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(LLVMContext &, EVT VT) const {
1148 if (!VT.isVector()) return getPointerTy();
1149 return VT.changeVectorElementTypeToInteger();
1152 /// getRegClassFor - Return the register class that should be used for the
1153 /// specified value type.
1154 const TargetRegisterClass *ARMTargetLowering::getRegClassFor(MVT VT) const {
1155 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
1156 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
1157 // load / store 4 to 8 consecutive D registers.
1158 if (Subtarget->hasNEON()) {
1159 if (VT == MVT::v4i64)
1160 return &ARM::QQPRRegClass;
1161 if (VT == MVT::v8i64)
1162 return &ARM::QQQQPRRegClass;
1164 return TargetLowering::getRegClassFor(VT);
1167 // memcpy, and other memory intrinsics, typically tries to use LDM/STM if the
1168 // source/dest is aligned and the copy size is large enough. We therefore want
1169 // to align such objects passed to memory intrinsics.
1170 bool ARMTargetLowering::shouldAlignPointerArgs(CallInst *CI, unsigned &MinSize,
1171 unsigned &PrefAlign) const {
1172 if (!isa<MemIntrinsic>(CI))
1175 // On ARM11 onwards (excluding M class) 8-byte aligned LDM is typically 1
1176 // cycle faster than 4-byte aligned LDM.
1177 PrefAlign = (Subtarget->hasV6Ops() && !Subtarget->isMClass() ? 8 : 4);
1181 // Create a fast isel object.
1183 ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
1184 const TargetLibraryInfo *libInfo) const {
1185 return ARM::createFastISel(funcInfo, libInfo);
1188 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
1189 unsigned NumVals = N->getNumValues();
1191 return Sched::RegPressure;
1193 for (unsigned i = 0; i != NumVals; ++i) {
1194 EVT VT = N->getValueType(i);
1195 if (VT == MVT::Glue || VT == MVT::Other)
1197 if (VT.isFloatingPoint() || VT.isVector())
1201 if (!N->isMachineOpcode())
1202 return Sched::RegPressure;
1204 // Load are scheduled for latency even if there instruction itinerary
1205 // is not available.
1206 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
1207 const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
1209 if (MCID.getNumDefs() == 0)
1210 return Sched::RegPressure;
1211 if (!Itins->isEmpty() &&
1212 Itins->getOperandCycle(MCID.getSchedClass(), 0) > 2)
1215 return Sched::RegPressure;
1218 //===----------------------------------------------------------------------===//
1220 //===----------------------------------------------------------------------===//
1222 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
1223 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
1225 default: llvm_unreachable("Unknown condition code!");
1226 case ISD::SETNE: return ARMCC::NE;
1227 case ISD::SETEQ: return ARMCC::EQ;
1228 case ISD::SETGT: return ARMCC::GT;
1229 case ISD::SETGE: return ARMCC::GE;
1230 case ISD::SETLT: return ARMCC::LT;
1231 case ISD::SETLE: return ARMCC::LE;
1232 case ISD::SETUGT: return ARMCC::HI;
1233 case ISD::SETUGE: return ARMCC::HS;
1234 case ISD::SETULT: return ARMCC::LO;
1235 case ISD::SETULE: return ARMCC::LS;
1239 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
1240 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
1241 ARMCC::CondCodes &CondCode2) {
1242 CondCode2 = ARMCC::AL;
1244 default: llvm_unreachable("Unknown FP condition!");
1246 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
1248 case ISD::SETOGT: CondCode = ARMCC::GT; break;
1250 case ISD::SETOGE: CondCode = ARMCC::GE; break;
1251 case ISD::SETOLT: CondCode = ARMCC::MI; break;
1252 case ISD::SETOLE: CondCode = ARMCC::LS; break;
1253 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
1254 case ISD::SETO: CondCode = ARMCC::VC; break;
1255 case ISD::SETUO: CondCode = ARMCC::VS; break;
1256 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
1257 case ISD::SETUGT: CondCode = ARMCC::HI; break;
1258 case ISD::SETUGE: CondCode = ARMCC::PL; break;
1260 case ISD::SETULT: CondCode = ARMCC::LT; break;
1262 case ISD::SETULE: CondCode = ARMCC::LE; break;
1264 case ISD::SETUNE: CondCode = ARMCC::NE; break;
1268 //===----------------------------------------------------------------------===//
1269 // Calling Convention Implementation
1270 //===----------------------------------------------------------------------===//
1272 #include "ARMGenCallingConv.inc"
1274 /// getEffectiveCallingConv - Get the effective calling convention, taking into
1275 /// account presence of floating point hardware and calling convention
1276 /// limitations, such as support for variadic functions.
1278 ARMTargetLowering::getEffectiveCallingConv(CallingConv::ID CC,
1279 bool isVarArg) const {
1282 llvm_unreachable("Unsupported calling convention");
1283 case CallingConv::ARM_AAPCS:
1284 case CallingConv::ARM_APCS:
1285 case CallingConv::GHC:
1287 case CallingConv::ARM_AAPCS_VFP:
1288 return isVarArg ? CallingConv::ARM_AAPCS : CallingConv::ARM_AAPCS_VFP;
1289 case CallingConv::C:
1290 if (!Subtarget->isAAPCS_ABI())
1291 return CallingConv::ARM_APCS;
1292 else if (Subtarget->hasVFP2() && !Subtarget->isThumb1Only() &&
1293 getTargetMachine().Options.FloatABIType == FloatABI::Hard &&
1295 return CallingConv::ARM_AAPCS_VFP;
1297 return CallingConv::ARM_AAPCS;
1298 case CallingConv::Fast:
1299 if (!Subtarget->isAAPCS_ABI()) {
1300 if (Subtarget->hasVFP2() && !Subtarget->isThumb1Only() && !isVarArg)
1301 return CallingConv::Fast;
1302 return CallingConv::ARM_APCS;
1303 } else if (Subtarget->hasVFP2() && !Subtarget->isThumb1Only() && !isVarArg)
1304 return CallingConv::ARM_AAPCS_VFP;
1306 return CallingConv::ARM_AAPCS;
1310 /// CCAssignFnForNode - Selects the correct CCAssignFn for the given
1311 /// CallingConvention.
1312 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
1314 bool isVarArg) const {
1315 switch (getEffectiveCallingConv(CC, isVarArg)) {
1317 llvm_unreachable("Unsupported calling convention");
1318 case CallingConv::ARM_APCS:
1319 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1320 case CallingConv::ARM_AAPCS:
1321 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1322 case CallingConv::ARM_AAPCS_VFP:
1323 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1324 case CallingConv::Fast:
1325 return (Return ? RetFastCC_ARM_APCS : FastCC_ARM_APCS);
1326 case CallingConv::GHC:
1327 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS_GHC);
1331 /// LowerCallResult - Lower the result values of a call into the
1332 /// appropriate copies out of appropriate physical registers.
1334 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1335 CallingConv::ID CallConv, bool isVarArg,
1336 const SmallVectorImpl<ISD::InputArg> &Ins,
1337 SDLoc dl, SelectionDAG &DAG,
1338 SmallVectorImpl<SDValue> &InVals,
1339 bool isThisReturn, SDValue ThisVal) const {
1341 // Assign locations to each value returned by this call.
1342 SmallVector<CCValAssign, 16> RVLocs;
1343 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
1344 *DAG.getContext(), Call);
1345 CCInfo.AnalyzeCallResult(Ins,
1346 CCAssignFnForNode(CallConv, /* Return*/ true,
1349 // Copy all of the result registers out of their specified physreg.
1350 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1351 CCValAssign VA = RVLocs[i];
1353 // Pass 'this' value directly from the argument to return value, to avoid
1354 // reg unit interference
1355 if (i == 0 && isThisReturn) {
1356 assert(!VA.needsCustom() && VA.getLocVT() == MVT::i32 &&
1357 "unexpected return calling convention register assignment");
1358 InVals.push_back(ThisVal);
1363 if (VA.needsCustom()) {
1364 // Handle f64 or half of a v2f64.
1365 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1367 Chain = Lo.getValue(1);
1368 InFlag = Lo.getValue(2);
1369 VA = RVLocs[++i]; // skip ahead to next loc
1370 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1372 Chain = Hi.getValue(1);
1373 InFlag = Hi.getValue(2);
1374 if (!Subtarget->isLittle())
1376 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1378 if (VA.getLocVT() == MVT::v2f64) {
1379 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1380 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1381 DAG.getConstant(0, dl, MVT::i32));
1383 VA = RVLocs[++i]; // skip ahead to next loc
1384 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1385 Chain = Lo.getValue(1);
1386 InFlag = Lo.getValue(2);
1387 VA = RVLocs[++i]; // skip ahead to next loc
1388 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1389 Chain = Hi.getValue(1);
1390 InFlag = Hi.getValue(2);
1391 if (!Subtarget->isLittle())
1393 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1394 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1395 DAG.getConstant(1, dl, MVT::i32));
1398 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
1400 Chain = Val.getValue(1);
1401 InFlag = Val.getValue(2);
1404 switch (VA.getLocInfo()) {
1405 default: llvm_unreachable("Unknown loc info!");
1406 case CCValAssign::Full: break;
1407 case CCValAssign::BCvt:
1408 Val = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), Val);
1412 InVals.push_back(Val);
1418 /// LowerMemOpCallTo - Store the argument to the stack.
1420 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
1421 SDValue StackPtr, SDValue Arg,
1422 SDLoc dl, SelectionDAG &DAG,
1423 const CCValAssign &VA,
1424 ISD::ArgFlagsTy Flags) const {
1425 unsigned LocMemOffset = VA.getLocMemOffset();
1426 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset, dl);
1427 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1428 return DAG.getStore(Chain, dl, Arg, PtrOff,
1429 MachinePointerInfo::getStack(LocMemOffset),
1433 void ARMTargetLowering::PassF64ArgInRegs(SDLoc dl, SelectionDAG &DAG,
1434 SDValue Chain, SDValue &Arg,
1435 RegsToPassVector &RegsToPass,
1436 CCValAssign &VA, CCValAssign &NextVA,
1438 SmallVectorImpl<SDValue> &MemOpChains,
1439 ISD::ArgFlagsTy Flags) const {
1441 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1442 DAG.getVTList(MVT::i32, MVT::i32), Arg);
1443 unsigned id = Subtarget->isLittle() ? 0 : 1;
1444 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd.getValue(id)));
1446 if (NextVA.isRegLoc())
1447 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1-id)));
1449 assert(NextVA.isMemLoc());
1450 if (!StackPtr.getNode())
1451 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1453 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1-id),
1459 /// LowerCall - Lowering a call into a callseq_start <-
1460 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
1463 ARMTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
1464 SmallVectorImpl<SDValue> &InVals) const {
1465 SelectionDAG &DAG = CLI.DAG;
1467 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
1468 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
1469 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
1470 SDValue Chain = CLI.Chain;
1471 SDValue Callee = CLI.Callee;
1472 bool &isTailCall = CLI.IsTailCall;
1473 CallingConv::ID CallConv = CLI.CallConv;
1474 bool doesNotRet = CLI.DoesNotReturn;
1475 bool isVarArg = CLI.IsVarArg;
1477 MachineFunction &MF = DAG.getMachineFunction();
1478 bool isStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1479 bool isThisReturn = false;
1480 bool isSibCall = false;
1481 auto Attr = MF.getFunction()->getFnAttribute("disable-tail-calls");
1483 // Disable tail calls if they're not supported.
1484 if (!Subtarget->supportsTailCall() || Attr.getValueAsString() == "true")
1488 // Check if it's really possible to do a tail call.
1489 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1490 isVarArg, isStructRet, MF.getFunction()->hasStructRetAttr(),
1491 Outs, OutVals, Ins, DAG);
1492 if (!isTailCall && CLI.CS && CLI.CS->isMustTailCall())
1493 report_fatal_error("failed to perform tail call elimination on a call "
1494 "site marked musttail");
1495 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1496 // detected sibcalls.
1503 // Analyze operands of the call, assigning locations to each operand.
1504 SmallVector<CCValAssign, 16> ArgLocs;
1505 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs,
1506 *DAG.getContext(), Call);
1507 CCInfo.AnalyzeCallOperands(Outs,
1508 CCAssignFnForNode(CallConv, /* Return*/ false,
1511 // Get a count of how many bytes are to be pushed on the stack.
1512 unsigned NumBytes = CCInfo.getNextStackOffset();
1514 // For tail calls, memory operands are available in our caller's stack.
1518 // Adjust the stack pointer for the new arguments...
1519 // These operations are automatically eliminated by the prolog/epilog pass
1521 Chain = DAG.getCALLSEQ_START(Chain,
1522 DAG.getIntPtrConstant(NumBytes, dl, true), dl);
1524 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1526 RegsToPassVector RegsToPass;
1527 SmallVector<SDValue, 8> MemOpChains;
1529 // Walk the register/memloc assignments, inserting copies/loads. In the case
1530 // of tail call optimization, arguments are handled later.
1531 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1533 ++i, ++realArgIdx) {
1534 CCValAssign &VA = ArgLocs[i];
1535 SDValue Arg = OutVals[realArgIdx];
1536 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1537 bool isByVal = Flags.isByVal();
1539 // Promote the value if needed.
1540 switch (VA.getLocInfo()) {
1541 default: llvm_unreachable("Unknown loc info!");
1542 case CCValAssign::Full: break;
1543 case CCValAssign::SExt:
1544 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1546 case CCValAssign::ZExt:
1547 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1549 case CCValAssign::AExt:
1550 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1552 case CCValAssign::BCvt:
1553 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1557 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1558 if (VA.needsCustom()) {
1559 if (VA.getLocVT() == MVT::v2f64) {
1560 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1561 DAG.getConstant(0, dl, MVT::i32));
1562 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1563 DAG.getConstant(1, dl, MVT::i32));
1565 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1566 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1568 VA = ArgLocs[++i]; // skip ahead to next loc
1569 if (VA.isRegLoc()) {
1570 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1571 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1573 assert(VA.isMemLoc());
1575 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1576 dl, DAG, VA, Flags));
1579 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1580 StackPtr, MemOpChains, Flags);
1582 } else if (VA.isRegLoc()) {
1583 if (realArgIdx == 0 && Flags.isReturned() && Outs[0].VT == MVT::i32) {
1584 assert(VA.getLocVT() == MVT::i32 &&
1585 "unexpected calling convention register assignment");
1586 assert(!Ins.empty() && Ins[0].VT == MVT::i32 &&
1587 "unexpected use of 'returned'");
1588 isThisReturn = true;
1590 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1591 } else if (isByVal) {
1592 assert(VA.isMemLoc());
1593 unsigned offset = 0;
1595 // True if this byval aggregate will be split between registers
1597 unsigned ByValArgsCount = CCInfo.getInRegsParamsCount();
1598 unsigned CurByValIdx = CCInfo.getInRegsParamsProcessed();
1600 if (CurByValIdx < ByValArgsCount) {
1602 unsigned RegBegin, RegEnd;
1603 CCInfo.getInRegsParamInfo(CurByValIdx, RegBegin, RegEnd);
1605 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1607 for (i = 0, j = RegBegin; j < RegEnd; i++, j++) {
1608 SDValue Const = DAG.getConstant(4*i, dl, MVT::i32);
1609 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
1610 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
1611 MachinePointerInfo(),
1612 false, false, false,
1613 DAG.InferPtrAlignment(AddArg));
1614 MemOpChains.push_back(Load.getValue(1));
1615 RegsToPass.push_back(std::make_pair(j, Load));
1618 // If parameter size outsides register area, "offset" value
1619 // helps us to calculate stack slot for remained part properly.
1620 offset = RegEnd - RegBegin;
1622 CCInfo.nextInRegsParam();
1625 if (Flags.getByValSize() > 4*offset) {
1626 unsigned LocMemOffset = VA.getLocMemOffset();
1627 SDValue StkPtrOff = DAG.getIntPtrConstant(LocMemOffset, dl);
1628 SDValue Dst = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr,
1630 SDValue SrcOffset = DAG.getIntPtrConstant(4*offset, dl);
1631 SDValue Src = DAG.getNode(ISD::ADD, dl, getPointerTy(), Arg, SrcOffset);
1632 SDValue SizeNode = DAG.getConstant(Flags.getByValSize() - 4*offset, dl,
1634 SDValue AlignNode = DAG.getConstant(Flags.getByValAlign(), dl,
1637 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
1638 SDValue Ops[] = { Chain, Dst, Src, SizeNode, AlignNode};
1639 MemOpChains.push_back(DAG.getNode(ARMISD::COPY_STRUCT_BYVAL, dl, VTs,
1642 } else if (!isSibCall) {
1643 assert(VA.isMemLoc());
1645 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1646 dl, DAG, VA, Flags));
1650 if (!MemOpChains.empty())
1651 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains);
1653 // Build a sequence of copy-to-reg nodes chained together with token chain
1654 // and flag operands which copy the outgoing args into the appropriate regs.
1656 // Tail call byval lowering might overwrite argument registers so in case of
1657 // tail call optimization the copies to registers are lowered later.
1659 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1660 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1661 RegsToPass[i].second, InFlag);
1662 InFlag = Chain.getValue(1);
1665 // For tail calls lower the arguments to the 'real' stack slot.
1667 // Force all the incoming stack arguments to be loaded from the stack
1668 // before any new outgoing arguments are stored to the stack, because the
1669 // outgoing stack slots may alias the incoming argument stack slots, and
1670 // the alias isn't otherwise explicit. This is slightly more conservative
1671 // than necessary, because it means that each store effectively depends
1672 // on every argument instead of just those arguments it would clobber.
1674 // Do not flag preceding copytoreg stuff together with the following stuff.
1676 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1677 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1678 RegsToPass[i].second, InFlag);
1679 InFlag = Chain.getValue(1);
1684 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1685 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1686 // node so that legalize doesn't hack it.
1687 bool isDirect = false;
1688 bool isARMFunc = false;
1689 bool isLocalARMFunc = false;
1690 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1692 if (Subtarget->genLongCalls()) {
1693 assert((Subtarget->isTargetWindows() ||
1694 getTargetMachine().getRelocationModel() == Reloc::Static) &&
1695 "long-calls with non-static relocation model!");
1696 // Handle a global address or an external symbol. If it's not one of
1697 // those, the target's already in a register, so we don't need to do
1699 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1700 const GlobalValue *GV = G->getGlobal();
1701 // Create a constant pool entry for the callee address
1702 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1703 ARMConstantPoolValue *CPV =
1704 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 0);
1706 // Get the address of the callee into a register
1707 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1708 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1709 Callee = DAG.getLoad(getPointerTy(), dl,
1710 DAG.getEntryNode(), CPAddr,
1711 MachinePointerInfo::getConstantPool(),
1712 false, false, false, 0);
1713 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1714 const char *Sym = S->getSymbol();
1716 // Create a constant pool entry for the callee address
1717 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1718 ARMConstantPoolValue *CPV =
1719 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1720 ARMPCLabelIndex, 0);
1721 // Get the address of the callee into a register
1722 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1723 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1724 Callee = DAG.getLoad(getPointerTy(), dl,
1725 DAG.getEntryNode(), CPAddr,
1726 MachinePointerInfo::getConstantPool(),
1727 false, false, false, 0);
1729 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1730 const GlobalValue *GV = G->getGlobal();
1732 bool isDef = GV->isStrongDefinitionForLinker();
1733 bool isStub = (!isDef && Subtarget->isTargetMachO()) &&
1734 getTargetMachine().getRelocationModel() != Reloc::Static;
1735 isARMFunc = !Subtarget->isThumb() || (isStub && !Subtarget->isMClass());
1736 // ARM call to a local ARM function is predicable.
1737 isLocalARMFunc = !Subtarget->isThumb() && (isDef || !ARMInterworking);
1738 // tBX takes a register source operand.
1739 if (isStub && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1740 assert(Subtarget->isTargetMachO() && "WrapperPIC use on non-MachO?");
1741 Callee = DAG.getNode(ARMISD::WrapperPIC, dl, getPointerTy(),
1742 DAG.getTargetGlobalAddress(GV, dl, getPointerTy(),
1743 0, ARMII::MO_NONLAZY));
1744 Callee = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Callee,
1745 MachinePointerInfo::getGOT(), false, false, true, 0);
1746 } else if (Subtarget->isTargetCOFF()) {
1747 assert(Subtarget->isTargetWindows() &&
1748 "Windows is the only supported COFF target");
1749 unsigned TargetFlags = GV->hasDLLImportStorageClass()
1750 ? ARMII::MO_DLLIMPORT
1751 : ARMII::MO_NO_FLAG;
1752 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), /*Offset=*/0,
1754 if (GV->hasDLLImportStorageClass())
1755 Callee = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
1756 DAG.getNode(ARMISD::Wrapper, dl, getPointerTy(),
1757 Callee), MachinePointerInfo::getGOT(),
1758 false, false, false, 0);
1760 // On ELF targets for PIC code, direct calls should go through the PLT
1761 unsigned OpFlags = 0;
1762 if (Subtarget->isTargetELF() &&
1763 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1764 OpFlags = ARMII::MO_PLT;
1765 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
1767 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1769 bool isStub = Subtarget->isTargetMachO() &&
1770 getTargetMachine().getRelocationModel() != Reloc::Static;
1771 isARMFunc = !Subtarget->isThumb() || (isStub && !Subtarget->isMClass());
1772 // tBX takes a register source operand.
1773 const char *Sym = S->getSymbol();
1774 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1775 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1776 ARMConstantPoolValue *CPV =
1777 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1778 ARMPCLabelIndex, 4);
1779 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1780 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1781 Callee = DAG.getLoad(getPointerTy(), dl,
1782 DAG.getEntryNode(), CPAddr,
1783 MachinePointerInfo::getConstantPool(),
1784 false, false, false, 0);
1785 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, dl, MVT::i32);
1786 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1787 getPointerTy(), Callee, PICLabel);
1789 unsigned OpFlags = 0;
1790 // On ELF targets for PIC code, direct calls should go through the PLT
1791 if (Subtarget->isTargetELF() &&
1792 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1793 OpFlags = ARMII::MO_PLT;
1794 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlags);
1798 // FIXME: handle tail calls differently.
1800 bool HasMinSizeAttr = MF.getFunction()->hasFnAttribute(Attribute::MinSize);
1801 if (Subtarget->isThumb()) {
1802 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1803 CallOpc = ARMISD::CALL_NOLINK;
1805 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1807 if (!isDirect && !Subtarget->hasV5TOps())
1808 CallOpc = ARMISD::CALL_NOLINK;
1809 else if (doesNotRet && isDirect && Subtarget->hasRAS() &&
1810 // Emit regular call when code size is the priority
1812 // "mov lr, pc; b _foo" to avoid confusing the RSP
1813 CallOpc = ARMISD::CALL_NOLINK;
1815 CallOpc = isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL;
1818 std::vector<SDValue> Ops;
1819 Ops.push_back(Chain);
1820 Ops.push_back(Callee);
1822 // Add argument registers to the end of the list so that they are known live
1824 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1825 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1826 RegsToPass[i].second.getValueType()));
1828 // Add a register mask operand representing the call-preserved registers.
1830 const uint32_t *Mask;
1831 const ARMBaseRegisterInfo *ARI = Subtarget->getRegisterInfo();
1833 // For 'this' returns, use the R0-preserving mask if applicable
1834 Mask = ARI->getThisReturnPreservedMask(MF, CallConv);
1836 // Set isThisReturn to false if the calling convention is not one that
1837 // allows 'returned' to be modeled in this way, so LowerCallResult does
1838 // not try to pass 'this' straight through
1839 isThisReturn = false;
1840 Mask = ARI->getCallPreservedMask(MF, CallConv);
1843 Mask = ARI->getCallPreservedMask(MF, CallConv);
1845 assert(Mask && "Missing call preserved mask for calling convention");
1846 Ops.push_back(DAG.getRegisterMask(Mask));
1849 if (InFlag.getNode())
1850 Ops.push_back(InFlag);
1852 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
1854 MF.getFrameInfo()->setHasTailCall();
1855 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, Ops);
1858 // Returns a chain and a flag for retval copy to use.
1859 Chain = DAG.getNode(CallOpc, dl, NodeTys, Ops);
1860 InFlag = Chain.getValue(1);
1862 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, dl, true),
1863 DAG.getIntPtrConstant(0, dl, true), InFlag, dl);
1865 InFlag = Chain.getValue(1);
1867 // Handle result values, copying them out of physregs into vregs that we
1869 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins, dl, DAG,
1870 InVals, isThisReturn,
1871 isThisReturn ? OutVals[0] : SDValue());
1874 /// HandleByVal - Every parameter *after* a byval parameter is passed
1875 /// on the stack. Remember the next parameter register to allocate,
1876 /// and then confiscate the rest of the parameter registers to insure
1878 void ARMTargetLowering::HandleByVal(CCState *State, unsigned &Size,
1879 unsigned Align) const {
1880 assert((State->getCallOrPrologue() == Prologue ||
1881 State->getCallOrPrologue() == Call) &&
1882 "unhandled ParmContext");
1884 // Byval (as with any stack) slots are always at least 4 byte aligned.
1885 Align = std::max(Align, 4U);
1887 unsigned Reg = State->AllocateReg(GPRArgRegs);
1891 unsigned AlignInRegs = Align / 4;
1892 unsigned Waste = (ARM::R4 - Reg) % AlignInRegs;
1893 for (unsigned i = 0; i < Waste; ++i)
1894 Reg = State->AllocateReg(GPRArgRegs);
1899 unsigned Excess = 4 * (ARM::R4 - Reg);
1901 // Special case when NSAA != SP and parameter size greater than size of
1902 // all remained GPR regs. In that case we can't split parameter, we must
1903 // send it to stack. We also must set NCRN to R4, so waste all
1904 // remained registers.
1905 const unsigned NSAAOffset = State->getNextStackOffset();
1906 if (NSAAOffset != 0 && Size > Excess) {
1907 while (State->AllocateReg(GPRArgRegs))
1912 // First register for byval parameter is the first register that wasn't
1913 // allocated before this method call, so it would be "reg".
1914 // If parameter is small enough to be saved in range [reg, r4), then
1915 // the end (first after last) register would be reg + param-size-in-regs,
1916 // else parameter would be splitted between registers and stack,
1917 // end register would be r4 in this case.
1918 unsigned ByValRegBegin = Reg;
1919 unsigned ByValRegEnd = std::min<unsigned>(Reg + Size / 4, ARM::R4);
1920 State->addInRegsParamInfo(ByValRegBegin, ByValRegEnd);
1921 // Note, first register is allocated in the beginning of function already,
1922 // allocate remained amount of registers we need.
1923 for (unsigned i = Reg + 1; i != ByValRegEnd; ++i)
1924 State->AllocateReg(GPRArgRegs);
1925 // A byval parameter that is split between registers and memory needs its
1926 // size truncated here.
1927 // In the case where the entire structure fits in registers, we set the
1928 // size in memory to zero.
1929 Size = std::max<int>(Size - Excess, 0);
1932 /// MatchingStackOffset - Return true if the given stack call argument is
1933 /// already available in the same position (relatively) of the caller's
1934 /// incoming argument stack.
1936 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1937 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1938 const TargetInstrInfo *TII) {
1939 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1941 if (Arg.getOpcode() == ISD::CopyFromReg) {
1942 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1943 if (!TargetRegisterInfo::isVirtualRegister(VR))
1945 MachineInstr *Def = MRI->getVRegDef(VR);
1948 if (!Flags.isByVal()) {
1949 if (!TII->isLoadFromStackSlot(Def, FI))
1954 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
1955 if (Flags.isByVal())
1956 // ByVal argument is passed in as a pointer but it's now being
1957 // dereferenced. e.g.
1958 // define @foo(%struct.X* %A) {
1959 // tail call @bar(%struct.X* byval %A)
1962 SDValue Ptr = Ld->getBasePtr();
1963 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
1966 FI = FINode->getIndex();
1970 assert(FI != INT_MAX);
1971 if (!MFI->isFixedObjectIndex(FI))
1973 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
1976 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
1977 /// for tail call optimization. Targets which want to do tail call
1978 /// optimization should implement this function.
1980 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1981 CallingConv::ID CalleeCC,
1983 bool isCalleeStructRet,
1984 bool isCallerStructRet,
1985 const SmallVectorImpl<ISD::OutputArg> &Outs,
1986 const SmallVectorImpl<SDValue> &OutVals,
1987 const SmallVectorImpl<ISD::InputArg> &Ins,
1988 SelectionDAG& DAG) const {
1989 const Function *CallerF = DAG.getMachineFunction().getFunction();
1990 CallingConv::ID CallerCC = CallerF->getCallingConv();
1991 bool CCMatch = CallerCC == CalleeCC;
1993 // Look for obvious safe cases to perform tail call optimization that do not
1994 // require ABI changes. This is what gcc calls sibcall.
1996 // Do not sibcall optimize vararg calls unless the call site is not passing
1998 if (isVarArg && !Outs.empty())
2001 // Exception-handling functions need a special set of instructions to indicate
2002 // a return to the hardware. Tail-calling another function would probably
2004 if (CallerF->hasFnAttribute("interrupt"))
2007 // Also avoid sibcall optimization if either caller or callee uses struct
2008 // return semantics.
2009 if (isCalleeStructRet || isCallerStructRet)
2012 // FIXME: Completely disable sibcall for Thumb1 since ThumbRegisterInfo::
2013 // emitEpilogue is not ready for them. Thumb tail calls also use t2B, as
2014 // the Thumb1 16-bit unconditional branch doesn't have sufficient relocation
2015 // support in the assembler and linker to be used. This would need to be
2016 // fixed to fully support tail calls in Thumb1.
2018 // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take
2019 // LR. This means if we need to reload LR, it takes an extra instructions,
2020 // which outweighs the value of the tail call; but here we don't know yet
2021 // whether LR is going to be used. Probably the right approach is to
2022 // generate the tail call here and turn it back into CALL/RET in
2023 // emitEpilogue if LR is used.
2025 // Thumb1 PIC calls to external symbols use BX, so they can be tail calls,
2026 // but we need to make sure there are enough registers; the only valid
2027 // registers are the 4 used for parameters. We don't currently do this
2029 if (Subtarget->isThumb1Only())
2032 // Externally-defined functions with weak linkage should not be
2033 // tail-called on ARM when the OS does not support dynamic
2034 // pre-emption of symbols, as the AAELF spec requires normal calls
2035 // to undefined weak functions to be replaced with a NOP or jump to the
2036 // next instruction. The behaviour of branch instructions in this
2037 // situation (as used for tail calls) is implementation-defined, so we
2038 // cannot rely on the linker replacing the tail call with a return.
2039 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2040 const GlobalValue *GV = G->getGlobal();
2041 const Triple &TT = getTargetMachine().getTargetTriple();
2042 if (GV->hasExternalWeakLinkage() &&
2043 (!TT.isOSWindows() || TT.isOSBinFormatELF() || TT.isOSBinFormatMachO()))
2047 // If the calling conventions do not match, then we'd better make sure the
2048 // results are returned in the same way as what the caller expects.
2050 SmallVector<CCValAssign, 16> RVLocs1;
2051 ARMCCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(), RVLocs1,
2052 *DAG.getContext(), Call);
2053 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
2055 SmallVector<CCValAssign, 16> RVLocs2;
2056 ARMCCState CCInfo2(CallerCC, false, DAG.getMachineFunction(), RVLocs2,
2057 *DAG.getContext(), Call);
2058 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
2060 if (RVLocs1.size() != RVLocs2.size())
2062 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
2063 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
2065 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
2067 if (RVLocs1[i].isRegLoc()) {
2068 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
2071 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
2077 // If Caller's vararg or byval argument has been split between registers and
2078 // stack, do not perform tail call, since part of the argument is in caller's
2080 const ARMFunctionInfo *AFI_Caller = DAG.getMachineFunction().
2081 getInfo<ARMFunctionInfo>();
2082 if (AFI_Caller->getArgRegsSaveSize())
2085 // If the callee takes no arguments then go on to check the results of the
2087 if (!Outs.empty()) {
2088 // Check if stack adjustment is needed. For now, do not do this if any
2089 // argument is passed on the stack.
2090 SmallVector<CCValAssign, 16> ArgLocs;
2091 ARMCCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
2092 *DAG.getContext(), Call);
2093 CCInfo.AnalyzeCallOperands(Outs,
2094 CCAssignFnForNode(CalleeCC, false, isVarArg));
2095 if (CCInfo.getNextStackOffset()) {
2096 MachineFunction &MF = DAG.getMachineFunction();
2098 // Check if the arguments are already laid out in the right way as
2099 // the caller's fixed stack objects.
2100 MachineFrameInfo *MFI = MF.getFrameInfo();
2101 const MachineRegisterInfo *MRI = &MF.getRegInfo();
2102 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
2103 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
2105 ++i, ++realArgIdx) {
2106 CCValAssign &VA = ArgLocs[i];
2107 EVT RegVT = VA.getLocVT();
2108 SDValue Arg = OutVals[realArgIdx];
2109 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
2110 if (VA.getLocInfo() == CCValAssign::Indirect)
2112 if (VA.needsCustom()) {
2113 // f64 and vector types are split into multiple registers or
2114 // register/stack-slot combinations. The types will not match
2115 // the registers; give up on memory f64 refs until we figure
2116 // out what to do about this.
2119 if (!ArgLocs[++i].isRegLoc())
2121 if (RegVT == MVT::v2f64) {
2122 if (!ArgLocs[++i].isRegLoc())
2124 if (!ArgLocs[++i].isRegLoc())
2127 } else if (!VA.isRegLoc()) {
2128 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
2140 ARMTargetLowering::CanLowerReturn(CallingConv::ID CallConv,
2141 MachineFunction &MF, bool isVarArg,
2142 const SmallVectorImpl<ISD::OutputArg> &Outs,
2143 LLVMContext &Context) const {
2144 SmallVector<CCValAssign, 16> RVLocs;
2145 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, Context);
2146 return CCInfo.CheckReturn(Outs, CCAssignFnForNode(CallConv, /*Return=*/true,
2150 static SDValue LowerInterruptReturn(SmallVectorImpl<SDValue> &RetOps,
2151 SDLoc DL, SelectionDAG &DAG) {
2152 const MachineFunction &MF = DAG.getMachineFunction();
2153 const Function *F = MF.getFunction();
2155 StringRef IntKind = F->getFnAttribute("interrupt").getValueAsString();
2157 // See ARM ARM v7 B1.8.3. On exception entry LR is set to a possibly offset
2158 // version of the "preferred return address". These offsets affect the return
2159 // instruction if this is a return from PL1 without hypervisor extensions.
2160 // IRQ/FIQ: +4 "subs pc, lr, #4"
2161 // SWI: 0 "subs pc, lr, #0"
2162 // ABORT: +4 "subs pc, lr, #4"
2163 // UNDEF: +4/+2 "subs pc, lr, #0"
2164 // UNDEF varies depending on where the exception came from ARM or Thumb
2165 // mode. Alongside GCC, we throw our hands up in disgust and pretend it's 0.
2168 if (IntKind == "" || IntKind == "IRQ" || IntKind == "FIQ" ||
2171 else if (IntKind == "SWI" || IntKind == "UNDEF")
2174 report_fatal_error("Unsupported interrupt attribute. If present, value "
2175 "must be one of: IRQ, FIQ, SWI, ABORT or UNDEF");
2177 RetOps.insert(RetOps.begin() + 1,
2178 DAG.getConstant(LROffset, DL, MVT::i32, false));
2180 return DAG.getNode(ARMISD::INTRET_FLAG, DL, MVT::Other, RetOps);
2184 ARMTargetLowering::LowerReturn(SDValue Chain,
2185 CallingConv::ID CallConv, bool isVarArg,
2186 const SmallVectorImpl<ISD::OutputArg> &Outs,
2187 const SmallVectorImpl<SDValue> &OutVals,
2188 SDLoc dl, SelectionDAG &DAG) const {
2190 // CCValAssign - represent the assignment of the return value to a location.
2191 SmallVector<CCValAssign, 16> RVLocs;
2193 // CCState - Info about the registers and stack slots.
2194 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
2195 *DAG.getContext(), Call);
2197 // Analyze outgoing return values.
2198 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
2202 SmallVector<SDValue, 4> RetOps;
2203 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
2204 bool isLittleEndian = Subtarget->isLittle();
2206 MachineFunction &MF = DAG.getMachineFunction();
2207 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2208 AFI->setReturnRegsCount(RVLocs.size());
2210 // Copy the result values into the output registers.
2211 for (unsigned i = 0, realRVLocIdx = 0;
2213 ++i, ++realRVLocIdx) {
2214 CCValAssign &VA = RVLocs[i];
2215 assert(VA.isRegLoc() && "Can only return in registers!");
2217 SDValue Arg = OutVals[realRVLocIdx];
2219 switch (VA.getLocInfo()) {
2220 default: llvm_unreachable("Unknown loc info!");
2221 case CCValAssign::Full: break;
2222 case CCValAssign::BCvt:
2223 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
2227 if (VA.needsCustom()) {
2228 if (VA.getLocVT() == MVT::v2f64) {
2229 // Extract the first half and return it in two registers.
2230 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
2231 DAG.getConstant(0, dl, MVT::i32));
2232 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
2233 DAG.getVTList(MVT::i32, MVT::i32), Half);
2235 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2236 HalfGPRs.getValue(isLittleEndian ? 0 : 1),
2238 Flag = Chain.getValue(1);
2239 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2240 VA = RVLocs[++i]; // skip ahead to next loc
2241 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2242 HalfGPRs.getValue(isLittleEndian ? 1 : 0),
2244 Flag = Chain.getValue(1);
2245 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2246 VA = RVLocs[++i]; // skip ahead to next loc
2248 // Extract the 2nd half and fall through to handle it as an f64 value.
2249 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
2250 DAG.getConstant(1, dl, MVT::i32));
2252 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
2254 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
2255 DAG.getVTList(MVT::i32, MVT::i32), Arg);
2256 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2257 fmrrd.getValue(isLittleEndian ? 0 : 1),
2259 Flag = Chain.getValue(1);
2260 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2261 VA = RVLocs[++i]; // skip ahead to next loc
2262 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2263 fmrrd.getValue(isLittleEndian ? 1 : 0),
2266 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
2268 // Guarantee that all emitted copies are
2269 // stuck together, avoiding something bad.
2270 Flag = Chain.getValue(1);
2271 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2274 // Update chain and glue.
2277 RetOps.push_back(Flag);
2279 // CPUs which aren't M-class use a special sequence to return from
2280 // exceptions (roughly, any instruction setting pc and cpsr simultaneously,
2281 // though we use "subs pc, lr, #N").
2283 // M-class CPUs actually use a normal return sequence with a special
2284 // (hardware-provided) value in LR, so the normal code path works.
2285 if (DAG.getMachineFunction().getFunction()->hasFnAttribute("interrupt") &&
2286 !Subtarget->isMClass()) {
2287 if (Subtarget->isThumb1Only())
2288 report_fatal_error("interrupt attribute is not supported in Thumb1");
2289 return LowerInterruptReturn(RetOps, dl, DAG);
2292 return DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, RetOps);
2295 bool ARMTargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
2296 if (N->getNumValues() != 1)
2298 if (!N->hasNUsesOfValue(1, 0))
2301 SDValue TCChain = Chain;
2302 SDNode *Copy = *N->use_begin();
2303 if (Copy->getOpcode() == ISD::CopyToReg) {
2304 // If the copy has a glue operand, we conservatively assume it isn't safe to
2305 // perform a tail call.
2306 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
2308 TCChain = Copy->getOperand(0);
2309 } else if (Copy->getOpcode() == ARMISD::VMOVRRD) {
2310 SDNode *VMov = Copy;
2311 // f64 returned in a pair of GPRs.
2312 SmallPtrSet<SDNode*, 2> Copies;
2313 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
2315 if (UI->getOpcode() != ISD::CopyToReg)
2319 if (Copies.size() > 2)
2322 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
2324 SDValue UseChain = UI->getOperand(0);
2325 if (Copies.count(UseChain.getNode()))
2329 // We are at the top of this chain.
2330 // If the copy has a glue operand, we conservatively assume it
2331 // isn't safe to perform a tail call.
2332 if (UI->getOperand(UI->getNumOperands()-1).getValueType() == MVT::Glue)
2338 } else if (Copy->getOpcode() == ISD::BITCAST) {
2339 // f32 returned in a single GPR.
2340 if (!Copy->hasOneUse())
2342 Copy = *Copy->use_begin();
2343 if (Copy->getOpcode() != ISD::CopyToReg || !Copy->hasNUsesOfValue(1, 0))
2345 // If the copy has a glue operand, we conservatively assume it isn't safe to
2346 // perform a tail call.
2347 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
2349 TCChain = Copy->getOperand(0);
2354 bool HasRet = false;
2355 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
2357 if (UI->getOpcode() != ARMISD::RET_FLAG &&
2358 UI->getOpcode() != ARMISD::INTRET_FLAG)
2370 bool ARMTargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
2371 if (!Subtarget->supportsTailCall())
2375 CI->getParent()->getParent()->getFnAttribute("disable-tail-calls");
2376 if (!CI->isTailCall() || Attr.getValueAsString() == "true")
2379 return !Subtarget->isThumb1Only();
2382 // Trying to write a 64 bit value so need to split into two 32 bit values first,
2383 // and pass the lower and high parts through.
2384 static SDValue LowerWRITE_REGISTER(SDValue Op, SelectionDAG &DAG) {
2386 SDValue WriteValue = Op->getOperand(2);
2388 // This function is only supposed to be called for i64 type argument.
2389 assert(WriteValue.getValueType() == MVT::i64
2390 && "LowerWRITE_REGISTER called for non-i64 type argument.");
2392 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, WriteValue,
2393 DAG.getConstant(0, DL, MVT::i32));
2394 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, WriteValue,
2395 DAG.getConstant(1, DL, MVT::i32));
2396 SDValue Ops[] = { Op->getOperand(0), Op->getOperand(1), Lo, Hi };
2397 return DAG.getNode(ISD::WRITE_REGISTER, DL, MVT::Other, Ops);
2400 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
2401 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
2402 // one of the above mentioned nodes. It has to be wrapped because otherwise
2403 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
2404 // be used to form addressing mode. These wrapped nodes will be selected
2406 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
2407 EVT PtrVT = Op.getValueType();
2408 // FIXME there is no actual debug info here
2410 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
2412 if (CP->isMachineConstantPoolEntry())
2413 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
2414 CP->getAlignment());
2416 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
2417 CP->getAlignment());
2418 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
2421 unsigned ARMTargetLowering::getJumpTableEncoding() const {
2422 return MachineJumpTableInfo::EK_Inline;
2425 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
2426 SelectionDAG &DAG) const {
2427 MachineFunction &MF = DAG.getMachineFunction();
2428 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2429 unsigned ARMPCLabelIndex = 0;
2431 EVT PtrVT = getPointerTy();
2432 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
2433 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2435 if (RelocM == Reloc::Static) {
2436 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
2438 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2439 ARMPCLabelIndex = AFI->createPICLabelUId();
2440 ARMConstantPoolValue *CPV =
2441 ARMConstantPoolConstant::Create(BA, ARMPCLabelIndex,
2442 ARMCP::CPBlockAddress, PCAdj);
2443 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2445 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
2446 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
2447 MachinePointerInfo::getConstantPool(),
2448 false, false, false, 0);
2449 if (RelocM == Reloc::Static)
2451 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, DL, MVT::i32);
2452 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
2455 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
2457 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
2458 SelectionDAG &DAG) const {
2460 EVT PtrVT = getPointerTy();
2461 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2462 MachineFunction &MF = DAG.getMachineFunction();
2463 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2464 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2465 ARMConstantPoolValue *CPV =
2466 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2467 ARMCP::CPValue, PCAdj, ARMCP::TLSGD, true);
2468 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2469 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
2470 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
2471 MachinePointerInfo::getConstantPool(),
2472 false, false, false, 0);
2473 SDValue Chain = Argument.getValue(1);
2475 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, dl, MVT::i32);
2476 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
2478 // call __tls_get_addr.
2481 Entry.Node = Argument;
2482 Entry.Ty = (Type *) Type::getInt32Ty(*DAG.getContext());
2483 Args.push_back(Entry);
2485 // FIXME: is there useful debug info available here?
2486 TargetLowering::CallLoweringInfo CLI(DAG);
2487 CLI.setDebugLoc(dl).setChain(Chain)
2488 .setCallee(CallingConv::C, Type::getInt32Ty(*DAG.getContext()),
2489 DAG.getExternalSymbol("__tls_get_addr", PtrVT), std::move(Args),
2492 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
2493 return CallResult.first;
2496 // Lower ISD::GlobalTLSAddress using the "initial exec" or
2497 // "local exec" model.
2499 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
2501 TLSModel::Model model) const {
2502 const GlobalValue *GV = GA->getGlobal();
2505 SDValue Chain = DAG.getEntryNode();
2506 EVT PtrVT = getPointerTy();
2507 // Get the Thread Pointer
2508 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2510 if (model == TLSModel::InitialExec) {
2511 MachineFunction &MF = DAG.getMachineFunction();
2512 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2513 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2514 // Initial exec model.
2515 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2516 ARMConstantPoolValue *CPV =
2517 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2518 ARMCP::CPValue, PCAdj, ARMCP::GOTTPOFF,
2520 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2521 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2522 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2523 MachinePointerInfo::getConstantPool(),
2524 false, false, false, 0);
2525 Chain = Offset.getValue(1);
2527 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, dl, MVT::i32);
2528 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
2530 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2531 MachinePointerInfo::getConstantPool(),
2532 false, false, false, 0);
2535 assert(model == TLSModel::LocalExec);
2536 ARMConstantPoolValue *CPV =
2537 ARMConstantPoolConstant::Create(GV, ARMCP::TPOFF);
2538 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2539 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2540 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2541 MachinePointerInfo::getConstantPool(),
2542 false, false, false, 0);
2545 // The address of the thread local variable is the add of the thread
2546 // pointer with the offset of the variable.
