6613 .addReg(NewVReg3, RegState::Kill)
6615 .addImm(36) // &jbuf[1] :: pc
6616 .addMemOperand(FIMMOSt));
6617 } else if (isThumb) {
6618 // Incoming value: jbuf
6619 // ldr.n r1, LCPI1_4
6623 // add r2, $jbuf, #+4 ; &jbuf[1]
6625 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6626 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tLDRpci), NewVReg1)
6627 .addConstantPoolIndex(CPI)
6628 .addMemOperand(CPMMO));
6629 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6630 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg2)
6631 .addReg(NewVReg1, RegState::Kill)
6633 // Set the low bit because of thumb mode.
6634 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6635 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tMOVi8), NewVReg3)
6636 .addReg(ARM::CPSR, RegState::Define)
6638 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6639 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tORR), NewVReg4)
6640 .addReg(ARM::CPSR, RegState::Define)
6641 .addReg(NewVReg2, RegState::Kill)
6642 .addReg(NewVReg3, RegState::Kill));
6643 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6644 BuildMI(*MBB, MI, dl, TII->get(ARM::tADDframe), NewVReg5)
6646 .addImm(36); // &jbuf[1] :: pc
6647 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tSTRi))
6648 .addReg(NewVReg4, RegState::Kill)
6649 .addReg(NewVReg5, RegState::Kill)
6651 .addMemOperand(FIMMOSt));
6653 // Incoming value: jbuf
6656 // str r1, [$jbuf, #+4] ; &jbuf[1]
6657 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6658 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::LDRi12), NewVReg1)
6659 .addConstantPoolIndex(CPI)
6661 .addMemOperand(CPMMO));
6662 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6663 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::PICADD), NewVReg2)
6664 .addReg(NewVReg1, RegState::Kill)
6665 .addImm(PCLabelId));
6666 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::STRi12))
6667 .addReg(NewVReg2, RegState::Kill)
6669 .addImm(36) // &jbuf[1] :: pc
6670 .addMemOperand(FIMMOSt));
6674 void ARMTargetLowering::EmitSjLjDispatchBlock(MachineInstr *MI,
6675 MachineBasicBlock *MBB) const {
6676 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
6677 DebugLoc dl = MI->getDebugLoc();
6678 MachineFunction *MF = MBB->getParent();
6679 MachineRegisterInfo *MRI = &MF->getRegInfo();
6680 MachineFrameInfo *MFI = MF->getFrameInfo();
6681 int FI = MFI->getFunctionContextIndex();
6683 const TargetRegisterClass *TRC = Subtarget->isThumb() ? &ARM::tGPRRegClass
6684 : &ARM::GPRnopcRegClass;
6686 // Get a mapping of the call site numbers to all of the landing pads they're
6688 DenseMap<unsigned, SmallVector<MachineBasicBlock*, 2> > CallSiteNumToLPad;
6689 unsigned MaxCSNum = 0;
6690 MachineModuleInfo &MMI = MF->getMMI();
6691 for (MachineFunction::iterator BB = MF->begin(), E = MF->end(); BB != E;
6693 if (!BB->isLandingPad()) continue;
6695 // FIXME: We should assert that the EH_LABEL is the first MI in the landing
6697 for (MachineBasicBlock::iterator
6698 II = BB->begin(), IE = BB->end(); II != IE; ++II) {
6699 if (!II->isEHLabel()) continue;
6701 MCSymbol *Sym = II->getOperand(0).getMCSymbol();
6702 if (!MMI.hasCallSiteLandingPad(Sym)) continue;
6704 SmallVectorImpl<unsigned> &CallSiteIdxs = MMI.getCallSiteLandingPad(Sym);
6705 for (SmallVectorImpl<unsigned>::iterator
6706 CSI = CallSiteIdxs.begin(), CSE = CallSiteIdxs.end();
6707 CSI != CSE; ++CSI) {
6708 CallSiteNumToLPad[*CSI].push_back(BB);
6709 MaxCSNum = std::max(MaxCSNum, *CSI);
6715 // Get an ordered list of the machine basic blocks for the jump table.
6716 std::vector<MachineBasicBlock*> LPadList;
6717 SmallPtrSet<MachineBasicBlock*, 64> InvokeBBs;
6718 LPadList.reserve(CallSiteNumToLPad.size());
6719 for (unsigned I = 1; I <= MaxCSNum; ++I) {
6720 SmallVectorImpl<MachineBasicBlock*> &MBBList = CallSiteNumToLPad[I];
6721 for (SmallVectorImpl<MachineBasicBlock*>::iterator
6722 II = MBBList.begin(), IE = MBBList.end(); II != IE; ++II) {
6723 LPadList.push_back(*II);
6724 InvokeBBs.insert((*II)->pred_begin(), (*II)->pred_end());
6728 assert(!LPadList.empty() &&
6729 "No landing pad destinations for the dispatch jump table!");
6731 // Create the jump table and associated information.
6732 MachineJumpTableInfo *JTI =
6733 MF->getOrCreateJumpTableInfo(MachineJumpTableInfo::EK_Inline);
6734 unsigned MJTI = JTI->createJumpTableIndex(LPadList);
6735 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
6737 // Create the MBBs for the dispatch code.
6739 // Shove the dispatch's address into the return slot in the function context.
6740 MachineBasicBlock *DispatchBB = MF->CreateMachineBasicBlock();
6741 DispatchBB->setIsLandingPad();
6743 MachineBasicBlock *TrapBB = MF->CreateMachineBasicBlock();
6744 unsigned trap_opcode;
6745 if (Subtarget->isThumb())
6746 trap_opcode = ARM::tTRAP;
6748 trap_opcode = Subtarget->useNaClTrap() ? ARM::TRAPNaCl : ARM::TRAP;
6750 BuildMI(TrapBB, dl, TII->get(trap_opcode));
6751 DispatchBB->addSuccessor(TrapBB);
6753 MachineBasicBlock *DispContBB = MF->CreateMachineBasicBlock();
6754 DispatchBB->addSuccessor(DispContBB);
6757 MF->insert(MF->end(), DispatchBB);
6758 MF->insert(MF->end(), DispContBB);
6759 MF->insert(MF->end(), TrapBB);
6761 // Insert code into the entry block that creates and registers the function
6763 SetupEntryBlockForSjLj(MI, MBB, DispatchBB, FI);
6765 MachineMemOperand *FIMMOLd =
6766 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
6767 MachineMemOperand::MOLoad |
6768 MachineMemOperand::MOVolatile, 4, 4);
6770 MachineInstrBuilder MIB;
6771 MIB = BuildMI(DispatchBB, dl, TII->get(ARM::Int_eh_sjlj_dispatchsetup));
6773 const ARMBaseInstrInfo *AII = static_cast<const ARMBaseInstrInfo*>(TII);
6774 const ARMBaseRegisterInfo &RI = AII->getRegisterInfo();
6776 // Add a register mask with no preserved registers. This results in all
6777 // registers being marked as clobbered.
6778 MIB.addRegMask(RI.getNoPreservedMask());
6780 unsigned NumLPads = LPadList.size();
6781 if (Subtarget->isThumb2()) {
6782 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6783 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2LDRi12), NewVReg1)
6786 .addMemOperand(FIMMOLd));
6788 if (NumLPads < 256) {
6789 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPri))
6791 .addImm(LPadList.size()));
6793 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6794 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVi16), VReg1)
6795 .addImm(NumLPads & 0xFFFF));
6797 unsigned VReg2 = VReg1;
6798 if ((NumLPads & 0xFFFF0000) != 0) {
6799 VReg2 = MRI->createVirtualRegister(TRC);
6800 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVTi16), VReg2)
6802 .addImm(NumLPads >> 16));
6805 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPrr))
6810 BuildMI(DispatchBB, dl, TII->get(ARM::t2Bcc))
6815 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6816 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::t2LEApcrelJT),NewVReg3)
6817 .addJumpTableIndex(MJTI));
6819 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6822 BuildMI(DispContBB, dl, TII->get(ARM::t2ADDrs), NewVReg4)
6823 .addReg(NewVReg3, RegState::Kill)
6825 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
6827 BuildMI(DispContBB, dl, TII->get(ARM::t2BR_JT))
6828 .addReg(NewVReg4, RegState::Kill)
6830 .addJumpTableIndex(MJTI);
6831 } else if (Subtarget->isThumb()) {
6832 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6833 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRspi), NewVReg1)
6836 .addMemOperand(FIMMOLd));
6838 if (NumLPads < 256) {
6839 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPi8))
6843 MachineConstantPool *ConstantPool = MF->getConstantPool();
6844 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
6845 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
6847 // MachineConstantPool wants an explicit alignment.
6848 unsigned Align = MF->getDataLayout().getPrefTypeAlignment(Int32Ty);
6850 Align = MF->getDataLayout().getTypeAllocSize(C->getType());
6851 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6853 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6854 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRpci))
6855 .addReg(VReg1, RegState::Define)
6856 .addConstantPoolIndex(Idx));
6857 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPr))
6862 BuildMI(DispatchBB, dl, TII->get(ARM::tBcc))
6867 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6868 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLSLri), NewVReg2)
6869 .addReg(ARM::CPSR, RegState::Define)
6873 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6874 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLEApcrelJT), NewVReg3)
6875 .addJumpTableIndex(MJTI));
6877 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6878 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg4)
6879 .addReg(ARM::CPSR, RegState::Define)
6880 .addReg(NewVReg2, RegState::Kill)
6883 MachineMemOperand *JTMMOLd =
6884 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
6885 MachineMemOperand::MOLoad, 4, 4);
6887 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6888 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLDRi), NewVReg5)
6889 .addReg(NewVReg4, RegState::Kill)
6891 .addMemOperand(JTMMOLd));
6893 unsigned NewVReg6 = NewVReg5;
6894 if (RelocM == Reloc::PIC_) {
6895 NewVReg6 = MRI->createVirtualRegister(TRC);
6896 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg6)
6897 .addReg(ARM::CPSR, RegState::Define)
6898 .addReg(NewVReg5, RegState::Kill)
6902 BuildMI(DispContBB, dl, TII->get(ARM::tBR_JTr))
6903 .addReg(NewVReg6, RegState::Kill)
6904 .addJumpTableIndex(MJTI);
6906 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6907 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRi12), NewVReg1)
6910 .addMemOperand(FIMMOLd));
6912 if (NumLPads < 256) {
6913 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPri))
6916 } else if (Subtarget->hasV6T2Ops() && isUInt<16>(NumLPads)) {
6917 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6918 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVi16), VReg1)
6919 .addImm(NumLPads & 0xFFFF));
6921 unsigned VReg2 = VReg1;
6922 if ((NumLPads & 0xFFFF0000) != 0) {
6923 VReg2 = MRI->createVirtualRegister(TRC);
6924 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVTi16), VReg2)
6926 .addImm(NumLPads >> 16));
6929 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
6933 MachineConstantPool *ConstantPool = MF->getConstantPool();
6934 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
6935 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
6937 // MachineConstantPool wants an explicit alignment.
6938 unsigned Align = MF->getDataLayout().getPrefTypeAlignment(Int32Ty);
6940 Align = MF->getDataLayout().getTypeAllocSize(C->getType());
6941 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6943 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6944 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRcp))
6945 .addReg(VReg1, RegState::Define)
6946 .addConstantPoolIndex(Idx)
6948 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
6950 .addReg(VReg1, RegState::Kill));
6953 BuildMI(DispatchBB, dl, TII->get(ARM::Bcc))
6958 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6960 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::MOVsi), NewVReg3)
6962 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
6963 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6964 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::LEApcrelJT), NewVReg4)
6965 .addJumpTableIndex(MJTI));
6967 MachineMemOperand *JTMMOLd =
6968 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
6969 MachineMemOperand::MOLoad, 4, 4);
6970 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6972 BuildMI(DispContBB, dl, TII->get(ARM::LDRrs), NewVReg5)
6973 .addReg(NewVReg3, RegState::Kill)
6976 .addMemOperand(JTMMOLd));
6978 if (RelocM == Reloc::PIC_) {
6979 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTadd))
6980 .addReg(NewVReg5, RegState::Kill)
6982 .addJumpTableIndex(MJTI);
6984 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTr))
6985 .addReg(NewVReg5, RegState::Kill)
6986 .addJumpTableIndex(MJTI);
6990 // Add the jump table entries as successors to the MBB.
6991 SmallPtrSet<MachineBasicBlock*, 8> SeenMBBs;
6992 for (std::vector<MachineBasicBlock*>::iterator
6993 I = LPadList.begin(), E = LPadList.end(); I != E; ++I) {
6994 MachineBasicBlock *CurMBB = *I;
6995 if (SeenMBBs.insert(CurMBB).second)
6996 DispContBB->addSuccessor(CurMBB);
6999 // N.B. the order the invoke BBs are processed in doesn't matter here.
7000 const MCPhysReg *SavedRegs = RI.getCalleeSavedRegs(MF);
7001 SmallVector<MachineBasicBlock*, 64> MBBLPads;
7002 for (MachineBasicBlock *BB : InvokeBBs) {
7004 // Remove the landing pad successor from the invoke block and replace it
7005 // with the new dispatch block.
7006 SmallVector<MachineBasicBlock*, 4> Successors(BB->succ_begin(),
7008 while (!Successors.empty()) {
7009 MachineBasicBlock *SMBB = Successors.pop_back_val();
7010 if (SMBB->isLandingPad()) {
7011 BB->removeSuccessor(SMBB);
7012 MBBLPads.push_back(SMBB);
7016 BB->addSuccessor(DispatchBB);
7018 // Find the invoke call and mark all of the callee-saved registers as
7019 // 'implicit defined' so that they're spilled. This prevents code from
7020 // moving instructions to before the EH block, where they will never be
7022 for (MachineBasicBlock::reverse_iterator
7023 II = BB->rbegin(), IE = BB->rend(); II != IE; ++II) {
7024 if (!II->isCall()) continue;
7026 DenseMap<unsigned, bool> DefRegs;
7027 for (MachineInstr::mop_iterator
7028 OI = II->operands_begin(), OE = II->operands_end();
7030 if (!OI->isReg()) continue;
7031 DefRegs[OI->getReg()] = true;
7034 MachineInstrBuilder MIB(*MF, &*II);
7036 for (unsigned i = 0; SavedRegs[i] != 0; ++i) {
7037 unsigned Reg = SavedRegs[i];
7038 if (Subtarget->isThumb2() &&
7039 !ARM::tGPRRegClass.contains(Reg) &&
7040 !ARM::hGPRRegClass.contains(Reg))
7042 if (Subtarget->isThumb1Only() && !ARM::tGPRRegClass.contains(Reg))
7044 if (!Subtarget->isThumb() && !ARM::GPRRegClass.contains(Reg))
7047 MIB.addReg(Reg, RegState::ImplicitDefine | RegState::Dead);
7054 // Mark all former landing pads as non-landing pads. The dispatch is the only
7056 for (SmallVectorImpl<MachineBasicBlock*>::iterator
7057 I = MBBLPads.begin(), E = MBBLPads.end(); I != E; ++I)
7058 (*I)->setIsLandingPad(false);
7060 // The instruction is gone now.
7061 MI->eraseFromParent();
7065 MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) {
7066 for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
7067 E = MBB->succ_end(); I != E; ++I)
7070 llvm_unreachable("Expecting a BB with two successors!");
7073 /// Return the load opcode for a given load size. If load size >= 8,
7074 /// neon opcode will be returned.
7075 static unsigned getLdOpcode(unsigned LdSize, bool IsThumb1, bool IsThumb2) {
7077 return LdSize == 16 ? ARM::VLD1q32wb_fixed
7078 : LdSize == 8 ? ARM::VLD1d32wb_fixed : 0;
7080 return LdSize == 4 ? ARM::tLDRi
7081 : LdSize == 2 ? ARM::tLDRHi
7082 : LdSize == 1 ? ARM::tLDRBi : 0;
7084 return LdSize == 4 ? ARM::t2LDR_POST
7085 : LdSize == 2 ? ARM::t2LDRH_POST
7086 : LdSize == 1 ? ARM::t2LDRB_POST : 0;
7087 return LdSize == 4 ? ARM::LDR_POST_IMM
7088 : LdSize == 2 ? ARM::LDRH_POST
7089 : LdSize == 1 ? ARM::LDRB_POST_IMM : 0;
7092 /// Return the store opcode for a given store size. If store size >= 8,
7093 /// neon opcode will be returned.
7094 static unsigned getStOpcode(unsigned StSize, bool IsThumb1, bool IsThumb2) {
7096 return StSize == 16 ? ARM::VST1q32wb_fixed
7097 : StSize == 8 ? ARM::VST1d32wb_fixed : 0;
7099 return StSize == 4 ? ARM::tSTRi
7100 : StSize == 2 ? ARM::tSTRHi
7101 : StSize == 1 ? ARM::tSTRBi : 0;
7103 return StSize == 4 ? ARM::t2STR_POST
7104 : StSize == 2 ? ARM::t2STRH_POST
7105 : StSize == 1 ? ARM::t2STRB_POST : 0;
7106 return StSize == 4 ? ARM::STR_POST_IMM
7107 : StSize == 2 ? ARM::STRH_POST
7108 : StSize == 1 ? ARM::STRB_POST_IMM : 0;
7111 /// Emit a post-increment load operation with given size. The instructions
7112 /// will be added to BB at Pos.
7113 static void emitPostLd(MachineBasicBlock *BB, MachineInstr *Pos,
7114 const TargetInstrInfo *TII, DebugLoc dl,
7115 unsigned LdSize, unsigned Data, unsigned AddrIn,
7116 unsigned AddrOut, bool IsThumb1, bool IsThumb2) {
7117 unsigned LdOpc = getLdOpcode(LdSize, IsThumb1, IsThumb2);
7118 assert(LdOpc != 0 && "Should have a load opcode");
7120 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
7121 .addReg(AddrOut, RegState::Define).addReg(AddrIn)
7123 } else if (IsThumb1) {
7124 // load + update AddrIn
7125 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
7126 .addReg(AddrIn).addImm(0));
7127 MachineInstrBuilder MIB =
7128 BuildMI(*BB, Pos, dl, TII->get(ARM::tADDi8), AddrOut);
7129 MIB = AddDefaultT1CC(MIB);
7130 MIB.addReg(AddrIn).addImm(LdSize);
7131 AddDefaultPred(MIB);
7132 } else if (IsThumb2) {
7133 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
7134 .addReg(AddrOut, RegState::Define).addReg(AddrIn)
7137 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
7138 .addReg(AddrOut, RegState::Define).addReg(AddrIn)
7139 .addReg(0).addImm(LdSize));
7143 /// Emit a post-increment store operation with given size. The instructions
7144 /// will be added to BB at Pos.
7145 static void emitPostSt(MachineBasicBlock *BB, MachineInstr *Pos,
7146 const TargetInstrInfo *TII, DebugLoc dl,
7147 unsigned StSize, unsigned Data, unsigned AddrIn,
7148 unsigned AddrOut, bool IsThumb1, bool IsThumb2) {
7149 unsigned StOpc = getStOpcode(StSize, IsThumb1, IsThumb2);
7150 assert(StOpc != 0 && "Should have a store opcode");
7152 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc), AddrOut)
7153 .addReg(AddrIn).addImm(0).addReg(Data));
7154 } else if (IsThumb1) {
7155 // store + update AddrIn
7156 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc)).addReg(Data)
7157 .addReg(AddrIn).addImm(0));
7158 MachineInstrBuilder MIB =
7159 BuildMI(*BB, Pos, dl, TII->get(ARM::tADDi8), AddrOut);
7160 MIB = AddDefaultT1CC(MIB);
7161 MIB.addReg(AddrIn).addImm(StSize);
7162 AddDefaultPred(MIB);
7163 } else if (IsThumb2) {
7164 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc), AddrOut)
7165 .addReg(Data).addReg(AddrIn).addImm(StSize));
7167 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc), AddrOut)
7168 .addReg(Data).addReg(AddrIn).addReg(0)
7174 ARMTargetLowering::EmitStructByval(MachineInstr *MI,
7175 MachineBasicBlock *BB) const {
7176 // This pseudo instruction has 3 operands: dst, src, size
7177 // We expand it to a loop if size > Subtarget->getMaxInlineSizeThreshold().
7178 // Otherwise, we will generate unrolled scalar copies.
7179 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
7180 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7181 MachineFunction::iterator It = BB;
7184 unsigned dest = MI->getOperand(0).getReg();
7185 unsigned src = MI->getOperand(1).getReg();
7186 unsigned SizeVal = MI->getOperand(2).getImm();
7187 unsigned Align = MI->getOperand(3).getImm();
7188 DebugLoc dl = MI->getDebugLoc();
7190 MachineFunction *MF = BB->getParent();
7191 MachineRegisterInfo &MRI = MF->getRegInfo();
7192 unsigned UnitSize = 0;
7193 const TargetRegisterClass *TRC = nullptr;
7194 const TargetRegisterClass *VecTRC = nullptr;
7196 bool IsThumb1 = Subtarget->isThumb1Only();
7197 bool IsThumb2 = Subtarget->isThumb2();
7201 } else if (Align & 2) {
7204 // Check whether we can use NEON instructions.
7205 if (!MF->getFunction()->hasFnAttribute(Attribute::NoImplicitFloat) &&
7206 Subtarget->hasNEON()) {
7207 if ((Align % 16 == 0) && SizeVal >= 16)
7209 else if ((Align % 8 == 0) && SizeVal >= 8)
7212 // Can't use NEON instructions.
7217 // Select the correct opcode and register class for unit size load/store
7218 bool IsNeon = UnitSize >= 8;
7219 TRC = (IsThumb1 || IsThumb2) ? &ARM::tGPRRegClass : &ARM::GPRRegClass;
7221 VecTRC = UnitSize == 16 ? &ARM::DPairRegClass
7222 : UnitSize == 8 ? &ARM::DPRRegClass
7225 unsigned BytesLeft = SizeVal % UnitSize;
7226 unsigned LoopSize = SizeVal - BytesLeft;
7228 if (SizeVal <= Subtarget->getMaxInlineSizeThreshold()) {
7229 // Use LDR and STR to copy.
7230 // [scratch, srcOut] = LDR_POST(srcIn, UnitSize)
7231 // [destOut] = STR_POST(scratch, destIn, UnitSize)
7232 unsigned srcIn = src;
7233 unsigned destIn = dest;
7234 for (unsigned i = 0; i < LoopSize; i+=UnitSize) {
7235 unsigned srcOut = MRI.createVirtualRegister(TRC);
7236 unsigned destOut = MRI.createVirtualRegister(TRC);
7237 unsigned scratch = MRI.createVirtualRegister(IsNeon ? VecTRC : TRC);
7238 emitPostLd(BB, MI, TII, dl, UnitSize, scratch, srcIn, srcOut,
7239 IsThumb1, IsThumb2);
7240 emitPostSt(BB, MI, TII, dl, UnitSize, scratch, destIn, destOut,
7241 IsThumb1, IsThumb2);
7246 // Handle the leftover bytes with LDRB and STRB.
