1 //===-- ARMISelLowering.cpp - ARM DAG Lowering Implementation -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the interfaces that ARM uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
15 #include "ARMISelLowering.h"
16 #include "ARMCallingConv.h"
17 #include "ARMConstantPoolValue.h"
18 #include "ARMMachineFunctionInfo.h"
19 #include "ARMPerfectShuffle.h"
20 #include "ARMSubtarget.h"
21 #include "ARMTargetMachine.h"
22 #include "ARMTargetObjectFile.h"
23 #include "MCTargetDesc/ARMAddressingModes.h"
24 #include "llvm/ADT/Statistic.h"
25 #include "llvm/ADT/StringExtras.h"
26 #include "llvm/CodeGen/CallingConvLower.h"
27 #include "llvm/CodeGen/IntrinsicLowering.h"
28 #include "llvm/CodeGen/MachineBasicBlock.h"
29 #include "llvm/CodeGen/MachineFrameInfo.h"
30 #include "llvm/CodeGen/MachineFunction.h"
31 #include "llvm/CodeGen/MachineInstrBuilder.h"
32 #include "llvm/CodeGen/MachineJumpTableInfo.h"
33 #include "llvm/CodeGen/MachineModuleInfo.h"
34 #include "llvm/CodeGen/MachineRegisterInfo.h"
35 #include "llvm/CodeGen/SelectionDAG.h"
36 #include "llvm/IR/CallingConv.h"
37 #include "llvm/IR/Constants.h"
38 #include "llvm/IR/Function.h"
39 #include "llvm/IR/GlobalValue.h"
40 #include "llvm/IR/IRBuilder.h"
41 #include "llvm/IR/Instruction.h"
42 #include "llvm/IR/Instructions.h"
43 #include "llvm/IR/Intrinsics.h"
44 #include "llvm/IR/Type.h"
45 #include "llvm/MC/MCSectionMachO.h"
46 #include "llvm/Support/CommandLine.h"
47 #include "llvm/Support/Debug.h"
48 #include "llvm/Support/ErrorHandling.h"
49 #include "llvm/Support/MathExtras.h"
50 #include "llvm/Target/TargetOptions.h"
54 #define DEBUG_TYPE "arm-isel"
56 STATISTIC(NumTailCalls, "Number of tail calls");
57 STATISTIC(NumMovwMovt, "Number of GAs materialized with movw + movt");
58 STATISTIC(NumLoopByVals, "Number of loops generated for byval arguments");
61 EnableARMLongCalls("arm-long-calls", cl::Hidden,
62 cl::desc("Generate calls via indirect call instructions"),
66 ARMInterworking("arm-interworking", cl::Hidden,
67 cl::desc("Enable / disable ARM interworking (for debugging only)"),
71 class ARMCCState : public CCState {
73 ARMCCState(CallingConv::ID CC, bool isVarArg, MachineFunction &MF,
74 SmallVectorImpl<CCValAssign> &locs, LLVMContext &C,
76 : CCState(CC, isVarArg, MF, locs, C) {
77 assert(((PC == Call) || (PC == Prologue)) &&
78 "ARMCCState users must specify whether their context is call"
79 "or prologue generation.");
85 // The APCS parameter registers.
86 static const MCPhysReg GPRArgRegs[] = {
87 ARM::R0, ARM::R1, ARM::R2, ARM::R3
90 void ARMTargetLowering::addTypeForNEON(MVT VT, MVT PromotedLdStVT,
91 MVT PromotedBitwiseVT) {
92 if (VT != PromotedLdStVT) {
93 setOperationAction(ISD::LOAD, VT, Promote);
94 AddPromotedToType (ISD::LOAD, VT, PromotedLdStVT);
96 setOperationAction(ISD::STORE, VT, Promote);
97 AddPromotedToType (ISD::STORE, VT, PromotedLdStVT);
100 MVT ElemTy = VT.getVectorElementType();
101 if (ElemTy != MVT::i64 && ElemTy != MVT::f64)
102 setOperationAction(ISD::SETCC, VT, Custom);
103 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
104 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
105 if (ElemTy == MVT::i32) {
106 setOperationAction(ISD::SINT_TO_FP, VT, Custom);
107 setOperationAction(ISD::UINT_TO_FP, VT, Custom);
108 setOperationAction(ISD::FP_TO_SINT, VT, Custom);
109 setOperationAction(ISD::FP_TO_UINT, VT, Custom);
111 setOperationAction(ISD::SINT_TO_FP, VT, Expand);
112 setOperationAction(ISD::UINT_TO_FP, VT, Expand);
113 setOperationAction(ISD::FP_TO_SINT, VT, Expand);
114 setOperationAction(ISD::FP_TO_UINT, VT, Expand);
116 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
117 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
118 setOperationAction(ISD::CONCAT_VECTORS, VT, Legal);
119 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Legal);
120 setOperationAction(ISD::SELECT, VT, Expand);
121 setOperationAction(ISD::SELECT_CC, VT, Expand);
122 setOperationAction(ISD::VSELECT, VT, Expand);
123 setOperationAction(ISD::SIGN_EXTEND_INREG, VT, Expand);
124 if (VT.isInteger()) {
125 setOperationAction(ISD::SHL, VT, Custom);
126 setOperationAction(ISD::SRA, VT, Custom);
127 setOperationAction(ISD::SRL, VT, Custom);
130 // Promote all bit-wise operations.
131 if (VT.isInteger() && VT != PromotedBitwiseVT) {
132 setOperationAction(ISD::AND, VT, Promote);
133 AddPromotedToType (ISD::AND, VT, PromotedBitwiseVT);
134 setOperationAction(ISD::OR, VT, Promote);
135 AddPromotedToType (ISD::OR, VT, PromotedBitwiseVT);
136 setOperationAction(ISD::XOR, VT, Promote);
137 AddPromotedToType (ISD::XOR, VT, PromotedBitwiseVT);
140 // Neon does not support vector divide/remainder operations.
141 setOperationAction(ISD::SDIV, VT, Expand);
142 setOperationAction(ISD::UDIV, VT, Expand);
143 setOperationAction(ISD::FDIV, VT, Expand);
144 setOperationAction(ISD::SREM, VT, Expand);
145 setOperationAction(ISD::UREM, VT, Expand);
146 setOperationAction(ISD::FREM, VT, Expand);
149 void ARMTargetLowering::addDRTypeForNEON(MVT VT) {
150 addRegisterClass(VT, &ARM::DPRRegClass);
151 addTypeForNEON(VT, MVT::f64, MVT::v2i32);
154 void ARMTargetLowering::addQRTypeForNEON(MVT VT) {
155 addRegisterClass(VT, &ARM::DPairRegClass);
156 addTypeForNEON(VT, MVT::v2f64, MVT::v4i32);
159 static TargetLoweringObjectFile *createTLOF(const Triple &TT) {
160 if (TT.isOSBinFormatMachO())
161 return new TargetLoweringObjectFileMachO();
162 if (TT.isOSWindows())
163 return new TargetLoweringObjectFileCOFF();
164 return new ARMElfTargetObjectFile();
167 ARMTargetLowering::ARMTargetLowering(const TargetMachine &TM)
168 : TargetLowering(TM, createTLOF(Triple(TM.getTargetTriple()))) {
169 Subtarget = &TM.getSubtarget<ARMSubtarget>();
170 RegInfo = TM.getSubtargetImpl()->getRegisterInfo();
171 Itins = TM.getSubtargetImpl()->getInstrItineraryData();
173 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
175 if (Subtarget->isTargetMachO()) {
176 // Uses VFP for Thumb libfuncs if available.
177 if (Subtarget->isThumb() && Subtarget->hasVFP2() &&
178 Subtarget->hasARMOps() && !TM.Options.UseSoftFloat) {
179 // Single-precision floating-point arithmetic.
180 setLibcallName(RTLIB::ADD_F32, "__addsf3vfp");
181 setLibcallName(RTLIB::SUB_F32, "__subsf3vfp");
182 setLibcallName(RTLIB::MUL_F32, "__mulsf3vfp");
183 setLibcallName(RTLIB::DIV_F32, "__divsf3vfp");
185 // Double-precision floating-point arithmetic.
186 setLibcallName(RTLIB::ADD_F64, "__adddf3vfp");
187 setLibcallName(RTLIB::SUB_F64, "__subdf3vfp");
188 setLibcallName(RTLIB::MUL_F64, "__muldf3vfp");
189 setLibcallName(RTLIB::DIV_F64, "__divdf3vfp");
191 // Single-precision comparisons.
192 setLibcallName(RTLIB::OEQ_F32, "__eqsf2vfp");
193 setLibcallName(RTLIB::UNE_F32, "__nesf2vfp");
194 setLibcallName(RTLIB::OLT_F32, "__ltsf2vfp");
195 setLibcallName(RTLIB::OLE_F32, "__lesf2vfp");
196 setLibcallName(RTLIB::OGE_F32, "__gesf2vfp");
197 setLibcallName(RTLIB::OGT_F32, "__gtsf2vfp");
198 setLibcallName(RTLIB::UO_F32, "__unordsf2vfp");
199 setLibcallName(RTLIB::O_F32, "__unordsf2vfp");
201 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
202 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETNE);
203 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
204 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
205 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
206 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
207 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
208 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
210 // Double-precision comparisons.
211 setLibcallName(RTLIB::OEQ_F64, "__eqdf2vfp");
212 setLibcallName(RTLIB::UNE_F64, "__nedf2vfp");
213 setLibcallName(RTLIB::OLT_F64, "__ltdf2vfp");
214 setLibcallName(RTLIB::OLE_F64, "__ledf2vfp");
215 setLibcallName(RTLIB::OGE_F64, "__gedf2vfp");
216 setLibcallName(RTLIB::OGT_F64, "__gtdf2vfp");
217 setLibcallName(RTLIB::UO_F64, "__unorddf2vfp");
218 setLibcallName(RTLIB::O_F64, "__unorddf2vfp");
220 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
221 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETNE);
222 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
223 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
224 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
225 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
226 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
227 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
229 // Floating-point to integer conversions.
230 // i64 conversions are done via library routines even when generating VFP
231 // instructions, so use the same ones.
232 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__fixdfsivfp");
233 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__fixunsdfsivfp");
234 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__fixsfsivfp");
235 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__fixunssfsivfp");
237 // Conversions between floating types.
238 setLibcallName(RTLIB::FPROUND_F64_F32, "__truncdfsf2vfp");
239 setLibcallName(RTLIB::FPEXT_F32_F64, "__extendsfdf2vfp");
241 // Integer to floating-point conversions.
242 // i64 conversions are done via library routines even when generating VFP
243 // instructions, so use the same ones.
244 // FIXME: There appears to be some naming inconsistency in ARM libgcc:
245 // e.g., __floatunsidf vs. __floatunssidfvfp.
246 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__floatsidfvfp");
247 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__floatunssidfvfp");
248 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__floatsisfvfp");
249 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__floatunssisfvfp");
253 // These libcalls are not available in 32-bit.
254 setLibcallName(RTLIB::SHL_I128, nullptr);
255 setLibcallName(RTLIB::SRL_I128, nullptr);
256 setLibcallName(RTLIB::SRA_I128, nullptr);
258 if (Subtarget->isAAPCS_ABI() && !Subtarget->isTargetMachO() &&
259 !Subtarget->isTargetWindows()) {
260 static const struct {
261 const RTLIB::Libcall Op;
262 const char * const Name;
263 const CallingConv::ID CC;
264 const ISD::CondCode Cond;
266 // Double-precision floating-point arithmetic helper functions
267 // RTABI chapter 4.1.2, Table 2
268 { RTLIB::ADD_F64, "__aeabi_dadd", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
269 { RTLIB::DIV_F64, "__aeabi_ddiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
270 { RTLIB::MUL_F64, "__aeabi_dmul", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
271 { RTLIB::SUB_F64, "__aeabi_dsub", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
273 // Double-precision floating-point comparison helper functions
274 // RTABI chapter 4.1.2, Table 3
275 { RTLIB::OEQ_F64, "__aeabi_dcmpeq", CallingConv::ARM_AAPCS, ISD::SETNE },
276 { RTLIB::UNE_F64, "__aeabi_dcmpeq", CallingConv::ARM_AAPCS, ISD::SETEQ },
277 { RTLIB::OLT_F64, "__aeabi_dcmplt", CallingConv::ARM_AAPCS, ISD::SETNE },
278 { RTLIB::OLE_F64, "__aeabi_dcmple", CallingConv::ARM_AAPCS, ISD::SETNE },
279 { RTLIB::OGE_F64, "__aeabi_dcmpge", CallingConv::ARM_AAPCS, ISD::SETNE },
280 { RTLIB::OGT_F64, "__aeabi_dcmpgt", CallingConv::ARM_AAPCS, ISD::SETNE },
281 { RTLIB::UO_F64, "__aeabi_dcmpun", CallingConv::ARM_AAPCS, ISD::SETNE },
282 { RTLIB::O_F64, "__aeabi_dcmpun", CallingConv::ARM_AAPCS, ISD::SETEQ },
284 // Single-precision floating-point arithmetic helper functions
285 // RTABI chapter 4.1.2, Table 4
286 { RTLIB::ADD_F32, "__aeabi_fadd", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
287 { RTLIB::DIV_F32, "__aeabi_fdiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
288 { RTLIB::MUL_F32, "__aeabi_fmul", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
289 { RTLIB::SUB_F32, "__aeabi_fsub", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
291 // Single-precision floating-point comparison helper functions
292 // RTABI chapter 4.1.2, Table 5
293 { RTLIB::OEQ_F32, "__aeabi_fcmpeq", CallingConv::ARM_AAPCS, ISD::SETNE },
294 { RTLIB::UNE_F32, "__aeabi_fcmpeq", CallingConv::ARM_AAPCS, ISD::SETEQ },
295 { RTLIB::OLT_F32, "__aeabi_fcmplt", CallingConv::ARM_AAPCS, ISD::SETNE },
296 { RTLIB::OLE_F32, "__aeabi_fcmple", CallingConv::ARM_AAPCS, ISD::SETNE },
297 { RTLIB::OGE_F32, "__aeabi_fcmpge", CallingConv::ARM_AAPCS, ISD::SETNE },
298 { RTLIB::OGT_F32, "__aeabi_fcmpgt", CallingConv::ARM_AAPCS, ISD::SETNE },
299 { RTLIB::UO_F32, "__aeabi_fcmpun", CallingConv::ARM_AAPCS, ISD::SETNE },
300 { RTLIB::O_F32, "__aeabi_fcmpun", CallingConv::ARM_AAPCS, ISD::SETEQ },
302 // Floating-point to integer conversions.
303 // RTABI chapter 4.1.2, Table 6
304 { RTLIB::FPTOSINT_F64_I32, "__aeabi_d2iz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
305 { RTLIB::FPTOUINT_F64_I32, "__aeabi_d2uiz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
306 { RTLIB::FPTOSINT_F64_I64, "__aeabi_d2lz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
307 { RTLIB::FPTOUINT_F64_I64, "__aeabi_d2ulz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
308 { RTLIB::FPTOSINT_F32_I32, "__aeabi_f2iz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
309 { RTLIB::FPTOUINT_F32_I32, "__aeabi_f2uiz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
310 { RTLIB::FPTOSINT_F32_I64, "__aeabi_f2lz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
311 { RTLIB::FPTOUINT_F32_I64, "__aeabi_f2ulz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
313 // Conversions between floating types.
314 // RTABI chapter 4.1.2, Table 7
315 { RTLIB::FPROUND_F64_F32, "__aeabi_d2f", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
316 { RTLIB::FPROUND_F64_F16, "__aeabi_d2h", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
317 { RTLIB::FPEXT_F32_F64, "__aeabi_f2d", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
319 // Integer to floating-point conversions.