2547 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
2551 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
2552 // TODO: implement the "local dynamic" model
2553 assert(Subtarget->isTargetELF() &&
2554 "TLS not implemented for non-ELF targets");
2555 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
2557 TLSModel::Model model = getTargetMachine().getTLSModel(GA->getGlobal());
2560 case TLSModel::GeneralDynamic:
2561 case TLSModel::LocalDynamic:
2562 return LowerToTLSGeneralDynamicModel(GA, DAG);
2563 case TLSModel::InitialExec:
2564 case TLSModel::LocalExec:
2565 return LowerToTLSExecModels(GA, DAG, model);
2567 llvm_unreachable("bogus TLS model");
2570 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
2571 SelectionDAG &DAG) const {
2572 EVT PtrVT = getPointerTy();
2574 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2575 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2576 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
2577 ARMConstantPoolValue *CPV =
2578 ARMConstantPoolConstant::Create(GV,
2579 UseGOTOFF ? ARMCP::GOTOFF : ARMCP::GOT);
2580 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2581 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2582 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
2584 MachinePointerInfo::getConstantPool(),
2585 false, false, false, 0);
2586 SDValue Chain = Result.getValue(1);
2587 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
2588 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
2590 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
2591 MachinePointerInfo::getGOT(),
2592 false, false, false, 0);
2596 // If we have T2 ops, we can materialize the address directly via movt/movw
2597 // pair. This is always cheaper.
2598 if (Subtarget->useMovt(DAG.getMachineFunction())) {
2600 // FIXME: Once remat is capable of dealing with instructions with register
2601 // operands, expand this into two nodes.
2602 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2603 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2605 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2606 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2607 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2608 MachinePointerInfo::getConstantPool(),
2609 false, false, false, 0);
2613 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
2614 SelectionDAG &DAG) const {
2615 EVT PtrVT = getPointerTy();
2617 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2618 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2620 if (Subtarget->useMovt(DAG.getMachineFunction()))
2623 // FIXME: Once remat is capable of dealing with instructions with register
2624 // operands, expand this into multiple nodes
2626 RelocM == Reloc::PIC_ ? ARMISD::WrapperPIC : ARMISD::Wrapper;
2628 SDValue G = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, ARMII::MO_NONLAZY);
2629 SDValue Result = DAG.getNode(Wrapper, dl, PtrVT, G);
2631 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2632 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
2633 MachinePointerInfo::getGOT(), false, false, false, 0);
2637 SDValue ARMTargetLowering::LowerGlobalAddressWindows(SDValue Op,
2638 SelectionDAG &DAG) const {
2639 assert(Subtarget->isTargetWindows() && "non-Windows COFF is not supported");
2640 assert(Subtarget->useMovt(DAG.getMachineFunction()) &&
2641 "Windows on ARM expects to use movw/movt");
2643 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2644 const ARMII::TOF TargetFlags =
2645 (GV->hasDLLImportStorageClass() ? ARMII::MO_DLLIMPORT : ARMII::MO_NO_FLAG);
2646 EVT PtrVT = getPointerTy();
2652 // FIXME: Once remat is capable of dealing with instructions with register
2653 // operands, expand this into two nodes.
2654 Result = DAG.getNode(ARMISD::Wrapper, DL, PtrVT,
2655 DAG.getTargetGlobalAddress(GV, DL, PtrVT, /*Offset=*/0,
2657 if (GV->hasDLLImportStorageClass())
2658 Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), Result,
2659 MachinePointerInfo::getGOT(), false, false, false, 0);
2663 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
2664 SelectionDAG &DAG) const {
2665 assert(Subtarget->isTargetELF() &&
2666 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
2667 MachineFunction &MF = DAG.getMachineFunction();
2668 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2669 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2670 EVT PtrVT = getPointerTy();
2672 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2673 ARMConstantPoolValue *CPV =
2674 ARMConstantPoolSymbol::Create(*DAG.getContext(), "_GLOBAL_OFFSET_TABLE_",
2675 ARMPCLabelIndex, PCAdj);
2676 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2677 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2678 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2679 MachinePointerInfo::getConstantPool(),
2680 false, false, false, 0);
2681 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, dl, MVT::i32);
2682 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2686 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
2688 SDValue Val = DAG.getConstant(0, dl, MVT::i32);
2689 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl,
2690 DAG.getVTList(MVT::i32, MVT::Other), Op.getOperand(0),
2691 Op.getOperand(1), Val);
2695 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
2697 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
2698 Op.getOperand(1), DAG.getConstant(0, dl, MVT::i32));
2702 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
2703 const ARMSubtarget *Subtarget) const {
2704 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2707 default: return SDValue(); // Don't custom lower most intrinsics.
2708 case Intrinsic::arm_rbit: {
2709 assert(Op.getOperand(1).getValueType() == MVT::i32 &&
2710 "RBIT intrinsic must have i32 type!");
2711 return DAG.getNode(ARMISD::RBIT, dl, MVT::i32, Op.getOperand(1));
2713 case Intrinsic::arm_thread_pointer: {
2714 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2715 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2717 case Intrinsic::eh_sjlj_lsda: {
2718 MachineFunction &MF = DAG.getMachineFunction();
2719 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2720 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2721 EVT PtrVT = getPointerTy();
2722 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2724 unsigned PCAdj = (RelocM != Reloc::PIC_)
2725 ? 0 : (Subtarget->isThumb() ? 4 : 8);
2726 ARMConstantPoolValue *CPV =
2727 ARMConstantPoolConstant::Create(MF.getFunction(), ARMPCLabelIndex,
2728 ARMCP::CPLSDA, PCAdj);
2729 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2730 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2732 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2733 MachinePointerInfo::getConstantPool(),
2734 false, false, false, 0);
2736 if (RelocM == Reloc::PIC_) {
2737 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, dl, MVT::i32);
2738 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2742 case Intrinsic::arm_neon_vmulls:
2743 case Intrinsic::arm_neon_vmullu: {
2744 unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vmulls)
2745 ? ARMISD::VMULLs : ARMISD::VMULLu;
2746 return DAG.getNode(NewOpc, SDLoc(Op), Op.getValueType(),
2747 Op.getOperand(1), Op.getOperand(2));
2752 static SDValue LowerATOMIC_FENCE(SDValue Op, SelectionDAG &DAG,
2753 const ARMSubtarget *Subtarget) {
2754 // FIXME: handle "fence singlethread" more efficiently.
2756 if (!Subtarget->hasDataBarrier()) {
2757 // Some ARMv6 cpus can support data barriers with an mcr instruction.
2758 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2760 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2761 "Unexpected ISD::ATOMIC_FENCE encountered. Should be libcall!");
2762 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2763 DAG.getConstant(0, dl, MVT::i32));
2766 ConstantSDNode *OrdN = cast<ConstantSDNode>(Op.getOperand(1));
2767 AtomicOrdering Ord = static_cast<AtomicOrdering>(OrdN->getZExtValue());
2768 ARM_MB::MemBOpt Domain = ARM_MB::ISH;
2769 if (Subtarget->isMClass()) {
2770 // Only a full system barrier exists in the M-class architectures.
2771 Domain = ARM_MB::SY;
2772 } else if (Subtarget->isSwift() && Ord == Release) {
2773 // Swift happens to implement ISHST barriers in a way that's compatible with
2774 // Release semantics but weaker than ISH so we'd be fools not to use
2775 // it. Beware: other processors probably don't!
2776 Domain = ARM_MB::ISHST;
2779 return DAG.getNode(ISD::INTRINSIC_VOID, dl, MVT::Other, Op.getOperand(0),
2780 DAG.getConstant(Intrinsic::arm_dmb, dl, MVT::i32),
2781 DAG.getConstant(Domain, dl, MVT::i32));
2784 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG,
2785 const ARMSubtarget *Subtarget) {
2786 // ARM pre v5TE and Thumb1 does not have preload instructions.
2787 if (!(Subtarget->isThumb2() ||
2788 (!Subtarget->isThumb1Only() && Subtarget->hasV5TEOps())))
2789 // Just preserve the chain.
2790 return Op.getOperand(0);
2793 unsigned isRead = ~cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() & 1;
2795 (!Subtarget->hasV7Ops() || !Subtarget->hasMPExtension()))
2796 // ARMv7 with MP extension has PLDW.
2797 return Op.getOperand(0);
2799 unsigned isData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
2800 if (Subtarget->isThumb()) {
2802 isRead = ~isRead & 1;
2803 isData = ~isData & 1;
2806 return DAG.getNode(ARMISD::PRELOAD, dl, MVT::Other, Op.getOperand(0),
2807 Op.getOperand(1), DAG.getConstant(isRead, dl, MVT::i32),
2808 DAG.getConstant(isData, dl, MVT::i32));
2811 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
2812 MachineFunction &MF = DAG.getMachineFunction();
2813 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
2815 // vastart just stores the address of the VarArgsFrameIndex slot into the
2816 // memory location argument.
2818 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2819 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2820 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
2821 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
2822 MachinePointerInfo(SV), false, false, 0);
2826 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
2827 SDValue &Root, SelectionDAG &DAG,
2829 MachineFunction &MF = DAG.getMachineFunction();
2830 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2832 const TargetRegisterClass *RC;
2833 if (AFI->isThumb1OnlyFunction())
2834 RC = &ARM::tGPRRegClass;
2836 RC = &ARM::GPRRegClass;
2838 // Transform the arguments stored in physical registers into virtual ones.
2839 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2840 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2843 if (NextVA.isMemLoc()) {
2844 MachineFrameInfo *MFI = MF.getFrameInfo();
2845 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true);
2847 // Create load node to retrieve arguments from the stack.
2848 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2849 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
2850 MachinePointerInfo::getFixedStack(FI),
2851 false, false, false, 0);
2853 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
2854 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2856 if (!Subtarget->isLittle())
2857 std::swap (ArgValue, ArgValue2);
2858 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
2861 // The remaining GPRs hold either the beginning of variable-argument
2862 // data, or the beginning of an aggregate passed by value (usually
2863 // byval). Either way, we allocate stack slots adjacent to the data
2864 // provided by our caller, and store the unallocated registers there.
2865 // If this is a variadic function, the va_list pointer will begin with
2866 // these values; otherwise, this reassembles a (byval) structure that
2867 // was split between registers and memory.
2868 // Return: The frame index registers were stored into.
2870 ARMTargetLowering::StoreByValRegs(CCState &CCInfo, SelectionDAG &DAG,
2871 SDLoc dl, SDValue &Chain,
2872 const Value *OrigArg,
2873 unsigned InRegsParamRecordIdx,
2875 unsigned ArgSize) const {
2876 // Currently, two use-cases possible:
2877 // Case #1. Non-var-args function, and we meet first byval parameter.
2878 // Setup first unallocated register as first byval register;
2879 // eat all remained registers
2880 // (these two actions are performed by HandleByVal method).
2881 // Then, here, we initialize stack frame with
2882 // "store-reg" instructions.
2883 // Case #2. Var-args function, that doesn't contain byval parameters.
2884 // The same: eat all remained unallocated registers,
2885 // initialize stack frame.
2887 MachineFunction &MF = DAG.getMachineFunction();
2888 MachineFrameInfo *MFI = MF.getFrameInfo();
2889 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2890 unsigned RBegin, REnd;
2891 if (InRegsParamRecordIdx < CCInfo.getInRegsParamsCount()) {
2892 CCInfo.getInRegsParamInfo(InRegsParamRecordIdx, RBegin, REnd);
2894 unsigned RBeginIdx = CCInfo.getFirstUnallocated(GPRArgRegs);
2895 RBegin = RBeginIdx == 4 ? (unsigned)ARM::R4 : GPRArgRegs[RBeginIdx];
2900 ArgOffset = -4 * (ARM::R4 - RBegin);
2902 int FrameIndex = MFI->CreateFixedObject(ArgSize, ArgOffset, false);
2903 SDValue FIN = DAG.getFrameIndex(FrameIndex, getPointerTy());
2905 SmallVector<SDValue, 4> MemOps;
2906 const TargetRegisterClass *RC =
2907 AFI->isThumb1OnlyFunction() ? &ARM::tGPRRegClass : &ARM::GPRRegClass;
2909 for (unsigned Reg = RBegin, i = 0; Reg < REnd; ++Reg, ++i) {
2910 unsigned VReg = MF.addLiveIn(Reg, RC);
2911 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2913 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2914 MachinePointerInfo(OrigArg, 4 * i), false, false, 0);
2915 MemOps.push_back(Store);
2916 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
2917 DAG.getConstant(4, dl, getPointerTy()));
2920 if (!MemOps.empty())
2921 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOps);
2925 // Setup stack frame, the va_list pointer will start from.
2927 ARMTargetLowering::VarArgStyleRegisters(CCState &CCInfo, SelectionDAG &DAG,
2928 SDLoc dl, SDValue &Chain,
2930 unsigned TotalArgRegsSaveSize,
2931 bool ForceMutable) const {
2932 MachineFunction &MF = DAG.getMachineFunction();
2933 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2935 // Try to store any remaining integer argument regs
2936 // to their spots on the stack so that they may be loaded by deferencing
2937 // the result of va_next.
2938 // If there is no regs to be stored, just point address after last
2939 // argument passed via stack.
2940 int FrameIndex = StoreByValRegs(CCInfo, DAG, dl, Chain, nullptr,
2941 CCInfo.getInRegsParamsCount(),
2942 CCInfo.getNextStackOffset(), 4);
2943 AFI->setVarArgsFrameIndex(FrameIndex);
2947 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
2948 CallingConv::ID CallConv, bool isVarArg,
2949 const SmallVectorImpl<ISD::InputArg>
2951 SDLoc dl, SelectionDAG &DAG,
2952 SmallVectorImpl<SDValue> &InVals)
2954 MachineFunction &MF = DAG.getMachineFunction();
2955 MachineFrameInfo *MFI = MF.getFrameInfo();
2957 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2959 // Assign locations to all of the incoming arguments.
2960 SmallVector<CCValAssign, 16> ArgLocs;
2961 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs,
2962 *DAG.getContext(), Prologue);
2963 CCInfo.AnalyzeFormalArguments(Ins,
2964 CCAssignFnForNode(CallConv, /* Return*/ false,
2967 SmallVector<SDValue, 16> ArgValues;
2969 Function::const_arg_iterator CurOrigArg = MF.getFunction()->arg_begin();
2970 unsigned CurArgIdx = 0;
2972 // Initially ArgRegsSaveSize is zero.
2973 // Then we increase this value each time we meet byval parameter.
2974 // We also increase this value in case of varargs function.
2975 AFI->setArgRegsSaveSize(0);
2977 // Calculate the amount of stack space that we need to allocate to store
2978 // byval and variadic arguments that are passed in registers.
2979 // We need to know this before we allocate the first byval or variadic
2980 // argument, as they will be allocated a stack slot below the CFA (Canonical
2981 // Frame Address, the stack pointer at entry to the function).
2982 unsigned ArgRegBegin = ARM::R4;
2983 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2984 if (CCInfo.getInRegsParamsProcessed() >= CCInfo.getInRegsParamsCount())
2987 CCValAssign &VA = ArgLocs[i];
2988 unsigned Index = VA.getValNo();
2989 ISD::ArgFlagsTy Flags = Ins[Index].Flags;
2990 if (!Flags.isByVal())
2993 assert(VA.isMemLoc() && "unexpected byval pointer in reg");
2994 unsigned RBegin, REnd;
2995 CCInfo.getInRegsParamInfo(CCInfo.getInRegsParamsProcessed(), RBegin, REnd);
2996 ArgRegBegin = std::min(ArgRegBegin, RBegin);
2998 CCInfo.nextInRegsParam();
3000 CCInfo.rewindByValRegsInfo();
3002 int lastInsIndex = -1;
3003 if (isVarArg && MFI->hasVAStart()) {
3004 unsigned RegIdx = CCInfo.getFirstUnallocated(GPRArgRegs);
3005 if (RegIdx != array_lengthof(GPRArgRegs))
3006 ArgRegBegin = std::min(ArgRegBegin, (unsigned)GPRArgRegs[RegIdx]);
3009 unsigned TotalArgRegsSaveSize = 4 * (ARM::R4 - ArgRegBegin);
3010 AFI->setArgRegsSaveSize(TotalArgRegsSaveSize);
3012 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3013 CCValAssign &VA = ArgLocs[i];
3014 if (Ins[VA.getValNo()].isOrigArg()) {
3015 std::advance(CurOrigArg,
3016 Ins[VA.getValNo()].getOrigArgIndex() - CurArgIdx);
3017 CurArgIdx = Ins[VA.getValNo()].getOrigArgIndex();
3019 // Arguments stored in registers.
3020 if (VA.isRegLoc()) {
3021 EVT RegVT = VA.getLocVT();
3023 if (VA.needsCustom()) {
3024 // f64 and vector types are split up into multiple registers or
3025 // combinations of registers and stack slots.
3026 if (VA.getLocVT() == MVT::v2f64) {
3027 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
3029 VA = ArgLocs[++i]; // skip ahead to next loc
3031 if (VA.isMemLoc()) {
3032 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true);
3033 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
3034 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
3035 MachinePointerInfo::getFixedStack(FI),
3036 false, false, false, 0);
3038 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
3041 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
3042 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
3043 ArgValue, ArgValue1,
3044 DAG.getIntPtrConstant(0, dl));
3045 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
3046 ArgValue, ArgValue2,
3047 DAG.getIntPtrConstant(1, dl));
3049 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
3052 const TargetRegisterClass *RC;
3054 if (RegVT == MVT::f32)
3055 RC = &ARM::SPRRegClass;
3056 else if (RegVT == MVT::f64)
3057 RC = &ARM::DPRRegClass;
3058 else if (RegVT == MVT::v2f64)
3059 RC = &ARM::QPRRegClass;
3060 else if (RegVT == MVT::i32)
3061 RC = AFI->isThumb1OnlyFunction() ? &ARM::tGPRRegClass
3062 : &ARM::GPRRegClass;
3064 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
3066 // Transform the arguments in physical registers into virtual ones.
3067 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
3068 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
3071 // If this is an 8 or 16-bit value, it is really passed promoted
3072 // to 32 bits. Insert an assert[sz]ext to capture this, then
3073 // truncate to the right size.
3074 switch (VA.getLocInfo()) {
3075 default: llvm_unreachable("Unknown loc info!");
3076 case CCValAssign::Full: break;
3077 case CCValAssign::BCvt:
3078 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
3080 case CCValAssign::SExt:
3081 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
3082 DAG.getValueType(VA.getValVT()));
3083 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
3085 case CCValAssign::ZExt:
3086 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
3087 DAG.getValueType(VA.getValVT()));
3088 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
3092 InVals.push_back(ArgValue);
3094 } else { // VA.isRegLoc()
3097 assert(VA.isMemLoc());
3098 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
3100 int index = VA.getValNo();
3102 // Some Ins[] entries become multiple ArgLoc[] entries.
3103 // Process them only once.
3104 if (index != lastInsIndex)
3106 ISD::ArgFlagsTy Flags = Ins[index].Flags;
3107 // FIXME: For now, all byval parameter objects are marked mutable.
3108 // This can be changed with more analysis.
3109 // In case of tail call optimization mark all arguments mutable.
3110 // Since they could be overwritten by lowering of arguments in case of
3112 if (Flags.isByVal()) {
3113 assert(Ins[index].isOrigArg() &&
3114 "Byval arguments cannot be implicit");
3115 unsigned CurByValIndex = CCInfo.getInRegsParamsProcessed();
3117 int FrameIndex = StoreByValRegs(CCInfo, DAG, dl, Chain, CurOrigArg,
3118 CurByValIndex, VA.getLocMemOffset(),
3119 Flags.getByValSize());
3120 InVals.push_back(DAG.getFrameIndex(FrameIndex, getPointerTy()));
3121 CCInfo.nextInRegsParam();
3123 unsigned FIOffset = VA.getLocMemOffset();
3124 int FI = MFI->CreateFixedObject(VA.getLocVT().getSizeInBits()/8,
3127 // Create load nodes to retrieve arguments from the stack.
3128 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
3129 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
3130 MachinePointerInfo::getFixedStack(FI),
3131 false, false, false, 0));
3133 lastInsIndex = index;
3139 if (isVarArg && MFI->hasVAStart())
3140 VarArgStyleRegisters(CCInfo, DAG, dl, Chain,
3141 CCInfo.getNextStackOffset(),
3142 TotalArgRegsSaveSize);
3144 AFI->setArgumentStackSize(CCInfo.getNextStackOffset());
3149 /// isFloatingPointZero - Return true if this is +0.0.
3150 static bool isFloatingPointZero(SDValue Op) {
3151 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
3152 return CFP->getValueAPF().isPosZero();
3153 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
3154 // Maybe this has already been legalized into the constant pool?
3155 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
3156 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
3157 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
3158 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
3159 return CFP->getValueAPF().isPosZero();
3161 } else if (Op->getOpcode() == ISD::BITCAST &&
3162 Op->getValueType(0) == MVT::f64) {
3163 // Handle (ISD::BITCAST (ARMISD::VMOVIMM (ISD::TargetConstant 0)) MVT::f64)
3164 // created by LowerConstantFP().
3165 SDValue BitcastOp = Op->getOperand(0);
3166 if (BitcastOp->getOpcode() == ARMISD::VMOVIMM) {
3167 SDValue MoveOp = BitcastOp->getOperand(0);
3168 if (MoveOp->getOpcode() == ISD::TargetConstant &&
3169 cast<ConstantSDNode>(MoveOp)->getZExtValue() == 0) {
3177 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
3178 /// the given operands.
3180 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
3181 SDValue &ARMcc, SelectionDAG &DAG,
3183 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
3184 unsigned C = RHSC->getZExtValue();
3185 if (!isLegalICmpImmediate(C)) {
3186 // Constant does not fit, try adjusting it by one?
3191 if (C != 0x80000000 && isLegalICmpImmediate(C-1)) {
3192 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
3193 RHS = DAG.getConstant(C - 1, dl, MVT::i32);
3198 if (C != 0 && isLegalICmpImmediate(C-1)) {
3199 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
3200 RHS = DAG.getConstant(C - 1, dl, MVT::i32);
3205 if (C != 0x7fffffff && isLegalICmpImmediate(C+1)) {
3206 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
3207 RHS = DAG.getConstant(C + 1, dl, MVT::i32);
3212 if (C != 0xffffffff && isLegalICmpImmediate(C+1)) {
3213 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
3214 RHS = DAG.getConstant(C + 1, dl, MVT::i32);
3221 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3222 ARMISD::NodeType CompareType;
3225 CompareType = ARMISD::CMP;
3230 CompareType = ARMISD::CMPZ;
3233 ARMcc = DAG.getConstant(CondCode, dl, MVT::i32);
3234 return DAG.getNode(CompareType, dl, MVT::Glue, LHS, RHS);
3237 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
3239 ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
3241 assert(!Subtarget->isFPOnlySP() || RHS.getValueType() != MVT::f64);
3243 if (!isFloatingPointZero(RHS))
3244 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Glue, LHS, RHS);
3246 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Glue, LHS);
3247 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Glue, Cmp);
3250 /// duplicateCmp - Glue values can have only one use, so this function
3251 /// duplicates a comparison node.
3253 ARMTargetLowering::duplicateCmp(SDValue Cmp, SelectionDAG &DAG) const {
3254 unsigned Opc = Cmp.getOpcode();
3256 if (Opc == ARMISD::CMP || Opc == ARMISD::CMPZ)
3257 return DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
3259 assert(Opc == ARMISD::FMSTAT && "unexpected comparison operation");
3260 Cmp = Cmp.getOperand(0);
3261 Opc = Cmp.getOpcode();
3262 if (Opc == ARMISD::CMPFP)
3263 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
3265 assert(Opc == ARMISD::CMPFPw0 && "unexpected operand of FMSTAT");
3266 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0));
3268 return DAG.getNode(ARMISD::FMSTAT, DL, MVT::Glue, Cmp);
3271 std::pair<SDValue, SDValue>
3272 ARMTargetLowering::getARMXALUOOp(SDValue Op, SelectionDAG &DAG,
3273 SDValue &ARMcc) const {
3274 assert(Op.getValueType() == MVT::i32 && "Unsupported value type");
3276 SDValue Value, OverflowCmp;
3277 SDValue LHS = Op.getOperand(0);
3278 SDValue RHS = Op.getOperand(1);
3281 // FIXME: We are currently always generating CMPs because we don't support
3282 // generating CMN through the backend. This is not as good as the natural
3283 // CMP case because it causes a register dependency and cannot be folded
3286 switch (Op.getOpcode()) {
3288 llvm_unreachable("Unknown overflow instruction!");
3290 ARMcc = DAG.getConstant(ARMCC::VC, dl, MVT::i32);
3291 Value = DAG.getNode(ISD::ADD, dl, Op.getValueType(), LHS, RHS);
3292 OverflowCmp = DAG.getNode(ARMISD::CMP, dl, MVT::Glue, Value, LHS);
3295 ARMcc = DAG.getConstant(ARMCC::HS, dl, MVT::i32);
3296 Value = DAG.getNode(ISD::ADD, dl, Op.getValueType(), LHS, RHS);
3297 OverflowCmp = DAG.getNode(ARMISD::CMP, dl, MVT::Glue, Value, LHS);
3300 ARMcc = DAG.getConstant(ARMCC::VC, dl, MVT::i32);
3301 Value = DAG.getNode(ISD::SUB, dl, Op.getValueType(), LHS, RHS);
3302 OverflowCmp = DAG.getNode(ARMISD::CMP, dl, MVT::Glue, LHS, RHS);
3305 ARMcc = DAG.getConstant(ARMCC::HS, dl, MVT::i32);
3306 Value = DAG.getNode(ISD::SUB, dl, Op.getValueType(), LHS, RHS);
3307 OverflowCmp = DAG.getNode(ARMISD::CMP, dl, MVT::Glue, LHS, RHS);
3311 return std::make_pair(Value, OverflowCmp);
3316 ARMTargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) const {
3317 // Let legalize expand this if it isn't a legal type yet.
3318 if (!DAG.getTargetLoweringInfo().isTypeLegal(Op.getValueType()))
3321 SDValue Value, OverflowCmp;
3323 std::tie(Value, OverflowCmp) = getARMXALUOOp(Op, DAG, ARMcc);
3324 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3326 // We use 0 and 1 as false and true values.
3327 SDValue TVal = DAG.getConstant(1, dl, MVT::i32);
3328 SDValue FVal = DAG.getConstant(0, dl, MVT::i32);
3329 EVT VT = Op.getValueType();
3331 SDValue Overflow = DAG.getNode(ARMISD::CMOV, dl, VT, TVal, FVal,
3332 ARMcc, CCR, OverflowCmp);
3334 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
3335 return DAG.getNode(ISD::MERGE_VALUES, dl, VTs, Value, Overflow);
3339 SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
3340 SDValue Cond = Op.getOperand(0);
3341 SDValue SelectTrue = Op.getOperand(1);
3342 SDValue SelectFalse = Op.getOperand(2);
3344 unsigned Opc = Cond.getOpcode();
3346 if (Cond.getResNo() == 1 &&
3347 (Opc == ISD::SADDO || Opc == ISD::UADDO || Opc == ISD::SSUBO ||
3348 Opc == ISD::USUBO)) {
3349 if (!DAG.getTargetLoweringInfo().isTypeLegal(Cond->getValueType(0)))
3352 SDValue Value, OverflowCmp;
3354 std::tie(Value, OverflowCmp) = getARMXALUOOp(Cond, DAG, ARMcc);
3355 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3356 EVT VT = Op.getValueType();
3358 return getCMOV(dl, VT, SelectTrue, SelectFalse, ARMcc, CCR,
3364 // (select (cmov 1, 0, cond), t, f) -> (cmov t, f, cond)
3365 // (select (cmov 0, 1, cond), t, f) -> (cmov f, t, cond)
3367 if (Cond.getOpcode() == ARMISD::CMOV && Cond.hasOneUse()) {
3368 const ConstantSDNode *CMOVTrue =
3369 dyn_cast<ConstantSDNode>(Cond.getOperand(0));
3370 const ConstantSDNode *CMOVFalse =
3371 dyn_cast<ConstantSDNode>(Cond.getOperand(1));
3373 if (CMOVTrue && CMOVFalse) {
3374 unsigned CMOVTrueVal = CMOVTrue->getZExtValue();
3375 unsigned CMOVFalseVal = CMOVFalse->getZExtValue();
3379 if (CMOVTrueVal == 1 && CMOVFalseVal == 0) {
3381 False = SelectFalse;
3382 } else if (CMOVTrueVal == 0 && CMOVFalseVal == 1) {
3387 if (True.getNode() && False.getNode()) {
3388 EVT VT = Op.getValueType();
3389 SDValue ARMcc = Cond.getOperand(2);
3390 SDValue CCR = Cond.getOperand(3);
3391 SDValue Cmp = duplicateCmp(Cond.getOperand(4), DAG);
3392 assert(True.getValueType() == VT);
3393 return getCMOV(dl, VT, True, False, ARMcc, CCR, Cmp, DAG);
3398 // ARM's BooleanContents value is UndefinedBooleanContent. Mask out the
3399 // undefined bits before doing a full-word comparison with zero.
3400 Cond = DAG.getNode(ISD::AND, dl, Cond.getValueType(), Cond,
3401 DAG.getConstant(1, dl, Cond.getValueType()));
3403 return DAG.getSelectCC(dl, Cond,
3404 DAG.getConstant(0, dl, Cond.getValueType()),
3405 SelectTrue, SelectFalse, ISD::SETNE);
3408 static void checkVSELConstraints(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
3409 bool &swpCmpOps, bool &swpVselOps) {
3410 // Start by selecting the GE condition code for opcodes that return true for
3412 if (CC == ISD::SETUGE || CC == ISD::SETOGE || CC == ISD::SETOLE ||
3414 CondCode = ARMCC::GE;
3416 // and GT for opcodes that return false for 'equality'.
3417 else if (CC == ISD::SETUGT || CC == ISD::SETOGT || CC == ISD::SETOLT ||
3419 CondCode = ARMCC::GT;
3421 // Since we are constrained to GE/GT, if the opcode contains 'less', we need
3422 // to swap the compare operands.
3423 if (CC == ISD::SETOLE || CC == ISD::SETULE || CC == ISD::SETOLT ||
3427 // Both GT and GE are ordered comparisons, and return false for 'unordered'.
3428 // If we have an unordered opcode, we need to swap the operands to the VSEL
3429 // instruction (effectively negating the condition).
3431 // This also has the effect of swapping which one of 'less' or 'greater'
3432 // returns true, so we also swap the compare operands. It also switches
3433 // whether we return true for 'equality', so we compensate by picking the
3434 // opposite condition code to our original choice.
3435 if (CC == ISD::SETULE || CC == ISD::SETULT || CC == ISD::SETUGE ||
3436 CC == ISD::SETUGT) {
3437 swpCmpOps = !swpCmpOps;
3438 swpVselOps = !swpVselOps;
3439 CondCode = CondCode == ARMCC::GT ? ARMCC::GE : ARMCC::GT;
3442 // 'ordered' is 'anything but unordered', so use the VS condition code and
3443 // swap the VSEL operands.
3444 if (CC == ISD::SETO) {
3445 CondCode = ARMCC::VS;
3449 // 'unordered or not equal' is 'anything but equal', so use the EQ condition
3450 // code and swap the VSEL operands.
3451 if (CC == ISD::SETUNE) {
3452 CondCode = ARMCC::EQ;
3457 SDValue ARMTargetLowering::getCMOV(SDLoc dl, EVT VT, SDValue FalseVal,
3458 SDValue TrueVal, SDValue ARMcc, SDValue CCR,
3459 SDValue Cmp, SelectionDAG &DAG) const {
3460 if (Subtarget->isFPOnlySP() && VT == MVT::f64) {
3461 FalseVal = DAG.getNode(ARMISD::VMOVRRD, dl,
3462 DAG.getVTList(MVT::i32, MVT::i32), FalseVal);
3463 TrueVal = DAG.getNode(ARMISD::VMOVRRD, dl,
3464 DAG.getVTList(MVT::i32, MVT::i32), TrueVal);
3466 SDValue TrueLow = TrueVal.getValue(0);
3467 SDValue TrueHigh = TrueVal.getValue(1);
3468 SDValue FalseLow = FalseVal.getValue(0);
3469 SDValue FalseHigh = FalseVal.getValue(1);
3471 SDValue Low = DAG.getNode(ARMISD::CMOV, dl, MVT::i32, FalseLow, TrueLow,
3473 SDValue High = DAG.getNode(ARMISD::CMOV, dl, MVT::i32, FalseHigh, TrueHigh,
3474 ARMcc, CCR, duplicateCmp(Cmp, DAG));
3476 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Low, High);
3478 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,
3483 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
3484 EVT VT = Op.getValueType();
3485 SDValue LHS = Op.getOperand(0);
3486 SDValue RHS = Op.getOperand(1);
3487 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
3488 SDValue TrueVal = Op.getOperand(2);
3489 SDValue FalseVal = Op.getOperand(3);
3492 if (Subtarget->isFPOnlySP() && LHS.getValueType() == MVT::f64) {
3493 DAG.getTargetLoweringInfo().softenSetCCOperands(DAG, MVT::f64, LHS, RHS, CC,
3496 // If softenSetCCOperands only returned one value, we should compare it to
3498 if (!RHS.getNode()) {
3499 RHS = DAG.getConstant(0, dl, LHS.getValueType());
3504 if (LHS.getValueType() == MVT::i32) {
3505 // Try to generate VSEL on ARMv8.
3506 // The VSEL instruction can't use all the usual ARM condition
3507 // codes: it only has two bits to select the condition code, so it's
3508 // constrained to use only GE, GT, VS and EQ.
3510 // To implement all the various ISD::SETXXX opcodes, we sometimes need to
3511 // swap the operands of the previous compare instruction (effectively
3512 // inverting the compare condition, swapping 'less' and 'greater') and
3513 // sometimes need to swap the operands to the VSEL (which inverts the
3514 // condition in the sense of firing whenever the previous condition didn't)
3515 if (Subtarget->hasFPARMv8() && (TrueVal.getValueType() == MVT::f32 ||
3516 TrueVal.getValueType() == MVT::f64)) {
3517 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3518 if (CondCode == ARMCC::LT || CondCode == ARMCC::LE ||
3519 CondCode == ARMCC::VC || CondCode == ARMCC::NE) {
3520 CC = ISD::getSetCCInverse(CC, true);
3521 std::swap(TrueVal, FalseVal);
3526 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3527 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3528 return getCMOV(dl, VT, FalseVal, TrueVal, ARMcc, CCR, Cmp, DAG);
3531 ARMCC::CondCodes CondCode, CondCode2;
3532 FPCCToARMCC(CC, CondCode, CondCode2);
3534 // Try to generate VMAXNM/VMINNM on ARMv8.
3535 if (Subtarget->hasFPARMv8() && (TrueVal.getValueType() == MVT::f32 ||
3536 TrueVal.getValueType() == MVT::f64)) {
3537 // We can use VMAXNM/VMINNM for a compare followed by a select with the
3538 // same operands, as follows:
3539 // c = fcmp [?gt, ?ge, ?lt, ?le] a, b
3541 // In NoNaNsFPMath the CC will have been changed from, e.g., 'ogt' to 'gt'.
3542 bool swapSides = false;
3543 if (!getTargetMachine().Options.NoNaNsFPMath) {
3544 // transformability may depend on which way around we compare
3552 // the non-NaN should be RHS
3553 swapSides = DAG.isKnownNeverNaN(LHS) && !DAG.isKnownNeverNaN(RHS);
3559 // the non-NaN should be LHS
3560 swapSides = DAG.isKnownNeverNaN(RHS) && !DAG.isKnownNeverNaN(LHS);
3564 swapSides = swapSides || (LHS == FalseVal && RHS == TrueVal);
3566 CC = ISD::getSetCCSwappedOperands(CC);
3567 std::swap(LHS, RHS);
3569 if (LHS == TrueVal && RHS == FalseVal) {
3570 bool canTransform = true;
3571 // FIXME: FastMathFlags::noSignedZeros() doesn't appear reachable from here
3572 if (!getTargetMachine().Options.UnsafeFPMath &&
3573 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
3574 const ConstantFPSDNode *Zero;
3581 // RHS must not be -0
3582 canTransform = (Zero = dyn_cast<ConstantFPSDNode>(RHS)) &&
3583 !Zero->isNegative();
3588 // LHS must not be -0
3589 canTransform = (Zero = dyn_cast<ConstantFPSDNode>(LHS)) &&
3590 !Zero->isNegative();
3595 // RHS must not be +0
3596 canTransform = (Zero = dyn_cast<ConstantFPSDNode>(RHS)) &&
3602 // LHS must not be +0
3603 canTransform = (Zero = dyn_cast<ConstantFPSDNode>(LHS)) &&
3609 // Note: If one of the elements in a pair is a number and the other
3610 // element is NaN, the corresponding result element is the number.
3611 // This is consistent with the IEEE 754-2008 standard.
3612 // Therefore, a > b ? a : b <=> vmax(a,b), if b is constant and a is NaN
3618 if (!DAG.isKnownNeverNaN(RHS))
3620 return DAG.getNode(ARMISD::VMAXNM, dl, VT, LHS, RHS);
3623 if (!DAG.isKnownNeverNaN(LHS))
3627 return DAG.getNode(ARMISD::VMAXNM, dl, VT, LHS, RHS);
3630 if (!DAG.isKnownNeverNaN(RHS))
3632 return DAG.getNode(ARMISD::VMINNM, dl, VT, LHS, RHS);
3635 if (!DAG.isKnownNeverNaN(LHS))
3639 return DAG.getNode(ARMISD::VMINNM, dl, VT, LHS, RHS);
3644 bool swpCmpOps = false;
3645 bool swpVselOps = false;
3646 checkVSELConstraints(CC, CondCode, swpCmpOps, swpVselOps);
3648 if (CondCode == ARMCC::GT || CondCode == ARMCC::GE ||
3649 CondCode == ARMCC::VS || CondCode == ARMCC::EQ) {
3651 std::swap(LHS, RHS);
3653 std::swap(TrueVal, FalseVal);
3657 SDValue ARMcc = DAG.getConstant(CondCode, dl, MVT::i32);
3658 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
3659 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3660 SDValue Result = getCMOV(dl, VT, FalseVal, TrueVal, ARMcc, CCR, Cmp, DAG);
3661 if (CondCode2 != ARMCC::AL) {
3662 SDValue ARMcc2 = DAG.getConstant(CondCode2, dl, MVT::i32);
3663 // FIXME: Needs another CMP because flag can have but one use.
3664 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
3665 Result = getCMOV(dl, VT, Result, TrueVal, ARMcc2, CCR, Cmp2, DAG);
3670 /// canChangeToInt - Given the fp compare operand, return true if it is suitable
3671 /// to morph to an integer compare sequence.
3672 static bool canChangeToInt(SDValue Op, bool &SeenZero,
3673 const ARMSubtarget *Subtarget) {
3674 SDNode *N = Op.getNode();
3675 if (!N->hasOneUse())
3676 // Otherwise it requires moving the value from fp to integer registers.
3678 if (!N->getNumValues())
3680 EVT VT = Op.getValueType();
3681 if (VT != MVT::f32 && !Subtarget->isFPBrccSlow())
3682 // f32 case is generally profitable. f64 case only makes sense when vcmpe +
3683 // vmrs are very slow, e.g. cortex-a8.
3686 if (isFloatingPointZero(Op)) {
3690 return ISD::isNormalLoad(N);
3693 static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) {
3694 if (isFloatingPointZero(Op))
3695 return DAG.getConstant(0, SDLoc(Op), MVT::i32);
3697 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op))
3698 return DAG.getLoad(MVT::i32, SDLoc(Op),
3699 Ld->getChain(), Ld->getBasePtr(), Ld->getPointerInfo(),
3700 Ld->isVolatile(), Ld->isNonTemporal(),
3701 Ld->isInvariant(), Ld->getAlignment());
3703 llvm_unreachable("Unknown VFP cmp argument!");
3706 static void expandf64Toi32(SDValue Op, SelectionDAG &DAG,
3707 SDValue &RetVal1, SDValue &RetVal2) {
3710 if (isFloatingPointZero(Op)) {
3711 RetVal1 = DAG.getConstant(0, dl, MVT::i32);
3712 RetVal2 = DAG.getConstant(0, dl, MVT::i32);
3716 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) {
3717 SDValue Ptr = Ld->getBasePtr();
3718 RetVal1 = DAG.getLoad(MVT::i32, dl,
3719 Ld->getChain(), Ptr,
3720 Ld->getPointerInfo(),
3721 Ld->isVolatile(), Ld->isNonTemporal(),
3722 Ld->isInvariant(), Ld->getAlignment());
3724 EVT PtrType = Ptr.getValueType();
3725 unsigned NewAlign = MinAlign(Ld->getAlignment(), 4);
3726 SDValue NewPtr = DAG.getNode(ISD::ADD, dl,
3727 PtrType, Ptr, DAG.getConstant(4, dl, PtrType));
3728 RetVal2 = DAG.getLoad(MVT::i32, dl,
3729 Ld->getChain(), NewPtr,
3730 Ld->getPointerInfo().getWithOffset(4),
3731 Ld->isVolatile(), Ld->isNonTemporal(),
3732 Ld->isInvariant(), NewAlign);
3736 llvm_unreachable("Unknown VFP cmp argument!");
3739 /// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some
3740 /// f32 and even f64 comparisons to integer ones.
3742 ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const {
3743 SDValue Chain = Op.getOperand(0);
3744 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3745 SDValue LHS = Op.getOperand(2);
3746 SDValue RHS = Op.getOperand(3);
3747 SDValue Dest = Op.getOperand(4);
3750 bool LHSSeenZero = false;
3751 bool LHSOk = canChangeToInt(LHS, LHSSeenZero, Subtarget);
3752 bool RHSSeenZero = false;
3753 bool RHSOk = canChangeToInt(RHS, RHSSeenZero, Subtarget);
3754 if (LHSOk && RHSOk && (LHSSeenZero || RHSSeenZero)) {
3755 // If unsafe fp math optimization is enabled and there are no other uses of
3756 // the CMP operands, and the condition code is EQ or NE, we can optimize it
3757 // to an integer comparison.
3758 if (CC == ISD::SETOEQ)
3760 else if (CC == ISD::SETUNE)
3763 SDValue Mask = DAG.getConstant(0x7fffffff, dl, MVT::i32);
3765 if (LHS.getValueType() == MVT::f32) {
3766 LHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3767 bitcastf32Toi32(LHS, DAG), Mask);
3768 RHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3769 bitcastf32Toi32(RHS, DAG), Mask);
3770 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3771 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3772 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3773 Chain, Dest, ARMcc, CCR, Cmp);
3778 expandf64Toi32(LHS, DAG, LHS1, LHS2);
3779 expandf64Toi32(RHS, DAG, RHS1, RHS2);
3780 LHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, LHS2, Mask);
3781 RHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, RHS2, Mask);
3782 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3783 ARMcc = DAG.getConstant(CondCode, dl, MVT::i32);
3784 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3785 SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest };
3786 return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops);
3792 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
3793 SDValue Chain = Op.getOperand(0);
3794 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3795 SDValue LHS = Op.getOperand(2);
3796 SDValue RHS = Op.getOperand(3);
3797 SDValue Dest = Op.getOperand(4);
3800 if (Subtarget->isFPOnlySP() && LHS.getValueType() == MVT::f64) {
3801 DAG.getTargetLoweringInfo().softenSetCCOperands(DAG, MVT::f64, LHS, RHS, CC,
3804 // If softenSetCCOperands only returned one value, we should compare it to
3806 if (!RHS.getNode()) {
3807 RHS = DAG.getConstant(0, dl, LHS.getValueType());
3812 if (LHS.getValueType() == MVT::i32) {
3814 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3815 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3816 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3817 Chain, Dest, ARMcc, CCR, Cmp);
3820 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3822 if (getTargetMachine().Options.UnsafeFPMath &&
3823 (CC == ISD::SETEQ || CC == ISD::SETOEQ ||
3824 CC == ISD::SETNE || CC == ISD::SETUNE)) {
3825 SDValue Result = OptimizeVFPBrcond(Op, DAG);
3826 if (Result.getNode())
3830 ARMCC::CondCodes CondCode, CondCode2;
3831 FPCCToARMCC(CC, CondCode, CondCode2);
3833 SDValue ARMcc = DAG.getConstant(CondCode, dl, MVT::i32);
3834 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
3835 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3836 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3837 SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp };
3838 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops);
3839 if (CondCode2 != ARMCC::AL) {
3840 ARMcc = DAG.getConstant(CondCode2, dl, MVT::i32);
3841 SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) };
3842 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops);
3847 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
3848 SDValue Chain = Op.getOperand(0);
3849 SDValue Table = Op.getOperand(1);
3850 SDValue Index = Op.getOperand(2);
3853 EVT PTy = getPointerTy();
3854 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
3855 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
3856 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI);
3857 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, dl, PTy));
3858 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
3859 if (Subtarget->isThumb2()) {
3860 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
3861 // which does another jump to the destination. This also makes it easier
3862 // to translate it to TBB / TBH later.
3863 // FIXME: This might not work if the function is extremely large.