7247 // [scratch, srcOut] = LDRB_POST(srcIn, 1)
7248 // [destOut] = STRB_POST(scratch, destIn, 1)
7249 for (unsigned i = 0; i < BytesLeft; i++) {
7250 unsigned srcOut = MRI.createVirtualRegister(TRC);
7251 unsigned destOut = MRI.createVirtualRegister(TRC);
7252 unsigned scratch = MRI.createVirtualRegister(TRC);
7253 emitPostLd(BB, MI, TII, dl, 1, scratch, srcIn, srcOut,
7254 IsThumb1, IsThumb2);
7255 emitPostSt(BB, MI, TII, dl, 1, scratch, destIn, destOut,
7256 IsThumb1, IsThumb2);
7260 MI->eraseFromParent(); // The instruction is gone now.
7264 // Expand the pseudo op to a loop.
7267 // movw varEnd, # --> with thumb2
7269 // ldrcp varEnd, idx --> without thumb2
7270 // fallthrough --> loopMBB
7272 // PHI varPhi, varEnd, varLoop
7273 // PHI srcPhi, src, srcLoop
7274 // PHI destPhi, dst, destLoop
7275 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
7276 // [destLoop] = STR_POST(scratch, destPhi, UnitSize)
7277 // subs varLoop, varPhi, #UnitSize
7279 // fallthrough --> exitMBB
7281 // epilogue to handle left-over bytes
7282 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
7283 // [destOut] = STRB_POST(scratch, destLoop, 1)
7284 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
7285 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
7286 MF->insert(It, loopMBB);
7287 MF->insert(It, exitMBB);
7289 // Transfer the remainder of BB and its successor edges to exitMBB.
7290 exitMBB->splice(exitMBB->begin(), BB,
7291 std::next(MachineBasicBlock::iterator(MI)), BB->end());
7292 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
7294 // Load an immediate to varEnd.
7295 unsigned varEnd = MRI.createVirtualRegister(TRC);
7296 if (Subtarget->useMovt(*MF)) {
7297 unsigned Vtmp = varEnd;
7298 if ((LoopSize & 0xFFFF0000) != 0)
7299 Vtmp = MRI.createVirtualRegister(TRC);
7300 AddDefaultPred(BuildMI(BB, dl,
7301 TII->get(IsThumb2 ? ARM::t2MOVi16 : ARM::MOVi16),
7302 Vtmp).addImm(LoopSize & 0xFFFF));
7304 if ((LoopSize & 0xFFFF0000) != 0)
7305 AddDefaultPred(BuildMI(BB, dl,
7306 TII->get(IsThumb2 ? ARM::t2MOVTi16 : ARM::MOVTi16),
7309 .addImm(LoopSize >> 16));
7311 MachineConstantPool *ConstantPool = MF->getConstantPool();
7312 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
7313 const Constant *C = ConstantInt::get(Int32Ty, LoopSize);
7315 // MachineConstantPool wants an explicit alignment.
7316 unsigned Align = MF->getDataLayout().getPrefTypeAlignment(Int32Ty);
7318 Align = MF->getDataLayout().getTypeAllocSize(C->getType());
7319 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
7322 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(ARM::tLDRpci)).addReg(
7323 varEnd, RegState::Define).addConstantPoolIndex(Idx));
7325 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(ARM::LDRcp)).addReg(
7326 varEnd, RegState::Define).addConstantPoolIndex(Idx).addImm(0));
7328 BB->addSuccessor(loopMBB);
7330 // Generate the loop body:
7331 // varPhi = PHI(varLoop, varEnd)
7332 // srcPhi = PHI(srcLoop, src)
7333 // destPhi = PHI(destLoop, dst)
7334 MachineBasicBlock *entryBB = BB;
7336 unsigned varLoop = MRI.createVirtualRegister(TRC);
7337 unsigned varPhi = MRI.createVirtualRegister(TRC);
7338 unsigned srcLoop = MRI.createVirtualRegister(TRC);
7339 unsigned srcPhi = MRI.createVirtualRegister(TRC);
7340 unsigned destLoop = MRI.createVirtualRegister(TRC);
7341 unsigned destPhi = MRI.createVirtualRegister(TRC);
7343 BuildMI(*BB, BB->begin(), dl, TII->get(ARM::PHI), varPhi)
7344 .addReg(varLoop).addMBB(loopMBB)
7345 .addReg(varEnd).addMBB(entryBB);
7346 BuildMI(BB, dl, TII->get(ARM::PHI), srcPhi)
7347 .addReg(srcLoop).addMBB(loopMBB)
7348 .addReg(src).addMBB(entryBB);
7349 BuildMI(BB, dl, TII->get(ARM::PHI), destPhi)
7350 .addReg(destLoop).addMBB(loopMBB)
7351 .addReg(dest).addMBB(entryBB);
7353 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
7354 // [destLoop] = STR_POST(scratch, destPhi, UnitSiz)
7355 unsigned scratch = MRI.createVirtualRegister(IsNeon ? VecTRC : TRC);
7356 emitPostLd(BB, BB->end(), TII, dl, UnitSize, scratch, srcPhi, srcLoop,
7357 IsThumb1, IsThumb2);
7358 emitPostSt(BB, BB->end(), TII, dl, UnitSize, scratch, destPhi, destLoop,
7359 IsThumb1, IsThumb2);
7361 // Decrement loop variable by UnitSize.
7363 MachineInstrBuilder MIB =
7364 BuildMI(*BB, BB->end(), dl, TII->get(ARM::tSUBi8), varLoop);
7365 MIB = AddDefaultT1CC(MIB);
7366 MIB.addReg(varPhi).addImm(UnitSize);
7367 AddDefaultPred(MIB);
7369 MachineInstrBuilder MIB =
7370 BuildMI(*BB, BB->end(), dl,
7371 TII->get(IsThumb2 ? ARM::t2SUBri : ARM::SUBri), varLoop);
7372 AddDefaultCC(AddDefaultPred(MIB.addReg(varPhi).addImm(UnitSize)));
7373 MIB->getOperand(5).setReg(ARM::CPSR);
7374 MIB->getOperand(5).setIsDef(true);
7376 BuildMI(*BB, BB->end(), dl,
7377 TII->get(IsThumb1 ? ARM::tBcc : IsThumb2 ? ARM::t2Bcc : ARM::Bcc))
7378 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
7380 // loopMBB can loop back to loopMBB or fall through to exitMBB.
7381 BB->addSuccessor(loopMBB);
7382 BB->addSuccessor(exitMBB);
7384 // Add epilogue to handle BytesLeft.
7386 MachineInstr *StartOfExit = exitMBB->begin();
7388 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
7389 // [destOut] = STRB_POST(scratch, destLoop, 1)
7390 unsigned srcIn = srcLoop;
7391 unsigned destIn = destLoop;
7392 for (unsigned i = 0; i < BytesLeft; i++) {
7393 unsigned srcOut = MRI.createVirtualRegister(TRC);
7394 unsigned destOut = MRI.createVirtualRegister(TRC);
7395 unsigned scratch = MRI.createVirtualRegister(TRC);
7396 emitPostLd(BB, StartOfExit, TII, dl, 1, scratch, srcIn, srcOut,
7397 IsThumb1, IsThumb2);
7398 emitPostSt(BB, StartOfExit, TII, dl, 1, scratch, destIn, destOut,
7399 IsThumb1, IsThumb2);
7404 MI->eraseFromParent(); // The instruction is gone now.
7409 ARMTargetLowering::EmitLowered__chkstk(MachineInstr *MI,
7410 MachineBasicBlock *MBB) const {
7411 const TargetMachine &TM = getTargetMachine();
7412 const TargetInstrInfo &TII = *Subtarget->getInstrInfo();
7413 DebugLoc DL = MI->getDebugLoc();
7415 assert(Subtarget->isTargetWindows() &&
7416 "__chkstk is only supported on Windows");
7417 assert(Subtarget->isThumb2() && "Windows on ARM requires Thumb-2 mode");
7419 // __chkstk takes the number of words to allocate on the stack in R4, and
7420 // returns the stack adjustment in number of bytes in R4. This will not
7421 // clober any other registers (other than the obvious lr).
7423 // Although, technically, IP should be considered a register which may be
7424 // clobbered, the call itself will not touch it. Windows on ARM is a pure
7425 // thumb-2 environment, so there is no interworking required. As a result, we
7426 // do not expect a veneer to be emitted by the linker, clobbering IP.
7428 // Each module receives its own copy of __chkstk, so no import thunk is
7429 // required, again, ensuring that IP is not clobbered.
7431 // Finally, although some linkers may theoretically provide a trampoline for
7432 // out of range calls (which is quite common due to a 32M range limitation of
7433 // branches for Thumb), we can generate the long-call version via
7434 // -mcmodel=large, alleviating the need for the trampoline which may clobber
7437 switch (TM.getCodeModel()) {
7438 case CodeModel::Small:
7439 case CodeModel::Medium:
7440 case CodeModel::Default:
7441 case CodeModel::Kernel:
7442 BuildMI(*MBB, MI, DL, TII.get(ARM::tBL))
7443 .addImm((unsigned)ARMCC::AL).addReg(0)
7444 .addExternalSymbol("__chkstk")
7445 .addReg(ARM::R4, RegState::Implicit | RegState::Kill)
7446 .addReg(ARM::R4, RegState::Implicit | RegState::Define)
7447 .addReg(ARM::R12, RegState::Implicit | RegState::Define | RegState::Dead);
7449 case CodeModel::Large:
7450 case CodeModel::JITDefault: {
7451 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
7452 unsigned Reg = MRI.createVirtualRegister(&ARM::rGPRRegClass);
7454 BuildMI(*MBB, MI, DL, TII.get(ARM::t2MOVi32imm), Reg)
7455 .addExternalSymbol("__chkstk");
7456 BuildMI(*MBB, MI, DL, TII.get(ARM::tBLXr))
7457 .addImm((unsigned)ARMCC::AL).addReg(0)
7458 .addReg(Reg, RegState::Kill)
7459 .addReg(ARM::R4, RegState::Implicit | RegState::Kill)
7460 .addReg(ARM::R4, RegState::Implicit | RegState::Define)
7461 .addReg(ARM::R12, RegState::Implicit | RegState::Define | RegState::Dead);
7466 AddDefaultCC(AddDefaultPred(BuildMI(*MBB, MI, DL, TII.get(ARM::t2SUBrr),
7468 .addReg(ARM::SP).addReg(ARM::R4)));
7470 MI->eraseFromParent();
7475 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
7476 MachineBasicBlock *BB) const {
7477 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
7478 DebugLoc dl = MI->getDebugLoc();
7479 bool isThumb2 = Subtarget->isThumb2();
7480 switch (MI->getOpcode()) {
7483 llvm_unreachable("Unexpected instr type to insert");
7485 // The Thumb2 pre-indexed stores have the same MI operands, they just
7486 // define them differently in the .td files from the isel patterns, so
7487 // they need pseudos.
7488 case ARM::t2STR_preidx:
7489 MI->setDesc(TII->get(ARM::t2STR_PRE));
7491 case ARM::t2STRB_preidx:
7492 MI->setDesc(TII->get(ARM::t2STRB_PRE));
7494 case ARM::t2STRH_preidx:
7495 MI->setDesc(TII->get(ARM::t2STRH_PRE));
7498 case ARM::STRi_preidx:
7499 case ARM::STRBi_preidx: {
7500 unsigned NewOpc = MI->getOpcode() == ARM::STRi_preidx ?
7501 ARM::STR_PRE_IMM : ARM::STRB_PRE_IMM;
7502 // Decode the offset.
7503 unsigned Offset = MI->getOperand(4).getImm();
7504 bool isSub = ARM_AM::getAM2Op(Offset) == ARM_AM::sub;
7505 Offset = ARM_AM::getAM2Offset(Offset);
7509 MachineMemOperand *MMO = *MI->memoperands_begin();
7510 BuildMI(*BB, MI, dl, TII->get(NewOpc))
7511 .addOperand(MI->getOperand(0)) // Rn_wb
7512 .addOperand(MI->getOperand(1)) // Rt
7513 .addOperand(MI->getOperand(2)) // Rn
7514 .addImm(Offset) // offset (skip GPR==zero_reg)
7515 .addOperand(MI->getOperand(5)) // pred
7516 .addOperand(MI->getOperand(6))
7517 .addMemOperand(MMO);
7518 MI->eraseFromParent();
7521 case ARM::STRr_preidx:
7522 case ARM::STRBr_preidx:
7523 case ARM::STRH_preidx: {
7525 switch (MI->getOpcode()) {
7526 default: llvm_unreachable("unexpected opcode!");
7527 case ARM::STRr_preidx: NewOpc = ARM::STR_PRE_REG; break;
7528 case ARM::STRBr_preidx: NewOpc = ARM::STRB_PRE_REG; break;
7529 case ARM::STRH_preidx: NewOpc = ARM::STRH_PRE; break;
7531 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(NewOpc));
7532 for (unsigned i = 0; i < MI->getNumOperands(); ++i)
7533 MIB.addOperand(MI->getOperand(i));
7534 MI->eraseFromParent();
7538 case ARM::tMOVCCr_pseudo: {
7539 // To "insert" a SELECT_CC instruction, we actually have to insert the
7540 // diamond control-flow pattern. The incoming instruction knows the
7541 // destination vreg to set, the condition code register to branch on, the
7542 // true/false values to select between, and a branch opcode to use.
7543 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7544 MachineFunction::iterator It = BB;
7550 // cmpTY ccX, r1, r2
7552 // fallthrough --> copy0MBB
7553 MachineBasicBlock *thisMBB = BB;
7554 MachineFunction *F = BB->getParent();
7555 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
7556 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
7557 F->insert(It, copy0MBB);
7558 F->insert(It, sinkMBB);
7560 // Transfer the remainder of BB and its successor edges to sinkMBB.
7561 sinkMBB->splice(sinkMBB->begin(), BB,
7562 std::next(MachineBasicBlock::iterator(MI)), BB->end());
7563 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
7565 BB->addSuccessor(copy0MBB);
7566 BB->addSuccessor(sinkMBB);
7568 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
7569 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
7572 // %FalseValue = ...
7573 // # fallthrough to sinkMBB
7576 // Update machine-CFG edges
7577 BB->addSuccessor(sinkMBB);
7580 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
7583 BuildMI(*BB, BB->begin(), dl,
7584 TII->get(ARM::PHI), MI->getOperand(0).getReg())
7585 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
7586 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
7588 MI->eraseFromParent(); // The pseudo instruction is gone now.
7593 case ARM::BCCZi64: {
7594 // If there is an unconditional branch to the other successor, remove it.
7595 BB->erase(std::next(MachineBasicBlock::iterator(MI)), BB->end());
7597 // Compare both parts that make up the double comparison separately for
7599 bool RHSisZero = MI->getOpcode() == ARM::BCCZi64;
7601 unsigned LHS1 = MI->getOperand(1).getReg();
7602 unsigned LHS2 = MI->getOperand(2).getReg();
7604 AddDefaultPred(BuildMI(BB, dl,
7605 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7606 .addReg(LHS1).addImm(0));
7607 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7608 .addReg(LHS2).addImm(0)
7609 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
7611 unsigned RHS1 = MI->getOperand(3).getReg();
7612 unsigned RHS2 = MI->getOperand(4).getReg();
7613 AddDefaultPred(BuildMI(BB, dl,
7614 TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
7615 .addReg(LHS1).addReg(RHS1));
7616 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
7617 .addReg(LHS2).addReg(RHS2)
7618 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
7621 MachineBasicBlock *destMBB = MI->getOperand(RHSisZero ? 3 : 5).getMBB();
7622 MachineBasicBlock *exitMBB = OtherSucc(BB, destMBB);
7623 if (MI->getOperand(0).getImm() == ARMCC::NE)
7624 std::swap(destMBB, exitMBB);
7626 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
7627 .addMBB(destMBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
7629 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2B)).addMBB(exitMBB));
7631 BuildMI(BB, dl, TII->get(ARM::B)) .addMBB(exitMBB);
7633 MI->eraseFromParent(); // The pseudo instruction is gone now.
7637 case ARM::Int_eh_sjlj_setjmp:
7638 case ARM::Int_eh_sjlj_setjmp_nofp:
7639 case ARM::tInt_eh_sjlj_setjmp:
7640 case ARM::t2Int_eh_sjlj_setjmp:
7641 case ARM::t2Int_eh_sjlj_setjmp_nofp:
7642 EmitSjLjDispatchBlock(MI, BB);
7647 // To insert an ABS instruction, we have to insert the
7648 // diamond control-flow pattern. The incoming instruction knows the
7649 // source vreg to test against 0, the destination vreg to set,
7650 // the condition code register to branch on, the
7651 // true/false values to select between, and a branch opcode to use.
7656 // BCC (branch to SinkBB if V0 >= 0)
7657 // RSBBB: V3 = RSBri V2, 0 (compute ABS if V2 < 0)
7658 // SinkBB: V1 = PHI(V2, V3)
7659 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7660 MachineFunction::iterator BBI = BB;
7662 MachineFunction *Fn = BB->getParent();
7663 MachineBasicBlock *RSBBB = Fn->CreateMachineBasicBlock(LLVM_BB);
7664 MachineBasicBlock *SinkBB = Fn->CreateMachineBasicBlock(LLVM_BB);
7665 Fn->insert(BBI, RSBBB);
7666 Fn->insert(BBI, SinkBB);
7668 unsigned int ABSSrcReg = MI->getOperand(1).getReg();
7669 unsigned int ABSDstReg = MI->getOperand(0).getReg();
7670 bool ABSSrcKIll = MI->getOperand(1).isKill();
7671 bool isThumb2 = Subtarget->isThumb2();
7672 MachineRegisterInfo &MRI = Fn->getRegInfo();
7673 // In Thumb mode S must not be specified if source register is the SP or
7674 // PC and if destination register is the SP, so restrict register class
7675 unsigned NewRsbDstReg =
7676 MRI.createVirtualRegister(isThumb2 ? &ARM::rGPRRegClass : &ARM::GPRRegClass);
7678 // Transfer the remainder of BB and its successor edges to sinkMBB.
7679 SinkBB->splice(SinkBB->begin(), BB,
7680 std::next(MachineBasicBlock::iterator(MI)), BB->end());
7681 SinkBB->transferSuccessorsAndUpdatePHIs(BB);
7683 BB->addSuccessor(RSBBB);
7684 BB->addSuccessor(SinkBB);
7686 // fall through to SinkMBB
7687 RSBBB->addSuccessor(SinkBB);
7689 // insert a cmp at the end of BB
7690 AddDefaultPred(BuildMI(BB, dl,
7691 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7692 .addReg(ABSSrcReg).addImm(0));
7694 // insert a bcc with opposite CC to ARMCC::MI at the end of BB
7696 TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)).addMBB(SinkBB)
7697 .addImm(ARMCC::getOppositeCondition(ARMCC::MI)).addReg(ARM::CPSR);
7699 // insert rsbri in RSBBB
7700 // Note: BCC and rsbri will be converted into predicated rsbmi
7701 // by if-conversion pass
7702 BuildMI(*RSBBB, RSBBB->begin(), dl,
7703 TII->get(isThumb2 ? ARM::t2RSBri : ARM::RSBri), NewRsbDstReg)
7704 .addReg(ABSSrcReg, ABSSrcKIll ? RegState::Kill : 0)
7705 .addImm(0).addImm((unsigned)ARMCC::AL).addReg(0).addReg(0);
7707 // insert PHI in SinkBB,
7708 // reuse ABSDstReg to not change uses of ABS instruction
7709 BuildMI(*SinkBB, SinkBB->begin(), dl,
7710 TII->get(ARM::PHI), ABSDstReg)
7711 .addReg(NewRsbDstReg).addMBB(RSBBB)
7712 .addReg(ABSSrcReg).addMBB(BB);
7714 // remove ABS instruction
7715 MI->eraseFromParent();
7717 // return last added BB
7720 case ARM::COPY_STRUCT_BYVAL_I32:
7722 return EmitStructByval(MI, BB);
7723 case ARM::WIN__CHKSTK:
7724 return EmitLowered__chkstk(MI, BB);
7728 void ARMTargetLowering::AdjustInstrPostInstrSelection(MachineInstr *MI,
7729 SDNode *Node) const {
7730 const MCInstrDesc *MCID = &MI->getDesc();
7731 // Adjust potentially 's' setting instructions after isel, i.e. ADC, SBC, RSB,
7732 // RSC. Coming out of isel, they have an implicit CPSR def, but the optional
7733 // operand is still set to noreg. If needed, set the optional operand's
7734 // register to CPSR, and remove the redundant implicit def.
7736 // e.g. ADCS (..., CPSR<imp-def>) -> ADC (... opt:CPSR<def>).
7738 // Rename pseudo opcodes.
7739 unsigned NewOpc = convertAddSubFlagsOpcode(MI->getOpcode());
7741 const ARMBaseInstrInfo *TII = Subtarget->getInstrInfo();
7742 MCID = &TII->get(NewOpc);
7744 assert(MCID->getNumOperands() == MI->getDesc().getNumOperands() + 1 &&
7745 "converted opcode should be the same except for cc_out");
7749 // Add the optional cc_out operand
7750 MI->addOperand(MachineOperand::CreateReg(0, /*isDef=*/true));
7752 unsigned ccOutIdx = MCID->getNumOperands() - 1;
7754 // Any ARM instruction that sets the 's' bit should specify an optional
7755 // "cc_out" operand in the last operand position.
7756 if (!MI->hasOptionalDef() || !MCID->OpInfo[ccOutIdx].isOptionalDef()) {
7757 assert(!NewOpc && "Optional cc_out operand required");
7760 // Look for an implicit def of CPSR added by MachineInstr ctor. Remove it
7761 // since we already have an optional CPSR def.
7762 bool definesCPSR = false;
7763 bool deadCPSR = false;
7764 for (unsigned i = MCID->getNumOperands(), e = MI->getNumOperands();
7766 const MachineOperand &MO = MI->getOperand(i);
7767 if (MO.isReg() && MO.isDef() && MO.getReg() == ARM::CPSR) {
7771 MI->RemoveOperand(i);
7776 assert(!NewOpc && "Optional cc_out operand required");
7779 assert(deadCPSR == !Node->hasAnyUseOfValue(1) && "inconsistent dead flag");
7781 assert(!MI->getOperand(ccOutIdx).getReg() &&
7782 "expect uninitialized optional cc_out operand");
7786 // If this instruction was defined with an optional CPSR def and its dag node
7787 // had a live implicit CPSR def, then activate the optional CPSR def.
7788 MachineOperand &MO = MI->getOperand(ccOutIdx);
7789 MO.setReg(ARM::CPSR);
7793 //===----------------------------------------------------------------------===//
7794 // ARM Optimization Hooks
7795 //===----------------------------------------------------------------------===//
7797 // Helper function that checks if N is a null or all ones constant.
7798 static inline bool isZeroOrAllOnes(SDValue N, bool AllOnes) {
7799 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N);
7802 return AllOnes ? C->isAllOnesValue() : C->isNullValue();
7805 // Return true if N is conditionally 0 or all ones.