320 // RTABI chapter 4.1.2, Table 8
321 { RTLIB::SINTTOFP_I32_F64, "__aeabi_i2d", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
322 { RTLIB::UINTTOFP_I32_F64, "__aeabi_ui2d", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
323 { RTLIB::SINTTOFP_I64_F64, "__aeabi_l2d", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
324 { RTLIB::UINTTOFP_I64_F64, "__aeabi_ul2d", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
325 { RTLIB::SINTTOFP_I32_F32, "__aeabi_i2f", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
326 { RTLIB::UINTTOFP_I32_F32, "__aeabi_ui2f", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
327 { RTLIB::SINTTOFP_I64_F32, "__aeabi_l2f", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
328 { RTLIB::UINTTOFP_I64_F32, "__aeabi_ul2f", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
330 // Long long helper functions
331 // RTABI chapter 4.2, Table 9
332 { RTLIB::MUL_I64, "__aeabi_lmul", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
333 { RTLIB::SHL_I64, "__aeabi_llsl", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
334 { RTLIB::SRL_I64, "__aeabi_llsr", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
335 { RTLIB::SRA_I64, "__aeabi_lasr", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
337 // Integer division functions
338 // RTABI chapter 4.3.1
339 { RTLIB::SDIV_I8, "__aeabi_idiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
340 { RTLIB::SDIV_I16, "__aeabi_idiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
341 { RTLIB::SDIV_I32, "__aeabi_idiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
342 { RTLIB::SDIV_I64, "__aeabi_ldivmod", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
343 { RTLIB::UDIV_I8, "__aeabi_uidiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
344 { RTLIB::UDIV_I16, "__aeabi_uidiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
345 { RTLIB::UDIV_I32, "__aeabi_uidiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
346 { RTLIB::UDIV_I64, "__aeabi_uldivmod", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
349 // RTABI chapter 4.3.4
350 { RTLIB::MEMCPY, "__aeabi_memcpy", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
351 { RTLIB::MEMMOVE, "__aeabi_memmove", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
352 { RTLIB::MEMSET, "__aeabi_memset", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
355 for (const auto &LC : LibraryCalls) {
356 setLibcallName(LC.Op, LC.Name);
357 setLibcallCallingConv(LC.Op, LC.CC);
358 if (LC.Cond != ISD::SETCC_INVALID)
359 setCmpLibcallCC(LC.Op, LC.Cond);
363 if (Subtarget->isTargetWindows()) {
364 static const struct {
365 const RTLIB::Libcall Op;
366 const char * const Name;
367 const CallingConv::ID CC;
369 { RTLIB::FPTOSINT_F32_I64, "__stoi64", CallingConv::ARM_AAPCS_VFP },
370 { RTLIB::FPTOSINT_F64_I64, "__dtoi64", CallingConv::ARM_AAPCS_VFP },
371 { RTLIB::FPTOUINT_F32_I64, "__stou64", CallingConv::ARM_AAPCS_VFP },
372 { RTLIB::FPTOUINT_F64_I64, "__dtou64", CallingConv::ARM_AAPCS_VFP },
373 { RTLIB::SINTTOFP_I64_F32, "__i64tos", CallingConv::ARM_AAPCS_VFP },
374 { RTLIB::SINTTOFP_I64_F64, "__i64tod", CallingConv::ARM_AAPCS_VFP },
375 { RTLIB::UINTTOFP_I64_F32, "__u64tos", CallingConv::ARM_AAPCS_VFP },
376 { RTLIB::UINTTOFP_I64_F64, "__u64tod", CallingConv::ARM_AAPCS_VFP },
379 for (const auto &LC : LibraryCalls) {
380 setLibcallName(LC.Op, LC.Name);
381 setLibcallCallingConv(LC.Op, LC.CC);
385 // Use divmod compiler-rt calls for iOS 5.0 and later.
386 if (Subtarget->getTargetTriple().isiOS() &&
387 !Subtarget->getTargetTriple().isOSVersionLT(5, 0)) {
388 setLibcallName(RTLIB::SDIVREM_I32, "__divmodsi4");
389 setLibcallName(RTLIB::UDIVREM_I32, "__udivmodsi4");
392 // The half <-> float conversion functions are always soft-float, but are
393 // needed for some targets which use a hard-float calling convention by
395 if (Subtarget->isAAPCS_ABI()) {
396 setLibcallCallingConv(RTLIB::FPROUND_F32_F16, CallingConv::ARM_AAPCS);
397 setLibcallCallingConv(RTLIB::FPROUND_F64_F16, CallingConv::ARM_AAPCS);
398 setLibcallCallingConv(RTLIB::FPEXT_F16_F32, CallingConv::ARM_AAPCS);
400 setLibcallCallingConv(RTLIB::FPROUND_F32_F16, CallingConv::ARM_APCS);
401 setLibcallCallingConv(RTLIB::FPROUND_F64_F16, CallingConv::ARM_APCS);
402 setLibcallCallingConv(RTLIB::FPEXT_F16_F32, CallingConv::ARM_APCS);
405 if (Subtarget->isThumb1Only())
406 addRegisterClass(MVT::i32, &ARM::tGPRRegClass);
408 addRegisterClass(MVT::i32, &ARM::GPRRegClass);
409 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
410 !Subtarget->isThumb1Only()) {
411 addRegisterClass(MVT::f32, &ARM::SPRRegClass);
412 addRegisterClass(MVT::f64, &ARM::DPRRegClass);
415 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
416 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
417 for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
418 InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
419 setTruncStoreAction((MVT::SimpleValueType)VT,
420 (MVT::SimpleValueType)InnerVT, Expand);
421 setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT, Expand);
422 setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT, Expand);
423 setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT, Expand);
425 setOperationAction(ISD::MULHS, (MVT::SimpleValueType)VT, Expand);
426 setOperationAction(ISD::SMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
427 setOperationAction(ISD::MULHU, (MVT::SimpleValueType)VT, Expand);
428 setOperationAction(ISD::UMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
430 setOperationAction(ISD::BSWAP, (MVT::SimpleValueType)VT, Expand);
433 setOperationAction(ISD::ConstantFP, MVT::f32, Custom);
434 setOperationAction(ISD::ConstantFP, MVT::f64, Custom);
436 if (Subtarget->hasNEON()) {
437 addDRTypeForNEON(MVT::v2f32);
438 addDRTypeForNEON(MVT::v8i8);
439 addDRTypeForNEON(MVT::v4i16);
440 addDRTypeForNEON(MVT::v2i32);
441 addDRTypeForNEON(MVT::v1i64);
443 addQRTypeForNEON(MVT::v4f32);
444 addQRTypeForNEON(MVT::v2f64);
445 addQRTypeForNEON(MVT::v16i8);
446 addQRTypeForNEON(MVT::v8i16);
447 addQRTypeForNEON(MVT::v4i32);
448 addQRTypeForNEON(MVT::v2i64);
450 // v2f64 is legal so that QR subregs can be extracted as f64 elements, but
451 // neither Neon nor VFP support any arithmetic operations on it.
452 // The same with v4f32. But keep in mind that vadd, vsub, vmul are natively
453 // supported for v4f32.
454 setOperationAction(ISD::FADD, MVT::v2f64, Expand);
455 setOperationAction(ISD::FSUB, MVT::v2f64, Expand);
456 setOperationAction(ISD::FMUL, MVT::v2f64, Expand);
457 // FIXME: Code duplication: FDIV and FREM are expanded always, see
458 // ARMTargetLowering::addTypeForNEON method for details.
459 setOperationAction(ISD::FDIV, MVT::v2f64, Expand);
460 setOperationAction(ISD::FREM, MVT::v2f64, Expand);
461 // FIXME: Create unittest.
462 // In another words, find a way when "copysign" appears in DAG with vector
464 setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand);
465 // FIXME: Code duplication: SETCC has custom operation action, see
466 // ARMTargetLowering::addTypeForNEON method for details.
467 setOperationAction(ISD::SETCC, MVT::v2f64, Expand);
468 // FIXME: Create unittest for FNEG and for FABS.
469 setOperationAction(ISD::FNEG, MVT::v2f64, Expand);
470 setOperationAction(ISD::FABS, MVT::v2f64, Expand);
471 setOperationAction(ISD::FSQRT, MVT::v2f64, Expand);
472 setOperationAction(ISD::FSIN, MVT::v2f64, Expand);
473 setOperationAction(ISD::FCOS, MVT::v2f64, Expand);
474 setOperationAction(ISD::FPOWI, MVT::v2f64, Expand);
475 setOperationAction(ISD::FPOW, MVT::v2f64, Expand);
476 setOperationAction(ISD::FLOG, MVT::v2f64, Expand);
477 setOperationAction(ISD::FLOG2, MVT::v2f64, Expand);
478 setOperationAction(ISD::FLOG10, MVT::v2f64, Expand);
479 setOperationAction(ISD::FEXP, MVT::v2f64, Expand);
480 setOperationAction(ISD::FEXP2, MVT::v2f64, Expand);
481 // FIXME: Create unittest for FCEIL, FTRUNC, FRINT, FNEARBYINT, FFLOOR.
482 setOperationAction(ISD::FCEIL, MVT::v2f64, Expand);
483 setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand);
484 setOperationAction(ISD::FRINT, MVT::v2f64, Expand);
485 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand);
486 setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand);
487 setOperationAction(ISD::FMA, MVT::v2f64, Expand);
489 setOperationAction(ISD::FSQRT, MVT::v4f32, Expand);
490 setOperationAction(ISD::FSIN, MVT::v4f32, Expand);
491 setOperationAction(ISD::FCOS, MVT::v4f32, Expand);
492 setOperationAction(ISD::FPOWI, MVT::v4f32, Expand);
493 setOperationAction(ISD::FPOW, MVT::v4f32, Expand);
494 setOperationAction(ISD::FLOG, MVT::v4f32, Expand);
495 setOperationAction(ISD::FLOG2, MVT::v4f32, Expand);
496 setOperationAction(ISD::FLOG10, MVT::v4f32, Expand);
497 setOperationAction(ISD::FEXP, MVT::v4f32, Expand);
498 setOperationAction(ISD::FEXP2, MVT::v4f32, Expand);
499 setOperationAction(ISD::FCEIL, MVT::v4f32, Expand);
500 setOperationAction(ISD::FTRUNC, MVT::v4f32, Expand);
501 setOperationAction(ISD::FRINT, MVT::v4f32, Expand);
502 setOperationAction(ISD::FNEARBYINT, MVT::v4f32, Expand);
503 setOperationAction(ISD::FFLOOR, MVT::v4f32, Expand);
505 // Mark v2f32 intrinsics.
506 setOperationAction(ISD::FSQRT, MVT::v2f32, Expand);
507 setOperationAction(ISD::FSIN, MVT::v2f32, Expand);
508 setOperationAction(ISD::FCOS, MVT::v2f32, Expand);
509 setOperationAction(ISD::FPOWI, MVT::v2f32, Expand);
510 setOperationAction(ISD::FPOW, MVT::v2f32, Expand);
511 setOperationAction(ISD::FLOG, MVT::v2f32, Expand);
512 setOperationAction(ISD::FLOG2, MVT::v2f32, Expand);
513 setOperationAction(ISD::FLOG10, MVT::v2f32, Expand);
514 setOperationAction(ISD::FEXP, MVT::v2f32, Expand);
515 setOperationAction(ISD::FEXP2, MVT::v2f32, Expand);
516 setOperationAction(ISD::FCEIL, MVT::v2f32, Expand);
517 setOperationAction(ISD::FTRUNC, MVT::v2f32, Expand);
518 setOperationAction(ISD::FRINT, MVT::v2f32, Expand);
519 setOperationAction(ISD::FNEARBYINT, MVT::v2f32, Expand);
520 setOperationAction(ISD::FFLOOR, MVT::v2f32, Expand);
522 // Neon does not support some operations on v1i64 and v2i64 types.
523 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
524 // Custom handling for some quad-vector types to detect VMULL.
525 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
526 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
527 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
528 // Custom handling for some vector types to avoid expensive expansions
529 setOperationAction(ISD::SDIV, MVT::v4i16, Custom);
530 setOperationAction(ISD::SDIV, MVT::v8i8, Custom);
531 setOperationAction(ISD::UDIV, MVT::v4i16, Custom);
532 setOperationAction(ISD::UDIV, MVT::v8i8, Custom);
533 setOperationAction(ISD::SETCC, MVT::v1i64, Expand);
534 setOperationAction(ISD::SETCC, MVT::v2i64, Expand);
535 // Neon does not have single instruction SINT_TO_FP and UINT_TO_FP with
536 // a destination type that is wider than the source, and nor does
537 // it have a FP_TO_[SU]INT instruction with a narrower destination than
539 setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Custom);
540 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
541 setOperationAction(ISD::FP_TO_UINT, MVT::v4i16, Custom);
542 setOperationAction(ISD::FP_TO_SINT, MVT::v4i16, Custom);
544 setOperationAction(ISD::FP_ROUND, MVT::v2f32, Expand);
545 setOperationAction(ISD::FP_EXTEND, MVT::v2f64, Expand);
547 // NEON does not have single instruction CTPOP for vectors with element
548 // types wider than 8-bits. However, custom lowering can leverage the
549 // v8i8/v16i8 vcnt instruction.
550 setOperationAction(ISD::CTPOP, MVT::v2i32, Custom);
551 setOperationAction(ISD::CTPOP, MVT::v4i32, Custom);
552 setOperationAction(ISD::CTPOP, MVT::v4i16, Custom);
553 setOperationAction(ISD::CTPOP, MVT::v8i16, Custom);
555 // NEON only has FMA instructions as of VFP4.
556 if (!Subtarget->hasVFP4()) {
557 setOperationAction(ISD::FMA, MVT::v2f32, Expand);
558 setOperationAction(ISD::FMA, MVT::v4f32, Expand);
561 setTargetDAGCombine(ISD::INTRINSIC_VOID);
562 setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN);
563 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
564 setTargetDAGCombine(ISD::SHL);
565 setTargetDAGCombine(ISD::SRL);
566 setTargetDAGCombine(ISD::SRA);
567 setTargetDAGCombine(ISD::SIGN_EXTEND);
568 setTargetDAGCombine(ISD::ZERO_EXTEND);
569 setTargetDAGCombine(ISD::ANY_EXTEND);
570 setTargetDAGCombine(ISD::SELECT_CC);
571 setTargetDAGCombine(ISD::BUILD_VECTOR);
572 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
573 setTargetDAGCombine(ISD::INSERT_VECTOR_ELT);
574 setTargetDAGCombine(ISD::STORE);
575 setTargetDAGCombine(ISD::FP_TO_SINT);
576 setTargetDAGCombine(ISD::FP_TO_UINT);
577 setTargetDAGCombine(ISD::FDIV);
579 // It is legal to extload from v4i8 to v4i16 or v4i32.
580 MVT Tys[6] = {MVT::v8i8, MVT::v4i8, MVT::v2i8,
581 MVT::v4i16, MVT::v2i16,
583 for (unsigned i = 0; i < 6; ++i) {
584 setLoadExtAction(ISD::EXTLOAD, Tys[i], Legal);
585 setLoadExtAction(ISD::ZEXTLOAD, Tys[i], Legal);
586 setLoadExtAction(ISD::SEXTLOAD, Tys[i], Legal);
590 // ARM and Thumb2 support UMLAL/SMLAL.
591 if (!Subtarget->isThumb1Only())
592 setTargetDAGCombine(ISD::ADDC);
594 if (Subtarget->isFPOnlySP()) {
595 // When targetting a floating-point unit with only single-precision
596 // operations, f64 is legal for the few double-precision instructions which
597 // are present However, no double-precision operations other than moves,
598 // loads and stores are provided by the hardware.