3864 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
3865 Addr, Op.getOperand(2), JTI);
3867 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
3868 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
3869 MachinePointerInfo::getJumpTable(),
3870 false, false, false, 0);
3871 Chain = Addr.getValue(1);
3872 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
3873 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI);
3875 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
3876 MachinePointerInfo::getJumpTable(),
3877 false, false, false, 0);
3878 Chain = Addr.getValue(1);
3879 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI);
3883 static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3884 EVT VT = Op.getValueType();
3887 if (Op.getValueType().getVectorElementType() == MVT::i32) {
3888 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::f32)
3890 return DAG.UnrollVectorOp(Op.getNode());
3893 assert(Op.getOperand(0).getValueType() == MVT::v4f32 &&
3894 "Invalid type for custom lowering!");
3895 if (VT != MVT::v4i16)
3896 return DAG.UnrollVectorOp(Op.getNode());
3898 Op = DAG.getNode(Op.getOpcode(), dl, MVT::v4i32, Op.getOperand(0));
3899 return DAG.getNode(ISD::TRUNCATE, dl, VT, Op);
3902 SDValue ARMTargetLowering::LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) const {
3903 EVT VT = Op.getValueType();
3905 return LowerVectorFP_TO_INT(Op, DAG);
3906 if (Subtarget->isFPOnlySP() && Op.getOperand(0).getValueType() == MVT::f64) {
3908 if (Op.getOpcode() == ISD::FP_TO_SINT)
3909 LC = RTLIB::getFPTOSINT(Op.getOperand(0).getValueType(),
3912 LC = RTLIB::getFPTOUINT(Op.getOperand(0).getValueType(),
3914 return makeLibCall(DAG, LC, Op.getValueType(), &Op.getOperand(0), 1,
3915 /*isSigned*/ false, SDLoc(Op)).first;
3921 static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3922 EVT VT = Op.getValueType();
3925 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::i32) {
3926 if (VT.getVectorElementType() == MVT::f32)
3928 return DAG.UnrollVectorOp(Op.getNode());
3931 assert(Op.getOperand(0).getValueType() == MVT::v4i16 &&
3932 "Invalid type for custom lowering!");
3933 if (VT != MVT::v4f32)
3934 return DAG.UnrollVectorOp(Op.getNode());
3938 switch (Op.getOpcode()) {
3939 default: llvm_unreachable("Invalid opcode!");
3940 case ISD::SINT_TO_FP:
3941 CastOpc = ISD::SIGN_EXTEND;
3942 Opc = ISD::SINT_TO_FP;
3944 case ISD::UINT_TO_FP:
3945 CastOpc = ISD::ZERO_EXTEND;
3946 Opc = ISD::UINT_TO_FP;
3950 Op = DAG.getNode(CastOpc, dl, MVT::v4i32, Op.getOperand(0));
3951 return DAG.getNode(Opc, dl, VT, Op);
3954 SDValue ARMTargetLowering::LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) const {
3955 EVT VT = Op.getValueType();
3957 return LowerVectorINT_TO_FP(Op, DAG);
3958 if (Subtarget->isFPOnlySP() && Op.getValueType() == MVT::f64) {
3960 if (Op.getOpcode() == ISD::SINT_TO_FP)
3961 LC = RTLIB::getSINTTOFP(Op.getOperand(0).getValueType(),
3964 LC = RTLIB::getUINTTOFP(Op.getOperand(0).getValueType(),
3966 return makeLibCall(DAG, LC, Op.getValueType(), &Op.getOperand(0), 1,
3967 /*isSigned*/ false, SDLoc(Op)).first;
3973 SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
3974 // Implement fcopysign with a fabs and a conditional fneg.
3975 SDValue Tmp0 = Op.getOperand(0);
3976 SDValue Tmp1 = Op.getOperand(1);
3978 EVT VT = Op.getValueType();
3979 EVT SrcVT = Tmp1.getValueType();
3980 bool InGPR = Tmp0.getOpcode() == ISD::BITCAST ||
3981 Tmp0.getOpcode() == ARMISD::VMOVDRR;
3982 bool UseNEON = !InGPR && Subtarget->hasNEON();
3985 // Use VBSL to copy the sign bit.
3986 unsigned EncodedVal = ARM_AM::createNEONModImm(0x6, 0x80);
3987 SDValue Mask = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v2i32,
3988 DAG.getTargetConstant(EncodedVal, dl, MVT::i32));
3989 EVT OpVT = (VT == MVT::f32) ? MVT::v2i32 : MVT::v1i64;
3991 Mask = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3992 DAG.getNode(ISD::BITCAST, dl, OpVT, Mask),
3993 DAG.getConstant(32, dl, MVT::i32));
3994 else /*if (VT == MVT::f32)*/
3995 Tmp0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp0);
3996 if (SrcVT == MVT::f32) {
3997 Tmp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp1);
3999 Tmp1 = DAG.getNode(ARMISD::VSHL, dl, OpVT,
4000 DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1),
4001 DAG.getConstant(32, dl, MVT::i32));
4002 } else if (VT == MVT::f32)
4003 Tmp1 = DAG.getNode(ARMISD::VSHRu, dl, MVT::v1i64,
4004 DAG.getNode(ISD::BITCAST, dl, MVT::v1i64, Tmp1),
4005 DAG.getConstant(32, dl, MVT::i32));
4006 Tmp0 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp0);
4007 Tmp1 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1);
4009 SDValue AllOnes = DAG.getTargetConstant(ARM_AM::createNEONModImm(0xe, 0xff),
4011 AllOnes = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v8i8, AllOnes);
4012 SDValue MaskNot = DAG.getNode(ISD::XOR, dl, OpVT, Mask,
4013 DAG.getNode(ISD::BITCAST, dl, OpVT, AllOnes));
4015 SDValue Res = DAG.getNode(ISD::OR, dl, OpVT,
4016 DAG.getNode(ISD::AND, dl, OpVT, Tmp1, Mask),
4017 DAG.getNode(ISD::AND, dl, OpVT, Tmp0, MaskNot));
4018 if (VT == MVT::f32) {
4019 Res = DAG.getNode(ISD::BITCAST, dl, MVT::v2f32, Res);
4020 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, Res,
4021 DAG.getConstant(0, dl, MVT::i32));
4023 Res = DAG.getNode(ISD::BITCAST, dl, MVT::f64, Res);
4029 // Bitcast operand 1 to i32.
4030 if (SrcVT == MVT::f64)
4031 Tmp1 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
4033 Tmp1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp1);
4035 // Or in the signbit with integer operations.
4036 SDValue Mask1 = DAG.getConstant(0x80000000, dl, MVT::i32);
4037 SDValue Mask2 = DAG.getConstant(0x7fffffff, dl, MVT::i32);
4038 Tmp1 = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp1, Mask1);
4039 if (VT == MVT::f32) {
4040 Tmp0 = DAG.getNode(ISD::AND, dl, MVT::i32,
4041 DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp0), Mask2);
4042 return DAG.getNode(ISD::BITCAST, dl, MVT::f32,
4043 DAG.getNode(ISD::OR, dl, MVT::i32, Tmp0, Tmp1));
4046 // f64: Or the high part with signbit and then combine two parts.
4047 Tmp0 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
4049 SDValue Lo = Tmp0.getValue(0);
4050 SDValue Hi = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp0.getValue(1), Mask2);
4051 Hi = DAG.getNode(ISD::OR, dl, MVT::i32, Hi, Tmp1);
4052 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
4055 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
4056 MachineFunction &MF = DAG.getMachineFunction();
4057 MachineFrameInfo *MFI = MF.getFrameInfo();
4058 MFI->setReturnAddressIsTaken(true);
4060 if (verifyReturnAddressArgumentIsConstant(Op, DAG))
4063 EVT VT = Op.getValueType();
4065 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
4067 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
4068 SDValue Offset = DAG.getConstant(4, dl, MVT::i32);
4069 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
4070 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
4071 MachinePointerInfo(), false, false, false, 0);
4074 // Return LR, which contains the return address. Mark it an implicit live-in.
4075 unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32));
4076 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
4079 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
4080 const ARMBaseRegisterInfo &ARI =
4081 *static_cast<const ARMBaseRegisterInfo*>(RegInfo);
4082 MachineFunction &MF = DAG.getMachineFunction();
4083 MachineFrameInfo *MFI = MF.getFrameInfo();
4084 MFI->setFrameAddressIsTaken(true);
4086 EVT VT = Op.getValueType();
4087 SDLoc dl(Op); // FIXME probably not meaningful
4088 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
4089 unsigned FrameReg = ARI.getFrameRegister(MF);
4090 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
4092 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
4093 MachinePointerInfo(),
4094 false, false, false, 0);
4098 // FIXME? Maybe this could be a TableGen attribute on some registers and
4099 // this table could be generated automatically from RegInfo.
4100 unsigned ARMTargetLowering::getRegisterByName(const char* RegName,
4102 unsigned Reg = StringSwitch<unsigned>(RegName)
4103 .Case("sp", ARM::SP)
4107 report_fatal_error(Twine("Invalid register name \""
4108 + StringRef(RegName) + "\"."));
4111 // Result is 64 bit value so split into two 32 bit values and return as a
4113 static void ExpandREAD_REGISTER(SDNode *N, SmallVectorImpl<SDValue> &Results,
4114 SelectionDAG &DAG) {
4117 // This function is only supposed to be called for i64 type destination.
4118 assert(N->getValueType(0) == MVT::i64
4119 && "ExpandREAD_REGISTER called for non-i64 type result.");
4121 SDValue Read = DAG.getNode(ISD::READ_REGISTER, DL,
4122 DAG.getVTList(MVT::i32, MVT::i32, MVT::Other),
4126 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Read.getValue(0),
4128 Results.push_back(Read.getOperand(0));
4131 /// ExpandBITCAST - If the target supports VFP, this function is called to
4132 /// expand a bit convert where either the source or destination type is i64 to
4133 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
4134 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
4135 /// vectors), since the legalizer won't know what to do with that.
4136 static SDValue ExpandBITCAST(SDNode *N, SelectionDAG &DAG) {
4137 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4139 SDValue Op = N->getOperand(0);
4141 // This function is only supposed to be called for i64 types, either as the
4142 // source or destination of the bit convert.
4143 EVT SrcVT = Op.getValueType();
4144 EVT DstVT = N->getValueType(0);
4145 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
4146 "ExpandBITCAST called for non-i64 type");
4148 // Turn i64->f64 into VMOVDRR.
4149 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
4150 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
4151 DAG.getConstant(0, dl, MVT::i32));
4152 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
4153 DAG.getConstant(1, dl, MVT::i32));
4154 return DAG.getNode(ISD::BITCAST, dl, DstVT,
4155 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
4158 // Turn f64->i64 into VMOVRRD.
4159 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
4161 if (DAG.getDataLayout().isBigEndian() && SrcVT.isVector() &&
4162 SrcVT.getVectorNumElements() > 1)
4163 Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
4164 DAG.getVTList(MVT::i32, MVT::i32),
4165 DAG.getNode(ARMISD::VREV64, dl, SrcVT, Op));
4167 Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
4168 DAG.getVTList(MVT::i32, MVT::i32), Op);
4169 // Merge the pieces into a single i64 value.
4170 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
4176 /// getZeroVector - Returns a vector of specified type with all zero elements.
4177 /// Zero vectors are used to represent vector negation and in those cases
4178 /// will be implemented with the NEON VNEG instruction. However, VNEG does
4179 /// not support i64 elements, so sometimes the zero vectors will need to be
4180 /// explicitly constructed. Regardless, use a canonical VMOV to create the
4182 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, SDLoc dl) {
4183 assert(VT.isVector() && "Expected a vector type");
4184 // The canonical modified immediate encoding of a zero vector is....0!
4185 SDValue EncodedVal = DAG.getTargetConstant(0, dl, MVT::i32);
4186 EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
4187 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal);
4188 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4191 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
4192 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
4193 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
4194 SelectionDAG &DAG) const {
4195 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4196 EVT VT = Op.getValueType();
4197 unsigned VTBits = VT.getSizeInBits();
4199 SDValue ShOpLo = Op.getOperand(0);
4200 SDValue ShOpHi = Op.getOperand(1);
4201 SDValue ShAmt = Op.getOperand(2);
4203 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
4205 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
4207 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
4208 DAG.getConstant(VTBits, dl, MVT::i32), ShAmt);
4209 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
4210 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
4211 DAG.getConstant(VTBits, dl, MVT::i32));
4212 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
4213 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
4214 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
4216 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
4217 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, dl, MVT::i32),
4218 ISD::SETGE, ARMcc, DAG, dl);
4219 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
4220 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc,
4223 SDValue Ops[2] = { Lo, Hi };
4224 return DAG.getMergeValues(Ops, dl);
4227 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
4228 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
4229 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
4230 SelectionDAG &DAG) const {
4231 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4232 EVT VT = Op.getValueType();
4233 unsigned VTBits = VT.getSizeInBits();
4235 SDValue ShOpLo = Op.getOperand(0);
4236 SDValue ShOpHi = Op.getOperand(1);
4237 SDValue ShAmt = Op.getOperand(2);
4240 assert(Op.getOpcode() == ISD::SHL_PARTS);
4241 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
4242 DAG.getConstant(VTBits, dl, MVT::i32), ShAmt);
4243 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
4244 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
4245 DAG.getConstant(VTBits, dl, MVT::i32));
4246 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
4247 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
4249 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
4250 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
4251 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, dl, MVT::i32),
4252 ISD::SETGE, ARMcc, DAG, dl);
4253 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
4254 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc,
4257 SDValue Ops[2] = { Lo, Hi };
4258 return DAG.getMergeValues(Ops, dl);
4261 SDValue ARMTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
4262 SelectionDAG &DAG) const {
4263 // The rounding mode is in bits 23:22 of the FPSCR.
4264 // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0
4265 // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3)
4266 // so that the shift + and get folded into a bitfield extract.
4268 SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::i32,
4269 DAG.getConstant(Intrinsic::arm_get_fpscr, dl,
4271 SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPSCR,
4272 DAG.getConstant(1U << 22, dl, MVT::i32));
4273 SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds,
4274 DAG.getConstant(22, dl, MVT::i32));
4275 return DAG.getNode(ISD::AND, dl, MVT::i32, RMODE,
4276 DAG.getConstant(3, dl, MVT::i32));
4279 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
4280 const ARMSubtarget *ST) {
4281 EVT VT = N->getValueType(0);
4284 if (!ST->hasV6T2Ops())
4287 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
4288 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
4291 /// getCTPOP16BitCounts - Returns a v8i8/v16i8 vector containing the bit-count
4292 /// for each 16-bit element from operand, repeated. The basic idea is to
4293 /// leverage vcnt to get the 8-bit counts, gather and add the results.
4295 /// Trace for v4i16:
4296 /// input = [v0 v1 v2 v3 ] (vi 16-bit element)
4297 /// cast: N0 = [w0 w1 w2 w3 w4 w5 w6 w7] (v0 = [w0 w1], wi 8-bit element)
4298 /// vcnt: N1 = [b0 b1 b2 b3 b4 b5 b6 b7] (bi = bit-count of 8-bit element wi)
4299 /// vrev: N2 = [b1 b0 b3 b2 b5 b4 b7 b6]
4300 /// [b0 b1 b2 b3 b4 b5 b6 b7]
4301 /// +[b1 b0 b3 b2 b5 b4 b7 b6]
4302 /// N3=N1+N2 = [k0 k0 k1 k1 k2 k2 k3 k3] (k0 = b0+b1 = bit-count of 16-bit v0,
4303 /// vuzp: = [k0 k1 k2 k3 k0 k1 k2 k3] each ki is 8-bits)
4304 static SDValue getCTPOP16BitCounts(SDNode *N, SelectionDAG &DAG) {
4305 EVT VT = N->getValueType(0);
4308 EVT VT8Bit = VT.is64BitVector() ? MVT::v8i8 : MVT::v16i8;
4309 SDValue N0 = DAG.getNode(ISD::BITCAST, DL, VT8Bit, N->getOperand(0));
4310 SDValue N1 = DAG.getNode(ISD::CTPOP, DL, VT8Bit, N0);
4311 SDValue N2 = DAG.getNode(ARMISD::VREV16, DL, VT8Bit, N1);
4312 SDValue N3 = DAG.getNode(ISD::ADD, DL, VT8Bit, N1, N2);
4313 return DAG.getNode(ARMISD::VUZP, DL, VT8Bit, N3, N3);
4316 /// lowerCTPOP16BitElements - Returns a v4i16/v8i16 vector containing the
4317 /// bit-count for each 16-bit element from the operand. We need slightly
4318 /// different sequencing for v4i16 and v8i16 to stay within NEON's available
4319 /// 64/128-bit registers.
4321 /// Trace for v4i16:
4322 /// input = [v0 v1 v2 v3 ] (vi 16-bit element)
4323 /// v8i8: BitCounts = [k0 k1 k2 k3 k0 k1 k2 k3 ] (ki is the bit-count of vi)
4324 /// v8i16:Extended = [k0 k1 k2 k3 k0 k1 k2 k3 ]
4325 /// v4i16:Extracted = [k0 k1 k2 k3 ]
4326 static SDValue lowerCTPOP16BitElements(SDNode *N, SelectionDAG &DAG) {
4327 EVT VT = N->getValueType(0);
4330 SDValue BitCounts = getCTPOP16BitCounts(N, DAG);
4331 if (VT.is64BitVector()) {
4332 SDValue Extended = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v8i16, BitCounts);
4333 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i16, Extended,
4334 DAG.getIntPtrConstant(0, DL));
4336 SDValue Extracted = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v8i8,
4337 BitCounts, DAG.getIntPtrConstant(0, DL));
4338 return DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v8i16, Extracted);
4342 /// lowerCTPOP32BitElements - Returns a v2i32/v4i32 vector containing the
4343 /// bit-count for each 32-bit element from the operand. The idea here is
4344 /// to split the vector into 16-bit elements, leverage the 16-bit count
4345 /// routine, and then combine the results.
4347 /// Trace for v2i32 (v4i32 similar with Extracted/Extended exchanged):
4348 /// input = [v0 v1 ] (vi: 32-bit elements)
4349 /// Bitcast = [w0 w1 w2 w3 ] (wi: 16-bit elements, v0 = [w0 w1])
4350 /// Counts16 = [k0 k1 k2 k3 ] (ki: 16-bit elements, bit-count of wi)
4351 /// vrev: N0 = [k1 k0 k3 k2 ]
4353 /// N1 =+[k1 k0 k3 k2 ]
4355 /// N2 =+[k1 k3 k0 k2 ]
4357 /// Extended =+[k1 k3 k0 k2 ]
4359 /// Extracted=+[k1 k3 ]
4361 static SDValue lowerCTPOP32BitElements(SDNode *N, SelectionDAG &DAG) {
4362 EVT VT = N->getValueType(0);
4365 EVT VT16Bit = VT.is64BitVector() ? MVT::v4i16 : MVT::v8i16;
4367 SDValue Bitcast = DAG.getNode(ISD::BITCAST, DL, VT16Bit, N->getOperand(0));
4368 SDValue Counts16 = lowerCTPOP16BitElements(Bitcast.getNode(), DAG);
4369 SDValue N0 = DAG.getNode(ARMISD::VREV32, DL, VT16Bit, Counts16);
4370 SDValue N1 = DAG.getNode(ISD::ADD, DL, VT16Bit, Counts16, N0);
4371 SDValue N2 = DAG.getNode(ARMISD::VUZP, DL, VT16Bit, N1, N1);
4373 if (VT.is64BitVector()) {
4374 SDValue Extended = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v4i32, N2);
4375 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i32, Extended,
4376 DAG.getIntPtrConstant(0, DL));
4378 SDValue Extracted = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i16, N2,
4379 DAG.getIntPtrConstant(0, DL));
4380 return DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v4i32, Extracted);
4384 static SDValue LowerCTPOP(SDNode *N, SelectionDAG &DAG,
4385 const ARMSubtarget *ST) {
4386 EVT VT = N->getValueType(0);
4388 assert(ST->hasNEON() && "Custom ctpop lowering requires NEON.");
4389 assert((VT == MVT::v2i32 || VT == MVT::v4i32 ||
4390 VT == MVT::v4i16 || VT == MVT::v8i16) &&
4391 "Unexpected type for custom ctpop lowering");
4393 if (VT.getVectorElementType() == MVT::i32)
4394 return lowerCTPOP32BitElements(N, DAG);
4396 return lowerCTPOP16BitElements(N, DAG);
4399 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
4400 const ARMSubtarget *ST) {
4401 EVT VT = N->getValueType(0);
4407 // Lower vector shifts on NEON to use VSHL.
4408 assert(ST->hasNEON() && "unexpected vector shift");
4410 // Left shifts translate directly to the vshiftu intrinsic.
4411 if (N->getOpcode() == ISD::SHL)
4412 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
4413 DAG.getConstant(Intrinsic::arm_neon_vshiftu, dl,
4415 N->getOperand(0), N->getOperand(1));
4417 assert((N->getOpcode() == ISD::SRA ||
4418 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
4420 // NEON uses the same intrinsics for both left and right shifts. For
4421 // right shifts, the shift amounts are negative, so negate the vector of
4423 EVT ShiftVT = N->getOperand(1).getValueType();
4424 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
4425 getZeroVector(ShiftVT, DAG, dl),
4427 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
4428 Intrinsic::arm_neon_vshifts :
4429 Intrinsic::arm_neon_vshiftu);
4430 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
4431 DAG.getConstant(vshiftInt, dl, MVT::i32),
4432 N->getOperand(0), NegatedCount);
4435 static SDValue Expand64BitShift(SDNode *N, SelectionDAG &DAG,
4436 const ARMSubtarget *ST) {
4437 EVT VT = N->getValueType(0);
4440 // We can get here for a node like i32 = ISD::SHL i32, i64
4444 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
4445 "Unknown shift to lower!");
4447 // We only lower SRA, SRL of 1 here, all others use generic lowering.
4448 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
4449 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
4452 // If we are in thumb mode, we don't have RRX.
4453 if (ST->isThumb1Only()) return SDValue();
4455 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
4456 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
4457 DAG.getConstant(0, dl, MVT::i32));
4458 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
4459 DAG.getConstant(1, dl, MVT::i32));
4461 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
4462 // captures the result into a carry flag.
4463 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
4464 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Glue), Hi);
4466 // The low part is an ARMISD::RRX operand, which shifts the carry in.
4467 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
4469 // Merge the pieces into a single i64 value.
4470 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
4473 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
4474 SDValue TmpOp0, TmpOp1;
4475 bool Invert = false;
4479 SDValue Op0 = Op.getOperand(0);
4480 SDValue Op1 = Op.getOperand(1);
4481 SDValue CC = Op.getOperand(2);
4482 EVT CmpVT = Op0.getValueType().changeVectorElementTypeToInteger();
4483 EVT VT = Op.getValueType();
4484 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
4487 if (Op1.getValueType().isFloatingPoint()) {
4488 switch (SetCCOpcode) {
4489 default: llvm_unreachable("Illegal FP comparison");
4491 case ISD::SETNE: Invert = true; // Fallthrough
4493 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
4495 case ISD::SETLT: Swap = true; // Fallthrough
4497 case ISD::SETGT: Opc = ARMISD::VCGT; break;
4499 case ISD::SETLE: Swap = true; // Fallthrough
4501 case ISD::SETGE: Opc = ARMISD::VCGE; break;
4502 case ISD::SETUGE: Swap = true; // Fallthrough
4503 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
4504 case ISD::SETUGT: Swap = true; // Fallthrough
4505 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
4506 case ISD::SETUEQ: Invert = true; // Fallthrough
4508 // Expand this to (OLT | OGT).
4512 Op0 = DAG.getNode(ARMISD::VCGT, dl, CmpVT, TmpOp1, TmpOp0);
4513 Op1 = DAG.getNode(ARMISD::VCGT, dl, CmpVT, TmpOp0, TmpOp1);
4515 case ISD::SETUO: Invert = true; // Fallthrough
4517 // Expand this to (OLT | OGE).
4521 Op0 = DAG.getNode(ARMISD::VCGT, dl, CmpVT, TmpOp1, TmpOp0);
4522 Op1 = DAG.getNode(ARMISD::VCGE, dl, CmpVT, TmpOp0, TmpOp1);
4526 // Integer comparisons.
4527 switch (SetCCOpcode) {
4528 default: llvm_unreachable("Illegal integer comparison");
4529 case ISD::SETNE: Invert = true;
4530 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
4531 case ISD::SETLT: Swap = true;
4532 case ISD::SETGT: Opc = ARMISD::VCGT; break;
4533 case ISD::SETLE: Swap = true;
4534 case ISD::SETGE: Opc = ARMISD::VCGE; break;
4535 case ISD::SETULT: Swap = true;
4536 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
4537 case ISD::SETULE: Swap = true;
4538 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
4541 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
4542 if (Opc == ARMISD::VCEQ) {
4545 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
4547 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
4550 // Ignore bitconvert.
4551 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BITCAST)
4552 AndOp = AndOp.getOperand(0);
4554 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
4556 Op0 = DAG.getNode(ISD::BITCAST, dl, CmpVT, AndOp.getOperand(0));
4557 Op1 = DAG.getNode(ISD::BITCAST, dl, CmpVT, AndOp.getOperand(1));
4564 std::swap(Op0, Op1);
4566 // If one of the operands is a constant vector zero, attempt to fold the
4567 // comparison to a specialized compare-against-zero form.
4569 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
4571 else if (ISD::isBuildVectorAllZeros(Op0.getNode())) {
4572 if (Opc == ARMISD::VCGE)
4573 Opc = ARMISD::VCLEZ;
4574 else if (Opc == ARMISD::VCGT)
4575 Opc = ARMISD::VCLTZ;
4580 if (SingleOp.getNode()) {
4583 Result = DAG.getNode(ARMISD::VCEQZ, dl, CmpVT, SingleOp); break;
4585 Result = DAG.getNode(ARMISD::VCGEZ, dl, CmpVT, SingleOp); break;
4587 Result = DAG.getNode(ARMISD::VCLEZ, dl, CmpVT, SingleOp); break;
4589 Result = DAG.getNode(ARMISD::VCGTZ, dl, CmpVT, SingleOp); break;
4591 Result = DAG.getNode(ARMISD::VCLTZ, dl, CmpVT, SingleOp); break;
4593 Result = DAG.getNode(Opc, dl, CmpVT, Op0, Op1);
4596 Result = DAG.getNode(Opc, dl, CmpVT, Op0, Op1);
4599 Result = DAG.getSExtOrTrunc(Result, dl, VT);
4602 Result = DAG.getNOT(dl, Result, VT);
4607 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
4608 /// valid vector constant for a NEON instruction with a "modified immediate"
4609 /// operand (e.g., VMOV). If so, return the encoded value.
4610 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
4611 unsigned SplatBitSize, SelectionDAG &DAG,
4612 SDLoc dl, EVT &VT, bool is128Bits,
4613 NEONModImmType type) {
4614 unsigned OpCmode, Imm;
4616 // SplatBitSize is set to the smallest size that splats the vector, so a
4617 // zero vector will always have SplatBitSize == 8. However, NEON modified
4618 // immediate instructions others than VMOV do not support the 8-bit encoding
4619 // of a zero vector, and the default encoding of zero is supposed to be the
4624 switch (SplatBitSize) {
4626 if (type != VMOVModImm)
4628 // Any 1-byte value is OK. Op=0, Cmode=1110.
4629 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
4632 VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
4636 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
4637 VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
4638 if ((SplatBits & ~0xff) == 0) {
4639 // Value = 0x00nn: Op=x, Cmode=100x.
4644 if ((SplatBits & ~0xff00) == 0) {
4645 // Value = 0xnn00: Op=x, Cmode=101x.
4647 Imm = SplatBits >> 8;
4653 // NEON's 32-bit VMOV supports splat values where:
4654 // * only one byte is nonzero, or
4655 // * the least significant byte is 0xff and the second byte is nonzero, or
4656 // * the least significant 2 bytes are 0xff and the third is nonzero.
4657 VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
4658 if ((SplatBits & ~0xff) == 0) {
4659 // Value = 0x000000nn: Op=x, Cmode=000x.
4664 if ((SplatBits & ~0xff00) == 0) {
4665 // Value = 0x0000nn00: Op=x, Cmode=001x.
4667 Imm = SplatBits >> 8;
4670 if ((SplatBits & ~0xff0000) == 0) {
4671 // Value = 0x00nn0000: Op=x, Cmode=010x.
4673 Imm = SplatBits >> 16;
4676 if ((SplatBits & ~0xff000000) == 0) {
4677 // Value = 0xnn000000: Op=x, Cmode=011x.
4679 Imm = SplatBits >> 24;
4683 // cmode == 0b1100 and cmode == 0b1101 are not supported for VORR or VBIC
4684 if (type == OtherModImm) return SDValue();
4686 if ((SplatBits & ~0xffff) == 0 &&
4687 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
4688 // Value = 0x0000nnff: Op=x, Cmode=1100.
4690 Imm = SplatBits >> 8;
4694 if ((SplatBits & ~0xffffff) == 0 &&
4695 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
4696 // Value = 0x00nnffff: Op=x, Cmode=1101.
4698 Imm = SplatBits >> 16;
4702 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
4703 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
4704 // VMOV.I32. A (very) minor optimization would be to replicate the value
4705 // and fall through here to test for a valid 64-bit splat. But, then the
4706 // caller would also need to check and handle the change in size.
4710 if (type != VMOVModImm)
4712 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
4713 uint64_t BitMask = 0xff;
4715 unsigned ImmMask = 1;
4717 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
4718 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
4721 } else if ((SplatBits & BitMask) != 0) {
4728 if (DAG.getDataLayout().isBigEndian())
4729 // swap higher and lower 32 bit word
4730 Imm = ((Imm & 0xf) << 4) | ((Imm & 0xf0) >> 4);
4732 // Op=1, Cmode=1110.
4734 VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
4739 llvm_unreachable("unexpected size for isNEONModifiedImm");
4742 unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm);
4743 return DAG.getTargetConstant(EncodedVal, dl, MVT::i32);
4746 SDValue ARMTargetLowering::LowerConstantFP(SDValue Op, SelectionDAG &DAG,
4747 const ARMSubtarget *ST) const {
4751 bool IsDouble = Op.getValueType() == MVT::f64;
4752 ConstantFPSDNode *CFP = cast<ConstantFPSDNode>(Op);
4754 // Use the default (constant pool) lowering for double constants when we have
4756 if (IsDouble && Subtarget->isFPOnlySP())
4759 // Try splatting with a VMOV.f32...
4760 APFloat FPVal = CFP->getValueAPF();
4761 int ImmVal = IsDouble ? ARM_AM::getFP64Imm(FPVal) : ARM_AM::getFP32Imm(FPVal);
4764 if (IsDouble || !ST->useNEONForSinglePrecisionFP()) {
4765 // We have code in place to select a valid ConstantFP already, no need to
4770 // It's a float and we are trying to use NEON operations where
4771 // possible. Lower it to a splat followed by an extract.
4773 SDValue NewVal = DAG.getTargetConstant(ImmVal, DL, MVT::i32);
4774 SDValue VecConstant = DAG.getNode(ARMISD::VMOVFPIMM, DL, MVT::v2f32,
4776 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecConstant,
4777 DAG.getConstant(0, DL, MVT::i32));
4780 // The rest of our options are NEON only, make sure that's allowed before
4782 if (!ST->hasNEON() || (!IsDouble && !ST->useNEONForSinglePrecisionFP()))
4786 uint64_t iVal = FPVal.bitcastToAPInt().getZExtValue();
4788 // It wouldn't really be worth bothering for doubles except for one very
4789 // important value, which does happen to match: 0.0. So make sure we don't do
4791 if (IsDouble && (iVal & 0xffffffff) != (iVal >> 32))
4794 // Try a VMOV.i32 (FIXME: i8, i16, or i64 could work too).
4795 SDValue NewVal = isNEONModifiedImm(iVal & 0xffffffffU, 0, 32, DAG, SDLoc(Op),
4796 VMovVT, false, VMOVModImm);
4797 if (NewVal != SDValue()) {
4799 SDValue VecConstant = DAG.getNode(ARMISD::VMOVIMM, DL, VMovVT,
4802 return DAG.getNode(ISD::BITCAST, DL, MVT::f64, VecConstant);
4804 // It's a float: cast and extract a vector element.
4805 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
4807 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
4808 DAG.getConstant(0, DL, MVT::i32));
4811 // Finally, try a VMVN.i32
4812 NewVal = isNEONModifiedImm(~iVal & 0xffffffffU, 0, 32, DAG, SDLoc(Op), VMovVT,
4814 if (NewVal != SDValue()) {
4816 SDValue VecConstant = DAG.getNode(ARMISD::VMVNIMM, DL, VMovVT, NewVal);
4819 return DAG.getNode(ISD::BITCAST, DL, MVT::f64, VecConstant);
4821 // It's a float: cast and extract a vector element.
4822 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
4824 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
4825 DAG.getConstant(0, DL, MVT::i32));
4831 // check if an VEXT instruction can handle the shuffle mask when the
4832 // vector sources of the shuffle are the same.
4833 static bool isSingletonVEXTMask(ArrayRef<int> M, EVT VT, unsigned &Imm) {
4834 unsigned NumElts = VT.getVectorNumElements();
4836 // Assume that the first shuffle index is not UNDEF. Fail if it is.
4842 // If this is a VEXT shuffle, the immediate value is the index of the first
4843 // element. The other shuffle indices must be the successive elements after
4845 unsigned ExpectedElt = Imm;
4846 for (unsigned i = 1; i < NumElts; ++i) {
4847 // Increment the expected index. If it wraps around, just follow it
4848 // back to index zero and keep going.
4850 if (ExpectedElt == NumElts)
4853 if (M[i] < 0) continue; // ignore UNDEF indices
4854 if (ExpectedElt != static_cast<unsigned>(M[i]))
4862 static bool isVEXTMask(ArrayRef<int> M, EVT VT,
4863 bool &ReverseVEXT, unsigned &Imm) {
4864 unsigned NumElts = VT.getVectorNumElements();
4865 ReverseVEXT = false;
4867 // Assume that the first shuffle index is not UNDEF. Fail if it is.
4873 // If this is a VEXT shuffle, the immediate value is the index of the first
4874 // element. The other shuffle indices must be the successive elements after
4876 unsigned ExpectedElt = Imm;
4877 for (unsigned i = 1; i < NumElts; ++i) {
4878 // Increment the expected index. If it wraps around, it may still be
4879 // a VEXT but the source vectors must be swapped.
4881 if (ExpectedElt == NumElts * 2) {
4886 if (M[i] < 0) continue; // ignore UNDEF indices
4887 if (ExpectedElt != static_cast<unsigned>(M[i]))
4891 // Adjust the index value if the source operands will be swapped.
4898 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
4899 /// instruction with the specified blocksize. (The order of the elements
4900 /// within each block of the vector is reversed.)
4901 static bool isVREVMask(ArrayRef<int> M, EVT VT, unsigned BlockSize) {
4902 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
4903 "Only possible block sizes for VREV are: 16, 32, 64");
4905 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4909 unsigned NumElts = VT.getVectorNumElements();
4910 unsigned BlockElts = M[0] + 1;
4911 // If the first shuffle index is UNDEF, be optimistic.
4913 BlockElts = BlockSize / EltSz;
4915 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
4918 for (unsigned i = 0; i < NumElts; ++i) {
4919 if (M[i] < 0) continue; // ignore UNDEF indices
4920 if ((unsigned) M[i] != (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
4927 static bool isVTBLMask(ArrayRef<int> M, EVT VT) {
4928 // We can handle <8 x i8> vector shuffles. If the index in the mask is out of
4929 // range, then 0 is placed into the resulting vector. So pretty much any mask
4930 // of 8 elements can work here.
4931 return VT == MVT::v8i8 && M.size() == 8;
4934 static bool isVTRNMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4935 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4939 unsigned NumElts = VT.getVectorNumElements();
4940 WhichResult = (M[0] == 0 ? 0 : 1);
4941 for (unsigned i = 0; i < NumElts; i += 2) {
4942 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
4943 (M[i+1] >= 0 && (unsigned) M[i+1] != i + NumElts + WhichResult))
4949 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
4950 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4951 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
4952 static bool isVTRN_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4953 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4957 unsigned NumElts = VT.getVectorNumElements();
4958 WhichResult = (M[0] == 0 ? 0 : 1);
4959 for (unsigned i = 0; i < NumElts; i += 2) {
4960 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
4961 (M[i+1] >= 0 && (unsigned) M[i+1] != i + WhichResult))
4967 static bool isVUZPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4968 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4972 unsigned NumElts = VT.getVectorNumElements();
4973 WhichResult = (M[0] == 0 ? 0 : 1);
4974 for (unsigned i = 0; i != NumElts; ++i) {
4975 if (M[i] < 0) continue; // ignore UNDEF indices
4976 if ((unsigned) M[i] != 2 * i + WhichResult)
4980 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4981 if (VT.is64BitVector() && EltSz == 32)
4987 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
4988 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4989 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
4990 static bool isVUZP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4991 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4995 unsigned Half = VT.getVectorNumElements() / 2;
4996 WhichResult = (M[0] == 0 ? 0 : 1);
4997 for (unsigned j = 0; j != 2; ++j) {
4998 unsigned Idx = WhichResult;
4999 for (unsigned i = 0; i != Half; ++i) {
5000 int MIdx = M[i + j * Half];
5001 if (MIdx >= 0 && (unsigned) MIdx != Idx)
5007 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
5008 if (VT.is64BitVector() && EltSz == 32)
5014 static bool isVZIPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
5015 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
5019 unsigned NumElts = VT.getVectorNumElements();
5020 WhichResult = (M[0] == 0 ? 0 : 1);
5021 unsigned Idx = WhichResult * NumElts / 2;
5022 for (unsigned i = 0; i != NumElts; i += 2) {
5023 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
5024 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx + NumElts))
5029 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
5030 if (VT.is64BitVector() && EltSz == 32)
5036 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
5037 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
5038 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
5039 static bool isVZIP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
5040 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
5044 unsigned NumElts = VT.getVectorNumElements();
5045 WhichResult = (M[0] == 0 ? 0 : 1);
5046 unsigned Idx = WhichResult * NumElts / 2;
5047 for (unsigned i = 0; i != NumElts; i += 2) {
5048 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
5049 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx))
5054 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
5055 if (VT.is64BitVector() && EltSz == 32)
5061 /// Check if \p ShuffleMask is a NEON two-result shuffle (VZIP, VUZP, VTRN),
5062 /// and return the corresponding ARMISD opcode if it is, or 0 if it isn't.
5063 static unsigned isNEONTwoResultShuffleMask(ArrayRef<int> ShuffleMask, EVT VT,
5064 unsigned &WhichResult,
5067 if (isVTRNMask(ShuffleMask, VT, WhichResult))
5068 return ARMISD::VTRN;
5069 if (isVUZPMask(ShuffleMask, VT, WhichResult))
5070 return ARMISD::VUZP;
5071 if (isVZIPMask(ShuffleMask, VT, WhichResult))
5072 return ARMISD::VZIP;
5075 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
5076 return ARMISD::VTRN;
5077 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
5078 return ARMISD::VUZP;
5079 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
5080 return ARMISD::VZIP;
5085 /// \return true if this is a reverse operation on an vector.
5086 static bool isReverseMask(ArrayRef<int> M, EVT VT) {
5087 unsigned NumElts = VT.getVectorNumElements();
5088 // Make sure the mask has the right size.
5089 if (NumElts != M.size())
5092 // Look for <15, ..., 3, -1, 1, 0>.
5093 for (unsigned i = 0; i != NumElts; ++i)
5094 if (M[i] >= 0 && M[i] != (int) (NumElts - 1 - i))
5100 // If N is an integer constant that can be moved into a register in one
5101 // instruction, return an SDValue of such a constant (will become a MOV
5102 // instruction). Otherwise return null.
5103 static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG,
5104 const ARMSubtarget *ST, SDLoc dl) {
5106 if (!isa<ConstantSDNode>(N))
5108 Val = cast<ConstantSDNode>(N)->getZExtValue();
5110 if (ST->isThumb1Only()) {
5111 if (Val <= 255 || ~Val <= 255)
5112 return DAG.getConstant(Val, dl, MVT::i32);
5114 if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1)
5115 return DAG.getConstant(Val, dl, MVT::i32);
5120 // If this is a case we can't handle, return null and let the default
5121 // expansion code take care of it.
5122 SDValue ARMTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
5123 const ARMSubtarget *ST) const {
5124 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
5126 EVT VT = Op.getValueType();
5128 APInt SplatBits, SplatUndef;
5129 unsigned SplatBitSize;
5131 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
5132 if (SplatBitSize <= 64) {
5133 // Check if an immediate VMOV works.
5135 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
5136 SplatUndef.getZExtValue(), SplatBitSize,
5137 DAG, dl, VmovVT, VT.is128BitVector(),
5139 if (Val.getNode()) {
5140 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
5141 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
5144 // Try an immediate VMVN.
5145 uint64_t NegatedImm = (~SplatBits).getZExtValue();
5146 Val = isNEONModifiedImm(NegatedImm,
5147 SplatUndef.getZExtValue(), SplatBitSize,
5148 DAG, dl, VmovVT, VT.is128BitVector(),
5150 if (Val.getNode()) {
5151 SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
5152 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
5155 // Use vmov.f32 to materialize other v2f32 and v4f32 splats.
5156 if ((VT == MVT::v2f32 || VT == MVT::v4f32) && SplatBitSize == 32) {
5157 int ImmVal = ARM_AM::getFP32Imm(SplatBits);
5159 SDValue Val = DAG.getTargetConstant(ImmVal, dl, MVT::i32);
5160 return DAG.getNode(ARMISD::VMOVFPIMM, dl, VT, Val);
5166 // Scan through the operands to see if only one value is used.
5168 // As an optimisation, even if more than one value is used it may be more
5169 // profitable to splat with one value then change some lanes.
5171 // Heuristically we decide to do this if the vector has a "dominant" value,
5172 // defined as splatted to more than half of the lanes.
5173 unsigned NumElts = VT.getVectorNumElements();
5174 bool isOnlyLowElement = true;
5175 bool usesOnlyOneValue = true;
5176 bool hasDominantValue = false;
5177 bool isConstant = true;
5179 // Map of the number of times a particular SDValue appears in the
5181 DenseMap<SDValue, unsigned> ValueCounts;
5183 for (unsigned i = 0; i < NumElts; ++i) {
5184 SDValue V = Op.getOperand(i);
5185 if (V.getOpcode() == ISD::UNDEF)
5188 isOnlyLowElement = false;
5189 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
5192 ValueCounts.insert(std::make_pair(V, 0));
5193 unsigned &Count = ValueCounts[V];
5195 // Is this value dominant? (takes up more than half of the lanes)
5196 if (++Count > (NumElts / 2)) {
5197 hasDominantValue = true;
5201 if (ValueCounts.size() != 1)
5202 usesOnlyOneValue = false;
5203 if (!Value.getNode() && ValueCounts.size() > 0)
5204 Value = ValueCounts.begin()->first;
5206 if (ValueCounts.size() == 0)
5207 return DAG.getUNDEF(VT);
5209 // Loads are better lowered with insert_vector_elt/ARMISD::BUILD_VECTOR.
5210 // Keep going if we are hitting this case.
5211 if (isOnlyLowElement && !ISD::isNormalLoad(Value.getNode()))
5212 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
5214 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5216 // Use VDUP for non-constant splats. For f32 constant splats, reduce to
5217 // i32 and try again.
5218 if (hasDominantValue && EltSize <= 32) {
5222 // If we are VDUPing a value that comes directly from a vector, that will
5223 // cause an unnecessary move to and from a GPR, where instead we could
5224 // just use VDUPLANE. We can only do this if the lane being extracted
5225 // is at a constant index, as the VDUP from lane instructions only have
5226 // constant-index forms.
5227 if (Value->getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5228 isa<ConstantSDNode>(Value->getOperand(1))) {
5229 // We need to create a new undef vector to use for the VDUPLANE if the
5230 // size of the vector from which we get the value is different than the
5231 // size of the vector that we need to create. We will insert the element
5232 // such that the register coalescer will remove unnecessary copies.
5233 if (VT != Value->getOperand(0).getValueType()) {
5234 ConstantSDNode *constIndex;
5235 constIndex = dyn_cast<ConstantSDNode>(Value->getOperand(1));
5236 assert(constIndex && "The index is not a constant!");
5237 unsigned index = constIndex->getAPIntValue().getLimitedValue() %
5238 VT.getVectorNumElements();
5239 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
5240 DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DAG.getUNDEF(VT),
5241 Value, DAG.getConstant(index, dl, MVT::i32)),
5242 DAG.getConstant(index, dl, MVT::i32));
5244 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
5245 Value->getOperand(0), Value->getOperand(1));
5247 N = DAG.getNode(ARMISD::VDUP, dl, VT, Value);
5249 if (!usesOnlyOneValue) {
5250 // The dominant value was splatted as 'N', but we now have to insert
5251 // all differing elements.
5252 for (unsigned I = 0; I < NumElts; ++I) {
5253 if (Op.getOperand(I) == Value)
5255 SmallVector<SDValue, 3> Ops;
5257 Ops.push_back(Op.getOperand(I));
5258 Ops.push_back(DAG.getConstant(I, dl, MVT::i32));
5259 N = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Ops);
5264 if (VT.getVectorElementType().isFloatingPoint()) {
5265 SmallVector<SDValue, 8> Ops;
5266 for (unsigned i = 0; i < NumElts; ++i)
5267 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, MVT::i32,
5269 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
5270 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, Ops);
5271 Val = LowerBUILD_VECTOR(Val, DAG, ST);
5273 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
5275 if (usesOnlyOneValue) {
5276 SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
5277 if (isConstant && Val.getNode())
5278 return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
5282 // If all elements are constants and the case above didn't get hit, fall back
5283 // to the default expansion, which will generate a load from the constant
5288 // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
5290 SDValue shuffle = ReconstructShuffle(Op, DAG);
5291 if (shuffle != SDValue())
5295 // Vectors with 32- or 64-bit elements can be built by directly assigning
5296 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
5297 // will be legalized.
5298 if (EltSize >= 32) {
5299 // Do the expansion with floating-point types, since that is what the VFP
5300 // registers are defined to use, and since i64 is not legal.