7806 // Detects these expressions where cc is an i1 value:
7808 // (select cc 0, y) [AllOnes=0]
7809 // (select cc y, 0) [AllOnes=0]
7810 // (zext cc) [AllOnes=0]
7811 // (sext cc) [AllOnes=0/1]
7812 // (select cc -1, y) [AllOnes=1]
7813 // (select cc y, -1) [AllOnes=1]
7815 // Invert is set when N is the null/all ones constant when CC is false.
7816 // OtherOp is set to the alternative value of N.
7817 static bool isConditionalZeroOrAllOnes(SDNode *N, bool AllOnes,
7818 SDValue &CC, bool &Invert,
7820 SelectionDAG &DAG) {
7821 switch (N->getOpcode()) {
7822 default: return false;
7824 CC = N->getOperand(0);
7825 SDValue N1 = N->getOperand(1);
7826 SDValue N2 = N->getOperand(2);
7827 if (isZeroOrAllOnes(N1, AllOnes)) {
7832 if (isZeroOrAllOnes(N2, AllOnes)) {
7839 case ISD::ZERO_EXTEND:
7840 // (zext cc) can never be the all ones value.
7844 case ISD::SIGN_EXTEND: {
7846 EVT VT = N->getValueType(0);
7847 CC = N->getOperand(0);
7848 if (CC.getValueType() != MVT::i1)
7852 // When looking for an AllOnes constant, N is an sext, and the 'other'
7854 OtherOp = DAG.getConstant(0, dl, VT);
7855 else if (N->getOpcode() == ISD::ZERO_EXTEND)
7856 // When looking for a 0 constant, N can be zext or sext.
7857 OtherOp = DAG.getConstant(1, dl, VT);
7859 OtherOp = DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), dl,
7866 // Combine a constant select operand into its use:
7868 // (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
7869 // (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
7870 // (and (select cc, -1, c), x) -> (select cc, x, (and, x, c)) [AllOnes=1]
7871 // (or (select cc, 0, c), x) -> (select cc, x, (or, x, c))
7872 // (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c))
7874 // The transform is rejected if the select doesn't have a constant operand that
7875 // is null, or all ones when AllOnes is set.
7877 // Also recognize sext/zext from i1:
7879 // (add (zext cc), x) -> (select cc (add x, 1), x)
7880 // (add (sext cc), x) -> (select cc (add x, -1), x)
7882 // These transformations eventually create predicated instructions.
7884 // @param N The node to transform.
7885 // @param Slct The N operand that is a select.
7886 // @param OtherOp The other N operand (x above).
7887 // @param DCI Context.
7888 // @param AllOnes Require the select constant to be all ones instead of null.
7889 // @returns The new node, or SDValue() on failure.
7891 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
7892 TargetLowering::DAGCombinerInfo &DCI,
7893 bool AllOnes = false) {
7894 SelectionDAG &DAG = DCI.DAG;
7895 EVT VT = N->getValueType(0);
7896 SDValue NonConstantVal;
7899 if (!isConditionalZeroOrAllOnes(Slct.getNode(), AllOnes, CCOp, SwapSelectOps,
7900 NonConstantVal, DAG))
7903 // Slct is now know to be the desired identity constant when CC is true.
7904 SDValue TrueVal = OtherOp;
7905 SDValue FalseVal = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
7906 OtherOp, NonConstantVal);
7907 // Unless SwapSelectOps says CC should be false.
7909 std::swap(TrueVal, FalseVal);
7911 return DAG.getNode(ISD::SELECT, SDLoc(N), VT,
7912 CCOp, TrueVal, FalseVal);
7915 // Attempt combineSelectAndUse on each operand of a commutative operator N.
7917 SDValue combineSelectAndUseCommutative(SDNode *N, bool AllOnes,
7918 TargetLowering::DAGCombinerInfo &DCI) {
7919 SDValue N0 = N->getOperand(0);
7920 SDValue N1 = N->getOperand(1);
7921 if (N0.getNode()->hasOneUse()) {
7922 SDValue Result = combineSelectAndUse(N, N0, N1, DCI, AllOnes);
7923 if (Result.getNode())
7926 if (N1.getNode()->hasOneUse()) {
7927 SDValue Result = combineSelectAndUse(N, N1, N0, DCI, AllOnes);
7928 if (Result.getNode())
7934 // AddCombineToVPADDL- For pair-wise add on neon, use the vpaddl instruction
7935 // (only after legalization).
7936 static SDValue AddCombineToVPADDL(SDNode *N, SDValue N0, SDValue N1,
7937 TargetLowering::DAGCombinerInfo &DCI,
7938 const ARMSubtarget *Subtarget) {
7940 // Only perform optimization if after legalize, and if NEON is available. We
7941 // also expected both operands to be BUILD_VECTORs.
7942 if (DCI.isBeforeLegalize() || !Subtarget->hasNEON()
7943 || N0.getOpcode() != ISD::BUILD_VECTOR
7944 || N1.getOpcode() != ISD::BUILD_VECTOR)
7947 // Check output type since VPADDL operand elements can only be 8, 16, or 32.
7948 EVT VT = N->getValueType(0);
7949 if (!VT.isInteger() || VT.getVectorElementType() == MVT::i64)
7952 // Check that the vector operands are of the right form.
7953 // N0 and N1 are BUILD_VECTOR nodes with N number of EXTRACT_VECTOR
7954 // operands, where N is the size of the formed vector.
7955 // Each EXTRACT_VECTOR should have the same input vector and odd or even
7956 // index such that we have a pair wise add pattern.
7958 // Grab the vector that all EXTRACT_VECTOR nodes should be referencing.
7959 if (N0->getOperand(0)->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
7961 SDValue Vec = N0->getOperand(0)->getOperand(0);
7962 SDNode *V = Vec.getNode();
7963 unsigned nextIndex = 0;
7965 // For each operands to the ADD which are BUILD_VECTORs,
7966 // check to see if each of their operands are an EXTRACT_VECTOR with
7967 // the same vector and appropriate index.
7968 for (unsigned i = 0, e = N0->getNumOperands(); i != e; ++i) {
7969 if (N0->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT
7970 && N1->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
7972 SDValue ExtVec0 = N0->getOperand(i);
7973 SDValue ExtVec1 = N1->getOperand(i);
7975 // First operand is the vector, verify its the same.
7976 if (V != ExtVec0->getOperand(0).getNode() ||
7977 V != ExtVec1->getOperand(0).getNode())
7980 // Second is the constant, verify its correct.
7981 ConstantSDNode *C0 = dyn_cast<ConstantSDNode>(ExtVec0->getOperand(1));
7982 ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(ExtVec1->getOperand(1));
7984 // For the constant, we want to see all the even or all the odd.
7985 if (!C0 || !C1 || C0->getZExtValue() != nextIndex
7986 || C1->getZExtValue() != nextIndex+1)
7995 // Create VPADDL node.
7996 SelectionDAG &DAG = DCI.DAG;
7997 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8001 // Build operand list.
8002 SmallVector<SDValue, 8> Ops;
8003 Ops.push_back(DAG.getConstant(Intrinsic::arm_neon_vpaddls, dl,
8004 TLI.getPointerTy(DAG.getDataLayout())));
8006 // Input is the vector.
8009 // Get widened type and narrowed type.
8011 unsigned numElem = VT.getVectorNumElements();
8013 EVT inputLaneType = Vec.getValueType().getVectorElementType();
8014 switch (inputLaneType.getSimpleVT().SimpleTy) {
8015 case MVT::i8: widenType = MVT::getVectorVT(MVT::i16, numElem); break;
8016 case MVT::i16: widenType = MVT::getVectorVT(MVT::i32, numElem); break;
8017 case MVT::i32: widenType = MVT::getVectorVT(MVT::i64, numElem); break;
8019 llvm_unreachable("Invalid vector element type for padd optimization.");
8022 SDValue tmp = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, widenType, Ops);
8023 unsigned ExtOp = VT.bitsGT(tmp.getValueType()) ? ISD::ANY_EXTEND : ISD::TRUNCATE;
8024 return DAG.getNode(ExtOp, dl, VT, tmp);
8027 static SDValue findMUL_LOHI(SDValue V) {
8028 if (V->getOpcode() == ISD::UMUL_LOHI ||
8029 V->getOpcode() == ISD::SMUL_LOHI)
8034 static SDValue AddCombineTo64bitMLAL(SDNode *AddcNode,
8035 TargetLowering::DAGCombinerInfo &DCI,
8036 const ARMSubtarget *Subtarget) {
8038 if (Subtarget->isThumb1Only()) return SDValue();
8040 // Only perform the checks after legalize when the pattern is available.
8041 if (DCI.isBeforeLegalize()) return SDValue();
8043 // Look for multiply add opportunities.
8044 // The pattern is a ISD::UMUL_LOHI followed by two add nodes, where
8045 // each add nodes consumes a value from ISD::UMUL_LOHI and there is
8046 // a glue link from the first add to the second add.
8047 // If we find this pattern, we can replace the U/SMUL_LOHI, ADDC, and ADDE by
8048 // a S/UMLAL instruction.
8051 // / \ [no multiline comment]
8057 assert(AddcNode->getOpcode() == ISD::ADDC && "Expect an ADDC");
8058 SDValue AddcOp0 = AddcNode->getOperand(0);
8059 SDValue AddcOp1 = AddcNode->getOperand(1);
8061 // Check if the two operands are from the same mul_lohi node.
8062 if (AddcOp0.getNode() == AddcOp1.getNode())
8065 assert(AddcNode->getNumValues() == 2 &&
8066 AddcNode->getValueType(0) == MVT::i32 &&
8067 "Expect ADDC with two result values. First: i32");
8069 // Check that we have a glued ADDC node.
8070 if (AddcNode->getValueType(1) != MVT::Glue)
8073 // Check that the ADDC adds the low result of the S/UMUL_LOHI.
8074 if (AddcOp0->getOpcode() != ISD::UMUL_LOHI &&
8075 AddcOp0->getOpcode() != ISD::SMUL_LOHI &&
8076 AddcOp1->getOpcode() != ISD::UMUL_LOHI &&
8077 AddcOp1->getOpcode() != ISD::SMUL_LOHI)
8080 // Look for the glued ADDE.
8081 SDNode* AddeNode = AddcNode->getGluedUser();
8085 // Make sure it is really an ADDE.
8086 if (AddeNode->getOpcode() != ISD::ADDE)
8089 assert(AddeNode->getNumOperands() == 3 &&
8090 AddeNode->getOperand(2).getValueType() == MVT::Glue &&
8091 "ADDE node has the wrong inputs");
8093 // Check for the triangle shape.
8094 SDValue AddeOp0 = AddeNode->getOperand(0);
8095 SDValue AddeOp1 = AddeNode->getOperand(1);
8097 // Make sure that the ADDE operands are not coming from the same node.
8098 if (AddeOp0.getNode() == AddeOp1.getNode())
8101 // Find the MUL_LOHI node walking up ADDE's operands.
8102 bool IsLeftOperandMUL = false;
8103 SDValue MULOp = findMUL_LOHI(AddeOp0);
8104 if (MULOp == SDValue())
8105 MULOp = findMUL_LOHI(AddeOp1);
8107 IsLeftOperandMUL = true;
8108 if (MULOp == SDValue())
8111 // Figure out the right opcode.
8112 unsigned Opc = MULOp->getOpcode();
8113 unsigned FinalOpc = (Opc == ISD::SMUL_LOHI) ? ARMISD::SMLAL : ARMISD::UMLAL;
8115 // Figure out the high and low input values to the MLAL node.
8116 SDValue* HiAdd = nullptr;
8117 SDValue* LoMul = nullptr;
8118 SDValue* LowAdd = nullptr;
8120 // Ensure that ADDE is from high result of ISD::SMUL_LOHI.
8121 if ((AddeOp0 != MULOp.getValue(1)) && (AddeOp1 != MULOp.getValue(1)))
8124 if (IsLeftOperandMUL)
8130 // Ensure that LoMul and LowAdd are taken from correct ISD::SMUL_LOHI node
8131 // whose low result is fed to the ADDC we are checking.
8133 if (AddcOp0 == MULOp.getValue(0)) {
8137 if (AddcOp1 == MULOp.getValue(0)) {
8145 // Create the merged node.
8146 SelectionDAG &DAG = DCI.DAG;
8148 // Build operand list.
8149 SmallVector<SDValue, 8> Ops;
8150 Ops.push_back(LoMul->getOperand(0));
8151 Ops.push_back(LoMul->getOperand(1));
8152 Ops.push_back(*LowAdd);
8153 Ops.push_back(*HiAdd);
8155 SDValue MLALNode = DAG.getNode(FinalOpc, SDLoc(AddcNode),
8156 DAG.getVTList(MVT::i32, MVT::i32), Ops);
8158 // Replace the ADDs' nodes uses by the MLA node's values.
8159 SDValue HiMLALResult(MLALNode.getNode(), 1);
8160 DAG.ReplaceAllUsesOfValueWith(SDValue(AddeNode, 0), HiMLALResult);
8162 SDValue LoMLALResult(MLALNode.getNode(), 0);
8163 DAG.ReplaceAllUsesOfValueWith(SDValue(AddcNode, 0), LoMLALResult);
8165 // Return original node to notify the driver to stop replacing.
8166 SDValue resNode(AddcNode, 0);
8170 /// PerformADDCCombine - Target-specific dag combine transform from
8171 /// ISD::ADDC, ISD::ADDE, and ISD::MUL_LOHI to MLAL.
8172 static SDValue PerformADDCCombine(SDNode *N,
8173 TargetLowering::DAGCombinerInfo &DCI,
8174 const ARMSubtarget *Subtarget) {
8176 return AddCombineTo64bitMLAL(N, DCI, Subtarget);
8180 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
8181 /// operands N0 and N1. This is a helper for PerformADDCombine that is
8182 /// called with the default operands, and if that fails, with commuted
8184 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
8185 TargetLowering::DAGCombinerInfo &DCI,
8186 const ARMSubtarget *Subtarget){
8188 // Attempt to create vpaddl for this add.
8189 SDValue Result = AddCombineToVPADDL(N, N0, N1, DCI, Subtarget);
8190 if (Result.getNode())
8193 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
8194 if (N0.getNode()->hasOneUse()) {
8195 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
8196 if (Result.getNode()) return Result;
8201 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
8203 static SDValue PerformADDCombine(SDNode *N,
8204 TargetLowering::DAGCombinerInfo &DCI,
8205 const ARMSubtarget *Subtarget) {
8206 SDValue N0 = N->getOperand(0);
8207 SDValue N1 = N->getOperand(1);
8209 // First try with the default operand order.
8210 SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI, Subtarget);
8211 if (Result.getNode())
8214 // If that didn't work, try again with the operands commuted.
8215 return PerformADDCombineWithOperands(N, N1, N0, DCI, Subtarget);
8218 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
8220 static SDValue PerformSUBCombine(SDNode *N,
8221 TargetLowering::DAGCombinerInfo &DCI) {
8222 SDValue N0 = N->getOperand(0);
8223 SDValue N1 = N->getOperand(1);
8225 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
8226 if (N1.getNode()->hasOneUse()) {
8227 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
8228 if (Result.getNode()) return Result;
8234 /// PerformVMULCombine
8235 /// Distribute (A + B) * C to (A * C) + (B * C) to take advantage of the
8236 /// special multiplier accumulator forwarding.
8242 // However, for (A + B) * (A + B),
8249 static SDValue PerformVMULCombine(SDNode *N,
8250 TargetLowering::DAGCombinerInfo &DCI,
8251 const ARMSubtarget *Subtarget) {
8252 if (!Subtarget->hasVMLxForwarding())
8255 SelectionDAG &DAG = DCI.DAG;
8256 SDValue N0 = N->getOperand(0);
8257 SDValue N1 = N->getOperand(1);
8258 unsigned Opcode = N0.getOpcode();
8259 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
8260 Opcode != ISD::FADD && Opcode != ISD::FSUB) {
8261 Opcode = N1.getOpcode();
8262 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
8263 Opcode != ISD::FADD && Opcode != ISD::FSUB)
8271 EVT VT = N->getValueType(0);
8273 SDValue N00 = N0->getOperand(0);
8274 SDValue N01 = N0->getOperand(1);
8275 return DAG.getNode(Opcode, DL, VT,
8276 DAG.getNode(ISD::MUL, DL, VT, N00, N1),
8277 DAG.getNode(ISD::MUL, DL, VT, N01, N1));
8280 static SDValue PerformMULCombine(SDNode *N,
8281 TargetLowering::DAGCombinerInfo &DCI,
8282 const ARMSubtarget *Subtarget) {
8283 SelectionDAG &DAG = DCI.DAG;
8285 if (Subtarget->isThumb1Only())
8288 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
8291 EVT VT = N->getValueType(0);
8292 if (VT.is64BitVector() || VT.is128BitVector())
8293 return PerformVMULCombine(N, DCI, Subtarget);
8297 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
8301 int64_t MulAmt = C->getSExtValue();
8302 unsigned ShiftAmt = countTrailingZeros<uint64_t>(MulAmt);
8304 ShiftAmt = ShiftAmt & (32 - 1);
8305 SDValue V = N->getOperand(0);
8309 MulAmt >>= ShiftAmt;
8312 if (isPowerOf2_32(MulAmt - 1)) {
8313 // (mul x, 2^N + 1) => (add (shl x, N), x)
8314 Res = DAG.getNode(ISD::ADD, DL, VT,
8316 DAG.getNode(ISD::SHL, DL, VT,
8318 DAG.getConstant(Log2_32(MulAmt - 1), DL,
8320 } else if (isPowerOf2_32(MulAmt + 1)) {
8321 // (mul x, 2^N - 1) => (sub (shl x, N), x)
8322 Res = DAG.getNode(ISD::SUB, DL, VT,
8323 DAG.getNode(ISD::SHL, DL, VT,
8325 DAG.getConstant(Log2_32(MulAmt + 1), DL,
8331 uint64_t MulAmtAbs = -MulAmt;
8332 if (isPowerOf2_32(MulAmtAbs + 1)) {
8333 // (mul x, -(2^N - 1)) => (sub x, (shl x, N))
8334 Res = DAG.getNode(ISD::SUB, DL, VT,
8336 DAG.getNode(ISD::SHL, DL, VT,
8338 DAG.getConstant(Log2_32(MulAmtAbs + 1), DL,
8340 } else if (isPowerOf2_32(MulAmtAbs - 1)) {
8341 // (mul x, -(2^N + 1)) => - (add (shl x, N), x)
8342 Res = DAG.getNode(ISD::ADD, DL, VT,
8344 DAG.getNode(ISD::SHL, DL, VT,
8346 DAG.getConstant(Log2_32(MulAmtAbs - 1), DL,
8348 Res = DAG.getNode(ISD::SUB, DL, VT,
8349 DAG.getConstant(0, DL, MVT::i32), Res);
8356 Res = DAG.getNode(ISD::SHL, DL, VT,
8357 Res, DAG.getConstant(ShiftAmt, DL, MVT::i32));
8359 // Do not add new nodes to DAG combiner worklist.
8360 DCI.CombineTo(N, Res, false);
8364 static SDValue PerformANDCombine(SDNode *N,
8365 TargetLowering::DAGCombinerInfo &DCI,
8366 const ARMSubtarget *Subtarget) {
8368 // Attempt to use immediate-form VBIC
8369 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
8371 EVT VT = N->getValueType(0);
8372 SelectionDAG &DAG = DCI.DAG;
8374 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8377 APInt SplatBits, SplatUndef;
8378 unsigned SplatBitSize;
8381 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
8382 if (SplatBitSize <= 64) {
8384 SDValue Val = isNEONModifiedImm((~SplatBits).getZExtValue(),
8385 SplatUndef.getZExtValue(), SplatBitSize,
8386 DAG, dl, VbicVT, VT.is128BitVector(),
8388 if (Val.getNode()) {
8390 DAG.getNode(ISD::BITCAST, dl, VbicVT, N->getOperand(0));
8391 SDValue Vbic = DAG.getNode(ARMISD::VBICIMM, dl, VbicVT, Input, Val);
8392 return DAG.getNode(ISD::BITCAST, dl, VT, Vbic);
8397 if (!Subtarget->isThumb1Only()) {
8398 // fold (and (select cc, -1, c), x) -> (select cc, x, (and, x, c))
8399 SDValue Result = combineSelectAndUseCommutative(N, true, DCI);
8400 if (Result.getNode())
8407 /// PerformORCombine - Target-specific dag combine xforms for ISD::OR
8408 static SDValue PerformORCombine(SDNode *N,
8409 TargetLowering::DAGCombinerInfo &DCI,
8410 const ARMSubtarget *Subtarget) {
8411 // Attempt to use immediate-form VORR
8412 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
8414 EVT VT = N->getValueType(0);
8415 SelectionDAG &DAG = DCI.DAG;
8417 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8420 APInt SplatBits, SplatUndef;
8421 unsigned SplatBitSize;
8423 if (BVN && Subtarget->hasNEON() &&
8424 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
8425 if (SplatBitSize <= 64) {
8427 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
8428 SplatUndef.getZExtValue(), SplatBitSize,
8429 DAG, dl, VorrVT, VT.is128BitVector(),
8431 if (Val.getNode()) {
8433 DAG.getNode(ISD::BITCAST, dl, VorrVT, N->getOperand(0));
8434 SDValue Vorr = DAG.getNode(ARMISD::VORRIMM, dl, VorrVT, Input, Val);
8435 return DAG.getNode(ISD::BITCAST, dl, VT, Vorr);
8440 if (!Subtarget->isThumb1Only()) {
8441 // fold (or (select cc, 0, c), x) -> (select cc, x, (or, x, c))
8442 SDValue Result = combineSelectAndUseCommutative(N, false, DCI);
8443 if (Result.getNode())
8447 // The code below optimizes (or (and X, Y), Z).
8448 // The AND operand needs to have a single user to make these optimizations
8450 SDValue N0 = N->getOperand(0);
8451 if (N0.getOpcode() != ISD::AND || !N0.hasOneUse())
8453 SDValue N1 = N->getOperand(1);
8455 // (or (and B, A), (and C, ~A)) => (VBSL A, B, C) when A is a constant.
8456 if (Subtarget->hasNEON() && N1.getOpcode() == ISD::AND && VT.isVector() &&
8457 DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
8459 unsigned SplatBitSize;
8462 APInt SplatBits0, SplatBits1;
8463 BuildVectorSDNode *BVN0 = dyn_cast<BuildVectorSDNode>(N0->getOperand(1));
8464 BuildVectorSDNode *BVN1 = dyn_cast<BuildVectorSDNode>(N1->getOperand(1));
8465 // Ensure that the second operand of both ands are constants
8466 if (BVN0 && BVN0->isConstantSplat(SplatBits0, SplatUndef, SplatBitSize,
8467 HasAnyUndefs) && !HasAnyUndefs) {
8468 if (BVN1 && BVN1->isConstantSplat(SplatBits1, SplatUndef, SplatBitSize,
8469 HasAnyUndefs) && !HasAnyUndefs) {
8470 // Ensure that the bit width of the constants are the same and that
8471 // the splat arguments are logical inverses as per the pattern we
8472 // are trying to simplify.