599 setOperationAction(ISD::FADD, MVT::f64, Expand);
600 setOperationAction(ISD::FSUB, MVT::f64, Expand);
601 setOperationAction(ISD::FMUL, MVT::f64, Expand);
602 setOperationAction(ISD::FMA, MVT::f64, Expand);
603 setOperationAction(ISD::FDIV, MVT::f64, Expand);
604 setOperationAction(ISD::FREM, MVT::f64, Expand);
605 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
606 setOperationAction(ISD::FGETSIGN, MVT::f64, Expand);
607 setOperationAction(ISD::FNEG, MVT::f64, Expand);
608 setOperationAction(ISD::FABS, MVT::f64, Expand);
609 setOperationAction(ISD::FSQRT, MVT::f64, Expand);
610 setOperationAction(ISD::FSIN, MVT::f64, Expand);
611 setOperationAction(ISD::FCOS, MVT::f64, Expand);
612 setOperationAction(ISD::FPOWI, MVT::f64, Expand);
613 setOperationAction(ISD::FPOW, MVT::f64, Expand);
614 setOperationAction(ISD::FLOG, MVT::f64, Expand);
615 setOperationAction(ISD::FLOG2, MVT::f64, Expand);
616 setOperationAction(ISD::FLOG10, MVT::f64, Expand);
617 setOperationAction(ISD::FEXP, MVT::f64, Expand);
618 setOperationAction(ISD::FEXP2, MVT::f64, Expand);
619 setOperationAction(ISD::FCEIL, MVT::f64, Expand);
620 setOperationAction(ISD::FTRUNC, MVT::f64, Expand);
621 setOperationAction(ISD::FRINT, MVT::f64, Expand);
622 setOperationAction(ISD::FNEARBYINT, MVT::f64, Expand);
623 setOperationAction(ISD::FFLOOR, MVT::f64, Expand);
624 setOperationAction(ISD::FP_ROUND, MVT::f32, Custom);
625 setOperationAction(ISD::FP_EXTEND, MVT::f64, Custom);
628 computeRegisterProperties();
630 // ARM does not have floating-point extending loads.
631 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
632 setLoadExtAction(ISD::EXTLOAD, 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 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
642 // ARM supports all 4 flavors of integer indexed load / store.
643 if (!Subtarget->isThumb1Only()) {
644 for (unsigned im = (unsigned)ISD::PRE_INC;
645 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
646 setIndexedLoadAction(im, MVT::i1, Legal);
647 setIndexedLoadAction(im, MVT::i8, Legal);
648 setIndexedLoadAction(im, MVT::i16, Legal);
649 setIndexedLoadAction(im, MVT::i32, Legal);
650 setIndexedStoreAction(im, MVT::i1, Legal);
651 setIndexedStoreAction(im, MVT::i8, Legal);
652 setIndexedStoreAction(im, MVT::i16, Legal);
653 setIndexedStoreAction(im, MVT::i32, Legal);
657 setOperationAction(ISD::SADDO, MVT::i32, Custom);
658 setOperationAction(ISD::UADDO, MVT::i32, Custom);
659 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
660 setOperationAction(ISD::USUBO, MVT::i32, Custom);
662 // i64 operation support.
663 setOperationAction(ISD::MUL, MVT::i64, Expand);
664 setOperationAction(ISD::MULHU, MVT::i32, Expand);
665 if (Subtarget->isThumb1Only()) {
666 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
667 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
669 if (Subtarget->isThumb1Only() || !Subtarget->hasV6Ops()
670 || (Subtarget->isThumb2() && !Subtarget->hasThumb2DSP()))
671 setOperationAction(ISD::MULHS, MVT::i32, Expand);
673 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
674 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
675 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
676 setOperationAction(ISD::SRL, MVT::i64, Custom);
677 setOperationAction(ISD::SRA, MVT::i64, Custom);
679 if (!Subtarget->isThumb1Only()) {
680 // FIXME: We should do this for Thumb1 as well.
681 setOperationAction(ISD::ADDC, MVT::i32, Custom);
682 setOperationAction(ISD::ADDE, MVT::i32, Custom);
683 setOperationAction(ISD::SUBC, MVT::i32, Custom);
684 setOperationAction(ISD::SUBE, MVT::i32, Custom);
687 // ARM does not have ROTL.
688 setOperationAction(ISD::ROTL, MVT::i32, Expand);
689 setOperationAction(ISD::CTTZ, MVT::i32, Custom);
690 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
691 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
692 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
694 // These just redirect to CTTZ and CTLZ on ARM.
695 setOperationAction(ISD::CTTZ_ZERO_UNDEF , MVT::i32 , Expand);
696 setOperationAction(ISD::CTLZ_ZERO_UNDEF , MVT::i32 , Expand);
698 setOperationAction(ISD::READCYCLECOUNTER, MVT::i64, Custom);
700 // Only ARMv6 has BSWAP.
701 if (!Subtarget->hasV6Ops())
702 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
704 if (!(Subtarget->hasDivide() && Subtarget->isThumb2()) &&
705 !(Subtarget->hasDivideInARMMode() && !Subtarget->isThumb())) {
706 // These are expanded into libcalls if the cpu doesn't have HW divider.
707 setOperationAction(ISD::SDIV, MVT::i32, Expand);
708 setOperationAction(ISD::UDIV, MVT::i32, Expand);
711 // FIXME: Also set divmod for SREM on EABI
712 setOperationAction(ISD::SREM, MVT::i32, Expand);
713 setOperationAction(ISD::UREM, MVT::i32, Expand);
714 // Register based DivRem for AEABI (RTABI 4.2)
715 if (Subtarget->isTargetAEABI()) {
716 setLibcallName(RTLIB::SDIVREM_I8, "__aeabi_idivmod");
717 setLibcallName(RTLIB::SDIVREM_I16, "__aeabi_idivmod");
718 setLibcallName(RTLIB::SDIVREM_I32, "__aeabi_idivmod");
719 setLibcallName(RTLIB::SDIVREM_I64, "__aeabi_ldivmod");
720 setLibcallName(RTLIB::UDIVREM_I8, "__aeabi_uidivmod");
721 setLibcallName(RTLIB::UDIVREM_I16, "__aeabi_uidivmod");
722 setLibcallName(RTLIB::UDIVREM_I32, "__aeabi_uidivmod");
723 setLibcallName(RTLIB::UDIVREM_I64, "__aeabi_uldivmod");
725 setLibcallCallingConv(RTLIB::SDIVREM_I8, CallingConv::ARM_AAPCS);
726 setLibcallCallingConv(RTLIB::SDIVREM_I16, CallingConv::ARM_AAPCS);
727 setLibcallCallingConv(RTLIB::SDIVREM_I32, CallingConv::ARM_AAPCS);
728 setLibcallCallingConv(RTLIB::SDIVREM_I64, CallingConv::ARM_AAPCS);
729 setLibcallCallingConv(RTLIB::UDIVREM_I8, CallingConv::ARM_AAPCS);
730 setLibcallCallingConv(RTLIB::UDIVREM_I16, CallingConv::ARM_AAPCS);
731 setLibcallCallingConv(RTLIB::UDIVREM_I32, CallingConv::ARM_AAPCS);
732 setLibcallCallingConv(RTLIB::UDIVREM_I64, CallingConv::ARM_AAPCS);
734 setOperationAction(ISD::SDIVREM, MVT::i32, Custom);
735 setOperationAction(ISD::UDIVREM, MVT::i32, Custom);
737 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
738 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
741 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
742 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
743 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
744 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
745 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
747 setOperationAction(ISD::TRAP, MVT::Other, Legal);
749 // Use the default implementation.
750 setOperationAction(ISD::VASTART, MVT::Other, Custom);
751 setOperationAction(ISD::VAARG, MVT::Other, Expand);
752 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
753 setOperationAction(ISD::VAEND, MVT::Other, Expand);
754 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
755 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
757 if (!Subtarget->isTargetMachO()) {
758 // Non-MachO platforms may return values in these registers via the
759 // personality function.
760 setExceptionPointerRegister(ARM::R0);
761 setExceptionSelectorRegister(ARM::R1);
764 if (Subtarget->getTargetTriple().isWindowsItaniumEnvironment())
765 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
767 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
769 // ARMv6 Thumb1 (except for CPUs that support dmb / dsb) and earlier use
770 // the default expansion. If we are targeting a single threaded system,
771 // then set them all for expand so we can lower them later into their
773 if (TM.Options.ThreadModel == ThreadModel::Single)
774 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Expand);
775 else if (Subtarget->hasAnyDataBarrier() && !Subtarget->isThumb1Only()) {
776 // ATOMIC_FENCE needs custom lowering; the others should have been expanded
777 // to ldrex/strex loops already.
778 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Custom);
780 // On v8, we have particularly efficient implementations of atomic fences
781 // if they can be combined with nearby atomic loads and stores.
782 if (!Subtarget->hasV8Ops()) {
783 // Automatically insert fences (dmb ish) around ATOMIC_SWAP etc.
784 setInsertFencesForAtomic(true);
787 // If there's anything we can use as a barrier, go through custom lowering
789 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other,
790 Subtarget->hasAnyDataBarrier() ? Custom : Expand);
792 // Set them all for expansion, which will force libcalls.
793 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Expand);
794 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Expand);
795 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Expand);
796 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Expand);
797 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Expand);
798 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Expand);
799 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Expand);
800 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Expand);
801 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i32, Expand);
802 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i32, Expand);
803 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i32, Expand);
804 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i32, Expand);
805 // Mark ATOMIC_LOAD and ATOMIC_STORE custom so we can handle the
806 // Unordered/Monotonic case.
807 setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Custom);
808 setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Custom);
811 setOperationAction(ISD::PREFETCH, MVT::Other, Custom);
813 // Requires SXTB/SXTH, available on v6 and up in both ARM and Thumb modes.
814 if (!Subtarget->hasV6Ops()) {
815 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
816 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
818 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
820 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
821 !Subtarget->isThumb1Only()) {
822 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
823 // iff target supports vfp2.
824 setOperationAction(ISD::BITCAST, MVT::i64, Custom);
825 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
828 // We want to custom lower some of our intrinsics.
829 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
830 if (Subtarget->isTargetDarwin()) {
831 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
832 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
833 setLibcallName(RTLIB::UNWIND_RESUME, "_Unwind_SjLj_Resume");
836 setOperationAction(ISD::SETCC, MVT::i32, Expand);
837 setOperationAction(ISD::SETCC, MVT::f32, Expand);
838 setOperationAction(ISD::SETCC, MVT::f64, Expand);
839 setOperationAction(ISD::SELECT, MVT::i32, Custom);
840 setOperationAction(ISD::SELECT, MVT::f32, Custom);
841 setOperationAction(ISD::SELECT, MVT::f64, Custom);
842 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
843 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
844 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
846 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
847 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
848 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
849 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
850 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
852 // We don't support sin/cos/fmod/copysign/pow
853 setOperationAction(ISD::FSIN, MVT::f64, Expand);
854 setOperationAction(ISD::FSIN, MVT::f32, Expand);
855 setOperationAction(ISD::FCOS, MVT::f32, Expand);
856 setOperationAction(ISD::FCOS, MVT::f64, Expand);
857 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
858 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
859 setOperationAction(ISD::FREM, MVT::f64, Expand);
860 setOperationAction(ISD::FREM, MVT::f32, Expand);
861 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
862 !Subtarget->isThumb1Only()) {
863 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
864 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
866 setOperationAction(ISD::FPOW, MVT::f64, Expand);
867 setOperationAction(ISD::FPOW, MVT::f32, Expand);
869 if (!Subtarget->hasVFP4()) {
870 setOperationAction(ISD::FMA, MVT::f64, Expand);
871 setOperationAction(ISD::FMA, MVT::f32, Expand);
874 // Various VFP goodness
875 if (!TM.Options.UseSoftFloat && !Subtarget->isThumb1Only()) {
876 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
877 if (Subtarget->hasVFP2()) {
878 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
879 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
880 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
881 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
884 // FP-ARMv8 adds f64 <-> f16 conversion. Before that it should be expanded.
885 if (!Subtarget->hasFPARMv8() || Subtarget->isFPOnlySP()) {
886 setOperationAction(ISD::FP16_TO_FP, MVT::f64, Expand);
887 setOperationAction(ISD::FP_TO_FP16, MVT::f64, Expand);
890 // fp16 is a special v7 extension that adds f16 <-> f32 conversions.
891 if (!Subtarget->hasFP16()) {
892 setOperationAction(ISD::FP16_TO_FP, MVT::f32, Expand);
893 setOperationAction(ISD::FP_TO_FP16, MVT::f32, Expand);
897 // Combine sin / cos into one node or libcall if possible.
898 if (Subtarget->hasSinCos()) {
899 setLibcallName(RTLIB::SINCOS_F32, "sincosf");
900 setLibcallName(RTLIB::SINCOS_F64, "sincos");
901 if (Subtarget->getTargetTriple().isiOS()) {
902 // For iOS, we don't want to the normal expansion of a libcall to
903 // sincos. We want to issue a libcall to __sincos_stret.
904 setOperationAction(ISD::FSINCOS, MVT::f64, Custom);
905 setOperationAction(ISD::FSINCOS, MVT::f32, Custom);
909 // FP-ARMv8 implements a lot of rounding-like FP operations.
910 if (Subtarget->hasFPARMv8()) {
911 setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
912 setOperationAction(ISD::FCEIL, MVT::f32, Legal);
913 setOperationAction(ISD::FROUND, MVT::f32, Legal);
914 setOperationAction(ISD::FTRUNC, MVT::f32, Legal);
915 setOperationAction(ISD::FNEARBYINT, MVT::f32, Legal);
916 setOperationAction(ISD::FRINT, MVT::f32, Legal);
917 if (!Subtarget->isFPOnlySP()) {
918 setOperationAction(ISD::FFLOOR, MVT::f64, Legal);
919 setOperationAction(ISD::FCEIL, MVT::f64, Legal);
920 setOperationAction(ISD::FROUND, MVT::f64, Legal);
921 setOperationAction(ISD::FTRUNC, MVT::f64, Legal);
922 setOperationAction(ISD::FNEARBYINT, MVT::f64, Legal);
923 setOperationAction(ISD::FRINT, MVT::f64, Legal);
926 // We have target-specific dag combine patterns for the following nodes:
927 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
928 setTargetDAGCombine(ISD::ADD);
929 setTargetDAGCombine(ISD::SUB);
930 setTargetDAGCombine(ISD::MUL);
931 setTargetDAGCombine(ISD::AND);
932 setTargetDAGCombine(ISD::OR);
933 setTargetDAGCombine(ISD::XOR);
935 if (Subtarget->hasV6Ops())
936 setTargetDAGCombine(ISD::SRL);
938 setStackPointerRegisterToSaveRestore(ARM::SP);
940 if (TM.Options.UseSoftFloat || Subtarget->isThumb1Only() ||
941 !Subtarget->hasVFP2())
942 setSchedulingPreference(Sched::RegPressure);
944 setSchedulingPreference(Sched::Hybrid);
946 //// temporary - rewrite interface to use type
947 MaxStoresPerMemset = 8;
948 MaxStoresPerMemsetOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
949 MaxStoresPerMemcpy = 4; // For @llvm.memcpy -> sequence of stores
950 MaxStoresPerMemcpyOptSize = Subtarget->isTargetDarwin() ? 4 : 2;
951 MaxStoresPerMemmove = 4; // For @llvm.memmove -> sequence of stores
952 MaxStoresPerMemmoveOptSize = Subtarget->isTargetDarwin() ? 4 : 2;
954 // On ARM arguments smaller than 4 bytes are extended, so all arguments
955 // are at least 4 bytes aligned.
956 setMinStackArgumentAlignment(4);
958 // Prefer likely predicted branches to selects on out-of-order cores.
959 PredictableSelectIsExpensive = Subtarget->isLikeA9();
961 setMinFunctionAlignment(Subtarget->isThumb() ? 1 : 2);
964 // FIXME: It might make sense to define the representative register class as the
965 // nearest super-register that has a non-null superset. For example, DPR_VFP2 is
966 // a super-register of SPR, and DPR is a superset if DPR_VFP2. Consequently,
967 // SPR's representative would be DPR_VFP2. This should work well if register
968 // pressure tracking were modified such that a register use would increment the
969 // pressure of the register class's representative and all of it's super
970 // classes' representatives transitively. We have not implemented this because
971 // of the difficulty prior to coalescing of modeling operand register classes
972 // due to the common occurrence of cross class copies and subregister insertions
974 std::pair<const TargetRegisterClass*, uint8_t>
975 ARMTargetLowering::findRepresentativeClass(MVT VT) const{
976 const TargetRegisterClass *RRC = nullptr;
978 switch (VT.SimpleTy) {
980 return TargetLowering::findRepresentativeClass(VT);
981 // Use DPR as representative register class for all floating point
982 // and vector types. Since there are 32 SPR registers and 32 DPR registers so
983 // the cost is 1 for both f32 and f64.