5301 EVT EltVT = EVT::getFloatingPointVT(EltSize);
5302 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
5303 SmallVector<SDValue, 8> Ops;
5304 for (unsigned i = 0; i < NumElts; ++i)
5305 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, EltVT, Op.getOperand(i)));
5306 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, Ops);
5307 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
5310 // If all else fails, just use a sequence of INSERT_VECTOR_ELT when we
5311 // know the default expansion would otherwise fall back on something even
5312 // worse. For a vector with one or two non-undef values, that's
5313 // scalar_to_vector for the elements followed by a shuffle (provided the
5314 // shuffle is valid for the target) and materialization element by element
5315 // on the stack followed by a load for everything else.
5316 if (!isConstant && !usesOnlyOneValue) {
5317 SDValue Vec = DAG.getUNDEF(VT);
5318 for (unsigned i = 0 ; i < NumElts; ++i) {
5319 SDValue V = Op.getOperand(i);
5320 if (V.getOpcode() == ISD::UNDEF)
5322 SDValue LaneIdx = DAG.getConstant(i, dl, MVT::i32);
5323 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Vec, V, LaneIdx);
5331 // Gather data to see if the operation can be modelled as a
5332 // shuffle in combination with VEXTs.
5333 SDValue ARMTargetLowering::ReconstructShuffle(SDValue Op,
5334 SelectionDAG &DAG) const {
5336 EVT VT = Op.getValueType();
5337 unsigned NumElts = VT.getVectorNumElements();
5339 SmallVector<SDValue, 2> SourceVecs;
5340 SmallVector<unsigned, 2> MinElts;
5341 SmallVector<unsigned, 2> MaxElts;
5343 for (unsigned i = 0; i < NumElts; ++i) {
5344 SDValue V = Op.getOperand(i);
5345 if (V.getOpcode() == ISD::UNDEF)
5347 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
5348 // A shuffle can only come from building a vector from various
5349 // elements of other vectors.
5351 } else if (V.getOperand(0).getValueType().getVectorElementType() !=
5352 VT.getVectorElementType()) {
5353 // This code doesn't know how to handle shuffles where the vector
5354 // element types do not match (this happens because type legalization
5355 // promotes the return type of EXTRACT_VECTOR_ELT).
5356 // FIXME: It might be appropriate to extend this code to handle
5357 // mismatched types.
5361 // Record this extraction against the appropriate vector if possible...
5362 SDValue SourceVec = V.getOperand(0);
5363 // If the element number isn't a constant, we can't effectively
5364 // analyze what's going on.
5365 if (!isa<ConstantSDNode>(V.getOperand(1)))
5367 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
5368 bool FoundSource = false;
5369 for (unsigned j = 0; j < SourceVecs.size(); ++j) {
5370 if (SourceVecs[j] == SourceVec) {
5371 if (MinElts[j] > EltNo)
5373 if (MaxElts[j] < EltNo)
5380 // Or record a new source if not...
5382 SourceVecs.push_back(SourceVec);
5383 MinElts.push_back(EltNo);
5384 MaxElts.push_back(EltNo);
5388 // Currently only do something sane when at most two source vectors
5390 if (SourceVecs.size() > 2)
5393 SDValue ShuffleSrcs[2] = {DAG.getUNDEF(VT), DAG.getUNDEF(VT) };
5394 int VEXTOffsets[2] = {0, 0};
5396 // This loop extracts the usage patterns of the source vectors
5397 // and prepares appropriate SDValues for a shuffle if possible.
5398 for (unsigned i = 0; i < SourceVecs.size(); ++i) {
5399 if (SourceVecs[i].getValueType() == VT) {
5400 // No VEXT necessary
5401 ShuffleSrcs[i] = SourceVecs[i];
5404 } else if (SourceVecs[i].getValueType().getVectorNumElements() < NumElts) {
5405 // It probably isn't worth padding out a smaller vector just to
5406 // break it down again in a shuffle.
5410 // Since only 64-bit and 128-bit vectors are legal on ARM and
5411 // we've eliminated the other cases...
5412 assert(SourceVecs[i].getValueType().getVectorNumElements() == 2*NumElts &&
5413 "unexpected vector sizes in ReconstructShuffle");
5415 if (MaxElts[i] - MinElts[i] >= NumElts) {
5416 // Span too large for a VEXT to cope
5420 if (MinElts[i] >= NumElts) {
5421 // The extraction can just take the second half
5422 VEXTOffsets[i] = NumElts;
5423 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5425 DAG.getIntPtrConstant(NumElts, dl));
5426 } else if (MaxElts[i] < NumElts) {
5427 // The extraction can just take the first half
5429 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5431 DAG.getIntPtrConstant(0, dl));
5433 // An actual VEXT is needed
5434 VEXTOffsets[i] = MinElts[i];
5435 SDValue VEXTSrc1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5437 DAG.getIntPtrConstant(0, dl));
5438 SDValue VEXTSrc2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5440 DAG.getIntPtrConstant(NumElts, dl));
5441 ShuffleSrcs[i] = DAG.getNode(ARMISD::VEXT, dl, VT, VEXTSrc1, VEXTSrc2,
5442 DAG.getConstant(VEXTOffsets[i], dl,
5447 SmallVector<int, 8> Mask;
5449 for (unsigned i = 0; i < NumElts; ++i) {
5450 SDValue Entry = Op.getOperand(i);
5451 if (Entry.getOpcode() == ISD::UNDEF) {
5456 SDValue ExtractVec = Entry.getOperand(0);
5457 int ExtractElt = cast<ConstantSDNode>(Op.getOperand(i)
5458 .getOperand(1))->getSExtValue();
5459 if (ExtractVec == SourceVecs[0]) {
5460 Mask.push_back(ExtractElt - VEXTOffsets[0]);
5462 Mask.push_back(ExtractElt + NumElts - VEXTOffsets[1]);
5466 // Final check before we try to produce nonsense...
5467 if (isShuffleMaskLegal(Mask, VT))
5468 return DAG.getVectorShuffle(VT, dl, ShuffleSrcs[0], ShuffleSrcs[1],
5474 /// isShuffleMaskLegal - Targets can use this to indicate that they only
5475 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
5476 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
5477 /// are assumed to be legal.
5479 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
5481 if (VT.getVectorNumElements() == 4 &&
5482 (VT.is128BitVector() || VT.is64BitVector())) {
5483 unsigned PFIndexes[4];
5484 for (unsigned i = 0; i != 4; ++i) {
5488 PFIndexes[i] = M[i];
5491 // Compute the index in the perfect shuffle table.
5492 unsigned PFTableIndex =
5493 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
5494 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
5495 unsigned Cost = (PFEntry >> 30);
5501 bool ReverseVEXT, isV_UNDEF;
5502 unsigned Imm, WhichResult;
5504 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5505 return (EltSize >= 32 ||
5506 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
5507 isVREVMask(M, VT, 64) ||
5508 isVREVMask(M, VT, 32) ||
5509 isVREVMask(M, VT, 16) ||
5510 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
5511 isVTBLMask(M, VT) ||
5512 isNEONTwoResultShuffleMask(M, VT, WhichResult, isV_UNDEF) ||
5513 ((VT == MVT::v8i16 || VT == MVT::v16i8) && isReverseMask(M, VT)));
5516 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
5517 /// the specified operations to build the shuffle.
5518 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
5519 SDValue RHS, SelectionDAG &DAG,
5521 unsigned OpNum = (PFEntry >> 26) & 0x0F;
5522 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
5523 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
5526 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
5535 OP_VUZPL, // VUZP, left result
5536 OP_VUZPR, // VUZP, right result
5537 OP_VZIPL, // VZIP, left result
5538 OP_VZIPR, // VZIP, right result
5539 OP_VTRNL, // VTRN, left result
5540 OP_VTRNR // VTRN, right result
5543 if (OpNum == OP_COPY) {
5544 if (LHSID == (1*9+2)*9+3) return LHS;
5545 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
5549 SDValue OpLHS, OpRHS;
5550 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
5551 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
5552 EVT VT = OpLHS.getValueType();
5555 default: llvm_unreachable("Unknown shuffle opcode!");
5557 // VREV divides the vector in half and swaps within the half.
5558 if (VT.getVectorElementType() == MVT::i32 ||
5559 VT.getVectorElementType() == MVT::f32)
5560 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
5561 // vrev <4 x i16> -> VREV32
5562 if (VT.getVectorElementType() == MVT::i16)
5563 return DAG.getNode(ARMISD::VREV32, dl, VT, OpLHS);
5564 // vrev <4 x i8> -> VREV16
5565 assert(VT.getVectorElementType() == MVT::i8);
5566 return DAG.getNode(ARMISD::VREV16, dl, VT, OpLHS);
5571 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
5572 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, dl, MVT::i32));
5576 return DAG.getNode(ARMISD::VEXT, dl, VT,
5578 DAG.getConstant(OpNum - OP_VEXT1 + 1, dl, MVT::i32));
5581 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
5582 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
5585 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
5586 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
5589 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
5590 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
5594 static SDValue LowerVECTOR_SHUFFLEv8i8(SDValue Op,
5595 ArrayRef<int> ShuffleMask,
5596 SelectionDAG &DAG) {
5597 // Check to see if we can use the VTBL instruction.
5598 SDValue V1 = Op.getOperand(0);
5599 SDValue V2 = Op.getOperand(1);
5602 SmallVector<SDValue, 8> VTBLMask;
5603 for (ArrayRef<int>::iterator
5604 I = ShuffleMask.begin(), E = ShuffleMask.end(); I != E; ++I)
5605 VTBLMask.push_back(DAG.getConstant(*I, DL, MVT::i32));
5607 if (V2.getNode()->getOpcode() == ISD::UNDEF)
5608 return DAG.getNode(ARMISD::VTBL1, DL, MVT::v8i8, V1,
5609 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8, VTBLMask));
5611 return DAG.getNode(ARMISD::VTBL2, DL, MVT::v8i8, V1, V2,
5612 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8, VTBLMask));
5615 static SDValue LowerReverse_VECTOR_SHUFFLEv16i8_v8i16(SDValue Op,
5616 SelectionDAG &DAG) {
5618 SDValue OpLHS = Op.getOperand(0);
5619 EVT VT = OpLHS.getValueType();
5621 assert((VT == MVT::v8i16 || VT == MVT::v16i8) &&
5622 "Expect an v8i16/v16i8 type");
5623 OpLHS = DAG.getNode(ARMISD::VREV64, DL, VT, OpLHS);
5624 // For a v16i8 type: After the VREV, we have got <8, ...15, 8, ..., 0>. Now,
5625 // extract the first 8 bytes into the top double word and the last 8 bytes
5626 // into the bottom double word. The v8i16 case is similar.
5627 unsigned ExtractNum = (VT == MVT::v16i8) ? 8 : 4;
5628 return DAG.getNode(ARMISD::VEXT, DL, VT, OpLHS, OpLHS,
5629 DAG.getConstant(ExtractNum, DL, MVT::i32));
5632 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
5633 SDValue V1 = Op.getOperand(0);
5634 SDValue V2 = Op.getOperand(1);
5636 EVT VT = Op.getValueType();
5637 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
5639 // Convert shuffles that are directly supported on NEON to target-specific
5640 // DAG nodes, instead of keeping them as shuffles and matching them again
5641 // during code selection. This is more efficient and avoids the possibility
5642 // of inconsistencies between legalization and selection.
5643 // FIXME: floating-point vectors should be canonicalized to integer vectors
5644 // of the same time so that they get CSEd properly.
5645 ArrayRef<int> ShuffleMask = SVN->getMask();
5647 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5648 if (EltSize <= 32) {
5649 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
5650 int Lane = SVN->getSplatIndex();
5651 // If this is undef splat, generate it via "just" vdup, if possible.
5652 if (Lane == -1) Lane = 0;
5654 // Test if V1 is a SCALAR_TO_VECTOR.
5655 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
5656 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
5658 // Test if V1 is a BUILD_VECTOR which is equivalent to a SCALAR_TO_VECTOR
5659 // (and probably will turn into a SCALAR_TO_VECTOR once legalization
5661 if (Lane == 0 && V1.getOpcode() == ISD::BUILD_VECTOR &&
5662 !isa<ConstantSDNode>(V1.getOperand(0))) {
5663 bool IsScalarToVector = true;
5664 for (unsigned i = 1, e = V1.getNumOperands(); i != e; ++i)
5665 if (V1.getOperand(i).getOpcode() != ISD::UNDEF) {
5666 IsScalarToVector = false;
5669 if (IsScalarToVector)
5670 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
5672 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
5673 DAG.getConstant(Lane, dl, MVT::i32));
5678 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
5681 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
5682 DAG.getConstant(Imm, dl, MVT::i32));
5685 if (isVREVMask(ShuffleMask, VT, 64))
5686 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
5687 if (isVREVMask(ShuffleMask, VT, 32))
5688 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
5689 if (isVREVMask(ShuffleMask, VT, 16))
5690 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
5692 if (V2->getOpcode() == ISD::UNDEF &&
5693 isSingletonVEXTMask(ShuffleMask, VT, Imm)) {
5694 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V1,
5695 DAG.getConstant(Imm, dl, MVT::i32));
5698 // Check for Neon shuffles that modify both input vectors in place.
5699 // If both results are used, i.e., if there are two shuffles with the same
5700 // source operands and with masks corresponding to both results of one of
5701 // these operations, DAG memoization will ensure that a single node is
5702 // used for both shuffles.
5703 unsigned WhichResult;
5705 if (unsigned ShuffleOpc = isNEONTwoResultShuffleMask(
5706 ShuffleMask, VT, WhichResult, isV_UNDEF)) {
5709 return DAG.getNode(ShuffleOpc, dl, DAG.getVTList(VT, VT), V1, V2)
5710 .getValue(WhichResult);
5713 // Also check for these shuffles through CONCAT_VECTORS: we canonicalize
5714 // shuffles that produce a result larger than their operands with:
5715 // shuffle(concat(v1, undef), concat(v2, undef))
5717 // shuffle(concat(v1, v2), undef)
5718 // because we can access quad vectors (see PerformVECTOR_SHUFFLECombine).
5720 // This is useful in the general case, but there are special cases where
5721 // native shuffles produce larger results: the two-result ops.
5723 // Look through the concat when lowering them:
5724 // shuffle(concat(v1, v2), undef)
5726 // concat(VZIP(v1, v2):0, :1)
5728 if (V1->getOpcode() == ISD::CONCAT_VECTORS &&
5729 V2->getOpcode() == ISD::UNDEF) {
5730 SDValue SubV1 = V1->getOperand(0);
5731 SDValue SubV2 = V1->getOperand(1);
5732 EVT SubVT = SubV1.getValueType();
5734 // We expect these to have been canonicalized to -1.
5735 assert(std::all_of(ShuffleMask.begin(), ShuffleMask.end(), [&](int i) {
5736 return i < (int)VT.getVectorNumElements();
5737 }) && "Unexpected shuffle index into UNDEF operand!");
5739 if (unsigned ShuffleOpc = isNEONTwoResultShuffleMask(
5740 ShuffleMask, SubVT, WhichResult, isV_UNDEF)) {
5743 assert((WhichResult == 0) &&
5744 "In-place shuffle of concat can only have one result!");
5745 SDValue Res = DAG.getNode(ShuffleOpc, dl, DAG.getVTList(SubVT, SubVT),
5747 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, Res.getValue(0),
5753 // If the shuffle is not directly supported and it has 4 elements, use
5754 // the PerfectShuffle-generated table to synthesize it from other shuffles.
5755 unsigned NumElts = VT.getVectorNumElements();
5757 unsigned PFIndexes[4];
5758 for (unsigned i = 0; i != 4; ++i) {
5759 if (ShuffleMask[i] < 0)
5762 PFIndexes[i] = ShuffleMask[i];
5765 // Compute the index in the perfect shuffle table.
5766 unsigned PFTableIndex =
5767 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
5768 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
5769 unsigned Cost = (PFEntry >> 30);
5772 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
5775 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
5776 if (EltSize >= 32) {
5777 // Do the expansion with floating-point types, since that is what the VFP
5778 // registers are defined to use, and since i64 is not legal.
5779 EVT EltVT = EVT::getFloatingPointVT(EltSize);
5780 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
5781 V1 = DAG.getNode(ISD::BITCAST, dl, VecVT, V1);
5782 V2 = DAG.getNode(ISD::BITCAST, dl, VecVT, V2);
5783 SmallVector<SDValue, 8> Ops;
5784 for (unsigned i = 0; i < NumElts; ++i) {
5785 if (ShuffleMask[i] < 0)
5786 Ops.push_back(DAG.getUNDEF(EltVT));
5788 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
5789 ShuffleMask[i] < (int)NumElts ? V1 : V2,
5790 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
5793 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, Ops);
5794 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
5797 if ((VT == MVT::v8i16 || VT == MVT::v16i8) && isReverseMask(ShuffleMask, VT))
5798 return LowerReverse_VECTOR_SHUFFLEv16i8_v8i16(Op, DAG);
5800 if (VT == MVT::v8i8) {
5801 SDValue NewOp = LowerVECTOR_SHUFFLEv8i8(Op, ShuffleMask, DAG);
5802 if (NewOp.getNode())
5809 static SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
5810 // INSERT_VECTOR_ELT is legal only for immediate indexes.
5811 SDValue Lane = Op.getOperand(2);
5812 if (!isa<ConstantSDNode>(Lane))
5818 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
5819 // EXTRACT_VECTOR_ELT is legal only for immediate indexes.
5820 SDValue Lane = Op.getOperand(1);
5821 if (!isa<ConstantSDNode>(Lane))
5824 SDValue Vec = Op.getOperand(0);
5825 if (Op.getValueType() == MVT::i32 &&
5826 Vec.getValueType().getVectorElementType().getSizeInBits() < 32) {
5828 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
5834 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
5835 // The only time a CONCAT_VECTORS operation can have legal types is when
5836 // two 64-bit vectors are concatenated to a 128-bit vector.
5837 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
5838 "unexpected CONCAT_VECTORS");
5840 SDValue Val = DAG.getUNDEF(MVT::v2f64);
5841 SDValue Op0 = Op.getOperand(0);
5842 SDValue Op1 = Op.getOperand(1);
5843 if (Op0.getOpcode() != ISD::UNDEF)
5844 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
5845 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op0),
5846 DAG.getIntPtrConstant(0, dl));
5847 if (Op1.getOpcode() != ISD::UNDEF)
5848 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
5849 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op1),
5850 DAG.getIntPtrConstant(1, dl));
5851 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Val);
5854 /// isExtendedBUILD_VECTOR - Check if N is a constant BUILD_VECTOR where each
5855 /// element has been zero/sign-extended, depending on the isSigned parameter,
5856 /// from an integer type half its size.
5857 static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG,
5859 // A v2i64 BUILD_VECTOR will have been legalized to a BITCAST from v4i32.
5860 EVT VT = N->getValueType(0);
5861 if (VT == MVT::v2i64 && N->getOpcode() == ISD::BITCAST) {
5862 SDNode *BVN = N->getOperand(0).getNode();
5863 if (BVN->getValueType(0) != MVT::v4i32 ||
5864 BVN->getOpcode() != ISD::BUILD_VECTOR)
5866 unsigned LoElt = DAG.getDataLayout().isBigEndian() ? 1 : 0;
5867 unsigned HiElt = 1 - LoElt;
5868 ConstantSDNode *Lo0 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt));
5869 ConstantSDNode *Hi0 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt));
5870 ConstantSDNode *Lo1 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt+2));
5871 ConstantSDNode *Hi1 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt+2));
5872 if (!Lo0 || !Hi0 || !Lo1 || !Hi1)
5875 if (Hi0->getSExtValue() == Lo0->getSExtValue() >> 32 &&
5876 Hi1->getSExtValue() == Lo1->getSExtValue() >> 32)
5879 if (Hi0->isNullValue() && Hi1->isNullValue())
5885 if (N->getOpcode() != ISD::BUILD_VECTOR)
5888 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
5889 SDNode *Elt = N->getOperand(i).getNode();
5890 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
5891 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5892 unsigned HalfSize = EltSize / 2;
5894 if (!isIntN(HalfSize, C->getSExtValue()))
5897 if (!isUIntN(HalfSize, C->getZExtValue()))
5908 /// isSignExtended - Check if a node is a vector value that is sign-extended
5909 /// or a constant BUILD_VECTOR with sign-extended elements.
5910 static bool isSignExtended(SDNode *N, SelectionDAG &DAG) {
5911 if (N->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N))
5913 if (isExtendedBUILD_VECTOR(N, DAG, true))
5918 /// isZeroExtended - Check if a node is a vector value that is zero-extended
5919 /// or a constant BUILD_VECTOR with zero-extended elements.
5920 static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) {
5921 if (N->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N))
5923 if (isExtendedBUILD_VECTOR(N, DAG, false))
5928 static EVT getExtensionTo64Bits(const EVT &OrigVT) {
5929 if (OrigVT.getSizeInBits() >= 64)
5932 assert(OrigVT.isSimple() && "Expecting a simple value type");
5934 MVT::SimpleValueType OrigSimpleTy = OrigVT.getSimpleVT().SimpleTy;
5935 switch (OrigSimpleTy) {
5936 default: llvm_unreachable("Unexpected Vector Type");
5945 /// AddRequiredExtensionForVMULL - Add a sign/zero extension to extend the total
5946 /// value size to 64 bits. We need a 64-bit D register as an operand to VMULL.
5947 /// We insert the required extension here to get the vector to fill a D register.
5948 static SDValue AddRequiredExtensionForVMULL(SDValue N, SelectionDAG &DAG,
5951 unsigned ExtOpcode) {
5952 // The vector originally had a size of OrigTy. It was then extended to ExtTy.
5953 // We expect the ExtTy to be 128-bits total. If the OrigTy is less than
5954 // 64-bits we need to insert a new extension so that it will be 64-bits.
5955 assert(ExtTy.is128BitVector() && "Unexpected extension size");
5956 if (OrigTy.getSizeInBits() >= 64)
5959 // Must extend size to at least 64 bits to be used as an operand for VMULL.
5960 EVT NewVT = getExtensionTo64Bits(OrigTy);
5962 return DAG.getNode(ExtOpcode, SDLoc(N), NewVT, N);
5965 /// SkipLoadExtensionForVMULL - return a load of the original vector size that
5966 /// does not do any sign/zero extension. If the original vector is less
5967 /// than 64 bits, an appropriate extension will be added after the load to
5968 /// reach a total size of 64 bits. We have to add the extension separately
5969 /// because ARM does not have a sign/zero extending load for vectors.
5970 static SDValue SkipLoadExtensionForVMULL(LoadSDNode *LD, SelectionDAG& DAG) {
5971 EVT ExtendedTy = getExtensionTo64Bits(LD->getMemoryVT());
5973 // The load already has the right type.
5974 if (ExtendedTy == LD->getMemoryVT())
5975 return DAG.getLoad(LD->getMemoryVT(), SDLoc(LD), LD->getChain(),
5976 LD->getBasePtr(), LD->getPointerInfo(), LD->isVolatile(),
5977 LD->isNonTemporal(), LD->isInvariant(),
5978 LD->getAlignment());
5980 // We need to create a zextload/sextload. We cannot just create a load
5981 // followed by a zext/zext node because LowerMUL is also run during normal
5982 // operation legalization where we can't create illegal types.
5983 return DAG.getExtLoad(LD->getExtensionType(), SDLoc(LD), ExtendedTy,
5984 LD->getChain(), LD->getBasePtr(), LD->getPointerInfo(),
5985 LD->getMemoryVT(), LD->isVolatile(), LD->isInvariant(),
5986 LD->isNonTemporal(), LD->getAlignment());
5989 /// SkipExtensionForVMULL - For a node that is a SIGN_EXTEND, ZERO_EXTEND,
5990 /// extending load, or BUILD_VECTOR with extended elements, return the
5991 /// unextended value. The unextended vector should be 64 bits so that it can
5992 /// be used as an operand to a VMULL instruction. If the original vector size
5993 /// before extension is less than 64 bits we add a an extension to resize
5994 /// the vector to 64 bits.
5995 static SDValue SkipExtensionForVMULL(SDNode *N, SelectionDAG &DAG) {
5996 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
5997 return AddRequiredExtensionForVMULL(N->getOperand(0), DAG,
5998 N->getOperand(0)->getValueType(0),
6002 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
6003 return SkipLoadExtensionForVMULL(LD, DAG);
6005 // Otherwise, the value must be a BUILD_VECTOR. For v2i64, it will
6006 // have been legalized as a BITCAST from v4i32.
6007 if (N->getOpcode() == ISD::BITCAST) {
6008 SDNode *BVN = N->getOperand(0).getNode();
6009 assert(BVN->getOpcode() == ISD::BUILD_VECTOR &&
6010 BVN->getValueType(0) == MVT::v4i32 && "expected v4i32 BUILD_VECTOR");
6011 unsigned LowElt = DAG.getDataLayout().isBigEndian() ? 1 : 0;
6012 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), MVT::v2i32,
6013 BVN->getOperand(LowElt), BVN->getOperand(LowElt+2));
6015 // Construct a new BUILD_VECTOR with elements truncated to half the size.
6016 assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR");
6017 EVT VT = N->getValueType(0);
6018 unsigned EltSize = VT.getVectorElementType().getSizeInBits() / 2;
6019 unsigned NumElts = VT.getVectorNumElements();
6020 MVT TruncVT = MVT::getIntegerVT(EltSize);
6021 SmallVector<SDValue, 8> Ops;
6023 for (unsigned i = 0; i != NumElts; ++i) {
6024 ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i));
6025 const APInt &CInt = C->getAPIntValue();
6026 // Element types smaller than 32 bits are not legal, so use i32 elements.
6027 // The values are implicitly truncated so sext vs. zext doesn't matter.
6028 Ops.push_back(DAG.getConstant(CInt.zextOrTrunc(32), dl, MVT::i32));
6030 return DAG.getNode(ISD::BUILD_VECTOR, dl,
6031 MVT::getVectorVT(TruncVT, NumElts), Ops);
6034 static bool isAddSubSExt(SDNode *N, SelectionDAG &DAG) {
6035 unsigned Opcode = N->getOpcode();
6036 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
6037 SDNode *N0 = N->getOperand(0).getNode();
6038 SDNode *N1 = N->getOperand(1).getNode();
6039 return N0->hasOneUse() && N1->hasOneUse() &&
6040 isSignExtended(N0, DAG) && isSignExtended(N1, DAG);
6045 static bool isAddSubZExt(SDNode *N, SelectionDAG &DAG) {
6046 unsigned Opcode = N->getOpcode();
6047 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
6048 SDNode *N0 = N->getOperand(0).getNode();
6049 SDNode *N1 = N->getOperand(1).getNode();
6050 return N0->hasOneUse() && N1->hasOneUse() &&
6051 isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG);
6056 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
6057 // Multiplications are only custom-lowered for 128-bit vectors so that
6058 // VMULL can be detected. Otherwise v2i64 multiplications are not legal.
6059 EVT VT = Op.getValueType();
6060 assert(VT.is128BitVector() && VT.isInteger() &&
6061 "unexpected type for custom-lowering ISD::MUL");
6062 SDNode *N0 = Op.getOperand(0).getNode();
6063 SDNode *N1 = Op.getOperand(1).getNode();
6064 unsigned NewOpc = 0;
6066 bool isN0SExt = isSignExtended(N0, DAG);
6067 bool isN1SExt = isSignExtended(N1, DAG);
6068 if (isN0SExt && isN1SExt)
6069 NewOpc = ARMISD::VMULLs;
6071 bool isN0ZExt = isZeroExtended(N0, DAG);
6072 bool isN1ZExt = isZeroExtended(N1, DAG);
6073 if (isN0ZExt && isN1ZExt)
6074 NewOpc = ARMISD::VMULLu;
6075 else if (isN1SExt || isN1ZExt) {
6076 // Look for (s/zext A + s/zext B) * (s/zext C). We want to turn these
6077 // into (s/zext A * s/zext C) + (s/zext B * s/zext C)
6078 if (isN1SExt && isAddSubSExt(N0, DAG)) {
6079 NewOpc = ARMISD::VMULLs;
6081 } else if (isN1ZExt && isAddSubZExt(N0, DAG)) {
6082 NewOpc = ARMISD::VMULLu;
6084 } else if (isN0ZExt && isAddSubZExt(N1, DAG)) {
6086 NewOpc = ARMISD::VMULLu;
6092 if (VT == MVT::v2i64)
6093 // Fall through to expand this. It is not legal.
6096 // Other vector multiplications are legal.
6101 // Legalize to a VMULL instruction.
6104 SDValue Op1 = SkipExtensionForVMULL(N1, DAG);
6106 Op0 = SkipExtensionForVMULL(N0, DAG);
6107 assert(Op0.getValueType().is64BitVector() &&
6108 Op1.getValueType().is64BitVector() &&
6109 "unexpected types for extended operands to VMULL");
6110 return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
6113 // Optimizing (zext A + zext B) * C, to (VMULL A, C) + (VMULL B, C) during
6114 // isel lowering to take advantage of no-stall back to back vmul + vmla.
6121 SDValue N00 = SkipExtensionForVMULL(N0->getOperand(0).getNode(), DAG);
6122 SDValue N01 = SkipExtensionForVMULL(N0->getOperand(1).getNode(), DAG);
6123 EVT Op1VT = Op1.getValueType();
6124 return DAG.getNode(N0->getOpcode(), DL, VT,
6125 DAG.getNode(NewOpc, DL, VT,
6126 DAG.getNode(ISD::BITCAST, DL, Op1VT, N00), Op1),
6127 DAG.getNode(NewOpc, DL, VT,
6128 DAG.getNode(ISD::BITCAST, DL, Op1VT, N01), Op1));
6132 LowerSDIV_v4i8(SDValue X, SDValue Y, SDLoc dl, SelectionDAG &DAG) {
6134 // float4 xf = vcvt_f32_s32(vmovl_s16(a.lo));
6135 // float4 yf = vcvt_f32_s32(vmovl_s16(b.lo));
6136 X = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, X);
6137 Y = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, Y);
6138 X = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, X);
6139 Y = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, Y);
6140 // Get reciprocal estimate.
6141 // float4 recip = vrecpeq_f32(yf);
6142 Y = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6143 DAG.getConstant(Intrinsic::arm_neon_vrecpe, dl, MVT::i32),
6145 // Because char has a smaller range than uchar, we can actually get away
6146 // without any newton steps. This requires that we use a weird bias
6147 // of 0xb000, however (again, this has been exhaustively tested).
6148 // float4 result = as_float4(as_int4(xf*recip) + 0xb000);
6149 X = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, X, Y);
6150 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, X);
6151 Y = DAG.getConstant(0xb000, dl, MVT::i32);
6152 Y = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Y, Y, Y, Y);
6153 X = DAG.getNode(ISD::ADD, dl, MVT::v4i32, X, Y);
6154 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, X);
6155 // Convert back to short.
6156 X = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, X);
6157 X = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, X);
6162 LowerSDIV_v4i16(SDValue N0, SDValue N1, SDLoc dl, SelectionDAG &DAG) {
6164 // Convert to float.
6165 // float4 yf = vcvt_f32_s32(vmovl_s16(y));
6166 // float4 xf = vcvt_f32_s32(vmovl_s16(x));
6167 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N0);
6168 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N1);
6169 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
6170 N1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
6172 // Use reciprocal estimate and one refinement step.
6173 // float4 recip = vrecpeq_f32(yf);
6174 // recip *= vrecpsq_f32(yf, recip);
6175 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6176 DAG.getConstant(Intrinsic::arm_neon_vrecpe, dl, MVT::i32),
6178 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6179 DAG.getConstant(Intrinsic::arm_neon_vrecps, dl, MVT::i32),
6181 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
6182 // Because short has a smaller range than ushort, we can actually get away
6183 // with only a single newton step. This requires that we use a weird bias
6184 // of 89, however (again, this has been exhaustively tested).
6185 // float4 result = as_float4(as_int4(xf*recip) + 0x89);
6186 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
6187 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
6188 N1 = DAG.getConstant(0x89, dl, MVT::i32);
6189 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
6190 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
6191 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
6192 // Convert back to integer and return.
6193 // return vmovn_s32(vcvt_s32_f32(result));
6194 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
6195 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
6199 static SDValue LowerSDIV(SDValue Op, SelectionDAG &DAG) {
6200 EVT VT = Op.getValueType();
6201 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
6202 "unexpected type for custom-lowering ISD::SDIV");
6205 SDValue N0 = Op.getOperand(0);
6206 SDValue N1 = Op.getOperand(1);
6209 if (VT == MVT::v8i8) {
6210 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N0);
6211 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N1);
6213 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
6214 DAG.getIntPtrConstant(4, dl));
6215 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
6216 DAG.getIntPtrConstant(4, dl));
6217 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
6218 DAG.getIntPtrConstant(0, dl));
6219 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
6220 DAG.getIntPtrConstant(0, dl));
6222 N0 = LowerSDIV_v4i8(N0, N1, dl, DAG); // v4i16
6223 N2 = LowerSDIV_v4i8(N2, N3, dl, DAG); // v4i16
6225 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
6226 N0 = LowerCONCAT_VECTORS(N0, DAG);
6228 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v8i8, N0);
6231 return LowerSDIV_v4i16(N0, N1, dl, DAG);
6234 static SDValue LowerUDIV(SDValue Op, SelectionDAG &DAG) {
6235 EVT VT = Op.getValueType();
6236 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
6237 "unexpected type for custom-lowering ISD::UDIV");
6240 SDValue N0 = Op.getOperand(0);
6241 SDValue N1 = Op.getOperand(1);
6244 if (VT == MVT::v8i8) {
6245 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N0);
6246 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N1);
6248 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
6249 DAG.getIntPtrConstant(4, dl));
6250 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
6251 DAG.getIntPtrConstant(4, dl));
6252 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
6253 DAG.getIntPtrConstant(0, dl));
6254 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
6255 DAG.getIntPtrConstant(0, dl));
6257 N0 = LowerSDIV_v4i16(N0, N1, dl, DAG); // v4i16
6258 N2 = LowerSDIV_v4i16(N2, N3, dl, DAG); // v4i16
6260 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
6261 N0 = LowerCONCAT_VECTORS(N0, DAG);
6263 N0 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v8i8,
6264 DAG.getConstant(Intrinsic::arm_neon_vqmovnsu, dl,
6270 // v4i16 sdiv ... Convert to float.
6271 // float4 yf = vcvt_f32_s32(vmovl_u16(y));
6272 // float4 xf = vcvt_f32_s32(vmovl_u16(x));
6273 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N0);
6274 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N1);
6275 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
6276 SDValue BN1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
6278 // Use reciprocal estimate and two refinement steps.
6279 // float4 recip = vrecpeq_f32(yf);
6280 // recip *= vrecpsq_f32(yf, recip);
6281 // recip *= vrecpsq_f32(yf, recip);
6282 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6283 DAG.getConstant(Intrinsic::arm_neon_vrecpe, dl, MVT::i32),
6285 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6286 DAG.getConstant(Intrinsic::arm_neon_vrecps, dl, MVT::i32),
6288 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
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 // Simply multiplying by the reciprocal estimate can leave us a few ulps
6294 // too low, so we add 2 ulps (exhaustive testing shows that this is enough,
6295 // and that it will never cause us to return an answer too large).
6296 // float4 result = as_float4(as_int4(xf*recip) + 2);
6297 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
6298 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
6299 N1 = DAG.getConstant(2, dl, MVT::i32);
6300 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
6301 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
6302 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
6303 // Convert back to integer and return.
6304 // return vmovn_u32(vcvt_s32_f32(result));
6305 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
6306 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
6310 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
6311 EVT VT = Op.getNode()->getValueType(0);
6312 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
6315 bool ExtraOp = false;
6316 switch (Op.getOpcode()) {
6317 default: llvm_unreachable("Invalid code");
6318 case ISD::ADDC: Opc = ARMISD::ADDC; break;
6319 case ISD::ADDE: Opc = ARMISD::ADDE; ExtraOp = true; break;
6320 case ISD::SUBC: Opc = ARMISD::SUBC; break;
6321 case ISD::SUBE: Opc = ARMISD::SUBE; ExtraOp = true; break;
6325 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
6327 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
6328 Op.getOperand(1), Op.getOperand(2));
6331 SDValue ARMTargetLowering::LowerFSINCOS(SDValue Op, SelectionDAG &DAG) const {
6332 assert(Subtarget->isTargetDarwin());
6334 // For iOS, we want to call an alternative entry point: __sincos_stret,
6335 // return values are passed via sret.
6337 SDValue Arg = Op.getOperand(0);
6338 EVT ArgVT = Arg.getValueType();
6339 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
6341 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
6342 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6344 // Pair of floats / doubles used to pass the result.
6345 StructType *RetTy = StructType::get(ArgTy, ArgTy, nullptr);
6347 // Create stack object for sret.
6348 const uint64_t ByteSize = TLI.getDataLayout()->getTypeAllocSize(RetTy);
6349 const unsigned StackAlign = TLI.getDataLayout()->getPrefTypeAlignment(RetTy);
6350 int FrameIdx = FrameInfo->CreateStackObject(ByteSize, StackAlign, false);
6351 SDValue SRet = DAG.getFrameIndex(FrameIdx, TLI.getPointerTy());
6357 Entry.Ty = RetTy->getPointerTo();
6358 Entry.isSExt = false;
6359 Entry.isZExt = false;
6360 Entry.isSRet = true;
6361 Args.push_back(Entry);
6365 Entry.isSExt = false;
6366 Entry.isZExt = false;
6367 Args.push_back(Entry);
6369 const char *LibcallName = (ArgVT == MVT::f64)
6370 ? "__sincos_stret" : "__sincosf_stret";
6371 SDValue Callee = DAG.getExternalSymbol(LibcallName, getPointerTy());
6373 TargetLowering::CallLoweringInfo CLI(DAG);
6374 CLI.setDebugLoc(dl).setChain(DAG.getEntryNode())
6375 .setCallee(CallingConv::C, Type::getVoidTy(*DAG.getContext()), Callee,
6377 .setDiscardResult();
6379 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
6381 SDValue LoadSin = DAG.getLoad(ArgVT, dl, CallResult.second, SRet,
6382 MachinePointerInfo(), false, false, false, 0);
6384 // Address of cos field.
6385 SDValue Add = DAG.getNode(ISD::ADD, dl, getPointerTy(), SRet,
6386 DAG.getIntPtrConstant(ArgVT.getStoreSize(), dl));
6387 SDValue LoadCos = DAG.getLoad(ArgVT, dl, LoadSin.getValue(1), Add,
6388 MachinePointerInfo(), false, false, false, 0);
6390 SDVTList Tys = DAG.getVTList(ArgVT, ArgVT);
6391 return DAG.getNode(ISD::MERGE_VALUES, dl, Tys,
6392 LoadSin.getValue(0), LoadCos.getValue(0));
6395 static SDValue LowerAtomicLoadStore(SDValue Op, SelectionDAG &DAG) {
6396 // Monotonic load/store is legal for all targets
6397 if (cast<AtomicSDNode>(Op)->getOrdering() <= Monotonic)
6400 // Acquire/Release load/store is not legal for targets without a
6401 // dmb or equivalent available.
6405 static void ReplaceREADCYCLECOUNTER(SDNode *N,
6406 SmallVectorImpl<SDValue> &Results,
6408 const ARMSubtarget *Subtarget) {
6410 SDValue Cycles32, OutChain;
6412 if (Subtarget->hasPerfMon()) {
6413 // Under Power Management extensions, the cycle-count is:
6414 // mrc p15, #0, <Rt>, c9, c13, #0
6415 SDValue Ops[] = { N->getOperand(0), // Chain
6416 DAG.getConstant(Intrinsic::arm_mrc, DL, MVT::i32),
6417 DAG.getConstant(15, DL, MVT::i32),
6418 DAG.getConstant(0, DL, MVT::i32),
6419 DAG.getConstant(9, DL, MVT::i32),
6420 DAG.getConstant(13, DL, MVT::i32),
6421 DAG.getConstant(0, DL, MVT::i32)
6424 Cycles32 = DAG.getNode(ISD::INTRINSIC_W_CHAIN, DL,
6425 DAG.getVTList(MVT::i32, MVT::Other), Ops);
6426 OutChain = Cycles32.getValue(1);
6428 // Intrinsic is defined to return 0 on unsupported platforms. Technically
6429 // there are older ARM CPUs that have implementation-specific ways of
6430 // obtaining this information (FIXME!).
6431 Cycles32 = DAG.getConstant(0, DL, MVT::i32);
6432 OutChain = DAG.getEntryNode();
6436 SDValue Cycles64 = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64,
6437 Cycles32, DAG.getConstant(0, DL, MVT::i32));
6438 Results.push_back(Cycles64);
6439 Results.push_back(OutChain);
6442 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
6443 switch (Op.getOpcode()) {
6444 default: llvm_unreachable("Don't know how to custom lower this!");
6445 case ISD::WRITE_REGISTER: return LowerWRITE_REGISTER(Op, DAG);
6446 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
6447 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
6448 case ISD::GlobalAddress:
6449 switch (Subtarget->getTargetTriple().getObjectFormat()) {
6450 default: llvm_unreachable("unknown object format");
6452 return LowerGlobalAddressWindows(Op, DAG);
6454 return LowerGlobalAddressELF(Op, DAG);
6456 return LowerGlobalAddressDarwin(Op, DAG);
6458 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
6459 case ISD::SELECT: return LowerSELECT(Op, DAG);
6460 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
6461 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
6462 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
6463 case ISD::VASTART: return LowerVASTART(Op, DAG);
6464 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, DAG, Subtarget);
6465 case ISD::PREFETCH: return LowerPREFETCH(Op, DAG, Subtarget);
6466 case ISD::SINT_TO_FP:
6467 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
6468 case ISD::FP_TO_SINT:
6469 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
6470 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
6471 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
6472 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
6473 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
6474 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
6475 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
6476 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
6478 case ISD::BITCAST: return ExpandBITCAST(Op.getNode(), DAG);
6481 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
6482 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
6483 case ISD::SRL_PARTS:
6484 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
6485 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
6486 case ISD::CTPOP: return LowerCTPOP(Op.getNode(), DAG, Subtarget);
6487 case ISD::SETCC: return LowerVSETCC(Op, DAG);
6488 case ISD::ConstantFP: return LowerConstantFP(Op, DAG, Subtarget);
6489 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG, Subtarget);
6490 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
6491 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
6492 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
6493 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
6494 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
6495 case ISD::MUL: return LowerMUL(Op, DAG);
6496 case ISD::SDIV: return LowerSDIV(Op, DAG);
6497 case ISD::UDIV: return LowerUDIV(Op, DAG);
6501 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
6506 return LowerXALUO(Op, DAG);
6507 case ISD::ATOMIC_LOAD:
6508 case ISD::ATOMIC_STORE: return LowerAtomicLoadStore(Op, DAG);
6509 case ISD::FSINCOS: return LowerFSINCOS(Op, DAG);
6511 case ISD::UDIVREM: return LowerDivRem(Op, DAG);
6512 case ISD::DYNAMIC_STACKALLOC:
6513 if (Subtarget->getTargetTriple().isWindowsItaniumEnvironment())
6514 return LowerDYNAMIC_STACKALLOC(Op, DAG);
6515 llvm_unreachable("Don't know how to custom lower this!");
6516 case ISD::FP_ROUND: return LowerFP_ROUND(Op, DAG);
6517 case ISD::FP_EXTEND: return LowerFP_EXTEND(Op, DAG);
6521 /// ReplaceNodeResults - Replace the results of node with an illegal result
6522 /// type with new values built out of custom code.
6523 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
6524 SmallVectorImpl<SDValue>&Results,
6525 SelectionDAG &DAG) const {
6527 switch (N->getOpcode()) {
6529 llvm_unreachable("Don't know how to custom expand this!");
6530 case ISD::READ_REGISTER:
6531 ExpandREAD_REGISTER(N, Results, DAG);
6534 Res = ExpandBITCAST(N, DAG);
6538 Res = Expand64BitShift(N, DAG, Subtarget);
6540 case ISD::READCYCLECOUNTER:
6541 ReplaceREADCYCLECOUNTER(N, Results, DAG, Subtarget);
6545 Results.push_back(Res);
6548 //===----------------------------------------------------------------------===//
6549 // ARM Scheduler Hooks
6550 //===----------------------------------------------------------------------===//
6552 /// SetupEntryBlockForSjLj - Insert code into the entry block that creates and
6553 /// registers the function context.
6554 void ARMTargetLowering::
6555 SetupEntryBlockForSjLj(MachineInstr *MI, MachineBasicBlock *MBB,
6556 MachineBasicBlock *DispatchBB, int FI) const {
6557 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
6558 DebugLoc dl = MI->getDebugLoc();
6559 MachineFunction *MF = MBB->getParent();
6560 MachineRegisterInfo *MRI = &MF->getRegInfo();
6561 MachineConstantPool *MCP = MF->getConstantPool();
6562 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
6563 const Function *F = MF->getFunction();
6565 bool isThumb = Subtarget->isThumb();
6566 bool isThumb2 = Subtarget->isThumb2();
6568 unsigned PCLabelId = AFI->createPICLabelUId();
6569 unsigned PCAdj = (isThumb || isThumb2) ? 4 : 8;
6570 ARMConstantPoolValue *CPV =
6571 ARMConstantPoolMBB::Create(F->getContext(), DispatchBB, PCLabelId, PCAdj);
6572 unsigned CPI = MCP->getConstantPoolIndex(CPV, 4);
6574 const TargetRegisterClass *TRC = isThumb ? &ARM::tGPRRegClass
6575 : &ARM::GPRRegClass;
6577 // Grab constant pool and fixed stack memory operands.
6578 MachineMemOperand *CPMMO =
6579 MF->getMachineMemOperand(MachinePointerInfo::getConstantPool(),
6580 MachineMemOperand::MOLoad, 4, 4);
6582 MachineMemOperand *FIMMOSt =
6583 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
6584 MachineMemOperand::MOStore, 4, 4);
6586 // Load the address of the dispatch MBB into the jump buffer.