8473 if (SplatBits0.getBitWidth() == SplatBits1.getBitWidth() &&
8474 SplatBits0 == ~SplatBits1) {
8475 // Canonicalize the vector type to make instruction selection
8477 EVT CanonicalVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
8478 SDValue Result = DAG.getNode(ARMISD::VBSL, dl, CanonicalVT,
8482 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
8488 // Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when
8491 // BFI is only available on V6T2+
8492 if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops())
8496 // 1) or (and A, mask), val => ARMbfi A, val, mask
8497 // iff (val & mask) == val
8499 // 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
8500 // 2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2)
8501 // && mask == ~mask2
8502 // 2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2)
8503 // && ~mask == mask2
8504 // (i.e., copy a bitfield value into another bitfield of the same width)
8509 SDValue N00 = N0.getOperand(0);
8511 // The value and the mask need to be constants so we can verify this is
8512 // actually a bitfield set. If the mask is 0xffff, we can do better
8513 // via a movt instruction, so don't use BFI in that case.
8514 SDValue MaskOp = N0.getOperand(1);
8515 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(MaskOp);
8518 unsigned Mask = MaskC->getZExtValue();
8522 // Case (1): or (and A, mask), val => ARMbfi A, val, mask
8523 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
8525 unsigned Val = N1C->getZExtValue();
8526 if ((Val & ~Mask) != Val)
8529 if (ARM::isBitFieldInvertedMask(Mask)) {
8530 Val >>= countTrailingZeros(~Mask);
8532 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00,
8533 DAG.getConstant(Val, DL, MVT::i32),
8534 DAG.getConstant(Mask, DL, MVT::i32));
8536 // Do not add new nodes to DAG combiner worklist.
8537 DCI.CombineTo(N, Res, false);
8540 } else if (N1.getOpcode() == ISD::AND) {
8541 // case (2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
8542 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
8545 unsigned Mask2 = N11C->getZExtValue();
8547 // Mask and ~Mask2 (or reverse) must be equivalent for the BFI pattern
8549 if (ARM::isBitFieldInvertedMask(Mask) &&
8551 // The pack halfword instruction works better for masks that fit it,
8552 // so use that when it's available.
8553 if (Subtarget->hasT2ExtractPack() &&
8554 (Mask == 0xffff || Mask == 0xffff0000))
8557 unsigned amt = countTrailingZeros(Mask2);
8558 Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0),
8559 DAG.getConstant(amt, DL, MVT::i32));
8560 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00, Res,
8561 DAG.getConstant(Mask, DL, MVT::i32));
8562 // Do not add new nodes to DAG combiner worklist.
8563 DCI.CombineTo(N, Res, false);
8565 } else if (ARM::isBitFieldInvertedMask(~Mask) &&
8567 // The pack halfword instruction works better for masks that fit it,
8568 // so use that when it's available.
8569 if (Subtarget->hasT2ExtractPack() &&
8570 (Mask2 == 0xffff || Mask2 == 0xffff0000))
8573 unsigned lsb = countTrailingZeros(Mask);
8574 Res = DAG.getNode(ISD::SRL, DL, VT, N00,
8575 DAG.getConstant(lsb, DL, MVT::i32));
8576 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res,
8577 DAG.getConstant(Mask2, DL, MVT::i32));
8578 // Do not add new nodes to DAG combiner worklist.
8579 DCI.CombineTo(N, Res, false);
8584 if (DAG.MaskedValueIsZero(N1, MaskC->getAPIntValue()) &&
8585 N00.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N00.getOperand(1)) &&
8586 ARM::isBitFieldInvertedMask(~Mask)) {
8587 // Case (3): or (and (shl A, #shamt), mask), B => ARMbfi B, A, ~mask
8588 // where lsb(mask) == #shamt and masked bits of B are known zero.
8589 SDValue ShAmt = N00.getOperand(1);
8590 unsigned ShAmtC = cast<ConstantSDNode>(ShAmt)->getZExtValue();
8591 unsigned LSB = countTrailingZeros(Mask);
8595 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1, N00.getOperand(0),
8596 DAG.getConstant(~Mask, DL, MVT::i32));
8598 // Do not add new nodes to DAG combiner worklist.
8599 DCI.CombineTo(N, Res, false);
8605 static SDValue PerformXORCombine(SDNode *N,
8606 TargetLowering::DAGCombinerInfo &DCI,
8607 const ARMSubtarget *Subtarget) {
8608 EVT VT = N->getValueType(0);
8609 SelectionDAG &DAG = DCI.DAG;
8611 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8614 if (!Subtarget->isThumb1Only()) {
8615 // fold (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c))
8616 SDValue Result = combineSelectAndUseCommutative(N, false, DCI);
8617 if (Result.getNode())
8624 /// PerformBFICombine - (bfi A, (and B, Mask1), Mask2) -> (bfi A, B, Mask2) iff
8625 /// the bits being cleared by the AND are not demanded by the BFI.
8626 static SDValue PerformBFICombine(SDNode *N,
8627 TargetLowering::DAGCombinerInfo &DCI) {
8628 SDValue N1 = N->getOperand(1);
8629 if (N1.getOpcode() == ISD::AND) {
8630 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
8633 unsigned InvMask = cast<ConstantSDNode>(N->getOperand(2))->getZExtValue();
8634 unsigned LSB = countTrailingZeros(~InvMask);
8635 unsigned Width = (32 - countLeadingZeros(~InvMask)) - LSB;
8637 static_cast<unsigned>(std::numeric_limits<unsigned>::digits) &&
8638 "undefined behavior");
8639 unsigned Mask = (1u << Width) - 1;
8640 unsigned Mask2 = N11C->getZExtValue();
8641 if ((Mask & (~Mask2)) == 0)
8642 return DCI.DAG.getNode(ARMISD::BFI, SDLoc(N), N->getValueType(0),
8643 N->getOperand(0), N1.getOperand(0),
8649 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
8650 /// ARMISD::VMOVRRD.
8651 static SDValue PerformVMOVRRDCombine(SDNode *N,
8652 TargetLowering::DAGCombinerInfo &DCI,
8653 const ARMSubtarget *Subtarget) {
8654 // vmovrrd(vmovdrr x, y) -> x,y
8655 SDValue InDouble = N->getOperand(0);
8656 if (InDouble.getOpcode() == ARMISD::VMOVDRR && !Subtarget->isFPOnlySP())
8657 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
8659 // vmovrrd(load f64) -> (load i32), (load i32)
8660 SDNode *InNode = InDouble.getNode();
8661 if (ISD::isNormalLoad(InNode) && InNode->hasOneUse() &&
8662 InNode->getValueType(0) == MVT::f64 &&
8663 InNode->getOperand(1).getOpcode() == ISD::FrameIndex &&
8664 !cast<LoadSDNode>(InNode)->isVolatile()) {
8665 // TODO: Should this be done for non-FrameIndex operands?
8666 LoadSDNode *LD = cast<LoadSDNode>(InNode);
8668 SelectionDAG &DAG = DCI.DAG;
8670 SDValue BasePtr = LD->getBasePtr();
8671 SDValue NewLD1 = DAG.getLoad(MVT::i32, DL, LD->getChain(), BasePtr,
8672 LD->getPointerInfo(), LD->isVolatile(),
8673 LD->isNonTemporal(), LD->isInvariant(),
8674 LD->getAlignment());
8676 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
8677 DAG.getConstant(4, DL, MVT::i32));
8678 SDValue NewLD2 = DAG.getLoad(MVT::i32, DL, NewLD1.getValue(1), OffsetPtr,
8679 LD->getPointerInfo(), LD->isVolatile(),
8680 LD->isNonTemporal(), LD->isInvariant(),
8681 std::min(4U, LD->getAlignment() / 2));
8683 DAG.ReplaceAllUsesOfValueWith(SDValue(LD, 1), NewLD2.getValue(1));
8684 if (DCI.DAG.getDataLayout().isBigEndian())
8685 std::swap (NewLD1, NewLD2);
8686 SDValue Result = DCI.CombineTo(N, NewLD1, NewLD2);
8693 /// PerformVMOVDRRCombine - Target-specific dag combine xforms for
8694 /// ARMISD::VMOVDRR. This is also used for BUILD_VECTORs with 2 operands.
8695 static SDValue PerformVMOVDRRCombine(SDNode *N, SelectionDAG &DAG) {
8696 // N=vmovrrd(X); vmovdrr(N:0, N:1) -> bit_convert(X)
8697 SDValue Op0 = N->getOperand(0);
8698 SDValue Op1 = N->getOperand(1);
8699 if (Op0.getOpcode() == ISD::BITCAST)
8700 Op0 = Op0.getOperand(0);
8701 if (Op1.getOpcode() == ISD::BITCAST)
8702 Op1 = Op1.getOperand(0);
8703 if (Op0.getOpcode() == ARMISD::VMOVRRD &&
8704 Op0.getNode() == Op1.getNode() &&
8705 Op0.getResNo() == 0 && Op1.getResNo() == 1)
8706 return DAG.getNode(ISD::BITCAST, SDLoc(N),
8707 N->getValueType(0), Op0.getOperand(0));
8711 /// hasNormalLoadOperand - Check if any of the operands of a BUILD_VECTOR node
8712 /// are normal, non-volatile loads. If so, it is profitable to bitcast an
8713 /// i64 vector to have f64 elements, since the value can then be loaded
8714 /// directly into a VFP register.
8715 static bool hasNormalLoadOperand(SDNode *N) {
8716 unsigned NumElts = N->getValueType(0).getVectorNumElements();
8717 for (unsigned i = 0; i < NumElts; ++i) {
8718 SDNode *Elt = N->getOperand(i).getNode();
8719 if (ISD::isNormalLoad(Elt) && !cast<LoadSDNode>(Elt)->isVolatile())
8725 /// PerformBUILD_VECTORCombine - Target-specific dag combine xforms for
8726 /// ISD::BUILD_VECTOR.
8727 static SDValue PerformBUILD_VECTORCombine(SDNode *N,
8728 TargetLowering::DAGCombinerInfo &DCI,
8729 const ARMSubtarget *Subtarget) {
8730 // build_vector(N=ARMISD::VMOVRRD(X), N:1) -> bit_convert(X):
8731 // VMOVRRD is introduced when legalizing i64 types. It forces the i64 value
8732 // into a pair of GPRs, which is fine when the value is used as a scalar,
8733 // but if the i64 value is converted to a vector, we need to undo the VMOVRRD.
8734 SelectionDAG &DAG = DCI.DAG;
8735 if (N->getNumOperands() == 2) {
8736 SDValue RV = PerformVMOVDRRCombine(N, DAG);
8741 // Load i64 elements as f64 values so that type legalization does not split
8742 // them up into i32 values.
8743 EVT VT = N->getValueType(0);
8744 if (VT.getVectorElementType() != MVT::i64 || !hasNormalLoadOperand(N))
8747 SmallVector<SDValue, 8> Ops;
8748 unsigned NumElts = VT.getVectorNumElements();
8749 for (unsigned i = 0; i < NumElts; ++i) {
8750 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(i));
8752 // Make the DAGCombiner fold the bitcast.
8753 DCI.AddToWorklist(V.getNode());
8755 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, NumElts);
8756 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, FloatVT, Ops);
8757 return DAG.getNode(ISD::BITCAST, dl, VT, BV);
8760 /// \brief Target-specific dag combine xforms for ARMISD::BUILD_VECTOR.
8762 PerformARMBUILD_VECTORCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
8763 // ARMISD::BUILD_VECTOR is introduced when legalizing ISD::BUILD_VECTOR.
8764 // At that time, we may have inserted bitcasts from integer to float.
8765 // If these bitcasts have survived DAGCombine, change the lowering of this
8766 // BUILD_VECTOR in something more vector friendly, i.e., that does not
8767 // force to use floating point types.
8769 // Make sure we can change the type of the vector.
8770 // This is possible iff:
8771 // 1. The vector is only used in a bitcast to a integer type. I.e.,
8772 // 1.1. Vector is used only once.
8773 // 1.2. Use is a bit convert to an integer type.
8774 // 2. The size of its operands are 32-bits (64-bits are not legal).
8775 EVT VT = N->getValueType(0);
8776 EVT EltVT = VT.getVectorElementType();
8778 // Check 1.1. and 2.
8779 if (EltVT.getSizeInBits() != 32 || !N->hasOneUse())
8782 // By construction, the input type must be float.
8783 assert(EltVT == MVT::f32 && "Unexpected type!");
8786 SDNode *Use = *N->use_begin();
8787 if (Use->getOpcode() != ISD::BITCAST ||
8788 Use->getValueType(0).isFloatingPoint())
8791 // Check profitability.
8792 // Model is, if more than half of the relevant operands are bitcast from
8793 // i32, turn the build_vector into a sequence of insert_vector_elt.
8794 // Relevant operands are everything that is not statically
8795 // (i.e., at compile time) bitcasted.
8796 unsigned NumOfBitCastedElts = 0;
8797 unsigned NumElts = VT.getVectorNumElements();
8798 unsigned NumOfRelevantElts = NumElts;
8799 for (unsigned Idx = 0; Idx < NumElts; ++Idx) {
8800 SDValue Elt = N->getOperand(Idx);
8801 if (Elt->getOpcode() == ISD::BITCAST) {
8802 // Assume only bit cast to i32 will go away.
8803 if (Elt->getOperand(0).getValueType() == MVT::i32)
8804 ++NumOfBitCastedElts;
8805 } else if (Elt.getOpcode() == ISD::UNDEF || isa<ConstantSDNode>(Elt))
8806 // Constants are statically casted, thus do not count them as
8807 // relevant operands.
8808 --NumOfRelevantElts;
8811 // Check if more than half of the elements require a non-free bitcast.
8812 if (NumOfBitCastedElts <= NumOfRelevantElts / 2)
8815 SelectionDAG &DAG = DCI.DAG;
8816 // Create the new vector type.
8817 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
8818 // Check if the type is legal.
8819 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8820 if (!TLI.isTypeLegal(VecVT))
8824 // ARMISD::BUILD_VECTOR E1, E2, ..., EN.
8825 // => BITCAST INSERT_VECTOR_ELT
8826 // (INSERT_VECTOR_ELT (...), (BITCAST EN-1), N-1),
8828 SDValue Vec = DAG.getUNDEF(VecVT);
8830 for (unsigned Idx = 0 ; Idx < NumElts; ++Idx) {
8831 SDValue V = N->getOperand(Idx);
8832 if (V.getOpcode() == ISD::UNDEF)
8834 if (V.getOpcode() == ISD::BITCAST &&
8835 V->getOperand(0).getValueType() == MVT::i32)
8836 // Fold obvious case.
8837 V = V.getOperand(0);
8839 V = DAG.getNode(ISD::BITCAST, SDLoc(V), MVT::i32, V);
8840 // Make the DAGCombiner fold the bitcasts.
8841 DCI.AddToWorklist(V.getNode());
8843 SDValue LaneIdx = DAG.getConstant(Idx, dl, MVT::i32);
8844 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VecVT, Vec, V, LaneIdx);
8846 Vec = DAG.getNode(ISD::BITCAST, dl, VT, Vec);
8847 // Make the DAGCombiner fold the bitcasts.
8848 DCI.AddToWorklist(Vec.getNode());
8852 /// PerformInsertEltCombine - Target-specific dag combine xforms for
8853 /// ISD::INSERT_VECTOR_ELT.
8854 static SDValue PerformInsertEltCombine(SDNode *N,
8855 TargetLowering::DAGCombinerInfo &DCI) {
8856 // Bitcast an i64 load inserted into a vector to f64.
8857 // Otherwise, the i64 value will be legalized to a pair of i32 values.
8858 EVT VT = N->getValueType(0);
8859 SDNode *Elt = N->getOperand(1).getNode();
8860 if (VT.getVectorElementType() != MVT::i64 ||
8861 !ISD::isNormalLoad(Elt) || cast<LoadSDNode>(Elt)->isVolatile())
8864 SelectionDAG &DAG = DCI.DAG;
8866 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
8867 VT.getVectorNumElements());
8868 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, N->getOperand(0));
8869 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(1));
8870 // Make the DAGCombiner fold the bitcasts.
8871 DCI.AddToWorklist(Vec.getNode());
8872 DCI.AddToWorklist(V.getNode());
8873 SDValue InsElt = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, FloatVT,
8874 Vec, V, N->getOperand(2));
8875 return DAG.getNode(ISD::BITCAST, dl, VT, InsElt);
8878 /// PerformVECTOR_SHUFFLECombine - Target-specific dag combine xforms for
8879 /// ISD::VECTOR_SHUFFLE.
8880 static SDValue PerformVECTOR_SHUFFLECombine(SDNode *N, SelectionDAG &DAG) {
8881 // The LLVM shufflevector instruction does not require the shuffle mask
8882 // length to match the operand vector length, but ISD::VECTOR_SHUFFLE does
8883 // have that requirement. When translating to ISD::VECTOR_SHUFFLE, if the
8884 // operands do not match the mask length, they are extended by concatenating
8885 // them with undef vectors. That is probably the right thing for other
8886 // targets, but for NEON it is better to concatenate two double-register
8887 // size vector operands into a single quad-register size vector. Do that
8888 // transformation here:
8889 // shuffle(concat(v1, undef), concat(v2, undef)) ->
8890 // shuffle(concat(v1, v2), undef)
8891 SDValue Op0 = N->getOperand(0);
8892 SDValue Op1 = N->getOperand(1);
8893 if (Op0.getOpcode() != ISD::CONCAT_VECTORS ||
8894 Op1.getOpcode() != ISD::CONCAT_VECTORS ||
8895 Op0.getNumOperands() != 2 ||
8896 Op1.getNumOperands() != 2)
8898 SDValue Concat0Op1 = Op0.getOperand(1);
8899 SDValue Concat1Op1 = Op1.getOperand(1);
8900 if (Concat0Op1.getOpcode() != ISD::UNDEF ||
8901 Concat1Op1.getOpcode() != ISD::UNDEF)
8903 // Skip the transformation if any of the types are illegal.
8904 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8905 EVT VT = N->getValueType(0);
8906 if (!TLI.isTypeLegal(VT) ||
8907 !TLI.isTypeLegal(Concat0Op1.getValueType()) ||
8908 !TLI.isTypeLegal(Concat1Op1.getValueType()))
8911 SDValue NewConcat = DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT,
8912 Op0.getOperand(0), Op1.getOperand(0));
8913 // Translate the shuffle mask.
8914 SmallVector<int, 16> NewMask;
8915 unsigned NumElts = VT.getVectorNumElements();
8916 unsigned HalfElts = NumElts/2;
8917 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
8918 for (unsigned n = 0; n < NumElts; ++n) {
8919 int MaskElt = SVN->getMaskElt(n);
8921 if (MaskElt < (int)HalfElts)
8923 else if (MaskElt >= (int)NumElts && MaskElt < (int)(NumElts + HalfElts))
8924 NewElt = HalfElts + MaskElt - NumElts;
8925 NewMask.push_back(NewElt);
8927 return DAG.getVectorShuffle(VT, SDLoc(N), NewConcat,
8928 DAG.getUNDEF(VT), NewMask.data());
8931 /// CombineBaseUpdate - Target-specific DAG combine function for VLDDUP,
8932 /// NEON load/store intrinsics, and generic vector load/stores, to merge
8933 /// base address updates.
8934 /// For generic load/stores, the memory type is assumed to be a vector.
8935 /// The caller is assumed to have checked legality.
8936 static SDValue CombineBaseUpdate(SDNode *N,
8937 TargetLowering::DAGCombinerInfo &DCI) {
8938 SelectionDAG &DAG = DCI.DAG;
8939 const bool isIntrinsic = (N->getOpcode() == ISD::INTRINSIC_VOID ||
8940 N->getOpcode() == ISD::INTRINSIC_W_CHAIN);
8941 const bool isStore = N->getOpcode() == ISD::STORE;
8942 const unsigned AddrOpIdx = ((isIntrinsic || isStore) ? 2 : 1);
8943 SDValue Addr = N->getOperand(AddrOpIdx);
8944 MemSDNode *MemN = cast<MemSDNode>(N);
8947 // Search for a use of the address operand that is an increment.
8948 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(),
8949 UE = Addr.getNode()->use_end(); UI != UE; ++UI) {
8951 if (User->getOpcode() != ISD::ADD ||
8952 UI.getUse().getResNo() != Addr.getResNo())
8955 // Check that the add is independent of the load/store. Otherwise, folding
8956 // it would create a cycle.
8957 if (User->isPredecessorOf(N) || N->isPredecessorOf(User))
8960 // Find the new opcode for the updating load/store.
8961 bool isLoadOp = true;
8962 bool isLaneOp = false;
8963 unsigned NewOpc = 0;
8964 unsigned NumVecs = 0;
8966 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
8968 default: llvm_unreachable("unexpected intrinsic for Neon base update");
8969 case Intrinsic::arm_neon_vld1: NewOpc = ARMISD::VLD1_UPD;
8971 case Intrinsic::arm_neon_vld2: NewOpc = ARMISD::VLD2_UPD;
8973 case Intrinsic::arm_neon_vld3: NewOpc = ARMISD::VLD3_UPD;
8975 case Intrinsic::arm_neon_vld4: NewOpc = ARMISD::VLD4_UPD;
8977 case Intrinsic::arm_neon_vld2lane: NewOpc = ARMISD::VLD2LN_UPD;
8978 NumVecs = 2; isLaneOp = true; break;
8979 case Intrinsic::arm_neon_vld3lane: NewOpc = ARMISD::VLD3LN_UPD;
8980 NumVecs = 3; isLaneOp = true; break;
8981 case Intrinsic::arm_neon_vld4lane: NewOpc = ARMISD::VLD4LN_UPD;
8982 NumVecs = 4; isLaneOp = true; break;
8983 case Intrinsic::arm_neon_vst1: NewOpc = ARMISD::VST1_UPD;
8984 NumVecs = 1; isLoadOp = false; break;
8985 case Intrinsic::arm_neon_vst2: NewOpc = ARMISD::VST2_UPD;
8986 NumVecs = 2; isLoadOp = false; break;
8987 case Intrinsic::arm_neon_vst3: NewOpc = ARMISD::VST3_UPD;
8988 NumVecs = 3; isLoadOp = false; break;
8989 case Intrinsic::arm_neon_vst4: NewOpc = ARMISD::VST4_UPD;
8990 NumVecs = 4; isLoadOp = false; break;
8991 case Intrinsic::arm_neon_vst2lane: NewOpc = ARMISD::VST2LN_UPD;
8992 NumVecs = 2; isLoadOp = false; isLaneOp = true; break;
8993 case Intrinsic::arm_neon_vst3lane: NewOpc = ARMISD::VST3LN_UPD;
8994 NumVecs = 3; isLoadOp = false; isLaneOp = true; break;
8995 case Intrinsic::arm_neon_vst4lane: NewOpc = ARMISD::VST4LN_UPD;
8996 NumVecs = 4; isLoadOp = false; isLaneOp = true; break;
9000 switch (N->getOpcode()) {
9001 default: llvm_unreachable("unexpected opcode for Neon base update");
9002 case ARMISD::VLD2DUP: NewOpc = ARMISD::VLD2DUP_UPD; NumVecs = 2; break;
9003 case ARMISD::VLD3DUP: NewOpc = ARMISD::VLD3DUP_UPD; NumVecs = 3; break;
9004 case ARMISD::VLD4DUP: NewOpc = ARMISD::VLD4DUP_UPD; NumVecs = 4; break;
9005 case ISD::LOAD: NewOpc = ARMISD::VLD1_UPD;
9006 NumVecs = 1; isLaneOp = false; break;
9007 case ISD::STORE: NewOpc = ARMISD::VST1_UPD;
9008 NumVecs = 1; isLaneOp = false; isLoadOp = false; break;
9012 // Find the size of memory referenced by the load/store.