984 case MVT::f32: case MVT::f64: case MVT::v8i8: case MVT::v4i16:
985 case MVT::v2i32: case MVT::v1i64: case MVT::v2f32:
986 RRC = &ARM::DPRRegClass;
987 // When NEON is used for SP, only half of the register file is available
988 // because operations that define both SP and DP results will be constrained
989 // to the VFP2 class (D0-D15). We currently model this constraint prior to
990 // coalescing by double-counting the SP regs. See the FIXME above.
991 if (Subtarget->useNEONForSinglePrecisionFP())
994 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
995 case MVT::v4f32: case MVT::v2f64:
996 RRC = &ARM::DPRRegClass;
1000 RRC = &ARM::DPRRegClass;
1004 RRC = &ARM::DPRRegClass;
1008 return std::make_pair(RRC, Cost);
1011 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
1013 default: return nullptr;
1014 case ARMISD::Wrapper: return "ARMISD::Wrapper";
1015 case ARMISD::WrapperPIC: return "ARMISD::WrapperPIC";
1016 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
1017 case ARMISD::CALL: return "ARMISD::CALL";
1018 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
1019 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
1020 case ARMISD::tCALL: return "ARMISD::tCALL";
1021 case ARMISD::BRCOND: return "ARMISD::BRCOND";
1022 case ARMISD::BR_JT: return "ARMISD::BR_JT";
1023 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
1024 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
1025 case ARMISD::INTRET_FLAG: return "ARMISD::INTRET_FLAG";
1026 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
1027 case ARMISD::CMP: return "ARMISD::CMP";
1028 case ARMISD::CMN: return "ARMISD::CMN";
1029 case ARMISD::CMPZ: return "ARMISD::CMPZ";
1030 case ARMISD::CMPFP: return "ARMISD::CMPFP";
1031 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
1032 case ARMISD::BCC_i64: return "ARMISD::BCC_i64";
1033 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
1035 case ARMISD::CMOV: return "ARMISD::CMOV";
1037 case ARMISD::RBIT: return "ARMISD::RBIT";
1039 case ARMISD::FTOSI: return "ARMISD::FTOSI";
1040 case ARMISD::FTOUI: return "ARMISD::FTOUI";
1041 case ARMISD::SITOF: return "ARMISD::SITOF";
1042 case ARMISD::UITOF: return "ARMISD::UITOF";
1044 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
1045 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
1046 case ARMISD::RRX: return "ARMISD::RRX";
1048 case ARMISD::ADDC: return "ARMISD::ADDC";
1049 case ARMISD::ADDE: return "ARMISD::ADDE";
1050 case ARMISD::SUBC: return "ARMISD::SUBC";
1051 case ARMISD::SUBE: return "ARMISD::SUBE";
1053 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
1054 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
1056 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
1057 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
1059 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN";
1061 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
1063 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
1065 case ARMISD::MEMBARRIER_MCR: return "ARMISD::MEMBARRIER_MCR";
1067 case ARMISD::PRELOAD: return "ARMISD::PRELOAD";
1069 case ARMISD::WIN__CHKSTK: return "ARMISD:::WIN__CHKSTK";
1071 case ARMISD::VCEQ: return "ARMISD::VCEQ";
1072 case ARMISD::VCEQZ: return "ARMISD::VCEQZ";
1073 case ARMISD::VCGE: return "ARMISD::VCGE";
1074 case ARMISD::VCGEZ: return "ARMISD::VCGEZ";
1075 case ARMISD::VCLEZ: return "ARMISD::VCLEZ";
1076 case ARMISD::VCGEU: return "ARMISD::VCGEU";
1077 case ARMISD::VCGT: return "ARMISD::VCGT";
1078 case ARMISD::VCGTZ: return "ARMISD::VCGTZ";
1079 case ARMISD::VCLTZ: return "ARMISD::VCLTZ";
1080 case ARMISD::VCGTU: return "ARMISD::VCGTU";
1081 case ARMISD::VTST: return "ARMISD::VTST";
1083 case ARMISD::VSHL: return "ARMISD::VSHL";
1084 case ARMISD::VSHRs: return "ARMISD::VSHRs";
1085 case ARMISD::VSHRu: return "ARMISD::VSHRu";
1086 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
1087 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
1088 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
1089 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
1090 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
1091 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
1092 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
1093 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
1094 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
1095 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
1096 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
1097 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
1098 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
1099 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
1100 case ARMISD::VMOVIMM: return "ARMISD::VMOVIMM";
1101 case ARMISD::VMVNIMM: return "ARMISD::VMVNIMM";
1102 case ARMISD::VMOVFPIMM: return "ARMISD::VMOVFPIMM";
1103 case ARMISD::VDUP: return "ARMISD::VDUP";
1104 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
1105 case ARMISD::VEXT: return "ARMISD::VEXT";
1106 case ARMISD::VREV64: return "ARMISD::VREV64";
1107 case ARMISD::VREV32: return "ARMISD::VREV32";
1108 case ARMISD::VREV16: return "ARMISD::VREV16";
1109 case ARMISD::VZIP: return "ARMISD::VZIP";
1110 case ARMISD::VUZP: return "ARMISD::VUZP";
1111 case ARMISD::VTRN: return "ARMISD::VTRN";
1112 case ARMISD::VTBL1: return "ARMISD::VTBL1";
1113 case ARMISD::VTBL2: return "ARMISD::VTBL2";
1114 case ARMISD::VMULLs: return "ARMISD::VMULLs";
1115 case ARMISD::VMULLu: return "ARMISD::VMULLu";
1116 case ARMISD::UMLAL: return "ARMISD::UMLAL";
1117 case ARMISD::SMLAL: return "ARMISD::SMLAL";
1118 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
1119 case ARMISD::FMAX: return "ARMISD::FMAX";
1120 case ARMISD::FMIN: return "ARMISD::FMIN";
1121 case ARMISD::VMAXNM: return "ARMISD::VMAX";
1122 case ARMISD::VMINNM: return "ARMISD::VMIN";
1123 case ARMISD::BFI: return "ARMISD::BFI";
1124 case ARMISD::VORRIMM: return "ARMISD::VORRIMM";
1125 case ARMISD::VBICIMM: return "ARMISD::VBICIMM";
1126 case ARMISD::VBSL: return "ARMISD::VBSL";
1127 case ARMISD::VLD2DUP: return "ARMISD::VLD2DUP";
1128 case ARMISD::VLD3DUP: return "ARMISD::VLD3DUP";
1129 case ARMISD::VLD4DUP: return "ARMISD::VLD4DUP";
1130 case ARMISD::VLD1_UPD: return "ARMISD::VLD1_UPD";
1131 case ARMISD::VLD2_UPD: return "ARMISD::VLD2_UPD";
1132 case ARMISD::VLD3_UPD: return "ARMISD::VLD3_UPD";
1133 case ARMISD::VLD4_UPD: return "ARMISD::VLD4_UPD";
1134 case ARMISD::VLD2LN_UPD: return "ARMISD::VLD2LN_UPD";
1135 case ARMISD::VLD3LN_UPD: return "ARMISD::VLD3LN_UPD";
1136 case ARMISD::VLD4LN_UPD: return "ARMISD::VLD4LN_UPD";
1137 case ARMISD::VLD2DUP_UPD: return "ARMISD::VLD2DUP_UPD";
1138 case ARMISD::VLD3DUP_UPD: return "ARMISD::VLD3DUP_UPD";
1139 case ARMISD::VLD4DUP_UPD: return "ARMISD::VLD4DUP_UPD";
1140 case ARMISD::VST1_UPD: return "ARMISD::VST1_UPD";
1141 case ARMISD::VST2_UPD: return "ARMISD::VST2_UPD";
1142 case ARMISD::VST3_UPD: return "ARMISD::VST3_UPD";
1143 case ARMISD::VST4_UPD: return "ARMISD::VST4_UPD";
1144 case ARMISD::VST2LN_UPD: return "ARMISD::VST2LN_UPD";
1145 case ARMISD::VST3LN_UPD: return "ARMISD::VST3LN_UPD";
1146 case ARMISD::VST4LN_UPD: return "ARMISD::VST4LN_UPD";
1150 EVT ARMTargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
1151 if (!VT.isVector()) return getPointerTy();
1152 return VT.changeVectorElementTypeToInteger();
1155 /// getRegClassFor - Return the register class that should be used for the
1156 /// specified value type.
1157 const TargetRegisterClass *ARMTargetLowering::getRegClassFor(MVT VT) const {
1158 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
1159 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
1160 // load / store 4 to 8 consecutive D registers.
1161 if (Subtarget->hasNEON()) {
1162 if (VT == MVT::v4i64)
1163 return &ARM::QQPRRegClass;
1164 if (VT == MVT::v8i64)
1165 return &ARM::QQQQPRRegClass;
1167 return TargetLowering::getRegClassFor(VT);
1170 // Create a fast isel object.
1172 ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
1173 const TargetLibraryInfo *libInfo) const {
1174 return ARM::createFastISel(funcInfo, libInfo);
1177 /// getMaximalGlobalOffset - Returns the maximal possible offset which can
1178 /// be used for loads / stores from the global.
1179 unsigned ARMTargetLowering::getMaximalGlobalOffset() const {
1180 return (Subtarget->isThumb1Only() ? 127 : 4095);
1183 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
1184 unsigned NumVals = N->getNumValues();
1186 return Sched::RegPressure;
1188 for (unsigned i = 0; i != NumVals; ++i) {
1189 EVT VT = N->getValueType(i);
1190 if (VT == MVT::Glue || VT == MVT::Other)
1192 if (VT.isFloatingPoint() || VT.isVector())
1196 if (!N->isMachineOpcode())
1197 return Sched::RegPressure;
1199 // Load are scheduled for latency even if there instruction itinerary
1200 // is not available.
1201 const TargetInstrInfo *TII =
1202 getTargetMachine().getSubtargetImpl()->getInstrInfo();
1203 const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
1205 if (MCID.getNumDefs() == 0)
1206 return Sched::RegPressure;
1207 if (!Itins->isEmpty() &&
1208 Itins->getOperandCycle(MCID.getSchedClass(), 0) > 2)
1211 return Sched::RegPressure;
1214 //===----------------------------------------------------------------------===//
1216 //===----------------------------------------------------------------------===//
1218 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
1219 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
1221 default: llvm_unreachable("Unknown condition code!");
1222 case ISD::SETNE: return ARMCC::NE;
1223 case ISD::SETEQ: return ARMCC::EQ;
1224 case ISD::SETGT: return ARMCC::GT;
1225 case ISD::SETGE: return ARMCC::GE;
1226 case ISD::SETLT: return ARMCC::LT;
1227 case ISD::SETLE: return ARMCC::LE;
1228 case ISD::SETUGT: return ARMCC::HI;
1229 case ISD::SETUGE: return ARMCC::HS;
1230 case ISD::SETULT: return ARMCC::LO;
1231 case ISD::SETULE: return ARMCC::LS;
1235 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
1236 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
1237 ARMCC::CondCodes &CondCode2) {
1238 CondCode2 = ARMCC::AL;
1240 default: llvm_unreachable("Unknown FP condition!");
1242 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
1244 case ISD::SETOGT: CondCode = ARMCC::GT; break;
1246 case ISD::SETOGE: CondCode = ARMCC::GE; break;
1247 case ISD::SETOLT: CondCode = ARMCC::MI; break;
1248 case ISD::SETOLE: CondCode = ARMCC::LS; break;
1249 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
1250 case ISD::SETO: CondCode = ARMCC::VC; break;
1251 case ISD::SETUO: CondCode = ARMCC::VS; break;
1252 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
1253 case ISD::SETUGT: CondCode = ARMCC::HI; break;
1254 case ISD::SETUGE: CondCode = ARMCC::PL; break;
1256 case ISD::SETULT: CondCode = ARMCC::LT; break;
1258 case ISD::SETULE: CondCode = ARMCC::LE; break;
1260 case ISD::SETUNE: CondCode = ARMCC::NE; break;
1264 //===----------------------------------------------------------------------===//
1265 // Calling Convention Implementation
1266 //===----------------------------------------------------------------------===//
1268 #include "ARMGenCallingConv.inc"
1270 /// getEffectiveCallingConv - Get the effective calling convention, taking into
1271 /// account presence of floating point hardware and calling convention
1272 /// limitations, such as support for variadic functions.
1274 ARMTargetLowering::getEffectiveCallingConv(CallingConv::ID CC,
1275 bool isVarArg) const {
1278 llvm_unreachable("Unsupported calling convention");
1279 case CallingConv::ARM_AAPCS:
1280 case CallingConv::ARM_APCS:
1281 case CallingConv::GHC:
1283 case CallingConv::ARM_AAPCS_VFP:
1284 return isVarArg ? CallingConv::ARM_AAPCS : CallingConv::ARM_AAPCS_VFP;
1285 case CallingConv::C:
1286 if (!Subtarget->isAAPCS_ABI())
1287 return CallingConv::ARM_APCS;
1288 else if (Subtarget->hasVFP2() && !Subtarget->isThumb1Only() &&
1289 getTargetMachine().Options.FloatABIType == FloatABI::Hard &&
1291 return CallingConv::ARM_AAPCS_VFP;
1293 return CallingConv::ARM_AAPCS;
1294 case CallingConv::Fast:
1295 if (!Subtarget->isAAPCS_ABI()) {
1296 if (Subtarget->hasVFP2() && !Subtarget->isThumb1Only() && !isVarArg)
1297 return CallingConv::Fast;
1298 return CallingConv::ARM_APCS;
1299 } else if (Subtarget->hasVFP2() && !Subtarget->isThumb1Only() && !isVarArg)
1300 return CallingConv::ARM_AAPCS_VFP;
1302 return CallingConv::ARM_AAPCS;
1306 /// CCAssignFnForNode - Selects the correct CCAssignFn for the given
1307 /// CallingConvention.
1308 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
1310 bool isVarArg) const {
1311 switch (getEffectiveCallingConv(CC, isVarArg)) {
1313 llvm_unreachable("Unsupported calling convention");
1314 case CallingConv::ARM_APCS:
1315 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1316 case CallingConv::ARM_AAPCS:
1317 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1318 case CallingConv::ARM_AAPCS_VFP:
1319 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1320 case CallingConv::Fast:
1321 return (Return ? RetFastCC_ARM_APCS : FastCC_ARM_APCS);
1322 case CallingConv::GHC:
1323 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS_GHC);
1327 /// LowerCallResult - Lower the result values of a call into the
1328 /// appropriate copies out of appropriate physical registers.
1330 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1331 CallingConv::ID CallConv, bool isVarArg,
1332 const SmallVectorImpl<ISD::InputArg> &Ins,
1333 SDLoc dl, SelectionDAG &DAG,
1334 SmallVectorImpl<SDValue> &InVals,
1335 bool isThisReturn, SDValue ThisVal) const {
1337 // Assign locations to each value returned by this call.
1338 SmallVector<CCValAssign, 16> RVLocs;
1339 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
1340 *DAG.getContext(), Call);
1341 CCInfo.AnalyzeCallResult(Ins,
1342 CCAssignFnForNode(CallConv, /* Return*/ true,
1345 // Copy all of the result registers out of their specified physreg.
1346 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1347 CCValAssign VA = RVLocs[i];
1349 // Pass 'this' value directly from the argument to return value, to avoid
1350 // reg unit interference
1351 if (i == 0 && isThisReturn) {
1352 assert(!VA.needsCustom() && VA.getLocVT() == MVT::i32 &&
1353 "unexpected return calling convention register assignment");
1354 InVals.push_back(ThisVal);
1359 if (VA.needsCustom()) {
1360 // Handle f64 or half of a v2f64.