6588 // Incoming value: jbuf
6589 // ldr.n r5, LCPI1_1
6592 // str r5, [$jbuf, #+4] ; &jbuf[1]
6593 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6594 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2LDRpci), NewVReg1)
6595 .addConstantPoolIndex(CPI)
6596 .addMemOperand(CPMMO));
6597 // Set the low bit because of thumb mode.
6598 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6600 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2ORRri), NewVReg2)
6601 .addReg(NewVReg1, RegState::Kill)
6603 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6604 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg3)
6605 .addReg(NewVReg2, RegState::Kill)
6607 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2STRi12))
6608 .addReg(NewVReg3, RegState::Kill)
6610 .addImm(36) // &jbuf[1] :: pc
6611 .addMemOperand(FIMMOSt));
6612 } else if (isThumb) {
6613 // Incoming value: jbuf
6614 // ldr.n r1, LCPI1_4
6618 // add r2, $jbuf, #+4 ; &jbuf[1]
6620 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6621 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tLDRpci), NewVReg1)
6622 .addConstantPoolIndex(CPI)
6623 .addMemOperand(CPMMO));
6624 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6625 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg2)
6626 .addReg(NewVReg1, RegState::Kill)
6628 // Set the low bit because of thumb mode.
6629 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6630 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tMOVi8), NewVReg3)
6631 .addReg(ARM::CPSR, RegState::Define)
6633 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6634 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tORR), NewVReg4)
6635 .addReg(ARM::CPSR, RegState::Define)
6636 .addReg(NewVReg2, RegState::Kill)
6637 .addReg(NewVReg3, RegState::Kill));
6638 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6639 BuildMI(*MBB, MI, dl, TII->get(ARM::tADDframe), NewVReg5)
6641 .addImm(36); // &jbuf[1] :: pc
6642 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tSTRi))
6643 .addReg(NewVReg4, RegState::Kill)
6644 .addReg(NewVReg5, RegState::Kill)
6646 .addMemOperand(FIMMOSt));
6648 // Incoming value: jbuf
6651 // str r1, [$jbuf, #+4] ; &jbuf[1]
6652 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6653 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::LDRi12), NewVReg1)
6654 .addConstantPoolIndex(CPI)
6656 .addMemOperand(CPMMO));
6657 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6658 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::PICADD), NewVReg2)
6659 .addReg(NewVReg1, RegState::Kill)
6660 .addImm(PCLabelId));
6661 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::STRi12))
6662 .addReg(NewVReg2, RegState::Kill)
6664 .addImm(36) // &jbuf[1] :: pc
6665 .addMemOperand(FIMMOSt));
6669 void ARMTargetLowering::EmitSjLjDispatchBlock(MachineInstr *MI,
6670 MachineBasicBlock *MBB) const {
6671 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
6672 DebugLoc dl = MI->getDebugLoc();
6673 MachineFunction *MF = MBB->getParent();
6674 MachineRegisterInfo *MRI = &MF->getRegInfo();
6675 MachineFrameInfo *MFI = MF->getFrameInfo();
6676 int FI = MFI->getFunctionContextIndex();
6678 const TargetRegisterClass *TRC = Subtarget->isThumb() ? &ARM::tGPRRegClass
6679 : &ARM::GPRnopcRegClass;
6681 // Get a mapping of the call site numbers to all of the landing pads they're
6683 DenseMap<unsigned, SmallVector<MachineBasicBlock*, 2> > CallSiteNumToLPad;
6684 unsigned MaxCSNum = 0;
6685 MachineModuleInfo &MMI = MF->getMMI();
6686 for (MachineFunction::iterator BB = MF->begin(), E = MF->end(); BB != E;
6688 if (!BB->isLandingPad()) continue;
6690 // FIXME: We should assert that the EH_LABEL is the first MI in the landing
6692 for (MachineBasicBlock::iterator
6693 II = BB->begin(), IE = BB->end(); II != IE; ++II) {
6694 if (!II->isEHLabel()) continue;
6696 MCSymbol *Sym = II->getOperand(0).getMCSymbol();
6697 if (!MMI.hasCallSiteLandingPad(Sym)) continue;
6699 SmallVectorImpl<unsigned> &CallSiteIdxs = MMI.getCallSiteLandingPad(Sym);
6700 for (SmallVectorImpl<unsigned>::iterator
6701 CSI = CallSiteIdxs.begin(), CSE = CallSiteIdxs.end();
6702 CSI != CSE; ++CSI) {
6703 CallSiteNumToLPad[*CSI].push_back(BB);
6704 MaxCSNum = std::max(MaxCSNum, *CSI);
6710 // Get an ordered list of the machine basic blocks for the jump table.
6711 std::vector<MachineBasicBlock*> LPadList;
6712 SmallPtrSet<MachineBasicBlock*, 64> InvokeBBs;
6713 LPadList.reserve(CallSiteNumToLPad.size());
6714 for (unsigned I = 1; I <= MaxCSNum; ++I) {
6715 SmallVectorImpl<MachineBasicBlock*> &MBBList = CallSiteNumToLPad[I];
6716 for (SmallVectorImpl<MachineBasicBlock*>::iterator
6717 II = MBBList.begin(), IE = MBBList.end(); II != IE; ++II) {
6718 LPadList.push_back(*II);
6719 InvokeBBs.insert((*II)->pred_begin(), (*II)->pred_end());
6723 assert(!LPadList.empty() &&
6724 "No landing pad destinations for the dispatch jump table!");
6726 // Create the jump table and associated information.
6727 MachineJumpTableInfo *JTI =
6728 MF->getOrCreateJumpTableInfo(MachineJumpTableInfo::EK_Inline);
6729 unsigned MJTI = JTI->createJumpTableIndex(LPadList);
6730 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
6732 // Create the MBBs for the dispatch code.
6734 // Shove the dispatch's address into the return slot in the function context.
6735 MachineBasicBlock *DispatchBB = MF->CreateMachineBasicBlock();
6736 DispatchBB->setIsLandingPad();
6738 MachineBasicBlock *TrapBB = MF->CreateMachineBasicBlock();
6739 unsigned trap_opcode;
6740 if (Subtarget->isThumb())
6741 trap_opcode = ARM::tTRAP;
6743 trap_opcode = Subtarget->useNaClTrap() ? ARM::TRAPNaCl : ARM::TRAP;
6745 BuildMI(TrapBB, dl, TII->get(trap_opcode));
6746 DispatchBB->addSuccessor(TrapBB);
6748 MachineBasicBlock *DispContBB = MF->CreateMachineBasicBlock();
6749 DispatchBB->addSuccessor(DispContBB);
6752 MF->insert(MF->end(), DispatchBB);
6753 MF->insert(MF->end(), DispContBB);
6754 MF->insert(MF->end(), TrapBB);
6756 // Insert code into the entry block that creates and registers the function
6758 SetupEntryBlockForSjLj(MI, MBB, DispatchBB, FI);
6760 MachineMemOperand *FIMMOLd =
6761 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
6762 MachineMemOperand::MOLoad |
6763 MachineMemOperand::MOVolatile, 4, 4);
6765 MachineInstrBuilder MIB;
6766 MIB = BuildMI(DispatchBB, dl, TII->get(ARM::Int_eh_sjlj_dispatchsetup));
6768 const ARMBaseInstrInfo *AII = static_cast<const ARMBaseInstrInfo*>(TII);
6769 const ARMBaseRegisterInfo &RI = AII->getRegisterInfo();
6771 // Add a register mask with no preserved registers. This results in all
6772 // registers being marked as clobbered.
6773 MIB.addRegMask(RI.getNoPreservedMask());
6775 unsigned NumLPads = LPadList.size();
6776 if (Subtarget->isThumb2()) {
6777 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6778 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2LDRi12), NewVReg1)
6781 .addMemOperand(FIMMOLd));
6783 if (NumLPads < 256) {
6784 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPri))
6786 .addImm(LPadList.size()));
6788 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6789 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVi16), VReg1)
6790 .addImm(NumLPads & 0xFFFF));
6792 unsigned VReg2 = VReg1;
6793 if ((NumLPads & 0xFFFF0000) != 0) {
6794 VReg2 = MRI->createVirtualRegister(TRC);
6795 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVTi16), VReg2)
6797 .addImm(NumLPads >> 16));
6800 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPrr))
6805 BuildMI(DispatchBB, dl, TII->get(ARM::t2Bcc))
6810 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6811 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::t2LEApcrelJT),NewVReg3)
6812 .addJumpTableIndex(MJTI));
6814 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6817 BuildMI(DispContBB, dl, TII->get(ARM::t2ADDrs), NewVReg4)
6818 .addReg(NewVReg3, RegState::Kill)
6820 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
6822 BuildMI(DispContBB, dl, TII->get(ARM::t2BR_JT))
6823 .addReg(NewVReg4, RegState::Kill)
6825 .addJumpTableIndex(MJTI);
6826 } else if (Subtarget->isThumb()) {
6827 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6828 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRspi), NewVReg1)
6831 .addMemOperand(FIMMOLd));
6833 if (NumLPads < 256) {
6834 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPi8))
6838 MachineConstantPool *ConstantPool = MF->getConstantPool();
6839 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
6840 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
6842 // MachineConstantPool wants an explicit alignment.
6843 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
6845 Align = getDataLayout()->getTypeAllocSize(C->getType());
6846 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6848 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6849 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRpci))
6850 .addReg(VReg1, RegState::Define)
6851 .addConstantPoolIndex(Idx));
6852 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPr))
6857 BuildMI(DispatchBB, dl, TII->get(ARM::tBcc))
6862 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6863 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLSLri), NewVReg2)
6864 .addReg(ARM::CPSR, RegState::Define)
6868 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6869 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLEApcrelJT), NewVReg3)
6870 .addJumpTableIndex(MJTI));
6872 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6873 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg4)
6874 .addReg(ARM::CPSR, RegState::Define)
6875 .addReg(NewVReg2, RegState::Kill)
6878 MachineMemOperand *JTMMOLd =
6879 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
6880 MachineMemOperand::MOLoad, 4, 4);
6882 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6883 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLDRi), NewVReg5)
6884 .addReg(NewVReg4, RegState::Kill)
6886 .addMemOperand(JTMMOLd));
6888 unsigned NewVReg6 = NewVReg5;
6889 if (RelocM == Reloc::PIC_) {
6890 NewVReg6 = MRI->createVirtualRegister(TRC);
6891 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg6)
6892 .addReg(ARM::CPSR, RegState::Define)
6893 .addReg(NewVReg5, RegState::Kill)
6897 BuildMI(DispContBB, dl, TII->get(ARM::tBR_JTr))
6898 .addReg(NewVReg6, RegState::Kill)
6899 .addJumpTableIndex(MJTI);
6901 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6902 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRi12), NewVReg1)
6905 .addMemOperand(FIMMOLd));
6907 if (NumLPads < 256) {
6908 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPri))
6911 } else if (Subtarget->hasV6T2Ops() && isUInt<16>(NumLPads)) {
6912 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6913 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVi16), VReg1)
6914 .addImm(NumLPads & 0xFFFF));
6916 unsigned VReg2 = VReg1;
6917 if ((NumLPads & 0xFFFF0000) != 0) {
6918 VReg2 = MRI->createVirtualRegister(TRC);
6919 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVTi16), VReg2)
6921 .addImm(NumLPads >> 16));
6924 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
6928 MachineConstantPool *ConstantPool = MF->getConstantPool();
6929 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
6930 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
6932 // MachineConstantPool wants an explicit alignment.
6933 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
6935 Align = getDataLayout()->getTypeAllocSize(C->getType());
6936 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6938 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6939 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRcp))
6940 .addReg(VReg1, RegState::Define)
6941 .addConstantPoolIndex(Idx)
6943 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
6945 .addReg(VReg1, RegState::Kill));
6948 BuildMI(DispatchBB, dl, TII->get(ARM::Bcc))
6953 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6955 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::MOVsi), NewVReg3)
6957 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
6958 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6959 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::LEApcrelJT), NewVReg4)
6960 .addJumpTableIndex(MJTI));
6962 MachineMemOperand *JTMMOLd =
6963 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
6964 MachineMemOperand::MOLoad, 4, 4);
6965 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6967 BuildMI(DispContBB, dl, TII->get(ARM::LDRrs), NewVReg5)
6968 .addReg(NewVReg3, RegState::Kill)
6971 .addMemOperand(JTMMOLd));
6973 if (RelocM == Reloc::PIC_) {
6974 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTadd))
6975 .addReg(NewVReg5, RegState::Kill)
6977 .addJumpTableIndex(MJTI);
6979 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTr))
6980 .addReg(NewVReg5, RegState::Kill)
6981 .addJumpTableIndex(MJTI);
6985 // Add the jump table entries as successors to the MBB.
6986 SmallPtrSet<MachineBasicBlock*, 8> SeenMBBs;
6987 for (std::vector<MachineBasicBlock*>::iterator
6988 I = LPadList.begin(), E = LPadList.end(); I != E; ++I) {
6989 MachineBasicBlock *CurMBB = *I;
6990 if (SeenMBBs.insert(CurMBB).second)
6991 DispContBB->addSuccessor(CurMBB);
6994 // N.B. the order the invoke BBs are processed in doesn't matter here.
6995 const MCPhysReg *SavedRegs = RI.getCalleeSavedRegs(MF);
6996 SmallVector<MachineBasicBlock*, 64> MBBLPads;
6997 for (MachineBasicBlock *BB : InvokeBBs) {
6999 // Remove the landing pad successor from the invoke block and replace it
7000 // with the new dispatch block.
7001 SmallVector<MachineBasicBlock*, 4> Successors(BB->succ_begin(),
7003 while (!Successors.empty()) {
7004 MachineBasicBlock *SMBB = Successors.pop_back_val();
7005 if (SMBB->isLandingPad()) {
7006 BB->removeSuccessor(SMBB);
7007 MBBLPads.push_back(SMBB);
7011 BB->addSuccessor(DispatchBB);
7013 // Find the invoke call and mark all of the callee-saved registers as
7014 // 'implicit defined' so that they're spilled. This prevents code from
7015 // moving instructions to before the EH block, where they will never be
7017 for (MachineBasicBlock::reverse_iterator
7018 II = BB->rbegin(), IE = BB->rend(); II != IE; ++II) {
7019 if (!II->isCall()) continue;
7021 DenseMap<unsigned, bool> DefRegs;
7022 for (MachineInstr::mop_iterator
7023 OI = II->operands_begin(), OE = II->operands_end();
7025 if (!OI->isReg()) continue;
7026 DefRegs[OI->getReg()] = true;
7029 MachineInstrBuilder MIB(*MF, &*II);
7031 for (unsigned i = 0; SavedRegs[i] != 0; ++i) {
7032 unsigned Reg = SavedRegs[i];
7033 if (Subtarget->isThumb2() &&
7034 !ARM::tGPRRegClass.contains(Reg) &&
7035 !ARM::hGPRRegClass.contains(Reg))
7037 if (Subtarget->isThumb1Only() && !ARM::tGPRRegClass.contains(Reg))
7039 if (!Subtarget->isThumb() && !ARM::GPRRegClass.contains(Reg))
7042 MIB.addReg(Reg, RegState::ImplicitDefine | RegState::Dead);
7049 // Mark all former landing pads as non-landing pads. The dispatch is the only
7051 for (SmallVectorImpl<MachineBasicBlock*>::iterator
7052 I = MBBLPads.begin(), E = MBBLPads.end(); I != E; ++I)
7053 (*I)->setIsLandingPad(false);
7055 // The instruction is gone now.
7056 MI->eraseFromParent();
7060 MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) {
7061 for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
7062 E = MBB->succ_end(); I != E; ++I)
7065 llvm_unreachable("Expecting a BB with two successors!");
7068 /// Return the load opcode for a given load size. If load size >= 8,
7069 /// neon opcode will be returned.
7070 static unsigned getLdOpcode(unsigned LdSize, bool IsThumb1, bool IsThumb2) {
7072 return LdSize == 16 ? ARM::VLD1q32wb_fixed
7073 : LdSize == 8 ? ARM::VLD1d32wb_fixed : 0;
7075 return LdSize == 4 ? ARM::tLDRi
7076 : LdSize == 2 ? ARM::tLDRHi
7077 : LdSize == 1 ? ARM::tLDRBi : 0;
7079 return LdSize == 4 ? ARM::t2LDR_POST
7080 : LdSize == 2 ? ARM::t2LDRH_POST
7081 : LdSize == 1 ? ARM::t2LDRB_POST : 0;
7082 return LdSize == 4 ? ARM::LDR_POST_IMM
7083 : LdSize == 2 ? ARM::LDRH_POST
7084 : LdSize == 1 ? ARM::LDRB_POST_IMM : 0;
7087 /// Return the store opcode for a given store size. If store size >= 8,
7088 /// neon opcode will be returned.
7089 static unsigned getStOpcode(unsigned StSize, bool IsThumb1, bool IsThumb2) {
7091 return StSize == 16 ? ARM::VST1q32wb_fixed
7092 : StSize == 8 ? ARM::VST1d32wb_fixed : 0;
7094 return StSize == 4 ? ARM::tSTRi
7095 : StSize == 2 ? ARM::tSTRHi
7096 : StSize == 1 ? ARM::tSTRBi : 0;
7098 return StSize == 4 ? ARM::t2STR_POST
7099 : StSize == 2 ? ARM::t2STRH_POST
7100 : StSize == 1 ? ARM::t2STRB_POST : 0;
7101 return StSize == 4 ? ARM::STR_POST_IMM
7102 : StSize == 2 ? ARM::STRH_POST
7103 : StSize == 1 ? ARM::STRB_POST_IMM : 0;
7106 /// Emit a post-increment load operation with given size. The instructions
7107 /// will be added to BB at Pos.
7108 static void emitPostLd(MachineBasicBlock *BB, MachineInstr *Pos,
7109 const TargetInstrInfo *TII, DebugLoc dl,
7110 unsigned LdSize, unsigned Data, unsigned AddrIn,
7111 unsigned AddrOut, bool IsThumb1, bool IsThumb2) {
7112 unsigned LdOpc = getLdOpcode(LdSize, IsThumb1, IsThumb2);
7113 assert(LdOpc != 0 && "Should have a load opcode");
7115 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
7116 .addReg(AddrOut, RegState::Define).addReg(AddrIn)
7118 } else if (IsThumb1) {
7119 // load + update AddrIn
7120 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
7121 .addReg(AddrIn).addImm(0));
7122 MachineInstrBuilder MIB =
7123 BuildMI(*BB, Pos, dl, TII->get(ARM::tADDi8), AddrOut);
7124 MIB = AddDefaultT1CC(MIB);
7125 MIB.addReg(AddrIn).addImm(LdSize);
7126 AddDefaultPred(MIB);
7127 } else if (IsThumb2) {
7128 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
7129 .addReg(AddrOut, RegState::Define).addReg(AddrIn)
7132 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
7133 .addReg(AddrOut, RegState::Define).addReg(AddrIn)
7134 .addReg(0).addImm(LdSize));
7138 /// Emit a post-increment store operation with given size. The instructions
7139 /// will be added to BB at Pos.
7140 static void emitPostSt(MachineBasicBlock *BB, MachineInstr *Pos,
7141 const TargetInstrInfo *TII, DebugLoc dl,
7142 unsigned StSize, unsigned Data, unsigned AddrIn,
7143 unsigned AddrOut, bool IsThumb1, bool IsThumb2) {
7144 unsigned StOpc = getStOpcode(StSize, IsThumb1, IsThumb2);
7145 assert(StOpc != 0 && "Should have a store opcode");
7147 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc), AddrOut)
7148 .addReg(AddrIn).addImm(0).addReg(Data));
7149 } else if (IsThumb1) {
7150 // store + update AddrIn
7151 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc)).addReg(Data)
7152 .addReg(AddrIn).addImm(0));
7153 MachineInstrBuilder MIB =
7154 BuildMI(*BB, Pos, dl, TII->get(ARM::tADDi8), AddrOut);
7155 MIB = AddDefaultT1CC(MIB);
7156 MIB.addReg(AddrIn).addImm(StSize);
7157 AddDefaultPred(MIB);
7158 } else if (IsThumb2) {
7159 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc), AddrOut)
7160 .addReg(Data).addReg(AddrIn).addImm(StSize));
7162 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc), AddrOut)
7163 .addReg(Data).addReg(AddrIn).addReg(0)
7169 ARMTargetLowering::EmitStructByval(MachineInstr *MI,
7170 MachineBasicBlock *BB) const {
7171 // This pseudo instruction has 3 operands: dst, src, size
7172 // We expand it to a loop if size > Subtarget->getMaxInlineSizeThreshold().
7173 // Otherwise, we will generate unrolled scalar copies.
7174 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
7175 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7176 MachineFunction::iterator It = BB;
7179 unsigned dest = MI->getOperand(0).getReg();
7180 unsigned src = MI->getOperand(1).getReg();
7181 unsigned SizeVal = MI->getOperand(2).getImm();
7182 unsigned Align = MI->getOperand(3).getImm();
7183 DebugLoc dl = MI->getDebugLoc();
7185 MachineFunction *MF = BB->getParent();
7186 MachineRegisterInfo &MRI = MF->getRegInfo();
7187 unsigned UnitSize = 0;
7188 const TargetRegisterClass *TRC = nullptr;
7189 const TargetRegisterClass *VecTRC = nullptr;
7191 bool IsThumb1 = Subtarget->isThumb1Only();
7192 bool IsThumb2 = Subtarget->isThumb2();
7196 } else if (Align & 2) {
7199 // Check whether we can use NEON instructions.
7200 if (!MF->getFunction()->hasFnAttribute(Attribute::NoImplicitFloat) &&
7201 Subtarget->hasNEON()) {
7202 if ((Align % 16 == 0) && SizeVal >= 16)
7204 else if ((Align % 8 == 0) && SizeVal >= 8)
7207 // Can't use NEON instructions.
7212 // Select the correct opcode and register class for unit size load/store
7213 bool IsNeon = UnitSize >= 8;
7214 TRC = (IsThumb1 || IsThumb2) ? &ARM::tGPRRegClass : &ARM::GPRRegClass;
7216 VecTRC = UnitSize == 16 ? &ARM::DPairRegClass
7217 : UnitSize == 8 ? &ARM::DPRRegClass
7220 unsigned BytesLeft = SizeVal % UnitSize;
7221 unsigned LoopSize = SizeVal - BytesLeft;
7223 if (SizeVal <= Subtarget->getMaxInlineSizeThreshold()) {
7224 // Use LDR and STR to copy.
7225 // [scratch, srcOut] = LDR_POST(srcIn, UnitSize)
7226 // [destOut] = STR_POST(scratch, destIn, UnitSize)
7227 unsigned srcIn = src;
7228 unsigned destIn = dest;
7229 for (unsigned i = 0; i < LoopSize; i+=UnitSize) {
7230 unsigned srcOut = MRI.createVirtualRegister(TRC);
7231 unsigned destOut = MRI.createVirtualRegister(TRC);
7232 unsigned scratch = MRI.createVirtualRegister(IsNeon ? VecTRC : TRC);
7233 emitPostLd(BB, MI, TII, dl, UnitSize, scratch, srcIn, srcOut,
7234 IsThumb1, IsThumb2);
7235 emitPostSt(BB, MI, TII, dl, UnitSize, scratch, destIn, destOut,
7236 IsThumb1, IsThumb2);
7241 // Handle the leftover bytes with LDRB and STRB.
7242 // [scratch, srcOut] = LDRB_POST(srcIn, 1)
7243 // [destOut] = STRB_POST(scratch, destIn, 1)
7244 for (unsigned i = 0; i < BytesLeft; i++) {
7245 unsigned srcOut = MRI.createVirtualRegister(TRC);
7246 unsigned destOut = MRI.createVirtualRegister(TRC);
7247 unsigned scratch = MRI.createVirtualRegister(TRC);
7248 emitPostLd(BB, MI, TII, dl, 1, scratch, srcIn, srcOut,
7249 IsThumb1, IsThumb2);
7250 emitPostSt(BB, MI, TII, dl, 1, scratch, destIn, destOut,
7251 IsThumb1, IsThumb2);
7255 MI->eraseFromParent(); // The instruction is gone now.
7259 // Expand the pseudo op to a loop.
7262 // movw varEnd, # --> with thumb2
7264 // ldrcp varEnd, idx --> without thumb2
7265 // fallthrough --> loopMBB
7267 // PHI varPhi, varEnd, varLoop
7268 // PHI srcPhi, src, srcLoop
7269 // PHI destPhi, dst, destLoop
7270 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
7271 // [destLoop] = STR_POST(scratch, destPhi, UnitSize)
7272 // subs varLoop, varPhi, #UnitSize
7274 // fallthrough --> exitMBB
7276 // epilogue to handle left-over bytes
7277 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
7278 // [destOut] = STRB_POST(scratch, destLoop, 1)
7279 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
7280 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
7281 MF->insert(It, loopMBB);
7282 MF->insert(It, exitMBB);
7284 // Transfer the remainder of BB and its successor edges to exitMBB.
7285 exitMBB->splice(exitMBB->begin(), BB,
7286 std::next(MachineBasicBlock::iterator(MI)), BB->end());
7287 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
7289 // Load an immediate to varEnd.
7290 unsigned varEnd = MRI.createVirtualRegister(TRC);
7291 if (Subtarget->useMovt(*MF)) {
7292 unsigned Vtmp = varEnd;
7293 if ((LoopSize & 0xFFFF0000) != 0)
7294 Vtmp = MRI.createVirtualRegister(TRC);
7295 AddDefaultPred(BuildMI(BB, dl,
7296 TII->get(IsThumb2 ? ARM::t2MOVi16 : ARM::MOVi16),
7297 Vtmp).addImm(LoopSize & 0xFFFF));
7299 if ((LoopSize & 0xFFFF0000) != 0)
7300 AddDefaultPred(BuildMI(BB, dl,
7301 TII->get(IsThumb2 ? ARM::t2MOVTi16 : ARM::MOVTi16),
7304 .addImm(LoopSize >> 16));
7306 MachineConstantPool *ConstantPool = MF->getConstantPool();
7307 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
7308 const Constant *C = ConstantInt::get(Int32Ty, LoopSize);
7310 // MachineConstantPool wants an explicit alignment.
7311 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
7313 Align = getDataLayout()->getTypeAllocSize(C->getType());
7314 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
7317 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(ARM::tLDRpci)).addReg(
7318 varEnd, RegState::Define).addConstantPoolIndex(Idx));
7320 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(ARM::LDRcp)).addReg(
7321 varEnd, RegState::Define).addConstantPoolIndex(Idx).addImm(0));
7323 BB->addSuccessor(loopMBB);
7325 // Generate the loop body:
7326 // varPhi = PHI(varLoop, varEnd)
7327 // srcPhi = PHI(srcLoop, src)
7328 // destPhi = PHI(destLoop, dst)
7329 MachineBasicBlock *entryBB = BB;
7331 unsigned varLoop = MRI.createVirtualRegister(TRC);
7332 unsigned varPhi = MRI.createVirtualRegister(TRC);
7333 unsigned srcLoop = MRI.createVirtualRegister(TRC);
7334 unsigned srcPhi = MRI.createVirtualRegister(TRC);
7335 unsigned destLoop = MRI.createVirtualRegister(TRC);
7336 unsigned destPhi = MRI.createVirtualRegister(TRC);
7338 BuildMI(*BB, BB->begin(), dl, TII->get(ARM::PHI), varPhi)
7339 .addReg(varLoop).addMBB(loopMBB)
7340 .addReg(varEnd).addMBB(entryBB);
7341 BuildMI(BB, dl, TII->get(ARM::PHI), srcPhi)
7342 .addReg(srcLoop).addMBB(loopMBB)
7343 .addReg(src).addMBB(entryBB);
7344 BuildMI(BB, dl, TII->get(ARM::PHI), destPhi)
7345 .addReg(destLoop).addMBB(loopMBB)
7346 .addReg(dest).addMBB(entryBB);
7348 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
7349 // [destLoop] = STR_POST(scratch, destPhi, UnitSiz)
7350 unsigned scratch = MRI.createVirtualRegister(IsNeon ? VecTRC : TRC);
7351 emitPostLd(BB, BB->end(), TII, dl, UnitSize, scratch, srcPhi, srcLoop,
7352 IsThumb1, IsThumb2);
7353 emitPostSt(BB, BB->end(), TII, dl, UnitSize, scratch, destPhi, destLoop,
7354 IsThumb1, IsThumb2);
7356 // Decrement loop variable by UnitSize.
7358 MachineInstrBuilder MIB =
7359 BuildMI(*BB, BB->end(), dl, TII->get(ARM::tSUBi8), varLoop);
7360 MIB = AddDefaultT1CC(MIB);
7361 MIB.addReg(varPhi).addImm(UnitSize);
7362 AddDefaultPred(MIB);
7364 MachineInstrBuilder MIB =
7365 BuildMI(*BB, BB->end(), dl,
7366 TII->get(IsThumb2 ? ARM::t2SUBri : ARM::SUBri), varLoop);
7367 AddDefaultCC(AddDefaultPred(MIB.addReg(varPhi).addImm(UnitSize)));
7368 MIB->getOperand(5).setReg(ARM::CPSR);
7369 MIB->getOperand(5).setIsDef(true);
7371 BuildMI(*BB, BB->end(), dl,
7372 TII->get(IsThumb1 ? ARM::tBcc : IsThumb2 ? ARM::t2Bcc : ARM::Bcc))
7373 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
7375 // loopMBB can loop back to loopMBB or fall through to exitMBB.
7376 BB->addSuccessor(loopMBB);
7377 BB->addSuccessor(exitMBB);
7379 // Add epilogue to handle BytesLeft.
7381 MachineInstr *StartOfExit = exitMBB->begin();
7383 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
7384 // [destOut] = STRB_POST(scratch, destLoop, 1)
7385 unsigned srcIn = srcLoop;
7386 unsigned destIn = destLoop;
7387 for (unsigned i = 0; i < BytesLeft; i++) {
7388 unsigned srcOut = MRI.createVirtualRegister(TRC);
7389 unsigned destOut = MRI.createVirtualRegister(TRC);
7390 unsigned scratch = MRI.createVirtualRegister(TRC);
7391 emitPostLd(BB, StartOfExit, TII, dl, 1, scratch, srcIn, srcOut,
7392 IsThumb1, IsThumb2);
7393 emitPostSt(BB, StartOfExit, TII, dl, 1, scratch, destIn, destOut,
7394 IsThumb1, IsThumb2);
7399 MI->eraseFromParent(); // The instruction is gone now.
7404 ARMTargetLowering::EmitLowered__chkstk(MachineInstr *MI,
7405 MachineBasicBlock *MBB) const {
7406 const TargetMachine &TM = getTargetMachine();
7407 const TargetInstrInfo &TII = *Subtarget->getInstrInfo();
7408 DebugLoc DL = MI->getDebugLoc();
7410 assert(Subtarget->isTargetWindows() &&
7411 "__chkstk is only supported on Windows");
7412 assert(Subtarget->isThumb2() && "Windows on ARM requires Thumb-2 mode");
7414 // __chkstk takes the number of words to allocate on the stack in R4, and
7415 // returns the stack adjustment in number of bytes in R4. This will not
7416 // clober any other registers (other than the obvious lr).
7418 // Although, technically, IP should be considered a register which may be
7419 // clobbered, the call itself will not touch it. Windows on ARM is a pure
7420 // thumb-2 environment, so there is no interworking required. As a result, we
7421 // do not expect a veneer to be emitted by the linker, clobbering IP.
7423 // Each module receives its own copy of __chkstk, so no import thunk is
7424 // required, again, ensuring that IP is not clobbered.
7426 // Finally, although some linkers may theoretically provide a trampoline for
7427 // out of range calls (which is quite common due to a 32M range limitation of
7428 // branches for Thumb), we can generate the long-call version via
7429 // -mcmodel=large, alleviating the need for the trampoline which may clobber
7432 switch (TM.getCodeModel()) {
7433 case CodeModel::Small:
7434 case CodeModel::Medium:
7435 case CodeModel::Default:
7436 case CodeModel::Kernel:
7437 BuildMI(*MBB, MI, DL, TII.get(ARM::tBL))
7438 .addImm((unsigned)ARMCC::AL).addReg(0)
7439 .addExternalSymbol("__chkstk")
7440 .addReg(ARM::R4, RegState::Implicit | RegState::Kill)
7441 .addReg(ARM::R4, RegState::Implicit | RegState::Define)
7442 .addReg(ARM::R12, RegState::Implicit | RegState::Define | RegState::Dead);
7444 case CodeModel::Large:
7445 case CodeModel::JITDefault: {
7446 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
7447 unsigned Reg = MRI.createVirtualRegister(&ARM::rGPRRegClass);
7449 BuildMI(*MBB, MI, DL, TII.get(ARM::t2MOVi32imm), Reg)
7450 .addExternalSymbol("__chkstk");
7451 BuildMI(*MBB, MI, DL, TII.get(ARM::tBLXr))
7452 .addImm((unsigned)ARMCC::AL).addReg(0)
7453 .addReg(Reg, RegState::Kill)
7454 .addReg(ARM::R4, RegState::Implicit | RegState::Kill)
7455 .addReg(ARM::R4, RegState::Implicit | RegState::Define)
7456 .addReg(ARM::R12, RegState::Implicit | RegState::Define | RegState::Dead);
7461 AddDefaultCC(AddDefaultPred(BuildMI(*MBB, MI, DL, TII.get(ARM::t2SUBrr),
7463 .addReg(ARM::SP).addReg(ARM::R4)));
7465 MI->eraseFromParent();
7470 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
7471 MachineBasicBlock *BB) const {
7472 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
7473 DebugLoc dl = MI->getDebugLoc();
7474 bool isThumb2 = Subtarget->isThumb2();
7475 switch (MI->getOpcode()) {
7478 llvm_unreachable("Unexpected instr type to insert");
7480 // The Thumb2 pre-indexed stores have the same MI operands, they just
7481 // define them differently in the .td files from the isel patterns, so
7482 // they need pseudos.
7483 case ARM::t2STR_preidx:
7484 MI->setDesc(TII->get(ARM::t2STR_PRE));
7486 case ARM::t2STRB_preidx:
7487 MI->setDesc(TII->get(ARM::t2STRB_PRE));
7489 case ARM::t2STRH_preidx:
7490 MI->setDesc(TII->get(ARM::t2STRH_PRE));
7493 case ARM::STRi_preidx:
7494 case ARM::STRBi_preidx: {
7495 unsigned NewOpc = MI->getOpcode() == ARM::STRi_preidx ?
7496 ARM::STR_PRE_IMM : ARM::STRB_PRE_IMM;
7497 // Decode the offset.
7498 unsigned Offset = MI->getOperand(4).getImm();
7499 bool isSub = ARM_AM::getAM2Op(Offset) == ARM_AM::sub;
7500 Offset = ARM_AM::getAM2Offset(Offset);
7504 MachineMemOperand *MMO = *MI->memoperands_begin();
7505 BuildMI(*BB, MI, dl, TII->get(NewOpc))
7506 .addOperand(MI->getOperand(0)) // Rn_wb
7507 .addOperand(MI->getOperand(1)) // Rt
7508 .addOperand(MI->getOperand(2)) // Rn
7509 .addImm(Offset) // offset (skip GPR==zero_reg)
7510 .addOperand(MI->getOperand(5)) // pred
7511 .addOperand(MI->getOperand(6))
7512 .addMemOperand(MMO);
7513 MI->eraseFromParent();
7516 case ARM::STRr_preidx:
7517 case ARM::STRBr_preidx:
7518 case ARM::STRH_preidx: {
7520 switch (MI->getOpcode()) {
7521 default: llvm_unreachable("unexpected opcode!");
7522 case ARM::STRr_preidx: NewOpc = ARM::STR_PRE_REG; break;
7523 case ARM::STRBr_preidx: NewOpc = ARM::STRB_PRE_REG; break;
7524 case ARM::STRH_preidx: NewOpc = ARM::STRH_PRE; break;
7526 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(NewOpc));
7527 for (unsigned i = 0; i < MI->getNumOperands(); ++i)
7528 MIB.addOperand(MI->getOperand(i));
7529 MI->eraseFromParent();
7533 case ARM::tMOVCCr_pseudo: {
7534 // To "insert" a SELECT_CC instruction, we actually have to insert the
7535 // diamond control-flow pattern. The incoming instruction knows the
7536 // destination vreg to set, the condition code register to branch on, the
7537 // true/false values to select between, and a branch opcode to use.
7538 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7539 MachineFunction::iterator It = BB;
7545 // cmpTY ccX, r1, r2
7547 // fallthrough --> copy0MBB
7548 MachineBasicBlock *thisMBB = BB;
7549 MachineFunction *F = BB->getParent();
7550 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
7551 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
7552 F->insert(It, copy0MBB);
7553 F->insert(It, sinkMBB);
7555 // Transfer the remainder of BB and its successor edges to sinkMBB.
7556 sinkMBB->splice(sinkMBB->begin(), BB,
7557 std::next(MachineBasicBlock::iterator(MI)), BB->end());
7558 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
7560 BB->addSuccessor(copy0MBB);
7561 BB->addSuccessor(sinkMBB);
7563 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
7564 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
7567 // %FalseValue = ...
7568 // # fallthrough to sinkMBB
7571 // Update machine-CFG edges
7572 BB->addSuccessor(sinkMBB);
7575 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
7578 BuildMI(*BB, BB->begin(), dl,
7579 TII->get(ARM::PHI), MI->getOperand(0).getReg())
7580 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
7581 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
7583 MI->eraseFromParent(); // The pseudo instruction is gone now.
7588 case ARM::BCCZi64: {
7589 // If there is an unconditional branch to the other successor, remove it.
7590 BB->erase(std::next(MachineBasicBlock::iterator(MI)), BB->end());
7592 // Compare both parts that make up the double comparison separately for
7594 bool RHSisZero = MI->getOpcode() == ARM::BCCZi64;
7596 unsigned LHS1 = MI->getOperand(1).getReg();
7597 unsigned LHS2 = MI->getOperand(2).getReg();
7599 AddDefaultPred(BuildMI(BB, dl,
7600 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7601 .addReg(LHS1).addImm(0));
7602 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7603 .addReg(LHS2).addImm(0)
7604 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
7606 unsigned RHS1 = MI->getOperand(3).getReg();
7607 unsigned RHS2 = MI->getOperand(4).getReg();
7608 AddDefaultPred(BuildMI(BB, dl,
7609 TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
7610 .addReg(LHS1).addReg(RHS1));
7611 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
7612 .addReg(LHS2).addReg(RHS2)
7613 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
7616 MachineBasicBlock *destMBB = MI->getOperand(RHSisZero ? 3 : 5).getMBB();
7617 MachineBasicBlock *exitMBB = OtherSucc(BB, destMBB);
7618 if (MI->getOperand(0).getImm() == ARMCC::NE)
7619 std::swap(destMBB, exitMBB);
7621 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
7622 .addMBB(destMBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
7624 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2B)).addMBB(exitMBB));
7626 BuildMI(BB, dl, TII->get(ARM::B)) .addMBB(exitMBB);
7628 MI->eraseFromParent(); // The pseudo instruction is gone now.
7632 case ARM::Int_eh_sjlj_setjmp:
7633 case ARM::Int_eh_sjlj_setjmp_nofp:
7634 case ARM::tInt_eh_sjlj_setjmp:
7635 case ARM::t2Int_eh_sjlj_setjmp:
7636 case ARM::t2Int_eh_sjlj_setjmp_nofp:
7637 EmitSjLjDispatchBlock(MI, BB);
7642 // To insert an ABS instruction, we have to insert the
7643 // diamond control-flow pattern. The incoming instruction knows the
7644 // source vreg to test against 0, the destination vreg to set,
7645 // the condition code register to branch on, the
7646 // true/false values to select between, and a branch opcode to use.
7651 // BCC (branch to SinkBB if V0 >= 0)
7652 // RSBBB: V3 = RSBri V2, 0 (compute ABS if V2 < 0)
7653 // SinkBB: V1 = PHI(V2, V3)
7654 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7655 MachineFunction::iterator BBI = BB;
7657 MachineFunction *Fn = BB->getParent();
7658 MachineBasicBlock *RSBBB = Fn->CreateMachineBasicBlock(LLVM_BB);
7659 MachineBasicBlock *SinkBB = Fn->CreateMachineBasicBlock(LLVM_BB);
7660 Fn->insert(BBI, RSBBB);
7661 Fn->insert(BBI, SinkBB);
7663 unsigned int ABSSrcReg = MI->getOperand(1).getReg();
7664 unsigned int ABSDstReg = MI->getOperand(0).getReg();
7665 bool ABSSrcKIll = MI->getOperand(1).isKill();
7666 bool isThumb2 = Subtarget->isThumb2();
7667 MachineRegisterInfo &MRI = Fn->getRegInfo();
7668 // In Thumb mode S must not be specified if source register is the SP or
7669 // PC and if destination register is the SP, so restrict register class
7670 unsigned NewRsbDstReg =
7671 MRI.createVirtualRegister(isThumb2 ? &ARM::rGPRRegClass : &ARM::GPRRegClass);
7673 // Transfer the remainder of BB and its successor edges to sinkMBB.
7674 SinkBB->splice(SinkBB->begin(), BB,
7675 std::next(MachineBasicBlock::iterator(MI)), BB->end());
7676 SinkBB->transferSuccessorsAndUpdatePHIs(BB);
7678 BB->addSuccessor(RSBBB);
7679 BB->addSuccessor(SinkBB);
7681 // fall through to SinkMBB
7682 RSBBB->addSuccessor(SinkBB);
7684 // insert a cmp at the end of BB
7685 AddDefaultPred(BuildMI(BB, dl,
7686 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7687 .addReg(ABSSrcReg).addImm(0));
7689 // insert a bcc with opposite CC to ARMCC::MI at the end of BB
7691 TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)).addMBB(SinkBB)
7692 .addImm(ARMCC::getOppositeCondition(ARMCC::MI)).addReg(ARM::CPSR);
7694 // insert rsbri in RSBBB
7695 // Note: BCC and rsbri will be converted into predicated rsbmi
7696 // by if-conversion pass
7697 BuildMI(*RSBBB, RSBBB->begin(), dl,
7698 TII->get(isThumb2 ? ARM::t2RSBri : ARM::RSBri), NewRsbDstReg)
7699 .addReg(ABSSrcReg, ABSSrcKIll ? RegState::Kill : 0)
7700 .addImm(0).addImm((unsigned)ARMCC::AL).addReg(0).addReg(0);
7702 // insert PHI in SinkBB,
7703 // reuse ABSDstReg to not change uses of ABS instruction
7704 BuildMI(*SinkBB, SinkBB->begin(), dl,
7705 TII->get(ARM::PHI), ABSDstReg)
7706 .addReg(NewRsbDstReg).addMBB(RSBBB)
7707 .addReg(ABSSrcReg).addMBB(BB);
7709 // remove ABS instruction
7710 MI->eraseFromParent();
7712 // return last added BB
7715 case ARM::COPY_STRUCT_BYVAL_I32:
7717 return EmitStructByval(MI, BB);
7718 case ARM::WIN__CHKSTK:
7719 return EmitLowered__chkstk(MI, BB);
7723 void ARMTargetLowering::AdjustInstrPostInstrSelection(MachineInstr *MI,
7724 SDNode *Node) const {
7725 const MCInstrDesc *MCID = &MI->getDesc();
7726 // Adjust potentially 's' setting instructions after isel, i.e. ADC, SBC, RSB,
7727 // RSC. Coming out of isel, they have an implicit CPSR def, but the optional
7728 // operand is still set to noreg. If needed, set the optional operand's
7729 // register to CPSR, and remove the redundant implicit def.
7731 // e.g. ADCS (..., CPSR<imp-def>) -> ADC (... opt:CPSR<def>).
7733 // Rename pseudo opcodes.
7734 unsigned NewOpc = convertAddSubFlagsOpcode(MI->getOpcode());
7736 const ARMBaseInstrInfo *TII = Subtarget->getInstrInfo();
7737 MCID = &TII->get(NewOpc);
7739 assert(MCID->getNumOperands() == MI->getDesc().getNumOperands() + 1 &&
7740 "converted opcode should be the same except for cc_out");
7744 // Add the optional cc_out operand
7745 MI->addOperand(MachineOperand::CreateReg(0, /*isDef=*/true));
7747 unsigned ccOutIdx = MCID->getNumOperands() - 1;
7749 // Any ARM instruction that sets the 's' bit should specify an optional
7750 // "cc_out" operand in the last operand position.
7751 if (!MI->hasOptionalDef() || !MCID->OpInfo[ccOutIdx].isOptionalDef()) {
7752 assert(!NewOpc && "Optional cc_out operand required");
7755 // Look for an implicit def of CPSR added by MachineInstr ctor. Remove it
7756 // since we already have an optional CPSR def.
7757 bool definesCPSR = false;
7758 bool deadCPSR = false;
7759 for (unsigned i = MCID->getNumOperands(), e = MI->getNumOperands();
7761 const MachineOperand &MO = MI->getOperand(i);
7762 if (MO.isReg() && MO.isDef() && MO.getReg() == ARM::CPSR) {
7766 MI->RemoveOperand(i);
7771 assert(!NewOpc && "Optional cc_out operand required");
7774 assert(deadCPSR == !Node->hasAnyUseOfValue(1) && "inconsistent dead flag");
7776 assert(!MI->getOperand(ccOutIdx).getReg() &&
7777 "expect uninitialized optional cc_out operand");
7781 // If this instruction was defined with an optional CPSR def and its dag node
7782 // had a live implicit CPSR def, then activate the optional CPSR def.
7783 MachineOperand &MO = MI->getOperand(ccOutIdx);
7784 MO.setReg(ARM::CPSR);
7788 //===----------------------------------------------------------------------===//
7789 // ARM Optimization Hooks
7790 //===----------------------------------------------------------------------===//
7792 // Helper function that checks if N is a null or all ones constant.
7793 static inline bool isZeroOrAllOnes(SDValue N, bool AllOnes) {
7794 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N);
7797 return AllOnes ? C->isAllOnesValue() : C->isNullValue();
7800 // Return true if N is conditionally 0 or all ones.