9015 VecTy = N->getValueType(0);
9016 } else if (isIntrinsic) {
9017 VecTy = N->getOperand(AddrOpIdx+1).getValueType();
9019 assert(isStore && "Node has to be a load, a store, or an intrinsic!");
9020 VecTy = N->getOperand(1).getValueType();
9023 unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8;
9025 NumBytes /= VecTy.getVectorNumElements();
9027 // If the increment is a constant, it must match the memory ref size.
9028 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
9029 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
9030 uint64_t IncVal = CInc->getZExtValue();
9031 if (IncVal != NumBytes)
9033 } else if (NumBytes >= 3 * 16) {
9034 // VLD3/4 and VST3/4 for 128-bit vectors are implemented with two
9035 // separate instructions that make it harder to use a non-constant update.
9039 // OK, we found an ADD we can fold into the base update.
9040 // Now, create a _UPD node, taking care of not breaking alignment.
9042 EVT AlignedVecTy = VecTy;
9043 unsigned Alignment = MemN->getAlignment();
9045 // If this is a less-than-standard-aligned load/store, change the type to
9046 // match the standard alignment.
9047 // The alignment is overlooked when selecting _UPD variants; and it's
9048 // easier to introduce bitcasts here than fix that.
9049 // There are 3 ways to get to this base-update combine:
9050 // - intrinsics: they are assumed to be properly aligned (to the standard
9051 // alignment of the memory type), so we don't need to do anything.
9052 // - ARMISD::VLDx nodes: they are only generated from the aforementioned
9053 // intrinsics, so, likewise, there's nothing to do.
9054 // - generic load/store instructions: the alignment is specified as an
9055 // explicit operand, rather than implicitly as the standard alignment
9056 // of the memory type (like the intrisics). We need to change the
9057 // memory type to match the explicit alignment. That way, we don't
9058 // generate non-standard-aligned ARMISD::VLDx nodes.
9059 if (isa<LSBaseSDNode>(N)) {
9062 if (Alignment < VecTy.getScalarSizeInBits() / 8) {
9063 MVT EltTy = MVT::getIntegerVT(Alignment * 8);
9064 assert(NumVecs == 1 && "Unexpected multi-element generic load/store.");
9065 assert(!isLaneOp && "Unexpected generic load/store lane.");
9066 unsigned NumElts = NumBytes / (EltTy.getSizeInBits() / 8);
9067 AlignedVecTy = MVT::getVectorVT(EltTy, NumElts);
9069 // Don't set an explicit alignment on regular load/stores that we want
9070 // to transform to VLD/VST 1_UPD nodes.
9071 // This matches the behavior of regular load/stores, which only get an
9072 // explicit alignment if the MMO alignment is larger than the standard
9073 // alignment of the memory type.
9074 // Intrinsics, however, always get an explicit alignment, set to the
9075 // alignment of the MMO.
9079 // Create the new updating load/store node.
9080 // First, create an SDVTList for the new updating node's results.
9082 unsigned NumResultVecs = (isLoadOp ? NumVecs : 0);
9084 for (n = 0; n < NumResultVecs; ++n)
9085 Tys[n] = AlignedVecTy;
9086 Tys[n++] = MVT::i32;
9087 Tys[n] = MVT::Other;
9088 SDVTList SDTys = DAG.getVTList(makeArrayRef(Tys, NumResultVecs+2));
9090 // Then, gather the new node's operands.
9091 SmallVector<SDValue, 8> Ops;
9092 Ops.push_back(N->getOperand(0)); // incoming chain
9093 Ops.push_back(N->getOperand(AddrOpIdx));
9096 if (StoreSDNode *StN = dyn_cast<StoreSDNode>(N)) {
9097 // Try to match the intrinsic's signature
9098 Ops.push_back(StN->getValue());
9100 // Loads (and of course intrinsics) match the intrinsics' signature,
9101 // so just add all but the alignment operand.
9102 for (unsigned i = AddrOpIdx + 1; i < N->getNumOperands() - 1; ++i)
9103 Ops.push_back(N->getOperand(i));
9106 // For all node types, the alignment operand is always the last one.
9107 Ops.push_back(DAG.getConstant(Alignment, dl, MVT::i32));
9109 // If this is a non-standard-aligned STORE, the penultimate operand is the
9110 // stored value. Bitcast it to the aligned type.
9111 if (AlignedVecTy != VecTy && N->getOpcode() == ISD::STORE) {
9112 SDValue &StVal = Ops[Ops.size()-2];
9113 StVal = DAG.getNode(ISD::BITCAST, dl, AlignedVecTy, StVal);
9116 SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, dl, SDTys,
9118 MemN->getMemOperand());
9121 SmallVector<SDValue, 5> NewResults;
9122 for (unsigned i = 0; i < NumResultVecs; ++i)
9123 NewResults.push_back(SDValue(UpdN.getNode(), i));
9125 // If this is an non-standard-aligned LOAD, the first result is the loaded
9126 // value. Bitcast it to the expected result type.
9127 if (AlignedVecTy != VecTy && N->getOpcode() == ISD::LOAD) {
9128 SDValue &LdVal = NewResults[0];
9129 LdVal = DAG.getNode(ISD::BITCAST, dl, VecTy, LdVal);
9132 NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs+1)); // chain
9133 DCI.CombineTo(N, NewResults);
9134 DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs));
9141 static SDValue PerformVLDCombine(SDNode *N,
9142 TargetLowering::DAGCombinerInfo &DCI) {
9143 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
9146 return CombineBaseUpdate(N, DCI);
9149 /// CombineVLDDUP - For a VDUPLANE node N, check if its source operand is a
9150 /// vldN-lane (N > 1) intrinsic, and if all the other uses of that intrinsic
9151 /// are also VDUPLANEs. If so, combine them to a vldN-dup operation and
9153 static bool CombineVLDDUP(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
9154 SelectionDAG &DAG = DCI.DAG;
9155 EVT VT = N->getValueType(0);
9156 // vldN-dup instructions only support 64-bit vectors for N > 1.
9157 if (!VT.is64BitVector())
9160 // Check if the VDUPLANE operand is a vldN-dup intrinsic.
9161 SDNode *VLD = N->getOperand(0).getNode();
9162 if (VLD->getOpcode() != ISD::INTRINSIC_W_CHAIN)
9164 unsigned NumVecs = 0;
9165 unsigned NewOpc = 0;
9166 unsigned IntNo = cast<ConstantSDNode>(VLD->getOperand(1))->getZExtValue();
9167 if (IntNo == Intrinsic::arm_neon_vld2lane) {
9169 NewOpc = ARMISD::VLD2DUP;
9170 } else if (IntNo == Intrinsic::arm_neon_vld3lane) {
9172 NewOpc = ARMISD::VLD3DUP;
9173 } else if (IntNo == Intrinsic::arm_neon_vld4lane) {
9175 NewOpc = ARMISD::VLD4DUP;
9180 // First check that all the vldN-lane uses are VDUPLANEs and that the lane
9181 // numbers match the load.
9182 unsigned VLDLaneNo =
9183 cast<ConstantSDNode>(VLD->getOperand(NumVecs+3))->getZExtValue();
9184 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
9186 // Ignore uses of the chain result.
9187 if (UI.getUse().getResNo() == NumVecs)
9190 if (User->getOpcode() != ARMISD::VDUPLANE ||
9191 VLDLaneNo != cast<ConstantSDNode>(User->getOperand(1))->getZExtValue())
9195 // Create the vldN-dup node.
9198 for (n = 0; n < NumVecs; ++n)
9200 Tys[n] = MVT::Other;
9201 SDVTList SDTys = DAG.getVTList(makeArrayRef(Tys, NumVecs+1));
9202 SDValue Ops[] = { VLD->getOperand(0), VLD->getOperand(2) };
9203 MemIntrinsicSDNode *VLDMemInt = cast<MemIntrinsicSDNode>(VLD);
9204 SDValue VLDDup = DAG.getMemIntrinsicNode(NewOpc, SDLoc(VLD), SDTys,
9205 Ops, VLDMemInt->getMemoryVT(),
9206 VLDMemInt->getMemOperand());
9209 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
9211 unsigned ResNo = UI.getUse().getResNo();
9212 // Ignore uses of the chain result.
9213 if (ResNo == NumVecs)
9216 DCI.CombineTo(User, SDValue(VLDDup.getNode(), ResNo));
9219 // Now the vldN-lane intrinsic is dead except for its chain result.
9220 // Update uses of the chain.
9221 std::vector<SDValue> VLDDupResults;
9222 for (unsigned n = 0; n < NumVecs; ++n)
9223 VLDDupResults.push_back(SDValue(VLDDup.getNode(), n));
9224 VLDDupResults.push_back(SDValue(VLDDup.getNode(), NumVecs));
9225 DCI.CombineTo(VLD, VLDDupResults);
9230 /// PerformVDUPLANECombine - Target-specific dag combine xforms for
9231 /// ARMISD::VDUPLANE.
9232 static SDValue PerformVDUPLANECombine(SDNode *N,
9233 TargetLowering::DAGCombinerInfo &DCI) {
9234 SDValue Op = N->getOperand(0);
9236 // If the source is a vldN-lane (N > 1) intrinsic, and all the other uses
9237 // of that intrinsic are also VDUPLANEs, combine them to a vldN-dup operation.
9238 if (CombineVLDDUP(N, DCI))
9239 return SDValue(N, 0);
9241 // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is
9242 // redundant. Ignore bit_converts for now; element sizes are checked below.
9243 while (Op.getOpcode() == ISD::BITCAST)
9244 Op = Op.getOperand(0);
9245 if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM)
9248 // Make sure the VMOV element size is not bigger than the VDUPLANE elements.
9249 unsigned EltSize = Op.getValueType().getVectorElementType().getSizeInBits();
9250 // The canonical VMOV for a zero vector uses a 32-bit element size.
9251 unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
9253 if (ARM_AM::decodeNEONModImm(Imm, EltBits) == 0)
9255 EVT VT = N->getValueType(0);
9256 if (EltSize > VT.getVectorElementType().getSizeInBits())
9259 return DCI.DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Op);
9262 static SDValue PerformLOADCombine(SDNode *N,
9263 TargetLowering::DAGCombinerInfo &DCI) {
9264 EVT VT = N->getValueType(0);
9266 // If this is a legal vector load, try to combine it into a VLD1_UPD.
9267 if (ISD::isNormalLoad(N) && VT.isVector() &&
9268 DCI.DAG.getTargetLoweringInfo().isTypeLegal(VT))
9269 return CombineBaseUpdate(N, DCI);
9274 /// PerformSTORECombine - Target-specific dag combine xforms for
9276 static SDValue PerformSTORECombine(SDNode *N,
9277 TargetLowering::DAGCombinerInfo &DCI) {
9278 StoreSDNode *St = cast<StoreSDNode>(N);
9279 if (St->isVolatile())
9282 // Optimize trunc store (of multiple scalars) to shuffle and store. First,
9283 // pack all of the elements in one place. Next, store to memory in fewer
9285 SDValue StVal = St->getValue();
9286 EVT VT = StVal.getValueType();
9287 if (St->isTruncatingStore() && VT.isVector()) {
9288 SelectionDAG &DAG = DCI.DAG;
9289 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9290 EVT StVT = St->getMemoryVT();
9291 unsigned NumElems = VT.getVectorNumElements();
9292 assert(StVT != VT && "Cannot truncate to the same type");
9293 unsigned FromEltSz = VT.getVectorElementType().getSizeInBits();
9294 unsigned ToEltSz = StVT.getVectorElementType().getSizeInBits();
9296 // From, To sizes and ElemCount must be pow of two
9297 if (!isPowerOf2_32(NumElems * FromEltSz * ToEltSz)) return SDValue();
9299 // We are going to use the original vector elt for storing.
9300 // Accumulated smaller vector elements must be a multiple of the store size.
9301 if (0 != (NumElems * FromEltSz) % ToEltSz) return SDValue();
9303 unsigned SizeRatio = FromEltSz / ToEltSz;
9304 assert(SizeRatio * NumElems * ToEltSz == VT.getSizeInBits());
9306 // Create a type on which we perform the shuffle.
9307 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(), StVT.getScalarType(),
9308 NumElems*SizeRatio);
9309 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
9312 SDValue WideVec = DAG.getNode(ISD::BITCAST, DL, WideVecVT, StVal);
9313 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
9314 for (unsigned i = 0; i < NumElems; ++i)
9315 ShuffleVec[i] = DAG.getDataLayout().isBigEndian()
9316 ? (i + 1) * SizeRatio - 1
9319 // Can't shuffle using an illegal type.
9320 if (!TLI.isTypeLegal(WideVecVT)) return SDValue();
9322 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, DL, WideVec,
9323 DAG.getUNDEF(WideVec.getValueType()),
9325 // At this point all of the data is stored at the bottom of the
9326 // register. We now need to save it to mem.
9328 // Find the largest store unit
9329 MVT StoreType = MVT::i8;
9330 for (MVT Tp : MVT::integer_valuetypes()) {
9331 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToEltSz)
9334 // Didn't find a legal store type.
9335 if (!TLI.isTypeLegal(StoreType))
9338 // Bitcast the original vector into a vector of store-size units
9339 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
9340 StoreType, VT.getSizeInBits()/EVT(StoreType).getSizeInBits());
9341 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
9342 SDValue ShuffWide = DAG.getNode(ISD::BITCAST, DL, StoreVecVT, Shuff);
9343 SmallVector<SDValue, 8> Chains;
9344 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits() / 8, DL,
9345 TLI.getPointerTy(DAG.getDataLayout()));
9346 SDValue BasePtr = St->getBasePtr();
9348 // Perform one or more big stores into memory.
9349 unsigned E = (ToEltSz*NumElems)/StoreType.getSizeInBits();
9350 for (unsigned I = 0; I < E; I++) {
9351 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
9352 StoreType, ShuffWide,
9353 DAG.getIntPtrConstant(I, DL));
9354 SDValue Ch = DAG.getStore(St->getChain(), DL, SubVec, BasePtr,
9355 St->getPointerInfo(), St->isVolatile(),
9356 St->isNonTemporal(), St->getAlignment());
9357 BasePtr = DAG.getNode(ISD::ADD, DL, BasePtr.getValueType(), BasePtr,
9359 Chains.push_back(Ch);
9361 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
9364 if (!ISD::isNormalStore(St))
9367 // Split a store of a VMOVDRR into two integer stores to avoid mixing NEON and
9368 // ARM stores of arguments in the same cache line.
9369 if (StVal.getNode()->getOpcode() == ARMISD::VMOVDRR &&
9370 StVal.getNode()->hasOneUse()) {
9371 SelectionDAG &DAG = DCI.DAG;
9372 bool isBigEndian = DAG.getDataLayout().isBigEndian();
9374 SDValue BasePtr = St->getBasePtr();
9375 SDValue NewST1 = DAG.getStore(St->getChain(), DL,
9376 StVal.getNode()->getOperand(isBigEndian ? 1 : 0 ),
9377 BasePtr, St->getPointerInfo(), St->isVolatile(),
9378 St->isNonTemporal(), St->getAlignment());
9380 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
9381 DAG.getConstant(4, DL, MVT::i32));
9382 return DAG.getStore(NewST1.getValue(0), DL,
9383 StVal.getNode()->getOperand(isBigEndian ? 0 : 1),
9384 OffsetPtr, St->getPointerInfo(), St->isVolatile(),
9385 St->isNonTemporal(),
9386 std::min(4U, St->getAlignment() / 2));
9389 if (StVal.getValueType() == MVT::i64 &&
9390 StVal.getNode()->getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
9392 // Bitcast an i64 store extracted from a vector to f64.
9393 // Otherwise, the i64 value will be legalized to a pair of i32 values.
9394 SelectionDAG &DAG = DCI.DAG;
9396 SDValue IntVec = StVal.getOperand(0);
9397 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
9398 IntVec.getValueType().getVectorNumElements());
9399 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, IntVec);
9400 SDValue ExtElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
9401 Vec, StVal.getOperand(1));
9403 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ExtElt);
9404 // Make the DAGCombiner fold the bitcasts.
9405 DCI.AddToWorklist(Vec.getNode());
9406 DCI.AddToWorklist(ExtElt.getNode());
9407 DCI.AddToWorklist(V.getNode());
9408 return DAG.getStore(St->getChain(), dl, V, St->getBasePtr(),
9409 St->getPointerInfo(), St->isVolatile(),
9410 St->isNonTemporal(), St->getAlignment(),
9414 // If this is a legal vector store, try to combine it into a VST1_UPD.
9415 if (ISD::isNormalStore(N) && VT.isVector() &&
9416 DCI.DAG.getTargetLoweringInfo().isTypeLegal(VT))
9417 return CombineBaseUpdate(N, DCI);
9422 // isConstVecPow2 - Return true if each vector element is a power of 2, all
9423 // elements are the same constant, C, and Log2(C) ranges from 1 to 32.
9424 static bool isConstVecPow2(SDValue ConstVec, bool isSigned, uint64_t &C)
9428 for (unsigned I = 0, E = ConstVec.getValueType().getVectorNumElements();
9430 ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(ConstVec.getOperand(I));
9435 APFloat APF = C->getValueAPF();
9436 if (APF.convertToInteger(&cN, 64, isSigned, APFloat::rmTowardZero, &isExact)
9437 != APFloat::opOK || !isExact)
9440 c0 = (I == 0) ? cN : c0;
9441 if (!isPowerOf2_64(cN) || c0 != cN || Log2_64(c0) < 1 || Log2_64(c0) > 32)
9448 /// PerformVCVTCombine - VCVT (floating-point to fixed-point, Advanced SIMD)
9449 /// can replace combinations of VMUL and VCVT (floating-point to integer)
9450 /// when the VMUL has a constant operand that is a power of 2.
9452 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
9453 /// vmul.f32 d16, d17, d16
9454 /// vcvt.s32.f32 d16, d16
9456 /// vcvt.s32.f32 d16, d16, #3
9457 static SDValue PerformVCVTCombine(SDNode *N,
9458 TargetLowering::DAGCombinerInfo &DCI,
9459 const ARMSubtarget *Subtarget) {
9460 SelectionDAG &DAG = DCI.DAG;
9461 SDValue Op = N->getOperand(0);
9463 if (!Subtarget->hasNEON() || !Op.getValueType().isVector() ||
9464 Op.getOpcode() != ISD::FMUL)
9468 SDValue N0 = Op->getOperand(0);
9469 SDValue ConstVec = Op->getOperand(1);
9470 bool isSigned = N->getOpcode() == ISD::FP_TO_SINT;
9472 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
9473 !isConstVecPow2(ConstVec, isSigned, C))
9476 MVT FloatTy = Op.getSimpleValueType().getVectorElementType();
9477 MVT IntTy = N->getSimpleValueType(0).getVectorElementType();
9478 unsigned NumLanes = Op.getValueType().getVectorNumElements();
9479 if (FloatTy.getSizeInBits() != 32 || IntTy.getSizeInBits() > 32 ||
9481 // These instructions only exist converting from f32 to i32. We can handle
9482 // smaller integers by generating an extra truncate, but larger ones would
9483 // be lossy. We also can't handle more then 4 lanes, since these intructions
9484 // only support v2i32/v4i32 types.
9489 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfp2fxs :
9490 Intrinsic::arm_neon_vcvtfp2fxu;
9491 SDValue FixConv = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl,
9492 NumLanes == 2 ? MVT::v2i32 : MVT::v4i32,
9493 DAG.getConstant(IntrinsicOpcode, dl, MVT::i32),
9495 DAG.getConstant(Log2_64(C), dl, MVT::i32));
9497 if (IntTy.getSizeInBits() < FloatTy.getSizeInBits())
9498 FixConv = DAG.getNode(ISD::TRUNCATE, dl, N->getValueType(0), FixConv);
9503 /// PerformVDIVCombine - VCVT (fixed-point to floating-point, Advanced SIMD)
9504 /// can replace combinations of VCVT (integer to floating-point) and VDIV
9505 /// when the VDIV has a constant operand that is a power of 2.
9507 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
9508 /// vcvt.f32.s32 d16, d16
9509 /// vdiv.f32 d16, d17, d16
9511 /// vcvt.f32.s32 d16, d16, #3
9512 static SDValue PerformVDIVCombine(SDNode *N,
9513 TargetLowering::DAGCombinerInfo &DCI,
9514 const ARMSubtarget *Subtarget) {
9515 SelectionDAG &DAG = DCI.DAG;
9516 SDValue Op = N->getOperand(0);
9517 unsigned OpOpcode = Op.getNode()->getOpcode();
9519 if (!Subtarget->hasNEON() || !N->getValueType(0).isVector() ||
9520 (OpOpcode != ISD::SINT_TO_FP && OpOpcode != ISD::UINT_TO_FP))
9524 SDValue ConstVec = N->getOperand(1);
9525 bool isSigned = OpOpcode == ISD::SINT_TO_FP;
9527 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
9528 !isConstVecPow2(ConstVec, isSigned, C))
9531 MVT FloatTy = N->getSimpleValueType(0).getVectorElementType();
9532 MVT IntTy = Op.getOperand(0).getSimpleValueType().getVectorElementType();
9533 if (FloatTy.getSizeInBits() != 32 || IntTy.getSizeInBits() > 32) {
9534 // These instructions only exist converting from i32 to f32. We can handle
9535 // smaller integers by generating an extra extend, but larger ones would
9541 SDValue ConvInput = Op.getOperand(0);
9542 unsigned NumLanes = Op.getValueType().getVectorNumElements();
9543 if (IntTy.getSizeInBits() < FloatTy.getSizeInBits())
9544 ConvInput = DAG.getNode(isSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND,
9545 dl, NumLanes == 2 ? MVT::v2i32 : MVT::v4i32,
9548 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfxs2fp :
9549 Intrinsic::arm_neon_vcvtfxu2fp;
9550 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl,
9552 DAG.getConstant(IntrinsicOpcode, dl, MVT::i32),
9553 ConvInput, DAG.getConstant(Log2_64(C), dl, MVT::i32));
9556 /// Getvshiftimm - Check if this is a valid build_vector for the immediate
9557 /// operand of a vector shift operation, where all the elements of the
9558 /// build_vector must have the same constant integer value.
9559 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
9560 // Ignore bit_converts.
9561 while (Op.getOpcode() == ISD::BITCAST)
9562 Op = Op.getOperand(0);
9563 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
9564 APInt SplatBits, SplatUndef;
9565 unsigned SplatBitSize;
9567 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
9568 HasAnyUndefs, ElementBits) ||
9569 SplatBitSize > ElementBits)
9571 Cnt = SplatBits.getSExtValue();
9575 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
9576 /// operand of a vector shift left operation. That value must be in the range:
9577 /// 0 <= Value < ElementBits for a left shift; or
9578 /// 0 <= Value <= ElementBits for a long left shift.