1361 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1363 Chain = Lo.getValue(1);
1364 InFlag = Lo.getValue(2);
1365 VA = RVLocs[++i]; // skip ahead to next loc
1366 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1368 Chain = Hi.getValue(1);
1369 InFlag = Hi.getValue(2);
1370 if (!Subtarget->isLittle())
1372 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1374 if (VA.getLocVT() == MVT::v2f64) {
1375 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1376 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1377 DAG.getConstant(0, MVT::i32));
1379 VA = RVLocs[++i]; // skip ahead to next loc
1380 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1381 Chain = Lo.getValue(1);
1382 InFlag = Lo.getValue(2);
1383 VA = RVLocs[++i]; // skip ahead to next loc
1384 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1385 Chain = Hi.getValue(1);
1386 InFlag = Hi.getValue(2);
1387 if (!Subtarget->isLittle())
1389 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1390 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1391 DAG.getConstant(1, MVT::i32));
1394 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
1396 Chain = Val.getValue(1);
1397 InFlag = Val.getValue(2);
1400 switch (VA.getLocInfo()) {
1401 default: llvm_unreachable("Unknown loc info!");
1402 case CCValAssign::Full: break;
1403 case CCValAssign::BCvt:
1404 Val = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), Val);
1408 InVals.push_back(Val);
1414 /// LowerMemOpCallTo - Store the argument to the stack.
1416 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
1417 SDValue StackPtr, SDValue Arg,
1418 SDLoc dl, SelectionDAG &DAG,
1419 const CCValAssign &VA,
1420 ISD::ArgFlagsTy Flags) const {
1421 unsigned LocMemOffset = VA.getLocMemOffset();
1422 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1423 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1424 return DAG.getStore(Chain, dl, Arg, PtrOff,
1425 MachinePointerInfo::getStack(LocMemOffset),
1429 void ARMTargetLowering::PassF64ArgInRegs(SDLoc dl, SelectionDAG &DAG,
1430 SDValue Chain, SDValue &Arg,
1431 RegsToPassVector &RegsToPass,
1432 CCValAssign &VA, CCValAssign &NextVA,
1434 SmallVectorImpl<SDValue> &MemOpChains,
1435 ISD::ArgFlagsTy Flags) const {
1437 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1438 DAG.getVTList(MVT::i32, MVT::i32), Arg);
1439 unsigned id = Subtarget->isLittle() ? 0 : 1;
1440 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd.getValue(id)));
1442 if (NextVA.isRegLoc())
1443 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1-id)));
1445 assert(NextVA.isMemLoc());
1446 if (!StackPtr.getNode())
1447 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1449 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1-id),
1455 /// LowerCall - Lowering a call into a callseq_start <-
1456 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
1459 ARMTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
1460 SmallVectorImpl<SDValue> &InVals) const {
1461 SelectionDAG &DAG = CLI.DAG;
1463 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
1464 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
1465 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
1466 SDValue Chain = CLI.Chain;
1467 SDValue Callee = CLI.Callee;
1468 bool &isTailCall = CLI.IsTailCall;
1469 CallingConv::ID CallConv = CLI.CallConv;
1470 bool doesNotRet = CLI.DoesNotReturn;
1471 bool isVarArg = CLI.IsVarArg;
1473 MachineFunction &MF = DAG.getMachineFunction();
1474 bool isStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1475 bool isThisReturn = false;
1476 bool isSibCall = false;
1478 // Disable tail calls if they're not supported.
1479 if (!Subtarget->supportsTailCall() || MF.getTarget().Options.DisableTailCalls)
1483 // Check if it's really possible to do a tail call.
1484 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1485 isVarArg, isStructRet, MF.getFunction()->hasStructRetAttr(),
1486 Outs, OutVals, Ins, DAG);
1487 if (!isTailCall && CLI.CS && CLI.CS->isMustTailCall())
1488 report_fatal_error("failed to perform tail call elimination on a call "
1489 "site marked musttail");
1490 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1491 // detected sibcalls.
1498 // Analyze operands of the call, assigning locations to each operand.
1499 SmallVector<CCValAssign, 16> ArgLocs;
1500 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs,
1501 *DAG.getContext(), Call);
1502 CCInfo.AnalyzeCallOperands(Outs,
1503 CCAssignFnForNode(CallConv, /* Return*/ false,
1506 // Get a count of how many bytes are to be pushed on the stack.
1507 unsigned NumBytes = CCInfo.getNextStackOffset();
1509 // For tail calls, memory operands are available in our caller's stack.
1513 // Adjust the stack pointer for the new arguments...
1514 // These operations are automatically eliminated by the prolog/epilog pass
1516 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true),
1519 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1521 RegsToPassVector RegsToPass;
1522 SmallVector<SDValue, 8> MemOpChains;
1524 // Walk the register/memloc assignments, inserting copies/loads. In the case
1525 // of tail call optimization, arguments are handled later.
1526 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1528 ++i, ++realArgIdx) {
1529 CCValAssign &VA = ArgLocs[i];
1530 SDValue Arg = OutVals[realArgIdx];
1531 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1532 bool isByVal = Flags.isByVal();
1534 // Promote the value if needed.
1535 switch (VA.getLocInfo()) {
1536 default: llvm_unreachable("Unknown loc info!");
1537 case CCValAssign::Full: break;
1538 case CCValAssign::SExt:
1539 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1541 case CCValAssign::ZExt:
1542 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1544 case CCValAssign::AExt:
1545 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1547 case CCValAssign::BCvt:
1548 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1552 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1553 if (VA.needsCustom()) {
1554 if (VA.getLocVT() == MVT::v2f64) {
1555 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1556 DAG.getConstant(0, MVT::i32));
1557 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1558 DAG.getConstant(1, MVT::i32));
1560 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1561 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1563 VA = ArgLocs[++i]; // skip ahead to next loc
1564 if (VA.isRegLoc()) {
1565 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1566 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1568 assert(VA.isMemLoc());
1570 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1571 dl, DAG, VA, Flags));
1574 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1575 StackPtr, MemOpChains, Flags);
1577 } else if (VA.isRegLoc()) {
1578 if (realArgIdx == 0 && Flags.isReturned() && Outs[0].VT == MVT::i32) {
1579 assert(VA.getLocVT() == MVT::i32 &&
1580 "unexpected calling convention register assignment");
1581 assert(!Ins.empty() && Ins[0].VT == MVT::i32 &&
1582 "unexpected use of 'returned'");
1583 isThisReturn = true;
1585 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1586 } else if (isByVal) {
1587 assert(VA.isMemLoc());
1588 unsigned offset = 0;
1590 // True if this byval aggregate will be split between registers
1592 unsigned ByValArgsCount = CCInfo.getInRegsParamsCount();
1593 unsigned CurByValIdx = CCInfo.getInRegsParamsProcessed();
1595 if (CurByValIdx < ByValArgsCount) {
1597 unsigned RegBegin, RegEnd;
1598 CCInfo.getInRegsParamInfo(CurByValIdx, RegBegin, RegEnd);
1600 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1602 for (i = 0, j = RegBegin; j < RegEnd; i++, j++) {
1603 SDValue Const = DAG.getConstant(4*i, MVT::i32);
1604 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
1605 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
1606 MachinePointerInfo(),
1607 false, false, false,
1608 DAG.InferPtrAlignment(AddArg));
1609 MemOpChains.push_back(Load.getValue(1));
1610 RegsToPass.push_back(std::make_pair(j, Load));
1613 // If parameter size outsides register area, "offset" value
1614 // helps us to calculate stack slot for remained part properly.
1615 offset = RegEnd - RegBegin;
1617 CCInfo.nextInRegsParam();
1620 if (Flags.getByValSize() > 4*offset) {
1621 unsigned LocMemOffset = VA.getLocMemOffset();
1622 SDValue StkPtrOff = DAG.getIntPtrConstant(LocMemOffset);
1623 SDValue Dst = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr,
1625 SDValue SrcOffset = DAG.getIntPtrConstant(4*offset);
1626 SDValue Src = DAG.getNode(ISD::ADD, dl, getPointerTy(), Arg, SrcOffset);
1627 SDValue SizeNode = DAG.getConstant(Flags.getByValSize() - 4*offset,
1629 SDValue AlignNode = DAG.getConstant(Flags.getByValAlign(), MVT::i32);
1631 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
1632 SDValue Ops[] = { Chain, Dst, Src, SizeNode, AlignNode};
1633 MemOpChains.push_back(DAG.getNode(ARMISD::COPY_STRUCT_BYVAL, dl, VTs,
1636 } else if (!isSibCall) {
1637 assert(VA.isMemLoc());
1639 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1640 dl, DAG, VA, Flags));
1644 if (!MemOpChains.empty())
1645 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains);
1647 // Build a sequence of copy-to-reg nodes chained together with token chain
1648 // and flag operands which copy the outgoing args into the appropriate regs.
1650 // Tail call byval lowering might overwrite argument registers so in case of
1651 // tail call optimization the copies to registers are lowered later.
1653 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1654 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1655 RegsToPass[i].second, InFlag);
1656 InFlag = Chain.getValue(1);
1659 // For tail calls lower the arguments to the 'real' stack slot.
1661 // Force all the incoming stack arguments to be loaded from the stack
1662 // before any new outgoing arguments are stored to the stack, because the
1663 // outgoing stack slots may alias the incoming argument stack slots, and
1664 // the alias isn't otherwise explicit. This is slightly more conservative
1665 // than necessary, because it means that each store effectively depends
1666 // on every argument instead of just those arguments it would clobber.
1668 // Do not flag preceding copytoreg stuff together with the following stuff.
1670 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1671 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1672 RegsToPass[i].second, InFlag);
1673 InFlag = Chain.getValue(1);
1678 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1679 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1680 // node so that legalize doesn't hack it.
1681 bool isDirect = false;
1682 bool isARMFunc = false;
1683 bool isLocalARMFunc = false;
1684 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1686 if (EnableARMLongCalls) {
1687 assert((Subtarget->isTargetWindows() ||
1688 getTargetMachine().getRelocationModel() == Reloc::Static) &&
1689 "long-calls with non-static relocation model!");
1690 // Handle a global address or an external symbol. If it's not one of
1691 // those, the target's already in a register, so we don't need to do
1693 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1694 const GlobalValue *GV = G->getGlobal();
1695 // Create a constant pool entry for the callee address
1696 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1697 ARMConstantPoolValue *CPV =
1698 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 0);
1700 // Get the address of the callee into a register
1701 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1702 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1703 Callee = DAG.getLoad(getPointerTy(), dl,
1704 DAG.getEntryNode(), CPAddr,
1705 MachinePointerInfo::getConstantPool(),
1706 false, false, false, 0);
1707 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1708 const char *Sym = S->getSymbol();
1710 // Create a constant pool entry for the callee address
1711 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1712 ARMConstantPoolValue *CPV =
1713 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1714 ARMPCLabelIndex, 0);
1715 // Get the address of the callee into a register
1716 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1717 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1718 Callee = DAG.getLoad(getPointerTy(), dl,
1719 DAG.getEntryNode(), CPAddr,
1720 MachinePointerInfo::getConstantPool(),
1721 false, false, false, 0);
1723 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1724 const GlobalValue *GV = G->getGlobal();
1726 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1727 bool isStub = (isExt && Subtarget->isTargetMachO()) &&
1728 getTargetMachine().getRelocationModel() != Reloc::Static;
1729 isARMFunc = !Subtarget->isThumb() || (isStub && !Subtarget->isMClass());
1730 // ARM call to a local ARM function is predicable.
1731 isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking);
1732 // tBX takes a register source operand.
1733 if (isStub && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1734 assert(Subtarget->isTargetMachO() && "WrapperPIC use on non-MachO?");
1735 Callee = DAG.getNode(ARMISD::WrapperPIC, dl, getPointerTy(),
1736 DAG.getTargetGlobalAddress(GV, dl, getPointerTy(),
1737 0, ARMII::MO_NONLAZY));
1738 Callee = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Callee,
1739 MachinePointerInfo::getGOT(), false, false, true, 0);
1740 } else if (Subtarget->isTargetCOFF()) {
1741 assert(Subtarget->isTargetWindows() &&
1742 "Windows is the only supported COFF target");
1743 unsigned TargetFlags = GV->hasDLLImportStorageClass()
1744 ? ARMII::MO_DLLIMPORT
1745 : ARMII::MO_NO_FLAG;
1746 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), /*Offset=*/0,
1748 if (GV->hasDLLImportStorageClass())
1749 Callee = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
1750 DAG.getNode(ARMISD::Wrapper, dl, getPointerTy(),
1751 Callee), MachinePointerInfo::getGOT(),
1752 false, false, false, 0);
1754 // On ELF targets for PIC code, direct calls should go through the PLT
1755 unsigned OpFlags = 0;
1756 if (Subtarget->isTargetELF() &&
1757 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1758 OpFlags = ARMII::MO_PLT;
1759 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
1761 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1763 bool isStub = Subtarget->isTargetMachO() &&
1764 getTargetMachine().getRelocationModel() != Reloc::Static;
1765 isARMFunc = !Subtarget->isThumb() || (isStub && !Subtarget->isMClass());
1766 // tBX takes a register source operand.
1767 const char *Sym = S->getSymbol();
1768 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1769 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1770 ARMConstantPoolValue *CPV =
1771 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1772 ARMPCLabelIndex, 4);
1773 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1774 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1775 Callee = DAG.getLoad(getPointerTy(), dl,
1776 DAG.getEntryNode(), CPAddr,
1777 MachinePointerInfo::getConstantPool(),
1778 false, false, false, 0);
1779 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1780 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1781 getPointerTy(), Callee, PICLabel);
1783 unsigned OpFlags = 0;
1784 // On ELF targets for PIC code, direct calls should go through the PLT
1785 if (Subtarget->isTargetELF() &&
1786 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1787 OpFlags = ARMII::MO_PLT;
1788 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlags);
1792 // FIXME: handle tail calls differently.
1794 bool HasMinSizeAttr = MF.getFunction()->getAttributes().hasAttribute(
1795 AttributeSet::FunctionIndex, Attribute::MinSize);
1796 if (Subtarget->isThumb()) {
1797 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1798 CallOpc = ARMISD::CALL_NOLINK;
1800 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1802 if (!isDirect && !Subtarget->hasV5TOps())
1803 CallOpc = ARMISD::CALL_NOLINK;
1804 else if (doesNotRet && isDirect && Subtarget->hasRAS() &&
1805 // Emit regular call when code size is the priority
1807 // "mov lr, pc; b _foo" to avoid confusing the RSP
1808 CallOpc = ARMISD::CALL_NOLINK;
1810 CallOpc = isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL;
1813 std::vector<SDValue> Ops;
1814 Ops.push_back(Chain);
1815 Ops.push_back(Callee);
1817 // Add argument registers to the end of the list so that they are known live
1819 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1820 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1821 RegsToPass[i].second.getValueType()));
1823 // Add a register mask operand representing the call-preserved registers.
1825 const uint32_t *Mask;
1826 const TargetRegisterInfo *TRI =
1827 getTargetMachine().getSubtargetImpl()->getRegisterInfo();
1828 const ARMBaseRegisterInfo *ARI = static_cast<const ARMBaseRegisterInfo*>(TRI);
1830 // For 'this' returns, use the R0-preserving mask if applicable
1831 Mask = ARI->getThisReturnPreservedMask(CallConv);
1833 // Set isThisReturn to false if the calling convention is not one that
1834 // allows 'returned' to be modeled in this way, so LowerCallResult does
1835 // not try to pass 'this' straight through
1836 isThisReturn = false;
1837 Mask = ARI->getCallPreservedMask(CallConv);
1840 Mask = ARI->getCallPreservedMask(CallConv);
1842 assert(Mask && "Missing call preserved mask for calling convention");
1843 Ops.push_back(DAG.getRegisterMask(Mask));
1846 if (InFlag.getNode())
1847 Ops.push_back(InFlag);
1849 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
1851 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, Ops);
1853 // Returns a chain and a flag for retval copy to use.