7801 // Detects these expressions where cc is an i1 value:
7803 // (select cc 0, y) [AllOnes=0]
7804 // (select cc y, 0) [AllOnes=0]
7805 // (zext cc) [AllOnes=0]
7806 // (sext cc) [AllOnes=0/1]
7807 // (select cc -1, y) [AllOnes=1]
7808 // (select cc y, -1) [AllOnes=1]
7810 // Invert is set when N is the null/all ones constant when CC is false.
7811 // OtherOp is set to the alternative value of N.
7812 static bool isConditionalZeroOrAllOnes(SDNode *N, bool AllOnes,
7813 SDValue &CC, bool &Invert,
7815 SelectionDAG &DAG) {
7816 switch (N->getOpcode()) {
7817 default: return false;
7819 CC = N->getOperand(0);
7820 SDValue N1 = N->getOperand(1);
7821 SDValue N2 = N->getOperand(2);
7822 if (isZeroOrAllOnes(N1, AllOnes)) {
7827 if (isZeroOrAllOnes(N2, AllOnes)) {
7834 case ISD::ZERO_EXTEND:
7835 // (zext cc) can never be the all ones value.
7839 case ISD::SIGN_EXTEND: {
7841 EVT VT = N->getValueType(0);
7842 CC = N->getOperand(0);
7843 if (CC.getValueType() != MVT::i1)
7847 // When looking for an AllOnes constant, N is an sext, and the 'other'
7849 OtherOp = DAG.getConstant(0, dl, VT);
7850 else if (N->getOpcode() == ISD::ZERO_EXTEND)
7851 // When looking for a 0 constant, N can be zext or sext.
7852 OtherOp = DAG.getConstant(1, dl, VT);
7854 OtherOp = DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), dl,
7861 // Combine a constant select operand into its use:
7863 // (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
7864 // (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
7865 // (and (select cc, -1, c), x) -> (select cc, x, (and, x, c)) [AllOnes=1]
7866 // (or (select cc, 0, c), x) -> (select cc, x, (or, x, c))
7867 // (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c))
7869 // The transform is rejected if the select doesn't have a constant operand that
7870 // is null, or all ones when AllOnes is set.
7872 // Also recognize sext/zext from i1:
7874 // (add (zext cc), x) -> (select cc (add x, 1), x)
7875 // (add (sext cc), x) -> (select cc (add x, -1), x)
7877 // These transformations eventually create predicated instructions.
7879 // @param N The node to transform.
7880 // @param Slct The N operand that is a select.
7881 // @param OtherOp The other N operand (x above).
7882 // @param DCI Context.
7883 // @param AllOnes Require the select constant to be all ones instead of null.
7884 // @returns The new node, or SDValue() on failure.
7886 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
7887 TargetLowering::DAGCombinerInfo &DCI,
7888 bool AllOnes = false) {
7889 SelectionDAG &DAG = DCI.DAG;
7890 EVT VT = N->getValueType(0);
7891 SDValue NonConstantVal;
7894 if (!isConditionalZeroOrAllOnes(Slct.getNode(), AllOnes, CCOp, SwapSelectOps,
7895 NonConstantVal, DAG))
7898 // Slct is now know to be the desired identity constant when CC is true.
7899 SDValue TrueVal = OtherOp;
7900 SDValue FalseVal = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
7901 OtherOp, NonConstantVal);
7902 // Unless SwapSelectOps says CC should be false.
7904 std::swap(TrueVal, FalseVal);
7906 return DAG.getNode(ISD::SELECT, SDLoc(N), VT,
7907 CCOp, TrueVal, FalseVal);
7910 // Attempt combineSelectAndUse on each operand of a commutative operator N.
7912 SDValue combineSelectAndUseCommutative(SDNode *N, bool AllOnes,
7913 TargetLowering::DAGCombinerInfo &DCI) {
7914 SDValue N0 = N->getOperand(0);
7915 SDValue N1 = N->getOperand(1);
7916 if (N0.getNode()->hasOneUse()) {
7917 SDValue Result = combineSelectAndUse(N, N0, N1, DCI, AllOnes);
7918 if (Result.getNode())
7921 if (N1.getNode()->hasOneUse()) {
7922 SDValue Result = combineSelectAndUse(N, N1, N0, DCI, AllOnes);
7923 if (Result.getNode())
7929 // AddCombineToVPADDL- For pair-wise add on neon, use the vpaddl instruction
7930 // (only after legalization).
7931 static SDValue AddCombineToVPADDL(SDNode *N, SDValue N0, SDValue N1,
7932 TargetLowering::DAGCombinerInfo &DCI,
7933 const ARMSubtarget *Subtarget) {
7935 // Only perform optimization if after legalize, and if NEON is available. We
7936 // also expected both operands to be BUILD_VECTORs.
7937 if (DCI.isBeforeLegalize() || !Subtarget->hasNEON()
7938 || N0.getOpcode() != ISD::BUILD_VECTOR
7939 || N1.getOpcode() != ISD::BUILD_VECTOR)
7942 // Check output type since VPADDL operand elements can only be 8, 16, or 32.
7943 EVT VT = N->getValueType(0);
7944 if (!VT.isInteger() || VT.getVectorElementType() == MVT::i64)
7947 // Check that the vector operands are of the right form.
7948 // N0 and N1 are BUILD_VECTOR nodes with N number of EXTRACT_VECTOR
7949 // operands, where N is the size of the formed vector.
7950 // Each EXTRACT_VECTOR should have the same input vector and odd or even
7951 // index such that we have a pair wise add pattern.
7953 // Grab the vector that all EXTRACT_VECTOR nodes should be referencing.
7954 if (N0->getOperand(0)->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
7956 SDValue Vec = N0->getOperand(0)->getOperand(0);
7957 SDNode *V = Vec.getNode();
7958 unsigned nextIndex = 0;
7960 // For each operands to the ADD which are BUILD_VECTORs,
7961 // check to see if each of their operands are an EXTRACT_VECTOR with
7962 // the same vector and appropriate index.
7963 for (unsigned i = 0, e = N0->getNumOperands(); i != e; ++i) {
7964 if (N0->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT
7965 && N1->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
7967 SDValue ExtVec0 = N0->getOperand(i);
7968 SDValue ExtVec1 = N1->getOperand(i);
7970 // First operand is the vector, verify its the same.
7971 if (V != ExtVec0->getOperand(0).getNode() ||
7972 V != ExtVec1->getOperand(0).getNode())
7975 // Second is the constant, verify its correct.
7976 ConstantSDNode *C0 = dyn_cast<ConstantSDNode>(ExtVec0->getOperand(1));
7977 ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(ExtVec1->getOperand(1));
7979 // For the constant, we want to see all the even or all the odd.
7980 if (!C0 || !C1 || C0->getZExtValue() != nextIndex
7981 || C1->getZExtValue() != nextIndex+1)
7990 // Create VPADDL node.
7991 SelectionDAG &DAG = DCI.DAG;
7992 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7996 // Build operand list.
7997 SmallVector<SDValue, 8> Ops;
7998 Ops.push_back(DAG.getConstant(Intrinsic::arm_neon_vpaddls, dl,
7999 TLI.getPointerTy()));
8001 // Input is the vector.
8004 // Get widened type and narrowed type.
8006 unsigned numElem = VT.getVectorNumElements();
8008 EVT inputLaneType = Vec.getValueType().getVectorElementType();
8009 switch (inputLaneType.getSimpleVT().SimpleTy) {
8010 case MVT::i8: widenType = MVT::getVectorVT(MVT::i16, numElem); break;
8011 case MVT::i16: widenType = MVT::getVectorVT(MVT::i32, numElem); break;
8012 case MVT::i32: widenType = MVT::getVectorVT(MVT::i64, numElem); break;
8014 llvm_unreachable("Invalid vector element type for padd optimization.");
8017 SDValue tmp = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, widenType, Ops);
8018 unsigned ExtOp = VT.bitsGT(tmp.getValueType()) ? ISD::ANY_EXTEND : ISD::TRUNCATE;
8019 return DAG.getNode(ExtOp, dl, VT, tmp);
8022 static SDValue findMUL_LOHI(SDValue V) {
8023 if (V->getOpcode() == ISD::UMUL_LOHI ||
8024 V->getOpcode() == ISD::SMUL_LOHI)
8029 static SDValue AddCombineTo64bitMLAL(SDNode *AddcNode,
8030 TargetLowering::DAGCombinerInfo &DCI,
8031 const ARMSubtarget *Subtarget) {
8033 if (Subtarget->isThumb1Only()) return SDValue();
8035 // Only perform the checks after legalize when the pattern is available.
8036 if (DCI.isBeforeLegalize()) return SDValue();
8038 // Look for multiply add opportunities.
8039 // The pattern is a ISD::UMUL_LOHI followed by two add nodes, where
8040 // each add nodes consumes a value from ISD::UMUL_LOHI and there is
8041 // a glue link from the first add to the second add.
8042 // If we find this pattern, we can replace the U/SMUL_LOHI, ADDC, and ADDE by
8043 // a S/UMLAL instruction.
8046 // / \ [no multiline comment]
8052 assert(AddcNode->getOpcode() == ISD::ADDC && "Expect an ADDC");
8053 SDValue AddcOp0 = AddcNode->getOperand(0);
8054 SDValue AddcOp1 = AddcNode->getOperand(1);
8056 // Check if the two operands are from the same mul_lohi node.
8057 if (AddcOp0.getNode() == AddcOp1.getNode())
8060 assert(AddcNode->getNumValues() == 2 &&
8061 AddcNode->getValueType(0) == MVT::i32 &&
8062 "Expect ADDC with two result values. First: i32");
8064 // Check that we have a glued ADDC node.
8065 if (AddcNode->getValueType(1) != MVT::Glue)
8068 // Check that the ADDC adds the low result of the S/UMUL_LOHI.
8069 if (AddcOp0->getOpcode() != ISD::UMUL_LOHI &&
8070 AddcOp0->getOpcode() != ISD::SMUL_LOHI &&
8071 AddcOp1->getOpcode() != ISD::UMUL_LOHI &&
8072 AddcOp1->getOpcode() != ISD::SMUL_LOHI)
8075 // Look for the glued ADDE.
8076 SDNode* AddeNode = AddcNode->getGluedUser();
8080 // Make sure it is really an ADDE.
8081 if (AddeNode->getOpcode() != ISD::ADDE)
8084 assert(AddeNode->getNumOperands() == 3 &&
8085 AddeNode->getOperand(2).getValueType() == MVT::Glue &&
8086 "ADDE node has the wrong inputs");
8088 // Check for the triangle shape.
8089 SDValue AddeOp0 = AddeNode->getOperand(0);
8090 SDValue AddeOp1 = AddeNode->getOperand(1);
8092 // Make sure that the ADDE operands are not coming from the same node.
8093 if (AddeOp0.getNode() == AddeOp1.getNode())
8096 // Find the MUL_LOHI node walking up ADDE's operands.
8097 bool IsLeftOperandMUL = false;
8098 SDValue MULOp = findMUL_LOHI(AddeOp0);
8099 if (MULOp == SDValue())
8100 MULOp = findMUL_LOHI(AddeOp1);
8102 IsLeftOperandMUL = true;
8103 if (MULOp == SDValue())
8106 // Figure out the right opcode.
8107 unsigned Opc = MULOp->getOpcode();
8108 unsigned FinalOpc = (Opc == ISD::SMUL_LOHI) ? ARMISD::SMLAL : ARMISD::UMLAL;
8110 // Figure out the high and low input values to the MLAL node.
8111 SDValue* HiAdd = nullptr;
8112 SDValue* LoMul = nullptr;
8113 SDValue* LowAdd = nullptr;
8115 // Ensure that ADDE is from high result of ISD::SMUL_LOHI.
8116 if ((AddeOp0 != MULOp.getValue(1)) && (AddeOp1 != MULOp.getValue(1)))
8119 if (IsLeftOperandMUL)
8125 // Ensure that LoMul and LowAdd are taken from correct ISD::SMUL_LOHI node
8126 // whose low result is fed to the ADDC we are checking.
8128 if (AddcOp0 == MULOp.getValue(0)) {
8132 if (AddcOp1 == MULOp.getValue(0)) {
8140 // Create the merged node.
8141 SelectionDAG &DAG = DCI.DAG;
8143 // Build operand list.
8144 SmallVector<SDValue, 8> Ops;
8145 Ops.push_back(LoMul->getOperand(0));
8146 Ops.push_back(LoMul->getOperand(1));
8147 Ops.push_back(*LowAdd);
8148 Ops.push_back(*HiAdd);
8150 SDValue MLALNode = DAG.getNode(FinalOpc, SDLoc(AddcNode),
8151 DAG.getVTList(MVT::i32, MVT::i32), Ops);
8153 // Replace the ADDs' nodes uses by the MLA node's values.
8154 SDValue HiMLALResult(MLALNode.getNode(), 1);
8155 DAG.ReplaceAllUsesOfValueWith(SDValue(AddeNode, 0), HiMLALResult);
8157 SDValue LoMLALResult(MLALNode.getNode(), 0);
8158 DAG.ReplaceAllUsesOfValueWith(SDValue(AddcNode, 0), LoMLALResult);
8160 // Return original node to notify the driver to stop replacing.
8161 SDValue resNode(AddcNode, 0);
8165 /// PerformADDCCombine - Target-specific dag combine transform from
8166 /// ISD::ADDC, ISD::ADDE, and ISD::MUL_LOHI to MLAL.
8167 static SDValue PerformADDCCombine(SDNode *N,
8168 TargetLowering::DAGCombinerInfo &DCI,
8169 const ARMSubtarget *Subtarget) {
8171 return AddCombineTo64bitMLAL(N, DCI, Subtarget);
8175 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
8176 /// operands N0 and N1. This is a helper for PerformADDCombine that is
8177 /// called with the default operands, and if that fails, with commuted
8179 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
8180 TargetLowering::DAGCombinerInfo &DCI,
8181 const ARMSubtarget *Subtarget){
8183 // Attempt to create vpaddl for this add.
8184 SDValue Result = AddCombineToVPADDL(N, N0, N1, DCI, Subtarget);
8185 if (Result.getNode())
8188 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
8189 if (N0.getNode()->hasOneUse()) {
8190 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
8191 if (Result.getNode()) return Result;
8196 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
8198 static SDValue PerformADDCombine(SDNode *N,
8199 TargetLowering::DAGCombinerInfo &DCI,
8200 const ARMSubtarget *Subtarget) {
8201 SDValue N0 = N->getOperand(0);
8202 SDValue N1 = N->getOperand(1);
8204 // First try with the default operand order.
8205 SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI, Subtarget);
8206 if (Result.getNode())
8209 // If that didn't work, try again with the operands commuted.
8210 return PerformADDCombineWithOperands(N, N1, N0, DCI, Subtarget);
8213 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
8215 static SDValue PerformSUBCombine(SDNode *N,
8216 TargetLowering::DAGCombinerInfo &DCI) {
8217 SDValue N0 = N->getOperand(0);
8218 SDValue N1 = N->getOperand(1);
8220 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
8221 if (N1.getNode()->hasOneUse()) {
8222 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
8223 if (Result.getNode()) return Result;
8229 /// PerformVMULCombine
8230 /// Distribute (A + B) * C to (A * C) + (B * C) to take advantage of the
8231 /// special multiplier accumulator forwarding.
8237 // However, for (A + B) * (A + B),
8244 static SDValue PerformVMULCombine(SDNode *N,
8245 TargetLowering::DAGCombinerInfo &DCI,
8246 const ARMSubtarget *Subtarget) {
8247 if (!Subtarget->hasVMLxForwarding())
8250 SelectionDAG &DAG = DCI.DAG;
8251 SDValue N0 = N->getOperand(0);
8252 SDValue N1 = N->getOperand(1);
8253 unsigned Opcode = N0.getOpcode();
8254 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
8255 Opcode != ISD::FADD && Opcode != ISD::FSUB) {
8256 Opcode = N1.getOpcode();
8257 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
8258 Opcode != ISD::FADD && Opcode != ISD::FSUB)
8266 EVT VT = N->getValueType(0);
8268 SDValue N00 = N0->getOperand(0);
8269 SDValue N01 = N0->getOperand(1);
8270 return DAG.getNode(Opcode, DL, VT,
8271 DAG.getNode(ISD::MUL, DL, VT, N00, N1),
8272 DAG.getNode(ISD::MUL, DL, VT, N01, N1));
8275 static SDValue PerformMULCombine(SDNode *N,
8276 TargetLowering::DAGCombinerInfo &DCI,
8277 const ARMSubtarget *Subtarget) {
8278 SelectionDAG &DAG = DCI.DAG;
8280 if (Subtarget->isThumb1Only())
8283 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
8286 EVT VT = N->getValueType(0);
8287 if (VT.is64BitVector() || VT.is128BitVector())
8288 return PerformVMULCombine(N, DCI, Subtarget);
8292 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
8296 int64_t MulAmt = C->getSExtValue();
8297 unsigned ShiftAmt = countTrailingZeros<uint64_t>(MulAmt);
8299 ShiftAmt = ShiftAmt & (32 - 1);
8300 SDValue V = N->getOperand(0);
8304 MulAmt >>= ShiftAmt;
8307 if (isPowerOf2_32(MulAmt - 1)) {
8308 // (mul x, 2^N + 1) => (add (shl x, N), x)
8309 Res = DAG.getNode(ISD::ADD, DL, VT,
8311 DAG.getNode(ISD::SHL, DL, VT,
8313 DAG.getConstant(Log2_32(MulAmt - 1), DL,
8315 } else if (isPowerOf2_32(MulAmt + 1)) {
8316 // (mul x, 2^N - 1) => (sub (shl x, N), x)
8317 Res = DAG.getNode(ISD::SUB, DL, VT,
8318 DAG.getNode(ISD::SHL, DL, VT,
8320 DAG.getConstant(Log2_32(MulAmt + 1), DL,
8326 uint64_t MulAmtAbs = -MulAmt;
8327 if (isPowerOf2_32(MulAmtAbs + 1)) {
8328 // (mul x, -(2^N - 1)) => (sub x, (shl x, N))
8329 Res = DAG.getNode(ISD::SUB, DL, VT,
8331 DAG.getNode(ISD::SHL, DL, VT,
8333 DAG.getConstant(Log2_32(MulAmtAbs + 1), DL,
8335 } else if (isPowerOf2_32(MulAmtAbs - 1)) {
8336 // (mul x, -(2^N + 1)) => - (add (shl x, N), x)
8337 Res = DAG.getNode(ISD::ADD, DL, VT,
8339 DAG.getNode(ISD::SHL, DL, VT,
8341 DAG.getConstant(Log2_32(MulAmtAbs - 1), DL,
8343 Res = DAG.getNode(ISD::SUB, DL, VT,
8344 DAG.getConstant(0, DL, MVT::i32), Res);
8351 Res = DAG.getNode(ISD::SHL, DL, VT,
8352 Res, DAG.getConstant(ShiftAmt, DL, MVT::i32));
8354 // Do not add new nodes to DAG combiner worklist.
8355 DCI.CombineTo(N, Res, false);
8359 static SDValue PerformANDCombine(SDNode *N,
8360 TargetLowering::DAGCombinerInfo &DCI,
8361 const ARMSubtarget *Subtarget) {
8363 // Attempt to use immediate-form VBIC
8364 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
8366 EVT VT = N->getValueType(0);
8367 SelectionDAG &DAG = DCI.DAG;
8369 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8372 APInt SplatBits, SplatUndef;
8373 unsigned SplatBitSize;
8376 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
8377 if (SplatBitSize <= 64) {
8379 SDValue Val = isNEONModifiedImm((~SplatBits).getZExtValue(),
8380 SplatUndef.getZExtValue(), SplatBitSize,
8381 DAG, dl, VbicVT, VT.is128BitVector(),
8383 if (Val.getNode()) {
8385 DAG.getNode(ISD::BITCAST, dl, VbicVT, N->getOperand(0));
8386 SDValue Vbic = DAG.getNode(ARMISD::VBICIMM, dl, VbicVT, Input, Val);
8387 return DAG.getNode(ISD::BITCAST, dl, VT, Vbic);
8392 if (!Subtarget->isThumb1Only()) {
8393 // fold (and (select cc, -1, c), x) -> (select cc, x, (and, x, c))
8394 SDValue Result = combineSelectAndUseCommutative(N, true, DCI);
8395 if (Result.getNode())
8402 /// PerformORCombine - Target-specific dag combine xforms for ISD::OR
8403 static SDValue PerformORCombine(SDNode *N,
8404 TargetLowering::DAGCombinerInfo &DCI,
8405 const ARMSubtarget *Subtarget) {
8406 // Attempt to use immediate-form VORR
8407 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
8409 EVT VT = N->getValueType(0);
8410 SelectionDAG &DAG = DCI.DAG;
8412 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8415 APInt SplatBits, SplatUndef;
8416 unsigned SplatBitSize;
8418 if (BVN && Subtarget->hasNEON() &&
8419 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
8420 if (SplatBitSize <= 64) {
8422 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
8423 SplatUndef.getZExtValue(), SplatBitSize,
8424 DAG, dl, VorrVT, VT.is128BitVector(),
8426 if (Val.getNode()) {
8428 DAG.getNode(ISD::BITCAST, dl, VorrVT, N->getOperand(0));
8429 SDValue Vorr = DAG.getNode(ARMISD::VORRIMM, dl, VorrVT, Input, Val);
8430 return DAG.getNode(ISD::BITCAST, dl, VT, Vorr);
8435 if (!Subtarget->isThumb1Only()) {
8436 // fold (or (select cc, 0, c), x) -> (select cc, x, (or, x, c))
8437 SDValue Result = combineSelectAndUseCommutative(N, false, DCI);
8438 if (Result.getNode())
8442 // The code below optimizes (or (and X, Y), Z).
8443 // The AND operand needs to have a single user to make these optimizations
8445 SDValue N0 = N->getOperand(0);
8446 if (N0.getOpcode() != ISD::AND || !N0.hasOneUse())
8448 SDValue N1 = N->getOperand(1);
8450 // (or (and B, A), (and C, ~A)) => (VBSL A, B, C) when A is a constant.
8451 if (Subtarget->hasNEON() && N1.getOpcode() == ISD::AND && VT.isVector() &&
8452 DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
8454 unsigned SplatBitSize;
8457 APInt SplatBits0, SplatBits1;
8458 BuildVectorSDNode *BVN0 = dyn_cast<BuildVectorSDNode>(N0->getOperand(1));
8459 BuildVectorSDNode *BVN1 = dyn_cast<BuildVectorSDNode>(N1->getOperand(1));
8460 // Ensure that the second operand of both ands are constants
8461 if (BVN0 && BVN0->isConstantSplat(SplatBits0, SplatUndef, SplatBitSize,
8462 HasAnyUndefs) && !HasAnyUndefs) {
8463 if (BVN1 && BVN1->isConstantSplat(SplatBits1, SplatUndef, SplatBitSize,
8464 HasAnyUndefs) && !HasAnyUndefs) {
8465 // Ensure that the bit width of the constants are the same and that
8466 // the splat arguments are logical inverses as per the pattern we
8467 // are trying to simplify.
8468 if (SplatBits0.getBitWidth() == SplatBits1.getBitWidth() &&
8469 SplatBits0 == ~SplatBits1) {
8470 // Canonicalize the vector type to make instruction selection
8472 EVT CanonicalVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
8473 SDValue Result = DAG.getNode(ARMISD::VBSL, dl, CanonicalVT,
8477 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
8483 // Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when
8486 // BFI is only available on V6T2+
8487 if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops())
8491 // 1) or (and A, mask), val => ARMbfi A, val, mask
8492 // iff (val & mask) == val
8494 // 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
8495 // 2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2)
8496 // && mask == ~mask2
8497 // 2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2)
8498 // && ~mask == mask2
8499 // (i.e., copy a bitfield value into another bitfield of the same width)
8504 SDValue N00 = N0.getOperand(0);
8506 // The value and the mask need to be constants so we can verify this is
8507 // actually a bitfield set. If the mask is 0xffff, we can do better
8508 // via a movt instruction, so don't use BFI in that case.
8509 SDValue MaskOp = N0.getOperand(1);
8510 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(MaskOp);
8513 unsigned Mask = MaskC->getZExtValue();
8517 // Case (1): or (and A, mask), val => ARMbfi A, val, mask
8518 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
8520 unsigned Val = N1C->getZExtValue();
8521 if ((Val & ~Mask) != Val)
8524 if (ARM::isBitFieldInvertedMask(Mask)) {
8525 Val >>= countTrailingZeros(~Mask);
8527 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00,
8528 DAG.getConstant(Val, DL, MVT::i32),
8529 DAG.getConstant(Mask, DL, MVT::i32));
8531 // Do not add new nodes to DAG combiner worklist.
8532 DCI.CombineTo(N, Res, false);
8535 } else if (N1.getOpcode() == ISD::AND) {
8536 // case (2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
8537 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
8540 unsigned Mask2 = N11C->getZExtValue();
8542 // Mask and ~Mask2 (or reverse) must be equivalent for the BFI pattern
8544 if (ARM::isBitFieldInvertedMask(Mask) &&
8546 // The pack halfword instruction works better for masks that fit it,
8547 // so use that when it's available.
8548 if (Subtarget->hasT2ExtractPack() &&
8549 (Mask == 0xffff || Mask == 0xffff0000))
8552 unsigned amt = countTrailingZeros(Mask2);
8553 Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0),
8554 DAG.getConstant(amt, DL, MVT::i32));
8555 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00, Res,
8556 DAG.getConstant(Mask, DL, MVT::i32));
8557 // Do not add new nodes to DAG combiner worklist.
8558 DCI.CombineTo(N, Res, false);
8560 } else if (ARM::isBitFieldInvertedMask(~Mask) &&
8562 // The pack halfword instruction works better for masks that fit it,
8563 // so use that when it's available.
8564 if (Subtarget->hasT2ExtractPack() &&
8565 (Mask2 == 0xffff || Mask2 == 0xffff0000))
8568 unsigned lsb = countTrailingZeros(Mask);
8569 Res = DAG.getNode(ISD::SRL, DL, VT, N00,
8570 DAG.getConstant(lsb, DL, MVT::i32));
8571 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res,
8572 DAG.getConstant(Mask2, DL, MVT::i32));
8573 // Do not add new nodes to DAG combiner worklist.
8574 DCI.CombineTo(N, Res, false);
8579 if (DAG.MaskedValueIsZero(N1, MaskC->getAPIntValue()) &&
8580 N00.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N00.getOperand(1)) &&
8581 ARM::isBitFieldInvertedMask(~Mask)) {
8582 // Case (3): or (and (shl A, #shamt), mask), B => ARMbfi B, A, ~mask
8583 // where lsb(mask) == #shamt and masked bits of B are known zero.
8584 SDValue ShAmt = N00.getOperand(1);
8585 unsigned ShAmtC = cast<ConstantSDNode>(ShAmt)->getZExtValue();
8586 unsigned LSB = countTrailingZeros(Mask);
8590 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1, N00.getOperand(0),
8591 DAG.getConstant(~Mask, DL, MVT::i32));
8593 // Do not add new nodes to DAG combiner worklist.
8594 DCI.CombineTo(N, Res, false);
8600 static SDValue PerformXORCombine(SDNode *N,
8601 TargetLowering::DAGCombinerInfo &DCI,
8602 const ARMSubtarget *Subtarget) {
8603 EVT VT = N->getValueType(0);
8604 SelectionDAG &DAG = DCI.DAG;
8606 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8609 if (!Subtarget->isThumb1Only()) {
8610 // fold (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c))
8611 SDValue Result = combineSelectAndUseCommutative(N, false, DCI);
8612 if (Result.getNode())
8619 /// PerformBFICombine - (bfi A, (and B, Mask1), Mask2) -> (bfi A, B, Mask2) iff
8620 /// the bits being cleared by the AND are not demanded by the BFI.
8621 static SDValue PerformBFICombine(SDNode *N,
8622 TargetLowering::DAGCombinerInfo &DCI) {
8623 SDValue N1 = N->getOperand(1);
8624 if (N1.getOpcode() == ISD::AND) {
8625 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
8628 unsigned InvMask = cast<ConstantSDNode>(N->getOperand(2))->getZExtValue();
8629 unsigned LSB = countTrailingZeros(~InvMask);
8630 unsigned Width = (32 - countLeadingZeros(~InvMask)) - LSB;
8632 static_cast<unsigned>(std::numeric_limits<unsigned>::digits) &&
8633 "undefined behavior");
8634 unsigned Mask = (1u << Width) - 1;
8635 unsigned Mask2 = N11C->getZExtValue();
8636 if ((Mask & (~Mask2)) == 0)
8637 return DCI.DAG.getNode(ARMISD::BFI, SDLoc(N), N->getValueType(0),
8638 N->getOperand(0), N1.getOperand(0),
8644 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
8645 /// ARMISD::VMOVRRD.
8646 static SDValue PerformVMOVRRDCombine(SDNode *N,
8647 TargetLowering::DAGCombinerInfo &DCI,
8648 const ARMSubtarget *Subtarget) {
8649 // vmovrrd(vmovdrr x, y) -> x,y
8650 SDValue InDouble = N->getOperand(0);
8651 if (InDouble.getOpcode() == ARMISD::VMOVDRR && !Subtarget->isFPOnlySP())
8652 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
8654 // vmovrrd(load f64) -> (load i32), (load i32)
8655 SDNode *InNode = InDouble.getNode();
8656 if (ISD::isNormalLoad(InNode) && InNode->hasOneUse() &&
8657 InNode->getValueType(0) == MVT::f64 &&
8658 InNode->getOperand(1).getOpcode() == ISD::FrameIndex &&
8659 !cast<LoadSDNode>(InNode)->isVolatile()) {
8660 // TODO: Should this be done for non-FrameIndex operands?
8661 LoadSDNode *LD = cast<LoadSDNode>(InNode);
8663 SelectionDAG &DAG = DCI.DAG;
8665 SDValue BasePtr = LD->getBasePtr();
8666 SDValue NewLD1 = DAG.getLoad(MVT::i32, DL, LD->getChain(), BasePtr,
8667 LD->getPointerInfo(), LD->isVolatile(),
8668 LD->isNonTemporal(), LD->isInvariant(),
8669 LD->getAlignment());
8671 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
8672 DAG.getConstant(4, DL, MVT::i32));
8673 SDValue NewLD2 = DAG.getLoad(MVT::i32, DL, NewLD1.getValue(1), OffsetPtr,
8674 LD->getPointerInfo(), LD->isVolatile(),
8675 LD->isNonTemporal(), LD->isInvariant(),
8676 std::min(4U, LD->getAlignment() / 2));
8678 DAG.ReplaceAllUsesOfValueWith(SDValue(LD, 1), NewLD2.getValue(1));
8679 if (DCI.DAG.getDataLayout().isBigEndian())
8680 std::swap (NewLD1, NewLD2);
8681 SDValue Result = DCI.CombineTo(N, NewLD1, NewLD2);
8688 /// PerformVMOVDRRCombine - Target-specific dag combine xforms for
8689 /// ARMISD::VMOVDRR. This is also used for BUILD_VECTORs with 2 operands.
8690 static SDValue PerformVMOVDRRCombine(SDNode *N, SelectionDAG &DAG) {
8691 // N=vmovrrd(X); vmovdrr(N:0, N:1) -> bit_convert(X)
8692 SDValue Op0 = N->getOperand(0);
8693 SDValue Op1 = N->getOperand(1);
8694 if (Op0.getOpcode() == ISD::BITCAST)
8695 Op0 = Op0.getOperand(0);
8696 if (Op1.getOpcode() == ISD::BITCAST)
8697 Op1 = Op1.getOperand(0);
8698 if (Op0.getOpcode() == ARMISD::VMOVRRD &&
8699 Op0.getNode() == Op1.getNode() &&
8700 Op0.getResNo() == 0 && Op1.getResNo() == 1)
8701 return DAG.getNode(ISD::BITCAST, SDLoc(N),
8702 N->getValueType(0), Op0.getOperand(0));
8706 /// hasNormalLoadOperand - Check if any of the operands of a BUILD_VECTOR node
8707 /// are normal, non-volatile loads. If so, it is profitable to bitcast an
8708 /// i64 vector to have f64 elements, since the value can then be loaded
8709 /// directly into a VFP register.
8710 static bool hasNormalLoadOperand(SDNode *N) {
8711 unsigned NumElts = N->getValueType(0).getVectorNumElements();
8712 for (unsigned i = 0; i < NumElts; ++i) {
8713 SDNode *Elt = N->getOperand(i).getNode();
8714 if (ISD::isNormalLoad(Elt) && !cast<LoadSDNode>(Elt)->isVolatile())
8720 /// PerformBUILD_VECTORCombine - Target-specific dag combine xforms for
8721 /// ISD::BUILD_VECTOR.
8722 static SDValue PerformBUILD_VECTORCombine(SDNode *N,
8723 TargetLowering::DAGCombinerInfo &DCI,
8724 const ARMSubtarget *Subtarget) {
8725 // build_vector(N=ARMISD::VMOVRRD(X), N:1) -> bit_convert(X):
8726 // VMOVRRD is introduced when legalizing i64 types. It forces the i64 value
8727 // into a pair of GPRs, which is fine when the value is used as a scalar,
8728 // but if the i64 value is converted to a vector, we need to undo the VMOVRRD.
8729 SelectionDAG &DAG = DCI.DAG;
8730 if (N->getNumOperands() == 2) {
8731 SDValue RV = PerformVMOVDRRCombine(N, DAG);
8736 // Load i64 elements as f64 values so that type legalization does not split
8737 // them up into i32 values.
8738 EVT VT = N->getValueType(0);
8739 if (VT.getVectorElementType() != MVT::i64 || !hasNormalLoadOperand(N))
8742 SmallVector<SDValue, 8> Ops;
8743 unsigned NumElts = VT.getVectorNumElements();
8744 for (unsigned i = 0; i < NumElts; ++i) {
8745 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(i));
8747 // Make the DAGCombiner fold the bitcast.
8748 DCI.AddToWorklist(V.getNode());
8750 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, NumElts);
8751 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, FloatVT, Ops);
8752 return DAG.getNode(ISD::BITCAST, dl, VT, BV);
8755 /// \brief Target-specific dag combine xforms for ARMISD::BUILD_VECTOR.
8757 PerformARMBUILD_VECTORCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
8758 // ARMISD::BUILD_VECTOR is introduced when legalizing ISD::BUILD_VECTOR.
8759 // At that time, we may have inserted bitcasts from integer to float.
8760 // If these bitcasts have survived DAGCombine, change the lowering of this
8761 // BUILD_VECTOR in something more vector friendly, i.e., that does not
8762 // force to use floating point types.
8764 // Make sure we can change the type of the vector.
8765 // This is possible iff:
8766 // 1. The vector is only used in a bitcast to a integer type. I.e.,
8767 // 1.1. Vector is used only once.
8768 // 1.2. Use is a bit convert to an integer type.
8769 // 2. The size of its operands are 32-bits (64-bits are not legal).
8770 EVT VT = N->getValueType(0);
8771 EVT EltVT = VT.getVectorElementType();
8773 // Check 1.1. and 2.
8774 if (EltVT.getSizeInBits() != 32 || !N->hasOneUse())
8777 // By construction, the input type must be float.
8778 assert(EltVT == MVT::f32 && "Unexpected type!");
8781 SDNode *Use = *N->use_begin();
8782 if (Use->getOpcode() != ISD::BITCAST ||
8783 Use->getValueType(0).isFloatingPoint())
8786 // Check profitability.
8787 // Model is, if more than half of the relevant operands are bitcast from
8788 // i32, turn the build_vector into a sequence of insert_vector_elt.
8789 // Relevant operands are everything that is not statically
8790 // (i.e., at compile time) bitcasted.
8791 unsigned NumOfBitCastedElts = 0;
8792 unsigned NumElts = VT.getVectorNumElements();
8793 unsigned NumOfRelevantElts = NumElts;
8794 for (unsigned Idx = 0; Idx < NumElts; ++Idx) {
8795 SDValue Elt = N->getOperand(Idx);
8796 if (Elt->getOpcode() == ISD::BITCAST) {
8797 // Assume only bit cast to i32 will go away.
8798 if (Elt->getOperand(0).getValueType() == MVT::i32)
8799 ++NumOfBitCastedElts;
8800 } else if (Elt.getOpcode() == ISD::UNDEF || isa<ConstantSDNode>(Elt))
8801 // Constants are statically casted, thus do not count them as
8802 // relevant operands.
8803 --NumOfRelevantElts;
8806 // Check if more than half of the elements require a non-free bitcast.
8807 if (NumOfBitCastedElts <= NumOfRelevantElts / 2)
8810 SelectionDAG &DAG = DCI.DAG;
8811 // Create the new vector type.
8812 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
8813 // Check if the type is legal.
8814 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8815 if (!TLI.isTypeLegal(VecVT))
8819 // ARMISD::BUILD_VECTOR E1, E2, ..., EN.
8820 // => BITCAST INSERT_VECTOR_ELT
8821 // (INSERT_VECTOR_ELT (...), (BITCAST EN-1), N-1),
8823 SDValue Vec = DAG.getUNDEF(VecVT);
8825 for (unsigned Idx = 0 ; Idx < NumElts; ++Idx) {
8826 SDValue V = N->getOperand(Idx);
8827 if (V.getOpcode() == ISD::UNDEF)
8829 if (V.getOpcode() == ISD::BITCAST &&
8830 V->getOperand(0).getValueType() == MVT::i32)
8831 // Fold obvious case.
8832 V = V.getOperand(0);
8834 V = DAG.getNode(ISD::BITCAST, SDLoc(V), MVT::i32, V);
8835 // Make the DAGCombiner fold the bitcasts.
8836 DCI.AddToWorklist(V.getNode());
8838 SDValue LaneIdx = DAG.getConstant(Idx, dl, MVT::i32);
8839 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VecVT, Vec, V, LaneIdx);
8841 Vec = DAG.getNode(ISD::BITCAST, dl, VT, Vec);
8842 // Make the DAGCombiner fold the bitcasts.
8843 DCI.AddToWorklist(Vec.getNode());
8847 /// PerformInsertEltCombine - Target-specific dag combine xforms for
8848 /// ISD::INSERT_VECTOR_ELT.
8849 static SDValue PerformInsertEltCombine(SDNode *N,
8850 TargetLowering::DAGCombinerInfo &DCI) {
8851 // Bitcast an i64 load inserted into a vector to f64.
8852 // Otherwise, the i64 value will be legalized to a pair of i32 values.
8853 EVT VT = N->getValueType(0);
8854 SDNode *Elt = N->getOperand(1).getNode();
8855 if (VT.getVectorElementType() != MVT::i64 ||
8856 !ISD::isNormalLoad(Elt) || cast<LoadSDNode>(Elt)->isVolatile())
8859 SelectionDAG &DAG = DCI.DAG;
8861 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
8862 VT.getVectorNumElements());
8863 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, N->getOperand(0));
8864 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(1));
8865 // Make the DAGCombiner fold the bitcasts.
8866 DCI.AddToWorklist(Vec.getNode());
8867 DCI.AddToWorklist(V.getNode());
8868 SDValue InsElt = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, FloatVT,
8869 Vec, V, N->getOperand(2));
8870 return DAG.getNode(ISD::BITCAST, dl, VT, InsElt);
8873 /// PerformVECTOR_SHUFFLECombine - Target-specific dag combine xforms for
8874 /// ISD::VECTOR_SHUFFLE.
8875 static SDValue PerformVECTOR_SHUFFLECombine(SDNode *N, SelectionDAG &DAG) {
8876 // The LLVM shufflevector instruction does not require the shuffle mask
8877 // length to match the operand vector length, but ISD::VECTOR_SHUFFLE does
8878 // have that requirement. When translating to ISD::VECTOR_SHUFFLE, if the
8879 // operands do not match the mask length, they are extended by concatenating
8880 // them with undef vectors. That is probably the right thing for other
8881 // targets, but for NEON it is better to concatenate two double-register
8882 // size vector operands into a single quad-register size vector. Do that
8883 // transformation here:
8884 // shuffle(concat(v1, undef), concat(v2, undef)) ->
8885 // shuffle(concat(v1, v2), undef)
8886 SDValue Op0 = N->getOperand(0);
8887 SDValue Op1 = N->getOperand(1);
8888 if (Op0.getOpcode() != ISD::CONCAT_VECTORS ||
8889 Op1.getOpcode() != ISD::CONCAT_VECTORS ||
8890 Op0.getNumOperands() != 2 ||
8891 Op1.getNumOperands() != 2)
8893 SDValue Concat0Op1 = Op0.getOperand(1);
8894 SDValue Concat1Op1 = Op1.getOperand(1);
8895 if (Concat0Op1.getOpcode() != ISD::UNDEF ||
8896 Concat1Op1.getOpcode() != ISD::UNDEF)
8898 // Skip the transformation if any of the types are illegal.
8899 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8900 EVT VT = N->getValueType(0);
8901 if (!TLI.isTypeLegal(VT) ||
8902 !TLI.isTypeLegal(Concat0Op1.getValueType()) ||
8903 !TLI.isTypeLegal(Concat1Op1.getValueType()))
8906 SDValue NewConcat = DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT,
8907 Op0.getOperand(0), Op1.getOperand(0));
8908 // Translate the shuffle mask.
8909 SmallVector<int, 16> NewMask;
8910 unsigned NumElts = VT.getVectorNumElements();
8911 unsigned HalfElts = NumElts/2;
8912 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
8913 for (unsigned n = 0; n < NumElts; ++n) {
8914 int MaskElt = SVN->getMaskElt(n);
8916 if (MaskElt < (int)HalfElts)
8918 else if (MaskElt >= (int)NumElts && MaskElt < (int)(NumElts + HalfElts))
8919 NewElt = HalfElts + MaskElt - NumElts;
8920 NewMask.push_back(NewElt);
8922 return DAG.getVectorShuffle(VT, SDLoc(N), NewConcat,
8923 DAG.getUNDEF(VT), NewMask.data());
8926 /// CombineBaseUpdate - Target-specific DAG combine function for VLDDUP,
8927 /// NEON load/store intrinsics, and generic vector load/stores, to merge
8928 /// base address updates.
8929 /// For generic load/stores, the memory type is assumed to be a vector.
8930 /// The caller is assumed to have checked legality.
8931 static SDValue CombineBaseUpdate(SDNode *N,
8932 TargetLowering::DAGCombinerInfo &DCI) {
8933 SelectionDAG &DAG = DCI.DAG;
8934 const bool isIntrinsic = (N->getOpcode() == ISD::INTRINSIC_VOID ||
8935 N->getOpcode() == ISD::INTRINSIC_W_CHAIN);
8936 const bool isStore = N->getOpcode() == ISD::STORE;
8937 const unsigned AddrOpIdx = ((isIntrinsic || isStore) ? 2 : 1);
8938 SDValue Addr = N->getOperand(AddrOpIdx);
8939 MemSDNode *MemN = cast<MemSDNode>(N);
8942 // Search for a use of the address operand that is an increment.
8943 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(),
8944 UE = Addr.getNode()->use_end(); UI != UE; ++UI) {
8946 if (User->getOpcode() != ISD::ADD ||
8947 UI.getUse().getResNo() != Addr.getResNo())
8950 // Check that the add is independent of the load/store. Otherwise, folding
8951 // it would create a cycle.
8952 if (User->isPredecessorOf(N) || N->isPredecessorOf(User))
8955 // Find the new opcode for the updating load/store.
8956 bool isLoadOp = true;
8957 bool isLaneOp = false;
8958 unsigned NewOpc = 0;
8959 unsigned NumVecs = 0;
8961 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
8963 default: llvm_unreachable("unexpected intrinsic for Neon base update");
8964 case Intrinsic::arm_neon_vld1: NewOpc = ARMISD::VLD1_UPD;
8966 case Intrinsic::arm_neon_vld2: NewOpc = ARMISD::VLD2_UPD;
8968 case Intrinsic::arm_neon_vld3: NewOpc = ARMISD::VLD3_UPD;
8970 case Intrinsic::arm_neon_vld4: NewOpc = ARMISD::VLD4_UPD;
8972 case Intrinsic::arm_neon_vld2lane: NewOpc = ARMISD::VLD2LN_UPD;
8973 NumVecs = 2; isLaneOp = true; break;
8974 case Intrinsic::arm_neon_vld3lane: NewOpc = ARMISD::VLD3LN_UPD;
8975 NumVecs = 3; isLaneOp = true; break;
8976 case Intrinsic::arm_neon_vld4lane: NewOpc = ARMISD::VLD4LN_UPD;
8977 NumVecs = 4; isLaneOp = true; break;
8978 case Intrinsic::arm_neon_vst1: NewOpc = ARMISD::VST1_UPD;
8979 NumVecs = 1; isLoadOp = false; break;
8980 case Intrinsic::arm_neon_vst2: NewOpc = ARMISD::VST2_UPD;
8981 NumVecs = 2; isLoadOp = false; break;
8982 case Intrinsic::arm_neon_vst3: NewOpc = ARMISD::VST3_UPD;
8983 NumVecs = 3; isLoadOp = false; break;
8984 case Intrinsic::arm_neon_vst4: NewOpc = ARMISD::VST4_UPD;
8985 NumVecs = 4; isLoadOp = false; break;
8986 case Intrinsic::arm_neon_vst2lane: NewOpc = ARMISD::VST2LN_UPD;
8987 NumVecs = 2; isLoadOp = false; isLaneOp = true; break;
8988 case Intrinsic::arm_neon_vst3lane: NewOpc = ARMISD::VST3LN_UPD;
8989 NumVecs = 3; isLoadOp = false; isLaneOp = true; break;
8990 case Intrinsic::arm_neon_vst4lane: NewOpc = ARMISD::VST4LN_UPD;
8991 NumVecs = 4; isLoadOp = false; isLaneOp = true; break;
8995 switch (N->getOpcode()) {
8996 default: llvm_unreachable("unexpected opcode for Neon base update");
8997 case ARMISD::VLD2DUP: NewOpc = ARMISD::VLD2DUP_UPD; NumVecs = 2; break;
8998 case ARMISD::VLD3DUP: NewOpc = ARMISD::VLD3DUP_UPD; NumVecs = 3; break;
8999 case ARMISD::VLD4DUP: NewOpc = ARMISD::VLD4DUP_UPD; NumVecs = 4; break;
9000 case ISD::LOAD: NewOpc = ARMISD::VLD1_UPD;
9001 NumVecs = 1; isLaneOp = false; break;
9002 case ISD::STORE: NewOpc = ARMISD::VST1_UPD;
9003 NumVecs = 1; isLaneOp = false; isLoadOp = false; break;
9007 // Find the size of memory referenced by the load/store.