9579 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
9580 assert(VT.isVector() && "vector shift count is not a vector type");
9581 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
9582 if (! getVShiftImm(Op, ElementBits, Cnt))
9584 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
9587 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
9588 /// operand of a vector shift right operation. For a shift opcode, the value
9589 /// is positive, but for an intrinsic the value count must be negative. The
9590 /// absolute value must be in the range:
9591 /// 1 <= |Value| <= ElementBits for a right shift; or
9592 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
9593 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
9595 assert(VT.isVector() && "vector shift count is not a vector type");
9596 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
9597 if (! getVShiftImm(Op, ElementBits, Cnt))
9601 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
9604 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
9605 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
9606 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
9609 // Don't do anything for most intrinsics.
9612 // Vector shifts: check for immediate versions and lower them.
9613 // Note: This is done during DAG combining instead of DAG legalizing because
9614 // the build_vectors for 64-bit vector element shift counts are generally
9615 // not legal, and it is hard to see their values after they get legalized to
9616 // loads from a constant pool.
9617 case Intrinsic::arm_neon_vshifts:
9618 case Intrinsic::arm_neon_vshiftu:
9619 case Intrinsic::arm_neon_vrshifts:
9620 case Intrinsic::arm_neon_vrshiftu:
9621 case Intrinsic::arm_neon_vrshiftn:
9622 case Intrinsic::arm_neon_vqshifts:
9623 case Intrinsic::arm_neon_vqshiftu:
9624 case Intrinsic::arm_neon_vqshiftsu:
9625 case Intrinsic::arm_neon_vqshiftns:
9626 case Intrinsic::arm_neon_vqshiftnu:
9627 case Intrinsic::arm_neon_vqshiftnsu:
9628 case Intrinsic::arm_neon_vqrshiftns:
9629 case Intrinsic::arm_neon_vqrshiftnu:
9630 case Intrinsic::arm_neon_vqrshiftnsu: {
9631 EVT VT = N->getOperand(1).getValueType();
9633 unsigned VShiftOpc = 0;
9636 case Intrinsic::arm_neon_vshifts:
9637 case Intrinsic::arm_neon_vshiftu:
9638 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
9639 VShiftOpc = ARMISD::VSHL;
9642 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
9643 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
9644 ARMISD::VSHRs : ARMISD::VSHRu);
9649 case Intrinsic::arm_neon_vrshifts:
9650 case Intrinsic::arm_neon_vrshiftu:
9651 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
9655 case Intrinsic::arm_neon_vqshifts:
9656 case Intrinsic::arm_neon_vqshiftu:
9657 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
9661 case Intrinsic::arm_neon_vqshiftsu:
9662 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
9664 llvm_unreachable("invalid shift count for vqshlu intrinsic");
9666 case Intrinsic::arm_neon_vrshiftn:
9667 case Intrinsic::arm_neon_vqshiftns:
9668 case Intrinsic::arm_neon_vqshiftnu:
9669 case Intrinsic::arm_neon_vqshiftnsu:
9670 case Intrinsic::arm_neon_vqrshiftns:
9671 case Intrinsic::arm_neon_vqrshiftnu:
9672 case Intrinsic::arm_neon_vqrshiftnsu:
9673 // Narrowing shifts require an immediate right shift.
9674 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
9676 llvm_unreachable("invalid shift count for narrowing vector shift "
9680 llvm_unreachable("unhandled vector shift");
9684 case Intrinsic::arm_neon_vshifts:
9685 case Intrinsic::arm_neon_vshiftu:
9686 // Opcode already set above.
9688 case Intrinsic::arm_neon_vrshifts:
9689 VShiftOpc = ARMISD::VRSHRs; break;
9690 case Intrinsic::arm_neon_vrshiftu:
9691 VShiftOpc = ARMISD::VRSHRu; break;
9692 case Intrinsic::arm_neon_vrshiftn:
9693 VShiftOpc = ARMISD::VRSHRN; break;
9694 case Intrinsic::arm_neon_vqshifts:
9695 VShiftOpc = ARMISD::VQSHLs; break;
9696 case Intrinsic::arm_neon_vqshiftu:
9697 VShiftOpc = ARMISD::VQSHLu; break;
9698 case Intrinsic::arm_neon_vqshiftsu:
9699 VShiftOpc = ARMISD::VQSHLsu; break;
9700 case Intrinsic::arm_neon_vqshiftns:
9701 VShiftOpc = ARMISD::VQSHRNs; break;
9702 case Intrinsic::arm_neon_vqshiftnu:
9703 VShiftOpc = ARMISD::VQSHRNu; break;
9704 case Intrinsic::arm_neon_vqshiftnsu:
9705 VShiftOpc = ARMISD::VQSHRNsu; break;
9706 case Intrinsic::arm_neon_vqrshiftns:
9707 VShiftOpc = ARMISD::VQRSHRNs; break;
9708 case Intrinsic::arm_neon_vqrshiftnu:
9709 VShiftOpc = ARMISD::VQRSHRNu; break;
9710 case Intrinsic::arm_neon_vqrshiftnsu:
9711 VShiftOpc = ARMISD::VQRSHRNsu; break;
9715 return DAG.getNode(VShiftOpc, dl, N->getValueType(0),
9716 N->getOperand(1), DAG.getConstant(Cnt, dl, MVT::i32));
9719 case Intrinsic::arm_neon_vshiftins: {
9720 EVT VT = N->getOperand(1).getValueType();
9722 unsigned VShiftOpc = 0;
9724 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
9725 VShiftOpc = ARMISD::VSLI;
9726 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
9727 VShiftOpc = ARMISD::VSRI;
9729 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
9733 return DAG.getNode(VShiftOpc, dl, N->getValueType(0),
9734 N->getOperand(1), N->getOperand(2),
9735 DAG.getConstant(Cnt, dl, MVT::i32));
9738 case Intrinsic::arm_neon_vqrshifts:
9739 case Intrinsic::arm_neon_vqrshiftu:
9740 // No immediate versions of these to check for.
9747 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
9748 /// lowers them. As with the vector shift intrinsics, this is done during DAG
9749 /// combining instead of DAG legalizing because the build_vectors for 64-bit
9750 /// vector element shift counts are generally not legal, and it is hard to see
9751 /// their values after they get legalized to loads from a constant pool.
9752 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
9753 const ARMSubtarget *ST) {
9754 EVT VT = N->getValueType(0);
9755 if (N->getOpcode() == ISD::SRL && VT == MVT::i32 && ST->hasV6Ops()) {
9756 // Canonicalize (srl (bswap x), 16) to (rotr (bswap x), 16) if the high
9757 // 16-bits of x is zero. This optimizes rev + lsr 16 to rev16.
9758 SDValue N1 = N->getOperand(1);
9759 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) {
9760 SDValue N0 = N->getOperand(0);
9761 if (C->getZExtValue() == 16 && N0.getOpcode() == ISD::BSWAP &&
9762 DAG.MaskedValueIsZero(N0.getOperand(0),
9763 APInt::getHighBitsSet(32, 16)))
9764 return DAG.getNode(ISD::ROTR, SDLoc(N), VT, N0, N1);
9768 // Nothing to be done for scalar shifts.
9769 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9770 if (!VT.isVector() || !TLI.isTypeLegal(VT))
9773 assert(ST->hasNEON() && "unexpected vector shift");
9776 switch (N->getOpcode()) {
9777 default: llvm_unreachable("unexpected shift opcode");
9780 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt)) {
9782 return DAG.getNode(ARMISD::VSHL, dl, VT, N->getOperand(0),
9783 DAG.getConstant(Cnt, dl, MVT::i32));
9789 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
9790 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
9791 ARMISD::VSHRs : ARMISD::VSHRu);
9793 return DAG.getNode(VShiftOpc, dl, VT, N->getOperand(0),
9794 DAG.getConstant(Cnt, dl, MVT::i32));
9800 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
9801 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
9802 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
9803 const ARMSubtarget *ST) {
9804 SDValue N0 = N->getOperand(0);
9806 // Check for sign- and zero-extensions of vector extract operations of 8-
9807 // and 16-bit vector elements. NEON supports these directly. They are
9808 // handled during DAG combining because type legalization will promote them
9809 // to 32-bit types and it is messy to recognize the operations after that.
9810 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
9811 SDValue Vec = N0.getOperand(0);
9812 SDValue Lane = N0.getOperand(1);
9813 EVT VT = N->getValueType(0);
9814 EVT EltVT = N0.getValueType();
9815 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9817 if (VT == MVT::i32 &&
9818 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
9819 TLI.isTypeLegal(Vec.getValueType()) &&
9820 isa<ConstantSDNode>(Lane)) {
9823 switch (N->getOpcode()) {
9824 default: llvm_unreachable("unexpected opcode");
9825 case ISD::SIGN_EXTEND:
9826 Opc = ARMISD::VGETLANEs;
9828 case ISD::ZERO_EXTEND:
9829 case ISD::ANY_EXTEND:
9830 Opc = ARMISD::VGETLANEu;
9833 return DAG.getNode(Opc, SDLoc(N), VT, Vec, Lane);
9840 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
9841 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
9842 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
9843 const ARMSubtarget *ST) {
9844 // If the target supports NEON, try to use vmax/vmin instructions for f32
9845 // selects like "x < y ? x : y". Unless the NoNaNsFPMath option is set,
9846 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
9847 // a NaN; only do the transformation when it matches that behavior.
9849 // For now only do this when using NEON for FP operations; if using VFP, it
9850 // is not obvious that the benefit outweighs the cost of switching to the
9852 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
9853 N->getValueType(0) != MVT::f32)
9856 SDValue CondLHS = N->getOperand(0);
9857 SDValue CondRHS = N->getOperand(1);
9858 SDValue LHS = N->getOperand(2);
9859 SDValue RHS = N->getOperand(3);
9860 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
9862 unsigned Opcode = 0;
9864 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
9865 IsReversed = false; // x CC y ? x : y
9866 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
9867 IsReversed = true ; // x CC y ? y : x
9881 // If LHS is NaN, an ordered comparison will be false and the result will
9882 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
9883 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
9884 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
9885 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
9887 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
9888 // will return -0, so vmin can only be used for unsafe math or if one of
9889 // the operands is known to be nonzero.
9890 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
9891 !DAG.getTarget().Options.UnsafeFPMath &&
9892 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9894 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
9903 // If LHS is NaN, an ordered comparison will be false and the result will
9904 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
9905 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
9906 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
9907 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
9909 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
9910 // will return +0, so vmax can only be used for unsafe math or if one of
9911 // the operands is known to be nonzero.
9912 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
9913 !DAG.getTarget().Options.UnsafeFPMath &&
9914 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9916 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
9922 return DAG.getNode(Opcode, SDLoc(N), N->getValueType(0), LHS, RHS);
9925 /// PerformCMOVCombine - Target-specific DAG combining for ARMISD::CMOV.
9927 ARMTargetLowering::PerformCMOVCombine(SDNode *N, SelectionDAG &DAG) const {
9928 SDValue Cmp = N->getOperand(4);
9929 if (Cmp.getOpcode() != ARMISD::CMPZ)
9930 // Only looking at EQ and NE cases.
9933 EVT VT = N->getValueType(0);
9935 SDValue LHS = Cmp.getOperand(0);
9936 SDValue RHS = Cmp.getOperand(1);
9937 SDValue FalseVal = N->getOperand(0);
9938 SDValue TrueVal = N->getOperand(1);
9939 SDValue ARMcc = N->getOperand(2);
9940 ARMCC::CondCodes CC =
9941 (ARMCC::CondCodes)cast<ConstantSDNode>(ARMcc)->getZExtValue();
9959 /// FIXME: Turn this into a target neutral optimization?
9961 if (CC == ARMCC::NE && FalseVal == RHS && FalseVal != LHS) {
9962 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, TrueVal, ARMcc,
9963 N->getOperand(3), Cmp);
9964 } else if (CC == ARMCC::EQ && TrueVal == RHS) {
9966 SDValue NewCmp = getARMCmp(LHS, RHS, ISD::SETNE, ARMcc, DAG, dl);
9967 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, FalseVal, ARMcc,
9968 N->getOperand(3), NewCmp);
9971 if (Res.getNode()) {
9972 APInt KnownZero, KnownOne;
9973 DAG.computeKnownBits(SDValue(N,0), KnownZero, KnownOne);
9974 // Capture demanded bits information that would be otherwise lost.
9975 if (KnownZero == 0xfffffffe)
9976 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
9977 DAG.getValueType(MVT::i1));
9978 else if (KnownZero == 0xffffff00)
9979 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
9980 DAG.getValueType(MVT::i8));
9981 else if (KnownZero == 0xffff0000)
9982 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
9983 DAG.getValueType(MVT::i16));
9989 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
9990 DAGCombinerInfo &DCI) const {
9991 switch (N->getOpcode()) {
9993 case ISD::ADDC: return PerformADDCCombine(N, DCI, Subtarget);
9994 case ISD::ADD: return PerformADDCombine(N, DCI, Subtarget);
9995 case ISD::SUB: return PerformSUBCombine(N, DCI);
9996 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
9997 case ISD::OR: return PerformORCombine(N, DCI, Subtarget);
9998 case ISD::XOR: return PerformXORCombine(N, DCI, Subtarget);
9999 case ISD::AND: return PerformANDCombine(N, DCI, Subtarget);
10000 case ARMISD::BFI: return PerformBFICombine(N, DCI);
10001 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI, Subtarget);
10002 case ARMISD::VMOVDRR: return PerformVMOVDRRCombine(N, DCI.DAG);
10003 case ISD::STORE: return PerformSTORECombine(N, DCI);
10004 case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DCI, Subtarget);
10005 case ISD::INSERT_VECTOR_ELT: return PerformInsertEltCombine(N, DCI);
10006 case ISD::VECTOR_SHUFFLE: return PerformVECTOR_SHUFFLECombine(N, DCI.DAG);
10007 case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI);
10008 case ISD::FP_TO_SINT:
10009 case ISD::FP_TO_UINT: return PerformVCVTCombine(N, DCI, Subtarget);
10010 case ISD::FDIV: return PerformVDIVCombine(N, DCI, Subtarget);
10011 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
10014 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
10015 case ISD::SIGN_EXTEND:
10016 case ISD::ZERO_EXTEND:
10017 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
10018 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
10019 case ARMISD::CMOV: return PerformCMOVCombine(N, DCI.DAG);
10020 case ISD::LOAD: return PerformLOADCombine(N, DCI);
10021 case ARMISD::VLD2DUP:
10022 case ARMISD::VLD3DUP:
10023 case ARMISD::VLD4DUP:
10024 return PerformVLDCombine(N, DCI);
10025 case ARMISD::BUILD_VECTOR:
10026 return PerformARMBUILD_VECTORCombine(N, DCI);
10027 case ISD::INTRINSIC_VOID:
10028 case ISD::INTRINSIC_W_CHAIN:
10029 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
10030 case Intrinsic::arm_neon_vld1:
10031 case Intrinsic::arm_neon_vld2:
10032 case Intrinsic::arm_neon_vld3:
10033 case Intrinsic::arm_neon_vld4:
10034 case Intrinsic::arm_neon_vld2lane:
10035 case Intrinsic::arm_neon_vld3lane:
10036 case Intrinsic::arm_neon_vld4lane:
10037 case Intrinsic::arm_neon_vst1:
10038 case Intrinsic::arm_neon_vst2:
10039 case Intrinsic::arm_neon_vst3:
10040 case Intrinsic::arm_neon_vst4:
10041 case Intrinsic::arm_neon_vst2lane:
10042 case Intrinsic::arm_neon_vst3lane:
10043 case Intrinsic::arm_neon_vst4lane:
10044 return PerformVLDCombine(N, DCI);
10052 bool ARMTargetLowering::isDesirableToTransformToIntegerOp(unsigned Opc,
10054 return (VT == MVT::f32) && (Opc == ISD::LOAD || Opc == ISD::STORE);
10057 bool ARMTargetLowering::allowsMisalignedMemoryAccesses(EVT VT,
10060 bool *Fast) const {
10061 // The AllowsUnaliged flag models the SCTLR.A setting in ARM cpus
10062 bool AllowsUnaligned = Subtarget->allowsUnalignedMem();
10064 switch (VT.getSimpleVT().SimpleTy) {
10070 // Unaligned access can use (for example) LRDB, LRDH, LDR
10071 if (AllowsUnaligned) {
10073 *Fast = Subtarget->hasV7Ops();
10080 // For any little-endian targets with neon, we can support unaligned ld/st
10081 // of D and Q (e.g. {D0,D1}) registers by using vld1.i8/vst1.i8.
10082 // A big-endian target may also explicitly support unaligned accesses
10083 if (Subtarget->hasNEON() && (AllowsUnaligned || Subtarget->isLittle())) {
10093 static bool memOpAlign(unsigned DstAlign, unsigned SrcAlign,
10094 unsigned AlignCheck) {
10095 return ((SrcAlign == 0 || SrcAlign % AlignCheck == 0) &&
10096 (DstAlign == 0 || DstAlign % AlignCheck == 0));
10099 EVT ARMTargetLowering::getOptimalMemOpType(uint64_t Size,
10100 unsigned DstAlign, unsigned SrcAlign,
10101 bool IsMemset, bool ZeroMemset,
10103 MachineFunction &MF) const {
10104 const Function *F = MF.getFunction();
10106 // See if we can use NEON instructions for this...
10107 if ((!IsMemset || ZeroMemset) && Subtarget->hasNEON() &&
10108 !F->hasFnAttribute(Attribute::NoImplicitFloat)) {
10111 (memOpAlign(SrcAlign, DstAlign, 16) ||
10112 (allowsMisalignedMemoryAccesses(MVT::v2f64, 0, 1, &Fast) && Fast))) {
10114 } else if (Size >= 8 &&
10115 (memOpAlign(SrcAlign, DstAlign, 8) ||
10116 (allowsMisalignedMemoryAccesses(MVT::f64, 0, 1, &Fast) &&
10122 // Lowering to i32/i16 if the size permits.
10125 else if (Size >= 2)
10128 // Let the target-independent logic figure it out.
10132 bool ARMTargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
10133 if (Val.getOpcode() != ISD::LOAD)
10136 EVT VT1 = Val.getValueType();
10137 if (!VT1.isSimple() || !VT1.isInteger() ||
10138 !VT2.isSimple() || !VT2.isInteger())
10141 switch (VT1.getSimpleVT().SimpleTy) {
10146 // 8-bit and 16-bit loads implicitly zero-extend to 32-bits.
10153 bool ARMTargetLowering::isVectorLoadExtDesirable(SDValue ExtVal) const {
10154 EVT VT = ExtVal.getValueType();
10156 if (!isTypeLegal(VT))
10159 // Don't create a loadext if we can fold the extension into a wide/long
10161 // If there's more than one user instruction, the loadext is desirable no
10162 // matter what. There can be two uses by the same instruction.
10163 if (ExtVal->use_empty() ||
10164 !ExtVal->use_begin()->isOnlyUserOf(ExtVal.getNode()))
10167 SDNode *U = *ExtVal->use_begin();
10168 if ((U->getOpcode() == ISD::ADD || U->getOpcode() == ISD::SUB ||
10169 U->getOpcode() == ISD::SHL || U->getOpcode() == ARMISD::VSHL))
10175 bool ARMTargetLowering::allowTruncateForTailCall(Type *Ty1, Type *Ty2) const {
10176 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
10179 if (!isTypeLegal(EVT::getEVT(Ty1)))
10182 assert(Ty1->getPrimitiveSizeInBits() <= 64 && "i128 is probably not a noop");
10184 // Assuming the caller doesn't have a zeroext or signext return parameter,
10185 // truncation all the way down to i1 is valid.
10190 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
10194 unsigned Scale = 1;
10195 switch (VT.getSimpleVT().SimpleTy) {
10196 default: return false;
10211 if ((V & (Scale - 1)) != 0)
10214 return V == (V & ((1LL << 5) - 1));
10217 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
10218 const ARMSubtarget *Subtarget) {
10219 bool isNeg = false;
10225 switch (VT.getSimpleVT().SimpleTy) {
10226 default: return false;
10231 // + imm12 or - imm8
10233 return V == (V & ((1LL << 8) - 1));
10234 return V == (V & ((1LL << 12) - 1));
10237 // Same as ARM mode. FIXME: NEON?
10238 if (!Subtarget->hasVFP2())
10243 return V == (V & ((1LL << 8) - 1));
10247 /// isLegalAddressImmediate - Return true if the integer value can be used
10248 /// as the offset of the target addressing mode for load / store of the
10250 static bool isLegalAddressImmediate(int64_t V, EVT VT,
10251 const ARMSubtarget *Subtarget) {
10255 if (!VT.isSimple())
10258 if (Subtarget->isThumb1Only())
10259 return isLegalT1AddressImmediate(V, VT);
10260 else if (Subtarget->isThumb2())
10261 return isLegalT2AddressImmediate(V, VT, Subtarget);
10266 switch (VT.getSimpleVT().SimpleTy) {
10267 default: return false;
10272 return V == (V & ((1LL << 12) - 1));
10275 return V == (V & ((1LL << 8) - 1));
10278 if (!Subtarget->hasVFP2()) // FIXME: NEON?
10283 return V == (V & ((1LL << 8) - 1));
10287 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
10289 int Scale = AM.Scale;
10293 switch (VT.getSimpleVT().SimpleTy) {
10294 default: return false;
10302 Scale = Scale & ~1;
10303 return Scale == 2 || Scale == 4 || Scale == 8;
10306 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
10310 // Note, we allow "void" uses (basically, uses that aren't loads or
10311 // stores), because arm allows folding a scale into many arithmetic
10312 // operations. This should be made more precise and revisited later.
10314 // Allow r << imm, but the imm has to be a multiple of two.
10315 if (Scale & 1) return false;
10316 return isPowerOf2_32(Scale);
10320 /// isLegalAddressingMode - Return true if the addressing mode represented
10321 /// by AM is legal for this target, for a load/store of the specified type.
10322 bool ARMTargetLowering::isLegalAddressingMode(const DataLayout &DL,
10323 const AddrMode &AM, Type *Ty,
10324 unsigned AS) const {
10325 EVT VT = getValueType(DL, Ty, true);
10326 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
10329 // Can never fold addr of global into load/store.
10333 switch (AM.Scale) {
10334 case 0: // no scale reg, must be "r+i" or "r", or "i".
10337 if (Subtarget->isThumb1Only())
10341 // ARM doesn't support any R+R*scale+imm addr modes.
10345 if (!VT.isSimple())
10348 if (Subtarget->isThumb2())
10349 return isLegalT2ScaledAddressingMode(AM, VT);
10351 int Scale = AM.Scale;
10352 switch (VT.getSimpleVT().SimpleTy) {
10353 default: return false;
10357 if (Scale < 0) Scale = -Scale;
10361 return isPowerOf2_32(Scale & ~1);
10365 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
10370 // Note, we allow "void" uses (basically, uses that aren't loads or
10371 // stores), because arm allows folding a scale into many arithmetic
10372 // operations. This should be made more precise and revisited later.
10374 // Allow r << imm, but the imm has to be a multiple of two.