1854 Chain = DAG.getNode(CallOpc, dl, NodeTys, Ops);
1855 InFlag = Chain.getValue(1);
1857 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1858 DAG.getIntPtrConstant(0, true), InFlag, dl);
1860 InFlag = Chain.getValue(1);
1862 // Handle result values, copying them out of physregs into vregs that we
1864 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins, dl, DAG,
1865 InVals, isThisReturn,
1866 isThisReturn ? OutVals[0] : SDValue());
1869 /// HandleByVal - Every parameter *after* a byval parameter is passed
1870 /// on the stack. Remember the next parameter register to allocate,
1871 /// and then confiscate the rest of the parameter registers to insure
1874 ARMTargetLowering::HandleByVal(
1875 CCState *State, unsigned &size, unsigned Align) const {
1876 unsigned reg = State->AllocateReg(GPRArgRegs, 4);
1877 assert((State->getCallOrPrologue() == Prologue ||
1878 State->getCallOrPrologue() == Call) &&
1879 "unhandled ParmContext");
1881 if ((ARM::R0 <= reg) && (reg <= ARM::R3)) {
1882 if (Subtarget->isAAPCS_ABI() && Align > 4) {
1883 unsigned AlignInRegs = Align / 4;
1884 unsigned Waste = (ARM::R4 - reg) % AlignInRegs;
1885 for (unsigned i = 0; i < Waste; ++i)
1886 reg = State->AllocateReg(GPRArgRegs, 4);
1889 unsigned excess = 4 * (ARM::R4 - reg);
1891 // Special case when NSAA != SP and parameter size greater than size of
1892 // all remained GPR regs. In that case we can't split parameter, we must
1893 // send it to stack. We also must set NCRN to R4, so waste all
1894 // remained registers.
1895 const unsigned NSAAOffset = State->getNextStackOffset();
1896 if (Subtarget->isAAPCS_ABI() && NSAAOffset != 0 && size > excess) {
1897 while (State->AllocateReg(GPRArgRegs, 4))
1902 // First register for byval parameter is the first register that wasn't
1903 // allocated before this method call, so it would be "reg".
1904 // If parameter is small enough to be saved in range [reg, r4), then
1905 // the end (first after last) register would be reg + param-size-in-regs,
1906 // else parameter would be splitted between registers and stack,
1907 // end register would be r4 in this case.
1908 unsigned ByValRegBegin = reg;
1909 unsigned ByValRegEnd = (size < excess) ? reg + size/4 : (unsigned)ARM::R4;
1910 State->addInRegsParamInfo(ByValRegBegin, ByValRegEnd);
1911 // Note, first register is allocated in the beginning of function already,
1912 // allocate remained amount of registers we need.
1913 for (unsigned i = reg+1; i != ByValRegEnd; ++i)
1914 State->AllocateReg(GPRArgRegs, 4);
1915 // A byval parameter that is split between registers and memory needs its
1916 // size truncated here.
1917 // In the case where the entire structure fits in registers, we set the
1918 // size in memory to zero.
1927 /// MatchingStackOffset - Return true if the given stack call argument is
1928 /// already available in the same position (relatively) of the caller's
1929 /// incoming argument stack.
1931 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1932 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1933 const TargetInstrInfo *TII) {
1934 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1936 if (Arg.getOpcode() == ISD::CopyFromReg) {
1937 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1938 if (!TargetRegisterInfo::isVirtualRegister(VR))
1940 MachineInstr *Def = MRI->getVRegDef(VR);
1943 if (!Flags.isByVal()) {
1944 if (!TII->isLoadFromStackSlot(Def, FI))
1949 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
1950 if (Flags.isByVal())
1951 // ByVal argument is passed in as a pointer but it's now being
1952 // dereferenced. e.g.
1953 // define @foo(%struct.X* %A) {
1954 // tail call @bar(%struct.X* byval %A)
1957 SDValue Ptr = Ld->getBasePtr();
1958 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
1961 FI = FINode->getIndex();
1965 assert(FI != INT_MAX);
1966 if (!MFI->isFixedObjectIndex(FI))
1968 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
1971 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
1972 /// for tail call optimization. Targets which want to do tail call
1973 /// optimization should implement this function.
1975 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1976 CallingConv::ID CalleeCC,
1978 bool isCalleeStructRet,
1979 bool isCallerStructRet,
1980 const SmallVectorImpl<ISD::OutputArg> &Outs,
1981 const SmallVectorImpl<SDValue> &OutVals,
1982 const SmallVectorImpl<ISD::InputArg> &Ins,
1983 SelectionDAG& DAG) const {
1984 const Function *CallerF = DAG.getMachineFunction().getFunction();
1985 CallingConv::ID CallerCC = CallerF->getCallingConv();
1986 bool CCMatch = CallerCC == CalleeCC;
1988 // Look for obvious safe cases to perform tail call optimization that do not
1989 // require ABI changes. This is what gcc calls sibcall.
1991 // Do not sibcall optimize vararg calls unless the call site is not passing
1993 if (isVarArg && !Outs.empty())
1996 // Exception-handling functions need a special set of instructions to indicate
1997 // a return to the hardware. Tail-calling another function would probably
1999 if (CallerF->hasFnAttribute("interrupt"))
2002 // Also avoid sibcall optimization if either caller or callee uses struct
2003 // return semantics.
2004 if (isCalleeStructRet || isCallerStructRet)
2007 // FIXME: Completely disable sibcall for Thumb1 since Thumb1RegisterInfo::
2008 // emitEpilogue is not ready for them. Thumb tail calls also use t2B, as
2009 // the Thumb1 16-bit unconditional branch doesn't have sufficient relocation
2010 // support in the assembler and linker to be used. This would need to be
2011 // fixed to fully support tail calls in Thumb1.
2013 // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take
2014 // LR. This means if we need to reload LR, it takes an extra instructions,
2015 // which outweighs the value of the tail call; but here we don't know yet
2016 // whether LR is going to be used. Probably the right approach is to
2017 // generate the tail call here and turn it back into CALL/RET in
2018 // emitEpilogue if LR is used.
2020 // Thumb1 PIC calls to external symbols use BX, so they can be tail calls,
2021 // but we need to make sure there are enough registers; the only valid
2022 // registers are the 4 used for parameters. We don't currently do this
2024 if (Subtarget->isThumb1Only())
2027 // Externally-defined functions with weak linkage should not be
2028 // tail-called on ARM when the OS does not support dynamic
2029 // pre-emption of symbols, as the AAELF spec requires normal calls
2030 // to undefined weak functions to be replaced with a NOP or jump to the
2031 // next instruction. The behaviour of branch instructions in this
2032 // situation (as used for tail calls) is implementation-defined, so we
2033 // cannot rely on the linker replacing the tail call with a return.
2034 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2035 const GlobalValue *GV = G->getGlobal();
2036 if (GV->hasExternalWeakLinkage())
2040 // If the calling conventions do not match, then we'd better make sure the
2041 // results are returned in the same way as what the caller expects.
2043 SmallVector<CCValAssign, 16> RVLocs1;
2044 ARMCCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(), RVLocs1,
2045 *DAG.getContext(), Call);
2046 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
2048 SmallVector<CCValAssign, 16> RVLocs2;
2049 ARMCCState CCInfo2(CallerCC, false, DAG.getMachineFunction(), RVLocs2,
2050 *DAG.getContext(), Call);
2051 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
2053 if (RVLocs1.size() != RVLocs2.size())
2055 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
2056 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
2058 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
2060 if (RVLocs1[i].isRegLoc()) {
2061 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
2064 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
2070 // If Caller's vararg or byval argument has been split between registers and
2071 // stack, do not perform tail call, since part of the argument is in caller's
2073 const ARMFunctionInfo *AFI_Caller = DAG.getMachineFunction().
2074 getInfo<ARMFunctionInfo>();
2075 if (AFI_Caller->getArgRegsSaveSize())
2078 // If the callee takes no arguments then go on to check the results of the
2080 if (!Outs.empty()) {
2081 // Check if stack adjustment is needed. For now, do not do this if any
2082 // argument is passed on the stack.
2083 SmallVector<CCValAssign, 16> ArgLocs;
2084 ARMCCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
2085 *DAG.getContext(), Call);
2086 CCInfo.AnalyzeCallOperands(Outs,
2087 CCAssignFnForNode(CalleeCC, false, isVarArg));
2088 if (CCInfo.getNextStackOffset()) {
2089 MachineFunction &MF = DAG.getMachineFunction();
2091 // Check if the arguments are already laid out in the right way as
2092 // the caller's fixed stack objects.
2093 MachineFrameInfo *MFI = MF.getFrameInfo();
2094 const MachineRegisterInfo *MRI = &MF.getRegInfo();
2095 const TargetInstrInfo *TII =
2096 getTargetMachine().getSubtargetImpl()->getInstrInfo();
2097 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
2099 ++i, ++realArgIdx) {
2100 CCValAssign &VA = ArgLocs[i];
2101 EVT RegVT = VA.getLocVT();
2102 SDValue Arg = OutVals[realArgIdx];
2103 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
2104 if (VA.getLocInfo() == CCValAssign::Indirect)
2106 if (VA.needsCustom()) {
2107 // f64 and vector types are split into multiple registers or
2108 // register/stack-slot combinations. The types will not match
2109 // the registers; give up on memory f64 refs until we figure
2110 // out what to do about this.
2113 if (!ArgLocs[++i].isRegLoc())
2115 if (RegVT == MVT::v2f64) {
2116 if (!ArgLocs[++i].isRegLoc())
2118 if (!ArgLocs[++i].isRegLoc())
2121 } else if (!VA.isRegLoc()) {
2122 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
2134 ARMTargetLowering::CanLowerReturn(CallingConv::ID CallConv,
2135 MachineFunction &MF, bool isVarArg,
2136 const SmallVectorImpl<ISD::OutputArg> &Outs,
2137 LLVMContext &Context) const {
2138 SmallVector<CCValAssign, 16> RVLocs;
2139 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, Context);
2140 return CCInfo.CheckReturn(Outs, CCAssignFnForNode(CallConv, /*Return=*/true,
2144 static SDValue LowerInterruptReturn(SmallVectorImpl<SDValue> &RetOps,
2145 SDLoc DL, SelectionDAG &DAG) {
2146 const MachineFunction &MF = DAG.getMachineFunction();
2147 const Function *F = MF.getFunction();
2149 StringRef IntKind = F->getFnAttribute("interrupt").getValueAsString();
2151 // See ARM ARM v7 B1.8.3. On exception entry LR is set to a possibly offset
2152 // version of the "preferred return address". These offsets affect the return
2153 // instruction if this is a return from PL1 without hypervisor extensions.
2154 // IRQ/FIQ: +4 "subs pc, lr, #4"
2155 // SWI: 0 "subs pc, lr, #0"
2156 // ABORT: +4 "subs pc, lr, #4"
2157 // UNDEF: +4/+2 "subs pc, lr, #0"
2158 // UNDEF varies depending on where the exception came from ARM or Thumb
2159 // mode. Alongside GCC, we throw our hands up in disgust and pretend it's 0.
2162 if (IntKind == "" || IntKind == "IRQ" || IntKind == "FIQ" ||
2165 else if (IntKind == "SWI" || IntKind == "UNDEF")
2168 report_fatal_error("Unsupported interrupt attribute. If present, value "
2169 "must be one of: IRQ, FIQ, SWI, ABORT or UNDEF");
2171 RetOps.insert(RetOps.begin() + 1, DAG.getConstant(LROffset, MVT::i32, false));
2173 return DAG.getNode(ARMISD::INTRET_FLAG, DL, MVT::Other, RetOps);
2177 ARMTargetLowering::LowerReturn(SDValue Chain,
2178 CallingConv::ID CallConv, bool isVarArg,
2179 const SmallVectorImpl<ISD::OutputArg> &Outs,
2180 const SmallVectorImpl<SDValue> &OutVals,
2181 SDLoc dl, SelectionDAG &DAG) const {
2183 // CCValAssign - represent the assignment of the return value to a location.
2184 SmallVector<CCValAssign, 16> RVLocs;
2186 // CCState - Info about the registers and stack slots.
2187 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
2188 *DAG.getContext(), Call);
2190 // Analyze outgoing return values.
2191 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
2195 SmallVector<SDValue, 4> RetOps;
2196 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
2197 bool isLittleEndian = Subtarget->isLittle();
2199 MachineFunction &MF = DAG.getMachineFunction();
2200 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2201 AFI->setReturnRegsCount(RVLocs.size());
2203 // Copy the result values into the output registers.
2204 for (unsigned i = 0, realRVLocIdx = 0;
2206 ++i, ++realRVLocIdx) {
2207 CCValAssign &VA = RVLocs[i];
2208 assert(VA.isRegLoc() && "Can only return in registers!");
2210 SDValue Arg = OutVals[realRVLocIdx];
2212 switch (VA.getLocInfo()) {
2213 default: llvm_unreachable("Unknown loc info!");
2214 case CCValAssign::Full: break;
2215 case CCValAssign::BCvt:
2216 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
2220 if (VA.needsCustom()) {
2221 if (VA.getLocVT() == MVT::v2f64) {
2222 // Extract the first half and return it in two registers.
2223 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
2224 DAG.getConstant(0, MVT::i32));
2225 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
2226 DAG.getVTList(MVT::i32, MVT::i32), Half);
2228 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2229 HalfGPRs.getValue(isLittleEndian ? 0 : 1),
2231 Flag = Chain.getValue(1);
2232 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2233 VA = RVLocs[++i]; // skip ahead to next loc
2234 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2235 HalfGPRs.getValue(isLittleEndian ? 1 : 0),
2237 Flag = Chain.getValue(1);
2238 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2239 VA = RVLocs[++i]; // skip ahead to next loc
2241 // Extract the 2nd half and fall through to handle it as an f64 value.
2242 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
2243 DAG.getConstant(1, MVT::i32));
2245 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
2247 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
2248 DAG.getVTList(MVT::i32, MVT::i32), Arg);
2249 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2250 fmrrd.getValue(isLittleEndian ? 0 : 1),
2252 Flag = Chain.getValue(1);
2253 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2254 VA = RVLocs[++i]; // skip ahead to next loc
2255 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2256 fmrrd.getValue(isLittleEndian ? 1 : 0),
2259 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
2261 // Guarantee that all emitted copies are
2262 // stuck together, avoiding something bad.
2263 Flag = Chain.getValue(1);
2264 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2267 // Update chain and glue.
2270 RetOps.push_back(Flag);
2272 // CPUs which aren't M-class use a special sequence to return from
2273 // exceptions (roughly, any instruction setting pc and cpsr simultaneously,
2274 // though we use "subs pc, lr, #N").
2276 // M-class CPUs actually use a normal return sequence with a special
2277 // (hardware-provided) value in LR, so the normal code path works.
2278 if (DAG.getMachineFunction().getFunction()->hasFnAttribute("interrupt") &&
2279 !Subtarget->isMClass()) {
2280 if (Subtarget->isThumb1Only())
2281 report_fatal_error("interrupt attribute is not supported in Thumb1");
2282 return LowerInterruptReturn(RetOps, dl, DAG);
2285 return DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, RetOps);
2288 bool ARMTargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
2289 if (N->getNumValues() != 1)
2291 if (!N->hasNUsesOfValue(1, 0))
2294 SDValue TCChain = Chain;
2295 SDNode *Copy = *N->use_begin();
2296 if (Copy->getOpcode() == ISD::CopyToReg) {
2297 // If the copy has a glue operand, we conservatively assume it isn't safe to
2298 // perform a tail call.
2299 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
2301 TCChain = Copy->getOperand(0);
2302 } else if (Copy->getOpcode() == ARMISD::VMOVRRD) {
2303 SDNode *VMov = Copy;
2304 // f64 returned in a pair of GPRs.