9010 VecTy = N->getValueType(0);
9011 } else if (isIntrinsic) {
9012 VecTy = N->getOperand(AddrOpIdx+1).getValueType();
9014 assert(isStore && "Node has to be a load, a store, or an intrinsic!");
9015 VecTy = N->getOperand(1).getValueType();
9018 unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8;
9020 NumBytes /= VecTy.getVectorNumElements();
9022 // If the increment is a constant, it must match the memory ref size.
9023 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
9024 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
9025 uint64_t IncVal = CInc->getZExtValue();
9026 if (IncVal != NumBytes)
9028 } else if (NumBytes >= 3 * 16) {
9029 // VLD3/4 and VST3/4 for 128-bit vectors are implemented with two
9030 // separate instructions that make it harder to use a non-constant update.
9034 // OK, we found an ADD we can fold into the base update.
9035 // Now, create a _UPD node, taking care of not breaking alignment.
9037 EVT AlignedVecTy = VecTy;
9038 unsigned Alignment = MemN->getAlignment();
9040 // If this is a less-than-standard-aligned load/store, change the type to
9041 // match the standard alignment.
9042 // The alignment is overlooked when selecting _UPD variants; and it's
9043 // easier to introduce bitcasts here than fix that.
9044 // There are 3 ways to get to this base-update combine:
9045 // - intrinsics: they are assumed to be properly aligned (to the standard
9046 // alignment of the memory type), so we don't need to do anything.
9047 // - ARMISD::VLDx nodes: they are only generated from the aforementioned
9048 // intrinsics, so, likewise, there's nothing to do.
9049 // - generic load/store instructions: the alignment is specified as an
9050 // explicit operand, rather than implicitly as the standard alignment
9051 // of the memory type (like the intrisics). We need to change the
9052 // memory type to match the explicit alignment. That way, we don't
9053 // generate non-standard-aligned ARMISD::VLDx nodes.
9054 if (isa<LSBaseSDNode>(N)) {
9057 if (Alignment < VecTy.getScalarSizeInBits() / 8) {
9058 MVT EltTy = MVT::getIntegerVT(Alignment * 8);
9059 assert(NumVecs == 1 && "Unexpected multi-element generic load/store.");
9060 assert(!isLaneOp && "Unexpected generic load/store lane.");
9061 unsigned NumElts = NumBytes / (EltTy.getSizeInBits() / 8);
9062 AlignedVecTy = MVT::getVectorVT(EltTy, NumElts);
9064 // Don't set an explicit alignment on regular load/stores that we want
9065 // to transform to VLD/VST 1_UPD nodes.
9066 // This matches the behavior of regular load/stores, which only get an
9067 // explicit alignment if the MMO alignment is larger than the standard
9068 // alignment of the memory type.
9069 // Intrinsics, however, always get an explicit alignment, set to the
9070 // alignment of the MMO.
9074 // Create the new updating load/store node.
9075 // First, create an SDVTList for the new updating node's results.
9077 unsigned NumResultVecs = (isLoadOp ? NumVecs : 0);
9079 for (n = 0; n < NumResultVecs; ++n)
9080 Tys[n] = AlignedVecTy;
9081 Tys[n++] = MVT::i32;
9082 Tys[n] = MVT::Other;
9083 SDVTList SDTys = DAG.getVTList(makeArrayRef(Tys, NumResultVecs+2));
9085 // Then, gather the new node's operands.
9086 SmallVector<SDValue, 8> Ops;
9087 Ops.push_back(N->getOperand(0)); // incoming chain
9088 Ops.push_back(N->getOperand(AddrOpIdx));
9091 if (StoreSDNode *StN = dyn_cast<StoreSDNode>(N)) {
9092 // Try to match the intrinsic's signature
9093 Ops.push_back(StN->getValue());
9095 // Loads (and of course intrinsics) match the intrinsics' signature,
9096 // so just add all but the alignment operand.
9097 for (unsigned i = AddrOpIdx + 1; i < N->getNumOperands() - 1; ++i)
9098 Ops.push_back(N->getOperand(i));
9101 // For all node types, the alignment operand is always the last one.
9102 Ops.push_back(DAG.getConstant(Alignment, dl, MVT::i32));
9104 // If this is a non-standard-aligned STORE, the penultimate operand is the
9105 // stored value. Bitcast it to the aligned type.
9106 if (AlignedVecTy != VecTy && N->getOpcode() == ISD::STORE) {
9107 SDValue &StVal = Ops[Ops.size()-2];
9108 StVal = DAG.getNode(ISD::BITCAST, dl, AlignedVecTy, StVal);
9111 SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, dl, SDTys,
9113 MemN->getMemOperand());
9116 SmallVector<SDValue, 5> NewResults;
9117 for (unsigned i = 0; i < NumResultVecs; ++i)
9118 NewResults.push_back(SDValue(UpdN.getNode(), i));
9120 // If this is an non-standard-aligned LOAD, the first result is the loaded
9121 // value. Bitcast it to the expected result type.
9122 if (AlignedVecTy != VecTy && N->getOpcode() == ISD::LOAD) {
9123 SDValue &LdVal = NewResults[0];
9124 LdVal = DAG.getNode(ISD::BITCAST, dl, VecTy, LdVal);
9127 NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs+1)); // chain
9128 DCI.CombineTo(N, NewResults);
9129 DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs));
9136 static SDValue PerformVLDCombine(SDNode *N,
9137 TargetLowering::DAGCombinerInfo &DCI) {
9138 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
9141 return CombineBaseUpdate(N, DCI);
9144 /// CombineVLDDUP - For a VDUPLANE node N, check if its source operand is a
9145 /// vldN-lane (N > 1) intrinsic, and if all the other uses of that intrinsic
9146 /// are also VDUPLANEs. If so, combine them to a vldN-dup operation and
9148 static bool CombineVLDDUP(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
9149 SelectionDAG &DAG = DCI.DAG;
9150 EVT VT = N->getValueType(0);
9151 // vldN-dup instructions only support 64-bit vectors for N > 1.
9152 if (!VT.is64BitVector())
9155 // Check if the VDUPLANE operand is a vldN-dup intrinsic.
9156 SDNode *VLD = N->getOperand(0).getNode();
9157 if (VLD->getOpcode() != ISD::INTRINSIC_W_CHAIN)
9159 unsigned NumVecs = 0;
9160 unsigned NewOpc = 0;
9161 unsigned IntNo = cast<ConstantSDNode>(VLD->getOperand(1))->getZExtValue();
9162 if (IntNo == Intrinsic::arm_neon_vld2lane) {
9164 NewOpc = ARMISD::VLD2DUP;
9165 } else if (IntNo == Intrinsic::arm_neon_vld3lane) {
9167 NewOpc = ARMISD::VLD3DUP;
9168 } else if (IntNo == Intrinsic::arm_neon_vld4lane) {
9170 NewOpc = ARMISD::VLD4DUP;
9175 // First check that all the vldN-lane uses are VDUPLANEs and that the lane
9176 // numbers match the load.
9177 unsigned VLDLaneNo =
9178 cast<ConstantSDNode>(VLD->getOperand(NumVecs+3))->getZExtValue();
9179 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
9181 // Ignore uses of the chain result.
9182 if (UI.getUse().getResNo() == NumVecs)
9185 if (User->getOpcode() != ARMISD::VDUPLANE ||
9186 VLDLaneNo != cast<ConstantSDNode>(User->getOperand(1))->getZExtValue())
9190 // Create the vldN-dup node.
9193 for (n = 0; n < NumVecs; ++n)
9195 Tys[n] = MVT::Other;
9196 SDVTList SDTys = DAG.getVTList(makeArrayRef(Tys, NumVecs+1));
9197 SDValue Ops[] = { VLD->getOperand(0), VLD->getOperand(2) };
9198 MemIntrinsicSDNode *VLDMemInt = cast<MemIntrinsicSDNode>(VLD);
9199 SDValue VLDDup = DAG.getMemIntrinsicNode(NewOpc, SDLoc(VLD), SDTys,
9200 Ops, VLDMemInt->getMemoryVT(),
9201 VLDMemInt->getMemOperand());
9204 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
9206 unsigned ResNo = UI.getUse().getResNo();
9207 // Ignore uses of the chain result.
9208 if (ResNo == NumVecs)
9211 DCI.CombineTo(User, SDValue(VLDDup.getNode(), ResNo));
9214 // Now the vldN-lane intrinsic is dead except for its chain result.
9215 // Update uses of the chain.
9216 std::vector<SDValue> VLDDupResults;
9217 for (unsigned n = 0; n < NumVecs; ++n)
9218 VLDDupResults.push_back(SDValue(VLDDup.getNode(), n));
9219 VLDDupResults.push_back(SDValue(VLDDup.getNode(), NumVecs));
9220 DCI.CombineTo(VLD, VLDDupResults);
9225 /// PerformVDUPLANECombine - Target-specific dag combine xforms for
9226 /// ARMISD::VDUPLANE.
9227 static SDValue PerformVDUPLANECombine(SDNode *N,
9228 TargetLowering::DAGCombinerInfo &DCI) {
9229 SDValue Op = N->getOperand(0);
9231 // If the source is a vldN-lane (N > 1) intrinsic, and all the other uses
9232 // of that intrinsic are also VDUPLANEs, combine them to a vldN-dup operation.
9233 if (CombineVLDDUP(N, DCI))
9234 return SDValue(N, 0);
9236 // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is
9237 // redundant. Ignore bit_converts for now; element sizes are checked below.
9238 while (Op.getOpcode() == ISD::BITCAST)
9239 Op = Op.getOperand(0);
9240 if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM)
9243 // Make sure the VMOV element size is not bigger than the VDUPLANE elements.
9244 unsigned EltSize = Op.getValueType().getVectorElementType().getSizeInBits();
9245 // The canonical VMOV for a zero vector uses a 32-bit element size.
9246 unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
9248 if (ARM_AM::decodeNEONModImm(Imm, EltBits) == 0)
9250 EVT VT = N->getValueType(0);
9251 if (EltSize > VT.getVectorElementType().getSizeInBits())
9254 return DCI.DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Op);
9257 static SDValue PerformLOADCombine(SDNode *N,
9258 TargetLowering::DAGCombinerInfo &DCI) {
9259 EVT VT = N->getValueType(0);
9261 // If this is a legal vector load, try to combine it into a VLD1_UPD.
9262 if (ISD::isNormalLoad(N) && VT.isVector() &&
9263 DCI.DAG.getTargetLoweringInfo().isTypeLegal(VT))
9264 return CombineBaseUpdate(N, DCI);
9269 /// PerformSTORECombine - Target-specific dag combine xforms for
9271 static SDValue PerformSTORECombine(SDNode *N,
9272 TargetLowering::DAGCombinerInfo &DCI) {
9273 StoreSDNode *St = cast<StoreSDNode>(N);
9274 if (St->isVolatile())
9277 // Optimize trunc store (of multiple scalars) to shuffle and store. First,
9278 // pack all of the elements in one place. Next, store to memory in fewer
9280 SDValue StVal = St->getValue();
9281 EVT VT = StVal.getValueType();
9282 if (St->isTruncatingStore() && VT.isVector()) {
9283 SelectionDAG &DAG = DCI.DAG;
9284 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9285 EVT StVT = St->getMemoryVT();
9286 unsigned NumElems = VT.getVectorNumElements();
9287 assert(StVT != VT && "Cannot truncate to the same type");
9288 unsigned FromEltSz = VT.getVectorElementType().getSizeInBits();
9289 unsigned ToEltSz = StVT.getVectorElementType().getSizeInBits();
9291 // From, To sizes and ElemCount must be pow of two
9292 if (!isPowerOf2_32(NumElems * FromEltSz * ToEltSz)) return SDValue();
9294 // We are going to use the original vector elt for storing.
9295 // Accumulated smaller vector elements must be a multiple of the store size.
9296 if (0 != (NumElems * FromEltSz) % ToEltSz) return SDValue();
9298 unsigned SizeRatio = FromEltSz / ToEltSz;
9299 assert(SizeRatio * NumElems * ToEltSz == VT.getSizeInBits());
9301 // Create a type on which we perform the shuffle.
9302 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(), StVT.getScalarType(),
9303 NumElems*SizeRatio);
9304 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
9307 SDValue WideVec = DAG.getNode(ISD::BITCAST, DL, WideVecVT, StVal);
9308 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
9309 for (unsigned i = 0; i < NumElems; ++i)
9310 ShuffleVec[i] = DAG.getDataLayout().isBigEndian()
9311 ? (i + 1) * SizeRatio - 1
9314 // Can't shuffle using an illegal type.
9315 if (!TLI.isTypeLegal(WideVecVT)) return SDValue();
9317 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, DL, WideVec,
9318 DAG.getUNDEF(WideVec.getValueType()),
9320 // At this point all of the data is stored at the bottom of the
9321 // register. We now need to save it to mem.
9323 // Find the largest store unit
9324 MVT StoreType = MVT::i8;
9325 for (MVT Tp : MVT::integer_valuetypes()) {
9326 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToEltSz)
9329 // Didn't find a legal store type.
9330 if (!TLI.isTypeLegal(StoreType))
9333 // Bitcast the original vector into a vector of store-size units
9334 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
9335 StoreType, VT.getSizeInBits()/EVT(StoreType).getSizeInBits());
9336 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
9337 SDValue ShuffWide = DAG.getNode(ISD::BITCAST, DL, StoreVecVT, Shuff);
9338 SmallVector<SDValue, 8> Chains;
9339 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits()/8, DL,
9340 TLI.getPointerTy());
9341 SDValue BasePtr = St->getBasePtr();
9343 // Perform one or more big stores into memory.
9344 unsigned E = (ToEltSz*NumElems)/StoreType.getSizeInBits();
9345 for (unsigned I = 0; I < E; I++) {
9346 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
9347 StoreType, ShuffWide,
9348 DAG.getIntPtrConstant(I, DL));
9349 SDValue Ch = DAG.getStore(St->getChain(), DL, SubVec, BasePtr,
9350 St->getPointerInfo(), St->isVolatile(),
9351 St->isNonTemporal(), St->getAlignment());
9352 BasePtr = DAG.getNode(ISD::ADD, DL, BasePtr.getValueType(), BasePtr,
9354 Chains.push_back(Ch);
9356 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
9359 if (!ISD::isNormalStore(St))
9362 // Split a store of a VMOVDRR into two integer stores to avoid mixing NEON and
9363 // ARM stores of arguments in the same cache line.
9364 if (StVal.getNode()->getOpcode() == ARMISD::VMOVDRR &&
9365 StVal.getNode()->hasOneUse()) {
9366 SelectionDAG &DAG = DCI.DAG;
9367 bool isBigEndian = DAG.getDataLayout().isBigEndian();
9369 SDValue BasePtr = St->getBasePtr();
9370 SDValue NewST1 = DAG.getStore(St->getChain(), DL,
9371 StVal.getNode()->getOperand(isBigEndian ? 1 : 0 ),
9372 BasePtr, St->getPointerInfo(), St->isVolatile(),
9373 St->isNonTemporal(), St->getAlignment());
9375 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
9376 DAG.getConstant(4, DL, MVT::i32));
9377 return DAG.getStore(NewST1.getValue(0), DL,
9378 StVal.getNode()->getOperand(isBigEndian ? 0 : 1),
9379 OffsetPtr, St->getPointerInfo(), St->isVolatile(),
9380 St->isNonTemporal(),
9381 std::min(4U, St->getAlignment() / 2));
9384 if (StVal.getValueType() == MVT::i64 &&
9385 StVal.getNode()->getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
9387 // Bitcast an i64 store extracted from a vector to f64.
9388 // Otherwise, the i64 value will be legalized to a pair of i32 values.
9389 SelectionDAG &DAG = DCI.DAG;
9391 SDValue IntVec = StVal.getOperand(0);
9392 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
9393 IntVec.getValueType().getVectorNumElements());
9394 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, IntVec);
9395 SDValue ExtElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
9396 Vec, StVal.getOperand(1));
9398 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ExtElt);
9399 // Make the DAGCombiner fold the bitcasts.
9400 DCI.AddToWorklist(Vec.getNode());
9401 DCI.AddToWorklist(ExtElt.getNode());
9402 DCI.AddToWorklist(V.getNode());
9403 return DAG.getStore(St->getChain(), dl, V, St->getBasePtr(),
9404 St->getPointerInfo(), St->isVolatile(),
9405 St->isNonTemporal(), St->getAlignment(),
9409 // If this is a legal vector store, try to combine it into a VST1_UPD.
9410 if (ISD::isNormalStore(N) && VT.isVector() &&
9411 DCI.DAG.getTargetLoweringInfo().isTypeLegal(VT))
9412 return CombineBaseUpdate(N, DCI);
9417 // isConstVecPow2 - Return true if each vector element is a power of 2, all
9418 // elements are the same constant, C, and Log2(C) ranges from 1 to 32.
9419 static bool isConstVecPow2(SDValue ConstVec, bool isSigned, uint64_t &C)
9423 for (unsigned I = 0, E = ConstVec.getValueType().getVectorNumElements();
9425 ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(ConstVec.getOperand(I));
9430 APFloat APF = C->getValueAPF();
9431 if (APF.convertToInteger(&cN, 64, isSigned, APFloat::rmTowardZero, &isExact)
9432 != APFloat::opOK || !isExact)
9435 c0 = (I == 0) ? cN : c0;
9436 if (!isPowerOf2_64(cN) || c0 != cN || Log2_64(c0) < 1 || Log2_64(c0) > 32)
9443 /// PerformVCVTCombine - VCVT (floating-point to fixed-point, Advanced SIMD)
9444 /// can replace combinations of VMUL and VCVT (floating-point to integer)
9445 /// when the VMUL has a constant operand that is a power of 2.
9447 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
9448 /// vmul.f32 d16, d17, d16
9449 /// vcvt.s32.f32 d16, d16
9451 /// vcvt.s32.f32 d16, d16, #3
9452 static SDValue PerformVCVTCombine(SDNode *N,
9453 TargetLowering::DAGCombinerInfo &DCI,
9454 const ARMSubtarget *Subtarget) {
9455 SelectionDAG &DAG = DCI.DAG;
9456 SDValue Op = N->getOperand(0);
9458 if (!Subtarget->hasNEON() || !Op.getValueType().isVector() ||
9459 Op.getOpcode() != ISD::FMUL)
9463 SDValue N0 = Op->getOperand(0);
9464 SDValue ConstVec = Op->getOperand(1);
9465 bool isSigned = N->getOpcode() == ISD::FP_TO_SINT;
9467 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
9468 !isConstVecPow2(ConstVec, isSigned, C))
9471 MVT FloatTy = Op.getSimpleValueType().getVectorElementType();
9472 MVT IntTy = N->getSimpleValueType(0).getVectorElementType();
9473 unsigned NumLanes = Op.getValueType().getVectorNumElements();
9474 if (FloatTy.getSizeInBits() != 32 || IntTy.getSizeInBits() > 32 ||
9476 // These instructions only exist converting from f32 to i32. We can handle
9477 // smaller integers by generating an extra truncate, but larger ones would
9478 // be lossy. We also can't handle more then 4 lanes, since these intructions
9479 // only support v2i32/v4i32 types.
9484 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfp2fxs :
9485 Intrinsic::arm_neon_vcvtfp2fxu;
9486 SDValue FixConv = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl,
9487 NumLanes == 2 ? MVT::v2i32 : MVT::v4i32,
9488 DAG.getConstant(IntrinsicOpcode, dl, MVT::i32),
9490 DAG.getConstant(Log2_64(C), dl, MVT::i32));
9492 if (IntTy.getSizeInBits() < FloatTy.getSizeInBits())
9493 FixConv = DAG.getNode(ISD::TRUNCATE, dl, N->getValueType(0), FixConv);
9498 /// PerformVDIVCombine - VCVT (fixed-point to floating-point, Advanced SIMD)
9499 /// can replace combinations of VCVT (integer to floating-point) and VDIV
9500 /// when the VDIV has a constant operand that is a power of 2.
9502 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
9503 /// vcvt.f32.s32 d16, d16
9504 /// vdiv.f32 d16, d17, d16
9506 /// vcvt.f32.s32 d16, d16, #3
9507 static SDValue PerformVDIVCombine(SDNode *N,
9508 TargetLowering::DAGCombinerInfo &DCI,
9509 const ARMSubtarget *Subtarget) {
9510 SelectionDAG &DAG = DCI.DAG;
9511 SDValue Op = N->getOperand(0);
9512 unsigned OpOpcode = Op.getNode()->getOpcode();
9514 if (!Subtarget->hasNEON() || !N->getValueType(0).isVector() ||
9515 (OpOpcode != ISD::SINT_TO_FP && OpOpcode != ISD::UINT_TO_FP))
9519 SDValue ConstVec = N->getOperand(1);
9520 bool isSigned = OpOpcode == ISD::SINT_TO_FP;
9522 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
9523 !isConstVecPow2(ConstVec, isSigned, C))
9526 MVT FloatTy = N->getSimpleValueType(0).getVectorElementType();
9527 MVT IntTy = Op.getOperand(0).getSimpleValueType().getVectorElementType();
9528 if (FloatTy.getSizeInBits() != 32 || IntTy.getSizeInBits() > 32) {
9529 // These instructions only exist converting from i32 to f32. We can handle
9530 // smaller integers by generating an extra extend, but larger ones would
9536 SDValue ConvInput = Op.getOperand(0);
9537 unsigned NumLanes = Op.getValueType().getVectorNumElements();
9538 if (IntTy.getSizeInBits() < FloatTy.getSizeInBits())
9539 ConvInput = DAG.getNode(isSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND,
9540 dl, NumLanes == 2 ? MVT::v2i32 : MVT::v4i32,
9543 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfxs2fp :
9544 Intrinsic::arm_neon_vcvtfxu2fp;
9545 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl,
9547 DAG.getConstant(IntrinsicOpcode, dl, MVT::i32),
9548 ConvInput, DAG.getConstant(Log2_64(C), dl, MVT::i32));
9551 /// Getvshiftimm - Check if this is a valid build_vector for the immediate
9552 /// operand of a vector shift operation, where all the elements of the
9553 /// build_vector must have the same constant integer value.
9554 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
9555 // Ignore bit_converts.
9556 while (Op.getOpcode() == ISD::BITCAST)
9557 Op = Op.getOperand(0);
9558 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
9559 APInt SplatBits, SplatUndef;
9560 unsigned SplatBitSize;
9562 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
9563 HasAnyUndefs, ElementBits) ||
9564 SplatBitSize > ElementBits)
9566 Cnt = SplatBits.getSExtValue();
9570 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
9571 /// operand of a vector shift left operation. That value must be in the range:
9572 /// 0 <= Value < ElementBits for a left shift; or
9573 /// 0 <= Value <= ElementBits for a long left shift.
9574 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
9575 assert(VT.isVector() && "vector shift count is not a vector type");
9576 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
9577 if (! getVShiftImm(Op, ElementBits, Cnt))
9579 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
9582 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
9583 /// operand of a vector shift right operation. For a shift opcode, the value
9584 /// is positive, but for an intrinsic the value count must be negative. The
9585 /// absolute value must be in the range:
9586 /// 1 <= |Value| <= ElementBits for a right shift; or
9587 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
9588 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
9590 assert(VT.isVector() && "vector shift count is not a vector type");
9591 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
9592 if (! getVShiftImm(Op, ElementBits, Cnt))
9596 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
9599 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
9600 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
9601 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
9604 // Don't do anything for most intrinsics.
9607 // Vector shifts: check for immediate versions and lower them.
9608 // Note: This is done during DAG combining instead of DAG legalizing because
9609 // the build_vectors for 64-bit vector element shift counts are generally
9610 // not legal, and it is hard to see their values after they get legalized to
9611 // loads from a constant pool.
9612 case Intrinsic::arm_neon_vshifts:
9613 case Intrinsic::arm_neon_vshiftu:
9614 case Intrinsic::arm_neon_vrshifts:
9615 case Intrinsic::arm_neon_vrshiftu:
9616 case Intrinsic::arm_neon_vrshiftn:
9617 case Intrinsic::arm_neon_vqshifts:
9618 case Intrinsic::arm_neon_vqshiftu:
9619 case Intrinsic::arm_neon_vqshiftsu:
9620 case Intrinsic::arm_neon_vqshiftns:
9621 case Intrinsic::arm_neon_vqshiftnu:
9622 case Intrinsic::arm_neon_vqshiftnsu:
9623 case Intrinsic::arm_neon_vqrshiftns:
9624 case Intrinsic::arm_neon_vqrshiftnu:
9625 case Intrinsic::arm_neon_vqrshiftnsu: {
9626 EVT VT = N->getOperand(1).getValueType();
9628 unsigned VShiftOpc = 0;
9631 case Intrinsic::arm_neon_vshifts:
9632 case Intrinsic::arm_neon_vshiftu:
9633 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
9634 VShiftOpc = ARMISD::VSHL;
9637 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
9638 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
9639 ARMISD::VSHRs : ARMISD::VSHRu);
9644 case Intrinsic::arm_neon_vrshifts:
9645 case Intrinsic::arm_neon_vrshiftu:
9646 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
9650 case Intrinsic::arm_neon_vqshifts:
9651 case Intrinsic::arm_neon_vqshiftu:
9652 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
9656 case Intrinsic::arm_neon_vqshiftsu:
9657 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
9659 llvm_unreachable("invalid shift count for vqshlu intrinsic");
9661 case Intrinsic::arm_neon_vrshiftn:
9662 case Intrinsic::arm_neon_vqshiftns:
9663 case Intrinsic::arm_neon_vqshiftnu:
9664 case Intrinsic::arm_neon_vqshiftnsu:
9665 case Intrinsic::arm_neon_vqrshiftns:
9666 case Intrinsic::arm_neon_vqrshiftnu:
9667 case Intrinsic::arm_neon_vqrshiftnsu:
9668 // Narrowing shifts require an immediate right shift.
9669 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
9671 llvm_unreachable("invalid shift count for narrowing vector shift "
9675 llvm_unreachable("unhandled vector shift");
9679 case Intrinsic::arm_neon_vshifts:
9680 case Intrinsic::arm_neon_vshiftu:
9681 // Opcode already set above.
9683 case Intrinsic::arm_neon_vrshifts:
9684 VShiftOpc = ARMISD::VRSHRs; break;
9685 case Intrinsic::arm_neon_vrshiftu:
9686 VShiftOpc = ARMISD::VRSHRu; break;
9687 case Intrinsic::arm_neon_vrshiftn:
9688 VShiftOpc = ARMISD::VRSHRN; break;
9689 case Intrinsic::arm_neon_vqshifts:
9690 VShiftOpc = ARMISD::VQSHLs; break;
9691 case Intrinsic::arm_neon_vqshiftu:
9692 VShiftOpc = ARMISD::VQSHLu; break;
9693 case Intrinsic::arm_neon_vqshiftsu:
9694 VShiftOpc = ARMISD::VQSHLsu; break;
9695 case Intrinsic::arm_neon_vqshiftns:
9696 VShiftOpc = ARMISD::VQSHRNs; break;
9697 case Intrinsic::arm_neon_vqshiftnu:
9698 VShiftOpc = ARMISD::VQSHRNu; break;
9699 case Intrinsic::arm_neon_vqshiftnsu:
9700 VShiftOpc = ARMISD::VQSHRNsu; break;
9701 case Intrinsic::arm_neon_vqrshiftns:
9702 VShiftOpc = ARMISD::VQRSHRNs; break;
9703 case Intrinsic::arm_neon_vqrshiftnu:
9704 VShiftOpc = ARMISD::VQRSHRNu; break;
9705 case Intrinsic::arm_neon_vqrshiftnsu:
9706 VShiftOpc = ARMISD::VQRSHRNsu; break;
9710 return DAG.getNode(VShiftOpc, dl, N->getValueType(0),
9711 N->getOperand(1), DAG.getConstant(Cnt, dl, MVT::i32));
9714 case Intrinsic::arm_neon_vshiftins: {
9715 EVT VT = N->getOperand(1).getValueType();
9717 unsigned VShiftOpc = 0;
9719 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
9720 VShiftOpc = ARMISD::VSLI;
9721 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
9722 VShiftOpc = ARMISD::VSRI;
9724 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
9728 return DAG.getNode(VShiftOpc, dl, N->getValueType(0),
9729 N->getOperand(1), N->getOperand(2),
9730 DAG.getConstant(Cnt, dl, MVT::i32));
9733 case Intrinsic::arm_neon_vqrshifts:
9734 case Intrinsic::arm_neon_vqrshiftu:
9735 // No immediate versions of these to check for.
9742 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
9743 /// lowers them. As with the vector shift intrinsics, this is done during DAG
9744 /// combining instead of DAG legalizing because the build_vectors for 64-bit
9745 /// vector element shift counts are generally not legal, and it is hard to see
9746 /// their values after they get legalized to loads from a constant pool.
9747 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
9748 const ARMSubtarget *ST) {
9749 EVT VT = N->getValueType(0);
9750 if (N->getOpcode() == ISD::SRL && VT == MVT::i32 && ST->hasV6Ops()) {
9751 // Canonicalize (srl (bswap x), 16) to (rotr (bswap x), 16) if the high
9752 // 16-bits of x is zero. This optimizes rev + lsr 16 to rev16.
9753 SDValue N1 = N->getOperand(1);
9754 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) {
9755 SDValue N0 = N->getOperand(0);
9756 if (C->getZExtValue() == 16 && N0.getOpcode() == ISD::BSWAP &&
9757 DAG.MaskedValueIsZero(N0.getOperand(0),
9758 APInt::getHighBitsSet(32, 16)))
9759 return DAG.getNode(ISD::ROTR, SDLoc(N), VT, N0, N1);
9763 // Nothing to be done for scalar shifts.
9764 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9765 if (!VT.isVector() || !TLI.isTypeLegal(VT))
9768 assert(ST->hasNEON() && "unexpected vector shift");
9771 switch (N->getOpcode()) {
9772 default: llvm_unreachable("unexpected shift opcode");
9775 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt)) {
9777 return DAG.getNode(ARMISD::VSHL, dl, VT, N->getOperand(0),
9778 DAG.getConstant(Cnt, dl, MVT::i32));
9784 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
9785 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
9786 ARMISD::VSHRs : ARMISD::VSHRu);
9788 return DAG.getNode(VShiftOpc, dl, VT, N->getOperand(0),
9789 DAG.getConstant(Cnt, dl, MVT::i32));
9795 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
9796 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
9797 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
9798 const ARMSubtarget *ST) {
9799 SDValue N0 = N->getOperand(0);
9801 // Check for sign- and zero-extensions of vector extract operations of 8-
9802 // and 16-bit vector elements. NEON supports these directly. They are
9803 // handled during DAG combining because type legalization will promote them
9804 // to 32-bit types and it is messy to recognize the operations after that.
9805 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
9806 SDValue Vec = N0.getOperand(0);
9807 SDValue Lane = N0.getOperand(1);
9808 EVT VT = N->getValueType(0);
9809 EVT EltVT = N0.getValueType();
9810 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9812 if (VT == MVT::i32 &&
9813 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
9814 TLI.isTypeLegal(Vec.getValueType()) &&
9815 isa<ConstantSDNode>(Lane)) {
9818 switch (N->getOpcode()) {
9819 default: llvm_unreachable("unexpected opcode");
9820 case ISD::SIGN_EXTEND:
9821 Opc = ARMISD::VGETLANEs;
9823 case ISD::ZERO_EXTEND:
9824 case ISD::ANY_EXTEND:
9825 Opc = ARMISD::VGETLANEu;
9828 return DAG.getNode(Opc, SDLoc(N), VT, Vec, Lane);
9835 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
9836 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
9837 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
9838 const ARMSubtarget *ST) {
9839 // If the target supports NEON, try to use vmax/vmin instructions for f32
9840 // selects like "x < y ? x : y". Unless the NoNaNsFPMath option is set,
9841 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
9842 // a NaN; only do the transformation when it matches that behavior.
9844 // For now only do this when using NEON for FP operations; if using VFP, it
9845 // is not obvious that the benefit outweighs the cost of switching to the
9847 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
9848 N->getValueType(0) != MVT::f32)
9851 SDValue CondLHS = N->getOperand(0);
9852 SDValue CondRHS = N->getOperand(1);
9853 SDValue LHS = N->getOperand(2);
9854 SDValue RHS = N->getOperand(3);
9855 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
9857 unsigned Opcode = 0;
9859 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
9860 IsReversed = false; // x CC y ? x : y
9861 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
9862 IsReversed = true ; // x CC y ? y : x
9876 // If LHS is NaN, an ordered comparison will be false and the result will
9877 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
9878 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
9879 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
9880 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
9882 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
9883 // will return -0, so vmin can only be used for unsafe math or if one of
9884 // the operands is known to be nonzero.
9885 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
9886 !DAG.getTarget().Options.UnsafeFPMath &&
9887 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9889 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
9898 // If LHS is NaN, an ordered comparison will be false and the result will
9899 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
9900 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
9901 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
9902 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
9904 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
9905 // will return +0, so vmax can only be used for unsafe math or if one of
9906 // the operands is known to be nonzero.
9907 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
9908 !DAG.getTarget().Options.UnsafeFPMath &&
9909 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9911 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
9917 return DAG.getNode(Opcode, SDLoc(N), N->getValueType(0), LHS, RHS);
9920 /// PerformCMOVCombine - Target-specific DAG combining for ARMISD::CMOV.
9922 ARMTargetLowering::PerformCMOVCombine(SDNode *N, SelectionDAG &DAG) const {
9923 SDValue Cmp = N->getOperand(4);
9924 if (Cmp.getOpcode() != ARMISD::CMPZ)
9925 // Only looking at EQ and NE cases.
9928 EVT VT = N->getValueType(0);
9930 SDValue LHS = Cmp.getOperand(0);
9931 SDValue RHS = Cmp.getOperand(1);
9932 SDValue FalseVal = N->getOperand(0);
9933 SDValue TrueVal = N->getOperand(1);
9934 SDValue ARMcc = N->getOperand(2);
9935 ARMCC::CondCodes CC =
9936 (ARMCC::CondCodes)cast<ConstantSDNode>(ARMcc)->getZExtValue();
9954 /// FIXME: Turn this into a target neutral optimization?
9956 if (CC == ARMCC::NE && FalseVal == RHS && FalseVal != LHS) {
9957 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, TrueVal, ARMcc,
9958 N->getOperand(3), Cmp);
9959 } else if (CC == ARMCC::EQ && TrueVal == RHS) {
9961 SDValue NewCmp = getARMCmp(LHS, RHS, ISD::SETNE, ARMcc, DAG, dl);
9962 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, FalseVal, ARMcc,
9963 N->getOperand(3), NewCmp);
9966 if (Res.getNode()) {
9967 APInt KnownZero, KnownOne;
9968 DAG.computeKnownBits(SDValue(N,0), KnownZero, KnownOne);
9969 // Capture demanded bits information that would be otherwise lost.
9970 if (KnownZero == 0xfffffffe)
9971 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
9972 DAG.getValueType(MVT::i1));
9973 else if (KnownZero == 0xffffff00)
9974 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
9975 DAG.getValueType(MVT::i8));
9976 else if (KnownZero == 0xffff0000)
9977 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
9978 DAG.getValueType(MVT::i16));
9984 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
9985 DAGCombinerInfo &DCI) const {
9986 switch (N->getOpcode()) {
9988 case ISD::ADDC: return PerformADDCCombine(N, DCI, Subtarget);
9989 case ISD::ADD: return PerformADDCombine(N, DCI, Subtarget);
9990 case ISD::SUB: return PerformSUBCombine(N, DCI);
9991 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
9992 case ISD::OR: return PerformORCombine(N, DCI, Subtarget);
9993 case ISD::XOR: return PerformXORCombine(N, DCI, Subtarget);
9994 case ISD::AND: return PerformANDCombine(N, DCI, Subtarget);
9995 case ARMISD::BFI: return PerformBFICombine(N, DCI);
9996 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI, Subtarget);
9997 case ARMISD::VMOVDRR: return PerformVMOVDRRCombine(N, DCI.DAG);
9998 case ISD::STORE: return PerformSTORECombine(N, DCI);
9999 case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DCI, Subtarget);
10000 case ISD::INSERT_VECTOR_ELT: return PerformInsertEltCombine(N, DCI);
10001 case ISD::VECTOR_SHUFFLE: return PerformVECTOR_SHUFFLECombine(N, DCI.DAG);
10002 case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI);
10003 case ISD::FP_TO_SINT:
10004 case ISD::FP_TO_UINT: return PerformVCVTCombine(N, DCI, Subtarget);
10005 case ISD::FDIV: return PerformVDIVCombine(N, DCI, Subtarget);
10006 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
10009 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
10010 case ISD::SIGN_EXTEND:
10011 case ISD::ZERO_EXTEND:
10012 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
10013 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
10014 case ARMISD::CMOV: return PerformCMOVCombine(N, DCI.DAG);
10015 case ISD::LOAD: return PerformLOADCombine(N, DCI);
10016 case ARMISD::VLD2DUP:
10017 case ARMISD::VLD3DUP:
10018 case ARMISD::VLD4DUP:
10019 return PerformVLDCombine(N, DCI);
10020 case ARMISD::BUILD_VECTOR:
10021 return PerformARMBUILD_VECTORCombine(N, DCI);
10022 case ISD::INTRINSIC_VOID:
10023 case ISD::INTRINSIC_W_CHAIN:
10024 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
10025 case Intrinsic::arm_neon_vld1:
10026 case Intrinsic::arm_neon_vld2:
10027 case Intrinsic::arm_neon_vld3:
10028 case Intrinsic::arm_neon_vld4:
10029 case Intrinsic::arm_neon_vld2lane:
10030 case Intrinsic::arm_neon_vld3lane:
10031 case Intrinsic::arm_neon_vld4lane:
10032 case Intrinsic::arm_neon_vst1:
10033 case Intrinsic::arm_neon_vst2:
10034 case Intrinsic::arm_neon_vst3:
10035 case Intrinsic::arm_neon_vst4:
10036 case Intrinsic::arm_neon_vst2lane:
10037 case Intrinsic::arm_neon_vst3lane:
10038 case Intrinsic::arm_neon_vst4lane:
10039 return PerformVLDCombine(N, DCI);
10047 bool ARMTargetLowering::isDesirableToTransformToIntegerOp(unsigned Opc,
10049 return (VT == MVT::f32) && (Opc == ISD::LOAD || Opc == ISD::STORE);
10052 bool ARMTargetLowering::allowsMisalignedMemoryAccesses(EVT VT,
10055 bool *Fast) const {
10056 // The AllowsUnaliged flag models the SCTLR.A setting in ARM cpus
10057 bool AllowsUnaligned = Subtarget->allowsUnalignedMem();
10059 switch (VT.getSimpleVT().SimpleTy) {
10065 // Unaligned access can use (for example) LRDB, LRDH, LDR
10066 if (AllowsUnaligned) {
10068 *Fast = Subtarget->hasV7Ops();
10075 // For any little-endian targets with neon, we can support unaligned ld/st
10076 // of D and Q (e.g. {D0,D1}) registers by using vld1.i8/vst1.i8.
10077 // A big-endian target may also explicitly support unaligned accesses
10078 if (Subtarget->hasNEON() && (AllowsUnaligned || Subtarget->isLittle())) {
10088 static bool memOpAlign(unsigned DstAlign, unsigned SrcAlign,
10089 unsigned AlignCheck) {
10090 return ((SrcAlign == 0 || SrcAlign % AlignCheck == 0) &&
10091 (DstAlign == 0 || DstAlign % AlignCheck == 0));
10094 EVT ARMTargetLowering::getOptimalMemOpType(uint64_t Size,
10095 unsigned DstAlign, unsigned SrcAlign,
10096 bool IsMemset, bool ZeroMemset,
10098 MachineFunction &MF) const {
10099 const Function *F = MF.getFunction();
10101 // See if we can use NEON instructions for this...
10102 if ((!IsMemset || ZeroMemset) && Subtarget->hasNEON() &&
10103 !F->hasFnAttribute(Attribute::NoImplicitFloat)) {
10106 (memOpAlign(SrcAlign, DstAlign, 16) ||
10107 (allowsMisalignedMemoryAccesses(MVT::v2f64, 0, 1, &Fast) && Fast))) {
10109 } else if (Size >= 8 &&
10110 (memOpAlign(SrcAlign, DstAlign, 8) ||
10111 (allowsMisalignedMemoryAccesses(MVT::f64, 0, 1, &Fast) &&
10117 // Lowering to i32/i16 if the size permits.
10120 else if (Size >= 2)
10123 // Let the target-independent logic figure it out.
10127 bool ARMTargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
10128 if (Val.getOpcode() != ISD::LOAD)
10131 EVT VT1 = Val.getValueType();
10132 if (!VT1.isSimple() || !VT1.isInteger() ||
10133 !VT2.isSimple() || !VT2.isInteger())
10136 switch (VT1.getSimpleVT().SimpleTy) {
10141 // 8-bit and 16-bit loads implicitly zero-extend to 32-bits.
10148 bool ARMTargetLowering::isVectorLoadExtDesirable(SDValue ExtVal) const {
10149 EVT VT = ExtVal.getValueType();
10151 if (!isTypeLegal(VT))
10154 // Don't create a loadext if we can fold the extension into a wide/long
10156 // If there's more than one user instruction, the loadext is desirable no
10157 // matter what. There can be two uses by the same instruction.
10158 if (ExtVal->use_empty() ||
10159 !ExtVal->use_begin()->isOnlyUserOf(ExtVal.getNode()))
10162 SDNode *U = *ExtVal->use_begin();
10163 if ((U->getOpcode() == ISD::ADD || U->getOpcode() == ISD::SUB ||
10164 U->getOpcode() == ISD::SHL || U->getOpcode() == ARMISD::VSHL))
10170 bool ARMTargetLowering::allowTruncateForTailCall(Type *Ty1, Type *Ty2) const {
10171 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
10174 if (!isTypeLegal(EVT::getEVT(Ty1)))
10177 assert(Ty1->getPrimitiveSizeInBits() <= 64 && "i128 is probably not a noop");
10179 // Assuming the caller doesn't have a zeroext or signext return parameter,
10180 // truncation all the way down to i1 is valid.
10185 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
10189 unsigned Scale = 1;
10190 switch (VT.getSimpleVT().SimpleTy) {
10191 default: return false;
10206 if ((V & (Scale - 1)) != 0)
10209 return V == (V & ((1LL << 5) - 1));
10212 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
10213 const ARMSubtarget *Subtarget) {
10214 bool isNeg = false;
10220 switch (VT.getSimpleVT().SimpleTy) {
10221 default: return false;
10226 // + imm12 or - imm8
10228 return V == (V & ((1LL << 8) - 1));
10229 return V == (V & ((1LL << 12) - 1));
10232 // Same as ARM mode. FIXME: NEON?
10233 if (!Subtarget->hasVFP2())
10238 return V == (V & ((1LL << 8) - 1));
10242 /// isLegalAddressImmediate - Return true if the integer value can be used
10243 /// as the offset of the target addressing mode for load / store of the
10245 static bool isLegalAddressImmediate(int64_t V, EVT VT,
10246 const ARMSubtarget *Subtarget) {
10250 if (!VT.isSimple())
10253 if (Subtarget->isThumb1Only())
10254 return isLegalT1AddressImmediate(V, VT);
10255 else if (Subtarget->isThumb2())
10256 return isLegalT2AddressImmediate(V, VT, Subtarget);
10261 switch (VT.getSimpleVT().SimpleTy) {
10262 default: return false;
10267 return V == (V & ((1LL << 12) - 1));
10270 return V == (V & ((1LL << 8) - 1));
10273 if (!Subtarget->hasVFP2()) // FIXME: NEON?
10278 return V == (V & ((1LL << 8) - 1));
10282 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
10284 int Scale = AM.Scale;
10288 switch (VT.getSimpleVT().SimpleTy) {
10289 default: return false;
10297 Scale = Scale & ~1;
10298 return Scale == 2 || Scale == 4 || Scale == 8;
10301 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
10305 // Note, we allow "void" uses (basically, uses that aren't loads or
10306 // stores), because arm allows folding a scale into many arithmetic
10307 // operations. This should be made more precise and revisited later.
10309 // Allow r << imm, but the imm has to be a multiple of two.
10310 if (Scale & 1) return false;
10311 return isPowerOf2_32(Scale);
10315 /// isLegalAddressingMode - Return true if the addressing mode represented
10316 /// by AM is legal for this target, for a load/store of the specified type.
10317 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
10319 unsigned AS) const {
10320 EVT VT = getValueType(Ty, true);
10321 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
10324 // Can never fold addr of global into load/store.
10328 switch (AM.Scale) {
10329 case 0: // no scale reg, must be "r+i" or "r", or "i".
10332 if (Subtarget->isThumb1Only())
10336 // ARM doesn't support any R+R*scale+imm addr modes.