10375 if (Scale & 1) return false;
10376 return isPowerOf2_32(Scale);
10382 /// isLegalICmpImmediate - Return true if the specified immediate is legal
10383 /// icmp immediate, that is the target has icmp instructions which can compare
10384 /// a register against the immediate without having to materialize the
10385 /// immediate into a register.
10386 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
10387 // Thumb2 and ARM modes can use cmn for negative immediates.
10388 if (!Subtarget->isThumb())
10389 return ARM_AM::getSOImmVal(std::abs(Imm)) != -1;
10390 if (Subtarget->isThumb2())
10391 return ARM_AM::getT2SOImmVal(std::abs(Imm)) != -1;
10392 // Thumb1 doesn't have cmn, and only 8-bit immediates.
10393 return Imm >= 0 && Imm <= 255;
10396 /// isLegalAddImmediate - Return true if the specified immediate is a legal add
10397 /// *or sub* immediate, that is the target has add or sub instructions which can
10398 /// add a register with the immediate without having to materialize the
10399 /// immediate into a register.
10400 bool ARMTargetLowering::isLegalAddImmediate(int64_t Imm) const {
10401 // Same encoding for add/sub, just flip the sign.
10402 int64_t AbsImm = std::abs(Imm);
10403 if (!Subtarget->isThumb())
10404 return ARM_AM::getSOImmVal(AbsImm) != -1;
10405 if (Subtarget->isThumb2())
10406 return ARM_AM::getT2SOImmVal(AbsImm) != -1;
10407 // Thumb1 only has 8-bit unsigned immediate.
10408 return AbsImm >= 0 && AbsImm <= 255;
10411 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
10412 bool isSEXTLoad, SDValue &Base,
10413 SDValue &Offset, bool &isInc,
10414 SelectionDAG &DAG) {
10415 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
10418 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
10419 // AddressingMode 3
10420 Base = Ptr->getOperand(0);
10421 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
10422 int RHSC = (int)RHS->getZExtValue();
10423 if (RHSC < 0 && RHSC > -256) {
10424 assert(Ptr->getOpcode() == ISD::ADD);
10426 Offset = DAG.getConstant(-RHSC, SDLoc(Ptr), RHS->getValueType(0));
10430 isInc = (Ptr->getOpcode() == ISD::ADD);
10431 Offset = Ptr->getOperand(1);
10433 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
10434 // AddressingMode 2
10435 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
10436 int RHSC = (int)RHS->getZExtValue();
10437 if (RHSC < 0 && RHSC > -0x1000) {
10438 assert(Ptr->getOpcode() == ISD::ADD);
10440 Offset = DAG.getConstant(-RHSC, SDLoc(Ptr), RHS->getValueType(0));
10441 Base = Ptr->getOperand(0);
10446 if (Ptr->getOpcode() == ISD::ADD) {
10448 ARM_AM::ShiftOpc ShOpcVal=
10449 ARM_AM::getShiftOpcForNode(Ptr->getOperand(0).getOpcode());
10450 if (ShOpcVal != ARM_AM::no_shift) {
10451 Base = Ptr->getOperand(1);
10452 Offset = Ptr->getOperand(0);
10454 Base = Ptr->getOperand(0);
10455 Offset = Ptr->getOperand(1);
10460 isInc = (Ptr->getOpcode() == ISD::ADD);
10461 Base = Ptr->getOperand(0);
10462 Offset = Ptr->getOperand(1);
10466 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
10470 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
10471 bool isSEXTLoad, SDValue &Base,
10472 SDValue &Offset, bool &isInc,
10473 SelectionDAG &DAG) {
10474 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
10477 Base = Ptr->getOperand(0);
10478 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
10479 int RHSC = (int)RHS->getZExtValue();
10480 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
10481 assert(Ptr->getOpcode() == ISD::ADD);
10483 Offset = DAG.getConstant(-RHSC, SDLoc(Ptr), RHS->getValueType(0));
10485 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
10486 isInc = Ptr->getOpcode() == ISD::ADD;
10487 Offset = DAG.getConstant(RHSC, SDLoc(Ptr), RHS->getValueType(0));
10495 /// getPreIndexedAddressParts - returns true by value, base pointer and
10496 /// offset pointer and addressing mode by reference if the node's address
10497 /// can be legally represented as pre-indexed load / store address.
10499 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
10501 ISD::MemIndexedMode &AM,
10502 SelectionDAG &DAG) const {
10503 if (Subtarget->isThumb1Only())
10508 bool isSEXTLoad = false;
10509 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
10510 Ptr = LD->getBasePtr();
10511 VT = LD->getMemoryVT();
10512 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
10513 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
10514 Ptr = ST->getBasePtr();
10515 VT = ST->getMemoryVT();
10520 bool isLegal = false;
10521 if (Subtarget->isThumb2())
10522 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
10523 Offset, isInc, DAG);
10525 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
10526 Offset, isInc, DAG);
10530 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
10534 /// getPostIndexedAddressParts - returns true by value, base pointer and
10535 /// offset pointer and addressing mode by reference if this node can be
10536 /// combined with a load / store to form a post-indexed load / store.
10537 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
10540 ISD::MemIndexedMode &AM,
10541 SelectionDAG &DAG) const {
10542 if (Subtarget->isThumb1Only())
10547 bool isSEXTLoad = false;
10548 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
10549 VT = LD->getMemoryVT();
10550 Ptr = LD->getBasePtr();
10551 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
10552 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
10553 VT = ST->getMemoryVT();
10554 Ptr = ST->getBasePtr();
10559 bool isLegal = false;
10560 if (Subtarget->isThumb2())
10561 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
10564 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
10570 // Swap base ptr and offset to catch more post-index load / store when
10571 // it's legal. In Thumb2 mode, offset must be an immediate.
10572 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
10573 !Subtarget->isThumb2())
10574 std::swap(Base, Offset);
10576 // Post-indexed load / store update the base pointer.
10581 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
10585 void ARMTargetLowering::computeKnownBitsForTargetNode(const SDValue Op,
10588 const SelectionDAG &DAG,
10589 unsigned Depth) const {
10590 unsigned BitWidth = KnownOne.getBitWidth();
10591 KnownZero = KnownOne = APInt(BitWidth, 0);
10592 switch (Op.getOpcode()) {
10598 // These nodes' second result is a boolean
10599 if (Op.getResNo() == 0)
10601 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - 1);
10603 case ARMISD::CMOV: {
10604 // Bits are known zero/one if known on the LHS and RHS.
10605 DAG.computeKnownBits(Op.getOperand(0), KnownZero, KnownOne, Depth+1);
10606 if (KnownZero == 0 && KnownOne == 0) return;
10608 APInt KnownZeroRHS, KnownOneRHS;
10609 DAG.computeKnownBits(Op.getOperand(1), KnownZeroRHS, KnownOneRHS, Depth+1);
10610 KnownZero &= KnownZeroRHS;
10611 KnownOne &= KnownOneRHS;
10614 case ISD::INTRINSIC_W_CHAIN: {
10615 ConstantSDNode *CN = cast<ConstantSDNode>(Op->getOperand(1));
10616 Intrinsic::ID IntID = static_cast<Intrinsic::ID>(CN->getZExtValue());
10619 case Intrinsic::arm_ldaex:
10620 case Intrinsic::arm_ldrex: {
10621 EVT VT = cast<MemIntrinsicSDNode>(Op)->getMemoryVT();
10622 unsigned MemBits = VT.getScalarType().getSizeInBits();
10623 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - MemBits);
10631 //===----------------------------------------------------------------------===//
10632 // ARM Inline Assembly Support
10633 //===----------------------------------------------------------------------===//
10635 bool ARMTargetLowering::ExpandInlineAsm(CallInst *CI) const {
10636 // Looking for "rev" which is V6+.
10637 if (!Subtarget->hasV6Ops())
10640 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
10641 std::string AsmStr = IA->getAsmString();
10642 SmallVector<StringRef, 4> AsmPieces;
10643 SplitString(AsmStr, AsmPieces, ";\n");
10645 switch (AsmPieces.size()) {
10646 default: return false;
10648 AsmStr = AsmPieces[0];
10650 SplitString(AsmStr, AsmPieces, " \t,");
10653 if (AsmPieces.size() == 3 &&
10654 AsmPieces[0] == "rev" && AsmPieces[1] == "$0" && AsmPieces[2] == "$1" &&
10655 IA->getConstraintString().compare(0, 4, "=l,l") == 0) {
10656 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
10657 if (Ty && Ty->getBitWidth() == 32)
10658 return IntrinsicLowering::LowerToByteSwap(CI);
10666 /// getConstraintType - Given a constraint letter, return the type of
10667 /// constraint it is for this target.
10668 ARMTargetLowering::ConstraintType
10669 ARMTargetLowering::getConstraintType(StringRef Constraint) const {
10670 if (Constraint.size() == 1) {
10671 switch (Constraint[0]) {
10673 case 'l': return C_RegisterClass;
10674 case 'w': return C_RegisterClass;
10675 case 'h': return C_RegisterClass;
10676 case 'x': return C_RegisterClass;
10677 case 't': return C_RegisterClass;
10678 case 'j': return C_Other; // Constant for movw.
10679 // An address with a single base register. Due to the way we
10680 // currently handle addresses it is the same as an 'r' memory constraint.
10681 case 'Q': return C_Memory;
10683 } else if (Constraint.size() == 2) {
10684 switch (Constraint[0]) {
10686 // All 'U+' constraints are addresses.
10687 case 'U': return C_Memory;
10690 return TargetLowering::getConstraintType(Constraint);
10693 /// Examine constraint type and operand type and determine a weight value.
10694 /// This object must already have been set up with the operand type
10695 /// and the current alternative constraint selected.
10696 TargetLowering::ConstraintWeight
10697 ARMTargetLowering::getSingleConstraintMatchWeight(
10698 AsmOperandInfo &info, const char *constraint) const {
10699 ConstraintWeight weight = CW_Invalid;
10700 Value *CallOperandVal = info.CallOperandVal;
10701 // If we don't have a value, we can't do a match,
10702 // but allow it at the lowest weight.
10703 if (!CallOperandVal)
10705 Type *type = CallOperandVal->getType();
10706 // Look at the constraint type.
10707 switch (*constraint) {
10709 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
10712 if (type->isIntegerTy()) {
10713 if (Subtarget->isThumb())
10714 weight = CW_SpecificReg;
10716 weight = CW_Register;
10720 if (type->isFloatingPointTy())
10721 weight = CW_Register;
10727 typedef std::pair<unsigned, const TargetRegisterClass*> RCPair;
10728 RCPair ARMTargetLowering::getRegForInlineAsmConstraint(
10729 const TargetRegisterInfo *TRI, StringRef Constraint, MVT VT) const {
10730 if (Constraint.size() == 1) {
10731 // GCC ARM Constraint Letters
10732 switch (Constraint[0]) {
10733 case 'l': // Low regs or general regs.
10734 if (Subtarget->isThumb())
10735 return RCPair(0U, &ARM::tGPRRegClass);
10736 return RCPair(0U, &ARM::GPRRegClass);
10737 case 'h': // High regs or no regs.
10738 if (Subtarget->isThumb())
10739 return RCPair(0U, &ARM::hGPRRegClass);
10742 if (Subtarget->isThumb1Only())
10743 return RCPair(0U, &ARM::tGPRRegClass);
10744 return RCPair(0U, &ARM::GPRRegClass);
10746 if (VT == MVT::Other)
10748 if (VT == MVT::f32)
10749 return RCPair(0U, &ARM::SPRRegClass);
10750 if (VT.getSizeInBits() == 64)
10751 return RCPair(0U, &ARM::DPRRegClass);
10752 if (VT.getSizeInBits() == 128)
10753 return RCPair(0U, &ARM::QPRRegClass);
10756 if (VT == MVT::Other)
10758 if (VT == MVT::f32)
10759 return RCPair(0U, &ARM::SPR_8RegClass);
10760 if (VT.getSizeInBits() == 64)
10761 return RCPair(0U, &ARM::DPR_8RegClass);
10762 if (VT.getSizeInBits() == 128)
10763 return RCPair(0U, &ARM::QPR_8RegClass);
10766 if (VT == MVT::f32)
10767 return RCPair(0U, &ARM::SPRRegClass);
10771 if (StringRef("{cc}").equals_lower(Constraint))
10772 return std::make_pair(unsigned(ARM::CPSR), &ARM::CCRRegClass);
10774 return TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
10777 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
10778 /// vector. If it is invalid, don't add anything to Ops.
10779 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
10780 std::string &Constraint,
10781 std::vector<SDValue>&Ops,
10782 SelectionDAG &DAG) const {
10785 // Currently only support length 1 constraints.
10786 if (Constraint.length() != 1) return;
10788 char ConstraintLetter = Constraint[0];
10789 switch (ConstraintLetter) {
10792 case 'I': case 'J': case 'K': case 'L':
10793 case 'M': case 'N': case 'O':
10794 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
10798 int64_t CVal64 = C->getSExtValue();
10799 int CVal = (int) CVal64;
10800 // None of these constraints allow values larger than 32 bits. Check
10801 // that the value fits in an int.
10802 if (CVal != CVal64)
10805 switch (ConstraintLetter) {
10807 // Constant suitable for movw, must be between 0 and
10809 if (Subtarget->hasV6T2Ops())
10810 if (CVal >= 0 && CVal <= 65535)
10814 if (Subtarget->isThumb1Only()) {
10815 // This must be a constant between 0 and 255, for ADD
10817 if (CVal >= 0 && CVal <= 255)
10819 } else if (Subtarget->isThumb2()) {
10820 // A constant that can be used as an immediate value in a
10821 // data-processing instruction.
10822 if (ARM_AM::getT2SOImmVal(CVal) != -1)
10825 // A constant that can be used as an immediate value in a
10826 // data-processing instruction.
10827 if (ARM_AM::getSOImmVal(CVal) != -1)
10833 if (Subtarget->isThumb()) { // FIXME thumb2
10834 // This must be a constant between -255 and -1, for negated ADD
10835 // immediates. This can be used in GCC with an "n" modifier that
10836 // prints the negated value, for use with SUB instructions. It is
10837 // not useful otherwise but is implemented for compatibility.
10838 if (CVal >= -255 && CVal <= -1)
10841 // This must be a constant between -4095 and 4095. It is not clear
10842 // what this constraint is intended for. Implemented for
10843 // compatibility with GCC.
10844 if (CVal >= -4095 && CVal <= 4095)
10850 if (Subtarget->isThumb1Only()) {
10851 // A 32-bit value where only one byte has a nonzero value. Exclude
10852 // zero to match GCC. This constraint is used by GCC internally for
10853 // constants that can be loaded with a move/shift combination.
10854 // It is not useful otherwise but is implemented for compatibility.
10855 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
10857 } else if (Subtarget->isThumb2()) {
10858 // A constant whose bitwise inverse can be used as an immediate
10859 // value in a data-processing instruction. This can be used in GCC
10860 // with a "B" modifier that prints the inverted value, for use with
10861 // BIC and MVN instructions. It is not useful otherwise but is
10862 // implemented for compatibility.
10863 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
10866 // A constant whose bitwise inverse can be used as an immediate
10867 // value in a data-processing instruction. This can be used in GCC
10868 // with a "B" modifier that prints the inverted value, for use with
10869 // BIC and MVN instructions. It is not useful otherwise but is
10870 // implemented for compatibility.
10871 if (ARM_AM::getSOImmVal(~CVal) != -1)
10877 if (Subtarget->isThumb1Only()) {
10878 // This must be a constant between -7 and 7,
10879 // for 3-operand ADD/SUB immediate instructions.
10880 if (CVal >= -7 && CVal < 7)
10882 } else if (Subtarget->isThumb2()) {
10883 // A constant whose negation can be used as an immediate value in a
10884 // data-processing instruction. This can be used in GCC with an "n"
10885 // modifier that prints the negated value, for use with SUB
10886 // instructions. It is not useful otherwise but is implemented for
10888 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
10891 // A constant whose negation can be used as an immediate value in a
10892 // data-processing instruction. This can be used in GCC with an "n"
10893 // modifier that prints the negated value, for use with SUB
10894 // instructions. It is not useful otherwise but is implemented for
10896 if (ARM_AM::getSOImmVal(-CVal) != -1)
10902 if (Subtarget->isThumb()) { // FIXME thumb2
10903 // This must be a multiple of 4 between 0 and 1020, for
10904 // ADD sp + immediate.
10905 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
10908 // A power of two or a constant between 0 and 32. This is used in
10909 // GCC for the shift amount on shifted register operands, but it is
10910 // useful in general for any shift amounts.
10911 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
10917 if (Subtarget->isThumb()) { // FIXME thumb2
10918 // This must be a constant between 0 and 31, for shift amounts.
10919 if (CVal >= 0 && CVal <= 31)
10925 if (Subtarget->isThumb()) { // FIXME thumb2
10926 // This must be a multiple of 4 between -508 and 508, for
10927 // ADD/SUB sp = sp + immediate.
10928 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
10933 Result = DAG.getTargetConstant(CVal, SDLoc(Op), Op.getValueType());
10937 if (Result.getNode()) {
10938 Ops.push_back(Result);
10941 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
10944 SDValue ARMTargetLowering::LowerDivRem(SDValue Op, SelectionDAG &DAG) const {
10945 assert(Subtarget->isTargetAEABI() && "Register-based DivRem lowering only");
10946 unsigned Opcode = Op->getOpcode();
10947 assert((Opcode == ISD::SDIVREM || Opcode == ISD::UDIVREM) &&
10948 "Invalid opcode for Div/Rem lowering");
10949 bool isSigned = (Opcode == ISD::SDIVREM);
10950 EVT VT = Op->getValueType(0);
10951 Type *Ty = VT.getTypeForEVT(*DAG.getContext());
10954 switch (VT.getSimpleVT().SimpleTy) {
10955 default: llvm_unreachable("Unexpected request for libcall!");
10956 case MVT::i8: LC = isSigned ? RTLIB::SDIVREM_I8 : RTLIB::UDIVREM_I8; break;
10957 case MVT::i16: LC = isSigned ? RTLIB::SDIVREM_I16 : RTLIB::UDIVREM_I16; break;
10958 case MVT::i32: LC = isSigned ? RTLIB::SDIVREM_I32 : RTLIB::UDIVREM_I32; break;
10959 case MVT::i64: LC = isSigned ? RTLIB::SDIVREM_I64 : RTLIB::UDIVREM_I64; break;
10962 SDValue InChain = DAG.getEntryNode();
10964 TargetLowering::ArgListTy Args;
10965 TargetLowering::ArgListEntry Entry;
10966 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i) {
10967 EVT ArgVT = Op->getOperand(i).getValueType();
10968 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
10969 Entry.Node = Op->getOperand(i);
10971 Entry.isSExt = isSigned;
10972 Entry.isZExt = !isSigned;
10973 Args.push_back(Entry);
10976 SDValue Callee = DAG.getExternalSymbol(getLibcallName(LC),
10977 getPointerTy(DAG.getDataLayout()));
10979 Type *RetTy = (Type*)StructType::get(Ty, Ty, nullptr);
10982 TargetLowering::CallLoweringInfo CLI(DAG);
10983 CLI.setDebugLoc(dl).setChain(InChain)
10984 .setCallee(getLibcallCallingConv(LC), RetTy, Callee, std::move(Args), 0)
10985 .setInRegister().setSExtResult(isSigned).setZExtResult(!isSigned);
10987 std::pair<SDValue, SDValue> CallInfo = LowerCallTo(CLI);
10988 return CallInfo.first;
10992 ARMTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const {
10993 assert(Subtarget->isTargetWindows() && "unsupported target platform");
10997 SDValue Chain = Op.getOperand(0);
10998 SDValue Size = Op.getOperand(1);
11000 SDValue Words = DAG.getNode(ISD::SRL, DL, MVT::i32, Size,
11001 DAG.getConstant(2, DL, MVT::i32));
11004 Chain = DAG.getCopyToReg(Chain, DL, ARM::R4, Words, Flag);
11005 Flag = Chain.getValue(1);
11007 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
11008 Chain = DAG.getNode(ARMISD::WIN__CHKSTK, DL, NodeTys, Chain, Flag);
11010 SDValue NewSP = DAG.getCopyFromReg(Chain, DL, ARM::SP, MVT::i32);
11011 Chain = NewSP.getValue(1);
11013 SDValue Ops[2] = { NewSP, Chain };
11014 return DAG.getMergeValues(Ops, DL);
11017 SDValue ARMTargetLowering::LowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) const {
11018 assert(Op.getValueType() == MVT::f64 && Subtarget->isFPOnlySP() &&
11019 "Unexpected type for custom-lowering FP_EXTEND");
11022 LC = RTLIB::getFPEXT(Op.getOperand(0).getValueType(), Op.getValueType());
11024 SDValue SrcVal = Op.getOperand(0);
11025 return makeLibCall(DAG, LC, Op.getValueType(), &SrcVal, 1,
11026 /*isSigned*/ false, SDLoc(Op)).first;
11029 SDValue ARMTargetLowering::LowerFP_ROUND(SDValue Op, SelectionDAG &DAG) const {
11030 assert(Op.getOperand(0).getValueType() == MVT::f64 &&
11031 Subtarget->isFPOnlySP() &&
11032 "Unexpected type for custom-lowering FP_ROUND");
11035 LC = RTLIB::getFPROUND(Op.getOperand(0).getValueType(), Op.getValueType());
11037 SDValue SrcVal = Op.getOperand(0);
11038 return makeLibCall(DAG, LC, Op.getValueType(), &SrcVal, 1,
11039 /*isSigned*/ false, SDLoc(Op)).first;
11043 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
11044 // The ARM target isn't yet aware of offsets.
11048 bool ARM::isBitFieldInvertedMask(unsigned v) {
11049 if (v == 0xffffffff)
11052 // there can be 1's on either or both "outsides", all the "inside"
11053 // bits must be 0's
11054 return isShiftedMask_32(~v);
11057 /// isFPImmLegal - Returns true if the target can instruction select the
11058 /// specified FP immediate natively. If false, the legalizer will
11059 /// materialize the FP immediate as a load from a constant pool.
11060 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
11061 if (!Subtarget->hasVFP3())
11063 if (VT == MVT::f32)
11064 return ARM_AM::getFP32Imm(Imm) != -1;
11065 if (VT == MVT::f64 && !Subtarget->isFPOnlySP())
11066 return ARM_AM::getFP64Imm(Imm) != -1;
11070 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
11071 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
11072 /// specified in the intrinsic calls.
11073 bool ARMTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
11075 unsigned Intrinsic) const {
11076 switch (Intrinsic) {
11077 case Intrinsic::arm_neon_vld1:
11078 case Intrinsic::arm_neon_vld2:
11079 case Intrinsic::arm_neon_vld3:
11080 case Intrinsic::arm_neon_vld4:
11081 case Intrinsic::arm_neon_vld2lane:
11082 case Intrinsic::arm_neon_vld3lane:
11083 case Intrinsic::arm_neon_vld4lane: {
11084 Info.opc = ISD::INTRINSIC_W_CHAIN;
11085 // Conservatively set memVT to the entire set of vectors loaded.