2305 SmallPtrSet<SDNode*, 2> Copies;
2306 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
2308 if (UI->getOpcode() != ISD::CopyToReg)
2312 if (Copies.size() > 2)
2315 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
2317 SDValue UseChain = UI->getOperand(0);
2318 if (Copies.count(UseChain.getNode()))
2322 // We are at the top of this chain.
2323 // If the copy has a glue operand, we conservatively assume it
2324 // isn't safe to perform a tail call.
2325 if (UI->getOperand(UI->getNumOperands()-1).getValueType() == MVT::Glue)
2331 } else if (Copy->getOpcode() == ISD::BITCAST) {
2332 // f32 returned in a single GPR.
2333 if (!Copy->hasOneUse())
2335 Copy = *Copy->use_begin();
2336 if (Copy->getOpcode() != ISD::CopyToReg || !Copy->hasNUsesOfValue(1, 0))
2338 // If the copy has a glue operand, we conservatively assume it isn't safe to
2339 // perform a tail call.
2340 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
2342 TCChain = Copy->getOperand(0);
2347 bool HasRet = false;
2348 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
2350 if (UI->getOpcode() != ARMISD::RET_FLAG &&
2351 UI->getOpcode() != ARMISD::INTRET_FLAG)
2363 bool ARMTargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
2364 if (!Subtarget->supportsTailCall())
2367 if (!CI->isTailCall() || getTargetMachine().Options.DisableTailCalls)
2370 return !Subtarget->isThumb1Only();
2373 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
2374 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
2375 // one of the above mentioned nodes. It has to be wrapped because otherwise
2376 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
2377 // be used to form addressing mode. These wrapped nodes will be selected
2379 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
2380 EVT PtrVT = Op.getValueType();
2381 // FIXME there is no actual debug info here
2383 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
2385 if (CP->isMachineConstantPoolEntry())
2386 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
2387 CP->getAlignment());
2389 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
2390 CP->getAlignment());
2391 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
2394 unsigned ARMTargetLowering::getJumpTableEncoding() const {
2395 return MachineJumpTableInfo::EK_Inline;
2398 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
2399 SelectionDAG &DAG) const {
2400 MachineFunction &MF = DAG.getMachineFunction();
2401 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2402 unsigned ARMPCLabelIndex = 0;
2404 EVT PtrVT = getPointerTy();
2405 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
2406 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2408 if (RelocM == Reloc::Static) {
2409 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
2411 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2412 ARMPCLabelIndex = AFI->createPICLabelUId();
2413 ARMConstantPoolValue *CPV =
2414 ARMConstantPoolConstant::Create(BA, ARMPCLabelIndex,
2415 ARMCP::CPBlockAddress, PCAdj);
2416 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2418 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
2419 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
2420 MachinePointerInfo::getConstantPool(),
2421 false, false, false, 0);
2422 if (RelocM == Reloc::Static)
2424 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2425 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
2428 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
2430 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
2431 SelectionDAG &DAG) const {
2433 EVT PtrVT = getPointerTy();
2434 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2435 MachineFunction &MF = DAG.getMachineFunction();
2436 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2437 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2438 ARMConstantPoolValue *CPV =
2439 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2440 ARMCP::CPValue, PCAdj, ARMCP::TLSGD, true);
2441 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2442 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
2443 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
2444 MachinePointerInfo::getConstantPool(),
2445 false, false, false, 0);
2446 SDValue Chain = Argument.getValue(1);
2448 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2449 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
2451 // call __tls_get_addr.
2454 Entry.Node = Argument;
2455 Entry.Ty = (Type *) Type::getInt32Ty(*DAG.getContext());
2456 Args.push_back(Entry);
2458 // FIXME: is there useful debug info available here?
2459 TargetLowering::CallLoweringInfo CLI(DAG);
2460 CLI.setDebugLoc(dl).setChain(Chain)
2461 .setCallee(CallingConv::C, Type::getInt32Ty(*DAG.getContext()),
2462 DAG.getExternalSymbol("__tls_get_addr", PtrVT), std::move(Args),
2465 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
2466 return CallResult.first;
2469 // Lower ISD::GlobalTLSAddress using the "initial exec" or
2470 // "local exec" model.
2472 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
2474 TLSModel::Model model) const {
2475 const GlobalValue *GV = GA->getGlobal();
2478 SDValue Chain = DAG.getEntryNode();
2479 EVT PtrVT = getPointerTy();
2480 // Get the Thread Pointer
2481 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2483 if (model == TLSModel::InitialExec) {
2484 MachineFunction &MF = DAG.getMachineFunction();
2485 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2486 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2487 // Initial exec model.
2488 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2489 ARMConstantPoolValue *CPV =
2490 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2491 ARMCP::CPValue, PCAdj, ARMCP::GOTTPOFF,
2493 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2494 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2495 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2496 MachinePointerInfo::getConstantPool(),
2497 false, false, false, 0);
2498 Chain = Offset.getValue(1);
2500 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2501 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
2503 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2504 MachinePointerInfo::getConstantPool(),
2505 false, false, false, 0);
2508 assert(model == TLSModel::LocalExec);
2509 ARMConstantPoolValue *CPV =
2510 ARMConstantPoolConstant::Create(GV, ARMCP::TPOFF);
2511 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2512 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2513 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2514 MachinePointerInfo::getConstantPool(),
2515 false, false, false, 0);
2518 // The address of the thread local variable is the add of the thread
2519 // pointer with the offset of the variable.
2520 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
2524 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
2525 // TODO: implement the "local dynamic" model
2526 assert(Subtarget->isTargetELF() &&
2527 "TLS not implemented for non-ELF targets");
2528 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
2530 TLSModel::Model model = getTargetMachine().getTLSModel(GA->getGlobal());
2533 case TLSModel::GeneralDynamic:
2534 case TLSModel::LocalDynamic:
2535 return LowerToTLSGeneralDynamicModel(GA, DAG);
2536 case TLSModel::InitialExec:
2537 case TLSModel::LocalExec:
2538 return LowerToTLSExecModels(GA, DAG, model);
2540 llvm_unreachable("bogus TLS model");
2543 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
2544 SelectionDAG &DAG) const {
2545 EVT PtrVT = getPointerTy();
2547 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2548 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2549 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
2550 ARMConstantPoolValue *CPV =
2551 ARMConstantPoolConstant::Create(GV,
2552 UseGOTOFF ? ARMCP::GOTOFF : ARMCP::GOT);
2553 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2554 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2555 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
2557 MachinePointerInfo::getConstantPool(),
2558 false, false, false, 0);
2559 SDValue Chain = Result.getValue(1);
2560 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
2561 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
2563 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
2564 MachinePointerInfo::getGOT(),
2565 false, false, false, 0);
2569 // If we have T2 ops, we can materialize the address directly via movt/movw
2570 // pair. This is always cheaper.
2571 if (Subtarget->useMovt(DAG.getMachineFunction())) {
2573 // FIXME: Once remat is capable of dealing with instructions with register
2574 // operands, expand this into two nodes.
2575 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2576 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2578 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2579 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2580 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2581 MachinePointerInfo::getConstantPool(),
2582 false, false, false, 0);
2586 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
2587 SelectionDAG &DAG) const {
2588 EVT PtrVT = getPointerTy();
2590 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2591 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2593 if (Subtarget->useMovt(DAG.getMachineFunction()))
2596 // FIXME: Once remat is capable of dealing with instructions with register
2597 // operands, expand this into multiple nodes
2599 RelocM == Reloc::PIC_ ? ARMISD::WrapperPIC : ARMISD::Wrapper;
2601 SDValue G = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, ARMII::MO_NONLAZY);
2602 SDValue Result = DAG.getNode(Wrapper, dl, PtrVT, G);
2604 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2605 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
2606 MachinePointerInfo::getGOT(), false, false, false, 0);
2610 SDValue ARMTargetLowering::LowerGlobalAddressWindows(SDValue Op,
2611 SelectionDAG &DAG) const {
2612 assert(Subtarget->isTargetWindows() && "non-Windows COFF is not supported");
2613 assert(Subtarget->useMovt(DAG.getMachineFunction()) &&
2614 "Windows on ARM expects to use movw/movt");
2616 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2617 const ARMII::TOF TargetFlags =
2618 (GV->hasDLLImportStorageClass() ? ARMII::MO_DLLIMPORT : ARMII::MO_NO_FLAG);
2619 EVT PtrVT = getPointerTy();
2625 // FIXME: Once remat is capable of dealing with instructions with register
2626 // operands, expand this into two nodes.
2627 Result = DAG.getNode(ARMISD::Wrapper, DL, PtrVT,
2628 DAG.getTargetGlobalAddress(GV, DL, PtrVT, /*Offset=*/0,
2630 if (GV->hasDLLImportStorageClass())
2631 Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), Result,
2632 MachinePointerInfo::getGOT(), false, false, false, 0);
2636 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
2637 SelectionDAG &DAG) const {
2638 assert(Subtarget->isTargetELF() &&
2639 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
2640 MachineFunction &MF = DAG.getMachineFunction();
2641 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2642 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2643 EVT PtrVT = getPointerTy();
2645 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2646 ARMConstantPoolValue *CPV =
2647 ARMConstantPoolSymbol::Create(*DAG.getContext(), "_GLOBAL_OFFSET_TABLE_",
2648 ARMPCLabelIndex, PCAdj);
2649 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2650 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2651 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2652 MachinePointerInfo::getConstantPool(),
2653 false, false, false, 0);
2654 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2655 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2659 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
2661 SDValue Val = DAG.getConstant(0, MVT::i32);
2662 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl,
2663 DAG.getVTList(MVT::i32, MVT::Other), Op.getOperand(0),
2664 Op.getOperand(1), Val);
2668 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
2670 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
2671 Op.getOperand(1), DAG.getConstant(0, MVT::i32));
2675 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
2676 const ARMSubtarget *Subtarget) const {
2677 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2680 default: return SDValue(); // Don't custom lower most intrinsics.
2681 case Intrinsic::arm_rbit: {
2682 assert(Op.getOperand(1).getValueType() == MVT::i32 &&
2683 "RBIT intrinsic must have i32 type!");
2684 return DAG.getNode(ARMISD::RBIT, dl, MVT::i32, Op.getOperand(1));
2686 case Intrinsic::arm_thread_pointer: {
2687 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2688 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2690 case Intrinsic::eh_sjlj_lsda: {
2691 MachineFunction &MF = DAG.getMachineFunction();
2692 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2693 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2694 EVT PtrVT = getPointerTy();
2695 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2697 unsigned PCAdj = (RelocM != Reloc::PIC_)
2698 ? 0 : (Subtarget->isThumb() ? 4 : 8);
2699 ARMConstantPoolValue *CPV =
2700 ARMConstantPoolConstant::Create(MF.getFunction(), ARMPCLabelIndex,
2701 ARMCP::CPLSDA, PCAdj);
2702 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2703 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2705 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2706 MachinePointerInfo::getConstantPool(),
2707 false, false, false, 0);
2709 if (RelocM == Reloc::PIC_) {
2710 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2711 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2715 case Intrinsic::arm_neon_vmulls:
2716 case Intrinsic::arm_neon_vmullu: {
2717 unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vmulls)
2718 ? ARMISD::VMULLs : ARMISD::VMULLu;
2719 return DAG.getNode(NewOpc, SDLoc(Op), Op.getValueType(),
2720 Op.getOperand(1), Op.getOperand(2));
2725 static SDValue LowerATOMIC_FENCE(SDValue Op, SelectionDAG &DAG,
2726 const ARMSubtarget *Subtarget) {
2727 // FIXME: handle "fence singlethread" more efficiently.
2729 if (!Subtarget->hasDataBarrier()) {
2730 // Some ARMv6 cpus can support data barriers with an mcr instruction.
2731 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2733 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2734 "Unexpected ISD::ATOMIC_FENCE encountered. Should be libcall!");
2735 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2736 DAG.getConstant(0, MVT::i32));
2739 ConstantSDNode *OrdN = cast<ConstantSDNode>(Op.getOperand(1));
2740 AtomicOrdering Ord = static_cast<AtomicOrdering>(OrdN->getZExtValue());
2741 ARM_MB::MemBOpt Domain = ARM_MB::ISH;
2742 if (Subtarget->isMClass()) {
2743 // Only a full system barrier exists in the M-class architectures.
2744 Domain = ARM_MB::SY;
2745 } else if (Subtarget->isSwift() && Ord == Release) {
2746 // Swift happens to implement ISHST barriers in a way that's compatible with
2747 // Release semantics but weaker than ISH so we'd be fools not to use
2748 // it. Beware: other processors probably don't!
2749 Domain = ARM_MB::ISHST;
2752 return DAG.getNode(ISD::INTRINSIC_VOID, dl, MVT::Other, Op.getOperand(0),
2753 DAG.getConstant(Intrinsic::arm_dmb, MVT::i32),
2754 DAG.getConstant(Domain, MVT::i32));
2757 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG,
2758 const ARMSubtarget *Subtarget) {
2759 // ARM pre v5TE and Thumb1 does not have preload instructions.
2760 if (!(Subtarget->isThumb2() ||
2761 (!Subtarget->isThumb1Only() && Subtarget->hasV5TEOps())))
2762 // Just preserve the chain.
2763 return Op.getOperand(0);
2766 unsigned isRead = ~cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() & 1;
2768 (!Subtarget->hasV7Ops() || !Subtarget->hasMPExtension()))
2769 // ARMv7 with MP extension has PLDW.
2770 return Op.getOperand(0);
2772 unsigned isData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
2773 if (Subtarget->isThumb()) {
2775 isRead = ~isRead & 1;
2776 isData = ~isData & 1;
2779 return DAG.getNode(ARMISD::PRELOAD, dl, MVT::Other, Op.getOperand(0),
2780 Op.getOperand(1), DAG.getConstant(isRead, MVT::i32),
2781 DAG.getConstant(isData, MVT::i32));
2784 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
2785 MachineFunction &MF = DAG.getMachineFunction();
2786 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
2788 // vastart just stores the address of the VarArgsFrameIndex slot into the
2789 // memory location argument.
2791 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2792 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2793 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
2794 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
2795 MachinePointerInfo(SV), false, false, 0);
2799 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
2800 SDValue &Root, SelectionDAG &DAG,
2802 MachineFunction &MF = DAG.getMachineFunction();
2803 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2805 const TargetRegisterClass *RC;
2806 if (AFI->isThumb1OnlyFunction())
2807 RC = &ARM::tGPRRegClass;
2809 RC = &ARM::GPRRegClass;
2811 // Transform the arguments stored in physical registers into virtual ones.
2812 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2813 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2816 if (NextVA.isMemLoc()) {
2817 MachineFrameInfo *MFI = MF.getFrameInfo();
2818 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true);
2820 // Create load node to retrieve arguments from the stack.
2821 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2822 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
2823 MachinePointerInfo::getFixedStack(FI),
2824 false, false, false, 0);
2826 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
2827 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2829 if (!Subtarget->isLittle())
2830 std::swap (ArgValue, ArgValue2);
2831 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
2835 ARMTargetLowering::computeRegArea(CCState &CCInfo, MachineFunction &MF,
2836 unsigned InRegsParamRecordIdx,
2838 unsigned &ArgRegsSize,
2839 unsigned &ArgRegsSaveSize)
2842 if (InRegsParamRecordIdx < CCInfo.getInRegsParamsCount()) {
2843 unsigned RBegin, REnd;
2844 CCInfo.getInRegsParamInfo(InRegsParamRecordIdx, RBegin, REnd);
2845 NumGPRs = REnd - RBegin;
2847 unsigned int firstUnalloced;
2848 firstUnalloced = CCInfo.getFirstUnallocated(GPRArgRegs,
2849 sizeof(GPRArgRegs) /
2850 sizeof(GPRArgRegs[0]));
2851 NumGPRs = (firstUnalloced <= 3) ? (4 - firstUnalloced) : 0;
2854 unsigned Align = MF.getTarget()
2856 ->getFrameLowering()
2857 ->getStackAlignment();
2858 ArgRegsSize = NumGPRs * 4;
2860 // If parameter is split between stack and GPRs...