10340 if (!VT.isSimple())
10343 if (Subtarget->isThumb2())
10344 return isLegalT2ScaledAddressingMode(AM, VT);
10346 int Scale = AM.Scale;
10347 switch (VT.getSimpleVT().SimpleTy) {
10348 default: return false;
10352 if (Scale < 0) Scale = -Scale;
10356 return isPowerOf2_32(Scale & ~1);
10360 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
10365 // Note, we allow "void" uses (basically, uses that aren't loads or
10366 // stores), because arm allows folding a scale into many arithmetic
10367 // operations. This should be made more precise and revisited later.
10369 // Allow r << imm, but the imm has to be a multiple of two.
10370 if (Scale & 1) return false;
10371 return isPowerOf2_32(Scale);
10377 /// isLegalICmpImmediate - Return true if the specified immediate is legal
10378 /// icmp immediate, that is the target has icmp instructions which can compare
10379 /// a register against the immediate without having to materialize the
10380 /// immediate into a register.
10381 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
10382 // Thumb2 and ARM modes can use cmn for negative immediates.
10383 if (!Subtarget->isThumb())
10384 return ARM_AM::getSOImmVal(std::abs(Imm)) != -1;
10385 if (Subtarget->isThumb2())
10386 return ARM_AM::getT2SOImmVal(std::abs(Imm)) != -1;
10387 // Thumb1 doesn't have cmn, and only 8-bit immediates.
10388 return Imm >= 0 && Imm <= 255;
10391 /// isLegalAddImmediate - Return true if the specified immediate is a legal add
10392 /// *or sub* immediate, that is the target has add or sub instructions which can
10393 /// add a register with the immediate without having to materialize the
10394 /// immediate into a register.
10395 bool ARMTargetLowering::isLegalAddImmediate(int64_t Imm) const {
10396 // Same encoding for add/sub, just flip the sign.
10397 int64_t AbsImm = std::abs(Imm);
10398 if (!Subtarget->isThumb())
10399 return ARM_AM::getSOImmVal(AbsImm) != -1;
10400 if (Subtarget->isThumb2())
10401 return ARM_AM::getT2SOImmVal(AbsImm) != -1;
10402 // Thumb1 only has 8-bit unsigned immediate.
10403 return AbsImm >= 0 && AbsImm <= 255;
10406 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
10407 bool isSEXTLoad, SDValue &Base,
10408 SDValue &Offset, bool &isInc,
10409 SelectionDAG &DAG) {
10410 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
10413 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
10414 // AddressingMode 3
10415 Base = Ptr->getOperand(0);
10416 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
10417 int RHSC = (int)RHS->getZExtValue();
10418 if (RHSC < 0 && RHSC > -256) {
10419 assert(Ptr->getOpcode() == ISD::ADD);
10421 Offset = DAG.getConstant(-RHSC, SDLoc(Ptr), RHS->getValueType(0));
10425 isInc = (Ptr->getOpcode() == ISD::ADD);
10426 Offset = Ptr->getOperand(1);
10428 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
10429 // AddressingMode 2
10430 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
10431 int RHSC = (int)RHS->getZExtValue();
10432 if (RHSC < 0 && RHSC > -0x1000) {
10433 assert(Ptr->getOpcode() == ISD::ADD);
10435 Offset = DAG.getConstant(-RHSC, SDLoc(Ptr), RHS->getValueType(0));
10436 Base = Ptr->getOperand(0);
10441 if (Ptr->getOpcode() == ISD::ADD) {
10443 ARM_AM::ShiftOpc ShOpcVal=
10444 ARM_AM::getShiftOpcForNode(Ptr->getOperand(0).getOpcode());
10445 if (ShOpcVal != ARM_AM::no_shift) {
10446 Base = Ptr->getOperand(1);
10447 Offset = Ptr->getOperand(0);
10449 Base = Ptr->getOperand(0);
10450 Offset = Ptr->getOperand(1);
10455 isInc = (Ptr->getOpcode() == ISD::ADD);
10456 Base = Ptr->getOperand(0);
10457 Offset = Ptr->getOperand(1);
10461 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
10465 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
10466 bool isSEXTLoad, SDValue &Base,
10467 SDValue &Offset, bool &isInc,
10468 SelectionDAG &DAG) {
10469 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
10472 Base = Ptr->getOperand(0);
10473 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
10474 int RHSC = (int)RHS->getZExtValue();
10475 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
10476 assert(Ptr->getOpcode() == ISD::ADD);
10478 Offset = DAG.getConstant(-RHSC, SDLoc(Ptr), RHS->getValueType(0));
10480 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
10481 isInc = Ptr->getOpcode() == ISD::ADD;
10482 Offset = DAG.getConstant(RHSC, SDLoc(Ptr), RHS->getValueType(0));
10490 /// getPreIndexedAddressParts - returns true by value, base pointer and
10491 /// offset pointer and addressing mode by reference if the node's address
10492 /// can be legally represented as pre-indexed load / store address.
10494 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
10496 ISD::MemIndexedMode &AM,
10497 SelectionDAG &DAG) const {
10498 if (Subtarget->isThumb1Only())
10503 bool isSEXTLoad = false;
10504 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
10505 Ptr = LD->getBasePtr();
10506 VT = LD->getMemoryVT();
10507 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
10508 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
10509 Ptr = ST->getBasePtr();
10510 VT = ST->getMemoryVT();
10515 bool isLegal = false;
10516 if (Subtarget->isThumb2())
10517 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
10518 Offset, isInc, DAG);
10520 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
10521 Offset, isInc, DAG);
10525 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
10529 /// getPostIndexedAddressParts - returns true by value, base pointer and
10530 /// offset pointer and addressing mode by reference if this node can be
10531 /// combined with a load / store to form a post-indexed load / store.
10532 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
10535 ISD::MemIndexedMode &AM,
10536 SelectionDAG &DAG) const {
10537 if (Subtarget->isThumb1Only())
10542 bool isSEXTLoad = false;
10543 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
10544 VT = LD->getMemoryVT();
10545 Ptr = LD->getBasePtr();
10546 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
10547 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
10548 VT = ST->getMemoryVT();
10549 Ptr = ST->getBasePtr();
10554 bool isLegal = false;
10555 if (Subtarget->isThumb2())
10556 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
10559 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
10565 // Swap base ptr and offset to catch more post-index load / store when
10566 // it's legal. In Thumb2 mode, offset must be an immediate.
10567 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
10568 !Subtarget->isThumb2())
10569 std::swap(Base, Offset);
10571 // Post-indexed load / store update the base pointer.
10576 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
10580 void ARMTargetLowering::computeKnownBitsForTargetNode(const SDValue Op,
10583 const SelectionDAG &DAG,
10584 unsigned Depth) const {
10585 unsigned BitWidth = KnownOne.getBitWidth();
10586 KnownZero = KnownOne = APInt(BitWidth, 0);
10587 switch (Op.getOpcode()) {
10593 // These nodes' second result is a boolean
10594 if (Op.getResNo() == 0)
10596 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - 1);
10598 case ARMISD::CMOV: {
10599 // Bits are known zero/one if known on the LHS and RHS.
10600 DAG.computeKnownBits(Op.getOperand(0), KnownZero, KnownOne, Depth+1);
10601 if (KnownZero == 0 && KnownOne == 0) return;
10603 APInt KnownZeroRHS, KnownOneRHS;
10604 DAG.computeKnownBits(Op.getOperand(1), KnownZeroRHS, KnownOneRHS, Depth+1);
10605 KnownZero &= KnownZeroRHS;
10606 KnownOne &= KnownOneRHS;
10609 case ISD::INTRINSIC_W_CHAIN: {
10610 ConstantSDNode *CN = cast<ConstantSDNode>(Op->getOperand(1));
10611 Intrinsic::ID IntID = static_cast<Intrinsic::ID>(CN->getZExtValue());
10614 case Intrinsic::arm_ldaex:
10615 case Intrinsic::arm_ldrex: {
10616 EVT VT = cast<MemIntrinsicSDNode>(Op)->getMemoryVT();
10617 unsigned MemBits = VT.getScalarType().getSizeInBits();
10618 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - MemBits);
10626 //===----------------------------------------------------------------------===//
10627 // ARM Inline Assembly Support
10628 //===----------------------------------------------------------------------===//
10630 bool ARMTargetLowering::ExpandInlineAsm(CallInst *CI) const {
10631 // Looking for "rev" which is V6+.
10632 if (!Subtarget->hasV6Ops())
10635 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
10636 std::string AsmStr = IA->getAsmString();
10637 SmallVector<StringRef, 4> AsmPieces;
10638 SplitString(AsmStr, AsmPieces, ";\n");
10640 switch (AsmPieces.size()) {
10641 default: return false;
10643 AsmStr = AsmPieces[0];
10645 SplitString(AsmStr, AsmPieces, " \t,");
10648 if (AsmPieces.size() == 3 &&
10649 AsmPieces[0] == "rev" && AsmPieces[1] == "$0" && AsmPieces[2] == "$1" &&
10650 IA->getConstraintString().compare(0, 4, "=l,l") == 0) {
10651 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
10652 if (Ty && Ty->getBitWidth() == 32)
10653 return IntrinsicLowering::LowerToByteSwap(CI);
10661 /// getConstraintType - Given a constraint letter, return the type of
10662 /// constraint it is for this target.
10663 ARMTargetLowering::ConstraintType
10664 ARMTargetLowering::getConstraintType(StringRef Constraint) const {
10665 if (Constraint.size() == 1) {
10666 switch (Constraint[0]) {
10668 case 'l': return C_RegisterClass;
10669 case 'w': return C_RegisterClass;
10670 case 'h': return C_RegisterClass;
10671 case 'x': return C_RegisterClass;
10672 case 't': return C_RegisterClass;
10673 case 'j': return C_Other; // Constant for movw.
10674 // An address with a single base register. Due to the way we
10675 // currently handle addresses it is the same as an 'r' memory constraint.
10676 case 'Q': return C_Memory;
10678 } else if (Constraint.size() == 2) {
10679 switch (Constraint[0]) {
10681 // All 'U+' constraints are addresses.
10682 case 'U': return C_Memory;
10685 return TargetLowering::getConstraintType(Constraint);
10688 /// Examine constraint type and operand type and determine a weight value.
10689 /// This object must already have been set up with the operand type
10690 /// and the current alternative constraint selected.
10691 TargetLowering::ConstraintWeight
10692 ARMTargetLowering::getSingleConstraintMatchWeight(
10693 AsmOperandInfo &info, const char *constraint) const {
10694 ConstraintWeight weight = CW_Invalid;
10695 Value *CallOperandVal = info.CallOperandVal;
10696 // If we don't have a value, we can't do a match,
10697 // but allow it at the lowest weight.
10698 if (!CallOperandVal)
10700 Type *type = CallOperandVal->getType();
10701 // Look at the constraint type.
10702 switch (*constraint) {
10704 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
10707 if (type->isIntegerTy()) {
10708 if (Subtarget->isThumb())
10709 weight = CW_SpecificReg;
10711 weight = CW_Register;
10715 if (type->isFloatingPointTy())
10716 weight = CW_Register;
10722 typedef std::pair<unsigned, const TargetRegisterClass*> RCPair;
10723 RCPair ARMTargetLowering::getRegForInlineAsmConstraint(
10724 const TargetRegisterInfo *TRI, StringRef Constraint, MVT VT) const {
10725 if (Constraint.size() == 1) {
10726 // GCC ARM Constraint Letters
10727 switch (Constraint[0]) {
10728 case 'l': // Low regs or general regs.
10729 if (Subtarget->isThumb())
10730 return RCPair(0U, &ARM::tGPRRegClass);
10731 return RCPair(0U, &ARM::GPRRegClass);
10732 case 'h': // High regs or no regs.
10733 if (Subtarget->isThumb())
10734 return RCPair(0U, &ARM::hGPRRegClass);
10737 if (Subtarget->isThumb1Only())
10738 return RCPair(0U, &ARM::tGPRRegClass);
10739 return RCPair(0U, &ARM::GPRRegClass);
10741 if (VT == MVT::Other)
10743 if (VT == MVT::f32)
10744 return RCPair(0U, &ARM::SPRRegClass);
10745 if (VT.getSizeInBits() == 64)
10746 return RCPair(0U, &ARM::DPRRegClass);
10747 if (VT.getSizeInBits() == 128)
10748 return RCPair(0U, &ARM::QPRRegClass);
10751 if (VT == MVT::Other)
10753 if (VT == MVT::f32)
10754 return RCPair(0U, &ARM::SPR_8RegClass);
10755 if (VT.getSizeInBits() == 64)
10756 return RCPair(0U, &ARM::DPR_8RegClass);
10757 if (VT.getSizeInBits() == 128)
10758 return RCPair(0U, &ARM::QPR_8RegClass);
10761 if (VT == MVT::f32)
10762 return RCPair(0U, &ARM::SPRRegClass);
10766 if (StringRef("{cc}").equals_lower(Constraint))
10767 return std::make_pair(unsigned(ARM::CPSR), &ARM::CCRRegClass);
10769 return TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
10772 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
10773 /// vector. If it is invalid, don't add anything to Ops.
10774 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
10775 std::string &Constraint,
10776 std::vector<SDValue>&Ops,
10777 SelectionDAG &DAG) const {
10780 // Currently only support length 1 constraints.
10781 if (Constraint.length() != 1) return;
10783 char ConstraintLetter = Constraint[0];
10784 switch (ConstraintLetter) {
10787 case 'I': case 'J': case 'K': case 'L':
10788 case 'M': case 'N': case 'O':
10789 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
10793 int64_t CVal64 = C->getSExtValue();
10794 int CVal = (int) CVal64;
10795 // None of these constraints allow values larger than 32 bits. Check
10796 // that the value fits in an int.
10797 if (CVal != CVal64)
10800 switch (ConstraintLetter) {
10802 // Constant suitable for movw, must be between 0 and
10804 if (Subtarget->hasV6T2Ops())
10805 if (CVal >= 0 && CVal <= 65535)
10809 if (Subtarget->isThumb1Only()) {
10810 // This must be a constant between 0 and 255, for ADD
10812 if (CVal >= 0 && CVal <= 255)
10814 } else if (Subtarget->isThumb2()) {
10815 // A constant that can be used as an immediate value in a
10816 // data-processing instruction.
10817 if (ARM_AM::getT2SOImmVal(CVal) != -1)
10820 // A constant that can be used as an immediate value in a
10821 // data-processing instruction.
10822 if (ARM_AM::getSOImmVal(CVal) != -1)
10828 if (Subtarget->isThumb()) { // FIXME thumb2
10829 // This must be a constant between -255 and -1, for negated ADD
10830 // immediates. This can be used in GCC with an "n" modifier that
10831 // prints the negated value, for use with SUB instructions. It is
10832 // not useful otherwise but is implemented for compatibility.
10833 if (CVal >= -255 && CVal <= -1)
10836 // This must be a constant between -4095 and 4095. It is not clear
10837 // what this constraint is intended for. Implemented for
10838 // compatibility with GCC.
10839 if (CVal >= -4095 && CVal <= 4095)
10845 if (Subtarget->isThumb1Only()) {
10846 // A 32-bit value where only one byte has a nonzero value. Exclude
10847 // zero to match GCC. This constraint is used by GCC internally for
10848 // constants that can be loaded with a move/shift combination.
10849 // It is not useful otherwise but is implemented for compatibility.
10850 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
10852 } else if (Subtarget->isThumb2()) {
10853 // A constant whose bitwise inverse can be used as an immediate
10854 // value in a data-processing instruction. This can be used in GCC
10855 // with a "B" modifier that prints the inverted value, for use with
10856 // BIC and MVN instructions. It is not useful otherwise but is
10857 // implemented for compatibility.
10858 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
10861 // A constant whose bitwise inverse can be used as an immediate
10862 // value in a data-processing instruction. This can be used in GCC
10863 // with a "B" modifier that prints the inverted value, for use with
10864 // BIC and MVN instructions. It is not useful otherwise but is
10865 // implemented for compatibility.
10866 if (ARM_AM::getSOImmVal(~CVal) != -1)
10872 if (Subtarget->isThumb1Only()) {
10873 // This must be a constant between -7 and 7,
10874 // for 3-operand ADD/SUB immediate instructions.
10875 if (CVal >= -7 && CVal < 7)
10877 } else if (Subtarget->isThumb2()) {
10878 // A constant whose negation can be used as an immediate value in a
10879 // data-processing instruction. This can be used in GCC with an "n"
10880 // modifier that prints the negated value, for use with SUB
10881 // instructions. It is not useful otherwise but is implemented for
10883 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
10886 // A constant whose negation can be used as an immediate value in a
10887 // data-processing instruction. This can be used in GCC with an "n"
10888 // modifier that prints the negated value, for use with SUB
10889 // instructions. It is not useful otherwise but is implemented for
10891 if (ARM_AM::getSOImmVal(-CVal) != -1)
10897 if (Subtarget->isThumb()) { // FIXME thumb2
10898 // This must be a multiple of 4 between 0 and 1020, for
10899 // ADD sp + immediate.
10900 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
10903 // A power of two or a constant between 0 and 32. This is used in
10904 // GCC for the shift amount on shifted register operands, but it is
10905 // useful in general for any shift amounts.
10906 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
10912 if (Subtarget->isThumb()) { // FIXME thumb2
10913 // This must be a constant between 0 and 31, for shift amounts.
10914 if (CVal >= 0 && CVal <= 31)
10920 if (Subtarget->isThumb()) { // FIXME thumb2
10921 // This must be a multiple of 4 between -508 and 508, for
10922 // ADD/SUB sp = sp + immediate.
10923 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
10928 Result = DAG.getTargetConstant(CVal, SDLoc(Op), Op.getValueType());
10932 if (Result.getNode()) {
10933 Ops.push_back(Result);
10936 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
10939 SDValue ARMTargetLowering::LowerDivRem(SDValue Op, SelectionDAG &DAG) const {
10940 assert(Subtarget->isTargetAEABI() && "Register-based DivRem lowering only");
10941 unsigned Opcode = Op->getOpcode();
10942 assert((Opcode == ISD::SDIVREM || Opcode == ISD::UDIVREM) &&
10943 "Invalid opcode for Div/Rem lowering");
10944 bool isSigned = (Opcode == ISD::SDIVREM);
10945 EVT VT = Op->getValueType(0);
10946 Type *Ty = VT.getTypeForEVT(*DAG.getContext());
10949 switch (VT.getSimpleVT().SimpleTy) {
10950 default: llvm_unreachable("Unexpected request for libcall!");
10951 case MVT::i8: LC = isSigned ? RTLIB::SDIVREM_I8 : RTLIB::UDIVREM_I8; break;
10952 case MVT::i16: LC = isSigned ? RTLIB::SDIVREM_I16 : RTLIB::UDIVREM_I16; break;
10953 case MVT::i32: LC = isSigned ? RTLIB::SDIVREM_I32 : RTLIB::UDIVREM_I32; break;
10954 case MVT::i64: LC = isSigned ? RTLIB::SDIVREM_I64 : RTLIB::UDIVREM_I64; break;
10957 SDValue InChain = DAG.getEntryNode();
10959 TargetLowering::ArgListTy Args;
10960 TargetLowering::ArgListEntry Entry;
10961 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i) {
10962 EVT ArgVT = Op->getOperand(i).getValueType();
10963 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
10964 Entry.Node = Op->getOperand(i);
10966 Entry.isSExt = isSigned;
10967 Entry.isZExt = !isSigned;
10968 Args.push_back(Entry);
10971 SDValue Callee = DAG.getExternalSymbol(getLibcallName(LC),
10974 Type *RetTy = (Type*)StructType::get(Ty, Ty, nullptr);
10977 TargetLowering::CallLoweringInfo CLI(DAG);
10978 CLI.setDebugLoc(dl).setChain(InChain)
10979 .setCallee(getLibcallCallingConv(LC), RetTy, Callee, std::move(Args), 0)
10980 .setInRegister().setSExtResult(isSigned).setZExtResult(!isSigned);
10982 std::pair<SDValue, SDValue> CallInfo = LowerCallTo(CLI);
10983 return CallInfo.first;
10987 ARMTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const {
10988 assert(Subtarget->isTargetWindows() && "unsupported target platform");
10992 SDValue Chain = Op.getOperand(0);
10993 SDValue Size = Op.getOperand(1);
10995 SDValue Words = DAG.getNode(ISD::SRL, DL, MVT::i32, Size,
10996 DAG.getConstant(2, DL, MVT::i32));
10999 Chain = DAG.getCopyToReg(Chain, DL, ARM::R4, Words, Flag);
11000 Flag = Chain.getValue(1);
11002 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
11003 Chain = DAG.getNode(ARMISD::WIN__CHKSTK, DL, NodeTys, Chain, Flag);
11005 SDValue NewSP = DAG.getCopyFromReg(Chain, DL, ARM::SP, MVT::i32);
11006 Chain = NewSP.getValue(1);
11008 SDValue Ops[2] = { NewSP, Chain };
11009 return DAG.getMergeValues(Ops, DL);
11012 SDValue ARMTargetLowering::LowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) const {
11013 assert(Op.getValueType() == MVT::f64 && Subtarget->isFPOnlySP() &&
11014 "Unexpected type for custom-lowering FP_EXTEND");
11017 LC = RTLIB::getFPEXT(Op.getOperand(0).getValueType(), Op.getValueType());
11019 SDValue SrcVal = Op.getOperand(0);
11020 return makeLibCall(DAG, LC, Op.getValueType(), &SrcVal, 1,
11021 /*isSigned*/ false, SDLoc(Op)).first;
11024 SDValue ARMTargetLowering::LowerFP_ROUND(SDValue Op, SelectionDAG &DAG) const {
11025 assert(Op.getOperand(0).getValueType() == MVT::f64 &&
11026 Subtarget->isFPOnlySP() &&
11027 "Unexpected type for custom-lowering FP_ROUND");
11030 LC = RTLIB::getFPROUND(Op.getOperand(0).getValueType(), Op.getValueType());
11032 SDValue SrcVal = Op.getOperand(0);
11033 return makeLibCall(DAG, LC, Op.getValueType(), &SrcVal, 1,
11034 /*isSigned*/ false, SDLoc(Op)).first;
11038 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
11039 // The ARM target isn't yet aware of offsets.
11043 bool ARM::isBitFieldInvertedMask(unsigned v) {
11044 if (v == 0xffffffff)
11047 // there can be 1's on either or both "outsides", all the "inside"
11048 // bits must be 0's
11049 return isShiftedMask_32(~v);
11052 /// isFPImmLegal - Returns true if the target can instruction select the
11053 /// specified FP immediate natively. If false, the legalizer will
11054 /// materialize the FP immediate as a load from a constant pool.
11055 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
11056 if (!Subtarget->hasVFP3())
11058 if (VT == MVT::f32)
11059 return ARM_AM::getFP32Imm(Imm) != -1;
11060 if (VT == MVT::f64 && !Subtarget->isFPOnlySP())
11061 return ARM_AM::getFP64Imm(Imm) != -1;
11065 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
11066 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
11067 /// specified in the intrinsic calls.
11068 bool ARMTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
11070 unsigned Intrinsic) const {
11071 switch (Intrinsic) {
11072 case Intrinsic::arm_neon_vld1:
11073 case Intrinsic::arm_neon_vld2:
11074 case Intrinsic::arm_neon_vld3:
11075 case Intrinsic::arm_neon_vld4:
11076 case Intrinsic::arm_neon_vld2lane:
11077 case Intrinsic::arm_neon_vld3lane:
11078 case Intrinsic::arm_neon_vld4lane: {
11079 Info.opc = ISD::INTRINSIC_W_CHAIN;
11080 // Conservatively set memVT to the entire set of vectors loaded.
11081 uint64_t NumElts = getDataLayout()->getTypeAllocSize(I.getType()) / 8;
11082 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
11083 Info.ptrVal = I.getArgOperand(0);
11085 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
11086 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
11087 Info.vol = false; // volatile loads with NEON intrinsics not supported
11088 Info.readMem = true;
11089 Info.writeMem = false;
11092 case Intrinsic::arm_neon_vst1:
11093 case Intrinsic::arm_neon_vst2:
11094 case Intrinsic::arm_neon_vst3:
11095 case Intrinsic::arm_neon_vst4:
11096 case Intrinsic::arm_neon_vst2lane:
11097 case Intrinsic::arm_neon_vst3lane:
11098 case Intrinsic::arm_neon_vst4lane: {
11099 Info.opc = ISD::INTRINSIC_VOID;
11100 // Conservatively set memVT to the entire set of vectors stored.
11101 unsigned NumElts = 0;
11102 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
11103 Type *ArgTy = I.getArgOperand(ArgI)->getType();
11104 if (!ArgTy->isVectorTy())
11106 NumElts += getDataLayout()->getTypeAllocSize(ArgTy) / 8;
11108 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
11109 Info.ptrVal = I.getArgOperand(0);
11111 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
11112 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
11113 Info.vol = false; // volatile stores with NEON intrinsics not supported
11114 Info.readMem = false;
11115 Info.writeMem = true;
11118 case Intrinsic::arm_ldaex:
11119 case Intrinsic::arm_ldrex: {
11120 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(0)->getType());
11121 Info.opc = ISD::INTRINSIC_W_CHAIN;
11122 Info.memVT = MVT::getVT(PtrTy->getElementType());
11123 Info.ptrVal = I.getArgOperand(0);
11125 Info.align = getDataLayout()->getABITypeAlignment(PtrTy->getElementType());
11127 Info.readMem = true;
11128 Info.writeMem = false;
11131 case Intrinsic::arm_stlex:
11132 case Intrinsic::arm_strex: {
11133 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(1)->getType());
11134 Info.opc = ISD::INTRINSIC_W_CHAIN;
11135 Info.memVT = MVT::getVT(PtrTy->getElementType());
11136 Info.ptrVal = I.getArgOperand(1);
11138 Info.align = getDataLayout()->getABITypeAlignment(PtrTy->getElementType());
11140 Info.readMem = false;
11141 Info.writeMem = true;
11144 case Intrinsic::arm_stlexd:
11145 case Intrinsic::arm_strexd: {
11146 Info.opc = ISD::INTRINSIC_W_CHAIN;
11147 Info.memVT = MVT::i64;
11148 Info.ptrVal = I.getArgOperand(2);
11152 Info.readMem = false;
11153 Info.writeMem = true;
11156 case Intrinsic::arm_ldaexd:
11157 case Intrinsic::arm_ldrexd: {
11158 Info.opc = ISD::INTRINSIC_W_CHAIN;
11159 Info.memVT = MVT::i64;
11160 Info.ptrVal = I.getArgOperand(0);
11164 Info.readMem = true;
11165 Info.writeMem = false;
11175 /// \brief Returns true if it is beneficial to convert a load of a constant
11176 /// to just the constant itself.
11177 bool ARMTargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
11179 assert(Ty->isIntegerTy());
11181 unsigned Bits = Ty->getPrimitiveSizeInBits();
11182 if (Bits == 0 || Bits > 32)
11187 bool ARMTargetLowering::hasLoadLinkedStoreConditional() const { return true; }
11189 Instruction* ARMTargetLowering::makeDMB(IRBuilder<> &Builder,
11190 ARM_MB::MemBOpt Domain) const {
11191 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
11193 // First, if the target has no DMB, see what fallback we can use.
11194 if (!Subtarget->hasDataBarrier()) {
11195 // Some ARMv6 cpus can support data barriers with an mcr instruction.
11196 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
11198 if (Subtarget->hasV6Ops() && !Subtarget->isThumb()) {
11199 Function *MCR = llvm::Intrinsic::getDeclaration(M, Intrinsic::arm_mcr);
11200 Value* args[6] = {Builder.getInt32(15), Builder.getInt32(0),
11201 Builder.getInt32(0), Builder.getInt32(7),
11202 Builder.getInt32(10), Builder.getInt32(5)};
11203 return Builder.CreateCall(MCR, args);
11205 // Instead of using barriers, atomic accesses on these subtargets use
11207 llvm_unreachable("makeDMB on a target so old that it has no barriers");
11210 Function *DMB = llvm::Intrinsic::getDeclaration(M, Intrinsic::arm_dmb);
11211 // Only a full system barrier exists in the M-class architectures.
11212 Domain = Subtarget->isMClass() ? ARM_MB::SY : Domain;
11213 Constant *CDomain = Builder.getInt32(Domain);
11214 return Builder.CreateCall(DMB, CDomain);
11218 // Based on http://www.cl.cam.ac.uk/~pes20/cpp/cpp0xmappings.html
11219 Instruction* ARMTargetLowering::emitLeadingFence(IRBuilder<> &Builder,
11220 AtomicOrdering Ord, bool IsStore,
11221 bool IsLoad) const {
11222 if (!getInsertFencesForAtomic())
11228 llvm_unreachable("Invalid fence: unordered/non-atomic");
11231 return nullptr; // Nothing to do
11232 case SequentiallyConsistent:
11234 return nullptr; // Nothing to do
11237 case AcquireRelease:
11238 if (Subtarget->isSwift())
11239 return makeDMB(Builder, ARM_MB::ISHST);
11240 // FIXME: add a comment with a link to documentation justifying this.
11242 return makeDMB(Builder, ARM_MB::ISH);
11244 llvm_unreachable("Unknown fence ordering in emitLeadingFence");
11247 Instruction* ARMTargetLowering::emitTrailingFence(IRBuilder<> &Builder,
11248 AtomicOrdering Ord, bool IsStore,
11249 bool IsLoad) const {
11250 if (!getInsertFencesForAtomic())
11256 llvm_unreachable("Invalid fence: unordered/not-atomic");
11259 return nullptr; // Nothing to do
11261 case AcquireRelease:
11262 case SequentiallyConsistent:
11263 return makeDMB(Builder, ARM_MB::ISH);
11265 llvm_unreachable("Unknown fence ordering in emitTrailingFence");
11268 // Loads and stores less than 64-bits are already atomic; ones above that
11269 // are doomed anyway, so defer to the default libcall and blame the OS when
11270 // things go wrong. Cortex M doesn't have ldrexd/strexd though, so don't emit
11271 // anything for those.
11272 bool ARMTargetLowering::shouldExpandAtomicStoreInIR(StoreInst *SI) const {
11273 unsigned Size = SI->getValueOperand()->getType()->getPrimitiveSizeInBits();
11274 return (Size == 64) && !Subtarget->isMClass();
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 // FIXME: ldrd and strd are atomic if the CPU has LPAE (e.g. A15 has that
11282 // guarantee, see DDI0406C ARM architecture reference manual,
11283 // sections A8.8.72-74 LDRD)
11284 bool ARMTargetLowering::shouldExpandAtomicLoadInIR(LoadInst *LI) const {
11285 unsigned Size = LI->getType()->getPrimitiveSizeInBits();
11286 return (Size == 64) && !Subtarget->isMClass();
11289 // For the real atomic operations, we have ldrex/strex up to 32 bits,
11290 // and up to 64 bits on the non-M profiles
11291 TargetLoweringBase::AtomicRMWExpansionKind
11292 ARMTargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
11293 unsigned Size = AI->getType()->getPrimitiveSizeInBits();
11294 return (Size <= (Subtarget->isMClass() ? 32U : 64U))
11295 ? AtomicRMWExpansionKind::LLSC
11296 : AtomicRMWExpansionKind::None;
11299 // This has so far only been implemented for MachO.
11300 bool ARMTargetLowering::useLoadStackGuardNode() const {
11301 return Subtarget->isTargetMachO();
11304 bool ARMTargetLowering::canCombineStoreAndExtract(Type *VectorTy, Value *Idx,
11305 unsigned &Cost) const {
11306 // If we do not have NEON, vector types are not natively supported.
11307 if (!Subtarget->hasNEON())
11310 // Floating point values and vector values map to the same register file.
11311 // Therefore, althought we could do a store extract of a vector type, this is
11312 // better to leave at float as we have more freedom in the addressing mode for
11314 if (VectorTy->isFPOrFPVectorTy())
11317 // If the index is unknown at compile time, this is very expensive to lower
11318 // and it is not possible to combine the store with the extract.
11319 if (!isa<ConstantInt>(Idx))
11322 assert(VectorTy->isVectorTy() && "VectorTy is not a vector type");
11323 unsigned BitWidth = cast<VectorType>(VectorTy)->getBitWidth();
11324 // We can do a store + vector extract on any vector that fits perfectly in a D
11326 if (BitWidth == 64 || BitWidth == 128) {
11333 Value *ARMTargetLowering::emitLoadLinked(IRBuilder<> &Builder, Value *Addr,
11334 AtomicOrdering Ord) const {
11335 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
11336 Type *ValTy = cast<PointerType>(Addr->getType())->getElementType();
11337 bool IsAcquire = isAtLeastAcquire(Ord);
11339 // Since i64 isn't legal and intrinsics don't get type-lowered, the ldrexd
11340 // intrinsic must return {i32, i32} and we have to recombine them into a
11341 // single i64 here.
11342 if (ValTy->getPrimitiveSizeInBits() == 64) {
11343 Intrinsic::ID Int =
11344 IsAcquire ? Intrinsic::arm_ldaexd : Intrinsic::arm_ldrexd;
11345 Function *Ldrex = llvm::Intrinsic::getDeclaration(M, Int);
11347 Addr = Builder.CreateBitCast(Addr, Type::getInt8PtrTy(M->getContext()));
11348 Value *LoHi = Builder.CreateCall(Ldrex, Addr, "lohi");
11350 Value *Lo = Builder.CreateExtractValue(LoHi, 0, "lo");
11351 Value *Hi = Builder.CreateExtractValue(LoHi, 1, "hi");
11352 if (!Subtarget->isLittle())
11353 std::swap (Lo, Hi);
11354 Lo = Builder.CreateZExt(Lo, ValTy, "lo64");
11355 Hi = Builder.CreateZExt(Hi, ValTy, "hi64");
11356 return Builder.CreateOr(
11357 Lo, Builder.CreateShl(Hi, ConstantInt::get(ValTy, 32)), "val64");
11360 Type *Tys[] = { Addr->getType() };
11361 Intrinsic::ID Int = IsAcquire ? Intrinsic::arm_ldaex : Intrinsic::arm_ldrex;
11362 Function *Ldrex = llvm::Intrinsic::getDeclaration(M, Int, Tys);
11364 return Builder.CreateTruncOrBitCast(
11365 Builder.CreateCall(Ldrex, Addr),
11366 cast<PointerType>(Addr->getType())->getElementType());
11369 Value *ARMTargetLowering::emitStoreConditional(IRBuilder<> &Builder, Value *Val,
11371 AtomicOrdering Ord) const {
11372 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
11373 bool IsRelease = isAtLeastRelease(Ord);
11375 // Since the intrinsics must have legal type, the i64 intrinsics take two
11376 // parameters: "i32, i32". We must marshal Val into the appropriate form
11377 // before the call.
11378 if (Val->getType()->getPrimitiveSizeInBits() == 64) {
11379 Intrinsic::ID Int =
11380 IsRelease ? Intrinsic::arm_stlexd : Intrinsic::arm_strexd;
11381 Function *Strex = Intrinsic::getDeclaration(M, Int);
11382 Type *Int32Ty = Type::getInt32Ty(M->getContext());
11384 Value *Lo = Builder.CreateTrunc(Val, Int32Ty, "lo");
11385 Value *Hi = Builder.CreateTrunc(Builder.CreateLShr(Val, 32), Int32Ty, "hi");
11386 if (!Subtarget->isLittle())
11387 std::swap (Lo, Hi);
11388 Addr = Builder.CreateBitCast(Addr, Type::getInt8PtrTy(M->getContext()));
11389 return Builder.CreateCall(Strex, {Lo, Hi, Addr});
11392 Intrinsic::ID Int = IsRelease ? Intrinsic::arm_stlex : Intrinsic::arm_strex;
11393 Type *Tys[] = { Addr->getType() };
11394 Function *Strex = Intrinsic::getDeclaration(M, Int, Tys);
11396 return Builder.CreateCall(
11397 Strex, {Builder.CreateZExtOrBitCast(
11398 Val, Strex->getFunctionType()->getParamType(0)),
11402 /// \brief Lower an interleaved load into a vldN intrinsic.
11404 /// E.g. Lower an interleaved load (Factor = 2):
11405 /// %wide.vec = load <8 x i32>, <8 x i32>* %ptr, align 4
11406 /// %v0 = shuffle %wide.vec, undef, <0, 2, 4, 6> ; Extract even elements
11407 /// %v1 = shuffle %wide.vec, undef, <1, 3, 5, 7> ; Extract odd elements
11410 /// %vld2 = { <4 x i32>, <4 x i32> } call llvm.arm.neon.vld2(%ptr, 4)
11411 /// %vec0 = extractelement { <4 x i32>, <4 x i32> } %vld2, i32 0
11412 /// %vec1 = extractelement { <4 x i32>, <4 x i32> } %vld2, i32 1
11413 bool ARMTargetLowering::lowerInterleavedLoad(
11414 LoadInst *LI, ArrayRef<ShuffleVectorInst *> Shuffles,
11415 ArrayRef<unsigned> Indices, unsigned Factor) const {
11416 assert(Factor >= 2 && Factor <= getMaxSupportedInterleaveFactor() &&
11417 "Invalid interleave factor");
11418 assert(!Shuffles.empty() && "Empty shufflevector input");
11419 assert(Shuffles.size() == Indices.size() &&
11420 "Unmatched number of shufflevectors and indices");
11422 VectorType *VecTy = Shuffles[0]->getType();
11423 Type *EltTy = VecTy->getVectorElementType();
11425 const DataLayout *DL = getDataLayout();
11426 unsigned VecSize = DL->getTypeAllocSizeInBits(VecTy);
11427 bool EltIs64Bits = DL->getTypeAllocSizeInBits(EltTy) == 64;
11429 // Skip illegal vector types and vector types of i64/f64 element (vldN doesn't
11430 // support i64/f64 element).
11431 if ((VecSize != 64 && VecSize != 128) || EltIs64Bits)
11434 // A pointer vector can not be the return type of the ldN intrinsics. Need to
11435 // load integer vectors first and then convert to pointer vectors.
11436 if (EltTy->isPointerTy())
11437 VecTy = VectorType::get(DL->getIntPtrType(EltTy),
11438 VecTy->getVectorNumElements());
11440 static const Intrinsic::ID LoadInts[3] = {Intrinsic::arm_neon_vld2,
11441 Intrinsic::arm_neon_vld3,
11442 Intrinsic::arm_neon_vld4};
11444 Function *VldnFunc =
11445 Intrinsic::getDeclaration(LI->getModule(), LoadInts[Factor - 2], VecTy);
11447 IRBuilder<> Builder(LI);
11448 SmallVector<Value *, 2> Ops;
11450 Type *Int8Ptr = Builder.getInt8PtrTy(LI->getPointerAddressSpace());
11451 Ops.push_back(Builder.CreateBitCast(LI->getPointerOperand(), Int8Ptr));
11452 Ops.push_back(Builder.getInt32(LI->getAlignment()));
11454 CallInst *VldN = Builder.CreateCall(VldnFunc, Ops, "vldN");
11456 // Replace uses of each shufflevector with the corresponding vector loaded
11458 for (unsigned i = 0; i < Shuffles.size(); i++) {
11459 ShuffleVectorInst *SV = Shuffles[i];
11460 unsigned Index = Indices[i];
11462 Value *SubVec = Builder.CreateExtractValue(VldN, Index);
11464 // Convert the integer vector to pointer vector if the element is pointer.
11465 if (EltTy->isPointerTy())
11466 SubVec = Builder.CreateIntToPtr(SubVec, SV->getType());
11468 SV->replaceAllUsesWith(SubVec);
11474 /// \brief Get a mask consisting of sequential integers starting from \p Start.
11476 /// I.e. <Start, Start + 1, ..., Start + NumElts - 1>
11477 static Constant *getSequentialMask(IRBuilder<> &Builder, unsigned Start,
11478 unsigned NumElts) {
11479 SmallVector<Constant *, 16> Mask;
11480 for (unsigned i = 0; i < NumElts; i++)
11481 Mask.push_back(Builder.getInt32(Start + i));
11483 return ConstantVector::get(Mask);
11486 /// \brief Lower an interleaved store into a vstN intrinsic.
11488 /// E.g. Lower an interleaved store (Factor = 3):
11489 /// %i.vec = shuffle <8 x i32> %v0, <8 x i32> %v1,
11490 /// <0, 4, 8, 1, 5, 9, 2, 6, 10, 3, 7, 11>
11491 /// store <12 x i32> %i.vec, <12 x i32>* %ptr, align 4
11494 /// %sub.v0 = shuffle <8 x i32> %v0, <8 x i32> v1, <0, 1, 2, 3>
11495 /// %sub.v1 = shuffle <8 x i32> %v0, <8 x i32> v1, <4, 5, 6, 7>
11496 /// %sub.v2 = shuffle <8 x i32> %v0, <8 x i32> v1, <8, 9, 10, 11>
11497 /// call void llvm.arm.neon.vst3(%ptr, %sub.v0, %sub.v1, %sub.v2, 4)
11499 /// Note that the new shufflevectors will be removed and we'll only generate one
11500 /// vst3 instruction in CodeGen.
11501 bool ARMTargetLowering::lowerInterleavedStore(StoreInst *SI,
11502 ShuffleVectorInst *SVI,
11503 unsigned Factor) const {
11504 assert(Factor >= 2 && Factor <= getMaxSupportedInterleaveFactor() &&
11505 "Invalid interleave factor");
11507 VectorType *VecTy = SVI->getType();
11508 assert(VecTy->getVectorNumElements() % Factor == 0 &&
11509 "Invalid interleaved store");
11511 unsigned NumSubElts = VecTy->getVectorNumElements() / Factor;
11512 Type *EltTy = VecTy->getVectorElementType();
11513 VectorType *SubVecTy = VectorType::get(EltTy, NumSubElts);
11515 const DataLayout *DL = getDataLayout();
11516 unsigned SubVecSize = DL->getTypeAllocSizeInBits(SubVecTy);
11517 bool EltIs64Bits = DL->getTypeAllocSizeInBits(EltTy) == 64;
11519 // Skip illegal sub vector types and vector types of i64/f64 element (vstN
11520 // doesn't support i64/f64 element).
11521 if ((SubVecSize != 64 && SubVecSize != 128) || EltIs64Bits)
11524 Value *Op0 = SVI->getOperand(0);
11525 Value *Op1 = SVI->getOperand(1);
11526 IRBuilder<> Builder(SI);
11528 // StN intrinsics don't support pointer vectors as arguments. Convert pointer
11529 // vectors to integer vectors.
11530 if (EltTy->isPointerTy()) {
11531 Type *IntTy = DL->getIntPtrType(EltTy);
11533 // Convert to the corresponding integer vector.
11535 VectorType::get(IntTy, Op0->getType()->getVectorNumElements());
11536 Op0 = Builder.CreatePtrToInt(Op0, IntVecTy);
11537 Op1 = Builder.CreatePtrToInt(Op1, IntVecTy);
11539 SubVecTy = VectorType::get(IntTy, NumSubElts);
11542 static Intrinsic::ID StoreInts[3] = {Intrinsic::arm_neon_vst2,
11543 Intrinsic::arm_neon_vst3,
11544 Intrinsic::arm_neon_vst4};
11545 Function *VstNFunc = Intrinsic::getDeclaration(
11546 SI->getModule(), StoreInts[Factor - 2], SubVecTy);
11548 SmallVector<Value *, 6> Ops;
11550 Type *Int8Ptr = Builder.getInt8PtrTy(SI->getPointerAddressSpace());
11551 Ops.push_back(Builder.CreateBitCast(SI->getPointerOperand(), Int8Ptr));
11553 // Split the shufflevector operands into sub vectors for the new vstN call.
11554 for (unsigned i = 0; i < Factor; i++)
11555 Ops.push_back(Builder.CreateShuffleVector(
11556 Op0, Op1, getSequentialMask(Builder, NumSubElts * i, NumSubElts)));
11558 Ops.push_back(Builder.getInt32(SI->getAlignment()));
11559 Builder.CreateCall(VstNFunc, Ops);
11571 static bool isHomogeneousAggregate(Type *Ty, HABaseType &Base,
11572 uint64_t &Members) {
11573 if (const StructType *ST = dyn_cast<StructType>(Ty)) {
11574 for (unsigned i = 0; i < ST->getNumElements(); ++i) {
11575 uint64_t SubMembers = 0;
11576 if (!isHomogeneousAggregate(ST->getElementType(i), Base, SubMembers))
11578 Members += SubMembers;
11580 } else if (const ArrayType *AT = dyn_cast<ArrayType>(Ty)) {
11581 uint64_t SubMembers = 0;
11582 if (!isHomogeneousAggregate(AT->getElementType(), Base, SubMembers))
11584 Members += SubMembers * AT->getNumElements();
11585 } else if (Ty->isFloatTy()) {
11586 if (Base != HA_UNKNOWN && Base != HA_FLOAT)
11590 } else if (Ty->isDoubleTy()) {
11591 if (Base != HA_UNKNOWN && Base != HA_DOUBLE)
11595 } else if (const VectorType *VT = dyn_cast<VectorType>(Ty)) {
11602 return VT->getBitWidth() == 64;
11604 return VT->getBitWidth() == 128;
11606 switch (VT->getBitWidth()) {
11619 return (Members > 0 && Members <= 4);
11622 /// \brief Return true if a type is an AAPCS-VFP homogeneous aggregate or one of
11623 /// [N x i32] or [N x i64]. This allows front-ends to skip emitting padding when
11624 /// passing according to AAPCS rules.
11625 bool ARMTargetLowering::functionArgumentNeedsConsecutiveRegisters(
11626 Type *Ty, CallingConv::ID CallConv, bool isVarArg) const {
11627 if (getEffectiveCallingConv(CallConv, isVarArg) !=
11628 CallingConv::ARM_AAPCS_VFP)
11631 HABaseType Base = HA_UNKNOWN;
11632 uint64_t Members = 0;
11633 bool IsHA = isHomogeneousAggregate(Ty, Base, Members);
11634 DEBUG(dbgs() << "isHA: " << IsHA << " "; Ty->dump());
11636 bool IsIntArray = Ty->isArrayTy() && Ty->getArrayElementType()->isIntegerTy();
11637 return IsHA || IsIntArray;