11086 auto &DL = I.getCalledFunction()->getParent()->getDataLayout();
11087 uint64_t NumElts = DL.getTypeAllocSize(I.getType()) / 8;
11088 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
11089 Info.ptrVal = I.getArgOperand(0);
11091 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
11092 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
11093 Info.vol = false; // volatile loads with NEON intrinsics not supported
11094 Info.readMem = true;
11095 Info.writeMem = false;
11098 case Intrinsic::arm_neon_vst1:
11099 case Intrinsic::arm_neon_vst2:
11100 case Intrinsic::arm_neon_vst3:
11101 case Intrinsic::arm_neon_vst4:
11102 case Intrinsic::arm_neon_vst2lane:
11103 case Intrinsic::arm_neon_vst3lane:
11104 case Intrinsic::arm_neon_vst4lane: {
11105 Info.opc = ISD::INTRINSIC_VOID;
11106 // Conservatively set memVT to the entire set of vectors stored.
11107 auto &DL = I.getCalledFunction()->getParent()->getDataLayout();
11108 unsigned NumElts = 0;
11109 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
11110 Type *ArgTy = I.getArgOperand(ArgI)->getType();
11111 if (!ArgTy->isVectorTy())
11113 NumElts += DL.getTypeAllocSize(ArgTy) / 8;
11115 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
11116 Info.ptrVal = I.getArgOperand(0);
11118 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
11119 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
11120 Info.vol = false; // volatile stores with NEON intrinsics not supported
11121 Info.readMem = false;
11122 Info.writeMem = true;
11125 case Intrinsic::arm_ldaex:
11126 case Intrinsic::arm_ldrex: {
11127 auto &DL = I.getCalledFunction()->getParent()->getDataLayout();
11128 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(0)->getType());
11129 Info.opc = ISD::INTRINSIC_W_CHAIN;
11130 Info.memVT = MVT::getVT(PtrTy->getElementType());
11131 Info.ptrVal = I.getArgOperand(0);
11133 Info.align = DL.getABITypeAlignment(PtrTy->getElementType());
11135 Info.readMem = true;
11136 Info.writeMem = false;
11139 case Intrinsic::arm_stlex:
11140 case Intrinsic::arm_strex: {
11141 auto &DL = I.getCalledFunction()->getParent()->getDataLayout();
11142 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(1)->getType());
11143 Info.opc = ISD::INTRINSIC_W_CHAIN;
11144 Info.memVT = MVT::getVT(PtrTy->getElementType());
11145 Info.ptrVal = I.getArgOperand(1);
11147 Info.align = DL.getABITypeAlignment(PtrTy->getElementType());
11149 Info.readMem = false;
11150 Info.writeMem = true;
11153 case Intrinsic::arm_stlexd:
11154 case Intrinsic::arm_strexd: {
11155 Info.opc = ISD::INTRINSIC_W_CHAIN;
11156 Info.memVT = MVT::i64;
11157 Info.ptrVal = I.getArgOperand(2);
11161 Info.readMem = false;
11162 Info.writeMem = true;
11165 case Intrinsic::arm_ldaexd:
11166 case Intrinsic::arm_ldrexd: {
11167 Info.opc = ISD::INTRINSIC_W_CHAIN;
11168 Info.memVT = MVT::i64;
11169 Info.ptrVal = I.getArgOperand(0);
11173 Info.readMem = true;
11174 Info.writeMem = false;
11184 /// \brief Returns true if it is beneficial to convert a load of a constant
11185 /// to just the constant itself.
11186 bool ARMTargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
11188 assert(Ty->isIntegerTy());
11190 unsigned Bits = Ty->getPrimitiveSizeInBits();
11191 if (Bits == 0 || Bits > 32)
11196 bool ARMTargetLowering::hasLoadLinkedStoreConditional() const { return true; }
11198 Instruction* ARMTargetLowering::makeDMB(IRBuilder<> &Builder,
11199 ARM_MB::MemBOpt Domain) const {
11200 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
11202 // First, if the target has no DMB, see what fallback we can use.
11203 if (!Subtarget->hasDataBarrier()) {
11204 // Some ARMv6 cpus can support data barriers with an mcr instruction.
11205 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
11207 if (Subtarget->hasV6Ops() && !Subtarget->isThumb()) {
11208 Function *MCR = llvm::Intrinsic::getDeclaration(M, Intrinsic::arm_mcr);
11209 Value* args[6] = {Builder.getInt32(15), Builder.getInt32(0),
11210 Builder.getInt32(0), Builder.getInt32(7),
11211 Builder.getInt32(10), Builder.getInt32(5)};
11212 return Builder.CreateCall(MCR, args);
11214 // Instead of using barriers, atomic accesses on these subtargets use
11216 llvm_unreachable("makeDMB on a target so old that it has no barriers");
11219 Function *DMB = llvm::Intrinsic::getDeclaration(M, Intrinsic::arm_dmb);
11220 // Only a full system barrier exists in the M-class architectures.
11221 Domain = Subtarget->isMClass() ? ARM_MB::SY : Domain;
11222 Constant *CDomain = Builder.getInt32(Domain);
11223 return Builder.CreateCall(DMB, CDomain);
11227 // Based on http://www.cl.cam.ac.uk/~pes20/cpp/cpp0xmappings.html
11228 Instruction* ARMTargetLowering::emitLeadingFence(IRBuilder<> &Builder,
11229 AtomicOrdering Ord, bool IsStore,
11230 bool IsLoad) const {
11231 if (!getInsertFencesForAtomic())
11237 llvm_unreachable("Invalid fence: unordered/non-atomic");
11240 return nullptr; // Nothing to do
11241 case SequentiallyConsistent:
11243 return nullptr; // Nothing to do
11246 case AcquireRelease:
11247 if (Subtarget->isSwift())
11248 return makeDMB(Builder, ARM_MB::ISHST);
11249 // FIXME: add a comment with a link to documentation justifying this.
11251 return makeDMB(Builder, ARM_MB::ISH);
11253 llvm_unreachable("Unknown fence ordering in emitLeadingFence");
11256 Instruction* ARMTargetLowering::emitTrailingFence(IRBuilder<> &Builder,
11257 AtomicOrdering Ord, bool IsStore,
11258 bool IsLoad) const {
11259 if (!getInsertFencesForAtomic())
11265 llvm_unreachable("Invalid fence: unordered/not-atomic");
11268 return nullptr; // Nothing to do
11270 case AcquireRelease:
11271 case SequentiallyConsistent:
11272 return makeDMB(Builder, ARM_MB::ISH);
11274 llvm_unreachable("Unknown fence ordering in emitTrailingFence");
11277 // Loads and stores less than 64-bits are already atomic; ones above that
11278 // are doomed anyway, so defer to the default libcall and blame the OS when
11279 // things go wrong. Cortex M doesn't have ldrexd/strexd though, so don't emit
11280 // anything for those.
11281 bool ARMTargetLowering::shouldExpandAtomicStoreInIR(StoreInst *SI) const {
11282 unsigned Size = SI->getValueOperand()->getType()->getPrimitiveSizeInBits();
11283 return (Size == 64) && !Subtarget->isMClass();
11286 // Loads and stores less than 64-bits are already atomic; ones above that
11287 // are doomed anyway, so defer to the default libcall and blame the OS when
11288 // things go wrong. Cortex M doesn't have ldrexd/strexd though, so don't emit
11289 // anything for those.
11290 // FIXME: ldrd and strd are atomic if the CPU has LPAE (e.g. A15 has that
11291 // guarantee, see DDI0406C ARM architecture reference manual,
11292 // sections A8.8.72-74 LDRD)
11293 bool ARMTargetLowering::shouldExpandAtomicLoadInIR(LoadInst *LI) const {
11294 unsigned Size = LI->getType()->getPrimitiveSizeInBits();
11295 return (Size == 64) && !Subtarget->isMClass();
11298 // For the real atomic operations, we have ldrex/strex up to 32 bits,
11299 // and up to 64 bits on the non-M profiles
11300 TargetLoweringBase::AtomicRMWExpansionKind
11301 ARMTargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
11302 unsigned Size = AI->getType()->getPrimitiveSizeInBits();
11303 return (Size <= (Subtarget->isMClass() ? 32U : 64U))
11304 ? AtomicRMWExpansionKind::LLSC
11305 : AtomicRMWExpansionKind::None;
11308 // This has so far only been implemented for MachO.
11309 bool ARMTargetLowering::useLoadStackGuardNode() const {
11310 return Subtarget->isTargetMachO();
11313 bool ARMTargetLowering::canCombineStoreAndExtract(Type *VectorTy, Value *Idx,
11314 unsigned &Cost) const {
11315 // If we do not have NEON, vector types are not natively supported.
11316 if (!Subtarget->hasNEON())
11319 // Floating point values and vector values map to the same register file.
11320 // Therefore, althought we could do a store extract of a vector type, this is
11321 // better to leave at float as we have more freedom in the addressing mode for
11323 if (VectorTy->isFPOrFPVectorTy())
11326 // If the index is unknown at compile time, this is very expensive to lower
11327 // and it is not possible to combine the store with the extract.
11328 if (!isa<ConstantInt>(Idx))
11331 assert(VectorTy->isVectorTy() && "VectorTy is not a vector type");
11332 unsigned BitWidth = cast<VectorType>(VectorTy)->getBitWidth();
11333 // We can do a store + vector extract on any vector that fits perfectly in a D
11335 if (BitWidth == 64 || BitWidth == 128) {
11342 Value *ARMTargetLowering::emitLoadLinked(IRBuilder<> &Builder, Value *Addr,
11343 AtomicOrdering Ord) const {
11344 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
11345 Type *ValTy = cast<PointerType>(Addr->getType())->getElementType();
11346 bool IsAcquire = isAtLeastAcquire(Ord);
11348 // Since i64 isn't legal and intrinsics don't get type-lowered, the ldrexd
11349 // intrinsic must return {i32, i32} and we have to recombine them into a
11350 // single i64 here.
11351 if (ValTy->getPrimitiveSizeInBits() == 64) {
11352 Intrinsic::ID Int =
11353 IsAcquire ? Intrinsic::arm_ldaexd : Intrinsic::arm_ldrexd;
11354 Function *Ldrex = llvm::Intrinsic::getDeclaration(M, Int);
11356 Addr = Builder.CreateBitCast(Addr, Type::getInt8PtrTy(M->getContext()));
11357 Value *LoHi = Builder.CreateCall(Ldrex, Addr, "lohi");
11359 Value *Lo = Builder.CreateExtractValue(LoHi, 0, "lo");
11360 Value *Hi = Builder.CreateExtractValue(LoHi, 1, "hi");
11361 if (!Subtarget->isLittle())
11362 std::swap (Lo, Hi);
11363 Lo = Builder.CreateZExt(Lo, ValTy, "lo64");
11364 Hi = Builder.CreateZExt(Hi, ValTy, "hi64");
11365 return Builder.CreateOr(
11366 Lo, Builder.CreateShl(Hi, ConstantInt::get(ValTy, 32)), "val64");
11369 Type *Tys[] = { Addr->getType() };
11370 Intrinsic::ID Int = IsAcquire ? Intrinsic::arm_ldaex : Intrinsic::arm_ldrex;
11371 Function *Ldrex = llvm::Intrinsic::getDeclaration(M, Int, Tys);
11373 return Builder.CreateTruncOrBitCast(
11374 Builder.CreateCall(Ldrex, Addr),
11375 cast<PointerType>(Addr->getType())->getElementType());
11378 Value *ARMTargetLowering::emitStoreConditional(IRBuilder<> &Builder, Value *Val,
11380 AtomicOrdering Ord) const {
11381 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
11382 bool IsRelease = isAtLeastRelease(Ord);
11384 // Since the intrinsics must have legal type, the i64 intrinsics take two
11385 // parameters: "i32, i32". We must marshal Val into the appropriate form
11386 // before the call.
11387 if (Val->getType()->getPrimitiveSizeInBits() == 64) {
11388 Intrinsic::ID Int =
11389 IsRelease ? Intrinsic::arm_stlexd : Intrinsic::arm_strexd;
11390 Function *Strex = Intrinsic::getDeclaration(M, Int);
11391 Type *Int32Ty = Type::getInt32Ty(M->getContext());
11393 Value *Lo = Builder.CreateTrunc(Val, Int32Ty, "lo");
11394 Value *Hi = Builder.CreateTrunc(Builder.CreateLShr(Val, 32), Int32Ty, "hi");
11395 if (!Subtarget->isLittle())
11396 std::swap (Lo, Hi);
11397 Addr = Builder.CreateBitCast(Addr, Type::getInt8PtrTy(M->getContext()));
11398 return Builder.CreateCall(Strex, {Lo, Hi, Addr});
11401 Intrinsic::ID Int = IsRelease ? Intrinsic::arm_stlex : Intrinsic::arm_strex;
11402 Type *Tys[] = { Addr->getType() };
11403 Function *Strex = Intrinsic::getDeclaration(M, Int, Tys);
11405 return Builder.CreateCall(
11406 Strex, {Builder.CreateZExtOrBitCast(
11407 Val, Strex->getFunctionType()->getParamType(0)),
11411 /// \brief Lower an interleaved load into a vldN intrinsic.
11413 /// E.g. Lower an interleaved load (Factor = 2):
11414 /// %wide.vec = load <8 x i32>, <8 x i32>* %ptr, align 4
11415 /// %v0 = shuffle %wide.vec, undef, <0, 2, 4, 6> ; Extract even elements
11416 /// %v1 = shuffle %wide.vec, undef, <1, 3, 5, 7> ; Extract odd elements
11419 /// %vld2 = { <4 x i32>, <4 x i32> } call llvm.arm.neon.vld2(%ptr, 4)
11420 /// %vec0 = extractelement { <4 x i32>, <4 x i32> } %vld2, i32 0
11421 /// %vec1 = extractelement { <4 x i32>, <4 x i32> } %vld2, i32 1
11422 bool ARMTargetLowering::lowerInterleavedLoad(
11423 LoadInst *LI, ArrayRef<ShuffleVectorInst *> Shuffles,
11424 ArrayRef<unsigned> Indices, unsigned Factor) const {
11425 assert(Factor >= 2 && Factor <= getMaxSupportedInterleaveFactor() &&
11426 "Invalid interleave factor");
11427 assert(!Shuffles.empty() && "Empty shufflevector input");
11428 assert(Shuffles.size() == Indices.size() &&
11429 "Unmatched number of shufflevectors and indices");
11431 VectorType *VecTy = Shuffles[0]->getType();
11432 Type *EltTy = VecTy->getVectorElementType();
11434 const DataLayout &DL = LI->getModule()->getDataLayout();
11435 unsigned VecSize = DL.getTypeAllocSizeInBits(VecTy);
11436 bool EltIs64Bits = DL.getTypeAllocSizeInBits(EltTy) == 64;
11438 // Skip illegal vector types and vector types of i64/f64 element (vldN doesn't
11439 // support i64/f64 element).
11440 if ((VecSize != 64 && VecSize != 128) || EltIs64Bits)
11443 // A pointer vector can not be the return type of the ldN intrinsics. Need to
11444 // load integer vectors first and then convert to pointer vectors.
11445 if (EltTy->isPointerTy())
11447 VectorType::get(DL.getIntPtrType(EltTy), VecTy->getVectorNumElements());
11449 static const Intrinsic::ID LoadInts[3] = {Intrinsic::arm_neon_vld2,
11450 Intrinsic::arm_neon_vld3,
11451 Intrinsic::arm_neon_vld4};
11453 Function *VldnFunc =
11454 Intrinsic::getDeclaration(LI->getModule(), LoadInts[Factor - 2], VecTy);
11456 IRBuilder<> Builder(LI);
11457 SmallVector<Value *, 2> Ops;
11459 Type *Int8Ptr = Builder.getInt8PtrTy(LI->getPointerAddressSpace());
11460 Ops.push_back(Builder.CreateBitCast(LI->getPointerOperand(), Int8Ptr));
11461 Ops.push_back(Builder.getInt32(LI->getAlignment()));
11463 CallInst *VldN = Builder.CreateCall(VldnFunc, Ops, "vldN");
11465 // Replace uses of each shufflevector with the corresponding vector loaded
11467 for (unsigned i = 0; i < Shuffles.size(); i++) {
11468 ShuffleVectorInst *SV = Shuffles[i];
11469 unsigned Index = Indices[i];
11471 Value *SubVec = Builder.CreateExtractValue(VldN, Index);
11473 // Convert the integer vector to pointer vector if the element is pointer.
11474 if (EltTy->isPointerTy())
11475 SubVec = Builder.CreateIntToPtr(SubVec, SV->getType());
11477 SV->replaceAllUsesWith(SubVec);
11483 /// \brief Get a mask consisting of sequential integers starting from \p Start.
11485 /// I.e. <Start, Start + 1, ..., Start + NumElts - 1>
11486 static Constant *getSequentialMask(IRBuilder<> &Builder, unsigned Start,
11487 unsigned NumElts) {
11488 SmallVector<Constant *, 16> Mask;
11489 for (unsigned i = 0; i < NumElts; i++)
11490 Mask.push_back(Builder.getInt32(Start + i));
11492 return ConstantVector::get(Mask);
11495 /// \brief Lower an interleaved store into a vstN intrinsic.
11497 /// E.g. Lower an interleaved store (Factor = 3):
11498 /// %i.vec = shuffle <8 x i32> %v0, <8 x i32> %v1,
11499 /// <0, 4, 8, 1, 5, 9, 2, 6, 10, 3, 7, 11>
11500 /// store <12 x i32> %i.vec, <12 x i32>* %ptr, align 4
11503 /// %sub.v0 = shuffle <8 x i32> %v0, <8 x i32> v1, <0, 1, 2, 3>
11504 /// %sub.v1 = shuffle <8 x i32> %v0, <8 x i32> v1, <4, 5, 6, 7>
11505 /// %sub.v2 = shuffle <8 x i32> %v0, <8 x i32> v1, <8, 9, 10, 11>
11506 /// call void llvm.arm.neon.vst3(%ptr, %sub.v0, %sub.v1, %sub.v2, 4)
11508 /// Note that the new shufflevectors will be removed and we'll only generate one
11509 /// vst3 instruction in CodeGen.
11510 bool ARMTargetLowering::lowerInterleavedStore(StoreInst *SI,
11511 ShuffleVectorInst *SVI,
11512 unsigned Factor) const {
11513 assert(Factor >= 2 && Factor <= getMaxSupportedInterleaveFactor() &&
11514 "Invalid interleave factor");
11516 VectorType *VecTy = SVI->getType();
11517 assert(VecTy->getVectorNumElements() % Factor == 0 &&
11518 "Invalid interleaved store");
11520 unsigned NumSubElts = VecTy->getVectorNumElements() / Factor;
11521 Type *EltTy = VecTy->getVectorElementType();
11522 VectorType *SubVecTy = VectorType::get(EltTy, NumSubElts);
11524 const DataLayout &DL = SI->getModule()->getDataLayout();
11525 unsigned SubVecSize = DL.getTypeAllocSizeInBits(SubVecTy);
11526 bool EltIs64Bits = DL.getTypeAllocSizeInBits(EltTy) == 64;
11528 // Skip illegal sub vector types and vector types of i64/f64 element (vstN
11529 // doesn't support i64/f64 element).
11530 if ((SubVecSize != 64 && SubVecSize != 128) || EltIs64Bits)
11533 Value *Op0 = SVI->getOperand(0);
11534 Value *Op1 = SVI->getOperand(1);
11535 IRBuilder<> Builder(SI);
11537 // StN intrinsics don't support pointer vectors as arguments. Convert pointer
11538 // vectors to integer vectors.
11539 if (EltTy->isPointerTy()) {
11540 Type *IntTy = DL.getIntPtrType(EltTy);
11542 // Convert to the corresponding integer vector.
11544 VectorType::get(IntTy, Op0->getType()->getVectorNumElements());
11545 Op0 = Builder.CreatePtrToInt(Op0, IntVecTy);
11546 Op1 = Builder.CreatePtrToInt(Op1, IntVecTy);
11548 SubVecTy = VectorType::get(IntTy, NumSubElts);
11551 static Intrinsic::ID StoreInts[3] = {Intrinsic::arm_neon_vst2,
11552 Intrinsic::arm_neon_vst3,
11553 Intrinsic::arm_neon_vst4};
11554 Function *VstNFunc = Intrinsic::getDeclaration(
11555 SI->getModule(), StoreInts[Factor - 2], SubVecTy);
11557 SmallVector<Value *, 6> Ops;
11559 Type *Int8Ptr = Builder.getInt8PtrTy(SI->getPointerAddressSpace());
11560 Ops.push_back(Builder.CreateBitCast(SI->getPointerOperand(), Int8Ptr));
11562 // Split the shufflevector operands into sub vectors for the new vstN call.
11563 for (unsigned i = 0; i < Factor; i++)
11564 Ops.push_back(Builder.CreateShuffleVector(
11565 Op0, Op1, getSequentialMask(Builder, NumSubElts * i, NumSubElts)));
11567 Ops.push_back(Builder.getInt32(SI->getAlignment()));
11568 Builder.CreateCall(VstNFunc, Ops);
11580 static bool isHomogeneousAggregate(Type *Ty, HABaseType &Base,
11581 uint64_t &Members) {
11582 if (const StructType *ST = dyn_cast<StructType>(Ty)) {
11583 for (unsigned i = 0; i < ST->getNumElements(); ++i) {
11584 uint64_t SubMembers = 0;
11585 if (!isHomogeneousAggregate(ST->getElementType(i), Base, SubMembers))
11587 Members += SubMembers;
11589 } else if (const ArrayType *AT = dyn_cast<ArrayType>(Ty)) {
11590 uint64_t SubMembers = 0;
11591 if (!isHomogeneousAggregate(AT->getElementType(), Base, SubMembers))
11593 Members += SubMembers * AT->getNumElements();
11594 } else if (Ty->isFloatTy()) {
11595 if (Base != HA_UNKNOWN && Base != HA_FLOAT)
11599 } else if (Ty->isDoubleTy()) {
11600 if (Base != HA_UNKNOWN && Base != HA_DOUBLE)
11604 } else if (const VectorType *VT = dyn_cast<VectorType>(Ty)) {
11611 return VT->getBitWidth() == 64;
11613 return VT->getBitWidth() == 128;
11615 switch (VT->getBitWidth()) {
11628 return (Members > 0 && Members <= 4);
11631 /// \brief Return true if a type is an AAPCS-VFP homogeneous aggregate or one of
11632 /// [N x i32] or [N x i64]. This allows front-ends to skip emitting padding when
11633 /// passing according to AAPCS rules.
11634 bool ARMTargetLowering::functionArgumentNeedsConsecutiveRegisters(
11635 Type *Ty, CallingConv::ID CallConv, bool isVarArg) const {
11636 if (getEffectiveCallingConv(CallConv, isVarArg) !=
11637 CallingConv::ARM_AAPCS_VFP)
11640 HABaseType Base = HA_UNKNOWN;
11641 uint64_t Members = 0;
11642 bool IsHA = isHomogeneousAggregate(Ty, Base, Members);
11643 DEBUG(dbgs() << "isHA: " << IsHA << " "; Ty->dump());
11645 bool IsIntArray = Ty->isArrayTy() && Ty->getArrayElementType()->isIntegerTy();
11646 return IsHA || IsIntArray;