2861 if (NumGPRs && Align > 4 &&
2862 (ArgRegsSize < ArgSize ||
2863 InRegsParamRecordIdx >= CCInfo.getInRegsParamsCount())) {
2864 // Add padding for part of param recovered from GPRs. For example,
2865 // if Align == 8, its last byte must be at address K*8 - 1.
2866 // We need to do it, since remained (stack) part of parameter has
2867 // stack alignment, and we need to "attach" "GPRs head" without gaps
2870 // |---- 8 bytes block ----| |---- 8 bytes block ----| |---- 8 bytes...
2871 // [ [padding] [GPRs head] ] [ Tail passed via stack ....
2873 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2875 OffsetToAlignment(ArgRegsSize + AFI->getArgRegsSaveSize(), Align);
2876 ArgRegsSaveSize = ArgRegsSize + Padding;
2878 // We don't need to extend regs save size for byval parameters if they
2879 // are passed via GPRs only.
2880 ArgRegsSaveSize = ArgRegsSize;
2883 // The remaining GPRs hold either the beginning of variable-argument
2884 // data, or the beginning of an aggregate passed by value (usually
2885 // byval). Either way, we allocate stack slots adjacent to the data
2886 // provided by our caller, and store the unallocated registers there.
2887 // If this is a variadic function, the va_list pointer will begin with
2888 // these values; otherwise, this reassembles a (byval) structure that
2889 // was split between registers and memory.
2890 // Return: The frame index registers were stored into.
2892 ARMTargetLowering::StoreByValRegs(CCState &CCInfo, SelectionDAG &DAG,
2893 SDLoc dl, SDValue &Chain,
2894 const Value *OrigArg,
2895 unsigned InRegsParamRecordIdx,
2896 unsigned OffsetFromOrigArg,
2900 unsigned ByValStoreOffset,
2901 unsigned TotalArgRegsSaveSize) const {
2903 // Currently, two use-cases possible:
2904 // Case #1. Non-var-args function, and we meet first byval parameter.
2905 // Setup first unallocated register as first byval register;
2906 // eat all remained registers
2907 // (these two actions are performed by HandleByVal method).
2908 // Then, here, we initialize stack frame with
2909 // "store-reg" instructions.
2910 // Case #2. Var-args function, that doesn't contain byval parameters.
2911 // The same: eat all remained unallocated registers,
2912 // initialize stack frame.
2914 MachineFunction &MF = DAG.getMachineFunction();
2915 MachineFrameInfo *MFI = MF.getFrameInfo();
2916 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2917 unsigned firstRegToSaveIndex, lastRegToSaveIndex;
2918 unsigned RBegin, REnd;
2919 if (InRegsParamRecordIdx < CCInfo.getInRegsParamsCount()) {
2920 CCInfo.getInRegsParamInfo(InRegsParamRecordIdx, RBegin, REnd);
2921 firstRegToSaveIndex = RBegin - ARM::R0;
2922 lastRegToSaveIndex = REnd - ARM::R0;
2924 firstRegToSaveIndex = CCInfo.getFirstUnallocated
2925 (GPRArgRegs, array_lengthof(GPRArgRegs));
2926 lastRegToSaveIndex = 4;
2929 unsigned ArgRegsSize, ArgRegsSaveSize;
2930 computeRegArea(CCInfo, MF, InRegsParamRecordIdx, ArgSize,
2931 ArgRegsSize, ArgRegsSaveSize);
2933 // Store any by-val regs to their spots on the stack so that they may be
2934 // loaded by deferencing the result of formal parameter pointer or va_next.
2935 // Note: once stack area for byval/varargs registers
2936 // was initialized, it can't be initialized again.
2937 if (ArgRegsSaveSize) {
2938 unsigned Padding = ArgRegsSaveSize - ArgRegsSize;
2941 assert(AFI->getStoredByValParamsPadding() == 0 &&
2942 "The only parameter may be padded.");
2943 AFI->setStoredByValParamsPadding(Padding);
2946 int FrameIndex = MFI->CreateFixedObject(ArgRegsSaveSize,
2949 (int64_t)TotalArgRegsSaveSize,
2951 SDValue FIN = DAG.getFrameIndex(FrameIndex, getPointerTy());
2953 MFI->CreateFixedObject(Padding,
2954 ArgOffset + ByValStoreOffset -
2955 (int64_t)ArgRegsSaveSize,
2959 SmallVector<SDValue, 4> MemOps;
2960 for (unsigned i = 0; firstRegToSaveIndex < lastRegToSaveIndex;
2961 ++firstRegToSaveIndex, ++i) {
2962 const TargetRegisterClass *RC;
2963 if (AFI->isThumb1OnlyFunction())
2964 RC = &ARM::tGPRRegClass;
2966 RC = &ARM::GPRRegClass;
2968 unsigned VReg = MF.addLiveIn(GPRArgRegs[firstRegToSaveIndex], RC);
2969 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2971 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2972 MachinePointerInfo(OrigArg, OffsetFromOrigArg + 4*i),
2974 MemOps.push_back(Store);
2975 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
2976 DAG.getConstant(4, getPointerTy()));
2979 AFI->setArgRegsSaveSize(ArgRegsSaveSize + AFI->getArgRegsSaveSize());
2981 if (!MemOps.empty())
2982 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOps);
2986 // We cannot allocate a zero-byte object for the first variadic argument,
2987 // so just make up a size.
2990 // This will point to the next argument passed via stack.
2991 return MFI->CreateFixedObject(
2992 ArgSize, ArgOffset, !ForceMutable);
2996 // Setup stack frame, the va_list pointer will start from.
2998 ARMTargetLowering::VarArgStyleRegisters(CCState &CCInfo, SelectionDAG &DAG,
2999 SDLoc dl, SDValue &Chain,
3001 unsigned TotalArgRegsSaveSize,
3002 bool ForceMutable) const {
3003 MachineFunction &MF = DAG.getMachineFunction();
3004 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
3006 // Try to store any remaining integer argument regs
3007 // to their spots on the stack so that they may be loaded by deferencing
3008 // the result of va_next.
3009 // If there is no regs to be stored, just point address after last
3010 // argument passed via stack.
3012 StoreByValRegs(CCInfo, DAG, dl, Chain, nullptr,
3013 CCInfo.getInRegsParamsCount(), 0, ArgOffset, 0, ForceMutable,
3014 0, TotalArgRegsSaveSize);
3016 AFI->setVarArgsFrameIndex(FrameIndex);
3020 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
3021 CallingConv::ID CallConv, bool isVarArg,
3022 const SmallVectorImpl<ISD::InputArg>
3024 SDLoc dl, SelectionDAG &DAG,
3025 SmallVectorImpl<SDValue> &InVals)
3027 MachineFunction &MF = DAG.getMachineFunction();
3028 MachineFrameInfo *MFI = MF.getFrameInfo();
3030 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
3032 // Assign locations to all of the incoming arguments.
3033 SmallVector<CCValAssign, 16> ArgLocs;
3034 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs,
3035 *DAG.getContext(), Prologue);
3036 CCInfo.AnalyzeFormalArguments(Ins,
3037 CCAssignFnForNode(CallConv, /* Return*/ false,
3040 SmallVector<SDValue, 16> ArgValues;
3041 int lastInsIndex = -1;
3043 Function::const_arg_iterator CurOrigArg = MF.getFunction()->arg_begin();
3044 unsigned CurArgIdx = 0;
3046 // Initially ArgRegsSaveSize is zero.
3047 // Then we increase this value each time we meet byval parameter.
3048 // We also increase this value in case of varargs function.
3049 AFI->setArgRegsSaveSize(0);
3051 unsigned ByValStoreOffset = 0;
3052 unsigned TotalArgRegsSaveSize = 0;
3053 unsigned ArgRegsSaveSizeMaxAlign = 4;
3055 // Calculate the amount of stack space that we need to allocate to store
3056 // byval and variadic arguments that are passed in registers.
3057 // We need to know this before we allocate the first byval or variadic
3058 // argument, as they will be allocated a stack slot below the CFA (Canonical
3059 // Frame Address, the stack pointer at entry to the function).
3060 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3061 CCValAssign &VA = ArgLocs[i];
3062 if (VA.isMemLoc()) {
3063 int index = VA.getValNo();
3064 if (index != lastInsIndex) {
3065 ISD::ArgFlagsTy Flags = Ins[index].Flags;
3066 if (Flags.isByVal()) {
3067 unsigned ExtraArgRegsSize;
3068 unsigned ExtraArgRegsSaveSize;
3069 computeRegArea(CCInfo, MF, CCInfo.getInRegsParamsProcessed(),
3070 Flags.getByValSize(),
3071 ExtraArgRegsSize, ExtraArgRegsSaveSize);
3073 TotalArgRegsSaveSize += ExtraArgRegsSaveSize;
3074 if (Flags.getByValAlign() > ArgRegsSaveSizeMaxAlign)
3075 ArgRegsSaveSizeMaxAlign = Flags.getByValAlign();
3076 CCInfo.nextInRegsParam();
3078 lastInsIndex = index;
3082 CCInfo.rewindByValRegsInfo();
3084 if (isVarArg && MFI->hasVAStart()) {
3085 unsigned ExtraArgRegsSize;
3086 unsigned ExtraArgRegsSaveSize;
3087 computeRegArea(CCInfo, MF, CCInfo.getInRegsParamsCount(), 0,
3088 ExtraArgRegsSize, ExtraArgRegsSaveSize);
3089 TotalArgRegsSaveSize += ExtraArgRegsSaveSize;
3091 // If the arg regs save area contains N-byte aligned values, the
3092 // bottom of it must be at least N-byte aligned.
3093 TotalArgRegsSaveSize = RoundUpToAlignment(TotalArgRegsSaveSize, ArgRegsSaveSizeMaxAlign);
3094 TotalArgRegsSaveSize = std::min(TotalArgRegsSaveSize, 16U);
3096 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3097 CCValAssign &VA = ArgLocs[i];
3098 std::advance(CurOrigArg, Ins[VA.getValNo()].OrigArgIndex - CurArgIdx);
3099 CurArgIdx = Ins[VA.getValNo()].OrigArgIndex;
3100 // Arguments stored in registers.
3101 if (VA.isRegLoc()) {
3102 EVT RegVT = VA.getLocVT();
3104 if (VA.needsCustom()) {
3105 // f64 and vector types are split up into multiple registers or
3106 // combinations of registers and stack slots.
3107 if (VA.getLocVT() == MVT::v2f64) {
3108 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
3110 VA = ArgLocs[++i]; // skip ahead to next loc
3112 if (VA.isMemLoc()) {
3113 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true);
3114 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
3115 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
3116 MachinePointerInfo::getFixedStack(FI),
3117 false, false, false, 0);
3119 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
3122 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
3123 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
3124 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
3125 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
3126 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
3128 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
3131 const TargetRegisterClass *RC;
3133 if (RegVT == MVT::f32)
3134 RC = &ARM::SPRRegClass;
3135 else if (RegVT == MVT::f64)
3136 RC = &ARM::DPRRegClass;
3137 else if (RegVT == MVT::v2f64)
3138 RC = &ARM::QPRRegClass;
3139 else if (RegVT == MVT::i32)
3140 RC = AFI->isThumb1OnlyFunction() ?
3141 (const TargetRegisterClass*)&ARM::tGPRRegClass :
3142 (const TargetRegisterClass*)&ARM::GPRRegClass;
3144 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
3146 // Transform the arguments in physical registers into virtual ones.
3147 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
3148 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
3151 // If this is an 8 or 16-bit value, it is really passed promoted
3152 // to 32 bits. Insert an assert[sz]ext to capture this, then
3153 // truncate to the right size.
3154 switch (VA.getLocInfo()) {
3155 default: llvm_unreachable("Unknown loc info!");
3156 case CCValAssign::Full: break;
3157 case CCValAssign::BCvt:
3158 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
3160 case CCValAssign::SExt:
3161 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
3162 DAG.getValueType(VA.getValVT()));
3163 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
3165 case CCValAssign::ZExt:
3166 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
3167 DAG.getValueType(VA.getValVT()));
3168 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
3172 InVals.push_back(ArgValue);
3174 } else { // VA.isRegLoc()
3177 assert(VA.isMemLoc());
3178 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
3180 int index = ArgLocs[i].getValNo();
3182 // Some Ins[] entries become multiple ArgLoc[] entries.
3183 // Process them only once.
3184 if (index != lastInsIndex)
3186 ISD::ArgFlagsTy Flags = Ins[index].Flags;
3187 // FIXME: For now, all byval parameter objects are marked mutable.
3188 // This can be changed with more analysis.
3189 // In case of tail call optimization mark all arguments mutable.
3190 // Since they could be overwritten by lowering of arguments in case of
3192 if (Flags.isByVal()) {
3193 unsigned CurByValIndex = CCInfo.getInRegsParamsProcessed();
3195 ByValStoreOffset = RoundUpToAlignment(ByValStoreOffset, Flags.getByValAlign());
3196 int FrameIndex = StoreByValRegs(
3197 CCInfo, DAG, dl, Chain, CurOrigArg,
3199 Ins[VA.getValNo()].PartOffset,
3200 VA.getLocMemOffset(),
3201 Flags.getByValSize(),
3202 true /*force mutable frames*/,
3204 TotalArgRegsSaveSize);
3205 ByValStoreOffset += Flags.getByValSize();
3206 ByValStoreOffset = std::min(ByValStoreOffset, 16U);
3207 InVals.push_back(DAG.getFrameIndex(FrameIndex, getPointerTy()));
3208 CCInfo.nextInRegsParam();
3210 unsigned FIOffset = VA.getLocMemOffset();
3211 int FI = MFI->CreateFixedObject(VA.getLocVT().getSizeInBits()/8,
3214 // Create load nodes to retrieve arguments from the stack.
3215 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
3216 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
3217 MachinePointerInfo::getFixedStack(FI),
3218 false, false, false, 0));
3220 lastInsIndex = index;
3226 if (isVarArg && MFI->hasVAStart())
3227 VarArgStyleRegisters(CCInfo, DAG, dl, Chain,
3228 CCInfo.getNextStackOffset(),
3229 TotalArgRegsSaveSize);
3231 AFI->setArgumentStackSize(CCInfo.getNextStackOffset());
3236 /// isFloatingPointZero - Return true if this is +0.0.
3237 static bool isFloatingPointZero(SDValue Op) {
3238 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
3239 return CFP->getValueAPF().isPosZero();
3240 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
3241 // Maybe this has already been legalized into the constant pool?
3242 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
3243 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
3244 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
3245 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
3246 return CFP->getValueAPF().isPosZero();
3248 } else if (Op->getOpcode() == ISD::BITCAST &&
3249 Op->getValueType(0) == MVT::f64) {
3250 // Handle (ISD::BITCAST (ARMISD::VMOVIMM (ISD::TargetConstant 0)) MVT::f64)
3251 // created by LowerConstantFP().
3252 SDValue BitcastOp = Op->getOperand(0);
3253 if (BitcastOp->getOpcode() == ARMISD::VMOVIMM) {
3254 SDValue MoveOp = BitcastOp->getOperand(0);
3255 if (MoveOp->getOpcode() == ISD::TargetConstant &&
3256 cast<ConstantSDNode>(MoveOp)->getZExtValue() == 0) {
3264 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
3265 /// the given operands.
3267 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
3268 SDValue &ARMcc, SelectionDAG &DAG,
3270 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
3271 unsigned C = RHSC->getZExtValue();
3272 if (!isLegalICmpImmediate(C)) {
3273 // Constant does not fit, try adjusting it by one?
3278 if (C != 0x80000000 && isLegalICmpImmediate(C-1)) {
3279 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
3280 RHS = DAG.getConstant(C-1, MVT::i32);
3285 if (C != 0 && isLegalICmpImmediate(C-1)) {
3286 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
3287 RHS = DAG.getConstant(C-1, MVT::i32);
3292 if (C != 0x7fffffff && isLegalICmpImmediate(C+1)) {
3293 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;